diff options
Diffstat (limited to 'lib/Transforms')
-rw-r--r-- | lib/Transforms/IPO/Internalize.cpp | 4 | ||||
-rw-r--r-- | lib/Transforms/Scalar/IndVarSimplify.cpp | 307 | ||||
-rw-r--r-- | lib/Transforms/Scalar/LoopUnswitch.cpp | 68 | ||||
-rw-r--r-- | lib/Transforms/Utils/SSAUpdater.cpp | 497 |
4 files changed, 396 insertions, 480 deletions
diff --git a/lib/Transforms/IPO/Internalize.cpp b/lib/Transforms/IPO/Internalize.cpp index 3d31932..47abb7d 100644 --- a/lib/Transforms/IPO/Internalize.cpp +++ b/lib/Transforms/IPO/Internalize.cpp @@ -156,6 +156,8 @@ bool InternalizePass::runOnModule(Module &M) { for (Module::global_iterator I = M.global_begin(), E = M.global_end(); I != E; ++I) if (!I->isDeclaration() && !I->hasLocalLinkage() && + // Available externally is really just a "declaration with a body". + !I->hasAvailableExternallyLinkage() && !ExternalNames.count(I->getName())) { I->setLinkage(GlobalValue::InternalLinkage); Changed = true; @@ -167,6 +169,8 @@ bool InternalizePass::runOnModule(Module &M) { for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); I != E; ++I) if (!I->isDeclaration() && !I->hasInternalLinkage() && + // Available externally is really just a "declaration with a body". + !I->hasAvailableExternallyLinkage() && !ExternalNames.count(I->getName())) { I->setLinkage(GlobalValue::InternalLinkage); Changed = true; diff --git a/lib/Transforms/Scalar/IndVarSimplify.cpp b/lib/Transforms/Scalar/IndVarSimplify.cpp index 988a4cb..6605666 100644 --- a/lib/Transforms/Scalar/IndVarSimplify.cpp +++ b/lib/Transforms/Scalar/IndVarSimplify.cpp @@ -510,6 +510,13 @@ void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) { // Now expand it into actual Instructions and patch it into place. Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt); + // Inform ScalarEvolution that this value is changing. The change doesn't + // affect its value, but it does potentially affect which use lists the + // value will be on after the replacement, which affects ScalarEvolution's + // ability to walk use lists and drop dangling pointers when a value is + // deleted. + SE->forgetValue(User); + // Patch the new value into place. if (Op->hasName()) NewVal->takeName(Op); @@ -616,36 +623,18 @@ void IndVarSimplify::SinkUnusedInvariants(Loop *L) { } } -/// Return true if it is OK to use SIToFPInst for an induction variable -/// with given initial and exit values. -static bool useSIToFPInst(ConstantFP &InitV, ConstantFP &ExitV, - uint64_t intIV, uint64_t intEV) { - - if (InitV.getValueAPF().isNegative() || ExitV.getValueAPF().isNegative()) - return true; - - // If the iteration range can be handled by SIToFPInst then use it. - APInt Max = APInt::getSignedMaxValue(32); - if (Max.getZExtValue() > static_cast<uint64_t>(abs64(intEV - intIV))) - return true; - - return false; -} - -/// convertToInt - Convert APF to an integer, if possible. -static bool convertToInt(const APFloat &APF, uint64_t *intVal) { - +/// ConvertToSInt - Convert APF to an integer, if possible. +static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { bool isExact = false; if (&APF.getSemantics() == &APFloat::PPCDoubleDouble) return false; - if (APF.convertToInteger(intVal, 32, APF.isNegative(), - APFloat::rmTowardZero, &isExact) - != APFloat::opOK) - return false; - if (!isExact) + // See if we can convert this to an int64_t + uint64_t UIntVal; + if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, + &isExact) != APFloat::opOK || !isExact) return false; + IntVal = UIntVal; return true; - } /// HandleFloatingPointIV - If the loop has floating induction variable @@ -657,144 +646,200 @@ static bool convertToInt(const APFloat &APF, uint64_t *intVal) { /// for(int i = 0; i < 10000; ++i) /// bar((double)i); /// -void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PH) { - - unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0)); +void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { + unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); unsigned BackEdge = IncomingEdge^1; // Check incoming value. - ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge)); - if (!InitValue) return; - uint64_t newInitValue = - Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits(); - if (!convertToInt(InitValue->getValueAPF(), &newInitValue)) + ConstantFP *InitValueVal = + dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); + + int64_t InitValue; + if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) return; - // Check IV increment. Reject this PH if increment operation is not + // Check IV increment. Reject this PN if increment operation is not // an add or increment value can not be represented by an integer. BinaryOperator *Incr = - dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge)); - if (!Incr) return; - if (Incr->getOpcode() != Instruction::FAdd) return; - ConstantFP *IncrValue = NULL; - unsigned IncrVIndex = 1; - if (Incr->getOperand(1) == PH) - IncrVIndex = 0; - IncrValue = dyn_cast<ConstantFP>(Incr->getOperand(IncrVIndex)); - if (!IncrValue) return; - uint64_t newIncrValue = - Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits(); - if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue)) + dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); + if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return; + + // If this is not an add of the PHI with a constantfp, or if the constant fp + // is not an integer, bail out. + ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1)); + int64_t IncValue; + if (IncValueVal == 0 || Incr->getOperand(0) != PN || + !ConvertToSInt(IncValueVal->getValueAPF(), IncValue)) return; - // Check Incr uses. One user is PH and the other users is exit condition used - // by the conditional terminator. + // Check Incr uses. One user is PN and the other user is an exit condition + // used by the conditional terminator. Value::use_iterator IncrUse = Incr->use_begin(); Instruction *U1 = cast<Instruction>(IncrUse++); if (IncrUse == Incr->use_end()) return; Instruction *U2 = cast<Instruction>(IncrUse++); if (IncrUse != Incr->use_end()) return; - // Find exit condition. - FCmpInst *EC = dyn_cast<FCmpInst>(U1); - if (!EC) - EC = dyn_cast<FCmpInst>(U2); - if (!EC) return; - - if (BranchInst *BI = dyn_cast<BranchInst>(EC->getParent()->getTerminator())) { - if (!BI->isConditional()) return; - if (BI->getCondition() != EC) return; - } - - // Find exit value. If exit value can not be represented as an integer then - // do not handle this floating point PH. - ConstantFP *EV = NULL; - unsigned EVIndex = 1; - if (EC->getOperand(1) == Incr) - EVIndex = 0; - EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex)); - if (!EV) return; - uint64_t intEV = Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits(); - if (!convertToInt(EV->getValueAPF(), &intEV)) + // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't + // only used by a branch, we can't transform it. + FCmpInst *Compare = dyn_cast<FCmpInst>(U1); + if (!Compare) + Compare = dyn_cast<FCmpInst>(U2); + if (Compare == 0 || !Compare->hasOneUse() || + !isa<BranchInst>(Compare->use_back())) return; - + + BranchInst *TheBr = cast<BranchInst>(Compare->use_back()); + + // We need to verify that the branch actually controls the iteration count + // of the loop. If not, the new IV can overflow and no one will notice. + // The branch block must be in the loop and one of the successors must be out + // of the loop. + assert(TheBr->isConditional() && "Can't use fcmp if not conditional"); + if (!L->contains(TheBr->getParent()) || + (L->contains(TheBr->getSuccessor(0)) && + L->contains(TheBr->getSuccessor(1)))) + return; + + + // If it isn't a comparison with an integer-as-fp (the exit value), we can't + // transform it. + ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1)); + int64_t ExitValue; + if (ExitValueVal == 0 || + !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue)) + return; + // Find new predicate for integer comparison. CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE; - switch (EC->getPredicate()) { + switch (Compare->getPredicate()) { + default: return; // Unknown comparison. case CmpInst::FCMP_OEQ: - case CmpInst::FCMP_UEQ: - NewPred = CmpInst::ICMP_EQ; - break; + case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break; + case CmpInst::FCMP_ONE: + case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break; case CmpInst::FCMP_OGT: - case CmpInst::FCMP_UGT: - NewPred = CmpInst::ICMP_UGT; - break; + case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break; case CmpInst::FCMP_OGE: - case CmpInst::FCMP_UGE: - NewPred = CmpInst::ICMP_UGE; - break; + case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break; case CmpInst::FCMP_OLT: - case CmpInst::FCMP_ULT: - NewPred = CmpInst::ICMP_ULT; - break; + case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break; case CmpInst::FCMP_OLE: - case CmpInst::FCMP_ULE: - NewPred = CmpInst::ICMP_ULE; - break; - default: - break; + case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break; } - if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return; + + // We convert the floating point induction variable to a signed i32 value if + // we can. This is only safe if the comparison will not overflow in a way + // that won't be trapped by the integer equivalent operations. Check for this + // now. + // TODO: We could use i64 if it is native and the range requires