diff options
| author | rdivacky <rdivacky@FreeBSD.org> | 2010-02-16 09:30:23 +0000 |
|---|---|---|
| committer | rdivacky <rdivacky@FreeBSD.org> | 2010-02-16 09:30:23 +0000 |
| commit | f25ddd991a5601d0101602c4c263a58c7af4b8a2 (patch) | |
| tree | 4cfca640904d1896e25032757a61f8959c066919 /lib/Transforms | |
| parent | 3fd58f91dd318518f7daa4ba64c0aaf31799d89b (diff) | |
| download | FreeBSD-src-f25ddd991a5601d0101602c4c263a58c7af4b8a2.zip FreeBSD-src-f25ddd991a5601d0101602c4c263a58c7af4b8a2.tar.gz | |
Update LLVM to r96341.
Diffstat (limited to 'lib/Transforms')
52 files changed, 4100 insertions, 3129 deletions
diff --git a/lib/Transforms/Hello/Makefile b/lib/Transforms/Hello/Makefile index 46f8098..c5e75d4 100644 --- a/lib/Transforms/Hello/Makefile +++ b/lib/Transforms/Hello/Makefile @@ -11,7 +11,6 @@ LEVEL = ../../.. LIBRARYNAME = LLVMHello LOADABLE_MODULE = 1 USEDLIBS = -CXXFLAGS = -fno-rtti include $(LEVEL)/Makefile.common diff --git a/lib/Transforms/IPO/ArgumentPromotion.cpp b/lib/Transforms/IPO/ArgumentPromotion.cpp index d8190a4..325d353 100644 --- a/lib/Transforms/IPO/ArgumentPromotion.cpp +++ b/lib/Transforms/IPO/ArgumentPromotion.cpp @@ -247,7 +247,7 @@ static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, return Low != Set.end() && IsPrefix(*Low, Indices); } -/// Mark the given indices (ToMark) as safe in the the given set of indices +/// Mark the given indices (ToMark) as safe in the given set of indices /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there /// is already a prefix of Indices in Safe, Indices are implicitely marked safe /// already. Furthermore, any indices that Indices is itself a prefix of, are diff --git a/lib/Transforms/IPO/ConstantMerge.cpp b/lib/Transforms/IPO/ConstantMerge.cpp index 4972687..3c05f88 100644 --- a/lib/Transforms/IPO/ConstantMerge.cpp +++ b/lib/Transforms/IPO/ConstantMerge.cpp @@ -19,10 +19,11 @@ #define DEBUG_TYPE "constmerge" #include "llvm/Transforms/IPO.h" +#include "llvm/DerivedTypes.h" #include "llvm/Module.h" #include "llvm/Pass.h" +#include "llvm/ADT/DenseMap.h" #include "llvm/ADT/Statistic.h" -#include <map> using namespace llvm; STATISTIC(NumMerged, "Number of global constants merged"); @@ -48,10 +49,10 @@ ModulePass *llvm::createConstantMergePass() { return new ConstantMerge(); } bool ConstantMerge::runOnModule(Module &M) { // Map unique constant/section pairs to globals. We don't want to merge // globals in different sections. - std::map<std::pair<Constant*, std::string>, GlobalVariable*> CMap; + DenseMap<Constant*, GlobalVariable*> CMap; // Replacements - This vector contains a list of replacements to perform. - std::vector<std::pair<GlobalVariable*, GlobalVariable*> > Replacements; + SmallVector<std::pair<GlobalVariable*, GlobalVariable*>, 32> Replacements; bool MadeChange = false; @@ -76,19 +77,21 @@ bool ConstantMerge::runOnModule(Module &M) { continue; } - // Only process constants with initializers. - if (GV->isConstant() && GV->hasDefinitiveInitializer()) { - Constant *Init = GV->getInitializer(); - - // Check to see if the initializer is already known. - GlobalVariable *&Slot = CMap[std::make_pair(Init, GV->getSection())]; - - if (Slot == 0) { // Nope, add it to the map. - Slot = GV; - } else if (GV->hasLocalLinkage()) { // Yup, this is a duplicate! - // Make all uses of the duplicate constant use the canonical version. - Replacements.push_back(std::make_pair(GV, Slot)); - } + // Only process constants with initializers in the default addres space. + if (!GV->isConstant() ||!GV->hasDefinitiveInitializer() || + GV->getType()->getAddressSpace() != 0 || !GV->getSection().empty()) + continue; + + Constant *Init = GV->getInitializer(); + + // Check to see if the initializer is already known. + GlobalVariable *&Slot = CMap[Init]; + + if (Slot == 0) { // Nope, add it to the map. + Slot = GV; + } else if (GV->hasLocalLinkage()) { // Yup, this is a duplicate! + // Make all uses of the duplicate constant use the canonical version. + Replacements.push_back(std::make_pair(GV, Slot)); } } @@ -100,11 +103,11 @@ bool ConstantMerge::runOnModule(Module &M) { // now. This avoid invalidating the pointers in CMap, which are unneeded // now. for (unsigned i = 0, e = Replacements.size(); i != e; ++i) { - // Eliminate any uses of the dead global... + // Eliminate any uses of the dead global. Replacements[i].first->replaceAllUsesWith(Replacements[i].second); - // Delete the global value from the module... - M.getGlobalList().erase(Replacements[i].first); + // Delete the global value from the module. + Replacements[i].first->eraseFromParent(); } NumMerged += Replacements.size(); diff --git a/lib/Transforms/IPO/DeadTypeElimination.cpp b/lib/Transforms/IPO/DeadTypeElimination.cpp index 025d77e..662fbb5 100644 --- a/lib/Transforms/IPO/DeadTypeElimination.cpp +++ b/lib/Transforms/IPO/DeadTypeElimination.cpp @@ -57,13 +57,13 @@ ModulePass *llvm::createDeadTypeEliminationPass() { // static inline bool ShouldNukeSymtabEntry(const Type *Ty){ // Nuke all names for primitive types! - if (Ty->isPrimitiveType() || Ty->isInteger()) + if (Ty->isPrimitiveType() || Ty->isIntegerTy()) return true; // Nuke all pointers to primitive types as well... if (const PointerType *PT = dyn_cast<PointerType>(Ty)) if (PT->getElementType()->isPrimitiveType() || - PT->getElementType()->isInteger()) + PT->getElementType()->isIntegerTy()) return true; return false; diff --git a/lib/Transforms/IPO/GlobalOpt.cpp b/lib/Transforms/IPO/GlobalOpt.cpp index ee260e9..df060eb 100644 --- a/lib/Transforms/IPO/GlobalOpt.cpp +++ b/lib/Transforms/IPO/GlobalOpt.cpp @@ -638,8 +638,8 @@ static bool AllUsesOfValueWillTrapIfNull(Value *V, } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) { // If we've already seen this phi node, ignore it, it has already been // checked. - if (PHIs.insert(PN)) - return AllUsesOfValueWillTrapIfNull(PN, PHIs); + if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs)) + return false; } else if (isa<ICmpInst>(*UI) && isa<ConstantPointerNull>(UI->getOperand(1))) { // Ignore setcc X, null @@ -1590,7 +1590,7 @@ static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { // simplification. In these cases, we typically end up with "cond ? v1 : v2" // where v1 and v2 both require constant pool loads, a big loss. if (GVElType == Type::getInt1Ty(GV->getContext()) || - GVElType->isFloatingPoint() || + GVElType->isFloatingPointTy() || isa<PointerType>(GVElType) || isa<VectorType>(GVElType)) return false; @@ -1925,7 +1925,7 @@ GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { if (!ATy) return 0; const StructType *STy = dyn_cast<StructType>(ATy->getElementType()); if (!STy || STy->getNumElements() != 2 || - !STy->getElementType(0)->isInteger(32)) return 0; + !STy->getElementType(0)->isIntegerTy(32)) return 0; const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1)); if (!PFTy) return 0; const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType()); diff --git a/lib/Transforms/IPO/Inliner.cpp b/lib/Transforms/IPO/Inliner.cpp index 0990278..752a97c 100644 --- a/lib/Transforms/IPO/Inliner.cpp +++ b/lib/Transforms/IPO/Inliner.cpp @@ -38,8 +38,15 @@ STATISTIC(NumDeleted, "Number of functions deleted because all callers found"); STATISTIC(NumMergedAllocas, "Number of allocas merged together"); static cl::opt<int> -InlineLimit("inline-threshold", cl::Hidden, cl::init(200), cl::ZeroOrMore, - cl::desc("Control the amount of inlining to perform (default = 200)")); +InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore, + cl::desc("Control the amount of inlining to perform (default = 225)")); + +static cl::opt<int> +HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325), + cl::desc("Threshold for inlining functions with inline hint")); + +// Threshold to use when optsize is specified (and there is no -inline-limit). +const int OptSizeThreshold = 75; Inliner::Inliner(void *ID) : CallGraphSCCPass(ID), InlineThreshold(InlineLimit) {} @@ -172,13 +179,23 @@ static bool InlineCallIfPossible(CallSite CS, CallGraph &CG, return true; } -unsigned Inliner::getInlineThreshold(Function* Caller) const { +unsigned Inliner::getInlineThreshold(CallSite CS) const { + int thres = InlineThreshold; + + // Listen to optsize when -inline-limit is not given. + Function *Caller = CS.getCaller(); if (Caller && !Caller->isDeclaration() && Caller->hasFnAttr(Attribute::OptimizeForSize) && InlineLimit.getNumOccurrences() == 0) - return 50; - else - return InlineThreshold; + thres = OptSizeThreshold; + + // Listen to inlinehint when it would increase the threshold. + Function *Callee = CS.getCalledFunction(); + if (HintThreshold > thres && Callee && !Callee->isDeclaration() && + Callee->hasFnAttr(Attribute::InlineHint)) + thres = HintThreshold; + + return thres; } /// shouldInline - Return true if the inliner should attempt to inline @@ -200,7 +217,7 @@ bool Inliner::shouldInline(CallSite CS) { int Cost = IC.getValue(); Function *Caller = CS.getCaller(); - int CurrentThreshold = getInlineThreshold(Caller); + int CurrentThreshold = getInlineThreshold(CS); float FudgeFactor = getInlineFudgeFactor(CS); if (Cost >= (int)(CurrentThreshold * FudgeFactor)) { DEBUG(dbgs() << " NOT Inlining: cost=" << Cost @@ -236,8 +253,7 @@ bool Inliner::shouldInline(CallSite CS) { outerCallsFound = true; int Cost2 = IC2.getValue(); - Function *Caller2 = CS2.getCaller(); - int CurrentThreshold2 = getInlineThreshold(Caller2); + int CurrentThreshold2 = getInlineThreshold(CS2); float FudgeFactor2 = getInlineFudgeFactor(CS2); if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2)) diff --git a/lib/Transforms/IPO/Makefile b/lib/Transforms/IPO/Makefile index fd018c4..5c42374 100644 --- a/lib/Transforms/IPO/Makefile +++ b/lib/Transforms/IPO/Makefile @@ -10,7 +10,6 @@ LEVEL = ../../.. LIBRARYNAME = LLVMipo BUILD_ARCHIVE = 1 -CXXFLAGS = -fno-rtti include $(LEVEL)/Makefile.common diff --git a/lib/Transforms/IPO/MergeFunctions.cpp b/lib/Transforms/IPO/MergeFunctions.cpp index fa8845b..b07e22c 100644 --- a/lib/Transforms/IPO/MergeFunctions.cpp +++ b/lib/Transforms/IPO/MergeFunctions.cpp @@ -467,7 +467,6 @@ static LinkageCategory categorize(const Function *F) { case GlobalValue::AppendingLinkage: case GlobalValue::DLLImportLinkage: case GlobalValue::DLLExportLinkage: - case GlobalValue::GhostLinkage: case GlobalValue::CommonLinkage: return ExternalStrong; } diff --git a/lib/Transforms/IPO/PartialInlining.cpp b/lib/Transforms/IPO/PartialInlining.cpp index f40902f..f8ec722 100644 --- a/lib/Transforms/IPO/PartialInlining.cpp +++ b/lib/Transforms/IPO/PartialInlining.cpp @@ -117,7 +117,7 @@ Function* PartialInliner::unswitchFunction(Function* F) { DominatorTree DT; DT.runOnFunction(*duplicateFunction); - // Extract the body of the the if. + // Extract the body of the if. Function* extractedFunction = ExtractCodeRegion(DT, toExtract); // Inline the top-level if test into all callers. diff --git a/lib/Transforms/IPO/StripSymbols.cpp b/lib/Transforms/IPO/StripSymbols.cpp index 0e0d83a..310e4a2 100644 --- a/lib/Transforms/IPO/StripSymbols.cpp +++ b/lib/Transforms/IPO/StripSymbols.cpp @@ -214,6 +214,15 @@ static bool StripDebugInfo(Module &M) { Changed = true; } + if (Function *DbgVal = M.getFunction("llvm.dbg.value")) { + while (!DbgVal->use_empty()) { + CallInst *CI = cast<CallInst>(DbgVal->use_back()); + CI->eraseFromParent(); + } + DbgVal->eraseFromParent(); + Changed = true; + } + NamedMDNode *NMD = M.getNamedMetadata("llvm.dbg.gv"); if (NMD) { Changed = true; diff --git a/lib/Transforms/InstCombine/InstCombine.h b/lib/Transforms/InstCombine/InstCombine.h index 5367900..09accb6 100644 --- a/lib/Transforms/InstCombine/InstCombine.h +++ b/lib/Transforms/InstCombine/InstCombine.h @@ -199,11 +199,12 @@ private: SmallVectorImpl<Value*> &NewIndices); Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI); - /// ValueRequiresCast - Return true if the cast from "V to Ty" actually - /// results in any code being generated. It does not require codegen if V is - /// simple enough or if the cast can be folded into other casts. - bool ValueRequiresCast(Instruction::CastOps opcode,const Value *V, - const Type *Ty); + /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually + /// results in any code being generated and is interesting to optimize out. If + /// the cast can be eliminated by some other simple transformation, we prefer + /// to do the simplification first. + bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V, + const Type *Ty); Instruction *visitCallSite(CallSite CS); bool transformConstExprCastCall(CallSite CS); diff --git a/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/lib/Transforms/InstCombine/InstCombineAddSub.cpp index 4891ff0..2da17f1 100644 --- a/lib/Transforms/InstCombine/InstCombineAddSub.cpp +++ b/lib/Transforms/InstCombine/InstCombineAddSub.cpp @@ -35,7 +35,7 @@ static Constant *SubOne(ConstantInt *C) { // Otherwise, return null. // static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) { - if (!V->hasOneUse() || !V->getType()->isInteger()) + if (!V->hasOneUse() || !V->getType()->isIntegerTy()) return 0; Instruction *I = dyn_cast<Instruction>(V); @@ -121,50 +121,34 @@ Instruction *InstCombiner::visitAdd(BinaryOperator &I) { match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) { uint32_t TySizeBits = I.getType()->getScalarSizeInBits(); const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue(); - - uint32_t Size = TySizeBits / 2; - APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1)); - APInt CFF80Val(-C0080Val); - do { - if (TySizeBits > Size) { - // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. - // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. - if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) || - (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) { - // This is a sign extend if the top bits are known zero. - if (!MaskedValueIsZero(XorLHS, - APInt::getHighBitsSet(TySizeBits, TySizeBits - Size))) - Size = 0; // Not a sign ext, but can't be any others either. - break; - } - } - Size >>= 1; - C0080Val = APIntOps::lshr(C0080Val, Size); - CFF80Val = APIntOps::ashr(CFF80Val, Size); - } while (Size >= 1); - - // FIXME: This shouldn't be necessary. When the backends can handle types - // with funny bit widths then this switch statement should be removed. It - // is just here to get the size of the "middle" type back up to something - // that the back ends can handle. - const Type *MiddleType = 0; - switch (Size) { - default: break; - case 32: - case 16: - case 8: MiddleType = IntegerType::get(I.getContext(), Size); break; + unsigned ExtendAmt = 0; + // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext. + // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext. + if (XorRHS->getValue() == -RHSVal) { + if (RHSVal.isPowerOf2()) + ExtendAmt = TySizeBits - RHSVal.logBase2() - 1; + else if (XorRHS->getValue().isPowerOf2()) + ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1; + } + + if (ExtendAmt) { + APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt); + if (!MaskedValueIsZero(XorLHS, Mask)) + ExtendAmt = 0; } - if (MiddleType) { - Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext"); - return new SExtInst(NewTrunc, I.getType(), I.getName()); + + if (ExtendAmt) { + Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt); + Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext"); + return BinaryOperator::CreateAShr(NewShl, ShAmt); } } } - if (I.getType()->isInteger(1)) + if (I.getType()->isIntegerTy(1)) return BinaryOperator::CreateXor(LHS, RHS); - if (I.getType()->isInteger()) { + if (I.getType()->isIntegerTy()) { // X + X --> X << 1 if (LHS == RHS) return BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1)); @@ -184,7 +168,7 @@ Instruction *InstCombiner::visitAdd(BinaryOperator &I) { // -A + B --> B - A // -A + -B --> -(A + B) if (Value *LHSV = dyn_castNegVal(LHS)) { - if (LHS->getType()->isIntOrIntVector()) { + if (LHS->getType()->isIntOrIntVectorTy()) { if (Value *RHSV = dyn_castNegVal(RHS)) { Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum"); return BinaryOperator::CreateNeg(NewAdd); @@ -238,7 +222,7 @@ Instruction *InstCombiner::visitAdd(BinaryOperator &I) { } // W*X + Y*Z --> W * (X+Z) iff W == Y - if (I.getType()->isIntOrIntVector()) { + if (I.getType()->isIntOrIntVectorTy()) { Value *W, *X, *Y, *Z; if (match(LHS, m_Mul(m_Value(W), m_Value(X))) && match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) { @@ -576,7 +560,7 @@ Instruction *InstCombiner::visitSub(BinaryOperator &I) { return ReplaceInstUsesWith(I, Op0); // undef - X -> undef if (isa<UndefValue>(Op1)) return ReplaceInstUsesWith(I, Op1); // X - undef -> undef - if (I.getType()->isInteger(1)) + if (I.getType()->isIntegerTy(1)) return BinaryOperator::CreateXor(Op0, Op1); if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) { @@ -676,6 +660,13 @@ Instruction *InstCombiner::visitSub(BinaryOperator &I) { return BinaryOperator::CreateSDiv(Op1I->getOperand(0), ConstantExpr::getNeg(DivRHS)); + // 0 - (C << X) -> (-C << X) + if (Op1I->getOpcode() == Instruction::Shl) + if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0)) + if (CSI->isZero()) + if (Value *ShlLHSNeg = dyn_castNegVal(Op1I->getOperand(0))) + return BinaryOperator::CreateShl(ShlLHSNeg, Op1I->getOperand(1)); + // X - X*C --> X * (1-C) ConstantInt *C2 = 0; if (dyn_castFoldableMul(Op1I, C2) == Op0) { diff --git a/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp index fa7bb12..5e47953 100644 --- a/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp +++ b/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp @@ -546,7 +546,7 @@ Instruction *InstCombiner::FoldAndOfICmps(Instruction &I, std::swap(LHSCC, RHSCC); } - // At this point, we know we have have two icmp instructions + // At this point, we know we have two icmp instructions // comparing a value against two constants and and'ing the result // together. Because of the above check, we know that we only have // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know @@ -932,24 +932,49 @@ Instruction *InstCombiner::visitAnd(BinaryOperator &I) { if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0)) if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS)) return Res; - + + // If and'ing two fcmp, try combine them into one. + if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) + if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) + if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS)) + return Res; + + // fold (and (cast A), (cast B)) -> (cast (and A, B)) if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) - if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) - if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ? - const Type *SrcTy = Op0C->getOperand(0)->getType(); - if (SrcTy == Op1C->getOperand(0)->getType() && - SrcTy->isIntOrIntVector() && - // Only do this if the casts both really cause code to be generated. - ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0), - I.getType()) && - ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0), - I.getType())) { - Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0), - Op1C->getOperand(0), I.getName()); + if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) { + const Type *SrcTy = Op0C->getOperand(0)->getType(); + if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ? + SrcTy == Op1C->getOperand(0)->getType() && + SrcTy->isIntOrIntVectorTy()) { + Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); + + // Only do this if the casts both really cause code to be generated. + if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && + ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { + Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName()); return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); } + + // If this is and(cast(icmp), cast(icmp)), try to fold this even if the + // cast is otherwise not optimizable. This happens for vector sexts. + if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) + if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) + if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS)) { + InsertNewInstBefore(Res, I); + return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); + } + + // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the + // cast is otherwise not optimizable. This happens for vector sexts. + if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) + if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) + if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS)) { + InsertNewInstBefore(Res, I); + return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); + } } + } // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { @@ -965,13 +990,6 @@ Instruction *InstCombiner::visitAnd(BinaryOperator &I) { } } - // If and'ing two fcmp, try combine them into one. - if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) { - if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) - if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS)) - return Res; - } - return Changed ? &I : 0; } @@ -1142,18 +1160,20 @@ static Instruction *MatchSelectFromAndOr(Value *A, Value *B, Value *C, Value *D) { // If A is not a select of -1/0, this cannot match. Value *Cond = 0; - if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond)))) + if (!match(A, m_SExt(m_Value(Cond))) || + !Cond->getType()->isIntegerTy(1)) return 0; // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B. - if (match(D, m_SelectCst<0, -1>(m_Specific(Cond)))) + if (match(D, m_Not(m_SExt(m_Specific(Cond))))) return SelectInst::Create(Cond, C, B); - if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))))) + if (match(D, m_SExt(m_Not(m_Specific(Cond))))) return SelectInst::Create(Cond, C, B); + // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D. - if (match(B, m_SelectCst<0, -1>(m_Specific(Cond)))) + if (match(B, m_Not(m_SExt(m_Specific(Cond))))) return SelectInst::Create(Cond, C, D); - if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))))) + if (match(B, m_SExt(m_Not(m_Specific(Cond))))) return SelectInst::Create(Cond, C, D); return 0; } @@ -1224,7 +1244,7 @@ Instruction *InstCombiner::FoldOrOfICmps(Instruction &I, std::swap(LHSCC, RHSCC); } - // At this point, we know we have have two icmp instructions + // At this point, we know we have two icmp instructions // comparing a value against two constants and or'ing the result // together. Because of the above check, we know that we only have // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the @@ -1595,15 +1615,19 @@ Instruction *InstCombiner::visitOr(BinaryOperator &I) { } } - // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants - if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D)) - return Match; - if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C)) - return Match; - if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D)) - return Match; - if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C)) - return Match; + // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants. + // Don't do this for vector select idioms, the code generator doesn't handle + // them well yet. + if (!isa<VectorType>(I.getType())) { + if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D)) + return Match; + if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C)) + return Match; + if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D)) + return Match; + if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C)) + return Match; + } // ((A&~B)|(~A&B)) -> A^B if ((match(C, m_Not(m_Specific(D))) && @@ -1663,37 +1687,51 @@ Instruction *InstCombiner::visitOr(BinaryOperator &I) { if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS)) return Res; + // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y) + if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) + if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) + if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS)) + return Res; + // fold (or (cast A), (cast B)) -> (cast (or A, B)) if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ? - if (!isa<ICmpInst>(Op0C->getOperand(0)) || - !isa<ICmpInst>(Op1C->getOperand(0))) { - const Type *SrcTy = Op0C->getOperand(0)->getType(); - if (SrcTy == Op1C->getOperand(0)->getType() && - SrcTy->isIntOrIntVector() && + const Type *SrcTy = Op0C->getOperand(0)->getType(); + if (SrcTy == Op1C->getOperand(0)->getType() && + SrcTy->isIntOrIntVectorTy()) { + Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); + + if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) && // Only do this if the casts both really cause code to be // generated. - ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0), - I.getType()) && - ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0), - I.getType())) { - Value *NewOp = Builder->CreateOr(Op0C->getOperand(0), - Op1C->getOperand(0), I.getName()); + ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && + ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { + Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName()); return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); } + + // If this is or(cast(icmp), cast(icmp)), try to fold this even if the + // cast is otherwise not optimizable. This happens for vector sexts. + if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) + if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) + if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS)) { + InsertNewInstBefore(Res, I); + return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); + } + + // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the + // cast is otherwise not optimizable. This happens for vector sexts. + if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) + if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) + if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS)) { + InsertNewInstBefore(Res, I); + return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); + } } } } - - // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y) - if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) { - if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) - if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS)) - return Res; - } - return Changed ? &I : 0; } @@ -1978,12 +2016,12 @@ Instruction *InstCombiner::visitXor(BinaryOperator &I) { if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind? const Type *SrcTy = Op0C->getOperand(0)->getType(); - if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() && + if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() && // Only do this if the casts both really cause code to be generated. - ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0), - I.getType()) && - ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0), - I.getType())) { + ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0), + I.getType()) && + ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0), + I.getType())) { Value *NewOp = Builder->CreateXor(Op0C->getOperand(0), Op1C->getOperand(0), I.getName()); return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp index 47c37c4..d7efdcf 100644 --- a/lib/Transforms/InstCombine/InstCombineCalls.cpp +++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp @@ -199,7 +199,7 @@ Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { // Extract the length and alignment and fill if they are constant. ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); - if (!LenC || !FillC || !FillC->getType()->isInteger(8)) + if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) return 0; uint64_t Len = LenC->getZExtValue(); Alignment = MI->getAlignment(); @@ -230,7 +230,6 @@ Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { return 0; } - /// visitCallInst - CallInst simplification. This mostly only handles folding /// of intrinsic instructions. For normal calls, it allows visitCallSite to do /// the heavy lifting. @@ -304,6 +303,60 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { switch (II->getIntrinsicID()) { default: break; + case Intrinsic::objectsize: { + const Type *ReturnTy = CI.getType(); + Value *Op1 = II->getOperand(1); + bool Min = (cast<ConstantInt>(II->getOperand(2))->getZExtValue() == 1); + + // We need target data for just about everything so depend on it. + if (!TD) break; + + // Get to the real allocated thing and offset as fast as possible. + Op1 = Op1->stripPointerCasts(); + + // If we've stripped down to a single global variable that we + // can know the size of then just return that. + if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { + if (GV->hasDefinitiveInitializer()) { + Constant *C = GV->getInitializer(); + size_t globalSize = TD->getTypeAllocSize(C->getType()); + return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, globalSize)); + } else { + Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); + return ReplaceInstUsesWith(CI, RetVal); + } + } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) { + + // Only handle constant GEPs here. + if (CE->getOpcode() != Instruction::GetElementPtr) break; + GEPOperator *GEP = cast<GEPOperator>(CE); + + // Make sure we're not a constant offset from an external + // global. + Value *Operand = GEP->getPointerOperand(); + Operand = Operand->stripPointerCasts(); + if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) + if (!GV->hasDefinitiveInitializer()) break; + + // Get what we're pointing to and its size. + const PointerType *BaseType = + cast<PointerType>(Operand->getType()); + size_t Size = TD->getTypeAllocSize(BaseType->getElementType()); + + // Get the current byte offset into the thing. Use the original + // operand in case we're looking through a bitcast. + SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end()); + const PointerType *OffsetType = + cast<PointerType>(GEP->getPointerOperand()->getType()); + size_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size()); + + assert(Size >= Offset); + + Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset); + return ReplaceInstUsesWith(CI, RetVal); + + } + } case Intrinsic::bswap: // bswap(bswap(x)) -> x if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1))) @@ -632,18 +685,6 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { return EraseInstFromFunction(CI); break; } - case Intrinsic::objectsize: { - ConstantInt *Const = cast<ConstantInt>(II->getOperand(2)); - const Type *Ty = CI.getType(); - - // 0 is maximum number of bytes left, 1 is minimum number of bytes left. - // TODO: actually add these values, the current return values are "don't - // know". - if (Const->getZExtValue() == 0) - return ReplaceInstUsesWith(CI, Constant::getAllOnesValue(Ty)); - else - return ReplaceInstUsesWith(CI, ConstantInt::get(Ty, 0)); - } } return visitCallSite(II); @@ -692,10 +733,14 @@ Instruction *InstCombiner::visitCallSite(CallSite CS) { Value *Callee = CS.getCalledValue(); if (Function *CalleeF = dyn_cast<Function>(Callee)) - if (CalleeF->getCallingConv() != CS.getCallingConv()) { + // If the call and callee calling conventions don't match, this call must + // be unreachable, as the call is undefined. + if (CalleeF->getCallingConv() != CS.getCallingConv() && + // Only do this for calls to a function with a body. A prototype may + // not actually end up matching the implementation's calling conv for a + // variety of reasons (e.g. it may be written in assembly). + !CalleeF->isDeclaration()) { Instruction *OldCall = CS.getInstruction(); - // If the call and callee calling conventions don't match, this call must - // be unreachable, as the call is undefined. new StoreInst(ConstantInt::getTrue(Callee->getContext()), UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), OldCall); @@ -703,8 +748,13 @@ Instruction *InstCombiner::visitCallSite(CallSite CS) { // This allows ValueHandlers and custom metadata to adjust itself. if (!OldCall->getType()->isVoidTy()) OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType())); - if (isa<CallInst>(OldCall)) // Not worth removing an invoke here. + if (isa<CallInst>(OldCall)) return EraseInstFromFunction(*OldCall); + + // We cannot remove an invoke, because it would change the CFG, just + // change the callee to a null pointer. + cast<InvokeInst>(OldCall)->setOperand(0, + Constant::getNullValue(CalleeF->getType())); return 0; } diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp index f25dd35..bb4a0e9 100644 --- a/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -23,7 +23,7 @@ using namespace PatternMatch; /// static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, int &Offset) { - assert(Val->getType()->isInteger(32) && "Unexpected allocation size type!"); + assert(Val->getType()->isIntegerTy(32) && "Unexpected allocation size type!"); if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { Offset = CI->getZExtValue(); Scale = 0; @@ -255,17 +255,26 @@ isEliminableCastPair( return Instruction::CastOps(Res); } -/// ValueRequiresCast - Return true if the cast from "V to Ty" actually results -/// in any code being generated. It does not require codegen if V is simple -/// enough or if the cast can be folded into other casts. -bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V, - const Type *Ty) { +/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually +/// results in any code being generated and is interesting to optimize out. If +/// the cast can be eliminated by some other simple transformation, we prefer +/// to do the simplification first. +bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, + const Type *Ty) { + // Noop casts and casts of constants should be eliminated trivially. if (V->getType() == Ty || isa<Constant>(V)) return false; - // If this is another cast that can be eliminated, it isn't codegen either. + // If this is another cast that can be eliminated, we prefer to have it + // eliminated. if (const CastInst *CI = dyn_cast<CastInst>(V)) - if (isEliminableCastPair(CI, opcode, Ty, TD)) + if (isEliminableCastPair(CI, opc, Ty, TD)) return false; + + // If this is a vector sext from a compare, then we don't want to break the + // idiom where each element of the extended vector is either zero or all ones. + if (opc == Instruction::SExt && isa<CmpInst>(V) && isa<VectorType>(Ty)) + return false; + return true; } @@ -828,7 +837,7 @@ Instruction *InstCombiner::visitZExt(ZExtInst &CI) { // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 Value *X; - if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isInteger(1) && + if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) { Value *New = Builder->CreateZExt(X, CI.getType()); @@ -923,12 +932,6 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { Value *Src = CI.getOperand(0); const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); - // Canonicalize sign-extend from i1 to a select. - if (Src->getType()->isInteger(1)) - return SelectInst::Create(Src, - Constant::getAllOnesValue(CI.getType()), - Constant::getNullValue(CI.getType())); - // Attempt to extend the entire input expression tree to the destination // type. Only do this if the dest type is a simple type, don't convert the // expression tree to something weird like i93 unless the source is also @@ -968,6 +971,30 @@ Instruction *InstCombiner::visitSExt(SExtInst &CI) { return BinaryOperator::CreateAShr(Res, ShAmt); } + + // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed + // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed + { + ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS; + if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) { + // sext (x <s 0) to i32 --> x>>s31 true if signbit set. + // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear. + if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) || + (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) { + Value *Sh = ConstantInt::get(CmpLHS->getType(), + CmpLHS->getType()->getScalarSizeInBits()-1); + Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit"); + if (In->getType() != CI.getType()) + In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp"); + + if (Pred == ICmpInst::ICMP_SGT) + In = Builder->CreateNot(In, In->getName()+".not"); + return ReplaceInstUsesWith(CI, In); + } + } + } + + // If the input is a shl/ashr pair of a same constant, then this is a sign // extension from a smaller value. If we could trust arbitrary bitwidth // integers, we could turn this into a truncate to the smaller bit and then @@ -1127,16 +1154,22 @@ Instruction *InstCombiner::visitSIToFP(CastInst &CI) { } Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { - // If the source integer type is larger than the intptr_t type for - // this target, do a trunc to the intptr_t type, then inttoptr of it. This - // allows the trunc to be exposed to other transforms. Don't do this for - // extending inttoptr's, because we don't know if the target sign or zero - // extends to pointers. - if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() > - TD->getPointerSizeInBits()) { - Value *P = Builder->CreateTrunc(CI.getOperand(0), - TD->getIntPtrType(CI.getContext()), "tmp"); - return new IntToPtrInst(P, CI.getType()); + // If the source integer type is not the intptr_t type for this target, do a + // trunc or zext to the intptr_t type, then inttoptr of it. This allows the + // cast to be exposed to other transforms. + if (TD) { + if (CI.getOperand(0)->getType()->getScalarSizeInBits() > + TD->getPointerSizeInBits()) { + Value *P = Builder->CreateTrunc(CI.getOperand(0), + TD->getIntPtrType(CI.getContext()), "tmp"); + return new IntToPtrInst(P, CI.getType()); + } + if (CI.getOperand(0)->getType()->getScalarSizeInBits() < + TD->getPointerSizeInBits()) { + Value *P = Builder->CreateZExt(CI.getOperand(0), + TD->getIntPtrType(CI.getContext()), "tmp"); + return new IntToPtrInst(P, CI.getType()); + } } if (Instruction *I = commonCastTransforms(CI)) @@ -1198,17 +1231,22 @@ Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { } Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { - // If the destination integer type is smaller than the intptr_t type for - // this target, do a ptrtoint to intptr_t then do a trunc. This allows the - // trunc to be exposed to other transforms. Don't do this for extending - // ptrtoint's, because we don't know if the target sign or zero extends its - // pointers. - if (TD && - CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { - Value *P = Builder->CreatePtrToInt(CI.getOperand(0), - TD->getIntPtrType(CI.getContext()), - "tmp"); - return new TruncInst(P, CI.getType()); + // If the destination integer type is not the intptr_t type for this target, + // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast + // to be exposed to other transforms. + if (TD) { + if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { + Value *P = Builder->CreatePtrToInt(CI.getOperand(0), + TD->getIntPtrType(CI.getContext()), + "tmp"); + return new TruncInst(P, CI.getType()); + } + if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { + Value *P = Builder->CreatePtrToInt(CI.getOperand(0), + TD->getIntPtrType(CI.getContext()), + "tmp"); + return new ZExtInst(P, CI.getType()); + } } return commonPointerCastTransforms(CI); diff --git a/lib/Transforms/InstCombine/InstCombineCompares.cpp b/lib/Transforms/InstCombine/InstCombineCompares.cpp index e59406c6..72af80f 100644 --- a/lib/Transforms/InstCombine/InstCombineCompares.cpp +++ b/lib/Transforms/InstCombine/InstCombineCompares.cpp @@ -1589,24 +1589,24 @@ Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { bool Changed = false; + Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); /// Orders the operands of the compare so that they are listed from most /// complex to least complex. This puts constants before unary operators, /// before binary operators. - if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { + if (getComplexity(Op0) < getComplexity(Op1)) { I.swapOperands(); + std::swap(Op0, Op1); Changed = true; } - Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); - if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) return ReplaceInstUsesWith(I, V); const Type *Ty = Op0->getType(); // icmp's with boolean values can always be turned into bitwise operations - if (Ty == Type::getInt1Ty(I.getContext())) { + if (Ty->isIntegerTy(1)) { switch (I.getPredicate()) { default: llvm_unreachable("Invalid icmp instruction!"); case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) @@ -1650,7 +1650,7 @@ Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { unsigned BitWidth = 0; if (TD) BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); - else if (Ty->isIntOrIntVector()) + else if (Ty->isIntOrIntVectorTy()) BitWidth = Ty->getScalarSizeInBits(); bool isSignBit = false; diff --git a/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp b/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp index ae728dd..e6c59c7 100644 --- a/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp +++ b/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp @@ -87,7 +87,7 @@ static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI, const Type *SrcPTy = SrcTy->getElementType(); - if (DestPTy->isInteger() || isa<PointerType>(DestPTy) || + if (DestPTy->isIntegerTy() || isa<PointerType>(DestPTy) || isa<VectorType>(DestPTy)) { // If the source is an array, the code below will not succeed. Check to // see if a trivial 'gep P, 0, 0' will help matters. Only do this for @@ -104,7 +104,7 @@ static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI, } if (IC.getTargetData() && - (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) || + (SrcPTy->isIntegerTy() || isa<PointerType>(SrcPTy) || isa<VectorType>(SrcPTy)) && // Do not allow turning this into a load of an integer, which is then // casted to a pointer, this pessimizes pointer analysis a lot. @@ -115,8 +115,9 @@ static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI, // 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 // the result of the loaded value. - Value *NewLoad = + LoadInst *NewLoad = IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName()); + NewLoad->setAlignment(LI.getAlignment()); // Now cast the result of the load. return new BitCastInst(NewLoad, LI.getType()); } @@ -199,12 +200,15 @@ Instruction *InstCombiner::visitLoadInst(LoadInst &LI) { // if (SelectInst *SI = dyn_cast<SelectInst>(Op)) { // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2). - if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) && - isSafeToLoadUnconditionally(SI->getOperand(2), SI)) { - Value *V1 = Builder->CreateLoad(SI->getOperand(1), - SI->getOperand(1)->getName()+".val"); - Value *V2 = Builder->CreateLoad(SI->getOperand(2), - SI->getOperand(2)->getName()+".val"); + unsigned Align = LI.getAlignment(); + if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, TD) && + isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, TD)) { + LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1), + SI->getOperand(1)->getName()+".val"); + LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2), + SI->getOperand(2)->getName()+".val"); + V1->setAlignment(Align); + V2->setAlignment(Align); return SelectInst::Create(SI->getCondition(), V1, V2); } @@ -239,7 +243,7 @@ static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) { const Type *SrcPTy = SrcTy->getElementType(); - if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy)) + if (!DestPTy->isIntegerTy() && !isa<PointerType>(DestPTy)) return 0; /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep" @@ -273,7 +277,7 @@ static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) { SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace()); } - if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy)) + if (!SrcPTy->isIntegerTy() && !isa<PointerType>(SrcPTy)) return 0; // If the pointers point into different address spaces or if they point to @@ -294,7 +298,7 @@ static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) { const Type* CastSrcTy = SIOp0->getType(); const Type* CastDstTy = SrcPTy; if (isa<PointerType>(CastDstTy)) { - if (CastSrcTy->isInteger()) + if (CastSrcTy->isIntegerTy()) opcode = Instruction::IntToPtr; } else if (isa<IntegerType>(CastDstTy)) { if (isa<PointerType>(SIOp0->getType())) diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp index 2e26a75..668c34f 100644 --- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -157,7 +157,7 @@ Instruction *InstCombiner::visitMul(BinaryOperator &I) { } /// i1 mul -> i1 and. - if (I.getType()->isInteger(1)) + if (I.getType()->isIntegerTy(1)) return BinaryOperator::CreateAnd(Op0, Op1); // X*(1 << Y) --> X << Y @@ -314,7 +314,7 @@ Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) { // undef / X -> 0 for integer. // undef / X -> undef for FP (the undef could be a snan). if (isa<UndefValue>(Op0)) { - if (Op0->getType()->isFPOrFPVector()) + if (Op0->getType()->isFPOrFPVectorTy()) return ReplaceInstUsesWith(I, Op0); return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); } @@ -386,7 +386,7 @@ Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) { return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); // It can't be division by zero, hence it must be division by one. - if (I.getType()->isInteger(1)) + if (I.getType()->isIntegerTy(1)) return ReplaceInstUsesWith(I, Op0); if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) { @@ -493,7 +493,7 @@ Instruction *InstCombiner::visitSDiv(BinaryOperator &I) { // If the sign bits of both operands are zero (i.e. we can prove they are // unsigned inputs), turn this into a udiv. - if (I.getType()->isInteger()) { + if (I.getType()->isIntegerTy()) { APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())); if (MaskedValueIsZero(Op0, Mask)) { if (MaskedValueIsZero(Op1, Mask)) { @@ -527,7 +527,7 @@ Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) { Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); if (isa<UndefValue>(Op0)) { // undef % X -> 0 - if (I.getType()->isFPOrFPVector()) + if (I.getType()->isFPOrFPVectorTy()) return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN) return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType())); } @@ -648,7 +648,7 @@ Instruction *InstCombiner::visitSRem(BinaryOperator &I) { // If the sign bits of both operands are zero (i.e. we can prove they are // unsigned inputs), turn this into a urem. - if (I.getType()->isInteger()) { + if (I.getType()->isIntegerTy()) { APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())); if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) { // X srem Y -> X urem Y, iff X and Y don't have sign bit set diff --git a/lib/Transforms/InstCombine/InstCombineSelect.cpp b/lib/Transforms/InstCombine/InstCombineSelect.cpp index 18b2dff..7807d9a 100644 --- a/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -326,44 +326,6 @@ Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI, break; } } - - // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed - // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed - CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; - if (match(TrueVal, m_ConstantInt<-1>()) && - match(FalseVal, m_ConstantInt<0>())) - Pred = ICI->getPredicate(); - else if (match(TrueVal, m_ConstantInt<0>()) && - match(FalseVal, m_ConstantInt<-1>())) - Pred = CmpInst::getInversePredicate(ICI->getPredicate()); - - if (Pred != CmpInst::BAD_ICMP_PREDICATE) { - // If we are just checking for a icmp eq of a single bit and zext'ing it - // to an integer, then shift the bit to the appropriate place and then - // cast to integer to avoid the comparison. - const APInt &Op1CV = CI->getValue(); - - // sext (x <s 0) to i32 --> x>>s31 true if signbit set. - // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear. - if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) || - (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) { - Value *In = ICI->getOperand(0); - Value *Sh = ConstantInt::get(In->getType(), - In->getType()->getScalarSizeInBits()-1); - In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh, - In->getName()+".lobit"), - *ICI); - if (In->getType() != SI.getType()) - In = CastInst::CreateIntegerCast(In, SI.getType(), - true/*SExt*/, "tmp", ICI); - - if (Pred == ICmpInst::ICMP_SGT) - In = InsertNewInstBefore(BinaryOperator::CreateNot(In, - In->getName()+".not"), *ICI); - - return ReplaceInstUsesWith(SI, In); - } - } } if (CmpLHS == TrueVal && CmpRHS == FalseVal) { @@ -479,7 +441,7 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { return ReplaceInstUsesWith(SI, FalseVal); } - if (SI.getType()->isInteger(1)) { + if (SI.getType()->isIntegerTy(1)) { if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) { if (C->getZExtValue()) { // Change: A = select B, true, C --> A = or B, C @@ -516,16 +478,25 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal)) if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) { // select C, 1, 0 -> zext C to int - if (FalseValC->isZero() && TrueValC->getValue() == 1) { - return CastInst::Create(Instruction::ZExt, CondVal, SI.getType()); - } else if (TrueValC->isZero() && FalseValC->getValue() == 1) { - // select C, 0, 1 -> zext !C to int - Value *NotCond = - InsertNewInstBefore(BinaryOperator::CreateNot(CondVal, - "not."