diff options
Diffstat (limited to 'lib/Transforms/Utils')
-rw-r--r-- | lib/Transforms/Utils/AddrModeMatcher.cpp | 2 | ||||
-rw-r--r-- | lib/Transforms/Utils/BypassSlowDivision.cpp | 2 | ||||
-rw-r--r-- | lib/Transforms/Utils/CMakeLists.txt | 1 | ||||
-rw-r--r-- | lib/Transforms/Utils/IntegerDivision.cpp | 312 | ||||
-rw-r--r-- | lib/Transforms/Utils/Local.cpp | 38 | ||||
-rw-r--r-- | lib/Transforms/Utils/MetaRenamer.cpp | 6 | ||||
-rw-r--r-- | lib/Transforms/Utils/SimplifyCFG.cpp | 509 |
7 files changed, 727 insertions, 143 deletions
diff --git a/lib/Transforms/Utils/AddrModeMatcher.cpp b/lib/Transforms/Utils/AddrModeMatcher.cpp index 1e6586bf0d..0f9dfddf57 100644 --- a/lib/Transforms/Utils/AddrModeMatcher.cpp +++ b/lib/Transforms/Utils/AddrModeMatcher.cpp @@ -55,7 +55,7 @@ void ExtAddrMode::print(raw_ostream &OS) const { OS << ']'; } -#ifndef NDEBUG +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) void ExtAddrMode::dump() const { print(dbgs()); dbgs() << '\n'; diff --git a/lib/Transforms/Utils/BypassSlowDivision.cpp b/lib/Transforms/Utils/BypassSlowDivision.cpp index 30d60be277..ac18b7d5a0 100644 --- a/lib/Transforms/Utils/BypassSlowDivision.cpp +++ b/lib/Transforms/Utils/BypassSlowDivision.cpp @@ -221,7 +221,7 @@ static bool reuseOrInsertFastDiv(Function &F, // be profitably bypassed and carried out with a shorter, faster divide. bool llvm::bypassSlowDivision(Function &F, Function::iterator &I, - const DenseMap<Type *, Type *> &BypassTypeMap) { + const DenseMap<Type*, Type*> &BypassTypeMap) { DivCacheTy DivCache; bool MadeChange = false; diff --git a/lib/Transforms/Utils/CMakeLists.txt b/lib/Transforms/Utils/CMakeLists.txt index aace75b745..c3f72b13af 100644 --- a/lib/Transforms/Utils/CMakeLists.txt +++ b/lib/Transforms/Utils/CMakeLists.txt @@ -11,6 +11,7 @@ add_llvm_library(LLVMTransformUtils DemoteRegToStack.cpp InlineFunction.cpp InstructionNamer.cpp + IntegerDivision.cpp LCSSA.cpp Local.cpp LoopSimplify.cpp diff --git a/lib/Transforms/Utils/IntegerDivision.cpp b/lib/Transforms/Utils/IntegerDivision.cpp new file mode 100644 index 0000000000..9d630349ab --- /dev/null +++ b/lib/Transforms/Utils/IntegerDivision.cpp @@ -0,0 +1,312 @@ +//===-- IntegerDivision.cpp - Expand integer division ---------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file contains an implementation of 32bit scalar integer division for +// targets that don't have native support. It's largely derived from +// compiler-rt's implementation of __udivsi3, but hand-tuned to reduce the +// amount of control flow +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "integer-division" +#include "llvm/Function.h" +#include "llvm/Instructions.h" +#include "llvm/Intrinsics.h" +#include "llvm/IRBuilder.h" +#include "llvm/Transforms/Utils/IntegerDivision.h" + +using namespace llvm; + +/// Generate code to divide two signed integers. Returns the quotient, rounded +/// towards 0. Builder's insert point should be pointing at the sdiv +/// instruction. This will generate a udiv in the process, and Builder's insert +/// point will be pointing at the udiv (if present, i.e. not folded), ready to +/// be expanded if the user wishes. +static Value *generateSignedDivisionCode(Value *Dividend, Value *Divisor, + IRBuilder<> &Builder) { + // Implementation taken from compiler-rt's __divsi3 + + ConstantInt *ThirtyOne = Builder.getInt32(31); + + // ; %tmp = ashr i32 %dividend, 31 + // ; %tmp1 = ashr i32 %divisor, 31 + // ; %tmp2 = xor i32 %tmp, %dividend + // ; %u_dvnd = sub nsw i32 %tmp2, %tmp + // ; %tmp3 = xor i32 %tmp1, %divisor + // ; %u_dvsr = sub nsw i32 %tmp3, %tmp1 + // ; %q_sgn = xor i32 %tmp1, %tmp + // ; %q_mag = udiv i32 %u_dvnd, %u_dvsr + // ; %tmp4 = xor i32 %q_mag, %q_sgn + // ; %q = sub i32 %tmp4, %q_sgn + Value *Tmp = Builder.CreateAShr(Dividend, ThirtyOne); + Value *Tmp1 = Builder.CreateAShr(Divisor, ThirtyOne); + Value *Tmp2 = Builder.CreateXor(Tmp, Dividend); + Value *U_Dvnd = Builder.CreateSub(Tmp2, Tmp); + Value *Tmp3 = Builder.CreateXor(Tmp1, Divisor); + Value *U_Dvsr = Builder.CreateSub(Tmp3, Tmp1); + Value *Q_Sgn = Builder.CreateXor(Tmp1, Tmp); + Value *Q_Mag = Builder.CreateUDiv(U_Dvnd, U_Dvsr); + Value *Tmp4 = Builder.CreateXor(Q_Mag, Q_Sgn); + Value *Q = Builder.CreateSub(Tmp4, Q_Sgn); + + if (Instruction *UDiv = dyn_cast<Instruction>(Q_Mag)) + Builder.SetInsertPoint(UDiv); + + return Q; +} + +/// Generates code to divide two unsigned scalar 32-bit integers. Returns the +/// quotient, rounded towards 0. Builder's insert point should be pointing at +/// the udiv instruction. +static Value *generateUnsignedDivisionCode(Value *Dividend, Value *Divisor, + IRBuilder<> &Builder) { + // The basic algorithm can be found in the compiler-rt project's + // implementation of __udivsi3.c. Here, we do a lower-level IR based approach + // that's been hand-tuned to lessen the amount of control flow involved. + + // Some helper values + IntegerType *I32Ty = Builder.getInt32Ty(); + + ConstantInt *Zero = Builder.getInt32(0); + ConstantInt *One = Builder.getInt32(1); + ConstantInt *ThirtyOne = Builder.getInt32(31); + ConstantInt *NegOne = ConstantInt::getSigned(I32Ty, -1); + ConstantInt *True = Builder.getTrue(); + + BasicBlock *IBB = Builder.GetInsertBlock(); + Function *F = IBB->getParent(); + Function *CTLZi32 = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz, + I32Ty); + + // Our CFG is going to look like: + // +---------------------+ + // | special-cases | + // | ... | + // +---------------------+ + // | | + // | +----------+ + // | | bb1 | + // | | ... | + // | +----------+ + // | | | + // | | +------------+ + // | | | preheader | + // | | | ... | + // | | +------------+ + // | | | + // | | | +---+ + // | | | | | + // | | +------------+ | + // | | | do-while | | + // | | | ... | | + // | | +------------+ | + // | | | | | + // | +-----------+ +---+ + // | | loop-exit | + // | | ... | + // | +-----------+ + // | | + // +-------+ + // | ... | + // | end | + // +-------+ + BasicBlock *SpecialCases = Builder.GetInsertBlock(); + SpecialCases->setName(Twine(SpecialCases->getName(), "_udiv-special-cases")); + BasicBlock *End = SpecialCases->splitBasicBlock(Builder.GetInsertPoint(), + "udiv-end"); + BasicBlock *LoopExit = BasicBlock::Create(Builder.getContext(), + "udiv-loop-exit", F, End); + BasicBlock *DoWhile = BasicBlock::Create(Builder.getContext(), + "udiv-do-while", F, End); + BasicBlock *Preheader = BasicBlock::Create(Builder.getContext(), + "udiv-preheader", F, End); + BasicBlock *BB1 = BasicBlock::Create(Builder.getContext(), + "udiv-bb1", F, End); + + // We'll be overwriting the terminator to insert our extra blocks + SpecialCases->getTerminator()->eraseFromParent(); + + // First