it. + + // The start/stride/exit values must all fit in signed i32. + if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue)) + return; + + // If not actually striding (add x, 0.0), avoid touching the code. + if (IncValue == 0) + return; + + // Positive and negative strides have different safety conditions. + if (IncValue > 0) { + // If we have a positive stride, we require the init to be less than the + // exit value and an equality or less than comparison. + if (InitValue >= ExitValue || + NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE) + return; + + uint32_t Range = uint32_t(ExitValue-InitValue); + if (NewPred == CmpInst::ICMP_SLE) { + // Normalize SLE -> SLT, check for infinite loop. + if (++Range == 0) return; // Range overflows. + } + + unsigned Leftover = Range % uint32_t(IncValue); + + // If this is an equality comparison, we require that the strided value + // exactly land on the exit value, otherwise the IV condition will wrap + // around and do things the fp IV wouldn't. + if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && + Leftover != 0) + return; + + // If the stride would wrap around the i32 before exiting, we can't + // transform the IV. + if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue) + return; + + } else { + // If we have a negative stride, we require the init to be greater than the + // exit value and an equality or greater than comparison. + if (InitValue >= ExitValue || + NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE) + return; + + uint32_t Range = uint32_t(InitValue-ExitValue); + if (NewPred == CmpInst::ICMP_SGE) { + // Normalize SGE -> SGT, check for infinite loop. + if (++Range == 0) return; // Range overflows. + } + + unsigned Leftover = Range % uint32_t(-IncValue); + + // If this is an equality comparison, we require that the strided value + // exactly land on the exit value, otherwise the IV condition will wrap + // around and do things the fp IV wouldn't. + if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && + Leftover != 0) + return; + + // If the stride would wrap around the i32 before exiting, we can't + // transform the IV. + if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue) + return; + } + + const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext()); // Insert new integer induction variable. - PHINode *NewPHI = PHINode::Create(Type::getInt32Ty(PH->getContext()), - PH->getName()+".int", PH); - NewPHI->addIncoming(ConstantInt::get(Type::getInt32Ty(PH->getContext()), - newInitValue), - PH->getIncomingBlock(IncomingEdge)); - - Value *NewAdd = BinaryOperator::CreateAdd(NewPHI, - ConstantInt::get(Type::getInt32Ty(PH->getContext()), - newIncrValue), - Incr->getName()+".int", Incr); - NewPHI->addIncoming(NewAdd, PH->getIncomingBlock(BackEdge)); - - // The back edge is edge 1 of newPHI, whatever it may have been in the - // original PHI. - ConstantInt *NewEV = ConstantInt::get(Type::getInt32Ty(PH->getContext()), - intEV); - Value *LHS = (EVIndex == 1 ? NewPHI->getIncomingValue(1) : NewEV); - Value *RHS = (EVIndex == 1 ? NewEV : NewPHI->getIncomingValue(1)); - ICmpInst *NewEC = new ICmpInst(EC->getParent()->getTerminator(), - NewPred, LHS, RHS, EC->getName()); - - // In the following deletions, PH may become dead and may be deleted. + PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN); + NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue), + PN->getIncomingBlock(IncomingEdge)); + + Value *NewAdd = + BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue), + Incr->getName()+".int", Incr); + NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge)); + + ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd, + ConstantInt::get(Int32Ty, ExitValue), + Compare->getName()); + + // In the following deletions, PN may become dead and may be deleted. // Use a WeakVH to observe whether this happens. - WeakVH WeakPH = PH; + WeakVH WeakPH = PN; - // Delete old, floating point, exit comparison instruction. - NewEC->takeName(EC); - EC->replaceAllUsesWith(NewEC); - RecursivelyDeleteTriviallyDeadInstructions(EC); + // Delete the old floating point exit comparison. The branch starts using the + // new comparison. + NewCompare->takeName(Compare); + Compare->replaceAllUsesWith(NewCompare); + RecursivelyDeleteTriviallyDeadInstructions(Compare); - // Delete old, floating point, increment instruction. + // Delete the old floating point increment. Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); RecursivelyDeleteTriviallyDeadInstructions(Incr); - // Replace floating induction variable, if it isn't already deleted. - // Give SIToFPInst preference over UIToFPInst because it is faster on - // platforms that are widely used. - if (WeakPH && !PH->use_empty()) { - if (useSIToFPInst(*InitValue, *EV, newInitValue, intEV)) { - SIToFPInst *Conv = new SIToFPInst(NewPHI, PH->getType(), "indvar.conv", - PH->getParent()->getFirstNonPHI()); - PH->replaceAllUsesWith(Conv); - } else { - UIToFPInst *Conv = new UIToFPInst(NewPHI, PH->getType(), "indvar.conv", - PH->getParent()->getFirstNonPHI()); - PH->replaceAllUsesWith(Conv); - } - RecursivelyDeleteTriviallyDeadInstructions(PH); + // If the FP induction variable still has uses, this is because something else + // in the loop uses its value. In order to canonicalize the induction + // variable, we chose to eliminate the IV and rewrite it in terms of an + // int->fp cast. + // + // We give preference to sitofp over uitofp because it is faster on most + // platforms. + if (WeakPH) { + Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", + PN->getParent()->getFirstNonPHI()); + PN->replaceAllUsesWith(Conv); + RecursivelyDeleteTriviallyDeadInstructions(PN); } // Add a new IVUsers entry for the newly-created integer PHI. diff --git a/lib/Transforms/Scalar/LoopUnswitch.cpp b/lib/Transforms/Scalar/LoopUnswitch.cpp index e3b809e..27fd2ef 100644 --- a/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -415,46 +415,44 @@ bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { Function *F = loopHeader->getParent(); - // If the condition is trivial, always unswitch. There is no code growth for - // this case. - if (!IsTrivialUnswitchCondition(LoopCond)) { - // Check to see if it would be profitable to unswitch current loop. + Constant *CondVal = 0; + BasicBlock *ExitBlock = 0; + if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { + // If the condition is trivial, always unswitch. There is no code growth + // for this case. + UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); + return true; + } - // Do not do non-trivial unswitch while optimizing for size. - if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) - return false; + // Check to see if it would be profitable to unswitch current loop. - // FIXME: This is overly conservative because it does not take into - // consideration code simplification opportunities and code that can - // be shared by the resultant unswitched loops. - CodeMetrics Metrics; - for (Loop::block_iterator I = currentLoop->block_begin(), - E = currentLoop->block_end(); - I != E; ++I) - Metrics.analyzeBasicBlock(*I); - - // Limit the number of instructions to avoid causing significant code - // expansion, and the number of basic blocks, to avoid loops with - // large numbers of branches which cause loop unswitching to go crazy. - // This is a very ad-hoc heuristic. - if (Metrics.NumInsts > Threshold || - Metrics.NumBlocks * 5 > Threshold || - Metrics.NeverInline) { - DEBUG(dbgs() << "NOT unswitching loop %" - << currentLoop->getHeader()->getName() << ", cost too high: " - << currentLoop->getBlocks().size() << "\n"); - return false; - } - } + // Do not do non-trivial unswitch while optimizing for size. + if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) + return false; - Constant *CondVal; - BasicBlock *ExitBlock; - if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { - UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); - } else { - UnswitchNontrivialCondition(LoopCond, Val, currentLoop); + // FIXME: This is overly conservative because it does not take into + // consideration code simplification opportunities and code that can + // be shared by the resultant unswitched loops. + CodeMetrics Metrics; + for (Loop::block_iterator I = currentLoop->block_begin(), + E = currentLoop->block_end(); + I != E; ++I) + Metrics.analyzeBasicBlock(*I); + + // Limit the number of instructions to avoid causing significant code + // expansion, and the number of basic blocks, to avoid loops with + // large numbers of branches which cause loop unswitching to go crazy. + // This is a very ad-hoc heuristic. + if (Metrics.NumInsts > Threshold || + Metrics.NumBlocks * 5 > Threshold || + Metrics.NeverInline) { + DEBUG(dbgs() << "NOT unswitching loop %" + << currentLoop->getHeader()->getName() << ", cost too high: " + << currentLoop->getBlocks().size() << "\n"); + return false; } + UnswitchNontrivialCondition(LoopCond, Val, currentLoop); return true; } diff --git a/lib/Transforms/Utils/SSAUpdater.cpp b/lib/Transforms/Utils/SSAUpdater.cpp index 292332e..a31235a 100644 --- a/lib/Transforms/Utils/SSAUpdater.cpp +++ b/lib/Transforms/Utils/SSAUpdater.cpp @@ -14,82 +14,31 @@ #include "llvm/Transforms/Utils/SSAUpdater.h" #include "llvm/Instructions.h" #include "llvm/ADT/DenseMap.h" -#include "llvm/Support/AlignOf.h" -#include "llvm/Support/Allocator.h" #include "llvm/Support/CFG.