+CondVal->getName()), SI); - return CastInst::Create(Instruction::ZExt, NotCond, SI.getType()); + if (FalseValC->isZero() && TrueValC->getValue() == 1) + return new ZExtInst(CondVal, SI.getType()); + + // select C, -1, 0 -> sext C to int + if (FalseValC->isZero() && TrueValC->isAllOnesValue()) + return new SExtInst(CondVal, SI.getType()); + + // select C, 0, 1 -> zext !C to int + if (TrueValC->isZero() && FalseValC->getValue() == 1) { + Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); + return new ZExtInst(NotCond, SI.getType()); } + // select C, 0, -1 -> sext !C to int + if (TrueValC->isZero() && FalseValC->isAllOnesValue()) { + Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); + return new SExtInst(NotCond, SI.getType()); + } + if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) { // If one of the constants is zero (we know they can't both be) and we // have an icmp instruction with zero, and we have an 'and' with the @@ -547,8 +518,7 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE; Value *V = ICA; if (ShouldNotVal) - V = InsertNewInstBefore(BinaryOperator::Create( - Instruction::Xor, V, ICA->getOperand(1)), SI); + V = Builder->CreateXor(V, ICA->getOperand(1)); return ReplaceInstUsesWith(SI, V); } } @@ -659,7 +629,7 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { } // See if we can fold the select into one of our operands. - if (SI.getType()->isInteger()) { + if (SI.getType()->isIntegerTy()) { if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal)) return FoldI; diff --git a/lib/Transforms/InstCombine/InstCombineShifts.cpp b/lib/Transforms/InstCombine/InstCombineShifts.cpp index 321c91d..836bda3 100644 --- a/lib/Transforms/InstCombine/InstCombineShifts.cpp +++ b/lib/Transforms/InstCombine/InstCombineShifts.cpp @@ -12,6 +12,7 @@ //===----------------------------------------------------------------------===// #include "InstCombine.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Support/PatternMatch.h" using namespace llvm; using namespace PatternMatch; @@ -69,10 +70,9 @@ Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1, if (Op1->uge(TypeBits)) { if (I.getOpcode() != Instruction::AShr) return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType())); - else { - I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1)); - return &I; - } + // ashr i32 X, 32 --> ashr i32 X, 31 + I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1)); + return &I; } // ((X*C1) << C2) == (X * (C1 << C2)) @@ -387,7 +387,29 @@ Instruction *InstCombiner::visitShl(BinaryOperator &I) { } Instruction *InstCombiner::visitLShr(BinaryOperator &I) { - return commonShiftTransforms(I); + if (Instruction *R = commonShiftTransforms(I)) + return R; + + Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); + + if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) { + unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); + // ctlz.i32(x)>>5 --> zext(x == 0) + // cttz.i32(x)>>5 --> zext(x == 0) + // ctpop.i32(x)>>5 --> zext(x == -1) + if ((II->getIntrinsicID() == Intrinsic::ctlz || + II->getIntrinsicID() == Intrinsic::cttz || + II->getIntrinsicID() == Intrinsic::ctpop) && + isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == Op1C->getZExtValue()){ + bool isCtPop = II->getIntrinsicID() == Intrinsic::ctpop; + Constant *RHS = ConstantInt::getSigned(Op0->getType(), isCtPop ? -1:0); + Value *Cmp = Builder->CreateICmpEQ(II->getOperand(1), RHS); + return new ZExtInst(Cmp, II->getType()); + } + } + + return 0; } Instruction *InstCombiner::visitAShr(BinaryOperator &I) { diff --git a/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp index 74a1b68..5e9a52f 100644 --- a/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp +++ b/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp @@ -107,7 +107,7 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, assert((TD || !isa<PointerType>(VTy)) && "SimplifyDemandedBits needs to know bit widths!"); assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) && - (!VTy->isIntOrIntVector() || + (!VTy->isIntOrIntVectorTy() || VTy->getScalarSizeInBits() == BitWidth) && KnownZero.getBitWidth() == BitWidth && KnownOne.getBitWidth() == BitWidth && @@ -138,11 +138,11 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, return 0; APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0); - APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne; + APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); Instruction *I = dyn_cast<Instruction>(V); if (!I) { - ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth); + ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth); return 0; // Only analyze instructions. } @@ -219,7 +219,7 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, switch (I->getOpcode()) { default: - ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth); + ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth); break; case Instruction::And: // If either the LHS or the RHS are Zero, the result is zero. @@ -249,9 +249,9 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, return I; // Output known-1 bits are only known if set in both the LHS & RHS. - RHSKnownOne &= LHSKnownOne; + KnownOne = RHSKnownOne & LHSKnownOne; // Output known-0 are known to be clear if zero in either the LHS | RHS. - RHSKnownZero |= LHSKnownZero; + KnownZero = RHSKnownZero | LHSKnownZero; break; case Instruction::Or: // If either the LHS or the RHS are One, the result is One. @@ -286,9 +286,9 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, return I; // Output known-0 bits are only known if clear in both the LHS & RHS. - RHSKnownZero &= LHSKnownZero; + KnownZero = RHSKnownZero & LHSKnownZero; // Output known-1 are known to be set if set in either the LHS | RHS. - RHSKnownOne |= LHSKnownOne; + KnownOne = RHSKnownOne | LHSKnownOne; break; case Instruction::Xor: { if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask, @@ -306,13 +306,6 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, if ((DemandedMask & LHSKnownZero) == DemandedMask) return I->getOperand(1); - // Output known-0 bits are known if clear or set in both the LHS & RHS. - APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) | - (RHSKnownOne & LHSKnownOne); - // Output known-1 are known to be set if set in only one of the LHS, RHS. - APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) | - (RHSKnownOne & LHSKnownZero); - // If all of the demanded bits are known to be zero on one side or the // other, turn this into an *inclusive* or. // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 @@ -368,10 +361,11 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, BinaryOperator::CreateXor(NewAnd, XorC, "tmp"); return InsertNewInstBefore(NewXor, *I); } - - - RHSKnownZero = KnownZeroOut; - RHSKnownOne = KnownOneOut; + + // Output known-0 bits are known if clear or set in both the LHS & RHS. + KnownZero= (RHSKnownZero & LHSKnownZero) | (RHSKnownOne & LHSKnownOne); + // Output known-1 are known to be set if set in only one of the LHS, RHS. + KnownOne = (RHSKnownZero & LHSKnownOne) | (RHSKnownOne & LHSKnownZero); break; } case Instruction::Select: @@ -389,61 +383,61 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, return I; // Only known if known in both the LHS and RHS. - RHSKnownOne &= LHSKnownOne; - RHSKnownZero &= LHSKnownZero; + KnownOne = RHSKnownOne & LHSKnownOne; + KnownZero = RHSKnownZero & LHSKnownZero; break; case Instruction::Trunc: { unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits(); DemandedMask.zext(truncBf); - RHSKnownZero.zext(truncBf); - RHSKnownOne.zext(truncBf); + KnownZero.zext(truncBf); + KnownOne.zext(truncBf); if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; DemandedMask.trunc(BitWidth); - RHSKnownZero.trunc(BitWidth); - RHSKnownOne.trunc(BitWidth); - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); + KnownZero.trunc(BitWidth); + KnownOne.trunc(BitWidth); + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); break; } case Instruction::BitCast: - if (!I->getOperand(0)->getType()->isIntOrIntVector()) - return false; // vector->int or fp->int? + if (!I->getOperand(0)->getType()->isIntOrIntVectorTy()) + return 0; // vector->int or fp->int? if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) { if (const VectorType *SrcVTy = dyn_cast<VectorType>(I->getOperand(0)->getType())) { if (DstVTy->getNumElements() != SrcVTy->getNumElements()) // Don't touch a bitcast between vectors of different element counts. - return false; + return 0; } else // Don't touch a scalar-to-vector bitcast. - return false; + return 0; } else if (isa<VectorType>(I->getOperand(0)->getType())) // Don't touch a vector-to-scalar bitcast. - return false; + return 0; if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); break; case Instruction::ZExt: { // Compute the bits in the result that are not present in the input. unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits(); DemandedMask.trunc(SrcBitWidth); - RHSKnownZero.trunc(SrcBitWidth); - RHSKnownOne.trunc(SrcBitWidth); + KnownZero.trunc(SrcBitWidth); + KnownOne.trunc(SrcBitWidth); if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; DemandedMask.zext(BitWidth); - RHSKnownZero.zext(BitWidth); - RHSKnownOne.zext(BitWidth); - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); + KnownZero.zext(BitWidth); + KnownOne.zext(BitWidth); + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); // The top bits are known to be zero. - RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); + KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); break; } case Instruction::SExt: { @@ -460,27 +454,27 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, InputDemandedBits.set(SrcBitWidth-1); InputDemandedBits.trunc(SrcBitWidth); - RHSKnownZero.trunc(SrcBitWidth); - RHSKnownOne.trunc(SrcBitWidth); + KnownZero.trunc(SrcBitWidth); + KnownOne.trunc(SrcBitWidth); if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; InputDemandedBits.zext(BitWidth); - RHSKnownZero.zext(BitWidth); - RHSKnownOne.zext(BitWidth); - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); + KnownZero.zext(BitWidth); + KnownOne.zext(BitWidth); + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); // If the sign bit of the input is known set or clear, then we know the // top bits of the result. // If the input sign bit is known zero, or if the NewBits are not demanded // convert this into a zero extension. - if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) { + if (KnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) { // Convert to ZExt cast CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName()); return InsertNewInstBefore(NewCast, *I); - } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set - RHSKnownOne |= NewBits; + } else if (KnownOne[SrcBitWidth-1]) { // Input sign bit known set + KnownOne |= NewBits; } break; } @@ -540,12 +534,12 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, // Bits are known one if they are known zero in one operand and one in the // other, and there is no input carry. - RHSKnownOne = ((LHSKnownZero & RHSVal) | - (LHSKnownOne & ~RHSVal)) & ~CarryBits; + KnownOne = ((LHSKnownZero & RHSVal) | + (LHSKnownOne & ~RHSVal)) & ~CarryBits; // Bits are known zero if they are known zero in both operands and there // is no input carry. - RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits; + KnownZero = LHSKnownZero & ~RHSVal & ~CarryBits; } else { // If the high-bits of this ADD are not demanded, then it does not demand // the high bits of its LHS or RHS. @@ -578,21 +572,21 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, } // Otherwise just hand the sub off to ComputeMaskedBits to fill in // the known zeros and ones. - ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth); + ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth); break; case Instruction::Shl: if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) { uint64_t ShiftAmt = SA->getLimitedValue(BitWidth); APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt)); if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); - RHSKnownZero <<= ShiftAmt; - RHSKnownOne <<= ShiftAmt; + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); + KnownZero <<= ShiftAmt; + KnownOne <<= ShiftAmt; // low bits known zero. if (ShiftAmt) - RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); + KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); } break; case Instruction::LShr: @@ -603,15 +597,15 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, // Unsigned shift right. APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt)); if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); - RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt); - RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt); + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); + KnownZero = APIntOps::lshr(KnownZero, ShiftAmt); + KnownOne = APIntOps::lshr(KnownOne, ShiftAmt); if (ShiftAmt) { // Compute the new bits that are at the top now. APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt)); - RHSKnownZero |= HighBits; // high bits known zero. + KnownZero |= HighBits; // high bits known zero. } } break; @@ -642,13 +636,13 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, if (DemandedMask.countLeadingZeros() <= ShiftAmt) DemandedMaskIn.set(BitWidth-1); if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn, - RHSKnownZero, RHSKnownOne, Depth+1)) + KnownZero, KnownOne, Depth+1)) return I; - assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?"); + assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); // Compute the new bits that are at the top now. APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt)); - RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt); - RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt); + KnownZero = APIntOps::lshr(KnownZero, ShiftAmt); + KnownOne = APIntOps::lshr(KnownOne, ShiftAmt); // Handle the sign bits. APInt SignBit(APInt::getSignBit(BitWidth)); @@ -657,14 +651,14 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, // If the input sign bit is known to be zero, or if none of the top bits // are demanded, turn this into an unsigned shift right. - if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] || + if (BitWidth <= ShiftAmt || KnownZero[BitWidth-ShiftAmt-1] || (HighBits & ~DemandedMask) == HighBits) { // Perform the logical shift right. Instruction *NewVal = BinaryOperator::CreateLShr( I->getOperand(0), SA, I->getName()); return InsertNewInstBefore(NewVal, *I); - } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one. - RHSKnownOne |= HighBits; + } else if ((KnownOne & SignBit) != 0) { // New bits are known one. + KnownOne |= HighBits; } } break; @@ -681,10 +675,19 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, LHSKnownZero, LHSKnownOne, Depth+1)) return I; + // The low bits of LHS are unchanged by the srem. + KnownZero = LHSKnownZero & LowBits; + KnownOne = LHSKnownOne & LowBits; + + // If LHS is non-negative or has all low bits zero, then the upper bits + // are all zero. if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits)) - LHSKnownZero |= ~LowBits; + KnownZero |= ~LowBits; - KnownZero |= LHSKnownZero & DemandedMask; + // If LHS is negative and not all low bits are zero, then the upper bits + // are all one. + if (LHSKnownOne[BitWidth-1] && ((LHSKnownOne & LowBits) != 0)) + KnownOne |= ~LowBits; assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?"); } @@ -743,15 +746,15 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, } } } - ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth); + ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth); break; } // If the client is only demanding bits that we know, return the known // constant. - if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) - return Constant::getIntegerValue(VTy, RHSKnownOne); - return false; + if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) + return Constant::getIntegerValue(VTy, KnownOne); + return 0; } @@ -764,7 +767,7 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, /// operation, the operation is simplified, then the resultant value is /// returned. This returns null if no change was made. Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, - APInt& UndefElts, + APInt &UndefElts, unsigned Depth) { unsigned VWidth = cast<VectorType>(V->getType())->getNumElements(); APInt EltMask(APInt::getAllOnesValue(VWidth)); @@ -774,13 +777,15 @@ Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, // If the entire vector is undefined, just return this info. UndefElts = EltMask; return 0; - } else if (DemandedElts == 0) { // If nothing is demanded, provide undef. + } + + if (DemandedElts == 0) { // If nothing is demanded, provide undef. UndefElts = EltMask; return UndefValue::get(V->getType()); } UndefElts = 0; - if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) { + if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) { const Type *EltTy = cast<VectorType>(V->getType())->getElementType(); Constant *Undef = UndefValue::get(EltTy); @@ -789,23 +794,25 @@ Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, if (!DemandedElts[i]) { // If not demanded, set to undef. Elts.push_back(Undef); UndefElts.set(i); - } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef. + } else if (isa<UndefValue>(CV->getOperand(i))) { // Already undef. Elts.push_back(Undef); UndefElts.set(i); } else { // Otherwise, defined. - Elts.push_back(CP->getOperand(i)); + Elts.push_back(CV->getOperand(i)); } // If we changed the constant, return it. Constant *NewCP = ConstantVector::get(Elts); - return NewCP != CP ? NewCP : 0; - } else if (isa<ConstantAggregateZero>(V)) { + return NewCP != CV ? NewCP : 0; + } + + if (isa<ConstantAggregateZero>(V)) { // Simplify the CAZ to a ConstantVector where the non-demanded elements are // set to undef. // Check if this is identity. If so, return 0 since we are not simplifying // anything. - if (DemandedElts == ((1ULL << VWidth) -1)) + if (DemandedElts.isAllOnesValue()) return 0; const Type *EltTy = cast<VectorType>(V->getType())->getElementType(); diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp index f11f557..20fda1a 100644 --- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp +++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp @@ -162,7 +162,8 @@ Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { // property. if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) { APInt UndefElts(VectorWidth, 0); - APInt DemandedMask(VectorWidth, 1 << IndexVal); + APInt DemandedMask(VectorWidth, 0); + DemandedMask.set(IndexVal); if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask, UndefElts)) { EI.setOperand(0, V); diff --git a/lib/Transforms/InstCombine/InstructionCombining.cpp b/lib/Transforms/InstCombine/InstructionCombining.cpp index 93b1961..96c0342 100644 --- a/lib/Transforms/InstCombine/InstructionCombining.cpp +++ b/lib/Transforms/InstCombine/InstructionCombining.cpp @@ -158,7 +158,7 @@ Value *InstCombiner::dyn_castNegVal(Value *V) const { return ConstantExpr::getNeg(C); if (ConstantVector *C = dyn_cast<ConstantVector>(V)) - if (C->getType()->getElementType()->isInteger()) + if (C->getType()->getElementType()->isIntegerTy()) return ConstantExpr::getNeg(C); return 0; @@ -177,7 +177,7 @@ Value *InstCombiner::dyn_castFNegVal(Value *V) const { return ConstantExpr::getFNeg(C); if (ConstantVector *C = dyn_cast<ConstantVector>(V)) - if (C->getType()->getElementType()->isFloatingPoint()) + if (C->getType()->getElementType()->isFloatingPointTy()) return ConstantExpr::getFNeg(C); return 0; @@ -226,7 +226,7 @@ Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) { if (isa<Constant>(TV) || isa<Constant>(FV)) { // Bool selects with constant operands can be folded to logical ops. - if (SI->getType()->isInteger(1)) return 0; + if (SI->getType()->isIntegerTy(1)) return 0; Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this); Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this); @@ -596,7 +596,7 @@ Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) { // (where tmp = 8*tmp2) into: // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast - if (TD && isa<ArrayType>(SrcElTy) && ResElTy->isInteger(8)) { + if (TD && isa<ArrayType>(SrcElTy) && ResElTy->isIntegerTy(8)) { uint64_t ArrayEltSize = TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()); diff --git a/lib/Transforms/InstCombine/Makefile b/lib/Transforms/InstCombine/Makefile index f9de42a..0c488e78 100644 --- a/lib/Transforms/InstCombine/Makefile +++ b/lib/Transforms/InstCombine/Makefile @@ -10,7 +10,6 @@ LEVEL = ../../.. LIBRARYNAME = LLVMInstCombine BUILD_ARCHIVE = 1 -CXXFLAGS = -fno-rtti include $(LEVEL)/Makefile.common diff --git a/lib/Transforms/Instrumentation/Makefile b/lib/Transforms/Instrumentation/Makefile index 1238896..6cbc7a9 100644 --- a/lib/Transforms/Instrumentation/Makefile +++ b/lib/Transforms/Instrumentation/Makefile @@ -10,7 +10,6 @@ LEVEL = ../../.. LIBRARYNAME = LLVMInstrumentation BUILD_ARCHIVE = 1 -CXXFLAGS = -fno-rtti include $(LEVEL)/Makefile.common diff --git a/lib/Transforms/Instrumentation/ProfilingUtils.cpp b/lib/Transforms/Instrumentation/ProfilingUtils.cpp index 3214c8c..8662a82 100644 --- a/lib/Transforms/Instrumentation/ProfilingUtils.cpp +++ b/lib/Transforms/Instrumentation/ProfilingUtils.cpp @@ -84,7 +84,7 @@ void llvm::InsertProfilingInitCall(Function *MainFn, const char *FnName, AI = MainFn->arg_begin(); // If the program looked at argc, have it look at the return value of the // init call instead. - if (!AI->getType()->isInteger(32)) { + if (!AI->getType()->isIntegerTy(32)) { Instruction::CastOps opcode; if (!AI->use_empty()) { opcode = CastInst::getCastOpcode(InitCall, true, AI->getType(), true); diff --git a/lib/Transforms/Scalar/CodeGenPrepare.cpp b/lib/Transforms/Scalar/CodeGenPrepare.cpp index c3139a5..21e6f89 100644 --- a/lib/Transforms/Scalar/CodeGenPrepare.cpp +++ b/lib/Transforms/Scalar/CodeGenPrepare.cpp @@ -32,7 +32,6 @@ #include "llvm/ADT/SmallSet.h" #include "llvm/Assembly/Writer.h" #include "llvm/Support/CallSite.h" -#include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/PatternMatch.h" @@ -40,9 +39,6 @@ using namespace llvm; using namespace llvm::PatternMatch; -static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak", - cl::init(false), cl::Hidden); - namespace { class CodeGenPrepare : public FunctionPass { /// TLI - Keep a pointer of a TargetLowering to consult for determining @@ -63,6 +59,10 @@ namespace { AU.addPreserved<ProfileInfo>(); } + virtual void releaseMemory() { + BackEdges.clear(); + } + private: bool EliminateMostlyEmptyBlocks(Function &F); bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; @@ -297,6 +297,70 @@ void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); } +/// FindReusablePredBB - Check all of the predecessors of the block DestPHI +/// lives in to see if there is a block that we can reuse as a critical edge +/// from TIBB. +static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) { + BasicBlock *Dest = DestPHI->getParent(); + + /// TIPHIValues - This array is lazily computed to determine the values of + /// PHIs in Dest that TI would provide. + SmallVector<Value*, 32> TIPHIValues; + + /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB. + unsigned TIBBEntryNo = 0; + + // Check to see if Dest has any blocks that can be used as a split edge for + // this terminator. + for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) { + BasicBlock *Pred = DestPHI->getIncomingBlock(pi); + // To be usable, the pred has to end with an uncond branch to the dest. + BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); + if (!PredBr || !PredBr->isUnconditional()) + continue; + // Must be empty other than the branch and debug info. + BasicBlock::iterator I = Pred->begin(); + while (isa<DbgInfoIntrinsic>(I)) + I++; + if (&*I != PredBr) + continue; + // Cannot be the entry block; its label does not get emitted. + if (Pred == &Dest->getParent()->getEntryBlock()) + continue; + + // Finally, since we know that Dest has phi nodes in it, we have to make + // sure that jumping to Pred will have the same effect as going to Dest in + // terms of PHI values. + PHINode *PN; + unsigned PHINo = 0; + unsigned PredEntryNo = pi; + + bool FoundMatch = true; + for (BasicBlock::iterator I = Dest->begin(); + (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { + if (PHINo == TIPHIValues.size()) { + if (PN->getIncomingBlock(TIBBEntryNo) != TIBB) + TIBBEntryNo = PN->getBasicBlockIndex(TIBB); + TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo)); + } + + // If the PHI entry doesn't work, we can't use this pred. + if (PN->getIncomingBlock(PredEntryNo) != Pred) + PredEntryNo = PN->getBasicBlockIndex(Pred); + + if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) { + FoundMatch = false; + break; + } + } + + // If we found a workable predecessor, change TI to branch to Succ. + if (FoundMatch) + return Pred; + } + return 0; +} + /// SplitEdgeNicely - Split the critical edge from TI to its specified /// successor if it will improve codegen. We only do this if the successor has @@ -311,13 +375,12 @@ static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, BasicBlock *Dest = TI->getSuccessor(SuccNum); assert(isa<PHINode>(Dest->begin()) && "This should only be called if Dest has a PHI!"); + PHINode *DestPHI = cast<PHINode>(Dest->begin()); // Do not split edges to EH landing pads. - if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) { + if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) if (Invoke->getSuccessor(1) == Dest) return; - } - // As a hack, never split backedges of loops. Even though the copy for any // PHIs inserted on the backedge would be dead for exits from the loop, we @@ -325,92 +388,16 @@ static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, if (BackEdges.count(std::make_pair(TIBB, Dest))) return; - if (!FactorCommonPreds) { - /// TIPHIValues - This array is lazily computed to determine the values of - /// PHIs in Dest that TI would provide. - SmallVector<Value*, 32> TIPHIValues; - - // Check to see if Dest has any blocks that can be used as a split edge for - // this terminator. - for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { - BasicBlock *Pred = *PI; - // To be usable, the pred has to end with an uncond branch to the dest. - BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); - if (!PredBr || !PredBr->isUnconditional()) - continue; - // Must be empty other than the branch and debug info. - BasicBlock::iterator I = Pred->begin(); - while (isa<DbgInfoIntrinsic>(I)) - I++; - if (dyn_cast<Instruction>(I) != PredBr) - continue; - // Cannot be the entry block; its label does not get emitted. - if (Pred == &(Dest->getParent()->getEntryBlock())) - continue; - - // Finally, since we know that Dest has phi nodes in it, we have to make - // sure that jumping to Pred will have the same effect as going to Dest in - // terms of PHI values. - PHINode *PN; - unsigned PHINo = 0; - bool FoundMatch = true; - for (BasicBlock::iterator I = Dest->begin(); - (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { - if (PHINo == TIPHIValues.size()) - TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); - - // If the PHI entry doesn't work, we can't use this pred. - if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { - FoundMatch = false; - break; - } - } - - // If we found a workable predecessor, change TI to branch to Succ. - if (FoundMatch) { - ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>(); - if (PFI) - PFI->splitEdge(TIBB, Dest, Pred); - Dest->removePredecessor(TIBB); - TI->setSuccessor(SuccNum, Pred); - return; - } - } - - SplitCriticalEdge(TI, SuccNum, P, true); + if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) { + ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>(); + if (PFI) + PFI->splitEdge(TIBB, Dest, ReuseBB); + Dest->removePredecessor(TIBB); + TI->setSuccessor(SuccNum, ReuseBB); return; } - PHINode *PN; - SmallVector<Value*, 8> TIPHIValues; - for (BasicBlock::iterator I = Dest->begin(); - (PN = dyn_cast<PHINode>(I)); ++I) - TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); - - SmallVector<BasicBlock*, 8> IdenticalPreds; - for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { - BasicBlock *Pred = *PI; - if (BackEdges.count(std::make_pair(Pred, Dest))) - continue; - if (PI == TIBB) - IdenticalPreds.push_back(Pred); - else { - bool Identical = true; - unsigned PHINo = 0; - for (BasicBlock::iterator I = Dest->begin(); - (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) - if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { - Identical = false; - break; - } - if (Identical) - IdenticalPreds.push_back(Pred); - } - } - - assert(!IdenticalPreds.empty()); - SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(), - ".critedge", P); + SplitCriticalEdge(TI, SuccNum, P, true); } diff --git a/lib/Transforms/Scalar/DeadStoreElimination.cpp b/lib/Transforms/Scalar/DeadStoreElimination.cpp index 320afa1..09c01d3 100644 --- a/lib/Transforms/Scalar/DeadStoreElimination.cpp +++ b/lib/Transforms/Scalar/DeadStoreElimination.cpp @@ -44,8 +44,14 @@ namespace { virtual bool runOnFunction(Function &F) { bool Changed = false; + + DominatorTree &DT = getAnalysis<DominatorTree>(); + for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) - Changed |= runOnBasicBlock(*I); + // Only check non-dead blocks. Dead blocks may have strange pointer + // cycles that will confuse alias analysis. + if (DT.isReachableFromEntry(I)) + Changed |= runOnBasicBlock(*I); return Changed; } diff --git a/lib/Transforms/Scalar/GVN.cpp b/lib/Transforms/Scalar/GVN.cpp index b29fe74..3ce7482 100644 --- a/lib/Transforms/Scalar/GVN.cpp +++ b/lib/Transforms/Scalar/GVN.cpp @@ -60,6 +60,7 @@ STATISTIC(NumPRELoad, "Number of loads PRE'd"); static cl::opt<bool> EnablePRE("enable-pre", cl::init(true), cl::Hidden); static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); +static cl::opt<bool> EnableFullLoadPRE("enable-full-load-pre", cl::init(false)); //===----------------------------------------------------------------------===// // ValueTable Class @@ -1522,8 +1523,6 @@ bool GVN::processNonLocalLoad(LoadInst *LI, while (TmpBB->getSinglePredecessor()) { isSinglePred = true; TmpBB = TmpBB->getSinglePredecessor(); - if (!TmpBB) // If haven't found any, bail now. - return false; if (TmpBB == LoadBB) // Infinite (unreachable) loop. return false; if (Blockers.count(TmpBB)) @@ -1539,10 +1538,12 @@ bool GVN::processNonLocalLoad(LoadInst *LI, // at least one of the values is LI. Since this means that we won't be able // to eliminate LI even if we insert uses in the other predecessors, we will // end up increasing code size. Reject this by scanning for LI. - for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) - if (ValuesPerBlock[i].isSimpleValue() && - ValuesPerBlock[i].getSimpleValue() == LI) - return false; + if (!EnableFullLoadPRE) { + for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) + if (ValuesPerBlock[i].isSimpleValue() && + ValuesPerBlock[i].getSimpleValue() == LI) + return false; + } // FIXME: It is extremely unclear what this loop is doing, other than // artificially restricting loadpre. @@ -1566,13 +1567,9 @@ bool GVN::processNonLocalLoad(LoadInst *LI, return false; } - // Okay, we have some hope :). Check to see if the loaded value is fully - // available in all but one predecessor. - // FIXME: If we could restructure the CFG, we could make a common pred with - // all the preds that don't have an available LI and insert a new load into - // that one block. - BasicBlock *UnavailablePred = 0; - + // Check to see how many predecessors have the loaded value fully + // available. + DenseMap<BasicBlock*, Value*> PredLoads; DenseMap<BasicBlock*, char> FullyAvailableBlocks; for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; @@ -1581,79 +1578,93 @@ bool GVN::processNonLocalLoad(LoadInst *LI, for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; ++PI) { - if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) + BasicBlock *Pred = *PI; + if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { continue; - - // If this load is not available in multiple predecessors, reject it. - if (UnavailablePred && UnavailablePred != *PI) + } + PredLoads[Pred] = 0; + // We don't currently handle critical edges :( + if (Pred->getTerminator()->getNumSuccessors() != 1) { + DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" + << Pred->getName() << "': " << *LI << '\n'); return false; - UnavailablePred = *PI; + } } - assert(UnavailablePred != 0 && + // Decide whether PRE is profitable for this load. + unsigned NumUnavailablePreds = PredLoads.size(); + assert(NumUnavailablePreds != 0 && "Fully available value should be eliminated above!"); - - // We don't currently handle critical edges :( - if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) { - DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '" - << UnavailablePred->getName() << "': " << *LI << '\n'); - return false; + if (!EnableFullLoadPRE) { + // If this load is unavailable in multiple predecessors, reject it. + // FIXME: If we could restructure the CFG, we could make a common pred with + // all the preds that don't have an available LI and insert a new load into + // that one block. + if (NumUnavailablePreds != 1) + return false; } - - // Do PHI translation to get its value in the predecessor if necessary. The - // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. - // + + // Check if the load can safely be moved to all the unavailable predecessors. + bool CanDoPRE = true; SmallVector<Instruction*, 8> NewInsts; - - // If all preds have a single successor, then we know it is safe to insert the - // load on the pred (?!?), so we can insert code to materialize the pointer if - // it is not available. - PHITransAddr Address(LI->getOperand(0), TD); - Value *LoadPtr = 0; - if (allSingleSucc) { - LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, - *DT, NewInsts); - } else { - Address.PHITranslateValue(LoadBB, UnavailablePred); - LoadPtr = Address.getAddr(); + for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), + E = PredLoads.end(); I != E; ++I) { + BasicBlock *UnavailablePred = I->first; + + // Do PHI translation to get its value in the predecessor if necessary. The + // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. + + // If all preds have a single successor, then we know it is safe to insert + // the load on the pred (?!?), so we can insert code to materialize the + // pointer if it is not available. + PHITransAddr Address(LI->getOperand(0), TD); + Value *LoadPtr = 0; + if (allSingleSucc) { + LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, + *DT, NewInsts); + } else { + Address.PHITranslateValue(LoadBB, UnavailablePred); + LoadPtr = Address.getAddr(); - // Make sure the value is live in the predecessor. - if (Instruction *Inst = dyn_cast_or_null<Instruction>(LoadPtr)) - if (!DT->dominates(Inst->getParent(), UnavailablePred)) - LoadPtr = 0; - } + // Make sure the value is live in the predecessor. + if (Instruction *Inst = dyn_cast_or_null<Instruction>(LoadPtr)) + if (!DT->dominates(Inst->getParent(), UnavailablePred)) + LoadPtr = 0; + } - // If we couldn't find or insert a computation of this phi translated value, - // we fail PRE. - if (LoadPtr == 0) { - assert(NewInsts.empty() && "Shouldn't insert insts on failure"); - DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " - << *LI->getOperand(0) << "\n"); - return false; - } + // If we couldn't find or insert a computation of this phi translated value, + // we fail PRE. + if (LoadPtr == 0) { + DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " + << *LI->getOperand(0) << "\n"); + CanDoPRE = false; + break; + } - // Assign value numbers to these new instructions. - for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { - // FIXME: We really _ought_ to insert these value numbers into their - // parent's availability map. However, in doing so, we risk getting into - // ordering issues. If a block hasn't been processed yet, we would be - // marking a value as AVAIL-IN, which isn't what we intend. - VN.lookup_or_add(NewInsts[i]); + // Make sure it is valid to move this load here. We have to watch out for: + // @1 = getelementptr (i8* p, ... + // test p and branch if == 0 + // load @1 + // It is valid to have the getelementptr before the test, even if p can be 0, + // as getelementptr only does address arithmetic. + // If we are not pushing the value through any multiple-successor blocks + // we do not have this case. Otherwise, check that the load is safe to + // put anywhere; this can be improved, but should be conservatively safe. + if (!allSingleSucc && + // FIXME: REEVALUTE THIS. + !isSafeToLoadUnconditionally(LoadPtr, + UnavailablePred->getTerminator(), + LI->getAlignment(), TD)) { + CanDoPRE = false; + break; + } + + I->second = LoadPtr; } - - // Make sure it is valid to move this load here. We have to watch out for: - // @1 = getelementptr (i8* p, ... - // test p and branch if == 0 - // load @1 - // It is valid to have the getelementptr before the test, even if p can be 0, - // as getelementptr only does address arithmetic. - // If we are not pushing the value through any multiple-successor blocks - // we do not have this case. Otherwise, check that the load is safe to - // put anywhere; this can be improved, but should be conservatively safe. - if (!allSingleSucc && - // FIXME: REEVALUTE THIS. - !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) { - assert(NewInsts.empty() && "Should not have inserted instructions"); + + if (!CanDoPRE) { + while (!NewInsts.empty()) + NewInsts.pop_back_val()->eraseFromParent(); return false; } @@ -1665,12 +1676,28 @@ bool GVN::processNonLocalLoad(LoadInst *LI, dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back() << '\n'); - Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, - LI->getAlignment(), - UnavailablePred->getTerminator()); + // Assign value numbers to the new instructions. + for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { + // FIXME: We really _ought_ to insert these value numbers into their + // parent's availability map. However, in doing so, we risk getting into + // ordering issues. If a block hasn't been processed yet, we would be + // marking a value as AVAIL-IN, which isn't what we intend. + VN.lookup_or_add(NewInsts[i]); + } - // Add the newly created load. - ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad)); + for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), + E = PredLoads.end(); I != E; ++I) { + BasicBlock *UnavailablePred = I->first; + Value *LoadPtr = I->second; + + Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, + LI->getAlignment(), + UnavailablePred->getTerminator()); + + // Add the newly created load. + ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, + NewLoad)); + } // Perform PHI construction. Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, @@ -1864,6 +1891,10 @@ Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) { /// by inserting it into the appropriate sets bool GVN::processInstruction(Instruction *I, SmallVectorImpl<Instruction*> &toErase) { + // Ignore dbg info intrinsics. + if (isa<DbgInfoIntrinsic>(I)) + return false; + if (LoadInst *LI = dyn_cast<LoadInst>(I)) { bool Changed = processLoad(LI, toErase); @@ -2075,7 +2106,7 @@ bool GVN::performPRE(Function &F) { for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); PI != PE; ++PI) { // We're not interested in PRE where the block is its - // own predecessor, on in blocks with predecessors + // own predecessor, or in blocks with predecessors // that are not reachable. if (*PI == CurrentBlock) { NumWithout = 2; @@ -2123,10 +2154,10 @@ bool GVN::performPRE(Function &F) { continue; } - // Instantiate the expression the in predecessor that lacked it. + // Instantiate the expression in the predecessor that lacked it. // Because we are going top-down through the block, all value numbers // will be available in the predecessor by the time we need them. Any - // that weren't original present will have been instantiated earlier + // that weren't originally present will have been instantiated earlier // in this loop. Instruction *PREInstr = CurInst->clone(); bool success = true; diff --git a/lib/Transforms/Scalar/IndVarSimplify.cpp b/lib/Transforms/Scalar/IndVarSimplify.cpp index 17f7d98..5302fdc 100644 --- a/lib/Transforms/Scalar/IndVarSimplify.cpp +++ b/lib/Transforms/Scalar/IndVarSimplify.cpp @@ -364,23 +364,17 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { if (ExitingBlock) NeedCannIV = true; } - for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { - const SCEV *Stride = IU->StrideOrder[i]; - const Type *Ty = SE->getEffectiveSCEVType(Stride->getType()); + for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) { + const Type *Ty = + SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType()); if (!LargestType || SE->getTypeSizeInBits(Ty) > SE->getTypeSizeInBits(LargestType)) LargestType = Ty; - - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[i]); - assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - - if (!SI->second->Users.empty()) - NeedCannIV = true; + NeedCannIV = true; } - // Now that we know the largest of of the induction variable expressions + // Now that we know the largest of the induction variable expressions // in this loop, insert a canonical induction variable of the largest size. Value *IndVar = 0; if (NeedCannIV) { @@ -455,72 +449,64 @@ void IndVarSimplify::RewriteIVExpressions(Loop *L, const Type *LargestType, // add the offsets to the primary induction variable and cast, avoiding // the need for the code evaluation methods to insert induction variables // of different sizes. - for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { - const SCEV *Stride = IU->StrideOrder[i]; - - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[i]); - assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - ilist<IVStrideUse> &List = SI->second->Users; - for (ilist<IVStrideUse>::iterator UI = List.begin(), - E = List.end(); UI != E; ++UI) { - Value *Op = UI->getOperandValToReplace(); - const Type *UseTy = Op->getType(); - Instruction *User = UI->getUser(); - - // Compute the final addrec to expand into code. - const SCEV *AR = IU->getReplacementExpr(*UI); - - // Evaluate the expression out of the loop, if possible. - if (!L->contains(UI->getUser())) { - const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop()); - if (ExitVal->isLoopInvariant(L)) - AR = ExitVal; - } + for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) { + const SCEV *Stride = UI->getStride(); + Value *Op = UI->getOperandValToReplace(); + const Type *UseTy = Op->getType(); + Instruction *User = UI->getUser(); + + // Compute the final addrec to expand into code. + const SCEV *AR = IU->getReplacementExpr(*UI); + + // Evaluate the expression out of the loop, if possible. + if (!L->contains(UI->getUser())) { + const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop()); + if (ExitVal->isLoopInvariant(L)) + AR = ExitVal; + } - // FIXME: It is an extremely bad idea to indvar substitute anything more - // complex than affine induction variables. Doing so will put expensive - // polynomial evaluations inside of the loop, and the str reduction pass - // currently can only reduce affine polynomials. For now just disable - // indvar subst on anything more complex than an affine addrec, unless - // it can be expanded to a trivial value. - if (!AR->isLoopInvariant(L) && !Stride->isLoopInvariant(L)) - continue; + // FIXME: It is an extremely bad idea to indvar substitute anything more + // complex than affine induction variables. Doing so will put expensive + // polynomial evaluations inside of the loop, and the str reduction pass + // currently can only reduce affine polynomials. For now just disable + // indvar subst on anything more complex than an affine addrec, unless + // it can be expanded to a trivial value. + if (!AR->isLoopInvariant(L) && !Stride->isLoopInvariant(L)) + continue; - // Determine the insertion point for this user. By default, insert - // immediately before the user. The SCEVExpander class will automatically - // hoist loop invariants out of the loop. For PHI nodes, there may be - // multiple uses, so compute the nearest common dominator for the - // incoming blocks. - Instruction *InsertPt = User; - if (PHINode *PHI = dyn_cast<PHINode>(InsertPt)) - for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) - if (PHI->getIncomingValue(i) == Op) { - if (InsertPt == User) - InsertPt = PHI->getIncomingBlock(i)->getTerminator(); - else - InsertPt = - DT->findNearestCommonDominator(InsertPt->getParent(), - PHI->getIncomingBlock(i)) - ->getTerminator(); - } - - // Now expand it into actual Instructions and patch it into place. - Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt); - - // Patch the new value into place. - if (Op->hasName()) - NewVal->takeName(Op); - User->replaceUsesOfWith(Op, NewVal); - UI->setOperandValToReplace(NewVal); - DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n' - << " into = " << *NewVal << "\n"); - ++NumRemoved; - Changed = true; - - // The old value may be dead now. - DeadInsts.push_back(Op); - } + // Determine the insertion point for this user. By default, insert + // immediately before the user. The SCEVExpander class will automatically + // hoist loop invariants out of the loop. For PHI nodes, there may be + // multiple uses, so compute the nearest common dominator for the + // incoming blocks. + Instruction *InsertPt = User; + if (PHINode *PHI = dyn_cast<PHINode>(InsertPt)) + for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) + if (PHI->getIncomingValue(i) == Op) { + if (InsertPt == User) + InsertPt = PHI->getIncomingBlock(i)->getTerminator(); + else + InsertPt = + DT->findNearestCommonDominator(InsertPt->getParent(), + PHI->getIncomingBlock(i)) + ->getTerminator(); + } + + // Now expand it into actual Instructions and patch it into place. + Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt); + + // Patch the new value into place. + if (Op->hasName()) + NewVal->takeName(Op); + User->replaceUsesOfWith(Op, NewVal); + UI->setOperandValToReplace(NewVal); + DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n' + << " into = " << *NewVal << "\n"); + ++NumRemoved; + Changed = true; + + // The old value may be dead now. + DeadInsts.push_back(Op); } // Clear the rewriter cache, because values that are in the rewriter's cache diff --git a/lib/Transforms/Scalar/JumpThreading.cpp b/lib/Transforms/Scalar/JumpThreading.cpp index 9531311..8f21aac 100644 --- a/lib/Transforms/Scalar/JumpThreading.cpp +++ b/lib/Transforms/Scalar/JumpThreading.cpp @@ -336,13 +336,18 @@ ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ else InterestingVal = ConstantInt::getFalse(I->getContext()); - // Scan for the sentinel. + // Scan for the sentinel. If we find an undef, force it to the + // interesting value: x|undef -> true and x&undef -> false. for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) - if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) + if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { Result.push_back(LHSVals[i]); + Result.back().first = InterestingVal; + } for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) - if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) + if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { Result.push_back(RHSVals[i]); + Result.back().first = InterestingVal; + } return !Result.empty(); } @@ -400,7 +405,7 @@ ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ // If comparing a live-in value against a constant, see if we know the // live-in value on any predecessors. if (LVI && isa<Constant>(Cmp->getOperand(1)) && - Cmp->getType()->isInteger() && // Not vector compare. + Cmp->getType()->isIntegerTy() && // Not vector compare. (!isa<Instruction>(Cmp->getOperand(0)) || cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); @@ -451,6 +456,12 @@ static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { /// ProcessBlock - If there are any predecessors whose control can be threaded /// through to a successor, transform them now. bool JumpThreading::ProcessBlock(BasicBlock *BB) { + // If the block is trivially dead, just return and let the caller nuke it. + // This simplifies other transformations. + if (pred_begin(BB) == pred_end(BB) && + BB != &BB->getParent()->getEntryBlock()) + return false; + // If this block has a single predecessor, and if that pred has a single // successor, merge the blocks. This encourages recursive jump threading // because now the condition in this block can be threaded through @@ -1117,6 +1128,11 @@ bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { isa<ConstantInt>(BO->getOperand(1))) return false; + // If the first instruction in BB isn't a phi, we won't be able to infer + // anything special about any particular predecessor. + if (!isa<PHINode>(BB->front())) + return false; + // If we have a xor as the branch input to this block, and we know that the // LHS or RHS of the xor in any predecessor is true/false, then we can clone // the condition into the predecessor and fix that value to true, saving some @@ -1174,6 +1190,26 @@ bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { BlocksToFoldInto.push_back(XorOpValues[i].second); } + // If we inferred a value for all of the predecessors, then duplication won't + // help us. However, we can just replace the LHS or RHS with the constant. + if (BlocksToFoldInto.size() == + cast<PHINode>(BB->front()).getNumIncomingValues()) { + if (SplitVal == 0) { + // If all preds provide undef, just nuke the xor, because it is undef too. + BO->replaceAllUsesWith(UndefValue::get(BO->getType())); + BO->eraseFromParent(); + } else if (SplitVal->isZero()) { + // If all preds provide 0, replace the xor with the other input. + BO->replaceAllUsesWith(BO->getOperand(isLHS)); + BO->eraseFromParent(); + } else { + // If all preds provide 1, set the computed value to 1. + BO->setOperand(!isLHS, SplitVal); + } + + return true; + } + // Try to duplicate BB into PredBB. return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); } @@ -1393,9 +1429,9 @@ bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, // Unless PredBB ends with an unconditional branch, split the edge so that we // can just clone the bits from BB into the end of the new PredBB. - BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); + BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); - if (!OldPredBranch->isUnconditional()) { + if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { PredBB = SplitEdge(PredBB, BB, this); OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); } diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp index fa820ed..240b298 100644 --- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp +++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp @@ -17,6 +17,40 @@ // available on the target, and it performs a variety of other optimizations // related to loop induction variables. // +// Terminology note: this code has a lot of handling for "post-increment" or +// "post-inc" users. This is not talking about post-increment addressing modes; +// it is instead talking about code like this: +// +// %i = phi [ 0, %entry ], [ %i.next, %latch ] +// ... +// %i.next = add %i, 1 +// %c = icmp eq %i.next, %n +// +// The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however +// it's useful to think about these as the same register, with some uses using +// the value of the register before the add and some using // it after. In this +// example, the icmp is a post-increment user, since it uses %i.next, which is +// the value of the induction variable after the increment. The other common +// case of post-increment users is users outside the loop. +// +// TODO: More sophistication in the way Formulae are generated and filtered. +// +// TODO: Handle multiple loops at a time. +// +// TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr +// instead of a GlobalValue? +// +// TODO: When truncation is free, truncate ICmp users' operands to make it a +// smaller encoding (on x86 at least). +// +// TODO: When a negated register is used by an add (such as in a list of +// multiple base registers, or as the increment expression in an addrec), +// we may not actually need both reg and (-1 * reg) in registers; the +// negation can be implemented by using a sub instead of an add. The +// lack of support for taking this into consideration when making +// register pressure decisions is partly worked around by the "Special" +// use kind. +// //===----------------------------------------------------------------------===// #define DEBUG_TYPE "loop-reduce" @@ -26,208 +60,401 @@ #include "llvm/IntrinsicInst.h" #include "llvm/DerivedTypes.h" #include "llvm/Analysis/IVUsers.h" +#include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/LoopPass.h" #include "llvm/Analysis/ScalarEvolutionExpander.h" -#include "llvm/Transforms/Utils/AddrModeMatcher.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/Local.h" -#include "llvm/ADT/Statistic.h" +#include "llvm/ADT/SmallBitVector.h" +#include "llvm/ADT/SetVector.h" +#include "llvm/ADT/DenseSet.h" #include "llvm/Support/Debug.h" -#include "llvm/Support/CommandLine.h" #include "llvm/Support/ValueHandle.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Target/TargetLowering.h" #include <algorithm> using namespace llvm; -STATISTIC(NumReduced , "Number of IV uses strength reduced"); -STATISTIC(NumInserted, "Number of PHIs inserted"); -STATISTIC(NumVariable, "Number of PHIs with variable strides"); -STATISTIC(NumEliminated, "Number of strides eliminated"); -STATISTIC(NumShadow, "Number of Shadow IVs optimized"); -STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses"); -STATISTIC(NumLoopCond, "Number of loop terminating conds optimized"); -STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero"); +namespace { + +/// RegSortData - This class holds data which is used to order reuse candidates. +class RegSortData { +public: + /// UsedByIndices - This represents the set of LSRUse indices which reference + /// a particular register. + SmallBitVector UsedByIndices; + + RegSortData() {} + + void print(raw_ostream &OS) const; + void dump() const; +}; -static cl::opt<bool> EnableFullLSRMode("enable-full-lsr", - cl::init(false), - cl::Hidden); +} + +void RegSortData::print(raw_ostream &OS) const { + OS << "[NumUses=" << UsedByIndices.count() << ']'; +} + +void RegSortData::dump() const { + print(errs()); errs() << '\n'; +} namespace { - struct BasedUser; +/// RegUseTracker - Map register candidates to information about how they are +/// used. +class RegUseTracker { + typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; - /// IVInfo - This structure keeps track of one IV expression inserted during - /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as - /// well as the PHI node and increment value created for rewrite. - struct IVExpr { - const SCEV *Stride; - const SCEV *Base; - PHINode *PHI; + RegUsesTy RegUses; + SmallVector<const SCEV *, 16> RegSequence; - IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi) - : Stride(stride), Base(base), PHI(phi) {} - }; +public: + void CountRegister(const SCEV *Reg, size_t LUIdx); + + bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; + + const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; + + void clear(); + + typedef SmallVectorImpl<const SCEV *>::iterator iterator; + typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; + iterator begin() { return RegSequence.begin(); } + iterator end() { return RegSequence.end(); } + const_iterator begin() const { return RegSequence.begin(); } + const_iterator end() const { return RegSequence.end(); } +}; + +} - /// IVsOfOneStride - This structure keeps track of all IV expression inserted - /// during StrengthReduceStridedIVUsers for a particular stride of the IV. - struct IVsOfOneStride { - std::vector<IVExpr> IVs; +void +RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { + std::pair<RegUsesTy::iterator, bool> Pair = + RegUses.insert(std::make_pair(Reg, RegSortData())); + RegSortData &RSD = Pair.first->second; + if (Pair.second) + RegSequence.push_back(Reg); + RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); + RSD.UsedByIndices.set(LUIdx); +} + +bool +RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { + if (!RegUses.count(Reg)) return false; + const SmallBitVector &UsedByIndices = + RegUses.find(Reg)->second.UsedByIndices; + int i = UsedByIndices.find_first(); + if (i == -1) return false; + if ((size_t)i != LUIdx) return true; + return UsedByIndices.find_next(i) != -1; +} + +const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { + RegUsesTy::const_iterator I = RegUses.find(Reg); + assert(I != RegUses.end() && "Unknown register!"); + return I->second.UsedByIndices; +} + +void RegUseTracker::clear() { + RegUses.clear(); + RegSequence.clear(); +} + +namespace { + +/// Formula - This class holds information that describes a formula for +/// computing satisfying a use. It may include broken-out immediates and scaled +/// registers. +struct Formula { + /// AM - This is used to represent complex addressing, as well as other kinds + /// of interesting uses. + TargetLowering::AddrMode AM; + + /// BaseRegs - The list of "base" registers for this use. When this is + /// non-empty, AM.HasBaseReg should be set to true. + SmallVector<const SCEV *, 2> BaseRegs; - void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) { - IVs.push_back(IVExpr(Stride, Base, PHI)); + /// ScaledReg - The 'scaled' register for this use. This should be non-null + /// when AM.Scale is not zero. + const SCEV *ScaledReg; + + Formula() : ScaledReg(0) {} + + void InitialMatch(const SCEV *S, Loop *L, + ScalarEvolution &SE, DominatorTree &DT); + + unsigned getNumRegs() const; + const Type *getType() const; + + bool referencesReg(const SCEV *S) const; + bool hasRegsUsedByUsesOtherThan(size_t LUIdx, + const RegUseTracker &RegUses) const; + + void print(raw_ostream &OS) const; + void dump() const; +}; + +} + +/// DoInitialMatch - Recurrsion helper for InitialMatch. +static void DoInitialMatch(const SCEV *S, Loop *L, + SmallVectorImpl<const SCEV *> &Good, + SmallVectorImpl<const SCEV *> &Bad, + ScalarEvolution &SE, DominatorTree &DT) { + // Collect expressions which properly dominate the loop header. + if (S->properlyDominates(L->getHeader(), &DT)) { + Good.push_back(S); + return; + } + + // Look at add operands. + if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { + for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); + I != E; ++I) + DoInitialMatch(*I, L, Good, Bad, SE, DT); + return; + } + + // Look at addrec operands. + if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) + if (!AR->getStart()->isZero()) { + DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT); + DoInitialMatch(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()), + AR->getStepRecurrence(SE), + AR->getLoop()), + L, Good, Bad, SE, DT); + return; } - }; - class LoopStrengthReduce : public LoopPass { - IVUsers *IU; - ScalarEvolution *SE; - bool Changed; - - /// IVsByStride - Keep track of all IVs that have been inserted for a - /// particular stride. - std::map<const SCEV *, IVsOfOneStride> IVsByStride; - - /// DeadInsts - Keep track of instructions we may have made dead, so that - /// we can remove them after we are done working. - SmallVector<WeakVH, 16> DeadInsts; - - /// TLI - Keep a pointer of a TargetLowering to consult for determining - /// transformation profitability. - const TargetLowering *TLI; - - public: - static char ID; // Pass ID, replacement for typeid - explicit LoopStrengthReduce(const TargetLowering *tli = NULL) : - LoopPass(&ID), TLI(tli) {} - - bool runOnLoop(Loop *L, LPPassManager &LPM); - - virtual void getAnalysisUsage(AnalysisUsage &AU) const { - // We split critical edges, so we change the CFG. However, we do update - // many analyses if they are around. - AU.addPreservedID(LoopSimplifyID); - AU.addPreserved("loops"); - AU.addPreserved("domfrontier"); - AU.addPreserved("domtree"); - - AU.addRequiredID(LoopSimplifyID); - AU.addRequired<ScalarEvolution>(); - AU.addPreserved<ScalarEvolution>(); - AU.addRequired<IVUsers>(); - AU.addPreserved<IVUsers>(); + // Handle a multiplication by -1 (negation) if it didn't fold. + if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) + if (Mul->getOperand(0)->isAllOnesValue()) { + SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); + const SCEV *NewMul = SE.getMulExpr(Ops); + + SmallVector<const SCEV *, 4> MyGood; + SmallVector<const SCEV *, 4> MyBad; + DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT); + const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( + SE.getEffectiveSCEVType(NewMul->getType()))); + for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), + E = MyGood.end(); I != E; ++I) + Good.push_back(SE.getMulExpr(NegOne, *I)); + for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), + E = MyBad.end(); I != E; ++I) + Bad.push_back(SE.getMulExpr(NegOne, *I)); + return; } - private: - void OptimizeIndvars(Loop *L); - - /// OptimizeLoopTermCond - Change loop terminating condition to use the - /// postinc iv when possible. - void OptimizeLoopTermCond(Loop *L); - - /// OptimizeShadowIV - If IV is used in a int-to-float cast - /// inside the loop then try to eliminate the cast opeation. - void OptimizeShadowIV(Loop *L); - - /// OptimizeMax - Rewrite the loop's terminating condition - /// if it uses a max computation. - ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond, - IVStrideUse* &CondUse); - - /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for - /// deciding when to exit the loop is used only for that purpose, try to - /// rearrange things so it counts down to a test against zero. - bool OptimizeLoopCountIV(Loop *L); - bool OptimizeLoopCountIVOfStride(const SCEV* &Stride, - IVStrideUse* &CondUse, Loop *L); - - /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a - /// single stride of IV. All of the users may have different starting - /// values, and this may not be the only stride. - void StrengthReduceIVUsersOfStride(const SCEV *Stride, - IVUsersOfOneStride &Uses, - Loop *L); - void StrengthReduceIVUsers(Loop *L); - - ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond, - IVStrideUse* &CondUse, - const SCEV* &CondStride, - bool PostPass = false); - - bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse, - const SCEV* &CondStride); - bool RequiresTypeConversion(const Type *Ty, const Type *NewTy); - const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *, - IVExpr&, const Type*, - const std::vector<BasedUser>& UsersToProcess); - bool ValidScale(bool, int64_t, - const std::vector<BasedUser>& UsersToProcess); - bool ValidOffset(bool, int64_t, int64_t, - const std::vector<BasedUser>& UsersToProcess); - const SCEV *CollectIVUsers(const SCEV *Stride, - IVUsersOfOneStride &Uses, - Loop *L, - bool &AllUsesAreAddresses, - bool &AllUsesAreOutsideLoop, - std::vector<BasedUser> &UsersToProcess); - bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc); - bool ShouldUseFullStrengthReductionMode( - const std::vector<BasedUser> &UsersToProcess, - const Loop *L, - bool AllUsesAreAddresses, - const SCEV *Stride); - void PrepareToStrengthReduceFully( - std::vector<BasedUser> &UsersToProcess, - const SCEV *Stride, - const SCEV *CommonExprs, - const Loop *L, - SCEVExpander &PreheaderRewriter); - void PrepareToStrengthReduceFromSmallerStride( - std::vector<BasedUser> &UsersToProcess, - Value *CommonBaseV, - const IVExpr &ReuseIV, - Instruction *PreInsertPt); - void PrepareToStrengthReduceWithNewPhi( - std::vector<BasedUser> &UsersToProcess, - const SCEV *Stride, - const SCEV *CommonExprs, - Value *CommonBaseV, - Instruction *IVIncInsertPt, - const Loop *L, - SCEVExpander &PreheaderRewriter); - - void DeleteTriviallyDeadInstructions(); - }; + // Ok, we can't do anything interesting. Just stuff the whole thing into a + // register and hope for the best. + Bad.push_back(S); } -char LoopStrengthReduce::ID = 0; -static RegisterPass<LoopStrengthReduce> -X("loop-reduce", "Loop Strength Reduction"); +/// InitialMatch - Incorporate loop-variant parts of S into this Formula, +/// attempting to keep all loop-invariant and loop-computable values in a +/// single base register. +void Formula::InitialMatch(const SCEV *S, Loop *L, + ScalarEvolution &SE, DominatorTree &DT) { + SmallVector<const SCEV *, 4> Good; + SmallVector<const SCEV *, 4> Bad; + DoInitialMatch(S, L, Good, Bad, SE, DT); + if (!Good.empty()) { + BaseRegs.push_back(SE.getAddExpr(Good)); + AM.HasBaseReg = true; + } + if (!Bad.empty()) { + BaseRegs.push_back(SE.getAddExpr(Bad)); + AM.HasBaseReg = true; + } +} -Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { - return new LoopStrengthReduce(TLI); +/// getNumRegs - Return the total number of register operands used by this +/// formula. This does not include register uses implied by non-constant +/// addrec strides. +unsigned Formula::getNumRegs() const { + return !!ScaledReg + BaseRegs.size(); } -/// DeleteTriviallyDeadInstructions - If any of the instructions is the -/// specified set are trivially dead, delete them and see if this makes any of -/// their operands subsequently dead. -void LoopStrengthReduce::DeleteTriviallyDeadInstructions() { - while (!DeadInsts.empty()) { - Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()); +/// getType - Return the type of this formula, if it has one, or null +/// otherwise. This type is meaningless except for the bit size. +const Type *Formula::getType() const { + return !BaseRegs.empty() ? BaseRegs.front()->getType() : + ScaledReg ? ScaledReg->getType() : + AM.BaseGV ? AM.BaseGV->getType() : + 0; +} - if (I == 0 || !isInstructionTriviallyDead(I)) - continue; +/// referencesReg - Test if this formula references the given register. +bool Formula::referencesReg(const SCEV *S) const { + return S == ScaledReg || + std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); +} - for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) - if (Instruction *U = dyn_cast<Instruction>(*OI)) { - *OI = 0; - if (U->use_empty()) - DeadInsts.push_back(U); +/// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers +/// which are used by uses other than the use with the given index. +bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, + const RegUseTracker &RegUses) const { + if (ScaledReg) + if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) + return true; + for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), + E = BaseRegs.end(); I != E; ++I) + if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) + return true; + return false; +} + +void Formula::print(raw_ostream &OS) const { + bool First = true; + if (AM.BaseGV) { + if (!First) OS << " + "; else First = false; + WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false); + } + if (AM.BaseOffs != 0) { + if (!First) OS << " + "; else First = false; + OS << AM.BaseOffs; + } + for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), + E = BaseRegs.end(); I != E; ++I) { + if (!First) OS << " + "; else First = false; + OS << "reg(" << **I << ')'; + } + if (AM.Scale != 0) { + if (!First) OS << " + "; else First = false; + OS << AM.Scale << "*reg("; + if (ScaledReg) + OS << *ScaledReg; + else + OS << "<unknown>"; + OS << ')'; + } +} + +void Formula::dump() const { + print(errs()); errs() << '\n'; +} + +/// getSDiv - Return an expression for LHS /s RHS, if it can be determined, +/// or null otherwise. If IgnoreSignificantBits is true, expressions like +/// (X * Y) /s Y are simplified to Y, ignoring that the multiplication may +/// overflow, which is useful when the result will be used in a context where +/// the most significant bits are ignored. +static const SCEV *getSDiv(const SCEV *LHS, const SCEV *RHS, + ScalarEvolution &SE, + bool IgnoreSignificantBits = false) { + // Handle the trivial case, which works for any SCEV type. + if (LHS == RHS) + return SE.getIntegerSCEV(1, LHS->getType()); + + // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some + // folding. + if (RHS->isAllOnesValue()) + return SE.getMulExpr(LHS, RHS); + + // Check for a division of a constant by a constant. + if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { + const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); + if (!RC) + return 0; + if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0) + return 0; + return SE.getConstant(C->getValue()->getValue() + .sdiv(RC->getValue()->getValue())); + } + + // Distribute the sdiv over addrec operands. + if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { + const SCEV *Start = getSDiv(AR->getStart(), RHS, SE, + IgnoreSignificantBits); + if (!Start) return 0; + const SCEV *Step = getSDiv(AR->getStepRecurrence(SE), RHS, SE, + IgnoreSignificantBits); + if (!Step) return 0; + return SE.getAddRecExpr(Start, Step, AR->getLoop()); + } + + // Distribute the sdiv over add operands. + if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { + SmallVector<const SCEV *, 8> Ops; + for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); + I != E; ++I) { + const SCEV *Op = getSDiv(*I, RHS, SE, + IgnoreSignificantBits); + if (!Op) return 0; + Ops.push_back(Op); + } + return SE.getAddExpr(Ops); + } + + // Check for a multiply operand that we can pull RHS out of. + if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) + if (IgnoreSignificantBits || Mul->hasNoSignedWrap()) { + SmallVector<const SCEV *, 4> Ops; + bool Found = false; + for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); + I != E; ++I) { + if (!Found) + if (const SCEV *Q = getSDiv(*I, RHS, SE, IgnoreSignificantBits)) { + Ops.push_back(Q); + Found = true; + continue; + } + Ops.push_back(*I); } + return Found ? SE.getMulExpr(Ops) : 0; + } - I->eraseFromParent(); - Changed = true; + // Otherwise we don't know. + return 0; +} + +/// ExtractImmediate - If S involves the addition of a constant integer value, +/// return that integer value, and mutate S to point to a new SCEV with that +/// value excluded. +static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { + if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { + if (C->getValue()->getValue().getMinSignedBits() <= 64) { + S = SE.getIntegerSCEV(0, C->getType()); + return C->getValue()->getSExtValue(); + } + } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { + SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); + int64_t Result = ExtractImmediate(NewOps.front(), SE); + S = SE.getAddExpr(NewOps); + return Result; + } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { + SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); + int64_t Result = ExtractImmediate(NewOps.front(), SE); + S = SE.getAddRecExpr(NewOps, AR->getLoop()); + return Result; + } + return 0; +} + +/// ExtractSymbol - If S involves the addition of a GlobalValue address, +/// return that symbol, and mutate S to point to a new SCEV with that +/// value excluded. +static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { + if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { + if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { + S = SE.getIntegerSCEV(0, GV->getType()); + return GV; + } + } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { + SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); + GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); + S = SE.getAddExpr(NewOps); + return Result; + } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { + SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); + GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); + S = SE.getAddRecExpr(NewOps, AR->getLoop()); + return Result; } + return 0; } /// isAddressUse - Returns true if the specified instruction is using the @@ -276,1775 +503,832 @@ static const Type *getAccessType(const Instruction *Inst) { break; } } - return AccessTy; -} - -namespace { - /// BasedUser - For a particular base value, keep information about how we've - /// partitioned the expression so far. - struct BasedUser { - /// Base - The Base value for the PHI node that needs to be inserted for - /// this use. As the use is processed, information gets moved from this - /// field to the Imm field (below). BasedUser values are sorted by this - /// field. - const SCEV *Base; - - /// Inst - The instruction using the induction variable. - Instruction *Inst; - - /// OperandValToReplace - The operand value of Inst to replace with the - /// EmittedBase. - Value *OperandValToReplace; - - /// Imm - The immediate value that should be added to the base immediately - /// before Inst, because it will be folded into the imm field of the - /// instruction. This is also sometimes used for loop-variant values that - /// must be added inside the loop. - const SCEV *Imm; - - /// Phi - The induction variable that performs the striding that - /// should be used for this user. - PHINode *Phi; - - // isUseOfPostIncrementedValue - True if this should use the - // post-incremented version of this IV, not the preincremented version. - // This can only be set in special cases, such as the terminating setcc - // instruction for a loop and uses outside the loop that are dominated by - // the loop. - bool isUseOfPostIncrementedValue; - - BasedUser(IVStrideUse &IVSU, ScalarEvolution *se) - : Base(IVSU.getOffset()), Inst(IVSU.getUser()), - OperandValToReplace(IVSU.getOperandValToReplace()), - Imm(se->getIntegerSCEV(0, Base->getType())), - isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {} - - // Once we rewrite the code to insert the new IVs we want, update the - // operands of Inst to use the new expression 'NewBase', with 'Imm' added - // to it. - void RewriteInstructionToUseNewBase(const SCEV *NewBase, - Instruction *InsertPt, - SCEVExpander &Rewriter, Loop *L, Pass *P, - SmallVectorImpl<WeakVH> &DeadInsts, - ScalarEvolution *SE); - - Value *InsertCodeForBaseAtPosition(const SCEV *NewBase, - const Type *Ty, - SCEVExpander &Rewriter, - Instruction *IP, - ScalarEvolution *SE); - void dump() const; - }; -} - -void BasedUser::dump() const { - dbgs() << " Base=" << *Base; - dbgs() << " Imm=" << *Imm; - dbgs() << " Inst: " << *Inst; -} - -Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *NewBase, - const Type *Ty, - SCEVExpander &Rewriter, - Instruction *IP, - ScalarEvolution *SE) { - Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP); - // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to - // re-analyze it. - const SCEV *NewValSCEV = SE->getUnknown(Base); + // All pointers have the same requirements, so canonicalize them to an + // arbitrary pointer type to minimize variation. + if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy)) + AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), + PTy->getAddressSpace()); - // Always emit the immediate into the same block as the user. - NewValSCEV = SE->getAddExpr(NewValSCEV, Imm); - - return Rewriter.expandCodeFor(NewValSCEV, Ty, IP); + return AccessTy; } +/// DeleteTriviallyDeadInstructions - If any of the instructions is the +/// specified set are trivially dead, delete them and see if this makes any of +/// their operands subsequently dead. +static bool +DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { + bool Changed = false; -// Once we rewrite the code to insert the new IVs we want, update the -// operands of Inst to use the new expression 'NewBase', with 'Imm' added -// to it. NewBasePt is the last instruction which contributes to the -// value of NewBase in the case that it's a diffferent instruction from -// the PHI that NewBase is computed from, or null otherwise. -// -void BasedUser::RewriteInstructionToUseNewBase(const SCEV *NewBase, - Instruction *NewBasePt, - SCEVExpander &Rewriter, Loop *L, Pass *P, - SmallVectorImpl<WeakVH> &DeadInsts, - ScalarEvolution *SE) { - if (!isa<PHINode>(Inst)) { - // By default, insert code at the user instruction. - BasicBlock::iterator InsertPt = Inst; - - // However, if the Operand is itself an instruction, the (potentially - // complex) inserted code may be shared by many users. Because of this, we - // want to emit code for the computation of the operand right before its old - // computation. This is usually safe, because we obviously used to use the - // computation when it was computed in its current block. However, in some - // cases (e.g. use of a post-incremented induction variable) the NewBase - // value will be pinned to live somewhere after the original computation. - // In this case, we have to back off. - // - // If this is a use outside the loop (which means after, since it is based - // on a loop indvar) we use the post-incremented value, so that we don't - // artificially make the preinc value live out the bottom of the loop. - if (!isUseOfPostIncrementedValue && L->contains(Inst)) { - if (NewBasePt && isa<PHINode>(OperandValToReplace)) { - InsertPt = NewBasePt; - ++InsertPt; - } else if (Instruction *OpInst - = dyn_cast<Instruction>(OperandValToReplace)) { - InsertPt = OpInst; - while (isa<PHINode>(InsertPt)) ++InsertPt; - } - } - Value *NewVal = InsertCodeForBaseAtPosition(NewBase, - OperandValToReplace->getType(), - Rewriter, InsertPt, SE); - // Replace the use of the operand Value with the new Phi we just created. - Inst->replaceUsesOfWith(OperandValToReplace, NewVal); - - DEBUG(dbgs() << " Replacing with "); - DEBUG(WriteAsOperand(dbgs(), NewVal, /*PrintType=*/false)); - DEBUG(dbgs() << ", which has value " << *NewBase << " plus IMM " - << *Imm << "\n"); - return; - } + while (!DeadInsts.empty()) { + Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()); - // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm - // expression into each operand block that uses it. Note that PHI nodes can - // have multiple entries for the same predecessor. We use a map to make sure - // that a PHI node only has a single Value* for each predecessor (which also - // prevents us from inserting duplicate code in some blocks). - DenseMap<BasicBlock*, Value*> InsertedCode; - PHINode *PN = cast<PHINode>(Inst); - for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { - if (PN->getIncomingValue(i) == OperandValToReplace) { - // If the original expression is outside the loop, put the replacement - // code in the same place as the original expression, - // which need not be an immediate predecessor of this PHI. This way we - // need only one copy of it even if it is referenced multiple times in - // the PHI. We don't do this when the original expression is inside the - // loop because multiple copies sometimes do useful sinking of code in - // that case(?). - Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace); - BasicBlock *PHIPred = PN->getIncomingBlock(i); - if (L->contains(OldLoc)) { - // If this is a critical edge, split the edge so that we do not insert - // the code on all predecessor/successor paths. We do this unless this - // is the canonical backedge for this loop, as this can make some - // inserted code be in an illegal position. - if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 && - !isa<IndirectBrInst>(PHIPred->getTerminator()) && - (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) { - - // First step, split the critical edge. - BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(), - P, false); - - // Next step: move the basic block. In particular, if the PHI node - // is outside of the loop, and PredTI is in the loop, we want to - // move the block to be immediately before the PHI block, not - // immediately after PredTI. - if (L->contains(PHIPred) && !L->contains(PN)) - NewBB->moveBefore(PN->getParent()); + if (I == 0 || !isInstructionTriviallyDead(I)) + continue; - // Splitting the edge can reduce the number of PHI entries we have. - e = PN->getNumIncomingValues(); - PHIPred = NewBB; - i = PN->getBasicBlockIndex(PHIPred); - } - } - Value *&Code = InsertedCode[PHIPred]; - if (!Code) { - // Insert the code into the end of the predecessor block. - Instruction *InsertPt = (L->contains(OldLoc)) ? - PHIPred->getTerminator() : - OldLoc->getParent()->getTerminator(); - Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(), - Rewriter, InsertPt, SE); - - DEBUG(dbgs() << " Changing PHI use to "); - DEBUG(WriteAsOperand(dbgs(), Code, /*PrintType=*/false)); - DEBUG(dbgs() << ", which has value " << *NewBase << " plus IMM " - << *Imm << "\n"); + for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) + if (Instruction *U = dyn_cast<Instruction>(*OI)) { + *OI = 0; + if (U->use_empty()) + DeadInsts.push_back(U); } - // Replace the use of the operand Value with the new Phi we just created. - PN->setIncomingValue(i, Code); - Rewriter.clear(); - } + I->eraseFromParent(); + Changed = true; } - // PHI node might have become a constant value after SplitCriticalEdge. - DeadInsts.push_back(Inst); + return Changed; } +namespace { -/// fitsInAddressMode - Return true if V can be subsumed within an addressing -/// mode, and does not need to be put in a register first. -static bool fitsInAddressMode(const SCEV *V, const Type *AccessTy, - const TargetLowering *TLI, bool HasBaseReg) { - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) { - int64_t VC = SC->getValue()->getSExtValue(); - if (TLI) { - TargetLowering::AddrMode AM; - AM.BaseOffs = VC; - AM.HasBaseReg = HasBaseReg; - return TLI->isLegalAddressingMode(AM, AccessTy); - } else { - // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field. - return (VC > -(1 << 16) && VC < (1 << 16)-1); - } - } - - if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) - if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) { - if (TLI) { - TargetLowering::AddrMode AM; - AM.BaseGV = GV; - AM.HasBaseReg = HasBaseReg; - return TLI->isLegalAddressingMode(AM, AccessTy); - } else { - // Default: assume global addresses are not legal. - } - } +/// Cost - This class is used to measure and compare candidate formulae. +class Cost { + /// TODO: Some of these could be merged. Also, a lexical ordering + /// isn't always optimal. + unsigned NumRegs; + unsigned AddRecCost; + unsigned NumIVMuls; + unsigned NumBaseAdds; + unsigned ImmCost; + unsigned SetupCost; + +public: + Cost() + : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), + SetupCost(0) {} + + unsigned getNumRegs() const { return NumRegs; } + + bool operator<(const Cost &Other) const; + + void Loose(); + + void RateFormula(const Formula &F, + SmallPtrSet<const SCEV *, 16> &Regs, + const DenseSet<const SCEV *> &VisitedRegs, + const Loop *L, + const SmallVectorImpl<int64_t> &Offsets, + ScalarEvolution &SE, DominatorTree &DT); + + void print(raw_ostream &OS) const; + void dump() const; + +private: + void RateRegister(const SCEV *Reg, + SmallPtrSet<const SCEV *, 16> &Regs, + const Loop *L, + ScalarEvolution &SE, DominatorTree &DT); + void RatePrimaryRegister(const SCEV *Reg, + SmallPtrSet<const SCEV *, 16> &Regs, + const Loop *L, + ScalarEvolution &SE, DominatorTree &DT); +}; - return false; } -/// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are -/// loop varying to the Imm operand. -static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm, - Loop *L, ScalarEvolution *SE) { - if (Val->isLoopInvariant(L)) return; // Nothing to do. - - if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) { - SmallVector<const SCEV *, 4> NewOps; - NewOps.reserve(SAE->getNumOperands()); +/// RateRegister - Tally up interesting quantities from the given register. +void Cost::RateRegister(const SCEV *Reg, + SmallPtrSet<const SCEV *, 16> &Regs, + const Loop *L, + ScalarEvolution &SE, DominatorTree &DT) { + if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { + if (AR->getLoop() == L) + AddRecCost += 1; /// TODO: This should be a function of the stride. + + // If this is an addrec for a loop that's already been visited by LSR, + // don't second-guess its addrec phi nodes. LSR isn't currently smart + // enough to reason about more than one loop at a time. Consider these + // registers free and leave them alone. + else if (L->contains(AR->getLoop()) || + (!AR->getLoop()->contains(L) && + DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) { + for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); + PHINode *PN = dyn_cast<PHINode>(I); ++I) + if (SE.isSCEVable(PN->getType()) && + (SE.getEffectiveSCEVType(PN->getType()) == + SE.getEffectiveSCEVType(AR->getType())) && + SE.getSCEV(PN) == AR) + return; - for (unsigned i = 0; i != SAE->getNumOperands(); ++i) - if (!SAE->getOperand(i)->isLoopInvariant(L)) { - // If this is a loop-variant expression, it must stay in the immediate - // field of the expression. - Imm = SE->getAddExpr(Imm, SAE->getOperand(i)); - } else { - NewOps.push_back(SAE->getOperand(i)); - } + // If this isn't one of the addrecs that the loop already has, it + // would require a costly new phi and add. TODO: This isn't + // precisely modeled right now. + ++NumBaseAdds; + if (!Regs.count(AR->getStart())) + RateRegister(AR->getStart(), Regs, L, SE, DT); + } - if (NewOps.empty()) - Val = SE->getIntegerSCEV(0, Val->getType()); - else - Val = SE->getAddExpr(NewOps); - } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) { - // Try to pull immediates out of the start value of nested addrec's. - const SCEV *Start = SARE->getStart(); - MoveLoopVariantsToImmediateField(Start, Imm, L, SE); - - SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end()); - Ops[0] = Start; - Val = SE->getAddRecExpr(Ops, SARE->getLoop()); - } else { - // Otherwise, all of Val is variant, move the whole thing over. - Imm = SE->getAddExpr(Imm, Val); - Val = SE->getIntegerSCEV(0, Val->getType()); + // Add the step value register, if it needs one. + // TODO: The non-affine case isn't precisely modeled here. + if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) + if (!Regs.count(AR->getStart())) + RateRegister(AR->getOperand(1), Regs, L, SE, DT); } + ++NumRegs; + + // Rough heuristic; favor registers which don't require extra setup + // instructions in the preheader. + if (!isa<SCEVUnknown>(Reg) && + !isa<SCEVConstant>(Reg) && + !