diff options
Diffstat (limited to 'contrib/llvm/lib/Transforms')
87 files changed, 12485 insertions, 2942 deletions
diff --git a/contrib/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp b/contrib/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp index b94dd69..be48b20 100644 --- a/contrib/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp +++ b/contrib/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp @@ -153,7 +153,8 @@ CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) { SmallPtrSet<Argument*, 8> ArgsToPromote; SmallPtrSet<Argument*, 8> ByValArgsToTransform; for (unsigned i = 0; i != PointerArgs.size(); ++i) { - bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); + bool isByVal=F->getParamAttributes(PointerArgs[i].second+1). + hasAttribute(Attributes::ByVal); Argument *PtrArg = PointerArgs[i].first; Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); @@ -517,8 +518,10 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, const AttrListPtr &PAL = F->getAttributes(); // Add any return attributes. - if (Attributes attrs = PAL.getRetAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); + Attributes attrs = PAL.getRetAttributes(); + if (attrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, + attrs)); // First, determine the new argument list unsigned ArgIndex = 1; @@ -534,7 +537,8 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, } else if (!ArgsToPromote.count(I)) { // Unchanged argument Params.push_back(I->getType()); - if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) + Attributes attrs = PAL.getParamAttributes(ArgIndex); + if (attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); } else if (I->use_empty()) { // Dead argument (which are always marked as promotable) @@ -587,19 +591,13 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, } // Add any function attributes. - if (Attributes attrs = PAL.getFnAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); + attrs = PAL.getFnAttributes(); + if (attrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + attrs)); Type *RetTy = FTy->getReturnType(); - // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which - // have zero fixed arguments. - bool ExtraArgHack = false; - if (Params.empty() && FTy->isVarArg()) { - ExtraArgHack = true; - Params.push_back(Type::getInt32Ty(F->getContext())); - } - // Construct the new function type using the new arguments. FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); @@ -613,7 +611,7 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, // Recompute the parameter attributes list based on the new arguments for // the function. - NF->setAttributes(AttrListPtr::get(AttributesVec)); + NF->setAttributes(AttrListPtr::get(F->getContext(), AttributesVec)); AttributesVec.clear(); F->getParent()->getFunctionList().insert(F, NF); @@ -641,8 +639,10 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, const AttrListPtr &CallPAL = CS.getAttributes(); // Add any return attributes. - if (Attributes attrs = CallPAL.getRetAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); + Attributes attrs = CallPAL.getRetAttributes(); + if (attrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, + attrs)); // Loop over the operands, inserting GEP and loads in the caller as // appropriate. @@ -653,7 +653,8 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { Args.push_back(*AI); // Unmodified argument - if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) + Attributes Attrs = CallPAL.getParamAttributes(ArgIndex); + if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } else if (ByValArgsToTransform.count(I)) { @@ -711,30 +712,32 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, } } - if (ExtraArgHack) - Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext()))); - // Push any varargs arguments on the list. for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { Args.push_back(*AI); - if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) + Attributes Attrs = CallPAL.getParamAttributes(ArgIndex); + if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } // Add any function attributes. - if (Attributes attrs = CallPAL.getFnAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); + attrs = CallPAL.getFnAttributes(); + if (attrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + attrs)); Instruction *New; if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), Args, "", Call); cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); - cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec)); + cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(II->getContext(), + AttributesVec)); } else { New = CallInst::Create(NF, Args, "", Call); cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); - cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec)); + cast<CallInst>(New)->setAttributes(AttrListPtr::get(New->getContext(), + AttributesVec)); if (cast<CallInst>(Call)->isTailCall()) cast<CallInst>(New)->setTailCall(); } @@ -870,16 +873,9 @@ CallGraphNode *ArgPromotion::DoPromotion(Function *F, } // Increment I2 past all of the arguments added for this promoted pointer. - for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) - ++I2; + std::advance(I2, ArgIndices.size()); } - // Notify the alias analysis implementation that we inserted a new argument. - if (ExtraArgHack) - AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())), - NF->arg_begin()); - - // Tell the alias analysis that the old function is about to disappear. AA.replaceWithNewValue(F, NF); diff --git a/contrib/llvm/lib/Transforms/IPO/BarrierNoopPass.cpp b/contrib/llvm/lib/Transforms/IPO/BarrierNoopPass.cpp new file mode 100644 index 0000000..2e32240 --- /dev/null +++ b/contrib/llvm/lib/Transforms/IPO/BarrierNoopPass.cpp @@ -0,0 +1,47 @@ +//===- BarrierNoopPass.cpp - A barrier pass for the pass manager ----------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// NOTE: DO NOT USE THIS IF AVOIDABLE +// +// This pass is a nonce pass intended to allow manipulation of the implicitly +// nesting pass manager. For example, it can be used to cause a CGSCC pass +// manager to be closed prior to running a new collection of function passes. +// +// FIXME: This is a huge HACK. This should be removed when the pass manager's +// nesting is made explicit instead of implicit. +// +//===----------------------------------------------------------------------===// + +#include "llvm/Pass.h" +#include "llvm/Transforms/IPO.h" +using namespace llvm; + +namespace { +/// \brief A nonce module pass used to place a barrier in a pass manager. +/// +/// There is no mechanism for ending a CGSCC pass manager once one is started. +/// This prevents extension points from having clear deterministic ordering +/// when they are phrased as non-module passes. +class BarrierNoop : public ModulePass { +public: + static char ID; // Pass identification. + + BarrierNoop() : ModulePass(ID) { + initializeBarrierNoopPass(*PassRegistry::getPassRegistry()); + } + + bool runOnModule(Module &M) { return false; } +}; +} + +ModulePass *llvm::createBarrierNoopPass() { return new BarrierNoop(); } + +char BarrierNoop::ID = 0; +INITIALIZE_PASS(BarrierNoop, "barrier", "A No-Op Barrier Pass", + false, false) diff --git a/contrib/llvm/lib/Transforms/IPO/ConstantMerge.cpp b/contrib/llvm/lib/Transforms/IPO/ConstantMerge.cpp index d8fae8a..e2f0126 100644 --- a/contrib/llvm/lib/Transforms/IPO/ConstantMerge.cpp +++ b/contrib/llvm/lib/Transforms/IPO/ConstantMerge.cpp @@ -23,7 +23,7 @@ #include "llvm/DerivedTypes.h" #include "llvm/Module.h" #include "llvm/Pass.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/PointerIntPair.h" #include "llvm/ADT/SmallPtrSet.h" @@ -50,7 +50,7 @@ namespace { // alignment to a concrete value. unsigned getAlignment(GlobalVariable *GV) const; - const TargetData *TD; + const DataLayout *TD; }; } @@ -98,7 +98,7 @@ unsigned ConstantMerge::getAlignment(GlobalVariable *GV) const { } bool ConstantMerge::runOnModule(Module &M) { - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); // Find all the globals that are marked "used". These cannot be merged. SmallPtrSet<const GlobalValue*, 8> UsedGlobals; @@ -107,7 +107,7 @@ bool ConstantMerge::runOnModule(Module &M) { // Map unique <constants, has-unknown-alignment> pairs to globals. We don't // want to merge globals of unknown alignment with those of explicit - // alignment. If we have TargetData, we always know the alignment. + // alignment. If we have DataLayout, we always know the alignment. DenseMap<PointerIntPair<Constant*, 1, bool>, GlobalVariable*> CMap; // Replacements - This vector contains a list of replacements to perform. diff --git a/contrib/llvm/lib/Transforms/IPO/DeadArgumentElimination.cpp b/contrib/llvm/lib/Transforms/IPO/DeadArgumentElimination.cpp index fd23a93..4cfd0b2 100644 --- a/contrib/llvm/lib/Transforms/IPO/DeadArgumentElimination.cpp +++ b/contrib/llvm/lib/Transforms/IPO/DeadArgumentElimination.cpp @@ -21,7 +21,9 @@ #include "llvm/Transforms/IPO.h" #include "llvm/CallingConv.h" #include "llvm/Constant.h" +#include "llvm/DebugInfo.h" #include "llvm/DerivedTypes.h" +#include "llvm/DIBuilder.h" #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" #include "llvm/LLVMContext.h" @@ -30,6 +32,7 @@ #include "llvm/Support/CallSite.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" +#include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringExtras.h" @@ -121,6 +124,15 @@ namespace { typedef SmallVector<RetOrArg, 5> UseVector; + // Map each LLVM function to corresponding metadata with debug info. If + // the function is replaced with another one, we should patch the pointer + // to LLVM function in metadata. + // As the code generation for module is finished (and DIBuilder is + // finalized) we assume that subprogram descriptors won't be changed, and + // they are stored in map for short duration anyway. + typedef DenseMap<Function*, DISubprogram> FunctionDIMap; + FunctionDIMap FunctionDIs; + protected: // DAH uses this to specify a different ID. explicit DAE(char &ID) : ModulePass(ID) {} @@ -141,6 +153,7 @@ namespace { unsigned RetValNum = 0); Liveness SurveyUses(const Value *V, UseVector &MaybeLiveUses); + void CollectFunctionDIs(Module &M); void SurveyFunction(const Function &F); void MarkValue(const RetOrArg &RA, Liveness L, const UseVector &MaybeLiveUses); @@ -180,6 +193,33 @@ INITIALIZE_PASS(DAH, "deadarghaX0r", ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } +/// CollectFunctionDIs - Map each function in the module to its debug info +/// descriptor. +void DAE::CollectFunctionDIs(Module &M) { + FunctionDIs.clear(); + + for (Module::named_metadata_iterator I = M.named_metadata_begin(), + E = M.named_metadata_end(); I != E; ++I) { + NamedMDNode &NMD = *I; + for (unsigned MDIndex = 0, MDNum = NMD.getNumOperands(); + MDIndex < MDNum; ++MDIndex) { + MDNode *Node = NMD.getOperand(MDIndex); + if (!DIDescriptor(Node).isCompileUnit()) + continue; + DICompileUnit CU(Node); + const DIArray &SPs = CU.getSubprograms(); + for (unsigned SPIndex = 0, SPNum = SPs.getNumElements(); + SPIndex < SPNum; ++SPIndex) { + DISubprogram SP(SPs.getElement(SPIndex)); + if (!SP.Verify()) + continue; + if (Function *F = SP.getFunction()) + FunctionDIs[F] = SP; + } + } + } +} + /// DeleteDeadVarargs - If this is an function that takes a ... list, and if /// llvm.vastart is never called, the varargs list is dead for the function. bool DAE::DeleteDeadVarargs(Function &Fn) { @@ -236,9 +276,11 @@ bool DAE::DeleteDeadVarargs(Function &Fn) { SmallVector<AttributeWithIndex, 8> AttributesVec; for (unsigned i = 0; PAL.getSlot(i).Index <= NumArgs; ++i) AttributesVec.push_back(PAL.getSlot(i)); - if (Attributes FnAttrs = PAL.getFnAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); - PAL = AttrListPtr::get(AttributesVec); + Attributes FnAttrs = PAL.getFnAttributes(); + if (FnAttrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + FnAttrs)); + PAL = AttrListPtr::get(Fn.getContext(), AttributesVec); } Instruction *New; @@ -284,6 +326,11 @@ bool DAE::DeleteDeadVarargs(Function &Fn) { I2->takeName(I); } + // Patch the pointer to LLVM function in debug info descriptor. + FunctionDIMap::iterator DI = FunctionDIs.find(&Fn); + if (DI != FunctionDIs.end()) + DI->second.replaceFunction(NF); + // Finally, nuke the old function. Fn.eraseFromParent(); return true; @@ -717,13 +764,17 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { // here. Currently, this should not be possible, but special handling might be // required when new return value attributes are added. if (NRetTy->isVoidTy()) - RAttrs &= ~Attribute::typeIncompatible(NRetTy); + RAttrs = + Attributes::get(NRetTy->getContext(), AttrBuilder(RAttrs). + removeAttributes(Attributes::typeIncompatible(NRetTy))); else - assert((RAttrs & Attribute::typeIncompatible(NRetTy)) == 0 - && "Return attributes no longer compatible?"); + assert(!AttrBuilder(RAttrs). + hasAttributes(Attributes::typeIncompatible(NRetTy)) && + "Return attributes no longer compatible?"); - if (RAttrs) - AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs)); + if (RAttrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, + RAttrs)); // Remember which arguments are still alive. SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); @@ -740,7 +791,8 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { // Get the original parameter attributes (skipping the first one, that is // for the return value. - if (Attributes Attrs = PAL.getParamAttributes(i + 1)) + Attributes Attrs = PAL.getParamAttributes(i + 1); + if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Params.size(), Attrs)); } else { ++NumArgumentsEliminated; @@ -749,11 +801,12 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { } } - if (FnAttrs != Attribute::None) - AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); + if (FnAttrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + FnAttrs)); // Reconstruct the AttributesList based on the vector we constructed. - AttrListPtr NewPAL = AttrListPtr::get(AttributesVec); + AttrListPtr NewPAL = AttrListPtr::get(F->getContext(), AttributesVec); // Create the new function type based on the recomputed parameters. FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); @@ -786,9 +839,12 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { Attributes RAttrs = CallPAL.getRetAttributes(); Attributes FnAttrs = CallPAL.getFnAttributes(); // Adjust in case the function was changed to return void. - RAttrs &= ~Attribute::typeIncompatible(NF->getReturnType()); - if (RAttrs) - AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs)); + RAttrs = + Attributes::get(NF->getContext(), AttrBuilder(RAttrs). + removeAttributes(Attributes::typeIncompatible(NF->getReturnType()))); + if (RAttrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, + RAttrs)); // Declare these outside of the loops, so we can reuse them for the second // loop, which loops the varargs. @@ -800,22 +856,25 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { if (ArgAlive[i]) { Args.push_back(*I); // Get original parameter attributes, but skip return attributes. - if (Attributes Attrs = CallPAL.getParamAttributes(i + 1)) + Attributes Attrs = CallPAL.getParamAttributes(i + 1); + if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } // Push any varargs arguments on the list. Don't forget their attributes. for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { Args.push_back(*I); - if (Attributes Attrs = CallPAL.getParamAttributes(i + 1)) + Attributes Attrs = CallPAL.getParamAttributes(i + 1); + if (Attrs.hasAttributes()) AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); } - if (FnAttrs != Attribute::None) - AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); + if (FnAttrs.hasAttributes()) + AttributesVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + FnAttrs)); // Reconstruct the AttributesList based on the vector we constructed. - AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec); + AttrListPtr NewCallPAL = AttrListPtr::get(F->getContext(), AttributesVec); Instruction *New; if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { @@ -952,6 +1011,11 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { BB->getInstList().erase(RI); } + // Patch the pointer to LLVM function in debug info descriptor. + FunctionDIMap::iterator DI = FunctionDIs.find(F); + if (DI != FunctionDIs.end()) + DI->second.replaceFunction(NF); + // Now that the old function is dead, delete it. F->eraseFromParent(); @@ -961,6 +1025,9 @@ bool DAE::RemoveDeadStuffFromFunction(Function *F) { bool DAE::runOnModule(Module &M) { bool Changed = false; + // Collect debug info descriptors for functions. + CollectFunctionDIs(M); + // First pass: Do a simple check to see if any functions can have their "..." // removed. We can do this if they never call va_start. This loop cannot be // fused with the next loop, because deleting a function invalidates diff --git a/contrib/llvm/lib/Transforms/IPO/ExtractGV.cpp b/contrib/llvm/lib/Transforms/IPO/ExtractGV.cpp index 4c7f0ed..6716deb 100644 --- a/contrib/llvm/lib/Transforms/IPO/ExtractGV.cpp +++ b/contrib/llvm/lib/Transforms/IPO/ExtractGV.cpp @@ -51,32 +51,75 @@ namespace { // Visit the GlobalVariables. for (Module::global_iterator I = M.global_begin(), E = M.global_end(); I != E; ++I) { - if (deleteStuff == (bool)Named.count(I) && !I->isDeclaration()) { - I->setInitializer(0); - } else { + bool Delete = + deleteStuff == (bool)Named.count(I) && !I->isDeclaration(); + if (!Delete) { if (I->hasAvailableExternallyLinkage()) continue; if (I->getName() == "llvm.global_ctors") continue; } - if (I->hasLocalLinkage()) + bool Local = I->hasLocalLinkage(); + if (Local) I->setVisibility(GlobalValue::HiddenVisibility); - I->setLinkage(GlobalValue::ExternalLinkage); + + if (Local || Delete) + I->setLinkage(GlobalValue::ExternalLinkage); + + if (Delete) + I->setInitializer(0); } // Visit the Functions. for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { - if (deleteStuff == (bool)Named.count(I) && !I->isDeclaration()) { - I->deleteBody(); - } else { + bool Delete = + deleteStuff == (bool)Named.count(I) && !I->isDeclaration(); + if (!Delete) { if (I->hasAvailableExternallyLinkage()) continue; } - if (I->hasLocalLinkage()) + bool Local = I->hasLocalLinkage(); + if (Local) I->setVisibility(GlobalValue::HiddenVisibility); - I->setLinkage(GlobalValue::ExternalLinkage); + + if (Local || Delete) + I->setLinkage(GlobalValue::ExternalLinkage); + + if (Delete) + I->deleteBody(); + } + + // Visit the Aliases. + for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); + I != E;) { + Module::alias_iterator CurI = I; + ++I; + + if (CurI->hasLocalLinkage()) { + CurI->setVisibility(GlobalValue::HiddenVisibility); + CurI->setLinkage(GlobalValue::ExternalLinkage); + } + + if (deleteStuff == (bool)Named.count(CurI)) { + Type *Ty = CurI->getType()->getElementType(); + + CurI->removeFromParent(); + llvm::Value *Declaration; + if (FunctionType *FTy = dyn_cast<FunctionType>(Ty)) { + Declaration = Function::Create(FTy, GlobalValue::ExternalLinkage, + CurI->getName(), &M); + + } else { + Declaration = + new GlobalVariable(M, Ty, false, GlobalValue::ExternalLinkage, + 0, CurI->getName()); + + } + CurI->replaceAllUsesWith(Declaration); + delete CurI; + } } return true; diff --git a/contrib/llvm/lib/Transforms/IPO/FunctionAttrs.cpp b/contrib/llvm/lib/Transforms/IPO/FunctionAttrs.cpp index f3f6228..18409f7 100644 --- a/contrib/llvm/lib/Transforms/IPO/FunctionAttrs.cpp +++ b/contrib/llvm/lib/Transforms/IPO/FunctionAttrs.cpp @@ -28,9 +28,9 @@ #include "llvm/Analysis/CallGraph.h" #include "llvm/Analysis/CaptureTracking.h" #include "llvm/ADT/SCCIterator.h" +#include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" -#include "llvm/ADT/UniqueVector.h" #include "llvm/Support/InstIterator.h" using namespace llvm; @@ -212,10 +212,17 @@ bool FunctionAttrs::AddReadAttrs(const CallGraphSCC &SCC) { MadeChange = true; // Clear out any existing attributes. - F->removeAttribute(~0, Attribute::ReadOnly | Attribute::ReadNone); + AttrBuilder B; + B.addAttribute(Attributes::ReadOnly) + .addAttribute(Attributes::ReadNone); + F->removeAttribute(AttrListPtr::FunctionIndex, + Attributes::get(F->getContext(), B)); // Add in the new attribute. - F->addAttribute(~0, ReadsMemory? Attribute::ReadOnly : Attribute::ReadNone); + B.clear(); + B.addAttribute(ReadsMemory ? Attributes::ReadOnly : Attributes::ReadNone); + F->addAttribute(AttrListPtr::FunctionIndex, + Attributes::get(F->getContext(), B)); if (ReadsMemory) ++NumReadOnly; @@ -276,8 +283,6 @@ namespace { void tooManyUses() { Captured = true; } - bool shouldExplore(Use *U) { return true; } - bool captured(Use *U) { CallSite CS(U->getUser()); if (!CS.getInstruction()) { Captured = true; return true; } @@ -352,6 +357,9 @@ bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) { ArgumentGraph AG; + AttrBuilder B; + B.addAttribute(Attributes::NoCapture); + // Check each function in turn, determining which pointer arguments are not // captured. for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { @@ -373,7 +381,7 @@ bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) { for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end(); A != E; ++A) { if (A->getType()->isPointerTy() && !A->hasNoCaptureAttr()) { - A->addAttr(Attribute::NoCapture); + A->addAttr(Attributes::get(F->getContext(), B)); ++NumNoCapture; Changed = true; } @@ -388,7 +396,7 @@ bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) { if (!Tracker.Captured) { if (Tracker.Uses.empty()) { // If it's trivially not captured, mark it nocapture now. - A->addAttr(Attribute::NoCapture); + A->addAttr(Attributes::get(F->getContext(), B)); ++NumNoCapture; Changed = true; } else { @@ -421,7 +429,9 @@ bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) { // eg. "void f(int* x) { if (...) f(x); }" if (ArgumentSCC[0]->Uses.size() == 1 && ArgumentSCC[0]->Uses[0] == ArgumentSCC[0]) { - ArgumentSCC[0]->Definition->addAttr(Attribute::NoCapture); + ArgumentSCC[0]-> + Definition-> + addAttr(Attributes::get(ArgumentSCC[0]->Definition->getContext(), B)); ++NumNoCapture; Changed = true; } @@ -463,7 +473,7 @@ bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) { for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) { Argument *A = ArgumentSCC[i]->Definition; - A->addAttr(Attribute::NoCapture); + A->addAttr(Attributes::get(A->getContext(), B)); ++NumNoCapture; Changed = true; } @@ -476,13 +486,13 @@ bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) { /// or a pointer that doesn't alias any other pointer visible to the caller. bool FunctionAttrs::IsFunctionMallocLike(Function *F, SmallPtrSet<Function*, 8> &SCCNodes) const { - UniqueVector<Value *> FlowsToReturn; + SmallSetVector<Value *, 8> FlowsToReturn; for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I) if (ReturnInst *Ret = dyn_cast<ReturnInst>(I->getTerminator())) FlowsToReturn.insert(Ret->getReturnValue()); for (unsigned i = 0; i != FlowsToReturn.size(); ++i) { - Value *RetVal = FlowsToReturn[i+1]; // UniqueVector[0] is reserved. + Value *RetVal = FlowsToReturn[i]; if (Constant *C = dyn_cast<Constant>(RetVal)) { if (!C->isNullValue() && !isa<UndefValue>(C)) @@ -520,7 +530,7 @@ bool FunctionAttrs::IsFunctionMallocLike(Function *F, case Instruction::Call: case Instruction::Invoke: { CallSite CS(RVI); - if (CS.paramHasAttr(0, Attribute::NoAlias)) + if (CS.paramHasAttr(0, Attributes::NoAlias)) break; if (CS.getCalledFunction() && SCCNodes.count(CS.getCalledFunction())) diff --git a/contrib/llvm/lib/Transforms/IPO/GlobalOpt.cpp b/contrib/llvm/lib/Transforms/IPO/GlobalOpt.cpp index 6d950d2..591278f 100644 --- a/contrib/llvm/lib/Transforms/IPO/GlobalOpt.cpp +++ b/contrib/llvm/lib/Transforms/IPO/GlobalOpt.cpp @@ -25,7 +25,7 @@ #include "llvm/Pass.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/MemoryBuiltins.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Support/CallSite.h" #include "llvm/Support/Debug.h" @@ -83,7 +83,7 @@ namespace { const GlobalStatus &GS); bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn); - TargetData *TD; + DataLayout *TD; TargetLibraryInfo *TLI; }; } @@ -225,6 +225,7 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, // Don't hack on volatile stores. if (SI->isVolatile()) return true; + GS.Ordering = StrongerOrdering(GS.Ordering, SI->getOrdering()); // If this is a direct store to the global (i.e., the global is a scalar @@ -234,6 +235,14 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, if (const GlobalVariable *GV = dyn_cast<GlobalVariable>( SI->getOperand(1))) { Value *StoredVal = SI->getOperand(0); + + if (Constant *C = dyn_cast<Constant>(StoredVal)) { + if (C->isThreadDependent()) { + // The stored value changes between threads; don't track it. + return true; + } + } + if (StoredVal == GV->getInitializer()) { if (GS.StoredType < GlobalStatus::isInitializerStored) GS.StoredType = GlobalStatus::isInitializerStored; @@ -346,7 +355,7 @@ static bool isLeakCheckerRoot(GlobalVariable *GV) { /// Given a value that is stored to a global but never read, determine whether /// it's safe to remove the store and the chain of computation that feeds the /// store. -static bool IsSafeComputationToRemove(Value *V) { +static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) { do { if (isa<Constant>(V)) return true; @@ -355,7 +364,7 @@ static bool IsSafeComputationToRemove(Value *V) { if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) || isa<GlobalValue>(V)) return false; - if (isAllocationFn(V)) + if (isAllocationFn(V, TLI)) return true; Instruction *I = cast<Instruction>(V); @@ -376,7 +385,8 @@ static bool IsSafeComputationToRemove(Value *V) { /// of the global and clean up any that obviously don't assign the global a /// value that isn't dynamically allocated. /// -static bool CleanupPointerRootUsers(GlobalVariable *GV) { +static bool CleanupPointerRootUsers(GlobalVariable *GV, + const TargetLibraryInfo *TLI) { // A brief explanation of leak checkers. The goal is to find bugs where // pointers are forgotten, causing an accumulating growth in memory // usage over time. The common strategy for leak checkers is to whitelist the @@ -432,18 +442,18 @@ static bool CleanupPointerRootUsers(GlobalVariable *GV) { C->destroyConstant(); // This could have invalidated UI, start over from scratch. Dead.clear(); - CleanupPointerRootUsers(GV); + CleanupPointerRootUsers(GV, TLI); return true; } } } for (int i = 0, e = Dead.size(); i != e; ++i) { - if (IsSafeComputationToRemove(Dead[i].first)) { + if (IsSafeComputationToRemove(Dead[i].first, TLI)) { Dead[i].second->eraseFromParent(); Instruction *I = Dead[i].first; do { - if (isAllocationFn(I)) + if (isAllocationFn(I, TLI)) break; Instruction *J = dyn_cast<Instruction>(I->getOperand(0)); if (!J) @@ -463,7 +473,7 @@ static bool CleanupPointerRootUsers(GlobalVariable *GV) { /// quick scan over the use list to clean up the easy and obvious cruft. This /// returns true if it made a change. static bool CleanupConstantGlobalUsers(Value *V, Constant *Init, - TargetData *TD, TargetLibraryInfo *TLI) { + DataLayout *TD, TargetLibraryInfo *TLI) { bool Changed = false; for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { User *U = *UI++; @@ -655,7 +665,7 @@ static bool GlobalUsersSafeToSRA(GlobalValue *GV) { /// behavior of the program in a more fine-grained way. We have determined that /// this transformation is safe already. We return the first global variable we /// insert so that the caller can reprocess it. -static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { +static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &TD) { // Make sure this global only has simple uses that we can SRA. if (!GlobalUsersSafeToSRA(GV)) return 0; @@ -931,7 +941,7 @@ static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { /// if the loaded value is dynamically null, then we know that they cannot be /// reachable with a null optimize away the load. static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV, - TargetData *TD, + DataLayout *TD, TargetLibraryInfo *TLI) { bool Changed = false; @@ -961,7 +971,9 @@ static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV, // If we get here we could have other crazy uses that are transitively // loaded. assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) || - isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser)) && + isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) || + isa<BitCastInst>(GlobalUser) || + isa<GetElementPtrInst>(GlobalUser)) && "Only expect load and stores!"); } } @@ -975,7 +987,7 @@ static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV, // nor is the global. if (AllNonStoreUsesGone) { if (isLeakCheckerRoot(GV)) { - Changed |= CleanupPointerRootUsers(GV); + Changed |= CleanupPointerRootUsers(GV, TLI); } else { Changed = true; CleanupConstantGlobalUsers(GV, 0, TD, TLI); @@ -993,7 +1005,7 @@ static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV, /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the /// instructions that are foldable. static void ConstantPropUsersOf(Value *V, - TargetData *TD, TargetLibraryInfo *TLI) { + DataLayout *TD, TargetLibraryInfo *TLI) { for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) if (Instruction *I = dyn_cast<Instruction>(*UI++)) if (Constant *NewC = ConstantFoldInstruction(I, TD, TLI)) { @@ -1016,7 +1028,7 @@ static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy, ConstantInt *NElements, - TargetData *TD, + DataLayout *TD, TargetLibraryInfo *TLI) { DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n'); @@ -1465,9 +1477,10 @@ static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break /// it up into multiple allocations of arrays of the fields. static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, - Value *NElems, TargetData *TD) { + Value *NElems, DataLayout *TD, + const TargetLibraryInfo *TLI) { DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n'); - Type *MAT = getMallocAllocatedType(CI); + Type *MAT = getMallocAllocatedType(CI, TLI); StructType *STy = cast<StructType>(MAT); // There is guaranteed to be at least one use of the malloc (storing @@ -1656,7 +1669,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, Type *AllocTy, AtomicOrdering Ordering, Module::global_iterator &GVI, - TargetData *TD, + DataLayout *TD, TargetLibraryInfo *TLI) { if (!TD) return false; @@ -1688,7 +1701,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // This eliminates dynamic allocation, avoids an indirection accessing the // data, and exposes the resultant global to further GlobalOpt. // We cannot optimize the malloc if we cannot determine malloc array size. - Value *NElems = getMallocArraySize(CI, TD, true); + Value *NElems = getMallocArraySize(CI, TD, TLI, true); if (!NElems) return false; @@ -1725,7 +1738,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // If this is a fixed size array, transform the Malloc to be an alloc of // structs. malloc [100 x struct],1 -> malloc struct, 100 - if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) { + if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) { Type *IntPtrTy = TD->getIntPtrType(CI->getContext()); unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes(); Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize); @@ -1742,7 +1755,8 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CI = cast<CallInst>(Malloc); } - GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true), TD); + GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, TLI, true), + TD, TLI); return true; } @@ -1754,7 +1768,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, AtomicOrdering Ordering, Module::global_iterator &GVI, - TargetData *TD, TargetLibraryInfo *TLI) { + DataLayout *TD, TargetLibraryInfo *TLI) { // Ignore no-op GEPs and bitcasts. StoredOnceVal = StoredOnceVal->stripPointerCasts(); @@ -1771,8 +1785,8 @@ static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, // Optimize away any trapping uses of the loaded value. if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, TD, TLI)) return true; - } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) { - Type *MallocType = getMallocAllocatedType(CI); + } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) { + Type *MallocType = getMallocAllocatedType(CI, TLI); if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI, TD, TLI)) @@ -1964,7 +1978,7 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, bool Changed; if (isLeakCheckerRoot(GV)) { // Delete any constant stores to the global. - Changed = CleanupPointerRootUsers(GV); + Changed = CleanupPointerRootUsers(GV, TLI); } else { // Delete any stores we can find to the global. We may not be able to // make it completely dead though. @@ -1997,7 +2011,7 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, ++NumMarked; return true; } else if (!GV->getInitializer()->getType()->isSingleValueType()) { - if (TargetData *TD = getAnalysisIfAvailable<TargetData>()) + if (DataLayout *TD = getAnalysisIfAvailable<DataLayout>()) if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) { GVI = FirstNewGV; // Don't skip the newly produced globals! return true; @@ -2056,25 +2070,26 @@ static void ChangeCalleesToFastCall(Function *F) { } } -static AttrListPtr StripNest(const AttrListPtr &Attrs) { +static AttrListPtr StripNest(LLVMContext &C, const AttrListPtr &Attrs) { for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { - if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) + if (!Attrs.getSlot(i).Attrs.hasAttribute(Attributes::Nest)) continue; // There can be only one. - return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); + return Attrs.removeAttr(C, Attrs.getSlot(i).Index, + Attributes::get(C, Attributes::Nest)); } return Attrs; } static void RemoveNestAttribute(Function *F) { - F->setAttributes(StripNest(F->getAttributes())); + F->setAttributes(StripNest(F->getContext(), F->getAttributes())); for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ if (isa<BlockAddress>(*UI)) continue; CallSite User(cast<Instruction>(*UI)); - User.setAttributes(StripNest(User.getAttributes())); + User.setAttributes(StripNest(F->getContext(), User.getAttributes())); } } @@ -2103,7 +2118,7 @@ bool GlobalOpt::OptimizeFunctions(Module &M) { Changed = true; } - if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) && + if (F->getAttributes().hasAttrSomewhere(Attributes::Nest) && !F->hasAddressTaken()) { // The function is not used by a trampoline intrinsic, so it is safe // to remove the 'nest' attribute. @@ -2251,7 +2266,7 @@ static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, static inline bool isSimpleEnoughValueToCommit(Constant *C, SmallPtrSet<Constant*, 8> &SimpleConstants, - const TargetData *TD); + const DataLayout *TD); /// isSimpleEnoughValueToCommit - Return true if the specified constant can be @@ -2264,7 +2279,7 @@ isSimpleEnoughValueToCommit(Constant *C, /// time. static bool isSimpleEnoughValueToCommitHelper(Constant *C, SmallPtrSet<Constant*, 8> &SimpleConstants, - const TargetData *TD) { + const DataLayout *TD) { // Simple integer, undef, constant aggregate zero, global addresses, etc are // all supported. if (C->getNumOperands() == 0 || isa<BlockAddress>(C) || @@ -2319,7 +2334,7 @@ static bool isSimpleEnoughValueToCommitHelper(Constant *C, static inline bool isSimpleEnoughValueToCommit(Constant *C, SmallPtrSet<Constant*, 8> &SimpleConstants, - const TargetData *TD) { + const DataLayout *TD) { // If we already checked this constant, we win. if (!SimpleConstants.insert(C)) return true; // Check the constant. @@ -2450,7 +2465,7 @@ namespace { /// Once an evaluation call fails, the evaluation object should not be reused. class Evaluator { public: - Evaluator(const TargetData *TD, const TargetLibraryInfo *TLI) + Evaluator(const DataLayout *TD, const TargetLibraryInfo *TLI) : TD(TD), TLI(TLI) { ValueStack.push_back(new DenseMap<Value*, Constant*>); } @@ -2531,7 +2546,7 @@ private: /// simple enough to live in a static initializer of a global. SmallPtrSet<Constant*, 8> SimpleConstants; - const TargetData *TD; + const DataLayout *TD; const TargetLibraryInfo *TLI; }; @@ -2869,7 +2884,7 @@ bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal, /// EvaluateStaticConstructor - Evaluate static constructors in the function, if /// we can. Return true if we can, false otherwise. -static bool EvaluateStaticConstructor(Function *F, const TargetData *TD, +static bool EvaluateStaticConstructor(Function *F, const DataLayout *TD, const TargetLibraryInfo *TLI) { // Call the function. Evaluator Eval(TD, TLI); @@ -3110,7 +3125,7 @@ bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) { bool GlobalOpt::runOnModule(Module &M) { bool Changed = false; - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); // Try to find the llvm.globalctors list. diff --git a/contrib/llvm/lib/Transforms/IPO/IPO.cpp b/contrib/llvm/lib/Transforms/IPO/IPO.cpp index 6233922..5d563d8 100644 --- a/contrib/llvm/lib/Transforms/IPO/IPO.cpp +++ b/contrib/llvm/lib/Transforms/IPO/IPO.cpp @@ -1,4 +1,4 @@ -//===-- Scalar.cpp --------------------------------------------------------===// +//===-- IPO.cpp -----------------------------------------------------------===// // // The LLVM Compiler Infrastructure // @@ -95,7 +95,10 @@ void LLVMAddIPSCCPPass(LLVMPassManagerRef PM) { } void LLVMAddInternalizePass(LLVMPassManagerRef PM, unsigned AllButMain) { - unwrap(PM)->add(createInternalizePass(AllButMain != 0)); + std::vector<const char *> Export; + if (AllButMain) + Export.push_back("main"); + unwrap(PM)->add(createInternalizePass(Export)); } void LLVMAddStripDeadPrototypesPass(LLVMPassManagerRef PM) { diff --git a/contrib/llvm/lib/Transforms/IPO/InlineAlways.cpp b/contrib/llvm/lib/Transforms/IPO/InlineAlways.cpp index 664ddf6..b1c36c1 100644 --- a/contrib/llvm/lib/Transforms/IPO/InlineAlways.cpp +++ b/contrib/llvm/lib/Transforms/IPO/InlineAlways.cpp @@ -23,7 +23,7 @@ #include "llvm/Support/CallSite.h" #include "llvm/Transforms/IPO.h" #include "llvm/Transforms/IPO/InlinerPass.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/ADT/SmallPtrSet.h" using namespace llvm; @@ -65,7 +65,7 @@ Pass *llvm::createAlwaysInlinerPass(bool InsertLifetime) { /// \brief Minimal filter to detect invalid constructs for inlining. static bool isInlineViable(Function &F) { - bool ReturnsTwice = F.hasFnAttr(Attribute::ReturnsTwice); + bool ReturnsTwice =F.getFnAttributes().hasAttribute(Attributes::ReturnsTwice); for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { // Disallow inlining of functions which contain an indirect branch. if (isa<IndirectBrInst>(BI->getTerminator())) @@ -114,7 +114,7 @@ InlineCost AlwaysInliner::getInlineCost(CallSite CS) { if (Callee->isDeclaration()) return InlineCost::getNever(); // Return never for anything not marked as always inline. - if (!Callee->hasFnAttr(Attribute::AlwaysInline)) + if (!Callee->getFnAttributes().hasAttribute(Attributes::AlwaysInline)) return InlineCost::getNever(); // Do some minimal analysis to preclude non-viable functions. diff --git a/contrib/llvm/lib/Transforms/IPO/InlineSimple.cpp b/contrib/llvm/lib/Transforms/IPO/InlineSimple.cpp index 50038d8..bf0b1f9 100644 --- a/contrib/llvm/lib/Transforms/IPO/InlineSimple.cpp +++ b/contrib/llvm/lib/Transforms/IPO/InlineSimple.cpp @@ -22,7 +22,7 @@ #include "llvm/Support/CallSite.h" #include "llvm/Transforms/IPO.h" #include "llvm/Transforms/IPO/InlinerPass.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" using namespace llvm; @@ -62,7 +62,7 @@ Pass *llvm::createFunctionInliningPass(int Threshold) { // doInitialization - Initializes the vector of functions that have been // annotated with the noinline attribute. bool SimpleInliner::doInitialization(CallGraph &CG) { - CA.setTargetData(getAnalysisIfAvailable<TargetData>()); + CA.setDataLayout(getAnalysisIfAvailable<DataLayout>()); return false; } diff --git a/contrib/llvm/lib/Transforms/IPO/Inliner.cpp b/contrib/llvm/lib/Transforms/IPO/Inliner.cpp index 712888a..abcb25f 100644 --- a/contrib/llvm/lib/Transforms/IPO/Inliner.cpp +++ b/contrib/llvm/lib/Transforms/IPO/Inliner.cpp @@ -19,7 +19,8 @@ #include "llvm/IntrinsicInst.h" #include "llvm/Analysis/CallGraph.h" #include "llvm/Analysis/InlineCost.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/IPO/InlinerPass.h" #include "llvm/Transforms/Utils/Cloning.h" #include "llvm/Transforms/Utils/Local.h" @@ -92,11 +93,11 @@ static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI, // If the inlined function had a higher stack protection level than the // calling function, then bump up the caller's stack protection level. - if (Callee->hasFnAttr(Attribute::StackProtectReq)) - Caller->addFnAttr(Attribute::StackProtectReq); - else if (Callee->hasFnAttr(Attribute::StackProtect) && - !Caller->hasFnAttr(Attribute::StackProtectReq)) - Caller->addFnAttr(Attribute::StackProtect); + if (Callee->getFnAttributes().hasAttribute(Attributes::StackProtectReq)) + Caller->addFnAttr(Attributes::StackProtectReq); + else if (Callee->getFnAttributes().hasAttribute(Attributes::StackProtect) && + !Caller->getFnAttributes().hasAttribute(Attributes::StackProtectReq)) + Caller->addFnAttr(Attributes::StackProtect); // Look at all of the allocas that we inlined through this call site. If we // have already inlined other allocas through other calls into this function, @@ -208,14 +209,15 @@ unsigned Inliner::getInlineThreshold(CallSite CS) const { // would decrease the threshold. Function *Caller = CS.getCaller(); bool OptSize = Caller && !Caller->isDeclaration() && - Caller->hasFnAttr(Attribute::OptimizeForSize); - if (!(InlineLimit.getNumOccurrences() > 0) && OptSize && OptSizeThreshold < thres) + Caller->getFnAttributes().hasAttribute(Attributes::OptimizeForSize); + if (!(InlineLimit.getNumOccurrences() > 0) && OptSize && + OptSizeThreshold < thres) thres = OptSizeThreshold; // Listen to the inlinehint attribute when it would increase the threshold. Function *Callee = CS.getCalledFunction(); bool InlineHint = Callee && !Callee->isDeclaration() && - Callee->hasFnAttr(Attribute::InlineHint); + Callee->getFnAttributes().hasAttribute(Attributes::InlineHint); if (InlineHint && HintThreshold > thres) thres = HintThreshold; @@ -338,7 +340,8 @@ static bool InlineHistoryIncludes(Function *F, int InlineHistoryID, bool Inliner::runOnSCC(CallGraphSCC &SCC) { CallGraph &CG = getAnalysis<CallGraph>(); - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); + const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); SmallPtrSet<Function*, 8> SCCFunctions; DEBUG(dbgs() << "Inliner visiting SCC:"); @@ -417,7 +420,7 @@ bool Inliner::runOnSCC(CallGraphSCC &SCC) { // just delete the call instead of trying to inline it, regardless of // size. This happens because IPSCCP propagates the result out of the // call and then we're left with the dead call. - if (isInstructionTriviallyDead(CS.getInstruction())) { + if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) { DEBUG(dbgs() << " -> Deleting dead call: " << *CS.getInstruction() << "\n"); // Update the call graph by deleting the edge from Callee to Caller. @@ -530,7 +533,8 @@ bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) { // Handle the case when this function is called and we only want to care // about always-inline functions. This is a bit of a hack to share code // between here and the InlineAlways pass. - if (AlwaysInlineOnly && !F->hasFnAttr(Attribute::AlwaysInline)) + if (AlwaysInlineOnly && + !F->getFnAttributes().hasAttribute(Attributes::AlwaysInline)) continue; // If the only remaining users of the function are dead constants, remove diff --git a/contrib/llvm/lib/Transforms/IPO/Internalize.cpp b/contrib/llvm/lib/Transforms/IPO/Internalize.cpp index fb5869e..aa629cc 100644 --- a/contrib/llvm/lib/Transforms/IPO/Internalize.cpp +++ b/contrib/llvm/lib/Transforms/IPO/Internalize.cpp @@ -7,9 +7,9 @@ // //===----------------------------------------------------------------------===// // -// This pass loops over all of the functions in the input module, looking for a -// main function. If a main function is found, all other functions and all -// global variables with initializers are marked as internal. +// This pass loops over all of the functions and variables in the input module. +// If the function or variable is not in the list of external names given to +// the pass it is marked as internal. // //===----------------------------------------------------------------------===// @@ -45,12 +45,9 @@ APIList("internalize-public-api-list", cl::value_desc("list"), namespace { class InternalizePass : public ModulePass { std::set<std::string> ExternalNames; - /// If no api symbols were specified and a main function is defined, - /// assume the main function is the only API - bool AllButMain; public: static char ID; // Pass identification, replacement for typeid - explicit InternalizePass(bool AllButMain = true); + explicit InternalizePass(); explicit InternalizePass(const std::vector <const char *>& exportList); void LoadFile(const char *Filename); virtual bool runOnModule(Module &M); @@ -66,8 +63,8 @@ char InternalizePass::ID = 0; INITIALIZE_PASS(InternalizePass, "internalize", "Internalize Global Symbols", false, false) -InternalizePass::InternalizePass(bool AllButMain) - : ModulePass(ID), AllButMain(AllButMain){ +InternalizePass::InternalizePass() + : ModulePass(ID) { initializeInternalizePassPass(*PassRegistry::getPassRegistry()); if (!APIFile.empty()) // If a filename is specified, use it. LoadFile(APIFile.c_str()); @@ -76,7 +73,7 @@ InternalizePass::InternalizePass(bool AllButMain) } InternalizePass::InternalizePass(const std::vector<const char *>&exportList) - : ModulePass(ID), AllButMain(false){ + : ModulePass(ID){ initializeInternalizePassPass(*PassRegistry::getPassRegistry()); for(std::vector<const char *>::const_iterator itr = exportList.begin(); itr != exportList.end(); itr++) { @@ -103,23 +100,6 @@ void InternalizePass::LoadFile(const char *Filename) { bool InternalizePass::runOnModule(Module &M) { CallGraph *CG = getAnalysisIfAvailable<CallGraph>(); CallGraphNode *ExternalNode = CG ? CG->getExternalCallingNode() : 0; - - if (ExternalNames.empty()) { - // Return if we're not in 'all but main' mode and have no external api - if (!AllButMain) - return false; - // If no list or file of symbols was specified, check to see if there is a - // "main" symbol defined in the module. If so, use it, otherwise do not - // internalize the module, it must be a library or something. - // - Function *MainFunc = M.getFunction("main"); - if (MainFunc == 0 || MainFunc->isDeclaration()) - return false; // No main found, must be a library... - - // Preserve main, internalize all else. - ExternalNames.insert(MainFunc->getName()); - } - bool Changed = false; // Never internalize functions which code-gen might insert. @@ -189,8 +169,8 @@ bool InternalizePass::runOnModule(Module &M) { return Changed; } -ModulePass *llvm::createInternalizePass(bool AllButMain) { - return new InternalizePass(AllButMain); +ModulePass *llvm::createInternalizePass() { + return new InternalizePass(); } ModulePass *llvm::createInternalizePass(const std::vector <const char *> &el) { diff --git a/contrib/llvm/lib/Transforms/IPO/MergeFunctions.cpp b/contrib/llvm/lib/Transforms/IPO/MergeFunctions.cpp index 9f70f66..44283dd 100644 --- a/contrib/llvm/lib/Transforms/IPO/MergeFunctions.cpp +++ b/contrib/llvm/lib/Transforms/IPO/MergeFunctions.cpp @@ -63,7 +63,7 @@ #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/ValueHandle.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include <vector> using namespace llvm; @@ -92,19 +92,19 @@ static unsigned profileFunction(const Function *F) { namespace { /// ComparableFunction - A struct that pairs together functions with a -/// TargetData so that we can keep them together as elements in the DenseSet. +/// DataLayout so that we can keep them together as elements in the DenseSet. class ComparableFunction { public: static const ComparableFunction EmptyKey; static const ComparableFunction TombstoneKey; - static TargetData * const LookupOnly; + static DataLayout * const LookupOnly; - ComparableFunction(Function *Func, TargetData *TD) + ComparableFunction(Function *Func, DataLayout *TD) : Func(Func), Hash(profileFunction(Func)), TD(TD) {} Function *getFunc() const { return Func; } unsigned getHash() const { return Hash; } - TargetData *getTD() const { return TD; } + DataLayout *getTD() const { return TD; } // Drops AssertingVH reference to the function. Outside of debug mode, this // does nothing. @@ -120,13 +120,13 @@ private: AssertingVH<Function> Func; unsigned Hash; - TargetData *TD; + DataLayout *TD; }; const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0); const ComparableFunction ComparableFunction::TombstoneKey = ComparableFunction(1); -TargetData *const ComparableFunction::LookupOnly = (TargetData*)(-1); +DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1); } @@ -150,12 +150,12 @@ namespace llvm { namespace { /// FunctionComparator - Compares two functions to determine whether or not -/// they will generate machine code with the same behaviour. TargetData is +/// they will generate machine code with the same behaviour. DataLayout is /// used if available. The comparator always fails conservatively (erring on the /// side of claiming that two functions are different). class FunctionComparator { public: - FunctionComparator(const TargetData *TD, const Function *F1, + FunctionComparator(const DataLayout *TD, const Function *F1, const Function *F2) : F1(F1), F2(F2), TD(TD) {} @@ -190,7 +190,7 @@ private: // The two functions undergoing comparison. const Function *F1, *F2; - const TargetData *TD; + const DataLayout *TD; DenseMap<const Value *, const Value *> id_map; DenseSet<const Value *> seen_values; @@ -591,8 +591,8 @@ private: /// to modify it. FnSetType FnSet; - /// TargetData for more accurate GEP comparisons. May be NULL. - TargetData *TD; + /// DataLayout for more accurate GEP comparisons. May be NULL. + DataLayout *TD; /// Whether or not the target supports global aliases. bool HasGlobalAliases; @@ -609,7 +609,7 @@ ModulePass *llvm::createMergeFunctionsPass() { bool MergeFunctions::runOnModule(Module &M) { bool Changed = false; - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage()) diff --git a/contrib/llvm/lib/Transforms/IPO/PassManagerBuilder.cpp b/contrib/llvm/lib/Transforms/IPO/PassManagerBuilder.cpp index 43b4ab5..05253fc 100644 --- a/contrib/llvm/lib/Transforms/IPO/PassManagerBuilder.cpp +++ b/contrib/llvm/lib/Transforms/IPO/PassManagerBuilder.cpp @@ -33,13 +33,21 @@ using namespace llvm; static cl::opt<bool> -RunVectorization("vectorize", cl::desc("Run vectorization passes")); +RunLoopVectorization("vectorize-loops", + cl::desc("Run the Loop vectorization passes")); + +static cl::opt<bool> +RunBBVectorization("vectorize", cl::desc("Run the BB vectorization passes")); static cl::opt<bool> UseGVNAfterVectorization("use-gvn-after-vectorization", cl::init(false), cl::Hidden, cl::desc("Run GVN instead of Early CSE after vectorization passes")); +static cl::opt<bool> UseNewSROA("use-new-sroa", + cl::init(true), cl::Hidden, + cl::desc("Enable the new, experimental SROA pass")); + PassManagerBuilder::PassManagerBuilder() { OptLevel = 2; SizeLevel = 0; @@ -48,7 +56,8 @@ PassManagerBuilder::PassManagerBuilder() { DisableSimplifyLibCalls = false; DisableUnitAtATime = false; DisableUnrollLoops = false; - Vectorize = RunVectorization; + Vectorize = RunBBVectorization; + LoopVectorize = RunLoopVectorization; } PassManagerBuilder::~PassManagerBuilder() { @@ -100,7 +109,10 @@ void PassManagerBuilder::populateFunctionPassManager(FunctionPassManager &FPM) { addInitialAliasAnalysisPasses(FPM); FPM.add(createCFGSimplificationPass()); - FPM.add(createScalarReplAggregatesPass()); + if (UseNewSROA) + FPM.add(createSROAPass()); + else + FPM.add(createScalarReplAggregatesPass()); FPM.add(createEarlyCSEPass()); FPM.add(createLowerExpectIntrinsicPass()); } @@ -112,6 +124,14 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) { MPM.add(Inliner); Inliner = 0; } + + // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC + // pass manager, but we don't want to add extensions into that pass manager. + // To prevent this we must insert a no-op module pass to reset the pass + // manager to get the same behavior as EP_OptimizerLast in non-O0 builds. + if (!GlobalExtensions->empty() || !Extensions.empty()) + MPM.add(createBarrierNoopPass()); + addExtensionsToPM(EP_EnabledOnOptLevel0, MPM); return; } @@ -147,7 +167,10 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) { // Start of function pass. // Break up aggregate allocas, using SSAUpdater. - MPM.add(createScalarReplAggregatesPass(-1, false)); + if (UseNewSROA) + MPM.add(createSROAPass(/*RequiresDomTree*/ false)); + else + MPM.add(createScalarReplAggregatesPass(-1, false)); MPM.add(createEarlyCSEPass()); // Catch trivial redundancies if (!DisableSimplifyLibCalls) MPM.add(createSimplifyLibCallsPass()); // Library Call Optimizations @@ -166,6 +189,12 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) { MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars MPM.add(createLoopIdiomPass()); // Recognize idioms like memset. MPM.add(createLoopDeletionPass()); // Delete dead loops + + if (LoopVectorize) { + MPM.add(createLoopVectorizePass()); + MPM.add(createLICMPass()); + } + if (!DisableUnrollLoops) MPM.add(createLoopUnrollPass()); // Unroll small loops addExtensionsToPM(EP_LoopOptimizerEnd, MPM); @@ -201,13 +230,12 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) { // FIXME: We shouldn't bother with this anymore. MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes - // GlobalOpt already deletes dead functions and globals, at -O3 try a + // GlobalOpt already deletes dead functions and globals, at -O2 try a // late pass of GlobalDCE. It is capable of deleting dead cycles. - if (OptLevel > 2) + if (OptLevel > 1) { MPM.add(createGlobalDCEPass()); // Remove dead fns and globals. - - if (OptLevel > 1) MPM.add(createConstantMergePass()); // Merge dup global constants + } } addExtensionsToPM(EP_OptimizerLast, MPM); } @@ -222,8 +250,11 @@ void PassManagerBuilder::populateLTOPassManager(PassManagerBase &PM, // Now that composite has been compiled, scan through the module, looking // for a main function. If main is defined, mark all other functions // internal. - if (Internalize) - PM.add(createInternalizePass(true)); + if (Internalize) { + std::vector<const char*> E; + E.push_back("main"); + PM.add(createInternalizePass(E)); + } // Propagate constants at call sites into the functions they call. This // opens opportunities for globalopt (and inlining) by substituting function @@ -265,7 +296,10 @@ void PassManagerBuilder::populateLTOPassManager(PassManagerBase &PM, PM.add(createInstructionCombiningPass()); PM.add(createJumpThreadingPass()); // Break up allocas - PM.add(createScalarReplAggregatesPass()); + if (UseNewSROA) + PM.add(createSROAPass()); + else + PM.add(createScalarReplAggregatesPass()); // Run a few AA driven optimizations here and now, to cleanup the code. PM.add(createFunctionAttrsPass()); // Add nocapture. diff --git a/contrib/llvm/lib/Transforms/IPO/PruneEH.cpp b/contrib/llvm/lib/Transforms/IPO/PruneEH.cpp index c8cc8fd..fb4ecbf 100644 --- a/contrib/llvm/lib/Transforms/IPO/PruneEH.cpp +++ b/contrib/llvm/lib/Transforms/IPO/PruneEH.cpp @@ -137,16 +137,18 @@ bool PruneEH::runOnSCC(CallGraphSCC &SCC) { // If the SCC doesn't unwind or doesn't throw, note this fact. if (!SCCMightUnwind || !SCCMightReturn) for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { - Attributes NewAttributes = Attribute::None; + AttrBuilder NewAttributes; if (!SCCMightUnwind) - NewAttributes |= Attribute::NoUnwind; + NewAttributes.addAttribute(Attributes::NoUnwind); if (!SCCMightReturn) - NewAttributes |= Attribute::NoReturn; + NewAttributes.addAttribute(Attributes::NoReturn); Function *F = (*I)->getFunction(); const AttrListPtr &PAL = F->getAttributes(); - const AttrListPtr &NPAL = PAL.addAttr(~0, NewAttributes); + const AttrListPtr &NPAL = PAL.addAttr(F->getContext(), ~0, + Attributes::get(F->getContext(), + NewAttributes)); if (PAL != NPAL) { MadeChange = true; F->setAttributes(NPAL); diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombine.h b/contrib/llvm/lib/Transforms/InstCombine/InstCombine.h index 0d5ef90..7467eca 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombine.h +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombine.h @@ -18,10 +18,11 @@ #include "llvm/Analysis/ValueTracking.h" #include "llvm/Support/InstVisitor.h" #include "llvm/Support/TargetFolder.h" +#include "llvm/Transforms/Utils/SimplifyLibCalls.h" namespace llvm { class CallSite; - class TargetData; + class DataLayout; class TargetLibraryInfo; class DbgDeclareInst; class MemIntrinsic; @@ -71,9 +72,10 @@ public: class LLVM_LIBRARY_VISIBILITY InstCombiner : public FunctionPass, public InstVisitor<InstCombiner, Instruction*> { - TargetData *TD; + DataLayout *TD; TargetLibraryInfo *TLI; bool MadeIRChange; + LibCallSimplifier *Simplifier; public: /// Worklist - All of the instructions that need to be simplified. InstCombineWorklist Worklist; @@ -95,7 +97,7 @@ public: virtual void getAnalysisUsage(AnalysisUsage &AU) const; - TargetData *getTargetData() const { return TD; } + DataLayout *getDataLayout() const { return TD; } TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; } @@ -218,7 +220,7 @@ private: Type *Ty); Instruction *visitCallSite(CallSite CS); - Instruction *tryOptimizeCall(CallInst *CI, const TargetData *TD); + Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *TD); bool transformConstExprCastCall(CallSite CS); Instruction *transformCallThroughTrampoline(CallSite CS, IntrinsicInst *Tramp); @@ -365,6 +367,10 @@ private: Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned); + + /// Descale - Return a value X such that Val = X * Scale, or null if none. If + /// the multiplication is known not to overflow then NoSignedWrap is set. + Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap); }; diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp index 99b62f8..d8257e6 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp @@ -13,7 +13,7 @@ #include "InstCombine.h" #include "llvm/Analysis/InstructionSimplify.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/PatternMatch.h" using namespace llvm; diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp index cbe1ca4..48f2704 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp @@ -13,7 +13,7 @@ #include "InstCombine.h" #include "llvm/Support/CallSite.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Analysis/MemoryBuiltins.h" #include "llvm/Transforms/Utils/BuildLibCalls.h" #include "llvm/Transforms/Utils/Local.h" @@ -29,6 +29,26 @@ static Type *getPromotedType(Type *Ty) { return Ty; } +/// reduceToSingleValueType - Given an aggregate type which ultimately holds a +/// single scalar element, like {{{type}}} or [1 x type], return type. +static Type *reduceToSingleValueType(Type *T) { + while (!T->isSingleValueType()) { + if (StructType *STy = dyn_cast<StructType>(T)) { + if (STy->getNumElements() == 1) + T = STy->getElementType(0); + else + break; + } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) { + if (ATy->getNumElements() == 1) + T = ATy->getElementType(); + else + break; + } else + break; + } + + return T; +} Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD); @@ -74,35 +94,37 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { // dest address will be promotable. See if we can find a better type than the // integer datatype. Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); + MDNode *CopyMD = 0; if (StrippedDest != MI->getArgOperand(0)) { Type *SrcETy = cast<PointerType>(StrippedDest->getType()) ->getElementType(); if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { // The SrcETy might be something like {{{double}}} or [1 x double]. Rip // down through these levels if so. - while (!SrcETy->isSingleValueType()) { - if (StructType *STy = dyn_cast<StructType>(SrcETy)) { - if (STy->getNumElements() == 1) - SrcETy = STy->getElementType(0); - else - break; - } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { - if (ATy->getNumElements() == 1) - SrcETy = ATy->getElementType(); - else - break; - } else - break; - } + SrcETy = reduceToSingleValueType(SrcETy); if (SrcETy->isSingleValueType()) { NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); + + // If the memcpy has metadata describing the members, see if we can + // get the TBAA tag describing our copy. + if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) { + if (M->getNumOperands() == 3 && + M->getOperand(0) && + isa<ConstantInt>(M->getOperand(0)) && + cast<ConstantInt>(M->getOperand(0))->isNullValue() && + M->getOperand(1) && + isa<ConstantInt>(M->getOperand(1)) && + cast<ConstantInt>(M->getOperand(1))->getValue() == Size && + M->getOperand(2) && + isa<MDNode>(M->getOperand(2))) + CopyMD = cast<MDNode>(M->getOperand(2)); + } } } } - // If the memcpy/memmove provides better alignment info than we can // infer, use it. SrcAlign = std::max(SrcAlign, CopyAlign); @@ -112,8 +134,12 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile()); L->setAlignment(SrcAlign); + if (CopyMD) + L->setMetadata(LLVMContext::MD_tbaa, CopyMD); StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile()); S->setAlignment(DstAlign); + if (CopyMD) + S->setMetadata(LLVMContext::MD_tbaa, CopyMD); // Set the size of the copy to 0, it will be deleted on the next iteration. MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); @@ -168,7 +194,7 @@ Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { /// the heavy lifting. /// Instruction *InstCombiner::visitCallInst(CallInst &CI) { - if (isFreeCall(&CI)) + if (isFreeCall(&CI, TLI)) return visitFree(CI); // If the caller function is nounwind, mark the call as nounwind, even if the @@ -243,7 +269,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { default: break; case Intrinsic::objectsize: { uint64_t Size; - if (getObjectSize(II->getArgOperand(0), Size, TD)) + if (getObjectSize(II->getArgOperand(0), Size, TD, TLI)) return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size)); return 0; } @@ -731,7 +757,7 @@ Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { /// passed through the varargs area, we can eliminate the use of the cast. static bool isSafeToEliminateVarargsCast(const CallSite CS, const CastInst * const CI, - const TargetData * const TD, + const DataLayout * const TD, const int ix) { if (!CI->isLosslessCast()) return false; @@ -752,49 +778,17 @@ static bool isSafeToEliminateVarargsCast(const CallSite CS, return true; } -namespace { -class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { - InstCombiner *IC; -protected: - void replaceCall(Value *With) { - NewInstruction = IC->ReplaceInstUsesWith(*CI, With); - } - bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { - if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) - return true; - if (ConstantInt *SizeCI = - dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { - if (SizeCI->isAllOnesValue()) - return true; - if (isString) { - uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); - // If the length is 0 we don't know how long it is and so we can't - // remove the check. - if (Len == 0) return false; - return SizeCI->getZExtValue() >= Len; - } - if (ConstantInt *Arg = dyn_cast<ConstantInt>( - CI->getArgOperand(SizeArgOp))) - return SizeCI->getZExtValue() >= Arg->getZExtValue(); - } - return false; - } -public: - InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } - Instruction *NewInstruction; -}; -} // end anonymous namespace - // Try to fold some different type of calls here. // Currently we're only working with the checking functions, memcpy_chk, // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, // strcat_chk and strncat_chk. -Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { +Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *TD) { if (CI->getCalledFunction() == 0) return 0; - InstCombineFortifiedLibCalls Simplifier(this); - Simplifier.fold(CI, TD, TLI); - return Simplifier.NewInstruction; + if (Value *With = Simplifier->optimizeCall(CI)) + return ReplaceInstUsesWith(*CI, With); + + return 0; } static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { @@ -877,7 +871,7 @@ static IntrinsicInst *FindInitTrampoline(Value *Callee) { // visitCallSite - Improvements for call and invoke instructions. // Instruction *InstCombiner::visitCallSite(CallSite CS) { - if (isAllocLikeFn(CS.getInstruction())) + if (isAllocLikeFn(CS.getInstruction(), TLI)) return visitAllocSite(*CS.getInstruction()); bool Changed = false; @@ -961,7 +955,7 @@ Instruction *InstCombiner::visitCallSite(CallSite CS) { Changed = true; } - // Try to optimize the call if possible, we require TargetData for most of + // Try to optimize the call if possible, we require DataLayout for most of // this. None of these calls are seen as possibly dead so go ahead and // delete the instruction now. if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { @@ -1013,8 +1007,8 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { return false; // Cannot transform this return value. if (!CallerPAL.isEmpty() && !Caller->use_empty()) { - Attributes RAttrs = CallerPAL.getRetAttributes(); - if (RAttrs & Attribute::typeIncompatible(NewRetTy)) + AttrBuilder RAttrs = CallerPAL.getRetAttributes(); + if (RAttrs.hasAttributes(Attributes::typeIncompatible(NewRetTy))) return false; // Attribute not compatible with transformed value. } @@ -1044,12 +1038,13 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { return false; // Cannot transform this parameter value. Attributes Attrs = CallerPAL.getParamAttributes(i + 1); - if (Attrs & Attribute::typeIncompatible(ParamTy)) + if (AttrBuilder(Attrs). + hasAttributes(Attributes::typeIncompatible(ParamTy))) return false; // Attribute not compatible with transformed value. // If the parameter is passed as a byval argument, then we have to have a // sized type and the sized type has to have the same size as the old type. - if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { + if (ParamTy != ActTy && Attrs.hasAttribute(Attributes::ByVal)) { PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) return false; @@ -1101,7 +1096,7 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) break; Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; - if (PAttrs & Attribute::VarArgsIncompatible) + if (PAttrs.hasIncompatibleWithVarArgsAttrs()) return false; } @@ -1114,15 +1109,17 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { attrVec.reserve(NumCommonArgs); // Get any return attributes. - Attributes RAttrs = CallerPAL.getRetAttributes(); + AttrBuilder RAttrs = CallerPAL.getRetAttributes(); // If the return value is not being used, the type may not be compatible // with the existing attributes. Wipe out any problematic attributes. - RAttrs &= ~Attribute::typeIncompatible(NewRetTy); + RAttrs.removeAttributes(Attributes::typeIncompatible(NewRetTy)); // Add the new return attributes. - if (RAttrs) - attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); + if (RAttrs.hasAttributes()) + attrVec.push_back( + AttributeWithIndex::get(AttrListPtr::ReturnIndex, + Attributes::get(FT->getContext(), RAttrs))); AI = CS.arg_begin(); for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { @@ -1136,7 +1133,8 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { } // Add any parameter attributes. - if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) + Attributes PAttrs = CallerPAL.getParamAttributes(i + 1); + if (PAttrs.hasAttributes()) attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); } @@ -1164,19 +1162,23 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { } // Add any parameter attributes. - if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) + Attributes PAttrs = CallerPAL.getParamAttributes(i + 1); + if (PAttrs.hasAttributes()) attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); } } } - if (Attributes FnAttrs = CallerPAL.getFnAttributes()) - attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); + Attributes FnAttrs = CallerPAL.getFnAttributes(); + if (FnAttrs.hasAttributes()) + attrVec.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + FnAttrs)); if (NewRetTy->isVoidTy()) Caller->setName(""); // Void type should not have a name. - const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec); + const AttrListPtr &NewCallerPAL = AttrListPtr::get(Callee->getContext(), + attrVec); Instruction *NC; if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { @@ -1240,8 +1242,9 @@ InstCombiner::transformCallThroughTrampoline(CallSite CS, // If the call already has the 'nest' attribute somewhere then give up - // otherwise 'nest' would occur twice after splicing in the chain. - if (Attrs.hasAttrSomewhere(Attribute::Nest)) - return 0; + for (unsigned I = 0, E = Attrs.getNumAttrs(); I != E; ++I) + if (Attrs.getAttributesAtIndex(I).hasAttribute(Attributes::Nest)) + return 0; assert(Tramp && "transformCallThroughTrampoline called with incorrect CallSite."); @@ -1254,12 +1257,12 @@ InstCombiner::transformCallThroughTrampoline(CallSite CS, if (!NestAttrs.isEmpty()) { unsigned NestIdx = 1; Type *NestTy = 0; - Attributes NestAttr = Attribute::None; + Attributes NestAttr; // Look for a parameter marked with the 'nest' attribute. for (FunctionType::param_iterator I = NestFTy->param_begin(), E = NestFTy->param_end(); I != E; ++NestIdx, ++I) - if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { + if (NestAttrs.getParamAttributes(NestIdx).hasAttribute(Attributes::Nest)){ // Record the parameter type and any other attributes. NestTy = *I; NestAttr = NestAttrs.getParamAttributes(NestIdx); @@ -1278,8 +1281,10 @@ InstCombiner::transformCallThroughTrampoline(CallSite CS, // mean appending it. Likewise for attributes. // Add any result attributes. - if (Attributes Attr = Attrs.getRetAttributes()) - NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); + Attributes Attr = Attrs.getRetAttributes(); + if (Attr.hasAttributes()) + NewAttrs.push_back(AttributeWithIndex::get(AttrListPtr::ReturnIndex, + Attr)); { unsigned Idx = 1; @@ -1299,7 +1304,8 @@ InstCombiner::transformCallThroughTrampoline(CallSite CS, // Add the original argument and attributes. NewArgs.push_back(*I); - if (Attributes Attr = Attrs.getParamAttributes(Idx)) + Attr = Attrs.getParamAttributes(Idx); + if (Attr.hasAttributes()) NewAttrs.push_back (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); @@ -1308,8 +1314,10 @@ InstCombiner::transformCallThroughTrampoline(CallSite CS, } // Add any function attributes. - if (Attributes Attr = Attrs.getFnAttributes()) - NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); + Attr = Attrs.getFnAttributes(); + if (Attr.hasAttributes()) + NewAttrs.push_back(AttributeWithIndex::get(AttrListPtr::FunctionIndex, + Attr)); // The trampoline may have been bitcast to a bogus type (FTy). // Handle this by synthesizing a new function type, equal to FTy @@ -1348,7 +1356,7 @@ InstCombiner::transformCallThroughTrampoline(CallSite CS, NestF->getType() == PointerType::getUnqual(NewFTy) ? NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy)); - const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs); + const AttrListPtr &NewPAL = AttrListPtr::get(FTy->getContext(), NewAttrs); Instruction *NewCaller; if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp index 555b442..bb59db8 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -13,7 +13,7 @@ #include "InstCombine.h" #include "llvm/Analysis/ConstantFolding.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Support/PatternMatch.h" using namespace llvm; @@ -78,7 +78,7 @@ static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, /// try to eliminate the cast by moving the type information into the alloc. Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI) { - // This requires TargetData to get the alloca alignment and size information. + // This requires DataLayout to get the alloca alignment and size information. if (!TD) return 0; PointerType *PTy = cast<PointerType>(CI.getType()); @@ -229,7 +229,7 @@ isEliminableCastPair( const CastInst *CI, ///< The first cast instruction unsigned opcode, ///< The opcode of the second cast instruction Type *DstTy, ///< The target type for the second cast instruction - TargetData *TD ///< The target data for pointer size + DataLayout *TD ///< The target data for pointer size ) { Type *SrcTy = CI->getOperand(0)->getType(); // A from above @@ -238,17 +238,20 @@ isEliminableCastPair( // Get the opcodes of the two Cast instructions Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); Instruction::CastOps secondOp = Instruction::CastOps(opcode); - + Type *SrcIntPtrTy = TD && SrcTy->isPtrOrPtrVectorTy() ? + TD->getIntPtrType(SrcTy) : 0; + Type *MidIntPtrTy = TD && MidTy->isPtrOrPtrVectorTy() ? + TD->getIntPtrType(MidTy) : 0; + Type *DstIntPtrTy = TD && DstTy->isPtrOrPtrVectorTy() ? + TD->getIntPtrType(DstTy) : 0; unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, - DstTy, - TD ? TD->getIntPtrType(CI->getContext()) : 0); - + DstTy, SrcIntPtrTy, MidIntPtrTy, + DstIntPtrTy); + // We don't want to form an inttoptr or ptrtoint that converts to an integer // type that differs from the pointer size. - if ((Res == Instruction::IntToPtr && - (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || - (Res == Instruction::PtrToInt && - (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) + if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) || + (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy)) Res = 0; return Instruction::CastOps(Res); diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp index bdd310e..7c3f8fe 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp @@ -16,7 +16,8 @@ #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/MemoryBuiltins.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Support/ConstantRange.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/PatternMatch.h" @@ -473,7 +474,7 @@ FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, /// If we can't emit an optimized form for this expression, this returns null. /// static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) { - TargetData &TD = *IC.getTargetData(); + DataLayout &TD = *IC.getDataLayout(); gep_type_iterator GTI = gep_type_begin(GEP); // Check to see if this gep only has a single variable index. If so, and if @@ -2355,8 +2356,25 @@ Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { // Try not to increase register pressure. BO0->hasOneUse() && BO1->hasOneUse()) { // Determine Y and Z in the form icmp (X+Y), (X+Z). - Value *Y = (A == C || A == D) ? B : A; - Value *Z = (C == A || C == B) ? D : C; + Value *Y, *Z; + if (A == C) { + // C + B == C + D -> B == D + Y = B; + Z = D; + } else if (A == D) { + // D + B == C + D -> B == C + Y = B; + Z = C; + } else if (B == C) { + // A + C == C + D -> A == D + Y = A; + Z = D; + } else { + assert(B == D); + // A + D == C + D -> A == C + Y = A; + Z = C; + } return new ICmpInst(Pred, Y, Z); } @@ -2894,10 +2912,6 @@ Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { if (!RHSF) break; - // We can't convert a PPC double double. - if (RHSF->getType()->isPPC_FP128Ty()) - break; - const fltSemantics *Sem; // FIXME: This shouldn't be here. if (LHSExt->getSrcTy()->isHalfTy()) @@ -2910,6 +2924,8 @@ Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { Sem = &APFloat::IEEEquad; else if (LHSExt->getSrcTy()->isX86_FP80Ty()) Sem = &APFloat::x87DoubleExtended; + else if (LHSExt->getSrcTy()->isPPC_FP128Ty()) + Sem = &APFloat::PPCDoubleDouble; else break; @@ -2985,6 +3001,44 @@ Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { return Res; } break; + case Instruction::Call: { + CallInst *CI = cast<CallInst>(LHSI); + LibFunc::Func Func; + // Various optimization for fabs compared with zero. + if (RHSC->isNullValue() && CI->getCalledFunction() && + TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) && + TLI->has(Func)) { + if (Func == LibFunc::fabs || Func == LibFunc::fabsf || + Func == LibFunc::fabsl) { + switch (I.getPredicate()) { + default: break; + // fabs(x) < 0 --> false + case FCmpInst::FCMP_OLT: + return ReplaceInstUsesWith(I, Builder->getFalse()); + // fabs(x) > 0 --> x != 0 + case FCmpInst::FCMP_OGT: + return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), + RHSC); + // fabs(x) <= 0 --> x == 0 + case FCmpInst::FCMP_OLE: + return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), + RHSC); + // fabs(x) >= 0 --> !isnan(x) + case FCmpInst::FCMP_OGE: + return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), + RHSC); + // fabs(x) == 0 --> x == 0 + // fabs(x) != 0 --> x != 0 + case FCmpInst::FCMP_OEQ: + case FCmpInst::FCMP_UEQ: + case FCmpInst::FCMP_ONE: + case FCmpInst::FCMP_UNE: + return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), + RHSC); + } + } + } + } } } diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp index c485844..4d106fc 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp @@ -14,13 +14,161 @@ #include "InstCombine.h" #include "llvm/IntrinsicInst.h" #include "llvm/Analysis/Loads.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/ADT/Statistic.h" using namespace llvm; -STATISTIC(NumDeadStore, "Number of dead stores eliminated"); +STATISTIC(NumDeadStore, "Number of dead stores eliminated"); +STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global"); + +/// pointsToConstantGlobal - Return true if V (possibly indirectly) points to +/// some part of a constant global variable. This intentionally only accepts +/// constant expressions because we can't rewrite arbitrary instructions. +static bool pointsToConstantGlobal(Value *V) { + if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) + return GV->isConstant(); + if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) + if (CE->getOpcode() == Instruction::BitCast || + CE->getOpcode() == Instruction::GetElementPtr) + return pointsToConstantGlobal(CE->getOperand(0)); + return false; +} + +/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) +/// pointer to an alloca. Ignore any reads of the pointer, return false if we +/// see any stores or other unknown uses. If we see pointer arithmetic, keep +/// track of whether it moves the pointer (with IsOffset) but otherwise traverse +/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to +/// the alloca, and if the source pointer is a pointer to a constant global, we +/// can optimize this. +static bool +isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, + SmallVectorImpl<Instruction *> &ToDelete, + bool IsOffset = false) { + // We track lifetime intrinsics as we encounter them. If we decide to go + // ahead and replace the value with the global, this lets the caller quickly + // eliminate the markers. + + for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { + User *U = cast<Instruction>(*UI); + + if (LoadInst *LI = dyn_cast<LoadInst>(U)) { + // Ignore non-volatile loads, they are always ok. + if (!LI->isSimple()) return false; + continue; + } + + if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { + // If uses of the bitcast are ok, we are ok. + if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, ToDelete, IsOffset)) + return false; + continue; + } + if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { + // If the GEP has all zero indices, it doesn't offset the pointer. If it + // doesn't, it does. + if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, ToDelete, + IsOffset || !GEP->hasAllZeroIndices())) + return false; + continue; + } + + if (CallSite CS = U) { + // If this is the function being called then we treat it like a load and + // ignore it. + if (CS.isCallee(UI)) + continue; + + // If this is a readonly/readnone call site, then we know it is just a + // load (but one that potentially returns the value itself), so we can + // ignore it if we know that the value isn't captured. + unsigned ArgNo = CS.getArgumentNo(UI); + if (CS.onlyReadsMemory() && + (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) + continue; + + // If this is being passed as a byval argument, the caller is making a + // copy, so it is only a read of the alloca. + if (CS.isByValArgument(ArgNo)) + continue; + } + + // Lifetime intrinsics can be handled by the caller. + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { + if (II->getIntrinsicID() == Intrinsic::lifetime_start || + II->getIntrinsicID() == Intrinsic::lifetime_end) { + assert(II->use_empty() && "Lifetime markers have no result to use!"); + ToDelete.push_back(II); + continue; + } + } + + // If this is isn't our memcpy/memmove, reject it as something we can't + // handle. + MemTransferInst *MI = dyn_cast<MemTransferInst>(U); + if (MI == 0) + return false; + + // If the transfer is using the alloca as a source of the transfer, then + // ignore it since it is a load (unless the transfer is volatile). + if (UI.getOperandNo() == 1) { + if (MI->isVolatile()) return false; + continue; + } + + // If we already have seen a copy, reject the second one. + if (TheCopy) return false; + + // If the pointer has been offset from the start of the alloca, we can't + // safely handle this. + if (IsOffset) return false; + + // If the memintrinsic isn't using the alloca as the dest, reject it. + if (UI.getOperandNo() != 0) return false; + + // If the source of the memcpy/move is not a constant global, reject it. + if (!pointsToConstantGlobal(MI->getSource())) + return false; + + // Otherwise, the transform is safe. Remember the copy instruction. + TheCopy = MI; + } + return true; +} + +/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only +/// modified by a copy from a constant global. If we can prove this, we can +/// replace any uses of the alloca with uses of the global directly. +static MemTransferInst * +isOnlyCopiedFromConstantGlobal(AllocaInst *AI, + SmallVectorImpl<Instruction *> &ToDelete) { + MemTransferInst *TheCopy = 0; + if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete)) + return TheCopy; + return 0; +} + +/// getPointeeAlignment - Compute the minimum alignment of the value pointed +/// to by the given pointer. +static unsigned getPointeeAlignment(Value *V, const DataLayout &TD) { + if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) + if (CE->getOpcode() == Instruction::BitCast || + (CE->getOpcode() == Instruction::GetElementPtr && + cast<GEPOperator>(CE)->hasAllZeroIndices())) + return getPointeeAlignment(CE->getOperand(0), TD); + + if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) + if (!GV->isDeclaration()) + return TD.getPreferredAlignment(GV); + + if (PointerType *PT = dyn_cast<PointerType>(V->getType())) + if (PT->getElementType()->isSized()) + return TD.getABITypeAlignment(PT->getElementType()); + + return 0; +} Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) { // Ensure that the alloca array size argument has type intptr_t, so that @@ -99,12 +247,16 @@ Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) { return &AI; } + // If the alignment of the entry block alloca is 0 (unspecified), + // assign it the preferred alignment. + if (EntryAI->getAlignment() == 0) + EntryAI->setAlignment( + TD->getPrefTypeAlignment(EntryAI->getAllocatedType())); // Replace this zero-sized alloca with the one at the start of the entry // block after ensuring that the address will be aligned enough for both // types. - unsigned MaxAlign = - std::max(TD->getPrefTypeAlignment(EntryAI->getAllocatedType()), - TD->getPrefTypeAlignment(AI.getAllocatedType())); + unsigned MaxAlign = std::max(EntryAI->getAlignment(), + AI.getAlignment()); EntryAI->setAlignment(MaxAlign); if (AI.getType() != EntryAI->getType()) return new BitCastInst(EntryAI, AI.getType()); @@ -113,6 +265,31 @@ Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) { } } + if (TD) { + // Check to see if this allocation is only modified by a memcpy/memmove from + // a constant global whose alignment is equal to or exceeds that of the + // allocation. If this is the case, we can change all users to use + // the constant global instead. This is commonly produced by the CFE by + // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' + // is only subsequently read. + SmallVector<Instruction *, 4> ToDelete; + if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) { + if (AI.getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) { + DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n'); + DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); + for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) + EraseInstFromFunction(*ToDelete[i]); + Constant *TheSrc = cast<Constant>(Copy->getSource()); + Instruction *NewI + = ReplaceInstUsesWith(AI, ConstantExpr::getBitCast(TheSrc, + AI.getType())); + EraseInstFromFunction(*Copy); + ++NumGlobalCopies; + return NewI; + } + } + } + // At last, use the generic allocation site handler to aggressively remove // unused allocas. return visitAllocSite(AI); @@ -121,7 +298,7 @@ Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) { /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible. static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI, - const TargetData *TD) { + const DataLayout *TD) { User *CI = cast<User>(LI.getOperand(0)); Value *CastOp = CI->getOperand(0); @@ -151,14 +328,14 @@ static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI, SrcPTy = SrcTy->getElementType(); } - if (IC.getTargetData() && + if (IC.getDataLayout() && (SrcPTy->isIntegerTy() || SrcPTy->isPointerTy() || SrcPTy->isVectorTy()) && // Do not allow turning this into a load of an integer, which is then // casted to a pointer, this pessimizes pointer analysis a lot. (SrcPTy->isPointerTy() == LI.getType()->isPointerTy()) && - IC.getTargetData()->getTypeSizeInBits(SrcPTy) == - IC.getTargetData()->getTypeSizeInBits(DestPTy)) { + IC.getDataLayout()->getTypeSizeInBits(SrcPTy) == + IC.getDataLayout()->getTypeSizeInBits(DestPTy)) { // Okay, we are casting from one integer or pointer type to another of // the same size. Instead of casting the pointer before the load, cast @@ -336,11 +513,11 @@ static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) { // If the pointers point into different address spaces or if they point to // values with different sizes, we can't do the transformation. - if (!IC.getTargetData() || + if (!IC.getDataLayout() || SrcTy->getAddressSpace() != cast<PointerType>(CI->getType())->getAddressSpace() || - IC.getTargetData()->getTypeSizeInBits(SrcPTy) != - IC.getTargetData()->getTypeSizeInBits(DestPTy)) + IC.getDataLayout()->getTypeSizeInBits(SrcPTy) != + IC.getDataLayout()->getTypeSizeInBits(DestPTy)) return 0; // Okay, we are casting from one integer or pointer type to another of diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp index 35a0bbb..cefe45e 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -37,7 +37,7 @@ static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC) { if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(PowerOf2), m_Value(A))), m_Value(B))) && // The "1" can be any value known to be a power of 2. - isPowerOfTwo(PowerOf2, IC.getTargetData())) { + isPowerOfTwo(PowerOf2, IC.getDataLayout())) { A = IC.Builder->CreateSub(A, B); return IC.Builder->CreateShl(PowerOf2, A); } @@ -46,7 +46,7 @@ static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC) { // inexact. Similarly for <<. if (BinaryOperator *I = dyn_cast<BinaryOperator>(V)) if (I->isLogicalShift() && - isPowerOfTwo(I->getOperand(0), IC.getTargetData())) { + isPowerOfTwo(I->getOperand(0), IC.getDataLayout())) { // We know that this is an exact/nuw shift and that the input is a // non-zero context as well. if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC)) { @@ -462,12 +462,23 @@ Instruction *InstCombiner::visitUDiv(BinaryOperator &I) { } } + // (x lshr C1) udiv C2 --> x udiv (C2 << C1) + if (ConstantInt *C2 = dyn_cast<ConstantInt>(Op1)) { + Value *X; + ConstantInt *C1; + if (match(Op0, m_LShr(m_Value(X), m_ConstantInt(C1)))) { + APInt NC = C2->getValue().shl(C1->getLimitedValue(C1->getBitWidth()-1)); + return BinaryOperator::CreateUDiv(X, Builder->getInt(NC)); + } + } + // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2) { const APInt *CI; Value *N; if (match(Op1, m_Shl(m_Power2(CI), m_Value(N))) || match(Op1, m_ZExt(m_Shl(m_Power2(CI), m_Value(N))))) { if (*CI != 1) - N = Builder->CreateAdd(N, ConstantInt::get(I.getType(),CI->logBase2())); + N = Builder->CreateAdd(N, + ConstantInt::get(N->getType(), CI->logBase2())); if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1)) N = Builder->CreateZExt(N, Z->getDestTy()); if (I.isExact()) diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp index 664546c..de9c77e 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp @@ -13,7 +13,7 @@ #include "InstCombine.h" #include "llvm/Analysis/InstructionSimplify.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/STLExtras.h" using namespace llvm; diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp index 291e800..a2d4c88 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -287,7 +287,7 @@ Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal, /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is /// replaced with RepOp. static Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, - const TargetData *TD, + const DataLayout *TD, const TargetLibraryInfo *TLI) { // Trivial replacement. if (V == Op) @@ -333,6 +333,10 @@ static Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, // All operands were constants, fold it. if (ConstOps.size() == I->getNumOperands()) { + if (CmpInst *C = dyn_cast<CmpInst>(I)) + return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], + ConstOps[1], TD, TLI); + if (LoadInst *LI = dyn_cast<LoadInst>(I)) if (!LI->isVolatile()) return ConstantFoldLoadFromConstPtr(ConstOps[0], TD); @@ -903,7 +907,7 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { return &SI; } - if (VectorType* VecTy = dyn_cast<VectorType>(SI.getType())) { + if (VectorType *VecTy = dyn_cast<VectorType>(SI.getType())) { unsigned VWidth = VecTy->getNumElements(); APInt UndefElts(VWidth, 0); APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); @@ -912,6 +916,28 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { return ReplaceInstUsesWith(SI, V); return &SI; } + + if (ConstantVector *CV = dyn_cast<ConstantVector>(CondVal)) { + // Form a shufflevector instruction. + SmallVector<Constant *, 8> Mask(VWidth); + Type *Int32Ty = Type::getInt32Ty(CV->getContext()); + for (unsigned i = 0; i != VWidth; ++i) { + Constant *Elem = cast<Constant>(CV->getOperand(i)); + if (ConstantInt *E = dyn_cast<ConstantInt>(Elem)) + Mask[i] = ConstantInt::get(Int32Ty, i + (E->isZero() ? VWidth : 0)); + else if (isa<UndefValue>(Elem)) + Mask[i] = UndefValue::get(Int32Ty); + else + return 0; + } + Constant *MaskVal = ConstantVector::get(Mask); + Value *V = Builder->CreateShuffleVector(TrueVal, FalseVal, MaskVal); + return ReplaceInstUsesWith(SI, V); + } + + if (isa<ConstantAggregateZero>(CondVal)) { + return ReplaceInstUsesWith(SI, FalseVal); + } } return 0; diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp index 4bb2403..57021f1 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp @@ -190,7 +190,7 @@ static Value *GetShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, V = IC.Builder->CreateLShr(C, NumBits); // If we got a constantexpr back, try to simplify it with TD info. if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) - V = ConstantFoldConstantExpression(CE, IC.getTargetData(), + V = ConstantFoldConstantExpression(CE, IC.getDataLayout(), IC.getTargetLibraryInfo()); return V; } diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp index 54be8ed..602b203 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp @@ -14,7 +14,7 @@ #include "InstCombine.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/IntrinsicInst.h" using namespace llvm; diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp index cf60f0f..dd7ea14 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp @@ -636,8 +636,11 @@ Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { // If LHS's width is changed, shift the mask value accordingly. // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any - // references to RHSOp0 to LHSOp0, so we don't need to shift the mask. - if (eltMask >= 0 && newRHS != NULL) + // references from RHSOp0 to LHSOp0, so we don't need to shift the mask. + // If newRHS == newLHS, we want to remap any references from newRHS to + // newLHS so that we can properly identify splats that may occur due to + // obfuscation accross the two vectors. + if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS) eltMask += newLHSWidth; } diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstCombineWorklist.h b/contrib/llvm/lib/Transforms/InstCombine/InstCombineWorklist.h index 99a02fc..ea654ae 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstCombineWorklist.h +++ b/contrib/llvm/lib/Transforms/InstCombine/InstCombineWorklist.h @@ -26,8 +26,8 @@ class LLVM_LIBRARY_VISIBILITY InstCombineWorklist { SmallVector<Instruction*, 256> Worklist; DenseMap<Instruction*, unsigned> WorklistMap; - void operator=(const InstCombineWorklist&RHS); // DO NOT IMPLEMENT - InstCombineWorklist(const InstCombineWorklist&); // DO NOT IMPLEMENT + void operator=(const InstCombineWorklist&RHS) LLVM_DELETED_FUNCTION; + InstCombineWorklist(const InstCombineWorklist&) LLVM_DELETED_FUNCTION; public: InstCombineWorklist() {} diff --git a/contrib/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/contrib/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp index 68ecd51..9a46f25 100644 --- a/contrib/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp +++ b/contrib/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp @@ -40,7 +40,7 @@ #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/MemoryBuiltins.