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/ValueHandle.h" #include "llvm/Support/raw_ostream.h" using namespace llvm; -/// BBInfo - Per-basic block information used internally by SSAUpdater. -/// The predecessors of each block are cached here since pred_iterator is -/// slow and we need to iterate over the blocks at least a few times. -class SSAUpdater::BBInfo { -public: - Value *AvailableVal; // Value to use in this block. - BasicBlock *DefBB; // Block that defines the available value. - unsigned NumPreds; // Number of predecessor blocks. - BasicBlock **Preds; // Array[NumPreds] of predecessor blocks. - unsigned Counter; // Marker to identify blocks already visited. - PHINode *PHITag; // Marker for existing PHIs that match. - - BBInfo(BasicBlock *BB, Value *V, BumpPtrAllocator *Allocator); -}; -typedef DenseMap<BasicBlock*, SSAUpdater::BBInfo*> BBMapTy; - -SSAUpdater::BBInfo::BBInfo(BasicBlock *BB, Value *V, - BumpPtrAllocator *Allocator) - : AvailableVal(V), DefBB(0), NumPreds(0), Preds(0), Counter(0), PHITag(0) { - // If this block has a known value, don't bother finding its predecessors. - if (V) { - DefBB = BB; - return; - } - - // 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 - // of them to get the predecessor list instead. - if (PHINode *SomePhi = dyn_cast<PHINode>(BB->begin())) { - NumPreds = SomePhi->getNumIncomingValues(); - Preds = static_cast<BasicBlock**> - (Allocator->Allocate(NumPreds * sizeof(BasicBlock*), - AlignOf<BasicBlock*>::Alignment)); - for (unsigned pi = 0; pi != NumPreds; ++pi) - Preds[pi] = SomePhi->getIncomingBlock(pi); - return; - } - - // Stash the predecessors in a temporary vector until we know how much space - // to allocate for them. - SmallVector<BasicBlock*, 10> TmpPreds; - for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { - TmpPreds.push_back(*PI); - ++NumPreds; - } - Preds = static_cast<BasicBlock**> - (Allocator->Allocate(NumPreds * sizeof(BasicBlock*), - AlignOf<BasicBlock*>::Alignment)); - memcpy(Preds, TmpPreds.data(), NumPreds * sizeof(BasicBlock*)); -} +typedef DenseMap<BasicBlock*, TrackingVH<Value> > AvailableValsTy; +typedef std::vector<std::pair<BasicBlock*, TrackingVH<Value> > > + IncomingPredInfoTy; -typedef DenseMap<BasicBlock*, Value*> AvailableValsTy; static AvailableValsTy &getAvailableVals(void *AV) { return *static_cast<AvailableValsTy*>(AV); } -static BBMapTy *getBBMap(void *BM) { - return static_cast<BBMapTy*>(BM); +static IncomingPredInfoTy &getIncomingPredInfo(void *IPI) { + return *static_cast<IncomingPredInfoTy*>(IPI); } -static BumpPtrAllocator *getAllocator(void *BPA) { - return static_cast<BumpPtrAllocator*>(BPA); -} SSAUpdater::SSAUpdater(SmallVectorImpl<PHINode*> *NewPHI) - : AV(0), PrototypeValue(0), BM(0), BPA(0), InsertedPHIs(NewPHI) {} + : AV(0), PrototypeValue(0), IPI(0), InsertedPHIs(NewPHI) {} SSAUpdater::~SSAUpdater() { delete &getAvailableVals(AV); + delete &getIncomingPredInfo(IPI); } /// Initialize - Reset this object to get ready for a new set of SSA @@ -99,6 +48,11 @@ void SSAUpdater::Initialize(Value *ProtoValue) { AV = new AvailableValsTy(); else getAvailableVals(AV).clear(); + + if (IPI == 0) + IPI = new IncomingPredInfoTy(); + else + getIncomingPredInfo(IPI).clear(); PrototypeValue = ProtoValue; } @@ -119,7 +73,7 @@ void SSAUpdater::AddAvailableValue(BasicBlock *BB, Value *V) { /// IsEquivalentPHI - Check if PHI has the same incoming value as specified /// in ValueMapping for each predecessor block. -static bool IsEquivalentPHI(PHINode *PHI, +static bool IsEquivalentPHI(PHINode *PHI, DenseMap<BasicBlock*, Value*> &ValueMapping) { unsigned PHINumValues = PHI->getNumIncomingValues(); if (PHINumValues != ValueMapping.size()) @@ -135,12 +89,38 @@ static bool IsEquivalentPHI(PHINode *PHI, 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) { - assert(BM == 0 && BPA == 0 && "Unexpected Internal State"); + assert(getIncomingPredInfo(IPI).empty() && "Unexpected Internal State"); Value *Res = GetValueAtEndOfBlockInternal(BB); - assert(BM == 0 && BPA == 0 && "Unexpected Internal State"); + assert(getIncomingPredInfo(IPI).empty() && "Unexpected Internal State"); return Res; } @@ -166,7 +146,7 @@ Value *SSAUpdater::GetValueAtEndOfBlock(BasicBlock *BB) { Value *SSAUpdater::GetValueInMiddleOfBlock(BasicBlock *BB) { // If there is no definition of the renamed variable in this block, just use // GetValueAtEndOfBlock to do our work. - if (!HasValueForBlock(BB)) + if (!getAvailableVals(AV).count(BB)) return GetValueAtEndOfBlock(BB); // Otherwise, we have the hard case. Get the live-in values for each @@ -213,18 +193,10 @@ 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) { - if (IsEquivalentPHI(SomePHI, ValueMapping)) - 