off, check for special cases: dividend or divisor is zero, divisor + // is greater than dividend, and divisor is 1. + // ; special-cases: + // ; %ret0_1 = icmp eq i32 %divisor, 0 + // ; %ret0_2 = icmp eq i32 %dividend, 0 + // ; %ret0_3 = or i1 %ret0_1, %ret0_2 + // ; %tmp0 = tail call i32 @llvm.ctlz.i32(i32 %divisor, i1 true) + // ; %tmp1 = tail call i32 @llvm.ctlz.i32(i32 %dividend, i1 true) + // ; %sr = sub nsw i32 %tmp0, %tmp1 + // ; %ret0_4 = icmp ugt i32 %sr, 31 + // ; %ret0 = or i1 %ret0_3, %ret0_4 + // ; %retDividend = icmp eq i32 %sr, 31 + // ; %retVal = select i1 %ret0, i32 0, i32 %dividend + // ; %earlyRet = or i1 %ret0, %retDividend + // ; br i1 %earlyRet, label %end, label %bb1 + Builder.SetInsertPoint(SpecialCases); + Value *Ret0_1 = Builder.CreateICmpEQ(Divisor, Zero); + Value *Ret0_2 = Builder.CreateICmpEQ(Dividend, Zero); + Value *Ret0_3 = Builder.CreateOr(Ret0_1, Ret0_2); + Value *Tmp0 = Builder.CreateCall2(CTLZi32, Divisor, True); + Value *Tmp1 = Builder.CreateCall2(CTLZi32, Dividend, True); + Value *SR = Builder.CreateSub(Tmp0, Tmp1); + Value *Ret0_4 = Builder.CreateICmpUGT(SR, ThirtyOne); + Value *Ret0 = Builder.CreateOr(Ret0_3, Ret0_4); + Value *RetDividend = Builder.CreateICmpEQ(SR, ThirtyOne); + Value *RetVal = Builder.CreateSelect(Ret0, Zero, Dividend); + Value *EarlyRet = Builder.CreateOr(Ret0, RetDividend); + Builder.CreateCondBr(EarlyRet, End, BB1); + + // ; bb1: ; preds = %special-cases + // ; %sr_1 = add i32 %sr, 1 + // ; %tmp2 = sub i32 31, %sr + // ; %q = shl i32 %dividend, %tmp2 + // ; %skipLoop = icmp eq i32 %sr_1, 0 + // ; br i1 %skipLoop, label %loop-exit, label %preheader + Builder.SetInsertPoint(BB1); + Value *SR_1 = Builder.CreateAdd(SR, One); + Value *Tmp2 = Builder.CreateSub(ThirtyOne, SR); + Value *Q = Builder.CreateShl(Dividend, Tmp2); + Value *SkipLoop = Builder.CreateICmpEQ(SR_1, Zero); + Builder.CreateCondBr(SkipLoop, LoopExit, Preheader); + + // ; preheader: ; preds = %bb1 + // ; %tmp3 = lshr i32 %dividend, %sr_1 + // ; %tmp4 = add i32 %divisor, -1 + // ; br label %do-while + Builder.SetInsertPoint(Preheader); + Value *Tmp3 = Builder.CreateLShr(Dividend, SR_1); + Value *Tmp4 = Builder.CreateAdd(Divisor, NegOne); + Builder.CreateBr(DoWhile); + + // ; do-while: ; preds = %do-while, %preheader + // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ] + // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ] + // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ] + // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ] + // ; %tmp5 = shl i32 %r_1, 1 + // ; %tmp6 = lshr i32 %q_2, 31 + // ; %tmp7 = or i32 %tmp5, %tmp6 + // ; %tmp8 = shl i32 %q_2, 1 + // ; %q_1 = or i32 %carry_1, %tmp8 + // ; %tmp9 = sub i32 %tmp4, %tmp7 + // ; %tmp10 = ashr i32 %tmp9, 31 + // ; %carry = and i32 %tmp10, 1 + // ; %tmp11 = and i32 %tmp10, %divisor + // ; %r = sub i32 %tmp7, %tmp11 + // ; %sr_2 = add i32 %sr_3, -1 + // ; %tmp12 = icmp eq i32 %sr_2, 0 + // ; br i1 %tmp12, label %loop-exit, label %do-while + Builder.SetInsertPoint(DoWhile); + PHINode *Carry_1 = Builder.CreatePHI(I32Ty, 2); + PHINode *SR_3 = Builder.CreatePHI(I32Ty, 2); + PHINode *R_1 = Builder.CreatePHI(I32Ty, 2); + PHINode *Q_2 = Builder.CreatePHI(I32Ty, 2); + Value *Tmp5 = Builder.CreateShl(R_1, One); + Value *Tmp6 = Builder.CreateLShr(Q_2, ThirtyOne); + Value *Tmp7 = Builder.CreateOr(Tmp5, Tmp6); + Value *Tmp8 = Builder.CreateShl(Q_2, One); + Value *Q_1 = Builder.CreateOr(Carry_1, Tmp8); + Value *Tmp9 = Builder.CreateSub(Tmp4, Tmp7); + Value *Tmp10 = Builder.CreateAShr(Tmp9, 31); + Value *Carry = Builder.CreateAnd(Tmp10, One); + Value *Tmp11 = Builder.CreateAnd(Tmp10, Divisor); + Value *R = Builder.CreateSub(Tmp7, Tmp11); + Value *SR_2 = Builder.CreateAdd(SR_3, NegOne); + Value *Tmp12 = Builder.CreateICmpEQ(SR_2, Zero); + Builder.CreateCondBr(Tmp12, LoopExit, DoWhile); + + // ; loop-exit: ; preds = %do-while, %bb1 + // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ] + // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ] + // ; %tmp13 = shl i32 %q_3, 1 + // ; %q_4 = or i32 %carry_2, %tmp13 + // ; br label %end + Builder.SetInsertPoint(LoopExit); + PHINode *Carry_2 = Builder.CreatePHI(I32Ty, 2); + PHINode *Q_3 = Builder.CreatePHI(I32Ty, 2); + Value *Tmp13 = Builder.CreateShl(Q_3, One); + Value *Q_4 = Builder.CreateOr(Carry_2, Tmp13); + Builder.CreateBr(End); + + // ; end: ; preds = %loop-exit, %special-cases + // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ] + // ; ret i32 %q_5 + Builder.SetInsertPoint(End, End->begin()); + PHINode *Q_5 = Builder.CreatePHI(I32Ty, 2); + + // Populate the Phis, since all values have now been created. Our Phis were: + // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ] + Carry_1->addIncoming(Zero, Preheader); + Carry_1->addIncoming(Carry, DoWhile); + // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ] + SR_3->addIncoming(SR_1, Preheader); + SR_3->addIncoming(SR_2, DoWhile); + // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ] + R_1->addIncoming(Tmp3, Preheader); + R_1->addIncoming(R, DoWhile); + // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ] + Q_2->addIncoming(Q, Preheader); + Q_2->addIncoming(Q_1, DoWhile); + // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ] + Carry_2->addIncoming(Zero, BB1); + Carry_2->addIncoming(Carry, DoWhile); + // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ] + Q_3->addIncoming(Q, BB1); + Q_3->addIncoming(Q_1, DoWhile); + // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ] + Q_5->addIncoming(Q_4, LoopExit); + Q_5->addIncoming(RetVal, SpecialCases); + + return Q_5; +} + +/// Generate code to divide two integers, replacing Div with the generated +/// code. This currently generates code similarly to compiler-rt's +/// implementations, but future work includes generating more specialized code +/// when more information about the operands are known. Currently only +/// implements 32bit scalar division, but future work is removing this +/// limitation. +/// +/// @brief Replace Div with generated code. +bool llvm::expandDivision(BinaryOperator *Div) { + assert((Div->getOpcode() == Instruction::SDiv || + Div->getOpcode() == Instruction::UDiv) && + "Trying to expand division from a non-division function"); + + IRBuilder<> Builder(Div); + + if (Div->getType()->isVectorTy()) + llvm_unreachable("Div over vectors not supported"); + + // First prepare the sign if it's a signed division + if (Div->getOpcode() == Instruction::SDiv) { + // Lower the code to unsigned division, and reset Div to point to the udiv. + Value *Quotient = generateSignedDivisionCode(Div->getOperand(0), + Div->getOperand(1), Builder); + Div->replaceAllUsesWith(Quotient); + Div->dropAllReferences(); + Div->eraseFromParent(); + + // If