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Support/CFG.h" @@ -88,7 +88,7 @@ void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const { Value *InstCombiner::EmitGEPOffset(User *GEP) { - return llvm::EmitGEPOffset(Builder, *getTargetData(), GEP); + return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP); } /// ShouldChangeType - Return true if it is desirable to convert a computation @@ -805,6 +805,244 @@ static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) { return true; } +/// Descale - Return a value X such that Val = X * Scale, or null if none. If +/// the multiplication is known not to overflow then NoSignedWrap is set. +Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) { + assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!"); + assert(cast<IntegerType>(Val->getType())->getBitWidth() == + Scale.getBitWidth() && "Scale not compatible with value!"); + + // If Val is zero or Scale is one then Val = Val * Scale. + if (match(Val, m_Zero()) || Scale == 1) { + NoSignedWrap = true; + return Val; + } + + // If Scale is zero then it does not divide Val. + if (Scale.isMinValue()) + return 0; + + // Look through chains of multiplications, searching for a constant that is + // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4 + // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by + // a factor of 4 will produce X*(Y*2). The principle of operation is to bore + // down from Val: + // + // Val = M1 * X || Analysis starts here and works down + // M1 = M2 * Y || Doesn't descend into terms with more + // M2 = Z * 4 \/ than one use + // + // Then to modify a term at the bottom: + // + // Val = M1 * X + // M1 = Z * Y || Replaced M2 with Z + // + // Then to work back up correcting nsw flags. + + // Op - the term we are currently analyzing. Starts at Val then drills down. + // Replaced with its descaled value before exiting from the drill down loop. + Value *Op = Val; + + // Parent - initially null, but after drilling down notes where Op came from. + // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the + // 0'th operand of Val. + std::pair<Instruction*, unsigned> Parent; + + // RequireNoSignedWrap - Set if the transform requires a descaling at deeper + // levels that doesn't overflow. + bool RequireNoSignedWrap = false; + + // logScale - log base 2 of the scale. Negative if not a power of 2. + int32_t logScale = Scale.exactLogBase2(); + + for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down + + if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { + // If Op is a constant divisible by Scale then descale to the quotient. + APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth. + APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder); + if (!Remainder.isMinValue()) + // Not divisible by Scale. + return 0; + // Replace with the quotient in the parent. + Op = ConstantInt::get(CI->getType(), Quotient); + NoSignedWrap = true; + break; + } + + if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) { + + if (BO->getOpcode() == Instruction::Mul) { + // Multiplication. + NoSignedWrap = BO->hasNoSignedWrap(); + if (RequireNoSignedWrap && !NoSignedWrap) + return 0; + + // There are three cases for multiplication: multiplication by exactly + // the scale, multiplication by a constant different to the scale, and + // multiplication by something else. + Value *LHS = BO->getOperand(0); + Value *RHS = BO->getOperand(1); + + if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { + // Multiplication by a constant. + if (CI->getValue() == Scale) { + // Multiplication by exactly the scale, replace the multiplication + // by its left-hand side in the parent. + Op = LHS; + break; + } + + // Otherwise drill down into the constant. + if (!Op->hasOneUse()) + return 0; + + Parent = std::make_pair(BO, 1); + continue; + } + + // Multiplication by something else. Drill down into the left-hand side + // since that's where the reassociate pass puts the good stuff. + if (!Op->hasOneUse()) + return 0; + + Parent = std::make_pair(BO, 0); + continue; + } + + if (logScale > 0 && BO->getOpcode() == Instruction::Shl && + isa<ConstantInt>(BO->getOperand(1))) { + // Multiplication by a power of 2. + NoSignedWrap = BO->hasNoSignedWrap(); + if (RequireNoSignedWrap && !NoSignedWrap) + return 0; + + Value *LHS = BO->getOperand(0); + int32_t Amt = cast<ConstantInt>(BO->getOperand(1))-> + getLimitedValue(Scale.getBitWidth()); + // Op = LHS << Amt. + + if (Amt == logScale) { + // Multiplication by exactly the scale, replace the multiplication + // by its left-hand side in the parent. + Op = LHS; + break; + } + if (Amt < logScale || !Op->hasOneUse()) + return 0; + + // Multiplication by more than the scale. Reduce the multiplying amount + // by the scale in the parent. + Parent = std::make_pair(BO, 1); + Op = ConstantInt::get(BO->getType(), Amt - logScale); + break; + } + } + + if (!Op->hasOneUse()) + return 0; + + if (CastInst *Cast = dyn_cast<CastInst>(Op)) { + if (Cast->getOpcode() == Instruction::SExt) { + // Op is sign-extended from a smaller type, descale in the smaller type. + unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); + APInt SmallScale = Scale.trunc(SmallSize); + // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to + // descale Op as (sext Y) * Scale. In order to have + // sext (Y * SmallScale) = (sext Y) * Scale + // some conditions need to hold however: SmallScale must sign-extend to + // Scale and the multiplication Y * SmallScale should not overflow. + if (SmallScale.sext(Scale.getBitWidth()) != Scale) + // SmallScale does not sign-extend to Scale. + return 0; + assert(SmallScale.exactLogBase2() == logScale); + // Require that Y * SmallScale must not overflow. + RequireNoSignedWrap = true; + + // Drill down through the cast. + Parent = std::make_pair(Cast, 0); + Scale = SmallScale; + continue; + } + + if (Cast->getOpcode() == Instruction::Trunc) { + // Op is truncated from a larger type, descale in the larger type. + // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then + // trunc (Y * sext Scale) = (trunc Y) * Scale + // always holds. However (trunc Y) * Scale may overflow even if + // trunc (Y * sext Scale) does not, so nsw flags need to be cleared + // from this point up in the expression (see later). + if (RequireNoSignedWrap) + return 0; + + // Drill down through the cast. + unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits(); + Parent = std::make_pair(Cast, 0); + Scale = Scale.sext(LargeSize); + if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits()) + logScale = -1; + assert(Scale.exactLogBase2() == logScale); + continue; + } + } + + // Unsupported expression, bail out. + return 0; + } + + // We know that we can successfully descale, so from here on we can safely + // modify the IR. Op holds the descaled version of the deepest term in the + // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known + // not to overflow. + + if (!Parent.first) + // The expression only had one term. + return Op; + + // Rewrite the parent using the descaled version of its operand. + assert(Parent.first->hasOneUse() && "Drilled down when more than one use!"); + assert(Op != Parent.first->getOperand(Parent.second) && + "Descaling was a no-op?"); + Parent.first->setOperand(Parent.second, Op); + Worklist.Add(Parent.first); + + // Now work back up the expression correcting nsw flags. The logic is based + // on the following observation: if X * Y is known not to overflow as a signed + // multiplication, and Y is replaced by a value Z with smaller absolute value, + // then X * Z will not overflow as a signed multiplication either. As we work + // our way up, having NoSignedWrap 'true' means that the descaled value at the + // current level has strictly smaller absolute value than the original. + Instruction *Ancestor = Parent.first; + do { + if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) { + // If the multiplication wasn't nsw then we can't say anything about the + // value of the descaled multiplication, and we have to clear nsw flags + // from this point on up. + bool OpNoSignedWrap = BO->hasNoSignedWrap(); + NoSignedWrap &= OpNoSignedWrap; + if (NoSignedWrap != OpNoSignedWrap) { + BO->setHasNoSignedWrap(NoSignedWrap); + Worklist.Add(Ancestor); + } + } else if (Ancestor->getOpcode() == Instruction::Trunc) { + // The fact that the descaled input to the trunc has smaller absolute + // value than the original input doesn't tell us anything useful about + // the absolute values of the truncations. + NoSignedWrap = false; + } + assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) && + "Failed to keep proper track of nsw flags while drilling down?"); + + if (Ancestor == Val) + // Got to the top, all done! + return Val; + + // Move up one level in the expression. + assert(Ancestor->hasOneUse() && "Drilled down when more than one use!"); + Ancestor = Ancestor->use_back(); + } while (1); +} + Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end()); @@ -817,7 +1055,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // by multiples of a zero size type with zero. if (TD) { bool MadeChange = false; - Type *IntPtrTy = TD->getIntPtrType(GEP.getContext()); + Type *IntPtrTy = TD->getIntPtrType(GEP.getPointerOperandType()); gep_type_iterator GTI = gep_type_begin(GEP); for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); @@ -836,7 +1074,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { } Type *IndexTy = (*I)->getType(); - if (IndexTy != IntPtrTy && !IndexTy->isVectorTy()) { + if (IndexTy != IntPtrTy) { // If we are using a wider index than needed for this platform, shrink // it to what we need. If narrower, sign-extend it to what we need. // This explicit cast can make subsequent optimizations more obvious. @@ -855,7 +1093,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src)) return 0; - // Note that if our source is a gep chain itself that we wait for that + // Note that if our source is a gep chain itself then we wait for that // chain to be resolved before we perform this transformation. This // avoids us creating a TON of code in some cases. if (GEPOperator *SrcGEP = @@ -987,63 +1225,74 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { } // Transform things like: + // %V = mul i64 %N, 4 + // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V + // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast + if (TD && ResElTy->isSized() && SrcElTy->isSized()) { + // Check that changing the type amounts to dividing the index by a scale + // factor. + uint64_t ResSize = TD->getTypeAllocSize(ResElTy); + uint64_t SrcSize = TD->getTypeAllocSize(SrcElTy); + if (ResSize && SrcSize % ResSize == 0) { + Value *Idx = GEP.getOperand(1); + unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); + uint64_t Scale = SrcSize / ResSize; + + // Earlier transforms ensure that the index has type IntPtrType, which + // considerably simplifies the logic by eliminating implicit casts. + assert(Idx->getType() == TD->getIntPtrType(GEP.getContext()) && + "Index not cast to pointer width?"); + + bool NSW; + if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { + // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. + // If the multiplication NewIdx * Scale may overflow then the new + // GEP may not be "inbounds". + Value *NewGEP = GEP.isInBounds() && NSW ? + Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) : + Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName()); + // The NewGEP must be pointer typed, so must the old one -> BitCast + return new BitCastInst(NewGEP, GEP.getType()); + } + } + } + + // Similarly, transform things like: // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp // (where tmp = 8*tmp2) into: // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast - - if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) { + if (TD && ResElTy->isSized() && SrcElTy->isSized() && + SrcElTy->isArrayTy()) { + // Check that changing to the array element type amounts to dividing the + // index by a scale factor. + uint64_t ResSize = TD->getTypeAllocSize(ResElTy); uint64_t ArrayEltSize = - TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()); - - // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We - // allow either a mul, shift, or constant here. - Value *NewIdx = 0; - ConstantInt *Scale = 0; - if (ArrayEltSize == 1) { - NewIdx = GEP.getOperand(1); - Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1); - } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) { - NewIdx = ConstantInt::get(CI->getType(), 1); - Scale = CI; - } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){ - if (Inst->getOpcode() == Instruction::Shl && - isa<ConstantInt>(Inst->getOperand(1))) { - ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1)); - uint32_t ShAmtVal = ShAmt->getLimitedValue(64); - Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()), - 1ULL << ShAmtVal); - NewIdx = Inst->getOperand(0); - } else if (Inst->getOpcode() == Instruction::Mul && - isa<ConstantInt>(Inst->getOperand(1))) { - Scale = cast<ConstantInt>(Inst->getOperand(1)); - NewIdx = Inst->getOperand(0); + TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()); + if (ResSize && ArrayEltSize % ResSize == 0) { + Value *Idx = GEP.getOperand(1); + unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits(); + uint64_t Scale = ArrayEltSize / ResSize; + + // Earlier transforms ensure that the index has type IntPtrType, which + // considerably simplifies the logic by eliminating implicit casts. + assert(Idx->getType() == TD->getIntPtrType(GEP.getContext()) && + "Index not cast to pointer width?"); + + bool NSW; + if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) { + // Successfully decomposed Idx as NewIdx * Scale, form a new GEP. + // If the multiplication NewIdx * Scale may overflow then the new + // GEP may not be "inbounds". + Value *Off[2]; + Off[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); + Off[1] = NewIdx; + Value *NewGEP = GEP.isInBounds() && NSW ? + Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) : + Builder->CreateGEP(StrippedPtr, Off, GEP.getName()); + // The NewGEP must be pointer typed, so must the old one -> BitCast + return new BitCastInst(NewGEP, GEP.getType()); } } - - // If the index will be to exactly the right offset with the scale taken - // out, perform the transformation. Note, we don't know whether Scale is - // signed or not. We'll use unsigned version of division/modulo - // operation after making sure Scale doesn't have the sign bit set. - if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL && - Scale->getZExtValue() % ArrayEltSize == 0) { - Scale = ConstantInt::get(Scale->getType(), - Scale->getZExtValue() / ArrayEltSize); - if (Scale->getZExtValue() != 1) { - Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(), - false /*ZExt*/); - NewIdx = Builder->CreateMul(NewIdx, C, "idxscale"); - } - - // Insert the new GEP instruction. - Value *Idx[2]; - Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext())); - Idx[1] = NewIdx; - Value *NewGEP = GEP.isInBounds() ? - Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()): - Builder->CreateGEP(StrippedPtr, Idx, GEP.getName()); - // The NewGEP must be pointer typed, so must the old one -> BitCast - return new BitCastInst(NewGEP, GEP.getType()); - } } } } @@ -1068,7 +1317,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // If the bitcast is of an allocation, and the allocation will be // converted to match the type of the cast, don't touch this. if (isa<AllocaInst>(BCI->getOperand(0)) || - isAllocationFn(BCI->getOperand(0))) { + isAllocationFn(BCI->getOperand(0), TLI)) { // See if the bitcast simplifies, if so, don't nuke this GEP yet. if (Instruction *I = visitBitCast(*BCI)) { if (I != BCI) { @@ -1107,7 +1356,8 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { static bool -isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users) { +isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users, + const TargetLibraryInfo *TLI) { SmallVector<Instruction*, 4> Worklist; Worklist.push_back(AI); @@ -1163,7 +1413,7 @@ isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users) { } } - if (isFreeCall(I)) { + if (isFreeCall(I, TLI)) { Users.push_back(I); continue; } @@ -1188,7 +1438,7 @@ Instruction *InstCombiner::visitAllocSite(Instruction &MI) { // to null and free calls, delete the calls and replace the comparisons with // true or false as appropriate. SmallVector<WeakVH, 64> Users; - if (isAllocSiteRemovable(&MI, Users)) { + if (isAllocSiteRemovable(&MI, Users, TLI)) { for (unsigned i = 0, e = Users.size(); i != e; ++i) { Instruction *I = cast_or_null<Instruction>(&*Users[i]); if (!I) continue; @@ -1853,7 +2103,7 @@ static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) { static bool AddReachableCodeToWorklist(BasicBlock *BB, SmallPtrSet<BasicBlock*, 64> &Visited, InstCombiner &IC, - const TargetData *TD, + const DataLayout *TD, const TargetLibraryInfo *TLI) { bool MadeIRChange = false; SmallVector<BasicBlock*, 256> Worklist; @@ -1872,7 +2122,7 @@ static bool AddReachableCodeToWorklist(BasicBlock *BB, Instruction *Inst = BBI++; // DCE instruction if trivially dead. - if (isInstructionTriviallyDead(Inst)) { + if (isInstructionTriviallyDead(Inst, TLI)) { ++NumDeadInst; DEBUG(errs() << "IC: DCE: " << *Inst << '\n'); Inst->eraseFromParent(); @@ -2002,7 +2252,7 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { if (I == 0) continue; // skip null values. // Check to see if we can DCE the instruction. - if (isInstructionTriviallyDead(I)) { + if (isInstructionTriviallyDead(I, TLI)) { DEBUG(errs() << "IC: DCE: " << *I << '\n'); EraseInstFromFunction(*I); ++NumDeadInst; @@ -2102,7 +2352,7 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { // If the instruction was modified, it's possible that it is now dead. // if so, remove it. - if (isInstructionTriviallyDead(I)) { + if (isInstructionTriviallyDead(I, TLI)) { EraseInstFromFunction(*I); } else { Worklist.Add(I); @@ -2117,9 +2367,27 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) { return MadeIRChange; } +namespace { +class InstCombinerLibCallSimplifier : public LibCallSimplifier { + InstCombiner *IC; +public: + InstCombinerLibCallSimplifier(const DataLayout *TD, + const TargetLibraryInfo *TLI, + InstCombiner *IC) + : LibCallSimplifier(TD, TLI) { + this->IC = IC; + } + + /// replaceAllUsesWith - override so that instruction replacement + /// can be defined in terms of the instruction combiner framework. + virtual void replaceAllUsesWith(Instruction *I, Value *With) const { + IC->ReplaceInstUsesWith(*I, With); + } +}; +} bool InstCombiner::runOnFunction(Function &F) { - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); /// Builder - This is an IRBuilder that automatically inserts new @@ -2129,6 +2397,9 @@ bool InstCombiner::runOnFunction(Function &F) { InstCombineIRInserter(Worklist)); Builder = &TheBuilder; + InstCombinerLibCallSimplifier TheSimplifier(TD, TLI, this); + Simplifier = &TheSimplifier; + bool EverMadeChange = false; // Lower dbg.declare intrinsics otherwise their value may be clobbered diff --git a/contrib/llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp b/contrib/llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp index 17b83ce..b7be462 100644 --- a/contrib/llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp +++ b/contrib/llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp @@ -15,7 +15,7 @@ #define DEBUG_TYPE "asan" -#include "FunctionBlackList.h" +#include "BlackList.h" #include "llvm/Function.h" #include "llvm/IRBuilder.h" #include "llvm/InlineAsm.h" @@ -35,7 +35,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Support/system_error.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Transforms/Instrumentation.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" @@ -61,6 +61,8 @@ static const int kAsanCtorAndCtorPriority = 1; static const char *kAsanReportErrorTemplate = "__asan_report_"; static const char *kAsanRegisterGlobalsName = "__asan_register_globals"; static const char *kAsanUnregisterGlobalsName = "__asan_unregister_globals"; +static const char *kAsanPoisonGlobalsName = "__asan_before_dynamic_init"; +static const char *kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init"; static const char *kAsanInitName = "__asan_init"; static const char *kAsanHandleNoReturnName = "__asan_handle_no_return"; static const char *kAsanMappingOffsetName = "__asan_mapping_offset"; @@ -106,6 +108,8 @@ static cl::opt<bool> ClUseAfterReturn("asan-use-after-return", // This flag may need to be replaced with -f[no]asan-globals. static cl::opt<bool> ClGlobals("asan-globals", cl::desc("Handle global objects"), cl::Hidden, cl::init(true)); +static cl::opt<bool> ClInitializers("asan-initialization-order", + cl::desc("Handle C++ initializer order"), cl::Hidden, cl::init(false)); static cl::opt<bool> ClMemIntrin("asan-memintrin", cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true)); // This flag may need to be replaced with -fasan-blacklist. @@ -144,41 +148,33 @@ static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"), cl::Hidden, cl::init(-1)); namespace { - -/// An object of this type is created while instrumenting every function. -struct AsanFunctionContext { - AsanFunctionContext(Function &Function) : F(Function) { } - - Function &F; -}; - /// AddressSanitizer: instrument the code in module to find memory bugs. -struct AddressSanitizer : public ModulePass { +struct AddressSanitizer : public FunctionPass { AddressSanitizer(); virtual const char *getPassName() const; - void instrumentMop(AsanFunctionContext &AFC, Instruction *I); - void instrumentAddress(AsanFunctionContext &AFC, - Instruction *OrigIns, IRBuilder<> &IRB, + void instrumentMop(Instruction *I); + void instrumentAddress(Instruction *OrigIns, IRBuilder<> &IRB, Value *Addr, uint32_t TypeSize, bool IsWrite); Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong, Value *ShadowValue, uint32_t TypeSize); Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr, bool IsWrite, size_t AccessSizeIndex); - bool instrumentMemIntrinsic(AsanFunctionContext &AFC, MemIntrinsic *MI); - void instrumentMemIntrinsicParam(AsanFunctionContext &AFC, - Instruction *OrigIns, Value *Addr, + bool instrumentMemIntrinsic(MemIntrinsic *MI); + void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite); Value *memToShadow(Value *Shadow, IRBuilder<> &IRB); - bool handleFunction(Module &M, Function &F); + bool runOnFunction(Function &F); + void createInitializerPoisonCalls(Module &M, + Value *FirstAddr, Value *LastAddr); bool maybeInsertAsanInitAtFunctionEntry(Function &F); - bool poisonStackInFunction(Module &M, Function &F); - virtual bool runOnModule(Module &M); + bool poisonStackInFunction(Function &F); + virtual bool doInitialization(Module &M); + virtual bool doFinalization(Module &M); bool insertGlobalRedzones(Module &M); static char ID; // Pass identification, replacement for typeid private: - uint64_t getAllocaSizeInBytes(AllocaInst *AI) { Type *Ty = AI->getAllocatedType(); uint64_t SizeInBytes = TD->getTypeAllocSize(Ty); @@ -194,12 +190,15 @@ struct AddressSanitizer : public ModulePass { } Function *checkInterfaceFunction(Constant *FuncOrBitcast); + bool ShouldInstrumentGlobal(GlobalVariable *G); void PoisonStack(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB, Value *ShadowBase, bool DoPoison); bool LooksLikeCodeInBug11395(Instruction *I); + void FindDynamicInitializers(Module &M); + bool HasDynamicInitializer(GlobalVariable *G); LLVMContext *C; - TargetData *TD; + DataLayout *TD; uint64_t MappingOffset; int MappingScale; size_t RedzoneSize; @@ -208,11 +207,15 @@ struct AddressSanitizer : public ModulePass { Type *IntptrPtrTy; Function *AsanCtorFunction; Function *AsanInitFunction; + Function *AsanStackMallocFunc, *AsanStackFreeFunc; + Function *AsanHandleNoReturnFunc; Instruction *CtorInsertBefore; - OwningPtr<FunctionBlackList> BL; + OwningPtr<BlackList> BL; // This array is indexed by AccessIsWrite and log2(AccessSize). Function *AsanErrorCallback[2][kNumberOfAccessSizes]; InlineAsm *EmptyAsm; + SmallSet<GlobalValue*, 32> DynamicallyInitializedGlobals; + SmallSet<GlobalValue*, 32> GlobalsCreatedByAsan; }; } // namespace @@ -221,8 +224,8 @@ char AddressSanitizer::ID = 0; INITIALIZE_PASS(AddressSanitizer, "asan", "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false, false) -AddressSanitizer::AddressSanitizer() : ModulePass(ID) { } -ModulePass *llvm::createAddressSanitizerPass() { +AddressSanitizer::AddressSanitizer() : FunctionPass(ID) { } +FunctionPass *llvm::createAddressSanitizerPass() { return new AddressSanitizer(); } @@ -243,38 +246,6 @@ static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) { GlobalValue::PrivateLinkage, StrConst, ""); } -// Split the basic block and insert an if-then code. -// Before: -// Head -// Cmp -// Tail -// After: -// Head -// if (Cmp) -// ThenBlock -// Tail -// -// ThenBlock block is created and its terminator is returned. -// If Unreachable, ThenBlock is terminated with UnreachableInst, otherwise -// it is terminated with BranchInst to Tail. -static TerminatorInst *splitBlockAndInsertIfThen(Value *Cmp, bool Unreachable) { - Instruction *SplitBefore = cast<Instruction>(Cmp)->getNextNode(); - BasicBlock *Head = SplitBefore->getParent(); - BasicBlock *Tail = Head->splitBasicBlock(SplitBefore); - TerminatorInst *HeadOldTerm = Head->getTerminator(); - LLVMContext &C = Head->getParent()->getParent()->getContext(); - BasicBlock *ThenBlock = BasicBlock::Create(C, "", Head->getParent(), Tail); - TerminatorInst *CheckTerm; - if (Unreachable) - CheckTerm = new UnreachableInst(C, ThenBlock); - else - CheckTerm = BranchInst::Create(Tail, ThenBlock); - BranchInst *HeadNewTerm = - BranchInst::Create(/*ifTrue*/ThenBlock, /*ifFalse*/Tail, Cmp); - ReplaceInstWithInst(HeadOldTerm, HeadNewTerm); - return CheckTerm; -} - Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) { // Shadow >> scale Shadow = IRB.CreateLShr(Shadow, MappingScale); @@ -286,12 +257,12 @@ Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) { } void AddressSanitizer::instrumentMemIntrinsicParam( - AsanFunctionContext &AFC, Instruction *OrigIns, + Instruction *OrigIns, Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) { // Check the first byte. { IRBuilder<> IRB(InsertBefore); - instrumentAddress(AFC, OrigIns, IRB, Addr, 8, IsWrite); + instrumentAddress(OrigIns, IRB, Addr, 8, IsWrite); } // Check the last byte. { @@ -301,13 +272,12 @@ void AddressSanitizer::instrumentMemIntrinsicParam( SizeMinusOne = IRB.CreateIntCast(SizeMinusOne, IntptrTy, false); Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); Value *AddrPlusSizeMinisOne = IRB.CreateAdd(AddrLong, SizeMinusOne); - instrumentAddress(AFC, OrigIns, IRB, AddrPlusSizeMinisOne, 8, IsWrite); + instrumentAddress(OrigIns, IRB, AddrPlusSizeMinisOne, 8, IsWrite); } } // Instrument memset/memmove/memcpy -bool AddressSanitizer::instrumentMemIntrinsic(AsanFunctionContext &AFC, - MemIntrinsic *MI) { +bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) { Value *Dst = MI->getDest(); MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI); Value *Src = MemTran ? MemTran->getSource() : 0; @@ -323,12 +293,12 @@ bool AddressSanitizer::instrumentMemIntrinsic(AsanFunctionContext &AFC, Value *Cmp = IRB.CreateICmpNE(Length, Constant::getNullValue(Length->getType())); - InsertBefore = splitBlockAndInsertIfThen(Cmp, false); + InsertBefore = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false); } - instrumentMemIntrinsicParam(AFC, MI, Dst, Length, InsertBefore, true); + instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true); if (Src) - instrumentMemIntrinsicParam(AFC, MI, Src, Length, InsertBefore, false); + instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false); return true; } @@ -358,14 +328,50 @@ static Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite) { return NULL; } -void AddressSanitizer::instrumentMop(AsanFunctionContext &AFC, Instruction *I) { - bool IsWrite; +void AddressSanitizer::FindDynamicInitializers(Module& M) { + // Clang generates metadata identifying all dynamically initialized globals. + NamedMDNode *DynamicGlobals = + M.getNamedMetadata("llvm.asan.dynamically_initialized_globals"); + if (!DynamicGlobals) + return; + for (int i = 0, n = DynamicGlobals->getNumOperands(); i < n; ++i) { + MDNode *MDN = DynamicGlobals->getOperand(i); + assert(MDN->getNumOperands() == 1); + Value *VG = MDN->getOperand(0); + // The optimizer may optimize away a global entirely, in which case we + // cannot instrument access to it. + if (!VG) + continue; + + GlobalVariable *G = cast<GlobalVariable>(VG); + DynamicallyInitializedGlobals.insert(G); + } +} +// Returns true if a global variable is initialized dynamically in this TU. +bool AddressSanitizer::HasDynamicInitializer(GlobalVariable *G) { + return DynamicallyInitializedGlobals.count(G); +} + +void AddressSanitizer::instrumentMop(Instruction *I) { + bool IsWrite = false; Value *Addr = isInterestingMemoryAccess(I, &IsWrite); assert(Addr); - if (ClOpt && ClOptGlobals && isa<GlobalVariable>(Addr)) { - // We are accessing a global scalar variable. Nothing to catch here. - return; + if (ClOpt && ClOptGlobals) { + if (GlobalVariable *G = dyn_cast<GlobalVariable>(Addr)) { + // If initialization order checking is disabled, a simple access to a + // dynamically initialized global is always valid. + if (!ClInitializers) + return; + // If a global variable does not have dynamic initialization we don't + // have to instrument it. However, if a global has external linkage, we + // assume it has dynamic initialization, as it may have an initializer + // in a different TU. + if (G->getLinkage() != GlobalVariable::ExternalLinkage && + !HasDynamicInitializer(G)) + return; + } } + Type *OrigPtrTy = Addr->getType(); Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType(); @@ -379,7 +385,7 @@ void AddressSanitizer::instrumentMop(AsanFunctionContext &AFC, Instruction *I) { } IRBuilder<> IRB(I); - instrumentAddress(AFC, I, IRB, Addr, TypeSize, IsWrite); + instrumentAddress(I, IRB, Addr, TypeSize, IsWrite); } // Validate the result of Module::getOrInsertFunction called for an interface @@ -424,8 +430,7 @@ Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong, return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue); } -void AddressSanitizer::instrumentAddress(AsanFunctionContext &AFC, - Instruction *OrigIns, +void AddressSanitizer::instrumentAddress(Instruction *OrigIns, IRBuilder<> &IRB, Value *Addr, uint32_t TypeSize, bool IsWrite) { Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); @@ -444,17 +449,19 @@ void AddressSanitizer::instrumentAddress(AsanFunctionContext &AFC, TerminatorInst *CrashTerm = 0; if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) { - TerminatorInst *CheckTerm = splitBlockAndInsertIfThen(Cmp, false); + TerminatorInst *CheckTerm = + SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false); assert(dyn_cast<BranchInst>(CheckTerm)->isUnconditional()); BasicBlock *NextBB = CheckTerm->getSuccessor(0); IRB.SetInsertPoint(CheckTerm); Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize); - BasicBlock *CrashBlock = BasicBlock::Create(*C, "", &AFC.F, NextBB); + BasicBlock *CrashBlock = + BasicBlock::Create(*C, "", NextBB->getParent(), NextBB); CrashTerm = new UnreachableInst(*C, CrashBlock); BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2); ReplaceInstWithInst(CheckTerm, NewTerm); } else { - CrashTerm = splitBlockAndInsertIfThen(Cmp, true); + CrashTerm = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), true); } Instruction *Crash = @@ -462,68 +469,108 @@ void AddressSanitizer::instrumentAddress(AsanFunctionContext &AFC, Crash->setDebugLoc(OrigIns->getDebugLoc()); } +void AddressSanitizer::createInitializerPoisonCalls(Module &M, + Value *FirstAddr, + Value *LastAddr) { + // We do all of our poisoning and unpoisoning within _GLOBAL__I_a. + Function *GlobalInit = M.getFunction("_GLOBAL__I_a"); + // If that function is not present, this TU contains no globals, or they have + // all been optimized away + if (!GlobalInit) + return; + + // Set up the arguments to our poison/unpoison functions. + IRBuilder<> IRB(GlobalInit->begin()->getFirstInsertionPt()); + + // Declare our poisoning and unpoisoning functions. + Function *AsanPoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction( + kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL)); + AsanPoisonGlobals->setLinkage(Function::ExternalLinkage); + Function *AsanUnpoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction( + kAsanUnpoisonGlobalsName, IRB.getVoidTy(), NULL)); + AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage); + + // Add a call to poison all external globals before the given function starts. + IRB.CreateCall2(AsanPoisonGlobals, FirstAddr, LastAddr); + + // Add calls to unpoison all globals before each return instruction. + for (Function::iterator I = GlobalInit->begin(), E = GlobalInit->end(); + I != E; ++I) { + if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator())) { + CallInst::Create(AsanUnpoisonGlobals, "", RI); + } + } +} + +bool AddressSanitizer::ShouldInstrumentGlobal(GlobalVariable *G) { + Type *Ty = cast<PointerType>(G->getType())->getElementType(); + DEBUG(dbgs() << "GLOBAL: " << *G << "\n"); + + if (BL->isIn(*G)) return false; + if (!Ty->isSized()) return false; + if (!G->hasInitializer()) return false; + if (GlobalsCreatedByAsan.count(G)) return false; // Our own global. + // Touch only those globals that will not be defined in other modules. + // Don't handle ODR type linkages since other modules may be built w/o asan. + if (G->getLinkage() != GlobalVariable::ExternalLinkage && + G->getLinkage() != GlobalVariable::PrivateLinkage && + G->getLinkage() != GlobalVariable::InternalLinkage) + return false; + // Two problems with thread-locals: + // - The address of the main thread's copy can't be computed at link-time. + // - Need to poison all copies, not just the main thread's one. + if (G->isThreadLocal()) + return false; + // For now, just ignore this Alloca if the alignment is large. + if (G->getAlignment() > RedzoneSize) return false; + + // Ignore all the globals with the names starting with "\01L_OBJC_". + // Many of those are put into the .cstring section. The linker compresses + // that section by removing the spare \0s after the string terminator, so + // our redzones get broken. + if ((G->getName().find("\01L_OBJC_") == 0) || + (G->getName().find("\01l_OBJC_") == 0)) { + DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G); + return false; + } + + if (G->hasSection()) { + StringRef Section(G->getSection()); + // Ignore the globals from the __OBJC section. The ObjC runtime assumes + // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to + // them. + if ((Section.find("__OBJC,") == 0) || + (Section.find("__DATA, __objc_") == 0)) { + DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G); + return false; + } + // See http://code.google.com/p/address-sanitizer/issues/detail?id=32 + // Constant CFString instances are compiled in the following way: + // -- the string buffer is emitted into + // __TEXT,__cstring,cstring_literals + // -- the constant NSConstantString structure referencing that buffer + // is placed into __DATA,__cfstring + // Therefore there's no point in placing redzones into __DATA,__cfstring. + // Moreover, it causes the linker to crash on OS X 10.7 + if (Section.find("__DATA,__cfstring") == 0) { + DEBUG(dbgs() << "Ignoring CFString: " << *G); + return false; + } + } + + return true; +} + // This function replaces all global variables with new variables that have // trailing redzones. It also creates a function that poisons // redzones and inserts this function into llvm.global_ctors. bool AddressSanitizer::insertGlobalRedzones(Module &M) { SmallVector<GlobalVariable *, 16> GlobalsToChange; - for (Module::GlobalListType::iterator G = M.getGlobalList().begin(), - E = M.getGlobalList().end(); G != E; ++G) { - Type *Ty = cast<PointerType>(G->getType())->getElementType(); - DEBUG(dbgs() << "GLOBAL: " << *G); - - if (!Ty->isSized()) continue; - if (!G->hasInitializer()) continue; - // Touch only those globals that will not be defined in other modules. - // Don't handle ODR type linkages since other modules may be built w/o asan. - if (G->getLinkage() != GlobalVariable::ExternalLinkage && - G->getLinkage() != GlobalVariable::PrivateLinkage && - G->getLinkage() != GlobalVariable::InternalLinkage) - continue; - // Two problems with thread-locals: - // - The address of the main thread's copy can't be computed at link-time. - // - Need to poison all copies, not just the main thread's one. - if (G->isThreadLocal()) - continue; - // For now, just ignore this Alloca if the alignment is large. - if (G->getAlignment() > RedzoneSize) continue; - - // Ignore all the globals with the names starting with "\01L_OBJC_". - // Many of those are put into the .cstring section. The linker compresses - // that section by removing the spare \0s after the string terminator, so - // our redzones get broken. - if ((G->getName().find("\01L_OBJC_") == 0) || - (G->getName().find("\01l_OBJC_") == 0)) { - DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G); - continue; - } - - if (G->hasSection()) { - StringRef Section(G->getSection()); - // Ignore the globals from the __OBJC section. The ObjC runtime assumes - // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to - // them. - if ((Section.find("__OBJC,") == 0) || - (Section.find("__DATA, __objc_") == 0)) { - DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G); - continue; - } - // See http://code.google.com/p/address-sanitizer/issues/detail?id=32 - // Constant CFString instances are compiled in the following way: - // -- the string buffer is emitted into - // __TEXT,__cstring,cstring_literals - // -- the constant NSConstantString structure referencing that buffer - // is placed into __DATA,__cfstring - // Therefore there's no point in placing redzones into __DATA,__cfstring. - // Moreover, it causes the linker to crash on OS X 10.7 - if (Section.find("__DATA,__cfstring") == 0) { - DEBUG(dbgs() << "Ignoring CFString: " << *G); - continue; - } - } - - GlobalsToChange.push_back(G); + for (Module::GlobalListType::iterator G = M.global_begin(), + E = M.global_end(); G != E; ++G) { + if (ShouldInstrumentGlobal(G)) + GlobalsToChange.push_back(G); } size_t n = GlobalsToChange.size(); @@ -534,13 +581,22 @@ bool AddressSanitizer::insertGlobalRedzones(Module &M) { // size_t size; // size_t size_with_redzone; // const char *name; + // size_t has_dynamic_init; // We initialize an array of such structures and pass it to a run-time call. StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy, - IntptrTy, IntptrTy, NULL); - SmallVector<Constant *, 16> Initializers(n); + IntptrTy, IntptrTy, + IntptrTy, NULL); + SmallVector<Constant *, 16> Initializers(n), DynamicInit; IRBuilder<> IRB(CtorInsertBefore); + if (ClInitializers) + FindDynamicInitializers(M); + + // The addresses of the first and last dynamically initialized globals in + // this TU. Used in initialization order checking. + Value *FirstDynamic = 0, *LastDynamic = 0; + for (size_t i = 0; i < n; i++) { GlobalVariable *G = GlobalsToChange[i]; PointerType *PtrTy = cast<PointerType>(G->getType()); @@ -549,6 +605,10 @@ bool AddressSanitizer::insertGlobalRedzones(Module &M) { uint64_t RightRedzoneSize = RedzoneSize + (RedzoneSize - (SizeInBytes % RedzoneSize)); Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize); + // Determine whether this global should be poisoned in initialization. + bool GlobalHasDynamicInitializer = HasDynamicInitializer(G); + // Don't check initialization order if this global is blacklisted. + GlobalHasDynamicInitializer &= !BL->isInInit(*G); StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL); Constant *NewInitializer = ConstantStruct::get( @@ -583,8 +643,17 @@ bool AddressSanitizer::insertGlobalRedzones(Module &M) { ConstantInt::get(IntptrTy, SizeInBytes), ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize), ConstantExpr::getPointerCast(Name, IntptrTy), + ConstantInt::get(IntptrTy, GlobalHasDynamicInitializer), NULL); - DEBUG(dbgs() << "NEW GLOBAL:\n" << *NewGlobal); + + // Populate the first and last globals declared in this TU. + if (ClInitializers && GlobalHasDynamicInitializer) { + LastDynamic = ConstantExpr::getPointerCast(NewGlobal, IntptrTy); + if (FirstDynamic == 0) + FirstDynamic = LastDynamic; + } + + DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n"); } ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n); @@ -592,8 +661,13 @@ bool AddressSanitizer::insertGlobalRedzones(Module &M) { M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage, ConstantArray::get(ArrayOfGlobalStructTy, Initializers), ""); + // Create calls for poisoning before initializers run and unpoisoning after. + if (ClInitializers && FirstDynamic && LastDynamic) + createInitializerPoisonCalls(M, FirstDynamic, LastDynamic); + Function *AsanRegisterGlobals = checkInterfaceFunction(M.getOrInsertFunction( - kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL)); + kAsanRegisterGlobalsName, IRB.getVoidTy(), + IntptrTy, IntptrTy, NULL)); AsanRegisterGlobals->setLinkage(Function::ExternalLinkage); IRB.CreateCall2(AsanRegisterGlobals, @@ -623,12 +697,13 @@ bool AddressSanitizer::insertGlobalRedzones(Module &M) { } // virtual -bool AddressSanitizer::runOnModule(Module &M) { +bool AddressSanitizer::doInitialization(Module &M) { // Initialize the private fields. No one has accessed them before. - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); + if (!TD) return false; - BL.reset(new FunctionBlackList(ClBlackListFile)); + BL.reset(new BlackList(ClBlackListFile)); C = &(M.getContext()); LongSize = TD->getPointerSizeInBits(); @@ -656,17 +731,27 @@ bool AddressSanitizer::runOnModule(Module &M) { std::string FunctionName = std::string(kAsanReportErrorTemplate) + (AccessIsWrite ? "store" : "load") + itostr(1 << AccessSizeIndex); // If we are merging crash callbacks, they have two parameters. - AsanErrorCallback[AccessIsWrite][AccessSizeIndex] = cast<Function>( - M.getOrInsertFunction(FunctionName, IRB.getVoidTy(), IntptrTy, NULL)); + AsanErrorCallback[AccessIsWrite][AccessSizeIndex] = + checkInterfaceFunction(M.getOrInsertFunction( + FunctionName, IRB.getVoidTy(), IntptrTy, NULL)); } } + + AsanStackMallocFunc = checkInterfaceFunction(M.getOrInsertFunction( + kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL)); + AsanStackFreeFunc = checkInterfaceFunction(M.getOrInsertFunction( + kAsanStackFreeName, IRB.getVoidTy(), + IntptrTy, IntptrTy, IntptrTy, NULL)); + AsanHandleNoReturnFunc = checkInterfaceFunction(M.getOrInsertFunction( + kAsanHandleNoReturnName, IRB.getVoidTy(), NULL)); + // We insert an empty inline asm after __asan_report* to avoid callback merge. EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false), StringRef(""), StringRef(""), /*hasSideEffects=*/true); llvm::Triple targetTriple(M.getTargetTriple()); - bool isAndroid = targetTriple.getEnvironment() == llvm::Triple::ANDROIDEABI; + bool isAndroid = targetTriple.getEnvironment() == llvm::Triple::Android; MappingOffset = isAndroid ? kDefaultShadowOffsetAndroid : (LongSize == 32 ? kDefaultShadowOffset32 : kDefaultShadowOffset64); @@ -686,10 +771,6 @@ bool AddressSanitizer::runOnModule(Module &M) { // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively. RedzoneSize = std::max(32, (int)(1 << MappingScale)); - bool Res = false; - - if (ClGlobals) - Res |= insertGlobalRedzones(M); if (ClMappingOffsetLog >= 0) { // Tell the run-time the current values of mapping offset and scale. @@ -709,17 +790,20 @@ bool AddressSanitizer::runOnModule(Module &M) { IRB.CreateLoad(asan_mapping_scale, true); } - - for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) { - if (F->isDeclaration()) continue; - Res |= handleFunction(M, *F); - } - appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority); - return Res; + return true; +} + +bool AddressSanitizer::doFinalization(Module &M) { + // We transform the globals at the very end so that the optimization analysis + // works on the original globals. + if (ClGlobals) + return insertGlobalRedzones(M); + return false; } + bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) { // For each NSObject descendant having a +load method, this method is invoked // by the ObjC runtime before any of the static constructors is called. @@ -736,19 +820,22 @@ bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) { return false; } -bool AddressSanitizer::handleFunction(Module &M, Function &F) { +bool AddressSanitizer::runOnFunction(Function &F) { if (BL->isIn(F)) return false; if (&F == AsanCtorFunction) return false; + DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n"); // If needed, insert __asan_init before checking for AddressSafety attr. maybeInsertAsanInitAtFunctionEntry(F); - if (!F.hasFnAttr(Attribute::AddressSafety)) return false; + if (!F.getFnAttributes().hasAttribute(Attributes::AddressSafety)) + return false; if (!ClDebugFunc.empty() && ClDebugFunc != F.getName()) return false; - // We want to instrument every address only once per basic block - // (unless there are calls between uses). + + // We want to instrument every address only once per basic block (unless there + // are calls between uses). SmallSet<Value*, 16> TempsToInstrument; SmallVector<Instruction*, 16> ToInstrument; SmallVector<Instruction*, 8> NoReturnCalls; @@ -786,8 +873,6 @@ bool AddressSanitizer::handleFunction(Module &M, Function &F) { } } - AsanFunctionContext AFC(F); - // Instrument. int NumInstrumented = 0; for (size_t i = 0, n = ToInstrument.size(); i != n; i++) { @@ -795,25 +880,23 @@ bool AddressSanitizer::handleFunction(Module &M, Function &F) { if (ClDebugMin < 0 || ClDebugMax < 0 || (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) { if (isInterestingMemoryAccess(Inst, &IsWrite)) - instrumentMop(AFC, Inst); + instrumentMop(Inst); else - instrumentMemIntrinsic(AFC, cast<MemIntrinsic>(Inst)); + instrumentMemIntrinsic(cast<MemIntrinsic>(Inst)); } NumInstrumented++; } - DEBUG(dbgs() << F); - - bool ChangedStack = poisonStackInFunction(M, F); + bool ChangedStack = poisonStackInFunction(F); // We must unpoison the stack before every NoReturn call (throw, _exit, etc). // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37 for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) { Instruction *CI = NoReturnCalls[i]; IRBuilder<> IRB(CI); - IRB.CreateCall(M.getOrInsertFunction(kAsanHandleNoReturnName, - IRB.getVoidTy(), NULL)); + IRB.CreateCall(AsanHandleNoReturnFunc); } + DEBUG(dbgs() << "ASAN done instrumenting:\n" << F << "\n"); return NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty(); } @@ -926,7 +1009,7 @@ bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) { // compiler hoists the load of the shadow value somewhere too high. // This causes asan to report a non-existing bug on 453.povray. // It sounds like an LLVM bug. -bool AddressSanitizer::poisonStackInFunction(Module &M, Function &F) { +bool AddressSanitizer::poisonStackInFunction(Function &F) { if (!ClStack) return false; SmallVector<AllocaInst*, 16> AllocaVec; SmallVector<Instruction*, 8> RetVec; @@ -976,8 +1059,6 @@ bool AddressSanitizer::poisonStackInFunction(Module &M, Function &F) { Value *LocalStackBase = OrigStackBase; if (DoStackMalloc) { - Value *AsanStackMallocFunc = M.getOrInsertFunction( - kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL); LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc, ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase); } @@ -1012,22 +1093,16 @@ bool AddressSanitizer::poisonStackInFunction(Module &M, Function &F) { Value *BasePlus1 = IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, LongSize/8)); BasePlus1 = IRB.CreateIntToPtr(BasePlus1, IntptrPtrTy); - Value *Description = IRB.CreatePointerCast( - createPrivateGlobalForString(M, StackDescription.str()), - IntptrTy); + GlobalVariable *StackDescriptionGlobal = + createPrivateGlobalForString(*F.getParent(), StackDescription.str()); + GlobalsCreatedByAsan.insert(StackDescriptionGlobal); + Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal, IntptrTy); IRB.CreateStore(Description, BasePlus1); // Poison the stack redzones at the entry. Value *ShadowBase = memToShadow(LocalStackBase, IRB); PoisonStack(ArrayRef<AllocaInst*>(AllocaVec), IRB, ShadowBase, true); - Value *AsanStackFreeFunc = NULL; - if (DoStackMalloc) { - AsanStackFreeFunc = M.getOrInsertFunction( - kAsanStackFreeName, IRB.getVoidTy(), - IntptrTy, IntptrTy, IntptrTy, NULL); - } - // Unpoison the stack before all ret instructions. for (size_t i = 0, n = RetVec.size(); i < n; i++) { Instruction *Ret = RetVec[i]; @@ -1046,6 +1121,10 @@ bool AddressSanitizer::poisonStackInFunction(Module &M, Function &F) { } } + // We are done. Remove the old unused alloca instructions. + for (size_t i = 0, n = AllocaVec.size(); i < n; i++) + AllocaVec[i]->eraseFromParent(); + if (ClDebugStack) { DEBUG(dbgs() << F); } diff --git a/contrib/llvm/lib/Transforms/Instrumentation/BlackList.cpp b/contrib/llvm/lib/Transforms/Instrumentation/BlackList.cpp new file mode 100644 index 0000000..ef34b8a --- /dev/null +++ b/contrib/llvm/lib/Transforms/Instrumentation/BlackList.cpp @@ -0,0 +1,105 @@ +//===-- BlackList.cpp - blacklist for sanitizers --------------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This is a utility class for instrumentation passes (like AddressSanitizer +// or ThreadSanitizer) to avoid instrumenting some functions or global +// variables based on a user-supplied blacklist. +// +//===----------------------------------------------------------------------===// + +#include <utility> +#include <string> + +#include "BlackList.h" +#include "llvm/ADT/OwningPtr.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/StringExtras.h" +#include "llvm/Function.h" +#include "llvm/GlobalVariable.h" +#include "llvm/Module.h" +#include "llvm/Support/MemoryBuffer.h" +#include "llvm/Support/Regex.h" +#include "llvm/Support/raw_ostream.h" +#include "llvm/Support/system_error.h" + +namespace llvm { + +BlackList::BlackList(const StringRef Path) { + // Validate and open blacklist file. + if (!Path.size()) return; + OwningPtr<MemoryBuffer> File; + if (error_code EC = MemoryBuffer::getFile(Path, File)) { + report_fatal_error("Can't open blacklist file: " + Path + ": " + + EC.message()); + } + + // Iterate through each line in the blacklist file. + SmallVector<StringRef, 16> Lines; + SplitString(File.take()->getBuffer(), Lines, "\n\r"); + StringMap<std::string> Regexps; + for (SmallVector<StringRef, 16>::iterator I = Lines.begin(), E = Lines.end(); + I != E; ++I) { + // Ignore empty lines and lines starting with "#" + if (I->empty() || I->startswith("#")) + continue; + // Get our prefix and unparsed regexp. + std::pair<StringRef, StringRef> SplitLine = I->split(":"); + StringRef Prefix = SplitLine.first; + std::string Regexp = SplitLine.second; + + // Replace * with .* + for (size_t pos = 0; (pos = Regexp.find("*", pos)) != std::string::npos; + pos += strlen(".*")) { + Regexp.replace(pos, strlen("*"), ".*"); + } + + // Check that the regexp is valid. + Regex CheckRE(Regexp); + std::string Error; + if (!CheckRE.isValid(Error)) { + report_fatal_error("malformed blacklist regex: " + SplitLine.second + + ": " + Error); + } + + // Add this regexp into the proper group by its prefix. + if (Regexps[Prefix].size()) + Regexps[Prefix] += "|"; + Regexps[Prefix] += Regexp; + } + + // Iterate through each of the prefixes, and create Regexs for them. + for (StringMap<std::string>::iterator I = Regexps.begin(), E = Regexps.end(); + I != E; ++I) { + Entries[I->getKey()] = new Regex(I->getValue()); + } +} + +bool BlackList::isIn(const Function &F) { + return isIn(*F.getParent()) || inSection("fun", F.getName()); +} + +bool BlackList::isIn(const GlobalVariable &G) { + return isIn(*G.getParent()) || inSection("global", G.getName()); +} + +bool BlackList::isIn(const Module &M) { + return inSection("src", M.getModuleIdentifier()); +} + +bool BlackList::isInInit(const GlobalVariable &G) { + return isIn(*G.getParent()) || inSection("global-init", G.getName()); +} + +bool BlackList::inSection(const StringRef Section, + const StringRef Query) { + Regex *FunctionRegex = Entries[Section]; + return FunctionRegex ? FunctionRegex->match(Query) : false; +} + +} // namespace llvm diff --git a/contrib/llvm/lib/Transforms/Instrumentation/BlackList.h b/contrib/llvm/lib/Transforms/Instrumentation/BlackList.h new file mode 100644 index 0000000..f3c05a5 --- /dev/null +++ b/contrib/llvm/lib/Transforms/Instrumentation/BlackList.h @@ -0,0 +1,57 @@ +//===-- BlackList.h - blacklist for sanitizers ------------------*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +//===----------------------------------------------------------------------===// +// +// This is a utility class for instrumentation passes (like AddressSanitizer +// or ThreadSanitizer) to avoid instrumenting some functions or global +// variables based on a user-supplied blacklist. +// +// The blacklist disables instrumentation of various functions and global +// variables. Each line contains a prefix, followed by a wild card expression. +// Empty lines and lines starting with "#" are ignored. +// --- +// # Blacklisted items: +// fun:*_ZN4base6subtle* +// global:*global_with_bad_access_or_initialization* +// global-init:*global_with_initialization_issues* +// src:file_with_tricky_code.cc +// --- +// Note that the wild card is in fact an llvm::Regex, but * is automatically +// replaced with .* +// This is similar to the "ignore" feature of ThreadSanitizer. +// http://code.google.com/p/data-race-test/wiki/ThreadSanitizerIgnores +// +//===----------------------------------------------------------------------===// +// + +#include "llvm/ADT/StringMap.h" + +namespace llvm { +class Function; +class GlobalVariable; +class Module; +class Regex; +class StringRef; + +class BlackList { + public: + BlackList(const StringRef Path); + // Returns whether either this function or it's source file are blacklisted. + bool isIn(const Function &F); + // Returns whether either this global or it's source file are blacklisted. + bool isIn(const GlobalVariable &G); + // Returns whether this module is blacklisted by filename. + bool isIn(const Module &M); + // Returns whether a global should be excluded from initialization checking. + bool isInInit(const GlobalVariable &G); + private: + StringMap<Regex*> Entries; + + bool inSection(const StringRef Section, const StringRef Query); +}; + +} // namespace llvm diff --git a/contrib/llvm/lib/Transforms/Instrumentation/BoundsChecking.cpp b/contrib/llvm/lib/Transforms/Instrumentation/BoundsChecking.cpp index 09e0f14..7810b1b 100644 --- a/contrib/llvm/lib/Transforms/Instrumentation/BoundsChecking.cpp +++ b/contrib/llvm/lib/Transforms/Instrumentation/BoundsChecking.cpp @@ -23,7 +23,8 @@ #include "llvm/Support/InstIterator.h" #include "llvm/Support/TargetFolder.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Instrumentation.h" using namespace llvm; @@ -47,11 +48,13 @@ namespace { virtual bool runOnFunction(Function &F); virtual void getAnalysisUsage(AnalysisUsage &AU) const { - AU.addRequired<TargetData>(); + AU.addRequired<DataLayout>(); + AU.addRequired<TargetLibraryInfo>(); } private: - const TargetData *TD; + const DataLayout *TD; + const TargetLibraryInfo *TLI; ObjectSizeOffsetEvaluator *ObjSizeEval; BuilderTy *Builder; Instruction *Inst; @@ -140,7 +143,7 @@ bool BoundsChecking::instrument(Value *Ptr, Value *InstVal) { Value *Offset = SizeOffset.second; ConstantInt *SizeCI = dyn_cast<ConstantInt>(Size); - IntegerType *IntTy = TD->getIntPtrType(Inst->getContext()); + Type *IntTy = TD->getIntPtrType(Ptr->getType()); Value *NeededSizeVal = ConstantInt::get(IntTy, NeededSize); // three checks are required to ensure safety: @@ -165,12 +168,13 @@ bool BoundsChecking::instrument(Value *Ptr, Value *InstVal) { } bool BoundsChecking::runOnFunction(Function &F) { - TD = &getAnalysis<TargetData>(); + TD = &getAnalysis<DataLayout>(); + TLI = &getAnalysis<TargetLibraryInfo>(); TrapBB = 0; BuilderTy TheBuilder(F.getContext(), TargetFolder(TD)); Builder = &TheBuilder; - ObjectSizeOffsetEvaluator TheObjSizeEval(TD, F.getContext()); + ObjectSizeOffsetEvaluator TheObjSizeEval(TD, TLI, F.getContext()); ObjSizeEval = &TheObjSizeEval; // check HANDLE_MEMORY_INST in include/llvm/Instruction.def for memory diff --git a/contrib/llvm/lib/Transforms/Instrumentation/FunctionBlackList.cpp b/contrib/llvm/lib/Transforms/Instrumentation/FunctionBlackList.cpp deleted file mode 100644 index 188ea4d..0000000 --- a/contrib/llvm/lib/Transforms/Instrumentation/FunctionBlackList.cpp +++ /dev/null @@ -1,79 +0,0 @@ -//===-- FunctionBlackList.cpp - blacklist of functions --------------------===// -// -// The LLVM Compiler Infrastructure -// -// This file is distributed under the University of Illinois Open Source -// License. See LICENSE.TXT for details. -// -//===----------------------------------------------------------------------===// -// -// This is a utility class for instrumentation passes (like AddressSanitizer -// or ThreadSanitizer) to avoid instrumenting some functions based on -// user-supplied blacklist. -// -//===----------------------------------------------------------------------===// - -#include "FunctionBlackList.h" -#include "llvm/ADT/OwningPtr.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/StringExtras.h" -#include "llvm/Function.h" -#include "llvm/Support/MemoryBuffer.h" -#include "llvm/Support/Regex.h" -#include "llvm/Support/raw_ostream.h" -#include "llvm/Support/system_error.h" - -namespace llvm { - -FunctionBlackList::FunctionBlackList(const std::string &Path) { - Functions = NULL; - const char *kFunPrefix = "fun:"; - if (!Path.size()) return; - std::string Fun; - - OwningPtr<MemoryBuffer> File; - if (error_code EC = MemoryBuffer::getFile(Path.c_str(), File)) { - report_fatal_error("Can't open blacklist file " + Path + ": " + - EC.message()); - } - MemoryBuffer *Buff = File.take(); - const char *Data = Buff->getBufferStart(); - size_t DataLen = Buff->getBufferSize(); - SmallVector<StringRef, 16> Lines; - SplitString(StringRef(Data, DataLen), Lines, "\n\r"); - for (size_t i = 0, numLines = Lines.size(); i < numLines; i++) { - if (Lines[i].startswith(kFunPrefix)) { - std::string ThisFunc = Lines[i].substr(strlen(kFunPrefix)); - std::string ThisFuncRE; - // add ThisFunc replacing * with .* - for (size_t j = 0, n = ThisFunc.size(); j < n; j++) { - if (ThisFunc[j] == '*') - ThisFuncRE += '.'; - ThisFuncRE += ThisFunc[j]; - } - // Check that the regexp is valid. - Regex CheckRE(ThisFuncRE); - std::string Error; - if (!CheckRE.isValid(Error)) - report_fatal_error("malformed blacklist regex: " + ThisFunc + - ": " + Error); - // Append to the final regexp. - if (Fun.size()) - Fun += "|"; - Fun += ThisFuncRE; - } - } - if (Fun.size()) { - Functions = new Regex(Fun); - } -} - -bool FunctionBlackList::isIn(const Function &F) { - if (Functions) { - bool Res = Functions->match(F.getName()); - return Res; - } - return false; -} - -} // namespace llvm diff --git a/contrib/llvm/lib/Transforms/Instrumentation/FunctionBlackList.h b/contrib/llvm/lib/Transforms/Instrumentation/FunctionBlackList.h deleted file mode 100644 index c1239b9..0000000 --- a/contrib/llvm/lib/Transforms/Instrumentation/FunctionBlackList.h +++ /dev/null @@ -1,37 +0,0 @@ -//===-- FunctionBlackList.cpp - blacklist of functions ----------*- C++ -*-===// -// -// The LLVM Compiler Infrastructure -// -// This file is distributed under the University of Illinois Open Source -// License. See LICENSE.TXT for details. -//===----------------------------------------------------------------------===// -// -// This is a utility class for instrumentation passes (like AddressSanitizer -// or ThreadSanitizer) to avoid instrumenting some functions based on -// user-supplied blacklist. -// -//===----------------------------------------------------------------------===// -// - -#include <string> - -namespace llvm { -class Function; -class Regex; - -// Blacklisted functions are not instrumented. -// The blacklist file contains one or more lines like this: -// --- -// fun:FunctionWildCard -// --- -// This is similar to the "ignore" feature of ThreadSanitizer. -// http://code.google.com/p/data-race-test/wiki/ThreadSanitizerIgnores -class FunctionBlackList { - public: - FunctionBlackList(const std::string &Path); - bool isIn(const Function &F); - private: - Regex *Functions; -}; - -} // namespace llvm diff --git a/contrib/llvm/lib/Transforms/Instrumentation/GCOVProfiling.cpp b/contrib/llvm/lib/Transforms/Instrumentation/GCOVProfiling.cpp index 264a6a6..e9192e5 100644 --- a/contrib/llvm/lib/Transforms/Instrumentation/GCOVProfiling.cpp +++ b/contrib/llvm/lib/Transforms/Instrumentation/GCOVProfiling.cpp @@ -88,11 +88,11 @@ namespace { // Add the function to write out all our counters to the global destructor // list. - void insertCounterWriteout(SmallVector<std::pair<GlobalVariable *, - MDNode *>, 8> &); + void insertCounterWriteout(ArrayRef<std::pair<GlobalVariable*, MDNode*> >); void insertIndirectCounterIncrement(); + void insertFlush(ArrayRef<std::pair<GlobalVariable*, MDNode*> >); - std::string mangleName(DICompileUnit CU, std::string NewStem); + std::string mangleName(DICompileUnit CU, const char *NewStem); bool EmitNotes; bool EmitData; @@ -329,7 +329,7 @@ namespace { }; } -std::string GCOVProfiler::mangleName(DICompileUnit CU, std::string NewStem) { +std::string GCOVProfiler::mangleName(DICompileUnit CU, const char *NewStem) { if (NamedMDNode *GCov = M->getNamedMetadata("llvm.gcov")) { for (int i = 0, e = GCov->getNumOperands(); i != e; ++i) { MDNode *N = GCov->getOperand(i); @@ -519,6 +519,7 @@ bool GCOVProfiler::emitProfileArcs() { } insertCounterWriteout(CountersBySP); + insertFlush(CountersBySP); } if (InsertIndCounterIncrCode) @@ -630,14 +631,15 @@ GlobalVariable *GCOVProfiler::getEdgeStateValue() { } void GCOVProfiler::insertCounterWriteout( - SmallVector<std::pair<GlobalVariable *, MDNode *>, 8> &CountersBySP) { - FunctionType *WriteoutFTy = - FunctionType::get(Type::getVoidTy(*Ctx), false); - Function *WriteoutF = Function::Create(WriteoutFTy, - GlobalValue::InternalLinkage, - "__llvm_gcov_writeout", M); + ArrayRef<std::pair<GlobalVariable *, MDNode *> > CountersBySP) { + FunctionType *WriteoutFTy = FunctionType::get(Type::getVoidTy(*Ctx), false); + Function *WriteoutF = M->getFunction("__llvm_gcov_writeout"); + if (!WriteoutF) + WriteoutF = Function::Create(WriteoutFTy, GlobalValue::InternalLinkage, + "__llvm_gcov_writeout", M); WriteoutF->setUnnamedAddr(true); - BasicBlock *BB = BasicBlock::Create(*Ctx, "", WriteoutF); + + BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", WriteoutF); IRBuilder<> Builder(BB); Constant *StartFile = getStartFileFunc(); @@ -648,11 +650,11 @@ void GCOVProfiler::insertCounterWriteout( NamedMDNode *CU_Nodes = M->getNamedMetadata("llvm.dbg.cu"); if (CU_Nodes) { for (unsigned i = 0, e = CU_Nodes->getNumOperands(); i != e; ++i) { - DICompileUnit compile_unit(CU_Nodes->getOperand(i)); - std::string FilenameGcda = mangleName(compile_unit, "gcda"); + DICompileUnit CU(CU_Nodes->getOperand(i)); + std::string FilenameGcda = mangleName(CU, "gcda"); Builder.CreateCall(StartFile, Builder.CreateGlobalStringPtr(FilenameGcda)); - for (SmallVector<std::pair<GlobalVariable *, MDNode *>, 8>::iterator + for (ArrayRef<std::pair<GlobalVariable *, MDNode *> >::iterator I = CountersBySP.begin(), E = CountersBySP.end(); I != E; ++I) { DISubprogram SP(I->second); @@ -680,7 +682,7 @@ void GCOVProfiler::insertCounterWriteout( "__llvm_gcov_init", M); F->setUnnamedAddr(true); F->setLinkage(GlobalValue::InternalLinkage); - F->addFnAttr(Attribute::NoInline); + F->addFnAttr(Attributes::NoInline); BB = BasicBlock::Create(*Ctx, "entry", F); Builder.SetInsertPoint(BB); @@ -699,7 +701,7 @@ void GCOVProfiler::insertIndirectCounterIncrement() { cast<Function>(GCOVProfiler::getIncrementIndirectCounterFunc()); Fn->setUnnamedAddr(true); Fn->setLinkage(GlobalValue::InternalLinkage); - Fn->addFnAttr(Attribute::NoInline); + Fn->addFnAttr(Attributes::NoInline); Type *Int32Ty = Type::getInt32Ty(*Ctx); Type *Int64Ty = Type::getInt64Ty(*Ctx); @@ -745,3 +747,42 @@ void GCOVProfiler::insertIndirectCounterIncrement() { Builder.SetInsertPoint(Exit); Builder.CreateRetVoid(); } + +void GCOVProfiler:: +insertFlush(ArrayRef<std::pair<GlobalVariable*, MDNode*> > CountersBySP) { + FunctionType *FTy = FunctionType::get(Type::getVoidTy(*Ctx), false); + Function *FlushF = M->getFunction("__gcov_flush"); + if (!FlushF) + FlushF = Function::Create(FTy, GlobalValue::InternalLinkage, + "__gcov_flush", M); + else + FlushF->setLinkage(GlobalValue::InternalLinkage); + FlushF->setUnnamedAddr(true); + + BasicBlock *Entry = BasicBlock::Create(*Ctx, "entry", FlushF); + + // Write out the current counters. + Constant *WriteoutF = M->getFunction("__llvm_gcov_writeout"); + assert(WriteoutF && "Need to create the writeout function first!"); + + IRBuilder<> Builder(Entry); + Builder.CreateCall(WriteoutF); + + // Zero out the counters. + for (ArrayRef<std::pair<GlobalVariable *, MDNode *> >::iterator + I = CountersBySP.begin(), E = CountersBySP.end(); + I != E; ++I) { + GlobalVariable *GV = I->first; + Constant *Null = Constant::getNullValue(GV->getType()->getElementType()); + Builder.CreateStore(Null, GV); + } + + Type *RetTy = FlushF->getReturnType(); + if (RetTy == Type::getVoidTy(*Ctx)) + Builder.CreateRetVoid(); + else if (RetTy->isIntegerTy()) + // Used if __gcov_flush was implicitly declared. + Builder.CreateRet(ConstantInt::get(RetTy, 0)); + else + report_fatal_error("invalid return type for __gcov_flush"); +} diff --git a/contrib/llvm/lib/Transforms/Instrumentation/MaximumSpanningTree.h b/contrib/llvm/lib/Transforms/Instrumentation/MaximumSpanningTree.h index f76c77e..a4bb5a6 100644 --- a/contrib/llvm/lib/Transforms/Instrumentation/MaximumSpanningTree.h +++ b/contrib/llvm/lib/Transforms/Instrumentation/MaximumSpanningTree.h @@ -26,30 +26,6 @@ namespace llvm { /// The type parameter T determines the type of the nodes of the graph. template <typename T> class MaximumSpanningTree { - - // A comparing class for comparing weighted edges. - template <typename CT> - struct EdgeWeightCompare { - bool operator()(typename MaximumSpanningTree<CT>::EdgeWeight X, - typename MaximumSpanningTree<CT>::EdgeWeight Y) const { - if (X.second > Y.second) return true; - if (X.second < Y.second) return false; - if (const BasicBlock *BBX = dyn_cast<BasicBlock>(X.first.first)) { - if (const BasicBlock *BBY = dyn_cast<BasicBlock>(Y.first.first)) { - if (BBX->size() > BBY->size()) return true; - if (BBX->size() < BBY->size()) return false; - } - } - if (const BasicBlock *BBX = dyn_cast<BasicBlock>(X.first.second)) { - if (const BasicBlock *BBY = dyn_cast<BasicBlock>(Y.first.second)) { - if (BBX->size() > BBY->size()) return true; - if (BBX->size() < BBY->size()) return false; - } - } - return false; - } - }; - public: typedef std::pair<const T*, const T*> Edge; typedef std::pair<Edge, double> EdgeWeight; @@ -59,6 +35,33 @@ namespace llvm { MaxSpanTree MST; + private: + // A comparing class for comparing weighted edges. + struct EdgeWeightCompare { + static bool getBlockSize(const T *X) { + const BasicBlock *BB = dyn_cast_or_null<BasicBlock>(X); + return BB ? BB->size() : 0; + } + + bool operator()(EdgeWeight X, EdgeWeight Y) const { + if (X.second > Y.second) return true; + if (X.second < Y.second) return false; + + // Equal edge weights: break ties by comparing block sizes. + size_t XSizeA = getBlockSize(X.first.first); + size_t YSizeA = getBlockSize(Y.first.first); + if (XSizeA > YSizeA) return true; + if (XSizeA < YSizeA) return false; + + size_t XSizeB = getBlockSize(X.first.second); + size_t YSizeB = getBlockSize(Y.first.second); + if (XSizeB > YSizeB) return true; + if (XSizeB < YSizeB) return false; + + return false; + } + }; + public: static char ID; // Class identification, replacement for typeinfo @@ -66,7 +69,7 @@ namespace llvm { /// spanning tree. MaximumSpanningTree(EdgeWeights &EdgeVector) { - std::stable_sort(EdgeVector.begin(), EdgeVector.end(), EdgeWeightCompare<T>()); + std::stable_sort(EdgeVector.begin(), EdgeVector.end(), EdgeWeightCompare()); // Create spanning tree, Forest contains a special data structure // that makes checking if two nodes are already in a common (sub-)tree diff --git a/contrib/llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp b/contrib/llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp index dc0fa71..9e10fc4 100644 --- a/contrib/llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp +++ b/contrib/llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp @@ -21,7 +21,7 @@ #define DEBUG_TYPE "tsan" -#include "FunctionBlackList.h" +#include "BlackList.h" #include "llvm/Function.h" #include "llvm/IRBuilder.h" #include "llvm/Intrinsics.h" @@ -38,7 +38,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Transforms/Instrumentation.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/ModuleUtils.h" @@ -47,10 +47,19 @@ using namespace llvm; static cl::opt<std::string> ClBlackListFile("tsan-blacklist", cl::desc("Blacklist file"), cl::Hidden); +static cl::opt<bool> ClInstrumentMemoryAccesses( + "tsan-instrument-memory-accesses", cl::init(true), + cl::desc("Instrument memory accesses"), cl::Hidden); +static cl::opt<bool> ClInstrumentFuncEntryExit( + "tsan-instrument-func-entry-exit", cl::init(true), + cl::desc("Instrument function entry and exit"), cl::Hidden); +static cl::opt<bool> ClInstrumentAtomics( + "tsan-instrument-atomics", cl::init(true), + cl::desc("Instrument atomics"), cl::Hidden); STATISTIC(NumInstrumentedReads, "Number of instrumented reads"); STATISTIC(NumInstrumentedWrites, "Number of instrumented writes"); -STATISTIC(NumOmittedReadsBeforeWrite, +STATISTIC(NumOmittedReadsBeforeWrite, "Number of reads ignored due to following writes"); STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size"); STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes"); @@ -76,8 +85,8 @@ struct ThreadSanitizer : public FunctionPass { bool addrPointsToConstantData(Value *Addr); int getMemoryAccessFuncIndex(Value *Addr); - TargetData *TD; - OwningPtr<FunctionBlackList> BL; + DataLayout *TD; + OwningPtr<BlackList> BL; IntegerType *OrdTy; // Callbacks to run-time library are computed in doInitialization. Function *TsanFuncEntry; @@ -88,6 +97,10 @@ struct ThreadSanitizer : public FunctionPass { Function *TsanWrite[kNumberOfAccessSizes]; Function *TsanAtomicLoad[kNumberOfAccessSizes]; Function *TsanAtomicStore[kNumberOfAccessSizes]; + Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes]; + Function *TsanAtomicCAS[kNumberOfAccessSizes]; + Function *TsanAtomicThreadFence; + Function *TsanAtomicSignalFence; Function *TsanVptrUpdate; }; } // namespace @@ -118,10 +131,10 @@ static Function *checkInterfaceFunction(Constant *FuncOrBitcast) { } bool ThreadSanitizer::doInitialization(Module &M) { - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); if (!TD) return false; - BL.reset(new FunctionBlackList(ClBlackListFile)); + BL.reset(new BlackList(ClBlackListFile)); // Always insert a call to __tsan_init into the module's CTORs. IRBuilder<> IRB(M.getContext()); @@ -158,10 +171,42 @@ bool ThreadSanitizer::doInitialization(Module &M) { TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction( AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy, NULL)); + + for (int op = AtomicRMWInst::FIRST_BINOP; + op <= AtomicRMWInst::LAST_BINOP; ++op) { + TsanAtomicRMW[op][i] = NULL; + const char *NamePart = NULL; + if (op == AtomicRMWInst::Xchg) + NamePart = "_exchange"; + else if (op == AtomicRMWInst::Add) + NamePart = "_fetch_add"; + else if (op == AtomicRMWInst::Sub) + NamePart = "_fetch_sub"; + else if (op == AtomicRMWInst::And) + NamePart = "_fetch_and"; + else if (op == AtomicRMWInst::Or) + NamePart = "_fetch_or"; + else if (op == AtomicRMWInst::Xor) + NamePart = "_fetch_xor"; + else + continue; + SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart); + TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction( + RMWName, Ty, PtrTy, Ty, OrdTy, NULL)); + } + + SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) + + "_compare_exchange_val"); + TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction( + AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, NULL)); } TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction( "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), NULL)); + TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction( + "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, NULL)); + TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction( + "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, NULL)); return true; } @@ -186,7 +231,7 @@ bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) { NumOmittedReadsFromConstantGlobals++; return true; } - } else if(LoadInst *L = dyn_cast<LoadInst>(Addr)) { + } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) { if (isVtableAccess(L)) { // Reads from a vtable pointer can not race with any writes. NumOmittedReadsFromVtable++; @@ -244,8 +289,8 @@ static bool isAtomic(Instruction *I) { return true; if (isa<AtomicCmpXchgInst>(I)) return true; - if (FenceInst *FI = dyn_cast<FenceInst>(I)) - return FI->getSynchScope() == CrossThread; + if (isa<FenceInst>(I)) + return true; return false; } @@ -284,17 +329,19 @@ bool ThreadSanitizer::runOnFunction(Function &F) { // (e.g. variables that do not escape, etc). // Instrument memory accesses. - for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) { - Res |= instrumentLoadOrStore(AllLoadsAndStores[i]); - } + if (ClInstrumentMemoryAccesses) + for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) { + Res |= instrumentLoadOrStore(AllLoadsAndStores[i]); + } // Instrument atomic memory accesses. - for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) { - Res |= instrumentAtomic(AtomicAccesses[i]); - } + if (ClInstrumentAtomics) + for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) { + Res |= instrumentAtomic(AtomicAccesses[i]); + } // Instrument function entry/exit points if there were instrumented accesses. - if (Res || HasCalls) { + if ((Res || HasCalls) && ClInstrumentFuncEntryExit) { IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI()); Value *ReturnAddress = IRB.CreateCall( Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress), @@ -343,12 +390,12 @@ static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) { switch (ord) { case NotAtomic: assert(false); case Unordered: // Fall-through. - case Monotonic: v = 1 << 0; break; - // case Consume: v = 1 << 1; break; // Not specified yet. - case Acquire: v = 1 << 2; break; - case Release: v = 1 << 3; break; - case AcquireRelease: v = 1 << 4; break; - case SequentiallyConsistent: v = 1 << 5; break; + case Monotonic: v = 0; break; + // case Consume: v = 1; break; // Not specified yet. + case Acquire: v = 2; break; + case Release: v = 3; break; + case AcquireRelease: v = 4; break; + case SequentiallyConsistent: v = 5; break; } return IRB->getInt32(v); } @@ -385,12 +432,44 @@ bool ThreadSanitizer::instrumentAtomic(Instruction *I) { CallInst *C = CallInst::Create(TsanAtomicStore[Idx], ArrayRef<Value*>(Args)); ReplaceInstWithInst(I, C); - } else if (isa<AtomicRMWInst>(I)) { - // FIXME: Not yet supported. - } else if (isa<AtomicCmpXchgInst>(I)) { - // FIXME: Not yet supported. - } else if (isa<FenceInst>(I)) { - // FIXME: Not yet supported. + } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) { + Value *Addr = RMWI->getPointerOperand(); + int Idx = getMemoryAccessFuncIndex(Addr); + if (Idx < 0) + return false; + Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx]; + if (F == NULL) + return false; + const size_t ByteSize = 1 << Idx; + const size_t BitSize = ByteSize * 8; + Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); + Type *PtrTy = Ty->getPointerTo(); + Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), + IRB.CreateIntCast(RMWI->getValOperand(), Ty, false), + createOrdering(&IRB, RMWI->getOrdering())}; + CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args)); + ReplaceInstWithInst(I, C); + } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) { + Value *Addr = CASI->getPointerOperand(); + int Idx = getMemoryAccessFuncIndex(Addr); + if (Idx < 0) + return false; + const size_t ByteSize = 1 << Idx; + const size_t BitSize = ByteSize * 8; + Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); + Type *PtrTy = Ty->getPointerTo(); + Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), + IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false), + IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false), + createOrdering(&IRB, CASI->getOrdering())}; + CallInst *C = CallInst::Create(TsanAtomicCAS[Idx], ArrayRef<Value*>(Args)); + ReplaceInstWithInst(I, C); + } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) { + Value *Args[] = {createOrdering(&IRB, FI->getOrdering())}; + Function *F = FI->getSynchScope() == SingleThread ? + TsanAtomicSignalFence : TsanAtomicThreadFence; + CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args)); + ReplaceInstWithInst(I, C); } return true; } diff --git a/contrib/llvm/lib/Transforms/Scalar/CodeGenPrepare.cpp b/contrib/llvm/lib/Transforms/Scalar/CodeGenPrepare.cpp index a3c426a..123ed0f 100644 --- a/contrib/llvm/lib/Transforms/Scalar/CodeGenPrepare.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/CodeGenPrepare.cpp @@ -27,6 +27,7 @@ #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/Analysis/Dominators.h" +#include "llvm/Analysis/DominatorInternals.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/ProfileInfo.h" #include "llvm/Assembly/Writer.h" @@ -37,12 +38,13 @@ #include "llvm/Support/PatternMatch.h" #include "llvm/Support/ValueHandle.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Target/TargetLowering.h" #include "llvm/Transforms/Utils/AddrModeMatcher.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/BuildLibCalls.h" +#include "llvm/Transforms/Utils/BypassSlowDivision.h" #include "llvm/Transforms/Utils/Local.h" using namespace llvm; using namespace llvm::PatternMatch; @@ -146,9 +148,18 @@ bool CodeGenPrepare::runOnFunction(Function &F) { TLInfo = &getAnalysis<TargetLibraryInfo>(); DT = getAnalysisIfAvailable<DominatorTree>(); PFI = getAnalysisIfAvailable<ProfileInfo>(); - OptSize = F.hasFnAttr(Attribute::OptimizeForSize); + OptSize = F.getFnAttributes().hasAttribute(Attributes::OptimizeForSize); + + /// This optimization identifies DIV instructions that can be + /// profitably bypassed and carried out with a shorter, faster divide. + if (TLI && TLI->isSlowDivBypassed()) { + const DenseMap<unsigned int, unsigned int> &BypassWidths = + TLI->getBypassSlowDivWidths(); + for (Function::iterator I = F.begin(); I != F.end(); I++) + EverMadeChange |= bypassSlowDivision(F, I, BypassWidths); + } - // First pass, eliminate blocks that contain only PHI nodes and an + // Eliminate blocks that contain only PHI nodes and an // unconditional branch. EverMadeChange |= EliminateMostlyEmptyBlocks(F); @@ -160,7 +171,7 @@ bool CodeGenPrepare::runOnFunction(Function &F) { bool MadeChange = true; while (MadeChange) { MadeChange = false; - for (Function::iterator I = F.begin(), E = F.end(); I != E; ) { + for (Function::iterator I = F.begin(); I != F.end(); ) { BasicBlock *BB = I++; MadeChange |= OptimizeBlock(*BB); } @@ -215,11 +226,13 @@ bool CodeGenPrepare::EliminateFallThrough(Function &F) { // edge, just collapse it. BasicBlock *SinglePred = BB->getSinglePredecessor(); - if (!SinglePred || SinglePred == BB) continue; + // Don't merge if BB's address is taken. + if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue; BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator()); if (Term && !Term->isConditional()) { Changed = true; + DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n"); // Remember if SinglePred was the entry block of the function. // If so, we will need to move BB back to the entry position. bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); @@ -230,7 +243,6 @@ bool CodeGenPrepare::EliminateFallThrough(Function &F) { // We have erased a block. Update the iterator. I = BB; - DEBUG(dbgs() << "Merged:\n"<< *SinglePred << "\n\n\n"); } } return Changed; @@ -610,7 +622,7 @@ bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { // happens. WeakVH IterHandle(CurInstIterator); - replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getTargetData() : 0, + replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0, TLInfo, ModifiedDT ? 0 : DT); // If the iterator instruction was recursively deleted, start over at the @@ -634,8 +646,8 @@ bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { // From here on out we're working with named functions. if (CI->getCalledFunction() == 0) return false; - // We'll need TargetData from here on out. - const TargetData *TD = TLI ? TLI->getTargetData() : 0; + // We'll need DataLayout from here on out. + const DataLayout *TD = TLI ? TLI->getDataLayout() : 0; if (!TD) return false; // Lower all default uses of _chk calls. This is very similar @@ -649,6 +661,7 @@ bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return /// instructions to the predecessor to enable tail call optimizations. The /// case it is currently looking for is: +/// @code /// bb0: /// %tmp0 = tail call i32 @f0() /// br label %return @@ -661,9 +674,11 @@ bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { /// return: /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ] /// ret i32 %retval +/// @endcode /// /// => /// +/// @code /// bb0: /// %tmp0 = tail call i32 @f0() /// ret i32 %tmp0 @@ -673,7 +688,7 @@ bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { /// bb2: /// %tmp2 = tail call i32 @f2() /// ret i32 %tmp2 -/// +/// @endcode bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) { if (!TLI) return false; @@ -699,7 +714,8 @@ bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) { // See llvm::isInTailCallPosition(). const Function *F = BB->getParent(); Attributes CallerRetAttr = F->getAttributes().getRetAttributes(); - if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt)) + if (CallerRetAttr.hasAttribute(Attributes::ZExt) || + CallerRetAttr.hasAttribute(Attributes::SExt)) return false; // Make sure there are no instructions between the PHI and return, or that the @@ -757,7 +773,10 @@ bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) { // Conservatively require the attributes of the call to match those of the // return. Ignore noalias because it doesn't affect the call sequence. Attributes CalleeRetAttr = CS.getAttributes().getRetAttributes(); - if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias) + if (AttrBuilder(CalleeRetAttr). + removeAttribute(Attributes::NoAlias) != + AttrBuilder(CallerRetAttr). + removeAttribute(Attributes::NoAlias)) continue; // Make sure the call instruction is followed by an unconditional branch to @@ -774,7 +793,7 @@ bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) { } // If we eliminated all predecessors of the block, delete the block now. - if (Changed && pred_begin(BB) == pred_end(BB)) + if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB)) BB->eraseFromParent(); return Changed; @@ -914,7 +933,7 @@ bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " << *MemoryInst); Type *IntPtrTy = - TLI->getTargetData()->getIntPtrType(AccessTy->getContext()); + TLI->getDataLayout()->getIntPtrType(AccessTy->getContext()); Value *Result = 0; @@ -988,7 +1007,7 @@ bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, WeakVH IterHandle(CurInstIterator); BasicBlock *BB = CurInstIterator->getParent(); - RecursivelyDeleteTriviallyDeadInstructions(Repl); + RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo); if (IterHandle != CurInstIterator) { // If the iterator instruction was recursively deleted, start over at the @@ -1174,17 +1193,32 @@ static bool isFormingBranchFromSelectProfitable(SelectInst *SI) { } +/// If we have a SelectInst that will likely profit from branch prediction, +/// turn it into a branch. bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) { - // If we have a SelectInst that will likely profit from branch prediction, - // turn it into a branch. - if (DisableSelectToBranch || OptSize || !TLI || - !TLI->isPredictableSelectExpensive()) - return false; + bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); - if (!SI->getCondition()->getType()->isIntegerTy(1) || - !isFormingBranchFromSelectProfitable(SI)) + // Can we convert the 'select' to CF ? + if (DisableSelectToBranch || OptSize || !TLI || VectorCond) return false; + TargetLowering::SelectSupportKind SelectKind; + if (VectorCond) + SelectKind = TargetLowering::VectorMaskSelect; + else if (SI->getType()->isVectorTy()) + SelectKind = TargetLowering::ScalarCondVectorVal; + else + SelectKind = TargetLowering::ScalarValSelect; + + // Do we have efficient codegen support for this kind of 'selects' ? + if (TLI->isSelectSupported(SelectKind)) { + // We have efficient codegen support for the select instruction. + // Check if it is profitable to keep this 'select'. + if (!TLI->isPredictableSelectExpensive() || + !isFormingBranchFromSelectProfitable(SI)) + return false; + } + ModifiedDT = true; // First, we split the block containing the select into 2 blocks. @@ -1302,7 +1336,7 @@ bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { bool MadeChange = false; CurInstIterator = BB.begin(); - for (BasicBlock::iterator E = BB.end(); CurInstIterator != E; ) + while (CurInstIterator != BB.end()) MadeChange |= OptimizeInst(CurInstIterator++); return MadeChange; diff --git a/contrib/llvm/lib/Transforms/Scalar/ConstantProp.cpp b/contrib/llvm/lib/Transforms/Scalar/ConstantProp.cpp index 5430f62..369720b 100644 --- a/contrib/llvm/lib/Transforms/Scalar/ConstantProp.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/ConstantProp.cpp @@ -24,7 +24,7 @@ #include "llvm/Constant.h" #include "llvm/Instruction.h" #include "llvm/Pass.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Support/InstIterator.h" #include "llvm/ADT/Statistic.h" @@ -67,7 +67,7 @@ bool ConstantPropagation::runOnFunction(Function &F) { WorkList.insert(&*i); } bool Changed = false; - TargetData *TD = getAnalysisIfAvailable<TargetData>(); + DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>(); while (!WorkList.empty()) { diff --git a/contrib/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp b/contrib/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp index 9b0aadb..3ec6f3d 100644 --- a/contrib/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp @@ -235,6 +235,11 @@ bool CorrelatedValuePropagation::processSwitch(SwitchInst *SI) { // This case never fires - remove it. CI.getCaseSuccessor()->removePredecessor(BB); SI->removeCase(CI); // Does not invalidate the iterator. + + // The condition can be modified by removePredecessor's PHI simplification + // logic. + Cond = SI->getCondition(); + ++NumDeadCases; Changed = true; } else if (State == LazyValueInfo::True) { diff --git a/contrib/llvm/lib/Transforms/Scalar/DCE.cpp b/contrib/llvm/lib/Transforms/Scalar/DCE.cpp index 8dbcc23..a2e074f 100644 --- a/contrib/llvm/lib/Transforms/Scalar/DCE.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/DCE.cpp @@ -22,6 +22,7 @@ #include "llvm/Instruction.h" #include "llvm/Pass.h" #include "llvm/Support/InstIterator.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/ADT/Statistic.h" using namespace llvm; @@ -38,10 +39,11 @@ namespace { initializeDeadInstEliminationPass(*PassRegistry::getPassRegistry()); } virtual bool runOnBasicBlock(BasicBlock &BB) { + TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); bool Changed = false; for (BasicBlock::iterator DI = BB.begin(); DI != BB.end(); ) { Instruction *Inst = DI++; - if (isInstructionTriviallyDead(Inst)) { + if (isInstructionTriviallyDead(Inst, TLI)) { Inst->eraseFromParent(); Changed = true; ++DIEEliminated; @@ -87,6 +89,8 @@ char DCE::ID = 0; INITIALIZE_PASS(DCE, "dce", "Dead Code Elimination", false, false) bool DCE::runOnFunction(Function &F) { + TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); + // Start out with all of the instructions in the worklist... std::vector<Instruction*> WorkList; for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) @@ -101,7 +105,7 @@ bool DCE::runOnFunction(Function &F) { Instruction *I = WorkList.back(); WorkList.pop_back(); - if (isInstructionTriviallyDead(I)) { // If the instruction is dead. + if (isInstructionTriviallyDead(I, TLI)) { // If the instruction is dead. // Loop over all of the values that the instruction uses, if there are // instructions being used, add them to the worklist, because they might // go dead after this one is removed. @@ -114,13 +118,8 @@ bool DCE::runOnFunction(Function &F) { I->eraseFromParent(); // Remove the instruction from the worklist if it still exists in it. - for (std::vector<Instruction*>::iterator WI = WorkList.begin(); - WI != WorkList.end(); ) { - if (*WI == I) - WI = WorkList.erase(WI); - else - ++WI; - } + WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I), + WorkList.end()); MadeChange = true; ++DCEEliminated; diff --git a/contrib/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp b/contrib/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp index 8b1283f..736cc05 100644 --- a/contrib/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp @@ -29,7 +29,8 @@ #include "llvm/Analysis/MemoryBuiltins.h" #include "llvm/Analysis/MemoryDependenceAnalysis.h" #include "llvm/Analysis/ValueTracking.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Support/Debug.h" #include "llvm/ADT/SetVector.h" @@ -45,6 +46,7 @@ namespace { AliasAnalysis *AA; MemoryDependenceAnalysis *MD; DominatorTree *DT; + const TargetLibraryInfo *TLI; static char ID; // Pass identification, replacement for typeid DSE() : FunctionPass(ID), AA(0), MD(0), DT(0) { @@ -55,6 +57,7 @@ namespace { AA = &getAnalysis<AliasAnalysis>(); MD = &getAnalysis<MemoryDependenceAnalysis>(); DT = &getAnalysis<DominatorTree>(); + TLI = AA->getTargetLibraryInfo(); bool Changed = false; for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) @@ -106,6 +109,7 @@ FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); } /// static void DeleteDeadInstruction(Instruction *I, MemoryDependenceAnalysis &MD, + const TargetLibraryInfo *TLI, SmallSetVector<Value*, 16> *ValueSet = 0) { SmallVector<Instruction*, 32> NowDeadInsts; @@ -130,7 +134,7 @@ static void DeleteDeadInstruction(Instruction *I, if (!Op->use_empty()) continue; if (Instruction *OpI = dyn_cast<Instruction>(Op)) - if (isInstructionTriviallyDead(OpI)) + if (isInstructionTriviallyDead(OpI, TLI)) NowDeadInsts.push_back(OpI); } @@ -143,7 +147,7 @@ static void DeleteDeadInstruction(Instruction *I, /// hasMemoryWrite - Does this instruction write some memory? This only returns /// true for things that we can analyze with other helpers below. -static bool hasMemoryWrite(Instruction *I) { +static bool hasMemoryWrite(Instruction *I, const TargetLibraryInfo *TLI) { if (isa<StoreInst>(I)) return true; if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { @@ -158,6 +162,26 @@ static bool hasMemoryWrite(Instruction *I) { return true; } } + if (CallSite CS = I) { + if (Function *F = CS.getCalledFunction()) { + if (TLI && TLI->has(LibFunc::strcpy) && + F->getName() == TLI->getName(LibFunc::strcpy)) { + return true; + } + if (TLI && TLI->has(LibFunc::strncpy) && + F->getName() == TLI->getName(LibFunc::strncpy)) { + return true; + } + if (TLI && TLI->has(LibFunc::strcat) && + F->getName() == TLI->getName(LibFunc::strcat)) { + return true; + } + if (TLI && TLI->has(LibFunc::strncat) && + F->getName() == TLI->getName(LibFunc::strncat)) { + return true; + } + } + } return false; } @@ -175,7 +199,7 @@ getLocForWrite(Instruction *Inst, AliasAnalysis &AA) { // If we don't have target data around, an unknown size in Location means // that we should use the size of the pointee type. This isn't valid for // memset/memcpy, which writes more than an i8. - if (Loc.Size == AliasAnalysis::UnknownSize && AA.getTargetData() == 0) + if (Loc.Size == AliasAnalysis::UnknownSize && AA.getDataLayout() == 0) return AliasAnalysis::Location(); return Loc; } @@ -189,7 +213,7 @@ getLocForWrite(Instruction *Inst, AliasAnalysis &AA) { // If we don't have target data around, an unknown size in Location means // that we should use the size of the pointee type. This isn't valid for // init.trampoline, which writes more than an i8. - if (AA.getTargetData() == 0) return AliasAnalysis::Location(); + if (AA.getDataLayout() == 0) return AliasAnalysis::Location(); // FIXME: We don't know the size of the trampoline, so we can't really // handle it here. @@ -205,7 +229,8 @@ getLocForWrite(Instruction *Inst, AliasAnalysis &AA) { /// instruction if any. static AliasAnalysis::Location getLocForRead(Instruction *Inst, AliasAnalysis &AA) { - assert(hasMemoryWrite(Inst) && "Unknown instruction case"); + assert(hasMemoryWrite(Inst, AA.getTargetLibraryInfo()) && + "Unknown instruction case"); // The only instructions that both read and write are the mem transfer // instructions (memcpy/memmove). @@ -222,23 +247,29 @@ static bool isRemovable(Instruction *I) { if (StoreInst *SI = dyn_cast<StoreInst>(I)) return SI->isUnordered(); - IntrinsicInst *II = cast<IntrinsicInst>(I); - switch (II->getIntrinsicID()) { - default: llvm_unreachable("doesn't pass 'hasMemoryWrite' predicate"); - case Intrinsic::lifetime_end: - // Never remove dead lifetime_end's, e.g. because it is followed by a - // free. - return false; - case Intrinsic::init_trampoline: - // Always safe to remove init_trampoline. - return true; + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { + switch (II->getIntrinsicID()) { + default: llvm_unreachable("doesn't pass 'hasMemoryWrite' predicate"); + case Intrinsic::lifetime_end: + // Never remove dead lifetime_end's, e.g. because it is followed by a + // free. + return false; + case Intrinsic::init_trampoline: + // Always safe to remove init_trampoline. + return true; - case Intrinsic::memset: - case Intrinsic::memmove: - case Intrinsic::memcpy: - // Don't remove volatile memory intrinsics. - return !cast<MemIntrinsic>(II)->isVolatile(); + case Intrinsic::memset: + case Intrinsic::memmove: + case Intrinsic::memcpy: + // Don't remove volatile memory intrinsics. + return !cast<MemIntrinsic>(II)->isVolatile(); + } } + + if (CallSite CS = I) + return CS.getInstruction()->use_empty(); + + return false; } @@ -249,14 +280,19 @@ static bool isShortenable(Instruction *I) { if (isa<StoreInst>(I)) return false; - IntrinsicInst *II = cast<IntrinsicInst>(I); - switch (II->getIntrinsicID()) { - default: return false; - case Intrinsic::memset: - case Intrinsic::memcpy: - // Do shorten memory intrinsics. - return true; + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { + switch (II->getIntrinsicID()) { + default: return false; + case Intrinsic::memset: + case Intrinsic::memcpy: + // Do shorten memory intrinsics. + return true; + } } + + // Don't shorten libcalls calls for now. + + return false; } /// getStoredPointerOperand - Return the pointer that is being written to. @@ -266,17 +302,23 @@ static Value *getStoredPointerOperand(Instruction *I) { if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) return MI->getDest(); - IntrinsicInst *II = cast<IntrinsicInst>(I); - switch (II->getIntrinsicID()) { - default: llvm_unreachable("Unexpected intrinsic!"); - case Intrinsic::init_trampoline: - return II->getArgOperand(0); + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { + switch (II->getIntrinsicID()) { + default: llvm_unreachable("Unexpected intrinsic!"); + case Intrinsic::init_trampoline: + return II->getArgOperand(0); + } } + + CallSite CS = I; + // All the supported functions so far happen to have dest as their first + // argument. + return CS.getArgument(0); } static uint64_t getPointerSize(const Value *V, AliasAnalysis &AA) { uint64_t Size; - if (getObjectSize(V, Size, AA.getTargetData())) + if (getObjectSize(V, Size, AA.getDataLayout(), AA.getTargetLibraryInfo())) return Size; return AliasAnalysis::UnknownSize; } @@ -309,10 +351,10 @@ static OverwriteResult isOverwrite(const AliasAnalysis::Location &Later, // comparison. if (Later.Size == AliasAnalysis::UnknownSize || Earlier.Size == AliasAnalysis::UnknownSize) { - // If we have no TargetData information around, then the size of the store + // If we have no DataLayout information around, then the size of the store // is inferrable from the pointee type. If they are the same type, then // we know that the store is safe. - if (AA.getTargetData() == 0 && + if (AA.getDataLayout() == 0 && Later.Ptr->getType() == Earlier.Ptr->getType()) return OverwriteComplete; @@ -328,13 +370,13 @@ static OverwriteResult isOverwrite(const AliasAnalysis::Location &Later, // larger than the earlier one. if (Later.Size == AliasAnalysis::UnknownSize || Earlier.Size == AliasAnalysis::UnknownSize || - AA.getTargetData() == 0) + AA.getDataLayout() == 0) return OverwriteUnknown; // Check to see if the later store is to the entire object (either a global, // an alloca, or a byval argument). If so, then it clearly overwrites any // other store to the same object. - const TargetData &TD = *AA.getTargetData(); + const DataLayout &TD = *AA.getDataLayout(); const Value *UO1 = GetUnderlyingObject(P1, &TD), *UO2 = GetUnderlyingObject(P2, &TD); @@ -454,13 +496,13 @@ bool DSE::runOnBasicBlock(BasicBlock &BB) { Instruction *Inst = BBI++; // Handle 'free' calls specially. - if (CallInst *F = isFreeCall(Inst)) { + if (CallInst *F = isFreeCall(Inst, TLI)) { MadeChange |= HandleFree(F); continue; } // If we find something that writes memory, get its memory dependence. - if (!hasMemoryWrite(Inst)) + if (!hasMemoryWrite(Inst, TLI)) continue; MemDepResult InstDep = MD->getDependency(Inst); @@ -483,7 +525,7 @@ bool DSE::runOnBasicBlock(BasicBlock &BB) { // in case we need it. WeakVH NextInst(BBI); - DeleteDeadInstruction(SI, *MD); + DeleteDeadInstruction(SI, *MD, TLI); if (NextInst == 0) // Next instruction deleted. BBI = BB.begin(); @@ -530,7 +572,7 @@ bool DSE::runOnBasicBlock(BasicBlock &BB) { << *DepWrite << "\n KILLER: " << *Inst << '\n'); // Delete the store and now-dead instructions that feed it. - DeleteDeadInstruction(DepWrite, *MD); + DeleteDeadInstruction(DepWrite, *MD, TLI); ++NumFastStores; MadeChange = true; @@ -627,7 +669,7 @@ bool DSE::HandleFree(CallInst *F) { MemDepResult Dep = MD->getPointerDependencyFrom(Loc, false, InstPt, BB); while (Dep.isDef() || Dep.isClobber()) { Instruction *Dependency = Dep.getInst(); - if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency)) + if (!hasMemoryWrite(Dependency, TLI) || !isRemovable(Dependency)) break; Value *DepPointer = @@ -640,7 +682,7 @@ bool DSE::HandleFree(CallInst *F) { Instruction *Next = llvm::next(BasicBlock::iterator(Dependency)); // DCE instructions only used to calculate that store - DeleteDeadInstruction(Dependency, *MD); + DeleteDeadInstruction(Dependency, *MD, TLI); ++NumFastStores; MadeChange = true; @@ -659,6 +701,22 @@ bool DSE::HandleFree(CallInst *F) { return MadeChange; } +namespace { + struct CouldRef { + typedef Value *argument_type; + const CallSite CS; + AliasAnalysis *AA; + + bool operator()(Value *I) { + // See if the call site touches the value. + AliasAnalysis::ModRefResult A = + AA->getModRefInfo(CS, I, getPointerSize(I, *AA)); + + return A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref; + } + }; +} + /// handleEndBlock - Remove dead stores to stack-allocated locations in the /// function end block. Ex: /// %A = alloca i32 @@ -680,7 +738,7 @@ bool DSE::handleEndBlock(BasicBlock &BB) { // Okay, so these are dead heap objects, but if the pointer never escapes // then it's leaked by this function anyways. - else if (isAllocLikeFn(I) && !PointerMayBeCaptured(I, true, true)) + else if (isAllocLikeFn(I, TLI) && !PointerMayBeCaptured(I, true, true)) DeadStackObjects.insert(I); } @@ -696,7 +754,7 @@ bool DSE::handleEndBlock(BasicBlock &BB) { --BBI; // If we find a store, check to see if it points into a dead stack value. - if (hasMemoryWrite(BBI) && isRemovable(BBI)) { + if (hasMemoryWrite(BBI, TLI) && isRemovable(BBI)) { // See through pointer-to-pointer bitcasts SmallVector<Value *, 4> Pointers; GetUnderlyingObjects(getStoredPointerOperand(BBI), Pointers); @@ -724,7 +782,7 @@ bool DSE::handleEndBlock(BasicBlock &BB) { dbgs() << '\n'); // DCE instructions only used to calculate that store. - DeleteDeadInstruction(Dead, *MD, &DeadStackObjects); + DeleteDeadInstruction(Dead, *MD, TLI, &DeadStackObjects); ++NumFastStores; MadeChange = true; continue; @@ -732,9 +790,9 @@ bool DSE::handleEndBlock(BasicBlock &BB) { } // Remove any dead non-memory-mutating instructions. - if (isInstructionTriviallyDead(BBI)) { + if (isInstructionTriviallyDead(BBI, TLI)) { Instruction *Inst = BBI++; - DeleteDeadInstruction(Inst, *MD, &DeadStackObjects); + DeleteDeadInstruction(Inst, *MD, TLI, &DeadStackObjects); ++NumFastOther; MadeChange = true; continue; @@ -750,7 +808,7 @@ bool DSE::handleEndBlock(BasicBlock &BB) { if (CallSite CS = cast<Value>(BBI)) { // Remove allocation function calls from the list of dead stack objects; // there can't be any references before the definition. - if (isAllocLikeFn(BBI)) + if (isAllocLikeFn(BBI, TLI)) DeadStackObjects.remove(BBI); // If this call does not access memory, it can't be loading any of our @@ -760,20 +818,8 @@ bool DSE::handleEndBlock(BasicBlock &BB) { // If the call might load from any of our allocas, then any store above // the call is live. - SmallVector<Value*, 8> LiveAllocas; - for (SmallSetVector<Value*, 16>::iterator I = DeadStackObjects.begin(), - E = DeadStackObjects.end(); I != E; ++I) { - // See if the call site touches it. - AliasAnalysis::ModRefResult A = - AA->getModRefInfo(CS, *I, getPointerSize(*I, *AA)); - - if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref) - LiveAllocas.push_back(*I); - } - - for (SmallVector<Value*, 8>::iterator I = LiveAllocas.begin(), - E = LiveAllocas.end(); I != E; ++I) - DeadStackObjects.remove(*I); + CouldRef Pred = { CS, AA }; + DeadStackObjects.remove_if(Pred); // If all of the allocas were clobbered by the call then we're not going // to find anything else to process. @@ -816,6 +862,20 @@ bool DSE::handleEndBlock(BasicBlock &BB) { return MadeChange; } +namespace { + struct CouldAlias { + typedef Value *argument_type; + const AliasAnalysis::Location &LoadedLoc; + AliasAnalysis *AA; + + bool operator()(Value *I) { + // See if the loaded location could alias the stack location. + AliasAnalysis::Location StackLoc(I, getPointerSize(I, *AA)); + return !AA->isNoAlias(StackLoc, LoadedLoc); + } + }; +} + /// RemoveAccessedObjects - Check to see if the specified location may alias any /// of the stack objects in the DeadStackObjects set. If so, they become live /// because the location is being loaded. @@ -834,16 +894,7 @@ void DSE::RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc, return; } - SmallVector<Value*, 16> NowLive; - for (SmallSetVector<Value*, 16>::iterator I = DeadStackObjects.begin(), - E = DeadStackObjects.end(); I != E; ++I) { - // See if the loaded location could alias the stack location. - AliasAnalysis::Location StackLoc(*I, getPointerSize(*I, *AA)); - if (!AA->isNoAlias(StackLoc, LoadedLoc)) - NowLive.push_back(*I); - } - - for (SmallVector<Value*, 16>::iterator I = NowLive.begin(), E = NowLive.end(); - I != E; ++I) - DeadStackObjects.remove(*I); + // Remove objects that could alias LoadedLoc. + CouldAlias Pred = { LoadedLoc, AA }; + DeadStackObjects.remove_if(Pred); } diff --git a/contrib/llvm/lib/Transforms/Scalar/EarlyCSE.cpp b/contrib/llvm/lib/Transforms/Scalar/EarlyCSE.cpp index 9759549..101009d 100644 --- a/contrib/llvm/lib/Transforms/Scalar/EarlyCSE.