(isa<SCEVAddRecExpr>(Reg) && + (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || + isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) + ++SetupCost; } +/// RatePrimaryRegister - Record this register in the set. If we haven't seen it +/// before, rate it. +void Cost::RatePrimaryRegister(const SCEV *Reg, + SmallPtrSet<const SCEV *, 16> &Regs, + const Loop *L, + ScalarEvolution &SE, DominatorTree &DT) { + if (Regs.insert(Reg)) + RateRegister(Reg, Regs, L, SE, DT); +} -/// MoveImmediateValues - Look at Val, and pull out any additions of constants -/// that can fit into the immediate field of instructions in the target. -/// Accumulate these immediate values into the Imm value. -static void MoveImmediateValues(const TargetLowering *TLI, - const Type *AccessTy, - const SCEV *&Val, const SCEV *&Imm, - bool isAddress, Loop *L, - ScalarEvolution *SE) { - if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) { - SmallVector<const SCEV *, 4> NewOps; - NewOps.reserve(SAE->getNumOperands()); - - for (unsigned i = 0; i != SAE->getNumOperands(); ++i) { - const SCEV *NewOp = SAE->getOperand(i); - MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE); - - if (!NewOp->isLoopInvariant(L)) { - // If this is a loop-variant expression, it must stay in the immediate - // field of the expression. - Imm = SE->getAddExpr(Imm, NewOp); - } else { - NewOps.push_back(NewOp); - } - } - - if (NewOps.empty()) - Val = SE->getIntegerSCEV(0, Val->getType()); - else - Val = SE->getAddExpr(NewOps); - return; - } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) { - // Try to pull immediates out of the start value of nested addrec's. - const SCEV *Start = SARE->getStart(); - MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE); - - if (Start != SARE->getStart()) { - SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end()); - Ops[0] = Start; - Val = SE->getAddRecExpr(Ops, SARE->getLoop()); +void Cost::RateFormula(const Formula &F, + SmallPtrSet<const SCEV *, 16> &Regs, + const DenseSet<const SCEV *> &VisitedRegs, + const Loop *L, + const SmallVectorImpl<int64_t> &Offsets, + ScalarEvolution &SE, DominatorTree &DT) { + // Tally up the registers. + if (const SCEV *ScaledReg = F.ScaledReg) { + if (VisitedRegs.count(ScaledReg)) { + Loose(); + return; } - return; - } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) { - // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field. - if (isAddress && - fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) && - SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) { - - const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType()); - const SCEV *NewOp = SME->getOperand(1); - MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE); - - // If we extracted something out of the subexpressions, see if we can - // simplify this! - if (NewOp != SME->getOperand(1)) { - // Scale SubImm up by "8". If the result is a target constant, we are - // good. - SubImm = SE->getMulExpr(SubImm, SME->getOperand(0)); - if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) { - // Accumulate the immediate. - Imm = SE->getAddExpr(Imm, SubImm); - - // Update what is left of 'Val'. - Val = SE->getMulExpr(SME->getOperand(0), NewOp); - return; - } - } + RatePrimaryRegister(ScaledReg, Regs, L, SE, DT); + } + for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), + E = F.BaseRegs.end(); I != E; ++I) { + const SCEV *BaseReg = *I; + if (VisitedRegs.count(BaseReg)) { + Loose(); + return; } + RatePrimaryRegister(BaseReg, Regs, L, SE, DT); + + NumIVMuls += isa<SCEVMulExpr>(BaseReg) && + BaseReg->hasComputableLoopEvolution(L); } - // Loop-variant expressions must stay in the immediate field of the - // expression. - if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) || - !Val->isLoopInvariant(L)) { - Imm = SE->getAddExpr(Imm, Val); - Val = SE->getIntegerSCEV(0, Val->getType()); - return; + if (F.BaseRegs.size() > 1) + NumBaseAdds += F.BaseRegs.size() - 1; + + // Tally up the non-zero immediates. + for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), + E = Offsets.end(); I != E; ++I) { + int64_t Offset = (uint64_t)*I + F.AM.BaseOffs; + if (F.AM.BaseGV) + ImmCost += 64; // Handle symbolic values conservatively. + // TODO: This should probably be the pointer size. + else if (Offset != 0) + ImmCost += APInt(64, Offset, true).getMinSignedBits(); } +} - // Otherwise, no immediates to move. +/// Loose - Set this cost to a loosing value. +void Cost::Loose() { + NumRegs = ~0u; + AddRecCost = ~0u; + NumIVMuls = ~0u; + NumBaseAdds = ~0u; + ImmCost = ~0u; + SetupCost = ~0u; } -static void MoveImmediateValues(const TargetLowering *TLI, - Instruction *User, - const SCEV *&Val, const SCEV *&Imm, - bool isAddress, Loop *L, - ScalarEvolution *SE) { - const Type *AccessTy = getAccessType(User); - MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE); +/// operator< - Choose the lower cost. +bool Cost::operator<(const Cost &Other) const { + if (NumRegs != Other.NumRegs) + return NumRegs < Other.NumRegs; + if (AddRecCost != Other.AddRecCost) + return AddRecCost < Other.AddRecCost; + if (NumIVMuls != Other.NumIVMuls) + return NumIVMuls < Other.NumIVMuls; + if (NumBaseAdds != Other.NumBaseAdds) + return NumBaseAdds < Other.NumBaseAdds; + if (ImmCost != Other.ImmCost) + return ImmCost < Other.ImmCost; + if (SetupCost != Other.SetupCost) + return SetupCost < Other.SetupCost; + return false; } -/// SeparateSubExprs - Decompose Expr into all of the subexpressions that are -/// added together. This is used to reassociate common addition subexprs -/// together for maximal sharing when rewriting bases. -static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs, - const SCEV *Expr, - ScalarEvolution *SE) { - if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) { - for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j) - SeparateSubExprs(SubExprs, AE->getOperand(j), SE); - } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) { - const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType()); - if (SARE->getOperand(0) == Zero) { - SubExprs.push_back(Expr); - } else { - // Compute the addrec with zero as its base. - SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end()); - Ops[0] = Zero; // Start with zero base. - SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop())); +void Cost::print(raw_ostream &OS) const { + OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); + if (AddRecCost != 0) + OS << ", with addrec cost " << AddRecCost; + if (NumIVMuls != 0) + OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); + if (NumBaseAdds != 0) + OS << ", plus " << NumBaseAdds << " base add" + << (NumBaseAdds == 1 ? "" : "s"); + if (ImmCost != 0) + OS << ", plus " << ImmCost << " imm cost"; + if (SetupCost != 0) + OS << ", plus " << SetupCost << " setup cost"; +} +void Cost::dump() const { + print(errs()); errs() << '\n'; +} - SeparateSubExprs(SubExprs, SARE->getOperand(0), SE); - } - } else if (!Expr->isZero()) { - // Do not add zero. - SubExprs.push_back(Expr); - } -} - -// This is logically local to the following function, but C++ says we have -// to make it file scope. -struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; }; - -/// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all -/// the Uses, removing any common subexpressions, except that if all such -/// subexpressions can be folded into an addressing mode for all uses inside -/// the loop (this case is referred to as "free" in comments herein) we do -/// not remove anything. This looks for things like (a+b+c) and -/// (a+c+d) and computes the common (a+c) subexpression. The common expression -/// is *removed* from the Bases and returned. -static const SCEV * -RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses, - ScalarEvolution *SE, Loop *L, - const TargetLowering *TLI) { - unsigned NumUses = Uses.size(); - - // Only one use? This is a very common case, so we handle it specially and - // cheaply. - const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType()); - const SCEV *Result = Zero; - const SCEV *FreeResult = Zero; - if (NumUses == 1) { - // If the use is inside the loop, use its base, regardless of what it is: - // it is clearly shared across all the IV's. If the use is outside the loop - // (which means after it) we don't want to factor anything *into* the loop, - // so just use 0 as the base. - if (L->contains(Uses[0].Inst)) - std::swap(Result, Uses[0].Base); - return Result; - } +namespace { - // To find common subexpressions, count how many of Uses use each expression. - // If any subexpressions are used Uses.size() times, they are common. - // Also track whether all uses of each expression can be moved into an - // an addressing mode "for free"; such expressions are left within the loop. - // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; }; - std::map<const SCEV *, SubExprUseData> SubExpressionUseData; - - // UniqueSubExprs - Keep track of all of the subexpressions we see in the - // order we see them. - SmallVector<const SCEV *, 16> UniqueSubExprs; - - SmallVector<const SCEV *, 16> SubExprs; - unsigned NumUsesInsideLoop = 0; - for (unsigned i = 0; i != NumUses; ++i) { - // If the user is outside the loop, just ignore it for base computation. - // Since the user is outside the loop, it must be *after* the loop (if it - // were before, it could not be based on the loop IV). We don't want users - // after the loop to affect base computation of values *inside* the loop, - // because we can always add their offsets to the result IV after the loop - // is done, ensuring we get good code inside the loop. - if (!L->contains(Uses[i].Inst)) - continue; - NumUsesInsideLoop++; +/// LSRFixup - An operand value in an instruction which is to be replaced +/// with some equivalent, possibly strength-reduced, replacement. +struct LSRFixup { + /// UserInst - The instruction which will be updated. + Instruction *UserInst; - // If the base is zero (which is common), return zero now, there are no - // CSEs we can find. - if (Uses[i].Base == Zero) return Zero; + /// OperandValToReplace - The operand of the instruction which will + /// be replaced. The operand may be used more than once; every instance + /// will be replaced. + Value *OperandValToReplace; - // If this use is as an address we may be able to put CSEs in the addressing - // mode rather than hoisting them. - bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace); - // We may need the AccessTy below, but only when isAddrUse, so compute it - // only in that case. - const Type *AccessTy = 0; - if (isAddrUse) - AccessTy = getAccessType(Uses[i].Inst); - - // Split the expression into subexprs. - SeparateSubExprs(SubExprs, Uses[i].Base, SE); - // Add one to SubExpressionUseData.Count for each subexpr present, and - // if the subexpr is not a valid immediate within an addressing mode use, - // set SubExpressionUseData.notAllUsesAreFree. We definitely want to - // hoist these out of the loop (if they are common to all uses). - for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) { - if (++SubExpressionUseData[SubExprs[j]].Count == 1) - UniqueSubExprs.push_back(SubExprs[j]); - if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false)) - SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true; - } - SubExprs.clear(); - } - - // Now that we know how many times each is used, build Result. Iterate over - // UniqueSubexprs so that we have a stable ordering. - for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) { - std::map<const SCEV *, SubExprUseData>::iterator I = - SubExpressionUseData.find(UniqueSubExprs[i]); - assert(I != SubExpressionUseData.end() && "Entry not found?"); - if (I->second.Count == NumUsesInsideLoop) { // Found CSE! - if (I->second.notAllUsesAreFree) - Result = SE->getAddExpr(Result, I->first); - else - FreeResult = SE->getAddExpr(FreeResult, I->first); - } else - // Remove non-cse's from SubExpressionUseData. - SubExpressionUseData.erase(I); - } - - if (FreeResult != Zero) { - // We have some subexpressions that can be subsumed into addressing - // modes in every use inside the loop. However, it's possible that - // there are so many of them that the combined FreeResult cannot - // be subsumed, or that the target cannot handle both a FreeResult - // and a Result in the same instruction (for example because it would - // require too many registers). Check this. - for (unsigned i=0; i<NumUses; ++i) { - if (!L->contains(Uses[i].Inst)) - continue; - // We know this is an addressing mode use; if there are any uses that - // are not, FreeResult would be Zero. - const Type *AccessTy = getAccessType(Uses[i].Inst); - if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) { - // FIXME: could split up FreeResult into pieces here, some hoisted - // and some not. There is no obvious advantage to this. - Result = SE->getAddExpr(Result, FreeResult); - FreeResult = Zero; - break; - } - } - } + /// PostIncLoop - If this user is to use the post-incremented value of an + /// induction variable, this variable is non-null and holds the loop + /// associated with the induction variable. + const Loop *PostIncLoop; - // If we found no CSE's, return now. - if (Result == Zero) return Result; + /// LUIdx - The index of the LSRUse describing the expression which + /// this fixup needs, minus an offset (below). + size_t LUIdx; - // If we still have a FreeResult, remove its subexpressions from - // SubExpressionUseData. This means they will remain in the use Bases. - if (FreeResult != Zero) { - SeparateSubExprs(SubExprs, FreeResult, SE); - for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) { - std::map<const SCEV *, SubExprUseData>::iterator I = - SubExpressionUseData.find(SubExprs[j]); - SubExpressionUseData.erase(I); - } - SubExprs.clear(); - } + /// Offset - A constant offset to be added to the LSRUse expression. + /// This allows multiple fixups to share the same LSRUse with different + /// offsets, for example in an unrolled loop. + int64_t Offset; - // Otherwise, remove all of the CSE's we found from each of the base values. - for (unsigned i = 0; i != NumUses; ++i) { - // Uses outside the loop don't necessarily include the common base, but - // the final IV value coming into those uses does. Instead of trying to - // remove the pieces of the common base, which might not be there, - // subtract off the base to compensate for this. - if (!L->contains(Uses[i].Inst)) { - Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result); - continue; - } + LSRFixup(); - // Split the expression into subexprs. - SeparateSubExprs(SubExprs, Uses[i].Base, SE); + void print(raw_ostream &OS) const; + void dump() const; +}; - // Remove any common subexpressions. - for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) - if (SubExpressionUseData.count(SubExprs[j])) { - SubExprs.erase(SubExprs.begin()+j); - --j; --e; - } +} - // Finally, add the non-shared expressions together. - if (SubExprs.empty()) - Uses[i].Base = Zero; - else - Uses[i].Base = SE->getAddExpr(SubExprs); - SubExprs.clear(); +LSRFixup::LSRFixup() + : UserInst(0), OperandValToReplace(0), PostIncLoop(0), + LUIdx(~size_t(0)), Offset(0) {} + +void LSRFixup::print(raw_ostream &OS) const { + OS << "UserInst="; + // Store is common and interesting enough to be worth special-casing. + if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { + OS << "store "; + WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); + } else if (UserInst->getType()->isVoidTy()) + OS << UserInst->getOpcodeName(); + else + WriteAsOperand(OS, UserInst, /*PrintType=*/false); + + OS << ", OperandValToReplace="; + WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); + + if (PostIncLoop) { + OS << ", PostIncLoop="; + WriteAsOperand(OS, PostIncLoop->getHeader(), /*PrintType=*/false); } - return Result; -} + if (LUIdx != ~size_t(0)) + OS << ", LUIdx=" << LUIdx; -/// ValidScale - Check whether the given Scale is valid for all loads and -/// stores in UsersToProcess. -/// -bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale, - const std::vector<BasedUser>& UsersToProcess) { - if (!TLI) - return true; + if (Offset != 0) + OS << ", Offset=" << Offset; +} - for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) { - // If this is a load or other access, pass the type of the access in. - const Type *AccessTy = - Type::getVoidTy(UsersToProcess[i].Inst->getContext()); - if (isAddressUse(UsersToProcess[i].Inst, - UsersToProcess[i].OperandValToReplace)) - AccessTy = getAccessType(UsersToProcess[i].Inst); - else if (isa<PHINode>(UsersToProcess[i].Inst)) - continue; +void LSRFixup::dump() const { + print(errs()); errs() << '\n'; +} - TargetLowering::AddrMode AM; - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm)) - AM.BaseOffs = SC->getValue()->getSExtValue(); - AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero(); - AM.Scale = Scale; +namespace { - // If load[imm+r*scale] is illegal, bail out. - if (!TLI->isLegalAddressingMode(AM, AccessTy)) - return false; +/// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding +/// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. +struct UniquifierDenseMapInfo { + static SmallVector<const SCEV *, 2> getEmptyKey() { + SmallVector<const SCEV *, 2> V; + V.push_back(reinterpret_cast<const SCEV *>(-1)); + return V; } - return true; -} - -/// ValidOffset - Check whether the given Offset is valid for all loads and -/// stores in UsersToProcess. -/// -bool LoopStrengthReduce::ValidOffset(bool HasBaseReg, - int64_t Offset, - int64_t Scale, - const std::vector<BasedUser>& UsersToProcess) { - if (!TLI) - return true; - for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) { - // If this is a load or other access, pass the type of the access in. - const Type *AccessTy = - Type::getVoidTy(UsersToProcess[i].Inst->getContext()); - if (isAddressUse(UsersToProcess[i].Inst, - UsersToProcess[i].OperandValToReplace)) - AccessTy = getAccessType(UsersToProcess[i].Inst); - else if (isa<PHINode>(UsersToProcess[i].Inst)) - continue; + static SmallVector<const SCEV *, 2> getTombstoneKey() { + SmallVector<const SCEV *, 2> V; + V.push_back(reinterpret_cast<const SCEV *>(-2)); + return V; + } - TargetLowering::AddrMode AM; - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm)) - AM.BaseOffs = SC->getValue()->getSExtValue(); - AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset; - AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero(); - AM.Scale = Scale; + static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) { + unsigned Result = 0; + for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), + E = V.end(); I != E; ++I) + Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); + return Result; + } - // If load[imm+r*scale] is illegal, bail out. - if (!TLI->isLegalAddressingMode(AM, AccessTy)) - return false; + static bool isEqual(const SmallVector<const SCEV *, 2> &LHS, + const SmallVector<const SCEV *, 2> &RHS) { + return LHS == RHS; } - return true; -} +}; + +/// LSRUse - This class holds the state that LSR keeps for each use in +/// IVUsers, as well as uses invented by LSR itself. It includes information +/// about what kinds of things can be folded into the user, information about +/// the user itself, and information about how the use may be satisfied. +/// TODO: Represent multiple users of the same expression in common? +class LSRUse { + DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier; + +public: + /// KindType - An enum for a kind of use, indicating what types of + /// scaled and immediate operands it might support. + enum KindType { + Basic, ///< A normal use, with no folding. + Special, ///< A special case of basic, allowing -1 scales. + Address, ///< An address use; folding according to TargetLowering + ICmpZero ///< An equality icmp with both operands folded into one. + // TODO: Add a generic icmp too? + }; -/// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not -/// a nop. -bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1, - const Type *Ty2) { - if (Ty1 == Ty2) - return false; - Ty1 = SE->getEffectiveSCEVType(Ty1); - Ty2 = SE->getEffectiveSCEVType(Ty2); - if (Ty1 == Ty2) - return false; - if (Ty1->canLosslesslyBitCastTo(Ty2)) - return false; - if (TLI && TLI->isTruncateFree(Ty1, Ty2)) - return false; - return true; -} + KindType Kind; + const Type *AccessTy; -/// CheckForIVReuse - Returns the multiple if the stride is the multiple -/// of a previous stride and it is a legal value for the target addressing -/// mode scale component and optional base reg. This allows the users of -/// this stride to be rewritten as prev iv * factor. It returns 0 if no -/// reuse is possible. Factors can be negative on same targets, e.g. ARM. -/// -/// If all uses are outside the loop, we don't require that all multiplies -/// be folded into the addressing mode, nor even that the factor be constant; -/// a multiply (executed once) outside the loop is better than another IV -/// within. Well, usually. -const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg, - bool AllUsesAreAddresses, - bool AllUsesAreOutsideLoop, - const SCEV *Stride, - IVExpr &IV, const Type *Ty, - const std::vector<BasedUser>& UsersToProcess) { - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) { - int64_t SInt = SC->getValue()->getSExtValue(); - for (unsigned NewStride = 0, e = IU->StrideOrder.size(); - NewStride != e; ++NewStride) { - std::map<const SCEV *, IVsOfOneStride>::iterator SI = - IVsByStride.find(IU->StrideOrder[NewStride]); - if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first)) - continue; - // The other stride has no uses, don't reuse it. - std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI = - IU->IVUsesByStride.find(IU->StrideOrder[NewStride]); - if (UI->second->Users.empty()) - continue; - int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); - if (SI->first != Stride && - (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0)) - continue; - int64_t Scale = SInt / SSInt; - // Check that this stride is valid for all the types used for loads and - // stores; if it can be used for some and not others, we might as well use - // the original stride everywhere, since we have to create the IV for it - // anyway. If the scale is 1, then we don't need to worry about folding - // multiplications. - if (Scale == 1 || - (AllUsesAreAddresses && - ValidScale(HasBaseReg, Scale, UsersToProcess))) { - // Prefer to reuse an IV with a base of zero. - for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), - IE = SI->second.IVs.end(); II != IE; ++II) - // Only reuse previous IV if it would not require a type conversion - // and if the base difference can be folded. - if (II->Base->isZero() && - !RequiresTypeConversion(II->Base->getType(), Ty)) { - IV = *II; - return SE->getIntegerSCEV(Scale, Stride->getType()); - } - // Otherwise, settle for an IV with a foldable base. - if (AllUsesAreAddresses) - for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), - IE = SI->second.IVs.end(); II != IE; ++II) - // Only reuse previous IV if it would not require a type conversion - // and if the base difference can be folded. - if (SE->getEffectiveSCEVType(II->Base->getType()) == - SE->getEffectiveSCEVType(Ty) && - isa<SCEVConstant>(II->Base)) { - int64_t Base = - cast<SCEVConstant>(II->Base)->getValue()->getSExtValue(); - if (Base > INT32_MIN && Base <= INT32_MAX && - ValidOffset(HasBaseReg, -Base * Scale, - Scale, UsersToProcess)) { - IV = *II; - return SE->getIntegerSCEV(Scale, Stride->getType()); - } - } - } - } - } else if (AllUsesAreOutsideLoop) { - // Accept nonconstant strides here; it is really really right to substitute - // an existing IV if we can. - for (unsigned NewStride = 0, e = IU->StrideOrder.size(); - NewStride != e; ++NewStride) { - std::map<const SCEV *, IVsOfOneStride>::iterator SI = - IVsByStride.find(IU->StrideOrder[NewStride]); - if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first)) - continue; - int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); - if (SI->first != Stride && SSInt != 1) - continue; - for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), - IE = SI->second.IVs.end(); II != IE; ++II) - // Accept nonzero base here. - // Only reuse previous IV if it would not require a type conversion. - if (!RequiresTypeConversion(II->Base->getType(), Ty)) { - IV = *II; - return Stride; - } - } - // Special case, old IV is -1*x and this one is x. Can treat this one as - // -1*old. - for (unsigned NewStride = 0, e = IU->StrideOrder.size(); - NewStride != e; ++NewStride) { - std::map<const SCEV *, IVsOfOneStride>::iterator SI = - IVsByStride.find(IU->StrideOrder[NewStride]); - if (SI == IVsByStride.end()) - continue; - if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first)) - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0))) - if (Stride == ME->getOperand(1) && - SC->getValue()->getSExtValue() == -1LL) - for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), - IE = SI->second.IVs.end(); II != IE; ++II) - // Accept nonzero base here. - // Only reuse previous IV if it would not require type conversion. - if (!RequiresTypeConversion(II->Base->getType(), Ty)) { - IV = *II; - return SE->getIntegerSCEV(-1LL, Stride->getType()); - } - } - } - return SE->getIntegerSCEV(0, Stride->getType()); -} - -/// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that -/// returns true if Val's isUseOfPostIncrementedValue is true. -static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) { - return Val.isUseOfPostIncrementedValue; -} - -/// isNonConstantNegative - Return true if the specified scev is negated, but -/// not a constant. -static bool isNonConstantNegative(const SCEV *Expr) { - const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr); - if (!Mul) return false; - - // If there is a constant factor, it will be first. - const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0)); - if (!SC) return false; - - // Return true if the value is negative, this matches things like (-42 * V). - return SC->getValue()->getValue().isNegative(); -} - -/// CollectIVUsers - Transform our list of users and offsets to a bit more -/// complex table. In this new vector, each 'BasedUser' contains 'Base', the -/// base of the strided accesses, as well as the old information from Uses. We -/// progressively move information from the Base field to the Imm field, until -/// we eventually have the full access expression to rewrite the use. -const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *Stride, - IVUsersOfOneStride &Uses, - Loop *L, - bool &AllUsesAreAddresses, - bool &AllUsesAreOutsideLoop, - std::vector<BasedUser> &UsersToProcess) { - // FIXME: Generalize to non-affine IV's. - if (!Stride->isLoopInvariant(L)) - return SE->getIntegerSCEV(0, Stride->getType()); - - UsersToProcess.reserve(Uses.Users.size()); - for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(), - E = Uses.Users.end(); I != E; ++I) { - UsersToProcess.push_back(BasedUser(*I, SE)); - - // Move any loop variant operands from the offset field to the immediate - // field of the use, so that we don't try to use something before it is - // computed. - MoveLoopVariantsToImmediateField(UsersToProcess.back().Base, - UsersToProcess.back().Imm, L, SE); - assert(UsersToProcess.back().Base->isLoopInvariant(L) && - "Base value is not loop invariant!"); - } - - // We now have a whole bunch of uses of like-strided induction variables, but - // they might all have different bases. We want to emit one PHI node for this - // stride which we fold as many common expressions (between the IVs) into as - // possible. Start by identifying the common expressions in the base values - // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find - // "A+B"), emit it to the preheader, then remove the expression from the - // UsersToProcess base values. - const SCEV *CommonExprs = - RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI); - - // Next, figure out what we can represent in the immediate fields of - // instructions. If we can represent anything there, move it to the imm - // fields of the BasedUsers. We do this so that it increases the commonality - // of the remaining uses. - unsigned NumPHI = 0; - bool HasAddress = false; - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { - // If the user is not in the current loop, this means it is using the exit - // value of the IV. Do not put anything in the base, make sure it's all in - // the immediate field to allow as much factoring as possible. - if (!L->contains(UsersToProcess[i].Inst)) { - UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, - UsersToProcess[i].Base); - UsersToProcess[i].Base = - SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType()); - } else { - // Not all uses are outside the loop. - AllUsesAreOutsideLoop = false; - - // Addressing modes can be folded into loads and stores. Be careful that - // the store is through the expression, not of the expression though. - bool isPHI = false; - bool isAddress = isAddressUse(UsersToProcess[i].Inst, - UsersToProcess[i].OperandValToReplace); - if (isa<PHINode>(UsersToProcess[i].Inst)) { - isPHI = true; - ++NumPHI; - } + SmallVector<int64_t, 8> Offsets; + int64_t MinOffset; + int64_t MaxOffset; - if (isAddress) - HasAddress = true; + /// AllFixupsOutsideLoop - This records whether all of the fixups using this + /// LSRUse are outside of the loop, in which case some special-case heuristics + /// may be used. + bool AllFixupsOutsideLoop; - // If this use isn't an address, then not all uses are addresses. - if (!isAddress && !isPHI) - AllUsesAreAddresses = false; + /// Formulae - A list of ways to build a value that can satisfy this user. + /// After the list is populated, one of these is selected heuristically and + /// used to formulate a replacement for OperandValToReplace in UserInst. + SmallVector<Formula, 12> Formulae; - MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base, - UsersToProcess[i].Imm, isAddress, L, SE); - } - } + /// Regs - The set of register candidates used by all formulae in this LSRUse. + SmallPtrSet<const SCEV *, 4> Regs; - // If one of the use is a PHI node and all other uses are addresses, still - // allow iv reuse. Essentially we are trading one constant multiplication - // for one fewer iv. - if (NumPHI > 1) - AllUsesAreAddresses = false; + LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T), + MinOffset(INT64_MAX), + MaxOffset(INT64_MIN), + AllFixupsOutsideLoop(true) {} - // There are no in-loop address uses. - if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop)) - AllUsesAreAddresses = false; + bool InsertFormula(size_t LUIdx, const Formula &F); - return CommonExprs; -} + void check() const; -/// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction -/// is valid and profitable for the given set of users of a stride. In -/// full strength-reduction mode, all addresses at the current stride are -/// strength-reduced all the way down to pointer arithmetic. -/// -bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode( - const std::vector<BasedUser> &UsersToProcess, - const Loop *L, - bool AllUsesAreAddresses, - const SCEV *Stride) { - if (!EnableFullLSRMode) - return false; + void print(raw_ostream &OS) const; + void dump() const; +}; - // The heuristics below aim to avoid increasing register pressure, but - // fully strength-reducing all the addresses increases the number of - // add instructions, so don't do this when optimizing for size. - // TODO: If the loop is large, the savings due to simpler addresses - // may oughtweight the costs of the extra increment instructions. - if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize)) - return false; +/// InsertFormula - If the given formula has not yet been inserted, add it to +/// the list, and return true. Return false otherwise. +bool LSRUse::InsertFormula(size_t LUIdx, const Formula &F) { + SmallVector<const SCEV *, 2> Key = F.BaseRegs; + if (F.ScaledReg) Key.push_back(F.ScaledReg); + // Unstable sort by host order ok, because this is only used for uniquifying. + std::sort(Key.begin(), Key.end()); - // TODO: For now, don't do full strength reduction if there could - // potentially be greater-stride multiples of the current stride - // which could reuse the current stride IV. - if (IU->StrideOrder.back() != Stride) + if (!Uniquifier.insert(Key).second) return false; - // Iterate through the uses to find conditions that automatically rule out - // full-lsr mode. - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) { - const SCEV *Base = UsersToProcess[i].Base; - const SCEV *Imm = UsersToProcess[i].Imm; - // If any users have a loop-variant component, they can't be fully - // strength-reduced. - if (Imm && !Imm->isLoopInvariant(L)) - return false; - // If there are to users with the same base and the difference between - // the two Imm values can't be folded into the address, full - // strength reduction would increase register pressure. - do { - const SCEV *CurImm = UsersToProcess[i].Imm; - if ((CurImm || Imm) && CurImm != Imm) { - if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType()); - if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType()); - const Instruction *Inst = UsersToProcess[i].Inst; - const Type *AccessTy = getAccessType(Inst); - const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm); - if (!Diff->isZero() && - (!AllUsesAreAddresses || - !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true))) - return false; - } - } while (++i != e && Base == UsersToProcess[i].Base); - } + // Using a register to hold the value of 0 is not profitable. + assert((!F.ScaledReg || !F.ScaledReg->isZero()) && + "Zero allocated in a scaled register!"); +#ifndef NDEBUG + for (SmallVectorImpl<const SCEV *>::const_iterator I = + F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) + assert(!(*I)->isZero() && "Zero allocated in a base register!"); +#endif - // If there's exactly one user in this stride, fully strength-reducing it - // won't increase register pressure. If it's starting from a non-zero base, - // it'll be simpler this way. - if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero()) - return true; + // Add the formula to the list. + Formulae.push_back(F); - // Otherwise, if there are any users in this stride that don't require - // a register for their base, full strength-reduction will increase - // register pressure. - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) - if (UsersToProcess[i].Base->isZero()) - return false; + // Record registers now being used by this use. + if (F.ScaledReg) Regs.insert(F.ScaledReg); + Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); - // Otherwise, go for it. return true; } -/// InsertAffinePhi Create and insert a PHI node for an induction variable -/// with the specified start and step values in the specified loop. -/// -/// If NegateStride is true, the stride should be negated by using a -/// subtract instead of an add. -/// -/// Return the created phi node. -/// -static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step, - Instruction *IVIncInsertPt, - const Loop *L, - SCEVExpander &Rewriter) { - assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!"); - assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!"); - - BasicBlock *Header = L->getHeader(); - BasicBlock *Preheader = L->getLoopPreheader(); - BasicBlock *LatchBlock = L->getLoopLatch(); - const Type *Ty = Start->getType(); - Ty = Rewriter.SE.getEffectiveSCEVType(Ty); - - PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin()); - PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()), - Preheader); - - // If the stride is negative, insert a sub instead of an add for the - // increment. - bool isNegative = isNonConstantNegative(Step); - const SCEV *IncAmount = Step; - if (isNegative) - IncAmount = Rewriter.SE.getNegativeSCEV(Step); - - // Insert an add instruction right before the terminator corresponding - // to the back-edge or just before the only use. The location is determined - // by the caller and passed in as IVIncInsertPt. - Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty, - Preheader->getTerminator()); - Instruction *IncV; - if (isNegative) { - IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next", - IVIncInsertPt); - } else { - IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next", - IVIncInsertPt); - } - if (!isa<ConstantInt>(StepV)) ++NumVariable; - - PN->addIncoming(IncV, LatchBlock); - - ++NumInserted; - return PN; -} - -static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) { - // We want to emit code for users inside the loop first. To do this, we - // rearrange BasedUser so that the entries at the end have - // isUseOfPostIncrementedValue = false, because we pop off the end of the - // vector (so we handle them first). - std::partition(UsersToProcess.begin(), UsersToProcess.end(), - PartitionByIsUseOfPostIncrementedValue); - - // Sort this by base, so that things with the same base are handled - // together. By partitioning first and stable-sorting later, we are - // guaranteed that within each base we will pop off users from within the - // loop before users outside of the loop with a particular base. - // - // We would like to use stable_sort here, but we can't. The problem is that - // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so - // we don't have anything to do a '<' comparison on. Because we think the - // number of uses is small, do a horrible bubble sort which just relies on - // ==. - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { - // Get a base value. - const SCEV *Base = UsersToProcess[i].Base; - - // Compact everything with this base to be consecutive with this one. - for (unsigned j = i+1; j != e; ++j) { - if (UsersToProcess[j].Base == Base) { - std::swap(UsersToProcess[i+1], UsersToProcess[j]); - ++i; - } - } +void LSRUse::print(raw_ostream &OS) const { + OS << "LSR Use: Kind="; + switch (Kind) { + case Basic: OS << "Basic"; break; + case Special: OS << "Special"; break; + case ICmpZero: OS << "ICmpZero"; break; + case Address: + OS << "Address of "; + if (isa<PointerType>(AccessTy)) + OS << "pointer"; // the full pointer type could be really verbose + else + OS << *AccessTy; } -} -/// PrepareToStrengthReduceFully - Prepare to fully strength-reduce -/// UsersToProcess, meaning lowering addresses all the way down to direct -/// pointer arithmetic. -/// -void -LoopStrengthReduce::PrepareToStrengthReduceFully( - std::vector<BasedUser> &UsersToProcess, - const SCEV *Stride, - const SCEV *CommonExprs, - const Loop *L, - SCEVExpander &PreheaderRewriter) { - DEBUG(dbgs() << " Fully reducing all users\n"); - - // Rewrite the UsersToProcess records, creating a separate PHI for each - // unique Base value. - Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator(); - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) { - // TODO: The uses are grouped by base, but not sorted. We arbitrarily - // pick the first Imm value here to start with, and adjust it for the - // other uses. - const SCEV *Imm = UsersToProcess[i].Imm; - const SCEV *Base = UsersToProcess[i].Base; - const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm); - PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L, - PreheaderRewriter); - // Loop over all the users with the same base. - do { - UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType()); - UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm); - UsersToProcess[i].Phi = Phi; - assert(UsersToProcess[i].Imm->isLoopInvariant(L) && - "ShouldUseFullStrengthReductionMode should reject this!"); - } while (++i != e && Base == UsersToProcess[i].Base); - } -} - -/// FindIVIncInsertPt - Return the location to insert the increment instruction. -/// If the only use if a use of postinc value, (must be the loop termination -/// condition), then insert it just before the use. -static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess, - const Loop *L) { - if (UsersToProcess.size() == 1 && - UsersToProcess[0].isUseOfPostIncrementedValue && - L->contains(UsersToProcess[0].Inst)) - return UsersToProcess[0].Inst; - return L->getLoopLatch()->getTerminator(); -} - -/// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the -/// given users to share. -/// -void -LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi( - std::vector<BasedUser> &UsersToProcess, - const SCEV *Stride, - const SCEV *CommonExprs, - Value *CommonBaseV, - Instruction *IVIncInsertPt, - const Loop *L, - SCEVExpander &PreheaderRewriter) { - DEBUG(dbgs() << " Inserting new PHI:\n"); - - PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV), - Stride, IVIncInsertPt, L, - PreheaderRewriter); - - // Remember this in case a later stride is multiple of this. - IVsByStride[Stride].addIV(Stride, CommonExprs, Phi); - - // All the users will share this new IV. - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) - UsersToProcess[i].Phi = Phi; - - DEBUG(dbgs() << " IV="); - DEBUG(WriteAsOperand(dbgs(), Phi, /*PrintType=*/false)); - DEBUG(dbgs() << "\n"); -} - -/// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to -/// reuse an induction variable with a stride that is a factor of the current -/// induction variable. -/// -void -LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride( - std::vector<BasedUser> &UsersToProcess, - Value *CommonBaseV, - const IVExpr &ReuseIV, - Instruction *PreInsertPt) { - DEBUG(dbgs() << " Rewriting in terms of existing IV of STRIDE " - << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n"); - - // All the users will share the reused IV. - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) - UsersToProcess[i].Phi = ReuseIV.PHI; - - Constant *C = dyn_cast<Constant>(CommonBaseV); - if (C && - (!C->isNullValue() && - !