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(), @@ -254,7 +226,7 @@ Value *SSAUpdater::GetValueInMiddleOfBlock(BasicBlock *BB) { /// which use their value in the corresponding predecessor. void SSAUpdater::RewriteUse(Use &U) { Instruction *User = cast<Instruction>(U.getUser()); - + Value *V; if (PHINode *UserPN = dyn_cast<PHINode>(User)) V = GetValueAtEndOfBlock(UserPN->getIncomingBlock(U)); @@ -264,264 +236,161 @@ void SSAUpdater::RewriteUse(Use &U) { U.set(V); } + /// GetValueAtEndOfBlockInternal - Check to see if AvailableVals has an entry /// for the specified BB and if so, return it. If not, construct SSA form by -/// first calculating the required placement of PHIs and then inserting new -/// PHIs where needed. +/// walking predecessors inserting PHI nodes as needed until we get to a block +/// where the value is available. +/// Value *SSAUpdater::GetValueAtEndOfBlockInternal(BasicBlock *BB) { AvailableValsTy &AvailableVals = getAvailableVals(AV); - if (Value *V = AvailableVals[BB]) - return V; - - // Pool allocation used internally by GetValueAtEndOfBlock. - BumpPtrAllocator AllocatorObj; - BBMapTy BBMapObj; - BPA = &AllocatorObj; - BM = &BBMapObj; - - BBInfo *Info = new (AllocatorObj) BBInfo(BB, 0, &AllocatorObj); - BBMapObj[BB] = Info; - - bool Changed; - unsigned Counter = 1; - do { - Changed = false; - FindPHIPlacement(BB, Info, Changed, Counter); - ++Counter; - } while (Changed); - - FindAvailableVal(BB, Info, Counter); - - BPA = 0; - BM = 0; - return Info->AvailableVal; -} -/// FindPHIPlacement - Recursively visit the predecessors of a block to find -/// the reaching definition for each predecessor and then determine whether -/// a PHI is needed in this block. -void SSAUpdater::FindPHIPlacement(BasicBlock *BB, BBInfo *Info, bool &Changed, - unsigned Counter) { - AvailableValsTy &AvailableVals = getAvailableVals(AV); - BBMapTy *BBMap = getBBMap(BM); - BumpPtrAllocator *Allocator = getAllocator(BPA); - bool BBNeedsPHI = false; - BasicBlock *SamePredDefBB = 0; - - // If there are no predecessors, then we must have found an unreachable - // block. Treat it as a definition with 'undef'. - if (Info->NumPreds == 0) { - Info->AvailableVal = UndefValue::get(PrototypeValue->getType()); - Info->DefBB = BB; - return; + // Query AvailableVals by doing an insertion of null. + std::pair<AvailableValsTy::iterator, bool> InsertRes = + AvailableVals.insert(std::make_pair(BB, TrackingVH<Value>())); + + // Handle the case when the insertion fails because we have already seen BB. + if (!InsertRes.second) { + // If the insertion failed, there are two cases. The first case is that the + // value is already available for the specified block. If we get this, just + // return the value. + if (InsertRes.first->second != 0) + return InsertRes.first->second; + + // Otherwise, if the value we find is null, then this is the value is not + // known but it is being computed elsewhere in our recursion. This means + // that we have a cycle. Handle this by inserting a PHI node and returning + // it. When we get back to the first instance of the recursion we will fill + // in the PHI node. + return InsertRes.first->second = + PHINode::Create(PrototypeValue->getType(), PrototypeValue->getName(), + &BB->front()); } - Info->Counter = Counter; - for (unsigned pi = 0; pi != Info->NumPreds; ++pi) { - BasicBlock *Pred = Info->Preds[pi]; - BBMapTy::value_type &BBMapBucket = BBMap->FindAndConstruct(Pred); - if (!BBMapBucket.second) { - Value *PredVal = AvailableVals.lookup(Pred); - BBMapBucket.second = new (*Allocator) BBInfo(Pred, PredVal, Allocator); - } - BBInfo *PredInfo = BBMapBucket.second; - BasicBlock *DefBB = 0; - if (!PredInfo->AvailableVal) { - if (PredInfo->Counter != Counter) - FindPHIPlacement(Pred, PredInfo, Changed, Counter); - - // Ignore back edges where the value is not yet known. - if (!PredInfo->DefBB) - continue; + // Okay, the value isn't in the map and we just inserted a null in the entry + // to indicate that we're processing the block. Since we have no idea what + // value is in this block, we have to recurse through our predecessors. + // + // While we're walking our predecessors, we keep track of them in a vector, + // then insert a PHI node in the end if we actually need one. We could use a + // smallvector here, but that would take a lot of stack space for every level + // of the recursion, just use IncomingPredInfo as an explicit stack. + IncomingPredInfoTy &IncomingPredInfo = getIncomingPredInfo(IPI); + unsigned FirstPredInfoEntry = IncomingPredInfo.size(); + + // As we're walking the predecessors, keep track of whether they are all + // 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> 