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/EarlyCSE.cpp @@ -18,11 +18,12 @@ #include "llvm/Pass.h" #include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/InstructionSimplify.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Support/Debug.h" #include "llvm/Support/RecyclingAllocator.h" +#include "llvm/ADT/Hashing.h" #include "llvm/ADT/ScopedHashTable.h" #include "llvm/ADT/Statistic.h" #include <deque> @@ -90,35 +91,56 @@ template<> struct DenseMapInfo<SimpleValue> { unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) { Instruction *Inst = Val.Inst; - // Hash in all of the operands as pointers. - unsigned Res = 0; - for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) - Res ^= getHash(Inst->getOperand(i)) << (i & 0xF); + if (BinaryOperator* BinOp = dyn_cast<BinaryOperator>(Inst)) { + Value *LHS = BinOp->getOperand(0); + Value *RHS = BinOp->getOperand(1); + if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1)) + std::swap(LHS, RHS); + + if (isa<OverflowingBinaryOperator>(BinOp)) { + // Hash the overflow behavior + unsigned Overflow = + BinOp->hasNoSignedWrap() * OverflowingBinaryOperator::NoSignedWrap | + BinOp->hasNoUnsignedWrap() * OverflowingBinaryOperator::NoUnsignedWrap; + return hash_combine(BinOp->getOpcode(), Overflow, LHS, RHS); + } - if (CastInst *CI = dyn_cast<CastInst>(Inst)) - Res ^= getHash(CI->getType()); - else if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) - Res ^= CI->getPredicate(); - else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst)) { - for (ExtractValueInst::idx_iterator I = EVI->idx_begin(), - E = EVI->idx_end(); I != E; ++I) - Res ^= *I; - } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst)) { - for (InsertValueInst::idx_iterator I = IVI->idx_begin(), - E = IVI->idx_end(); I != E; ++I) - Res ^= *I; - } else { - // nothing extra to hash in. - assert((isa<CallInst>(Inst) || - isa<BinaryOperator>(Inst) || isa<GetElementPtrInst>(Inst) || - isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) || - isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst)) && - "Invalid/unknown instruction"); + return hash_combine(BinOp->getOpcode(), LHS, RHS); } + if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) { + Value *LHS = CI->getOperand(0); + Value *RHS = CI->getOperand(1); + CmpInst::Predicate Pred = CI->getPredicate(); + if (Inst->getOperand(0) > Inst->getOperand(1)) { + std::swap(LHS, RHS); + Pred = CI->getSwappedPredicate(); + } + return hash_combine(Inst->getOpcode(), Pred, LHS, RHS); + } + + if (CastInst *CI = dyn_cast<CastInst>(Inst)) + return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0)); + + if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst)) + return hash_combine(EVI->getOpcode(), EVI->getOperand(0), + hash_combine_range(EVI->idx_begin(), EVI->idx_end())); + + if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst)) + return hash_combine(IVI->getOpcode(), IVI->getOperand(0), + IVI->getOperand(1), + hash_combine_range(IVI->idx_begin(), IVI->idx_end())); + + assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) || + isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) || + isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) || + isa<ShuffleVectorInst>(Inst)) && "Invalid/unknown instruction"); + // Mix in the opcode. - return (Res << 1) ^ Inst->getOpcode(); + return hash_combine(Inst->getOpcode(), + hash_combine_range(Inst->value_op_begin(), + Inst->value_op_end())); } bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) { @@ -128,7 +150,41 @@ bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) { return LHSI == RHSI; if (LHSI->getOpcode() != RHSI->getOpcode()) return false; - return LHSI->isIdenticalTo(RHSI); + if (LHSI->isIdenticalTo(RHSI)) return true; + + // If we're not strictly identical, we still might be a commutable instruction + if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) { + if (!LHSBinOp->isCommutative()) + return false; + + assert(isa<BinaryOperator>(RHSI) + && "same opcode, but different instruction type?"); + BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI); + + // Check overflow attributes + if (isa<OverflowingBinaryOperator>(LHSBinOp)) { + assert(isa<OverflowingBinaryOperator>(RHSBinOp) + && "same opcode, but different operator type?"); + if (LHSBinOp->hasNoUnsignedWrap() != RHSBinOp->hasNoUnsignedWrap() || + LHSBinOp->hasNoSignedWrap() != RHSBinOp->hasNoSignedWrap()) + return false; + } + + // Commuted equality + return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) && + LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0); + } + if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) { + assert(isa<CmpInst>(RHSI) + && "same opcode, but different instruction type?"); + CmpInst *RHSCmp = cast<CmpInst>(RHSI); + // Commuted equality + return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) && + LHSCmp->getOperand(1) == RHSCmp->getOperand(0) && + LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate(); + } + + return false; } //===----------------------------------------------------------------------===// @@ -216,7 +272,7 @@ namespace { /// cases. class EarlyCSE : public FunctionPass { public: - const TargetData *TD; + const DataLayout *TD; const TargetLibraryInfo *TLI; DominatorTree *DT; typedef RecyclingAllocator<BumpPtrAllocator, @@ -274,7 +330,8 @@ private: CallScope(*availableCalls) {} private: - NodeScope(const NodeScope&); // DO NOT IMPLEMENT + NodeScope(const NodeScope&) LLVM_DELETED_FUNCTION; + void operator=(const NodeScope&) LLVM_DELETED_FUNCTION; ScopedHTType::ScopeTy Scope; LoadHTType::ScopeTy LoadScope; @@ -313,7 +370,8 @@ private: void process() { Processed = true; } private: - StackNode(const StackNode&); // DO NOT IMPLEMENT + StackNode(const StackNode&) LLVM_DELETED_FUNCTION; + void operator=(const StackNode&) LLVM_DELETED_FUNCTION; // Members. unsigned CurrentGeneration; @@ -374,7 +432,7 @@ bool EarlyCSE::processNode(DomTreeNode *Node) { Instruction *Inst = I++; // Dead instructions should just be removed. - if (isInstructionTriviallyDead(Inst)) { + if (isInstructionTriviallyDead(Inst, TLI)) { DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n'); Inst->eraseFromParent(); Changed = true; @@ -506,7 +564,7 @@ bool EarlyCSE::processNode(DomTreeNode *Node) { bool EarlyCSE::runOnFunction(Function &F) { std::deque<StackNode *> nodesToProcess; - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); DT = &getAnalysis<DominatorTree>(); diff --git a/contrib/llvm/lib/Transforms/Scalar/GVN.cpp b/contrib/llvm/lib/Transforms/Scalar/GVN.cpp index 4822fd0..f003e06 100644 --- a/contrib/llvm/lib/Transforms/Scalar/GVN.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/GVN.cpp @@ -41,7 +41,7 @@ #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/PatternMatch.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/SSAUpdater.h" @@ -271,16 +271,16 @@ void ValueTable::add(Value *V, uint32_t num) { valueNumbering.insert(std::make_pair(V, num)); } -uint32_t ValueTable::lookup_or_add_call(CallInst* C) { +uint32_t ValueTable::lookup_or_add_call(CallInst *C) { if (AA->doesNotAccessMemory(C)) { Expression exp = create_expression(C); - uint32_t& e = expressionNumbering[exp]; + uint32_t &e = expressionNumbering[exp]; if (!e) e = nextValueNumber++; valueNumbering[C] = e; return e; } else if (AA->onlyReadsMemory(C)) { Expression exp = create_expression(C); - uint32_t& e = expressionNumbering[exp]; + uint32_t &e = expressionNumbering[exp]; if (!e) { e = nextValueNumber++; valueNumbering[C] = e; @@ -413,7 +413,7 @@ uint32_t ValueTable::lookup_or_add(Value *V) { case Instruction::LShr: case Instruction::AShr: case Instruction::And: - case Instruction::Or : + case Instruction::Or: case Instruction::Xor: case Instruction::ICmp: case Instruction::FCmp: @@ -503,7 +503,7 @@ namespace { bool NoLoads; MemoryDependenceAnalysis *MD; DominatorTree *DT; - const TargetData *TD; + const DataLayout *TD; const TargetLibraryInfo *TLI; ValueTable VN; @@ -535,7 +535,7 @@ namespace { InstrsToErase.push_back(I); } - const TargetData *getTargetData() const { return TD; } + const DataLayout *getDataLayout() const { return TD; } DominatorTree &getDominatorTree() const { return *DT; } AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); } MemoryDependenceAnalysis &getMemDep() const { return *MD; } @@ -632,6 +632,7 @@ INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) INITIALIZE_AG_DEPENDENCY(AliasAnalysis) INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void GVN::dump(DenseMap<uint32_t, Value*>& d) { errs() << "{\n"; for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), @@ -641,6 +642,7 @@ void GVN::dump(DenseMap<uint32_t, Value*>& d) { } errs() << "}\n"; } +#endif /// IsValueFullyAvailableInBlock - Return true if we can prove that the value /// we're analyzing is fully available in the specified block. As we go, keep @@ -728,7 +730,7 @@ SpeculationFailure: /// CoerceAvailableValueToLoadType will succeed. static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy, - const TargetData &TD) { + const DataLayout &TD) { // If the loaded or stored value is an first class array or struct, don't try // to transform them. We need to be able to bitcast to integer. if (LoadTy->isStructTy() || LoadTy->isArrayTy() || @@ -744,7 +746,6 @@ static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, return true; } - /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and /// then a load from a must-aliased pointer of a different type, try to coerce /// the stored value. LoadedTy is the type of the load we want to replace and @@ -754,7 +755,7 @@ static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy, Instruction *InsertPt, - const TargetData &TD) { + const DataLayout &TD) { if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) return 0; @@ -767,24 +768,25 @@ static Value *CoerceAvailableValueToLoadType(Value *StoredVal, // If the store and reload are the same size, we can always reuse it. if (StoreSize == LoadSize) { // Pointer to Pointer -> use bitcast. - if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) + if (StoredValTy->getScalarType()->isPointerTy() && + LoadedTy->getScalarType()->isPointerTy()) return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); // Convert source pointers to integers, which can be bitcast. - if (StoredValTy->isPointerTy()) { - StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); + if (StoredValTy->getScalarType()->isPointerTy()) { + StoredValTy = TD.getIntPtrType(StoredValTy); StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); } Type *TypeToCastTo = LoadedTy; - if (TypeToCastTo->isPointerTy()) - TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); + if (TypeToCastTo->getScalarType()->isPointerTy()) + TypeToCastTo = TD.getIntPtrType(TypeToCastTo); if (StoredValTy != TypeToCastTo) StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); // Cast to pointer if the load needs a pointer type. - if (LoadedTy->isPointerTy()) + if (LoadedTy->getScalarType()->isPointerTy()) StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); return StoredVal; @@ -796,8 +798,8 @@ static Value *CoerceAvailableValueToLoadType(Value *StoredVal, assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); // Convert source pointers to integers, which can be manipulated. - if (StoredValTy->isPointerTy()) { - StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); + if (StoredValTy->getScalarType()->isPointerTy()) { + StoredValTy = TD.getIntPtrType(StoredValTy); StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); } @@ -822,7 +824,7 @@ static Value *CoerceAvailableValueToLoadType(Value *StoredVal, return StoredVal; // If the result is a pointer, inttoptr. - if (LoadedTy->isPointerTy()) + if (LoadedTy->getScalarType()->isPointerTy()) return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); // Otherwise, bitcast. @@ -840,7 +842,7 @@ static Value *CoerceAvailableValueToLoadType(Value *StoredVal, static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, Value *WritePtr, uint64_t WriteSizeInBits, - const TargetData &TD) { + const DataLayout &TD) { // If the loaded or stored value is a first class array or struct, don't try // to transform them. We need to be able to bitcast to integer. if (LoadTy->isStructTy() || LoadTy->isArrayTy()) @@ -913,7 +915,7 @@ static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, /// memdep query of a load that ends up being a clobbering store. static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, StoreInst *DepSI, - const TargetData &TD) { + const DataLayout &TD) { // Cannot handle reading from store of first-class aggregate yet. if (DepSI->getValueOperand()->getType()->isStructTy() || DepSI->getValueOperand()->getType()->isArrayTy()) @@ -929,7 +931,7 @@ static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, /// memdep query of a load that ends up being clobbered by another load. See if /// the other load can feed into the second load. static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, - LoadInst *DepLI, const TargetData &TD){ + LoadInst *DepLI, const DataLayout &TD){ // Cannot handle reading from store of first-class aggregate yet. if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) return -1; @@ -957,7 +959,7 @@ static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, MemIntrinsic *MI, - const TargetData &TD) { + const DataLayout &TD) { // If the mem operation is a non-constant size, we can't handle it. ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); if (SizeCst == 0) return -1; @@ -1007,7 +1009,7 @@ static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, /// before we give up. static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, Type *LoadTy, - Instruction *InsertPt, const TargetData &TD){ + Instruction *InsertPt, const DataLayout &TD){ LLVMContext &Ctx = SrcVal->getType()->getContext(); uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; @@ -1017,8 +1019,9 @@ static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, // Compute which bits of the stored value are being used by the load. Convert // to an integer type to start with. - if (SrcVal->getType()->isPointerTy()) - SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx)); + if (SrcVal->getType()->getScalarType()->isPointerTy()) + SrcVal = Builder.CreatePtrToInt(SrcVal, + TD.getIntPtrType(SrcVal->getType())); if (!SrcVal->getType()->isIntegerTy()) SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8)); @@ -1046,7 +1049,7 @@ static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, GVN &gvn) { - const TargetData &TD = *gvn.getTargetData(); + const DataLayout &TD = *gvn.getDataLayout(); // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to // widen SrcVal out to a larger load. unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); @@ -1105,7 +1108,7 @@ static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, /// memdep query of a load that ends up being a clobbering mem intrinsic. static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, Instruction *InsertPt, - const TargetData &TD){ + const DataLayout &TD){ LLVMContext &Ctx = LoadTy->getContext(); uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; @@ -1229,7 +1232,7 @@ struct AvailableValueInBlock { if (isSimpleValue()) { Res = getSimpleValue(); if (Res->getType() != LoadTy) { - const TargetData *TD = gvn.getTargetData(); + const DataLayout *TD = gvn.getDataLayout(); assert(TD && "Need target data to handle type mismatch case"); Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), *TD); @@ -1251,7 +1254,7 @@ struct AvailableValueInBlock { << *Res << '\n' << "\n\n\n"); } } else { - const TargetData *TD = gvn.getTargetData(); + const DataLayout *TD = gvn.getDataLayout(); assert(TD && "Need target data to handle type mismatch case"); Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy, BB->getTerminator(), *TD); @@ -1299,7 +1302,7 @@ static Value *ConstructSSAForLoadSet(LoadInst *LI, Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); // If new PHI nodes were created, notify alias analysis. - if (V->getType()->isPointerTy()) { + if (V->getType()->getScalarType()->isPointerTy()) { AliasAnalysis *AA = gvn.getAliasAnalysis(); for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) @@ -1436,7 +1439,7 @@ bool GVN::processNonLocalLoad(LoadInst *LI) { Instruction *DepInst = DepInfo.getInst(); // Loading the allocation -> undef. - if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst) || + if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || // Loading immediately after lifetime begin -> undef. isLifetimeStart(DepInst)) { ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, @@ -1496,7 +1499,7 @@ bool GVN::processNonLocalLoad(LoadInst *LI) { if (isa<PHINode>(V)) V->takeName(LI); - if (V->getType()->isPointerTy()) + if (V->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(V); markInstructionForDeletion(LI); ++NumGVNLoad; @@ -1728,7 +1731,7 @@ bool GVN::processNonLocalLoad(LoadInst *LI) { LI->replaceAllUsesWith(V); if (isa<PHINode>(V)) V->takeName(LI); - if (V->getType()->isPointerTy()) + if (V->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(V); markInstructionForDeletion(LI); ++NumPRELoad; @@ -1855,7 +1858,7 @@ bool GVN::processLoad(LoadInst *L) { // Replace the load! L->replaceAllUsesWith(AvailVal); - if (AvailVal->getType()->isPointerTy()) + if (AvailVal->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(AvailVal); markInstructionForDeletion(L); ++NumGVNLoad; @@ -1912,7 +1915,7 @@ bool GVN::processLoad(LoadInst *L) { // Remove it! L->replaceAllUsesWith(StoredVal); - if (StoredVal->getType()->isPointerTy()) + if (StoredVal->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(StoredVal); markInstructionForDeletion(L); ++NumGVNLoad; @@ -1941,7 +1944,7 @@ bool GVN::processLoad(LoadInst *L) { // Remove it! patchAndReplaceAllUsesWith(AvailableVal, L); - if (DepLI->getType()->isPointerTy()) + if (DepLI->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(DepLI); markInstructionForDeletion(L); ++NumGVNLoad; @@ -1951,7 +1954,7 @@ bool GVN::processLoad(LoadInst *L) { // If this load really doesn't depend on anything, then we must be loading an // undef value. This can happen when loading for a fresh allocation with no // intervening stores, for example. - if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst)) { + if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) { L->replaceAllUsesWith(UndefValue::get(L->getType())); markInstructionForDeletion(L); ++NumGVNLoad; @@ -2182,7 +2185,7 @@ bool GVN::processInstruction(Instruction *I) { // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) { I->replaceAllUsesWith(V); - if (MD && V->getType()->isPointerTy()) + if (MD && V->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(V); markInstructionForDeletion(I); ++NumGVNSimpl; @@ -2231,12 +2234,20 @@ bool GVN::processInstruction(Instruction *I) { Value *SwitchCond = SI->getCondition(); BasicBlock *Parent = SI->getParent(); bool Changed = false; + + // Remember how many outgoing edges there are to every successor. + SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; + for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) + ++SwitchEdges[SI->getSuccessor(i)]; + for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) { BasicBlock *Dst = i.getCaseSuccessor(); - BasicBlockEdge E(Parent, Dst); - if (E.isSingleEdge()) + // If there is only a single edge, propagate the case value into it. + if (SwitchEdges.lookup(Dst) == 1) { + BasicBlockEdge E(Parent, Dst); Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E); + } } return Changed; } @@ -2274,7 +2285,7 @@ bool GVN::processInstruction(Instruction *I) { // Remove it! patchAndReplaceAllUsesWith(repl, I); - if (MD && repl->getType()->isPointerTy()) + if (MD && repl->getType()->getScalarType()->isPointerTy()) MD->invalidateCachedPointerInfo(repl); markInstructionForDeletion(I); return true; @@ -2285,7 +2296,7 @@ bool GVN::runOnFunction(Function& F) { if (!NoLoads) MD = &getAnalysis<MemoryDependenceAnalysis>(); DT = &getAnalysis<DominatorTree>(); - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); VN.setMemDep(MD); @@ -2522,7 +2533,7 @@ bool GVN::performPRE(Function &F) { addToLeaderTable(ValNo, Phi, CurrentBlock); Phi->setDebugLoc(CurInst->getDebugLoc()); CurInst->replaceAllUsesWith(Phi); - if (Phi->getType()->isPointerTy()) { + if (Phi->getType()->getScalarType()->isPointerTy()) { // Because we have added a PHI-use of the pointer value, it has now // "escaped" from alias analysis' perspective. We need to inform // AA of this. diff --git a/contrib/llvm/lib/Transforms/Scalar/GlobalMerge.cpp b/contrib/llvm/lib/Transforms/Scalar/GlobalMerge.cpp index b36a3cb..6301aad 100644 --- a/contrib/llvm/lib/Transforms/Scalar/GlobalMerge.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/GlobalMerge.cpp @@ -62,7 +62,7 @@ #include "llvm/Intrinsics.h" #include "llvm/Module.h" #include "llvm/Pass.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLowering.h" #include "llvm/Target/TargetLoweringObjectFile.h" #include "llvm/ADT/Statistic.h" @@ -98,9 +98,9 @@ namespace { } struct GlobalCmp { - const TargetData *TD; + const DataLayout *TD; - GlobalCmp(const TargetData *td) : TD(td) { } + GlobalCmp(const DataLayout *td) : TD(td) { } bool operator()(const GlobalVariable *GV1, const GlobalVariable *GV2) { Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType(); @@ -119,7 +119,7 @@ INITIALIZE_PASS(GlobalMerge, "global-merge", bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals, Module &M, bool isConst) const { - const TargetData *TD = TLI->getTargetData(); + const DataLayout *TD = TLI->getDataLayout(); // FIXME: Infer the maximum possible offset depending on the actual users // (these max offsets are different for the users inside Thumb or ARM @@ -170,7 +170,7 @@ bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals, bool GlobalMerge::doInitialization(Module &M) { SmallVector<GlobalVariable*, 16> Globals, ConstGlobals, BSSGlobals; - const TargetData *TD = TLI->getTargetData(); + const DataLayout *TD = TLI->getDataLayout(); unsigned MaxOffset = TLI->getMaximalGlobalOffset(); bool Changed = false; diff --git a/contrib/llvm/lib/Transforms/Scalar/IndVarSimplify.cpp b/contrib/llvm/lib/Transforms/Scalar/IndVarSimplify.cpp index 37f8bdf..310fd61 100644 --- a/contrib/llvm/lib/Transforms/Scalar/IndVarSimplify.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/IndVarSimplify.cpp @@ -43,7 +43,8 @@ #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/SimplifyIndVar.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/Target/TargetLibraryInfo.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" @@ -67,7 +68,8 @@ namespace { LoopInfo *LI; ScalarEvolution *SE; DominatorTree *DT; - TargetData *TD; + DataLayout *TD; + TargetLibraryInfo *TLI; SmallVector<WeakVH, 16> DeadInsts; bool Changed; @@ -218,8 +220,6 @@ static Instruction *getInsertPointForUses(Instruction *User, Value *Def, /// ConvertToSInt - Convert APF to an integer, if possible. static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { bool isExact = false; - if (&APF.getSemantics() == &APFloat::PPCDoubleDouble) - return false; // See if we can convert this to an int64_t uint64_t UIntVal; if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, @@ -414,11 +414,11 @@ void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { // new comparison. NewCompare->takeName(Compare); Compare->replaceAllUsesWith(NewCompare); - RecursivelyDeleteTriviallyDeadInstructions(Compare); + RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI); // Delete the old floating point increment. Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); - RecursivelyDeleteTriviallyDeadInstructions(Incr); + RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI); // If the FP induction variable still has uses, this is because something else // in the loop uses its value. In order to canonicalize the induction @@ -431,7 +431,7 @@ void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", PN->getParent()->getFirstInsertionPt()); PN->replaceAllUsesWith(Conv); - RecursivelyDeleteTriviallyDeadInstructions(PN); + RecursivelyDeleteTriviallyDeadInstructions(PN, TLI); } Changed = true; } @@ -549,15 +549,17 @@ void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { PN->setIncomingValue(i, ExitVal); - // If this instruction is dead now, delete it. - RecursivelyDeleteTriviallyDeadInstructions(Inst); + // If this instruction is dead now, delete it. Don't do it now to avoid + // invalidating iterators. + if (isInstructionTriviallyDead(Inst, TLI)) + DeadInsts.push_back(Inst); if (NumPreds == 1) { // Completely replace a single-pred PHI. This is safe, because the // NewVal won't be variant in the loop, so we don't need an LCSSA phi // node anymore. PN->replaceAllUsesWith(ExitVal); - RecursivelyDeleteTriviallyDeadInstructions(PN); + PN->eraseFromParent(); } } if (NumPreds != 1) { @@ -595,13 +597,13 @@ namespace { class WideIVVisitor : public IVVisitor { ScalarEvolution *SE; - const TargetData *TD; + const DataLayout *TD; public: WideIVInfo WI; WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV, - const TargetData *TData) : + const DataLayout *TData) : SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; } // Implement the interface used by simplifyUsersOfIV. @@ -1259,8 +1261,13 @@ static bool needsLFTR(Loop *L, DominatorTree *DT) { if (!Phi) return true; + // Do LFTR if PHI node is defined in the loop, but is *not* a counter. + int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); + if (Idx < 0) + return true; + // Do LFTR if the exit condition's IV is *not* a simple counter. - Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch()); + Value *IncV = Phi->getIncomingValue(Idx); return Phi != getLoopPhiForCounter(IncV, L, DT); } @@ -1339,7 +1346,7 @@ static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { /// could at least handle constant BECounts. static PHINode * FindLoopCounter(Loop *L, const SCEV *BECount, - ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) { + ScalarEvolution *SE, DominatorTree *DT, const DataLayout *TD) { uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); Value *Cond = @@ -1696,7 +1703,8 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { LI = &getAnalysis<LoopInfo>(); SE = &getAnalysis<ScalarEvolution>(); DT = &getAnalysis<DominatorTree>(); - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); + TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); DeadInsts.clear(); Changed = false; @@ -1763,7 +1771,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { while (!DeadInsts.empty()) if (Instruction *Inst = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) - RecursivelyDeleteTriviallyDeadInstructions(Inst); + RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); // The Rewriter may not be used from this point on. @@ -1772,7 +1780,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { SinkUnusedInvariants(L); // Clean up dead instructions. - Changed |= DeleteDeadPHIs(L->getHeader()); + Changed |= DeleteDeadPHIs(L->getHeader(), TLI); // Check a post-condition. assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!"); diff --git a/contrib/llvm/lib/Transforms/Scalar/JumpThreading.cpp b/contrib/llvm/lib/Transforms/Scalar/JumpThreading.cpp index dd42c59..e7ffa09 100644 --- a/contrib/llvm/lib/Transforms/Scalar/JumpThreading.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/JumpThreading.cpp @@ -23,7 +23,7 @@ #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/SSAUpdater.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/DenseSet.h" @@ -75,7 +75,7 @@ namespace { /// revectored to the false side of the second if. /// class JumpThreading : public FunctionPass { - TargetData *TD; + DataLayout *TD; TargetLibraryInfo *TLI; LazyValueInfo *LVI; #ifdef NDEBUG @@ -147,7 +147,7 @@ FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } /// bool JumpThreading::runOnFunction(Function &F) { DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); LVI = &getAnalysis<LazyValueInfo>(); @@ -1455,7 +1455,7 @@ bool JumpThreading::ThreadEdge(BasicBlock *BB, // At this point, the IR is fully up to date and consistent. Do a quick scan // over the new instructions and zap any that are constants or dead. This // frequently happens because of phi translation. - SimplifyInstructionsInBlock(NewBB, TD); + SimplifyInstructionsInBlock(NewBB, TD, TLI); // Threaded an edge! ++NumThreads; diff --git a/contrib/llvm/lib/Transforms/Scalar/LICM.cpp b/contrib/llvm/lib/Transforms/Scalar/LICM.cpp index 0192e92..4818437 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LICM.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LICM.cpp @@ -46,7 +46,7 @@ #include "llvm/Analysis/ValueTracking.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/SSAUpdater.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Support/CFG.h" #include "llvm/Support/CommandLine.h" @@ -100,7 +100,7 @@ namespace { LoopInfo *LI; // Current LoopInfo DominatorTree *DT; // Dominator Tree for the current Loop. - TargetData *TD; // TargetData for constant folding. + DataLayout *TD; // DataLayout for constant folding. TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding. // State that is updated as we process loops. @@ -108,6 +108,9 @@ namespace { BasicBlock *Preheader; // The preheader block of the current loop... Loop *CurLoop; // The current loop we are working on... AliasSetTracker *CurAST; // AliasSet information for the current loop... + bool MayThrow; // The current loop contains an instruction which + // may throw, thus preventing code motion of + // instructions with side effects. DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap; /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. @@ -204,7 +207,7 @@ bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { AA = &getAnalysis<AliasAnalysis>(); DT = &getAnalysis<DominatorTree>(); - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); CurAST = new AliasSetTracker(*AA); @@ -240,6 +243,15 @@ bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { CurAST->add(*BB); // Incorporate the specified basic block } + MayThrow = false; + // TODO: We've already searched for instructions which may throw in subloops. + // We may want to reuse this information. + for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end(); + (BB != BBE) && !MayThrow ; ++BB) + for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); + (I != E) && !MayThrow; ++I) + MayThrow |= I->mayThrow(); + // We want to visit all of the instructions in this loop... that are not parts // of our subloops (they have already had their invariants hoisted out of // their loop, into this loop, so there is no need to process the BODIES of @@ -307,7 +319,7 @@ void LICM::SinkRegion(DomTreeNode *N) { // If the instruction is dead, we would try to sink it because it isn't used // in the loop, instead, just delete it. - if (isInstructionTriviallyDead(&I)) { + if (isInstructionTriviallyDead(&I, TLI)) { DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); ++II; CurAST->deleteValue(&I); @@ -418,17 +430,22 @@ bool LICM::canSinkOrHoistInst(Instruction &I) { if (!FoundMod) return true; } - // FIXME: This should use mod/ref information to see if we can hoist or sink - // the call. + // FIXME: This should use mod/ref information to see if we can hoist or + // sink the call. return false; } - // Otherwise these instructions are hoistable/sinkable - return isa<BinaryOperator>(I) || isa<CastInst>(I) || - isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || - isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || - isa<ShuffleVectorInst>(I); + // Only these instructions are hoistable/sinkable. + bool HoistableKind = (isa<BinaryOperator>(I) || isa<CastInst>(I) || + isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || + isa<CmpInst>(I) || isa<InsertElementInst>(I) || + isa<ExtractElementInst>(I) || + isa<ShuffleVectorInst>(I)); + if (!HoistableKind) + return false; + + return isSafeToExecuteUnconditionally(I); } /// isNotUsedInLoop - Return true if the only users of this instruction are @@ -604,6 +621,12 @@ bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { } bool LICM::isGuaranteedToExecute(Instruction &Inst) { + + // Somewhere in this loop there is an instruction which may throw and make us + // exit the loop. + if (MayThrow) + return false; + // Otherwise we have to check to make sure that the instruction dominates all // of the exit blocks. If it doesn't, then there is a path out of the loop // which does not execute this instruction, so we can't hoist it. diff --git a/contrib/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/contrib/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp index ac1082c..a44e798 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp @@ -54,7 +54,7 @@ #include "llvm/Analysis/ValueTracking.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" using namespace llvm; @@ -65,7 +65,7 @@ STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores"); namespace { class LoopIdiomRecognize : public LoopPass { Loop *CurLoop; - const TargetData *TD; + const DataLayout *TD; DominatorTree *DT; ScalarEvolution *SE; TargetLibraryInfo *TLI; @@ -132,7 +132,8 @@ Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); } /// and zero out all the operands of this instruction. If any of them become /// dead, delete them and the computation tree that feeds them. /// -static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE) { +static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE, + const TargetLibraryInfo *TLI) { SmallVector<Instruction*, 32> NowDeadInsts; NowDeadInsts.push_back(I); @@ -153,7 +154,7 @@ static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE) { if (!Op->use_empty()) continue; if (Instruction *OpI = dyn_cast<Instruction>(Op)) - if (isInstructionTriviallyDead(OpI)) + if (isInstructionTriviallyDead(OpI, TLI)) NowDeadInsts.push_back(OpI); } @@ -164,15 +165,21 @@ static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE) { /// deleteIfDeadInstruction - If the specified value is a dead instruction, /// delete it and any recursively used instructions. -static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE) { +static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE, + const TargetLibraryInfo *TLI) { if (Instruction *I = dyn_cast<Instruction>(V)) - if (isInstructionTriviallyDead(I)) - deleteDeadInstruction(I, SE); + if (isInstructionTriviallyDead(I, TLI)) + deleteDeadInstruction(I, SE, TLI); } bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { CurLoop = L; + // If the loop could not be converted to canonical form, it must have an + // indirectbr in it, just give up. + if (!L->getLoopPreheader()) + return false; + // Disable loop idiom recognition if the function's name is a common idiom. StringRef Name = L->getHeader()->getParent()->getName(); if (Name == "memset" || Name == "memcpy") @@ -192,7 +199,7 @@ bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { return false; // We require target data for now. - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); if (TD == 0) return false; DT = &getAnalysis<DominatorTree>(); @@ -401,7 +408,7 @@ static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access, /// /// Note that we don't ever attempt to use memset_pattern8 or 4, because these /// just replicate their input array and then pass on to memset_pattern16. -static Constant *getMemSetPatternValue(Value *V, const TargetData &TD) { +static Constant *getMemSetPatternValue(Value *V, const DataLayout &TD) { // If the value isn't a constant, we can't promote it to being in a constant // array. We could theoretically do a store to an alloca or something, but // that doesn't seem worthwhile. @@ -490,7 +497,7 @@ processLoopStridedStore(Value *DestPtr, unsigned StoreSize, StoreSize, getAnalysis<AliasAnalysis>(), TheStore)){ Expander.clear(); // If we generated new code for the base pointer, clean up. - deleteIfDeadInstruction(BasePtr, *SE); + deleteIfDeadInstruction(BasePtr, *SE, TLI); return false; } @@ -538,7 +545,7 @@ processLoopStridedStore(Value *DestPtr, unsigned StoreSize, // Okay, the memset has been formed. Zap the original store and anything that // feeds into it. - deleteDeadInstruction(TheStore, *SE); + deleteDeadInstruction(TheStore, *SE, TLI); ++NumMemSet; return true; } @@ -579,7 +586,7 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, getAnalysis<AliasAnalysis>(), SI)) { Expander.clear(); // If we generated new code for the base pointer, clean up. - deleteIfDeadInstruction(StoreBasePtr, *SE); + deleteIfDeadInstruction(StoreBasePtr, *SE, TLI); return false; } @@ -594,8 +601,8 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, StoreSize, getAnalysis<AliasAnalysis>(), SI)) { Expander.clear(); // If we generated new code for the base pointer, clean up. - deleteIfDeadInstruction(LoadBasePtr, *SE); - deleteIfDeadInstruction(StoreBasePtr, *SE); + deleteIfDeadInstruction(LoadBasePtr, *SE, TLI); + deleteIfDeadInstruction(StoreBasePtr, *SE, TLI); return false; } @@ -628,7 +635,7 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, // Okay, the memset has been formed. Zap the original store and anything that // feeds into it. - deleteDeadInstruction(SI, *SE); + deleteDeadInstruction(SI, *SE, TLI); ++NumMemCpy; return true; } diff --git a/contrib/llvm/lib/Transforms/Scalar/LoopInstSimplify.cpp b/contrib/llvm/lib/Transforms/Scalar/LoopInstSimplify.cpp index 982400c..558f62e 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LoopInstSimplify.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LoopInstSimplify.cpp @@ -18,7 +18,7 @@ #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/LoopPass.h" #include "llvm/Support/Debug.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Scalar.h" #include "llvm/Transforms/Utils/Local.h" @@ -66,7 +66,7 @@ Pass *llvm::createLoopInstSimplifyPass() { bool LoopInstSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>(); LoopInfo *LI = &getAnalysis<LoopInfo>(); - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>(); SmallVector<BasicBlock*, 8> ExitBlocks; @@ -120,7 +120,7 @@ bool LoopInstSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { ++NumSimplified; } } - LocalChanged |= RecursivelyDeleteTriviallyDeadInstructions(I); + LocalChanged |= RecursivelyDeleteTriviallyDeadInstructions(I, TLI); if (IsSubloopHeader && !isa<PHINode>(I)) break; diff --git a/contrib/llvm/lib/Transforms/Scalar/LoopRotation.cpp b/contrib/llvm/lib/Transforms/Scalar/LoopRotation.cpp index 7eeb152..abe07aa 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LoopRotation.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LoopRotation.cpp @@ -24,6 +24,7 @@ #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/SSAUpdater.h" #include "llvm/Transforms/Utils/ValueMapper.h" +#include "llvm/Support/CFG.h" #include "llvm/Support/Debug.h" #include "llvm/ADT/Statistic.h" using namespace llvm; @@ -256,6 +257,7 @@ bool LoopRotate::rotateLoop(Loop *L) { return false; BasicBlock *OrigHeader = L->getHeader(); + BasicBlock *OrigLatch = L->getLoopLatch(); BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator()); if (BI == 0 || BI->isUnconditional()) @@ -267,13 +269,9 @@ bool LoopRotate::rotateLoop(Loop *L) { if (!L->isLoopExiting(OrigHeader)) return false; - // Updating PHInodes in loops with multiple exits adds complexity. - // Keep it simple, and restrict loop rotation to loops with one exit only. - // In future, lift this restriction and support for multiple exits if - // required. - SmallVector<BasicBlock*, 8> ExitBlocks; - L->getExitBlocks(ExitBlocks); - if (ExitBlocks.size() > 1) + // If the loop latch already contains a branch that leaves the loop then the + // loop is already rotated. + if (OrigLatch == 0 || L->isLoopExiting(OrigLatch)) return false; // Check size of original header and reject loop if it is very big. @@ -286,11 +284,10 @@ bool LoopRotate::rotateLoop(Loop *L) { // Now, this loop is suitable for rotation. BasicBlock *OrigPreheader = L->getLoopPreheader(); - BasicBlock *OrigLatch = L->getLoopLatch(); // If the loop could not be converted to canonical form, it must have an // indirectbr in it, just give up. - if (OrigPreheader == 0 || OrigLatch == 0) + if (OrigPreheader == 0) return false; // Anything ScalarEvolution may know about this loop or the PHI nodes @@ -298,6 +295,8 @@ bool LoopRotate::rotateLoop(Loop *L) { if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>()) SE->forgetLoop(L); + DEBUG(dbgs() << "LoopRotation: rotating "; L->dump()); + // Find new Loop header. NewHeader is a Header's one and only successor // that is inside loop. Header's other successor is outside the // loop. Otherwise loop is not suitable for rotation. @@ -408,10 +407,19 @@ bool LoopRotate::rotateLoop(Loop *L) { // Update DominatorTree to reflect the CFG change we just made. Then split // edges as necessary to preserve LoopSimplify form. if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) { - // Since OrigPreheader now has the conditional branch to Exit block, it is - // the dominator of Exit. - DT->changeImmediateDominator(Exit, OrigPreheader); - DT->changeImmediateDominator(NewHeader, OrigPreheader); + // Everything that was dominated by the old loop header is now dominated + // by the original loop preheader. Conceptually the header was merged + // into the preheader, even though we reuse the actual block as a new + // loop latch. + DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader); + SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(), + OrigHeaderNode->end()); + DomTreeNode *OrigPreheaderNode = DT->getNode(OrigPreheader); + for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) + DT->changeImmediateDominator(HeaderChildren[I], OrigPreheaderNode); + + assert(DT->getNode(Exit)->getIDom() == OrigPreheaderNode); + assert(DT->getNode(NewHeader)->getIDom() == OrigPreheaderNode); // Update OrigHeader to be dominated by the new header block. DT->changeImmediateDominator(OrigHeader, OrigLatch); @@ -440,6 +448,35 @@ bool LoopRotate::rotateLoop(Loop *L) { // Update OrigHeader to be dominated by the new header block. DT->changeImmediateDominator(NewHeader, OrigPreheader); DT->changeImmediateDominator(OrigHeader, OrigLatch); + + // Brute force incremental dominator tree update. Call + // findNearestCommonDominator on all CFG predecessors of each child of the + // original header. + DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader); + SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(), + OrigHeaderNode->end()); + bool Changed; + do { + Changed = false; + for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) { + DomTreeNode *Node = HeaderChildren[I]; + BasicBlock *BB = Node->getBlock(); + + pred_iterator PI = pred_begin(BB); + BasicBlock *NearestDom = *PI; + for (pred_iterator PE = pred_end(BB); PI != PE; ++PI) + NearestDom = DT->findNearestCommonDominator(NearestDom, *PI); + + // Remember if this changes the DomTree. + if (Node->getIDom()->getBlock() != NearestDom) { + DT->changeImmediateDominator(BB, NearestDom); + Changed = true; + } + } + + // If the dominator changed, this may have an effect on other + // predecessors, continue until we reach a fixpoint. + } while (Changed); } } @@ -452,6 +489,8 @@ bool LoopRotate::rotateLoop(Loop *L) { // emitted code isn't too gross in this common case. MergeBlockIntoPredecessor(OrigHeader, this); + DEBUG(dbgs() << "LoopRotation: into "; L->dump()); + ++NumRotated; return true; } diff --git a/contrib/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/contrib/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp index b14a713..958348d 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp @@ -54,7 +54,7 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "loop-reduce" -#include "llvm/Transforms/Scalar.h" +#include "llvm/AddressingMode.h" #include "llvm/Constants.h" #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" @@ -64,6 +64,7 @@ #include "llvm/Analysis/LoopPass.h" #include "llvm/Analysis/ScalarEvolutionExpander.h" #include "llvm/Assembly/Writer.h" +#include "llvm/Transforms/Scalar.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/ADT/SmallBitVector.h" @@ -121,9 +122,11 @@ void RegSortData::print(raw_ostream &OS) const { OS << "[NumUses=" << UsedByIndices.count() << ']'; } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void RegSortData::dump() const { print(errs()); errs() << '\n'; } +#endif namespace { @@ -223,7 +226,7 @@ namespace { struct Formula { /// AM - This is used to represent complex addressing, as well as other kinds /// of interesting uses. - TargetLowering::AddrMode AM; + AddrMode AM; /// BaseRegs - The list of "base" registers for this use. When this is /// non-empty, AM.HasBaseReg should be set to true. @@ -414,9 +417,11 @@ void Formula::print(raw_ostream &OS) const { } } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void Formula::dump() const { print(errs()); errs() << '\n'; } +#endif /// isAddRecSExtable - Return true if the given addrec can be sign-extended /// without changing its value. @@ -738,7 +743,8 @@ DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { bool Changed = false; while (!DeadInsts.empty()) { - Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()); + Value *V = DeadInsts.pop_back_val(); + Instruction *I = dyn_cast_or_null<Instruction>(V); if (I == 0 || !isInstructionTriviallyDead(I)) continue; @@ -973,9 +979,11 @@ void Cost::print(raw_ostream &OS) const { OS << ", plus " << SetupCost << " setup cost"; } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void Cost::dump() const { print(errs()); errs() << '\n'; } +#endif namespace { @@ -1059,9 +1067,11 @@ void LSRFixup::print(raw_ostream &OS) const { OS << ", Offset=" << Offset; } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void LSRFixup::dump() const { print(errs()); errs() << '\n'; } +#endif namespace { @@ -1251,14 +1261,16 @@ void LSRUse::print(raw_ostream &OS) const { OS << ", widest fixup type: " << *WidestFixupType; } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void LSRUse::dump() const { print(errs()); errs() << '\n'; } +#endif /// isLegalUse - Test whether the use described by AM is "legal", meaning it can /// be completely folded into the user instruction at isel time. This includes /// address-mode folding and special icmp tricks. -static bool isLegalUse(const TargetLowering::AddrMode &AM, +static bool isLegalUse(const AddrMode &AM, LSRUse::KindType Kind, Type *AccessTy, const TargetLowering *TLI) { switch (Kind) { @@ -1315,7 +1327,7 @@ static bool isLegalUse(const TargetLowering::AddrMode &AM, llvm_unreachable("Invalid LSRUse Kind!"); } -static bool isLegalUse(TargetLowering::AddrMode AM, +static bool isLegalUse(AddrMode AM, int64_t MinOffset, int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy, const TargetLowering *TLI) { @@ -1346,7 +1358,7 @@ static bool isAlwaysFoldable(int64_t BaseOffs, // Conservatively, create an address with an immediate and a // base and a scale. - TargetLowering::AddrMode AM; + AddrMode AM; AM.BaseOffs = BaseOffs; AM.BaseGV = BaseGV; AM.HasBaseReg = HasBaseReg; @@ -1384,7 +1396,7 @@ static bool isAlwaysFoldable(const SCEV *S, // Conservatively, create an address with an immediate and a // base and a scale. - TargetLowering::AddrMode AM; + AddrMode AM; AM.BaseOffs = BaseOffs; AM.BaseGV = BaseGV; AM.HasBaseReg = HasBaseReg; @@ -2009,7 +2021,7 @@ LSRInstance::OptimizeLoopTermCond() { goto decline_post_inc; // Check for possible scaled-address reuse. Type *AccessTy = getAccessType(UI->getUser()); - TargetLowering::AddrMode AM; + AddrMode AM; AM.Scale = C->getSExtValue(); if (TLI->isLegalAddressingMode(AM, AccessTy)) goto decline_post_inc; @@ -3435,9 +3447,11 @@ void WorkItem::print(raw_ostream &OS) const { << " , add offset " << Imm; } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void WorkItem::dump() const { print(errs()); errs() << '\n'; } +#endif /// GenerateCrossUseConstantOffsets - Look for registers which are a constant /// distance apart and try to form reuse opportunities between them. @@ -4451,17 +4465,21 @@ void LSRInstance::RewriteForPHI(PHINode *PN, SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); NewBB = NewBBs[0]; } - - // If PN is outside of the loop and BB is in the loop, we want to - // move the block to be immediately before the PHI block, not - // immediately after BB. - if (L->contains(BB) && !L->contains(PN)) - NewBB->moveBefore(PN->getParent()); - - // Splitting the edge can reduce the number of PHI entries we have. - e = PN->getNumIncomingValues(); - BB = NewBB; - i = PN->getBasicBlockIndex(BB); + // If NewBB==NULL, then SplitCriticalEdge refused to split because all + // phi predecessors are identical. The simple thing to do is skip + // splitting in this case rather than complicate the API. + if (NewBB) { + // If PN is outside of the loop and BB is in the loop, we want to + // move the block to be immediately before the PHI block, not + // immediately after BB. + if (L->contains(BB) && !L->contains(PN)) + NewBB->moveBefore(PN->getParent()); + + // Splitting the edge can reduce the number of PHI entries we have. + e = PN->getNumIncomingValues(); + BB = NewBB; + i = PN->getBasicBlockIndex(BB); + } } } @@ -4730,9 +4748,11 @@ void LSRInstance::print(raw_ostream &OS) const { print_uses(OS); } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void LSRInstance::dump() const { print(errs()); errs() << '\n'; } +#endif namespace { diff --git a/contrib/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp b/contrib/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp index 09a186f..0d781ac 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp @@ -22,7 +22,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/UnrollLoop.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include <climits> using namespace llvm; @@ -113,7 +113,7 @@ Pass *llvm::createLoopUnrollPass(int Threshold, int Count, int AllowPartial) { /// ApproximateLoopSize - Approximate the size of the loop. static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls, - const TargetData *TD) { + const DataLayout *TD) { CodeMetrics Metrics; for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E; ++I) @@ -145,7 +145,8 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { // not user specified. unsigned Threshold = CurrentThreshold; if (!UserThreshold && - Header->getParent()->hasFnAttr(Attribute::OptimizeForSize)) + Header->getParent()->getFnAttributes(). + hasAttribute(Attributes::OptimizeForSize)) Threshold = OptSizeUnrollThreshold; // Find trip count and trip multiple if count is not available @@ -178,7 +179,7 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { // Enforce the threshold. if (Threshold != NoThreshold) { - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); unsigned NumInlineCandidates; unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates, TD); DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); diff --git a/contrib/llvm/lib/Transforms/Scalar/LoopUnswitch.cpp b/contrib/llvm/lib/Transforms/Scalar/LoopUnswitch.cpp index 58f7739..047b43e 100644 --- a/contrib/llvm/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -638,7 +638,8 @@ bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { // Check to see if it would be profitable to unswitch current loop. // Do not do non-trivial unswitch while optimizing for size. - if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) + if (OptimizeForSize || + F->getFnAttributes().hasAttribute(Attributes::OptimizeForSize)) return false; UnswitchNontrivialCondition(LoopCond, Val, currentLoop); @@ -906,13 +907,9 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, /// specified. static void RemoveFromWorklist(Instruction *I, std::vector<Instruction*> &Worklist) { - std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), - Worklist.end(), I); - while (WI != Worklist.end()) { - unsigned Offset = WI-Worklist.begin(); - Worklist.erase(WI); - WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); - } + + Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I), + Worklist.end()); } /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the diff --git a/contrib/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/contrib/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp index 2a5ee33..517657cf 100644 --- a/contrib/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp @@ -27,7 +27,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" #include <list> @@ -38,8 +38,8 @@ STATISTIC(NumMemSetInfer, "Number of memsets inferred"); STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy"); STATISTIC(NumCpyToSet, "Number of memcpys converted to memset"); -static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, - bool &VariableIdxFound, const TargetData &TD){ +static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, + bool &VariableIdxFound, const DataLayout &TD){ // Skip over the first indices. gep_type_iterator GTI = gep_type_begin(GEP); for (unsigned i = 1; i != Idx; ++i, ++GTI) @@ -72,11 +72,11 @@ static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, /// constant offset, and return that constant offset. For example, Ptr1 might /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, - const TargetData &TD) { + const DataLayout &TD) { Ptr1 = Ptr1->stripPointerCasts(); Ptr2 = Ptr2->stripPointerCasts(); - GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); - GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); + GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1); + GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2); bool VariableIdxFound = false; @@ -141,12 +141,12 @@ struct MemsetRange { /// TheStores - The actual stores that make up this range. SmallVector<Instruction*, 16> TheStores; - bool isProfitableToUseMemset(const TargetData &TD) const; + bool isProfitableToUseMemset(const DataLayout &TD) const; }; } // end anon namespace -bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { +bool MemsetRange::isProfitableToUseMemset(const DataLayout &TD) const { // If we found more than 4 stores to merge or 16 bytes, use memset. if (TheStores.size() >= 4 || End-Start >= 16) return true; @@ -192,9 +192,9 @@ class MemsetRanges { /// because each element is relatively large and expensive to copy. std::list<MemsetRange> Ranges; typedef std::list<MemsetRange>::iterator range_iterator; - const TargetData &TD; + const DataLayout &TD; public: - MemsetRanges(const TargetData &td) : TD(td) {} + MemsetRanges(const DataLayout &td) : TD(td) {} typedef std::list<MemsetRange>::const_iterator const_iterator; const_iterator begin() const { return Ranges.begin(); } @@ -302,7 +302,7 @@ namespace { class MemCpyOpt : public FunctionPass { MemoryDependenceAnalysis *MD; TargetLibraryInfo *TLI; - const TargetData *TD; + const DataLayout *TD; public: static char ID; // Pass identification, replacement for typeid MemCpyOpt() : FunctionPass(ID) { @@ -332,7 +332,7 @@ namespace { bool processMemCpy(MemCpyInst *M); bool processMemMove(MemMoveInst *M); bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc, - uint64_t cpyLen, CallInst *C); + uint64_t cpyLen, unsigned cpyAlign, CallInst *C); bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, uint64_t MSize); bool processByValArgument(CallSite CS, unsigned ArgNo); @@ -509,10 +509,18 @@ bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { } if (C) { + unsigned storeAlign = SI->getAlignment(); + if (!storeAlign) + storeAlign = TD->getABITypeAlignment(SI->getOperand(0)->getType()); + unsigned loadAlign = LI->getAlignment(); + if (!loadAlign) + loadAlign = TD->getABITypeAlignment(LI->getType()); + bool changed = performCallSlotOptzn(LI, SI->getPointerOperand()->stripPointerCasts(), LI->getPointerOperand()->stripPointerCasts(), - TD->getTypeStoreSize(SI->getOperand(0)->getType()), C); + TD->getTypeStoreSize(SI->getOperand(0)->getType()), + std::min(storeAlign, loadAlign), C); if (changed) { MD->removeInstruction(SI); SI->eraseFromParent(); @@ -559,7 +567,8 @@ bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) { /// the call write its result directly into the destination of the memcpy. bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, Value *cpyDest, Value *cpySrc, - uint64_t cpyLen, CallInst *C) { + uint64_t cpyLen, unsigned cpyAlign, + CallInst *C) { // The general transformation to keep in mind is // // call @func(..., src, ...) @@ -625,6 +634,16 @@ bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, return false; } + // Check that dest points to memory that is at least as aligned as src. + unsigned srcAlign = srcAlloca->getAlignment(); + if (!srcAlign) + srcAlign = TD->getABITypeAlignment(srcAlloca->getAllocatedType()); + bool isDestSufficientlyAligned = srcAlign <= cpyAlign; + // If dest is not aligned enough and we can't increase its alignment then + // bail out. + if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest)) + return false; + // Check that src is not accessed except via the call and the memcpy. This // guarantees that it holds only undefined values when passed in (so the final // memcpy can be dropped), that it is not read or written between the call and @@ -673,20 +692,26 @@ bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, bool changedArgument = false; for (unsigned i = 0; i < CS.arg_size(); ++i) if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { - if (cpySrc->getType() != cpyDest->getType()) - cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), - cpyDest->getName(), C); + Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest + : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), + cpyDest->getName(), C); changedArgument = true; - if (CS.getArgument(i)->getType() == cpyDest->getType()) - CS.setArgument(i, cpyDest); + if (CS.getArgument(i)->getType() == Dest->getType()) + CS.setArgument(i, Dest); else - CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, - CS.getArgument(i)->getType(), cpyDest->getName(), C)); + CS.setArgument(i, CastInst::CreatePointerCast(Dest, + CS.getArgument(i)->getType(), Dest->getName(), C)); } if (!changedArgument) return false; + // If the destination wasn't sufficiently aligned then increase its alignment. + if (!isDestSufficientlyAligned) { + assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!"); + cast<AllocaInst>(cpyDest)->setAlignment(srcAlign); + } + // Drop any cached information about the call, because we may have changed // its dependence information by changing its parameter. MD->removeInstruction(C); @@ -813,7 +838,8 @@ bool MemCpyOpt::processMemCpy(MemCpyInst *M) { if (DepInfo.isClobber()) { if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) { if (performCallSlotOptzn(M, M->getDest(), M->getSource(), - CopySize->getZExtValue(), C)) { + CopySize->getZExtValue(), M->getAlignment(), + C)) { MD->removeInstruction(M); M->eraseFromParent(); return true; @@ -974,7 +1000,7 @@ bool MemCpyOpt::iterateOnFunction(Function &F) { bool MemCpyOpt::runOnFunction(Function &F) { bool MadeChange = false; MD = &getAnalysis<MemoryDependenceAnalysis>(); - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); TLI = &getAnalysis<TargetLibraryInfo>(); // If we don't have at least memset and memcpy, there is little point of doing diff --git a/contrib/llvm/lib/Transforms/Scalar/ObjCARC.cpp b/contrib/llvm/lib/Transforms/Scalar/ObjCARC.cpp index 3222f20..dfdf505 100644 --- a/contrib/llvm/lib/Transforms/Scalar/ObjCARC.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/ObjCARC.cpp @@ -29,6 +29,7 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "objc-arc" +#include "llvm/Support/raw_ostream.h" #include "llvm/Support/CommandLine.h" #include "llvm/ADT/DenseMap.h" using namespace llvm; @@ -1120,9 +1121,8 @@ namespace { bool relatedSelect(const SelectInst *A, const Value *B); bool relatedPHI(const PHINode *A, const Value *B); - // Do not implement. - void operator=(const ProvenanceAnalysis &); - ProvenanceAnalysis(const ProvenanceAnalysis &); + void operator=(const ProvenanceAnalysis &) LLVM_DELETED_FUNCTION; + ProvenanceAnalysis(const ProvenanceAnalysis &) LLVM_DELETED_FUNCTION; public: ProvenanceAnalysis() {} @@ -1236,16 +1236,19 @@ bool ProvenanceAnalysis::relatedCheck(const Value *A, const Value *B) { // An ObjC-Identified object can't alias a load if it is never locally stored. if (AIsIdentified) { + // Check for an obvious escape. + if (isa<LoadInst>(B)) + return isStoredObjCPointer(A); if (BIsIdentified) { - // If both pointers have provenance, they can be directly compared. - if (A != B) - return false; - } else { - if (isa<LoadInst>(B)) - return isStoredObjCPointer(A); + // Check for an obvious escape. + if (isa<LoadInst>(A)) + return isStoredObjCPointer(B); + // Both pointers are identified and escapes aren't an evident problem. + return false; } - } else { - if (BIsIdentified && isa<LoadInst>(A)) + } else if (BIsIdentified) { + // Check for an obvious escape. + if (isa<LoadInst>(A)) return isStoredObjCPointer(B); } @@ -1381,9 +1384,6 @@ namespace { /// PtrState - This class summarizes several per-pointer runtime properties /// which are propogated through the flow graph. class PtrState { - /// NestCount - The known minimum level of retain+release nesting. - unsigned NestCount; - /// KnownPositiveRefCount - True if the reference count is known to /// be incremented. bool KnownPositiveRefCount; @@ -1401,7 +1401,7 @@ namespace { /// TODO: Encapsulate this better. RRInfo RRI; - PtrState() : NestCount(0), KnownPositiveRefCount(false), Partial(false), + PtrState() : KnownPositiveRefCount(false), Partial(false), Seq(S_None) {} void SetKnownPositiveRefCount() { @@ -1416,18 +1416,6 @@ namespace { return KnownPositiveRefCount; } - void IncrementNestCount() { - if (NestCount != UINT_MAX) ++NestCount; - } - - void DecrementNestCount() { - if (NestCount != 0) --NestCount; - } - - bool IsKnownNested() const { - return NestCount > 0; - } - void SetSeq(Sequence NewSeq) { Seq = NewSeq; } @@ -1454,7 +1442,6 @@ void PtrState::Merge(const PtrState &Other, bool TopDown) { Seq = MergeSeqs(Seq, Other.Seq, TopDown); KnownPositiveRefCount = KnownPositiveRefCount && Other.KnownPositiveRefCount; - NestCount = std::min(NestCount, Other.NestCount); // We can't merge a plain objc_retain with an objc_retainBlock. if (RRI.IsRetainBlock != Other.RRI.IsRetainBlock) @@ -1610,6 +1597,12 @@ void BBState::MergePred(const BBState &Other) { // loop backedge. Loop backedges are special. TopDownPathCount += Other.TopDownPathCount; + // Check for overflow. If we have overflow, fall back to conservative behavior. + if (TopDownPathCount < Other.TopDownPathCount) { + clearTopDownPointers(); + return; + } + // For each entry in the other set, if our set has an entry with the same key, // merge the entries. Otherwise, copy the entry and merge it with an empty // entry. @@ -1635,6 +1628,12 @@ void BBState::MergeSucc(const BBState &Other) { // loop backedge. Loop backedges are special. BottomUpPathCount += Other.BottomUpPathCount; + // Check for overflow. If we have overflow, fall back to conservative behavior. + if (BottomUpPathCount < Other.BottomUpPathCount) { + clearBottomUpPointers(); + return; + } + // For each entry in the other set, if our set has an entry with the // same key, merge the entries. Otherwise, copy the entry and merge // it with an empty entry. @@ -1789,7 +1788,9 @@ Constant *ObjCARCOpt::getRetainRVCallee(Module *M) { Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); Type *Params[] = { I8X }; FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false); - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); RetainRVCallee = M->getOrInsertFunction("objc_retainAutoreleasedReturnValue", FTy, Attributes); @@ -1803,7 +1804,9 @@ Constant *ObjCARCOpt::getAutoreleaseRVCallee(Module *M) { Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); Type *Params[] = { I8X }; FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false); - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); AutoreleaseRVCallee = M->getOrInsertFunction("objc_autoreleaseReturnValue", FTy, Attributes); @@ -1815,7 +1818,9 @@ Constant *ObjCARCOpt::getReleaseCallee(Module *M) { if (!ReleaseCallee) { LLVMContext &C = M->getContext(); Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) }; - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); ReleaseCallee = M->getOrInsertFunction( "objc_release", @@ -1829,7 +1834,9 @@ Constant *ObjCARCOpt::getRetainCallee(Module *M) { if (!RetainCallee) { LLVMContext &C = M->getContext(); Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) }; - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); RetainCallee = M->getOrInsertFunction( "objc_retain", @@ -1858,7 +1865,9 @@ Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) { if (!AutoreleaseCallee) { LLVMContext &C = M->getContext(); Type *Params[] = { PointerType::getUnqual(Type::getInt8Ty(C)) }; - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); AutoreleaseCallee = M->getOrInsertFunction( "objc_autorelease", @@ -1868,6 +1877,26 @@ Constant *ObjCARCOpt::getAutoreleaseCallee(Module *M) { return AutoreleaseCallee; } +/// IsPotentialUse - Test whether the given value is possible a +/// reference-counted pointer, including tests which utilize AliasAnalysis. +static bool IsPotentialUse(const Value *Op, AliasAnalysis &AA) { + // First make the rudimentary check. + if (!IsPotentialUse(Op)) + return false; + + // Objects in constant memory are not reference-counted. + if (AA.pointsToConstantMemory(Op)) + return false; + + // Pointers in constant memory are not pointing to reference-counted objects. + if (const LoadInst *LI = dyn_cast<LoadInst>(Op)) + if (AA.pointsToConstantMemory(LI->getPointerOperand())) + return false; + + // Otherwise assume the worst. + return true; +} + /// CanAlterRefCount - Test whether the given instruction can result in a /// reference count modification (positive or negative) for the pointer's /// object. @@ -1894,7 +1923,7 @@ CanAlterRefCount(const Instruction *Inst, const Value *Ptr, for (ImmutableCallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E; ++I) { const Value *Op = *I; - if (IsPotentialUse(Op) && PA.related(Ptr, Op)) + if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op)) return true; } return false; @@ -1919,14 +1948,14 @@ CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA, // Comparing a pointer with null, or any other constant, isn't really a use, // because we don't care what the pointer points to, or about the values // of any other dynamic reference-counted pointers. - if (!IsPotentialUse(ICI->getOperand(1))) + if (!IsPotentialUse(ICI->getOperand(1), *PA.getAA())) return false; } else if (ImmutableCallSite CS = static_cast<const Value *>(Inst)) { // For calls, just check the arguments (and not the callee operand). for (ImmutableCallSite::arg_iterator OI = CS.arg_begin(), OE = CS.arg_end(); OI != OE; ++OI) { const Value *Op = *OI; - if (IsPotentialUse(Op) && PA.related(Ptr, Op)) + if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op)) return true; } return false; @@ -1936,14 +1965,14 @@ CanUse(const Instruction *Inst, const Value *Ptr, ProvenanceAnalysis &PA, const Value *Op = GetUnderlyingObjCPtr(SI->getPointerOperand()); // If we can't tell what the underlying object was, assume there is a // dependence. - return IsPotentialUse(Op) && PA.related(Op, Ptr); + return IsPotentialUse(Op, *PA.getAA()) && PA.related(Op, Ptr); } // Check each operand for a match. for (User::const_op_iterator OI = Inst->op_begin(), OE = Inst->op_end(); OI != OE; ++OI) { const Value *Op = *OI; - if (IsPotentialUse(Op) && PA.related(Ptr, Op)) + if (IsPotentialUse(Op, *PA.getAA()) && PA.related(Ptr, Op)) return true; } return false; @@ -2612,11 +2641,11 @@ ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst, MDNode *ReleaseMetadata = Inst->getMetadata(ImpreciseReleaseMDKind); S.ResetSequenceProgress(ReleaseMetadata ? S_MovableRelease : S_Release); S.RRI.ReleaseMetadata = ReleaseMetadata; - S.RRI.KnownSafe = S.IsKnownNested() || S.IsKnownIncremented(); + S.RRI.KnownSafe = S.IsKnownIncremented(); S.RRI.IsTailCallRelease = cast<CallInst>(Inst)->isTailCall(); S.RRI.Calls.insert(Inst); - S.IncrementNestCount(); + S.SetKnownPositiveRefCount(); break; } case IC_RetainBlock: @@ -2631,7 +2660,6 @@ ObjCARCOpt::VisitInstructionBottomUp(Instruction *Inst, PtrState &S = MyStates.getPtrBottomUpState(Arg); S.SetKnownPositiveRefCount(); - S.DecrementNestCount(); switch (S.GetSeq()) { case S_Stop: @@ -2747,8 +2775,9 @@ ObjCARCOpt::VisitBottomUp(BasicBlock *BB, // Merge the states from each successor to compute the initial state // for the current block. - for (BBState::edge_iterator SI(MyStates.succ_begin()), - SE(MyStates.succ_end()); SI != SE; ++SI) { + BBState::edge_iterator SI(MyStates.succ_begin()), + SE(MyStates.succ_end()); + if (SI != SE) { const BasicBlock *Succ = *SI; DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Succ); assert(I != BBStates.end()); @@ -2760,7 +2789,6 @@ ObjCARCOpt::VisitBottomUp(BasicBlock *BB, assert(I != BBStates.end()); MyStates.MergeSucc(I->second); } - break; } // Visit all the instructions, bottom-up. @@ -2823,12 +2851,11 @@ ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst, S.ResetSequenceProgress(S_Retain); S.RRI.IsRetainBlock = Class == IC_RetainBlock; - // Don't check S.IsKnownIncremented() here because it's not sufficient. - S.RRI.KnownSafe = S.IsKnownNested(); + S.RRI.KnownSafe = S.IsKnownIncremented(); S.RRI.Calls.insert(Inst); } - S.IncrementNestCount(); + S.SetKnownPositiveRefCount(); // A retain can be a potential use; procede to the generic checking // code below. @@ -2838,7 +2865,7 @@ ObjCARCOpt::VisitInstructionTopDown(Instruction *Inst, Arg = GetObjCArg(Inst); PtrState &S = MyStates.getPtrTopDownState(Arg); - S.DecrementNestCount(); + S.ClearRefCount(); switch (S.GetSeq()) { case S_Retain: @@ -2935,8 +2962,9 @@ ObjCARCOpt::VisitTopDown(BasicBlock *BB, // Merge the states from each predecessor to compute the initial state // for the current block. - for (BBState::edge_iterator PI(MyStates.pred_begin()), - PE(MyStates.pred_end()); PI != PE; ++PI) { + BBState::edge_iterator PI(MyStates.pred_begin()), + PE(MyStates.pred_end()); + if (PI != PE) { const BasicBlock *Pred = *PI; DenseMap<const BasicBlock *, BBState>::iterator I = BBStates.find(Pred); assert(I != BBStates.end()); @@ -2948,7 +2976,6 @@ ObjCARCOpt::VisitTopDown(BasicBlock *BB, assert(I != BBStates.end()); MyStates.MergePred(I->second); } - break; } // Visit all the instructions, top-down. @@ -3532,19 +3559,19 @@ bool ObjCARCOpt::OptimizeSequences(Function &F) { } /// OptimizeReturns - Look for this pattern: -/// +/// \code /// %call = call i8* @something(...) /// %2 = call i8* @objc_retain(i8* %call) /// %3 = call i8* @objc_autorelease(i8* %2) /// ret i8* %3 -/// +/// \endcode /// And delete the retain and autorelease. /// /// Otherwise if it's just this: -/// +/// \code /// %3 = call i8* @objc_autorelease(i8* %2) /// ret i8* %3 -/// +/// \endcode /// convert the autorelease to autoreleaseRV. void ObjCARCOpt::OptimizeReturns(Function &F) { if (!F.getReturnType()->isPointerTy()) @@ -3814,8 +3841,9 @@ Constant *ObjCARCContract::getStoreStrongCallee(Module *M) { Type *Params[] = { I8XX, I8X }; AttrListPtr Attributes = AttrListPtr() - .addAttr(~0u, Attribute::NoUnwind) - .addAttr(1, Attribute::NoCapture); + .addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)) + .addAttr(M->getContext(), 1, Attributes::get(C, Attributes::NoCapture)); StoreStrongCallee = M->getOrInsertFunction( @@ -3832,7 +3860,9 @@ Constant *ObjCARCContract::getRetainAutoreleaseCallee(Module *M) { Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); Type *Params[] = { I8X }; FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false); - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); RetainAutoreleaseCallee = M->getOrInsertFunction("objc_retainAutorelease", FTy, Attributes); } @@ -3845,7 +3875,9 @@ Constant *ObjCARCContract::getRetainAutoreleaseRVCallee(Module *M) { Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C)); Type *Params[] = { I8X }; FunctionType *FTy = FunctionType::get(I8X, Params, /*isVarArg=*/false); - AttrListPtr Attributes = AttrListPtr().addAttr(~0u, Attribute::NoUnwind); + AttrListPtr Attributes = + AttrListPtr().addAttr(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::get(C, Attributes::NoUnwind)); RetainAutoreleaseRVCallee = M->getOrInsertFunction("objc_retainAutoreleaseReturnValue", FTy, Attributes); diff --git a/contrib/llvm/lib/Transforms/Scalar/Reassociate.cpp b/contrib/llvm/lib/Transforms/Scalar/Reassociate.cpp index 09687d8..7a40797 100644 --- a/contrib/llvm/lib/Transforms/Scalar/Reassociate.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/Reassociate.cpp @@ -339,36 +339,6 @@ static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { } } -/// EvaluateRepeatedConstant - Compute C op C op ... op C where the constant C -/// is repeated Weight times. -static Constant *EvaluateRepeatedConstant(unsigned Opcode, Constant *C, - APInt Weight) { - // For addition the result can be efficiently computed as the product of the - // constant and the weight. - if (Opcode == Instruction::Add) - return ConstantExpr::getMul(C, ConstantInt::get(C->getContext(), Weight)); - - // The weight might be huge, so compute by repeated squaring to ensure that - // compile time is proportional to the logarithm of the weight. - Constant *Result = 0; - Constant *Power = C; // Successively C, C op C, (C op C) op (C op C) etc. - // Visit the bits in Weight. - while (Weight != 0) { - // If the current bit in Weight is non-zero do Result = Result op Power. - if (Weight[0]) - Result = Result ? ConstantExpr::get(Opcode, Result, Power) : Power; - // Move on to the next bit if any more are non-zero. - Weight = Weight.lshr(1); - if (Weight.isMinValue()) - break; - // Square the power. - Power = ConstantExpr::get(Opcode, Power, Power); - } - - assert(Result && "Only positive weights supported!"); - return Result; -} - typedef std::pair<Value*, APInt> RepeatedValue; /// LinearizeExprTree - Given an associative binary expression, return the leaf @@ -382,9 +352,7 @@ typedef std::pair<Value*, APInt> RepeatedValue; /// op /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times /// -/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct, and -/// they are all non-constant except possibly for the last one, which if it is -/// constant will have weight one (Ops[N].second === 1). +/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. /// /// This routine may modify the function, in which case it returns 'true'. The /// changes it makes may well be destructive, changing the value computed by 'I' @@ -604,7 +572,6 @@ static bool LinearizeExprTree(BinaryOperator *I, // The leaves, repeated according to their weights, represent the linearized // form of the expression. - Constant *Cst = 0; // Accumulate constants here. for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { Value *V = LeafOrder[i]; LeafMap::iterator It = Leaves.find(V); @@ -618,31 +585,14 @@ static bool LinearizeExprTree(BinaryOperator *I, continue; // Ensure the leaf is only output once. It->second = 0; - // Glob all constants together into Cst. - if (Constant *C = dyn_cast<Constant>(V)) { - C = EvaluateRepeatedConstant(Opcode, C, Weight); - Cst = Cst ? ConstantExpr::get(Opcode, Cst, C) : C; - continue; - } - // Add non-constant Ops.push_back(std::make_pair(V, Weight)); } - // Add any constants back into Ops, all globbed together and reduced to having - // weight 1 for the convenience of users. - Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); - if (Cst && Cst != Identity) { - // If combining multiple constants resulted in the absorber then the entire - // expression must evaluate to the absorber. - if (Cst == Absorber) - Ops.clear(); - Ops.push_back(std::make_pair(Cst, APInt(Bitwidth, 1))); - } - // For nilpotent operations or addition there may be no operands, for example // because the expression was "X xor X" or consisted of 2^Bitwidth additions: // in both cases the weight reduces to 0 causing the value to be skipped. if (Ops.empty()) { + Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); assert(Identity && "Associative operation without identity!"); Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1))); } @@ -656,8 +606,8 @@ void Reassociate::RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops) { assert(Ops.size() > 1 && "Single values should be used directly!"); - // Since our optimizations never increase the number of operations, the new - // expression can always be written by reusing the existing binary operators + // Since our optimizations should never increase the number of operations, the + // new expression can usually be written reusing the existing binary operators // from the original expression tree, without creating any new instructions, // though the rewritten expression may have a completely different topology. // We take care to not change anything if the new expression will be the same @@ -671,6 +621,20 @@ void Reassociate::RewriteExprTree(BinaryOperator *I, unsigned Opcode = I->getOpcode(); BinaryOperator *Op = I; + /// NotRewritable - The operands being written will be the leaves of the new + /// expression and must not be used as inner nodes (via NodesToRewrite) by + /// mistake. Inner nodes are always reassociable, and usually leaves are not + /// (if they were they would have been incorporated into the expression and so + /// would not be leaves), so most of the time there is no danger of this. But + /// in rare cases a leaf may become reassociable if an optimization kills uses + /// of it, or it may momentarily become reassociable during rewriting (below) + /// due it being removed as an operand of one of its uses. Ensure that misuse + /// of leaf nodes as inner nodes cannot occur by remembering all of the future + /// leaves and refusing to reuse any of them as inner nodes. + SmallPtrSet<Value*, 8> NotRewritable; + for (unsigned i = 0, e = Ops.size(); i != e; ++i) + NotRewritable.insert(Ops[i].Op); + // ExpressionChanged - Non-null if the rewritten expression differs from the // original in some non-trivial way, requiring the clearing of optional flags. // Flags are cleared from the operator in ExpressionChanged up to I inclusive. @@ -703,12 +667,14 @@ void Reassociate::RewriteExprTree(BinaryOperator *I, // the old operands with the new ones. DEBUG(dbgs() << "RA: " << *Op << '\n'); if (NewLHS != OldLHS) { - if (BinaryOperator *BO = isReassociableOp(OldLHS, Opcode)) + BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); + if (BO && !NotRewritable.count(BO)) NodesToRewrite.push_back(BO); Op->setOperand(0, NewLHS); } if (NewRHS != OldRHS) { - if (BinaryOperator *BO = isReassociableOp(OldRHS, Opcode)) + BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); + if (BO && !NotRewritable.count(BO)) NodesToRewrite.push_back(BO); Op->setOperand(1, NewRHS); } @@ -732,7 +698,8 @@ void Reassociate::RewriteExprTree(BinaryOperator *I, Op->swapOperands(); } else { // Overwrite with the new right-hand side. - if (BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode)) + BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); + if (BO && !NotRewritable.count(BO)) NodesToRewrite.push_back(BO); Op->setOperand(1, NewRHS); ExpressionChanged = Op; @@ -745,7 +712,8 @@ void Reassociate::RewriteExprTree(BinaryOperator *I, // Now deal with the left-hand side. If this is already an operation node // from the original expression then just rewrite the rest of the expression // into it. - if (BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode)) { + BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); + if (BO && !NotRewritable.count(BO)) { Op = BO; continue; } @@ -1446,9 +1414,26 @@ Value *Reassociate::OptimizeExpression(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops) { // Now that we have the linearized expression tree, try to optimize it. // Start by folding any constants that we found. - if (Ops.size() == 1) return Ops[0].Op; - + Constant *Cst = 0; unsigned Opcode = I->getOpcode(); + while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { + Constant *C = cast<Constant>(Ops.pop_back_val().Op); + Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; + } + // If there was nothing but constants then we are done. + if (Ops.empty()) + return Cst; + + // Put the combined constant back at the end of the operand list, except if + // there is no point. For example, an add of 0 gets dropped here, while a + // multiplication by zero turns the whole expression into zero. + if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { + if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) + return Cst; + Ops.push_back(ValueEntry(0, Cst)); + } + + if (Ops.size() == 1) return Ops[0].Op; // Handle destructive annihilation due to identities between elements in the // argument list here. diff --git a/contrib/llvm/lib/Transforms/Scalar/SCCP.cpp b/contrib/llvm/lib/Transforms/Scalar/SCCP.cpp index 2c39aab..686520e 100644 --- a/contrib/llvm/lib/Transforms/Scalar/SCCP.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/SCCP.cpp @@ -26,7 +26,7 @@ #include "llvm/Pass.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Transforms/Utils/Local.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Support/CallSite.h" #include "llvm/Support/Debug.h" @@ -153,7 +153,7 @@ namespace { /// Constant Propagation. /// class SCCPSolver : public InstVisitor<SCCPSolver> { - const TargetData *TD; + const DataLayout *TD; const TargetLibraryInfo *TLI; SmallPtrSet<BasicBlock*, 8> BBExecutable; // The BBs that are executable. DenseMap<Value*, LatticeVal> ValueState; // The state each value is in. @@ -205,7 +205,7 @@ class SCCPSolver : public InstVisitor<SCCPSolver> { typedef std::pair<BasicBlock*, BasicBlock*> Edge; DenseSet<Edge> KnownFeasibleEdges; public: - SCCPSolver(const TargetData *td, const TargetLibraryInfo *tli) + SCCPSolver(const DataLayout *td, const TargetLibraryInfo *tli) : TD(td), TLI(tli) {} /// MarkBlockExecutable - This method can be used by clients to mark all of @@ -1564,7 +1564,7 @@ static void DeleteInstructionInBlock(BasicBlock *BB) { // bool SCCP::runOnFunction(Function &F) { DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n"); - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>(); SCCPSolver Solver(TD, TLI); @@ -1693,7 +1693,7 @@ static bool AddressIsTaken(const GlobalValue *GV) { } bool IPSCCP::runOnModule(Module &M) { - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>(); SCCPSolver Solver(TD, TLI); diff --git a/contrib/llvm/lib/Transforms/Scalar/SROA.cpp b/contrib/llvm/lib/Transforms/Scalar/SROA.cpp new file mode 100644 index 0000000..ccc2f7a --- /dev/null +++ b/contrib/llvm/lib/Transforms/Scalar/SROA.cpp @@ -0,0 +1,3697 @@ +//===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +/// \file +/// This transformation implements the well known scalar replacement of +/// aggregates transformation. It tries to identify promotable elements of an +/// aggregate alloca, and promote them to registers. It will also try to +/// convert uses of an element (or set of elements) of an alloca into a vector +/// or bitfield-style integer scalar if appropriate. +/// +/// It works to do this with minimal slicing of the alloca so that regions +/// which are merely transferred in and out of external memory remain unchanged +/// and are not decomposed to scalar code. +/// +/// Because this also performs alloca promotion, it can be thought of as also +/// serving the purpose of SSA formation. The algorithm iterates on the +/// function until all opportunities for promotion have been realized. +/// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "sroa" +#include "llvm/Transforms/Scalar.h" +#include "llvm/Constants.h" +#include "llvm/DIBuilder.h" +#include "llvm/DebugInfo.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Function.h" +#include "llvm/IRBuilder.h" +#include "llvm/Instructions.h" +#include "llvm/IntrinsicInst.h" +#include "llvm/LLVMContext.h" +#include "llvm/Module.h" +#include "llvm/Operator.h" +#include "llvm/Pass.h" +#include "llvm/ADT/SetVector.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/Analysis/Dominators.h" +#include "llvm/Analysis/Loads.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/GetElementPtrTypeIterator.h" +#include "llvm/Support/InstVisitor.h" +#include "llvm/Support/MathExtras.h" +#include "llvm/Support/raw_ostream.h" +#include "llvm/DataLayout.h" +#include "llvm/Transforms/Utils/Local.h" +#include "llvm/Transforms/Utils/PromoteMemToReg.h" +#include "llvm/Transforms/Utils/SSAUpdater.h" +using namespace llvm; + +STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement"); +STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced"); +STATISTIC(NumPromoted, "Number of allocas promoted to SSA values"); +STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion"); +STATISTIC(NumDeleted, "Number of instructions deleted"); +STATISTIC(NumVectorized, "Number of vectorized aggregates"); + +/// Hidden option to force the pass to not use DomTree and mem2reg, instead +/// forming SSA values through the SSAUpdater infrastructure. +static cl::opt<bool> +ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden); + +namespace { +/// \brief Alloca partitioning representation. +/// +/// This class represents a partitioning of an alloca into slices, and +/// information about the nature of uses of each slice of the alloca. The goal +/// is that this information is sufficient to decide if and how to split the +/// alloca apart and replace slices with scalars. It is also intended that this +/// structure can capture the relevant information needed both to decide about +/// and to enact these transformations. +class AllocaPartitioning { +public: + /// \brief A common base class for representing a half-open byte range. + struct ByteRange { + /// \brief The beginning offset of the range. + uint64_t BeginOffset; + + /// \brief The ending offset, not included in the range. + uint64_t EndOffset; + + ByteRange() : BeginOffset(), EndOffset() {} + ByteRange(uint64_t BeginOffset, uint64_t EndOffset) + : BeginOffset(BeginOffset), EndOffset(EndOffset) {} + + /// \brief Support for ordering ranges. + /// + /// This provides an ordering over ranges such that start offsets are + /// always increasing, and within equal start offsets, the end offsets are + /// decreasing. Thus the spanning range comes first in a cluster with the + /// same start position. + bool operator<(const ByteRange &RHS) const { + if (BeginOffset < RHS.BeginOffset) return true; + if (BeginOffset > RHS.BeginOffset) return false; + if (EndOffset > RHS.EndOffset) return true; + return false; + } + + /// \brief Support comparison with a single offset to allow binary searches. + friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) { + return LHS.BeginOffset < RHSOffset; + } + + friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset, + const ByteRange &RHS) { + return LHSOffset < RHS.BeginOffset; + } + + bool operator==(const ByteRange &RHS) const { + return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset; + } + bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); } + }; + + /// \brief A partition of an alloca. + /// + /// This structure represents a contiguous partition of the alloca. These are + /// formed by examining the uses of the alloca. During formation, they may + /// overlap but once an AllocaPartitioning is built, the Partitions within it + /// are all disjoint. + struct Partition : public ByteRange { + /// \brief Whether this partition is splittable into smaller partitions. + /// + /// We flag partitions as splittable when they are formed entirely due to + /// accesses by trivially splittable operations such as memset and memcpy. + bool IsSplittable; + + /// \brief Test whether a partition has been marked as dead. + bool isDead() const { + if (BeginOffset == UINT64_MAX) { + assert(EndOffset == UINT64_MAX); + return true; + } + return false; + } + + /// \brief Kill a partition. + /// This is accomplished by setting both its beginning and end offset to + /// the maximum possible value. + void kill() { + assert(!isDead() && "He's Dead, Jim!"); + BeginOffset = EndOffset = UINT64_MAX; + } + + Partition() : ByteRange(), IsSplittable() {} + Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable) + : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {} + }; + + /// \brief A particular use of a partition of the alloca. + /// + /// This structure is used to associate uses of a partition with it. They + /// mark the range of bytes which are referenced by a particular instruction, + /// and includes a handle to the user itself and the pointer value in use. + /// The bounds of these uses are determined by intersecting the bounds of the + /// memory use itself with a particular partition. As a consequence there is + /// intentionally overlap between various uses of the same partition. + struct PartitionUse : public ByteRange { + /// \brief The use in question. Provides access to both user and used value. + /// + /// Note that this may be null if the partition use is *dead*, that is, it + /// should be ignored. + Use *U; + + PartitionUse() : ByteRange(), U() {} + PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U) + : ByteRange(BeginOffset, EndOffset), U(U) {} + }; + + /// \brief Construct a partitioning of a particular alloca. + /// + /// Construction does most of the work for partitioning the alloca. This + /// performs the necessary walks of users and builds a partitioning from it. + AllocaPartitioning(const DataLayout &TD, AllocaInst &AI); + + /// \brief Test whether a pointer to the allocation escapes our analysis. + /// + /// If this is true, the partitioning is never fully built and should be + /// ignored. + bool isEscaped() const { return PointerEscapingInstr; } + + /// \brief Support for iterating over the partitions. + /// @{ + typedef SmallVectorImpl<Partition>::iterator iterator; + iterator begin() { return Partitions.begin(); } + iterator end() { return Partitions.end(); } + + typedef SmallVectorImpl<Partition>::const_iterator const_iterator; + const_iterator begin() const { return Partitions.begin(); } + const_iterator end() const { return Partitions.end(); } + /// @} + + /// \brief Support for iterating over and manipulating a particular + /// partition's uses. + /// + /// The iteration support provided for uses is more limited, but also + /// includes some manipulation routines to support rewriting the uses of + /// partitions during SROA. + /// @{ + typedef SmallVectorImpl<PartitionUse>::iterator use_iterator; + use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); } + use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); } + use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); } + use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); } + + typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator; + const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); } + const_use_iterator use_begin(const_iterator I) const { + return Uses[I - begin()].begin(); + } + const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); } + const_use_iterator use_end(const_iterator I) const { + return Uses[I - begin()].end(); + } + + unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); } + unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); } + const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const { + return Uses[PIdx][UIdx]; + } + const PartitionUse &getUse(const_iterator I, unsigned UIdx) const { + return Uses[I - begin()][UIdx]; + } + + void use_push_back(unsigned Idx, const PartitionUse &PU) { + Uses[Idx].push_back(PU); + } + void use_push_back(const_iterator I, const PartitionUse &PU) { + Uses[I - begin()].push_back(PU); + } + /// @} + + /// \brief Allow iterating the dead users for this alloca. + /// + /// These are instructions which will never actually use the alloca as they + /// are outside the allocated range. They are safe to replace with undef and + /// delete. + /// @{ + typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator; + dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); } + dead_user_iterator dead_user_end() const { return DeadUsers.end(); } + /// @} + + /// \brief Allow iterating the dead expressions referring to this alloca. + /// + /// These are operands which have cannot actually be used to refer to the + /// alloca as they are outside its range and the user doesn't correct for + /// that. These mostly consist of PHI node inputs and the like which we just + /// need to replace with undef. + /// @{ + typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator; + dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); } + dead_op_iterator dead_op_end() const { return DeadOperands.end(); } + /// @} + + /// \brief MemTransferInst auxiliary data. + /// This struct provides some auxiliary data about memory transfer + /// intrinsics such as memcpy and memmove. These intrinsics can use two + /// different ranges within the same alloca, and provide other challenges to + /// correctly represent. We stash extra data to help us untangle this + /// after the partitioning is complete. + struct MemTransferOffsets { + /// The destination begin and end offsets when the destination is within + /// this alloca. If the end offset is zero the destination is not within + /// this alloca. + uint64_t DestBegin, DestEnd; + + /// The source begin and end offsets when the source is within this alloca. + /// If the end offset is zero, the source is not within this alloca. + uint64_t SourceBegin, SourceEnd; + + /// Flag for whether an alloca is splittable. + bool IsSplittable; + }; + MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const { + return MemTransferInstData.lookup(&II); + } + + /// \brief Map from a PHI or select operand back to a partition. + /// + /// When manipulating PHI nodes or selects, they can use more than one + /// partition of an alloca. We store a special mapping to allow finding the + /// partition referenced by each of these operands, if any. + iterator findPartitionForPHIOrSelectOperand(Use *U) { + SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt + = PHIOrSelectOpMap.find(U); + if (MapIt == PHIOrSelectOpMap.end()) + return end(); + + return begin() + MapIt->second.first; + } + + /// \brief Map from a PHI or select operand back to the specific use of + /// a partition. + /// + /// Similar to mapping these operands back to the partitions, this maps + /// directly to the use structure of that partition. + use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) { + SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt + = PHIOrSelectOpMap.find(U); + assert(MapIt != PHIOrSelectOpMap.end()); + return Uses[MapIt->second.first].begin() + MapIt->second.second; + } + + /// \brief Compute a common type among the uses of a particular partition. + /// + /// This routines walks all of the uses of a particular partition and tries + /// to find a common type between them. Untyped operations such as memset and + /// memcpy are ignored. + Type *getCommonType(iterator I) const; + +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) + void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const; + void printUsers(raw_ostream &OS, const_iterator I, + StringRef Indent = " ") const; + void print(raw_ostream &OS) const; + void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const; + void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const; +#endif + +private: + template <typename DerivedT, typename RetT = void> class BuilderBase; + class PartitionBuilder; + friend class AllocaPartitioning::PartitionBuilder; + class UseBuilder; + friend class AllocaPartitioning::UseBuilder; + +#ifndef NDEBUG + /// \brief Handle to alloca instruction to simplify method interfaces. + AllocaInst &AI; +#endif + + /// \brief The instruction responsible for this alloca having no partitioning. + /// + /// When an instruction (potentially) escapes the pointer to the alloca, we + /// store a pointer to that here and abort trying to partition the alloca. + /// This will be null if the alloca is partitioned successfully. + Instruction *PointerEscapingInstr; + + /// \brief The partitions of the alloca. + /// + /// We store a vector of the partitions over the alloca here. This vector is + /// sorted by increasing begin offset, and then by decreasing end offset. See + /// the Partition inner class for more details. Initially (during + /// construction) there are overlaps, but we form a disjoint sequence of + /// partitions while finishing construction and a fully constructed object is + /// expected to always have this as a disjoint space. + SmallVector<Partition, 8> Partitions; + + /// \brief The uses of the partitions. + /// + /// This is essentially a mapping from each partition to a list of uses of + /// that partition. The mapping is done with a Uses vector that has the exact + /// same number of entries as the partition vector. Each entry is itself + /// a vector of the uses. + SmallVector<SmallVector<PartitionUse, 2>, 8> Uses; + + /// \brief Instructions which will become dead if we rewrite the alloca. + /// + /// Note that these are not separated by partition. This is because we expect + /// a partitioned alloca to be completely rewritten or not rewritten at all. + /// If rewritten, all these instructions can simply be removed and replaced + /// with undef as they come from outside of the allocated space. + SmallVector<Instruction *, 8> DeadUsers; + + /// \brief Operands which will become dead if we rewrite the alloca. + /// + /// These are operands that in their particular use can be replaced with + /// undef when we rewrite the alloca. These show up in out-of-bounds inputs + /// to PHI nodes and the like. They aren't entirely dead (there might be + /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we + /// want to swap this particular input for undef to simplify the use lists of + /// the alloca. + SmallVector<Use *, 8> DeadOperands; + + /// \brief The underlying storage for auxiliary memcpy and memset info. + SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData; + + /// \brief A side datastructure used when building up the partitions and uses. + /// + /// This mapping is only really used during the initial building of the + /// partitioning so that we can retain information about PHI and select nodes + /// processed. + SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes; + + /// \brief Auxiliary information for particular PHI or select operands. + SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap; + + /// \brief A utility routine called from the constructor. + /// + /// This does what it says on the tin. It is the key of the alloca partition + /// splitting and merging. After it is called we have the desired disjoint + /// collection of partitions. + void splitAndMergePartitions(); +}; +} + +template <typename DerivedT, typename RetT> +class AllocaPartitioning::BuilderBase + : public InstVisitor<DerivedT, RetT> { +public: + BuilderBase(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P) + : TD(TD), + AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())), + P(P) { + enqueueUsers(AI, 0); + } + +protected: + const DataLayout &TD; + const uint64_t AllocSize; + AllocaPartitioning &P; + + SmallPtrSet<Use *, 8> VisitedUses; + + struct OffsetUse { + Use *U; + int64_t Offset; + }; + SmallVector<OffsetUse, 8> Queue; + + // The active offset and use while visiting. + Use *U; + int64_t Offset; + + void enqueueUsers(Instruction &I, int64_t UserOffset) { + for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); + UI != UE; ++UI) { + if (VisitedUses.insert(&UI.getUse())) { + OffsetUse OU = { &UI.getUse(), UserOffset }; + Queue.push_back(OU); + } + } + } + + bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) { + GEPOffset = Offset; + for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI); + GTI != GTE; ++GTI) { + ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand()); + if (!OpC) + return false; + if (OpC->isZero()) + continue; + + // Handle a struct index, which adds its field offset to the pointer. + if (StructType *STy = dyn_cast<StructType>(*GTI)) { + unsigned ElementIdx = OpC->getZExtValue(); + const StructLayout *SL = TD.getStructLayout(STy); + uint64_t ElementOffset = SL->getElementOffset(ElementIdx); + // Check that we can continue to model this GEP in a signed 64-bit offset. + if (ElementOffset > INT64_MAX || + (GEPOffset >= 0 && + ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) { + DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding " + << "what can be represented in an int64_t!