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(), - TLI, false))) - // We want the common base emitted into the preheader! This is just - // using cast as a copy so BitCast (no-op cast) is appropriate - CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(), - "commonbase", PreInsertPt); -} - -static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset, - const Type *AccessTy, - std::vector<BasedUser> &UsersToProcess, - const TargetLowering *TLI) { - SmallVector<Instruction*, 16> AddrModeInsts; - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { - if (UsersToProcess[i].isUseOfPostIncrementedValue) - continue; - ExtAddrMode AddrMode = - AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace, - AccessTy, UsersToProcess[i].Inst, - AddrModeInsts, *TLI); - if (GV && GV != AddrMode.BaseGV) - return false; - if (Offset && !AddrMode.BaseOffs) - // FIXME: How to accurate check it's immediate offset is folded. - return false; - AddrModeInsts.clear(); + OS << ", Offsets={"; + for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), + E = Offsets.end(); I != E; ++I) { + OS << *I; + if (next(I) != E) + OS << ','; } - return true; -} + OS << '}'; -/// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single -/// stride of IV. All of the users may have different starting values, and this -/// may not be the only stride. -void -LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *Stride, - IVUsersOfOneStride &Uses, - Loop *L) { - // If all the users are moved to another stride, then there is nothing to do. - if (Uses.Users.empty()) - return; + if (AllFixupsOutsideLoop) + OS << ", all-fixups-outside-loop"; +} - // Keep track if every use in UsersToProcess is an address. If they all are, - // we may be able to rewrite the entire collection of them in terms of a - // smaller-stride IV. - bool AllUsesAreAddresses = true; - - // Keep track if every use of a single stride is outside the loop. If so, - // we want to be more aggressive about reusing a smaller-stride IV; a - // multiply outside the loop is better than another IV inside. Well, usually. - bool AllUsesAreOutsideLoop = true; - - // Transform our list of users and offsets to a bit more complex table. In - // this new vector, each 'BasedUser' contains 'Base' the base of the - // strided accessas well as the old information from Uses. We progressively - // move information from the Base field to the Imm field, until we eventually - // have the full access expression to rewrite the use. - std::vector<BasedUser> UsersToProcess; - const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses, - AllUsesAreOutsideLoop, - UsersToProcess); - - // Sort the UsersToProcess array so that users with common bases are - // next to each other. - SortUsersToProcess(UsersToProcess); - - // If we managed to find some expressions in common, we'll need to carry - // their value in a register and add it in for each use. This will take up - // a register operand, which potentially restricts what stride values are - // valid. - bool HaveCommonExprs = !CommonExprs->isZero(); - const Type *ReplacedTy = CommonExprs->getType(); - - // If all uses are addresses, consider sinking the immediate part of the - // common expression back into uses if they can fit in the immediate fields. - if (TLI && HaveCommonExprs && AllUsesAreAddresses) { - const SCEV *NewCommon = CommonExprs; - const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy); - MoveImmediateValues(TLI, Type::getVoidTy( - L->getLoopPreheader()->getContext()), - NewCommon, Imm, true, L, SE); - if (!Imm->isZero()) { - bool DoSink = true; - - // If the immediate part of the common expression is a GV, check if it's - // possible to fold it into the target addressing mode. - GlobalValue *GV = 0; - if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm)) - GV = dyn_cast<GlobalValue>(SU->getValue()); - int64_t Offset = 0; - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm)) - Offset = SC->getValue()->getSExtValue(); - if (GV || Offset) - // Pass VoidTy as the AccessTy to be conservative, because - // there could be multiple access types among all the uses. - DoSink = IsImmFoldedIntoAddrMode(GV, Offset, - Type::getVoidTy(L->getLoopPreheader()->getContext()), - UsersToProcess, TLI); - - if (DoSink) { - DEBUG(dbgs() << " Sinking " << *Imm << " back down into uses\n"); - for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) - UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm); - CommonExprs = NewCommon; - HaveCommonExprs = !CommonExprs->isZero(); - ++NumImmSunk; - } - } - } +void LSRUse::dump() const { + print(errs()); errs() << '\n'; +} - // Now that we know what we need to do, insert the PHI node itself. - // - DEBUG(dbgs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE " - << *Stride << ":\n" - << " Common base: " << *CommonExprs << "\n"); +/// 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, + LSRUse::KindType Kind, const Type *AccessTy, + const TargetLowering *TLI) { + switch (Kind) { + case LSRUse::Address: + // If we have low-level target information, ask the target if it can + // completely fold this address. + if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy); + + // Otherwise, just guess that reg+reg addressing is legal. + return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1; + + case LSRUse::ICmpZero: + // There's not even a target hook for querying whether it would be legal to + // fold a GV into an ICmp. + if (AM.BaseGV) + return false; - SCEVExpander Rewriter(*SE); - SCEVExpander PreheaderRewriter(*SE); + // ICmp only has two operands; don't allow more than two non-trivial parts. + if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0) + return false; - BasicBlock *Preheader = L->getLoopPreheader(); - Instruction *PreInsertPt = Preheader->getTerminator(); - BasicBlock *LatchBlock = L->getLoopLatch(); - Instruction *IVIncInsertPt = LatchBlock->getTerminator(); - - Value *CommonBaseV = Constant::getNullValue(ReplacedTy); - - const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy); - IVExpr ReuseIV(SE->getIntegerSCEV(0, - Type::getInt32Ty(Preheader->getContext())), - SE->getIntegerSCEV(0, - Type::getInt32Ty(Preheader->getContext())), - 0); - - // Choose a strength-reduction strategy and prepare for it by creating - // the necessary PHIs and adjusting the bookkeeping. - if (ShouldUseFullStrengthReductionMode(UsersToProcess, L, - AllUsesAreAddresses, Stride)) { - PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L, - PreheaderRewriter); - } else { - // Emit the initial base value into the loop preheader. - CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy, - PreInsertPt); - - // If all uses are addresses, check if it is possible to reuse an IV. The - // new IV must have a stride that is a multiple of the old stride; the - // multiple must be a number that can be encoded in the scale field of the - // target addressing mode; and we must have a valid instruction after this - // substitution, including the immediate field, if any. - RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses, - AllUsesAreOutsideLoop, - Stride, ReuseIV, ReplacedTy, - UsersToProcess); - if (!RewriteFactor->isZero()) - PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV, - ReuseIV, PreInsertPt); - else { - IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L); - PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs, - CommonBaseV, IVIncInsertPt, - L, PreheaderRewriter); - } - } + // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by + // putting the scaled register in the other operand of the icmp. + if (AM.Scale != 0 && AM.Scale != -1) + return false; - // Process all the users now, replacing their strided uses with - // strength-reduced forms. This outer loop handles all bases, the inner - // loop handles all users of a particular base. - while (!UsersToProcess.empty()) { - const SCEV *Base = UsersToProcess.back().Base; - Instruction *Inst = UsersToProcess.back().Inst; - - // Emit the code for Base into the preheader. - Value *BaseV = 0; - if (!Base->isZero()) { - BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt); - - DEBUG(dbgs() << " INSERTING code for BASE = " << *Base << ":"); - if (BaseV->hasName()) - DEBUG(dbgs() << " Result value name = %" << BaseV->getName()); - DEBUG(dbgs() << "\n"); - - // If BaseV is a non-zero constant, make sure that it gets inserted into - // the preheader, instead of being forward substituted into the uses. We - // do this by forcing a BitCast (noop cast) to be inserted into the - // preheader in this case. - if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) && - isa<Constant>(BaseV)) { - // We want this constant emitted into the preheader! This is just - // using cast as a copy so BitCast (no-op cast) is appropriate - BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert", - PreInsertPt); - } + // If we have low-level target information, ask the target if it can fold an + // integer immediate on an icmp. + if (AM.BaseOffs != 0) { + if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs); + return false; } - // Emit the code to add the immediate offset to the Phi value, just before - // the instructions that we identified as using this stride and base. - do { - // FIXME: Use emitted users to emit other users. - BasedUser &User = UsersToProcess.back(); - - DEBUG(dbgs() << " Examining "); - if (User.isUseOfPostIncrementedValue) - DEBUG(dbgs() << "postinc"); - else - DEBUG(dbgs() << "preinc"); - DEBUG(dbgs() << " use "); - DEBUG(WriteAsOperand(dbgs(), UsersToProcess.back().OperandValToReplace, - /*PrintType=*/false)); - DEBUG(dbgs() << " in Inst: " << *User.Inst); - - // If this instruction wants to use the post-incremented value, move it - // after the post-inc and use its value instead of the PHI. - Value *RewriteOp = User.Phi; - if (User.isUseOfPostIncrementedValue) { - RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock); - // If this user is in the loop, make sure it is the last thing in the - // loop to ensure it is dominated by the increment. In case it's the - // only use of the iv, the increment instruction is already before the - // use. - if (L->contains(User.Inst) && User.Inst != IVIncInsertPt) - User.Inst->moveBefore(IVIncInsertPt); - } - - const SCEV *RewriteExpr = SE->getUnknown(RewriteOp); - - if (SE->getEffectiveSCEVType(RewriteOp->getType()) != - SE->getEffectiveSCEVType(ReplacedTy)) { - assert(SE->getTypeSizeInBits(RewriteOp->getType()) > - SE->getTypeSizeInBits(ReplacedTy) && - "Unexpected widening cast!"); - RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy); - } + return true; - // If we had to insert new instructions for RewriteOp, we have to - // consider that they may not have been able to end up immediately - // next to RewriteOp, because non-PHI instructions may never precede - // PHI instructions in a block. In this case, remember where the last - // instruction was inserted so that if we're replacing a different - // PHI node, we can use the later point to expand the final - // RewriteExpr. - Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp); - if (RewriteOp == User.Phi) NewBasePt = 0; - - // Clear the SCEVExpander's expression map so that we are guaranteed - // to have the code emitted where we expect it. - Rewriter.clear(); - - // If we are reusing the iv, then it must be multiplied by a constant - // factor to take advantage of the addressing mode scale component. - if (!RewriteFactor->isZero()) { - // If we're reusing an IV with a nonzero base (currently this happens - // only when all reuses are outside the loop) subtract that base here. - // The base has been used to initialize the PHI node but we don't want - // it here. - if (!ReuseIV.Base->isZero()) { - const SCEV *typedBase = ReuseIV.Base; - if (SE->getEffectiveSCEVType(RewriteExpr->getType()) != - SE->getEffectiveSCEVType(ReuseIV.Base->getType())) { - // It's possible the original IV is a larger type than the new IV, - // in which case we have to truncate the Base. We checked in - // RequiresTypeConversion that this is valid. - assert(SE->getTypeSizeInBits(RewriteExpr->getType()) < - SE->getTypeSizeInBits(ReuseIV.Base->getType()) && - "Unexpected lengthening conversion!"); - typedBase = SE->getTruncateExpr(ReuseIV.Base, - RewriteExpr->getType()); - } - RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase); - } + case LSRUse::Basic: + // Only handle single-register values. + return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0; - // Multiply old variable, with base removed, by new scale factor. - RewriteExpr = SE->getMulExpr(RewriteFactor, - RewriteExpr); - - // The common base is emitted in the loop preheader. But since we - // are reusing an IV, it has not been used to initialize the PHI node. - // Add it to the expression used to rewrite the uses. - // When this use is outside the loop, we earlier subtracted the - // common base, and are adding it back here. Use the same expression - // as before, rather than CommonBaseV, so DAGCombiner will zap it. - if (!CommonExprs->isZero()) { - if (L->contains(User.Inst)) - RewriteExpr = SE->getAddExpr(RewriteExpr, - SE->getUnknown(CommonBaseV)); - else - RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs); - } - } - - // Now that we know what we need to do, insert code before User for the - // immediate and any loop-variant expressions. - if (BaseV) - // Add BaseV to the PHI value if needed. - RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV)); - - User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt, - Rewriter, L, this, - DeadInsts, SE); - - // Mark old value we replaced as possibly dead, so that it is eliminated - // if we just replaced the last use of that value. - DeadInsts.push_back(User.OperandValToReplace); - - UsersToProcess.pop_back(); - ++NumReduced; - - // If there are any more users to process with the same base, process them - // now. We sorted by base above, so we just have to check the last elt. - } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base); - // TODO: Next, find out which base index is the most common, pull it out. - } - - // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but - // different starting values, into different PHIs. -} - -void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) { - // Note: this processes each stride/type pair individually. All users - // passed into StrengthReduceIVUsersOfStride have the same type AND stride. - // Also, note that we iterate over IVUsesByStride indirectly by using - // StrideOrder. This extra layer of indirection makes the ordering of - // strides deterministic - not dependent on map order. - for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) { - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[Stride]); - assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - // FIXME: Generalize to non-affine IV's. - if (!SI->first->isLoopInvariant(L)) - continue; - StrengthReduceIVUsersOfStride(SI->first, *SI->second, L); + case LSRUse::Special: + // Only handle -1 scales, or no scale. + return AM.Scale == 0 || AM.Scale == -1; } + + return false; } -/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, -/// set the IV user and stride information and return true, otherwise return -/// false. -bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, - IVStrideUse *&CondUse, - const SCEV* &CondStride) { - for (unsigned Stride = 0, e = IU->StrideOrder.size(); - Stride != e && !CondUse; ++Stride) { - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[Stride]); - assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - - for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(), - E = SI->second->Users.end(); UI != E; ++UI) - if (UI->getUser() == Cond) { - // NOTE: we could handle setcc instructions with multiple uses here, but - // InstCombine does it as well for simple uses, it's not clear that it - // occurs enough in real life to handle. - CondUse = UI; - CondStride = SI->first; - return true; - } +static bool isLegalUse(TargetLowering::AddrMode AM, + int64_t MinOffset, int64_t MaxOffset, + LSRUse::KindType Kind, const Type *AccessTy, + const TargetLowering *TLI) { + // Check for overflow. + if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) != + (MinOffset > 0)) + return false; + AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset; + if (isLegalUse(AM, Kind, AccessTy, TLI)) { + AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset; + // Check for overflow. + if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) != + (MaxOffset > 0)) + return false; + AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset; + return isLegalUse(AM, Kind, AccessTy, TLI); } return false; } -namespace { - // Constant strides come first which in turns are sorted by their absolute - // values. If absolute values are the same, then positive strides comes first. - // e.g. - // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X - struct StrideCompare { - const ScalarEvolution *SE; - explicit StrideCompare(const ScalarEvolution *se) : SE(se) {} - - bool operator()(const SCEV *LHS, const SCEV *RHS) { - const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS); - const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS); - if (LHSC && RHSC) { - int64_t LV = LHSC->getValue()->getSExtValue(); - int64_t RV = RHSC->getValue()->getSExtValue(); - uint64_t ALV = (LV < 0) ? -LV : LV; - uint64_t ARV = (RV < 0) ? -RV : RV; - if (ALV == ARV) { - if (LV != RV) - return LV > RV; - } else { - return ALV < ARV; - } +static bool isAlwaysFoldable(int64_t BaseOffs, + GlobalValue *BaseGV, + bool HasBaseReg, + LSRUse::KindType Kind, const Type *AccessTy, + const TargetLowering *TLI, + ScalarEvolution &SE) { + // Fast-path: zero is always foldable. + if (BaseOffs == 0 && !BaseGV) return true; + + // Conservatively, create an address with an immediate and a + // base and a scale. + TargetLowering::AddrMode AM; + AM.BaseOffs = BaseOffs; + AM.BaseGV = BaseGV; + AM.HasBaseReg = HasBaseReg; + AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; + + return isLegalUse(AM, Kind, AccessTy, TLI); +} - // If it's the same value but different type, sort by bit width so - // that we emit larger induction variables before smaller - // ones, letting the smaller be re-written in terms of larger ones. - return SE->getTypeSizeInBits(RHS->getType()) < - SE->getTypeSizeInBits(LHS->getType()); - } - return LHSC && !RHSC; - } - }; +static bool isAlwaysFoldable(const SCEV *S, + int64_t MinOffset, int64_t MaxOffset, + bool HasBaseReg, + LSRUse::KindType Kind, const Type *AccessTy, + const TargetLowering *TLI, + ScalarEvolution &SE) { + // Fast-path: zero is always foldable. + if (S->isZero()) return true; + + // Conservatively, create an address with an immediate and a + // base and a scale. + int64_t BaseOffs = ExtractImmediate(S, SE); + GlobalValue *BaseGV = ExtractSymbol(S, SE); + + // If there's anything else involved, it's not foldable. + if (!S->isZero()) return false; + + // Fast-path: zero is always foldable. + if (BaseOffs == 0 && !BaseGV) return true; + + // Conservatively, create an address with an immediate and a + // base and a scale. + TargetLowering::AddrMode AM; + AM.BaseOffs = BaseOffs; + AM.BaseGV = BaseGV; + AM.HasBaseReg = HasBaseReg; + AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; + + return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI); } -/// ChangeCompareStride - If a loop termination compare instruction is the -/// only use of its stride, and the compaison is against a constant value, -/// try eliminate the stride by moving the compare instruction to another -/// stride and change its constant operand accordingly. e.g. -/// -/// loop: -/// ... -/// v1 = v1 + 3 -/// v2 = v2 + 1 -/// if (v2 < 10) goto loop -/// => -/// loop: -/// ... -/// v1 = v1 + 3 -/// if (v1 < 30) goto loop -ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond, - IVStrideUse* &CondUse, - const SCEV* &CondStride, - bool PostPass) { - // If there's only one stride in the loop, there's nothing to do here. - if (IU->StrideOrder.size() < 2) - return Cond; - // If there are other users of the condition's stride, don't bother - // trying to change the condition because the stride will still - // remain. - std::map<const SCEV *, IVUsersOfOneStride *>::iterator I = - IU->IVUsesByStride.find(CondStride); - if (I == IU->IVUsesByStride.end()) - return Cond; - if (I->second->Users.size() > 1) { - for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(), - EE = I->second->Users.end(); II != EE; ++II) { - if (II->getUser() == Cond) - continue; - if (!isInstructionTriviallyDead(II->getUser())) - return Cond; - } +/// FormulaSorter - This class implements an ordering for formulae which sorts +/// the by their standalone cost. +class FormulaSorter { + /// These two sets are kept empty, so that we compute standalone costs. + DenseSet<const SCEV *> VisitedRegs; + SmallPtrSet<const SCEV *, 16> Regs; + Loop *L; + LSRUse *LU; + ScalarEvolution &SE; + DominatorTree &DT; + +public: + FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt) + : L(l), LU(&lu), SE(se), DT(dt) {} + + bool operator()(const Formula &A, const Formula &B) { + Cost CostA; + CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT); + Regs.clear(); + Cost CostB; + CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT); + Regs.clear(); + return CostA < CostB; + } +}; + +/// LSRInstance - This class holds state for the main loop strength reduction +/// logic. +class LSRInstance { + IVUsers &IU; + ScalarEvolution &SE; + DominatorTree &DT; + const TargetLowering *const TLI; + Loop *const L; + bool Changed; + + /// IVIncInsertPos - This is the insert position that the current loop's + /// induction variable increment should be placed. In simple loops, this is + /// the latch block's terminator. But in more complicated cases, this is a + /// position which will dominate all the in-loop post-increment users. + Instruction *IVIncInsertPos; + + /// Factors - Interesting factors between use strides. + SmallSetVector<int64_t, 8> Factors; + + /// Types - Interesting use types, to facilitate truncation reuse. + SmallSetVector<const Type *, 4> Types; + + /// Fixups - The list of operands which are to be replaced. + SmallVector<LSRFixup, 16> Fixups; + + /// Uses - The list of interesting uses. + SmallVector<LSRUse, 16> Uses; + + /// RegUses - Track which uses use which register candidates. + RegUseTracker RegUses; + + void OptimizeShadowIV(); + bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); + ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); + bool OptimizeLoopTermCond(); + + void CollectInterestingTypesAndFactors(); + void CollectFixupsAndInitialFormulae(); + + LSRFixup &getNewFixup() { + Fixups.push_back(LSRFixup()); + return Fixups.back(); } - // Only handle constant strides for now. - const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride); - if (!SC) return Cond; - - ICmpInst::Predicate Predicate = Cond->getPredicate(); - int64_t CmpSSInt = SC->getValue()->getSExtValue(); - unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType()); - uint64_t SignBit = 1ULL << (BitWidth-1); - const Type *CmpTy = Cond->getOperand(0)->getType(); - const Type *NewCmpTy = NULL; - unsigned TyBits = SE->getTypeSizeInBits(CmpTy); - unsigned NewTyBits = 0; - const SCEV *NewStride = NULL; - Value *NewCmpLHS = NULL; - Value *NewCmpRHS = NULL; - int64_t Scale = 1; - const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy); - - if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) { - int64_t CmpVal = C->getValue().getSExtValue(); - - // Check the relevant induction variable for conformance to - // the pattern. - const SCEV *IV = SE->getSCEV(Cond->getOperand(0)); - const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); - if (!AR || !AR->isAffine()) - return Cond; - - const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart()); - // Check stride constant and the comparision constant signs to detect - // overflow. - if (StartC) { - if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) || - (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0)) - return Cond; - } else { - // More restrictive check for the other cases. - if ((CmpVal & SignBit) != (CmpSSInt & SignBit)) - return Cond; - } - - // Look for a suitable stride / iv as replacement. - for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[i]); - if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty()) - continue; - int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); - if (SSInt == CmpSSInt || - abs64(SSInt) < abs64(CmpSSInt) || - (SSInt % CmpSSInt) != 0) - continue; - Scale = SSInt / CmpSSInt; - int64_t NewCmpVal = CmpVal * Scale; + // Support for sharing of LSRUses between LSRFixups. + typedef DenseMap<const SCEV *, size_t> UseMapTy; + UseMapTy UseMap; + + bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, + LSRUse::KindType Kind, const Type *AccessTy); + + std::pair<size_t, int64_t> getUse(const SCEV *&Expr, + LSRUse::KindType Kind, + const Type *AccessTy); + +public: + void InsertInitialFormula(const SCEV *S, Loop *L, LSRUse &LU, size_t LUIdx); + void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); + void CountRegisters(const Formula &F, size_t LUIdx); + bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); + + void CollectLoopInvariantFixupsAndFormulae(); + + void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, + unsigned Depth = 0); + void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); + void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); + void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); + void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); + void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); + void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); + void GenerateCrossUseConstantOffsets(); + void GenerateAllReuseFormulae(); + + void FilterOutUndesirableDedicatedRegisters(); + void NarrowSearchSpaceUsingHeuristics(); + + void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, + Cost &SolutionCost, + SmallVectorImpl<const Formula *> &Workspace, + const Cost &CurCost, + const SmallPtrSet<const SCEV *, 16> &CurRegs, + DenseSet<const SCEV *> &VisitedRegs) const; + void Solve(SmallVectorImpl<const Formula *> &Solution) const; + + Value *Expand(const LSRFixup &LF, + const Formula &F, + BasicBlock::iterator IP, Loop *L, Instruction *IVIncInsertPos, + SCEVExpander &Rewriter, + SmallVectorImpl<WeakVH> &DeadInsts, + ScalarEvolution &SE, DominatorTree &DT) const; + void Rewrite(const LSRFixup &LF, + const Formula &F, + Loop *L, Instruction *IVIncInsertPos, + SCEVExpander &Rewriter, + SmallVectorImpl<WeakVH> &DeadInsts, + ScalarEvolution &SE, DominatorTree &DT, + Pass *P) const; + void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, + Pass *P); + + LSRInstance(const TargetLowering *tli, Loop *l, Pass *P); + + bool getChanged() const { return Changed; } + + void print_factors_and_types(raw_ostream &OS) const; + void print_fixups(raw_ostream &OS) const; + void print_uses(raw_ostream &OS) const; + void print(raw_ostream &OS) const; + void dump() const; +}; - // If old icmp value fits in icmp immediate field, but the new one doesn't - // try something else. - if (TLI && - TLI->isLegalICmpImmediate(CmpVal) && - !TLI->isLegalICmpImmediate(NewCmpVal)) - continue; +} - APInt Mul = APInt(BitWidth*2, CmpVal, true); - Mul = Mul * APInt(BitWidth*2, Scale, true); - // Check for overflow. - if (!Mul.isSignedIntN(BitWidth)) - continue; - // Check for overflow in the stride's type too. - if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType()))) - continue; +/// OptimizeShadowIV - If IV is used in a int-to-float cast +/// inside the loop then try to eliminate the cast opeation. +void LSRInstance::OptimizeShadowIV() { + const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); + if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) + return; - // Watch out for overflow. - if (ICmpInst::isSigned(Predicate) && - (CmpVal & SignBit) != (NewCmpVal & SignBit)) - continue; + for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); + UI != E; /* empty */) { + IVUsers::const_iterator CandidateUI = UI; + ++UI; + Instruction *ShadowUse = CandidateUI->getUser(); + const Type *DestTy = NULL; - // Pick the best iv to use trying to avoid a cast. - NewCmpLHS = NULL; - for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(), - E = SI->second->Users.end(); UI != E; ++UI) { - Value *Op = UI->getOperandValToReplace(); - - // If the IVStrideUse implies a cast, check for an actual cast which - // can be used to find the original IV expression. - if (SE->getEffectiveSCEVType(Op->getType()) != - SE->getEffectiveSCEVType(SI->first->getType())) { - CastInst *CI = dyn_cast<CastInst>(Op); - // If it's not a simple cast, it's complicated. - if (!CI) - continue; - // If it's a cast from a type other than the stride type, - // it's complicated. - if (CI->getOperand(0)->getType() != SI->first->getType()) - continue; - // Ok, we found the IV expression in the stride's type. - Op = CI->getOperand(0); - } + /* If shadow use is a int->float cast then insert a second IV + to eliminate this cast. - NewCmpLHS = Op; - if (NewCmpLHS->getType() == CmpTy) - break; - } - if (!NewCmpLHS) - continue; + for (unsigned i = 0; i < n; ++i) + foo((double)i); - NewCmpTy = NewCmpLHS->getType(); - NewTyBits = SE->getTypeSizeInBits(NewCmpTy); - const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits); - if (RequiresTypeConversion(NewCmpTy, CmpTy)) { - // Check if it is possible to rewrite it using - // an iv / stride of a smaller integer type. - unsigned Bits = NewTyBits; - if (ICmpInst::isSigned(Predicate)) - --Bits; - uint64_t Mask = (1ULL << Bits) - 1; - if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal) - continue; - } + is transformed into - // Don't rewrite if use offset is non-constant and the new type is - // of a different type. - // FIXME: too conservative? - if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset())) - continue; + double d = 0.0; + for (unsigned i = 0; i < n; ++i, ++d) + foo(d); + */ + if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) + DestTy = UCast->getDestTy(); + else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) + DestTy = SCast->getDestTy(); + if (!DestTy) continue; - if (!PostPass) { - bool AllUsesAreAddresses = true; - bool AllUsesAreOutsideLoop = true; - std::vector<BasedUser> UsersToProcess; - const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L, - AllUsesAreAddresses, - AllUsesAreOutsideLoop, - UsersToProcess); - // Avoid rewriting the compare instruction with an iv of new stride - // if it's likely the new stride uses will be rewritten using the - // stride of the compare instruction. - if (AllUsesAreAddresses && - ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) - continue; - } + if (TLI) { + // If target does not support DestTy natively then do not apply + // this transformation. + EVT DVT = TLI->getValueType(DestTy); + if (!TLI->isTypeLegal(DVT)) continue; + } - // Avoid rewriting the compare instruction with an iv which has - // implicit extension or truncation built into it. - // TODO: This is over-conservative. - if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits) - continue; + PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); + if (!PH) continue; + if (PH->getNumIncomingValues() != 2) continue; - // If scale is negative, use swapped predicate unless it's testing - // for equality. - if (Scale < 0 && !Cond->isEquality()) - Predicate = ICmpInst::getSwappedPredicate(Predicate); + const Type *SrcTy = PH->getType(); + int Mantissa = DestTy->getFPMantissaWidth(); + if (Mantissa == -1) continue; + if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) + continue; - NewStride = IU->StrideOrder[i]; - if (!isa<PointerType>(NewCmpTy)) - NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal); - else { - Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal); - NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy); - } - NewOffset = TyBits == NewTyBits - ? SE->getMulExpr(CondUse->getOffset(), - SE->getConstant(CmpTy, Scale)) - : SE->getConstant(NewCmpIntTy, - cast<SCEVConstant>(CondUse->getOffset())->getValue() - ->getSExtValue()*Scale); - break; + unsigned Entry, Latch; + if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { + Entry = 0; + Latch = 1; + } else { + Entry = 1; + Latch = 0; } - } - // Forgo this transformation if it the increment happens to be - // unfortunately positioned after the condition, and the condition - // has multiple uses which prevent it from being moved immediately - // before the branch. See - // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll - // for an example of this situation. - if (!Cond->hasOneUse()) { - for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end(); - I != E; ++I) - if (I == NewCmpLHS) - return Cond; - } + ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); + if (!Init) continue; + Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue()); - if (NewCmpRHS) { - // Create a new compare instruction using new stride / iv. - ICmpInst *OldCond = Cond; - // Insert new compare instruction. - Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS, - L->getHeader()->getName() + ".termcond"); + BinaryOperator *Incr = + dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); + if (!Incr) continue; + if (Incr->getOpcode() != Instruction::Add + && Incr->getOpcode() != Instruction::Sub) + continue; + + /* Initialize new IV, double d = 0.0 in above example. */ + ConstantInt *C = NULL; + if (Incr->getOperand(0) == PH) + C = dyn_cast<ConstantInt>(Incr->getOperand(1)); + else if (Incr->getOperand(1) == PH) + C = dyn_cast<ConstantInt>(Incr->getOperand(0)); + else + continue; - DEBUG(dbgs() << " Change compare stride in Inst " << *OldCond); - DEBUG(dbgs() << " to " << *Cond << '\n'); + if (!C) continue; - // Remove the old compare instruction. The old indvar is probably dead too. - DeadInsts.push_back(CondUse->getOperandValToReplace()); - OldCond->replaceAllUsesWith(Cond); - OldCond->eraseFromParent(); + // Ignore negative constants, as the code below doesn't handle them + // correctly. TODO: Remove this restriction. + if (!C->getValue().isStrictlyPositive()) continue; - IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS); - CondUse = &IU->IVUsesByStride[NewStride]->Users.back(); - CondStride = NewStride; - ++NumEliminated; - Changed = true; + /* Add new PHINode. */ + PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH); + + /* create new increment. '++d' in above example. */ + Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); + BinaryOperator *NewIncr = + BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? + Instruction::FAdd : Instruction::FSub, + NewPH, CFP, "IV.S.next.", Incr); + + NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); + NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); + + /* Remove cast operation */ + ShadowUse->replaceAllUsesWith(NewPH); + ShadowUse->eraseFromParent(); + break; } +} - return Cond; +/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, +/// set the IV user and stride information and return true, otherwise return +/// false. +bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, + IVStrideUse *&CondUse) { + for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) + if (UI->getUser() == Cond) { + // NOTE: we could handle setcc instructions with multiple uses here, but + // InstCombine does it as well for simple uses, it's not clear that it + // occurs enough in real life to handle. + CondUse = UI; + return true; + } + return false; } /// OptimizeMax - Rewrite the loop's terminating condition if it uses @@ -2087,7 +1371,7 @@ ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond, /// are designed around them. The most obvious example of this is the /// LoopInfo analysis, which doesn't remember trip count values. It /// expects to be able to rediscover the trip count each time it is -/// needed, and it does this using a simple analyis that only succeeds if +/// needed, and it does this using a simple analysis that only succeeds if /// the loop has a canonical induction variable. /// /// However, when it comes time to generate code, the maximum operation @@ -2097,8 +1381,7 @@ ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond, /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting /// the instructions for the maximum computation. /// -ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond, - IVStrideUse* &CondUse) { +ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { // Check that the loop matches the pattern we're looking for. if (Cond->getPredicate() != CmpInst::ICMP_EQ && Cond->getPredicate() != CmpInst::ICMP_NE) @@ -2107,19 +1390,19 @@ ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond, SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); if (!Sel || !Sel->hasOneUse()) return Cond; - const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); + const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) return Cond; - const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType()); + const SCEV *One = SE.getIntegerSCEV(1, BackedgeTakenCount->getType()); // Add one to the backedge-taken count to get the trip count. - const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One); + const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One); // Check for a max calculation that matches the pattern. if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount)) return Cond; const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount); - if (Max != SE->getSCEV(Sel)) return Cond; + if (Max != SE.getSCEV(Sel)) return Cond; // To handle a max with more than two operands, this optimization would // require additional checking and setup. @@ -2129,14 +1412,13 @@ ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond, const SCEV *MaxLHS = Max->getOperand(0); const SCEV *MaxRHS = Max->getOperand(1); if (!MaxLHS || MaxLHS != One) return Cond; - // Check the relevant induction variable for conformance to // the pattern. - const SCEV *IV = SE->getSCEV(Cond->getOperand(0)); + const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); if (!AR || !AR->isAffine() || AR->getStart() != One || - AR->getStepRecurrence(*SE) != One) + AR->getStepRecurrence(SE) != One) return Cond; assert(AR->getLoop() == L && @@ -2145,9 +1427,9 @@ ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond, // Check the right operand of the select, and remember it, as it will // be used in the new comparison instruction. Value *NewRHS = 0; - if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS) + if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) NewRHS = Sel->getOperand(1); - else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS) + else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) NewRHS = Sel->getOperand(2); if (!NewRHS) return Cond; @@ -2174,552 +1456,1764 @@ ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond, return NewCond; } -/// OptimizeShadowIV - If IV is used in a int-to-float cast -/// inside the loop then try to eliminate the cast opeation. -void LoopStrengthReduce::OptimizeShadowIV(Loop *L) { +/// OptimizeLoopTermCond - Change loop terminating condition to use the +/// postinc iv when possible. +bool +LSRInstance::OptimizeLoopTermCond() { + SmallPtrSet<Instruction *, 4> PostIncs; - const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); - if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) - return; + BasicBlock *LatchBlock = L->getLoopLatch(); + SmallVector<BasicBlock*, 8> ExitingBlocks; + L->getExitingBlocks(ExitingBlocks); + + for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { + BasicBlock *ExitingBlock = ExitingBlocks[i]; + + // Get the terminating condition for the loop if possible. If we + // can, we want to change it to use a post-incremented version of its + // induction variable, to allow coalescing the live ranges for the IV into + // one register value. + + BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); + if (!TermBr) + continue; + // FIXME: Overly conservative, termination condition could be an 'or' etc.. + if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) + continue; - for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; - ++Stride) { - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[Stride]); - assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - if (!isa<SCEVConstant>(SI->first)) + // Search IVUsesByStride to find Cond's IVUse if there is one. + IVStrideUse *CondUse = 0; + ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); + if (!FindIVUserForCond(Cond, CondUse)) continue; - for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(), - E = SI->second->Users.end(); UI != E; /* empty */) { - ilist<IVStrideUse>::iterator CandidateUI = UI; - ++UI; - Instruction *ShadowUse = CandidateUI->getUser(); - const Type *DestTy = NULL; - - /* If shadow use is a int->float cast then insert a second IV - to eliminate this cast. - - for (unsigned i = 0; i < n; ++i) - foo((double)i); - - is transformed into - - double d = 0.0; - for (unsigned i = 0; i < n; ++i, ++d) - foo(d); - */ - if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) - DestTy = UCast->getDestTy(); - else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) - DestTy = SCast->getDestTy(); - if (!DestTy) continue; - - if (TLI) { - // If target does not support DestTy natively then do not apply - // this transformation. - EVT DVT = TLI->getValueType(DestTy); - if (!TLI->isTypeLegal(DVT)) continue; - } + // If the trip count is computed in terms of a max (due to ScalarEvolution + // being unable to find a sufficient guard, for example), change the loop + // comparison to use SLT or ULT instead of NE. + // One consequence of doing this now is that it disrupts the count-down + // optimization. That's not always a bad thing though, because in such + // cases it may still be worthwhile to avoid a max. + Cond = OptimizeMax(Cond, CondUse); + + // If this exiting block dominates the latch block, it may also use + // the post-inc value if it won't be shared with other uses. + // Check for dominance. + if (!DT.dominates(ExitingBlock, LatchBlock)) + continue; - PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); - if (!PH) continue; - if (PH->getNumIncomingValues() != 2) continue; + // Conservatively avoid trying to use the post-inc value in non-latch + // exits if there may be pre-inc users in intervening blocks. + if (LatchBlock != ExitingBlock) + for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) + // Test if the use is reachable from the exiting block. This dominator + // query is a conservative approximation of reachability. + if (&*UI != CondUse && + !