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 + // of them to get the predecessor list instead. + if (PHINode *SomePhi = dyn_cast<PHINode>(BB->begin())) { + for (unsigned i = 0, e = SomePhi->getNumIncomingValues(); i != e; ++i) { + BasicBlock *PredBB = SomePhi->getIncomingBlock(i); + Value *PredVal = GetValueAtEndOfBlockInternal(PredBB); + IncomingPredInfo.push_back(std::make_pair(PredBB, PredVal)); + + // Set ExistingValue to singular value from all predecessors so far. + if (i == 0) + ExistingValue = PredVal; + else if (PredVal != ExistingValue) + ExistingValue = 0; } - DefBB = PredInfo->DefBB; + } else { + bool isFirstPred = true; + for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { + BasicBlock *PredBB = *PI; + Value *PredVal = GetValueAtEndOfBlockInternal(PredBB); + IncomingPredInfo.push_back(std::make_pair(PredBB, PredVal)); - if (!SamePredDefBB) - SamePredDefBB = DefBB; - else if (DefBB != SamePredDefBB) - BBNeedsPHI = true; + // Set ExistingValue to singular value from all predecessors so far. + if (isFirstPred) { + ExistingValue = PredVal; + isFirstPred = false; + } else if (PredVal != ExistingValue) + ExistingValue = 0; + } } - BasicBlock *NewDefBB = (BBNeedsPHI ? BB : SamePredDefBB); - if (Info->DefBB != NewDefBB) { - Changed = true; - Info->DefBB = NewDefBB; - } -} + // If there are no predecessors, then we must have found an unreachable block + // just return 'undef'. Since there are no predecessors, InsertRes must not + // be invalidated. + if (IncomingPredInfo.size() == FirstPredInfoEntry) + return InsertRes.first->second = UndefValue::get(PrototypeValue->getType()); + + /// Look up BB's entry in AvailableVals. 'InsertRes' may be invalidated. If + /// this block is involved in a loop, a no-entry PHI node will have been + /// inserted as InsertedVal. Otherwise, we'll still have the null we inserted + /// above. + TrackingVH<Value> &InsertedVal = AvailableVals[BB]; + + // 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 (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 != ExistingValue) + OldVal->replaceAllUsesWith(ExistingValue); + else + OldVal->replaceAllUsesWith(UndefValue::get(InsertedVal->getType())); + OldVal->eraseFromParent(); + } else { + InsertedVal = ExistingValue; + } -/// FindAvailableVal - If this block requires a PHI, first check if an existing -/// PHI matches the PHI placement and reaching definitions computed earlier, -/// and if not, create a new PHI. Visit all the block's predecessors to -/// calculate the available value for each one and fill in the incoming values -/// for a new PHI. -void SSAUpdater::FindAvailableVal(BasicBlock *BB, BBInfo *Info, - unsigned Counter) { - if (Info->AvailableVal || Info->Counter == Counter) - return; + // Either path through the 'if' should have set InsertedVal -> ExistingVal. + assert((InsertedVal == ExistingValue || isa<UndefValue>(InsertedVal)) && + "RAUW didn't change InsertedVal to be ExistingValue"); - AvailableValsTy &AvailableVals = getAvailableVals(AV); - BBMapTy *BBMap = getBBMap(BM); - - // Check if there needs to be a PHI in BB. - PHINode *NewPHI = 0; - if (Info->DefBB == BB) { - // Look for an existing PHI. - FindExistingPHI(BB); - if (!Info->AvailableVal) { - NewPHI = PHINode::Create(PrototypeValue->getType(), - PrototypeValue->getName(), &BB->front()); - NewPHI->reserveOperandSpace(Info->NumPreds); - Info->AvailableVal = NewPHI; - AvailableVals[BB] = NewPHI; - } + // Drop the entries we added in IncomingPredInfo to restore the stack. + IncomingPredInfo.erase(IncomingPredInfo.begin()+FirstPredInfoEntry, + IncomingPredInfo.end()); + return ExistingValue; } - // Iterate through the block's predecessors. - Info->Counter = Counter; - for (unsigned pi = 0; pi != Info->NumPreds; ++pi) { - BasicBlock *Pred = Info->Preds[pi]; - BBInfo *PredInfo = (*BBMap)[Pred]; - FindAvailableVal(Pred, PredInfo, Counter); - if (NewPHI) { - // Skip to the nearest preceding definition. - if (PredInfo->DefBB != Pred) - PredInfo = (*BBMap)[PredInfo->DefBB]; - NewPHI->addIncoming(PredInfo->AvailableVal, Pred); - } else if (!Info->AvailableVal) - Info->AvailableVal = PredInfo->AvailableVal; - } + // Otherwise, we do need a PHI: insert one now if we don't already have one. + if (InsertedVal == 0) + InsertedVal = PHINode::Create(PrototypeValue->getType(), + PrototypeValue->getName(), &BB->front()); - if (NewPHI) { - DEBUG(dbgs() << " Inserted PHI: " << *NewPHI << "\n"); + PHINode *InsertedPHI = cast<PHINode>(InsertedVal); + InsertedPHI->reserveOperandSpace(IncomingPredInfo.size()-FirstPredInfoEntry); - // If the client wants to know about all new instructions, tell it. - if (InsertedPHIs) InsertedPHIs->push_back(NewPHI); - } -} + // Fill in all