\n" + << " alloca: " << P.AI << "\n"); + return false; + } + if (GEPOffset < 0) + GEPOffset = ElementOffset + (uint64_t)-GEPOffset; + else + GEPOffset += ElementOffset; + continue; + } + + APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits()); + Index *= APInt(Index.getBitWidth(), + TD.getTypeAllocSize(GTI.getIndexedType())); + Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset, + /*isSigned*/true); + // Check if the result can be stored in our int64_t offset. + if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) { + DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding " + << "what can be represented in an int64_t!\n" + << " alloca: " << P.AI << "\n"); + return false; + } + + GEPOffset = Index.getSExtValue(); + } + return true; + } + + Value *foldSelectInst(SelectInst &SI) { + // If the condition being selected on is a constant or the same value is + // being selected between, fold the select. Yes this does (rarely) happen + // early on. + if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition())) + return SI.getOperand(1+CI->isZero()); + if (SI.getOperand(1) == SI.getOperand(2)) { + assert(*U == SI.getOperand(1)); + return SI.getOperand(1); + } + return 0; + } +}; + +/// \brief Builder for the alloca partitioning. +/// +/// This class builds an alloca partitioning by recursively visiting the uses +/// of an alloca and splitting the partitions for each load and store at each +/// offset. +class AllocaPartitioning::PartitionBuilder + : public BuilderBase<PartitionBuilder, bool> { + friend class InstVisitor<PartitionBuilder, bool>; + + SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap; + +public: + PartitionBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P) + : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {} + + /// \brief Run the builder over the allocation. + bool operator()() { + // Note that we have to re-evaluate size on each trip through the loop as + // the queue grows at the tail. + for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) { + U = Queue[Idx].U; + Offset = Queue[Idx].Offset; + if (!visit(cast<Instruction>(U->getUser()))) + return false; + } + return true; + } + +private: + bool markAsEscaping(Instruction &I) { + P.PointerEscapingInstr = &I; + return false; + } + + void insertUse(Instruction &I, int64_t Offset, uint64_t Size, + bool IsSplittable = false) { + // Completely skip uses which have a zero size or don't overlap the + // allocation. + if (Size == 0 || + (Offset >= 0 && (uint64_t)Offset >= AllocSize) || + (Offset < 0 && (uint64_t)-Offset >= Size)) { + DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset + << " which starts past the end of the " << AllocSize + << " byte alloca:\n" + << " alloca: " << P.AI << "\n" + << " use: " << I << "\n"); + return; + } + + // Clamp the start to the beginning of the allocation. + if (Offset < 0) { + DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset + << " to start at the beginning of the alloca:\n" + << " alloca: " << P.AI << "\n" + << " use: " << I << "\n"); + Size -= (uint64_t)-Offset; + Offset = 0; + } + + uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size; + + // Clamp the end offset to the end of the allocation. Note that this is + // formulated to handle even the case where "BeginOffset + Size" overflows. + // NOTE! This may appear superficially to be something we could ignore + // entirely, but that is not so! There may be PHI-node uses where some + // instructions are dead but not others. We can't completely ignore the + // PHI node, and so have to record at least the information here. + assert(AllocSize >= BeginOffset); // Established above. + if (Size > AllocSize - BeginOffset) { + DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset + << " to remain within the " << AllocSize << " byte alloca:\n" + << " alloca: " << P.AI << "\n" + << " use: " << I << "\n"); + EndOffset = AllocSize; + } + + Partition New(BeginOffset, EndOffset, IsSplittable); + P.Partitions.push_back(New); + } + + bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset, + bool IsVolatile) { + uint64_t Size = TD.getTypeStoreSize(Ty); + + // If this memory access can be shown to *statically* extend outside the + // bounds of of the allocation, it's behavior is undefined, so simply + // ignore it. Note that this is more strict than the generic clamping + // behavior of insertUse. We also try to handle cases which might run the + // risk of overflow. + // FIXME: We should instead consider the pointer to have escaped if this + // function is being instrumented for addressing bugs or race conditions. + if (Offset < 0 || (uint64_t)Offset >= AllocSize || + Size > (AllocSize - (uint64_t)Offset)) { + DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte " + << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset + << " which extends past the end of the " << AllocSize + << " byte alloca:\n" + << " alloca: " << P.AI << "\n" + << " use: " << I << "\n"); + return true; + } + + // We allow splitting of loads and stores where the type is an integer type + // and which cover the entire alloca. Such integer loads and stores + // often require decomposition into fine grained loads and stores. + bool IsSplittable = false; + if (IntegerType *ITy = dyn_cast<IntegerType>(Ty)) + IsSplittable = !IsVolatile && ITy->getBitWidth() == AllocSize*8; + + insertUse(I, Offset, Size, IsSplittable); + return true; + } + + bool visitBitCastInst(BitCastInst &BC) { + enqueueUsers(BC, Offset); + return true; + } + + bool visitGetElementPtrInst(GetElementPtrInst &GEPI) { + int64_t GEPOffset; + if (!computeConstantGEPOffset(GEPI, GEPOffset)) + return markAsEscaping(GEPI); + + enqueueUsers(GEPI, GEPOffset); + return true; + } + + bool visitLoadInst(LoadInst &LI) { + assert((!LI.isSimple() || LI.getType()->isSingleValueType()) && + "All simple FCA loads should have been pre-split"); + return handleLoadOrStore(LI.getType(), LI, Offset, LI.isVolatile()); + } + + bool visitStoreInst(StoreInst &SI) { + Value *ValOp = SI.getValueOperand(); + if (ValOp == *U) + return markAsEscaping(SI); + + assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) && + "All simple FCA stores should have been pre-split"); + return handleLoadOrStore(ValOp->getType(), SI, Offset, SI.isVolatile()); + } + + + bool visitMemSetInst(MemSetInst &II) { + assert(II.getRawDest() == *U && "Pointer use is not the destination?"); + ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); + uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset; + insertUse(II, Offset, Size, Length); + return true; + } + + bool visitMemTransferInst(MemTransferInst &II) { + ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); + uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset; + if (!Size) + // Zero-length mem transfer intrinsics can be ignored entirely. + return true; + + MemTransferOffsets &Offsets = P.MemTransferInstData[&II]; + + // Only intrinsics with a constant length can be split. + Offsets.IsSplittable = Length; + + if (*U == II.getRawDest()) { + Offsets.DestBegin = Offset; + Offsets.DestEnd = Offset + Size; + } + if (*U == II.getRawSource()) { + Offsets.SourceBegin = Offset; + Offsets.SourceEnd = Offset + Size; + } + + // If we have set up end offsets for both the source and the destination, + // we have found both sides of this transfer pointing at the same alloca. + bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd; + if (SeenBothEnds && II.getRawDest() != II.getRawSource()) { + unsigned PrevIdx = MemTransferPartitionMap[&II]; + + // Check if the begin offsets match and this is a non-volatile transfer. + // In that case, we can completely elide the transfer. + if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) { + P.Partitions[PrevIdx].kill(); + return true; + } + + // Otherwise we have an offset transfer within the same alloca. We can't + // split those. + P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false; + } else if (SeenBothEnds) { + // Handle the case where this exact use provides both ends of the + // operation. + assert(II.getRawDest() == II.getRawSource()); + + // For non-volatile transfers this is a no-op. + if (!II.isVolatile()) + return true; + + // Otherwise just suppress splitting. + Offsets.IsSplittable = false; + } + + + // Insert the use now that we've fixed up the splittable nature. + insertUse(II, Offset, Size, Offsets.IsSplittable); + + // Setup the mapping from intrinsic to partition of we've not seen both + // ends of this transfer. + if (!SeenBothEnds) { + unsigned NewIdx = P.Partitions.size() - 1; + bool Inserted + = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second; + assert(Inserted && + "Already have intrinsic in map but haven't seen both ends"); + (void)Inserted; + } + + return true; + } + + // Disable SRoA for any intrinsics except for lifetime invariants. + // FIXME: What about debug instrinsics? This matches old behavior, but + // doesn't make sense. + bool visitIntrinsicInst(IntrinsicInst &II) { + if (II.getIntrinsicID() == Intrinsic::lifetime_start || + II.getIntrinsicID() == Intrinsic::lifetime_end) { + ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0)); + uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue()); + insertUse(II, Offset, Size, true); + return true; + } + + return markAsEscaping(II); + } + + Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) { + // We consider any PHI or select that results in a direct load or store of + // the same offset to be a viable use for partitioning purposes. These uses + // are considered unsplittable and the size is the maximum loaded or stored + // size. + SmallPtrSet<Instruction *, 4> Visited; + SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses; + Visited.insert(Root); + Uses.push_back(std::make_pair(cast<Instruction>(*U), Root)); + // If there are no loads or stores, the access is dead. We mark that as + // a size zero access. + Size = 0; + do { + Instruction *I, *UsedI; + llvm::tie(UsedI, I) = Uses.pop_back_val(); + + if (LoadInst *LI = dyn_cast<LoadInst>(I)) { + Size = std::max(Size, TD.getTypeStoreSize(LI->getType())); + continue; + } + if (StoreInst *SI = dyn_cast<StoreInst>(I)) { + Value *Op = SI->getOperand(0); + if (Op == UsedI) + return SI; + Size = std::max(Size, TD.getTypeStoreSize(Op->getType())); + continue; + } + + if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { + if (!GEP->hasAllZeroIndices()) + return GEP; + } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) && + !isa<SelectInst>(I)) { + return I; + } + + for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE; + ++UI) + if (Visited.insert(cast<Instruction>(*UI))) + Uses.push_back(std::make_pair(I, cast<Instruction>(*UI))); + } while (!Uses.empty()); + + return 0; + } + + bool visitPHINode(PHINode &PN) { + // See if we already have computed info on this node. + std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN]; + if (PHIInfo.first) { + PHIInfo.second = true; + insertUse(PN, Offset, PHIInfo.first); + return true; + } + + // Check for an unsafe use of the PHI node. + if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first)) + return markAsEscaping(*EscapingI); + + insertUse(PN, Offset, PHIInfo.first); + return true; + } + + bool visitSelectInst(SelectInst &SI) { + if (Value *Result = foldSelectInst(SI)) { + if (Result == *U) + // If the result of the constant fold will be the pointer, recurse + // through the select as if we had RAUW'ed it. + enqueueUsers(SI, Offset); + + return true; + } + + // See if we already have computed info on this node. + std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI]; + if (SelectInfo.first) { + SelectInfo.second = true; + insertUse(SI, Offset, SelectInfo.first); + return true; + } + + // Check for an unsafe use of the PHI node. + if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first)) + return markAsEscaping(*EscapingI); + + insertUse(SI, Offset, SelectInfo.first); + return true; + } + + /// \brief Disable SROA entirely if there are unhandled users of the alloca. + bool visitInstruction(Instruction &I) { return markAsEscaping(I); } +}; + + +/// \brief Use adder for the alloca partitioning. +/// +/// This class adds the uses of an alloca to all of the partitions which they +/// use. For splittable partitions, this can end up doing essentially a linear +/// walk of the partitions, but the number of steps remains bounded by the +/// total result instruction size: +/// - The number of partitions is a result of the number unsplittable +/// instructions using the alloca. +/// - The number of users of each partition is at worst the total number of +/// splittable instructions using the alloca. +/// Thus we will produce N * M instructions in the end, where N are the number +/// of unsplittable uses and M are the number of splittable. This visitor does +/// the exact same number of updates to the partitioning. +/// +/// In the more common case, this visitor will leverage the fact that the +/// partition space is pre-sorted, and do a logarithmic search for the +/// partition needed, making the total visit a classical ((N + M) * log(N)) +/// complexity operation. +class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> { + friend class InstVisitor<UseBuilder>; + + /// \brief Set to de-duplicate dead instructions found in the use walk. + SmallPtrSet<Instruction *, 4> VisitedDeadInsts; + +public: + UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P) + : BuilderBase<UseBuilder>(TD, AI, P) {} + + /// \brief Run the builder over the allocation. + void operator()() { + // Note that we have to re-evaluate size on each trip through the loop as + // the queue grows at the tail. + for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) { + U = Queue[Idx].U; + Offset = Queue[Idx].Offset; + this->visit(cast<Instruction>(U->getUser())); + } + } + +private: + void markAsDead(Instruction &I) { + if (VisitedDeadInsts.insert(&I)) + P.DeadUsers.push_back(&I); + } + + void insertUse(Instruction &User, int64_t Offset, uint64_t Size) { + // If the use has a zero size or extends outside of the allocation, record + // it as a dead use for elimination later. + if (Size == 0 || (uint64_t)Offset >= AllocSize || + (Offset < 0 && (uint64_t)-Offset >= Size)) + return markAsDead(User); + + // Clamp the start to the beginning of the allocation. + if (Offset < 0) { + Size -= (uint64_t)-Offset; + Offset = 0; + } + + uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size; + + // Clamp the end offset to the end of the allocation. Note that this is + // formulated to handle even the case where "BeginOffset + Size" overflows. + assert(AllocSize >= BeginOffset); // Established above. + if (Size > AllocSize - BeginOffset) + EndOffset = AllocSize; + + // NB: This only works if we have zero overlapping partitions. + iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset); + if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset) + B = llvm::prior(B); + for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset; + ++I) { + PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset), + std::min(I->EndOffset, EndOffset), U); + P.use_push_back(I, NewPU); + if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser())) + P.PHIOrSelectOpMap[U] + = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1); + } + } + + void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) { + uint64_t Size = TD.getTypeStoreSize(Ty); + + // If this memory access can be shown to *statically* extend outside the + // bounds of of the allocation, it's behavior is undefined, so simply + // ignore it. Note that this is more strict than the generic clamping + // behavior of insertUse. + if (Offset < 0 || (uint64_t)Offset >= AllocSize || + Size > (AllocSize - (uint64_t)Offset)) + return markAsDead(I); + + insertUse(I, Offset, Size); + } + + void visitBitCastInst(BitCastInst &BC) { + if (BC.use_empty()) + return markAsDead(BC); + + enqueueUsers(BC, Offset); + } + + void visitGetElementPtrInst(GetElementPtrInst &GEPI) { + if (GEPI.use_empty()) + return markAsDead(GEPI); + + int64_t GEPOffset; + if (!computeConstantGEPOffset(GEPI, GEPOffset)) + llvm_unreachable("Unable to compute constant offset for use"); + + enqueueUsers(GEPI, GEPOffset); + } + + void visitLoadInst(LoadInst &LI) { + handleLoadOrStore(LI.getType(), LI, Offset); + } + + void visitStoreInst(StoreInst &SI) { + handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset); + } + + void visitMemSetInst(MemSetInst &II) { + ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); + uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset; + insertUse(II, Offset, Size); + } + + void visitMemTransferInst(MemTransferInst &II) { + ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); + uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset; + if (!Size) + return markAsDead(II); + + MemTransferOffsets &Offsets = P.MemTransferInstData[&II]; + if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd && + Offsets.DestBegin == Offsets.SourceBegin) + return markAsDead(II); // Skip identity transfers without side-effects. + + insertUse(II, Offset, Size); + } + + void visitIntrinsicInst(IntrinsicInst &II) { + assert(II.getIntrinsicID() == Intrinsic::lifetime_start || + II.getIntrinsicID() == Intrinsic::lifetime_end); + + ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0)); + insertUse(II, Offset, + std::min(AllocSize - Offset, Length->getLimitedValue())); + } + + void insertPHIOrSelect(Instruction &User, uint64_t Offset) { + uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first; + + // For PHI and select operands outside the alloca, we can't nuke the entire + // phi or select -- the other side might still be relevant, so we special + // case them here and use a separate structure to track the operands + // themselves which should be replaced with undef. + if (Offset >= AllocSize) { + P.DeadOperands.push_back(U); + return; + } + + insertUse(User, Offset, Size); + } + void visitPHINode(PHINode &PN) { + if (PN.use_empty()) + return markAsDead(PN); + + insertPHIOrSelect(PN, Offset); + } + void visitSelectInst(SelectInst &SI) { + if (SI.use_empty()) + return markAsDead(SI); + + if (Value *Result = foldSelectInst(SI)) { + if (Result == *U) + // If the result of the constant fold will be the pointer, recurse + // through the select as if we had RAUW'ed it. + enqueueUsers(SI, Offset); + else + // Otherwise the operand to the select is dead, and we can replace it + // with undef. + P.DeadOperands.push_back(U); + + return; + } + + insertPHIOrSelect(SI, Offset); + } + + /// \brief Unreachable, we've already visited the alloca once. + void visitInstruction(Instruction &I) { + llvm_unreachable("Unhandled instruction in use builder."); + } +}; + +void AllocaPartitioning::splitAndMergePartitions() { + size_t NumDeadPartitions = 0; + + // Track the range of splittable partitions that we pass when accumulating + // overlapping unsplittable partitions. + uint64_t SplitEndOffset = 0ull; + + Partition New(0ull, 0ull, false); + + for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) { + ++j; + + if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) { + assert(New.BeginOffset == New.EndOffset); + New = Partitions[i]; + } else { + assert(New.IsSplittable); + New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset); + } + assert(New.BeginOffset != New.EndOffset); + + // Scan the overlapping partitions. + while (j != e && New.EndOffset > Partitions[j].BeginOffset) { + // If the new partition we are forming is splittable, stop at the first + // unsplittable partition. + if (New.IsSplittable && !Partitions[j].IsSplittable) + break; + + // Grow the new partition to include any equally splittable range. 'j' is + // always equally splittable when New is splittable, but when New is not + // splittable, we may subsume some (or part of some) splitable partition + // without growing the new one. + if (New.IsSplittable == Partitions[j].IsSplittable) { + New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset); + } else { + assert(!New.IsSplittable); + assert(Partitions[j].IsSplittable); + SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset); + } + + Partitions[j].kill(); + ++NumDeadPartitions; + ++j; + } + + // If the new partition is splittable, chop off the end as soon as the + // unsplittable subsequent partition starts and ensure we eventually cover + // the splittable area. + if (j != e && New.IsSplittable) { + SplitEndOffset = std::max(SplitEndOffset, New.EndOffset); + New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset); + } + + // Add the new partition if it differs from the original one and is + // non-empty. We can end up with an empty partition here if it was + // splittable but there is an unsplittable one that starts at the same + // offset. + if (New != Partitions[i]) { + if (New.BeginOffset != New.EndOffset) + Partitions.push_back(New); + // Mark the old one for removal. + Partitions[i].kill(); + ++NumDeadPartitions; + } + + New.BeginOffset = New.EndOffset; + if (!New.IsSplittable) { + New.EndOffset = std::max(New.EndOffset, SplitEndOffset); + if (j != e && !Partitions[j].IsSplittable) + New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset); + New.IsSplittable = true; + // If there is a trailing splittable partition which won't be fused into + // the next splittable partition go ahead and add it onto the partitions + // list. + if (New.BeginOffset < New.EndOffset && + (j == e || !Partitions[j].IsSplittable || + New.EndOffset < Partitions[j].BeginOffset)) { + Partitions.push_back(New); + New.BeginOffset = New.EndOffset = 0ull; + } + } + } + + // Re-sort the partitions now that they have been split and merged into + // disjoint set of partitions. Also remove any of the dead partitions we've + // replaced in the process. + std::sort(Partitions.begin(), Partitions.end()); + if (NumDeadPartitions) { + assert(Partitions.back().isDead()); + assert((ptrdiff_t)NumDeadPartitions == + std::count(Partitions.begin(), Partitions.end(), Partitions.back())); + } + Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end()); +} + +AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI) + : +#ifndef NDEBUG + AI(AI), +#endif + PointerEscapingInstr(0) { + PartitionBuilder PB(TD, AI, *this); + if (!PB()) + return; + + // Sort the uses. This arranges for the offsets to be in ascending order, + // and the sizes to be in descending order. + std::sort(Partitions.begin(), Partitions.end()); + + // Remove any partitions from the back which are marked as dead. + while (!Partitions.empty() && Partitions.back().isDead()) + Partitions.pop_back(); + + if (Partitions.size() > 1) { + // Intersect splittability for all partitions with equal offsets and sizes. + // Then remove all but the first so that we have a sequence of non-equal but + // potentially overlapping partitions. + for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E; + I = J) { + ++J; + while (J != E && *I == *J) { + I->IsSplittable &= J->IsSplittable; + ++J; + } + } + Partitions.erase(std::unique(Partitions.begin(), Partitions.end()), + Partitions.end()); + + // Split splittable and merge unsplittable partitions into a disjoint set + // of partitions over the used space of the allocation. + splitAndMergePartitions(); + } + + // Now build up the user lists for each of these disjoint partitions by + // re-walking the recursive users of the alloca. + Uses.resize(Partitions.size()); + UseBuilder UB(TD, AI, *this); + UB(); +} + +Type *AllocaPartitioning::getCommonType(iterator I) const { + Type *Ty = 0; + for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) { + if (!UI->U) + continue; // Skip dead uses. + if (isa<IntrinsicInst>(*UI->U->getUser())) + continue; + if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset) + continue; + + Type *UserTy = 0; + if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) { + UserTy = LI->getType(); + } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) { + UserTy = SI->getValueOperand()->getType(); + } else { + return 0; // Bail if we have weird uses. + } + + if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) { + // If the type is larger than the partition, skip it. We only encounter + // this for split integer operations where we want to use the type of the + // entity causing the split. + if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8) + continue; + + // If we have found an integer type use covering the alloca, use that + // regardless of the other types, as integers are often used for a "bucket + // of bits" type. + return ITy; + } + + if (Ty && Ty != UserTy) + return 0; + + Ty = UserTy; + } + return Ty; +} + +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) + +void AllocaPartitioning::print(raw_ostream &OS, const_iterator I, + StringRef Indent) const { + OS << Indent << "partition #" << (I - begin()) + << " [" << I->BeginOffset << "," << I->EndOffset << ")" + << (I->IsSplittable ? " (splittable)" : "") + << (Uses[I - begin()].empty() ? " (zero uses)" : "") + << "\n"; +} + +void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I, + StringRef Indent) const { + for (const_use_iterator UI = use_begin(I), UE = use_end(I); + UI != UE; ++UI) { + if (!UI->U) + continue; // Skip dead uses. + OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") " + << "used by: " << *UI->U->getUser() << "\n"; + if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) { + const MemTransferOffsets &MTO = MemTransferInstData.lookup(II); + bool IsDest; + if (!MTO.IsSplittable) + IsDest = UI->BeginOffset == MTO.DestBegin; + else + IsDest = MTO.DestBegin != 0u; + OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": " + << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin) + << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n"; + } + } +} + +void AllocaPartitioning::print(raw_ostream &OS) const { + if (PointerEscapingInstr) { + OS << "No partitioning for alloca: " << AI << "\n" + << " A pointer to this alloca escaped by:\n" + << " " << *PointerEscapingInstr << "\n"; + return; + } + + OS << "Partitioning of alloca: " << AI << "\n"; + unsigned Num = 0; + for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) { + print(OS, I); + printUsers(OS, I); + } +} + +void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); } +void AllocaPartitioning::dump() const { print(dbgs()); } + +#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) + + +namespace { +/// \brief Implementation of LoadAndStorePromoter for promoting allocas. +/// +/// This subclass of LoadAndStorePromoter adds overrides to handle promoting +/// the loads and stores of an alloca instruction, as well as updating its +/// debug information. This is used when a domtree is unavailable and thus +/// mem2reg in its full form can't be used to handle promotion of allocas to +/// scalar values. +class AllocaPromoter : public LoadAndStorePromoter { + AllocaInst &AI; + DIBuilder &DIB; + + SmallVector<DbgDeclareInst *, 4> DDIs; + SmallVector<DbgValueInst *, 4> DVIs; + +public: + AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, + AllocaInst &AI, DIBuilder &DIB) + : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {} + + void run(const SmallVectorImpl<Instruction*> &Insts) { + // Remember which alloca we're promoting (for isInstInList). + if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) { + for (Value::use_iterator UI = DebugNode->use_begin(), + UE = DebugNode->use_end(); + UI != UE; ++UI) + if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) + DDIs.push_back(DDI); + else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) + DVIs.push_back(DVI); + } + + LoadAndStorePromoter::run(Insts); + AI.eraseFromParent(); + while (!DDIs.empty()) + DDIs.pop_back_val()->eraseFromParent(); + while (!DVIs.empty()) + DVIs.pop_back_val()->eraseFromParent(); + } + + virtual bool isInstInList(Instruction *I, + const SmallVectorImpl<Instruction*> &Insts) const { + if (LoadInst *LI = dyn_cast<LoadInst>(I)) + return LI->getOperand(0) == &AI; + return cast<StoreInst>(I)->getPointerOperand() == &AI; + } + + virtual void updateDebugInfo(Instruction *Inst) const { + for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), + E = DDIs.end(); I != E; ++I) { + DbgDeclareInst *DDI = *I; + if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) + ConvertDebugDeclareToDebugValue(DDI, SI, DIB); + else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) + ConvertDebugDeclareToDebugValue(DDI, LI, DIB); + } + for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), + E = DVIs.end(); I != E; ++I) { + DbgValueInst *DVI = *I; + Value *Arg = NULL; + if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { + // If an argument is zero extended then use argument directly. The ZExt + // may be zapped by an optimization pass in future. + if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) + Arg = dyn_cast<Argument>(ZExt->getOperand(0)); + if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) + Arg = dyn_cast<Argument>(SExt->getOperand(0)); + if (!Arg) + Arg = SI->getOperand(0); + } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { + Arg = LI->getOperand(0); + } else { + continue; + } + Instruction *DbgVal = + DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), + Inst); + DbgVal->setDebugLoc(DVI->getDebugLoc()); + } + } +}; +} // end anon namespace + + +namespace { +/// \brief An optimization pass providing Scalar Replacement of Aggregates. +/// +/// This pass takes allocations which can be completely analyzed (that is, they +/// don't escape) and tries to turn them into scalar SSA values. There are +/// a few steps to this process. +/// +/// 1) It takes allocations of aggregates and analyzes the ways in which they +/// are used to try to split them into smaller allocations, ideally of +/// a single scalar data type. It will split up memcpy and memset accesses +/// as necessary and try to isolate invidual scalar accesses. +/// 2) It will transform accesses into forms which are suitable for SSA value +/// promotion. This can be replacing a memset with a scalar store of an +/// integer value, or it can involve speculating operations on a PHI or +/// select to be a PHI or select of the results. +/// 3) Finally, this will try to detect a pattern of accesses which map cleanly +/// onto insert and extract operations on a vector value, and convert them to +/// this form. By doing so, it will enable promotion of vector aggregates to +/// SSA vector values. +class SROA : public FunctionPass { + const bool RequiresDomTree; + + LLVMContext *C; + const DataLayout *TD; + DominatorTree *DT; + + /// \brief Worklist of alloca instructions to simplify. + /// + /// Each alloca in the function is added to this. Each new alloca formed gets + /// added to it as well to recursively simplify unless that alloca can be + /// directly promoted. Finally, each time we rewrite a use of an alloca other + /// the one being actively rewritten, we add it back onto the list if not + /// already present to ensure it is re-visited. + SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist; + + /// \brief A collection of instructions to delete. + /// We try to batch deletions to simplify code and make things a bit more + /// efficient. + SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts; + + /// \brief Post-promotion worklist. + /// + /// Sometimes we discover an alloca which has a high probability of becoming + /// viable for SROA after a round of promotion takes place. In those cases, + /// the alloca is enqueued here for re-processing. + /// + /// Note that we have to be very careful to clear allocas out of this list in + /// the event they are deleted. + SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist; + + /// \brief A collection of alloca instructions we can directly promote. + std::vector<AllocaInst *> PromotableAllocas; + +public: + SROA(bool RequiresDomTree = true) + : FunctionPass(ID), RequiresDomTree(RequiresDomTree), + C(0), TD(0), DT(0) { + initializeSROAPass(*PassRegistry::getPassRegistry()); + } + bool runOnFunction(Function &F); + void getAnalysisUsage(AnalysisUsage &AU) const; + + const char *getPassName() const { return "SROA"; } + static char ID; + +private: + friend class PHIOrSelectSpeculator; + friend class AllocaPartitionRewriter; + friend class AllocaPartitionVectorRewriter; + + bool rewriteAllocaPartition(AllocaInst &AI, + AllocaPartitioning &P, + AllocaPartitioning::iterator PI); + bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P); + bool runOnAlloca(AllocaInst &AI); + void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas); + bool promoteAllocas(Function &F); +}; +} + +char SROA::ID = 0; + +FunctionPass *llvm::createSROAPass(bool RequiresDomTree) { + return new SROA(RequiresDomTree); +} + +INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", + false, false) +INITIALIZE_PASS_DEPENDENCY(DominatorTree) +INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", + false, false) + +namespace { +/// \brief Visitor to speculate PHIs and Selects where possible. +class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> { + // Befriend the base class so it can delegate to private visit methods. + friend class llvm::InstVisitor<PHIOrSelectSpeculator>; + + const DataLayout &TD; + AllocaPartitioning &P; + SROA &Pass; + +public: + PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass) + : TD(TD), P(P), Pass(Pass) {} + + /// \brief Visit the users of an alloca partition and rewrite them. + void visitUsers(AllocaPartitioning::const_iterator PI) { + // Note that we need to use an index here as the underlying vector of uses + // may be grown during speculation. However, we never need to re-visit the + // new uses, and so we can use the initial size bound. + for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) { + const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx); + if (!PU.U) + continue; // Skip dead use. + + visit(cast<Instruction>(PU.U->getUser())); + } + } + +private: + // By default, skip this instruction. + void visitInstruction(Instruction &I) {} + + /// PHI instructions that use an alloca and are subsequently loaded can be + /// rewritten to load both input pointers in the pred blocks and then PHI the + /// results, allowing the load of the alloca to be promoted. + /// From this: + /// %P2 = phi [i32* %Alloca, i32* %Other] + /// %V = load i32* %P2 + /// to: + /// %V1 = load i32* %Alloca -> will be mem2reg'd + /// ... + /// %V2 = load i32* %Other + /// ... + /// %V = phi [i32 %V1, i32 %V2] + /// + /// We can do this to a select if its only uses are loads and if the operands + /// to the select can be loaded unconditionally. + /// + /// FIXME: This should be hoisted into a generic utility, likely in + /// Transforms/Util/Local.h + bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) { + // For now, we can only do this promotion if the load is in the same block + // as the PHI, and if there are no stores between the phi and load. + // TODO: Allow recursive phi users. + // TODO: Allow stores. + BasicBlock *BB = PN.getParent(); + unsigned MaxAlign = 0; + for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); + UI != UE; ++UI) { + LoadInst *LI = dyn_cast<LoadInst>(*UI); + if (LI == 0 || !LI->isSimple()) return false; + + // For now we only allow loads in the same block as the PHI. This is + // a common case that happens when instcombine merges two loads through + // a PHI. + if (LI->getParent() != BB) return false; + + // Ensure that there are no instructions between the PHI and the load that + // could store. + for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI) + if (BBI->mayWriteToMemory()) + return false; + + MaxAlign = std::max(MaxAlign, LI->getAlignment()); + Loads.push_back(LI); + } + + // We can only transform this if it is safe to push the loads into the + // predecessor blocks. The only thing to watch out for is that we can't put + // a possibly trapping load in the predecessor if it is a critical edge. + for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; + ++Idx) { + TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator(); + Value *InVal = PN.getIncomingValue(Idx); + + // If the value is produced by the terminator of the predecessor (an + // invoke) or it has side-effects, there is no valid place to put a load + // in the predecessor. + if (TI == InVal || TI->mayHaveSideEffects()) + return false; + + // If the predecessor has a single successor, then the edge isn't + // critical. + if (TI->getNumSuccessors() == 1) + continue; + + // If this pointer is always safe to load, or if we can prove that there + // is already a load in the block, then we can move the load to the pred + // block. + if (InVal->isDereferenceablePointer() || + isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD)) + continue; + + return false; + } + + return true; + } + + void visitPHINode(PHINode &PN) { + DEBUG(dbgs() << " original: " << PN << "\n"); + + SmallVector<LoadInst *, 4> Loads; + if (!isSafePHIToSpeculate(PN, Loads)) + return; + + assert(!Loads.empty()); + + Type *LoadTy = cast<PointerType>(PN.getType())->getElementType(); + IRBuilder<> PHIBuilder(&PN); + PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(), + PN.getName() + ".sroa.speculated"); + + // Get the TBAA tag and alignment to use from one of the loads. It doesn't + // matter which one we get and if any differ, it doesn't matter. + LoadInst *SomeLoad = cast<LoadInst>(Loads.back()); + MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); + unsigned Align = SomeLoad->getAlignment(); + + // Rewrite all loads of the PN to use the new PHI. + do { + LoadInst *LI = Loads.pop_back_val(); + LI->replaceAllUsesWith(NewPN); + Pass.DeadInsts.insert(LI); + } while (!Loads.empty()); + + // Inject loads into all of the pred blocks. + for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) { + BasicBlock *Pred = PN.getIncomingBlock(Idx); + TerminatorInst *TI = Pred->getTerminator(); + Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx)); + Value *InVal = PN.getIncomingValue(Idx); + IRBuilder<> PredBuilder(TI); + + LoadInst *Load + = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." + + Pred->getName())); + ++NumLoadsSpeculated; + Load->setAlignment(Align); + if (TBAATag) + Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); + NewPN->addIncoming(Load, Pred); + + Instruction *Ptr = dyn_cast<Instruction>(InVal); + if (!Ptr) + // No uses to rewrite. + continue; + + // Try to lookup and rewrite any partition uses corresponding to this phi + // input. + AllocaPartitioning::iterator PI + = P.findPartitionForPHIOrSelectOperand(InUse); + if (PI == P.end()) + continue; + + // Replace the Use in the PartitionUse for this operand with the Use + // inside the load. + AllocaPartitioning::use_iterator UI + = P.findPartitionUseForPHIOrSelectOperand(InUse); + assert(isa<PHINode>(*UI->U->getUser())); + UI->U = &Load->getOperandUse(Load->getPointerOperandIndex()); + } + DEBUG(dbgs() << " speculated to: " << *NewPN << "\n"); + } + + /// Select instructions that use an alloca and are subsequently loaded can be + /// rewritten to load both input pointers and then select between the result, + /// allowing the load of the alloca to be promoted. + /// From this: + /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other + /// %V = load i32* %P2 + /// to: + /// %V1 = load i32* %Alloca -> will be mem2reg'd + /// %V2 = load i32* %Other + /// %V = select i1 %cond, i32 %V1, i32 %V2 + /// + /// We can do this to a select if its only uses are loads and if the operand + /// to the select can be loaded unconditionally. + bool isSafeSelectToSpeculate(SelectInst &SI, + SmallVectorImpl<LoadInst *> &Loads) { + Value *TValue = SI.getTrueValue(); + Value *FValue = SI.getFalseValue(); + bool TDerefable = TValue->isDereferenceablePointer(); + bool FDerefable = FValue->isDereferenceablePointer(); + + for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); + UI != UE; ++UI) { + LoadInst *LI = dyn_cast<LoadInst>(*UI); + if (LI == 0 || !LI->isSimple()) return false; + + // Both operands to the select need to be dereferencable, either + // absolutely (e.g. allocas) or at this point because we can see other + // accesses to it. + if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI, + LI->getAlignment(), &TD)) + return false; + if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI, + LI->getAlignment(), &TD)) + return false; + Loads.push_back(LI); + } + + return true; + } + + void visitSelectInst(SelectInst &SI) { + DEBUG(dbgs() << " original: " << SI << "\n"); + IRBuilder<> IRB(&SI); + + // If the select isn't safe to speculate, just use simple logic to emit it. + SmallVector<LoadInst *, 4> Loads; + if (!isSafeSelectToSpeculate(SI, Loads)) + return; + + Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) }; + AllocaPartitioning::iterator PIs[2]; + AllocaPartitioning::PartitionUse PUs[2]; + for (unsigned i = 0, e = 2; i != e; ++i) { + PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]); + if (PIs[i] != P.end()) { + // If the pointer is within the partitioning, remove the select from + // its uses. We'll add in the new loads below. + AllocaPartitioning::use_iterator UI + = P.findPartitionUseForPHIOrSelectOperand(Ops[i]); + PUs[i] = *UI; + // Clear out the use here so that the offsets into the use list remain + // stable but this use is ignored when rewriting. + UI->U = 0; + } + } + + Value *TV = SI.getTrueValue(); + Value *FV = SI.getFalseValue(); + // Replace the loads of the select with a select of two loads. + while (!Loads.empty()) { + LoadInst *LI = Loads.pop_back_val(); + + IRB.SetInsertPoint(LI); + LoadInst *TL = + IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true"); + LoadInst *FL = + IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false"); + NumLoadsSpeculated += 2; + + // Transfer alignment and TBAA info if present. + TL->setAlignment(LI->getAlignment()); + FL->setAlignment(LI->getAlignment()); + if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { + TL->setMetadata(LLVMContext::MD_tbaa, Tag); + FL->setMetadata(LLVMContext::MD_tbaa, Tag); + } + + Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL, + LI->getName() + ".sroa.speculated"); + + LoadInst *Loads[2] = { TL, FL }; + for (unsigned i = 0, e = 2; i != e; ++i) { + if (PIs[i] != P.end()) { + Use *LoadUse = &Loads[i]->getOperandUse(0); + assert(PUs[i].U->get() == LoadUse->get()); + PUs[i].U = LoadUse; + P.use_push_back(PIs[i], PUs[i]); + } + } + + DEBUG(dbgs() << " speculated to: " << *V << "\n"); + LI->replaceAllUsesWith(V); + Pass.DeadInsts.insert(LI); + } + } +}; +} + +/// \brief Accumulate the constant offsets in a GEP into a single APInt offset. +/// +/// If the provided GEP is all-constant, the total byte offset formed by the +/// GEP is computed and Offset is set to it. If the GEP has any non-constant +/// operands, the function returns false and the value of Offset is unmodified. +static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP, + APInt &Offset) { + APInt GEPOffset(Offset.getBitWidth(), 0); + for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); + GTI != GTE; ++GTI) { + ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand()); + if (!OpC) + return false; + if (OpC->isZero()) continue; + + // Handle a struct index, which adds its field offset to the pointer. + if (StructType *STy = dyn_cast<StructType>(*GTI)) { + unsigned ElementIdx = OpC->getZExtValue(); + const StructLayout *SL = TD.getStructLayout(STy); + GEPOffset += APInt(Offset.getBitWidth(), + SL->getElementOffset(ElementIdx)); + continue; + } + + APInt TypeSize(Offset.getBitWidth(), + TD.getTypeAllocSize(GTI.getIndexedType())); + if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) { + assert((VTy->getScalarSizeInBits() % 8) == 0 && + "vector element size is not a multiple of 8, cannot GEP over it"); + TypeSize = VTy->getScalarSizeInBits() / 8; + } + + GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize; + } + Offset = GEPOffset; + return true; +} + +/// \brief Build a GEP out of a base pointer and indices. +/// +/// This will return the BasePtr if that is valid, or build a new GEP +/// instruction using the IRBuilder if GEP-ing is needed. +static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr, + SmallVectorImpl<Value *> &Indices, + const Twine &Prefix) { + if (Indices.empty()) + return BasePtr; + + // A single zero index is a no-op, so check for this and avoid building a GEP + // in that case. + if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero()) + return BasePtr; + + return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx"); +} + +/// \brief Get a natural GEP off of the BasePtr walking through Ty toward +/// TargetTy without changing the offset of the pointer. +/// +/// This routine assumes we've already established a properly offset GEP with +/// Indices, and arrived at the Ty type. The goal is to continue to GEP with +/// zero-indices down through type layers until we find one the same as +/// TargetTy. If we can't find one with the same type, we at least try to use +/// one with the same size. If none of that works, we just produce the GEP as +/// indicated by Indices to have the correct offset. +static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD, + Value *BasePtr, Type *Ty, Type *TargetTy, + SmallVectorImpl<Value *> &Indices, + const Twine &Prefix) { + if (Ty == TargetTy) + return buildGEP(IRB, BasePtr, Indices, Prefix); + + // See if we can descend into a struct and locate a field with the correct + // type. + unsigned NumLayers = 0; + Type *ElementTy = Ty; + do { + if (ElementTy->isPointerTy()) + break; + if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) { + ElementTy = SeqTy->getElementType(); + // Note that we use the default address space as this index is over an + // array or a vector, not a pointer. + Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0))); + } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) { + if (STy->element_begin() == STy->element_end()) + break; // Nothing left to descend into. + ElementTy = *STy->element_begin(); + Indices.push_back(IRB.getInt32(0)); + } else { + break; + } + ++NumLayers; + } while (ElementTy != TargetTy); + if (ElementTy != TargetTy) + Indices.erase(Indices.end() - NumLayers, Indices.end()); + + return buildGEP(IRB, BasePtr, Indices, Prefix); +} + +/// \brief Recursively compute indices for a natural GEP. +/// +/// This is the recursive step for getNaturalGEPWithOffset that walks down the +/// element types adding appropriate indices for the GEP. +static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD, + Value *Ptr, Type *Ty, APInt &Offset, + Type *TargetTy, + SmallVectorImpl<Value *> &Indices, + const Twine &Prefix) { + if (Offset == 0) + return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix); + + // We can't recurse through pointer types. + if (Ty->isPointerTy()) + return 0; + + // We try to analyze GEPs over vectors here, but note that these GEPs are + // extremely poorly defined currently. The long-term goal is to remove GEPing + // over a vector from the IR completely. + if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) { + unsigned ElementSizeInBits = VecTy->getScalarSizeInBits(); + if (ElementSizeInBits % 8) + return 0; // GEPs over non-multiple of 8 size vector elements are invalid. + APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8); + APInt NumSkippedElements = Offset.sdiv(ElementSize); + if (NumSkippedElements.ugt(VecTy->getNumElements())) + return 0; + Offset -= NumSkippedElements * ElementSize; + Indices.push_back(IRB.getInt(NumSkippedElements)); + return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(), + Offset, TargetTy, Indices, Prefix); + } + + if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) { + Type *ElementTy = ArrTy->getElementType(); + APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy)); + APInt NumSkippedElements = Offset.sdiv(ElementSize); + if (NumSkippedElements.ugt(ArrTy->getNumElements())) + return 0; + + Offset -= NumSkippedElements * ElementSize; + Indices.push_back(IRB.getInt(NumSkippedElements)); + return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy, + Indices, Prefix); + } + + StructType *STy = dyn_cast<StructType>(Ty); + if (!STy) + return 0; + + const StructLayout *SL = TD.getStructLayout(STy); + uint64_t StructOffset = Offset.getZExtValue(); + if (StructOffset >= SL->getSizeInBytes()) + return 0; + unsigned Index = SL->getElementContainingOffset(StructOffset); + Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index)); + Type *ElementTy = STy->getElementType(Index); + if (Offset.uge(TD.getTypeAllocSize(ElementTy))) + return 0; // The offset points into alignment padding. + + Indices.push_back(IRB.getInt32(Index)); + return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy, + Indices, Prefix); +} + +/// \brief Get a natural GEP from a base pointer to a particular offset and +/// resulting in a particular type. +/// +/// The goal is to produce a "natural" looking GEP that works with the existing +/// composite types to arrive at the appropriate offset and element type for +/// a pointer. TargetTy is the element type the returned GEP should point-to if +/// possible. We recurse by decreasing Offset, adding the appropriate index to +/// Indices, and setting Ty to the result subtype. +/// +/// If no natural GEP can be constructed, this function returns null. +static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD, + Value *Ptr, APInt Offset, Type *TargetTy, + SmallVectorImpl<Value *> &Indices, + const Twine &Prefix) { + PointerType *Ty = cast<PointerType>(Ptr->getType()); + + // Don't consider any GEPs through an i8* as natural unless the TargetTy is + // an i8. + if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8)) + return 0; + + Type *ElementTy = Ty->getElementType(); + if (!ElementTy->isSized()) + return 0; // We can't GEP through an unsized element. + APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy)); + if (ElementSize == 0) + return 0; // Zero-length arrays can't help us build a natural GEP. + APInt NumSkippedElements = Offset.sdiv(ElementSize); + + Offset -= NumSkippedElements * ElementSize; + Indices.push_back(IRB.getInt(NumSkippedElements)); + return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy, + Indices, Prefix); +} + +/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the +/// resulting pointer has PointerTy. +/// +/// This tries very hard to compute a "natural" GEP which arrives at the offset +/// and produces the pointer type desired. Where it cannot, it will try to use +/// the natural GEP to arrive at the offset and bitcast to the type. Where that +/// fails, it will try to use an existing i8* and GEP to the byte offset and +/// bitcast to the type. +/// +/// The strategy for finding the more natural GEPs is to peel off layers of the +/// pointer, walking back through bit casts and GEPs, searching for a base +/// pointer from which we can compute a natural GEP with the desired +/// properities. The algorithm tries to fold as many constant indices into +/// a single GEP as possible, thus making each GEP more independent of the +/// surrounding code. +static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD, + Value *Ptr, APInt Offset, Type *PointerTy, + const Twine &Prefix) { + // Even though we don't look through PHI nodes, we could be called on an + // instruction in an unreachable block, which may be on a cycle. + SmallPtrSet<Value *, 4> Visited; + Visited.insert(Ptr); + SmallVector<Value *, 4> Indices; + + // We may end up computing an offset pointer that has the wrong type. If we + // never are able to compute one directly that has the correct type, we'll + // fall back to it, so keep it around here. + Value *OffsetPtr = 0; + + // Remember any i8 pointer we come across to re-use if we need to do a raw + // byte offset. + Value *Int8Ptr = 0; + APInt Int8PtrOffset(Offset.getBitWidth(), 0); + + Type *TargetTy = PointerTy->getPointerElementType(); + + do { + // First fold any existing GEPs into the offset. + while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { + APInt GEPOffset(Offset.getBitWidth(), 0); + if (!accumulateGEPOffsets(TD, *GEP, GEPOffset)) + break; + Offset += GEPOffset; + Ptr = GEP->getPointerOperand(); + if (!Visited.insert(Ptr)) + break; + } + + // See if we can perform a natural GEP here. + Indices.clear(); + if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy, + Indices, Prefix)) { + if (P->getType() == PointerTy) { + // Zap any offset pointer that we ended up computing in previous rounds. + if (OffsetPtr && OffsetPtr->use_empty()) + if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) + I->eraseFromParent(); + return P; + } + if (!OffsetPtr) { + OffsetPtr = P; + } + } + + // Stash this pointer if we've found an i8*. + if (Ptr->getType()->isIntegerTy(8)) { + Int8Ptr = Ptr; + Int8PtrOffset = Offset; + } + + // Peel off a layer of the pointer and update the offset appropriately. + if (Operator::getOpcode(Ptr) == Instruction::BitCast) { + Ptr = cast<Operator>(Ptr)->getOperand(0); + } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) { + if (GA->mayBeOverridden()) + break; + Ptr = GA->getAliasee(); + } else { + break; + } + assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!"); + } while (Visited.insert(Ptr)); + + if (!OffsetPtr) { + if (!Int8Ptr) { + Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(), + Prefix + ".raw_cast"); + Int8PtrOffset = Offset; + } + + OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr : + IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset), + Prefix + ".raw_idx"); + } + Ptr = OffsetPtr; + + // On the off chance we were targeting i8*, guard the bitcast here. + if (Ptr->getType() != PointerTy) + Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast"); + + return Ptr; +} + +/// \brief Test whether we can convert a value from the old to the new type. +/// +/// This predicate should be used to guard calls to convertValue in order to +/// ensure that we only try to convert viable values. The strategy is that we +/// will peel off single element struct and array wrappings to get to an +/// underlying value, and convert that value. +static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) { + if (OldTy == NewTy) + return true; + if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy)) + return false; + if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType()) + return false; + + if (NewTy->isPointerTy() || OldTy->isPointerTy()) { + if (NewTy->isPointerTy() && OldTy->isPointerTy()) + return true; + if (NewTy->isIntegerTy() || OldTy->isIntegerTy()) + return true; + return false; + } + + return true; +} + +/// \brief Generic routine to convert an SSA value to a value of a different +/// type. +/// +/// This will try various different casting techniques, such as bitcasts, +/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test +/// two types for viability with this routine. +static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V, + Type *Ty) { + assert(canConvertValue(DL, V->getType(), Ty) && + "Value not convertable to type"); + if (V->getType() == Ty) + return V; + if (V->getType()->isIntegerTy() && Ty->isPointerTy()) + return IRB.CreateIntToPtr(V, Ty); + if (V->getType()->isPointerTy() && Ty->isIntegerTy()) + return IRB.CreatePtrToInt(V, Ty); + + return IRB.CreateBitCast(V, Ty); +} + +/// \brief Test whether the given alloca partition can be promoted to a vector. +/// +/// This is a quick test to check whether we can rewrite a particular alloca +/// partition (and its newly formed alloca) into a vector alloca with only +/// whole-vector loads and stores such that it could be promoted to a vector +/// SSA value. We only can ensure this for a limited set of operations, and we +/// don't want to do the rewrites unless we are confident that the result will +/// be promotable, so we have an early test here. +static bool isVectorPromotionViable(const DataLayout &TD, + Type *AllocaTy, + AllocaPartitioning &P, + uint64_t PartitionBeginOffset, + uint64_t PartitionEndOffset, + AllocaPartitioning::const_use_iterator I, + AllocaPartitioning::const_use_iterator E) { + VectorType *Ty = dyn_cast<VectorType>(AllocaTy); + if (!Ty) + return false; + + uint64_t VecSize = TD.getTypeSizeInBits(Ty); + uint64_t ElementSize = Ty->getScalarSizeInBits(); + + // While the definition of LLVM vectors is bitpacked, we don't support sizes + // that aren't byte sized. + if (ElementSize % 8) + return false; + assert((VecSize % 8) == 0 && "vector size not a multiple of element size?"); + VecSize /= 8; + ElementSize /= 8; + + for (; I != E; ++I) { + if (!I->U) + continue; // Skip dead use. + + uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset; + uint64_t BeginIndex = BeginOffset / ElementSize; + if (BeginIndex * ElementSize != BeginOffset || + BeginIndex >= Ty->getNumElements()) + return false; + uint64_t EndOffset = I->EndOffset - PartitionBeginOffset; + uint64_t EndIndex = EndOffset / ElementSize; + if (EndIndex * ElementSize != EndOffset || + EndIndex > Ty->getNumElements()) + return false; + + // FIXME: We should build shuffle vector instructions to handle + // non-element-sized accesses. + if ((EndOffset - BeginOffset) != ElementSize && + (EndOffset - BeginOffset) != VecSize) + return false; + + if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) { + if (MI->isVolatile()) + return false; + if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) { + const AllocaPartitioning::MemTransferOffsets &MTO + = P.getMemTransferOffsets(*MTI); + if (!MTO.IsSplittable) + return false; + } + } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) { + // Disable vector promotion when there are loads or stores of an FCA. + return false; + } else if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) { + if (LI->isVolatile()) + return false; + } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) { + if (SI->isVolatile()) + return false; + } else { + return false; + } + } + return true; +} + +/// \brief Test whether the given alloca partition's integer operations can be +/// widened to promotable ones. +/// +/// This is a quick test to check whether we can rewrite the integer loads and +/// stores to a particular alloca into wider loads and stores and be able to +/// promote the resulting alloca. +static bool isIntegerWideningViable(const DataLayout &TD, + Type *AllocaTy, + uint64_t AllocBeginOffset, + AllocaPartitioning &P, + AllocaPartitioning::const_use_iterator I, + AllocaPartitioning::const_use_iterator E) { + uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy); + + // Don't try to handle allocas with bit-padding. + if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy)) + return false; + + // We need to ensure that an integer type with the appropriate bitwidth can + // be converted to the alloca type, whatever that is. We don't want to force + // the alloca itself to have an integer type if there is a more suitable one. + Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits); + if (!canConvertValue(TD, AllocaTy, IntTy) || + !canConvertValue(TD, IntTy, AllocaTy)) + return false; + + uint64_t Size = TD.getTypeStoreSize(AllocaTy); + + // Check the uses to ensure the uses are (likely) promoteable integer uses. + // Also ensure that the alloca has a covering load or store. We don't want + // to widen the integer operotains only to fail to promote due to some other + // unsplittable entry (which we may make splittable later). + bool WholeAllocaOp = false; + for (; I != E; ++I) { + if (!I->U) + continue; // Skip dead use. + + uint64_t RelBegin = I->BeginOffset - AllocBeginOffset; + uint64_t RelEnd = I->EndOffset - AllocBeginOffset; + + // We can't reasonably handle cases where the load or store extends past + // the end of the aloca's type and into its padding. + if (RelEnd > Size) + return false; + + if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) { + if (LI->isVolatile()) + return false; + if (RelBegin == 0 && RelEnd == Size) + WholeAllocaOp = true; + if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) { + if (ITy->getBitWidth() < TD.getTypeStoreSize(ITy)) + return false; + continue; + } + // Non-integer loads need to be convertible from the alloca type so that + // they are promotable. + if (RelBegin != 0 || RelEnd != Size || + !canConvertValue(TD, AllocaTy, LI->getType())) + return false; + } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) { + Type *ValueTy = SI->getValueOperand()->getType(); + if (SI->isVolatile()) + return false; + if (RelBegin == 0 && RelEnd == Size) + WholeAllocaOp = true; + if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) { + if (ITy->getBitWidth() < TD.getTypeStoreSize(ITy)) + return false; + continue; + } + // Non-integer stores need to be convertible to the alloca type so that + // they are promotable. + if (RelBegin != 0 || RelEnd != Size || + !canConvertValue(TD, ValueTy, AllocaTy)) + return false; + } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) { + if (MI->isVolatile()) + return false; + if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) { + const AllocaPartitioning::MemTransferOffsets &MTO + = P.getMemTransferOffsets(*MTI); + if (!MTO.IsSplittable) + return false; + } + } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->U->getUser())) { + if (II->getIntrinsicID() != Intrinsic::lifetime_start && + II->getIntrinsicID() != Intrinsic::lifetime_end) + return false; + } else { + return false; + } + } + return WholeAllocaOp; +} + +static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V, + IntegerType *Ty, uint64_t Offset, + const Twine &Name) { + DEBUG(dbgs() << " start: " << *V << "\n"); + IntegerType *IntTy = cast<IntegerType>(V->getType()); + assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && + "Element extends past full value"); + uint64_t ShAmt = 8*Offset; + if (DL.isBigEndian()) + ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset); + if (ShAmt) { + V = IRB.CreateLShr(V, ShAmt, Name + ".shift"); + DEBUG(dbgs() << " shifted: " << *V << "\n"); + } + assert(Ty->getBitWidth() <= IntTy->getBitWidth() && + "Cannot extract to a larger integer!"); + if (Ty != IntTy) { + V = IRB.CreateTrunc(V, Ty, Name + ".trunc"); + DEBUG(dbgs() << " trunced: " << *V << "\n"); + } + return V; +} + +static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old, + Value *V, uint64_t Offset, const Twine &Name) { + IntegerType *IntTy = cast<IntegerType>(Old->getType()); + IntegerType *Ty = cast<IntegerType>(V->getType()); + assert(Ty->getBitWidth() <= IntTy->getBitWidth() && + "Cannot insert a larger integer!"); + DEBUG(dbgs() << " start: " << *V << "\n"); + if (Ty != IntTy) { + V = IRB.CreateZExt(V, IntTy, Name + ".ext"); + DEBUG(dbgs() << " extended: " << *V << "\n"); + } + assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && + "Element store outside of alloca store"); + uint64_t ShAmt = 8*Offset; + if (DL.isBigEndian()) + ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset); + if (ShAmt) { + V = IRB.CreateShl(V, ShAmt, Name + ".shift"); + DEBUG(dbgs() << " shifted: " << *V << "\n"); + } + + if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) { + APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt); + Old = IRB.CreateAnd(Old, Mask, Name + ".mask"); + DEBUG(dbgs() << " masked: " << *Old << "\n"); + V = IRB.CreateOr(Old, V, Name + ".insert"); + DEBUG(dbgs() << " inserted: " << *V << "\n"); + } + return V; +} + +namespace { +/// \brief Visitor to rewrite instructions using a partition of an alloca to +/// use a new alloca. +/// +/// Also implements the rewriting to vector-based accesses when the partition +/// passes the isVectorPromotionViable predicate. Most of the rewriting logic +/// lives here. +class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter, + bool> { + // Befriend the base class so it can delegate to private visit methods. + friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>; + + const DataLayout &TD; + AllocaPartitioning &P; + SROA &Pass; + AllocaInst &OldAI, &NewAI; + const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset; + Type *NewAllocaTy; + + // If we are rewriting an alloca partition which can be written as pure + // vector operations, we stash extra information here. When VecTy is + // non-null, we have some strict guarantees about the rewriten alloca: + // - The new alloca is exactly the size of the vector type here. + // - The accesses all either map to the entire vector or to a single + // element. + // - The set of accessing instructions is only one of those handled above + // in isVectorPromotionViable. Generally these are the same access kinds + // which are promotable via mem2reg. + VectorType *VecTy; + Type *ElementTy; + uint64_t ElementSize; + + // This is a convenience and flag variable that will be null unless the new + // alloca's integer operations should be widened to this integer type due to + // passing isIntegerWideningViable above. If it is non-null, the desired + // integer type will be stored here for easy access during rewriting. + IntegerType *IntTy; + + // The offset of the partition user currently being rewritten. + uint64_t BeginOffset, EndOffset; + Use *OldUse; + Instruction *OldPtr; + + // The name prefix to use when rewriting instructions for this alloca. + std::string NamePrefix; + +public: + AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P, + AllocaPartitioning::iterator PI, + SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI, + uint64_t NewBeginOffset, uint64_t NewEndOffset) + : TD(TD), P(P), Pass(Pass), + OldAI(OldAI), NewAI(NewAI), + NewAllocaBeginOffset(NewBeginOffset), + NewAllocaEndOffset(NewEndOffset), + NewAllocaTy(NewAI.getAllocatedType()), + VecTy(), ElementTy(), ElementSize(), IntTy(), + BeginOffset(), EndOffset() { + } + + /// \brief Visit the users of the alloca partition and rewrite them. + bool visitUsers(AllocaPartitioning::const_use_iterator I, + AllocaPartitioning::const_use_iterator E) { + if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P, + NewAllocaBeginOffset, NewAllocaEndOffset, + I, E)) { + ++NumVectorized; + VecTy = cast<VectorType>(NewAI.getAllocatedType()); + ElementTy = VecTy->getElementType(); + assert((VecTy->getScalarSizeInBits() % 8) == 0 && + "Only multiple-of-8 sized vector elements are viable"); + ElementSize = VecTy->getScalarSizeInBits() / 8; + } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(), + NewAllocaBeginOffset, P, I, E)) { + IntTy = Type::getIntNTy(NewAI.getContext(), + TD.getTypeSizeInBits(NewAI.getAllocatedType())); + } + bool CanSROA = true; + for (; I != E; ++I) { + if (!I->U) + continue; // Skip dead uses. + BeginOffset = I->BeginOffset; + EndOffset = I->EndOffset; + OldUse = I->U; + OldPtr = cast<Instruction>(I->U->get()); + NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str(); + CanSROA &= visit(cast<Instruction>(I->U->getUser())); + } + if (VecTy) { + assert(CanSROA); + VecTy = 0; + ElementTy = 0; + ElementSize = 0; + } + if (IntTy) { + assert(CanSROA); + IntTy = 0; + } + return CanSROA; + } + +private: + // Every instruction which can end up as a user must have a rewrite rule. + bool visitInstruction(Instruction &I) { + DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n"); + llvm_unreachable("No rewrite rule for this instruction!"); + } + + Twine getName(const Twine &Suffix) { + return NamePrefix + Suffix; + } + + Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) { + assert(BeginOffset >= NewAllocaBeginOffset); + APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset); + return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName("")); + } + + /// \brief Compute suitable alignment to access an offset into the new alloca. + unsigned getOffsetAlign(uint64_t Offset) { + unsigned NewAIAlign = NewAI.getAlignment(); + if (!NewAIAlign) + NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType()); + return MinAlign(NewAIAlign, Offset); + } + + /// \brief Compute suitable alignment to access this partition of the new + /// alloca. + unsigned getPartitionAlign() { + return getOffsetAlign(BeginOffset - NewAllocaBeginOffset); + } + + /// \brief Compute suitable alignment to access a type at an offset of the + /// new alloca. + /// + /// \returns zero if the type's ABI alignment is a suitable alignment, + /// otherwise returns the maximal suitable alignment. + unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) { + unsigned Align = getOffsetAlign(Offset); + return Align == TD.getABITypeAlignment(Ty) ? 0 : Align; + } + + /// \brief Compute suitable alignment to access a type at the beginning of + /// this partition of the new alloca. + /// + /// See \c getOffsetTypeAlign for details; this routine delegates to it. + unsigned getPartitionTypeAlign(Type *Ty) { + return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset); + } + + ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) { + assert(VecTy && "Can only call getIndex when rewriting a vector"); + uint64_t RelOffset = Offset - NewAllocaBeginOffset; + assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds"); + uint32_t Index = RelOffset / ElementSize; + assert(Index * ElementSize == RelOffset); + return IRB.getInt32(Index); + } + + void deleteIfTriviallyDead(Value *V) { + Instruction *I = cast<Instruction>(V); + if (isInstructionTriviallyDead(I)) + Pass.DeadInsts.insert(I); + } + + Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) { + Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".load")); + if (LI.getType() == VecTy->getElementType() || + BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) { + V = IRB.CreateExtractElement(V, getIndex(IRB, BeginOffset), + getName(".extract")); + } + return V; + } + + Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) { + assert(IntTy && "We cannot insert an integer to the alloca"); + assert(!LI.isVolatile()); + Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".load")); + V = convertValue(TD, IRB, V, IntTy); + assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); + uint64_t Offset = BeginOffset - NewAllocaBeginOffset; + if (Offset > 0 || EndOffset < NewAllocaEndOffset) + V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset, + getName(".extract")); + return V; + } + + bool visitLoadInst(LoadInst &LI) { + DEBUG(dbgs() << " original: " << LI << "\n"); + Value *OldOp = LI.getOperand(0); + assert(OldOp == OldPtr); + IRBuilder<> IRB(&LI); + + uint64_t Size = EndOffset - BeginOffset; + bool IsSplitIntLoad = Size < TD.getTypeStoreSize(LI.getType()); + + // If this memory access can be shown to *statically* extend outside the + // bounds of the original allocation it's behavior is undefined. Rather + // than trying to transform it, just replace it with undef. + // FIXME: We should do something more clever for functions being + // instrumented by asan. + // FIXME: Eventually, once ASan and friends can flush out bugs here, this + // should be transformed to a load of null making it unreachable. + uint64_t OldAllocSize = TD.getTypeAllocSize(OldAI.getAllocatedType()); + if (TD.getTypeStoreSize(LI.getType()) > OldAllocSize) { + LI.replaceAllUsesWith(UndefValue::get(LI.getType())); + Pass.DeadInsts.insert(&LI); + deleteIfTriviallyDead(OldOp); + DEBUG(dbgs() << " to: undef!!\n"); + return true; + } + + Type *TargetTy = IsSplitIntLoad ? Type::getIntNTy(LI.getContext(), Size * 8) + : LI.getType(); + bool IsPtrAdjusted = false; + Value *V; + if (VecTy) { + V = rewriteVectorizedLoadInst(IRB, LI, OldOp); + } else if (IntTy && LI.getType()->isIntegerTy()) { + V = rewriteIntegerLoad(IRB, LI); + } else if (BeginOffset == NewAllocaBeginOffset && + canConvertValue(TD, NewAllocaTy, LI.getType())) { + V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + LI.isVolatile(), getName(".load")); + } else { + Type *LTy = TargetTy->getPointerTo(); + V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy), + getPartitionTypeAlign(TargetTy), + LI.isVolatile(), getName(".load")); + IsPtrAdjusted = true; + } + V = convertValue(TD, IRB, V, TargetTy); + + if (IsSplitIntLoad) { + assert(!LI.isVolatile()); + assert(LI.getType()->isIntegerTy() && + "Only integer type loads and stores are split"); + assert(LI.getType()->getIntegerBitWidth() == + TD.getTypeStoreSizeInBits(LI.getType()) && + "Non-byte-multiple bit width"); + assert(LI.getType()->getIntegerBitWidth() == + TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) && + "Only alloca-wide loads can be split and recomposed"); + // Move the insertion point just past the load so that we can refer to it. + IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI))); + // Create a placeholder value with the same type as LI to use as the + // basis for the new value. This allows us to replace the uses of LI with + // the computed value, and then replace the placeholder with LI, leaving + // LI only used for this computation. + Value *Placeholder + = new LoadInst(UndefValue::get(LI.getType()->getPointerTo())); + V = insertInteger(TD, IRB, Placeholder, V, BeginOffset, + getName(".insert")); + LI.replaceAllUsesWith(V); + Placeholder->replaceAllUsesWith(&LI); + delete Placeholder; + } else { + LI.replaceAllUsesWith(V); + } + + Pass.DeadInsts.insert(&LI); + deleteIfTriviallyDead(OldOp); + DEBUG(dbgs() << " to: " << *V << "\n"); + return !LI.isVolatile() && !IsPtrAdjusted; + } + + bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V, + StoreInst &SI, Value *OldOp) { + if (V->getType() == ElementTy || + BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) { + if (V->getType() != ElementTy) + V = convertValue(TD, IRB, V, ElementTy); + LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".load")); + V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset), + getName(".insert")); + } else if (V->getType() != VecTy) { + V = convertValue(TD, IRB, V, VecTy); + } + StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment()); + Pass.DeadInsts.insert(&SI); + + (void)Store; + DEBUG(dbgs() << " to: " << *Store << "\n"); + return true; + } + + bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) { + assert(IntTy && "We cannot extract an integer from the alloca"); + assert(!SI.isVolatile()); + if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) { + Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".oldload")); + Old = convertValue(TD, IRB, Old, IntTy); + assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); + uint64_t Offset = BeginOffset - NewAllocaBeginOffset; + V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset, + getName(".insert")); + } + V = convertValue(TD, IRB, V, NewAllocaTy); + StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment()); + Pass.DeadInsts.insert(&SI); + (void)Store; + DEBUG(dbgs() << " to: " << *Store << "\n"); + return true; + } + + bool visitStoreInst(StoreInst &SI) { + DEBUG(dbgs() << " original: " << SI << "\n"); + Value *OldOp = SI.getOperand(1); + assert(OldOp == OldPtr); + IRBuilder<> IRB(&SI); + + Value *V = SI.getValueOperand(); + + // Strip all inbounds GEPs and pointer casts to try to dig out any root + // alloca that should be re-examined after promoting this alloca. + if (V->getType()->isPointerTy()) + if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets())) + Pass.PostPromotionWorklist.insert(AI); + + uint64_t Size = EndOffset - BeginOffset; + if (Size < TD.getTypeStoreSize(V->getType())) { + assert(!SI.isVolatile()); + assert(V->getType()->isIntegerTy() && + "Only integer type loads and stores are split"); + assert(V->getType()->getIntegerBitWidth() == + TD.getTypeStoreSizeInBits(V->getType()) && + "Non-byte-multiple bit width"); + assert(V->getType()->getIntegerBitWidth() == + TD.getTypeSizeInBits(OldAI.getAllocatedType()) && + "Only alloca-wide stores can be split and recomposed"); + IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8); + V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset, + getName(".extract")); + } + + if (VecTy) + return rewriteVectorizedStoreInst(IRB, V, SI, OldOp); + if (IntTy && V->getType()->isIntegerTy()) + return rewriteIntegerStore(IRB, V, SI); + + StoreInst *NewSI; + if (BeginOffset == NewAllocaBeginOffset && + canConvertValue(TD, V->getType(), NewAllocaTy)) { + V = convertValue(TD, IRB, V, NewAllocaTy); + NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(), + SI.isVolatile()); + } else { + Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo()); + NewSI = IRB.CreateAlignedStore(V, NewPtr, + getPartitionTypeAlign(V->getType()), + SI.isVolatile()); + } + (void)NewSI; + Pass.DeadInsts.insert(&SI); + deleteIfTriviallyDead(OldOp); + + DEBUG(dbgs() << " to: " << *NewSI << "\n"); + return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile(); + } + + bool visitMemSetInst(MemSetInst &II) { + DEBUG(dbgs() << " original: " << II << "\n"); + IRBuilder<> IRB(&II); + assert(II.getRawDest() == OldPtr); + + // If the memset has a variable size, it cannot be split, just adjust the + // pointer to the new alloca. + if (!isa<Constant>(II.getLength())) { + II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType())); + Type *CstTy = II.getAlignmentCst()->getType(); + II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign())); + + deleteIfTriviallyDead(OldPtr); + return false; + } + + // Record this instruction for deletion. + Pass.DeadInsts.insert(&II); + + Type *AllocaTy = NewAI.getAllocatedType(); + Type *ScalarTy = AllocaTy->getScalarType(); + + // If this doesn't map cleanly onto the alloca type, and that type isn't + // a single value type, just emit a memset. + if (!VecTy && !IntTy && + (BeginOffset != NewAllocaBeginOffset || + EndOffset != NewAllocaEndOffset || + !AllocaTy->isSingleValueType() || + !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) { + Type *SizeTy = II.getLength()->getType(); + Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset); + CallInst *New + = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB, + II.getRawDest()->getType()), + II.getValue(), Size, getPartitionAlign(), + II.isVolatile()); + (void)New; + DEBUG(dbgs() << " to: " << *New << "\n"); + return false; + } + + // If we can represent this as a simple value, we have to build the actual + // value to store, which requires expanding the byte present in memset to + // a sensible representation for the alloca type. This is essentially + // splatting the byte to a sufficiently wide integer, bitcasting to the + // desired scalar type, and splatting it across any desired vector type. + uint64_t Size = EndOffset - BeginOffset; + Value *V = II.getValue(); + IntegerType *VTy = cast<IntegerType>(V->getType()); + Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8); + if (Size*8 > VTy->getBitWidth()) + V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")), + ConstantExpr::getUDiv( + Constant::getAllOnesValue(SplatIntTy), + ConstantExpr::getZExt( + Constant::getAllOnesValue(V->getType()), + SplatIntTy)), + getName(".isplat")); + + // If this is an element-wide memset of a vectorizable alloca, insert it. + if (VecTy && (BeginOffset > NewAllocaBeginOffset || + EndOffset < NewAllocaEndOffset)) { + if (V->getType() != ScalarTy) + V = convertValue(TD, IRB, V, ScalarTy); + StoreInst *Store = IRB.CreateAlignedStore( + IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI, + NewAI.getAlignment(), + getName(".load")), + V, getIndex(IRB, BeginOffset), + getName(".insert")), + &NewAI, NewAI.getAlignment()); + (void)Store; + DEBUG(dbgs() << " to: " << *Store << "\n"); + return true; + } + + // If this is a memset on an alloca where we can widen stores, insert the + // set integer. + if (IntTy && (BeginOffset > NewAllocaBeginOffset || + EndOffset < NewAllocaEndOffset)) { + assert(!II.isVolatile()); + Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".oldload")); + Old = convertValue(TD, IRB, Old, IntTy); + assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); + uint64_t Offset = BeginOffset - NewAllocaBeginOffset; + V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert")); + } + + if (V->getType() != AllocaTy) + V = convertValue(TD, IRB, V, AllocaTy); + + Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(), + II.isVolatile()); + (void)New; + DEBUG(dbgs() << " to: " << *New << "\n"); + return !II.isVolatile(); + } + + bool visitMemTransferInst(MemTransferInst &II) { + // Rewriting of memory transfer instructions can be a bit tricky. We break + // them into two categories: split intrinsics and unsplit intrinsics. + + DEBUG(dbgs() << " original: " << II << "\n"); + IRBuilder<> IRB(&II); + + assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr); + bool IsDest = II.getRawDest() == OldPtr; + + const AllocaPartitioning::MemTransferOffsets &MTO + = P.getMemTransferOffsets(II); + + // Compute the relative offset within the transfer. + unsigned IntPtrWidth = TD.getPointerSizeInBits(); + APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin + : MTO.SourceBegin)); + + unsigned Align = II.getAlignment(); + if (Align > 1) + Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(), + MinAlign(II.getAlignment(), getPartitionAlign())); + + // For unsplit intrinsics, we simply modify the source and destination + // pointers in place. This isn't just an optimization, it is a matter of + // correctness. With unsplit intrinsics we may be dealing with transfers + // within a single alloca before SROA ran, or with transfers that have + // a variable length. We may also be dealing with memmove instead of + // memcpy, and so simply updating the pointers is the necessary for us to + // update both source and dest of a single call. + if (!MTO.IsSplittable) { + Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource(); + if (IsDest) + II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType())); + else + II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType())); + + Type *CstTy = II.getAlignmentCst()->getType(); + II.setAlignment(ConstantInt::get(CstTy, Align)); + + DEBUG(dbgs() << " to: " << II << "\n"); + deleteIfTriviallyDead(OldOp); + return false; + } + // For split transfer intrinsics we have an incredibly useful assurance: + // the source and destination do not reside within the same alloca, and at + // least one of them does not escape. This means that we can replace + // memmove with memcpy, and we don't need to worry about all manner of + // downsides to splitting and transforming the operations. + + // If this doesn't map cleanly onto the alloca type, and that type isn't + // a single value type, just emit a memcpy. + bool EmitMemCpy + = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset || + EndOffset != NewAllocaEndOffset || + !NewAI.getAllocatedType()->isSingleValueType()); + + // If we're just going to emit a memcpy, the alloca hasn't changed, and the + // size hasn't been shrunk based on analysis of the viable range, this is + // a no-op. + if (EmitMemCpy && &OldAI == &NewAI) { + uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin; + uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd; + // Ensure the start lines up. + assert(BeginOffset == OrigBegin); + (void)OrigBegin; + + // Rewrite the size as needed. + if (EndOffset != OrigEnd) + II.setLength(ConstantInt::get(II.getLength()->getType(), + EndOffset - BeginOffset)); + return false; + } + // Record this instruction for deletion. + Pass.DeadInsts.insert(&II); + + bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset && + EndOffset == NewAllocaEndOffset; + bool IsVectorElement = VecTy && !IsWholeAlloca; + uint64_t Size = EndOffset - BeginOffset; + IntegerType *SubIntTy + = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0; + + Type *OtherPtrTy = IsDest ? II.getRawSource()->getType() + : II.getRawDest()->getType(); + if (!EmitMemCpy) { + if (IsVectorElement) + OtherPtrTy = VecTy->getElementType()->getPointerTo(); + else if (IntTy && !IsWholeAlloca) + OtherPtrTy = SubIntTy->getPointerTo(); + else + OtherPtrTy = NewAI.getType(); + } + + // Compute the other pointer, folding as much as possible to produce + // a single, simple GEP in most cases. + Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest(); + OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy, + getName("." + OtherPtr->getName())); + + // Strip all inbounds GEPs and pointer casts to try to dig out any root + // alloca that should be re-examined after rewriting this instruction. + if (AllocaInst *AI + = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) + Pass.Worklist.insert(AI); + + if (EmitMemCpy) { + Value *OurPtr + = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType() + : II.getRawSource()->getType()); + Type *SizeTy = II.getLength()->getType(); + Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset); + + CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr, + IsDest ? OtherPtr : OurPtr, + Size, Align, II.isVolatile()); + (void)New; + DEBUG(dbgs() << " to: " << *New << "\n"); + return false; + } + + // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy + // is equivalent to 1, but that isn't true if we end up rewriting this as + // a load or store. + if (!Align) + Align = 1; + + Value *SrcPtr = OtherPtr; + Value *DstPtr = &NewAI; + if (!IsDest) + std::swap(SrcPtr, DstPtr); + + Value *Src; + if (IsVectorElement && !IsDest) { + // We have to extract rather than load. + Src = IRB.CreateExtractElement( + IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")), + getIndex(IRB, BeginOffset), + getName(".copyextract")); + } else if (IntTy && !IsWholeAlloca && !IsDest) { + Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".load")); + Src = convertValue(TD, IRB, Src, IntTy); + assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); + uint64_t Offset = BeginOffset - NewAllocaBeginOffset; + Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract")); + } else { + Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(), + getName(".copyload")); + } + + if (IntTy && !IsWholeAlloca && IsDest) { + Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), + getName(".oldload")); + Old = convertValue(TD, IRB, Old, IntTy); + assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); + uint64_t Offset = BeginOffset - NewAllocaBeginOffset; + Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert")); + Src = convertValue(TD, IRB, Src, NewAllocaTy); + } + + if (IsVectorElement && IsDest) { + // We have to insert into a loaded copy before storing. + Src = IRB.CreateInsertElement( + IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")), + Src, getIndex(IRB, BeginOffset), + getName(".insert")); + } + + StoreInst *Store = cast<StoreInst>( + IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile())); + (void)Store; + DEBUG(dbgs() << " to: " << *Store << "\n"); + return !II.isVolatile(); + } + + bool visitIntrinsicInst(IntrinsicInst &II) { + assert(II.getIntrinsicID() == Intrinsic::lifetime_start || + II.getIntrinsicID() == Intrinsic::lifetime_end); + DEBUG(dbgs() << " original: " << II << "\n"); + IRBuilder<> IRB(&II); + assert(II.getArgOperand(1) == OldPtr); + + // Record this instruction for deletion. + Pass.DeadInsts.insert(&II); + + ConstantInt *Size + = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()), + EndOffset - BeginOffset); + Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType()); + Value *New; + if (II.getIntrinsicID() == Intrinsic::lifetime_start) + New = IRB.CreateLifetimeStart(Ptr, Size); + else + New = IRB.CreateLifetimeEnd(Ptr, Size); + + DEBUG(dbgs() << " to: " << *New << "\n"); + return true; + } + + bool visitPHINode(PHINode &PN) { + DEBUG(dbgs() << " original: " << PN << "\n"); + + // We would like to compute a new pointer in only one place, but have it be + // as local as possible to the PHI. To do that, we re-use the location of + // the old pointer, which necessarily must be in the right position to + // dominate the PHI. + IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr)); + + Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType()); + // Replace the operands which were using the old pointer. + std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr); + + DEBUG(dbgs() << " to: " << PN << "\n"); + deleteIfTriviallyDead(OldPtr); + return false; + } + + bool visitSelectInst(SelectInst &SI) { + DEBUG(dbgs() << " original: " << SI << "\n"); + IRBuilder<> IRB(&SI); + + // Find the operand we need to rewrite here. + bool IsTrueVal = SI.getTrueValue() == OldPtr; + if (IsTrueVal) + assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!"); + else + assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!"); + + Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType()); + SI.setOperand(IsTrueVal ? 1 : 2, NewPtr); + DEBUG(dbgs() << " to: " << SI << "\n"); + deleteIfTriviallyDead(OldPtr); + return false; + } + +}; +} + +namespace { +/// \brief Visitor to rewrite aggregate loads and stores as scalar. +/// +/// This pass aggressively rewrites all aggregate loads and stores on +/// a particular pointer (or any pointer derived from it which we can identify) +/// with scalar loads and stores. +class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> { + // Befriend the base class so it can delegate to private visit methods. + friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>; + + const DataLayout &TD; + + /// Queue of pointer uses to analyze and potentially rewrite. + SmallVector<Use *, 8> Queue; + + /// Set to prevent us from cycling with phi nodes and loops. + SmallPtrSet<User *, 8> Visited; + + /// The current pointer use being rewritten. This is used to dig up the used + /// value (as opposed to the user). + Use *U; + +public: + AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {} + + /// Rewrite loads and stores through a pointer and all pointers derived from + /// it. + bool rewrite(Instruction &I) { + DEBUG(dbgs() << " Rewriting FCA loads and stores...