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { + // Conservatively assume there may be reuse if the quotient of their + // strides could be a legal scale. + const SCEV *A = CondUse->getStride(); + const SCEV *B = UI->getStride(); + if (SE.getTypeSizeInBits(A->getType()) != + SE.getTypeSizeInBits(B->getType())) { + if (SE.getTypeSizeInBits(A->getType()) > + SE.getTypeSizeInBits(B->getType())) + B = SE.getSignExtendExpr(B, A->getType()); + else + A = SE.getSignExtendExpr(A, B->getType()); + } + if (const SCEVConstant *D = + dyn_cast_or_null<SCEVConstant>(getSDiv(B, A, SE))) { + // Stride of one or negative one can have reuse with non-addresses. + if (D->getValue()->isOne() || + D->getValue()->isAllOnesValue()) + goto decline_post_inc; + // Avoid weird situations. + if (D->getValue()->getValue().getMinSignedBits() >= 64 || + D->getValue()->getValue().isMinSignedValue()) + goto decline_post_inc; + // Without TLI, assume that any stride might be valid, and so any + // use might be shared. + if (!TLI) + goto decline_post_inc; + // Check for possible scaled-address reuse. + const Type *AccessTy = getAccessType(UI->getUser()); + TargetLowering::AddrMode AM; + AM.Scale = D->getValue()->getSExtValue(); + if (TLI->isLegalAddressingMode(AM, AccessTy)) + goto decline_post_inc; + AM.Scale = -AM.Scale; + if (TLI->isLegalAddressingMode(AM, AccessTy)) + goto decline_post_inc; + } + } - const Type *SrcTy = PH->getType(); - int Mantissa = DestTy->getFPMantissaWidth(); - if (Mantissa == -1) continue; - if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa) - continue; + DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " + << *Cond << '\n'); - unsigned Entry, Latch; - if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { - Entry = 0; - Latch = 1; + // It's possible for the setcc instruction to be anywhere in the loop, and + // possible for it to have multiple users. If it is not immediately before + // the exiting block branch, move it. + if (&*++BasicBlock::iterator(Cond) != TermBr) { + if (Cond->hasOneUse()) { + Cond->moveBefore(TermBr); } else { - Entry = 1; - Latch = 0; + // Clone the terminating condition and insert into the loopend. + ICmpInst *OldCond = Cond; + Cond = cast<ICmpInst>(Cond->clone()); + Cond->setName(L->getHeader()->getName() + ".termcond"); + ExitingBlock->getInstList().insert(TermBr, Cond); + + // Clone the IVUse, as the old use still exists! + CondUse = &IU.AddUser(CondUse->getStride(), CondUse->getOffset(), + Cond, CondUse->getOperandValToReplace()); + TermBr->replaceUsesOfWith(OldCond, Cond); } + } - ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); - if (!Init) continue; - Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue()); + // If we get to here, we know that we can transform the setcc instruction to + // use the post-incremented version of the IV, allowing us to coalesce the + // live ranges for the IV correctly. + CondUse->setOffset(SE.getMinusSCEV(CondUse->getOffset(), + CondUse->getStride())); + CondUse->setIsUseOfPostIncrementedValue(true); + Changed = true; - BinaryOperator *Incr = - dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); - if (!Incr) continue; - if (Incr->getOpcode() != Instruction::Add - && Incr->getOpcode() != Instruction::Sub) - continue; + PostIncs.insert(Cond); + decline_post_inc:; + } - /* Initialize new IV, double d = 0.0 in above example. */ - ConstantInt *C = NULL; - if (Incr->getOperand(0) == PH) - C = dyn_cast<ConstantInt>(Incr->getOperand(1)); - else if (Incr->getOperand(1) == PH) - C = dyn_cast<ConstantInt>(Incr->getOperand(0)); - else - continue; + // Determine an insertion point for the loop induction variable increment. It + // must dominate all the post-inc comparisons we just set up, and it must + // dominate the loop latch edge. + IVIncInsertPos = L->getLoopLatch()->getTerminator(); + for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), + E = PostIncs.end(); I != E; ++I) { + BasicBlock *BB = + DT.findNearestCommonDominator(IVIncInsertPos->getParent(), + (*I)->getParent()); + if (BB == (*I)->getParent()) + IVIncInsertPos = *I; + else if (BB != IVIncInsertPos->getParent()) + IVIncInsertPos = BB->getTerminator(); + } - if (!C) continue; + return Changed; +} + +bool +LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, + LSRUse::KindType Kind, const Type *AccessTy) { + int64_t NewMinOffset = LU.MinOffset; + int64_t NewMaxOffset = LU.MaxOffset; + const Type *NewAccessTy = AccessTy; + + // Check for a mismatched kind. It's tempting to collapse mismatched kinds to + // something conservative, however this can pessimize in the case that one of + // the uses will have all its uses outside the loop, for example. + if (LU.Kind != Kind) + return false; + // Conservatively assume HasBaseReg is true for now. + if (NewOffset < LU.MinOffset) { + if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true, + Kind, AccessTy, TLI, SE)) + return false; + NewMinOffset = NewOffset; + } else if (NewOffset > LU.MaxOffset) { + if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true, + Kind, AccessTy, TLI, SE)) + return false; + NewMaxOffset = NewOffset; + } + // Check for a mismatched access type, and fall back conservatively as needed. + if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) + NewAccessTy = Type::getVoidTy(AccessTy->getContext()); + + // Update the use. + LU.MinOffset = NewMinOffset; + LU.MaxOffset = NewMaxOffset; + LU.AccessTy = NewAccessTy; + if (NewOffset != LU.Offsets.back()) + LU.Offsets.push_back(NewOffset); + return true; +} - // Ignore negative constants, as the code below doesn't handle them - // correctly. TODO: Remove this restriction. - if (!C->getValue().isStrictlyPositive()) continue; +/// getUse - Return an LSRUse index and an offset value for a fixup which +/// needs the given expression, with the given kind and optional access type. +/// Either reuse an exisitng use or create a new one, as needed. +std::pair<size_t, int64_t> +LSRInstance::getUse(const SCEV *&Expr, + LSRUse::KindType Kind, const Type *AccessTy) { + const SCEV *Copy = Expr; + int64_t Offset = ExtractImmediate(Expr, SE); + + // Basic uses can't accept any offset, for example. + if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, + Kind, AccessTy, TLI, SE)) { + Expr = Copy; + Offset = 0; + } - /* Add new PHINode. */ - PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH); + std::pair<UseMapTy::iterator, bool> P = + UseMap.insert(std::make_pair(Expr, 0)); + if (!P.second) { + // A use already existed with this base. + size_t LUIdx = P.first->second; + LSRUse &LU = Uses[LUIdx]; + if (reconcileNewOffset(LU, Offset, Kind, AccessTy)) + // Reuse this use. + return std::make_pair(LUIdx, Offset); + } - /* create new increment. '++d' in above example. */ - Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); - BinaryOperator *NewIncr = - BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? - Instruction::FAdd : Instruction::FSub, - NewPH, CFP, "IV.S.next.", Incr); + // Create a new use. + size_t LUIdx = Uses.size(); + P.first->second = LUIdx; + Uses.push_back(LSRUse(Kind, AccessTy)); + LSRUse &LU = Uses[LUIdx]; - NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); - NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); + // We don't need to track redundant offsets, but we don't need to go out + // of our way here to avoid them. + if (LU.Offsets.empty() || Offset != LU.Offsets.back()) + LU.Offsets.push_back(Offset); - /* Remove cast operation */ - ShadowUse->replaceAllUsesWith(NewPH); - ShadowUse->eraseFromParent(); - NumShadow++; - break; + LU.MinOffset = Offset; + LU.MaxOffset = Offset; + return std::make_pair(LUIdx, Offset); +} + +void LSRInstance::CollectInterestingTypesAndFactors() { + SmallSetVector<const SCEV *, 4> Strides; + + // Collect interesting types and factors. + for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { + const SCEV *Stride = UI->getStride(); + + // Collect interesting types. + Types.insert(SE.getEffectiveSCEVType(Stride->getType())); + + // Collect interesting factors. + for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = + Strides.begin(), SEnd = Strides.end(); NewStrideIter != SEnd; + ++NewStrideIter) { + const SCEV *OldStride = Stride; + const SCEV *NewStride = *NewStrideIter; + if (OldStride == NewStride) + continue; + + if (SE.getTypeSizeInBits(OldStride->getType()) != + SE.getTypeSizeInBits(NewStride->getType())) { + if (SE.getTypeSizeInBits(OldStride->getType()) > + SE.getTypeSizeInBits(NewStride->getType())) + NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); + else + OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); + } + if (const SCEVConstant *Factor = + dyn_cast_or_null<SCEVConstant>(getSDiv(NewStride, OldStride, + SE, true))) { + if (Factor->getValue()->getValue().getMinSignedBits() <= 64) + Factors.insert(Factor->getValue()->getValue().getSExtValue()); + } else if (const SCEVConstant *Factor = + dyn_cast_or_null<SCEVConstant>(getSDiv(OldStride, NewStride, + SE, true))) { + if (Factor->getValue()->getValue().getMinSignedBits() <= 64) + Factors.insert(Factor->getValue()->getValue().getSExtValue()); + } } + Strides.insert(Stride); } -} -/// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar -/// uses in the loop, look to see if we can eliminate some, in favor of using -/// common indvars for the different uses. -void LoopStrengthReduce::OptimizeIndvars(Loop *L) { - // TODO: implement optzns here. + // If all uses use the same type, don't bother looking for truncation-based + // reuse. + if (Types.size() == 1) + Types.clear(); - OptimizeShadowIV(L); + DEBUG(print_factors_and_types(dbgs())); } -bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L, - bool CheckPreInc) { - int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue(); - for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(IU->StrideOrder[i]); - const SCEV *Share = SI->first; - if (!isa<SCEVConstant>(SI->first) || Share == Stride) - continue; - int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue(); - if (SSInt == SInt) - return true; // This can definitely be reused. - if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0) - continue; - int64_t Scale = SSInt / SInt; - bool AllUsesAreAddresses = true; - bool AllUsesAreOutsideLoop = true; - std::vector<BasedUser> UsersToProcess; - const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L, - AllUsesAreAddresses, - AllUsesAreOutsideLoop, - UsersToProcess); - if (AllUsesAreAddresses && - ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) { - if (!CheckPreInc) - return true; - // Any pre-inc iv use? - IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share]; - for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(), - E = StrideUses.Users.end(); I != E; ++I) { - if (!I->isUseOfPostIncrementedValue()) - return true; +void LSRInstance::CollectFixupsAndInitialFormulae() { + for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { + // Record the uses. + LSRFixup &LF = getNewFixup(); + LF.UserInst = UI->getUser(); + LF.OperandValToReplace = UI->getOperandValToReplace(); + if (UI->isUseOfPostIncrementedValue()) + LF.PostIncLoop = L; + + LSRUse::KindType Kind = LSRUse::Basic; + const Type *AccessTy = 0; + if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { + Kind = LSRUse::Address; + AccessTy = getAccessType(LF.UserInst); + } + + const SCEV *S = IU.getCanonicalExpr(*UI); + + // Equality (== and !=) ICmps are special. We can rewrite (i == N) as + // (N - i == 0), and this allows (N - i) to be the expression that we work + // with rather than just N or i, so we can consider the register + // requirements for both N and i at the same time. Limiting this code to + // equality icmps is not a problem because all interesting loops use + // equality icmps, thanks to IndVarSimplify. + if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) + if (CI->isEquality()) { + // Swap the operands if needed to put the OperandValToReplace on the + // left, for consistency. + Value *NV = CI->getOperand(1); + if (NV == LF.OperandValToReplace) { + CI->setOperand(1, CI->getOperand(0)); + CI->setOperand(0, NV); + } + + // x == y --> x - y == 0 + const SCEV *N = SE.getSCEV(NV); + if (N->isLoopInvariant(L)) { + Kind = LSRUse::ICmpZero; + S = SE.getMinusSCEV(N, S); + } + + // -1 and the negations of all interesting strides (except the negation + // of -1) are now also interesting. + for (size_t i = 0, e = Factors.size(); i != e; ++i) + if (Factors[i] != -1) + Factors.insert(-(uint64_t)Factors[i]); + Factors.insert(-1); } + + // Set up the initial formula for this use. + std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); + LF.LUIdx = P.first; + LF.Offset = P.second; + LSRUse &LU = Uses[LF.LUIdx]; + LU.AllFixupsOutsideLoop &= !L->contains(LF.UserInst); + + // If this is the first use of this LSRUse, give it a formula. + if (LU.Formulae.empty()) { + InsertInitialFormula(S, L, LU, LF.LUIdx); + CountRegisters(LU.Formulae.back(), LF.LUIdx); } } - return false; + + DEBUG(print_fixups(dbgs())); } -/// isUsedByExitBranch - Return true if icmp is used by a loop terminating -/// conditional branch or it's and / or with other conditions before being used -/// as the condition. -static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) { - BasicBlock *CondBB = Cond->getParent(); - if (!L->isLoopExiting(CondBB)) - return false; - BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator()); - if (!TermBr || !TermBr->isConditional()) +void +LSRInstance::InsertInitialFormula(const SCEV *S, Loop *L, + LSRUse &LU, size_t LUIdx) { + Formula F; + F.InitialMatch(S, L, SE, DT); + bool Inserted = InsertFormula(LU, LUIdx, F); + assert(Inserted && "Initial formula already exists!"); (void)Inserted; +} + +void +LSRInstance::InsertSupplementalFormula(const SCEV *S, + LSRUse &LU, size_t LUIdx) { + Formula F; + F.BaseRegs.push_back(S); + F.AM.HasBaseReg = true; + bool Inserted = InsertFormula(LU, LUIdx, F); + assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; +} + +/// CountRegisters - Note which registers are used by the given formula, +/// updating RegUses. +void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { + if (F.ScaledReg) + RegUses.CountRegister(F.ScaledReg, LUIdx); + for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), + E = F.BaseRegs.end(); I != E; ++I) + RegUses.CountRegister(*I, LUIdx); +} + +/// InsertFormula - If the given formula has not yet been inserted, add it to +/// the list, and return true. Return false otherwise. +bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { + if (!LU.InsertFormula(LUIdx, F)) return false; - Value *User = *Cond->use_begin(); - Instruction *UserInst = dyn_cast<Instruction>(User); - while (UserInst && - (UserInst->getOpcode() == Instruction::And || - UserInst->getOpcode() == Instruction::Or)) { - if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB) - return false; - User = *User->use_begin(); - UserInst = dyn_cast<Instruction>(User); + CountRegisters(F, LUIdx); + return true; +} + +/// CollectLoopInvariantFixupsAndFormulae - Check for other uses of +/// loop-invariant values which we're tracking. These other uses will pin these +/// values in registers, making them less profitable for elimination. +/// TODO: This currently misses non-constant addrec step registers. +/// TODO: Should this give more weight to users inside the loop? +void +LSRInstance::CollectLoopInvariantFixupsAndFormulae() { + SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); + SmallPtrSet<const SCEV *, 8> Inserted; + + while (!Worklist.empty()) { + const SCEV *S = Worklist.pop_back_val(); + + if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) + Worklist.insert(Worklist.end(), N->op_begin(), N->op_end()); + else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) + Worklist.push_back(C->getOperand()); + else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { + Worklist.push_back(D->getLHS()); + Worklist.push_back(D->getRHS()); + } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { + if (!Inserted.insert(U)) continue; + const Value *V = U->getValue(); + if (const Instruction *Inst = dyn_cast<Instruction>(V)) + if (L->contains(Inst)) continue; + for (Value::use_const_iterator UI = V->use_begin(), UE = V->use_end(); + UI != UE; ++UI) { + const Instruction *UserInst = dyn_cast<Instruction>(*UI); + // Ignore non-instructions. + if (!UserInst) + continue; + // Ignore instructions in other functions (as can happen with + // Constants). + if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) + continue; + // Ignore instructions not dominated by the loop. + const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? + UserInst->getParent() : + cast<PHINode>(UserInst)->getIncomingBlock( + PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); + if (!DT.dominates(L->getHeader(), UseBB)) + continue; + // Ignore uses which are part of other SCEV expressions, to avoid + // analyzing them multiple times. + if (SE.isSCEVable(UserInst->getType()) && + !isa<SCEVUnknown>(SE.getSCEV(const_cast<Instruction *>(UserInst)))) + continue; + // Ignore icmp instructions which are already being analyzed. + if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { + unsigned OtherIdx = !UI.getOperandNo(); + Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); + if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L)) + continue; + } + + LSRFixup &LF = getNewFixup(); + LF.UserInst = const_cast<Instruction *>(UserInst); + LF.OperandValToReplace = UI.getUse(); + std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); + LF.LUIdx = P.first; + LF.Offset = P.second; + LSRUse &LU = Uses[LF.LUIdx]; + LU.AllFixupsOutsideLoop &= L->contains(LF.UserInst); + InsertSupplementalFormula(U, LU, LF.LUIdx); + CountRegisters(LU.Formulae.back(), Uses.size() - 1); + break; + } + } } - return User == TermBr; } -static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse, - ScalarEvolution *SE, Loop *L, - const TargetLowering *TLI = 0) { - if (!L->contains(Cond)) - return false; +/// CollectSubexprs - Split S into subexpressions which can be pulled out into +/// separate registers. If C is non-null, multiply each subexpression by C. +static void CollectSubexprs(const SCEV *S, const SCEVConstant *C, + SmallVectorImpl<const SCEV *> &Ops, + ScalarEvolution &SE) { + if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { + // Break out add operands. + for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); + I != E; ++I) + CollectSubexprs(*I, C, Ops, SE); + return; + } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { + // Split a non-zero base out of an addrec. + if (!AR->getStart()->isZero()) { + CollectSubexprs(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()), + AR->getStepRecurrence(SE), + AR->getLoop()), C, Ops, SE); + CollectSubexprs(AR->getStart(), C, Ops, SE); + return; + } + } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { + // Break (C * (a + b + c)) into C*a + C*b + C*c. + if (Mul->getNumOperands() == 2) + if (const SCEVConstant *Op0 = + dyn_cast<SCEVConstant>(Mul->getOperand(0))) { + CollectSubexprs(Mul->getOperand(1), + C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0, + Ops, SE); + return; + } + } - if (!isa<SCEVConstant>(CondUse->getOffset())) - return false; + // Otherwise use the value itself. + Ops.push_back(C ? SE.getMulExpr(C, S) : S); +} - // Handle only tests for equality for the moment. - if (!Cond->isEquality() || !Cond->hasOneUse()) - return false; - if (!isUsedByExitBranch(Cond, L)) - return false; +/// GenerateReassociations - Split out subexpressions from adds and the bases of +/// addrecs. +void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, + Formula Base, + unsigned Depth) { + // Arbitrarily cap recursion to protect compile time. + if (Depth >= 3) return; + + for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { + const SCEV *BaseReg = Base.BaseRegs[i]; + + SmallVector<const SCEV *, 8> AddOps; + CollectSubexprs(BaseReg, 0, AddOps, SE); + if (AddOps.size() == 1) continue; + + for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), + JE = AddOps.end(); J != JE; ++J) { + // Don't pull a constant into a register if the constant could be folded + // into an immediate field. + if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset, + Base.getNumRegs() > 1, + LU.Kind, LU.AccessTy, TLI, SE)) + continue; - Value *CondOp0 = Cond->getOperand(0); - const SCEV *IV = SE->getSCEV(CondOp0); - const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); - if (!AR || !AR->isAffine()) - return false; + // Collect all operands except *J. + SmallVector<const SCEV *, 8> InnerAddOps; + for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(), + KE = AddOps.end(); K != KE; ++K) + if (K != J) + InnerAddOps.push_back(*K); + + // Don't leave just a constant behind in a register if the constant could + // be folded into an immediate field. + if (InnerAddOps.size() == 1 && + isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset, + Base.getNumRegs() > 1, + LU.Kind, LU.AccessTy, TLI, SE)) + continue; - const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); - if (!SC || SC->getValue()->getSExtValue() < 0) - // If it's already counting down, don't do anything. - return false; + Formula F = Base; + F.BaseRegs[i] = SE.getAddExpr(InnerAddOps); + F.BaseRegs.push_back(*J); + if (InsertFormula(LU, LUIdx, F)) + // If that formula hadn't been seen before, recurse to find more like + // it. + GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); + } + } +} - // If the RHS of the comparison is not an loop invariant, the rewrite - // cannot be done. Also bail out if it's already comparing against a zero. - // If we are checking this before cmp stride optimization, check if it's - // comparing against a already legal immediate. - Value *RHS = Cond->getOperand(1); - ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS); - if (!L->isLoopInvariant(RHS) || - (RHSC && RHSC->isZero()) || - (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue()))) - return false; +/// GenerateCombinations - Generate a formula consisting of all of the +/// loop-dominating registers added into a single register. +void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, + Formula Base) { + // This method is only intersting on a plurality of registers. + if (Base.BaseRegs.size() <= 1) return; + + Formula F = Base; + F.BaseRegs.clear(); + SmallVector<const SCEV *, 4> Ops; + for (SmallVectorImpl<const SCEV *>::const_iterator + I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { + const SCEV *BaseReg = *I; + if (BaseReg->properlyDominates(L->getHeader(), &DT) && + !BaseReg->hasComputableLoopEvolution(L)) + Ops.push_back(BaseReg); + else + F.BaseRegs.push_back(BaseReg); + } + if (Ops.size() > 1) { + const SCEV *Sum = SE.getAddExpr(Ops); + // TODO: If Sum is zero, it probably means ScalarEvolution missed an + // opportunity to fold something. For now, just ignore such cases + // rather than procede with zero in a register. + if (!Sum->isZero()) { + F.BaseRegs.push_back(Sum); + (void)InsertFormula(LU, LUIdx, F); + } + } +} - // Make sure the IV is only used for counting. Value may be preinc or - // postinc; 2 uses in either case. - if (!CondOp0->hasNUses(2)) - return false; +/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. +void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, + Formula Base) { + // We can't add a symbolic offset if the address already contains one. + if (Base.AM.BaseGV) return; - return true; + for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { + const SCEV *G = Base.BaseRegs[i]; + GlobalValue *GV = ExtractSymbol(G, SE); + if (G->isZero() || !GV) + continue; + Formula F = Base; + F.AM.BaseGV = GV; + if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, + LU.Kind, LU.AccessTy, TLI)) + continue; + F.BaseRegs[i] = G; + (void)InsertFormula(LU, LUIdx, F); + } } -/// OptimizeLoopTermCond - Change loop terminating condition to use the -/// postinc iv when possible. -void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) { - BasicBlock *LatchBlock = L->getLoopLatch(); - bool LatchExit = L->isLoopExiting(LatchBlock); - SmallVector<BasicBlock*, 8> ExitingBlocks; - L->getExitingBlocks(ExitingBlocks); +/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. +void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, + Formula Base) { + // TODO: For now, just add the min and max offset, because it usually isn't + // worthwhile looking at everything inbetween. + SmallVector<int64_t, 4> Worklist; + Worklist.push_back(LU.MinOffset); + if (LU.MaxOffset != LU.MinOffset) + Worklist.push_back(LU.MaxOffset); + + for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { + const SCEV *G = Base.BaseRegs[i]; + + for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), + E = Worklist.end(); I != E; ++I) { + Formula F = Base; + F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I; + if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I, + LU.Kind, LU.AccessTy, TLI)) { + F.BaseRegs[i] = SE.getAddExpr(G, SE.getIntegerSCEV(*I, G->getType())); + + (void)InsertFormula(LU, LUIdx, F); + } + } - for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { - BasicBlock *ExitingBlock = ExitingBlocks[i]; + int64_t Imm = ExtractImmediate(G, SE); + if (G->isZero() || Imm == 0) + continue; + Formula F = Base; + F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm; + if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, + LU.Kind, LU.AccessTy, TLI)) + continue; + F.BaseRegs[i] = G; + (void)InsertFormula(LU, LUIdx, F); + } +} - // Finally, get the terminating condition for the loop if possible. If we - // can, we want to change it to use a post-incremented version of its - // induction variable, to allow coalescing the live ranges for the IV into - // one register value. +/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up +/// the comparison. For example, x == y -> x*c == y*c. +void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, + Formula Base) { + if (LU.Kind != LSRUse::ICmpZero) return; - BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); - if (!TermBr) + // Determine the integer type for the base formula. + const Type *IntTy = Base.getType(); + if (!IntTy) return; + if (SE.getTypeSizeInBits(IntTy) > 64) return; + + // Don't do this if there is more than one offset. + if (LU.MinOffset != LU.MaxOffset) return; + + assert(!Base.AM.BaseGV && "ICmpZero use is not legal!"); + + // Check each interesting stride. + for (SmallSetVector<int64_t, 8>::const_iterator + I = Factors.begin(), E = Factors.end(); I != E; ++I) { + int64_t Factor = *I; + Formula F = Base; + + // Check that the multiplication doesn't overflow. + F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor; + if ((int64_t)F.AM.BaseOffs / Factor != Base.AM.BaseOffs) continue; - // FIXME: Overly conservative, termination condition could be an 'or' etc.. - if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) + + // Check that multiplying with the use offset doesn't overflow. + int64_t Offset = LU.MinOffset; + Offset = (uint64_t)Offset * Factor; + if ((int64_t)Offset / Factor != LU.MinOffset) continue; - // Search IVUsesByStride to find Cond's IVUse if there is one. - IVStrideUse *CondUse = 0; - const SCEV *CondStride = 0; - ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); - if (!FindIVUserForCond(Cond, CondUse, CondStride)) + // Check that this scale is legal. + if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI)) continue; - // If the latch block is exiting and it's not a single block loop, it's - // not safe to use postinc iv in other exiting blocks. FIXME: overly - // conservative? How about icmp stride optimization? - bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock); - if (UsePostInc && ExitingBlock != LatchBlock) { - if (!Cond->hasOneUse()) - // See below, we don't want the condition to be cloned. - UsePostInc = false; - else { - // If exiting block is the latch block, we know it's safe and profitable - // to transform the icmp to use post-inc iv. Otherwise do so only if it - // would not reuse another iv and its iv would be reused by other uses. - // We are optimizing for the case where the icmp is the only use of the - // iv. - IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride]; - for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(), - E = StrideUses.Users.end(); I != E; ++I) { - if (I->getUser() == Cond) - continue; - if (!I->isUseOfPostIncrementedValue()) { - UsePostInc = false; - break; - } + // Compensate for the use having MinOffset built into it. + F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset; + + const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy); + + // Check that multiplying with each base register doesn't overflow. + for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { + F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); + if (getSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) + goto next; + } + + // Check that multiplying with the scaled register doesn't overflow. + if (F.ScaledReg) { + F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); + if (getSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) + continue; + } + + // If we make it here and it's legal, add it. + (void)InsertFormula(LU, LUIdx, F); + next:; + } +} + +/// GenerateScales - Generate stride factor reuse formulae by making use of +/// scaled-offset address modes, for example. +void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, + Formula Base) { + // Determine the integer type for the base formula. + const Type *IntTy = Base.getType(); + if (!IntTy) return; + + // If this Formula already has a scaled register, we can't add another one. + if (Base.AM.Scale != 0) return; + + // Check each interesting stride. + for (SmallSetVector<int64_t, 8>::const_iterator + I = Factors.begin(), E = Factors.end(); I != E; ++I) { + int64_t Factor = *I; + + Base.AM.Scale = Factor; + Base.AM.HasBaseReg = Base.BaseRegs.size() > 1; + // Check whether this scale is going to be legal. + if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, + LU.Kind, LU.AccessTy, TLI)) { + // As a special-case, handle special out-of-loop Basic users specially. + // TODO: Reconsider this special case. + if (LU.Kind == LSRUse::Basic && + isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, + LSRUse::Special, LU.AccessTy, TLI) && + LU.AllFixupsOutsideLoop) + LU.Kind = LSRUse::Special; + else + continue; + } + // For an ICmpZero, negating a solitary base register won't lead to + // new solutions. + if (LU.Kind == LSRUse::ICmpZero && + !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV) + continue; + // For each addrec base reg, apply the scale, if possible. + for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) + if (const SCEVAddRecExpr *AR = + dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { + const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy); + if (FactorS->isZero()) + continue; + // Divide out the factor, ignoring high bits, since we'll be + // scaling the value back up in the end. + if (const SCEV *Quotient = getSDiv(AR, FactorS, SE, true)) { + // TODO: This could be optimized to avoid all the copying. + Formula F = Base; + F.ScaledReg = Quotient; + std::swap(F.BaseRegs[i], F.BaseRegs.back()); + F.BaseRegs.pop_back(); + (void)InsertFormula(LU, LUIdx, F); } } + } +} - // If iv for the stride might be shared and any of the users use pre-inc - // iv might be used, then it's not safe to use post-inc iv. - if (UsePostInc && - isa<SCEVConstant>(CondStride) && - StrideMightBeShared(CondStride, L, true)) - UsePostInc = false; - } +/// GenerateTruncates - Generate reuse formulae from different IV types. +void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, + Formula Base) { + // This requires TargetLowering to tell us which truncates are free. + if (!TLI) return; + + // Don't bother truncating symbolic values. + if (Base.AM.BaseGV) return; + + // Determine the integer type for the base formula. + const Type *DstTy = Base.getType(); + if (!DstTy) return; + DstTy = SE.getEffectiveSCEVType(DstTy); + + for (SmallSetVector<const Type *, 4>::const_iterator + I = Types.begin(), E = Types.end(); I != E; ++I) { + const Type *SrcTy = *I; + if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) { + Formula F = Base; + + if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); + for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), + JE = F.BaseRegs.end(); J != JE; ++J) + *J = SE.getAnyExtendExpr(*J, SrcTy); + + // TODO: This assumes we've done basic processing on all uses and + // have an idea what the register usage is. + if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) + continue; - // If the trip count is computed in terms of a max (due to ScalarEvolution - // being unable to find a sufficient guard, for example), change the loop - // comparison to use SLT or ULT instead of NE. - Cond = OptimizeMax(L, Cond, CondUse); - - // If possible, change stride and operands of the compare instruction to - // eliminate one stride. However, avoid rewriting the compare instruction - // with an iv of new stride if it's likely the new stride uses will be - // rewritten using the stride of the compare instruction. - if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) { - // If the condition stride is a constant and it's the only use, we might - // want to optimize it first by turning it to count toward zero. - if (!StrideMightBeShared(CondStride, L, false) && - !ShouldCountToZero(Cond, CondUse, SE, L, TLI)) - Cond = ChangeCompareStride(L, Cond, CondUse, CondStride); + (void)InsertFormula(LU, LUIdx, F); } + } +} + +namespace { + +/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to +/// defer modifications so that the search phase doesn't have to worry about +/// the data structures moving underneath it. +struct WorkItem { + size_t LUIdx; + int64_t Imm; + const SCEV *OrigReg; + + WorkItem(size_t LI, int64_t I, const SCEV *R) + : LUIdx(LI), Imm(I), OrigReg(R) {} + + void print(raw_ostream &OS) const; + void dump() const; +}; + +} + +void WorkItem::print(raw_ostream &OS) const { + OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx + << " , add offset " << Imm; +} + +void WorkItem::dump() const { + print(errs()); errs() << '\n'; +} + +/// GenerateCrossUseConstantOffsets - Look for registers which are a constant +/// distance apart and try to form reuse opportunities between them. +void LSRInstance::GenerateCrossUseConstantOffsets() { + // Group the registers by their value without any added constant offset. + typedef std::map<int64_t, const SCEV *> ImmMapTy; + typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; + RegMapTy Map; + DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; + SmallVector<const SCEV *, 8> Sequence; + for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); + I != E; ++I) { + const SCEV *Reg = *I; + int64_t Imm = ExtractImmediate(Reg, SE); + std::pair<RegMapTy::iterator, bool> Pair = + Map.insert(std::make_pair(Reg, ImmMapTy())); + if (Pair.second) + Sequence.push_back(Reg); + Pair.first->second.insert(std::make_pair(Imm, *I)); + UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); + } - if (!UsePostInc) + // Now examine each set of registers with the same base value. Build up + // a list of work to do and do the work in a separate step so that we're + // not adding formulae and register counts while we're searching. + SmallVector<WorkItem, 32> WorkItems; + SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; + for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), + E = Sequence.end(); I != E; ++I) { + const SCEV *Reg = *I; + const ImmMapTy &Imms = Map.find(Reg)->second; + + // It's not worthwhile looking for reuse if there's only one offset. + if (Imms.size() == 1) continue; - DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " - << *Cond << '\n'); + DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; + for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); + J != JE; ++J) + dbgs() << ' ' << J->first; + dbgs() << '\n'); - // It's possible for the setcc instruction to be anywhere in the loop, and - // possible for it to have multiple users. If it is not immediately before - // the exiting block branch, move it. - if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) { - if (Cond->hasOneUse()) { // Condition has a single use, just move it. - Cond->moveBefore(TermBr); - } else { - // Otherwise, clone the terminating condition and insert into the - // loopend. - Cond = cast<ICmpInst>(Cond->clone()); - Cond->setName(L->getHeader()->getName() + ".termcond"); - ExitingBlock->getInstList().insert(TermBr, Cond); + // Examine each offset. + for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); + J != JE; ++J) { + const SCEV *OrigReg = J->second; - // Clone the IVUse, as the old use still exists! - IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond, - CondUse->getOperandValToReplace()); - CondUse = &IU->IVUsesByStride[CondStride]->Users.back(); + int64_t JImm = J->first; + const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); + + if (!isa<SCEVConstant>(OrigReg) && + UsedByIndicesMap[Reg].count() == 1) { + DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); + continue; + } + + // Conservatively examine offsets between this orig reg a few selected + // other orig regs. + ImmMapTy::const_iterator OtherImms[] = { + Imms.begin(), prior(Imms.end()), + Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) + }; + for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { + ImmMapTy::const_iterator M = OtherImms[i]; + if (M == J || M == JE) continue; + + // Compute the difference between the two. + int64_t Imm = (uint64_t)JImm - M->first; + for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; + LUIdx = UsedByIndices.find_next(LUIdx)) + // Make a memo of this use, offset, and register tuple. + if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) + WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); } } + } - // If we get to here, we know that we can transform the setcc instruction to - // use the post-incremented version of the IV, allowing us to coalesce the - // live ranges for the IV correctly. - CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride)); - CondUse->setIsUseOfPostIncrementedValue(true); - Changed = true; + Map.clear(); + Sequence.clear(); + UsedByIndicesMap.clear(); + UniqueItems.clear(); + + // Now iterate through the worklist and add new formulae. + for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), + E = WorkItems.end(); I != E; ++I) { + const WorkItem &WI = *I; + size_t LUIdx = WI.LUIdx; + LSRUse &LU = Uses[LUIdx]; + int64_t Imm = WI.Imm; + const SCEV *OrigReg = WI.OrigReg; + + const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); + const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); + unsigned BitWidth = SE.getTypeSizeInBits(IntTy); + + // TODO: Use a more targetted data structure. + for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { + Formula F = LU.Formulae[L]; + // Use the immediate in the scaled register. + if (F.ScaledReg == OrigReg) { + int64_t Offs = (uint64_t)F.AM.BaseOffs + + Imm * (uint64_t)F.AM.Scale; + // Don't create 50 + reg(-50). + if (F.referencesReg(SE.getSCEV( + ConstantInt::get(IntTy, -(uint64_t)Offs)))) + continue; + Formula NewF = F; + NewF.AM.BaseOffs = Offs; + if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, + LU.Kind, LU.AccessTy, TLI)) + continue; + NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); + + // If the new scale is a constant in a register, and adding the constant + // value to the immediate would produce a value closer to zero than the + // immediate itself, then the formula isn't worthwhile. + if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) + if (C->getValue()->getValue().isNegative() != + (NewF.AM.BaseOffs < 0) && + (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale)) + .ule(APInt(BitWidth, NewF.AM.BaseOffs).abs())) + continue; - ++NumLoopCond; + // OK, looks good. + (void)InsertFormula(LU, LUIdx, NewF); + } else { + // Use the immediate in a base register. + for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { + const SCEV *BaseReg = F.BaseRegs[N]; + if (BaseReg != OrigReg) + continue; + Formula NewF = F; + NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm; + if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, + LU.Kind, LU.AccessTy, TLI)) + continue; + NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); + + // If the new formula has a constant in a register, and adding the + // constant value to the immediate would produce a value closer to + // zero than the immediate itself, then the formula isn't worthwhile. + for (SmallVectorImpl<const SCEV *>::const_iterator + J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); + J != JE; ++J) + if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) + if (C->getValue()->getValue().isNegative() != + (NewF.AM.BaseOffs < 0) && + C->getValue()->getValue().abs() + .ule(APInt(BitWidth, NewF.AM.BaseOffs).abs())) + goto skip_formula; + + // Ok, looks good. + (void)InsertFormula(LU, LUIdx, NewF); + break; + skip_formula:; + } + } + } } } -bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride, - IVStrideUse* &CondUse, - Loop *L) { - // If the only use is an icmp of a loop exiting conditional branch, then - // attempt the optimization. - BasedUser User = BasedUser(*CondUse, SE); - assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!"); - ICmpInst *Cond = cast<ICmpInst>(User.Inst); +/// GenerateAllReuseFormulae - Generate formulae for each use. +void +LSRInstance::GenerateAllReuseFormulae() { + // This is split into multiple loops so that hasRegsUsedByUsesOtherThan + // queries are more precise. + for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { + LSRUse &LU = Uses[LUIdx]; + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateReassociations(LU, LUIdx, LU.Formulae[i]); + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateCombinations(LU, LUIdx, LU.Formulae[i]); + } + for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { + LSRUse &LU = Uses[LUIdx]; + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateScales(LU, LUIdx, LU.Formulae[i]); + } + for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { + LSRUse &LU = Uses[LUIdx]; + for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) + GenerateTruncates(LU, LUIdx, LU.Formulae[i]); + } - // Less strict check now that compare stride optimization is done. - if (!ShouldCountToZero(Cond, CondUse, SE, L)) - return false; + GenerateCrossUseConstantOffsets(); +} - Value *CondOp0 = Cond->getOperand(0); - PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0); - Instruction *Incr; - if (!PHIExpr) { - // Value tested is postinc. Find the phi node. - Incr = dyn_cast<BinaryOperator>(CondOp0); - // FIXME: Just use User.OperandValToReplace here? - if (!Incr || Incr->getOpcode() != Instruction::Add) - return false; +/// If their are multiple formulae with the same set of registers used +/// by other uses, pick the best one and delete the others. +void LSRInstance::FilterOutUndesirableDedicatedRegisters() { +#ifndef NDEBUG + bool Changed = false; +#endif + + // Collect the best formula for each unique set of shared registers. This + // is reset for each use. + typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo> + BestFormulaeTy; + BestFormulaeTy BestFormulae; + + for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { + LSRUse &LU = Uses[LUIdx]; + FormulaSorter Sorter(L, LU, SE, DT); + + // Clear out the set of used regs; it will be recomputed. + LU.Regs.clear(); + + for (size_t FIdx = 0, NumForms = LU.Formulae.size(); + FIdx != NumForms; ++FIdx) { + Formula &F = LU.Formulae[FIdx]; + + SmallVector<const SCEV *, 2> Key; + for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), + JE = F.BaseRegs.end(); J != JE; ++J) { + const SCEV *Reg = *J; + if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) + Key.push_back(Reg); + } + if (F.ScaledReg && + RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) + Key.push_back(F.ScaledReg); + // Unstable sort by host order ok, because this is only used for + // uniquifying. + std::sort(Key.begin(), Key.end()); + + std::pair<BestFormulaeTy::const_iterator, bool> P = + BestFormulae.insert(std::make_pair(Key, FIdx)); + if (!P.second) { + Formula &Best = LU.Formulae[P.first->second]; + if (Sorter.operator()(F, Best)) + std::swap(F, Best); + DEBUG(dbgs() << "Filtering out "; F.print(dbgs()); + dbgs() << "\n" + " in favor of "; Best.print(dbgs()); + dbgs() << '\n'); +#ifndef NDEBUG + Changed = true; +#endif + std::swap(F, LU.Formulae.back()); + LU.Formulae.pop_back(); + --FIdx; + --NumForms; + continue; + } + if (F.ScaledReg) LU.Regs.insert(F.ScaledReg); + LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); + } + BestFormulae.clear(); + } - PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0)); - if (!PHIExpr) - return false; - // 1 use for preinc value, the increment. - if (!PHIExpr->hasOneUse()) - return false; - } else { - assert(isa<PHINode>(CondOp0) && - "Unexpected loop exiting counting instruction sequence!"); - PHIExpr = cast<PHINode>(CondOp0); - // Value tested is preinc. Find the increment. - // A CmpInst is not a BinaryOperator; we depend on this. - Instruction::use_iterator UI = PHIExpr->use_begin(); - Incr = dyn_cast<BinaryOperator>(UI); - if (!Incr) - Incr = dyn_cast<BinaryOperator>(++UI); - // One use for postinc value, the phi. Unnecessarily conservative? - if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add) - return false; + DEBUG(if (Changed) { + dbgs() << "\n" + "After filtering out undesirable candidates:\n"; + print_uses(dbgs()); + }); +} + +/// NarrowSearchSpaceUsingHeuristics - If there are an extrordinary number of +/// formulae to choose from, use some rough heuristics to prune down the number +/// of formulae. This keeps the main solver from taking an extrordinary amount +/// of time in some worst-case scenarios. +void LSRInstance::NarrowSearchSpaceUsingHeuristics() { + // This is a rough guess that seems to work fairly well. + const size_t Limit = UINT16_MAX; + + SmallPtrSet<const SCEV *, 4> Taken; + for (;;) { + // Estimate the worst-case number of solutions we might consider. We almost + // never consider this many solutions because we prune the search space, + // but the pruning isn't always sufficient. + uint32_t Power = 1; + for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), + E = Uses.end(); I != E; ++I) { + size_t FSize = I->Formulae.size(); + if (FSize >= Limit) { + Power = Limit; + break; + } + Power *= FSize; + if (Power >= Limit) + break; + } + if (Power < Limit) + break; + + // Ok, we have too many of formulae on our hands to conveniently handle. + // Use a rough heuristic to thin out the list. + + // Pick the register which is used by the most LSRUses, which is likely + // to be a good reuse register candidate. + const SCEV *Best = 0; + unsigned BestNum = 0; + for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); + I != E; ++I) { + const SCEV *Reg = *I; + if (Taken.count(Reg)) + continue; + if (!Best) + Best = Reg; + else { + unsigned Count = RegUses.getUsedByIndices(Reg).count(); + if (Count > BestNum) { + Best = Reg; + BestNum = Count; + } + } + } + + DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best + << " will yeild profitable reuse.