the predecessors of the PHI. + for (IncomingPredInfoTy::iterator I = + IncomingPredInfo.begin()+FirstPredInfoEntry, + E = IncomingPredInfo.end(); I != E; ++I) + InsertedPHI->addIncoming(I->second, I->first); -/// FindExistingPHI - Look through the PHI nodes in a block to see if any of -/// them match what is needed. -void SSAUpdater::FindExistingPHI(BasicBlock *BB) { - PHINode *SomePHI; - for (BasicBlock::iterator It = BB->begin(); - (SomePHI = dyn_cast<PHINode>(It)); ++It) { - if (CheckIfPHIMatches(SomePHI)) { - RecordMatchingPHI(SomePHI); - break; - } - ClearPHITags(SomePHI); - } -} + // Drop the entries we added in IncomingPredInfo to restore the stack. + IncomingPredInfo.erase(IncomingPredInfo.begin()+FirstPredInfoEntry, + IncomingPredInfo.end()); -/// CheckIfPHIMatches - Check if a PHI node matches the placement and values -/// in the BBMap. -bool SSAUpdater::CheckIfPHIMatches(PHINode *PHI) { - BBMapTy *BBMap = getBBMap(BM); - SmallVector<PHINode*, 20> WorkList; - WorkList.push_back(PHI); - - // Mark that the block containing this PHI has been visited. - (*BBMap)[PHI->getParent()]->PHITag = PHI; - - while (!WorkList.empty()) { - PHI = WorkList.pop_back_val(); - - // Iterate through the PHI's incoming values. - for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { - Value *IncomingVal = PHI->getIncomingValue(i); - BasicBlock *Pred = PHI->getIncomingBlock(i); - BBInfo *PredInfo = (*BBMap)[Pred]; - // Skip to the nearest preceding definition. - if (PredInfo->DefBB != Pred) { - Pred = PredInfo->DefBB; - PredInfo = (*BBMap)[Pred]; - } - - // Check if it matches the expected value. - if (PredInfo->AvailableVal) { - if (IncomingVal == PredInfo->AvailableVal) - continue; - return false; - } - - // Check if the value is a PHI in the correct block. - PHINode *IncomingPHIVal = dyn_cast<PHINode>(IncomingVal); - if (!IncomingPHIVal || IncomingPHIVal->getParent() != Pred) - return false; - - // If this block has already been visited, check if this PHI matches. - if (PredInfo->PHITag) { - if (IncomingPHIVal == PredInfo->PHITag) - continue; - return false; - } - PredInfo->PHITag = IncomingPHIVal; - - WorkList.push_back(IncomingPHIVal); - } - } - return true; -} + // See if the PHI node can be merged to a single value. This can happen in + // loop cases when we get a PHI of itself and one other value. + if (Value *ConstVal = InsertedPHI->hasConstantValue()) { + InsertedPHI->replaceAllUsesWith(ConstVal); + InsertedPHI->eraseFromParent(); + InsertedVal = ConstVal; + } else { + DEBUG(dbgs() << " Inserted PHI: " << *InsertedPHI << "\n"); -/// RecordMatchingPHI - For a PHI node that matches, record it and its input -/// PHIs in both the BBMap and the AvailableVals mapping. -void SSAUpdater::RecordMatchingPHI(PHINode *PHI) { - BBMapTy *BBMap = getBBMap(BM); - AvailableValsTy &AvailableVals = getAvailableVals(AV); - SmallVector<PHINode*, 20> WorkList; - WorkList.push_back(PHI); - - // Record this PHI. - BasicBlock *BB = PHI->getParent(); - AvailableVals[BB] = PHI; - (*BBMap)[BB]->AvailableVal = PHI; - - while (!WorkList.empty()) { - PHI = WorkList.pop_back_val(); - - // Iterate through the PHI's incoming values. - for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { - PHINode *IncomingPHIVal = dyn_cast<PHINode>(PHI->getIncomingValue(i)); - if (!IncomingPHIVal) continue; - BB = IncomingPHIVal->getParent(); - BBInfo *Info = (*BBMap)[BB]; - if (!Info || Info->AvailableVal) - continue; - - // Record the PHI and add it to the worklist. - AvailableVals[BB] = IncomingPHIVal; - Info->AvailableVal = IncomingPHIVal; - WorkList.push_back(IncomingPHIVal); - } + // If the client wants to know about all new instructions, tell it. + if (InsertedPHIs) InsertedPHIs->push_back(InsertedPHI); } -} -/// ClearPHITags - When one of the existing PHI nodes fails to match, clear -/// the PHITag values that were stored in the BBMap when checking to see if -/// it matched. -void SSAUpdater::ClearPHITags(PHINode *PHI) { - BBMapTy *BBMap = getBBMap(BM); - SmallVector<PHINode*, 20> WorkList; - WorkList.push_back(PHI); - - // Clear the tag for this PHI. - (*BBMap)[PHI->getParent()]->PHITag = 0; - - while (!WorkList.empty()) { - PHI = WorkList.pop_back_val(); - - // Iterate through the PHI's incoming values. - for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { - PHINode *IncomingPHIVal = dyn_cast<PHINode>(PHI->getIncomingValue(i)); - if (!IncomingPHIVal) continue; - BasicBlock *BB = IncomingPHIVal->getParent(); - BBInfo *Info = (*BBMap)[BB]; - if (!Info || Info->AvailableVal || !Info->PHITag) - continue; - - // Clear the tag and add the PHI to the worklist. - Info->PHITag = 0; - WorkList.push_back(IncomingPHIVal); - } - } + return InsertedVal; } |