\n"); + enqueueUsers(I); + bool Changed = false; + while (!Queue.empty()) { + U = Queue.pop_back_val(); + Changed |= visit(cast<Instruction>(U->getUser())); + } + return Changed; + } + +private: + /// Enqueue all the users of the given instruction for further processing. + /// This uses a set to de-duplicate users. + void enqueueUsers(Instruction &I) { + for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; + ++UI) + if (Visited.insert(*UI)) + Queue.push_back(&UI.getUse()); + } + + // Conservative default is to not rewrite anything. + bool visitInstruction(Instruction &I) { return false; } + + /// \brief Generic recursive split emission class. + template <typename Derived> + class OpSplitter { + protected: + /// The builder used to form new instructions. + IRBuilder<> IRB; + /// The indices which to be used with insert- or extractvalue to select the + /// appropriate value within the aggregate. + SmallVector<unsigned, 4> Indices; + /// The indices to a GEP instruction which will move Ptr to the correct slot + /// within the aggregate. + SmallVector<Value *, 4> GEPIndices; + /// The base pointer of the original op, used as a base for GEPing the + /// split operations. + Value *Ptr; + + /// Initialize the splitter with an insertion point, Ptr and start with a + /// single zero GEP index. + OpSplitter(Instruction *InsertionPoint, Value *Ptr) + : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {} + + public: + /// \brief Generic recursive split emission routine. + /// + /// This method recursively splits an aggregate op (load or store) into + /// scalar or vector ops. It splits recursively until it hits a single value + /// and emits that single value operation via the template argument. + /// + /// The logic of this routine relies on GEPs and insertvalue and + /// extractvalue all operating with the same fundamental index list, merely + /// formatted differently (GEPs need actual values). + /// + /// \param Ty The type being split recursively into smaller ops. + /// \param Agg The aggregate value being built up or stored, depending on + /// whether this is splitting a load or a store respectively. + void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) { + if (Ty->isSingleValueType()) + return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name); + + if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { + unsigned OldSize = Indices.size(); + (void)OldSize; + for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size; + ++Idx) { + assert(Indices.size() == OldSize && "Did not return to the old size"); + Indices.push_back(Idx); + GEPIndices.push_back(IRB.getInt32(Idx)); + emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx)); + GEPIndices.pop_back(); + Indices.pop_back(); + } + return; + } + + if (StructType *STy = dyn_cast<StructType>(Ty)) { + unsigned OldSize = Indices.size(); + (void)OldSize; + for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size; + ++Idx) { + assert(Indices.size() == OldSize && "Did not return to the old size"); + Indices.push_back(Idx); + GEPIndices.push_back(IRB.getInt32(Idx)); + emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx)); + GEPIndices.pop_back(); + Indices.pop_back(); + } + return; + } + + llvm_unreachable("Only arrays and structs are aggregate loadable types"); + } + }; + + struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> { + LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr) + : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {} + + /// Emit a leaf load of a single value. This is called at the leaves of the + /// recursive emission to actually load values. + void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) { + assert(Ty->isSingleValueType()); + // Load the single value and insert it using the indices. + Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices, + Name + ".gep"), + Name + ".load"); + Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert"); + DEBUG(dbgs() << " to: " << *Load << "\n"); + } + }; + + bool visitLoadInst(LoadInst &LI) { + assert(LI.getPointerOperand() == *U); + if (!LI.isSimple() || LI.getType()->isSingleValueType()) + return false; + + // We have an aggregate being loaded, split it apart. + DEBUG(dbgs() << " original: " << LI << "\n"); + LoadOpSplitter Splitter(&LI, *U); + Value *V = UndefValue::get(LI.getType()); + Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca"); + LI.replaceAllUsesWith(V); + LI.eraseFromParent(); + return true; + } + + struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> { + StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr) + : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {} + + /// Emit a leaf store of a single value. This is called at the leaves of the + /// recursive emission to actually produce stores. + void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) { + assert(Ty->isSingleValueType()); + // Extract the single value and store it using the indices. + Value *Store = IRB.CreateStore( + IRB.CreateExtractValue(Agg, Indices, Name + ".extract"), + IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep")); + (void)Store; + DEBUG(dbgs() << " to: " << *Store << "\n"); + } + }; + + bool visitStoreInst(StoreInst &SI) { + if (!SI.isSimple() || SI.getPointerOperand() != *U) + return false; + Value *V = SI.getValueOperand(); + if (V->getType()->isSingleValueType()) + return false; + + // We have an aggregate being stored, split it apart. + DEBUG(dbgs() << " original: " << SI << "\n"); + StoreOpSplitter Splitter(&SI, *U); + Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca"); + SI.eraseFromParent(); + return true; + } + + bool visitBitCastInst(BitCastInst &BC) { + enqueueUsers(BC); + return false; + } + + bool visitGetElementPtrInst(GetElementPtrInst &GEPI) { + enqueueUsers(GEPI); + return false; + } + + bool visitPHINode(PHINode &PN) { + enqueueUsers(PN); + return false; + } + + bool visitSelectInst(SelectInst &SI) { + enqueueUsers(SI); + return false; + } +}; +} + +/// \brief Strip aggregate type wrapping. +/// +/// This removes no-op aggregate types wrapping an underlying type. It will +/// strip as many layers of types as it can without changing either the type +/// size or the allocated size. +static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) { + if (Ty->isSingleValueType()) + return Ty; + + uint64_t AllocSize = DL.getTypeAllocSize(Ty); + uint64_t TypeSize = DL.getTypeSizeInBits(Ty); + + Type *InnerTy; + if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) { + InnerTy = ArrTy->getElementType(); + } else if (StructType *STy = dyn_cast<StructType>(Ty)) { + const StructLayout *SL = DL.getStructLayout(STy); + unsigned Index = SL->getElementContainingOffset(0); + InnerTy = STy->getElementType(Index); + } else { + return Ty; + } + + if (AllocSize > DL.getTypeAllocSize(InnerTy) || + TypeSize > DL.getTypeSizeInBits(InnerTy)) + return Ty; + + return stripAggregateTypeWrapping(DL, InnerTy); +} + +/// \brief Try to find a partition of the aggregate type passed in for a given +/// offset and size. +/// +/// This recurses through the aggregate type and tries to compute a subtype +/// based on the offset and size. When the offset and size span a sub-section +/// of an array, it will even compute a new array type for that sub-section, +/// and the same for structs. +/// +/// Note that this routine is very strict and tries to find a partition of the +/// type which produces the *exact* right offset and size. It is not forgiving +/// when the size or offset cause either end of type-based partition to be off. +/// Also, this is a best-effort routine. It is reasonable to give up and not +/// return a type if necessary. +static Type *getTypePartition(const DataLayout &TD, Type *Ty, + uint64_t Offset, uint64_t Size) { + if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size) + return stripAggregateTypeWrapping(TD, Ty); + if (Offset > TD.getTypeAllocSize(Ty) || + (TD.getTypeAllocSize(Ty) - Offset) < Size) + return 0; + + if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) { + // We can't partition pointers... + if (SeqTy->isPointerTy()) + return 0; + + Type *ElementTy = SeqTy->getElementType(); + uint64_t ElementSize = TD.getTypeAllocSize(ElementTy); + uint64_t NumSkippedElements = Offset / ElementSize; + if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) + if (NumSkippedElements >= ArrTy->getNumElements()) + return 0; + if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) + if (NumSkippedElements >= VecTy->getNumElements()) + return 0; + Offset -= NumSkippedElements * ElementSize; + + // First check if we need to recurse. + if (Offset > 0 || Size < ElementSize) { + // Bail if the partition ends in a different array element. + if ((Offset + Size) > ElementSize) + return 0; + // Recurse through the element type trying to peel off offset bytes. + return getTypePartition(TD, ElementTy, Offset, Size); + } + assert(Offset == 0); + + if (Size == ElementSize) + return stripAggregateTypeWrapping(TD, ElementTy); + assert(Size > ElementSize); + uint64_t NumElements = Size / ElementSize; + if (NumElements * ElementSize != Size) + return 0; + return ArrayType::get(ElementTy, NumElements); + } + + StructType *STy = dyn_cast<StructType>(Ty); + if (!STy) + return 0; + + const StructLayout *SL = TD.getStructLayout(STy); + if (Offset >= SL->getSizeInBytes()) + return 0; + uint64_t EndOffset = Offset + Size; + if (EndOffset > SL->getSizeInBytes()) + return 0; + + unsigned Index = SL->getElementContainingOffset(Offset); + Offset -= SL->getElementOffset(Index); + + Type *ElementTy = STy->getElementType(Index); + uint64_t ElementSize = TD.getTypeAllocSize(ElementTy); + if (Offset >= ElementSize) + return 0; // The offset points into alignment padding. + + // See if any partition must be contained by the element. + if (Offset > 0 || Size < ElementSize) { + if ((Offset + Size) > ElementSize) + return 0; + return getTypePartition(TD, ElementTy, Offset, Size); + } + assert(Offset == 0); + + if (Size == ElementSize) + return stripAggregateTypeWrapping(TD, ElementTy); + + StructType::element_iterator EI = STy->element_begin() + Index, + EE = STy->element_end(); + if (EndOffset < SL->getSizeInBytes()) { + unsigned EndIndex = SL->getElementContainingOffset(EndOffset); + if (Index == EndIndex) + return 0; // Within a single element and its padding. + + // Don't try to form "natural" types if the elements don't line up with the + // expected size. + // FIXME: We could potentially recurse down through the last element in the + // sub-struct to find a natural end point. + if (SL->getElementOffset(EndIndex) != EndOffset) + return 0; + + assert(Index < EndIndex); + EE = STy->element_begin() + EndIndex; + } + + // Try to build up a sub-structure. + StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE), + STy->isPacked()); + const StructLayout *SubSL = TD.getStructLayout(SubTy); + if (Size != SubSL->getSizeInBytes()) + return 0; // The sub-struct doesn't have quite the size needed. + + return SubTy; +} + +/// \brief Rewrite an alloca partition's users. +/// +/// This routine drives both of the rewriting goals of the SROA pass. It tries +/// to rewrite uses of an alloca partition to be conducive for SSA value +/// promotion. If the partition needs a new, more refined alloca, this will +/// build that new alloca, preserving as much type information as possible, and +/// rewrite the uses of the old alloca to point at the new one and have the +/// appropriate new offsets. It also evaluates how successful the rewrite was +/// at enabling promotion and if it was successful queues the alloca to be +/// promoted. +bool SROA::rewriteAllocaPartition(AllocaInst &AI, + AllocaPartitioning &P, + AllocaPartitioning::iterator PI) { + uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset; + bool IsLive = false; + for (AllocaPartitioning::use_iterator UI = P.use_begin(PI), + UE = P.use_end(PI); + UI != UE && !IsLive; ++UI) + if (UI->U) + IsLive = true; + if (!IsLive) + return false; // No live uses left of this partition. + + DEBUG(dbgs() << "Speculating PHIs and selects in partition " + << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n"); + + PHIOrSelectSpeculator Speculator(*TD, P, *this); + DEBUG(dbgs() << " speculating "); + DEBUG(P.print(dbgs(), PI, "")); + Speculator.visitUsers(PI); + + // Try to compute a friendly type for this partition of the alloca. This + // won't always succeed, in which case we fall back to a legal integer type + // or an i8 array of an appropriate size. + Type *AllocaTy = 0; + if (Type *PartitionTy = P.getCommonType(PI)) + if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize) + AllocaTy = PartitionTy; + if (!AllocaTy) + if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(), + PI->BeginOffset, AllocaSize)) + AllocaTy = PartitionTy; + if ((!AllocaTy || + (AllocaTy->isArrayTy() && + AllocaTy->getArrayElementType()->isIntegerTy())) && + TD->isLegalInteger(AllocaSize * 8)) + AllocaTy = Type::getIntNTy(*C, AllocaSize * 8); + if (!AllocaTy) + AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize); + assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize); + + // Check for the case where we're going to rewrite to a new alloca of the + // exact same type as the original, and with the same access offsets. In that + // case, re-use the existing alloca, but still run through the rewriter to + // performe phi and select speculation. + AllocaInst *NewAI; + if (AllocaTy == AI.getAllocatedType()) { + assert(PI->BeginOffset == 0 && + "Non-zero begin offset but same alloca type"); + assert(PI == P.begin() && "Begin offset is zero on later partition"); + NewAI = &AI; + } else { + unsigned Alignment = AI.getAlignment(); + if (!Alignment) { + // The minimum alignment which users can rely on when the explicit + // alignment is omitted or zero is that required by the ABI for this + // type. + Alignment = TD->getABITypeAlignment(AI.getAllocatedType()); + } + Alignment = MinAlign(Alignment, PI->BeginOffset); + // If we will get at least this much alignment from the type alone, leave + // the alloca's alignment unconstrained. + if (Alignment <= TD->getABITypeAlignment(AllocaTy)) + Alignment = 0; + NewAI = new AllocaInst(AllocaTy, 0, Alignment, + AI.getName() + ".sroa." + Twine(PI - P.begin()), + &AI); + ++NumNewAllocas; + } + + DEBUG(dbgs() << "Rewriting alloca partition " + << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: " + << *NewAI << "\n"); + + // Track the high watermark of the post-promotion worklist. We will reset it + // to this point if the alloca is not in fact scheduled for promotion. + unsigned PPWOldSize = PostPromotionWorklist.size(); + + AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI, + PI->BeginOffset, PI->EndOffset); + DEBUG(dbgs() << " rewriting "); + DEBUG(P.print(dbgs(), PI, "")); + bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI)); + if (Promotable) { + DEBUG(dbgs() << " and queuing for promotion\n"); + PromotableAllocas.push_back(NewAI); + } else if (NewAI != &AI) { + // If we can't promote the alloca, iterate on it to check for new + // refinements exposed by splitting the current alloca. Don't iterate on an + // alloca which didn't actually change and didn't get promoted. + Worklist.insert(NewAI); + } + + // Drop any post-promotion work items if promotion didn't happen. + if (!Promotable) + while (PostPromotionWorklist.size() > PPWOldSize) + PostPromotionWorklist.pop_back(); + + return true; +} + +/// \brief Walks the partitioning of an alloca rewriting uses of each partition. +bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) { + bool Changed = false; + for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE; + ++PI) + Changed |= rewriteAllocaPartition(AI, P, PI); + + return Changed; +} + +/// \brief Analyze an alloca for SROA. +/// +/// This analyzes the alloca to ensure we can reason about it, builds +/// a partitioning of the alloca, and then hands it off to be split and +/// rewritten as needed. +bool SROA::runOnAlloca(AllocaInst &AI) { + DEBUG(dbgs() << "SROA alloca: " << AI << "\n"); + ++NumAllocasAnalyzed; + + // Special case dead allocas, as they're trivial. + if (AI.use_empty()) { + AI.eraseFromParent(); + return true; + } + + // Skip alloca forms that this analysis can't handle. + if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() || + TD->getTypeAllocSize(AI.getAllocatedType()) == 0) + return false; + + bool Changed = false; + + // First, split any FCA loads and stores touching this alloca to promote + // better splitting and promotion opportunities. + AggLoadStoreRewriter AggRewriter(*TD); + Changed |= AggRewriter.rewrite(AI); + + // Build the partition set using a recursive instruction-visiting builder. + AllocaPartitioning P(*TD, AI); + DEBUG(P.print(dbgs())); + if (P.isEscaped()) + return Changed; + + // Delete all the dead users of this alloca before splitting and rewriting it. + for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(), + DE = P.dead_user_end(); + DI != DE; ++DI) { + Changed = true; + (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType())); + DeadInsts.insert(*DI); + } + for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(), + DE = P.dead_op_end(); + DO != DE; ++DO) { + Value *OldV = **DO; + // Clobber the use with an undef value. + **DO = UndefValue::get(OldV->getType()); + if (Instruction *OldI = dyn_cast<Instruction>(OldV)) + if (isInstructionTriviallyDead(OldI)) { + Changed = true; + DeadInsts.insert(OldI); + } + } + + // No partitions to split. Leave the dead alloca for a later pass to clean up. + if (P.begin() == P.end()) + return Changed; + + return splitAlloca(AI, P) || Changed; +} + +/// \brief Delete the dead instructions accumulated in this run. +/// +/// Recursively deletes the dead instructions we've accumulated. This is done +/// at the very end to maximize locality of the recursive delete and to +/// minimize the problems of invalidated instruction pointers as such pointers +/// are used heavily in the intermediate stages of the algorithm. +/// +/// We also record the alloca instructions deleted here so that they aren't +/// subsequently handed to mem2reg to promote. +void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) { + while (!DeadInsts.empty()) { + Instruction *I = DeadInsts.pop_back_val(); + DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n"); + + I->replaceAllUsesWith(UndefValue::get(I->getType())); + + for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) + if (Instruction *U = dyn_cast<Instruction>(*OI)) { + // Zero out the operand and see if it becomes trivially dead. + *OI = 0; + if (isInstructionTriviallyDead(U)) + DeadInsts.insert(U); + } + + if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) + DeletedAllocas.insert(AI); + + ++NumDeleted; + I->eraseFromParent(); + } +} + +/// \brief Promote the allocas, using the best available technique. +/// +/// This attempts to promote whatever allocas have been identified as viable in +/// the PromotableAllocas list. If that list is empty, there is nothing to do. +/// If there is a domtree available, we attempt to promote using the full power +/// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is +/// based on the SSAUpdater utilities. This function returns whether any +/// promotion occured. +bool SROA::promoteAllocas(Function &F) { + if (PromotableAllocas.empty()) + return false; + + NumPromoted += PromotableAllocas.size(); + + if (DT && !ForceSSAUpdater) { + DEBUG(dbgs() << "Promoting allocas with mem2reg...\n"); + PromoteMemToReg(PromotableAllocas, *DT); + PromotableAllocas.clear(); + return true; + } + + DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n"); + SSAUpdater SSA; + DIBuilder DIB(*F.getParent()); + SmallVector<Instruction*, 64> Insts; + + for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) { + AllocaInst *AI = PromotableAllocas[Idx]; + for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); + UI != UE;) { + Instruction *I = cast<Instruction>(*UI++); + // FIXME: Currently the SSAUpdater infrastructure doesn't reason about + // lifetime intrinsics and so we strip them (and the bitcasts+GEPs + // leading to them) here. Eventually it should use them to optimize the + // scalar values produced. + if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { + assert(onlyUsedByLifetimeMarkers(I) && + "Found a bitcast used outside of a lifetime marker."); + while (!I->use_empty()) + cast<Instruction>(*I->use_begin())->eraseFromParent(); + I->eraseFromParent(); + continue; + } + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { + assert(II->getIntrinsicID() == Intrinsic::lifetime_start || + II->getIntrinsicID() == Intrinsic::lifetime_end); + II->eraseFromParent(); + continue; + } + + Insts.push_back(I); + } + AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts); + Insts.clear(); + } + + PromotableAllocas.clear(); + return true; +} + +namespace { + /// \brief A predicate to test whether an alloca belongs to a set. + class IsAllocaInSet { + typedef SmallPtrSet<AllocaInst *, 4> SetType; + const SetType &Set; + + public: + typedef AllocaInst *argument_type; + + IsAllocaInSet(const SetType &Set) : Set(Set) {} + bool operator()(AllocaInst *AI) const { return Set.count(AI); } + }; +} + +bool SROA::runOnFunction(Function &F) { + DEBUG(dbgs() << "SROA function: " << F.getName() << "\n"); + C = &F.getContext(); + TD = getAnalysisIfAvailable<DataLayout>(); + if (!TD) { + DEBUG(dbgs() << " Skipping SROA -- no target data!\n"); + return false; + } + DT = getAnalysisIfAvailable<DominatorTree>(); + + BasicBlock &EntryBB = F.getEntryBlock(); + for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end()); + I != E; ++I) + if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) + Worklist.insert(AI); + + bool Changed = false; + // A set of deleted alloca instruction pointers which should be removed from + // the list of promotable allocas. + SmallPtrSet<AllocaInst *, 4> DeletedAllocas; + + do { + while (!Worklist.empty()) { + Changed |= runOnAlloca(*Worklist.pop_back_val()); + deleteDeadInstructions(DeletedAllocas); + + // Remove the deleted allocas from various lists so that we don't try to + // continue processing them. + if (!DeletedAllocas.empty()) { + Worklist.remove_if(IsAllocaInSet(DeletedAllocas)); + PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas)); + PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(), + PromotableAllocas.end(), + IsAllocaInSet(DeletedAllocas)), + PromotableAllocas.end()); + DeletedAllocas.clear(); + } + } + + Changed |= promoteAllocas(F); + + Worklist = PostPromotionWorklist; + PostPromotionWorklist.clear(); + } while (!Worklist.empty()); + + return Changed; +} + +void SROA::getAnalysisUsage(AnalysisUsage &AU) const { + if (RequiresDomTree) + AU.addRequired<DominatorTree>(); + AU.setPreservesCFG(); +} diff --git a/contrib/llvm/lib/Transforms/Scalar/Scalar.cpp b/contrib/llvm/lib/Transforms/Scalar/Scalar.cpp index 48318c8..39630fd 100644 --- a/contrib/llvm/lib/Transforms/Scalar/Scalar.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/Scalar.cpp @@ -19,7 +19,7 @@ #include "llvm/PassManager.h" #include "llvm/Analysis/Passes.h" #include "llvm/Analysis/Verifier.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Transforms/Scalar.h" using namespace llvm; @@ -59,6 +59,7 @@ void llvm::initializeScalarOpts(PassRegistry &Registry) { initializeRegToMemPass(Registry); initializeSCCPPass(Registry); initializeIPSCCPPass(Registry); + initializeSROAPass(Registry); initializeSROA_DTPass(Registry); initializeSROA_SSAUpPass(Registry); initializeCFGSimplifyPassPass(Registry); diff --git a/contrib/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/contrib/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp index 6637126..a46d09c 100644 --- a/contrib/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp @@ -46,7 +46,7 @@ #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/PromoteMemToReg.h" #include "llvm/Transforms/Utils/SSAUpdater.h" @@ -56,7 +56,6 @@ STATISTIC(NumReplaced, "Number of allocas broken up"); STATISTIC(NumPromoted, "Number of allocas promoted"); STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); STATISTIC(NumConverted, "Number of aggregates converted to scalar"); -STATISTIC(NumGlobals, "Number of allocas copied from constant global"); namespace { struct SROA : public FunctionPass { @@ -88,7 +87,7 @@ namespace { private: bool HasDomTree; - TargetData *TD; + DataLayout *TD; /// DeadInsts - Keep track of instructions we have made dead, so that /// we can remove them after we are done working. @@ -183,9 +182,6 @@ namespace { void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, SmallVector<AllocaInst*, 32> &NewElts); bool ShouldAttemptScalarRepl(AllocaInst *AI); - - static MemTransferInst *isOnlyCopiedFromConstantGlobal( - AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); }; // SROA_DT - SROA that uses DominatorTree. @@ -262,7 +258,7 @@ namespace { class ConvertToScalarInfo { /// AllocaSize - The size of the alloca being considered in bytes. unsigned AllocaSize; - const TargetData &TD; + const DataLayout &TD; unsigned ScalarLoadThreshold; /// IsNotTrivial - This is set to true if there is some access to the object @@ -305,7 +301,7 @@ class ConvertToScalarInfo { bool HadDynamicAccess; public: - explicit ConvertToScalarInfo(unsigned Size, const TargetData &td, + explicit ConvertToScalarInfo(unsigned Size, const DataLayout &td, unsigned SLT) : AllocaSize(Size), TD(td), ScalarLoadThreshold(SLT), IsNotTrivial(false), ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false), @@ -1024,11 +1020,11 @@ ConvertScalar_InsertValue(Value *SV, Value *Old, bool SROA::runOnFunction(Function &F) { - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); bool Changed = performPromotion(F); - // FIXME: ScalarRepl currently depends on TargetData more than it + // FIXME: ScalarRepl currently depends on DataLayout more than it // theoretically needs to. It should be refactored in order to support // target-independent IR. Until this is done, just skip the actual // scalar-replacement portion of this pass. @@ -1138,7 +1134,7 @@ public: /// /// We can do this to a select if its only uses are loads and if the operand to /// the select can be loaded unconditionally. -static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { +static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *TD) { bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); @@ -1176,7 +1172,7 @@ static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { /// /// We can do this to a select if its only uses are loads and if the operand to /// the select can be loaded unconditionally. -static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { +static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *TD) { // For now, we can only do this promotion if the load is in the same block as // the PHI, and if there are no stores between the phi and load. // TODO: Allow recursive phi users. @@ -1240,7 +1236,7 @@ static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { /// direct (non-volatile) loads and stores to it. If the alloca is close but /// not quite there, this will transform the code to allow promotion. As such, /// it is a non-pure predicate. -static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { +static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *TD) { SetVector<Instruction*, SmallVector<Instruction*, 4>, SmallPtrSet<Instruction*, 4> > InstsToRewrite; @@ -1465,26 +1461,6 @@ bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { return false; } -/// getPointeeAlignment - Compute the minimum alignment of the value pointed -/// to by the given pointer. -static unsigned getPointeeAlignment(Value *V, const TargetData &TD) { - if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) - if (CE->getOpcode() == Instruction::BitCast || - (CE->getOpcode() == Instruction::GetElementPtr && - cast<GEPOperator>(CE)->hasAllZeroIndices())) - return getPointeeAlignment(CE->getOperand(0), TD); - - if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) - if (!GV->isDeclaration()) - return TD.getPreferredAlignment(GV); - - if (PointerType *PT = dyn_cast<PointerType>(V->getType())) - return TD.getABITypeAlignment(PT->getElementType()); - - return 0; -} - - // performScalarRepl - This algorithm is a simple worklist driven algorithm, // which runs on all of the alloca instructions in the function, removing them // if they are only used by getelementptr instructions. @@ -1516,29 +1492,6 @@ bool SROA::performScalarRepl(Function &F) { if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) continue; - // Check to see if this allocation is only modified by a memcpy/memmove from - // a constant global whose alignment is equal to or exceeds that of the - // allocation. If this is the case, we can change all users to use - // the constant global instead. This is commonly produced by the CFE by - // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' - // is only subsequently read. - SmallVector<Instruction *, 4> ToDelete; - if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { - if (AI->getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) { - DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); - DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); - for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) - ToDelete[i]->eraseFromParent(); - Constant *TheSrc = cast<Constant>(Copy->getSource()); - AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); - Copy->eraseFromParent(); // Don't mutate the global. - AI->eraseFromParent(); - ++NumGlobals; - Changed = true; - continue; - } - } - // Check to see if we can perform the core SROA transformation. We cannot // transform the allocation instruction if it is an array allocation // (allocations OF arrays are ok though), and an allocation of a scalar @@ -2584,7 +2537,7 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, /// HasPadding - Return true if the specified type has any structure or /// alignment padding in between the elements that would be split apart /// by SROA; return false otherwise. -static bool HasPadding(Type *Ty, const TargetData &TD) { +static bool HasPadding(Type *Ty, const DataLayout &TD) { if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { Ty = ATy->getElementType(); return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); @@ -2656,134 +2609,3 @@ bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { return true; } - - - -/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to -/// some part of a constant global variable. This intentionally only accepts -/// constant expressions because we don't can't rewrite arbitrary instructions. -static bool PointsToConstantGlobal(Value *V) { - if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) - return GV->isConstant(); - if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) - if (CE->getOpcode() == Instruction::BitCast || - CE->getOpcode() == Instruction::GetElementPtr) - return PointsToConstantGlobal(CE->getOperand(0)); - return false; -} - -/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) -/// pointer to an alloca. Ignore any reads of the pointer, return false if we -/// see any stores or other unknown uses. If we see pointer arithmetic, keep -/// track of whether it moves the pointer (with isOffset) but otherwise traverse -/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to -/// the alloca, and if the source pointer is a pointer to a constant global, we -/// can optimize this. -static bool -isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, - bool isOffset, - SmallVector<Instruction *, 4> &LifetimeMarkers) { - // We track lifetime intrinsics as we encounter them. If we decide to go - // ahead and replace the value with the global, this lets the caller quickly - // eliminate the markers. - - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { - User *U = cast<Instruction>(*UI); - - if (LoadInst *LI = dyn_cast<LoadInst>(U)) { - // Ignore non-volatile loads, they are always ok. - if (!LI->isSimple()) return false; - continue; - } - - if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { - // If uses of the bitcast are ok, we are ok. - if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, - LifetimeMarkers)) - return false; - continue; - } - if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { - // If the GEP has all zero indices, it doesn't offset the pointer. If it - // doesn't, it does. - if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, - isOffset || !GEP->hasAllZeroIndices(), - LifetimeMarkers)) - return false; - continue; - } - - if (CallSite CS = U) { - // If this is the function being called then we treat it like a load and - // ignore it. - if (CS.isCallee(UI)) - continue; - - // If this is a readonly/readnone call site, then we know it is just a - // load (but one that potentially returns the value itself), so we can - // ignore it if we know that the value isn't captured. - unsigned ArgNo = CS.getArgumentNo(UI); - if (CS.onlyReadsMemory() && - (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) - continue; - - // If this is being passed as a byval argument, the caller is making a - // copy, so it is only a read of the alloca. - if (CS.isByValArgument(ArgNo)) - continue; - } - - // Lifetime intrinsics can be handled by the caller. - if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { - if (II->getIntrinsicID() == Intrinsic::lifetime_start || - II->getIntrinsicID() == Intrinsic::lifetime_end) { - assert(II->use_empty() && "Lifetime markers have no result to use!"); - LifetimeMarkers.push_back(II); - continue; - } - } - - // If this is isn't our memcpy/memmove, reject it as something we can't - // handle. - MemTransferInst *MI = dyn_cast<MemTransferInst>(U); - if (MI == 0) - return false; - - // If the transfer is using the alloca as a source of the transfer, then - // ignore it since it is a load (unless the transfer is volatile). - if (UI.getOperandNo() == 1) { - if (MI->isVolatile()) return false; - continue; - } - - // If we already have seen a copy, reject the second one. - if (TheCopy) return false; - - // If the pointer has been offset from the start of the alloca, we can't - // safely handle this. - if (isOffset) return false; - - // If the memintrinsic isn't using the alloca as the dest, reject it. - if (UI.getOperandNo() != 0) return false; - - // If the source of the memcpy/move is not a constant global, reject it. - if (!PointsToConstantGlobal(MI->getSource())) - return false; - - // Otherwise, the transform is safe. Remember the copy instruction. - TheCopy = MI; - } - return true; -} - -/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only -/// modified by a copy from a constant global. If we can prove this, we can -/// replace any uses of the alloca with uses of the global directly. -MemTransferInst * -SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, - SmallVector<Instruction*, 4> &ToDelete) { - MemTransferInst *TheCopy = 0; - if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) - return TheCopy; - return 0; -} diff --git a/contrib/llvm/lib/Transforms/Scalar/SimplifyCFGPass.cpp b/contrib/llvm/lib/Transforms/Scalar/SimplifyCFGPass.cpp index d13e4ab..9f24bb6 100644 --- a/contrib/llvm/lib/Transforms/Scalar/SimplifyCFGPass.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/SimplifyCFGPass.cpp @@ -31,10 +31,11 @@ #include "llvm/Attributes.h" #include "llvm/Support/CFG.h" #include "llvm/Pass.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/Statistic.h" +#include "llvm/TargetTransformInfo.h" using namespace llvm; STATISTIC(NumSimpl, "Number of blocks simplified"); @@ -59,9 +60,9 @@ FunctionPass *llvm::createCFGSimplificationPass() { return new CFGSimplifyPass(); } -/// ChangeToUnreachable - Insert an unreachable instruction before the specified +/// changeToUnreachable - Insert an unreachable instruction before the specified /// instruction, making it and the rest of the code in the block dead. -static void ChangeToUnreachable(Instruction *I, bool UseLLVMTrap) { +static void changeToUnreachable(Instruction *I, bool UseLLVMTrap) { BasicBlock *BB = I->getParent(); // Loop over all of the successors, removing BB's entry from any PHI // nodes. @@ -87,8 +88,8 @@ static void ChangeToUnreachable(Instruction *I, bool UseLLVMTrap) { } } -/// ChangeToCall - Convert the specified invoke into a normal call. -static void ChangeToCall(InvokeInst *II) { +/// changeToCall - Convert the specified invoke into a normal call. +static void changeToCall(InvokeInst *II) { SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, "", II); NewCall->takeName(II); @@ -105,7 +106,7 @@ static void ChangeToCall(InvokeInst *II) { II->eraseFromParent(); } -static bool MarkAliveBlocks(BasicBlock *BB, +static bool markAliveBlocks(BasicBlock *BB, SmallPtrSet<BasicBlock*, 128> &Reachable) { SmallVector<BasicBlock*, 128> Worklist; @@ -129,7 +130,7 @@ static bool MarkAliveBlocks(BasicBlock *BB, ++BBI; if (!isa<UnreachableInst>(BBI)) { // Don't insert a call to llvm.trap right before the unreachable. - ChangeToUnreachable(BBI, false); + changeToUnreachable(BBI, false); Changed = true; } break; @@ -148,7 +149,7 @@ static bool MarkAliveBlocks(BasicBlock *BB, if (isa<UndefValue>(Ptr) || (isa<ConstantPointerNull>(Ptr) && SI->getPointerAddressSpace() == 0)) { - ChangeToUnreachable(SI, true); + changeToUnreachable(SI, true); Changed = true; break; } @@ -159,7 +160,7 @@ static bool MarkAliveBlocks(BasicBlock *BB, if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) { Value *Callee = II->getCalledValue(); if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { - ChangeToUnreachable(II, true); + changeToUnreachable(II, true); Changed = true; } else if (II->doesNotThrow()) { if (II->use_empty() && II->onlyReadsMemory()) { @@ -168,7 +169,7 @@ static bool MarkAliveBlocks(BasicBlock *BB, II->getUnwindDest()->removePredecessor(II->getParent()); II->eraseFromParent(); } else - ChangeToCall(II); + changeToCall(II); Changed = true; } } @@ -180,12 +181,12 @@ static bool MarkAliveBlocks(BasicBlock *BB, return Changed; } -/// RemoveUnreachableBlocksFromFn - Remove blocks that are not reachable, even +/// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even /// if they are in a dead cycle. Return true if a change was made, false /// otherwise. -static bool RemoveUnreachableBlocksFromFn(Function &F) { +static bool removeUnreachableBlocksFromFn(Function &F) { SmallPtrSet<BasicBlock*, 128> Reachable; - bool Changed = MarkAliveBlocks(F.begin(), Reachable); + bool Changed = markAliveBlocks(F.begin(), Reachable); // If there are unreachable blocks in the CFG... if (Reachable.size() == F.size()) @@ -215,9 +216,9 @@ static bool RemoveUnreachableBlocksFromFn(Function &F) { return true; } -/// MergeEmptyReturnBlocks - If we have more than one empty (other than phi +/// mergeEmptyReturnBlocks - If we have more than one empty (other than phi /// node) return blocks, merge them together to promote recursive block merging. -static bool MergeEmptyReturnBlocks(Function &F) { +static bool mergeEmptyReturnBlocks(Function &F) { bool Changed = false; BasicBlock *RetBlock = 0; @@ -291,9 +292,10 @@ static bool MergeEmptyReturnBlocks(Function &F) { return Changed; } -/// IterativeSimplifyCFG - Call SimplifyCFG on all the blocks in the function, +/// iterativelySimplifyCFG - Call SimplifyCFG on all the blocks in the function, /// iterating until no more changes are made. -static bool IterativeSimplifyCFG(Function &F, const TargetData *TD) { +static bool iterativelySimplifyCFG(Function &F, const DataLayout *TD, + const TargetTransformInfo *TTI) { bool Changed = false; bool LocalChange = true; while (LocalChange) { @@ -302,7 +304,7 @@ static bool IterativeSimplifyCFG(Function &F, const TargetData *TD) { // Loop over all of the basic blocks and remove them if they are unneeded... // for (Function::iterator BBIt = F.begin(); BBIt != F.end(); ) { - if (SimplifyCFG(BBIt++, TD)) { + if (SimplifyCFG(BBIt++, TD, TTI)) { LocalChange = true; ++NumSimpl; } @@ -316,25 +318,27 @@ static bool IterativeSimplifyCFG(Function &F, const TargetData *TD) { // simplify the CFG. // bool CFGSimplifyPass::runOnFunction(Function &F) { - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); - bool EverChanged = RemoveUnreachableBlocksFromFn(F); - EverChanged |= MergeEmptyReturnBlocks(F); - EverChanged |= IterativeSimplifyCFG(F, TD); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); + const TargetTransformInfo *TTI = + getAnalysisIfAvailable<TargetTransformInfo>(); + bool EverChanged = removeUnreachableBlocksFromFn(F); + EverChanged |= mergeEmptyReturnBlocks(F); + EverChanged |= iterativelySimplifyCFG(F, TD, TTI); // If neither pass changed anything, we're done. if (!EverChanged) return false; - // IterativeSimplifyCFG can (rarely) make some loops dead. If this happens, - // RemoveUnreachableBlocksFromFn is needed to nuke them, which means we should + // iterativelySimplifyCFG can (rarely) make some loops dead. If this happens, + // removeUnreachableBlocksFromFn is needed to nuke them, which means we should // iterate between the two optimizations. We structure the code like this to - // avoid reruning IterativeSimplifyCFG if the second pass of - // RemoveUnreachableBlocksFromFn doesn't do anything. - if (!RemoveUnreachableBlocksFromFn(F)) + // avoid reruning iterativelySimplifyCFG if the second pass of + // removeUnreachableBlocksFromFn doesn't do anything. + if (!removeUnreachableBlocksFromFn(F)) return true; do { - EverChanged = IterativeSimplifyCFG(F, TD); - EverChanged |= RemoveUnreachableBlocksFromFn(F); + EverChanged = iterativelySimplifyCFG(F, TD, TTI); + EverChanged |= removeUnreachableBlocksFromFn(F); } while (EverChanged); return true; diff --git a/contrib/llvm/lib/Transforms/Scalar/SimplifyLibCalls.cpp b/contrib/llvm/lib/Transforms/Scalar/SimplifyLibCalls.cpp index f110320..17d07cd 100644 --- a/contrib/llvm/lib/Transforms/Scalar/SimplifyLibCalls.cpp +++ b/contrib/llvm/lib/Transforms/Scalar/SimplifyLibCalls.cpp @@ -28,9 +28,10 @@ #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringMap.h" #include "llvm/Analysis/ValueTracking.h" +#include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Config/config.h" // FIXME: Shouldn't depend on host! using namespace llvm; @@ -38,6 +39,10 @@ using namespace llvm; STATISTIC(NumSimplified, "Number of library calls simplified"); STATISTIC(NumAnnotated, "Number of attributes added to library functions"); +static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden, + cl::init(false), + cl::desc("Enable unsafe double to float " + "shrinking for math lib calls")); //===----------------------------------------------------------------------===// // Optimizer Base Class //===----------------------------------------------------------------------===// @@ -48,7 +53,7 @@ namespace { class LibCallOptimization { protected: Function *Caller; - const TargetData *TD; + const DataLayout *TD; const TargetLibraryInfo *TLI; LLVMContext* Context; public: @@ -63,7 +68,7 @@ public: virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) =0; - Value *OptimizeCall(CallInst *CI, const TargetData *TD, + Value *OptimizeCall(CallInst *CI, const DataLayout *TD, const TargetLibraryInfo *TLI, IRBuilder<> &B) { Caller = CI->getParent()->getParent(); this->TD = TD; @@ -85,22 +90,6 @@ public: // Helper Functions //===----------------------------------------------------------------------===// -/// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the -/// value is equal or not-equal to zero. -static bool IsOnlyUsedInZeroEqualityComparison(Value *V) { - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); - UI != E; ++UI) { - if (ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) - if (IC->isEquality()) - if (Constant *C = dyn_cast<Constant>(IC->getOperand(1))) - if (C->isNullValue()) - continue; - // Unknown instruction. - return false; - } - return true; -} - static bool CallHasFloatingPointArgument(const CallInst *CI) { for (CallInst::const_op_iterator it = CI->op_begin(), e = CI->op_end(); it != e; ++it) { @@ -110,799 +99,62 @@ static bool CallHasFloatingPointArgument(const CallInst *CI) { return false; } -/// IsOnlyUsedInEqualityComparison - Return true if it is only used in equality -/// comparisons with With. -static bool IsOnlyUsedInEqualityComparison(Value *V, Value *With) { - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); - UI != E; ++UI) { - if (ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) - if (IC->isEquality() && IC->getOperand(1) == With) - continue; - // Unknown instruction. - return false; - } - return true; -} - +namespace { //===----------------------------------------------------------------------===// -// String and Memory LibCall Optimizations +// Math Library Optimizations //===----------------------------------------------------------------------===// //===---------------------------------------===// -// 'strcat' Optimizations -namespace { -struct StrCatOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strcat" function prototype. - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - FT->getReturnType() != B.getInt8PtrTy() || - FT->getParamType(0) != FT->getReturnType() || - FT->getParamType(1) != FT->getReturnType()) - return 0; - - // Extract some information from the instruction - Value *Dst = CI->getArgOperand(0); - Value *Src = CI->getArgOperand(1); - - // See if we can get the length of the input string. - uint64_t Len = GetStringLength(Src); - if (Len == 0) return 0; - --Len; // Unbias length. - - // Handle the simple, do-nothing case: strcat(x, "") -> x - if (Len == 0) - return Dst; - - // These optimizations require TargetData. - if (!TD) return 0; - - return EmitStrLenMemCpy(Src, Dst, Len, B); - } - - Value *EmitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len, IRBuilder<> &B) { - // We need to find the end of the destination string. That's where the - // memory is to be moved to. We just generate a call to strlen. - Value *DstLen = EmitStrLen(Dst, B, TD, TLI); - if (!DstLen) - return 0; - - // Now that we have the destination's length, we must index into the - // destination's pointer to get the actual memcpy destination (end of - // the string .. we're concatenating). - Value *CpyDst = B.CreateGEP(Dst, DstLen, "endptr"); - - // We have enough information to now generate the memcpy call to do the - // concatenation for us. Make a memcpy to copy the nul byte with align = 1. - B.CreateMemCpy(CpyDst, Src, - ConstantInt::get(TD->getIntPtrType(*Context), Len + 1), 1); - return Dst; - } -}; - -//===---------------------------------------===// -// 'strncat' Optimizations - -struct StrNCatOpt : public StrCatOpt { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strncat" function prototype. - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || - FT->getReturnType() != B.getInt8PtrTy() || - FT->getParamType(0) != FT->getReturnType() || - FT->getParamType(1) != FT->getReturnType() || - !FT->getParamType(2)->isIntegerTy()) - return 0; - - // Extract some information from the instruction - Value *Dst = CI->getArgOperand(0); - Value *Src = CI->getArgOperand(1); - uint64_t Len; - - // We don't do anything if length is not constant - if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2))) - Len = LengthArg->getZExtValue(); - else - return 0; - - // See if we can get the length of the input string. - uint64_t SrcLen = GetStringLength(Src); - if (SrcLen == 0) return 0; - --SrcLen; // Unbias length. - - // Handle the simple, do-nothing cases: - // strncat(x, "", c) -> x - // strncat(x, c, 0) -> x - if (SrcLen == 0 || Len == 0) return Dst; - - // These optimizations require TargetData. - if (!TD) return 0; - - // We don't optimize this case - if (Len < SrcLen) return 0; - - // strncat(x, s, c) -> strcat(x, s) - // s is constant so the strcat can be optimized further - return EmitStrLenMemCpy(Src, Dst, SrcLen, B); - } -}; - -//===---------------------------------------===// -// 'strchr' Optimizations - -struct StrChrOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strchr" function prototype. - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - FT->getReturnType() != B.getInt8PtrTy() || - FT->getParamType(0) != FT->getReturnType() || - !FT->getParamType(1)->isIntegerTy(32)) - return 0; - - Value *SrcStr = CI->getArgOperand(0); - - // If the second operand is non-constant, see if we can compute the length - // of the input string and turn this into memchr. - ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); - if (CharC == 0) { - // These optimizations require TargetData. - if (!TD) return 0; - - uint64_t Len = GetStringLength(SrcStr); - if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32))// memchr needs i32. - return 0; - - return EmitMemChr(SrcStr, CI->getArgOperand(1), // include nul. - ConstantInt::get(TD->getIntPtrType(*Context), Len), - B, TD, TLI); - } - - // Otherwise, the character is a constant, see if the first argument is - // a string literal. If so, we can constant fold. - StringRef Str; - if (!getConstantStringInfo(SrcStr, Str)) - return 0; - - // Compute the offset, make sure to handle the case when we're searching for - // zero (a weird way to spell strlen). - size_t I = CharC->getSExtValue() == 0 ? - Str.size() : Str.find(CharC->getSExtValue()); - if (I == StringRef::npos) // Didn't find the char. strchr returns null. - return Constant::getNullValue(CI->getType()); - - // strchr(s+n,c) -> gep(s+n+i,c) - return B.CreateGEP(SrcStr, B.getInt64(I), "strchr"); - } -}; - -//===---------------------------------------===// -// 'strrchr' Optimizations - -struct StrRChrOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strrchr" function prototype. - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - FT->getReturnType() != B.getInt8PtrTy() || - FT->getParamType(0) != FT->getReturnType() || - !FT->getParamType(1)->isIntegerTy(32)) - return 0; - - Value *SrcStr = CI->getArgOperand(0); - ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); - - // Cannot fold anything if we're not looking for a constant. - if (!CharC) - return 0; - - StringRef Str; - if (!getConstantStringInfo(SrcStr, Str)) { - // strrchr(s, 0) -> strchr(s, 0) - if (TD && CharC->isZero()) - return EmitStrChr(SrcStr, '\0', B, TD, TLI); - return 0; - } - - // Compute the offset. - size_t I = CharC->getSExtValue() == 0 ? - Str.size() : Str.rfind(CharC->getSExtValue()); - if (I == StringRef::npos) // Didn't find the char. Return null. - return Constant::getNullValue(CI->getType()); - - // strrchr(s+n,c) -> gep(s+n+i,c) - return B.CreateGEP(SrcStr, B.getInt64(I), "strrchr"); - } -}; - -//===---------------------------------------===// -// 'strcmp' Optimizations - -struct StrCmpOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strcmp" function prototype. - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - !FT->getReturnType()->isIntegerTy(32) || - FT->getParamType(0) != FT->getParamType(1) || - FT->getParamType(0) != B.getInt8PtrTy()) - return 0; - - Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1); - if (Str1P == Str2P) // strcmp(x,x) -> 0 - return ConstantInt::get(CI->getType(), 0); - - StringRef Str1, Str2; - bool HasStr1 = getConstantStringInfo(Str1P, Str1); - bool HasStr2 = getConstantStringInfo(Str2P, Str2); - - // strcmp(x, y) -> cnst (if both x and y are constant strings) - if (HasStr1 && HasStr2) - return ConstantInt::get(CI->getType(), Str1.compare(Str2)); - - if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x - return B.CreateNeg(B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), - CI->getType())); - - if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x - return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType()); - - // strcmp(P, "x") -> memcmp(P, "x", 2) - uint64_t Len1 = GetStringLength(Str1P); - uint64_t Len2 = GetStringLength(Str2P); - if (Len1 && Len2) { - // These optimizations require TargetData. - if (!TD) return 0; - - return EmitMemCmp(Str1P, Str2P, - ConstantInt::get(TD->getIntPtrType(*Context), - std::min(Len1, Len2)), B, TD, TLI); - } - - return 0; - } -}; - -//===---------------------------------------===// -// 'strncmp' Optimizations - -struct StrNCmpOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strncmp" function prototype. - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || - !FT->getReturnType()->isIntegerTy(32) || - FT->getParamType(0) != FT->getParamType(1) || - FT->getParamType(0) != B.getInt8PtrTy() || - !FT->getParamType(2)->isIntegerTy()) - return 0; - - Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1); - if (Str1P == Str2P) // strncmp(x,x,n) -> 0 - return ConstantInt::get(CI->getType(), 0); - - // Get the length argument if it is constant. - uint64_t Length; - if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2))) - Length = LengthArg->getZExtValue(); - else - return 0; - - if (Length == 0) // strncmp(x,y,0) -> 0 - return ConstantInt::get(CI->getType(), 0); - - if (TD && Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1) - return EmitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, TD, TLI); - - StringRef Str1, Str2; - bool HasStr1 = getConstantStringInfo(Str1P, Str1); - bool HasStr2 = getConstantStringInfo(Str2P, Str2); - - // strncmp(x, y) -> cnst (if both x and y are constant strings) - if (HasStr1 && HasStr2) { - StringRef SubStr1 = Str1.substr(0, Length); - StringRef SubStr2 = Str2.substr(0, Length); - return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2)); - } - - if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x - return B.CreateNeg(B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), - CI->getType())); - - if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x - return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType()); - - return 0; - } -}; - - -//===---------------------------------------===// -// 'strcpy' Optimizations - -struct StrCpyOpt : public LibCallOptimization { - bool OptChkCall; // True if it's optimizing a __strcpy_chk libcall. - - StrCpyOpt(bool c) : OptChkCall(c) {} - - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "strcpy" function prototype. - unsigned NumParams = OptChkCall ? 3 : 2; - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != NumParams || - FT->getReturnType() != FT->getParamType(0) || - FT->getParamType(0) != FT->getParamType(1) || - FT->getParamType(0) != B.getInt8PtrTy()) - return 0; - - Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); - if (Dst == Src) // strcpy(x,x) -> x - return Src; - - // These optimizations require TargetData. - if (!TD) return 0; - - // See if we can get the length of the input string. - uint64_t Len = GetStringLength(Src); - if (Len == 0) return 0; - - // We have enough information to now generate the memcpy call to do the - // concatenation for us. Make a memcpy to copy the nul byte with align = 1. - if (!OptChkCall || - !EmitMemCpyChk(Dst, Src, - ConstantInt::get(TD->getIntPtrType(*Context), Len), - CI->getArgOperand(2), B, TD, TLI)) - B.CreateMemCpy(Dst, Src, - ConstantInt::get(TD->getIntPtrType(*Context), Len), 1); - return Dst; - } -}; - -//===---------------------------------------===// -// 'stpcpy' Optimizations - -struct StpCpyOpt: public LibCallOptimization { - bool OptChkCall; // True if it's optimizing a __stpcpy_chk libcall. - - StpCpyOpt(bool c) : OptChkCall(c) {} - - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // Verify the "stpcpy" function prototype. - unsigned NumParams = OptChkCall ? 3 : 2; - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != NumParams || - FT->getReturnType() != FT->getParamType(0) || - FT->getParamType(0) != FT->getParamType(1) || - FT->getParamType(0) != B.getInt8PtrTy()) - return 0; - - // These optimizations require TargetData. - if (!TD) return 0; - - Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); - if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x) - Value *StrLen = EmitStrLen(Src, B, TD, TLI); - return StrLen ? B.CreateInBoundsGEP(Dst, StrLen) : 0; - } - - // See if we can get the length of the input string. - uint64_t Len = GetStringLength(Src); - if (Len == 0) return 0; - - Value *LenV = ConstantInt::get(TD->getIntPtrType(*Context), Len); - Value *DstEnd = B.CreateGEP(Dst, - ConstantInt::get(TD->getIntPtrType(*Context), - Len - 1)); - - // We have enough information to now generate the memcpy call to do the - // copy for us. Make a memcpy to copy the nul byte with align = 1. - if (!OptChkCall || !EmitMemCpyChk(Dst, Src, LenV, CI->getArgOperand(2), B, - TD, TLI)) - B.CreateMemCpy(Dst, Src, LenV, 1); - return DstEnd; - } -}; - -//===---------------------------------------===// -// 'strncpy' Optimizations - -struct StrNCpyOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || - FT->getParamType(0) != FT->getParamType(1) || - FT->getParamType(0) != B.getInt8PtrTy() || - !FT->getParamType(2)->isIntegerTy()) - return 0; - - Value *Dst = CI->getArgOperand(0); - Value *Src = CI->getArgOperand(1); - Value *LenOp = CI->getArgOperand(2); - - // See if we can get the length of the input string. - uint64_t SrcLen = GetStringLength(Src); - if (SrcLen == 0) return 0; - --SrcLen; - - if (SrcLen == 0) { - // strncpy(x, "", y) -> memset(x, '\0', y, 1) - B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1); - return Dst; - } - - uint64_t Len; - if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp)) - Len = LengthArg->getZExtValue(); - else - return 0; - - if (Len == 0) return Dst; // strncpy(x, y, 0) -> x - - // These optimizations require TargetData. - if (!TD) return 0; - - // Let strncpy handle the zero padding - if (Len > SrcLen+1) return 0; - - // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant] - B.CreateMemCpy(Dst, Src, - ConstantInt::get(TD->getIntPtrType(*Context), Len), 1); - - return Dst; - } -}; - -//===---------------------------------------===// -// 'strlen' Optimizations - -struct StrLenOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 1 || - FT->getParamType(0) != B.getInt8PtrTy() || - !FT->getReturnType()->isIntegerTy()) - return 0; - - Value *Src = CI->getArgOperand(0); - - // Constant folding: strlen("xyz") -> 3 - if (uint64_t Len = GetStringLength(Src)) - return ConstantInt::get(CI->getType(), Len-1); - - // strlen(x) != 0 --> *x != 0 - // strlen(x) == 0 --> *x == 0 - if (IsOnlyUsedInZeroEqualityComparison(CI)) - return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType()); - return 0; - } -}; - - -//===---------------------------------------===// -// 'strpbrk' Optimizations - -struct StrPBrkOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - FT->getParamType(0) != B.getInt8PtrTy() || - FT->getParamType(1) != FT->getParamType(0) || - FT->getReturnType() != FT->getParamType(0)) - return 0; - - StringRef S1, S2; - bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); - bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); - - // strpbrk(s, "") -> NULL - // strpbrk("", s) -> NULL - if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) - return Constant::getNullValue(CI->getType()); - - // Constant folding. - if (HasS1 && HasS2) { - size_t I = S1.find_first_of(S2); - if (I == std::string::npos) // No match. - return Constant::getNullValue(CI->getType()); - - return B.CreateGEP(CI->getArgOperand(0), B.getInt64(I), "strpbrk"); - } - - // strpbrk(s, "a") -> strchr(s, 'a') - if (TD && HasS2 && S2.size() == 1) - return EmitStrChr(CI->getArgOperand(0), S2[0], B, TD, TLI); - - return 0; - } -}; - -//===---------------------------------------===// -// 'strto*' Optimizations. This handles strtol, strtod, strtof, strtoul, etc. - -struct StrToOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if ((FT->getNumParams() != 2 && FT->getNumParams() != 3) || - !FT->getParamType(0)->isPointerTy() || - !FT->getParamType(1)->isPointerTy()) - return 0; - - Value *EndPtr = CI->getArgOperand(1); - if (isa<ConstantPointerNull>(EndPtr)) { - // With a null EndPtr, this function won't capture the main argument. - // It would be readonly too, except that it still may write to errno. - CI->addAttribute(1, Attribute::NoCapture); - } - - return 0; - } -}; - -//===---------------------------------------===// -// 'strspn' Optimizations - -struct StrSpnOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - FT->getParamType(0) != B.getInt8PtrTy() || - FT->getParamType(1) != FT->getParamType(0) || - !FT->getReturnType()->isIntegerTy()) - return 0; - - StringRef S1, S2; - bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); - bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); - - // strspn(s, "") -> 0 - // strspn("", s) -> 0 - if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) - return Constant::getNullValue(CI->getType()); - - // Constant folding. - if (HasS1 && HasS2) { - size_t Pos = S1.find_first_not_of(S2); - if (Pos == StringRef::npos) Pos = S1.size(); - return ConstantInt::get(CI->getType(), Pos); - } - - return 0; - } -}; - -//===---------------------------------------===// -// 'strcspn' Optimizations - -struct StrCSpnOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - FT->getParamType(0) != B.getInt8PtrTy() || - FT->getParamType(1) != FT->getParamType(0) || - !FT->getReturnType()->isIntegerTy()) - return 0; - - StringRef S1, S2; - bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); - bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); - - // strcspn("", s) -> 0 - if (HasS1 && S1.empty()) - return Constant::getNullValue(CI->getType()); - - // Constant folding. - if (HasS1 && HasS2) { - size_t Pos = S1.find_first_of(S2); - if (Pos == StringRef::npos) Pos = S1.size(); - return ConstantInt::get(CI->getType(), Pos); - } - - // strcspn(s, "") -> strlen(s) - if (TD && HasS2 && S2.empty()) - return EmitStrLen(CI->getArgOperand(0), B, TD, TLI); - - return 0; - } -}; - -//===---------------------------------------===// -// 'strstr' Optimizations +// Double -> Float Shrinking Optimizations for Unary Functions like 'floor' -struct StrStrOpt : public LibCallOptimization { +struct UnaryDoubleFPOpt : public LibCallOptimization { + bool CheckRetType; + UnaryDoubleFPOpt(bool CheckReturnType): CheckRetType(CheckReturnType) {} virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 2 || - !FT->getParamType(0)->isPointerTy() || - !FT->getParamType(1)->isPointerTy() || - !FT->getReturnType()->isPointerTy()) + if (FT->getNumParams() != 1 || !FT->getReturnType()->isDoubleTy() || + !FT->getParamType(0)->isDoubleTy()) return 0; - // fold strstr(x, x) -> x. - if (CI->getArgOperand(0) == CI->getArgOperand(1)) - return B.CreateBitCast(CI->getArgOperand(0), CI->getType()); - - // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0 - if (TD && IsOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) { - Value *StrLen = EmitStrLen(CI->getArgOperand(1), B, TD, TLI); - if (!StrLen) - return 0; - Value *StrNCmp = EmitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1), - StrLen, B, TD, TLI); - if (!StrNCmp) - return 0; - for (Value::use_iterator UI = CI->use_begin(), UE = CI->use_end(); - UI != UE; ) { - ICmpInst *Old = cast<ICmpInst>(*UI++); - Value *Cmp = B.CreateICmp(Old->getPredicate(), StrNCmp, - ConstantInt::getNullValue(StrNCmp->getType()), - "cmp"); - Old->replaceAllUsesWith(Cmp); - Old->eraseFromParent(); + if (CheckRetType) { + // Check if all the uses for function like 'sin' are converted to float. + for (Value::use_iterator UseI = CI->use_begin(); UseI != CI->use_end(); + ++UseI) { + FPTruncInst *Cast = dyn_cast<FPTruncInst>(*UseI); + if (Cast == 0 || !Cast->getType()->isFloatTy()) + return 0; } - return CI; - } - - // See if either input string is a constant string. - StringRef SearchStr, ToFindStr; - bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr); - bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr); - - // fold strstr(x, "") -> x. - if (HasStr2 && ToFindStr.empty()) - return B.CreateBitCast(CI->getArgOperand(0), CI->getType()); - - // If both strings are known, constant fold it. - if (HasStr1 && HasStr2) { - std::string::size_type Offset = SearchStr.find(ToFindStr); - - if (Offset == StringRef::npos) // strstr("foo", "bar") -> null - return Constant::getNullValue(CI->getType()); - - // strstr("abcd", "bc") -> gep((char*)"abcd", 1) - Value *Result = CastToCStr(CI->getArgOperand(0), B); - Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr"); - return B.CreateBitCast(Result, CI->getType()); - } - - // fold strstr(x, "y") -> strchr(x, 'y'). - if (HasStr2 && ToFindStr.size() == 1) { - Value *StrChr= EmitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TD, TLI); - return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : 0; } - return 0; - } -}; - - -//===---------------------------------------===// -// 'memcmp' Optimizations - -struct MemCmpOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || !FT->getParamType(0)->isPointerTy() || - !FT->getParamType(1)->isPointerTy() || - !FT->getReturnType()->isIntegerTy(32)) - return 0; - Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1); - - if (LHS == RHS) // memcmp(s,s,x) -> 0 - return Constant::getNullValue(CI->getType()); - - // Make sure we have a constant length. - ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2)); - if (!LenC) return 0; - uint64_t Len = LenC->getZExtValue(); - - if (Len == 0) // memcmp(s1,s2,0) -> 0 - return Constant::getNullValue(CI->getType()); - - // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS - if (Len == 1) { - Value *LHSV = B.CreateZExt(B.CreateLoad(CastToCStr(LHS, B), "lhsc"), - CI->getType(), "lhsv"); - Value *RHSV = B.CreateZExt(B.CreateLoad(CastToCStr(RHS, B), "rhsc"), - CI->getType(), "rhsv"); - return B.CreateSub(LHSV, RHSV, "chardiff"); - } - - // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant) - StringRef LHSStr, RHSStr; - if (getConstantStringInfo(LHS, LHSStr) && - getConstantStringInfo(RHS, RHSStr)) { - // Make sure we're not reading out-of-bounds memory. - if (Len > LHSStr.size() || Len > RHSStr.size()) - return 0; - uint64_t Ret = memcmp(LHSStr.data(), RHSStr.data(), Len); - return ConstantInt::get(CI->getType(), Ret); - } - - return 0; - } -}; - -//===---------------------------------------===// -// 'memcpy' Optimizations - -struct MemCpyOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // These optimizations require TargetData. - if (!TD) return 0; - - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || - !FT->getParamType(0)->isPointerTy() || - !FT->getParamType(1)->isPointerTy() || - FT->getParamType(2) != TD->getIntPtrType(*Context)) - return 0; - - // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1) - B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1), - CI->getArgOperand(2), 1); - return CI->getArgOperand(0); - } -}; - -//===---------------------------------------===// -// 'memmove' Optimizations - -struct MemMoveOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // These optimizations require TargetData. - if (!TD) return 0; - - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || - !FT->getParamType(0)->isPointerTy() || - !FT->getParamType(1)->isPointerTy() || - FT->getParamType(2) != TD->getIntPtrType(*Context)) - return 0; - - // memmove(x, y, n) -> llvm.memmove(x, y, n, 1) - B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1), - CI->getArgOperand(2), 1); - return CI->getArgOperand(0); - } -}; - -//===---------------------------------------===// -// 'memset' Optimizations - -struct MemSetOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // These optimizations require TargetData. - if (!TD) return 0; - - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || - !FT->getParamType(0)->isPointerTy() || - !FT->getParamType(1)->isIntegerTy() || - FT->getParamType(2) != TD->getIntPtrType(*Context)) + // If this is something like 'floor((double)floatval)', convert to floorf. + FPExtInst *Cast = dyn_cast<FPExtInst>(CI->getArgOperand(0)); + if (Cast == 0 || !Cast->getOperand(0)->getType()->isFloatTy()) return 0; - // memset(p, v, n) -> llvm.memset(p, v, n, 1) - Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false); - B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1); - return CI->getArgOperand(0); + // floor((double)floatval) -> (double)floorf(floatval) + Value *V = Cast->getOperand(0); + V = EmitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes()); + return B.CreateFPExt(V, B.getDoubleTy()); } }; -//===----------------------------------------------------------------------===// -// Math Library Optimizations -//===----------------------------------------------------------------------===// - //===---------------------------------------===// // 'cos*' Optimizations - struct CosOpt : public LibCallOptimization { virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + Value *Ret = NULL; + if (UnsafeFPShrink && Callee->getName() == "cos" && + TLI->has(LibFunc::cosf)) { + UnaryDoubleFPOpt UnsafeUnaryDoubleFP(true); + Ret = UnsafeUnaryDoubleFP.CallOptimizer(Callee, CI, B); + } + FunctionType *FT = Callee->getFunctionType(); // Just make sure this has 1 argument of FP type, which matches the // result type. if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || !FT->getParamType(0)->isFloatingPointTy()) - return 0; + return Ret; // cos(-x) -> cos(x) Value *Op1 = CI->getArgOperand(0); @@ -910,7 +162,7 @@ struct CosOpt : public LibCallOptimization { BinaryOperator *BinExpr = cast<BinaryOperator>(Op1); return B.CreateCall(Callee, BinExpr->getOperand(1), "cos"); } - return 0; + return Ret; } }; @@ -919,13 +171,20 @@ struct CosOpt : public LibCallOptimization { struct PowOpt : public LibCallOptimization { virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + Value *Ret = NULL; + if (UnsafeFPShrink && Callee->getName() == "pow" && + TLI->has(LibFunc::powf)) { + UnaryDoubleFPOpt UnsafeUnaryDoubleFP(true); + Ret = UnsafeUnaryDoubleFP.CallOptimizer(Callee, CI, B); + } + FunctionType *FT = Callee->getFunctionType(); // Just make sure this has 2 arguments of the same FP type, which match the // result type. if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) || FT->getParamType(0) != FT->getParamType(1) || !FT->getParamType(0)->isFloatingPointTy()) - return 0; + return Ret; Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1); if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) { @@ -936,7 +195,7 @@ struct PowOpt : public LibCallOptimization { } ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2); - if (Op2C == 0) return 0; + if (Op2C == 0) return Ret; if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0 return ConstantFP::get(CI->getType(), 1.0); @@ -974,12 +233,19 @@ struct PowOpt : public LibCallOptimization { struct Exp2Opt : public LibCallOptimization { virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + Value *Ret = NULL; + if (UnsafeFPShrink && Callee->getName() == "exp2" && + TLI->has(LibFunc::exp2)) { + UnaryDoubleFPOpt UnsafeUnaryDoubleFP(true); + Ret = UnsafeUnaryDoubleFP.CallOptimizer(Callee, CI, B); + } + FunctionType *FT = Callee->getFunctionType(); // Just make sure this has 1 argument of FP type, which matches the // result type. if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || !FT->getParamType(0)->isFloatingPointTy()) - return 0; + return Ret; Value *Op = CI->getArgOperand(0); // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32 @@ -1016,29 +282,7 @@ struct Exp2Opt : public LibCallOptimization { return CI; } - return 0; - } -}; - -//===---------------------------------------===// -// Double -> Float Shrinking Optimizations for Unary Functions like 'floor' - -struct UnaryDoubleFPOpt : public LibCallOptimization { - virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - FunctionType *FT = Callee->getFunctionType(); - if (FT->getNumParams() != 1 || !FT->getReturnType()->isDoubleTy() || - !FT->getParamType(0)->isDoubleTy()) - return 0; - - // If this is something like 'floor((double)floatval)', convert to floorf. - FPExtInst *Cast = dyn_cast<FPExtInst>(CI->getArgOperand(0)); - if (Cast == 0 || !Cast->getOperand(0)->getType()->isFloatTy()) - return 0; - - // floor((double)floatval) -> (double)floorf(floatval) - Value *V = Cast->getOperand(0); - V = EmitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes()); - return B.CreateFPExt(V, B.getDoubleTy()); + return Ret; } }; @@ -1063,8 +307,8 @@ struct FFSOpt : public LibCallOptimization { // Constant fold. if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { - if (CI->getValue() == 0) // ffs(0) -> 0. - return Constant::getNullValue(CI->getType()); + if (CI->isZero()) // ffs(0) -> 0. + return B.getInt32(0); // ffs(c) -> cttz(c)+1 return B.getInt32(CI->getValue().countTrailingZeros() + 1); } @@ -1267,7 +511,7 @@ struct SPrintFOpt : public LibCallOptimization { if (FormatStr[i] == '%') return 0; // we found a format specifier, bail out. - // These optimizations require TargetData. + // These optimizations require DataLayout. if (!TD) return 0; // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1) @@ -1297,7 +541,7 @@ struct SPrintFOpt : public LibCallOptimization { } if (FormatStr[1] == 's') { - // These optimizations require TargetData. + // These optimizations require DataLayout. if (!TD) return 0; // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1) @@ -1385,7 +629,7 @@ struct FWriteOpt : public LibCallOptimization { struct FPutsOpt : public LibCallOptimization { virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // These optimizations require TargetData. + // These optimizations require DataLayout. if (!TD) return 0; // Require two pointers. Also, we can't optimize if return value is used. @@ -1422,7 +666,7 @@ struct FPrintFOpt : public LibCallOptimization { if (FormatStr[i] == '%') // Could handle %% -> % if we cared. return 0; // We found a format specifier. - // These optimizations require TargetData. + // These optimizations require DataLayout. if (!TD) return 0; Value *NewCI = EmitFWrite(CI->getArgOperand(1), @@ -1524,17 +768,9 @@ namespace { TargetLibraryInfo *TLI; StringMap<LibCallOptimization*> Optimizations; - // String and Memory LibCall Optimizations - StrCatOpt StrCat; StrNCatOpt StrNCat; StrChrOpt StrChr; StrRChrOpt StrRChr; - StrCmpOpt StrCmp; StrNCmpOpt StrNCmp; - StrCpyOpt StrCpy; StrCpyOpt StrCpyChk; - StpCpyOpt StpCpy; StpCpyOpt StpCpyChk; - StrNCpyOpt StrNCpy; - StrLenOpt StrLen; StrPBrkOpt StrPBrk; - StrToOpt StrTo; StrSpnOpt StrSpn; StrCSpnOpt StrCSpn; StrStrOpt StrStr; - MemCmpOpt MemCmp; MemCpyOpt MemCpy; MemMoveOpt MemMove; MemSetOpt MemSet; // Math Library Optimizations - CosOpt Cos; PowOpt Pow; Exp2Opt Exp2; UnaryDoubleFPOpt UnaryDoubleFP; + CosOpt Cos; PowOpt Pow; Exp2Opt Exp2; + UnaryDoubleFPOpt UnaryDoubleFP, UnsafeUnaryDoubleFP; // Integer Optimizations FFSOpt FFS; AbsOpt Abs; IsDigitOpt IsDigit; IsAsciiOpt IsAscii; ToAsciiOpt ToAscii; @@ -1546,11 +782,13 @@ namespace { bool Modified; // This is only used by doInitialization. public: static char ID; // Pass identification - SimplifyLibCalls() : FunctionPass(ID), StrCpy(false), StrCpyChk(true), - StpCpy(false), StpCpyChk(true) { + SimplifyLibCalls() : FunctionPass(ID), UnaryDoubleFP(false), + UnsafeUnaryDoubleFP(true) { initializeSimplifyLibCallsPass(*PassRegistry::getPassRegistry()); } void AddOpt(LibFunc::Func F, LibCallOptimization* Opt); + void AddOpt(LibFunc::Func F1, LibFunc::Func F2, LibCallOptimization* Opt); + void InitOptimizations(); bool runOnFunction(Function &F); @@ -1586,40 +824,15 @@ void SimplifyLibCalls::AddOpt(LibFunc::Func F, LibCallOptimization* Opt) { Optimizations[TLI->getName(F)] = Opt; } +void SimplifyLibCalls::AddOpt(LibFunc::Func F1, LibFunc::Func F2, + LibCallOptimization* Opt) { + if (TLI->has(F1) && TLI->has(F2)) + Optimizations[TLI->getName(F1)] = Opt; +} + /// Optimizations - Populate the Optimizations map with all the optimizations /// we know. void SimplifyLibCalls::InitOptimizations() { - // String and Memory LibCall Optimizations - Optimizations["strcat"] = &StrCat; - Optimizations["strncat"] = &StrNCat; - Optimizations["strchr"] = &StrChr; - Optimizations["strrchr"] = &StrRChr; - Optimizations["strcmp"] = &StrCmp; - Optimizations["strncmp"] = &StrNCmp; - Optimizations["strcpy"] = &StrCpy; - Optimizations["strncpy"] = &StrNCpy; - Optimizations["stpcpy"] = &StpCpy; - Optimizations["strlen"] = &StrLen; - Optimizations["strpbrk"] = &StrPBrk; - Optimizations["strtol"] = &StrTo; - Optimizations["strtod"] = &StrTo; - Optimizations["strtof"] = &StrTo; - Optimizations["strtoul"] = &StrTo; - Optimizations["strtoll"] = &StrTo; - Optimizations["strtold"] = &StrTo; - Optimizations["strtoull"] = &StrTo; - Optimizations["strspn"] = &StrSpn; - Optimizations["strcspn"] = &StrCSpn; - Optimizations["strstr"] = &StrStr; - Optimizations["memcmp"] = &MemCmp; - AddOpt(LibFunc::memcpy, &MemCpy); - Optimizations["memmove"] = &MemMove; - AddOpt(LibFunc::memset, &MemSet); - - // _chk variants of String and Memory LibCall Optimizations. - Optimizations["__strcpy_chk"] = &StrCpyChk; - Optimizations["__stpcpy_chk"] = &StpCpyChk; - // Math Library Optimizations Optimizations["cosf"] = &Cos; Optimizations["cos"] = &Cos; @@ -1641,16 +854,37 @@ void SimplifyLibCalls::InitOptimizations() { Optimizations["llvm.exp2.f64"] = &Exp2; Optimizations["llvm.exp2.f32"] = &Exp2; - if (TLI->has(LibFunc::floor) && TLI->has(LibFunc::floorf)) - Optimizations["floor"] = &UnaryDoubleFP; - if (TLI->has(LibFunc::ceil) && TLI->has(LibFunc::ceilf)) - Optimizations["ceil"] = &UnaryDoubleFP; - if (TLI->has(LibFunc::round) && TLI->has(LibFunc::roundf)) - Optimizations["round"] = &UnaryDoubleFP; - if (TLI->has(LibFunc::rint) && TLI->has(LibFunc::rintf)) - Optimizations["rint"] = &UnaryDoubleFP; - if (TLI->has(LibFunc::nearbyint) && TLI->has(LibFunc::nearbyintf)) - Optimizations["nearbyint"] = &UnaryDoubleFP; + AddOpt(LibFunc::ceil, LibFunc::ceilf, &UnaryDoubleFP); + AddOpt(LibFunc::fabs, LibFunc::fabsf, &UnaryDoubleFP); + AddOpt(LibFunc::floor, LibFunc::floorf, &UnaryDoubleFP); + AddOpt(LibFunc::rint, LibFunc::rintf, &UnaryDoubleFP); + AddOpt(LibFunc::round, LibFunc::roundf, &UnaryDoubleFP); + AddOpt(LibFunc::nearbyint, LibFunc::nearbyintf, &UnaryDoubleFP); + AddOpt(LibFunc::trunc, LibFunc::truncf, &UnaryDoubleFP); + + if(UnsafeFPShrink) { + AddOpt(LibFunc::acos, LibFunc::acosf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::acosh, LibFunc::acoshf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::asin, LibFunc::asinf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::asinh, LibFunc::asinhf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::atan, LibFunc::atanf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::atanh, LibFunc::atanhf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::cbrt, LibFunc::cbrtf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::cosh, LibFunc::coshf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::exp, LibFunc::expf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::exp10, LibFunc::exp10f, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::expm1, LibFunc::expm1f, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::log, LibFunc::logf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::log10, LibFunc::log10f, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::log1p, LibFunc::log1pf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::log2, LibFunc::log2f, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::logb, LibFunc::logbf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::sin, LibFunc::sinf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::sinh, LibFunc::sinhf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::sqrt, LibFunc::sqrtf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::tan, LibFunc::tanf, &UnsafeUnaryDoubleFP); + AddOpt(LibFunc::tanh, LibFunc::tanhf, &UnsafeUnaryDoubleFP); + } // Integer Optimizations Optimizations["ffs"] = &FFS; @@ -1681,7 +915,7 @@ bool SimplifyLibCalls::runOnFunction(Function &F) { if (Optimizations.empty()) InitOptimizations(); - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); IRBuilder<> Builder(F.getContext()); diff --git a/contrib/llvm/lib/Transforms/Utils/AddrModeMatcher.cpp b/contrib/llvm/lib/Transforms/Utils/AddrModeMatcher.cpp index d831452..6815e41 100644 --- a/contrib/llvm/lib/Transforms/Utils/AddrModeMatcher.cpp +++ b/contrib/llvm/lib/Transforms/Utils/AddrModeMatcher.cpp @@ -16,7 +16,7 @@ #include "llvm/GlobalValue.h" #include "llvm/Instruction.h" #include "llvm/Assembly/Writer.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Support/Debug.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/PatternMatch.h" @@ -55,10 +55,12 @@ void ExtAddrMode::print(raw_ostream &OS) const { OS << ']'; } +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void ExtAddrMode::dump() const { print(dbgs()); dbgs() << '\n'; } +#endif /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode. @@ -219,7 +221,7 @@ bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode, unsigned VariableScale = 0; int64_t ConstantOffset = 0; - const TargetData *TD = TLI.getTargetData(); + const DataLayout *TD = TLI.getDataLayout(); gep_type_iterator GTI = gep_type_begin(AddrInst); for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { if (StructType *STy = dyn_cast<StructType>(*GTI)) { diff --git a/contrib/llvm/lib/Transforms/Utils/BasicBlockUtils.cpp b/contrib/llvm/lib/Transforms/Utils/BasicBlockUtils.cpp index 2679b93..9fea113 100644 --- a/contrib/llvm/lib/Transforms/Utils/BasicBlockUtils.cpp +++ b/contrib/llvm/lib/Transforms/Utils/BasicBlockUtils.cpp @@ -22,7 +22,7 @@ #include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/MemoryDependenceAnalysis.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Scalar.h" #include "llvm/Support/ErrorHandling.h" @@ -94,7 +94,7 @@ void llvm::FoldSingleEntryPHINodes(BasicBlock *BB, Pass *P) { /// is dead. Also recursively delete any operands that become dead as /// a result. This includes tracing the def-use list from the PHI to see if /// it is ultimately unused or if it reaches an unused cycle. -bool llvm::DeleteDeadPHIs(BasicBlock *BB) { +bool llvm::DeleteDeadPHIs(BasicBlock *BB, const TargetLibraryInfo *TLI) { // Recursively deleting a PHI may cause multiple PHIs to be deleted // or RAUW'd undef, so use an array of WeakVH for the PHIs to delete. SmallVector<WeakVH, 8> PHIs; @@ -105,7 +105,7 @@ bool llvm::DeleteDeadPHIs(BasicBlock *BB) { bool Changed = false; for (unsigned i = 0, e = PHIs.size(); i != e; ++i) if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i].operator Value*())) - Changed |= RecursivelyDeleteDeadPHINode(PN); + Changed |= RecursivelyDeleteDeadPHINode(PN, TLI); return Changed; } @@ -687,3 +687,42 @@ ReturnInst *llvm::FoldReturnIntoUncondBranch(ReturnInst *RI, BasicBlock *BB, return cast<ReturnInst>(NewRet); } +/// SplitBlockAndInsertIfThen - Split the containing block at the +/// specified instruction - everything before and including Cmp stays +/// in the old basic block, and everything after Cmp is moved to a +/// new block. The two blocks are connected by a conditional branch +/// (with value of Cmp being the condition). +/// Before: +/// Head +/// Cmp +/// Tail +/// After: +/// Head +/// Cmp +/// if (Cmp) +/// ThenBlock +/// Tail +/// +/// If Unreachable is true, then ThenBlock ends with +/// UnreachableInst, otherwise it branches to Tail. +/// Returns the NewBasicBlock's terminator. + +TerminatorInst *llvm::SplitBlockAndInsertIfThen(Instruction *Cmp, + bool Unreachable, MDNode *BranchWeights) { + Instruction *SplitBefore = Cmp->getNextNode(); + BasicBlock *Head = SplitBefore->getParent(); + BasicBlock *Tail = Head->splitBasicBlock(SplitBefore); + TerminatorInst *HeadOldTerm = Head->getTerminator(); + LLVMContext &C = Head->getContext(); + BasicBlock *ThenBlock = BasicBlock::Create(C, "", Head->getParent(), Tail); + TerminatorInst *CheckTerm; + if (Unreachable) + CheckTerm = new UnreachableInst(C, ThenBlock); + else + CheckTerm = BranchInst::Create(Tail, ThenBlock); + BranchInst *HeadNewTerm = + BranchInst::Create(/*ifTrue*/ThenBlock, /*ifFalse*/Tail, Cmp); + HeadNewTerm->setMetadata(LLVMContext::MD_prof, BranchWeights); + ReplaceInstWithInst(HeadOldTerm, HeadNewTerm); + return CheckTerm; +} diff --git a/contrib/llvm/lib/Transforms/Utils/BuildLibCalls.cpp b/contrib/llvm/lib/Transforms/Utils/BuildLibCalls.cpp index e13fd71..74b2ee1 100644 --- a/contrib/llvm/lib/Transforms/Utils/BuildLibCalls.cpp +++ b/contrib/llvm/lib/Transforms/Utils/BuildLibCalls.cpp @@ -22,7 +22,7 @@ #include "llvm/Module.h" #include "llvm/Type.h" #include "llvm/ADT/SmallString.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" using namespace llvm; @@ -34,19 +34,22 @@ Value *llvm::CastToCStr(Value *V, IRBuilder<> &B) { /// EmitStrLen - Emit a call to the strlen function to the builder, for the /// specified pointer. This always returns an integer value of size intptr_t. -Value *llvm::EmitStrLen(Value *Ptr, IRBuilder<> &B, const TargetData *TD, +Value *llvm::EmitStrLen(Value *Ptr, IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::strlen)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[2]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(~0u, Attribute::ReadOnly | - Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + Attributes::AttrVal AVs[2] = { Attributes::ReadOnly, Attributes::NoUnwind }; + AWI[1] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + ArrayRef<Attributes::AttrVal>(AVs, 2)); LLVMContext &Context = B.GetInsertBlock()->getContext(); - Constant *StrLen = M->getOrInsertFunction("strlen", AttrListPtr::get(AWI), + Constant *StrLen = M->getOrInsertFunction("strlen", + AttrListPtr::get(M->getContext(), + AWI), TD->getIntPtrType(Context), B.getInt8PtrTy(), NULL); @@ -61,18 +64,21 @@ Value *llvm::EmitStrLen(Value *Ptr, IRBuilder<> &B, const TargetData *TD, /// specified pointer. Ptr is required to be some pointer type, MaxLen must /// be of size_t type, and the return value has 'intptr_t' type. Value *llvm::EmitStrNLen(Value *Ptr, Value *MaxLen, IRBuilder<> &B, - const TargetData *TD, const TargetLibraryInfo *TLI) { + const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::strnlen)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[2]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(~0u, Attribute::ReadOnly | - Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + Attributes::AttrVal AVs[2] = { Attributes::ReadOnly, Attributes::NoUnwind }; + AWI[1] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + ArrayRef<Attributes::AttrVal>(AVs, 2)); LLVMContext &Context = B.GetInsertBlock()->getContext(); - Constant *StrNLen = M->getOrInsertFunction("strnlen", AttrListPtr::get(AWI), + Constant *StrNLen = M->getOrInsertFunction("strnlen", + AttrListPtr::get(M->getContext(), + AWI), TD->getIntPtrType(Context), B.getInt8PtrTy(), TD->getIntPtrType(Context), @@ -88,17 +94,21 @@ Value *llvm::EmitStrNLen(Value *Ptr, Value *MaxLen, IRBuilder<> &B, /// specified pointer and character. Ptr is required to be some pointer type, /// and the return value has 'i8*' type. Value *llvm::EmitStrChr(Value *Ptr, char C, IRBuilder<> &B, - const TargetData *TD, const TargetLibraryInfo *TLI) { + const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::strchr)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); + Attributes::AttrVal AVs[2] = { Attributes::ReadOnly, Attributes::NoUnwind }; AttributeWithIndex AWI = - AttributeWithIndex::get(~0u, Attribute::ReadOnly | Attribute::NoUnwind); + AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + ArrayRef<Attributes::AttrVal>(AVs, 2)); Type *I8Ptr = B.getInt8PtrTy(); Type *I32Ty = B.getInt32Ty(); - Constant *StrChr = M->getOrInsertFunction("strchr", AttrListPtr::get(AWI), + Constant *StrChr = M->getOrInsertFunction("strchr", + AttrListPtr::get(M->getContext(), + AWI), I8Ptr, I8Ptr, I32Ty, NULL); CallInst *CI = B.CreateCall2(StrChr, CastToCStr(Ptr, B), ConstantInt::get(I32Ty, C), "strchr"); @@ -109,20 +119,23 @@ Value *llvm::EmitStrChr(Value *Ptr, char C, IRBuilder<> &B, /// EmitStrNCmp - Emit a call to the strncmp function to the builder. Value *llvm::EmitStrNCmp(Value *Ptr1, Value *Ptr2, Value *Len, - IRBuilder<> &B, const TargetData *TD, + IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::strncmp)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[3]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(2, Attribute::NoCapture); - AWI[2] = AttributeWithIndex::get(~0u, Attribute::ReadOnly | - Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), 2, Attributes::NoCapture); + Attributes::AttrVal AVs[2] = { Attributes::ReadOnly, Attributes::NoUnwind }; + AWI[2] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + ArrayRef<Attributes::AttrVal>(AVs, 2)); LLVMContext &Context = B.GetInsertBlock()->getContext(); - Value *StrNCmp = M->getOrInsertFunction("strncmp", AttrListPtr::get(AWI), + Value *StrNCmp = M->getOrInsertFunction("strncmp", + AttrListPtr::get(M->getContext(), + AWI), B.getInt32Ty(), B.getInt8PtrTy(), B.getInt8PtrTy(), @@ -139,17 +152,19 @@ Value *llvm::EmitStrNCmp(Value *Ptr1, Value *Ptr2, Value *Len, /// EmitStrCpy - Emit a call to the strcpy function to the builder, for the /// specified pointer arguments. Value *llvm::EmitStrCpy(Value *Dst, Value *Src, IRBuilder<> &B, - const TargetData *TD, const TargetLibraryInfo *TLI, + const DataLayout *TD, const TargetLibraryInfo *TLI, StringRef Name) { if (!TLI->has(LibFunc::strcpy)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[2]; - AWI[0] = AttributeWithIndex::get(2, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 2, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); Type *I8Ptr = B.getInt8PtrTy(); - Value *StrCpy = M->getOrInsertFunction(Name, AttrListPtr::get(AWI), + Value *StrCpy = M->getOrInsertFunction(Name, + AttrListPtr::get(M->getContext(), AWI), I8Ptr, I8Ptr, I8Ptr, NULL); CallInst *CI = B.CreateCall2(StrCpy, CastToCStr(Dst, B), CastToCStr(Src, B), Name); @@ -161,17 +176,20 @@ Value *llvm::EmitStrCpy(Value *Dst, Value *Src, IRBuilder<> &B, /// EmitStrNCpy - Emit a call to the strncpy function to the builder, for the /// specified pointer arguments. Value *llvm::EmitStrNCpy(Value *Dst, Value *Src, Value *Len, - IRBuilder<> &B, const TargetData *TD, + IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI, StringRef Name) { if (!TLI->has(LibFunc::strncpy)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[2]; - AWI[0] = AttributeWithIndex::get(2, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 2, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); Type *I8Ptr = B.getInt8PtrTy(); - Value *StrNCpy = M->getOrInsertFunction(Name, AttrListPtr::get(AWI), + Value *StrNCpy = M->getOrInsertFunction(Name, + AttrListPtr::get(M->getContext(), + AWI), I8Ptr, I8Ptr, I8Ptr, Len->getType(), NULL); CallInst *CI = B.CreateCall3(StrNCpy, CastToCStr(Dst, B), CastToCStr(Src, B), @@ -185,17 +203,18 @@ Value *llvm::EmitStrNCpy(Value *Dst, Value *Src, Value *Len, /// This expects that the Len and ObjSize have type 'intptr_t' and Dst/Src /// are pointers. Value *llvm::EmitMemCpyChk(Value *Dst, Value *Src, Value *Len, Value *ObjSize, - IRBuilder<> &B, const TargetData *TD, + IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::memcpy_chk)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI; - AWI = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); LLVMContext &Context = B.GetInsertBlock()->getContext(); Value *MemCpy = M->getOrInsertFunction("__memcpy_chk", - AttrListPtr::get(AWI), + AttrListPtr::get(M->getContext(), AWI), B.getInt8PtrTy(), B.getInt8PtrTy(), B.getInt8PtrTy(), @@ -212,16 +231,19 @@ Value *llvm::EmitMemCpyChk(Value *Dst, Value *Src, Value *Len, Value *ObjSize, /// EmitMemChr - Emit a call to the memchr function. This assumes that Ptr is /// a pointer, Val is an i32 value, and Len is an 'intptr_t' value. Value *llvm::EmitMemChr(Value *Ptr, Value *Val, - Value *Len, IRBuilder<> &B, const TargetData *TD, + Value *Len, IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::memchr)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI; - AWI = AttributeWithIndex::get(~0u, Attribute::ReadOnly | Attribute::NoUnwind); + Attributes::AttrVal AVs[2] = { Attributes::ReadOnly, Attributes::NoUnwind }; + AWI = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + ArrayRef<Attributes::AttrVal>(AVs, 2)); LLVMContext &Context = B.GetInsertBlock()->getContext(); - Value *MemChr = M->getOrInsertFunction("memchr", AttrListPtr::get(AWI), + Value *MemChr = M->getOrInsertFunction("memchr", + AttrListPtr::get(M->getContext(), AWI), B.getInt8PtrTy(), B.getInt8PtrTy(), B.getInt32Ty(), @@ -237,20 +259,22 @@ Value *llvm::EmitMemChr(Value *Ptr, Value *Val, /// EmitMemCmp - Emit a call to the memcmp function. Value *llvm::EmitMemCmp(Value *Ptr1, Value *Ptr2, - Value *Len, IRBuilder<> &B, const TargetData *TD, + Value *Len, IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::memcmp)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[3]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(2, Attribute::NoCapture); - AWI[2] = AttributeWithIndex::get(~0u, Attribute::ReadOnly | - Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), 2, Attributes::NoCapture); + Attributes::AttrVal AVs[2] = { Attributes::ReadOnly, Attributes::NoUnwind }; + AWI[2] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + ArrayRef<Attributes::AttrVal>(AVs, 2)); LLVMContext &Context = B.GetInsertBlock()->getContext(); - Value *MemCmp = M->getOrInsertFunction("memcmp", AttrListPtr::get(AWI), + Value *MemCmp = M->getOrInsertFunction("memcmp", + AttrListPtr::get(M->getContext(), AWI), B.getInt32Ty(), B.getInt8PtrTy(), B.getInt8PtrTy(), @@ -294,7 +318,7 @@ Value *llvm::EmitUnaryFloatFnCall(Value *Op, StringRef Name, IRBuilder<> &B, /// EmitPutChar - Emit a call to the putchar function. This assumes that Char /// is an integer. -Value *llvm::EmitPutChar(Value *Char, IRBuilder<> &B, const TargetData *TD, +Value *llvm::EmitPutChar(Value *Char, IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::putchar)) return 0; @@ -316,17 +340,19 @@ Value *llvm::EmitPutChar(Value *Char, IRBuilder<> &B, const TargetData *TD, /// EmitPutS - Emit a call to the puts function. This assumes that Str is /// some pointer. -Value *llvm::EmitPutS(Value *Str, IRBuilder<> &B, const TargetData *TD, +Value *llvm::EmitPutS(Value *Str, IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::puts)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[2]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); - Value *PutS = M->getOrInsertFunction("puts", AttrListPtr::get(AWI), + Value *PutS = M->getOrInsertFunction("puts", + AttrListPtr::get(M->getContext(), AWI), B.getInt32Ty(), B.getInt8PtrTy(), NULL); @@ -339,17 +365,19 @@ Value *llvm::EmitPutS(Value *Str, IRBuilder<> &B, const TargetData *TD, /// EmitFPutC - Emit a call to the fputc function. This assumes that Char is /// an integer and File is a pointer to FILE. Value *llvm::EmitFPutC(Value *Char, Value *File, IRBuilder<> &B, - const TargetData *TD, const TargetLibraryInfo *TLI) { + const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::fputc)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[2]; - AWI[0] = AttributeWithIndex::get(2, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 2, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); Constant *F; if (File->getType()->isPointerTy()) - F = M->getOrInsertFunction("fputc", AttrListPtr::get(AWI), + F = M->getOrInsertFunction("fputc", + AttrListPtr::get(M->getContext(), AWI), B.getInt32Ty(), B.getInt32Ty(), File->getType(), NULL); @@ -370,19 +398,21 @@ Value *llvm::EmitFPutC(Value *Char, Value *File, IRBuilder<> &B, /// EmitFPutS - Emit a call to the puts function. Str is required to be a /// pointer and File is a pointer to FILE. Value *llvm::EmitFPutS(Value *Str, Value *File, IRBuilder<> &B, - const TargetData *TD, const TargetLibraryInfo *TLI) { + const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::fputs)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[3]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(2, Attribute::NoCapture); - AWI[2] = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), 2, Attributes::NoCapture); + AWI[2] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); StringRef FPutsName = TLI->getName(LibFunc::fputs); Constant *F; if (File->getType()->isPointerTy()) - F = M->getOrInsertFunction(FPutsName, AttrListPtr::get(AWI), + F = M->getOrInsertFunction(FPutsName, + AttrListPtr::get(M->getContext(), AWI), B.getInt32Ty(), B.getInt8PtrTy(), File->getType(), NULL); @@ -400,21 +430,23 @@ Value *llvm::EmitFPutS(Value *Str, Value *File, IRBuilder<> &B, /// EmitFWrite - Emit a call to the fwrite function. This assumes that Ptr is /// a pointer, Size is an 'intptr_t', and File is a pointer to FILE. Value *llvm::EmitFWrite(Value *Ptr, Value *Size, Value *File, - IRBuilder<> &B, const TargetData *TD, + IRBuilder<> &B, const DataLayout *TD, const TargetLibraryInfo *TLI) { if (!TLI->has(LibFunc::fwrite)) return 0; Module *M = B.GetInsertBlock()->getParent()->getParent(); AttributeWithIndex AWI[3]; - AWI[0] = AttributeWithIndex::get(1, Attribute::NoCapture); - AWI[1] = AttributeWithIndex::get(4, Attribute::NoCapture); - AWI[2] = AttributeWithIndex::get(~0u, Attribute::NoUnwind); + AWI[0] = AttributeWithIndex::get(M->getContext(), 1, Attributes::NoCapture); + AWI[1] = AttributeWithIndex::get(M->getContext(), 4, Attributes::NoCapture); + AWI[2] = AttributeWithIndex::get(M->getContext(), AttrListPtr::FunctionIndex, + Attributes::NoUnwind); LLVMContext &Context = B.GetInsertBlock()->getContext(); StringRef FWriteName = TLI->getName(LibFunc::fwrite); Constant *F; if (File->getType()->isPointerTy()) - F = M->getOrInsertFunction(FWriteName, AttrListPtr::get(AWI), + F = M->getOrInsertFunction(FWriteName, + AttrListPtr::get(M->getContext(), AWI), TD->getIntPtrType(Context), B.getInt8PtrTy(), TD->getIntPtrType(Context), @@ -436,9 +468,9 @@ Value *llvm::EmitFWrite(Value *Ptr, Value *Size, Value *File, SimplifyFortifiedLibCalls::~SimplifyFortifiedLibCalls() { } -bool SimplifyFortifiedLibCalls::fold(CallInst *CI, const TargetData *TD, +bool SimplifyFortifiedLibCalls::fold(CallInst *CI, const DataLayout *TD, const TargetLibraryInfo *TLI) { - // We really need TargetData for later. + // We really need DataLayout for later. if (!TD) return false; this->CI = CI; diff --git a/contrib/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp b/contrib/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp new file mode 100644 index 0000000..bee2f7b --- /dev/null +++ b/contrib/llvm/lib/Transforms/Utils/BypassSlowDivision.cpp @@ -0,0 +1,262 @@ +//===-- BypassSlowDivision.cpp - Bypass slow division ---------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file contains an optimization for div and rem on architectures that +// execute short instructions significantly faster than longer instructions. +// For example, on Intel Atom 32-bit divides are slow enough that during +// runtime it is profitable to check the value of the operands, and if they are +// positive and less than 256 use an unsigned 8-bit divide. +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "bypass-slow-division" +#include "llvm/Instructions.h" +#include "llvm/Function.h" +#include "llvm/IRBuilder.