\n"); + Taken.insert(Best); + + // In any use with formulae which references this register, delete formulae + // which don't reference it. + for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(), + E = Uses.end(); I != E; ++I) { + LSRUse &LU = *I; + if (!LU.Regs.count(Best)) continue; + + // Clear out the set of used regs; it will be recomputed. + LU.Regs.clear(); + + for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { + Formula &F = LU.Formulae[i]; + if (!F.referencesReg(Best)) { + DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); + std::swap(LU.Formulae.back(), F); + LU.Formulae.pop_back(); + --e; + --i; + continue; + } + + if (F.ScaledReg) LU.Regs.insert(F.ScaledReg); + LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); + } + } + + DEBUG(dbgs() << "After pre-selection:\n"; + print_uses(dbgs())); } +} - // Replace the increment with a decrement. - DEBUG(dbgs() << "LSR: Examining use "); - DEBUG(WriteAsOperand(dbgs(), CondOp0, /*PrintType=*/false)); - DEBUG(dbgs() << " in Inst: " << *Cond << '\n'); - BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub, - Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr); - Incr->replaceAllUsesWith(Decr); - Incr->eraseFromParent(); - - // Substitute endval-startval for the original startval, and 0 for the - // original endval. Since we're only testing for equality this is OK even - // if the computation wraps around. - BasicBlock *Preheader = L->getLoopPreheader(); - Instruction *PreInsertPt = Preheader->getTerminator(); - unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0; - Value *StartVal = PHIExpr->getIncomingValue(InBlock); - Value *EndVal = Cond->getOperand(1); - DEBUG(dbgs() << " Optimize loop counting iv to count down [" - << *EndVal << " .. " << *StartVal << "]\n"); - - // FIXME: check for case where both are constant. - Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0); - BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub, - EndVal, StartVal, "tmp", PreInsertPt); - PHIExpr->setIncomingValue(InBlock, NewStartVal); - Cond->setOperand(1, Zero); - DEBUG(dbgs() << " New icmp: " << *Cond << "\n"); - - int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue(); - const SCEV *NewStride = 0; - bool Found = false; - for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { - const SCEV *OldStride = IU->StrideOrder[i]; - if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride)) - if (SC->getValue()->getSExtValue() == -SInt) { - Found = true; - NewStride = OldStride; +/// SolveRecurse - This is the recursive solver. +void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, + Cost &SolutionCost, + SmallVectorImpl<const Formula *> &Workspace, + const Cost &CurCost, + const SmallPtrSet<const SCEV *, 16> &CurRegs, + DenseSet<const SCEV *> &VisitedRegs) const { + // Some ideas: + // - prune more: + // - use more aggressive filtering + // - sort the formula so that the most profitable solutions are found first + // - sort the uses too + // - search faster: + // - dont compute a cost, and then compare. compare while computing a cost + // and bail early. + // - track register sets with SmallBitVector + + const LSRUse &LU = Uses[Workspace.size()]; + + // If this use references any register that's already a part of the + // in-progress solution, consider it a requirement that a formula must + // reference that register in order to be considered. This prunes out + // unprofitable searching. + SmallSetVector<const SCEV *, 4> ReqRegs; + for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), + E = CurRegs.end(); I != E; ++I) + if (LU.Regs.count(*I)) + ReqRegs.insert(*I); + + bool AnySatisfiedReqRegs = false; + SmallPtrSet<const SCEV *, 16> NewRegs; + Cost NewCost; +retry: + for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), + E = LU.Formulae.end(); I != E; ++I) { + const Formula &F = *I; + + // Ignore formulae which do not use any of the required registers. + for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), + JE = ReqRegs.end(); J != JE; ++J) { + const SCEV *Reg = *J; + if ((!F.ScaledReg || F.ScaledReg != Reg) && + std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == + F.BaseRegs.end()) + goto skip; + } + AnySatisfiedReqRegs = true; + + // Evaluate the cost of the current formula. If it's already worse than + // the current best, prune the search at that point. + NewCost = CurCost; + NewRegs = CurRegs; + NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); + if (NewCost < SolutionCost) { + Workspace.push_back(&F); + if (Workspace.size() != Uses.size()) { + SolveRecurse(Solution, SolutionCost, Workspace, NewCost, + NewRegs, VisitedRegs); + if (F.getNumRegs() == 1 && Workspace.size() == 1) + VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); + } else { + DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); + dbgs() << ". Regs:"; + for (SmallPtrSet<const SCEV *, 16>::const_iterator + I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) + dbgs() << ' ' << **I; + dbgs() << '\n'); + + SolutionCost = NewCost; + Solution = Workspace; + } + Workspace.pop_back(); + } + skip:; + } + + // If none of the formulae had all of the required registers, relax the + // constraint so that we don't exclude all formulae. + if (!AnySatisfiedReqRegs) { + ReqRegs.clear(); + goto retry; + } +} + +void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { + SmallVector<const Formula *, 8> Workspace; + Cost SolutionCost; + SolutionCost.Loose(); + Cost CurCost; + SmallPtrSet<const SCEV *, 16> CurRegs; + DenseSet<const SCEV *> VisitedRegs; + Workspace.reserve(Uses.size()); + + SolveRecurse(Solution, SolutionCost, Workspace, CurCost, + CurRegs, VisitedRegs); + + // Ok, we've now made all our decisions. + DEBUG(dbgs() << "\n" + "The chosen solution requires "; SolutionCost.print(dbgs()); + dbgs() << ":\n"; + for (size_t i = 0, e = Uses.size(); i != e; ++i) { + dbgs() << " "; + Uses[i].print(dbgs()); + dbgs() << "\n" + " "; + Solution[i]->print(dbgs()); + dbgs() << '\n'; + }); +} + +/// getImmediateDominator - A handy utility for the specific DominatorTree +/// query that we need here. +/// +static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) { + DomTreeNode *Node = DT.getNode(BB); + if (!Node) return 0; + Node = Node->getIDom(); + if (!Node) return 0; + return Node->getBlock(); +} + +Value *LSRInstance::Expand(const LSRFixup &LF, + const Formula &F, + BasicBlock::iterator IP, + Loop *L, Instruction *IVIncInsertPos, + SCEVExpander &Rewriter, + SmallVectorImpl<WeakVH> &DeadInsts, + ScalarEvolution &SE, DominatorTree &DT) const { + const LSRUse &LU = Uses[LF.LUIdx]; + + // Then, collect some instructions which we will remain dominated by when + // expanding the replacement. These must be dominated by any operands that + // will be required in the expansion. + SmallVector<Instruction *, 4> Inputs; + if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) + Inputs.push_back(I); + if (LU.Kind == LSRUse::ICmpZero) + if (Instruction *I = + dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) + Inputs.push_back(I); + if (LF.PostIncLoop && !L->contains(LF.UserInst)) + Inputs.push_back(L->getLoopLatch()->getTerminator()); + + // Then, climb up the immediate dominator tree as far as we can go while + // still being dominated by the input positions. + for (;;) { + bool AllDominate = true; + Instruction *BetterPos = 0; + BasicBlock *IDom = getImmediateDominator(IP->getParent(), DT); + if (!IDom) break; + Instruction *Tentative = IDom->getTerminator(); + for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), + E = Inputs.end(); I != E; ++I) { + Instruction *Inst = *I; + if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { + AllDominate = false; break; } + if (IDom == Inst->getParent() && + (!BetterPos || DT.dominates(BetterPos, Inst))) + BetterPos = next(BasicBlock::iterator(Inst)); + } + if (!AllDominate) + break; + if (BetterPos) + IP = BetterPos; + else + IP = Tentative; } + while (isa<PHINode>(IP)) ++IP; + + // Inform the Rewriter if we have a post-increment use, so that it can + // perform an advantageous expansion. + Rewriter.setPostInc(LF.PostIncLoop); + + // This is the type that the user actually needs. + const Type *OpTy = LF.OperandValToReplace->getType(); + // This will be the type that we'll initially expand to. + const Type *Ty = F.getType(); + if (!Ty) + // No type known; just expand directly to the ultimate type. + Ty = OpTy; + else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) + // Expand directly to the ultimate type if it's the right size. + Ty = OpTy; + // This is the type to do integer arithmetic in. + const Type *IntTy = SE.getEffectiveSCEVType(Ty); + + // Build up a list of operands to add together to form the full base. + SmallVector<const SCEV *, 8> Ops; + + // Expand the BaseRegs portion. + for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), + E = F.BaseRegs.end(); I != E; ++I) { + const SCEV *Reg = *I; + assert(!Reg->isZero() && "Zero allocated in a base register!"); + + // If we're expanding for a post-inc user for the add-rec's loop, make the + // post-inc adjustment. + const SCEV *Start = Reg; + while (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Start)) { + if (AR->getLoop() == LF.PostIncLoop) { + Reg = SE.getAddExpr(Reg, AR->getStepRecurrence(SE)); + // If the user is inside the loop, insert the code after the increment + // so that it is dominated by its operand. + if (L->contains(LF.UserInst)) + IP = IVIncInsertPos; + break; + } + Start = AR->getStart(); + } - if (!Found) - NewStride = SE->getIntegerSCEV(-SInt, Stride->getType()); - IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0)); - IU->IVUsesByStride[Stride]->removeUser(CondUse); + Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); + } - CondUse = &IU->IVUsesByStride[NewStride]->Users.back(); - Stride = NewStride; + // Expand the ScaledReg portion. + Value *ICmpScaledV = 0; + if (F.AM.Scale != 0) { + const SCEV *ScaledS = F.ScaledReg; + + // If we're expanding for a post-inc user for the add-rec's loop, make the + // post-inc adjustment. + if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(ScaledS)) + if (AR->getLoop() == LF.PostIncLoop) + ScaledS = SE.getAddExpr(ScaledS, AR->getStepRecurrence(SE)); + + if (LU.Kind == LSRUse::ICmpZero) { + // An interesting way of "folding" with an icmp is to use a negated + // scale, which we'll implement by inserting it into the other operand + // of the icmp. + assert(F.AM.Scale == -1 && + "The only scale supported by ICmpZero uses is -1!"); + ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); + } else { + // Otherwise just expand the scaled register and an explicit scale, + // which is expected to be matched as part of the address. + ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); + ScaledS = SE.getMulExpr(ScaledS, + SE.getIntegerSCEV(F.AM.Scale, + ScaledS->getType())); + Ops.push_back(ScaledS); + } + } - ++NumCountZero; + // Expand the immediate portions. + if (F.AM.BaseGV) + Ops.push_back(SE.getSCEV(F.AM.BaseGV)); + int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset; + if (Offset != 0) { + if (LU.Kind == LSRUse::ICmpZero) { + // The other interesting way of "folding" with an ICmpZero is to use a + // negated immediate. + if (!ICmpScaledV) + ICmpScaledV = ConstantInt::get(IntTy, -Offset); + else { + Ops.push_back(SE.getUnknown(ICmpScaledV)); + ICmpScaledV = ConstantInt::get(IntTy, Offset); + } + } else { + // Just add the immediate values. These again are expected to be matched + // as part of the address. + Ops.push_back(SE.getIntegerSCEV(Offset, IntTy)); + } + } - return true; + // Emit instructions summing all the operands. + const SCEV *FullS = Ops.empty() ? + SE.getIntegerSCEV(0, IntTy) : + SE.getAddExpr(Ops); + Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); + + // We're done expanding now, so reset the rewriter. + Rewriter.setPostInc(0); + + // An ICmpZero Formula represents an ICmp which we're handling as a + // comparison against zero. Now that we've expanded an expression for that + // form, update the ICmp's other operand. + if (LU.Kind == LSRUse::ICmpZero) { + ICmpInst *CI = cast<ICmpInst>(LF.UserInst); + DeadInsts.push_back(CI->getOperand(1)); + assert(!F.AM.BaseGV && "ICmp does not support folding a global value and " + "a scale at the same time!"); + if (F.AM.Scale == -1) { + if (ICmpScaledV->getType() != OpTy) { + Instruction *Cast = + CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, + OpTy, false), + ICmpScaledV, OpTy, "tmp", CI); + ICmpScaledV = Cast; + } + CI->setOperand(1, ICmpScaledV); + } else { + assert(F.AM.Scale == 0 && + "ICmp does not support folding a global value and " + "a scale at the same time!"); + Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), + -(uint64_t)Offset); + if (C->getType() != OpTy) + C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, + OpTy, false), + C, OpTy); + + CI->setOperand(1, C); + } + } + + return FullV; } -/// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding -/// when to exit the loop is used only for that purpose, try to rearrange things -/// so it counts down to a test against zero. -bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) { - bool ThisChanged = false; - for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { - const SCEV *Stride = IU->StrideOrder[i]; - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = - IU->IVUsesByStride.find(Stride); - assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - // FIXME: Generalize to non-affine IV's. - if (!SI->first->isLoopInvariant(L)) - continue; - // If stride is a constant and it has an icmpinst use, check if we can - // optimize the loop to count down. - if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) { - Instruction *User = SI->second->Users.begin()->getUser(); - if (!isa<ICmpInst>(User)) - continue; - const SCEV *CondStride = Stride; - IVStrideUse *Use = &*SI->second->Users.begin(); - if (!OptimizeLoopCountIVOfStride(CondStride, Use, L)) - continue; - ThisChanged = true; +/// Rewrite - Emit instructions for the leading candidate expression for this +/// LSRUse (this is called "expanding"), and update the UserInst to reference +/// the newly expanded value. +void LSRInstance::Rewrite(const LSRFixup &LF, + const Formula &F, + Loop *L, Instruction *IVIncInsertPos, + SCEVExpander &Rewriter, + SmallVectorImpl<WeakVH> &DeadInsts, + ScalarEvolution &SE, DominatorTree &DT, + Pass *P) const { + const Type *OpTy = LF.OperandValToReplace->getType(); + + // First, find an insertion point that dominates UserInst. For PHI nodes, + // find the nearest block which dominates all the relevant uses. + if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { + DenseMap<BasicBlock *, Value *> Inserted; + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) + if (PN->getIncomingValue(i) == LF.OperandValToReplace) { + BasicBlock *BB = PN->getIncomingBlock(i); - // Now check if it's possible to reuse this iv for other stride uses. - for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) { - const SCEV *SStride = IU->StrideOrder[j]; - if (SStride == CondStride) - continue; - std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII = - IU->IVUsesByStride.find(SStride); - assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!"); - // FIXME: Generalize to non-affine IV's. - if (!SII->first->isLoopInvariant(L)) - continue; - // FIXME: Rewrite other stride using CondStride. + // If this is a critical edge, split the edge so that we do not insert + // the code on all predecessor/successor paths. We do this unless this + // is the canonical backedge for this loop, which complicates post-inc + // users. + if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && + !isa<IndirectBrInst>(BB->getTerminator()) && + (PN->getParent() != L->getHeader() || !L->contains(BB))) { + // Split the critical edge. + BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P); + + // 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); + } + + std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = + Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); + if (!Pair.second) + PN->setIncomingValue(i, Pair.first->second); + else { + Value *FullV = Expand(LF, F, BB->getTerminator(), L, IVIncInsertPos, + Rewriter, DeadInsts, SE, DT); + + // If this is reuse-by-noop-cast, insert the noop cast. + if (FullV->getType() != OpTy) + FullV = + CastInst::Create(CastInst::getCastOpcode(FullV, false, + OpTy, false), + FullV, LF.OperandValToReplace->getType(), + "tmp", BB->getTerminator()); + + PN->setIncomingValue(i, FullV); + Pair.first->second = FullV; + } } + } else { + Value *FullV = Expand(LF, F, LF.UserInst, L, IVIncInsertPos, + Rewriter, DeadInsts, SE, DT); + + // If this is reuse-by-noop-cast, insert the noop cast. + if (FullV->getType() != OpTy) { + Instruction *Cast = + CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), + FullV, OpTy, "tmp", LF.UserInst); + FullV = Cast; } + + // Update the user. ICmpZero is handled specially here (for now) because + // Expand may have updated one of the operands of the icmp already, and + // its new value may happen to be equal to LF.OperandValToReplace, in + // which case doing replaceUsesOfWith leads to replacing both operands + // with the same value. TODO: Reorganize this. + if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) + LF.UserInst->setOperand(0, FullV); + else + LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); } - Changed |= ThisChanged; - return ThisChanged; + DeadInsts.push_back(LF.OperandValToReplace); } -bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) { - IU = &getAnalysis<IVUsers>(); - SE = &getAnalysis<ScalarEvolution>(); - Changed = false; +void +LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, + Pass *P) { + // Keep track of instructions we may have made dead, so that + // we can remove them after we are done working. + SmallVector<WeakVH, 16> DeadInsts; + + SCEVExpander Rewriter(SE); + Rewriter.disableCanonicalMode(); + Rewriter.setIVIncInsertPos(L, IVIncInsertPos); - // If LoopSimplify form is not available, stay out of trouble. - if (!L->getLoopPreheader() || !L->getLoopLatch()) - return false; + // Expand the new value definitions and update the users. + for (size_t i = 0, e = Fixups.size(); i != e; ++i) { + size_t LUIdx = Fixups[i].LUIdx; + + Rewrite(Fixups[i], *Solution[LUIdx], L, IVIncInsertPos, Rewriter, + DeadInsts, SE, DT, P); + + Changed = true; + } - if (!IU->IVUsesByStride.empty()) { - DEBUG(dbgs() << "\nLSR on \"" << L->getHeader()->getParent()->getName() - << "\" "; - L->print(dbgs())); + // Clean up after ourselves. This must be done before deleting any + // instructions. + Rewriter.clear(); - // Sort the StrideOrder so we process larger strides first. - std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(), - StrideCompare(SE)); + Changed |= DeleteTriviallyDeadInstructions(DeadInsts); +} - // Optimize induction variables. Some indvar uses can be transformed to use - // strides that will be needed for other purposes. A common example of this - // is the exit test for the loop, which can often be rewritten to use the - // computation of some other indvar to decide when to terminate the loop. - OptimizeIndvars(L); +LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P) + : IU(P->getAnalysis<IVUsers>()), + SE(P->getAnalysis<ScalarEvolution>()), + DT(P->getAnalysis<DominatorTree>()), + TLI(tli), L(l), Changed(false), IVIncInsertPos(0) { - // Change loop terminating condition to use the postinc iv when possible - // and optimize loop terminating compare. FIXME: Move this after - // StrengthReduceIVUsersOfStride? - OptimizeLoopTermCond(L); + // If LoopSimplify form is not available, stay out of trouble. + if (!L->isLoopSimplifyForm()) return; + + // If there's no interesting work to be done, bail early. + if (IU.empty()) return; + + DEBUG(dbgs() << "\nLSR on loop "; + WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); + dbgs() << ":\n"); + + /// OptimizeShadowIV - If IV is used in a int-to-float cast + /// inside the loop then try to eliminate the cast opeation. + OptimizeShadowIV(); + + // Change loop terminating condition to use the postinc iv when possible. + Changed |= OptimizeLoopTermCond(); + + CollectInterestingTypesAndFactors(); + CollectFixupsAndInitialFormulae(); + CollectLoopInvariantFixupsAndFormulae(); + + DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; + print_uses(dbgs())); + + // Now use the reuse data to generate a bunch of interesting ways + // to formulate the values needed for the uses. + GenerateAllReuseFormulae(); + + DEBUG(dbgs() << "\n" + "After generating reuse formulae:\n"; + print_uses(dbgs())); + + FilterOutUndesirableDedicatedRegisters(); + NarrowSearchSpaceUsingHeuristics(); + + SmallVector<const Formula *, 8> Solution; + Solve(Solution); + assert(Solution.size() == Uses.size() && "Malformed solution!"); + + // Release memory that is no longer needed. + Factors.clear(); + Types.clear(); + RegUses.clear(); + +#ifndef NDEBUG + // Formulae should be legal. + for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), + E = Uses.end(); I != E; ++I) { + const LSRUse &LU = *I; + for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), + JE = LU.Formulae.end(); J != JE; ++J) + assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset, + LU.Kind, LU.AccessTy, TLI) && + "Illegal formula generated!"); + }; +#endif - // FIXME: We can shrink overlarge IV's here. e.g. if the code has - // computation in i64 values and the target doesn't support i64, demote - // the computation to 32-bit if safe. + // Now that we've decided what we want, make it so. + ImplementSolution(Solution, P); +} - // FIXME: Attempt to reuse values across multiple IV's. In particular, we - // could have something like "for(i) { foo(i*8); bar(i*16) }", which should - // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. - // Need to be careful that IV's are all the same type. Only works for - // intptr_t indvars. +void LSRInstance::print_factors_and_types(raw_ostream &OS) const { + if (Factors.empty() && Types.empty()) return; - // IVsByStride keeps IVs for one particular loop. - assert(IVsByStride.empty() && "Stale entries in IVsByStride?"); + OS << "LSR has identified the following interesting factors and types: "; + bool First = true; - StrengthReduceIVUsers(L); + for (SmallSetVector<int64_t, 8>::const_iterator + I = Factors.begin(), E = Factors.end(); I != E; ++I) { + if (!First) OS << ", "; + First = false; + OS << '*' << *I; + } - // After all sharing is done, see if we can adjust the loop to test against - // zero instead of counting up to a maximum. This is usually faster. - OptimizeLoopCountIV(L); + for (SmallSetVector<const Type *, 4>::const_iterator + I = Types.begin(), E = Types.end(); I != E; ++I) { + if (!First) OS << ", "; + First = false; + OS << '(' << **I << ')'; + } + OS << '\n'; +} - // We're done analyzing this loop; release all the state we built up for it. - IVsByStride.clear(); +void LSRInstance::print_fixups(raw_ostream &OS) const { + OS << "LSR is examining the following fixup sites:\n"; + for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), + E = Fixups.end(); I != E; ++I) { + const LSRFixup &LF = *I; + dbgs() << " "; + LF.print(OS); + OS << '\n'; + } +} - // Clean up after ourselves - DeleteTriviallyDeadInstructions(); +void LSRInstance::print_uses(raw_ostream &OS) const { + OS << "LSR is examining the following uses:\n"; + for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), + E = Uses.end(); I != E; ++I) { + const LSRUse &LU = *I; + dbgs() << " "; + LU.print(OS); + OS << '\n'; + for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), + JE = LU.Formulae.end(); J != JE; ++J) { + OS << " "; + J->print(OS); + OS << '\n'; + } } +} + +void LSRInstance::print(raw_ostream &OS) const { + print_factors_and_types(OS); + print_fixups(OS); + print_uses(OS); +} + +void LSRInstance::dump() const { + print(errs()); errs() << '\n'; +} + +namespace { + +class LoopStrengthReduce : public LoopPass { + /// TLI - Keep a pointer of a TargetLowering to consult for determining + /// transformation profitability. + const TargetLowering *const TLI; + +public: + static char ID; // Pass ID, replacement for typeid + explicit LoopStrengthReduce(const TargetLowering *tli = 0); + +private: + bool runOnLoop(Loop *L, LPPassManager &LPM); + void getAnalysisUsage(AnalysisUsage &AU) const; +}; + +} + +char LoopStrengthReduce::ID = 0; +static RegisterPass<LoopStrengthReduce> +X("loop-reduce", "Loop Strength Reduction"); + +Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { + return new LoopStrengthReduce(TLI); +} + +LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli) + : LoopPass(&ID), TLI(tli) {} + +void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { + // We split critical edges, so we change the CFG. However, we do update + // many analyses if they are around. + AU.addPreservedID(LoopSimplifyID); + AU.addPreserved<LoopInfo>(); + AU.addPreserved("domfrontier"); + + AU.addRequiredID(LoopSimplifyID); + AU.addRequired<DominatorTree>(); + AU.addPreserved<DominatorTree>(); + AU.addRequired<ScalarEvolution>(); + AU.addPreserved<ScalarEvolution>(); + AU.addRequired<IVUsers>(); + AU.addPreserved<IVUsers>(); +} + +bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { + bool Changed = false; + + // Run the main LSR transformation. + Changed |= LSRInstance(TLI, L, this).getChanged(); // At this point, it is worth checking to see if any recurrence PHIs are also // dead, so that we can remove them as well. diff --git a/lib/Transforms/Scalar/LoopUnrollPass.cpp b/lib/Transforms/Scalar/LoopUnrollPass.cpp index ee8cb4f..a355ec3 100644 --- a/lib/Transforms/Scalar/LoopUnrollPass.cpp +++ b/lib/Transforms/Scalar/LoopUnrollPass.cpp @@ -76,11 +76,12 @@ static RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops"); Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); } /// ApproximateLoopSize - Approximate the size of the loop. -static unsigned ApproximateLoopSize(const Loop *L) { +static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) { CodeMetrics Metrics; for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E; ++I) Metrics.analyzeBasicBlock(*I); + NumCalls = Metrics.NumCalls; return Metrics.NumInsts; } @@ -110,8 +111,13 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) { // Enforce the threshold. if (UnrollThreshold != NoThreshold) { - unsigned LoopSize = ApproximateLoopSize(L); + unsigned NumCalls; + unsigned LoopSize = ApproximateLoopSize(L, NumCalls); DEBUG(dbgs() << " Loop Size = " << LoopSize << "\n"); + if (NumCalls != 0) { + DEBUG(dbgs() << " Not unrolling loop with function calls.\n"); + return false; + } uint64_t Size = (uint64_t)LoopSize*Count; if (TripCount != 1 && Size > UnrollThreshold) { DEBUG(dbgs() << " Too large to fully unroll with count: " << Count diff --git a/lib/Transforms/Scalar/LoopUnswitch.cpp b/lib/Transforms/Scalar/LoopUnswitch.cpp index 527a7b5..990e0c4 100644 --- a/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -169,6 +169,10 @@ Pass *llvm::createLoopUnswitchPass(bool Os) { /// invariant in the loop, or has an invariant piece, return the invariant. /// Otherwise, return null. static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { + // We can never unswitch on vector conditions. + if (isa<VectorType>(Cond->getType())) + return 0; + // Constants should be folded, not unswitched on! if (isa<Constant>(Cond)) return 0; @@ -401,7 +405,7 @@ bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when /// LoopCond == Val to simplify the loop. If we decide that this is profitable, /// unswitch the loop, reprocess the pieces, then return true. -bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){ +bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { initLoopData(); @@ -867,7 +871,7 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC // in the loop with the appropriate one directly. if (IsEqual || (isa<ConstantInt>(Val) && - Val->getType()->isInteger(1))) { + Val->getType()->isIntegerTy(1))) { Value *Replacement; if (IsEqual) Replacement = Val; @@ -993,10 +997,10 @@ void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { case Instruction::And: if (isa<ConstantInt>(I->getOperand(0)) && // constant -> RHS - I->getOperand(0)->getType()->isInteger(1)) + I->getOperand(0)->getType()->isIntegerTy(1)) cast<BinaryOperator>(I)->swapOperands(); if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) - if (CB->getType()->isInteger(1)) { + if (CB->getType()->isIntegerTy(1)) { if (CB->isOne()) // X & 1 -> X ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); else // X & 0 -> 0 @@ -1007,10 +1011,10 @@ void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { case Instruction::Or: if (isa<ConstantInt>(I->getOperand(0)) && // constant -> RHS - I->getOperand(0)->getType()->isInteger(1)) + I->getOperand(0)->getType()->isIntegerTy(1)) cast<BinaryOperator>(I)->swapOperands(); if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) - if (CB->getType()->isInteger(1)) { + if (CB->getType()->isIntegerTy(1)) { if (CB->isOne()) // X | 1 -> 1 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); else // X | 0 -> X diff --git a/lib/Transforms/Scalar/Makefile b/lib/Transforms/Scalar/Makefile index e18f30f..cc42fd0 100644 --- a/lib/Transforms/Scalar/Makefile +++ b/lib/Transforms/Scalar/Makefile @@ -10,7 +10,6 @@ LEVEL = ../../.. LIBRARYNAME = LLVMScalarOpts BUILD_ARCHIVE = 1 -CXXFLAGS = -fno-rtti include $(LEVEL)/Makefile.common diff --git a/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/lib/Transforms/Scalar/MemCpyOptimizer.cpp index e0aa491..62e2977 100644 --- a/lib/Transforms/Scalar/MemCpyOptimizer.cpp +++ b/lib/Transforms/Scalar/MemCpyOptimizer.cpp @@ -42,7 +42,7 @@ static Value *isBytewiseValue(Value *V) { LLVMContext &Context = V->getContext(); // All byte-wide stores are splatable, even of arbitrary variables. - if (V->getType()->isInteger(8)) return V; + if (V->getType()->isIntegerTy(8)) return V; // Constant float and double values can be handled as integer values if the // corresponding integer value is "byteable". An important case is 0.0. diff --git a/lib/Transforms/Scalar/Reassociate.cpp b/lib/Transforms/Scalar/Reassociate.cpp index 4a99f4a..187216a 100644 --- a/lib/Transforms/Scalar/Reassociate.cpp +++ b/lib/Transforms/Scalar/Reassociate.cpp @@ -182,7 +182,7 @@ unsigned Reassociate::getRank(Value *V) { // If this is a not or neg instruction, do not count it for rank. This // assures us that X and ~X will have the same rank. - if (!I->getType()->isInteger() || + if (!I->getType()->isIntegerTy() || (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) ++Rank; @@ -249,7 +249,7 @@ void Reassociate::LinearizeExpr(BinaryOperator *I) { /// LinearizeExprTree - Given an associative binary expression tree, traverse /// all of the uses putting it into canonical form. This forces a left-linear -/// form of the the expression (((a+b)+c)+d), and collects information about the +/// form of the expression (((a+b)+c)+d), and collects information about the /// rank of the non-tree operands. /// /// NOTE: These intentionally destroys the expression tree operands (turning @@ -299,7 +299,7 @@ void Reassociate::LinearizeExprTree(BinaryOperator *I, Success = false; MadeChange = true; } else if (RHSBO) { - // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not + // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not // part of the expression tree. LinearizeExpr(I); LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); @@ -929,10 +929,19 @@ void Reassociate::ReassociateBB(BasicBlock *BB) { } // Reject cases where it is pointless to do this. - if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || + if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() || isa<VectorType>(BI->getType())) continue; // Floating point ops are not associative. + // Do not reassociate boolean (i1) expressions. We want to preserve the + // original order of evaluation for short-circuited comparisons that + // SimplifyCFG has folded to AND/OR expressions. If the expression + // is not further optimized, it is likely to be transformed back to a + // short-circuited form for code gen, and the source order may have been + // optimized for the most likely conditions. + if (BI->getType()->isIntegerTy(1)) + continue; + // If this is a subtract instruction which is not already in negate form, // see if we can convert it to X+-Y. if (BI->getOpcode() == Instruction::Sub) { diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp index f473480..822712e 100644 --- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp +++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp @@ -202,12 +202,18 @@ bool SROA::performPromotion(Function &F) { return Changed; } -/// getNumSAElements - Return the number of elements in the specific struct or -/// array. -static uint64_t getNumSAElements(const Type *T) { +/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for +/// SROA. It must be a struct or array type with a small number of elements. +static bool ShouldAttemptScalarRepl(AllocaInst *AI) { + const Type *T = AI->getAllocatedType(); + // Do not promote any struct into more than 32 separate vars. if (const StructType *ST = dyn_cast<StructType>(T)) - return ST->getNumElements(); - return cast<ArrayType>(T)->getNumElements(); + return ST->getNumElements() <= 32; + // Arrays are much less likely to be safe for SROA; only consider + // them if they are very small. + if (const ArrayType *AT = dyn_cast<ArrayType>(T)) + return AT->getNumElements() <= 8; + return false; } // performScalarRepl - This algorithm is a simple worklist driven algorithm, @@ -266,22 +272,18 @@ bool SROA::performScalarRepl(Function &F) { // Do not promote [0 x %struct]. if (AllocaSize == 0) continue; + // If the alloca looks like a good candidate for scalar replacement, and if + // all its users can be transformed, then split up the aggregate into its + // separate elements. + if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { + DoScalarReplacement(AI, WorkList); + Changed = true; + continue; + } + // Do not promote any struct whose size is too big. if (AllocaSize > SRThreshold) continue; - if ((isa<StructType>(AI->getAllocatedType()) || - isa<ArrayType>(AI->getAllocatedType())) && - // Do not promote any struct into more than "32" separate vars. - getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { - // Check that all of the users of the allocation are capable of being - // transformed. - if (isSafeAllocaToScalarRepl(AI)) { - DoScalarReplacement(AI, WorkList); - Changed = true; - continue; - } - } - // If we can turn this aggregate value (potentially with casts) into a // simple scalar value that can be mem2reg'd into a register value. // IsNotTrivial tracks whether this is something that mem2reg could have @@ -681,7 +683,7 @@ void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, Val->takeName(GEPI); } if (Val->getType() != GEPI->getType()) - Val = new BitCastInst(Val, GEPI->getType(), Val->getNameStr(), GEPI); + Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); GEPI->replaceAllUsesWith(Val); DeadInsts.push_back(GEPI); } @@ -769,7 +771,7 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, Value *Idx[2] = { Zero, ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, - OtherPtr->getNameStr()+"."+Twine(i), + OtherPtr->getName()+"."+Twine(i), MI); uint64_t EltOffset; const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); @@ -833,7 +835,7 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, StoreVal = ConstantInt::get(Context, TotalVal); if (isa<PointerType>(ValTy)) StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); - else if (ValTy->isFloatingPoint()) + else if (ValTy->isFloatingPointTy()) StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); assert(StoreVal->getType() == ValTy && "Type mismatch!"); @@ -853,12 +855,11 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, // Cast the element pointer to BytePtrTy. if (EltPtr->getType() != BytePtrTy) - EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); + EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); // Cast the other pointer (if we have one) to BytePtrTy. if (OtherElt && OtherElt->getType() != BytePtrTy) - OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), - MI); + OtherElt = new BitCastInst(OtherElt, BytePtrTy, OtherElt->getName(), MI); unsigned EltSize = TD->getTypeAllocSize(EltTy); @@ -938,7 +939,7 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, Value *DestField = NewElts[i]; if (EltVal->getType() == FieldTy) { // Storing to an integer field of this size, just do it. - } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { + } else if (FieldTy->isFloatingPointTy() || isa<VectorType>(FieldTy)) { // Bitcast to the right element type (for fp/vector values). EltVal = new BitCastInst(EltVal, FieldTy, "", SI); } else { @@ -982,7 +983,8 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, Value *DestField = NewElts[i]; if (EltVal->getType() == ArrayEltTy) { // Storing to an integer field of this size, just do it. - } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { + } else if (ArrayEltTy->isFloatingPointTy() || + isa<VectorType>(ArrayEltTy)) { // Bitcast to the right element type (for fp/vector values). EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); } else { @@ -1042,7 +1044,7 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), FieldSizeBits); - if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && + if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPointTy() && !isa<VectorType>(FieldTy)) SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy), @@ -1384,9 +1386,9 @@ void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { // If the source and destination are both to the same alloca, then this is // a noop copy-to-self, just delete it. Otherwise, emit a load and store // as appropriate. - AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); + AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); - if (MTI->getSource()->getUnderlyingObject() != OrigAI) { + if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { // Dest must be OrigAI, change this to be a load from the original // pointer (bitcasted), then a store to our new alloca. assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); @@ -1396,7 +1398,7 @@ void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); SrcVal->setAlignment(MTI->getAlignment()); Builder.CreateStore(SrcVal, NewAI); - } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { + } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { // Src must be OrigAI, change this to be a load from NewAI then a store // through the original dest pointer (bitcasted). assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); @@ -1521,7 +1523,7 @@ Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, // If the result is an integer, this is a trunc or bitcast. if (isa<IntegerType>(ToType)) { // Should be done. - } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { + } else if (ToType->isFloatingPointTy() || isa<VectorType>(ToType)) { // Just do a bitcast, we know the sizes match up. FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); } else { @@ -1599,7 +1601,7 @@ Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); - if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) + if (SV->getType()->isFloatingPointTy() || isa<VectorType>(SV->getType())) SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth), "tmp"); else if (isa<PointerType>(SV->getType())) diff --git a/lib/Transforms/Scalar/SimplifyCFGPass.cpp b/lib/Transforms/Scalar/SimplifyCFGPass.cpp index 43447de..62f34a2 100644 --- a/lib/Transforms/Scalar/SimplifyCFGPass.cpp +++ b/lib/Transforms/Scalar/SimplifyCFGPass.cpp @@ -30,6 +30,7 @@ #include "llvm/Attributes.h" #include "llvm/Support/CFG.h" #include "llvm/Pass.h" +#include "llvm/Target/TargetData.