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/Transforms/Utils/BypassSlowDivision.h" + +using namespace llvm; + +namespace { + struct DivOpInfo { + bool SignedOp; + Value *Dividend; + Value *Divisor; + + DivOpInfo(bool InSignedOp, Value *InDividend, Value *InDivisor) + : SignedOp(InSignedOp), Dividend(InDividend), Divisor(InDivisor) {} + }; + + struct DivPhiNodes { + PHINode *Quotient; + PHINode *Remainder; + + DivPhiNodes(PHINode *InQuotient, PHINode *InRemainder) + : Quotient(InQuotient), Remainder(InRemainder) {} + }; +} + +namespace llvm { + template<> + struct DenseMapInfo<DivOpInfo> { + static bool isEqual(const DivOpInfo &Val1, const DivOpInfo &Val2) { + return Val1.SignedOp == Val2.SignedOp && + Val1.Dividend == Val2.Dividend && + Val1.Divisor == Val2.Divisor; + } + + static DivOpInfo getEmptyKey() { + return DivOpInfo(false, 0, 0); + } + + static DivOpInfo getTombstoneKey() { + return DivOpInfo(true, 0, 0); + } + + static unsigned getHashValue(const DivOpInfo &Val) { + return (unsigned)(reinterpret_cast<uintptr_t>(Val.Dividend) ^ + reinterpret_cast<uintptr_t>(Val.Divisor)) ^ + (unsigned)Val.SignedOp; + } + }; + + typedef DenseMap<DivOpInfo, DivPhiNodes> DivCacheTy; +} + +// insertFastDiv - Substitutes the div/rem instruction with code that checks the +// value of the operands and uses a shorter-faster div/rem instruction when +// possible and the longer-slower div/rem instruction otherwise. +static bool insertFastDiv(Function &F, + Function::iterator &I, + BasicBlock::iterator &J, + IntegerType *BypassType, + bool UseDivOp, + bool UseSignedOp, + DivCacheTy &PerBBDivCache) { + // Get instruction operands + Instruction *Instr = J; + Value *Dividend = Instr->getOperand(0); + Value *Divisor = Instr->getOperand(1); + + if (isa<ConstantInt>(Divisor) || + (isa<ConstantInt>(Dividend) && isa<ConstantInt>(Divisor))) { + // Operations with immediate values should have + // been solved and replaced during compile time. + return false; + } + + // Basic Block is split before divide + BasicBlock *MainBB = I; + BasicBlock *SuccessorBB = I->splitBasicBlock(J); + ++I; //advance iterator I to successorBB + + // Add new basic block for slow divide operation + BasicBlock *SlowBB = BasicBlock::Create(F.getContext(), "", + MainBB->getParent(), SuccessorBB); + SlowBB->moveBefore(SuccessorBB); + IRBuilder<> SlowBuilder(SlowBB, SlowBB->begin()); + Value *SlowQuotientV; + Value *SlowRemainderV; + if (UseSignedOp) { + SlowQuotientV = SlowBuilder.CreateSDiv(Dividend, Divisor); + SlowRemainderV = SlowBuilder.CreateSRem(Dividend, Divisor); + } else { + SlowQuotientV = SlowBuilder.CreateUDiv(Dividend, Divisor); + SlowRemainderV = SlowBuilder.CreateURem(Dividend, Divisor); + } + SlowBuilder.CreateBr(SuccessorBB); + + // Add new basic block for fast divide operation + BasicBlock *FastBB = BasicBlock::Create(F.getContext(), "", + MainBB->getParent(), SuccessorBB); + FastBB->moveBefore(SlowBB); + IRBuilder<> FastBuilder(FastBB, FastBB->begin()); + Value *ShortDivisorV = FastBuilder.CreateCast(Instruction::Trunc, Divisor, + BypassType); + Value *ShortDividendV = FastBuilder.CreateCast(Instruction::Trunc, Dividend, + BypassType); + + // udiv/urem because optimization only handles positive numbers + Value *ShortQuotientV = FastBuilder.CreateExactUDiv(ShortDividendV, + ShortDivisorV); + Value *ShortRemainderV = FastBuilder.CreateURem(ShortDividendV, + ShortDivisorV); + Value *FastQuotientV = FastBuilder.CreateCast(Instruction::ZExt, + ShortQuotientV, + Dividend->getType()); + Value *FastRemainderV = FastBuilder.CreateCast(Instruction::ZExt, + ShortRemainderV, + Dividend->getType()); + FastBuilder.CreateBr(SuccessorBB); + + // Phi nodes for result of div and rem + IRBuilder<> SuccessorBuilder(SuccessorBB, SuccessorBB->begin()); + PHINode *QuoPhi = SuccessorBuilder.CreatePHI(Instr->getType(), 2); + QuoPhi->addIncoming(SlowQuotientV, SlowBB); + QuoPhi->addIncoming(FastQuotientV, FastBB); + PHINode *RemPhi = SuccessorBuilder.CreatePHI(Instr->getType(), 2); + RemPhi->addIncoming(SlowRemainderV, SlowBB); + RemPhi->addIncoming(FastRemainderV, FastBB); + + // Replace Instr with appropriate phi node + if (UseDivOp) + Instr->replaceAllUsesWith(QuoPhi); + else + Instr->replaceAllUsesWith(RemPhi); + Instr->eraseFromParent(); + + // Combine operands into a single value with OR for value testing below + MainBB->getInstList().back().eraseFromParent(); + IRBuilder<> MainBuilder(MainBB, MainBB->end()); + Value *OrV = MainBuilder.CreateOr(Dividend, Divisor); + + // BitMask is inverted to check if the operands are + // larger than the bypass type + uint64_t BitMask = ~BypassType->getBitMask(); + Value *AndV = MainBuilder.CreateAnd(OrV, BitMask); + + // Compare operand values and branch + Value *ZeroV = MainBuilder.getInt32(0); + Value *CmpV = MainBuilder.CreateICmpEQ(AndV, ZeroV); + MainBuilder.CreateCondBr(CmpV, FastBB, SlowBB); + + // point iterator J at first instruction of successorBB + J = I->begin(); + + // Cache phi nodes to be used later in place of other instances + // of div or rem with the same sign, dividend, and divisor + DivOpInfo Key(UseSignedOp, Dividend, Divisor); + DivPhiNodes Value(QuoPhi, RemPhi); + PerBBDivCache.insert(std::pair<DivOpInfo, DivPhiNodes>(Key, Value)); + return true; +} + +// reuseOrInsertFastDiv - Reuses previously computed dividend or remainder if +// operands and operation are identical. Otherwise call insertFastDiv to perform +// the optimization and cache the resulting dividend and remainder. +static bool reuseOrInsertFastDiv(Function &F, + Function::iterator &I, + BasicBlock::iterator &J, + IntegerType *BypassType, + bool UseDivOp, + bool UseSignedOp, + DivCacheTy &PerBBDivCache) { + // Get instruction operands + Instruction *Instr = J; + DivOpInfo Key(UseSignedOp, Instr->getOperand(0), Instr->getOperand(1)); + DivCacheTy::iterator CacheI = PerBBDivCache.find(Key); + + if (CacheI == PerBBDivCache.end()) { + // If previous instance does not exist, insert fast div + return insertFastDiv(F, I, J, BypassType, UseDivOp, UseSignedOp, + PerBBDivCache); + } + + // Replace operation value with previously generated phi node + DivPhiNodes &Value = CacheI->second; + if (UseDivOp) { + // Replace all uses of div instruction with quotient phi node + J->replaceAllUsesWith(Value.Quotient); + } else { + // Replace all uses of rem instruction with remainder phi node + J->replaceAllUsesWith(Value.Remainder); + } + + // Advance to next operation + ++J; + + // Remove redundant operation + Instr->eraseFromParent(); + return true; +} + +// bypassSlowDivision - This optimization identifies DIV instructions that can +// be profitably bypassed and carried out with a shorter, faster divide. +bool llvm::bypassSlowDivision(Function &F, + Function::iterator &I, + const DenseMap<unsigned int, unsigned int> &BypassWidths) { + DivCacheTy DivCache; + + bool MadeChange = false; + for (BasicBlock::iterator J = I->begin(); J != I->end(); J++) { + + // Get instruction details + unsigned Opcode = J->getOpcode(); + bool UseDivOp = Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; + bool UseRemOp = Opcode == Instruction::SRem || Opcode == Instruction::URem; + bool UseSignedOp = Opcode == Instruction::SDiv || + Opcode == Instruction::SRem; + + // Only optimize div or rem ops + if (!UseDivOp && !UseRemOp) + continue; + + // Skip division on vector types, only optimize integer instructions + if (!J->getType()->isIntegerTy()) + continue; + + // Get bitwidth of div/rem instruction + IntegerType *T = cast<IntegerType>(J->getType()); + int bitwidth = T->getBitWidth(); + + // Continue if bitwidth is not bypassed + DenseMap<unsigned int, unsigned int>::const_iterator BI = BypassWidths.find(bitwidth); + if (BI == BypassWidths.end()) + continue; + + // Get type for div/rem instruction with bypass bitwidth + IntegerType *BT = IntegerType::get(J->getContext(), BI->second); + + MadeChange |= reuseOrInsertFastDiv(F, I, J, BT, UseDivOp, + UseSignedOp, DivCache); + } + + return MadeChange; +} diff --git a/contrib/llvm/lib/Transforms/Utils/CloneFunction.cpp b/contrib/llvm/lib/Transforms/Utils/CloneFunction.cpp index 99237b8..7ba9f6d 100644 --- a/contrib/llvm/lib/Transforms/Utils/CloneFunction.cpp +++ b/contrib/llvm/lib/Transforms/Utils/CloneFunction.cpp @@ -98,10 +98,14 @@ void llvm::CloneFunctionInto(Function *NewFunc, const Function *OldFunc, Anew->addAttr( OldFunc->getAttributes() .getParamAttributes(I->getArgNo() + 1)); NewFunc->setAttributes(NewFunc->getAttributes() - .addAttr(0, OldFunc->getAttributes() + .addAttr(NewFunc->getContext(), + AttrListPtr::ReturnIndex, + OldFunc->getAttributes() .getRetAttributes())); NewFunc->setAttributes(NewFunc->getAttributes() - .addAttr(~0, OldFunc->getAttributes() + .addAttr(NewFunc->getContext(), + AttrListPtr::FunctionIndex, + OldFunc->getAttributes() .getFnAttributes())); } @@ -202,14 +206,14 @@ namespace { bool ModuleLevelChanges; const char *NameSuffix; ClonedCodeInfo *CodeInfo; - const TargetData *TD; + const DataLayout *TD; public: PruningFunctionCloner(Function *newFunc, const Function *oldFunc, ValueToValueMapTy &valueMap, bool moduleLevelChanges, const char *nameSuffix, ClonedCodeInfo *codeInfo, - const TargetData *td) + const DataLayout *td) : NewFunc(newFunc), OldFunc(oldFunc), VMap(valueMap), ModuleLevelChanges(moduleLevelChanges), NameSuffix(nameSuffix), CodeInfo(codeInfo), TD(td) { @@ -365,7 +369,7 @@ void llvm::CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc, SmallVectorImpl<ReturnInst*> &Returns, const char *NameSuffix, ClonedCodeInfo *CodeInfo, - const TargetData *TD, + const DataLayout *TD, Instruction *TheCall) { assert(NameSuffix && "NameSuffix cannot be null!"); diff --git a/contrib/llvm/lib/Transforms/Utils/CodeExtractor.cpp b/contrib/llvm/lib/Transforms/Utils/CodeExtractor.cpp index c545cd6..281714f 100644 --- a/contrib/llvm/lib/Transforms/Utils/CodeExtractor.cpp +++ b/contrib/llvm/lib/Transforms/Utils/CodeExtractor.cpp @@ -346,7 +346,7 @@ Function *CodeExtractor::constructFunction(const ValueSet &inputs, header->getName(), M); // If the old function is no-throw, so is the new one. if (oldFunction->doesNotThrow()) - newFunction->setDoesNotThrow(true); + newFunction->setDoesNotThrow(); newFunction->getBasicBlockList().push_back(newRootNode); diff --git a/contrib/llvm/lib/Transforms/Utils/InlineFunction.cpp b/contrib/llvm/lib/Transforms/Utils/InlineFunction.cpp index 89e89e7..009847f 100644 --- a/contrib/llvm/lib/Transforms/Utils/InlineFunction.cpp +++ b/contrib/llvm/lib/Transforms/Utils/InlineFunction.cpp @@ -27,7 +27,7 @@ #include "llvm/Analysis/CallGraph.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Support/CallSite.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Transforms/Utils/Local.h" using namespace llvm; @@ -357,7 +357,7 @@ static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, Type *VoidPtrTy = Type::getInt8PtrTy(Context); - // Create the alloca. If we have TargetData, use nice alignment. + // Create the alloca. If we have DataLayout, use nice alignment. unsigned Align = 1; if (IFI.TD) Align = IFI.TD->getPrefTypeAlignment(AggTy); diff --git a/contrib/llvm/lib/Transforms/Utils/IntegerDivision.cpp b/contrib/llvm/lib/Transforms/Utils/IntegerDivision.cpp new file mode 100644 index 0000000..55227e2 --- /dev/null +++ b/contrib/llvm/lib/Transforms/Utils/IntegerDivision.cpp @@ -0,0 +1,420 @@ +//===-- IntegerDivision.cpp - Expand integer division ---------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file contains an implementation of 32bit scalar integer division for +// targets that don't have native support. It's largely derived from +// compiler-rt's implementation of __udivsi3, but hand-tuned to reduce the +// amount of control flow +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "integer-division" +#include "llvm/Function.h" +#include "llvm/Instructions.h" +#include "llvm/Intrinsics.h" +#include "llvm/IRBuilder.h" +#include "llvm/Transforms/Utils/IntegerDivision.h" + +using namespace llvm; + +/// Generate code to compute the remainder of two signed integers. Returns the +/// remainder, which will have the sign of the dividend. Builder's insert point +/// should be pointing where the caller wants code generated, e.g. at the srem +/// instruction. This will generate a urem in the process, and Builder's insert +/// point will be pointing at the uren (if present, i.e. not folded), ready to +/// be expanded if the user wishes +static Value *generateSignedRemainderCode(Value *Dividend, Value *Divisor, + IRBuilder<> &Builder) { + ConstantInt *ThirtyOne = Builder.getInt32(31); + + // ; %dividend_sgn = ashr i32 %dividend, 31 + // ; %divisor_sgn = ashr i32 %divisor, 31 + // ; %dvd_xor = xor i32 %dividend, %dividend_sgn + // ; %dvs_xor = xor i32 %divisor, %divisor_sgn + // ; %u_dividend = sub i32 %dvd_xor, %dividend_sgn + // ; %u_divisor = sub i32 %dvs_xor, %divisor_sgn + // ; %urem = urem i32 %dividend, %divisor + // ; %xored = xor i32 %urem, %dividend_sgn + // ; %srem = sub i32 %xored, %dividend_sgn + Value *DividendSign = Builder.CreateAShr(Dividend, ThirtyOne); + Value *DivisorSign = Builder.CreateAShr(Divisor, ThirtyOne); + Value *DvdXor = Builder.CreateXor(Dividend, DividendSign); + Value *DvsXor = Builder.CreateXor(Divisor, DivisorSign); + Value *UDividend = Builder.CreateSub(DvdXor, DividendSign); + Value *UDivisor = Builder.CreateSub(DvsXor, DivisorSign); + Value *URem = Builder.CreateURem(UDividend, UDivisor); + Value *Xored = Builder.CreateXor(URem, DividendSign); + Value *SRem = Builder.CreateSub(Xored, DividendSign); + + if (Instruction *URemInst = dyn_cast<Instruction>(URem)) + Builder.SetInsertPoint(URemInst); + + return SRem; +} + + +/// Generate code to compute the remainder of two unsigned integers. Returns the +/// remainder. Builder's insert point should be pointing where the caller wants +/// code generated, e.g. at the urem instruction. This will generate a udiv in +/// the process, and Builder's insert point will be pointing at the udiv (if +/// present, i.e. not folded), ready to be expanded if the user wishes +static Value *generatedUnsignedRemainderCode(Value *Dividend, Value *Divisor, + IRBuilder<> &Builder) { + // Remainder = Dividend - Quotient*Divisor + + // ; %quotient = udiv i32 %dividend, %divisor + // ; %product = mul i32 %divisor, %quotient + // ; %remainder = sub i32 %dividend, %product + Value *Quotient = Builder.CreateUDiv(Dividend, Divisor); + Value *Product = Builder.CreateMul(Divisor, Quotient); + Value *Remainder = Builder.CreateSub(Dividend, Product); + + if (Instruction *UDiv = dyn_cast<Instruction>(Quotient)) + Builder.SetInsertPoint(UDiv); + + return Remainder; +} + +/// Generate code to divide two signed integers. Returns the quotient, rounded +/// towards 0. Builder's insert point should be pointing where the caller wants +/// code generated, e.g. at the sdiv instruction. This will generate a udiv in +/// the process, and Builder's insert point will be pointing at the udiv (if +/// present, i.e. not folded), ready to be expanded if the user wishes. +static Value *generateSignedDivisionCode(Value *Dividend, Value *Divisor, + IRBuilder<> &Builder) { + // Implementation taken from compiler-rt's __divsi3 + + ConstantInt *ThirtyOne = Builder.getInt32(31); + + // ; %tmp = ashr i32 %dividend, 31 + // ; %tmp1 = ashr i32 %divisor, 31 + // ; %tmp2 = xor i32 %tmp, %dividend + // ; %u_dvnd = sub nsw i32 %tmp2, %tmp + // ; %tmp3 = xor i32 %tmp1, %divisor + // ; %u_dvsr = sub nsw i32 %tmp3, %tmp1 + // ; %q_sgn = xor i32 %tmp1, %tmp + // ; %q_mag = udiv i32 %u_dvnd, %u_dvsr + // ; %tmp4 = xor i32 %q_mag, %q_sgn + // ; %q = sub i32 %tmp4, %q_sgn + Value *Tmp = Builder.CreateAShr(Dividend, ThirtyOne); + Value *Tmp1 = Builder.CreateAShr(Divisor, ThirtyOne); + Value *Tmp2 = Builder.CreateXor(Tmp, Dividend); + Value *U_Dvnd = Builder.CreateSub(Tmp2, Tmp); + Value *Tmp3 = Builder.CreateXor(Tmp1, Divisor); + Value *U_Dvsr = Builder.CreateSub(Tmp3, Tmp1); + Value *Q_Sgn = Builder.CreateXor(Tmp1, Tmp); + Value *Q_Mag = Builder.CreateUDiv(U_Dvnd, U_Dvsr); + Value *Tmp4 = Builder.CreateXor(Q_Mag, Q_Sgn); + Value *Q = Builder.CreateSub(Tmp4, Q_Sgn); + + if (Instruction *UDiv = dyn_cast<Instruction>(Q_Mag)) + Builder.SetInsertPoint(UDiv); + + return Q; +} + +/// Generates code to divide two unsigned scalar 32-bit integers. Returns the +/// quotient, rounded towards 0. Builder's insert point should be pointing where +/// the caller wants code generated, e.g. at the udiv instruction. +static Value *generateUnsignedDivisionCode(Value *Dividend, Value *Divisor, + IRBuilder<> &Builder) { + // The basic algorithm can be found in the compiler-rt project's + // implementation of __udivsi3.c. Here, we do a lower-level IR based approach + // that's been hand-tuned to lessen the amount of control flow involved. + + // Some helper values + IntegerType *I32Ty = Builder.getInt32Ty(); + + ConstantInt *Zero = Builder.getInt32(0); + ConstantInt *One = Builder.getInt32(1); + ConstantInt *ThirtyOne = Builder.getInt32(31); + ConstantInt *NegOne = ConstantInt::getSigned(I32Ty, -1); + ConstantInt *True = Builder.getTrue(); + + BasicBlock *IBB = Builder.GetInsertBlock(); + Function *F = IBB->getParent(); + Function *CTLZi32 = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz, + I32Ty); + + // Our CFG is going to look like: + // +---------------------+ + // | special-cases | + // | ... | + // +---------------------+ + // | | + // | +----------+ + // | | bb1 | + // | | ... | + // | +----------+ + // | | | + // | | +------------+ + // | | | preheader | + // | | | ... | + // | | +------------+ + // | | | + // | | | +---+ + // | | | | | + // | | +------------+ | + // | | | do-while | | + // | | | ... | | + // | | +------------+ | + // | | | | | + // | +-----------+ +---+ + // | | loop-exit | + // | | ... | + // | +-----------+ + // | | + // +-------+ + // | ... | + // | end | + // +-------+ + BasicBlock *SpecialCases = Builder.GetInsertBlock(); + SpecialCases->setName(Twine(SpecialCases->getName(), "_udiv-special-cases")); + BasicBlock *End = SpecialCases->splitBasicBlock(Builder.GetInsertPoint(), + "udiv-end"); + BasicBlock *LoopExit = BasicBlock::Create(Builder.getContext(), + "udiv-loop-exit", F, End); + BasicBlock *DoWhile = BasicBlock::Create(Builder.getContext(), + "udiv-do-while", F, End); + BasicBlock *Preheader = BasicBlock::Create(Builder.getContext(), + "udiv-preheader", F, End); + BasicBlock *BB1 = BasicBlock::Create(Builder.getContext(), + "udiv-bb1", F, End); + + // We'll be overwriting the terminator to insert our extra blocks + SpecialCases->getTerminator()->eraseFromParent(); + + // First off, check for special cases: dividend or divisor is zero, divisor + // is greater than dividend, and divisor is 1. + // ; special-cases: + // ; %ret0_1 = icmp eq i32 %divisor, 0 + // ; %ret0_2 = icmp eq i32 %dividend, 0 + // ; %ret0_3 = or i1 %ret0_1, %ret0_2 + // ; %tmp0 = tail call i32 @llvm.ctlz.i32(i32 %divisor, i1 true) + // ; %tmp1 = tail call i32 @llvm.ctlz.i32(i32 %dividend, i1 true) + // ; %sr = sub nsw i32 %tmp0, %tmp1 + // ; %ret0_4 = icmp ugt i32 %sr, 31 + // ; %ret0 = or i1 %ret0_3, %ret0_4 + // ; %retDividend = icmp eq i32 %sr, 31 + // ; %retVal = select i1 %ret0, i32 0, i32 %dividend + // ; %earlyRet = or i1 %ret0, %retDividend + // ; br i1 %earlyRet, label %end, label %bb1 + Builder.SetInsertPoint(SpecialCases); + Value *Ret0_1 = Builder.CreateICmpEQ(Divisor, Zero); + Value *Ret0_2 = Builder.CreateICmpEQ(Dividend, Zero); + Value *Ret0_3 = Builder.CreateOr(Ret0_1, Ret0_2); + Value *Tmp0 = Builder.CreateCall2(CTLZi32, Divisor, True); + Value *Tmp1 = Builder.CreateCall2(CTLZi32, Dividend, True); + Value *SR = Builder.CreateSub(Tmp0, Tmp1); + Value *Ret0_4 = Builder.CreateICmpUGT(SR, ThirtyOne); + Value *Ret0 = Builder.CreateOr(Ret0_3, Ret0_4); + Value *RetDividend = Builder.CreateICmpEQ(SR, ThirtyOne); + Value *RetVal = Builder.CreateSelect(Ret0, Zero, Dividend); + Value *EarlyRet = Builder.CreateOr(Ret0, RetDividend); + Builder.CreateCondBr(EarlyRet, End, BB1); + + // ; bb1: ; preds = %special-cases + // ; %sr_1 = add i32 %sr, 1 + // ; %tmp2 = sub i32 31, %sr + // ; %q = shl i32 %dividend, %tmp2 + // ; %skipLoop = icmp eq i32 %sr_1, 0 + // ; br i1 %skipLoop, label %loop-exit, label %preheader + Builder.SetInsertPoint(BB1); + Value *SR_1 = Builder.CreateAdd(SR, One); + Value *Tmp2 = Builder.CreateSub(ThirtyOne, SR); + Value *Q = Builder.CreateShl(Dividend, Tmp2); + Value *SkipLoop = Builder.CreateICmpEQ(SR_1, Zero); + Builder.CreateCondBr(SkipLoop, LoopExit, Preheader); + + // ; preheader: ; preds = %bb1 + // ; %tmp3 = lshr i32 %dividend, %sr_1 + // ; %tmp4 = add i32 %divisor, -1 + // ; br label %do-while + Builder.SetInsertPoint(Preheader); + Value *Tmp3 = Builder.CreateLShr(Dividend, SR_1); + Value *Tmp4 = Builder.CreateAdd(Divisor, NegOne); + Builder.CreateBr(DoWhile); + + // ; do-while: ; preds = %do-while, %preheader + // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ] + // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ] + // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ] + // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ] + // ; %tmp5 = shl i32 %r_1, 1 + // ; %tmp6 = lshr i32 %q_2, 31 + // ; %tmp7 = or i32 %tmp5, %tmp6 + // ; %tmp8 = shl i32 %q_2, 1 + // ; %q_1 = or i32 %carry_1, %tmp8 + // ; %tmp9 = sub i32 %tmp4, %tmp7 + // ; %tmp10 = ashr i32 %tmp9, 31 + // ; %carry = and i32 %tmp10, 1 + // ; %tmp11 = and i32 %tmp10, %divisor + // ; %r = sub i32 %tmp7, %tmp11 + // ; %sr_2 = add i32 %sr_3, -1 + // ; %tmp12 = icmp eq i32 %sr_2, 0 + // ; br i1 %tmp12, label %loop-exit, label %do-while + Builder.SetInsertPoint(DoWhile); + PHINode *Carry_1 = Builder.CreatePHI(I32Ty, 2); + PHINode *SR_3 = Builder.CreatePHI(I32Ty, 2); + PHINode *R_1 = Builder.CreatePHI(I32Ty, 2); + PHINode *Q_2 = Builder.CreatePHI(I32Ty, 2); + Value *Tmp5 = Builder.CreateShl(R_1, One); + Value *Tmp6 = Builder.CreateLShr(Q_2, ThirtyOne); + Value *Tmp7 = Builder.CreateOr(Tmp5, Tmp6); + Value *Tmp8 = Builder.CreateShl(Q_2, One); + Value *Q_1 = Builder.CreateOr(Carry_1, Tmp8); + Value *Tmp9 = Builder.CreateSub(Tmp4, Tmp7); + Value *Tmp10 = Builder.CreateAShr(Tmp9, 31); + Value *Carry = Builder.CreateAnd(Tmp10, One); + Value *Tmp11 = Builder.CreateAnd(Tmp10, Divisor); + Value *R = Builder.CreateSub(Tmp7, Tmp11); + Value *SR_2 = Builder.CreateAdd(SR_3, NegOne); + Value *Tmp12 = Builder.CreateICmpEQ(SR_2, Zero); + Builder.CreateCondBr(Tmp12, LoopExit, DoWhile); + + // ; loop-exit: ; preds = %do-while, %bb1 + // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ] + // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ] + // ; %tmp13 = shl i32 %q_3, 1 + // ; %q_4 = or i32 %carry_2, %tmp13 + // ; br label %end + Builder.SetInsertPoint(LoopExit); + PHINode *Carry_2 = Builder.CreatePHI(I32Ty, 2); + PHINode *Q_3 = Builder.CreatePHI(I32Ty, 2); + Value *Tmp13 = Builder.CreateShl(Q_3, One); + Value *Q_4 = Builder.CreateOr(Carry_2, Tmp13); + Builder.CreateBr(End); + + // ; end: ; preds = %loop-exit, %special-cases + // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ] + // ; ret i32 %q_5 + Builder.SetInsertPoint(End, End->begin()); + PHINode *Q_5 = Builder.CreatePHI(I32Ty, 2); + + // Populate the Phis, since all values have now been created. Our Phis were: + // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ] + Carry_1->addIncoming(Zero, Preheader); + Carry_1->addIncoming(Carry, DoWhile); + // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ] + SR_3->addIncoming(SR_1, Preheader); + SR_3->addIncoming(SR_2, DoWhile); + // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ] + R_1->addIncoming(Tmp3, Preheader); + R_1->addIncoming(R, DoWhile); + // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ] + Q_2->addIncoming(Q, Preheader); + Q_2->addIncoming(Q_1, DoWhile); + // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ] + Carry_2->addIncoming(Zero, BB1); + Carry_2->addIncoming(Carry, DoWhile); + // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ] + Q_3->addIncoming(Q, BB1); + Q_3->addIncoming(Q_1, DoWhile); + // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ] + Q_5->addIncoming(Q_4, LoopExit); + Q_5->addIncoming(RetVal, SpecialCases); + + return Q_5; +} + +/// Generate code to calculate the remainder of two integers, replacing Rem with +/// the generated code. This currently generates code using the udiv expansion, +/// but future work includes generating more specialized code, e.g. when more +/// information about the operands are known. Currently only implements 32bit +/// scalar division (due to udiv's limitation), but future work is removing this +/// limitation. +/// +/// @brief Replace Rem with generated code. +bool llvm::expandRemainder(BinaryOperator *Rem) { + assert((Rem->getOpcode() == Instruction::SRem || + Rem->getOpcode() == Instruction::URem) && + "Trying to expand remainder from a non-remainder function"); + + IRBuilder<> Builder(Rem); + + // First prepare the sign if it's a signed remainder + if (Rem->getOpcode() == Instruction::SRem) { + Value *Remainder = generateSignedRemainderCode(Rem->getOperand(0), + Rem->getOperand(1), Builder); + + Rem->replaceAllUsesWith(Remainder); + Rem->dropAllReferences(); + Rem->eraseFromParent(); + + // If we didn't actually generate a udiv instruction, we're done + BinaryOperator *BO = dyn_cast<BinaryOperator>(Builder.GetInsertPoint()); + if (!BO || BO->getOpcode() != Instruction::URem) + return true; + + Rem = BO; + } + + Value *Remainder = generatedUnsignedRemainderCode(Rem->getOperand(0), + Rem->getOperand(1), + Builder); + + Rem->replaceAllUsesWith(Remainder); + Rem->dropAllReferences(); + Rem->eraseFromParent(); + + // Expand the udiv + if (BinaryOperator *UDiv = dyn_cast<BinaryOperator>(Builder.GetInsertPoint())) { + assert(UDiv->getOpcode() == Instruction::UDiv && "Non-udiv in expansion?"); + expandDivision(UDiv); + } + + return true; +} + + +/// Generate code to divide two integers, replacing Div with the generated +/// code. This currently generates code similarly to compiler-rt's +/// implementations, but future work includes generating more specialized code +/// when more information about the operands are known. Currently only +/// implements 32bit scalar division, but future work is removing this +/// limitation. +/// +/// @brief Replace Div with generated code. +bool llvm::expandDivision(BinaryOperator *Div) { + assert((Div->getOpcode() == Instruction::SDiv || + Div->getOpcode() == Instruction::UDiv) && + "Trying to expand division from a non-division function"); + + IRBuilder<> Builder(Div); + + if (Div->getType()->isVectorTy()) + llvm_unreachable("Div over vectors not supported"); + + // First prepare the sign if it's a signed division + if (Div->getOpcode() == Instruction::SDiv) { + // Lower the code to unsigned division, and reset Div to point to the udiv. + Value *Quotient = generateSignedDivisionCode(Div->getOperand(0), + Div->getOperand(1), Builder); + Div->replaceAllUsesWith(Quotient); + Div->dropAllReferences(); + Div->eraseFromParent(); + + // If we didn't actually generate a udiv instruction, we're done + BinaryOperator *BO = dyn_cast<BinaryOperator>(Builder.GetInsertPoint()); + if (!BO || BO->getOpcode() != Instruction::UDiv) + return true; + + Div = BO; + } + + // Insert the unsigned division code + Value *Quotient = generateUnsignedDivisionCode(Div->getOperand(0), + Div->getOperand(1), + Builder); + Div->replaceAllUsesWith(Quotient); + Div->dropAllReferences(); + Div->eraseFromParent(); + + return true; +} diff --git a/contrib/llvm/lib/Transforms/Utils/LCSSA.cpp b/contrib/llvm/lib/Transforms/Utils/LCSSA.cpp index b654111..5e05c83 100644 --- a/contrib/llvm/lib/Transforms/Utils/LCSSA.cpp +++ b/contrib/llvm/lib/Transforms/Utils/LCSSA.cpp @@ -53,6 +53,8 @@ namespace { // Cached analysis information for the current function. DominatorTree *DT; + LoopInfo *LI; + ScalarEvolution *SE; std::vector<BasicBlock*> LoopBlocks; PredIteratorCache PredCache; Loop *L; @@ -117,6 +119,8 @@ bool LCSSA::runOnLoop(Loop *TheLoop, LPPassManager &LPM) { L = TheLoop; DT = &getAnalysis<DominatorTree>(); + LI = &getAnalysis<LoopInfo>(); + SE = getAnalysisIfAvailable<ScalarEvolution>(); // Get the set of exiting blocks. SmallVector<BasicBlock*, 8> ExitBlocks; @@ -156,6 +160,12 @@ bool LCSSA::runOnLoop(Loop *TheLoop, LPPassManager &LPM) { MadeChange |= ProcessInstruction(I, ExitBlocks); } } + + // If we modified the code, remove any caches about the loop from SCEV to + // avoid dangling entries. + // FIXME: This is a big hammer, can we clear the cache more selectively? + if (SE && MadeChange) + SE->forgetLoop(L); assert(L->isLCSSAForm(*DT)); PredCache.clear(); @@ -245,7 +255,7 @@ bool LCSSA::ProcessInstruction(Instruction *Inst, // Remember that this phi makes the value alive in this block. SSAUpdate.AddAvailableValue(ExitBB, PN); } - + // Rewrite all uses outside the loop in terms of the new PHIs we just // inserted. for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) { @@ -260,6 +270,9 @@ bool LCSSA::ProcessInstruction(Instruction *Inst, if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) { + // Tell the VHs that the uses changed. This updates SCEV's caches. + if (UsesToRewrite[i]->get()->hasValueHandle()) + ValueHandleBase::ValueIsRAUWd(*UsesToRewrite[i], UserBB->begin()); UsesToRewrite[i]->set(UserBB->begin()); continue; } diff --git a/contrib/llvm/lib/Transforms/Utils/Local.cpp b/contrib/llvm/lib/Transforms/Utils/Local.cpp index bed7d72..a954d82 100644 --- a/contrib/llvm/lib/Transforms/Utils/Local.cpp +++ b/contrib/llvm/lib/Transforms/Utils/Local.cpp @@ -23,6 +23,7 @@ #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" #include "llvm/Intrinsics.h" +#include "llvm/MDBuilder.h" #include "llvm/Metadata.h" #include "llvm/Operator.h" #include "llvm/ADT/DenseMap.h" @@ -38,7 +39,7 @@ #include "llvm/Support/MathExtras.h" #include "llvm/Support/ValueHandle.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" using namespace llvm; //===----------------------------------------------------------------------===// @@ -52,7 +53,8 @@ using namespace llvm; /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch /// conditions and indirectbr addresses this might make dead if /// DeleteDeadConditions is true. -bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { +bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, + const TargetLibraryInfo *TLI) { TerminatorInst *T = BB->getTerminator(); IRBuilder<> Builder(T); @@ -96,7 +98,7 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { Value *Cond = BI->getCondition(); BI->eraseFromParent(); if (DeleteDeadConditions) - RecursivelyDeleteTriviallyDeadInstructions(Cond); + RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); return true; } return false; @@ -121,6 +123,27 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { // Check to see if this branch is going to the same place as the default // dest. If so, eliminate it as an explicit compare. if (i.getCaseSuccessor() == DefaultDest) { + MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); + // MD should have 2 + NumCases operands. + if (MD && MD->getNumOperands() == 2 + SI->getNumCases()) { + // Collect branch weights into a vector. + SmallVector<uint32_t, 8> Weights; + for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; + ++MD_i) { + ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); + assert(CI); + Weights.push_back(CI->getValue().getZExtValue()); + } + // Merge weight of this case to the default weight. + unsigned idx = i.getCaseIndex(); + Weights[0] += Weights[idx+1]; + // Remove weight for this case. + std::swap(Weights[idx+1], Weights.back()); + Weights.pop_back(); + SI->setMetadata(LLVMContext::MD_prof, + MDBuilder(BB->getContext()). + createBranchWeights(Weights)); + } // Remove this entry. DefaultDest->removePredecessor(SI->getParent()); SI->removeCase(i); @@ -161,7 +184,7 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { Value *Cond = SI->getCondition(); SI->eraseFromParent(); if (DeleteDeadConditions) - RecursivelyDeleteTriviallyDeadInstructions(Cond); + RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); return true; } @@ -177,8 +200,20 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { "cond"); // Insert the new branch. - Builder.CreateCondBr(Cond, FirstCase.getCaseSuccessor(), - SI->getDefaultDest()); + BranchInst *NewBr = Builder.CreateCondBr(Cond, + FirstCase.getCaseSuccessor(), + SI->getDefaultDest()); + MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); + if (MD && MD->getNumOperands() == 3) { + ConstantInt *SICase = dyn_cast<ConstantInt>(MD->getOperand(2)); + ConstantInt *SIDef = dyn_cast<ConstantInt>(MD->getOperand(1)); + assert(SICase && SIDef); + // The TrueWeight should be the weight for the single case of SI. + NewBr->setMetadata(LLVMContext::MD_prof, + MDBuilder(BB->getContext()). + createBranchWeights(SICase->getValue().getZExtValue(), + SIDef->getValue().getZExtValue())); + } // Delete the old switch. SI->eraseFromParent(); @@ -205,7 +240,7 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { Value *Address = IBI->getAddress(); IBI->eraseFromParent(); if (DeleteDeadConditions) - RecursivelyDeleteTriviallyDeadInstructions(Address); + RecursivelyDeleteTriviallyDeadInstructions(Address, TLI); // If we didn't find our destination in the IBI successor list, then we // have undefined behavior. Replace the unconditional branch with an @@ -230,7 +265,8 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions) { /// isInstructionTriviallyDead - Return true if the result produced by the /// instruction is not used, and the instruction has no side effects. /// -bool llvm::isInstructionTriviallyDead(Instruction *I) { +bool llvm::isInstructionTriviallyDead(Instruction *I, + const TargetLibraryInfo *TLI) { if (!I->use_empty() || isa<TerminatorInst>(I)) return false; // We don't want the landingpad instruction removed by anything this general. @@ -265,9 +301,9 @@ bool llvm::isInstructionTriviallyDead(Instruction *I) { return isa<UndefValue>(II->getArgOperand(1)); } - if (isAllocLikeFn(I)) return true; + if (isAllocLikeFn(I, TLI)) return true; - if (CallInst *CI = isFreeCall(I)) + if (CallInst *CI = isFreeCall(I, TLI)) if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0))) return C->isNullValue() || isa<UndefValue>(C); @@ -278,9 +314,11 @@ bool llvm::isInstructionTriviallyDead(Instruction *I) { /// trivially dead instruction, delete it. If that makes any of its operands /// trivially dead, delete them too, recursively. Return true if any /// instructions were deleted. -bool llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V) { +bool +llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V, + const TargetLibraryInfo *TLI) { Instruction *I = dyn_cast<Instruction>(V); - if (!I || !I->use_empty() || !isInstructionTriviallyDead(I)) + if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI)) return false; SmallVector<Instruction*, 16> DeadInsts; @@ -301,7 +339,7 @@ bool llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V) { // operand, and if it is 'trivially' dead, delete it in a future loop // iteration. if (Instruction *OpI = dyn_cast<Instruction>(OpV)) - if (isInstructionTriviallyDead(OpI)) + if (isInstructionTriviallyDead(OpI, TLI)) DeadInsts.push_back(OpI); } @@ -334,19 +372,20 @@ static bool areAllUsesEqual(Instruction *I) { /// either forms a cycle or is terminated by a trivially dead instruction, /// delete it. If that makes any of its operands trivially dead, delete them /// too, recursively. Return true if a change was made. -bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN) { +bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN, + const TargetLibraryInfo *TLI) { SmallPtrSet<Instruction*, 4> Visited; for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects(); I = cast<Instruction>(*I->use_begin())) { if (I->use_empty()) - return RecursivelyDeleteTriviallyDeadInstructions(I); + return RecursivelyDeleteTriviallyDeadInstructions(I, TLI); // If we find an instruction more than once, we're on a cycle that // won't prove fruitful. if (!Visited.insert(I)) { // Break the cycle and delete the instruction and its operands. I->replaceAllUsesWith(UndefValue::get(I->getType())); - (void)RecursivelyDeleteTriviallyDeadInstructions(I); + (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI); return true; } } @@ -358,7 +397,8 @@ bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN) { /// /// This returns true if it changed the code, note that it can delete /// instructions in other blocks as well in this block. -bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, const TargetData *TD) { +bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, const DataLayout *TD, + const TargetLibraryInfo *TLI) { bool MadeChange = false; #ifndef NDEBUG @@ -381,7 +421,7 @@ bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, const TargetData *TD) { continue; } - MadeChange |= RecursivelyDeleteTriviallyDeadInstructions(Inst); + MadeChange |= RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); if (BIHandle != BI) BI = BB->begin(); } @@ -405,7 +445,7 @@ bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, const TargetData *TD) { /// .. and delete the predecessor corresponding to the '1', this will attempt to /// recursively fold the and to 0. void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, - TargetData *TD) { + DataLayout *TD) { // This only adjusts blocks with PHI nodes. if (!isa<PHINode>(BB->begin())) return; @@ -720,7 +760,7 @@ bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) { /// their preferred alignment from the beginning. /// static unsigned enforceKnownAlignment(Value *V, unsigned Align, - unsigned PrefAlign, const TargetData *TD) { + unsigned PrefAlign, const DataLayout *TD) { V = V->stripPointerCasts(); if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { @@ -763,7 +803,7 @@ static unsigned enforceKnownAlignment(Value *V, unsigned Align, /// and it is more than the alignment of the ultimate object, see if we can /// increase the alignment of the ultimate object, making this check succeed. unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign, - const TargetData *TD) { + const DataLayout *TD) { assert(V->getType()->isPointerTy() && "getOrEnforceKnownAlignment expects a pointer!"); unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 64; diff --git a/contrib/llvm/lib/Transforms/Utils/LoopSimplify.cpp b/contrib/llvm/lib/Transforms/Utils/LoopSimplify.cpp index 0bc185d..9d9e201 100644 --- a/contrib/llvm/lib/Transforms/Utils/LoopSimplify.cpp +++ b/contrib/llvm/lib/Transforms/Utils/LoopSimplify.cpp @@ -46,6 +46,7 @@ #include "llvm/LLVMContext.h" #include "llvm/Type.h" #include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/DependenceAnalysis.h" #include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/LoopPass.h" @@ -89,6 +90,7 @@ namespace { AU.addPreserved<AliasAnalysis>(); AU.addPreserved<ScalarEvolution>(); + AU.addPreserved<DependenceAnalysis>(); AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. } @@ -194,6 +196,11 @@ ReprocessLoop: BI->setCondition(ConstantInt::get(Cond->getType(), !L->contains(BI->getSuccessor(0)))); + + // This may make the loop analyzable, force SCEV recomputation. + if (SE) + SE->forgetLoop(L); + Changed = true; } } diff --git a/contrib/llvm/lib/Transforms/Utils/MetaRenamer.cpp b/contrib/llvm/lib/Transforms/Utils/MetaRenamer.cpp new file mode 100644 index 0000000..233bc12 --- /dev/null +++ b/contrib/llvm/lib/Transforms/Utils/MetaRenamer.cpp @@ -0,0 +1,132 @@ +//===- MetaRenamer.cpp - Rename everything with metasyntatic names --------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This pass renames everything with metasyntatic names. The intent is to use +// this pass after bugpoint reduction to conceal the nature of the original +// program. +// +//===----------------------------------------------------------------------===// + +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallString.h" +#include "llvm/Transforms/IPO.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Function.h" +#include "llvm/Module.h" +#include "llvm/Pass.h" +#include "llvm/Type.h" +#include "llvm/TypeFinder.h" + +using namespace llvm; + +namespace { + + // This PRNG is from the ISO C spec. It is intentionally simple and + // unsuitable for cryptographic use. We're just looking for enough + // variety to surprise and delight users. + struct PRNG { + unsigned long next; + + void srand(unsigned int seed) { + next = seed; + } + + int rand(void) { + next = next * 1103515245 + 12345; + return (unsigned int)(next / 65536) % 32768; + } + }; + + struct MetaRenamer : public ModulePass { + static char ID; // Pass identification, replacement for typeid + MetaRenamer() : ModulePass(ID) { + initializeMetaRenamerPass(*PassRegistry::getPassRegistry()); + } + + void getAnalysisUsage(AnalysisUsage &AU) const { + AU.setPreservesAll(); + } + + bool runOnModule(Module &M) { + static const char *metaNames[] = { + // See http://en.wikipedia.org/wiki/Metasyntactic_variable + "foo", "bar", "baz", "quux", "barney", "snork", "zot", "blam", "hoge", + "wibble", "wobble", "widget", "wombat", "ham", "eggs", "pluto", "spam" + }; + + // Seed our PRNG with simple additive sum of ModuleID. We're looking to + // simply avoid always having the same function names, and we need to + // remain deterministic. + unsigned int randSeed = 0; + for (std::string::const_iterator I = M.getModuleIdentifier().begin(), + E = M.getModuleIdentifier().end(); I != E; ++I) + randSeed += *I; + + PRNG prng; + prng.srand(randSeed); + + // Rename all aliases + for (Module::alias_iterator AI = M.alias_begin(), AE = M.alias_end(); + AI != AE; ++AI) + AI->setName("alias"); + + // Rename all global variables + for (Module::global_iterator GI = M.global_begin(), GE = M.global_end(); + GI != GE; ++GI) + GI->setName("global"); + + // Rename all struct types + TypeFinder StructTypes; + StructTypes.run(M, true); + for (unsigned i = 0, e = StructTypes.size(); i != e; ++i) { + StructType *STy = StructTypes[i]; + if (STy->isLiteral() || STy->getName().empty()) continue; + + SmallString<128> NameStorage; + STy->setName((Twine("struct.") + metaNames[prng.rand() % + array_lengthof(metaNames)]).toStringRef(NameStorage)); + } + + // Rename all functions + for (Module::iterator FI = M.begin(), FE = M.end(); + FI != FE; ++FI) { + FI->setName(metaNames[prng.rand() % array_lengthof(metaNames)]); + runOnFunction(*FI); + } + return true; + } + + bool runOnFunction(Function &F) { + for (Function::arg_iterator AI = F.arg_begin(), AE = F.arg_end(); + AI != AE; ++AI) + if (!AI->getType()->isVoidTy()) + AI->setName("arg"); + + for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { + BB->setName("bb"); + + for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) + if (!I->getType()->isVoidTy()) + I->setName("tmp"); + } + return true; + } + }; +} + +char MetaRenamer::ID = 0; +INITIALIZE_PASS(MetaRenamer, "metarenamer", + "Assign new names to everything", false, false) +//===----------------------------------------------------------------------===// +// +// MetaRenamer - Rename everything with metasyntactic names. +// +ModulePass *llvm::createMetaRenamerPass() { + return new MetaRenamer(); +} diff --git a/contrib/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp b/contrib/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp index dd5e20e..558de9d 100644 --- a/contrib/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp +++ b/contrib/llvm/lib/Transforms/Utils/PromoteMemoryToRegister.cpp @@ -212,9 +212,13 @@ namespace { /// DenseMap<AllocaInst*, unsigned> AllocaLookup; - /// NewPhiNodes - The PhiNodes we're adding. + /// NewPhiNodes - The PhiNodes we're adding. That map is used to simplify + /// some Phi nodes as we iterate over it, so it should have deterministic + /// iterators. We could use a MapVector, but since we already maintain a + /// map from BasicBlock* to a stable numbering (BBNumbers), the DenseMap is + /// more efficient (also supports removal). /// - DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes; + DenseMap<std::pair<unsigned, unsigned>, PHINode*> NewPhiNodes; /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas /// it corresponds to. @@ -588,7 +592,11 @@ void PromoteMem2Reg::run() { while (EliminatedAPHI) { EliminatedAPHI = false; - for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I = + // Iterating over NewPhiNodes is deterministic, so it is safe to try to + // simplify and RAUW them as we go. If it was not, we could add uses to + // the values we replace with in a non deterministic order, thus creating + // non deterministic def->use chains. + for (DenseMap<std::pair<unsigned, unsigned>, PHINode*>::iterator I = NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) { PHINode *PN = I->second; @@ -612,7 +620,7 @@ void PromoteMem2Reg::run() { // have incoming values for all predecessors. Loop over all PHI nodes we have // created, inserting undef values if they are missing any incoming values. // - for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I = + for (DenseMap<std::pair<unsigned, unsigned>, PHINode*>::iterator I = NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) { // We want to do this once per basic block. As such, only process a block // when we find the PHI that is the first entry in the block. @@ -992,7 +1000,7 @@ void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info, bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo, unsigned &Version) { // Look up the basic-block in question. - PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)]; + PHINode *&PN = NewPhiNodes[std::make_pair(BBNumbers[BB], AllocaNo)]; // If the BB already has a phi node added for the i'th alloca then we're done! if (PN) return false; diff --git a/contrib/llvm/lib/Transforms/Utils/SSAUpdater.cpp b/contrib/llvm/lib/Transforms/Utils/SSAUpdater.cpp index e568a61..72d4199 100644 --- a/contrib/llvm/lib/Transforms/Utils/SSAUpdater.cpp +++ b/contrib/llvm/lib/Transforms/Utils/SSAUpdater.cpp @@ -39,7 +39,7 @@ SSAUpdater::SSAUpdater(SmallVectorImpl<PHINode*> *NewPHI) : AV(0), ProtoType(0), ProtoName(), InsertedPHIs(NewPHI) {} SSAUpdater::~SSAUpdater() { - delete &getAvailableVals(AV); + delete static_cast<AvailableValsTy*>(AV); } /// Initialize - Reset this object to get ready for a new set of SSA diff --git a/contrib/llvm/lib/Transforms/Utils/SimplifyCFG.cpp b/contrib/llvm/lib/Transforms/Utils/SimplifyCFG.cpp index 518df7c..c767da6 100644 --- a/contrib/llvm/lib/Transforms/Utils/SimplifyCFG.cpp +++ b/contrib/llvm/lib/Transforms/Utils/SimplifyCFG.cpp @@ -14,6 +14,7 @@ #define DEBUG_TYPE "simplifycfg" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Constants.h" +#include "llvm/DataLayout.h" #include "llvm/DerivedTypes.h" #include "llvm/GlobalVariable.h" #include "llvm/IRBuilder.h" @@ -22,6 +23,7 @@ #include "llvm/LLVMContext.h" #include "llvm/MDBuilder.h" #include "llvm/Metadata.h" +#include "llvm/Module.h" #include "llvm/Operator.h" #include "llvm/Type.h" #include "llvm/ADT/DenseMap.h" @@ -38,7 +40,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/NoFolder.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Target/TargetData.h" +#include "llvm/TargetTransformInfo.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include <algorithm> #include <set> @@ -53,6 +55,13 @@ static cl::opt<bool> DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), cl::desc("Duplicate return instructions into unconditional branches")); +static cl::opt<bool> +SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), + cl::desc("Sink common instructions down to the end block")); + +STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); +STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); +STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); STATISTIC(NumSpeculations, "Number of speculative executed instructions"); namespace { @@ -68,10 +77,13 @@ namespace { // Comparing pointers is ok as we only rely on the order for uniquing. return Value < RHS.Value; } + + bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; } }; class SimplifyCFGOpt { - const TargetData *const TD; + const DataLayout *const TD; + const TargetTransformInfo *const TTI; Value *isValueEqualityComparison(TerminatorInst *TI); BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, @@ -91,7 +103,8 @@ class SimplifyCFGOpt { bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder); public: - explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} + SimplifyCFGOpt(const DataLayout *td, const TargetTransformInfo *tti) + : TD(td), TTI(tti) {} bool run(BasicBlock *BB); }; } @@ -101,14 +114,14 @@ public: /// static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { if (SI1 == SI2) return false; // Can't merge with self! - + // It is not safe to merge these two switch instructions if they have a common // successor, and if that successor has a PHI node, and if *that* PHI node has // conflicting incoming values from the two switch blocks. BasicBlock *SI1BB = SI1->getParent(); BasicBlock *SI2BB = SI2->getParent(); SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); - + for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) if (SI1Succs.count(*I)) for (BasicBlock::iterator BBI = (*I)->begin(); @@ -118,7 +131,7 @@ static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { PN->getIncomingValueForBlock(SI2BB)) return false; } - + return true; } @@ -135,7 +148,7 @@ static bool isProfitableToFoldUnconditional(BranchInst *SI1, assert(SI1->isUnconditional() && SI2->isConditional()); // We fold the unconditional branch if we can easily update all PHI nodes in - // common successors: + // common successors: // 1> We have a constant incoming value for the conditional branch; // 2> We have "Cond" as the incoming value for the unconditional branch; // 3> SI2->getCondition() and Cond have same operands. @@ -170,7 +183,7 @@ static bool isProfitableToFoldUnconditional(BranchInst *SI1, static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, BasicBlock *ExistPred) { if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do - + PHINode *PN; for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I) @@ -222,7 +235,7 @@ static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, // doesn't dominate BB. if (Pred2->getSinglePredecessor() == 0) return 0; - + // If we found a conditional branch predecessor, make sure that it branches // to BB and Pred2Br. If it doesn't, this isn't an "if statement". if (Pred1Br->getSuccessor(0) == BB && @@ -252,7 +265,7 @@ static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, // Otherwise, if this is a conditional branch, then we can use it! BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); if (BI == 0) return 0; - + assert(BI->isConditional() && "Two successors but not conditional?"); if (BI->getSuccessor(0) == Pred1) { IfTrue = Pred1; @@ -345,7 +358,7 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, // If we aren't allowing aggressive promotion anymore, then don't consider // instructions in the 'if region'. if (AggressiveInsts == 0) return false; - + // If we have seen this instruction before, don't count it again. if (AggressiveInsts->count(I)) return true; @@ -374,7 +387,7 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr /// and PointerNullValue. Return NULL if value is not a constant int. -static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { +static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) { // Normal constant int. ConstantInt *CI = dyn_cast<ConstantInt>(V); if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy()) @@ -382,7 +395,7 @@ static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { // This is some kind of pointer constant. Turn it into a pointer-sized // ConstantInt if possible. - IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); + IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType())); // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). if (isa<ConstantPointerNull>(V)) @@ -408,10 +421,10 @@ static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) { /// Values vector. static Value * GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, - const TargetData *TD, bool isEQ, unsigned &UsedICmps) { + const DataLayout *TD, bool isEQ, unsigned &UsedICmps) { Instruction *I = dyn_cast<Instruction>(V); if (I == 0) return 0; - + // If this is an icmp against a constant, handle this as one of the cases. if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { @@ -420,21 +433,21 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, Vals.push_back(C); return I->getOperand(0); } - + // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to // the set. ConstantRange Span = ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); - + // If this is an and/!= check then we want to optimize "x ugt 2" into // x != 0 && x != 1. if (!isEQ) Span = Span.inverse(); - + // If there are a ton of values, we don't want to make a ginormous switch. if (Span.getSetSize().ugt(8) || Span.isEmptySet()) return 0; - + for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); UsedICmps++; @@ -442,11 +455,11 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, } return 0; } - + // Otherwise, we can only handle an | or &, depending on isEQ. if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And)) return 0; - + unsigned NumValsBeforeLHS = Vals.size(); unsigned UsedICmpsBeforeLHS = UsedICmps; if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD, @@ -467,12 +480,12 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, Extra = I->getOperand(1); return LHS; } - + Vals.resize(NumValsBeforeLHS); UsedICmps = UsedICmpsBeforeLHS; return 0; } - + // If the LHS can't be folded in, but Extra is available and RHS can, try to // use LHS as Extra. if (Extra == 0 || Extra == I->getOperand(0)) { @@ -484,7 +497,7 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, assert(Vals.size() == NumValsBeforeLHS); Extra = OldExtra; } - + return 0; } @@ -556,11 +569,7 @@ GetValueEqualityComparisonCases(TerminatorInst *TI, /// in the list that match the specified block. static void EliminateBlockCases(BasicBlock *BB, std::vector<ValueEqualityComparisonCase> &Cases) { - for (unsigned i = 0, e = Cases.size(); i != e; ++i) - if (Cases[i].Dest == BB) { - Cases.erase(Cases.begin()+i); - --i; --e; - } + Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end()); } /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as @@ -615,6 +624,9 @@ SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, assert(ThisVal && "This isn't a value comparison!!"); if (ThisVal != PredVal) return false; // Different predicates. + // TODO: Preserve branch weight metadata, similarly to how + // FoldValueComparisonIntoPredecessors preserves it. + // Find out information about when control will move from Pred to TI's block. std::vector<ValueEqualityComparisonCase> PredCases; BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), @@ -634,7 +646,7 @@ SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, // can simplify TI. if (!ValuesOverlap(PredCases, ThisCases)) return false; - + if (isa<BranchInst>(TI)) { // Okay, one of the successors of this condbr is dead. Convert it to a // uncond br. @@ -652,7 +664,7 @@ SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, EraseTerminatorInstAndDCECond(TI); return true; } - + SwitchInst *SI = cast<SwitchInst>(TI); // Okay, TI has cases that are statically dead, prune them away. SmallPtrSet<Constant*, 16> DeadCases; @@ -662,18 +674,37 @@ SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() << "Through successor TI: " << *TI); + // Collect branch weights into a vector. + SmallVector<uint32_t, 8> Weights; + MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); + bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); + if (HasWeight) + for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; + ++MD_i) { + ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); + assert(CI); + Weights.push_back(CI->getValue().getZExtValue()); + } for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { --i; if (DeadCases.count(i.getCaseValue())) { + if (HasWeight) { + std::swap(Weights[i.getCaseIndex()+1], Weights.back()); + Weights.pop_back(); + } i.getCaseSuccessor()->removePredecessor(TI->getParent()); SI->removeCase(i); } } + if (HasWeight && Weights.size() >= 2) + SI->setMetadata(LLVMContext::MD_prof, + MDBuilder(SI->getParent()->getContext()). + createBranchWeights(Weights)); DEBUG(dbgs() << "Leaving: " << *TI << "\n"); return true; } - + // Otherwise, TI's block must correspond to some matched value. Find out // which value (or set of values) this is. ConstantInt *TIV = 0; @@ -729,8 +760,8 @@ namespace { } static int ConstantIntSortPredicate(const void *P1, const void *P2) { - const ConstantInt *LHS = *(const ConstantInt**)P1; - const ConstantInt *RHS = *(const ConstantInt**)P2; + const ConstantInt *LHS = *(const ConstantInt*const*)P1; + const ConstantInt *RHS = *(const ConstantInt*const*)P2; if (LHS->getValue().ult(RHS->getValue())) return 1; if (LHS->getValue() == RHS->getValue()) @@ -738,6 +769,56 @@ static int ConstantIntSortPredicate(const void *P1, const void *P2) { return -1; } +static inline bool HasBranchWeights(const Instruction* I) { + MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof); + if (ProfMD && ProfMD->getOperand(0)) + if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) + return MDS->getString().equals("branch_weights"); + + return false; +} + +/// Get Weights of a given TerminatorInst, the default weight is at the front +/// of the vector. If TI is a conditional eq, we need to swap the branch-weight +/// metadata. +static void GetBranchWeights(TerminatorInst *TI, + SmallVectorImpl<uint64_t> &Weights) { + MDNode* MD = TI->getMetadata(LLVMContext::MD_prof); + assert(MD); + for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) { + ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i)); + assert(CI); + Weights.push_back(CI->getValue().getZExtValue()); + } + + // If TI is a conditional eq, the default case is the false case, + // and the corresponding branch-weight data is at index 2. We swap the + // default weight to be the first entry. + if (BranchInst* BI = dyn_cast<BranchInst>(TI)) { + assert(Weights.size() == 2); + ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); + if (ICI->getPredicate() == ICmpInst::ICMP_EQ) + std::swap(Weights.front(), Weights.back()); + } +} + +/// Sees if any of the weights are too big for a uint32_t, and halves all the +/// weights if any are. +static void FitWeights(MutableArrayRef<uint64_t> Weights) { + bool Halve = false; + for (unsigned i = 0; i < Weights.size(); ++i) + if (Weights[i] > UINT_MAX) { + Halve = true; + break; + } + + if (! Halve) + return; + + for (unsigned i = 0; i < Weights.size(); ++i) + Weights[i] /= 2; +} + /// FoldValueComparisonIntoPredecessors - The specified terminator is a value /// equality comparison instruction (either a switch or a branch on "X == c"). /// See if any of the predecessors of the terminator block are value comparisons @@ -770,6 +851,31 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, // build. SmallVector<BasicBlock*, 8> NewSuccessors; + // Update the branch weight metadata along the way + SmallVector<uint64_t, 8> Weights; + bool PredHasWeights = HasBranchWeights(PTI); + bool SuccHasWeights = HasBranchWeights(TI); + + if (PredHasWeights) { + GetBranchWeights(PTI, Weights); + // branch-weight metadata is inconsistant here. + if (Weights.size() != 1 + PredCases.size()) + PredHasWeights = SuccHasWeights = false; + } else if (SuccHasWeights) + // If there are no predecessor weights but there are successor weights, + // populate Weights with 1, which will later be scaled to the sum of + // successor's weights + Weights.assign(1 + PredCases.size(), 1); + + SmallVector<uint64_t, 8> SuccWeights; + if (SuccHasWeights) { + GetBranchWeights(TI, SuccWeights); + // branch-weight metadata is inconsistant here. + if (SuccWeights.size() != 1 + BBCases.size()) + PredHasWeights = SuccHasWeights = false; + } else if (PredHasWeights) + SuccWeights.assign(1 + BBCases.size(), 1); + if (PredDefault == BB) { // If this is the default destination from PTI, only the edges in TI // that don't occur in PTI, or that branch to BB will be activated. @@ -780,6 +886,14 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, else { // The default destination is BB, we don't need explicit targets. std::swap(PredCases[i], PredCases.back()); + + if (PredHasWeights || SuccHasWeights) { + // Increase weight for the default case. + Weights[0] += Weights[i+1]; + std::swap(Weights[i+1], Weights.back()); + Weights.pop_back(); + } + PredCases.pop_back(); --i; --e; } @@ -790,21 +904,47 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, PredDefault = BBDefault; NewSuccessors.push_back(BBDefault); } + + unsigned CasesFromPred = Weights.size(); + uint64_t ValidTotalSuccWeight = 0; for (unsigned i = 0, e = BBCases.size(); i != e; ++i) if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) { PredCases.push_back(BBCases[i]); NewSuccessors.push_back(BBCases[i].Dest); + if (SuccHasWeights || PredHasWeights) { + // The default weight is at index 0, so weight for the ith case + // should be at index i+1. Scale the cases from successor by + // PredDefaultWeight (Weights[0]). + Weights.push_back(Weights[0] * SuccWeights[i+1]); + ValidTotalSuccWeight += SuccWeights[i+1]; + } } + if (SuccHasWeights || PredHasWeights) { + ValidTotalSuccWeight += SuccWeights[0]; + // Scale the cases from predecessor by ValidTotalSuccWeight. + for (unsigned i = 1; i < CasesFromPred; ++i) + Weights[i] *= ValidTotalSuccWeight; + // Scale the default weight by SuccDefaultWeight (SuccWeights[0]). + Weights[0] *= SuccWeights[0]; + } } else { // If this is not the default destination from PSI, only the edges // in SI that occur in PSI with a destination of BB will be // activated. std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; + std::map<ConstantInt*, uint64_t> WeightsForHandled; for (unsigned i = 0, e = PredCases.size(); i != e; ++i) if (PredCases[i].Dest == BB) { PTIHandled.insert(PredCases[i].Value); + + if (PredHasWeights || SuccHasWeights) { + WeightsForHandled[PredCases[i].Value] = Weights[i+1]; + std::swap(Weights[i+1], Weights.back()); + Weights.pop_back(); + } + std::swap(PredCases[i], PredCases.back()); PredCases.pop_back(); --i; --e; @@ -815,6 +955,8 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, for (unsigned i = 0, e = BBCases.size(); i != e; ++i) if (PTIHandled.count(BBCases[i].Value)) { // If this is one we are capable of getting... + if (PredHasWeights || SuccHasWeights) + Weights.push_back(WeightsForHandled[BBCases[i].Value]); PredCases.push_back(BBCases[i]); NewSuccessors.push_back(BBCases[i].Dest); PTIHandled.erase(BBCases[i].Value);// This constant is taken care of @@ -822,9 +964,11 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, // If there are any constants vectored to BB that TI doesn't handle, // they must go to the default destination of TI. - for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = + for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = PTIHandled.begin(), E = PTIHandled.end(); I != E; ++I) { + if (PredHasWeights || SuccHasWeights) + Weights.push_back(WeightsForHandled[*I]); PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); NewSuccessors.push_back(BBDefault); } @@ -839,7 +983,7 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, Builder.SetInsertPoint(PTI); // Convert pointer to int before we switch. if (CV->getType()->isPointerTy()) { - assert(TD && "Cannot switch on pointer without TargetData"); + assert(TD && "Cannot switch on pointer without DataLayout"); CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()), "magicptr"); } @@ -851,6 +995,17 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, for (unsigned i = 0, e = PredCases.size(); i != e; ++i) NewSI->addCase(PredCases[i].Value, PredCases[i].Dest); + if (PredHasWeights || SuccHasWeights) { + // Halve the weights if any of them cannot fit in an uint32_t + FitWeights(Weights); + + SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); + + NewSI->setMetadata(LLVMContext::MD_prof, + MDBuilder(BB->getContext()). + createBranchWeights(MDWeights)); + } + EraseTerminatorInstAndDCECond(PTI); // Okay, last check. If BB is still a successor of PSI, then we must @@ -984,11 +1139,11 @@ HoistTerminator: Value *BB1V = PN->getIncomingValueForBlock(BB1); Value *BB2V = PN->getIncomingValueForBlock(BB2); if (BB1V == BB2V) continue; - + // These values do not agree. Insert a select instruction before NT // that determines the right value. SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; - if (SI == 0) + if (SI == 0) SI = cast<SelectInst> (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V, BB1V->getName()+"."+BB2V->getName())); @@ -1008,6 +1163,175 @@ HoistTerminator: return true; } +/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, +/// check whether BBEnd has only two predecessors and the other predecessor +/// ends with an unconditional branch. If it is true, sink any common code +/// in the two predecessors to BBEnd. +static bool SinkThenElseCodeToEnd(BranchInst *BI1) { + assert(BI1->isUnconditional()); + BasicBlock *BB1 = BI1->getParent(); + BasicBlock *BBEnd = BI1->getSuccessor(0); + + // Check that BBEnd has two predecessors and the other predecessor ends with + // an unconditional branch. + pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd); + BasicBlock *Pred0 = *PI++; + if (PI == PE) // Only one predecessor. + return false; + BasicBlock *Pred1 = *PI++; + if (PI != PE) // More than two predecessors. + return false; + BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0; + BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); + if (!BI2 || !BI2->isUnconditional()) + return false; + + // Gather the PHI nodes in BBEnd. + std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; + Instruction *FirstNonPhiInBBEnd = 0; + for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); + I != E; ++I) { + if (PHINode *PN = dyn_cast<PHINode>(I)) { + Value *BB1V = PN->getIncomingValueForBlock(BB1); + Value *BB2V = PN->getIncomingValueForBlock(BB2); + MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); + } else { + FirstNonPhiInBBEnd = &*I; + break; + } + } + if (!FirstNonPhiInBBEnd) + return false; + + + // This does very trivial matching, with limited scanning, to find identical + // instructions in the two blocks. We scan backward for obviously identical + // instructions in an identical order. + BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), + RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), + RE2 = BB2->getInstList().rend(); + // Skip debug info. + while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; + if (RI1 == RE1) + return false; + while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; + if (RI2 == RE2) + return false; + // Skip the unconditional branches. + ++RI1; + ++RI2; + + bool Changed = false; + while (RI1 != RE1 && RI2 != RE2) { + // Skip debug info. + while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; + if (RI1 == RE1) + return Changed; + while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; + if (RI2 == RE2) + return Changed; + + Instruction *I1 = &*RI1, *I2 = &*RI2; + // I1 and I2 should have a single use in the same PHI node, and they + // perform the same operation. + // Cannot move control-flow-involving, volatile loads, vaarg, etc. + if (isa<PHINode>(I1) || isa<PHINode>(I2) || + isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || + isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || + isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || + I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || + I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || + !I1->hasOneUse() || !I2->hasOneUse() || + MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || + MapValueFromBB1ToBB2[I1].first != I2) + return Changed; + + // Check whether we should swap the operands of ICmpInst. + ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); + bool SwapOpnds = false; + if (ICmp1 && ICmp2 && + ICmp1->getOperand(0) != ICmp2->getOperand(0) && + ICmp1->getOperand(1) != ICmp2->getOperand(1) && + (ICmp1->getOperand(0) == ICmp2->getOperand(1) || + ICmp1->getOperand(1) == ICmp2->getOperand(0))) { + ICmp2->swapOperands(); + SwapOpnds = true; + } + if (!I1->isSameOperationAs(I2)) { + if (SwapOpnds) + ICmp2->swapOperands(); + return Changed; + } + + // The operands should be either the same or they need to be generated + // with a PHI node after sinking. We only handle the case where there is + // a single pair of different operands. + Value *DifferentOp1 = 0, *DifferentOp2 = 0; + unsigned Op1Idx = 0; + for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { + if (I1->getOperand(I) == I2->getOperand(I)) + continue; + // Early exit if we have more-than one pair of different operands or + // the different operand is already in MapValueFromBB1ToBB2. + // Early exit if we need a PHI node to replace a constant. + if (DifferentOp1 || + MapValueFromBB1ToBB2.find(I1->getOperand(I)) != + MapValueFromBB1ToBB2.end() || + isa<Constant>(I1->getOperand(I)) || + isa<Constant>(I2->getOperand(I))) { + // If we can't sink the instructions, undo the swapping. + if (SwapOpnds) + ICmp2->swapOperands(); + return Changed; + } + DifferentOp1 = I1->getOperand(I); + Op1Idx = I; + DifferentOp2 = I2->getOperand(I); + } + + // We insert the pair of different operands to MapValueFromBB1ToBB2 and + // remove (I1, I2) from MapValueFromBB1ToBB2. + if (DifferentOp1) { + PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, + DifferentOp1->getName() + ".sink", + BBEnd->begin()); + MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); + // I1 should use NewPN instead of DifferentOp1. + I1->setOperand(Op1Idx, NewPN); + NewPN->addIncoming(DifferentOp1, BB1); + NewPN->addIncoming(DifferentOp2, BB2); + DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); + } + PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; + MapValueFromBB1ToBB2.erase(I1); + + DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); + DEBUG(dbgs() << " " << *I2 << "\n";); + // We need to update RE1 and RE2 if we are going to sink the first + // instruction in the basic block down. + bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); + // Sink the instruction. + BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); + if (!OldPN->use_empty()) + OldPN->replaceAllUsesWith(I1); + OldPN->eraseFromParent(); + + if (!I2->use_empty()) + I2->replaceAllUsesWith(I1); + I1->intersectOptionalDataWith(I2); + I2->eraseFromParent(); + + if (UpdateRE1) + RE1 = BB1->getInstList().rend(); + if (UpdateRE2) + RE2 = BB2->getInstList().rend(); + FirstNonPhiInBBEnd = I1; + NumSinkCommons++; + Changed = true; + } + return Changed; +} + /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code /// (for now, restricted to a single instruction that's side effect free) from @@ -1056,7 +1380,7 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { // Do not hoist the instruction if any of its operands are defined but not // used in this BB. The transformation will prevent the operand from // being sunk into the use block. - for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); + for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); i != e; ++i) { Instruction *OpI = dyn_cast<Instruction>(*i); if (OpI && OpI->getParent() == BIParent && @@ -1112,7 +1436,7 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { // as well. if (PHIs.empty()) return false; - + // If we get here, we can hoist the instruction and if-convert. DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";); @@ -1162,13 +1486,13 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { BranchInst *BI = cast<BranchInst>(BB->getTerminator()); unsigned Size = 0; - + for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { if (isa<DbgInfoIntrinsic>(BBI)) continue; if (Size > 10) return false; // Don't clone large BB's. ++Size; - + // We can only support instructions that do not define values that are // live outside of the current basic block. for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); @@ -1176,7 +1500,7 @@ static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { Instruction *U = cast<Instruction>(*UI); if (U->getParent() != BB || isa<PHINode>(U)) return false; } - + // Looks ok, continue checking. } @@ -1187,38 +1511,38 @@ static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { /// that is defined in the same block as the branch and if any PHI entries are /// constants, thread edges corresponding to that entry to be branches to their /// ultimate destination. -static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { +static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) { BasicBlock *BB = BI->getParent(); PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); // NOTE: we currently cannot transform this case if the PHI node is used // outside of the block. if (!PN || PN->getParent() != BB || !PN->hasOneUse()) return false; - + // Degenerate case of a single entry PHI. if (PN->getNumIncomingValues() == 1) { FoldSingleEntryPHINodes(PN->getParent()); - return true; + return true; } // Now we know that this block has multiple preds and two succs. if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; - + // Okay, this is a simple enough basic block. See if any phi values are // constants. for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i)); if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue; - + // Okay, we now know that all edges from PredBB should be revectored to // branch to RealDest. BasicBlock *PredBB = PN->getIncomingBlock(i); BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); - + if (RealDest == BB) continue; // Skip self loops. // Skip if the predecessor's terminator is an indirect branch. if (isa<IndirectBrInst>(PredBB->getTerminator())) continue; - + // The dest block might have PHI nodes, other predecessors and other // difficult cases. Instead of being smart about this, just insert a new // block that jumps to the destination block, effectively splitting @@ -1227,7 +1551,7 @@ static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { RealDest->getName()+".critedge", RealDest->getParent(), RealDest); BranchInst::Create(RealDest, EdgeBB); - + // Update PHI nodes. AddPredecessorToBlock(RealDest, EdgeBB, BB); @@ -1244,7 +1568,7 @@ static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { // Clone the instruction. Instruction *N = BBI->clone(); if (BBI->hasName()) N->setName(BBI->getName()+".c"); - + // Update operands due to translation. for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) { @@ -1252,7 +1576,7 @@ static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { if (PI != TranslateMap.end()) *i = PI->second; } - + // Check for trivial simplification. if (Value *V = SimplifyInstruction(N, TD)) { TranslateMap[BBI] = V; @@ -1283,7 +1607,7 @@ static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) { /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry /// PHI node, see if we can eliminate it. -static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { +static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) { // Ok, this is a two entry PHI node. Check to see if this is a simple "if // statement", which has a very simple dominance structure. Basically, we // are trying to find the condition that is being branched on, which @@ -1297,7 +1621,7 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { // Don't bother if the branch will be constant folded trivially. isa<ConstantInt>(IfCond)) return false; - + // Okay, we found that we can merge this two-entry phi node into a select. // Doing so would require us to fold *all* two entry phi nodes in this block. // At some point this becomes non-profitable (particularly if the target @@ -1307,14 +1631,14 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) if (NumPhis > 2) return false; - + // Loop over the PHI's seeing if we can promote them all to select // instructions. While we are at it, keep track of the instructions // that need to be moved to the dominating block. SmallPtrSet<Instruction*, 4> AggressiveInsts; unsigned MaxCostVal0 = PHINodeFoldingThreshold, MaxCostVal1 = PHINodeFoldingThreshold; - + for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) { PHINode *PN = cast<PHINode>(II++); if (Value *V = SimplifyInstruction(PN, TD)) { @@ -1322,19 +1646,19 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { PN->eraseFromParent(); continue; } - + if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts, MaxCostVal0) || !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts, MaxCostVal1)) return false; } - + // If we folded the first phi, PN dangles at this point. Refresh it. If // we ran out of PHIs then we simplified them all. PN = dyn_cast<PHINode>(BB->begin()); if (PN == 0) return true; - + // Don't fold i1 branches on PHIs which contain binary operators. These can // often be turned into switches and other things. if (PN->getType()->isIntegerTy(1) && @@ -1342,7 +1666,7 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { isa<BinaryOperator>(PN->getIncomingValue(1)) || isa<BinaryOperator>(IfCond))) return false; - + // If we all PHI nodes are promotable, check to make sure that all // instructions in the predecessor blocks can be promoted as well. If // not, we won't be able to get rid of the control flow, so it's not @@ -1362,7 +1686,7 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { return false; } } - + if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) { IfBlock2 = 0; } else { @@ -1375,15 +1699,15 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { return false; } } - + DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: " << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); - + // If we can still promote the PHI nodes after this gauntlet of tests, // do all of the PHI's now. Instruction *InsertPt = DomBlock->getTerminator(); IRBuilder<true, NoFolder> Builder(InsertPt); - + // Move all 'aggressive' instructions, which are defined in the // conditional parts of the if's up to the dominating block. if (IfBlock1) @@ -1394,19 +1718,19 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { DomBlock->getInstList().splice(InsertPt, IfBlock2->getInstList(), IfBlock2->begin(), IfBlock2->getTerminator()); - + while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { // Change the PHI node into a select instruction. Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); - - SelectInst *NV = + + SelectInst *NV = cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, "")); PN->replaceAllUsesWith(NV); NV->takeName(PN); PN->eraseFromParent(); } - + // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement // has been flattened. Change DomBlock to jump directly to our new block to // avoid other simplifycfg's kicking in on the diamond. @@ -1420,14 +1744,14 @@ static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) { /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes /// to two returning blocks, try to merge them together into one return, /// introducing a select if the return values disagree. -static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, +static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, IRBuilder<> &Builder) { assert(BI->isConditional() && "Must be a conditional branch"); BasicBlock *TrueSucc = BI->getSuccessor(0); BasicBlock *FalseSucc = BI->getSuccessor(1); ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); - + // Check to ensure both blocks are empty (just a return) or optionally empty // with PHI nodes. If there are other instructions, merging would cause extra // computation on one path or the other. @@ -1447,12 +1771,12 @@ static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, EraseTerminatorInstAndDCECond(BI); return true; } - + // Otherwise, figure out what the true and false return values are // so we can insert a new select instruction. Value *TrueValue = TrueRet->getReturnValue(); Value *FalseValue = FalseRet->getReturnValue(); - + // Unwrap any PHI nodes in the return blocks. if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) if (TVPN->getParent() == TrueSucc) @@ -1460,7 +1784,7 @@ static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) if (FVPN->getParent() == FalseSucc) FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); - + // In order for this transformation to be safe, we must be able to // unconditionally execute both operands to the return. This is // normally the case, but we could have a potentially-trapping @@ -1472,12 +1796,12 @@ static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) if (FCV->canTrap()) return false; - + // Okay, we collected all the mapped values and checked them for sanity, and // defined to really do this transformation. First, update the CFG. TrueSucc->removePredecessor(BI->getParent()); FalseSucc->removePredecessor(BI->getParent()); - + // Insert select instructions where needed. Value *BrCond = BI->getCondition(); if (TrueValue) { @@ -1491,15 +1815,15 @@ static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, } } - Value *RI = !TrueValue ? + Value *RI = !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue); (void) RI; - + DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc); - + EraseTerminatorInstAndDCECond(BI); return true; @@ -1510,7 +1834,7 @@ static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, /// parameters and return true, or returns false if no or invalid metadata was /// found. static bool ExtractBranchMetadata(BranchInst *BI, - APInt &ProbTrue, APInt &ProbFalse) { + uint64_t &ProbTrue, uint64_t &ProbFalse) { assert(BI->isConditional() && "Looking for probabilities on unconditional branch?"); MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); @@ -1518,35 +1842,11 @@ static bool ExtractBranchMetadata(BranchInst *BI, ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); if (!CITrue || !CIFalse) return false; - ProbTrue = CITrue->getValue(); - ProbFalse = CIFalse->getValue(); - assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 && - "Branch probability metadata must be 32-bit integers"); + ProbTrue = CITrue->getValue().getZExtValue(); + ProbFalse = CIFalse->getValue().getZExtValue(); return true; } -/// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In -/// the event of overflow, logically-shifts all four inputs right until the -/// multiply fits. -static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D, - unsigned &BitsLost) { - BitsLost = 0; - bool Overflow = false; - APInt Result = A.umul_ov(B, Overflow); - if (Overflow) { - APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A); - do { - B = B.lshr(1); - ++BitsLost; - } while (B.ugt(MaxB)); - A = A.lshr(BitsLost); - C = C.lshr(BitsLost); - D = D.lshr(BitsLost); - Result = A * B; - } - return Result; -} - /// checkCSEInPredecessor - Return true if the given instruction is available /// in its predecessor block. If yes, the instruction will be removed. /// @@ -1600,7 +1900,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { if (Cond == 0) return false; } - + if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || Cond->getParent() != BB || !Cond->hasOneUse()) return false; @@ -1623,7 +1923,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { isSafeToSpeculativelyExecute(FrontIt)) { BonusInst = &*FrontIt; ++FrontIt; - + // Ignore dbg intrinsics. while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt; } @@ -1631,13 +1931,13 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { // Only a single bonus inst is allowed. if (&*FrontIt != Cond) return false; - + // Make sure the instruction after the condition is the cond branch. BasicBlock::iterator CondIt = Cond; ++CondIt; // Ingore dbg intrinsics. while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt; - + if (&*CondIt != BI) return false; @@ -1649,7 +1949,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) if (CE->canTrap()) return false; - + // Finally, don't infinitely unroll conditional loops. BasicBlock *TrueDest = BI->getSuccessor(0); BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0; @@ -1659,22 +1959,22 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { BasicBlock *PredBlock = *PI; BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); - + // Check that we have two conditional branches. If there is a PHI node in // the common successor, verify that the same value flows in from both // blocks. SmallVector<PHINode*, 4> PHIs; if (PBI == 0 || PBI->isUnconditional() || - (BI->isConditional() && + (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) || (!BI->isConditional() && !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs))) continue; - + // Determine if the two branches share a common destination. - Instruction::BinaryOps Opc; + Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; bool InvertPredCond = false; - + if (BI->isConditional()) { if (PBI->getSuccessor(0) == TrueDest) Opc = Instruction::Or; @@ -1693,7 +1993,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { // Ensure that any values used in the bonus instruction are also used // by the terminator of the predecessor. This means that those values - // must already have been resolved, so we won't be inhibiting the + // must already have been resolved, so we won't be inhibiting the // out-of-order core by speculating them earlier. if (BonusInst) { // Collect the values used by the bonus inst @@ -1707,47 +2007,47 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { SmallVector<std::pair<Value*, unsigned>, 4> Worklist; Worklist.push_back(std::make_pair(PBI->getOperand(0), 0)); - + // Walk up to four levels back up the use-def chain of the predecessor's // terminator to see if all those values were used. The choice of four // levels is arbitrary, to provide a compile-time-cost bound. while (!Worklist.empty()) { std::pair<Value*, unsigned> Pair = Worklist.back(); Worklist.pop_back(); - + if (Pair.second >= 4) continue; UsedValues.erase(Pair.first); if (UsedValues.empty()) break; - + if (Instruction *I = dyn_cast<Instruction>(Pair.first)) { for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) Worklist.push_back(std::make_pair(OI->get(), Pair.second+1)); - } + } } - + if (!UsedValues.empty()) return false; } DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB); - IRBuilder<> Builder(PBI); + IRBuilder<> Builder(PBI); // If we need to invert the condition in the pred block to match, do so now. if (InvertPredCond) { Value *NewCond = PBI->getCondition(); - + if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { CmpInst *CI = cast<CmpInst>(NewCond); CI->setPredicate(CI->getInversePredicate()); } else { - NewCond = Builder.CreateNot(NewCond, + NewCond = Builder.CreateNot(NewCond, PBI->getCondition()->getName()+".not"); } - + PBI->setCondition(NewCond); PBI->swapSuccessors(); } - + // If we have a bonus inst, clone it into the predecessor block. Instruction *NewBonus = 0; if (BonusInst) { @@ -1756,7 +2056,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { NewBonus->takeName(BonusInst); BonusInst->setName(BonusInst->getName()+".old"); } - + // Clone Cond into the predecessor basic block, and or/and the // two conditions together. Instruction *New = Cond->clone(); @@ -1764,21 +2064,60 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { PredBlock->getInstList().insert(PBI, New); New->takeName(Cond); Cond->setName(New->getName()+".old"); - + if (BI->isConditional()) { - Instruction *NewCond = + Instruction *NewCond = cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond")); PBI->setCondition(NewCond); + uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; + bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, + PredFalseWeight); + bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, + SuccFalseWeight); + SmallVector<uint64_t, 8> NewWeights; + if (PBI->getSuccessor(0) == BB) { + if (PredHasWeights && SuccHasWeights) { + // PBI: br i1 %x, BB, FalseDest + // BI: br i1 %y, TrueDest, FalseDest + //TrueWeight is TrueWeight for PBI * TrueWeight for BI. + NewWeights.push_back(PredTrueWeight * SuccTrueWeight); + //FalseWeight is FalseWeight for PBI * TotalWeight for BI + + // TrueWeight for PBI * FalseWeight for BI. + // We assume that total weights of a BranchInst can fit into 32 bits. + // Therefore, we will not have overflow using 64-bit arithmetic. + NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + + SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); + } AddPredecessorToBlock(TrueDest, PredBlock, BB); PBI->setSuccessor(0, TrueDest); } if (PBI->getSuccessor(1) == BB) { + if (PredHasWeights && SuccHasWeights) { + // PBI: br i1 %x, TrueDest, BB + // BI: br i1 %y, TrueDest, FalseDest + //TrueWeight is TrueWeight for PBI * TotalWeight for BI + + // FalseWeight for PBI * TrueWeight for BI. + NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + + SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); + //FalseWeight is FalseWeight for PBI * FalseWeight for BI. + NewWeights.push_back(PredFalseWeight * SuccFalseWeight); + } AddPredecessorToBlock(FalseDest, PredBlock, BB); PBI->setSuccessor(1, FalseDest); } + if (NewWeights.size() == 2) { + // Halve the weights if any of them cannot fit in an uint32_t + FitWeights(NewWeights); + + SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); + PBI->setMetadata(LLVMContext::MD_prof, + MDBuilder(BI->getContext()). + createBranchWeights(MDWeights)); + } else + PBI->setMetadata(LLVMContext::MD_prof, NULL); } else { // Update PHI nodes in the common successors. for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { @@ -1806,7 +2145,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C) // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond) // is false: PBI_Cond and BI_Value - MergedCond = + MergedCond = cast<Instruction>(Builder.CreateBinOp(Instruction::And, PBI->getCondition(), New, "and.cond")); @@ -1814,7 +2153,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { Instruction *NotCond = cast<Instruction>(Builder.CreateNot(PBI->getCondition(), "not.cond")); - MergedCond = + MergedCond = cast<Instruction>(Builder.CreateBinOp(Instruction::Or, NotCond, MergedCond, "or.cond")); @@ -1833,95 +2172,11 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { // TODO: If BB is reachable from all paths through PredBlock, then we // could replace PBI's branch probabilities with BI's. - // Merge probability data into PredBlock's branch. - APInt A, B, C, D; - if (PBI->isConditional() && BI->isConditional() && - ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) { - // Given IR which does: - // bbA: - // br i1 %x, label %bbB, label %bbC - // bbB: - // br i1 %y, label %bbD, label %bbC - // Let's call the probability that we take the edge from %bbA to %bbB - // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to - // %bbC probability 'd'. - // - // We transform the IR into: - // bbA: - // br i1 %z, label %bbD, label %bbC - // where the probability of going to %bbD is (a*c) and going to bbC is - // (b+a*d). - // - // Probabilities aren't stored as ratios directly. Using branch weights, - // we get: - // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D. - - // In the event of overflow, we want to drop the LSB of the input - // probabilities. - unsigned BitsLost; - - // Ignore overflow result on ProbTrue. - APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost); - - APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost); - if (BitsLost) { - ProbTrue = ProbTrue.lshr(BitsLost*2); - } - - APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost); - if (BitsLost) { - ProbTrue = ProbTrue.lshr(BitsLost*2); - Tmp1 = Tmp1.lshr(BitsLost*2); - } - - APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost); - if (BitsLost) { - ProbTrue = ProbTrue.lshr(BitsLost*2); - Tmp1 = Tmp1.lshr(BitsLost*2); - Tmp2 = Tmp2.lshr(BitsLost*2); - } - - bool Overflow1 = false, Overflow2 = false; - APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1); - APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2); - - if (Overflow1 || Overflow2) { - ProbTrue = ProbTrue.lshr(1); - Tmp1 = Tmp1.lshr(1); - Tmp2 = Tmp2.lshr(1); - Tmp3 = Tmp3.lshr(1); - Tmp4 = Tmp2 + Tmp3; - ProbFalse = Tmp4 + Tmp1; - } - - // The sum of branch weights must fit in 32-bits. - if (ProbTrue.isNegative() && ProbFalse.isNegative()) { - ProbTrue = ProbTrue.lshr(1); - ProbFalse = ProbFalse.lshr(1); - } - - if (ProbTrue != ProbFalse) { - // Normalize the result. - APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse); - ProbTrue = ProbTrue.udiv(GCD); - ProbFalse = ProbFalse.udiv(GCD); - - MDBuilder MDB(BI->getContext()); - MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(), - ProbFalse.getZExtValue()); - PBI->setMetadata(LLVMContext::MD_prof, N); - } else { - PBI->setMetadata(LLVMContext::MD_prof, NULL); - } - } else { - PBI->setMetadata(LLVMContext::MD_prof, NULL); - } - // Copy any debug value intrinsics into the end of PredBlock. for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) if (isa<DbgInfoIntrinsic>(*I)) I->clone()->insertBefore(PBI); - + return true; } return false; @@ -1936,7 +2191,7 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { BasicBlock *BB = BI->getParent(); // If this block ends with a branch instruction, and if there is a - // predecessor that ends on a branch of the same condition, make + // predecessor that ends on a branch of the same condition, make // this conditional branch redundant. if (PBI->getCondition() == BI->getCondition() && PBI->getSuccessor(0) != PBI->getSuccessor(1)) { @@ -1945,11 +2200,11 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { if (BB->getSinglePredecessor()) { // Turn this into a branch on constant. bool CondIsTrue = PBI->getSuccessor(0) == BB; - BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), + BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue)); return true; // Nuke the branch on constant. } - + // Otherwise, if there are multiple predecessors, insert a PHI that merges // in the constant and simplify the block result. Subsequent passes of // simplifycfg will thread the block. @@ -1969,18 +2224,18 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { PBI->getCondition() == BI->getCondition() && PBI->getSuccessor(0) != PBI->getSuccessor(1)) { bool CondIsTrue = PBI->getSuccessor(0) == BB; - NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), + NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue), P); } else { NewPN->addIncoming(BI->getCondition(), P); } } - + BI->setCondition(NewPN); return true; } } - + // If this is a conditional branch in an empty block, and if any // predecessors is a conditional branch to one of our destinations, // fold the conditions into logical ops and one cond br. @@ -1991,11 +2246,11 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { if (&*BBI != BI) return false; - + if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) if (CE->canTrap()) return false; - + int PBIOp, BIOp; if (PBI->getSuccessor(0) == BI->getSuccessor(0)) PBIOp = BIOp = 0; @@ -2007,31 +2262,31 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { PBIOp = BIOp = 1; else return false; - + // Check to make sure that the other destination of this branch // isn't BB itself. If so, this is an infinite loop that will // keep getting unwound. if (PBI->getSuccessor(PBIOp) == BB) return false; - - // Do not perform this transformation if it would require + + // Do not perform this transformation if it would require // insertion of a large number of select instructions. For targets // without predication/cmovs, this is a big pessimization. BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); - + unsigned NumPhis = 0; for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II); ++II, ++NumPhis) if (NumPhis > 2) // Disable this xform. return false; - + // Finally, if everything is ok, fold the branches to logical ops. BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); - + DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent() << "AND: " << *BI->getParent()); - - + + // If OtherDest *is* BB, then BB is a basic block with a single conditional // branch in it, where one edge (OtherDest) goes back to itself but the other // exits. We don't *know* that the program avoids the infinite loop @@ -2046,13 +2301,13 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { "infloop", BB->getParent()); BranchInst::Create(InfLoopBlock, InfLoopBlock); OtherDest = InfLoopBlock; - } - + } + DEBUG(dbgs() << *PBI->getParent()->getParent()); // BI may have other predecessors. Because of this, we leave // it alone, but modify PBI. - + // Make sure we get to CommonDest on True&True directions. Value *PBICond = PBI->getCondition(); IRBuilder<true, NoFolder> Builder(PBI); @@ -2065,16 +2320,43 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { // Merge the conditions. Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge"); - + // Modify PBI to branch on the new condition to the new dests. PBI->setCondition(Cond); PBI->setSuccessor(0, CommonDest); PBI->setSuccessor(1, OtherDest); - + + // Update branch weight for PBI. + uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; + bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, + PredFalseWeight); + bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, + SuccFalseWeight); + if (PredHasWeights && SuccHasWeights) { + uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight; + uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight; + uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight; + uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight; + // The weight to CommonDest should be PredCommon * SuccTotal + + // PredOther * SuccCommon. + // The weight to OtherDest should be PredOther * SuccOther. + SmallVector<uint64_t, 2> NewWeights; + NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) + + PredOther * SuccCommon); + NewWeights.push_back(PredOther * SuccOther); + // Halve the weights if any of them cannot fit in an uint32_t + FitWeights(NewWeights); + + SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end()); + PBI->setMetadata(LLVMContext::MD_prof, + MDBuilder(BI->getContext()). + createBranchWeights(MDWeights)); + } + // OtherDest may have phi nodes. If so, add an entry from PBI's // block that are identical to the entries for BI's block. AddPredecessorToBlock(OtherDest, PBI->getParent(), BB); - + // We know that the CommonDest already had an edge from PBI to // it. If it has PHIs though, the PHIs may have different // entries for BB and PBI's BB. If so, insert a select to make @@ -2092,10 +2374,10 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { PN->setIncomingValue(PBBIdx, NV); } } - + DEBUG(dbgs() << "INTO: " << *PBI->getParent()); DEBUG(dbgs() << *PBI->getParent()->getParent()); - + // This basic block is probably dead. We know it has at least // one fewer predecessor. return true; @@ -2107,7 +2389,9 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { // Also makes sure not to introduce new successors by assuming that edges to // non-successor TrueBBs and FalseBBs aren't reachable. static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, - BasicBlock *TrueBB, BasicBlock *FalseBB){ + BasicBlock *TrueBB, BasicBlock *FalseBB, + uint32_t TrueWeight, + uint32_t FalseWeight){ // Remove any superfluous successor edges from the CFG. // First, figure out which successors to preserve. // If TrueBB and FalseBB are equal, only try to preserve one copy of that @@ -2136,10 +2420,15 @@ static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond, // We were only looking for one successor, and it was present. // Create an unconditional branch to it. Builder.CreateBr(TrueBB); - else + else { // We found both of the successors we were looking for. // Create a conditional branch sharing the condition of the select. - Builder.CreateCondBr(Cond, TrueBB, FalseBB); + BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB); + if (TrueWeight != FalseWeight) + NewBI->setMetadata(LLVMContext::MD_prof, + MDBuilder(OldTerm->getContext()). + createBranchWeights(TrueWeight, FalseWeight)); + } } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) { // Neither of the selected blocks were successors, so this // terminator must be unreachable. @@ -2176,8 +2465,23 @@ static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) { BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor(); BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor(); + // Get weight for TrueBB and FalseBB. + uint32_t TrueWeight = 0, FalseWeight = 0; + SmallVector<uint64_t, 8> Weights; + bool HasWeights = HasBranchWeights(SI); + if (HasWeights) { + GetBranchWeights(SI, Weights); + if (Weights.size() == 1 + SI->getNumCases()) { + TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal). + getSuccessorIndex()]; + FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal). + getSuccessorIndex()]; + } + } + // Perform the actual simplification. - return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB); + return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, + TrueWeight, FalseWeight); } // SimplifyIndirectBrOnSelect - Replaces @@ -2197,7 +2501,8 @@ static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { BasicBlock *FalseBB = FBA->getBasicBlock(); // Perform the actual simplification. - return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB); + return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, + 0, 0); } /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp @@ -2214,11 +2519,11 @@ static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) { /// br label %end /// end: /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ] -/// +/// /// We prefer to split the edge to 'end' so that there is a true/false entry to /// the PHI, merging the third icmp into the switch. static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, - const TargetData *TD, + const DataLayout *TD, IRBuilder<> &Builder) { BasicBlock *BB = ICI->getParent(); @@ -2228,17 +2533,17 @@ static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, Value *V = ICI->getOperand(0); ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1)); - + // The pattern we're looking for is where our only predecessor is a switch on // 'V' and this block is the default case for the switch. In this case we can // fold the compared value into the switch to simplify things. BasicBlock *Pred = BB->getSinglePredecessor(); if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false; - + SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator()); if (SI->getCondition() != V) return false; - + // If BB is reachable on a non-default case, then we simply know the value of // V in this block. Substitute it and constant fold the icmp instruction // away. @@ -2246,7 +2551,7 @@ static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, ConstantInt *VVal = SI->findCaseDest(BB); assert(VVal && "Should have a unique destination value"); ICI->setOperand(0, VVal); - + if (Value *V = SimplifyInstruction(ICI, TD)) { ICI->replaceAllUsesWith(V); ICI->eraseFromParent(); @@ -2254,7 +2559,7 @@ static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, // BB is now empty, so it is likely to simplify away. return SimplifyCFG(BB) | true; } - + // Ok, the block is reachable from the default dest. If the constant we're // comparing exists in one of the other edges, then we can constant fold ICI // and zap it. @@ -2264,13 +2569,13 @@ static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, V = ConstantInt::getFalse(BB->getContext()); else V = ConstantInt::getTrue(BB->getContext()); - + ICI->replaceAllUsesWith(V); ICI->eraseFromParent(); // BB is now empty, so it is likely to simplify away. return SimplifyCFG(BB) | true; } - + // The use of the icmp has to be in the 'end' block, by the only PHI node in // the block. BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0); @@ -2296,8 +2601,23 @@ static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, // the switch to the merge point on the compared value. BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB); + SmallVector<uint64_t, 8> Weights; + bool HasWeights = HasBranchWeights(SI); + if (HasWeights) { + GetBranchWeights(SI, Weights); + if (Weights.size() == 1 + SI->getNumCases()) { + // Split weight for default case to case for "Cst". + Weights[0] = (Weights[0]+1) >> 1; + Weights.push_back(Weights[0]); + + SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); + SI->setMetadata(LLVMContext::MD_prof, + MDBuilder(SI->getContext()). + createBranchWeights(MDWeights)); + } + } SI->addCase(Cst, NewBB); - + // NewBB branches to the phi block, add the uncond branch and the phi entry. Builder.SetInsertPoint(NewBB); Builder.SetCurrentDebugLocation(SI->getDebugLoc()); @@ -2309,12 +2629,12 @@ static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI, /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch. /// Check to see if it is branching on an or/and chain of icmp instructions, and /// fold it into a switch instruction if so. -static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, +static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD, IRBuilder<> &Builder) { Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); if (Cond == 0) return false; - - + + // Change br (X == 0 | X == 1), T, F into a switch instruction. // If this is a bunch of seteq's or'd together, or if it's a bunch of // 'setne's and'ed together, collect them. @@ -2323,7 +2643,7 @@ static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, bool TrueWhenEqual = true; Value *ExtraCase = 0; unsigned UsedICmps = 0; - + if (Cond->getOpcode() == Instruction::Or) { CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true, UsedICmps); @@ -2332,7 +2652,7 @@ static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, UsedICmps); TrueWhenEqual = false; } - + // If we didn't have a multiply compared value, fail. if (CompVal == 0) return false; @@ -2344,21 +2664,24 @@ static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, // instruction can't handle, remove them now. array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate); Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); - + // If Extra was used, we require at least two switch values to do the // transformation. A switch with one value is just an cond branch. if (ExtraCase && Values.size() < 2) return false; - + + // TODO: Preserve branch weight metadata, similarly to how + // FoldValueComparisonIntoPredecessors preserves it. + // Figure out which block is which destination. BasicBlock *DefaultBB = BI->getSuccessor(1); BasicBlock *EdgeBB = BI->getSuccessor(0); if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); - + BasicBlock *BB = BI->getParent(); - + DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size() << " cases into SWITCH. BB is:\n" << *BB); - + // If there are any extra values that couldn't be folded into the switch // then we evaluate them with an explicit branch first. Split the block // right before the condbr to handle it. @@ -2372,13 +2695,13 @@ static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB); else Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB); - + OldTI->eraseFromParent(); - + // If there are PHI nodes in EdgeBB, then we need to add a new entry to them // for the edge we just added. AddPredecessorToBlock(EdgeBB, BB, NewBB); - + DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase << "\nEXTRABB = " << *BB); BB = NewBB; @@ -2387,19 +2710,19 @@ static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, Builder.SetInsertPoint(BI); // Convert pointer to int before we switch. if (CompVal->getType()->isPointerTy()) { - assert(TD && "Cannot switch on pointer without TargetData"); + assert(TD && "Cannot switch on pointer without DataLayout"); CompVal = Builder.CreatePtrToInt(CompVal, TD->getIntPtrType(CompVal->getContext()), "magicptr"); } - + // Create the new switch instruction now. SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size()); // Add all of the 'cases' to the switch instruction. for (unsigned i = 0, e = Values.size(); i != e; ++i) New->addCase(Values[i], EdgeBB); - + // We added edges from PI to the EdgeBB. As such, if there were any // PHI nodes in EdgeBB, they need entries to be added corresponding to // the number of edges added. @@ -2410,10 +2733,10 @@ static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD, for (unsigned i = 0, e = Values.size()-1; i != e; ++i) PN->addIncoming(InVal, BB); } - + // Erase the old branch instruction. EraseTerminatorInstAndDCECond(BI); - + DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n'); return true; } @@ -2467,7 +2790,7 @@ bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) { bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { BasicBlock *BB = RI->getParent(); if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false; - + // Find predecessors that end with branches. SmallVector<BasicBlock*, 8> UncondBranchPreds; SmallVector<BranchInst*, 8> CondBranchPreds; @@ -2481,7 +2804,7 @@ bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { CondBranchPreds.push_back(BI); } } - + // If we found some, do the transformation! if (!UncondBranchPreds.empty() && DupRet) { while (!UncondBranchPreds.empty()) { @@ -2490,21 +2813,21 @@ bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { << "INTO UNCOND BRANCH PRED: " << *Pred); (void)FoldReturnIntoUncondBranch(RI, BB, Pred); } - + // If we eliminated all predecessors of the block, delete the block now. if (pred_begin(BB) == pred_end(BB)) // We know there are no successors, so just nuke the block. BB->eraseFromParent(); - + return true; } - + // Check out all of the conditional branches going to this return // instruction. If any of them just select between returns, change the // branch itself into a select/return pair. while (!CondBranchPreds.empty()) { BranchInst *BI = CondBranchPreds.pop_back_val(); - + // Check to see if the non-BB successor is also a return block. if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && @@ -2516,9 +2839,9 @@ bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) { bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { BasicBlock *BB = UI->getParent(); - + bool Changed = false; - + // If there are any instructions immediately before the unreachable that can // be removed, do so. while (UI != BB->begin()) { @@ -2558,11 +2881,11 @@ bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { BBI->eraseFromParent(); Changed = true; } - + // If the unreachable instruction is the first in the block, take a gander // at all of the predecessors of this instruction, and simplify them. if (&BB->front() != UI) return Changed; - + SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); for (unsigned i = 0, e = Preds.size(); i != e; ++i) { TerminatorInst *TI = Preds[i]->getTerminator(); @@ -2615,7 +2938,7 @@ bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { BasicBlock *MaxBlock = 0; for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { - if (I->second.first > MaxPop || + if (I->second.first > MaxPop || (I->second.first == MaxPop && MaxIndex > I->second.second)) { MaxPop = I->second.first; MaxIndex = I->second.second; @@ -2627,13 +2950,13 @@ bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { // edges to it. SI->setDefaultDest(MaxBlock); Changed = true; - + // If MaxBlock has phinodes in it, remove MaxPop-1 entries from // it. if (isa<PHINode>(MaxBlock->begin())) for (unsigned i = 0; i != MaxPop-1; ++i) MaxBlock->removePredecessor(SI->getParent()); - + for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i) if (i.getCaseSuccessor() == MaxBlock) { @@ -2648,7 +2971,7 @@ bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { // place to note that the call does not throw though. BranchInst *BI = Builder.CreateBr(II->getNormalDest()); II->removeFromParent(); // Take out of symbol table - + // Insert the call now... SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3); Builder.SetInsertPoint(BI); @@ -2663,7 +2986,7 @@ bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) { } } } - + // If this block is now dead, remove it. if (pred_begin(BB) == pred_end(BB) && BB != &BB->getParent()->getEntryBlock()) { @@ -2706,9 +3029,28 @@ static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { if (!Offset->isNullValue()) Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); - Builder.CreateCondBr( + BranchInst *NewBI = Builder.CreateCondBr( Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); + // Update weight for the newly-created conditional branch. + SmallVector<uint64_t, 8> Weights; + bool HasWeights = HasBranchWeights(SI); + if (HasWeights) { + GetBranchWeights(SI, Weights); + if (Weights.size() == 1 + SI->getNumCases()) { + // Combine all weights for the cases to be the true weight of NewBI. + // We assume that the sum of all weights for a Terminator can fit into 32 + // bits. + uint32_t NewTrueWeight = 0; + for (unsigned I = 1, E = Weights.size(); I != E; ++I) + NewTrueWeight += (uint32_t)Weights[I]; + NewBI->setMetadata(LLVMContext::MD_prof, + MDBuilder(SI->getContext()). + createBranchWeights(NewTrueWeight, + (uint32_t)Weights[0])); + } + } + // Prune obsolete incoming values off the successor's PHI nodes. for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin(); isa<PHINode>(BBI); ++BBI) { @@ -2739,15 +3081,33 @@ static bool EliminateDeadSwitchCases(SwitchInst *SI) { } } + SmallVector<uint64_t, 8> Weights; + bool HasWeight = HasBranchWeights(SI); + if (HasWeight) { + GetBranchWeights(SI, Weights); + HasWeight = (Weights.size() == 1 + SI->getNumCases()); + } + // Remove dead cases from the switch. for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) { SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]); assert(Case != SI->case_default() && "Case was not found. Probably mistake in DeadCases forming."); + if (HasWeight) { + std::swap(Weights[Case.getCaseIndex()+1], Weights.back()); + Weights.pop_back(); + } + // Prune unused values from PHI nodes. Case.getCaseSuccessor()->removePredecessor(SI->getParent()); SI->removeCase(Case); } + if (HasWeight) { + SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end()); + SI->setMetadata(LLVMContext::MD_prof, + MDBuilder(SI->getParent()->getContext()). + createBranchWeights(MDWeights)); + } return !DeadCases.empty(); } @@ -2823,33 +3183,512 @@ static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { return Changed; } -bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { - // If this switch is too complex to want to look at, ignore it. - if (!isValueEqualityComparison(SI)) +/// ValidLookupTableConstant - Return true if the backend will be able to handle +/// initializing an array of constants like C. +static bool ValidLookupTableConstant(Constant *C) { + if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) + return CE->isGEPWithNoNotionalOverIndexing(); + + return isa<ConstantFP>(C) || + isa<ConstantInt>(C) || + isa<ConstantPointerNull>(C) || + isa<GlobalValue>(C) || + isa<UndefValue>(C); +} + +/// LookupConstant - If V is a Constant, return it. Otherwise, try to look up +/// its constant value in ConstantPool, returning 0 if it's not there. +static Constant *LookupConstant(Value *V, + const SmallDenseMap<Value*, Constant*>& ConstantPool) { + if (Constant *C = dyn_cast<Constant>(V)) + return C; + return ConstantPool.lookup(V); +} + +/// ConstantFold - Try to fold instruction I into a constant. This works for +/// simple instructions such as binary operations where both operands are +/// constant or can be replaced by constants from the ConstantPool. Returns the +/// resulting constant on success, 0 otherwise. +static Constant *ConstantFold(Instruction *I, + const SmallDenseMap<Value*, Constant*>& ConstantPool) { + if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { + Constant *A = LookupConstant(BO->getOperand(0), ConstantPool); + if (!A) + return 0; + Constant *B = LookupConstant(BO->getOperand(1), ConstantPool); + if (!B) + return 0; + return ConstantExpr::get(BO->getOpcode(), A, B); + } + + if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { + Constant *A = LookupConstant(I->getOperand(0), ConstantPool); + if (!A) + return 0; + Constant *B = LookupConstant(I->getOperand(1), ConstantPool); + if (!B) + return 0; + return ConstantExpr::getCompare(Cmp->getPredicate(), A, B); + } + + if (SelectInst *Select = dyn_cast<SelectInst>(I)) { + Constant *A = LookupConstant(Select->getCondition(), ConstantPool); + if (!A) + return 0; + if (A->isAllOnesValue()) + return LookupConstant(Select->getTrueValue(), ConstantPool); + if (A->isNullValue()) + return LookupConstant(Select->getFalseValue(), ConstantPool); + return 0; + } + + if (CastInst *Cast = dyn_cast<CastInst>(I)) { + Constant *A = LookupConstant(I->getOperand(0), ConstantPool); + if (!A) + return 0; + return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy()); + } + + return 0; +} + +/// GetCaseResults - Try to determine the resulting constant values in phi nodes +/// at the common destination basic block, *CommonDest, for one of the case +/// destionations CaseDest corresponding to value CaseVal (0 for the default +/// case), of a switch instruction SI. +static bool GetCaseResults(SwitchInst *SI, + ConstantInt *CaseVal, + BasicBlock *CaseDest, + BasicBlock **CommonDest, + SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { + // The block from which we enter the common destination. + BasicBlock *Pred = SI->getParent(); + + // If CaseDest is empty except for some side-effect free instructions through + // which we can constant-propagate the CaseVal, continue to its successor. + SmallDenseMap<Value*, Constant*> ConstantPool; + ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal)); + for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E; + ++I) { + if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) { + // If the terminator is a simple branch, continue to the next block. + if (T->getNumSuccessors() != 1) + return false; + Pred = CaseDest; + CaseDest = T->getSuccessor(0); + } else if (isa<DbgInfoIntrinsic>(I)) { + // Skip debug intrinsic. + continue; + } else if (Constant *C = ConstantFold(I, ConstantPool)) { + // Instruction is side-effect free and constant. + ConstantPool.insert(std::make_pair(I, C)); + } else { + break; + } + } + + // If we did not have a CommonDest before, use the current one. + if (!*CommonDest) + *CommonDest = CaseDest; + // If the destination isn't the common one, abort. + if (CaseDest != *CommonDest) + return false; + + // Get the values for this case from phi nodes in the destination block. + BasicBlock::iterator I = (*CommonDest)->begin(); + while (PHINode *PHI = dyn_cast<PHINode>(I++)) { + int Idx = PHI->getBasicBlockIndex(Pred); + if (Idx == -1) + continue; + + Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx), + ConstantPool); + if (!ConstVal) + return false; + + // Note: If the constant comes from constant-propagating the case value + // through the CaseDest basic block, it will be safe to remove the + // instructions in that block. They cannot be used (except in the phi nodes + // we visit) outside CaseDest, because that block does not dominate its + // successor. If it did, we would not be in this phi node. + + // Be conservative about which kinds of constants we support. + if (!ValidLookupTableConstant(ConstVal)) + return false; + + Res.push_back(std::make_pair(PHI, ConstVal)); + } + + return true; +} + +namespace { + /// SwitchLookupTable - This class represents a lookup table that can be used + /// to replace a switch. + class SwitchLookupTable { + public: + /// SwitchLookupTable - Create a lookup table to use as a switch replacement + /// with the contents of Values, using DefaultValue to fill any holes in the + /// table. + SwitchLookupTable(Module &M, + uint64_t TableSize, + ConstantInt *Offset, + const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, + Constant *DefaultValue, + const DataLayout *TD); + + /// BuildLookup - Build instructions with Builder to retrieve the value at + /// the position given by Index in the lookup table. + Value *BuildLookup(Value *Index, IRBuilder<> &Builder); + + /// WouldFitInRegister - Return true if a table with TableSize elements of + /// type ElementType would fit in a target-legal register. + static bool WouldFitInRegister(const DataLayout *TD, + uint64_t TableSize, + const Type *ElementType); + + private: + // Depending on the contents of the table, it can be represented in + // different ways. + enum { + // For tables where each element contains the same value, we just have to + // store that single value and return it for each lookup. + SingleValueKind, + + // For small tables with integer elements, we can pack them into a bitmap + // that fits into a target-legal register. Values are retrieved by + // shift and mask operations. + BitMapKind, + + // The table is stored as an array of values. Values are retrieved by load + // instructions from the table. + ArrayKind + } Kind; + + // For SingleValueKind, this is the single value. + Constant *SingleValue; + + // For BitMapKind, this is the bitmap. + ConstantInt *BitMap; + IntegerType *BitMapElementTy; + + // For ArrayKind, this is the array. + GlobalVariable *Array; + }; +} + +SwitchLookupTable::SwitchLookupTable(Module &M, + uint64_t TableSize, + ConstantInt *Offset, + const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, + Constant *DefaultValue, + const DataLayout *TD) { + assert(Values.size() && "Can't build lookup table without values!"); + assert(TableSize >= Values.size() && "Can't fit values in table!"); + + // If all values in the table are equal, this is that value. + SingleValue = Values.begin()->second; + + // Build up the table contents. + SmallVector<Constant*, 64> TableContents(TableSize); + for (size_t I = 0, E = Values.size(); I != E; ++I) { + ConstantInt *CaseVal = Values[I].first; + Constant *CaseRes = Values[I].second; + assert(CaseRes->getType() == DefaultValue->getType()); + + uint64_t Idx = (CaseVal->getValue() - Offset->getValue()) + .getLimitedValue(); + TableContents[Idx] = CaseRes; + + if (CaseRes != SingleValue) + SingleValue = 0; + } + + // Fill in any holes in the table with the default result. + if (Values.size() < TableSize) { + for (uint64_t I = 0; I < TableSize; ++I) { + if (!TableContents[I]) + TableContents[I] = DefaultValue; + } + + if (DefaultValue != SingleValue) + SingleValue = 0; + } + + // If each element in the table contains the same value, we only need to store + // that single value. + if (SingleValue) { + Kind = SingleValueKind; + return; + } + + // If the type is integer and the table fits in a register, build a bitmap. + if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) { + IntegerType *IT = cast<IntegerType>(DefaultValue->getType()); + APInt TableInt(TableSize * IT->getBitWidth(), 0); + for (uint64_t I = TableSize; I > 0; --I) { + TableInt <<= IT->getBitWidth(); + // Insert values into the bitmap. Undef values are set to zero. + if (!isa<UndefValue>(TableContents[I - 1])) { + ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]); + TableInt |= Val->getValue().zext(TableInt.getBitWidth()); + } + } + BitMap = ConstantInt::get(M.getContext(), TableInt); + BitMapElementTy = IT; + Kind = BitMapKind; + ++NumBitMaps; + return; + } + + // Store the table in an array. + ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize); + Constant *Initializer = ConstantArray::get(ArrayTy, TableContents); + + Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true, + GlobalVariable::PrivateLinkage, + Initializer, + "switch.table"); + Array->setUnnamedAddr(true); + Kind = ArrayKind; +} + +Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) { + switch (Kind) { + case SingleValueKind: + return SingleValue; + case BitMapKind: { + // Type of the bitmap (e.g. i59). + IntegerType *MapTy = BitMap->getType(); + + // Cast Index to the same type as the bitmap. + // Note: The Index is <= the number of elements in the table, so + // truncating it to the width of the bitmask is safe. + Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast"); + + // Multiply the shift amount by the element width. + ShiftAmt = Builder.CreateMul(ShiftAmt, + ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()), + "switch.shiftamt"); + + // Shift down. + Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt, + "switch.downshift"); + // Mask off. + return Builder.CreateTrunc(DownShifted, BitMapElementTy, + "switch.masked"); + } + case ArrayKind: { + Value *GEPIndices[] = { Builder.getInt32(0), Index }; + Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices, + "switch.gep"); + return Builder.CreateLoad(GEP, "switch.load"); + } + } + llvm_unreachable("Unknown lookup table kind!"); +} + +bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, + uint64_t TableSize, + const Type *ElementType) { + if (!TD) + return false; + const IntegerType *IT = dyn_cast<IntegerType>(ElementType); + if (!IT) + return false; + // FIXME: If the type is wider than it needs to be, e.g. i8 but all values + // are <= 15, we could try to narrow the type. + + // Avoid overflow, fitsInLegalInteger uses unsigned int for the width. + if (TableSize >= UINT_MAX/IT->getBitWidth()) + return false; + return TD->fitsInLegalInteger(TableSize * IT->getBitWidth()); +} + +/// ShouldBuildLookupTable - Determine whether a lookup table should be built +/// for this switch, based on the number of caes, size of the table and the +/// types of the results. +static bool ShouldBuildLookupTable(SwitchInst *SI, + uint64_t TableSize, + const DataLayout *TD, + const SmallDenseMap<PHINode*, Type*>& ResultTypes) { + // The table density should be at least 40%. This is the same criterion as for + // jump tables, see SelectionDAGBuilder::handleJTSwitchCase. + // FIXME: Find the best cut-off. + if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10) + return false; // TableSize overflowed, or mul below might overflow. + if (SI->getNumCases() * 10 >= TableSize * 4) + return true; + + // If each table would fit in a register, we should build it anyway. + for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(), + E = ResultTypes.end(); I != E; ++I) { + if (!SwitchLookupTable::WouldFitInRegister(TD, TableSize, I->second)) + return false; + } + return true; +} + +/// SwitchToLookupTable - If the switch is only used to initialize one or more +/// phi nodes in a common successor block with different constant values, +/// replace the switch with lookup tables. +static bool SwitchToLookupTable(SwitchInst *SI, + IRBuilder<> &Builder, + const DataLayout* TD, + const TargetTransformInfo *TTI) { + assert(SI->getNumCases() > 1 && "Degenerate switch?"); + + // Only build lookup table when we have a target that supports it. + if (!TTI || !TTI->getScalarTargetTransformInfo() || + !TTI->getScalarTargetTransformInfo()->shouldBuildLookupTables()) return false; + // FIXME: If the switch is too sparse for a lookup table, perhaps we could + // split off a dense part and build a lookup table for that. + + // FIXME: This creates arrays of GEPs to constant strings, which means each + // GEP needs a runtime relocation in PIC code. We should just build one big + // string and lookup indices into that. + + // Ignore the switch if the number of cases is too small. + // This is similar to the check when building jump tables in + // SelectionDAGBuilder::handleJTSwitchCase. + // FIXME: Determine the best cut-off. + if (SI->getNumCases() < 4) + return false; + + // Figure out the corresponding result for each case value and phi node in the + // common destination, as well as the the min and max case values. + assert(SI->case_begin() != SI->case_end()); + SwitchInst::CaseIt CI = SI->case_begin(); + ConstantInt *MinCaseVal = CI.getCaseValue(); + ConstantInt *MaxCaseVal = CI.getCaseValue(); + + BasicBlock *CommonDest = 0; + typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy; + SmallDenseMap<PHINode*, ResultListTy> ResultLists; + SmallDenseMap<PHINode*, Constant*> DefaultResults; + SmallDenseMap<PHINode*, Type*> ResultTypes; + SmallVector<PHINode*, 4> PHIs; + + for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) { + ConstantInt *CaseVal = CI.getCaseValue(); + if (CaseVal->getValue().slt(MinCaseVal->getValue())) + MinCaseVal = CaseVal; + if (CaseVal->getValue().sgt(MaxCaseVal->getValue())) + MaxCaseVal = CaseVal; + + // Resulting value at phi nodes for this case value. + typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy; + ResultsTy Results; + if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest, + Results)) + return false; + + // Append the result from this case to the list for each phi. + for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) { + if (!ResultLists.count(I->first)) + PHIs.push_back(I->first); + ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second)); + } + } + + // Get the resulting values for the default case. + SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList; + if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest, + DefaultResultsList)) + return false; + for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) { + PHINode *PHI = DefaultResultsList[I].first; + Constant *Result = DefaultResultsList[I].second; + DefaultResults[PHI] = Result; + ResultTypes[PHI] = Result->getType(); + } + + APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue(); + uint64_t TableSize = RangeSpread.getLimitedValue() + 1; + if (!ShouldBuildLookupTable(SI, TableSize, TD, ResultTypes)) + return false; + + // Create the BB that does the lookups. + Module &Mod = *CommonDest->getParent()->getParent(); + BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(), + "switch.lookup", + CommonDest->getParent(), + CommonDest); + + // Check whether the condition value is within the case range, and branch to + // the new BB. + Builder.SetInsertPoint(SI); + Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal, + "switch.tableidx"); + Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get( + MinCaseVal->getType(), TableSize)); + Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest()); + + // Populate the BB that does the lookups. + Builder.SetInsertPoint(LookupBB); + bool ReturnedEarly = false; + for (size_t I = 0, E = PHIs.size(); I != E; ++I) { + PHINode *PHI = PHIs[I]; + + SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI], + DefaultResults[PHI], TD); + + Value *Result = Table.BuildLookup(TableIndex, Builder); + + // If the result is used to return immediately from the function, we want to + // do that right here. + if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) && + *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) { + Builder.CreateRet(Result); + ReturnedEarly = true; + break; + } + + PHI->addIncoming(Result, LookupBB); + } + + if (!ReturnedEarly) + Builder.CreateBr(CommonDest); + + // Remove the switch. + for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) { + BasicBlock *Succ = SI->getSuccessor(i); + if (Succ == SI->getDefaultDest()) continue; + Succ->removePredecessor(SI->getParent()); + } + SI->eraseFromParent(); + + ++NumLookupTables; + return true; +} + +bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { BasicBlock *BB = SI->getParent(); - // If we only have one predecessor, and if it is a branch on this value, - // see if that predecessor totally determines the outcome of this switch. - if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) - if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) - return SimplifyCFG(BB) | true; + if (isValueEqualityComparison(SI)) { + // If we only have one predecessor, and if it is a branch on this value, + // see if that predecessor totally determines the outcome of this switch. + if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) + if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder)) + return SimplifyCFG(BB) | true; - Value *Cond = SI->getCondition(); - if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) - if (SimplifySwitchOnSelect(SI, Select)) - return SimplifyCFG(BB) | true; + Value *Cond = SI->getCondition(); + if (SelectInst *Select = dyn_cast<SelectInst>(Cond)) + if (SimplifySwitchOnSelect(SI, Select)) + return SimplifyCFG(BB) | true; - // If the block only contains the switch, see if we can fold the block - // away into any preds. - BasicBlock::iterator BBI = BB->begin(); - // Ignore dbg intrinsics. - while (isa<DbgInfoIntrinsic>(BBI)) - ++BBI; - if (SI == &*BBI) - if (FoldValueComparisonIntoPredecessors(SI, Builder)) - return SimplifyCFG(BB) | true; + // If the block only contains the switch, see if we can fold the block + // away into any preds. + BasicBlock::iterator BBI = BB->begin(); + // Ignore dbg intrinsics. + while (isa<DbgInfoIntrinsic>(BBI)) + ++BBI; + if (SI == &*BBI) + if (FoldValueComparisonIntoPredecessors(SI, Builder)) + return SimplifyCFG(BB) | true; + } // Try to transform the switch into an icmp and a branch. if (TurnSwitchRangeIntoICmp(SI, Builder)) @@ -2862,13 +3701,16 @@ bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) { if (ForwardSwitchConditionToPHI(SI)) return SimplifyCFG(BB) | true; + if (SwitchToLookupTable(SI, Builder, TD, TTI)) + return SimplifyCFG(BB) | true; + return false; } bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { BasicBlock *BB = IBI->getParent(); bool Changed = false; - + // Eliminate redundant destinations. SmallPtrSet<Value *, 8> Succs; for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { @@ -2879,7 +3721,7 @@ bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { --i; --e; Changed = true; } - } + } if (IBI->getNumDestinations() == 0) { // If the indirectbr has no successors, change it to unreachable. @@ -2887,14 +3729,14 @@ bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { EraseTerminatorInstAndDCECond(IBI); return true; } - + if (IBI->getNumDestinations() == 1) { // If the indirectbr has one successor, change it to a direct branch. BranchInst::Create(IBI->getDestination(0), IBI); EraseTerminatorInstAndDCECond(IBI); return true; } - + if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) { if (SimplifyIndirectBrOnSelect(IBI, SI)) return SimplifyCFG(BB) | true; @@ -2904,13 +3746,16 @@ bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) { bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ BasicBlock *BB = BI->getParent(); - + + if (SinkCommon && SinkThenElseCodeToEnd(BI)) + return true; + // If the Terminator is the only non-phi instruction, simplify the block. BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime(); if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() && TryToSimplifyUncondBranchFromEmptyBlock(BB)) return true; - + // If the only instruction in the block is a seteq/setne comparison // against a constant, try to simplify the block. if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) @@ -2921,7 +3766,7 @@ bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder)) return true; } - + // If this basic block is ONLY a compare and a branch, and if a predecessor // branches to us and our successor, fold the comparison into the // predecessor and use logical operations to update the incoming value @@ -2934,7 +3779,7 @@ bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){ bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { BasicBlock *BB = BI->getParent(); - + // Conditional branch if (isValueEqualityComparison(BI)) { // If we only have one predecessor, and if it is a branch on this value, @@ -2943,7 +3788,7 @@ bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder)) return SimplifyCFG(BB) | true; - + // This block must be empty, except for the setcond inst, if it exists. // Ignore dbg intrinsics. BasicBlock::iterator I = BB->begin(); @@ -2962,17 +3807,17 @@ bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { return SimplifyCFG(BB) | true; } } - + // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction. if (SimplifyBranchOnICmpChain(BI, TD, Builder)) return true; - + // If this basic block is ONLY a compare and a branch, and if a predecessor // branches to us and one of our successors, fold the comparison into the // predecessor and use logical operations to pick the right destination. if (FoldBranchToCommonDest(BI)) return SimplifyCFG(BB) | true; - + // We have a conditional branch to two blocks that are only reachable // from BI. We know that the condbr dominates the two blocks, so see if // there is any identical code in the "then" and "else" blocks. If so, we @@ -2999,14 +3844,14 @@ bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) { if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1))) return SimplifyCFG(BB) | true; } - + // If this is a branch on a phi node in the current block, thread control // through this block if any PHI node entries are constants. if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) if (PN->getParent() == BI->getParent()) if (FoldCondBranchOnPHI(BI, TD)) return SimplifyCFG(BB) | true; - + // Scan predecessor blocks for conditional branches. for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) @@ -3023,11 +3868,12 @@ static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) { if (!C) return false; - if (!I->hasOneUse()) // Only look at single-use instructions, for compile time + if (I->use_empty()) return false; if (C->isNullValue()) { - Instruction *Use = I->use_back(); + // Only look at the first use, avoid hurting compile time with long uselists + User *Use = *I->use_begin(); // Now make sure that there are no instructions in between that can alter // control flow (eg. calls) @@ -3114,7 +3960,7 @@ bool SimplifyCFGOpt::run(BasicBlock *BB) { // if (MergeBlockIntoPredecessor(BB)) return true; - + IRBuilder<> Builder(BB); // If there is a trivial two-entry PHI node in this basic block, and we can @@ -3152,6 +3998,7 @@ bool SimplifyCFGOpt::run(BasicBlock *BB) { /// eliminates unreachable basic blocks, and does other "peephole" optimization /// of the CFG. It returns true if a modification was made. /// -bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { - return SimplifyCFGOpt(TD).run(BB); +bool llvm::SimplifyCFG(BasicBlock *BB, const DataLayout *TD, + const TargetTransformInfo *TTI) { + return SimplifyCFGOpt(TD, TTI).run(BB); } diff --git a/contrib/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp b/contrib/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp index 5d673f1..110f3808 100644 --- a/contrib/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp +++ b/contrib/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp @@ -24,7 +24,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/SimplifyIndVar.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" @@ -44,7 +44,7 @@ namespace { Loop *L; LoopInfo *LI; ScalarEvolution *SE; - const TargetData *TD; // May be NULL + const DataLayout *TD; // May be NULL SmallVectorImpl<WeakVH> &DeadInsts; @@ -56,7 +56,7 @@ namespace { L(Loop), LI(LPM->getAnalysisIfAvailable<LoopInfo>()), SE(SE), - TD(LPM->getAnalysisIfAvailable<TargetData>()), + TD(LPM->getAnalysisIfAvailable<DataLayout>()), DeadInsts(Dead), Changed(false) { assert(LI && "IV simplification requires LoopInfo"); diff --git a/contrib/llvm/lib/Transforms/Utils/SimplifyInstructions.cpp b/contrib/llvm/lib/Transforms/Utils/SimplifyInstructions.cpp index 81eb9e0..65353dc 100644 --- a/contrib/llvm/lib/Transforms/Utils/SimplifyInstructions.cpp +++ b/contrib/llvm/lib/Transforms/Utils/SimplifyInstructions.cpp @@ -23,7 +23,7 @@ #include "llvm/ADT/Statistic.h" #include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/InstructionSimplify.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Scalar.h" #include "llvm/Transforms/Utils/Local.h" @@ -46,7 +46,7 @@ namespace { /// runOnFunction - Remove instructions that simplify. bool runOnFunction(Function &F) { const DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>(); - const TargetData *TD = getAnalysisIfAvailable<TargetData>(); + const DataLayout *TD = getAnalysisIfAvailable<DataLayout>(); const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>(); SmallPtrSet<const Instruction*, 8> S1, S2, *ToSimplify = &S1, *Next = &S2; bool Changed = false; @@ -72,7 +72,7 @@ namespace { ++NumSimplified; Changed = true; } - Changed |= RecursivelyDeleteTriviallyDeadInstructions(I); + Changed |= RecursivelyDeleteTriviallyDeadInstructions(I, TLI); } // Place the list of instructions to simplify on the next loop iteration diff --git a/contrib/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/contrib/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp new file mode 100644 index 0000000..c3ea638 --- /dev/null +++ b/contrib/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp @@ -0,0 +1,1149 @@ +//===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This is a utility pass used for testing the InstructionSimplify analysis. +// The analysis is applied to every instruction, and if it simplifies then the +// instruction is replaced by the simplification. If you are looking for a pass +// that performs serious instruction folding, use the instcombine pass instead. +// +//===----------------------------------------------------------------------===// + +#include "llvm/Transforms/Utils/SimplifyLibCalls.h" +#include "llvm/DataLayout.h" +#include "llvm/ADT/StringMap.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/Function.h" +#include "llvm/IRBuilder.h" +#include "llvm/LLVMContext.h" +#include "llvm/Target/TargetLibraryInfo.h" +#include "llvm/Transforms/Utils/BuildLibCalls.h" + +using namespace llvm; + +/// This class is the abstract base class for the set of optimizations that +/// corresponds to one library call. +namespace { +class LibCallOptimization { +protected: + Function *Caller; + const DataLayout *TD; + const TargetLibraryInfo *TLI; + const LibCallSimplifier *LCS; + LLVMContext* Context; +public: + LibCallOptimization() { } + virtual ~LibCallOptimization() {} + + /// callOptimizer - This pure virtual method is implemented by base classes to + /// do various optimizations. If this returns null then no transformation was + /// performed. If it returns CI, then it transformed the call and CI is to be + /// deleted. If it returns something else, replace CI with the new value and + /// delete CI. + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) + =0; + + Value *optimizeCall(CallInst *CI, const DataLayout *TD, + const TargetLibraryInfo *TLI, + const LibCallSimplifier *LCS, IRBuilder<> &B) { + Caller = CI->getParent()->getParent(); + this->TD = TD; + this->TLI = TLI; + this->LCS = LCS; + if (CI->getCalledFunction()) + Context = &CI->getCalledFunction()->getContext(); + + // We never change the calling convention. + if (CI->getCallingConv() != llvm::CallingConv::C) + return NULL; + + return callOptimizer(CI->getCalledFunction(), CI, B); + } +}; + +//===----------------------------------------------------------------------===// +// Helper Functions +//===----------------------------------------------------------------------===// + +/// isOnlyUsedInZeroEqualityComparison - Return true if it only matters that the +/// value is equal or not-equal to zero. +static bool isOnlyUsedInZeroEqualityComparison(Value *V) { + for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); + UI != E; ++UI) { + if (ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) + if (IC->isEquality()) + if (Constant *C = dyn_cast<Constant>(IC->getOperand(1))) + if (C->isNullValue()) + continue; + // Unknown instruction. + return false; + } + return true; +} + +/// isOnlyUsedInEqualityComparison - Return true if it is only used in equality +/// comparisons with With. +static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) { + for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); + UI != E; ++UI) { + if (ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) + if (IC->isEquality() && IC->getOperand(1) == With) + continue; + // Unknown instruction. + return false; + } + return true; +} + +//===----------------------------------------------------------------------===// +// Fortified Library Call Optimizations +//===----------------------------------------------------------------------===// + +struct FortifiedLibCallOptimization : public LibCallOptimization { +protected: + virtual bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, + bool isString) const = 0; +}; + +struct InstFortifiedLibCallOptimization : public FortifiedLibCallOptimization { + CallInst *CI; + + bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { + if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) + return true; + if (ConstantInt *SizeCI = + dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { + if (SizeCI->isAllOnesValue()) + return true; + if (isString) { + uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); + // If the length is 0 we don't know how long it is and so we can't + // remove the check. + if (Len == 0) return false; + return SizeCI->getZExtValue() >= Len; + } + if (ConstantInt *Arg = dyn_cast<ConstantInt>( + CI->getArgOperand(SizeArgOp))) + return SizeCI->getZExtValue() >= Arg->getZExtValue(); + } + return false; + } +}; + +struct MemCpyChkOpt : public InstFortifiedLibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + this->CI = CI; + FunctionType *FT = Callee->getFunctionType(); + LLVMContext &Context = CI->getParent()->getContext(); + + // Check if this has the right signature. + if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy() || + FT->getParamType(2) != TD->getIntPtrType(Context) || + FT->getParamType(3) != TD->getIntPtrType(Context)) + return 0; + + if (isFoldable(3, 2, false)) { + B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1), + CI->getArgOperand(2), 1); + return CI->getArgOperand(0); + } + return 0; + } +}; + +struct MemMoveChkOpt : public InstFortifiedLibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + this->CI = CI; + FunctionType *FT = Callee->getFunctionType(); + LLVMContext &Context = CI->getParent()->getContext(); + + // Check if this has the right signature. + if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy() || + FT->getParamType(2) != TD->getIntPtrType(Context) || + FT->getParamType(3) != TD->getIntPtrType(Context)) + return 0; + + if (isFoldable(3, 2, false)) { + B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1), + CI->getArgOperand(2), 1); + return CI->getArgOperand(0); + } + return 0; + } +}; + +struct MemSetChkOpt : public InstFortifiedLibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + this->CI = CI; + FunctionType *FT = Callee->getFunctionType(); + LLVMContext &Context = CI->getParent()->getContext(); + + // Check if this has the right signature. + if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isIntegerTy() || + FT->getParamType(2) != TD->getIntPtrType(Context) || + FT->getParamType(3) != TD->getIntPtrType(Context)) + return 0; + + if (isFoldable(3, 2, false)) { + Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), + false); + B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1); + return CI->getArgOperand(0); + } + return 0; + } +}; + +struct StrCpyChkOpt : public InstFortifiedLibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + this->CI = CI; + StringRef Name = Callee->getName(); + FunctionType *FT = Callee->getFunctionType(); + LLVMContext &Context = CI->getParent()->getContext(); + + // Check if this has the right signature. + if (FT->getNumParams() != 3 || + FT->getReturnType() != FT->getParamType(0) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != Type::getInt8PtrTy(Context) || + FT->getParamType(2) != TD->getIntPtrType(Context)) + return 0; + + Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); + if (Dst == Src) // __strcpy_chk(x,x) -> x + return Src; + + // If a) we don't have any length information, or b) we know this will + // fit then just lower to a plain strcpy. Otherwise we'll keep our + // strcpy_chk call which may fail at runtime if the size is too long. + // TODO: It might be nice to get a maximum length out of the possible + // string lengths for varying. + if (isFoldable(2, 1, true)) { + Value *Ret = EmitStrCpy(Dst, Src, B, TD, TLI, Name.substr(2, 6)); + return Ret; + } else { + // Maybe we can stil fold __strcpy_chk to __memcpy_chk. + uint64_t Len = GetStringLength(Src); + if (Len == 0) return 0; + + // This optimization require DataLayout. + if (!TD) return 0; + + Value *Ret = + EmitMemCpyChk(Dst, Src, + ConstantInt::get(TD->getIntPtrType(Context), Len), + CI->getArgOperand(2), B, TD, TLI); + return Ret; + } + return 0; + } +}; + +struct StpCpyChkOpt : public InstFortifiedLibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + this->CI = CI; + StringRef Name = Callee->getName(); + FunctionType *FT = Callee->getFunctionType(); + LLVMContext &Context = CI->getParent()->getContext(); + + // Check if this has the right signature. + if (FT->getNumParams() != 3 || + FT->getReturnType() != FT->getParamType(0) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != Type::getInt8PtrTy(Context) || + FT->getParamType(2) != TD->getIntPtrType(FT->getParamType(0))) + return 0; + + Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); + if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x) + Value *StrLen = EmitStrLen(Src, B, TD, TLI); + return StrLen ? B.CreateInBoundsGEP(Dst, StrLen) : 0; + } + + // If a) we don't have any length information, or b) we know this will + // fit then just lower to a plain stpcpy. Otherwise we'll keep our + // stpcpy_chk call which may fail at runtime if the size is too long. + // TODO: It might be nice to get a maximum length out of the possible + // string lengths for varying. + if (isFoldable(2, 1, true)) { + Value *Ret = EmitStrCpy(Dst, Src, B, TD, TLI, Name.substr(2, 6)); + return Ret; + } else { + // Maybe we can stil fold __stpcpy_chk to __memcpy_chk. + uint64_t Len = GetStringLength(Src); + if (Len == 0) return 0; + + // This optimization require DataLayout. + if (!TD) return 0; + + Type *PT = FT->getParamType(0); + Value *LenV = ConstantInt::get(TD->getIntPtrType(PT), Len); + Value *DstEnd = B.CreateGEP(Dst, + ConstantInt::get(TD->getIntPtrType(PT), + Len - 1)); + if (!EmitMemCpyChk(Dst, Src, LenV, CI->getArgOperand(2), B, TD, TLI)) + return 0; + return DstEnd; + } + return 0; + } +}; + +struct StrNCpyChkOpt : public InstFortifiedLibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + this->CI = CI; + StringRef Name = Callee->getName(); + FunctionType *FT = Callee->getFunctionType(); + LLVMContext &Context = CI->getParent()->getContext(); + + // Check if this has the right signature. + if (FT->getNumParams() != 4 || FT->getReturnType() != FT->getParamType(0) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != Type::getInt8PtrTy(Context) || + !FT->getParamType(2)->isIntegerTy() || + FT->getParamType(3) != TD->getIntPtrType(Context)) + return 0; + + if (isFoldable(3, 2, false)) { + Value *Ret = EmitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1), + CI->getArgOperand(2), B, TD, TLI, + Name.substr(2, 7)); + return Ret; + } + return 0; + } +}; + +//===----------------------------------------------------------------------===// +// String and Memory Library Call Optimizations +//===----------------------------------------------------------------------===// + +struct StrCatOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strcat" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getReturnType() != B.getInt8PtrTy() || + FT->getParamType(0) != FT->getReturnType() || + FT->getParamType(1) != FT->getReturnType()) + return 0; + + // Extract some information from the instruction + Value *Dst = CI->getArgOperand(0); + Value *Src = CI->getArgOperand(1); + + // See if we can get the length of the input string. + uint64_t Len = GetStringLength(Src); + if (Len == 0) return 0; + --Len; // Unbias length. + + // Handle the simple, do-nothing case: strcat(x, "") -> x + if (Len == 0) + return Dst; + + // These optimizations require DataLayout. + if (!TD) return 0; + + return emitStrLenMemCpy(Src, Dst, Len, B); + } + + Value *emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len, + IRBuilder<> &B) { + // We need to find the end of the destination string. That's where the + // memory is to be moved to. We just generate a call to strlen. + Value *DstLen = EmitStrLen(Dst, B, TD, TLI); + if (!DstLen) + return 0; + + // Now that we have the destination's length, we must index into the + // destination's pointer to get the actual memcpy destination (end of + // the string .. we're concatenating). + Value *CpyDst = B.CreateGEP(Dst, DstLen, "endptr"); + + // We have enough information to now generate the memcpy call to do the + // concatenation for us. Make a memcpy to copy the nul byte with align = 1. + B.CreateMemCpy(CpyDst, Src, + ConstantInt::get(TD->getIntPtrType(*Context), Len + 1), 1); + return Dst; + } +}; + +struct StrNCatOpt : public StrCatOpt { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strncat" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || + FT->getReturnType() != B.getInt8PtrTy() || + FT->getParamType(0) != FT->getReturnType() || + FT->getParamType(1) != FT->getReturnType() || + !FT->getParamType(2)->isIntegerTy()) + return 0; + + // Extract some information from the instruction + Value *Dst = CI->getArgOperand(0); + Value *Src = CI->getArgOperand(1); + uint64_t Len; + + // We don't do anything if length is not constant + if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2))) + Len = LengthArg->getZExtValue(); + else + return 0; + + // See if we can get the length of the input string. + uint64_t SrcLen = GetStringLength(Src); + if (SrcLen == 0) return 0; + --SrcLen; // Unbias length. + + // Handle the simple, do-nothing cases: + // strncat(x, "", c) -> x + // strncat(x, c, 0) -> x + if (SrcLen == 0 || Len == 0) return Dst; + + // These optimizations require DataLayout. + if (!TD) return 0; + + // We don't optimize this case + if (Len < SrcLen) return 0; + + // strncat(x, s, c) -> strcat(x, s) + // s is constant so the strcat can be optimized further + return emitStrLenMemCpy(Src, Dst, SrcLen, B); + } +}; + +struct StrChrOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strchr" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getReturnType() != B.getInt8PtrTy() || + FT->getParamType(0) != FT->getReturnType() || + !FT->getParamType(1)->isIntegerTy(32)) + return 0; + + Value *SrcStr = CI->getArgOperand(0); + + // If the second operand is non-constant, see if we can compute the length + // of the input string and turn this into memchr. + ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); + if (CharC == 0) { + // These optimizations require DataLayout. + if (!TD) return 0; + + uint64_t Len = GetStringLength(SrcStr); + if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32))// memchr needs i32. + return 0; + + return EmitMemChr(SrcStr, CI->getArgOperand(1), // include nul. + ConstantInt::get(TD->getIntPtrType(*Context), Len), + B, TD, TLI); + } + + // Otherwise, the character is a constant, see if the first argument is + // a string literal. If so, we can constant fold. + StringRef Str; + if (!getConstantStringInfo(SrcStr, Str)) + return 0; + + // Compute the offset, make sure to handle the case when we're searching for + // zero (a weird way to spell strlen). + size_t I = CharC->getSExtValue() == 0 ? + Str.size() : Str.find(CharC->getSExtValue()); + if (I == StringRef::npos) // Didn't find the char. strchr returns null. + return Constant::getNullValue(CI->getType()); + + // strchr(s+n,c) -> gep(s+n+i,c) + return B.CreateGEP(SrcStr, B.getInt64(I), "strchr"); + } +}; + +struct StrRChrOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strrchr" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getReturnType() != B.getInt8PtrTy() || + FT->getParamType(0) != FT->getReturnType() || + !FT->getParamType(1)->isIntegerTy(32)) + return 0; + + Value *SrcStr = CI->getArgOperand(0); + ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); + + // Cannot fold anything if we're not looking for a constant. + if (!CharC) + return 0; + + StringRef Str; + if (!getConstantStringInfo(SrcStr, Str)) { + // strrchr(s, 0) -> strchr(s, 0) + if (TD && CharC->isZero()) + return EmitStrChr(SrcStr, '\0', B, TD, TLI); + return 0; + } + + // Compute the offset. + size_t I = CharC->getSExtValue() == 0 ? + Str.size() : Str.rfind(CharC->getSExtValue()); + if (I == StringRef::npos) // Didn't find the char. Return null. + return Constant::getNullValue(CI->getType()); + + // strrchr(s+n,c) -> gep(s+n+i,c) + return B.CreateGEP(SrcStr, B.getInt64(I), "strrchr"); + } +}; + +struct StrCmpOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strcmp" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + !FT->getReturnType()->isIntegerTy(32) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != B.getInt8PtrTy()) + return 0; + + Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1); + if (Str1P == Str2P) // strcmp(x,x) -> 0 + return ConstantInt::get(CI->getType(), 0); + + StringRef Str1, Str2; + bool HasStr1 = getConstantStringInfo(Str1P, Str1); + bool HasStr2 = getConstantStringInfo(Str2P, Str2); + + // strcmp(x, y) -> cnst (if both x and y are constant strings) + if (HasStr1 && HasStr2) + return ConstantInt::get(CI->getType(), Str1.compare(Str2)); + + if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x + return B.CreateNeg(B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), + CI->getType())); + + if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x + return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType()); + + // strcmp(P, "x") -> memcmp(P, "x", 2) + uint64_t Len1 = GetStringLength(Str1P); + uint64_t Len2 = GetStringLength(Str2P); + if (Len1 && Len2) { + // These optimizations require DataLayout. + if (!TD) return 0; + + return EmitMemCmp(Str1P, Str2P, + ConstantInt::get(TD->getIntPtrType(*Context), + std::min(Len1, Len2)), B, TD, TLI); + } + + return 0; + } +}; + +struct StrNCmpOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strncmp" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || + !FT->getReturnType()->isIntegerTy(32) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != B.getInt8PtrTy() || + !FT->getParamType(2)->isIntegerTy()) + return 0; + + Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1); + if (Str1P == Str2P) // strncmp(x,x,n) -> 0 + return ConstantInt::get(CI->getType(), 0); + + // Get the length argument if it is constant. + uint64_t Length; + if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2))) + Length = LengthArg->getZExtValue(); + else + return 0; + + if (Length == 0) // strncmp(x,y,0) -> 0 + return ConstantInt::get(CI->getType(), 0); + + if (TD && Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1) + return EmitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, TD, TLI); + + StringRef Str1, Str2; + bool HasStr1 = getConstantStringInfo(Str1P, Str1); + bool HasStr2 = getConstantStringInfo(Str2P, Str2); + + // strncmp(x, y) -> cnst (if both x and y are constant strings) + if (HasStr1 && HasStr2) { + StringRef SubStr1 = Str1.substr(0, Length); + StringRef SubStr2 = Str2.substr(0, Length); + return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2)); + } + + if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x + return B.CreateNeg(B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), + CI->getType())); + + if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x + return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType()); + + return 0; + } +}; + +struct StrCpyOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "strcpy" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getReturnType() != FT->getParamType(0) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != B.getInt8PtrTy()) + return 0; + + Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); + if (Dst == Src) // strcpy(x,x) -> x + return Src; + + // These optimizations require DataLayout. + if (!TD) return 0; + + // See if we can get the length of the input string. + uint64_t Len = GetStringLength(Src); + if (Len == 0) return 0; + + // We have enough information to now generate the memcpy call to do the + // copy for us. Make a memcpy to copy the nul byte with align = 1. + B.CreateMemCpy(Dst, Src, + ConstantInt::get(TD->getIntPtrType(*Context), Len), 1); + return Dst; + } +}; + +struct StpCpyOpt: public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // Verify the "stpcpy" function prototype. + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getReturnType() != FT->getParamType(0) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != B.getInt8PtrTy()) + return 0; + + // These optimizations require DataLayout. + if (!TD) return 0; + + Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); + if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x) + Value *StrLen = EmitStrLen(Src, B, TD, TLI); + return StrLen ? B.CreateInBoundsGEP(Dst, StrLen) : 0; + } + + // See if we can get the length of the input string. + uint64_t Len = GetStringLength(Src); + if (Len == 0) return 0; + + Type *PT = FT->getParamType(0); + Value *LenV = ConstantInt::get(TD->getIntPtrType(PT), Len); + Value *DstEnd = B.CreateGEP(Dst, + ConstantInt::get(TD->getIntPtrType(PT), + Len - 1)); + + // We have enough information to now generate the memcpy call to do the + // copy for us. Make a memcpy to copy the nul byte with align = 1. + B.CreateMemCpy(Dst, Src, LenV, 1); + return DstEnd; + } +}; + +struct StrNCpyOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || + FT->getParamType(0) != FT->getParamType(1) || + FT->getParamType(0) != B.getInt8PtrTy() || + !FT->getParamType(2)->isIntegerTy()) + return 0; + + Value *Dst = CI->getArgOperand(0); + Value *Src = CI->getArgOperand(1); + Value *LenOp = CI->getArgOperand(2); + + // See if we can get the length of the input string. + uint64_t SrcLen = GetStringLength(Src); + if (SrcLen == 0) return 0; + --SrcLen; + + if (SrcLen == 0) { + // strncpy(x, "", y) -> memset(x, '\0', y, 1) + B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1); + return Dst; + } + + uint64_t Len; + if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp)) + Len = LengthArg->getZExtValue(); + else + return 0; + + if (Len == 0) return Dst; // strncpy(x, y, 0) -> x + + // These optimizations require DataLayout. + if (!TD) return 0; + + // Let strncpy handle the zero padding + if (Len > SrcLen+1) return 0; + + Type *PT = FT->getParamType(0); + // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant] + B.CreateMemCpy(Dst, Src, + ConstantInt::get(TD->getIntPtrType(PT), Len), 1); + + return Dst; + } +}; + +struct StrLenOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 1 || + FT->getParamType(0) != B.getInt8PtrTy() || + !FT->getReturnType()->isIntegerTy()) + return 0; + + Value *Src = CI->getArgOperand(0); + + // Constant folding: strlen("xyz") -> 3 + if (uint64_t Len = GetStringLength(Src)) + return ConstantInt::get(CI->getType(), Len-1); + + // strlen(x) != 0 --> *x != 0 + // strlen(x) == 0 --> *x == 0 + if (isOnlyUsedInZeroEqualityComparison(CI)) + return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType()); + return 0; + } +}; + +struct StrPBrkOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getParamType(0) != B.getInt8PtrTy() || + FT->getParamType(1) != FT->getParamType(0) || + FT->getReturnType() != FT->getParamType(0)) + return 0; + + StringRef S1, S2; + bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); + bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); + + // strpbrk(s, "") -> NULL + // strpbrk("", s) -> NULL + if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) + return Constant::getNullValue(CI->getType()); + + // Constant folding. + if (HasS1 && HasS2) { + size_t I = S1.find_first_of(S2); + if (I == std::string::npos) // No match. + return Constant::getNullValue(CI->getType()); + + return B.CreateGEP(CI->getArgOperand(0), B.getInt64(I), "strpbrk"); + } + + // strpbrk(s, "a") -> strchr(s, 'a') + if (TD && HasS2 && S2.size() == 1) + return EmitStrChr(CI->getArgOperand(0), S2[0], B, TD, TLI); + + return 0; + } +}; + +struct StrToOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if ((FT->getNumParams() != 2 && FT->getNumParams() != 3) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy()) + return 0; + + Value *EndPtr = CI->getArgOperand(1); + if (isa<ConstantPointerNull>(EndPtr)) { + // With a null EndPtr, this function won't capture the main argument. + // It would be readonly too, except that it still may write to errno. + CI->addAttribute(1, Attributes::get(Callee->getContext(), + Attributes::NoCapture)); + } + + return 0; + } +}; + +struct StrSpnOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getParamType(0) != B.getInt8PtrTy() || + FT->getParamType(1) != FT->getParamType(0) || + !FT->getReturnType()->isIntegerTy()) + return 0; + + StringRef S1, S2; + bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); + bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); + + // strspn(s, "") -> 0 + // strspn("", s) -> 0 + if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) + return Constant::getNullValue(CI->getType()); + + // Constant folding. + if (HasS1 && HasS2) { + size_t Pos = S1.find_first_not_of(S2); + if (Pos == StringRef::npos) Pos = S1.size(); + return ConstantInt::get(CI->getType(), Pos); + } + + return 0; + } +}; + +struct StrCSpnOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + FT->getParamType(0) != B.getInt8PtrTy() || + FT->getParamType(1) != FT->getParamType(0) || + !FT->getReturnType()->isIntegerTy()) + return 0; + + StringRef S1, S2; + bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); + bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); + + // strcspn("", s) -> 0 + if (HasS1 && S1.empty()) + return Constant::getNullValue(CI->getType()); + + // Constant folding. + if (HasS1 && HasS2) { + size_t Pos = S1.find_first_of(S2); + if (Pos == StringRef::npos) Pos = S1.size(); + return ConstantInt::get(CI->getType(), Pos); + } + + // strcspn(s, "") -> strlen(s) + if (TD && HasS2 && S2.empty()) + return EmitStrLen(CI->getArgOperand(0), B, TD, TLI); + + return 0; + } +}; + +struct StrStrOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 2 || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy() || + !FT->getReturnType()->isPointerTy()) + return 0; + + // fold strstr(x, x) -> x. + if (CI->getArgOperand(0) == CI->getArgOperand(1)) + return B.CreateBitCast(CI->getArgOperand(0), CI->getType()); + + // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0 + if (TD && isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) { + Value *StrLen = EmitStrLen(CI->getArgOperand(1), B, TD, TLI); + if (!StrLen) + return 0; + Value *StrNCmp = EmitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1), + StrLen, B, TD, TLI); + if (!StrNCmp) + return 0; + for (Value::use_iterator UI = CI->use_begin(), UE = CI->use_end(); + UI != UE; ) { + ICmpInst *Old = cast<ICmpInst>(*UI++); + Value *Cmp = B.CreateICmp(Old->getPredicate(), StrNCmp, + ConstantInt::getNullValue(StrNCmp->getType()), + "cmp"); + LCS->replaceAllUsesWith(Old, Cmp); + } + return CI; + } + + // See if either input string is a constant string. + StringRef SearchStr, ToFindStr; + bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr); + bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr); + + // fold strstr(x, "") -> x. + if (HasStr2 && ToFindStr.empty()) + return B.CreateBitCast(CI->getArgOperand(0), CI->getType()); + + // If both strings are known, constant fold it. + if (HasStr1 && HasStr2) { + std::string::size_type Offset = SearchStr.find(ToFindStr); + + if (Offset == StringRef::npos) // strstr("foo", "bar") -> null + return Constant::getNullValue(CI->getType()); + + // strstr("abcd", "bc") -> gep((char*)"abcd", 1) + Value *Result = CastToCStr(CI->getArgOperand(0), B); + Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr"); + return B.CreateBitCast(Result, CI->getType()); + } + + // fold strstr(x, "y") -> strchr(x, 'y'). + if (HasStr2 && ToFindStr.size() == 1) { + Value *StrChr= EmitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TD, TLI); + return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : 0; + } + return 0; + } +}; + +struct MemCmpOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy() || + !FT->getReturnType()->isIntegerTy(32)) + return 0; + + Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1); + + if (LHS == RHS) // memcmp(s,s,x) -> 0 + return Constant::getNullValue(CI->getType()); + + // Make sure we have a constant length. + ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2)); + if (!LenC) return 0; + uint64_t Len = LenC->getZExtValue(); + + if (Len == 0) // memcmp(s1,s2,0) -> 0 + return Constant::getNullValue(CI->getType()); + + // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS + if (Len == 1) { + Value *LHSV = B.CreateZExt(B.CreateLoad(CastToCStr(LHS, B), "lhsc"), + CI->getType(), "lhsv"); + Value *RHSV = B.CreateZExt(B.CreateLoad(CastToCStr(RHS, B), "rhsc"), + CI->getType(), "rhsv"); + return B.CreateSub(LHSV, RHSV, "chardiff"); + } + + // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant) + StringRef LHSStr, RHSStr; + if (getConstantStringInfo(LHS, LHSStr) && + getConstantStringInfo(RHS, RHSStr)) { + // Make sure we're not reading out-of-bounds memory. + if (Len > LHSStr.size() || Len > RHSStr.size()) + return 0; + uint64_t Ret = memcmp(LHSStr.data(), RHSStr.data(), Len); + return ConstantInt::get(CI->getType(), Ret); + } + + return 0; + } +}; + +struct MemCpyOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // These optimizations require DataLayout. + if (!TD) return 0; + + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy() || + FT->getParamType(2) != TD->getIntPtrType(*Context)) + return 0; + + // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1) + B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1), + CI->getArgOperand(2), 1); + return CI->getArgOperand(0); + } +}; + +struct MemMoveOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // These optimizations require DataLayout. + if (!TD) return 0; + + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isPointerTy() || + FT->getParamType(2) != TD->getIntPtrType(*Context)) + return 0; + + // memmove(x, y, n) -> llvm.memmove(x, y, n, 1) + B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1), + CI->getArgOperand(2), 1); + return CI->getArgOperand(0); + } +}; + +struct MemSetOpt : public LibCallOptimization { + virtual Value *callOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { + // These optimizations require DataLayout. + if (!TD) return 0; + + FunctionType *FT = Callee->getFunctionType(); + if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || + !FT->getParamType(0)->isPointerTy() || + !FT->getParamType(1)->isIntegerTy() || + FT->getParamType(2) != TD->getIntPtrType(*Context)) + return 0; + + // memset(p, v, n) -> llvm.memset(p, v, n, 1) + Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false); + B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1); + return CI->getArgOperand(0); + } +}; + +} // End anonymous namespace. + +namespace llvm { + +class LibCallSimplifierImpl { + const DataLayout *TD; + const TargetLibraryInfo *TLI; + const LibCallSimplifier *LCS; + StringMap<LibCallOptimization*> Optimizations; + + // Fortified library call optimizations. + MemCpyChkOpt MemCpyChk; + MemMoveChkOpt MemMoveChk; + MemSetChkOpt MemSetChk; + StrCpyChkOpt StrCpyChk; + StpCpyChkOpt StpCpyChk; + StrNCpyChkOpt StrNCpyChk; + + // String library call optimizations. + StrCatOpt StrCat; + StrNCatOpt StrNCat; + StrChrOpt StrChr; + StrRChrOpt StrRChr; + StrCmpOpt StrCmp; + StrNCmpOpt StrNCmp; + StrCpyOpt StrCpy; + StpCpyOpt StpCpy; + StrNCpyOpt StrNCpy; + StrLenOpt StrLen; + StrPBrkOpt StrPBrk; + StrToOpt StrTo; + StrSpnOpt StrSpn; + StrCSpnOpt StrCSpn; + StrStrOpt StrStr; + + // Memory library call optimizations. + MemCmpOpt MemCmp; + MemCpyOpt MemCpy; + MemMoveOpt MemMove; + MemSetOpt MemSet; + + void initOptimizations(); + void addOpt(LibFunc::Func F, LibCallOptimization* Opt); +public: + LibCallSimplifierImpl(const DataLayout *TD, const TargetLibraryInfo *TLI, + const LibCallSimplifier *LCS) { + this->TD = TD; + this->TLI = TLI; + this->LCS = LCS; + } + + Value *optimizeCall(CallInst *CI); +}; + +void LibCallSimplifierImpl::initOptimizations() { + // Fortified library call optimizations. + Optimizations["__memcpy_chk"] = &MemCpyChk; + Optimizations["__memmove_chk"] = &MemMoveChk; + Optimizations["__memset_chk"] = &MemSetChk; + Optimizations["__strcpy_chk"] = &StrCpyChk; + Optimizations["__stpcpy_chk"] = &StpCpyChk; + Optimizations["__strncpy_chk"] = &StrNCpyChk; + Optimizations["__stpncpy_chk"] = &StrNCpyChk; + + // String library call optimizations. + addOpt(LibFunc::strcat, &StrCat); + addOpt(LibFunc::strncat, &StrNCat); + addOpt(LibFunc::strchr, &StrChr); + addOpt(LibFunc::strrchr, &StrRChr); + addOpt(LibFunc::strcmp, &StrCmp); + addOpt(LibFunc::strncmp, &StrNCmp); + addOpt(LibFunc::strcpy, &StrCpy); + addOpt(LibFunc::stpcpy, &StpCpy); + addOpt(LibFunc::strncpy, &StrNCpy); + addOpt(LibFunc::strlen, &StrLen); + addOpt(LibFunc::strpbrk, &StrPBrk); + addOpt(LibFunc::strtol, &StrTo); + addOpt(LibFunc::strtod, &StrTo); + addOpt(LibFunc::strtof, &StrTo); + addOpt(LibFunc::strtoul, &StrTo); + addOpt(LibFunc::strtoll, &StrTo); + addOpt(LibFunc::strtold, &StrTo); + addOpt(LibFunc::strtoull, &StrTo); + addOpt(LibFunc::strspn, &StrSpn); + addOpt(LibFunc::strcspn, &StrCSpn); + addOpt(LibFunc::strstr, &StrStr); + + // Memory library call optimizations. + addOpt(LibFunc::memcmp, &MemCmp); + addOpt(LibFunc::memcpy, &MemCpy); + addOpt(LibFunc::memmove, &MemMove); + addOpt(LibFunc::memset, &MemSet); +} + +Value *LibCallSimplifierImpl::optimizeCall(CallInst *CI) { + if (Optimizations.empty()) + initOptimizations(); + + Function *Callee = CI->getCalledFunction(); + LibCallOptimization *LCO = Optimizations.lookup(Callee->getName()); + if (LCO) { + IRBuilder<> Builder(CI); + return LCO->optimizeCall(CI, TD, TLI, LCS, Builder); + } + return 0; +} + +void LibCallSimplifierImpl::addOpt(LibFunc::Func F, LibCallOptimization* Opt) { + if (TLI->has(F)) + Optimizations[TLI->getName(F)] = Opt; +} + +LibCallSimplifier::LibCallSimplifier(const DataLayout *TD, + const TargetLibraryInfo *TLI) { + Impl = new LibCallSimplifierImpl(TD, TLI, this); +} + +LibCallSimplifier::~LibCallSimplifier() { + delete Impl; +} + +Value *LibCallSimplifier::optimizeCall(CallInst *CI) { + return Impl->optimizeCall(CI); +} + +void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) const { + I->replaceAllUsesWith(With); + I->eraseFromParent(); +} + +} diff --git a/contrib/llvm/lib/Transforms/Utils/Utils.cpp b/contrib/llvm/lib/Transforms/Utils/Utils.cpp index 24e8c8f..5812d46 100644 --- a/contrib/llvm/lib/Transforms/Utils/Utils.cpp +++ b/contrib/llvm/lib/Transforms/Utils/Utils.cpp @@ -29,6 +29,7 @@ void llvm::initializeTransformUtils(PassRegistry &Registry) { initializePromotePassPass(Registry); initializeUnifyFunctionExitNodesPass(Registry); initializeInstSimplifierPass(Registry); + initializeMetaRenamerPass(Registry); } /// LLVMInitializeTransformUtils - C binding for initializeTransformUtilsPasses. diff --git a/contrib/llvm/lib/Transforms/Utils/ValueMapper.cpp b/contrib/llvm/lib/Transforms/Utils/ValueMapper.cpp index fc2538d..a30b093 100644 --- a/contrib/llvm/lib/Transforms/Utils/ValueMapper.cpp +++ b/contrib/llvm/lib/Transforms/Utils/ValueMapper.cpp @@ -21,7 +21,7 @@ using namespace llvm; // Out of line method to get vtable etc for class. -void ValueMapTypeRemapper::Anchor() {} +void ValueMapTypeRemapper::anchor() {} Value *llvm::MapValue(const Value *V, ValueToValueMapTy &VM, RemapFlags Flags, ValueMapTypeRemapper *TypeMapper) { diff --git a/contrib/llvm/lib/Transforms/Vectorize/BBVectorize.cpp b/contrib/llvm/lib/Transforms/Vectorize/BBVectorize.cpp index 62d23cb..f7be3e3 100644 --- a/contrib/llvm/lib/Transforms/Vectorize/BBVectorize.cpp +++ b/contrib/llvm/lib/Transforms/Vectorize/BBVectorize.cpp @@ -28,12 +28,14 @@ #include "llvm/Type.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/DenseSet.h" +#include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/StringExtras.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/AliasSetTracker.h" +#include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Analysis/ValueTracking.h" @@ -41,17 +43,27 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Support/ValueHandle.h" -#include "llvm/Target/TargetData.h" +#include "llvm/DataLayout.h" +#include "llvm/TargetTransformInfo.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Vectorize.h" #include <algorithm> #include <map> using namespace llvm; +static cl::opt<bool> +IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false), + cl::Hidden, cl::desc("Ignore target information")); + static cl::opt<unsigned> ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden, cl::desc("The required chain depth for vectorization")); +static cl::opt<bool> +UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false), + cl::Hidden, cl::desc("Use the chain depth requirement with" + " target information")); + static cl::opt<unsigned> SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden, cl::desc("The maximum search distance for instruction pairs")); @@ -93,8 +105,9 @@ static cl::opt<bool> NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, cl::desc("Don't try to vectorize floating-point values")); +// FIXME: This should default to false once pointer vector support works. static cl::opt<bool> -NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden, +NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden, cl::desc("Don't try to vectorize pointer values")); static cl::opt<bool> @@ -159,6 +172,12 @@ DebugCycleCheck("bb-vectorize-debug-cycle-check", cl::init(false), cl::Hidden, cl::desc("When debugging is enabled, output information on the" " cycle-checking process")); + +static cl::opt<bool> +PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair", + cl::init(false), cl::Hidden, + cl::desc("When debugging is enabled, dump the basic block after" + " every pair is fused")); #endif STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize"); @@ -177,13 +196,19 @@ namespace { BBVectorize(Pass *P, const VectorizeConfig &C) : BasicBlockPass(ID), Config(C) { AA = &P->getAnalysis<AliasAnalysis>(); + DT = &P->getAnalysis<DominatorTree>(); SE = &P->getAnalysis<ScalarEvolution>(); - TD = P->getAnalysisIfAvailable<TargetData>(); + TD = P->getAnalysisIfAvailable<DataLayout>(); + TTI = IgnoreTargetInfo ? 0 : + P->getAnalysisIfAvailable<TargetTransformInfo>(); + VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; } typedef std::pair<Value *, Value *> ValuePair; + typedef std::pair<ValuePair, int> ValuePairWithCost; typedef std::pair<ValuePair, size_t> ValuePairWithDepth; typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair + typedef std::pair<VPPair, unsigned> VPPairWithType; typedef std::pair<std::multimap<Value *, Value *>::iterator, std::multimap<Value *, Value *>::iterator> VPIteratorPair; typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator, @@ -191,8 +216,11 @@ namespace { VPPIteratorPair; AliasAnalysis *AA; + DominatorTree *DT; ScalarEvolution *SE; - TargetData *TD; + DataLayout *TD; + TargetTransformInfo *TTI; + const VectorTargetTransformInfo *VTTI; // FIXME: const correct? @@ -201,11 +229,23 @@ namespace { bool getCandidatePairs(BasicBlock &BB, BasicBlock::iterator &Start, std::multimap<Value *, Value *> &CandidatePairs, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<ValuePair, int> &CandidatePairCostSavings, std::vector<Value *> &PairableInsts, bool NonPow2Len); + // FIXME: The current implementation does not account for pairs that + // are connected in multiple ways. For example: + // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap) + enum PairConnectionType { + PairConnectionDirect, + PairConnectionSwap, + PairConnectionSplat + }; + void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs, std::vector<Value *> &PairableInsts, - std::multimap<ValuePair, ValuePair> &ConnectedPairs); + std::multimap<ValuePair, ValuePair> &ConnectedPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes); void buildDepMap(BasicBlock &BB, std::multimap<Value *, Value *> &CandidatePairs, @@ -213,19 +253,29 @@ namespace { DenseSet<ValuePair> &PairableInstUsers); void choosePairs(std::multimap<Value *, Value *> &CandidatePairs, + DenseMap<ValuePair, int> &CandidatePairCostSavings, std::vector<Value *> &PairableInsts, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, std::multimap<ValuePair, ValuePair> &ConnectedPairs, + std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, DenseSet<ValuePair> &PairableInstUsers, DenseMap<Value *, Value *>& ChosenPairs); void fuseChosenPairs(BasicBlock &BB, std::vector<Value *> &PairableInsts, - DenseMap<Value *, Value *>& ChosenPairs); + DenseMap<Value *, Value *>& ChosenPairs, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, + std::multimap<ValuePair, ValuePair> &ConnectedPairs, + std::multimap<ValuePair, ValuePair> &ConnectedPairDeps); + bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore); bool areInstsCompatible(Instruction *I, Instruction *J, - bool IsSimpleLoadStore, bool NonPow2Len); + bool IsSimpleLoadStore, bool NonPow2Len, + int &CostSavings, int &FixedOrder); bool trackUsesOfI(DenseSet<Value *> &Users, AliasSetTracker &WriteSet, Instruction *I, @@ -236,6 +286,7 @@ namespace { std::multimap<Value *, Value *> &CandidatePairs, std::vector<Value *> &PairableInsts, std::multimap<ValuePair, ValuePair> &ConnectedPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, ValuePair P); bool pairsConflict(ValuePair P, ValuePair Q, @@ -267,17 +318,21 @@ namespace { void findBestTreeFor( std::multimap<Value *, Value *> &CandidatePairs, + DenseMap<ValuePair, int> &CandidatePairCostSavings, std::vector<Value *> &PairableInsts, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, std::multimap<ValuePair, ValuePair> &ConnectedPairs, + std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, DenseSet<ValuePair> &PairableInstUsers, std::multimap<ValuePair, ValuePair> &PairableInstUserMap, DenseMap<Value *, Value *> &ChosenPairs, DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, - size_t &BestEffSize, VPIteratorPair ChoiceRange, + int &BestEffSize, VPIteratorPair ChoiceRange, bool UseCycleCheck); Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I, - Instruction *J, unsigned o, bool FlipMemInputs); + Instruction *J, unsigned o); void fillNewShuffleMask(LLVMContext& Context, Instruction *J, unsigned MaskOffset, unsigned NumInElem, @@ -289,20 +344,20 @@ namespace { bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J, unsigned o, Value *&LOp, unsigned numElemL, - Type *ArgTypeL, Type *ArgTypeR, + Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ, unsigned IdxOff = 0); Value *getReplacementInput(LLVMContext& Context, Instruction *I, - Instruction *J, unsigned o, bool FlipMemInputs); + Instruction *J, unsigned o, bool IBeforeJ); void getReplacementInputsForPair(LLVMContext& Context, Instruction *I, Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, - bool FlipMemInputs); + bool IBeforeJ); void replaceOutputsOfPair(LLVMContext& Context, Instruction *I, Instruction *J, Instruction *K, Instruction *&InsertionPt, Instruction *&K1, - Instruction *&K2, bool FlipMemInputs); + Instruction *&K2); void collectPairLoadMoveSet(BasicBlock &BB, DenseMap<Value *, Value *> &ChosenPairs, @@ -314,10 +369,6 @@ namespace { DenseMap<Value *, Value *> &ChosenPairs, std::multimap<Value *, Value *> &LoadMoveSet); - void collectPtrInfo(std::vector<Value *> &PairableInsts, - DenseMap<Value *, Value *> &ChosenPairs, - DenseSet<Value *> &LowPtrInsts); - bool canMoveUsesOfIAfterJ(BasicBlock &BB, std::multimap<Value *, Value *> &LoadMoveSet, Instruction *I, Instruction *J); @@ -330,13 +381,22 @@ namespace { void combineMetadata(Instruction *K, const Instruction *J); bool vectorizeBB(BasicBlock &BB) { + if (!DT->isReachableFromEntry(&BB)) { + DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() << + " in " << BB.getParent()->getName() << "\n"); + return false; + } + + DEBUG(if (VTTI) dbgs() << "BBV: using target information\n"); + bool changed = false; // Iterate a sufficient number of times to merge types of size 1 bit, // then 2 bits, then 4, etc. up to half of the target vector width of the // target vector register. unsigned n = 1; for (unsigned v = 2; - v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter); + (VTTI || v <= Config.VectorBits) && + (!Config.MaxIter || n <= Config.MaxIter); v *= 2, ++n) { DEBUG(dbgs() << "BBV: fusing loop #" << n << " for " << BB.getName() << " in " << @@ -363,8 +423,12 @@ namespace { virtual bool runOnBasicBlock(BasicBlock &BB) { AA = &getAnalysis<AliasAnalysis>(); + DT = &getAnalysis<DominatorTree>(); SE = &getAnalysis<ScalarEvolution>(); - TD = getAnalysisIfAvailable<TargetData>(); + TD = getAnalysisIfAvailable<DataLayout>(); + TTI = IgnoreTargetInfo ? 0 : + getAnalysisIfAvailable<TargetTransformInfo>(); + VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0; return vectorizeBB(BB); } @@ -372,8 +436,10 @@ namespace { virtual void getAnalysisUsage(AnalysisUsage &AU) const { BasicBlockPass::getAnalysisUsage(AU); AU.addRequired<AliasAnalysis>(); + AU.addRequired<DominatorTree>(); AU.addRequired<ScalarEvolution>(); AU.addPreserved<AliasAnalysis>(); + AU.addPreserved<DominatorTree>(); AU.addPreserved<ScalarEvolution>(); AU.setPreservesCFG(); } @@ -415,6 +481,14 @@ namespace { T2 = cast<CastInst>(I)->getSrcTy(); else T2 = T1; + + if (SelectInst *SI = dyn_cast<SelectInst>(I)) { + T2 = SI->getCondition()->getType(); + } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) { + T2 = SI->getOperand(0)->getType(); + } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { + T2 = CI->getOperand(0)->getType(); + } } // Returns the weight associated with the provided value. A chain of @@ -446,6 +520,62 @@ namespace { return 1; } + // Returns the cost of the provided instruction using VTTI. + // This does not handle loads and stores. + unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { + switch (Opcode) { + default: break; + case Instruction::GetElementPtr: + // We mark this instruction as zero-cost because scalar GEPs are usually + // lowered to the intruction addressing mode. At the moment we don't + // generate vector GEPs. + return 0; + case Instruction::Br: + return VTTI->getCFInstrCost(Opcode); + case Instruction::PHI: + return 0; + case Instruction::Add: + case Instruction::FAdd: + case Instruction::Sub: + case Instruction::FSub: + case Instruction::Mul: + case Instruction::FMul: + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + return VTTI->getArithmeticInstrCost(Opcode, T1); + case Instruction::Select: + case Instruction::ICmp: + case Instruction::FCmp: + return VTTI->getCmpSelInstrCost(Opcode, T1, T2); + case Instruction::ZExt: + case Instruction::SExt: + case Instruction::FPToUI: + case Instruction::FPToSI: + case Instruction::FPExt: + case Instruction::PtrToInt: + case Instruction::IntToPtr: + case Instruction::SIToFP: + case Instruction::UIToFP: + case Instruction::Trunc: + case Instruction::FPTrunc: + case Instruction::BitCast: + case Instruction::ShuffleVector: + return VTTI->getCastInstrCost(Opcode, T1, T2); + } + + return 1; + } + // This determines the relative offset of two loads or stores, returning // true if the offset could be determined to be some constant value. // For example, if OffsetInElmts == 1, then J accesses the memory directly @@ -453,20 +583,30 @@ namespace { // directly after J. bool getPairPtrInfo(Instruction *I, Instruction *J, Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment, - int64_t &OffsetInElmts) { + unsigned &IAddressSpace, unsigned &JAddressSpace, + int64_t &OffsetInElmts, bool ComputeOffset = true) { OffsetInElmts = 0; - if (isa<LoadInst>(I)) { - IPtr = cast<LoadInst>(I)->getPointerOperand(); - JPtr = cast<LoadInst>(J)->getPointerOperand(); - IAlignment = cast<LoadInst>(I)->getAlignment(); - JAlignment = cast<LoadInst>(J)->getAlignment(); + if (LoadInst *LI = dyn_cast<LoadInst>(I)) { + LoadInst *LJ = cast<LoadInst>(J); + IPtr = LI->getPointerOperand(); + JPtr = LJ->getPointerOperand(); + IAlignment = LI->getAlignment(); + JAlignment = LJ->getAlignment(); + IAddressSpace = LI->getPointerAddressSpace(); + JAddressSpace = LJ->getPointerAddressSpace(); } else { - IPtr = cast<StoreInst>(I)->getPointerOperand(); - JPtr = cast<StoreInst>(J)->getPointerOperand(); - IAlignment = cast<StoreInst>(I)->getAlignment(); - JAlignment = cast<StoreInst>(J)->getAlignment(); + StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J); + IPtr = SI->getPointerOperand(); + JPtr = SJ->getPointerOperand(); + IAlignment = SI->getAlignment(); + JAlignment = SJ->getAlignment(); + IAddressSpace = SI->getPointerAddressSpace(); + JAddressSpace = SJ->getPointerAddressSpace(); } + if (!ComputeOffset) + return true; + const SCEV *IPtrSCEV = SE->getSCEV(IPtr); const SCEV *JPtrSCEV = SE->getSCEV(JPtr); @@ -536,6 +676,19 @@ namespace { return false; } + + bool isPureIEChain(InsertElementInst *IE) { + InsertElementInst *IENext = IE; + do { + if (!isa<UndefValue>(IENext->getOperand(0)) && + !isa<InsertElementInst>(IENext->getOperand(0))) { + return false; + } + } while ((IENext = + dyn_cast<InsertElementInst>(IENext->getOperand(0)))); + + return true; + } }; // This function implements one vectorization iteration on the provided @@ -546,11 +699,18 @@ namespace { std::vector<Value *> AllPairableInsts; DenseMap<Value *, Value *> AllChosenPairs; + DenseSet<ValuePair> AllFixedOrderPairs; + DenseMap<VPPair, unsigned> AllPairConnectionTypes; + std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps; do { std::vector<Value *> PairableInsts; std::multimap<Value *, Value *> CandidatePairs; + DenseSet<ValuePair> FixedOrderPairs; + DenseMap<ValuePair, int> CandidatePairCostSavings; ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs, + FixedOrderPairs, + CandidatePairCostSavings, PairableInsts, NonPow2Len); if (PairableInsts.empty()) continue; @@ -563,10 +723,18 @@ namespace { // Note that it only matters that both members of the second pair use some // element of the first pair (to allow for splatting). - std::multimap<ValuePair, ValuePair> ConnectedPairs; - computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs); + std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps; + DenseMap<VPPair, unsigned> PairConnectionTypes; + computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs, + PairConnectionTypes); if (ConnectedPairs.empty()) continue; + for (std::multimap<ValuePair, ValuePair>::iterator + I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); + I != IE; ++I) { + ConnectedPairDeps.insert(VPPair(I->second, I->first)); + } + // Build the pairable-instruction dependency map DenseSet<ValuePair> PairableInstUsers; buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers); @@ -578,13 +746,48 @@ namespace { // variables. DenseMap<Value *, Value *> ChosenPairs; - choosePairs(CandidatePairs, PairableInsts, ConnectedPairs, + choosePairs(CandidatePairs, CandidatePairCostSavings, + PairableInsts, FixedOrderPairs, PairConnectionTypes, + ConnectedPairs, ConnectedPairDeps, PairableInstUsers, ChosenPairs); if (ChosenPairs.empty()) continue; AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(), PairableInsts.end()); AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end()); + + // Only for the chosen pairs, propagate information on fixed-order pairs, + // pair connections, and their types to the data structures used by the + // pair fusion procedures. + for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(), + IE = ChosenPairs.end(); I != IE; ++I) { + if (FixedOrderPairs.count(*I)) + AllFixedOrderPairs.insert(*I); + else if (FixedOrderPairs.count(ValuePair(I->second, I->first))) + AllFixedOrderPairs.insert(ValuePair(I->second, I->first)); + + for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin(); + J != IE; ++J) { + DenseMap<VPPair, unsigned>::iterator K = + PairConnectionTypes.find(VPPair(*I, *J)); + if (K != PairConnectionTypes.end()) { + AllPairConnectionTypes.insert(*K); + } else { + K = PairConnectionTypes.find(VPPair(*J, *I)); + if (K != PairConnectionTypes.end()) + AllPairConnectionTypes.insert(*K); + } + } + } + + for (std::multimap<ValuePair, ValuePair>::iterator + I = ConnectedPairs.begin(), IE = ConnectedPairs.end(); + I != IE; ++I) { + if (AllPairConnectionTypes.count(*I)) { + AllConnectedPairs.insert(*I); + AllConnectedPairDeps.insert(VPPair(I->second, I->first)); + } + } } while (ShouldContinue); if (AllChosenPairs.empty()) return false; @@ -597,11 +800,13 @@ namespace { // replaced with a vector_extract on the result. Subsequent optimization // passes should coalesce the build/extract combinations. - fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs); + fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs, + AllPairConnectionTypes, + AllConnectedPairs, AllConnectedPairDeps); // It is important to cleanup here so that future iterations of this // function have less work to do. - (void) SimplifyInstructionsInBlock(&BB, TD); + (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo()); return true; } @@ -667,15 +872,22 @@ namespace { !(VectorType::isValidElementType(T2) || T2->isVectorTy())) return false; - if (T1->getScalarSizeInBits() == 1 && T2->getScalarSizeInBits() == 1) { + if (T1->getScalarSizeInBits() == 1) { if (!Config.VectorizeBools) return false; } else { - if (!Config.VectorizeInts - && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy())) + if (!Config.VectorizeInts && T1->isIntOrIntVectorTy()) return false; } - + + if (T2->getScalarSizeInBits() == 1) { + if (!Config.VectorizeBools) + return false; + } else { + if (!Config.VectorizeInts && T2->isIntOrIntVectorTy()) + return false; + } + if (!Config.VectorizeFloats && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) return false; @@ -691,8 +903,8 @@ namespace { T2->getScalarType()->isPointerTy())) return false; - if (T1->getPrimitiveSizeInBits() >= Config.VectorBits || - T2->getPrimitiveSizeInBits() >= Config.VectorBits) + if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits || + T2->getPrimitiveSizeInBits() >= Config.VectorBits)) return false; return true; @@ -703,10 +915,14 @@ namespace { // that I has already been determined to be vectorizable and that J is not // in the use tree of I. bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J, - bool IsSimpleLoadStore, bool NonPow2Len) { + bool IsSimpleLoadStore, bool NonPow2Len, + int &CostSavings, int &FixedOrder) { DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I << " <-> " << *J << "\n"); + CostSavings = 0; + FixedOrder = 0; + // Loads and stores can be merged if they have different alignments, // but are otherwise the same. if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment | @@ -719,38 +935,84 @@ namespace { unsigned MaxTypeBits = std::max( IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(), IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits()); - if (MaxTypeBits > Config.VectorBits) + if (!VTTI && MaxTypeBits > Config.VectorBits) return false; // FIXME: handle addsub-type operations! if (IsSimpleLoadStore) { Value *IPtr, *JPtr; - unsigned IAlignment, JAlignment; + unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; int64_t OffsetInElmts = 0; if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, + IAddressSpace, JAddressSpace, OffsetInElmts) && abs64(OffsetInElmts) == 1) { - if (Config.AlignedOnly) { - Type *aTypeI = isa<StoreInst>(I) ? - cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); - Type *aTypeJ = isa<StoreInst>(J) ? - cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); + FixedOrder = (int) OffsetInElmts; + unsigned BottomAlignment = IAlignment; + if (OffsetInElmts < 0) BottomAlignment = JAlignment; + + Type *aTypeI = isa<StoreInst>(I) ? + cast<StoreInst>(I)->getValueOperand()->getType() : I->getType(); + Type *aTypeJ = isa<StoreInst>(J) ? + cast<StoreInst>(J)->getValueOperand()->getType() : J->getType(); + Type *VType = getVecTypeForPair(aTypeI, aTypeJ); + if (Config.AlignedOnly) { // An aligned load or store is possible only if the instruction // with the lower offset has an alignment suitable for the // vector type. - unsigned BottomAlignment = IAlignment; - if (OffsetInElmts < 0) BottomAlignment = JAlignment; - - Type *VType = getVecTypeForPair(aTypeI, aTypeJ); unsigned VecAlignment = TD->getPrefTypeAlignment(VType); if (BottomAlignment < VecAlignment) return false; } + + if (VTTI) { + unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(), + IAlignment, IAddressSpace); + unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(), + JAlignment, JAddressSpace); + unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType, + BottomAlignment, + IAddressSpace); + if (VCost > ICost + JCost) + return false; + + // We don't want to fuse to a type that will be split, even + // if the two input types will also be split and there is no other + // associated cost. + unsigned VParts = VTTI->getNumberOfParts(VType); + if (VParts > 1) + return false; + else if (!VParts && VCost == ICost + JCost) + return false; + + CostSavings = ICost + JCost - VCost; + } } else { return false; } + } else if (VTTI) { + unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2); + unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); + Type *VT1 = getVecTypeForPair(IT1, JT1), + *VT2 = getVecTypeForPair(IT2, JT2); + unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); + + if (VCost > ICost + JCost) + return false; + + // We don't want to fuse to a type that will be split, even + // if the two input types will also be split and there is no other + // associated cost. + unsigned VParts1 = VTTI->getNumberOfParts(VT1), + VParts2 = VTTI->getNumberOfParts(VT2); + if (VParts1 > 1 || VParts2 > 1) + return false; + else if ((!VParts1 || !VParts2) && VCost == ICost + JCost) + return false; + + CostSavings = ICost + JCost - VCost; } // The powi intrinsic is special because only the first argument is @@ -833,6 +1095,8 @@ namespace { bool BBVectorize::getCandidatePairs(BasicBlock &BB, BasicBlock::iterator &Start, std::multimap<Value *, Value *> &CandidatePairs, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<ValuePair, int> &CandidatePairCostSavings, std::vector<Value *> &PairableInsts, bool NonPow2Len) { BasicBlock::iterator E = BB.end(); if (Start == E) return false; @@ -869,7 +1133,9 @@ namespace { // J does not use I, and comes before the first use of I, so it can be // merged with I if the instructions are compatible. - if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len)) continue; + int CostSavings, FixedOrder; + if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len, + CostSavings, FixedOrder)) continue; // J is a candidate for merging with I. if (!PairableInsts.size() || @@ -878,6 +1144,14 @@ namespace { } CandidatePairs.insert(ValuePair(I, J)); + if (VTTI) + CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J), + CostSavings)); + + if (FixedOrder == 1) + FixedOrderPairs.insert(ValuePair(I, J)); + else if (FixedOrder == -1) + FixedOrderPairs.insert(ValuePair(J, I)); // The next call to this function must start after the last instruction // selected during this invocation. @@ -887,7 +1161,8 @@ namespace { } DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair " - << *I << " <-> " << *J << "\n"); + << *I << " <-> " << *J << " (cost savings: " << + CostSavings << ")\n"); // If we have already found too many pairs, break here and this function // will be called again starting after the last instruction selected @@ -915,6 +1190,7 @@ namespace { std::multimap<Value *, Value *> &CandidatePairs, std::vector<Value *> &PairableInsts, std::multimap<ValuePair, ValuePair> &ConnectedPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, ValuePair P) { StoreInst *SI, *SJ; @@ -946,12 +1222,18 @@ namespace { VPIteratorPair JPairRange = CandidatePairs.equal_range(*J); // Look for <I, J>: - if (isSecondInIteratorPair<Value*>(*J, IPairRange)) - ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); + if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { + VPPair VP(P, ValuePair(*I, *J)); + ConnectedPairs.insert(VP); + PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect)); + } // Look for <J, I>: - if (isSecondInIteratorPair<Value*>(*I, JPairRange)) - ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I))); + if (isSecondInIteratorPair<Value*>(*I, JPairRange)) { + VPPair VP(P, ValuePair(*J, *I)); + ConnectedPairs.insert(VP); + PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap)); + } } if (Config.SplatBreaksChain) continue; @@ -962,8 +1244,11 @@ namespace { P.first == SJ->getPointerOperand()) continue; - if (isSecondInIteratorPair<Value*>(*J, IPairRange)) - ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); + if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { + VPPair VP(P, ValuePair(*I, *J)); + ConnectedPairs.insert(VP); + PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); + } } } @@ -985,8 +1270,11 @@ namespace { P.second == SJ->getPointerOperand()) continue; - if (isSecondInIteratorPair<Value*>(*J, IPairRange)) - ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J))); + if (isSecondInIteratorPair<Value*>(*J, IPairRange)) { + VPPair VP(P, ValuePair(*I, *J)); + ConnectedPairs.insert(VP); + PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat)); + } } } } @@ -997,7 +1285,8 @@ namespace { void BBVectorize::computeConnectedPairs( std::multimap<Value *, Value *> &CandidatePairs, std::vector<Value *> &PairableInsts, - std::multimap<ValuePair, ValuePair> &ConnectedPairs) { + std::multimap<ValuePair, ValuePair> &ConnectedPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes) { for (std::vector<Value *>::iterator PI = PairableInsts.begin(), PE = PairableInsts.end(); PI != PE; ++PI) { @@ -1006,7 +1295,7 @@ namespace { for (std::multimap<Value *, Value *>::iterator P = choiceRange.first; P != choiceRange.second; ++P) computePairsConnectedTo(CandidatePairs, PairableInsts, - ConnectedPairs, *P); + ConnectedPairs, PairConnectionTypes, *P); } DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size() @@ -1196,7 +1485,7 @@ namespace { PrunedTree.insert(QTop.first); // Visit each child, pruning as necessary... - DenseMap<ValuePair, size_t> BestChildren; + SmallVector<ValuePairWithDepth, 8> BestChildren; VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first); for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first; K != QTopRange.second; ++K) { @@ -1228,7 +1517,7 @@ namespace { DenseSet<ValuePair> CurrentPairs; bool CanAdd = true; - for (DenseMap<ValuePair, size_t>::iterator C2 + for (SmallVector<ValuePairWithDepth, 8>::iterator C2 = BestChildren.begin(), E2 = BestChildren.end(); C2 != E2; ++C2) { if (C2->first.first == C->first.first || @@ -1313,22 +1602,22 @@ namespace { // to an already-selected child. Check for this here, and if a // conflict is found, then remove the previously-selected child // before adding this one in its place. - for (DenseMap<ValuePair, size_t>::iterator C2 + for (SmallVector<ValuePairWithDepth, 8>::iterator C2 = BestChildren.begin(); C2 != BestChildren.end();) { if (C2->first.first == C->first.first || C2->first.first == C->first.second || C2->first.second == C->first.first || C2->first.second == C->first.second || pairsConflict(C2->first, C->first, PairableInstUsers)) - BestChildren.erase(C2++); + C2 = BestChildren.erase(C2); else ++C2; } - BestChildren.insert(ValuePairWithDepth(C->first, C->second)); + BestChildren.push_back(ValuePairWithDepth(C->first, C->second)); } - for (DenseMap<ValuePair, size_t>::iterator C + for (SmallVector<ValuePairWithDepth, 8>::iterator C = BestChildren.begin(), E2 = BestChildren.end(); C != E2; ++C) { size_t DepthF = getDepthFactor(C->first.first); @@ -1341,13 +1630,17 @@ namespace { // pairs, given the choice of root pairs as an iterator range. void BBVectorize::findBestTreeFor( std::multimap<Value *, Value *> &CandidatePairs, + DenseMap<ValuePair, int> &CandidatePairCostSavings, std::vector<Value *> &PairableInsts, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, std::multimap<ValuePair, ValuePair> &ConnectedPairs, + std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, DenseSet<ValuePair> &PairableInstUsers, std::multimap<ValuePair, ValuePair> &PairableInstUserMap, DenseMap<Value *, Value *> &ChosenPairs, DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth, - size_t &BestEffSize, VPIteratorPair ChoiceRange, + int &BestEffSize, VPIteratorPair ChoiceRange, bool UseCycleCheck) { for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first; J != ChoiceRange.second; ++J) { @@ -1397,17 +1690,289 @@ namespace { PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree, PrunedTree, *J, UseCycleCheck); - size_t EffSize = 0; - for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), - E = PrunedTree.end(); S != E; ++S) - EffSize += getDepthFactor(S->first); + int EffSize = 0; + if (VTTI) { + DenseSet<Value *> PrunedTreeInstrs; + for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), + E = PrunedTree.end(); S != E; ++S) { + PrunedTreeInstrs.insert(S->first); + PrunedTreeInstrs.insert(S->second); + } + + // The set of pairs that have already contributed to the total cost. + DenseSet<ValuePair> IncomingPairs; + + // If the cost model were perfect, this might not be necessary; but we + // need to make sure that we don't get stuck vectorizing our own + // shuffle chains. + bool HasNontrivialInsts = false; + + // The node weights represent the cost savings associated with + // fusing the pair of instructions. + for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), + E = PrunedTree.end(); S != E; ++S) { + if (!isa<ShuffleVectorInst>(S->first) && + !isa<InsertElementInst>(S->first) && + !isa<ExtractElementInst>(S->first)) + HasNontrivialInsts = true; + + bool FlipOrder = false; + + if (getDepthFactor(S->first)) { + int ESContrib = CandidatePairCostSavings.find(*S)->second; + DEBUG(if (DebugPairSelection) dbgs() << "\tweight {" + << *S->first << " <-> " << *S->second << "} = " << + ESContrib << "\n"); + EffSize += ESContrib; + } + + // The edge weights contribute in a negative sense: they represent + // the cost of shuffles. + VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S); + if (IP.first != ConnectedPairDeps.end()) { + unsigned NumDepsDirect = 0, NumDepsSwap = 0; + for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; + Q != IP.second; ++Q) { + if (!PrunedTree.count(Q->second)) + continue; + DenseMap<VPPair, unsigned>::iterator R = + PairConnectionTypes.find(VPPair(Q->second, Q->first)); + assert(R != PairConnectionTypes.end() && + "Cannot find pair connection type"); + if (R->second == PairConnectionDirect) + ++NumDepsDirect; + else if (R->second == PairConnectionSwap) + ++NumDepsSwap; + } + + // If there are more swaps than direct connections, then + // the pair order will be flipped during fusion. So the real + // number of swaps is the minimum number. + FlipOrder = !FixedOrderPairs.count(*S) && + ((NumDepsSwap > NumDepsDirect) || + FixedOrderPairs.count(ValuePair(S->second, S->first))); + + for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; + Q != IP.second; ++Q) { + if (!PrunedTree.count(Q->second)) + continue; + DenseMap<VPPair, unsigned>::iterator R = + PairConnectionTypes.find(VPPair(Q->second, Q->first)); + assert(R != PairConnectionTypes.end() && + "Cannot find pair connection type"); + Type *Ty1 = Q->second.first->getType(), + *Ty2 = Q->second.second->getType(); + Type *VTy = getVecTypeForPair(Ty1, Ty2); + if ((R->second == PairConnectionDirect && FlipOrder) || + (R->second == PairConnectionSwap && !FlipOrder) || + R->second == PairConnectionSplat) { + int ESContrib = (int) getInstrCost(Instruction::ShuffleVector, + VTy, VTy); + DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << + *Q->second.first << " <-> " << *Q->second.second << + "} -> {" << + *S->first << " <-> " << *S->second << "} = " << + ESContrib << "\n"); + EffSize -= ESContrib; + } + } + } + + // Compute the cost of outgoing edges. We assume that edges outgoing + // to shuffles, inserts or extracts can be merged, and so contribute + // no additional cost. + if (!S->first->getType()->isVoidTy()) { + Type *Ty1 = S->first->getType(), + *Ty2 = S->second->getType(); + Type *VTy = getVecTypeForPair(Ty1, Ty2); + + bool NeedsExtraction = false; + for (Value::use_iterator I = S->first->use_begin(), + IE = S->first->use_end(); I != IE; ++I) { + if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { + // Shuffle can be folded if it has no other input + if (isa<UndefValue>(SI->getOperand(1))) + continue; + } + if (isa<ExtractElementInst>(*I)) + continue; + if (PrunedTreeInstrs.count(*I)) + continue; + NeedsExtraction = true; + break; + } + + if (NeedsExtraction) { + int ESContrib; + if (Ty1->isVectorTy()) + ESContrib = (int) getInstrCost(Instruction::ShuffleVector, + Ty1, VTy); + else + ESContrib = (int) VTTI->getVectorInstrCost( + Instruction::ExtractElement, VTy, 0); + + DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << + *S->first << "} = " << ESContrib << "\n"); + EffSize -= ESContrib; + } + + NeedsExtraction = false; + for (Value::use_iterator I = S->second->use_begin(), + IE = S->second->use_end(); I != IE; ++I) { + if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) { + // Shuffle can be folded if it has no other input + if (isa<UndefValue>(SI->getOperand(1))) + continue; + } + if (isa<ExtractElementInst>(*I)) + continue; + if (PrunedTreeInstrs.count(*I)) + continue; + NeedsExtraction = true; + break; + } + + if (NeedsExtraction) { + int ESContrib; + if (Ty2->isVectorTy()) + ESContrib = (int) getInstrCost(Instruction::ShuffleVector, + Ty2, VTy); + else + ESContrib = (int) VTTI->getVectorInstrCost( + Instruction::ExtractElement, VTy, 1); + DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" << + *S->second << "} = " << ESContrib << "\n"); + EffSize -= ESContrib; + } + } + + // Compute the cost of incoming edges. + if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) { + Instruction *S1 = cast<Instruction>(S->first), + *S2 = cast<Instruction>(S->second); + for (unsigned o = 0; o < S1->getNumOperands(); ++o) { + Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o); + + // Combining constants into vector constants (or small vector + // constants into larger ones are assumed free). + if (isa<Constant>(O1) && isa<Constant>(O2)) + continue; + + if (FlipOrder) + std::swap(O1, O2); + + ValuePair VP = ValuePair(O1, O2); + ValuePair VPR = ValuePair(O2, O1); + + // Internal edges are not handled here. + if (PrunedTree.count(VP) || PrunedTree.count(VPR)) + continue; + + Type *Ty1 = O1->getType(), + *Ty2 = O2->getType(); + Type *VTy = getVecTypeForPair(Ty1, Ty2); + + // Combining vector operations of the same type is also assumed + // folded with other operations. + if (Ty1 == Ty2) { + // If both are insert elements, then both can be widened. + InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1), + *IEO2 = dyn_cast<InsertElementInst>(O2); + if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2)) + continue; + // If both are extract elements, and both have the same input + // type, then they can be replaced with a shuffle + ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1), + *EIO2 = dyn_cast<ExtractElementInst>(O2); + if (EIO1 && EIO2 && + EIO1->getOperand(0)->getType() == + EIO2->getOperand(0)->getType()) + continue; + // If both are a shuffle with equal operand types and only two + // unqiue operands, then they can be replaced with a single + // shuffle + ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1), + *SIO2 = dyn_cast<ShuffleVectorInst>(O2); + if (SIO1 && SIO2 && + SIO1->getOperand(0)->getType() == + SIO2->getOperand(0)->getType()) { + SmallSet<Value *, 4> SIOps; + SIOps.insert(SIO1->getOperand(0)); + SIOps.insert(SIO1->getOperand(1)); + SIOps.insert(SIO2->getOperand(0)); + SIOps.insert(SIO2->getOperand(1)); + if (SIOps.size() <= 2) + continue; + } + } + + int ESContrib; + // This pair has already been formed. + if (IncomingPairs.count(VP)) { + continue; + } else if (IncomingPairs.count(VPR)) { + ESContrib = (int) getInstrCost(Instruction::ShuffleVector, + VTy, VTy); + } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) { + ESContrib = (int) VTTI->getVectorInstrCost( + Instruction::InsertElement, VTy, 0); + ESContrib += (int) VTTI->getVectorInstrCost( + Instruction::InsertElement, VTy, 1); + } else if (!Ty1->isVectorTy()) { + // O1 needs to be inserted into a vector of size O2, and then + // both need to be shuffled together. + ESContrib = (int) VTTI->getVectorInstrCost( + Instruction::InsertElement, Ty2, 0); + ESContrib += (int) getInstrCost(Instruction::ShuffleVector, + VTy, Ty2); + } else if (!Ty2->isVectorTy()) { + // O2 needs to be inserted into a vector of size O1, and then + // both need to be shuffled together. + ESContrib = (int) VTTI->getVectorInstrCost( + Instruction::InsertElement, Ty1, 0); + ESContrib += (int) getInstrCost(Instruction::ShuffleVector, + VTy, Ty1); + } else { + Type *TyBig = Ty1, *TySmall = Ty2; + if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements()) + std::swap(TyBig, TySmall); + + ESContrib = (int) getInstrCost(Instruction::ShuffleVector, + VTy, TyBig); + if (TyBig != TySmall) + ESContrib += (int) getInstrCost(Instruction::ShuffleVector, + TyBig, TySmall); + } + + DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" + << *O1 << " <-> " << *O2 << "} = " << + ESContrib << "\n"); + EffSize -= ESContrib; + IncomingPairs.insert(VP); + } + } + } + + if (!HasNontrivialInsts) { + DEBUG(if (DebugPairSelection) dbgs() << + "\tNo non-trivial instructions in tree;" + " override to zero effective size\n"); + EffSize = 0; + } + } else { + for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(), + E = PrunedTree.end(); S != E; ++S) + EffSize += (int) getDepthFactor(S->first); + } DEBUG(if (DebugPairSelection) dbgs() << "BBV: found pruned Tree for pair {" << *J->first << " <-> " << *J->second << "} of depth " << MaxDepth << " and size " << PrunedTree.size() << " (effective size: " << EffSize << ")\n"); - if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) { + if (((VTTI && !UseChainDepthWithTI) || + MaxDepth >= Config.ReqChainDepth) && + EffSize > 0 && EffSize > BestEffSize) { BestMaxDepth = MaxDepth; BestEffSize = EffSize; BestTree = PrunedTree; @@ -1419,8 +1984,12 @@ namespace { // that will be fused into vector instructions. void BBVectorize::choosePairs( std::multimap<Value *, Value *> &CandidatePairs, + DenseMap<ValuePair, int> &CandidatePairCostSavings, std::vector<Value *> &PairableInsts, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, std::multimap<ValuePair, ValuePair> &ConnectedPairs, + std::multimap<ValuePair, ValuePair> &ConnectedPairDeps, DenseSet<ValuePair> &PairableInstUsers, DenseMap<Value *, Value *>& ChosenPairs) { bool UseCycleCheck = @@ -1435,9 +2004,12 @@ namespace { VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I); // The best pair to choose and its tree: - size_t BestMaxDepth = 0, BestEffSize = 0; + size_t BestMaxDepth = 0; + int BestEffSize = 0; DenseSet<ValuePair> BestTree; - findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs, + findBestTreeFor(CandidatePairs, CandidatePairCostSavings, + PairableInsts, FixedOrderPairs, PairConnectionTypes, + ConnectedPairs, ConnectedPairDeps, PairableInstUsers, PairableInstUserMap, ChosenPairs, BestTree, BestMaxDepth, BestEffSize, ChoiceRange, UseCycleCheck); @@ -1490,24 +2062,19 @@ namespace { // Returns the value that is to be used as the pointer input to the vector // instruction that fuses I with J. Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context, - Instruction *I, Instruction *J, unsigned o, - bool FlipMemInputs) { + Instruction *I, Instruction *J, unsigned o) { Value *IPtr, *JPtr; - unsigned IAlignment, JAlignment; + unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace; int64_t OffsetInElmts; - // Note: the analysis might fail here, that is why FlipMemInputs has + // Note: the analysis might fail here, that is why the pair order has // been precomputed (OffsetInElmts must be unused here). (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, - OffsetInElmts); + IAddressSpace, JAddressSpace, + OffsetInElmts, false); // The pointer value is taken to be the one with the lowest offset. - Value *VPtr; - if (!FlipMemInputs) { - VPtr = IPtr; - } else { - VPtr = JPtr; - } + Value *VPtr = IPtr; Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType(); Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType(); @@ -1515,7 +2082,7 @@ namespace { Type *VArgPtrType = PointerType::get(VArgType, cast<PointerType>(IPtr->getType())->getAddressSpace()); return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o), - /* insert before */ FlipMemInputs ? J : I); + /* insert before */ I); } void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J, @@ -1585,23 +2152,12 @@ namespace { Instruction *J, unsigned o, Value *&LOp, unsigned numElemL, Type *ArgTypeL, Type *ArgTypeH, - unsigned IdxOff) { + bool IBeforeJ, unsigned IdxOff) { bool ExpandedIEChain = false; if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) { // If we have a pure insertelement chain, then this can be rewritten // into a chain that directly builds the larger type. - bool PureChain = true; - InsertElementInst *LIENext = LIE; - do { - if (!isa<UndefValue>(LIENext->getOperand(0)) && - !isa<InsertElementInst>(LIENext->getOperand(0))) { - PureChain = false; - break; - } - } while ((LIENext = - dyn_cast<InsertElementInst>(LIENext->getOperand(0)))); - - if (PureChain) { + if (isPureIEChain(LIE)) { SmallVector<Value *, 8> VectElemts(numElemL, UndefValue::get(ArgTypeL->getScalarType())); InsertElementInst *LIENext = LIE; @@ -1619,8 +2175,9 @@ namespace { LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i], ConstantInt::get(Type::getInt32Ty(Context), i + IdxOff), - getReplacementName(I, true, o, i+1)); - LIENext->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, + true, o, i+1)); + LIENext->insertBefore(IBeforeJ ? J : I); LIEPrev = LIENext; } @@ -1635,7 +2192,7 @@ namespace { // Returns the value to be used as the specified operand of the vector // instruction that fuses I with J. Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I, - Instruction *J, unsigned o, bool FlipMemInputs) { + Instruction *J, unsigned o, bool IBeforeJ) { Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1); @@ -1646,12 +2203,6 @@ namespace { Instruction *L = I, *H = J; Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ; - if (FlipMemInputs) { - L = J; - H = I; - ArgTypeL = ArgTypeJ; - ArgTypeH = ArgTypeI; - } unsigned numElemL; if (ArgTypeL->isVectorTy()) @@ -1804,8 +2355,9 @@ namespace { Instruction *S = new ShuffleVectorInst(I1, UndefValue::get(I1T), ConstantVector::get(Mask), - getReplacementName(I, true, o)); - S->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, + true, o)); + S->insertBefore(IBeforeJ ? J : I); return S; } @@ -1826,8 +2378,9 @@ namespace { Instruction *NewI1 = new ShuffleVectorInst(I1, UndefValue::get(I1T), ConstantVector::get(Mask), - getReplacementName(I, true, o, 1)); - NewI1->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); + NewI1->insertBefore(IBeforeJ ? J : I); I1 = NewI1; I1T = I2T; I1Elem = I2Elem; @@ -1842,8 +2395,9 @@ namespace { Instruction *NewI2 = new ShuffleVectorInst(I2, UndefValue::get(I2T), ConstantVector::get(Mask), - getReplacementName(I, true, o, 1)); - NewI2->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); + NewI2->insertBefore(IBeforeJ ? J : I); I2 = NewI2; I2T = I1T; I2Elem = I1Elem; @@ -1863,8 +2417,8 @@ namespace { Instruction *NewOp = new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask), - getReplacementName(I, true, o)); - NewOp->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, true, o)); + NewOp->insertBefore(IBeforeJ ? J : I); return NewOp; } } @@ -1872,17 +2426,17 @@ namespace { Type *ArgType = ArgTypeL; if (numElemL < numElemH) { if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH, - ArgTypeL, VArgType, 1)) { + ArgTypeL, VArgType, IBeforeJ, 1)) { // This is another short-circuit case: we're combining a scalar into // a vector that is formed by an IE chain. We've just expanded the IE // chain, now insert the scalar and we're done. Instruction *S = InsertElementInst::Create(HOp, LOp, CV0, - getReplacementName(I, true, o)); - S->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, true, o)); + S->insertBefore(IBeforeJ ? J : I); return S; } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL, - ArgTypeH)) { + ArgTypeH, IBeforeJ)) { // The two vector inputs to the shuffle must be the same length, // so extend the smaller vector to be the same length as the larger one. Instruction *NLOp; @@ -1897,29 +2451,32 @@ namespace { NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL), ConstantVector::get(Mask), - getReplacementName(I, true, o, 1)); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); } else { NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0, - getReplacementName(I, true, o, 1)); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); } - NLOp->insertBefore(J); + NLOp->insertBefore(IBeforeJ ? J : I); LOp = NLOp; } ArgType = ArgTypeH; } else if (numElemL > numElemH) { if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL, - ArgTypeH, VArgType)) { + ArgTypeH, VArgType, IBeforeJ)) { Instruction *S = InsertElementInst::Create(LOp, HOp, ConstantInt::get(Type::getInt32Ty(Context), numElemL), - getReplacementName(I, true, o)); - S->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, + true, o)); + S->insertBefore(IBeforeJ ? J : I); return S; } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH, - ArgTypeL)) { + ArgTypeL, IBeforeJ)) { Instruction *NHOp; if (numElemH > 1) { std::vector<Constant *> Mask(numElemL); @@ -1931,13 +2488,15 @@ namespace { NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH), ConstantVector::get(Mask), - getReplacementName(I, true, o, 1)); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); } else { NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0, - getReplacementName(I, true, o, 1)); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); } - NHOp->insertBefore(J); + NHOp->insertBefore(IBeforeJ ? J : I); HOp = NHOp; } } @@ -1955,19 +2514,21 @@ namespace { } Instruction *BV = new ShuffleVectorInst(LOp, HOp, - ConstantVector::get(Mask), - getReplacementName(I, true, o)); - BV->insertBefore(J); + ConstantVector::get(Mask), + getReplacementName(IBeforeJ ? I : J, true, o)); + BV->insertBefore(IBeforeJ ? J : I); return BV; } Instruction *BV1 = InsertElementInst::Create( UndefValue::get(VArgType), LOp, CV0, - getReplacementName(I, true, o, 1)); - BV1->insertBefore(I); + getReplacementName(IBeforeJ ? I : J, + true, o, 1)); + BV1->insertBefore(IBeforeJ ? J : I); Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1, - getReplacementName(I, true, o, 2)); - BV2->insertBefore(J); + getReplacementName(IBeforeJ ? I : J, + true, o, 2)); + BV2->insertBefore(IBeforeJ ? J : I); return BV2; } @@ -1976,7 +2537,7 @@ namespace { void BBVectorize::getReplacementInputsForPair(LLVMContext& Context, Instruction *I, Instruction *J, SmallVector<Value *, 3> &ReplacedOperands, - bool FlipMemInputs) { + bool IBeforeJ) { unsigned NumOperands = I->getNumOperands(); for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) { @@ -1985,8 +2546,7 @@ namespace { if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) { // This is the pointer for a load/store instruction. - ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o, - FlipMemInputs); + ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o); continue; } else if (isa<CallInst>(I)) { Function *F = cast<CallInst>(I)->getCalledFunction(); @@ -2014,8 +2574,7 @@ namespace { continue; } - ReplacedOperands[o] = - getReplacementInput(Context, I, J, o, FlipMemInputs); + ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ); } } @@ -2026,8 +2585,7 @@ namespace { void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I, Instruction *J, Instruction *K, Instruction *&InsertionPt, - Instruction *&K1, Instruction *&K2, - bool FlipMemInputs) { + Instruction *&K1, Instruction *&K2) { if (isa<StoreInst>(I)) { AA->replaceWithNewValue(I, K); AA->replaceWithNewValue(J, K); @@ -2057,13 +2615,11 @@ namespace { } K1 = new ShuffleVectorInst(K, UndefValue::get(VType), - ConstantVector::get( - FlipMemInputs ? Mask2 : Mask1), + ConstantVector::get( Mask1), getReplacementName(K, false, 1)); } else { Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); - Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); - K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0, + K1 = ExtractElementInst::Create(K, CV0, getReplacementName(K, false, 1)); } @@ -2075,13 +2631,11 @@ namespace { } K2 = new ShuffleVectorInst(K, UndefValue::get(VType), - ConstantVector::get( - FlipMemInputs ? Mask1 : Mask2), + ConstantVector::get( Mask2), getReplacementName(K, false, 2)); } else { - Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0); Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1); - K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1, + K2 = ExtractElementInst::Create(K, CV1, getReplacementName(K, false, 2)); } @@ -2181,36 +2735,6 @@ namespace { } } - // As with the aliasing information, SCEV can also change because of - // vectorization. This information is used to compute relative pointer - // offsets; the necessary information will be cached here prior to - // fusion. - void BBVectorize::collectPtrInfo(std::vector<Value *> &PairableInsts, - DenseMap<Value *, Value *> &ChosenPairs, - DenseSet<Value *> &LowPtrInsts) { - for (std::vector<Value *>::iterator PI = PairableInsts.begin(), - PIE = PairableInsts.end(); PI != PIE; ++PI) { - DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI); - if (P == ChosenPairs.end()) continue; - - Instruction *I = cast<Instruction>(P->first); - Instruction *J = cast<Instruction>(P->second); - - if (!isa<LoadInst>(I) && !isa<StoreInst>(I)) - continue; - - Value *IPtr, *JPtr; - unsigned IAlignment, JAlignment; - int64_t OffsetInElmts; - if (!getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment, - OffsetInElmts) || abs64(OffsetInElmts) != 1) - llvm_unreachable("Pre-fusion pointer analysis failed"); - - Value *LowPI = (OffsetInElmts > 0) ? I : J; - LowPtrInsts.insert(LowPI); - } - } - // When the first instruction in each pair is cloned, it will inherit its // parent's metadata. This metadata must be combined with that of the other // instruction in a safe way. @@ -2244,27 +2768,27 @@ namespace { // second member). void BBVectorize::fuseChosenPairs(BasicBlock &BB, std::vector<Value *> &PairableInsts, - DenseMap<Value *, Value *> &ChosenPairs) { + DenseMap<Value *, Value *> &ChosenPairs, + DenseSet<ValuePair> &FixedOrderPairs, + DenseMap<VPPair, unsigned> &PairConnectionTypes, + std::multimap<ValuePair, ValuePair> &ConnectedPairs, + std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) { LLVMContext& Context = BB.getContext(); // During the vectorization process, the order of the pairs to be fused // could be flipped. So we'll add each pair, flipped, into the ChosenPairs // list. After a pair is fused, the flipped pair is removed from the list. - std::vector<ValuePair> FlippedPairs; - FlippedPairs.reserve(ChosenPairs.size()); + DenseSet<ValuePair> FlippedPairs; for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(), E = ChosenPairs.end(); P != E; ++P) - FlippedPairs.push_back(ValuePair(P->second, P->first)); - for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(), + FlippedPairs.insert(ValuePair(P->second, P->first)); + for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(), E = FlippedPairs.end(); P != E; ++P) ChosenPairs.insert(*P); std::multimap<Value *, Value *> LoadMoveSet; collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet); - DenseSet<Value *> LowPtrInsts; - collectPtrInfo(PairableInsts, ChosenPairs, LowPtrInsts); - DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n"); for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) { @@ -2304,44 +2828,92 @@ namespace { continue; } - bool FlipMemInputs = false; - if (isa<LoadInst>(I) || isa<StoreInst>(I)) - FlipMemInputs = (LowPtrInsts.find(I) == LowPtrInsts.end()); + // If the pair must have the other order, then flip it. + bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I)); + if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) { + // This pair does not have a fixed order, and so we might want to + // flip it if that will yield fewer shuffles. We count the number + // of dependencies connected via swaps, and those directly connected, + // and flip the order if the number of swaps is greater. + bool OrigOrder = true; + VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J)); + if (IP.first == ConnectedPairDeps.end()) { + IP = ConnectedPairDeps.equal_range(ValuePair(J, I)); + OrigOrder = false; + } + if (IP.first != ConnectedPairDeps.end()) { + unsigned NumDepsDirect = 0, NumDepsSwap = 0; + for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; + Q != IP.second; ++Q) { + DenseMap<VPPair, unsigned>::iterator R = + PairConnectionTypes.find(VPPair(Q->second, Q->first)); + assert(R != PairConnectionTypes.end() && + "Cannot find pair connection type"); + if (R->second == PairConnectionDirect) + ++NumDepsDirect; + else if (R->second == PairConnectionSwap) + ++NumDepsSwap; + } + + if (!OrigOrder) + std::swap(NumDepsDirect, NumDepsSwap); + + if (NumDepsSwap > NumDepsDirect) { + FlipPairOrder = true; + DEBUG(dbgs() << "BBV: reordering pair: " << *I << + " <-> " << *J << "\n"); + } + } + } + + Instruction *L = I, *H = J; + if (FlipPairOrder) + std::swap(H, L); + + // If the pair being fused uses the opposite order from that in the pair + // connection map, then we need to flip the types. + VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L)); + for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first; + Q != IP.second; ++Q) { + DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q); + assert(R != PairConnectionTypes.end() && + "Cannot find pair connection type"); + if (R->second == PairConnectionDirect) + R->second = PairConnectionSwap; + else if (R->second == PairConnectionSwap) + R->second = PairConnectionDirect; + } + + bool LBeforeH = !FlipPairOrder; unsigned NumOperands = I->getNumOperands(); SmallVector<Value *, 3> ReplacedOperands(NumOperands); - getReplacementInputsForPair(Context, I, J, ReplacedOperands, - FlipMemInputs); + getReplacementInputsForPair(Context, L, H, ReplacedOperands, + LBeforeH); // Make a copy of the original operation, change its type to the vector // type and replace its operands with the vector operands. - Instruction *K = I->clone(); - if (I->hasName()) K->takeName(I); + Instruction *K = L->clone(); + if (L->hasName()) + K->takeName(L); + else if (H->hasName()) + K->takeName(H); if (!isa<StoreInst>(K)) - K->mutateType(getVecTypeForPair(I->getType(), J->getType())); + K->mutateType(getVecTypeForPair(L->getType(), H->getType())); - combineMetadata(K, J); + combineMetadata(K, H); + K->intersectOptionalDataWith(H); for (unsigned o = 0; o < NumOperands; ++o) K->setOperand(o, ReplacedOperands[o]); - // If we've flipped the memory inputs, make sure that we take the correct - // alignment. - if (FlipMemInputs) { - if (isa<StoreInst>(K)) - cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment()); - else - cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment()); - } - K->insertAfter(J); // Instruction insertion point: Instruction *InsertionPt = K; Instruction *K1 = 0, *K2 = 0; - replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2, - FlipMemInputs); + replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2); // The use tree of the first original instruction must be moved to after // the location of the second instruction. The entire use tree of the @@ -2351,10 +2923,10 @@ namespace { moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J); if (!isa<StoreInst>(I)) { - I->replaceAllUsesWith(K1); - J->replaceAllUsesWith(K2); - AA->replaceWithNewValue(I, K1); - AA->replaceWithNewValue(J, K2); + L->replaceAllUsesWith(K1); + H->replaceAllUsesWith(K2); + AA->replaceWithNewValue(L, K1); + AA->replaceWithNewValue(H, K2); } // Instructions that may read from memory may be in the load move set. @@ -2387,6 +2959,9 @@ namespace { SE->forgetValue(J); I->eraseFromParent(); J->eraseFromParent(); + + DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" << + BB << "\n"); } DEBUG(dbgs() << "BBV: final: \n" << BB << "\n"); @@ -2397,6 +2972,7 @@ char BBVectorize::ID = 0; static const char bb_vectorize_name[] = "Basic-Block Vectorization"; INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_DEPENDENCY(DominatorTree) INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false) diff --git a/contrib/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/contrib/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp new file mode 100644 index 0000000..a7ef248 --- /dev/null +++ b/contrib/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp @@ -0,0 +1,1941 @@ +//===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops +// and generates target-independent LLVM-IR. Legalization of the IR is done +// in the codegen. However, the vectorizes uses (will use) the codegen +// interfaces to generate IR that is likely to result in an optimal binary. +// +// The loop vectorizer combines consecutive loop iteration into a single +// 'wide' iteration. After this transformation the index is incremented +// by the SIMD vector width, and not by one. +// +// This pass has three parts: +// 1. The main loop pass that drives the different parts. +// 2. LoopVectorizationLegality - A unit that checks for the legality +// of the vectorization. +// 3. SingleBlockLoopVectorizer - A unit that performs the actual +// widening of instructions. +// 4. LoopVectorizationCostModel - A unit that checks for the profitability +// of vectorization. It decides on the optimal vector width, which +// can be one, if vectorization is not profitable. +//===----------------------------------------------------------------------===// +// +// The reduction-variable vectorization is based on the paper: +// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. +// +// Variable uniformity checks are inspired by: +// Karrenberg, R. and Hack, S. Whole Function Vectorization. +// +// Other ideas/concepts are from: +// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. +// +//===----------------------------------------------------------------------===// +#define LV_NAME "loop-vectorize" +#define DEBUG_TYPE LV_NAME +#include "llvm/Constants.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Instructions.h" +#include "llvm/LLVMContext.h" +#include "llvm/Pass.h" +#include "llvm/Analysis/LoopPass.h" +#include "llvm/Value.h" +#include "llvm/Function.h" +#include "llvm/Analysis/Verifier.h" +#include "llvm/Module.h" +#include "llvm/Type.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/StringExtras.h" +#include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/AliasSetTracker.h" +#include "llvm/Analysis/ScalarEvolution.h" +#include "llvm/Analysis/Dominators.h" +#include "llvm/Analysis/ScalarEvolutionExpressions.h" +#include "llvm/Analysis/ScalarEvolutionExpander.h" +#include "llvm/Analysis/LoopInfo.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/Transforms/Scalar.h" +#include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/TargetTransformInfo.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/raw_ostream.h" +#include "llvm/DataLayout.h" +#include "llvm/Transforms/Utils/Local.h" +#include <algorithm> +using namespace llvm; + +static cl::opt<unsigned> +VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, + cl::desc("Set the default vectorization width. Zero is autoselect.")); + +/// We don't vectorize loops with a known constant trip count below this number. +const unsigned TinyTripCountThreshold = 16; + +/// When performing a runtime memory check, do not check more than this +/// number of pointers. Notice that the check is quadratic! +const unsigned RuntimeMemoryCheckThreshold = 2; + +namespace { + +// Forward declarations. +class LoopVectorizationLegality; +class LoopVectorizationCostModel; + +/// SingleBlockLoopVectorizer vectorizes loops which contain only one basic +/// block to a specified vectorization factor (VF). +/// This class performs the widening of scalars into vectors, or multiple +/// scalars. This class also implements the following features: +/// * It inserts an epilogue loop for handling loops that don't have iteration +/// counts that are known to be a multiple of the vectorization factor. +/// * It handles the code generation for reduction variables. +/// * Scalarization (implementation using scalars) of un-vectorizable +/// instructions. +/// SingleBlockLoopVectorizer does not perform any vectorization-legality +/// checks, and relies on the caller to check for the different legality +/// aspects. The SingleBlockLoopVectorizer relies on the +/// LoopVectorizationLegality class to provide information about the induction +/// and reduction variables that were found to a given vectorization factor. +class SingleBlockLoopVectorizer { +public: + /// Ctor. + SingleBlockLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li, + DominatorTree *dt, LPPassManager *Lpm, + unsigned VecWidth): + OrigLoop(Orig), SE(Se), LI(Li), DT(dt), LPM(Lpm), VF(VecWidth), + Builder(Se->getContext()), Induction(0), OldInduction(0) { } + + // Perform the actual loop widening (vectorization). + void vectorize(LoopVectorizationLegality *Legal) { + ///Create a new empty loop. Unlink the old loop and connect the new one. + createEmptyLoop(Legal); + /// Widen each instruction in the old loop to a new one in the new loop. + /// Use the Legality module to find the induction and reduction variables. + vectorizeLoop(Legal); + // Register the new loop and update the analysis passes. + updateAnalysis(); + } + +private: + /// Create an empty loop, based on the loop ranges of the old loop. + void createEmptyLoop(LoopVectorizationLegality *Legal); + /// Copy and widen the instructions from the old loop. + void vectorizeLoop(LoopVectorizationLegality *Legal); + /// Insert the new loop to the loop hierarchy and pass manager + /// and update the analysis passes. + void updateAnalysis(); + + /// This instruction is un-vectorizable. Implement it as a sequence + /// of scalars. + void scalarizeInstruction(Instruction *Instr); + + /// Create a broadcast instruction. This method generates a broadcast + /// instruction (shuffle) for loop invariant values and for the induction + /// value. If this is the induction variable then we extend it to N, N+1, ... + /// this is needed because each iteration in the loop corresponds to a SIMD + /// element. + Value *getBroadcastInstrs(Value *V); + + /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 .. + /// for each element in the vector. Starting from zero. + Value *getConsecutiveVector(Value* Val); + + /// When we go over instructions in the basic block we rely on previous + /// values within the current basic block or on loop invariant values. + /// When we widen (vectorize) values we place them in the map. If the values + /// are not within the map, they have to be loop invariant, so we simply + /// broadcast them into a vector. + Value *getVectorValue(Value *V); + + /// Get a uniform vector of constant integers. We use this to get + /// vectors of ones and zeros for the reduction code. + Constant* getUniformVector(unsigned Val, Type* ScalarTy); + + typedef DenseMap<Value*, Value*> ValueMap; + + /// The original loop. + Loop *OrigLoop; + // Scev analysis to use. + ScalarEvolution *SE; + // Loop Info. + LoopInfo *LI; + // Dominator Tree. + DominatorTree *DT; + // Loop Pass Manager; + LPPassManager *LPM; + // The vectorization factor to use. + unsigned VF; + + // The builder that we use + IRBuilder<> Builder; + + // --- Vectorization state --- + + /// The vector-loop preheader. + BasicBlock *LoopVectorPreHeader; + /// The scalar-loop preheader. + BasicBlock *LoopScalarPreHeader; + /// Middle Block between the vector and the scalar. + BasicBlock *LoopMiddleBlock; + ///The ExitBlock of the scalar loop. + BasicBlock *LoopExitBlock; + ///The vector loop body. + BasicBlock *LoopVectorBody; + ///The scalar loop body. + BasicBlock *LoopScalarBody; + ///The first bypass block. + BasicBlock *LoopBypassBlock; + + /// The new Induction variable which was added to the new block. + PHINode *Induction; + /// The induction variable of the old basic block. + PHINode *OldInduction; + // Maps scalars to widened vectors. + ValueMap WidenMap; +}; + +/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and +/// to what vectorization factor. +/// This class does not look at the profitability of vectorization, only the +/// legality. This class has two main kinds of checks: +/// * Memory checks - The code in canVectorizeMemory checks if vectorization +/// will change the order of memory accesses in a way that will change the +/// correctness of the program. +/// * Scalars checks - The code in canVectorizeBlock checks for a number +/// of different conditions, such as the availability of a single induction +/// variable, that all types are supported and vectorize-able, etc. +/// This code reflects the capabilities of SingleBlockLoopVectorizer. +/// This class is also used by SingleBlockLoopVectorizer for identifying +/// induction variable and the different reduction variables. +class LoopVectorizationLegality { +public: + LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl): + TheLoop(Lp), SE(Se), DL(Dl), Induction(0) { } + + /// This represents the kinds of reductions that we support. + enum ReductionKind { + NoReduction, /// Not a reduction. + IntegerAdd, /// Sum of numbers. + IntegerMult, /// Product of numbers. + IntegerOr, /// Bitwise or logical OR of numbers. + IntegerAnd, /// Bitwise or logical AND of numbers. + IntegerXor /// Bitwise or logical XOR of numbers. + }; + + /// This POD struct holds information about reduction variables. + struct ReductionDescriptor { + // Default C'tor + ReductionDescriptor(): + StartValue(0), LoopExitInstr(0), Kind(NoReduction) {} + + // C'tor. + ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K): + StartValue(Start), LoopExitInstr(Exit), Kind(K) {} + + // The starting value of the reduction. + // It does not have to be zero! + Value *StartValue; + // The instruction who's value is used outside the loop. + Instruction *LoopExitInstr; + // The kind of the reduction. + ReductionKind Kind; + }; + + // This POD struct holds information about the memory runtime legality + // check that a group of pointers do not overlap. + struct RuntimePointerCheck { + /// This flag indicates if we need to add the runtime check. + bool Need; + /// Holds the pointers that we need to check. + SmallVector<Value*, 2> Pointers; + }; + + /// ReductionList contains the reduction descriptors for all + /// of the reductions that were found in the loop. + typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList; + + /// Returns true if it is legal to vectorize this loop. + /// This does not mean that it is profitable to vectorize this + /// loop, only that it is legal to do so. + bool canVectorize(); + + /// Returns the Induction variable. + PHINode *getInduction() {return Induction;} + + /// Returns the reduction variables found in the loop. + ReductionList *getReductionVars() { return &Reductions; } + + /// Check if the pointer returned by this GEP is consecutive + /// when the index is vectorized. This happens when the last + /// index of the GEP is consecutive, like the induction variable. + /// This check allows us to vectorize A[idx] into a wide load/store. + bool isConsecutiveGep(Value *Ptr); + + /// Returns true if the value V is uniform within the loop. + bool isUniform(Value *V); + + /// Returns true if this instruction will remain scalar after vectorization. + bool isUniformAfterVectorization(Instruction* I) {return Uniforms.count(I);} + + /// Returns the information that we collected about runtime memory check. + RuntimePointerCheck *getRuntimePointerCheck() {return &PtrRtCheck; } +private: + /// Check if a single basic block loop is vectorizable. + /// At this point we know that this is a loop with a constant trip count + /// and we only need to check individual instructions. + bool canVectorizeBlock(BasicBlock &BB); + + /// When we vectorize loops we may change the order in which + /// we read and write from memory. This method checks if it is + /// legal to vectorize the code, considering only memory constrains. + /// Returns true if BB is vectorizable + bool canVectorizeMemory(BasicBlock &BB); + + /// Returns True, if 'Phi' is the kind of reduction variable for type + /// 'Kind'. If this is a reduction variable, it adds it to ReductionList. + bool AddReductionVar(PHINode *Phi, ReductionKind Kind); + /// Returns true if the instruction I can be a reduction variable of type + /// 'Kind'. + bool isReductionInstr(Instruction *I, ReductionKind Kind); + /// Returns True, if 'Phi' is an induction variable. + bool isInductionVariable(PHINode *Phi); + /// Return true if can compute the address bounds of Ptr within the loop. + bool hasComputableBounds(Value *Ptr); + + /// The loop that we evaluate. + Loop *TheLoop; + /// Scev analysis. + ScalarEvolution *SE; + /// DataLayout analysis. + DataLayout *DL; + + // --- vectorization state --- // + + /// Holds the induction variable. + PHINode *Induction; + /// Holds the reduction variables. + ReductionList Reductions; + /// Allowed outside users. This holds the reduction + /// vars which can be accessed from outside the loop. + SmallPtrSet<Value*, 4> AllowedExit; + /// This set holds the variables which are known to be uniform after + /// vectorization. + SmallPtrSet<Instruction*, 4> Uniforms; + /// We need to check that all of the pointers in this list are disjoint + /// at runtime. + RuntimePointerCheck PtrRtCheck; +}; + +/// LoopVectorizationCostModel - estimates the expected speedups due to +/// vectorization. +/// In many cases vectorization is not profitable. This can happen because +/// of a number of reasons. In this class we mainly attempt to predict +/// the expected speedup/slowdowns due to the supported instruction set. +/// We use the VectorTargetTransformInfo to query the different backends +/// for the cost of different operations. +class LoopVectorizationCostModel { +public: + /// C'tor. + LoopVectorizationCostModel(Loop *Lp, ScalarEvolution *Se, + LoopVectorizationLegality *Leg, + const VectorTargetTransformInfo *Vtti): + TheLoop(Lp), SE(Se), Legal(Leg), VTTI(Vtti) { } + + /// Returns the most profitable vectorization factor for the loop that is + /// smaller or equal to the VF argument. This method checks every power + /// of two up to VF. + unsigned findBestVectorizationFactor(unsigned VF = 8); + +private: + /// Returns the expected execution cost. The unit of the cost does + /// not matter because we use the 'cost' units to compare different + /// vector widths. The cost that is returned is *not* normalized by + /// the factor width. + unsigned expectedCost(unsigned VF); + + /// Returns the execution time cost of an instruction for a given vector + /// width. Vector width of one means scalar. + unsigned getInstructionCost(Instruction *I, unsigned VF); + + /// A helper function for converting Scalar types to vector types. + /// If the incoming type is void, we return void. If the VF is 1, we return + /// the scalar type. + static Type* ToVectorTy(Type *Scalar, unsigned VF); + + /// The loop that we evaluate. + Loop *TheLoop; + /// Scev analysis. + ScalarEvolution *SE; + + /// Vectorization legality. + LoopVectorizationLegality *Legal; + /// Vector target information. + const VectorTargetTransformInfo *VTTI; +}; + +struct LoopVectorize : public LoopPass { + static char ID; // Pass identification, replacement for typeid + + LoopVectorize() : LoopPass(ID) { + initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); + } + + ScalarEvolution *SE; + DataLayout *DL; + LoopInfo *LI; + TargetTransformInfo *TTI; + DominatorTree *DT; + + virtual bool runOnLoop(Loop *L, LPPassManager &LPM) { + // We only vectorize innermost loops. + if (!L->empty()) + return false; + + SE = &getAnalysis<ScalarEvolution>(); + DL = getAnalysisIfAvailable<DataLayout>(); + LI = &getAnalysis<LoopInfo>(); + TTI = getAnalysisIfAvailable<TargetTransformInfo>(); + DT = &getAnalysis<DominatorTree>(); + + DEBUG(dbgs() << "LV: Checking a loop in \"" << + L->getHeader()->getParent()->getName() << "\"\n"); + + // Check if it is legal to vectorize the loop. + LoopVectorizationLegality LVL(L, SE, DL); + if (!LVL.canVectorize()) { + DEBUG(dbgs() << "LV: Not vectorizing.\n"); + return false; + } + + // Select the preffered vectorization factor. + unsigned VF = 1; + if (VectorizationFactor == 0) { + const VectorTargetTransformInfo *VTTI = 0; + if (TTI) + VTTI = TTI->getVectorTargetTransformInfo(); + // Use the cost model. + LoopVectorizationCostModel CM(L, SE, &LVL, VTTI); + VF = CM.findBestVectorizationFactor(); + + if (VF == 1) { + DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); + return false; + } + + } else { + // Use the user command flag. + VF = VectorizationFactor; + } + + DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<< + L->getHeader()->getParent()->getParent()->getModuleIdentifier()<< + "\n"); + + // If we decided that it is *legal* to vectorizer the loop then do it. + SingleBlockLoopVectorizer LB(L, SE, LI, DT, &LPM, VF); + LB.vectorize(&LVL); + + DEBUG(verifyFunction(*L->getHeader()->getParent())); + return true; + } + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + LoopPass::getAnalysisUsage(AU); + AU.addRequiredID(LoopSimplifyID); + AU.addRequiredID(LCSSAID); + AU.addRequired<LoopInfo>(); + AU.addRequired<ScalarEvolution>(); + AU.addRequired<DominatorTree>(); + AU.addPreserved<LoopInfo>(); + AU.addPreserved<DominatorTree>(); + } + +}; + +Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) { + // Instructions that access the old induction variable + // actually want to get the new one. + if (V == OldInduction) + V = Induction; + // Create the types. + LLVMContext &C = V->getContext(); + Type *VTy = VectorType::get(V->getType(), VF); + Type *I32 = IntegerType::getInt32Ty(C); + Constant *Zero = ConstantInt::get(I32, 0); + Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF)); + Value *UndefVal = UndefValue::get(VTy); + // Insert the value into a new vector. + Value *SingleElem = Builder.CreateInsertElement(UndefVal, V, Zero); + // Broadcast the scalar into all locations in the vector. + Value *Shuf = Builder.CreateShuffleVector(SingleElem, UndefVal, Zeros, + "broadcast"); + // We are accessing the induction variable. Make sure to promote the + // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes. + if (V == Induction) + return getConsecutiveVector(Shuf); + return Shuf; +} + +Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) { + assert(Val->getType()->isVectorTy() && "Must be a vector"); + assert(Val->getType()->getScalarType()->isIntegerTy() && + "Elem must be an integer"); + // Create the types. + Type *ITy = Val->getType()->getScalarType(); + VectorType *Ty = cast<VectorType>(Val->getType()); + unsigned VLen = Ty->getNumElements(); + SmallVector<Constant*, 8> Indices; + + // Create a vector of consecutive numbers from zero to VF. + for (unsigned i = 0; i < VLen; ++i) + Indices.push_back(ConstantInt::get(ITy, i)); + + // Add the consecutive indices to the vector value. + Constant *Cv = ConstantVector::get(Indices); + assert(Cv->getType() == Val->getType() && "Invalid consecutive vec"); + return Builder.CreateAdd(Val, Cv, "induction"); +} + +bool LoopVectorizationLegality::isConsecutiveGep(Value *Ptr) { + GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr); + if (!Gep) + return false; + + unsigned NumOperands = Gep->getNumOperands(); + Value *LastIndex = Gep->getOperand(NumOperands - 1); + + // Check that all of the gep indices are uniform except for the last. + for (unsigned i = 0; i < NumOperands - 1; ++i) + if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop)) + return false; + + // We can emit wide load/stores only of the last index is the induction + // variable. + const SCEV *Last = SE->getSCEV(LastIndex); + if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) { + const SCEV *Step = AR->getStepRecurrence(*SE); + + // The memory is consecutive because the last index is consecutive + // and all other indices are loop invariant. + if (Step->isOne()) + return true; + } + + return false; +} + +bool LoopVectorizationLegality::isUniform(Value *V) { + return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); +} + +Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) { + assert(!V->getType()->isVectorTy() && "Can't widen a vector"); + // If we saved a vectorized copy of V, use it. + Value *&MapEntry = WidenMap[V]; + if (MapEntry) + return MapEntry; + + // Broadcast V and save the value for future uses. + Value *B = getBroadcastInstrs(V); + MapEntry = B; + return B; +} + +Constant* +SingleBlockLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) { + SmallVector<Constant*, 8> Indices; + // Create a vector of consecutive numbers from zero to VF. + for (unsigned i = 0; i < VF; ++i) + Indices.push_back(ConstantInt::get(ScalarTy, Val, true)); + + // Add the consecutive indices to the vector value. + return ConstantVector::get(Indices); +} + +void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) { + assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); + // Holds vector parameters or scalars, in case of uniform vals. + SmallVector<Value*, 8> Params; + + // Find all of the vectorized parameters. + for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { + Value *SrcOp = Instr->getOperand(op); + + // If we are accessing the old induction variable, use the new one. + if (SrcOp == OldInduction) { + Params.push_back(getBroadcastInstrs(Induction)); + continue; + } + + // Try using previously calculated values. + Instruction *SrcInst = dyn_cast<Instruction>(SrcOp); + + // If the src is an instruction that appeared earlier in the basic block + // then it should already be vectorized. + if (SrcInst && SrcInst->getParent() == Instr->getParent()) { + assert(WidenMap.count(SrcInst) && "Source operand is unavailable"); + // The parameter is a vector value from earlier. + Params.push_back(WidenMap[SrcInst]); + } else { + // The parameter is a scalar from outside the loop. Maybe even a constant. + Params.push_back(SrcOp); + } + } + + assert(Params.size() == Instr->getNumOperands() && + "Invalid number of operands"); + + // Does this instruction return a value ? + bool IsVoidRetTy = Instr->getType()->isVoidTy(); + Value *VecResults = 0; + + // If we have a return value, create an empty vector. We place the scalarized + // instructions in this vector. + if (!IsVoidRetTy) + VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF)); + + // For each scalar that we create: + for (unsigned i = 0; i < VF; ++i) { + Instruction *Cloned = Instr->clone(); + if (!IsVoidRetTy) + Cloned->setName(Instr->getName() + ".cloned"); + // Replace the operands of the cloned instrucions with extracted scalars. + for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { + Value *Op = Params[op]; + // Param is a vector. Need to extract the right lane. + if (Op->getType()->isVectorTy()) + Op = Builder.CreateExtractElement(Op, Builder.getInt32(i)); + Cloned->setOperand(op, Op); + } + + // Place the cloned scalar in the new loop. + Builder.Insert(Cloned); + + // If the original scalar returns a value we need to place it in a vector + // so that future users will be able to use it. + if (!IsVoidRetTy) + VecResults = Builder.CreateInsertElement(VecResults, Cloned, + Builder.getInt32(i)); + } + + if (!IsVoidRetTy) + WidenMap[Instr] = VecResults; +} + +void +SingleBlockLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) { + /* + In this function we generate a new loop. The new loop will contain + the vectorized instructions while the old loop will continue to run the + scalar remainder. + + [ ] <-- vector loop bypass. + / | + / v +| [ ] <-- vector pre header. +| | +| v +| [ ] \ +| [ ]_| <-- vector loop. +| | + \ v + >[ ] <--- middle-block. + / | + / v +| [ ] <--- new preheader. +| | +| v +| [ ] \ +| [ ]_| <-- old scalar loop to handle remainder. + \ | + \ v + >[ ] <-- exit block. + ... + */ + + OldInduction = Legal->getInduction(); + assert(OldInduction && "We must have a single phi node."); + Type *IdxTy = OldInduction->getType(); + + // Find the loop boundaries. + const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getHeader()); + assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count"); + + // Get the total trip count from the count by adding 1. + ExitCount = SE->getAddExpr(ExitCount, + SE->getConstant(ExitCount->getType(), 1)); + // We may need to extend the index in case there is a type mismatch. + // We know that the count starts at zero and does not overflow. + // We are using Zext because it should be less expensive. + if (ExitCount->getType() != IdxTy) + ExitCount = SE->getZeroExtendExpr(ExitCount, IdxTy); + + // This is the original scalar-loop preheader. + BasicBlock *BypassBlock = OrigLoop->getLoopPreheader(); + BasicBlock *ExitBlock = OrigLoop->getExitBlock(); + assert(ExitBlock && "Must have an exit block"); + + // The loop index does not have to start at Zero. It starts with this value. + Value *StartIdx = OldInduction->getIncomingValueForBlock(BypassBlock); + + assert(OrigLoop->getNumBlocks() == 1 && "Invalid loop"); + assert(BypassBlock && "Invalid loop structure"); + + BasicBlock *VectorPH = + BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph"); + BasicBlock *VecBody = VectorPH->splitBasicBlock(VectorPH->getTerminator(), + "vector.body"); + + BasicBlock *MiddleBlock = VecBody->splitBasicBlock(VecBody->getTerminator(), + "middle.block"); + BasicBlock *ScalarPH = + MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), + "scalar.preheader"); + // Find the induction variable. + BasicBlock *OldBasicBlock = OrigLoop->getHeader(); + + // Use this IR builder to create the loop instructions (Phi, Br, Cmp) + // inside the loop. + Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); + + // Generate the induction variable. + Induction = Builder.CreatePHI(IdxTy, 2, "index"); + Constant *Step = ConstantInt::get(IdxTy, VF); + + // Expand the trip count and place the new instructions in the preheader. + // Notice that the pre-header does not change, only the loop body. + SCEVExpander Exp(*SE, "induction"); + Instruction *Loc = BypassBlock->getTerminator(); + + // Count holds the overall loop count (N). + Value *Count = Exp.expandCodeFor(ExitCount, Induction->getType(), Loc); + + // Add the start index to the loop count to get the new end index. + Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc); + + // Now we need to generate the expression for N - (N % VF), which is + // the part that the vectorized body will execute. + Constant *CIVF = ConstantInt::get(IdxTy, VF); + Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc); + Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc); + Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx, + "end.idx.rnd.down", Loc); + + // Now, compare the new count to zero. If it is zero, jump to the scalar part. + Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, + IdxEndRoundDown, + StartIdx, + "cmp.zero", Loc); + + LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck = + Legal->getRuntimePointerCheck(); + Value *MemoryRuntimeCheck = 0; + if (PtrRtCheck->Need) { + unsigned NumPointers = PtrRtCheck->Pointers.size(); + SmallVector<Value* , 2> Starts; + SmallVector<Value* , 2> Ends; + + // Use this type for pointer arithmetic. + Type* PtrArithTy = PtrRtCheck->Pointers[0]->getType(); + + for (unsigned i=0; i < NumPointers; ++i) { + Value *Ptr = PtrRtCheck->Pointers[i]; + const SCEV *Sc = SE->getSCEV(Ptr); + + if (SE->isLoopInvariant(Sc, OrigLoop)) { + DEBUG(dbgs() << "LV1: Adding RT check for a loop invariant ptr:" << + *Ptr <<"\n"); + Starts.push_back(Ptr); + Ends.push_back(Ptr); + } else { + DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr <<"\n"); + const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); + Value *Start = Exp.expandCodeFor(AR->getStart(), PtrArithTy, Loc); + const SCEV *Ex = SE->getExitCount(OrigLoop, OrigLoop->getHeader()); + const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE); + assert(!isa<SCEVCouldNotCompute>(ScEnd) && "Invalid scev range."); + Value *End = Exp.expandCodeFor(ScEnd, PtrArithTy, Loc); + Starts.push_back(Start); + Ends.push_back(End); + } + } + + for (unsigned i=0; i < NumPointers; ++i) { + for (unsigned j=i+1; j < NumPointers; ++j) { + Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE, + Starts[0], Ends[1], "bound0", Loc); + Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE, + Starts[1], Ends[0], "bound1", Loc); + Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1, + "found.conflict", Loc); + if (MemoryRuntimeCheck) { + MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or, + MemoryRuntimeCheck, + IsConflict, + "conflict.rdx", Loc); + } else { + MemoryRuntimeCheck = IsConflict; + } + } + } + }// end of need-runtime-check code. + + // If we are using memory runtime checks, include them in. + if (MemoryRuntimeCheck) { + Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck, + "CntOrMem", Loc); + } + + BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc); + // Remove the old terminator. + Loc->eraseFromParent(); + + // We are going to resume the execution of the scalar loop. + // This PHI decides on what number to start. If we come from the + // vector loop then we need to start with the end index minus the + // index modulo VF. If we come from a bypass edge then we need to start + // from the real start. + PHINode* ResumeIndex = PHINode::Create(IdxTy, 2, "resume.idx", + MiddleBlock->getTerminator()); + ResumeIndex->addIncoming(StartIdx, BypassBlock); + ResumeIndex->addIncoming(IdxEndRoundDown, VecBody); + + // Add a check in the middle block to see if we have completed + // all of the iterations in the first vector loop. + // If (N - N%VF) == N, then we *don't* need to run the remainder. + Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd, + ResumeIndex, "cmp.n", + MiddleBlock->getTerminator()); + + BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator()); + // Remove the old terminator. + MiddleBlock->getTerminator()->eraseFromParent(); + + // Create i+1 and fill the PHINode. + Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next"); + Induction->addIncoming(StartIdx, VectorPH); + Induction->addIncoming(NextIdx, VecBody); + // Create the compare. + Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown); + Builder.CreateCondBr(ICmp, MiddleBlock, VecBody); + + // Now we have two terminators. Remove the old one from the block. + VecBody->getTerminator()->eraseFromParent(); + + // Fix the scalar body iteration count. + unsigned BlockIdx = OldInduction->getBasicBlockIndex(ScalarPH); + OldInduction->setIncomingValue(BlockIdx, ResumeIndex); + + // Get ready to start creating new instructions into the vectorized body. + Builder.SetInsertPoint(VecBody->getFirstInsertionPt()); + + // Register the new loop. + Loop* Lp = new Loop(); + LPM->insertLoop(Lp, OrigLoop->getParentLoop()); + + Lp->addBasicBlockToLoop(VecBody, LI->getBase()); + + Loop *ParentLoop = OrigLoop->getParentLoop(); + if (ParentLoop) { + ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase()); + ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase()); + ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase()); + } + + // Save the state. + LoopVectorPreHeader = VectorPH; + LoopScalarPreHeader = ScalarPH; + LoopMiddleBlock = MiddleBlock; + LoopExitBlock = ExitBlock; + LoopVectorBody = VecBody; + LoopScalarBody = OldBasicBlock; + LoopBypassBlock = BypassBlock; +} + +/// This function returns the identity element (or neutral element) for +/// the operation K. +static unsigned +getReductionIdentity(LoopVectorizationLegality::ReductionKind K) { + switch (K) { + case LoopVectorizationLegality::IntegerXor: + case LoopVectorizationLegality::IntegerAdd: + case LoopVectorizationLegality::IntegerOr: + // Adding, Xoring, Oring zero to a number does not change it. + return 0; + case LoopVectorizationLegality::IntegerMult: + // Multiplying a number by 1 does not change it. + return 1; + case LoopVectorizationLegality::IntegerAnd: + // AND-ing a number with an all-1 value does not change it. + return -1; + default: + llvm_unreachable("Unknown reduction kind"); + } +} + +void +SingleBlockLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) { + //===------------------------------------------------===// + // + // Notice: any optimization or new instruction that go + // into the code below should be also be implemented in + // the cost-model. + // + //===------------------------------------------------===// + typedef SmallVector<PHINode*, 4> PhiVector; + BasicBlock &BB = *OrigLoop->getHeader(); + Constant *Zero = ConstantInt::get( + IntegerType::getInt32Ty(BB.getContext()), 0); + + // In order to support reduction variables we need to be able to vectorize + // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two + // steages. First, we create a new vector PHI node with no incoming edges. + // We use this value when we vectorize all of the instructions that use the + // PHI. Next, after all of the instructions in the block are complete we + // add the new incoming edges to the PHI. At this point all of the + // instructions in the basic block are vectorized, so we can use them to + // construct the PHI. + PhiVector PHIsToFix; + + // For each instruction in the old loop. + for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) { + Instruction *Inst = it; + + switch (Inst->getOpcode()) { + case Instruction::Br: + // Nothing to do for PHIs and BR, since we already took care of the + // loop control flow instructions. + continue; + case Instruction::PHI:{ + PHINode* P = cast<PHINode>(Inst); + // Special handling for the induction var. + if (OldInduction == Inst) + continue; + // This is phase one of vectorizing PHIs. + // This has to be a reduction variable. + assert(Legal->getReductionVars()->count(P) && "Not a Reduction"); + Type *VecTy = VectorType::get(Inst->getType(), VF); + WidenMap[Inst] = Builder.CreatePHI(VecTy, 2, "vec.phi"); + PHIsToFix.push_back(P); + continue; + } + case Instruction::Add: + case Instruction::FAdd: + case Instruction::Sub: + case Instruction::FSub: + case Instruction::Mul: + case Instruction::FMul: + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: { + // Just widen binops. + BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst); + Value *A = getVectorValue(Inst->getOperand(0)); + Value *B = getVectorValue(Inst->getOperand(1)); + + // Use this vector value for all users of the original instruction. + Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B); + WidenMap[Inst] = V; + + // Update the NSW, NUW and Exact flags. + BinaryOperator *VecOp = cast<BinaryOperator>(V); + if (isa<OverflowingBinaryOperator>(BinOp)) { + VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); + VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); + } + if (isa<PossiblyExactOperator>(VecOp)) + VecOp->setIsExact(BinOp->isExact()); + break; + } + case Instruction::Select: { + // Widen selects. + // If the selector is loop invariant we can create a select + // instruction with a scalar condition. Otherwise, use vector-select. + Value *Cond = Inst->getOperand(0); + bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop); + + // The condition can be loop invariant but still defined inside the + // loop. This means that we can't just use the original 'cond' value. + // We have to take the 'vectorized' value and pick the first lane. + // Instcombine will make this a no-op. + Cond = getVectorValue(Cond); + if (InvariantCond) + Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0)); + + Value *Op0 = getVectorValue(Inst->getOperand(1)); + Value *Op1 = getVectorValue(Inst->getOperand(2)); + WidenMap[Inst] = Builder.CreateSelect(Cond, Op0, Op1); + break; + } + + case Instruction::ICmp: + case Instruction::FCmp: { + // Widen compares. Generate vector compares. + bool FCmp = (Inst->getOpcode() == Instruction::FCmp); + CmpInst *Cmp = dyn_cast<CmpInst>(Inst); + Value *A = getVectorValue(Inst->getOperand(0)); + Value *B = getVectorValue(Inst->getOperand(1)); + if (FCmp) + WidenMap[Inst] = Builder.CreateFCmp(Cmp->getPredicate(), A, B); + else + WidenMap[Inst] = Builder.CreateICmp(Cmp->getPredicate(), A, B); + break; + } + + case Instruction::Store: { + // Attempt to issue a wide store. + StoreInst *SI = dyn_cast<StoreInst>(Inst); + Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF); + Value *Ptr = SI->getPointerOperand(); + unsigned Alignment = SI->getAlignment(); + + assert(!Legal->isUniform(Ptr) && + "We do not allow storing to uniform addresses"); + + GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); + + // This store does not use GEPs. + if (!Legal->isConsecutiveGep(Gep)) { + scalarizeInstruction(Inst); + break; + } + + // The last index does not have to be the induction. It can be + // consecutive and be a function of the index. For example A[I+1]; + unsigned NumOperands = Gep->getNumOperands(); + Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1)); + LastIndex = Builder.CreateExtractElement(LastIndex, Zero); + + // Create the new GEP with the new induction variable. + GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); + Gep2->setOperand(NumOperands - 1, LastIndex); + Ptr = Builder.Insert(Gep2); + Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo()); + Value *Val = getVectorValue(SI->getValueOperand()); + Builder.CreateStore(Val, Ptr)->setAlignment(Alignment); + break; + } + case Instruction::Load: { + // Attempt to issue a wide load. + LoadInst *LI = dyn_cast<LoadInst>(Inst); + Type *RetTy = VectorType::get(LI->getType(), VF); + Value *Ptr = LI->getPointerOperand(); + unsigned Alignment = LI->getAlignment(); + GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); + + // If we don't have a gep, or that the pointer is loop invariant, + // scalarize the load. + if (!Gep || Legal->isUniform(Gep) || !Legal->isConsecutiveGep(Gep)) { + scalarizeInstruction(Inst); + break; + } + + // The last index does not have to be the induction. It can be + // consecutive and be a function of the index. For example A[I+1]; + unsigned NumOperands = Gep->getNumOperands(); + Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1)); + LastIndex = Builder.CreateExtractElement(LastIndex, Zero); + + // Create the new GEP with the new induction variable. + GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone()); + Gep2->setOperand(NumOperands - 1, LastIndex); + Ptr = Builder.Insert(Gep2); + Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo()); + LI = Builder.CreateLoad(Ptr); + LI->setAlignment(Alignment); + // Use this vector value for all users of the load. + WidenMap[Inst] = LI; + break; + } + case Instruction::ZExt: + case Instruction::SExt: + case Instruction::FPToUI: + case Instruction::FPToSI: + case Instruction::FPExt: + case Instruction::PtrToInt: + case Instruction::IntToPtr: + case Instruction::SIToFP: + case Instruction::UIToFP: + case Instruction::Trunc: + case Instruction::FPTrunc: + case Instruction::BitCast: { + /// Vectorize bitcasts. + CastInst *CI = dyn_cast<CastInst>(Inst); + Value *A = getVectorValue(Inst->getOperand(0)); + Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); + WidenMap[Inst] = Builder.CreateCast(CI->getOpcode(), A, DestTy); + break; + } + + default: + /// All other instructions are unsupported. Scalarize them. + scalarizeInstruction(Inst); + break; + }// end of switch. + }// end of for_each instr. + + // At this point every instruction in the original loop is widended to + // a vector form. We are almost done. Now, we need to fix the PHI nodes + // that we vectorized. The PHI nodes are currently empty because we did + // not want to introduce cycles. Notice that the remaining PHI nodes + // that we need to fix are reduction variables. + + // Create the 'reduced' values for each of the induction vars. + // The reduced values are the vector values that we scalarize and combine + // after the loop is finished. + for (PhiVector::iterator it = PHIsToFix.begin(), e = PHIsToFix.end(); + it != e; ++it) { + PHINode *RdxPhi = *it; + PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]); + assert(RdxPhi && "Unable to recover vectorized PHI"); + + // Find the reduction variable descriptor. + assert(Legal->getReductionVars()->count(RdxPhi) && + "Unable to find the reduction variable"); + LoopVectorizationLegality::ReductionDescriptor RdxDesc = + (*Legal->getReductionVars())[RdxPhi]; + + // We need to generate a reduction vector from the incoming scalar. + // To do so, we need to generate the 'identity' vector and overide + // one of the elements with the incoming scalar reduction. We need + // to do it in the vector-loop preheader. + Builder.SetInsertPoint(LoopBypassBlock->getTerminator()); + + // This is the vector-clone of the value that leaves the loop. + Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr); + Type *VecTy = VectorExit->getType(); + + // Find the reduction identity variable. Zero for addition, or, xor, + // one for multiplication, -1 for And. + Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind), + VecTy->getScalarType()); + + // This vector is the Identity vector where the first element is the + // incoming scalar reduction. + Value *VectorStart = Builder.CreateInsertElement(Identity, + RdxDesc.StartValue, Zero); + + + // Fix the vector-loop phi. + // We created the induction variable so we know that the + // preheader is the first entry. + BasicBlock *VecPreheader = Induction->getIncomingBlock(0); + + // Reductions do not have to start at zero. They can start with + // any loop invariant values. + VecRdxPhi->addIncoming(VectorStart, VecPreheader); + unsigned SelfEdgeIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody); + Value *Val = getVectorValue(RdxPhi->getIncomingValue(SelfEdgeIdx)); + VecRdxPhi->addIncoming(Val, LoopVectorBody); + + // Before each round, move the insertion point right between + // the PHIs and the values we are going to write. + // This allows us to write both PHINodes and the extractelement + // instructions. + Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); + + // This PHINode contains the vectorized reduction variable, or + // the initial value vector, if we bypass the vector loop. + PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); + NewPhi->addIncoming(VectorStart, LoopBypassBlock); + NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody); + + // Extract the first scalar. + Value *Scalar0 = + Builder.CreateExtractElement(NewPhi, Builder.getInt32(0)); + // Extract and reduce the remaining vector elements. + for (unsigned i=1; i < VF; ++i) { + Value *Scalar1 = + Builder.CreateExtractElement(NewPhi, Builder.getInt32(i)); + switch (RdxDesc.Kind) { + case LoopVectorizationLegality::IntegerAdd: + Scalar0 = Builder.CreateAdd(Scalar0, Scalar1); + break; + case LoopVectorizationLegality::IntegerMult: + Scalar0 = Builder.CreateMul(Scalar0, Scalar1); + break; + case LoopVectorizationLegality::IntegerOr: + Scalar0 = Builder.CreateOr(Scalar0, Scalar1); + break; + case LoopVectorizationLegality::IntegerAnd: + Scalar0 = Builder.CreateAnd(Scalar0, Scalar1); + break; + case LoopVectorizationLegality::IntegerXor: + Scalar0 = Builder.CreateXor(Scalar0, Scalar1); + break; + default: + llvm_unreachable("Unknown reduction operation"); + } + } + + // Now, we need to fix the users of the reduction variable + // inside and outside of the scalar remainder loop. + // We know that the loop is in LCSSA form. We need to update the + // PHI nodes in the exit blocks. + for (BasicBlock::iterator LEI = LoopExitBlock->begin(), + LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) { + PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI); + if (!LCSSAPhi) continue; + + // All PHINodes need to have a single entry edge, or two if + // we already fixed them. + assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI"); + + // We found our reduction value exit-PHI. Update it with the + // incoming bypass edge. + if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) { + // Add an edge coming from the bypass. + LCSSAPhi->addIncoming(Scalar0, LoopMiddleBlock); + break; + } + }// end of the LCSSA phi scan. + + // Fix the scalar loop reduction variable with the incoming reduction sum + // from the vector body and from the backedge value. + int IncomingEdgeBlockIdx = (RdxPhi)->getBasicBlockIndex(LoopScalarBody); + int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); // The other block. + (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, Scalar0); + (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr); + }// end of for each redux variable. +} + +void SingleBlockLoopVectorizer::updateAnalysis() { + // The original basic block. + SE->forgetLoop(OrigLoop); + + // Update the dominator tree information. + assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) && + "Entry does not dominate exit."); + + DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock); + DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader); + DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock); + DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock); + DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader); + DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); + + DEBUG(DT->verifyAnalysis()); +} + +bool LoopVectorizationLegality::canVectorize() { + if (!TheLoop->getLoopPreheader()) { + assert(false && "No preheader!!"); + DEBUG(dbgs() << "LV: Loop not normalized." << "\n"); + return false; + } + + // We can only vectorize single basic block loops. + unsigned NumBlocks = TheLoop->getNumBlocks(); + if (NumBlocks != 1) { + DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n"); + return false; + } + + // We need to have a loop header. + BasicBlock *BB = TheLoop->getHeader(); + DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n"); + + // ScalarEvolution needs to be able to find the exit count. + const SCEV *ExitCount = SE->getExitCount(TheLoop, BB); + if (ExitCount == SE->getCouldNotCompute()) { + DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n"); + return false; + } + + // Do not loop-vectorize loops with a tiny trip count. + unsigned TC = SE->getSmallConstantTripCount(TheLoop, BB); + if (TC > 0u && TC < TinyTripCountThreshold) { + DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << + "This loop is not worth vectorizing.\n"); + return false; + } + + // Go over each instruction and look at memory deps. + if (!canVectorizeBlock(*BB)) { + DEBUG(dbgs() << "LV: Can't vectorize this loop header\n"); + return false; + } + + DEBUG(dbgs() << "LV: We can vectorize this loop" << + (PtrRtCheck.Need ? " (with a runtime bound check)" : "") + <<"!\n"); + + // Okay! We can vectorize. At this point we don't have any other mem analysis + // which may limit our maximum vectorization factor, so just return true with + // no restrictions. + return true; +} + +bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) { + // Scan the instructions in the block and look for hazards. + for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) { + Instruction *I = it; + + PHINode *Phi = dyn_cast<PHINode>(I); + if (Phi) { + // This should not happen because the loop should be normalized. + if (Phi->getNumIncomingValues() != 2) { + DEBUG(dbgs() << "LV: Found an invalid PHI.\n"); + return false; + } + // We only look at integer phi nodes. + if (!Phi->getType()->isIntegerTy()) { + DEBUG(dbgs() << "LV: Found an non-int PHI.\n"); + return false; + } + + if (isInductionVariable(Phi)) { + if (Induction) { + DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n"); + return false; + } + DEBUG(dbgs() << "LV: Found the induction PHI."<< *Phi <<"\n"); + Induction = Phi; + continue; + } + if (AddReductionVar(Phi, IntegerAdd)) { + DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, IntegerMult)) { + DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, IntegerOr)) { + DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, IntegerAnd)) { + DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n"); + continue; + } + if (AddReductionVar(Phi, IntegerXor)) { + DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n"); + continue; + } + + DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n"); + return false; + }// end of PHI handling + + // We still don't handle functions. + CallInst *CI = dyn_cast<CallInst>(I); + if (CI) { + DEBUG(dbgs() << "LV: Found a call site.\n"); + return false; + } + + // We do not re-vectorize vectors. + if (!VectorType::isValidElementType(I->getType()) && + !I->getType()->isVoidTy()) { + DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n"); + return false; + } + + // Reduction instructions are allowed to have exit users. + // All other instructions must not have external users. + if (!AllowedExit.count(I)) + //Check that all of the users of the loop are inside the BB. + for (Value::use_iterator it = I->use_begin(), e = I->use_end(); + it != e; ++it) { + Instruction *U = cast<Instruction>(*it); + // This user may be a reduction exit value. + BasicBlock *Parent = U->getParent(); + if (Parent != &BB) { + DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n"); + return false; + } + } + } // next instr. + + if (!Induction) { + DEBUG(dbgs() << "LV: Did not find an induction var.\n"); + return false; + } + + // Don't vectorize if the memory dependencies do not allow vectorization. + if (!canVectorizeMemory(BB)) + return false; + + // We now know that the loop is vectorizable! + // Collect variables that will remain uniform after vectorization. + std::vector<Value*> Worklist; + + // Start with the conditional branch and walk up the block. + Worklist.push_back(BB.getTerminator()->getOperand(0)); + + while (Worklist.size()) { + Instruction *I = dyn_cast<Instruction>(Worklist.back()); + Worklist.pop_back(); + // Look at instructions inside this block. + if (!I) continue; + if (I->getParent() != &BB) continue; + + // Stop when reaching PHI nodes. + if (isa<PHINode>(I)) { + assert(I == Induction && "Found a uniform PHI that is not the induction"); + break; + } + + // This is a known uniform. + Uniforms.insert(I); + + // Insert all operands. + for (int i=0, Op = I->getNumOperands(); i < Op; ++i) { + Worklist.push_back(I->getOperand(i)); + } + } + + return true; +} + +bool LoopVectorizationLegality::canVectorizeMemory(BasicBlock &BB) { + typedef SmallVector<Value*, 16> ValueVector; + typedef SmallPtrSet<Value*, 16> ValueSet; + // Holds the Load and Store *instructions*. + ValueVector Loads; + ValueVector Stores; + PtrRtCheck.Pointers.clear(); + PtrRtCheck.Need = false; + + // Scan the BB and collect legal loads and stores. + for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) { + Instruction *I = it; + + // If this is a load, save it. If this instruction can read from memory + // but is not a load, then we quit. Notice that we don't handle function + // calls that read or write. + if (I->mayReadFromMemory()) { + LoadInst *Ld = dyn_cast<LoadInst>(I); + if (!Ld) return false; + if (!Ld->isSimple()) { + DEBUG(dbgs() << "LV: Found a non-simple load.\n"); + return false; + } + Loads.push_back(Ld); + continue; + } + + // Save store instructions. Abort if other instructions write to memory. + if (I->mayWriteToMemory()) { + StoreInst *St = dyn_cast<StoreInst>(I); + if (!St) return false; + if (!St->isSimple()) { + DEBUG(dbgs() << "LV: Found a non-simple store.\n"); + return false; + } + Stores.push_back(St); + } + } // next instr. + + // Now we have two lists that hold the loads and the stores. + // Next, we find the pointers that they use. + + // Check if we see any stores. If there are no stores, then we don't + // care if the pointers are *restrict*. + if (!Stores.size()) { + DEBUG(dbgs() << "LV: Found a read-only loop!\n"); + return true; + } + + // Holds the read and read-write *pointers* that we find. + ValueVector Reads; + ValueVector ReadWrites; + + // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects + // multiple times on the same object. If the ptr is accessed twice, once + // for read and once for write, it will only appear once (on the write + // list). This is okay, since we are going to check for conflicts between + // writes and between reads and writes, but not between reads and reads. + ValueSet Seen; + + ValueVector::iterator I, IE; + for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) { + StoreInst *ST = dyn_cast<StoreInst>(*I); + assert(ST && "Bad StoreInst"); + Value* Ptr = ST->getPointerOperand(); + + if (isUniform(Ptr)) { + DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n"); + return false; + } + + // If we did *not* see this pointer before, insert it to + // the read-write list. At this phase it is only a 'write' list. + if (Seen.insert(Ptr)) + ReadWrites.push_back(Ptr); + } + + for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) { + LoadInst *LD = dyn_cast<LoadInst>(*I); + assert(LD && "Bad LoadInst"); + Value* Ptr = LD->getPointerOperand(); + // If we did *not* see this pointer before, insert it to the + // read list. If we *did* see it before, then it is already in + // the read-write list. This allows us to vectorize expressions + // such as A[i] += x; Because the address of A[i] is a read-write + // pointer. This only works if the index of A[i] is consecutive. + // If the address of i is unknown (for example A[B[i]]) then we may + // read a few words, modify, and write a few words, and some of the + // words may be written to the same address. + if (Seen.insert(Ptr) || !isConsecutiveGep(Ptr)) + Reads.push_back(Ptr); + } + + // If we write (or read-write) to a single destination and there are no + // other reads in this loop then is it safe to vectorize. + if (ReadWrites.size() == 1 && Reads.size() == 0) { + DEBUG(dbgs() << "LV: Found a write-only loop!\n"); + return true; + } + + // Find pointers with computable bounds. We are going to use this information + // to place a runtime bound check. + bool RT = true; + for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) + if (hasComputableBounds(*I)) { + PtrRtCheck.Pointers.push_back(*I); + DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n"); + } else { + RT = false; + break; + } + for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) + if (hasComputableBounds(*I)) { + PtrRtCheck.Pointers.push_back(*I); + DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n"); + } else { + RT = false; + break; + } + + // Check that we did not collect too many pointers or found a + // unsizeable pointer. + if (!RT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) { + PtrRtCheck.Pointers.clear(); + RT = false; + } + + PtrRtCheck.Need = RT; + + if (RT) { + DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n"); + } + + // Now that the pointers are in two lists (Reads and ReadWrites), we + // can check that there are no conflicts between each of the writes and + // between the writes to the reads. + ValueSet WriteObjects; + ValueVector TempObjects; + + // Check that the read-writes do not conflict with other read-write + // pointers. + for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) { + GetUnderlyingObjects(*I, TempObjects, DL); + for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end(); + it != e; ++it) { + if (!isIdentifiedObject(*it)) { + DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n"); + return RT; + } + if (!WriteObjects.insert(*it)) { + DEBUG(dbgs() << "LV: Found a possible write-write reorder:" + << **it <<"\n"); + return RT; + } + } + TempObjects.clear(); + } + + /// Check that the reads don't conflict with the read-writes. + for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) { + GetUnderlyingObjects(*I, TempObjects, DL); + for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end(); + it != e; ++it) { + if (!isIdentifiedObject(*it)) { + DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n"); + return RT; + } + if (WriteObjects.count(*it)) { + DEBUG(dbgs() << "LV: Found a possible read/write reorder:" + << **it <<"\n"); + return RT; + } + } + TempObjects.clear(); + } + + // It is safe to vectorize and we don't need any runtime checks. + DEBUG(dbgs() << "LV: We don't need a runtime memory check.\n"); + PtrRtCheck.Pointers.clear(); + PtrRtCheck.Need = false; + return true; +} + +bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi, + ReductionKind Kind) { + if (Phi->getNumIncomingValues() != 2) + return false; + + // Find the possible incoming reduction variable. + BasicBlock *BB = Phi->getParent(); + int SelfEdgeIdx = Phi->getBasicBlockIndex(BB); + int InEdgeBlockIdx = (SelfEdgeIdx ? 0 : 1); // The other entry. + Value *RdxStart = Phi->getIncomingValue(InEdgeBlockIdx); + + // ExitInstruction is the single value which is used outside the loop. + // We only allow for a single reduction value to be used outside the loop. + // This includes users of the reduction, variables (which form a cycle + // which ends in the phi node). + Instruction *ExitInstruction = 0; + + // Iter is our iterator. We start with the PHI node and scan for all of the + // users of this instruction. All users must be instructions which can be + // used as reduction variables (such as ADD). We may have a single + // out-of-block user. They cycle must end with the original PHI. + // Also, we can't have multiple block-local users. + Instruction *Iter = Phi; + while (true) { + // Any reduction instr must be of one of the allowed kinds. + if (!isReductionInstr(Iter, Kind)) + return false; + + // Did we found a user inside this block ? + bool FoundInBlockUser = false; + // Did we reach the initial PHI node ? + bool FoundStartPHI = false; + + // If the instruction has no users then this is a broken + // chain and can't be a reduction variable. + if (Iter->use_empty()) + return false; + + // For each of the *users* of iter. + for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end(); + it != e; ++it) { + Instruction *U = cast<Instruction>(*it); + // We already know that the PHI is a user. + if (U == Phi) { + FoundStartPHI = true; + continue; + } + // Check if we found the exit user. + BasicBlock *Parent = U->getParent(); + if (Parent != BB) { + // We must have a single exit instruction. + if (ExitInstruction != 0) + return false; + ExitInstruction = Iter; + } + // We can't have multiple inside users. + if (FoundInBlockUser) + return false; + FoundInBlockUser = true; + Iter = U; + } + + // We found a reduction var if we have reached the original + // phi node and we only have a single instruction with out-of-loop + // users. + if (FoundStartPHI && ExitInstruction) { + // This instruction is allowed to have out-of-loop users. + AllowedExit.insert(ExitInstruction); + + // Save the description of this reduction variable. + ReductionDescriptor RD(RdxStart, ExitInstruction, Kind); + Reductions[Phi] = RD; + return true; + } + } +} + +bool +LoopVectorizationLegality::isReductionInstr(Instruction *I, + ReductionKind Kind) { + switch (I->getOpcode()) { + default: + return false; + case Instruction::PHI: + // possibly. + return true; + case Instruction::Add: + case Instruction::Sub: + return Kind == IntegerAdd; + case Instruction::Mul: + case Instruction::UDiv: + case Instruction::SDiv: + return Kind == IntegerMult; + case Instruction::And: + return Kind == IntegerAnd; + case Instruction::Or: + return Kind == IntegerOr; + case Instruction::Xor: + return Kind == IntegerXor; + } +} + +bool LoopVectorizationLegality::isInductionVariable(PHINode *Phi) { + // Check that the PHI is consecutive and starts at zero. + const SCEV *PhiScev = SE->getSCEV(Phi); + const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); + if (!AR) { + DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n"); + return false; + } + const SCEV *Step = AR->getStepRecurrence(*SE); + + if (!Step->isOne()) { + DEBUG(dbgs() << "LV: PHI stride does not equal one.\n"); + return false; + } + return true; +} + +bool LoopVectorizationLegality::hasComputableBounds(Value *Ptr) { + const SCEV *PhiScev = SE->getSCEV(Ptr); + const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev); + if (!AR) + return false; + + return AR->isAffine(); +} + +unsigned +LoopVectorizationCostModel::findBestVectorizationFactor(unsigned VF) { + if (!VTTI) { + DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n"); + return 1; + } + + float Cost = expectedCost(1); + unsigned Width = 1; + DEBUG(dbgs() << "LV: Scalar loop costs: "<< (int)Cost << ".\n"); + for (unsigned i=2; i <= VF; i*=2) { + // Notice that the vector loop needs to be executed less times, so + // we need to divide the cost of the vector loops by the width of + // the vector elements. + float VectorCost = expectedCost(i) / (float)i; + DEBUG(dbgs() << "LV: Vector loop of width "<< i << " costs: " << + (int)VectorCost << ".\n"); + if (VectorCost < Cost) { + Cost = VectorCost; + Width = i; + } + } + + DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n"); + return Width; +} + +unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) { + // We can only estimate the cost of single basic block loops. + assert(1 == TheLoop->getNumBlocks() && "Too many blocks in loop"); + + BasicBlock *BB = TheLoop->getHeader(); + unsigned Cost = 0; + + // For each instruction in the old loop. + for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { + Instruction *Inst = it; + unsigned C = getInstructionCost(Inst, VF); + Cost += C; + DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF "<< VF << + " For instruction: "<< *Inst << "\n"); + } + + return Cost; +} + +unsigned +LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) { + assert(VTTI && "Invalid vector target transformation info"); + + // If we know that this instruction will remain uniform, check the cost of + // the scalar version. + if (Legal->isUniformAfterVectorization(I)) + VF = 1; + + Type *RetTy = I->getType(); + Type *VectorTy = ToVectorTy(RetTy, VF); + + + // TODO: We need to estimate the cost of intrinsic calls. + switch (I->getOpcode()) { + case Instruction::GetElementPtr: + // We mark this instruction as zero-cost because scalar GEPs are usually + // lowered to the intruction addressing mode. At the moment we don't + // generate vector geps. + return 0; + case Instruction::Br: { + return VTTI->getCFInstrCost(I->getOpcode()); + } + case Instruction::PHI: + return 0; + case Instruction::Add: + case Instruction::FAdd: + case Instruction::Sub: + case Instruction::FSub: + case Instruction::Mul: + case Instruction::FMul: + case Instruction::UDiv: + case Instruction::SDiv: + case Instruction::FDiv: + case Instruction::URem: + case Instruction::SRem: + case Instruction::FRem: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: { + return VTTI->getArithmeticInstrCost(I->getOpcode(), VectorTy); + } + case Instruction::Select: { + SelectInst *SI = cast<SelectInst>(I); + const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); + bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); + Type *CondTy = SI->getCondition()->getType(); + if (ScalarCond) + CondTy = VectorType::get(CondTy, VF); + + return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy); + } + case Instruction::ICmp: + case Instruction::FCmp: { + Type *ValTy = I->getOperand(0)->getType(); + VectorTy = ToVectorTy(ValTy, VF); + return VTTI->getCmpSelInstrCost(I->getOpcode(), VectorTy); + } + case Instruction::Store: { + StoreInst *SI = cast<StoreInst>(I); + Type *ValTy = SI->getValueOperand()->getType(); + VectorTy = ToVectorTy(ValTy, VF); + + if (VF == 1) + return VTTI->getMemoryOpCost(I->getOpcode(), ValTy, + SI->getAlignment(), SI->getPointerAddressSpace()); + + // Scalarized stores. + if (!Legal->isConsecutiveGep(SI->getPointerOperand())) { + unsigned Cost = 0; + unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement, + ValTy); + // The cost of extracting from the value vector. + Cost += VF * (ExtCost); + // The cost of the scalar stores. + Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(), + ValTy->getScalarType(), + SI->getAlignment(), + SI->getPointerAddressSpace()); + return Cost; + } + + // Wide stores. + return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, SI->getAlignment(), + SI->getPointerAddressSpace()); + } + case Instruction::Load: { + LoadInst *LI = cast<LoadInst>(I); + + if (VF == 1) + return VTTI->getMemoryOpCost(I->getOpcode(), RetTy, + LI->getAlignment(), + LI->getPointerAddressSpace()); + + // Scalarized loads. + if (!Legal->isConsecutiveGep(LI->getPointerOperand())) { + unsigned Cost = 0; + unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy); + // The cost of inserting the loaded value into the result vector. + Cost += VF * (InCost); + // The cost of the scalar stores. + Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(), + RetTy->getScalarType(), + LI->getAlignment(), + LI->getPointerAddressSpace()); + return Cost; + } + + // Wide loads. + return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy, LI->getAlignment(), + LI->getPointerAddressSpace()); + } + case Instruction::ZExt: + case Instruction::SExt: + case Instruction::FPToUI: + case Instruction::FPToSI: + case Instruction::FPExt: + case Instruction::PtrToInt: + case Instruction::IntToPtr: + case Instruction::SIToFP: + case Instruction::UIToFP: + case Instruction::Trunc: + case Instruction::FPTrunc: + case Instruction::BitCast: { + Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF); + return VTTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy); + } + default: { + // We are scalarizing the instruction. Return the cost of the scalar + // instruction, plus the cost of insert and extract into vector + // elements, times the vector width. + unsigned Cost = 0; + + bool IsVoid = RetTy->isVoidTy(); + + unsigned InsCost = (IsVoid ? 0 : + VTTI->getInstrCost(Instruction::InsertElement, + VectorTy)); + + unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement, + VectorTy); + + // The cost of inserting the results plus extracting each one of the + // operands. + Cost += VF * (InsCost + ExtCost * I->getNumOperands()); + + // The cost of executing VF copies of the scalar instruction. + Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy); + return Cost; + } + }// end of switch. +} + +Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) { + if (Scalar->isVoidTy() || VF == 1) + return Scalar; + return VectorType::get(Scalar, VF); +} + +} // namespace + +char LoopVectorize::ID = 0; +static const char lv_name[] = "Loop Vectorization"; +INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false) +INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) +INITIALIZE_PASS_DEPENDENCY(LoopSimplify) +INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false) + +namespace llvm { + Pass *createLoopVectorizePass() { + return new LoopVectorize(); + } +} + diff --git a/contrib/llvm/lib/Transforms/Vectorize/Vectorize.cpp b/contrib/llvm/lib/Transforms/Vectorize/Vectorize.cpp index 1ef6002..d26973a 100644 --- a/contrib/llvm/lib/Transforms/Vectorize/Vectorize.cpp +++ b/contrib/llvm/lib/Transforms/Vectorize/Vectorize.cpp @@ -7,7 +7,7 @@ // //===----------------------------------------------------------------------===// // -// This file implements common infrastructure for libLLVMVectorizeOpts.a, which +// This file implements common infrastructure for libLLVMVectorizeOpts.a, which // implements several vectorization transformations over the LLVM intermediate // representation, including the C bindings for that library. // @@ -23,10 +23,11 @@ using namespace llvm; -/// initializeVectorizationPasses - Initialize all passes linked into the +/// initializeVectorizationPasses - Initialize all passes linked into the /// Vectorization library. void llvm::initializeVectorization(PassRegistry &Registry) { initializeBBVectorizePass(Registry); + initializeLoopVectorizePass(Registry); } void LLVMInitializeVectorization(LLVMPassRegistryRef R) { @@ -37,3 +38,6 @@ void LLVMAddBBVectorizePass(LLVMPassManagerRef PM) { unwrap(PM)->add(createBBVectorizePass()); } +void LLVMAddLoopVectorizePass(LLVMPassManagerRef PM) { + unwrap(PM)->add(createLoopVectorizePass()); +} |