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/Statistic.h" @@ -261,7 +262,7 @@ static bool MergeEmptyReturnBlocks(Function &F) { /// IterativeSimplifyCFG - Call SimplifyCFG on all the blocks in the function, /// iterating until no more changes are made. -static bool IterativeSimplifyCFG(Function &F) { +static bool IterativeSimplifyCFG(Function &F, const TargetData *TD) { bool Changed = false; bool LocalChange = true; while (LocalChange) { @@ -271,7 +272,7 @@ static bool IterativeSimplifyCFG(Function &F) { // if they are unneeded... // for (Function::iterator BBIt = ++F.begin(); BBIt != F.end(); ) { - if (SimplifyCFG(BBIt++)) { + if (SimplifyCFG(BBIt++, TD)) { LocalChange = true; ++NumSimpl; } @@ -285,10 +286,11 @@ static bool IterativeSimplifyCFG(Function &F) { // simplify the CFG. // bool CFGSimplifyPass::runOnFunction(Function &F) { + const TargetData *TD = getAnalysisIfAvailable<TargetData>(); bool EverChanged = RemoveUnreachableBlocksFromFn(F); EverChanged |= MergeEmptyReturnBlocks(F); - EverChanged |= IterativeSimplifyCFG(F); - + EverChanged |= IterativeSimplifyCFG(F, TD); + // If neither pass changed anything, we're done. if (!EverChanged) return false; @@ -299,11 +301,11 @@ bool CFGSimplifyPass::runOnFunction(Function &F) { // RemoveUnreachableBlocksFromFn doesn't do anything. if (!RemoveUnreachableBlocksFromFn(F)) return true; - + do { - EverChanged = IterativeSimplifyCFG(F); + EverChanged = IterativeSimplifyCFG(F, TD); EverChanged |= RemoveUnreachableBlocksFromFn(F); } while (EverChanged); - + return true; } diff --git a/lib/Transforms/Scalar/SimplifyHalfPowrLibCalls.cpp b/lib/Transforms/Scalar/SimplifyHalfPowrLibCalls.cpp index 5acd6aa..4464961 100644 --- a/lib/Transforms/Scalar/SimplifyHalfPowrLibCalls.cpp +++ b/lib/Transforms/Scalar/SimplifyHalfPowrLibCalls.cpp @@ -68,7 +68,7 @@ InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs, Function *Callee = Call->getCalledFunction(); // Minimally sanity-check the CFG of half_powr to ensure that it contains - // the the kind of code we expect. If we're running this pass, we have + // the kind of code we expect. If we're running this pass, we have // reason to believe it will be what we expect. Function::iterator I = Callee->begin(); BasicBlock *Prologue = I++; diff --git a/lib/Transforms/Scalar/SimplifyLibCalls.cpp b/lib/Transforms/Scalar/SimplifyLibCalls.cpp index a49da9c..54b4380 100644 --- a/lib/Transforms/Scalar/SimplifyLibCalls.cpp +++ b/lib/Transforms/Scalar/SimplifyLibCalls.cpp @@ -152,7 +152,7 @@ Value *LibCallOptimization::EmitStrLen(Value *Ptr, IRBuilder<> &B) { Constant *StrLen =M->getOrInsertFunction("strlen", AttrListPtr::get(AWI, 2), TD->getIntPtrType(*Context), - Type::getInt8PtrTy(*Context), + Type::getInt8PtrTy(*Context), NULL); CallInst *CI = B.CreateCall(StrLen, CastToCStr(Ptr, B), "strlen"); if (const Function *F = dyn_cast<Function>(StrLen->stripPointerCasts())) @@ -232,10 +232,10 @@ Value *LibCallOptimization::EmitMemChr(Value *Ptr, Value *Val, AWI = AttributeWithIndex::get(~0u, Attribute::ReadOnly | Attribute::NoUnwind); Value *MemChr = M->getOrInsertFunction("memchr", AttrListPtr::get(&AWI, 1), - Type::getInt8PtrTy(*Context), - Type::getInt8PtrTy(*Context), + Type::getInt8PtrTy(*Context), + Type::getInt8PtrTy(*Context), Type::getInt32Ty(*Context), - TD->getIntPtrType(*Context), + TD->getIntPtrType(*Context), NULL); CallInst *CI = B.CreateCall3(MemChr, CastToCStr(Ptr, B), Val, Len, "memchr"); @@ -321,9 +321,9 @@ Value *LibCallOptimization::EmitPutChar(Value *Char, IRBuilder<> &B) { Type::getInt32Ty(*Context), NULL); CallInst *CI = B.CreateCall(PutChar, B.CreateIntCast(Char, - Type::getInt32Ty(*Context), - /*isSigned*/true, - "chari"), + Type::getInt32Ty(*Context), + /*isSigned*/true, + "chari"), "putchar"); if (const Function *F = dyn_cast<Function>(PutChar->stripPointerCasts())) @@ -341,7 +341,7 @@ void LibCallOptimization::EmitPutS(Value *Str, IRBuilder<> &B) { Value *PutS = M->getOrInsertFunction("puts", AttrListPtr::get(AWI, 2), Type::getInt32Ty(*Context), - Type::getInt8PtrTy(*Context), + Type::getInt8PtrTy(*Context), NULL); CallInst *CI = B.CreateCall(PutS, CastToCStr(Str, B), "puts"); if (const Function *F = dyn_cast<Function>(PutS->stripPointerCasts())) @@ -359,13 +359,13 @@ void LibCallOptimization::EmitFPutC(Value *Char, Value *File, IRBuilder<> &B) { Constant *F; if (isa<PointerType>(File->getType())) F = M->getOrInsertFunction("fputc", AttrListPtr::get(AWI, 2), - Type::getInt32Ty(*Context), + Type::getInt32Ty(*Context), Type::getInt32Ty(*Context), File->getType(), - NULL); + NULL); else F = M->getOrInsertFunction("fputc", - Type::getInt32Ty(*Context), - Type::getInt32Ty(*Context), + Type::getInt32Ty(*Context), + Type::getInt32Ty(*Context), File->getType(), NULL); Char = B.CreateIntCast(Char, Type::getInt32Ty(*Context), /*isSigned*/true, "chari"); @@ -386,7 +386,7 @@ void LibCallOptimization::EmitFPutS(Value *Str, Value *File, IRBuilder<> &B) { Constant *F; if (isa<PointerType>(File->getType())) F = M->getOrInsertFunction("fputs", AttrListPtr::get(AWI, 3), - Type::getInt32Ty(*Context), + Type::getInt32Ty(*Context), Type::getInt8PtrTy(*Context), File->getType(), NULL); else @@ -414,13 +414,13 @@ void LibCallOptimization::EmitFWrite(Value *Ptr, Value *Size, Value *File, TD->getIntPtrType(*Context), Type::getInt8PtrTy(*Context), TD->getIntPtrType(*Context), - TD->getIntPtrType(*Context), + TD->getIntPtrType(*Context), File->getType(), NULL); else F = M->getOrInsertFunction("fwrite", TD->getIntPtrType(*Context), Type::getInt8PtrTy(*Context), TD->getIntPtrType(*Context), - TD->getIntPtrType(*Context), + TD->getIntPtrType(*Context), File->getType(), NULL); CallInst *CI = B.CreateCall4(F, CastToCStr(Ptr, B), Size, ConstantInt::get(TD->getIntPtrType(*Context), 1), File); @@ -525,7 +525,7 @@ static uint64_t GetStringLengthH(Value *V, SmallPtrSet<PHINode*, 32> &PHIs) { // Must be a Constant Array ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit); - if (!Array || !Array->getType()->getElementType()->isInteger(8)) + if (!Array || !Array->getType()->getElementType()->isIntegerTy(8)) return false; // Get the number of elements in the array @@ -697,7 +697,7 @@ struct StrChrOpt : public LibCallOptimization { if (!TD) return 0; uint64_t Len = GetStringLength(SrcStr); - if (Len == 0 || !FT->getParamType(1)->isInteger(32)) // memchr needs i32. + if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32))// memchr needs i32. return 0; return EmitMemChr(SrcStr, CI->getOperand(2), // include nul. @@ -739,7 +739,7 @@ struct StrCmpOpt : public LibCallOptimization { // Verify the "strcmp" function prototype. const FunctionType *FT = Callee->getFunctionType(); if (FT->getNumParams() != 2 || - !FT->getReturnType()->isInteger(32) || + !FT->getReturnType()->isIntegerTy(32) || FT->getParamType(0) != FT->getParamType(1) || FT->getParamType(0) != Type::getInt8PtrTy(*Context)) return 0; @@ -787,7 +787,7 @@ struct StrNCmpOpt : public LibCallOptimization { // Verify the "strncmp" function prototype. const FunctionType *FT = Callee->getFunctionType(); if (FT->getNumParams() != 3 || - !FT->getReturnType()->isInteger(32) || + !FT->getReturnType()->isIntegerTy(32) || FT->getParamType(0) != FT->getParamType(1) || FT->getParamType(0) != Type::getInt8PtrTy(*Context) || !isa<IntegerType>(FT->getParamType(2))) @@ -1008,7 +1008,7 @@ struct MemCmpOpt : public LibCallOptimization { const FunctionType *FT = Callee->getFunctionType(); if (FT->getNumParams() != 3 || !isa<PointerType>(FT->getParamType(0)) || !isa<PointerType>(FT->getParamType(1)) || - !FT->getReturnType()->isInteger(32)) + !FT->getReturnType()->isIntegerTy(32)) return 0; Value *LHS = CI->getOperand(1), *RHS = CI->getOperand(2); @@ -1203,22 +1203,23 @@ struct MemMoveChkOpt : public LibCallOptimization { struct StrCpyChkOpt : public LibCallOptimization { virtual Value *CallOptimizer(Function *Callee, CallInst *CI, IRBuilder<> &B) { - // These optimizations require TargetData. - if (!TD) return 0; - const FunctionType *FT = Callee->getFunctionType(); if (FT->getNumParams() != 3 || FT->getReturnType() != FT->getParamType(0) || !isa<PointerType>(FT->getParamType(0)) || - !isa<PointerType>(FT->getParamType(1)) || - !isa<IntegerType>(FT->getParamType(2))) + !isa<PointerType>(FT->getParamType(1))) return 0; ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getOperand(3)); if (!SizeCI) return 0; - // We don't have any length information, just lower to a plain strcpy. - if (SizeCI->isAllOnesValue()) + // 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 (SizeCI->isAllOnesValue() || + SizeCI->getZExtValue() >= GetStringLength(CI->getOperand(2))) return EmitStrCpy(CI->getOperand(1), CI->getOperand(2), B); return 0; @@ -1240,7 +1241,7 @@ struct PowOpt : public LibCallOptimization { // result type. if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) || FT->getParamType(0) != FT->getParamType(1) || - !FT->getParamType(0)->isFloatingPoint()) + !FT->getParamType(0)->isFloatingPointTy()) return 0; Value *Op1 = CI->getOperand(1), *Op2 = CI->getOperand(2); @@ -1294,7 +1295,7 @@ struct Exp2Opt : public LibCallOptimization { // 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)->isFloatingPoint()) + !FT->getParamType(0)->isFloatingPointTy()) return 0; Value *Op = CI->getOperand(1); @@ -1327,7 +1328,7 @@ struct Exp2Opt : public LibCallOptimization { Module *M = Caller->getParent(); Value *Callee = M->getOrInsertFunction(Name, Op->getType(), Op->getType(), - Type::getInt32Ty(*Context),NULL); + Type::getInt32Ty(*Context),NULL); CallInst *CI = B.CreateCall2(Callee, One, LdExpArg); if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts())) CI->setCallingConv(F->getCallingConv()); @@ -1374,7 +1375,7 @@ struct FFSOpt : public LibCallOptimization { // Just make sure this has 2 arguments of the same FP type, which match the // result type. if (FT->getNumParams() != 1 || - !FT->getReturnType()->isInteger(32) || + !FT->getReturnType()->isIntegerTy(32) || !isa<IntegerType>(FT->getParamType(0))) return 0; @@ -1410,7 +1411,7 @@ struct IsDigitOpt : public LibCallOptimization { const FunctionType *FT = Callee->getFunctionType(); // We require integer(i32) if (FT->getNumParams() != 1 || !isa<IntegerType>(FT->getReturnType()) || - !FT->getParamType(0)->isInteger(32)) + !FT->getParamType(0)->isIntegerTy(32)) return 0; // isdigit(c) -> (c-'0') <u 10 @@ -1431,7 +1432,7 @@ struct IsAsciiOpt : public LibCallOptimization { const FunctionType *FT = Callee->getFunctionType(); // We require integer(i32) if (FT->getNumParams() != 1 || !isa<IntegerType>(FT->getReturnType()) || - !FT->getParamType(0)->isInteger(32)) + !FT->getParamType(0)->isIntegerTy(32)) return 0; // isascii(c) -> c <u 128 @@ -1472,7 +1473,7 @@ struct ToAsciiOpt : public LibCallOptimization { const FunctionType *FT = Callee->getFunctionType(); // We require i32(i32) if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || - !FT->getParamType(0)->isInteger(32)) + !FT->getParamType(0)->isIntegerTy(32)) return 0; // isascii(c) -> c & 0x7f diff --git a/lib/Transforms/Scalar/TailRecursionElimination.cpp b/lib/Transforms/Scalar/TailRecursionElimination.cpp index 4119cb9..162d902 100644 --- a/lib/Transforms/Scalar/TailRecursionElimination.cpp +++ b/lib/Transforms/Scalar/TailRecursionElimination.cpp @@ -211,7 +211,8 @@ bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { // FIXME: Writes to memory only matter if they may alias the pointer // being loaded from. if (CI->mayWriteToMemory() || - !isSafeToLoadUnconditionally(L->getPointerOperand(), L)) + !isSafeToLoadUnconditionally(L->getPointerOperand(), L, + L->getAlignment())) return false; } } diff --git a/lib/Transforms/Utils/BreakCriticalEdges.cpp b/lib/Transforms/Utils/BreakCriticalEdges.cpp index 19c7206..3657390 100644 --- a/lib/Transforms/Utils/BreakCriticalEdges.cpp +++ b/lib/Transforms/Utils/BreakCriticalEdges.cpp @@ -179,7 +179,7 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, // Create a new basic block, linking it into the CFG. BasicBlock *NewBB = BasicBlock::Create(TI->getContext(), TIBB->getName() + "." + DestBB->getName() + "_crit_edge"); - // Create our unconditional branch... + // Create our unconditional branch. BranchInst::Create(DestBB, NewBB); // Branch to the new block, breaking the edge. @@ -192,16 +192,47 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, // If there are any PHI nodes in DestBB, we need to update them so that they // merge incoming values from NewBB instead of from TIBB. - // - for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) { - PHINode *PN = cast<PHINode>(I); - // We no longer enter through TIBB, now we come in through NewBB. Revector - // exactly one entry in the PHI node that used to come from TIBB to come - // from NewBB. - int BBIdx = PN->getBasicBlockIndex(TIBB); - PN->setIncomingBlock(BBIdx, NewBB); + if (PHINode *APHI = dyn_cast<PHINode>(DestBB->begin())) { + // This conceptually does: + // foreach (PHINode *PN in DestBB) + // PN->setIncomingBlock(PN->getIncomingBlock(TIBB), NewBB); + // but is optimized for two cases. + + if (APHI->getNumIncomingValues() <= 8) { // Small # preds case. + unsigned BBIdx = 0; + for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) { + // We no longer enter through TIBB, now we come in through NewBB. + // Revector exactly one entry in the PHI node that used to come from + // TIBB to come from NewBB. + PHINode *PN = cast<PHINode>(I); + + // Reuse the previous value of BBIdx if it lines up. In cases where we + // have multiple phi nodes with *lots* of predecessors, this is a speed + // win because we don't have to scan the PHI looking for TIBB. This + // happens because the BB list of PHI nodes are usually in the same + // order. + if (PN->getIncomingBlock(BBIdx) != TIBB) + BBIdx = PN->getBasicBlockIndex(TIBB); + PN->setIncomingBlock(BBIdx, NewBB); + } + } else { + // However, the foreach loop is slow for blocks with lots of predecessors + // because PHINode::getIncomingBlock is O(n) in # preds. Instead, walk + // the user list of TIBB to find the PHI nodes. + SmallPtrSet<PHINode*, 16> UpdatedPHIs; + + for (Value::use_iterator UI = TIBB->use_begin(), E = TIBB->use_end(); + UI != E; ) { + Value::use_iterator Use = UI++; + if (PHINode *PN = dyn_cast<PHINode>(Use)) { + // Remove one entry from each PHI. + if (PN->getParent() == DestBB && UpdatedPHIs.insert(PN)) + PN->setOperand(Use.getOperandNo(), NewBB); + } + } + } } - + // If there are any other edges from TIBB to DestBB, update those to go // through the split block, making those edges non-critical as well (and // reducing the number of phi entries in the DestBB if relevant). @@ -221,6 +252,15 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, // If we don't have a pass object, we can't update anything... if (P == 0) return NewBB; + + DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>(); + DominanceFrontier *DF = P->getAnalysisIfAvailable<DominanceFrontier>(); + LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>(); + ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>(); + + // If we have nothing to update, just return. + if (DT == 0 && DF == 0 && LI == 0 && PI == 0) + return NewBB; // Now update analysis information. Since the only predecessor of NewBB is // the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate @@ -229,14 +269,23 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, // loop header) then NewBB dominates DestBB. SmallVector<BasicBlock*, 8> OtherPreds; - for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB); I != E; ++I) - if (*I != NewBB) - OtherPreds.push_back(*I); + // If there is a PHI in the block, loop over predecessors with it, which is + // faster than iterating pred_begin/end. + if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) + if (PN->getIncomingBlock(i) != NewBB) + OtherPreds.push_back(PN->getIncomingBlock(i)); + } else { + for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB); + I != E; ++I) + if (*I != NewBB) + OtherPreds.push_back(*I); + } bool NewBBDominatesDestBB = true; // Should we update DominatorTree information? - if (DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>()) { + if (DT) { DomTreeNode *TINode = DT->getNode(TIBB); // The new block is not the immediate dominator for any other nodes, but @@ -267,7 +316,7 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, } // Should we update DominanceFrontier information? - if (DominanceFrontier *DF = P->getAnalysisIfAvailable<DominanceFrontier>()) { + if (DF) { // If NewBBDominatesDestBB hasn't been computed yet, do so with DF. if (!OtherPreds.empty()) { // FIXME: IMPLEMENT THIS! @@ -301,7 +350,7 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, } // Update LoopInfo if it is around. - if (LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>()) { + if (LI) { if (Loop *TIL = LI->getLoopFor(TIBB)) { // If one or the other blocks were not in a loop, the new block is not // either, and thus LI doesn't need to be updated. @@ -382,9 +431,8 @@ BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, } // Update ProfileInfo if it is around. - if (ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>()) { - PI->splitEdge(TIBB,DestBB,NewBB,MergeIdenticalEdges); - } + if (PI) + PI->splitEdge(TIBB, DestBB, NewBB, MergeIdenticalEdges); return NewBB; } diff --git a/lib/Transforms/Utils/CloneFunction.cpp b/lib/Transforms/Utils/CloneFunction.cpp index bd750cc..c80827d 100644 --- a/lib/Transforms/Utils/CloneFunction.cpp +++ b/lib/Transforms/Utils/CloneFunction.cpp @@ -33,7 +33,7 @@ using namespace llvm; // CloneBasicBlock - See comments in Cloning.h BasicBlock *llvm::CloneBasicBlock(const BasicBlock *BB, DenseMap<const Value*, Value*> &ValueMap, - const char *NameSuffix, Function *F, + const Twine &NameSuffix, Function *F, ClonedCodeInfo *CodeInfo) { BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "", F); if (BB->hasName()) NewBB->setName(BB->getName()+NameSuffix); diff --git a/lib/Transforms/Utils/Local.cpp b/lib/Transforms/Utils/Local.cpp index 92bdf2d..57ad459 100644 --- a/lib/Transforms/Utils/Local.cpp +++ b/lib/Transforms/Utils/Local.cpp @@ -38,20 +38,82 @@ using namespace llvm; // Local analysis. // +/// getUnderlyingObjectWithOffset - Strip off up to MaxLookup GEPs and +/// bitcasts to get back to the underlying object being addressed, keeping +/// track of the offset in bytes from the GEPs relative to the result. +/// This is closely related to Value::getUnderlyingObject but is located +/// here to avoid making VMCore depend on TargetData. +static Value *getUnderlyingObjectWithOffset(Value *V, const TargetData *TD, + uint64_t &ByteOffset, + unsigned MaxLookup = 6) { + if (!isa<PointerType>(V->getType())) + return V; + for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { + if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { + if (!GEP->hasAllConstantIndices()) + return V; + SmallVector<Value*, 8> Indices(GEP->op_begin() + 1, GEP->op_end()); + ByteOffset += TD->getIndexedOffset(GEP->getPointerOperandType(), + &Indices[0], Indices.size()); + V = GEP->getPointerOperand(); + } else if (Operator::getOpcode(V) == Instruction::BitCast) { + V = cast<Operator>(V)->getOperand(0); + } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { + if (GA->mayBeOverridden()) + return V; + V = GA->getAliasee(); + } else { + return V; + } + assert(isa<PointerType>(V->getType()) && "Unexpected operand type!"); + } + return V; +} + /// isSafeToLoadUnconditionally - Return true if we know that executing a load /// from this value cannot trap. If it is not obviously safe to load from the /// specified pointer, we do a quick local scan of the basic block containing /// ScanFrom, to determine if the address is already accessed. -bool llvm::isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) { - // If it is an alloca it is always safe to load from. - if (isa<AllocaInst>(V)) return true; +bool llvm::isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom, + unsigned Align, const TargetData *TD) { + uint64_t ByteOffset = 0; + Value *Base = V; + if (TD) + Base = getUnderlyingObjectWithOffset(V, TD, ByteOffset); + + const Type *BaseType = 0; + unsigned BaseAlign = 0; + if (const AllocaInst *AI = dyn_cast<AllocaInst>(Base)) { + // An alloca is safe to load from as load as it is suitably aligned. + BaseType = AI->getAllocatedType(); + BaseAlign = AI->getAlignment(); + } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Base)) { + // Global variables are safe to load from but their size cannot be + // guaranteed if they are overridden. + if (!isa<GlobalAlias>(GV) && !GV->mayBeOverridden()) { + BaseType = GV->getType()->getElementType(); + BaseAlign = GV->getAlignment(); + } + } - // If it is a global variable it is mostly safe to load from. - if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V)) - // Don't try to evaluate aliases. External weak GV can be null. - return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage(); + if (BaseType && BaseType->isSized()) { + if (TD && BaseAlign == 0) + BaseAlign = TD->getPrefTypeAlignment(BaseType); - // Otherwise, be a little bit agressive by scanning the local block where we + if (Align <= BaseAlign) { + if (!TD) + return true; // Loading directly from an alloca or global is OK. + + // Check if the load is within the bounds of the underlying object. + const PointerType *AddrTy = cast<PointerType>(V->getType()); + uint64_t LoadSize = TD->getTypeStoreSize(AddrTy->getElementType()); + if (ByteOffset + LoadSize <= TD->getTypeAllocSize(BaseType) && + (Align == 0 || (ByteOffset % Align) == 0)) + return true; + } + } + + // Otherwise, be a little bit aggressive by scanning the local block where we // want to check to see if the pointer is already being loaded or stored // from/to. If so, the previous load or store would have already trapped, // so there is no harm doing an extra load (also, CSE will later eliminate @@ -428,6 +490,17 @@ void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, Pass *P) { // Splice all the instructions from PredBB to DestBB. PredBB->getTerminator()->eraseFromParent(); DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList()); + + // Zap anything that took the address of DestBB. Not doing this will give the + // address an invalid value. + if (DestBB->hasAddressTaken()) { + BlockAddress *BA = BlockAddress::get(DestBB); + Constant *Replacement = + ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1); + BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement, + BA->getType())); + BA->destroyConstant(); + } // Anything that branched to PredBB now branches to DestBB. PredBB->replaceAllUsesWith(DestBB); diff --git a/lib/Transforms/Utils/LoopSimplify.cpp b/lib/Transforms/Utils/LoopSimplify.cpp index e81b779..57bab60 100644 --- a/lib/Transforms/Utils/LoopSimplify.cpp +++ b/lib/Transforms/Utils/LoopSimplify.cpp @@ -176,8 +176,9 @@ ReprocessLoop: SmallVector<BasicBlock*, 8> ExitBlocks; L->getExitBlocks(ExitBlocks); - SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end()); - for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(), + SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(), + ExitBlocks.end()); + for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(), E = ExitBlockSet.end(); I != E; ++I) { BasicBlock *ExitBlock = *I; for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); diff --git a/lib/Transforms/Utils/LoopUnroll.cpp b/lib/Transforms/Utils/LoopUnroll.cpp index 53117a0..e47c86d 100644 --- a/lib/Transforms/Utils/LoopUnroll.cpp +++ b/lib/Transforms/Utils/LoopUnroll.cpp @@ -29,7 +29,6 @@ #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/Cloning.h" #include "llvm/Transforms/Utils/Local.h" -#include <cstdio> using namespace llvm; @@ -204,15 +203,12 @@ bool llvm::UnrollLoop(Loop *L, unsigned Count, LoopInfo* LI, LPPassManager* LPM) Latches.push_back(LatchBlock); for (unsigned It = 1; It != Count; ++It) { - char SuffixBuffer[100]; - sprintf(SuffixBuffer, ".%d", It); - std::vector<BasicBlock*> NewBlocks; for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(), E = LoopBlocks.end(); BB != E; ++BB) { ValueMapTy ValueMap; - BasicBlock *New = CloneBasicBlock(*BB, ValueMap, SuffixBuffer); + BasicBlock *New = CloneBasicBlock(*BB, ValueMap, "." + Twine(It)); Header->getParent()->getBasicBlockList().push_back(New); // Loop over all of the PHI nodes in the block, changing them to use the diff --git a/lib/Transforms/Utils/Makefile b/lib/Transforms/Utils/Makefile index b9761df..d1e9336 100644 --- a/lib/Transforms/Utils/Makefile +++ b/lib/Transforms/Utils/Makefile @@ -10,7 +10,6 @@ LEVEL = ../../.. LIBRARYNAME = LLVMTransformUtils BUILD_ARCHIVE = 1 -CXXFLAGS = -fno-rtti include $(LEVEL)/Makefile.common diff --git a/lib/Transforms/Utils/PromoteMemoryToRegister.cpp b/lib/Transforms/Utils/PromoteMemoryToRegister.cpp index d9261ac..544e20b 100644 --- a/lib/Transforms/Utils/PromoteMemoryToRegister.cpp +++ b/lib/Transforms/Utils/PromoteMemoryToRegister.cpp @@ -23,6 +23,7 @@ #include "llvm/Function.h" #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" +#include "llvm/Metadata.h" #include "llvm/Analysis/DebugInfo.h" #include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/AliasSetTracker.h" @@ -84,6 +85,18 @@ bool llvm::isAllocaPromotable(const AllocaInst *AI) { return true; } +/// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the +/// alloca 'V', if any. +static DbgDeclareInst *FindAllocaDbgDeclare(Value *V) { + if (MDNode *DebugNode = MDNode::getIfExists(V->getContext(), &V, 1)) + for (Value::use_iterator UI = DebugNode->use_begin(), + E = DebugNode->use_end(); UI != E; ++UI) + if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) + return DDI; + + return 0; +} + namespace { struct AllocaInfo; @@ -188,6 +201,11 @@ namespace { /// std::vector<Value*> PointerAllocaValues; + /// AllocaDbgDeclares - For each alloca, we keep track of the dbg.declare + /// intrinsic that describes it, if any, so that we can convert it to a + /// dbg.value intrinsic if the alloca gets promoted. + SmallVector<DbgDeclareInst*, 8> AllocaDbgDeclares; + /// Visited - The set of basic blocks the renamer has already visited. /// SmallPtrSet<BasicBlock*, 16> Visited; @@ -202,6 +220,9 @@ namespace { PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt, DominanceFrontier &df, AliasSetTracker *ast) : Allocas(A), DT(dt), DF(df), DIF(0), AST(ast) {} + ~PromoteMem2Reg() { + delete DIF; + } void run(); @@ -243,9 +264,9 @@ namespace { LargeBlockInfo &LBI); void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info, LargeBlockInfo &LBI); - void ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, StoreInst* SI, - uint64_t Offset); - + void ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, StoreInst *SI); + + void RenamePass(BasicBlock *BB, BasicBlock *Pred, RenamePassData::ValVector &IncVals, std::vector<RenamePassData> &Worklist); @@ -262,6 +283,7 @@ namespace { bool OnlyUsedInOneBlock; Value *AllocaPointerVal; + DbgDeclareInst *DbgDeclare; void clear() { DefiningBlocks.clear(); @@ -270,6 +292,7 @@ namespace { OnlyBlock = 0; OnlyUsedInOneBlock = true; AllocaPointerVal = 0; + DbgDeclare = 0; } /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our @@ -304,28 +327,18 @@ namespace { OnlyUsedInOneBlock = false; } } + + DbgDeclare = FindAllocaDbgDeclare(AI); } }; } // end of anonymous namespace -/// Finds the llvm.dbg.declare intrinsic corresponding to an alloca if any. -static DbgDeclareInst *findDbgDeclare(AllocaInst *AI) { - Function *F = AI->getParent()->getParent(); - for (Function::iterator FI = F->begin(), FE = F->end(); FI != FE; ++FI) - for (BasicBlock::iterator BI = (*FI).begin(), BE = (*FI).end(); - BI != BE; ++BI) - if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI)) - if (DDI->getAddress() == AI) - return DDI; - - return 0; -} - void PromoteMem2Reg::run() { Function &F = *DF.getRoot()->getParent(); if (AST) PointerAllocaValues.resize(Allocas.size()); + AllocaDbgDeclares.resize(Allocas.size()); AllocaInfo Info; LargeBlockInfo LBI; @@ -360,8 +373,11 @@ void PromoteMem2Reg::run() { // Finally, after the scan, check to see if the store is all that is left. if (Info.UsingBlocks.empty()) { - // Record debuginfo for the store before removing it. - ConvertDebugDeclareToDebugValue(findDbgDeclare(AI), Info.OnlyStore, 0); + // Record debuginfo for the store and remove the declaration's debuginfo. + if (DbgDeclareInst *DDI = Info.DbgDeclare) { + ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore); + DDI->eraseFromParent(); + } // Remove the (now dead) store and alloca. Info.OnlyStore->eraseFromParent(); LBI.deleteValue(Info.OnlyStore); @@ -388,11 +404,11 @@ void PromoteMem2Reg::run() { if (Info.UsingBlocks.empty()) { // Remove the (now dead) stores and alloca. - DbgDeclareInst *DDI = findDbgDeclare(AI); while (!AI->use_empty()) { StoreInst *SI = cast<StoreInst>(AI->use_back()); // Record debuginfo for the store before removing it. - ConvertDebugDeclareToDebugValue(DDI, SI, 0); + if (DbgDeclareInst *DDI = Info.DbgDeclare) + ConvertDebugDeclareToDebugValue(DDI, SI); SI->eraseFromParent(); LBI.deleteValue(SI); } @@ -404,6 +420,10 @@ void PromoteMem2Reg::run() { // The alloca has been processed, move on. RemoveFromAllocasList(AllocaNum); + // The alloca's debuginfo can be removed as well. + if (DbgDeclareInst *DDI = Info.DbgDeclare) + DDI->eraseFromParent(); + ++NumLocalPromoted; continue; } @@ -421,6 +441,9 @@ void PromoteMem2Reg::run() { // stored into the alloca. if (AST) PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal; + + // Remember the dbg.declare intrinsic describing this alloca, if any. + if (Info.DbgDeclare) AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare; // Keep the reverse mapping of the 'Allocas' array for the rename pass. AllocaLookup[Allocas[AllocaNum]] = AllocaNum; @@ -476,7 +499,11 @@ void PromoteMem2Reg::run() { A->eraseFromParent(); } - + // Remove alloca's dbg.declare instrinsics from the function. + for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i) + if (DbgDeclareInst *DDI = AllocaDbgDeclares[i]) + DDI->eraseFromParent(); + // Loop over all of the PHI nodes and see if there are any that we can get // rid of because they merge all of the same incoming values. This can // happen due to undef values coming into the PHI nodes. This process is @@ -857,14 +884,19 @@ void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info, // Inserts a llvm.dbg.value instrinsic before the stores to an alloca'd value // that has an associated llvm.dbg.decl intrinsic. void PromoteMem2Reg::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, - StoreInst* SI, - uint64_t Offset) { - if (!DDI) return; + StoreInst *SI) { + DIVariable DIVar(DDI->getVariable()); + if (!DIVar.getNode()) + return; if (!DIF) DIF = new DIFactory(*SI->getParent()->getParent()->getParent()); - DIF->InsertDbgValueIntrinsic(SI->getOperand(0), Offset, - DIVariable(DDI->getVariable()), SI); + Instruction *DbgVal = DIF->InsertDbgValueIntrinsic(SI->getOperand(0), 0, + DIVar, SI); + + // Propagate any debug metadata from the store onto the dbg.value. + if (MDNode *SIMD = SI->getMetadata("dbg")) + DbgVal->setMetadata("dbg", SIMD); } // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific @@ -980,7 +1012,8 @@ NextIteration: // what value were we writing? IncomingVals[ai->second] = SI->getOperand(0); // Record debuginfo for the store before removing it. - ConvertDebugDeclareToDebugValue(findDbgDeclare(Dest), SI, 0); + if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second]) + ConvertDebugDeclareToDebugValue(DDI, SI); BB->getInstList().erase(SI); } } diff --git a/lib/Transforms/Utils/SSAUpdater.cpp b/lib/Transforms/Utils/SSAUpdater.cpp index 161bf21..a31235a 100644 --- a/lib/Transforms/Utils/SSAUpdater.cpp +++ b/lib/Transforms/Utils/SSAUpdater.cpp @@ -71,6 +71,50 @@ void SSAUpdater::AddAvailableValue(BasicBlock *BB, Value *V) { getAvailableVals(AV)[BB] = V; } +/// IsEquivalentPHI - Check if PHI has the same incoming value as specified +/// in ValueMapping for each predecessor block. +static bool IsEquivalentPHI(PHINode *PHI, + DenseMap<BasicBlock*, Value*> &ValueMapping) { + unsigned PHINumValues = PHI->getNumIncomingValues(); + if (PHINumValues != ValueMapping.size()) + return false; + + // Scan the phi to see if it matches. + for (unsigned i = 0, e = PHINumValues; i != e; ++i) + if (ValueMapping[PHI->getIncomingBlock(i)] != + PHI->getIncomingValue(i)) { + return false; + } + + return true; +} + +/// GetExistingPHI - Check if BB already contains a phi node that is equivalent +/// to the specified mapping from predecessor blocks to incoming values. +static Value *GetExistingPHI(BasicBlock *BB, + DenseMap<BasicBlock*, Value*> &ValueMapping) { + PHINode *SomePHI; + for (BasicBlock::iterator It = BB->begin(); + (SomePHI = dyn_cast<PHINode>(It)); ++It) { + if (IsEquivalentPHI(SomePHI, ValueMapping)) + return SomePHI; + } + return 0; +} + +/// GetExistingPHI - Check if BB already contains an equivalent phi node. +/// The InputIt type must be an iterator over std::pair<BasicBlock*, Value*> +/// objects that specify the mapping from predecessor blocks to incoming values. +template<typename InputIt> +static Value *GetExistingPHI(BasicBlock *BB, const InputIt &I, + const InputIt &E) { + // Avoid create the mapping if BB has no phi nodes at all. + if (!isa<PHINode>(BB->begin())) + return 0; + DenseMap<BasicBlock*, Value*> ValueMapping(I, E); + return GetExistingPHI(BB, ValueMapping); +} + /// GetValueAtEndOfBlock - Construct SSA form, materializing a value that is /// live at the end of the specified block. Value *SSAUpdater::GetValueAtEndOfBlock(BasicBlock *BB) { @@ -149,28 +193,11 @@ Value *SSAUpdater::GetValueInMiddleOfBlock(BasicBlock *BB) { if (SingularValue != 0) return SingularValue; - // Otherwise, we do need a PHI: check to see if we already have one available - // in this block that produces the right value. - if (isa<PHINode>(BB->begin())) { - DenseMap<BasicBlock*, Value*> ValueMapping(PredValues.begin(), - PredValues.end()); - PHINode *SomePHI; - for (BasicBlock::iterator It = BB->begin(); - (SomePHI = dyn_cast<PHINode>(It)); ++It) { - // Scan this phi to see if it is what we need. - bool Equal = true; - for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) - if (ValueMapping[SomePHI->getIncomingBlock(i)] != - SomePHI->getIncomingValue(i)) { - Equal = false; - break; - } - - if (Equal) - return SomePHI; - } - } - + // Otherwise, we do need a PHI. + if (Value *ExistingPHI = GetExistingPHI(BB, PredValues.begin(), + PredValues.end())) + return ExistingPHI; + // Ok, we have no way out, insert a new one now. PHINode *InsertedPHI = PHINode::Create(PrototypeValue->getType(), PrototypeValue->getName(), @@ -255,7 +282,7 @@ Value *SSAUpdater::GetValueAtEndOfBlockInternal(BasicBlock *BB) { // producing the same value. If so, this value will capture it, if not, it // will get reset to null. We distinguish the no-predecessor case explicitly // below. - TrackingVH<Value> SingularValue; + TrackingVH<Value> ExistingValue; // We can get our predecessor info by walking the pred_iterator list, but it // is relatively slow. If we already have PHI nodes in this block, walk one @@ -266,11 +293,11 @@ Value *SSAUpdater::GetValueAtEndOfBlockInternal(BasicBlock *BB) { Value *PredVal = GetValueAtEndOfBlockInternal(PredBB); IncomingPredInfo.push_back(std::make_pair(PredBB, PredVal)); - // Compute SingularValue. + // Set ExistingValue to singular value from all predecessors so far. if (i == 0) - SingularValue = PredVal; - else if (PredVal != SingularValue) - SingularValue = 0; + ExistingValue = PredVal; + else if (PredVal != ExistingValue) + ExistingValue = 0; } } else { bool isFirstPred = true; @@ -279,12 +306,12 @@ Value *SSAUpdater::GetValueAtEndOfBlockInternal(BasicBlock *BB) { Value *PredVal = GetValueAtEndOfBlockInternal(PredBB); IncomingPredInfo.push_back(std::make_pair(PredBB, PredVal)); - // Compute SingularValue. + // Set ExistingValue to singular value from all predecessors so far. if (isFirstPred) { - SingularValue = PredVal; + ExistingValue = PredVal; isFirstPred = false; - } else if (PredVal != SingularValue) - SingularValue = 0; + } else if (PredVal != ExistingValue) + ExistingValue = 0; } } @@ -300,31 +327,38 @@ Value *SSAUpdater::GetValueAtEndOfBlockInternal(BasicBlock *BB) { /// above. TrackingVH<Value> &InsertedVal = AvailableVals[BB]; - // If all the predecessor values are the same then we don't need to insert a + // If the predecessor values are not all the same, then check to see if there + // is an existing PHI that can be used. + if (!ExistingValue) + ExistingValue = GetExistingPHI(BB, + IncomingPredInfo.begin()+FirstPredInfoEntry, + IncomingPredInfo.end()); + + // If there is an existing value we can use, then we don't need to insert a // PHI. This is the simple and common case. - if (SingularValue) { - // If a PHI node got inserted, replace it with the singlar value and delete + if (ExistingValue) { + // If a PHI node got inserted, replace it with the existing value and delete // it. if (InsertedVal) { PHINode *OldVal = cast<PHINode>(InsertedVal); // Be careful about dead loops. These RAUW's also update InsertedVal. - if (InsertedVal != SingularValue) - OldVal->replaceAllUsesWith(SingularValue); + if (InsertedVal != ExistingValue) + OldVal->replaceAllUsesWith(ExistingValue); else OldVal->replaceAllUsesWith(UndefValue::get(InsertedVal->getType())); OldVal->eraseFromParent(); } else { - InsertedVal = SingularValue; + InsertedVal = ExistingValue; } - // Either path through the 'if' should have set insertedVal -> SingularVal. - assert((InsertedVal == SingularValue || isa<UndefValue>(InsertedVal)) && - "RAUW didn't change InsertedVal to be SingularVal"); + // Either path through the 'if' should have set InsertedVal -> ExistingVal. + assert((InsertedVal == ExistingValue || isa<UndefValue>(InsertedVal)) && + "RAUW didn't change InsertedVal to be ExistingValue"); // Drop the entries we added in IncomingPredInfo to restore the stack. IncomingPredInfo.erase(IncomingPredInfo.begin()+FirstPredInfoEntry, IncomingPredInfo.end()); - return SingularValue; + return ExistingValue; } // Otherwise, we do need a PHI: insert one now if we don't already have one. diff --git a/lib/Transforms/Utils/SimplifyCFG.cpp b/lib/Transforms/Utils/SimplifyCFG.cpp index cb53296..2215059 100644 --- a/lib/Transforms/Utils/SimplifyCFG.cpp +++ b/lib/Transforms/Utils/SimplifyCFG.cpp @@ -23,6 +23,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Analysis/ConstantFolding.h" +#include "llvm/Target/TargetData.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallVector.h" @@ -36,6 +37,28 @@ using namespace llvm; STATISTIC(NumSpeculations, "Number of speculative executed instructions"); +namespace { +class SimplifyCFGOpt { + const TargetData *const TD; + + ConstantInt *GetConstantInt(Value *V); + Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values); + Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values); + bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, + std::vector<ConstantInt*> &Values); + Value *isValueEqualityComparison(TerminatorInst *TI); + BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, + std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases); + bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, + BasicBlock *Pred); + bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI); + +public: + explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {} + bool run(BasicBlock *BB); +}; +} + /// SafeToMergeTerminators - Return true if it is safe to merge these two /// terminator instructions together. /// @@ -243,17 +266,48 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, return true; } +/// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr +/// and PointerNullValue. Return NULL if value is not a constant int. +ConstantInt *SimplifyCFGOpt::GetConstantInt(Value *V) { + // Normal constant int. + ConstantInt *CI = dyn_cast<ConstantInt>(V); + if (CI || !TD || !isa<Constant>(V) || !isa<PointerType>(V->getType())) + return CI; + + // This is some kind of pointer constant. Turn it into a pointer-sized + // ConstantInt if possible. + const IntegerType *PtrTy = TD->getIntPtrType(V->getContext()); + + // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*). + if (isa<ConstantPointerNull>(V)) + return ConstantInt::get(PtrTy, 0); + + // IntToPtr const int. + if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) + if (CE->getOpcode() == Instruction::IntToPtr) + if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) { + // The constant is very likely to have the right type already. + if (CI->getType() == PtrTy) + return CI; + else + return cast<ConstantInt> + (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false)); + } + return 0; +} + /// GatherConstantSetEQs - Given a potentially 'or'd together collection of /// icmp_eq instructions that compare a value against a constant, return the /// value being compared, and stick the constant into the Values vector. -static Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values){ +Value *SimplifyCFGOpt:: +GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values) { if (Instruction *Inst = dyn_cast<Instruction>(V)) { if (Inst->getOpcode() == Instruction::ICmp && cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_EQ) { - if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { + if (ConstantInt *C = GetConstantInt(Inst->getOperand(1))) { Values.push_back(C); return Inst->getOperand(0); - } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { + } else if (ConstantInt *C = GetConstantInt(Inst->getOperand(0))) { Values.push_back(C); return Inst->getOperand(1); } @@ -270,14 +324,15 @@ static Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values){ /// GatherConstantSetNEs - Given a potentially 'and'd together collection of /// setne instructions that compare a value against a constant, return the value /// being compared, and stick the constant into the Values vector. -static Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values){ +Value *SimplifyCFGOpt:: +GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values) { if (Instruction *Inst = dyn_cast<Instruction>(V)) { if (Inst->getOpcode() == Instruction::ICmp && cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_NE) { - if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { + if (ConstantInt *C = GetConstantInt(Inst->getOperand(1))) { Values.push_back(C); return Inst->getOperand(0); - } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { + } else if (ConstantInt *C = GetConstantInt(Inst->getOperand(0))) { Values.push_back(C); return Inst->getOperand(1); } @@ -294,8 +349,8 @@ static Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values){ /// GatherValueComparisons - If the specified Cond is an 'and' or 'or' of a /// bunch of comparisons of one value against constants, return the value and /// the constants being compared. -static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, - std::vector<ConstantInt*> &Values) { +bool SimplifyCFGOpt::GatherValueComparisons(Instruction *Cond, Value *&CompVal, + std::vector<ConstantInt*> &Values) { if (Cond->getOpcode() == Instruction::Or) { CompVal = GatherConstantSetEQs(Cond, Values); @@ -327,29 +382,32 @@ static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { /// isValueEqualityComparison - Return true if the specified terminator checks /// to see if a value is equal to constant integer value. -static Value *isValueEqualityComparison(TerminatorInst *TI) { +Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { + Value *CV = 0; if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { // Do not permit merging of large switch instructions into their // predecessors unless there is only one predecessor. - if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()), - pred_end(SI->getParent())) > 128) - return 0; - - return SI->getCondition(); - } - if (BranchInst *BI = dyn_cast<BranchInst>(TI)) + if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()), + pred_end(SI->getParent())) <= 128) + CV = SI->getCondition(); + } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) if (BI->isConditional() && BI->getCondition()->hasOneUse()) if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || ICI->getPredicate() == ICmpInst::ICMP_NE) && - isa<ConstantInt>(ICI->getOperand(1))) - return ICI->getOperand(0); - return 0; + GetConstantInt(ICI->getOperand(1))) + CV = ICI->getOperand(0); + + // Unwrap any lossless ptrtoint cast. + if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext())) + if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) + CV = PTII->getOperand(0); + return CV; } /// GetValueEqualityComparisonCases - Given a value comparison instruction, /// decode all of the 'cases' that it represents and return the 'default' block. -static BasicBlock * +BasicBlock *SimplifyCFGOpt:: GetValueEqualityComparisonCases(TerminatorInst *TI, std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { @@ -362,7 +420,7 @@ GetValueEqualityComparisonCases(TerminatorInst *TI, BranchInst *BI = cast<BranchInst>(TI); ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); - Cases.push_back(std::make_pair(cast<ConstantInt>(ICI->getOperand(1)), + Cases.push_back(std::make_pair(GetConstantInt(ICI->getOperand(1)), BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE))); return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); @@ -421,8 +479,9 @@ ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, /// comparison with the same value, and if that comparison determines the /// outcome of this comparison. If so, simplify TI. This does a very limited /// form of jump threading. -static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, - BasicBlock *Pred) { +bool SimplifyCFGOpt:: +SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, + BasicBlock *Pred) { Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); if (!PredVal) return false; // Not a value comparison in predecessor. @@ -548,7 +607,7 @@ namespace { /// 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 /// on the same value. If so, and if safe to do so, fold them together. -static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { +bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { BasicBlock *BB = TI->getParent(); Value *CV = isValueEqualityComparison(TI); // CondVal assert(CV && "Not a comparison?"); @@ -641,6 +700,13 @@ static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) AddPredecessorToBlock(NewSuccessors[i], Pred, BB); + // Convert pointer to int before we switch. + if (isa<PointerType>(CV->getType())) { + assert(TD && "Cannot switch on pointer without TargetData"); + CV = new PtrToIntInst(CV, TD->getIntPtrType(CV->getContext()), + "magicptr", PTI); + } + // Now that the successors are updated, create the new Switch instruction. SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, PredCases.size(), PTI); @@ -1011,7 +1077,7 @@ static bool FoldCondBranchOnPHI(BranchInst *BI) { for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { ConstantInt *CB; if ((CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i))) && - CB->getType()->isInteger(1)) { + CB->getType()->isIntegerTy(1)) { // Okay, we now know that all edges from PredBB should be revectored to // branch to RealDest. BasicBlock *PredBB = PN->getIncomingBlock(i); @@ -1589,14 +1655,7 @@ static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { return true; } -/// SimplifyCFG - This function is used to do simplification of a CFG. For -/// example, it adjusts branches to branches to eliminate the extra hop, it -/// eliminates unreachable basic blocks, and does other "peephole" optimization -/// of the CFG. It returns true if a modification was made. -/// -/// WARNING: The entry node of a function may not be simplified. -/// -bool llvm::SimplifyCFG(BasicBlock *BB) { +bool SimplifyCFGOpt::run(BasicBlock *BB) { bool Changed = false; Function *M = BB->getParent(); @@ -1997,7 +2056,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { Value *CompVal = 0; std::vector<ConstantInt*> Values; bool TrueWhenEqual = GatherValueComparisons(Cond, CompVal, Values); - if (CompVal && CompVal->getType()->isInteger()) { + if (CompVal) { // There might be duplicate constants in the list, which the switch // instruction can't handle, remove them now. std::sort(Values.begin(), Values.end(), ConstantIntOrdering()); @@ -2008,6 +2067,14 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { BasicBlock *EdgeBB = BI->getSuccessor(0); if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); + // Convert pointer to int before we switch. + if (isa<PointerType>(CompVal->getType())) { + assert(TD && "Cannot switch on pointer without TargetData"); + CompVal = new PtrToIntInst(CompVal, + TD->getIntPtrType(CompVal->getContext()), + "magicptr", BI); + } + // Create the new switch instruction now. SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, Values.size(), BI); @@ -2035,3 +2102,14 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { return Changed; } + +/// SimplifyCFG - This function is used to do simplification of a CFG. For +/// example, it adjusts branches to branches to eliminate the extra hop, it +/// eliminates unreachable basic blocks, and does other "peephole" optimization +/// of the CFG. It returns true if a modification was made. +/// +/// WARNING: The entry node of a function may not be simplified. +/// +bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) { + return SimplifyCFGOpt(TD).run(BB); +} diff --git a/lib/Transforms/Utils/ValueMapper.cpp b/lib/Transforms/Utils/ValueMapper.cpp index a6e6701..6045048 100644 --- a/lib/Transforms/Utils/ValueMapper.cpp +++ b/lib/Transforms/Utils/ValueMapper.cpp @@ -35,7 +35,7 @@ Value *llvm::MapValue(const Value *V, ValueMapTy &VM) { if (const MDNode *MD = dyn_cast<MDNode>(V)) { SmallVector<Value*, 4> Elts; - for (unsigned i = 0; i != MD->getNumOperands(); i++) + for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i) Elts.push_back(MD->getOperand(i) ? MapValue(MD->getOperand(i), VM) : 0); return VM[V] = MDNode::get(V->getContext(), Elts.data(), Elts.size()); } |