we didn't actually generate a udiv instruction, we're done + BinaryOperator *BO = dyn_cast<BinaryOperator>(Builder.GetInsertPoint()); + if (!BO || BO->getOpcode() != Instruction::UDiv) + return true; + + Div = BO; + } + + // Insert the unsigned division code + Value *Quotient = generateUnsignedDivisionCode(Div->getOperand(0), + Div->getOperand(1), + Builder); + Div->replaceAllUsesWith(Quotient); + Div->dropAllReferences(); + Div->eraseFromParent(); + + return true; +} diff --git a/lib/Transforms/Utils/Local.cpp b/lib/Transforms/Utils/Local.cpp index 0601433565..2d89516c43 100644 --- a/lib/Transforms/Utils/Local.cpp +++ b/lib/Transforms/Utils/Local.cpp @@ -23,6 +23,7 @@ #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" #include "llvm/Intrinsics.h" +#include "llvm/MDBuilder.h" #include "llvm/Metadata.h" #include "llvm/Operator.h" #include "llvm/ADT/DenseMap.h" @@ -122,6 +123,27 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, // Check to see if this branch is going to the same place as the default // dest. If so, eliminate it as an explicit compare. if (i.getCaseSuccessor() == DefaultDest) { + MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); + // MD should have 2 + NumCases operands. + if (MD && MD->getNumOperands() == 2 + SI->getNumCases()) { + // Collect branch weights into a vector. + SmallVector<uint32_t, 8> Weights; + for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; + ++MD_i) { + ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); + assert(CI); + Weights.push_back(CI->getValue().getZExtValue()); + } + // Merge weight of this case to the default weight. + unsigned idx = i.getCaseIndex(); + Weights[0] += Weights[idx+1]; + // Remove weight for this case. + std::swap(Weights[idx+1], Weights.back()); + Weights.pop_back(); + SI->setMetadata(LLVMContext::MD_prof, + MDBuilder(BB->getContext()). + createBranchWeights(Weights)); + } // Remove this entry. DefaultDest->removePredecessor(SI->getParent()); SI->removeCase(i); @@ -178,8 +200,20 @@ bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, "cond"); // Insert the new branch. - Builder.CreateCondBr(Cond, FirstCase.getCaseSuccessor(), - SI->getDefaultDest()); + BranchInst *NewBr = Builder.CreateCondBr(Cond, + FirstCase.getCaseSuccessor(), + SI->getDefaultDest()); + MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); + if (MD && MD->getNumOperands() == 3) { + ConstantInt *SICase = dyn_cast<ConstantInt>(MD->getOperand(2)); + ConstantInt *SIDef = dyn_cast<ConstantInt>(MD->getOperand(1)); + assert(SICase && SIDef); + // The TrueWeight should be the weight for the single case of SI. + NewBr->setMetadata(LLVMContext::MD_prof, + MDBuilder(BB->getContext()). + createBranchWeights(SICase->getValue().getZExtValue(), + SIDef->getValue().getZExtValue())); + } // Delete the old switch. SI->eraseFromParent(); diff --git a/lib/Transforms/Utils/MetaRenamer.cpp b/lib/Transforms/Utils/MetaRenamer.cpp index 60f031e16f..233bc12d3c 100644 --- a/lib/Transforms/Utils/MetaRenamer.cpp +++ b/lib/Transforms/Utils/MetaRenamer.cpp @@ -13,15 +13,15 @@ // //===----------------------------------------------------------------------===// +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallString.h" #include "llvm/Transforms/IPO.h" +#include "llvm/DerivedTypes.h" #include "llvm/Function.h" #include "llvm/Module.h" #include "llvm/Pass.h" #include "llvm/Type.h" #include "llvm/TypeFinder.h" -#include "llvm/DerivedTypes.h" -#include "llvm/ADT/STLExtras.h" -#include "llvm/ADT/SmallString.h" using namespace llvm; diff --git a/lib/Transforms/Utils/SimplifyCFG.cpp b/lib/Transforms/Utils/SimplifyCFG.cpp index 142f157c3c..876ff2c337 100644 --- a/lib/Transforms/Utils/SimplifyCFG.cpp +++ b/lib/Transforms/Utils/SimplifyCFG.cpp @@ -54,8 +54,13 @@ static cl::opt<bool> DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false), cl::desc("Duplicate return instructions into unconditional branches")); +static cl::opt<bool> +SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), + cl::desc("Sink common instructions down to the end block")); + STATISTIC(NumSpeculations, "Number of speculative executed instructions"); STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); +STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); namespace { /// ValueEqualityComparisonCase - Represents a case of a switch. @@ -667,13 +672,32 @@ SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator() << "Through successor TI: " << *TI); + // Collect branch weights into a vector. + SmallVector<uint32_t, 8> Weights; + MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); + bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases()); + if (HasWeight) + for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; + ++MD_i) { + ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); + assert(CI); + Weights.push_back(CI->getValue().getZExtValue()); + } for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) { --i; if (DeadCases.count(i.getCaseValue())) { + if (HasWeight) { + std::swap(Weights[i.getCaseIndex()+1], Weights.back()); + Weights.pop_back(); + } i.getCaseSuccessor()->removePredecessor(TI->getParent()); SI->removeCase(i); } } + if (HasWeight) + SI->setMetadata(LLVMContext::MD_prof, + MDBuilder(SI->getParent()->getContext()). + createBranchWeights(Weights)); DEBUG(dbgs() << "Leaving: " << *TI << "\n"); return true; @@ -830,18 +854,24 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, bool PredHasWeights = HasBranchWeights(PTI); bool SuccHasWeights = HasBranchWeights(TI); - if (PredHasWeights) + if (PredHasWeights) { GetBranchWeights(PTI, Weights); - else if (SuccHasWeights) + // branch-weight metadata is inconsistant here. + if (Weights.size() != 1 + PredCases.size()) + PredHasWeights = SuccHasWeights = false; + } else if (SuccHasWeights) // If there are no predecessor weights but there are successor weights, // populate Weights with 1, which will later be scaled to the sum of // successor's weights Weights.assign(1 + PredCases.size(), 1); SmallVector<uint64_t, 8> SuccWeights; - if (SuccHasWeights) + if (SuccHasWeights) { GetBranchWeights(TI, SuccWeights); - else if (PredHasWeights) + // branch-weight metadata is inconsistant here. + if (SuccWeights.size() != 1 + BBCases.size()) + PredHasWeights = SuccHasWeights = false; + } else if (PredHasWeights) SuccWeights.assign(1 + BBCases.size(), 1); if (PredDefault == BB) { @@ -898,18 +928,21 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, Weights[0] *= SuccWeights[0]; } } else { - // FIXME: preserve branch weight metadata, similarly to the 'then' - // above. For now, drop it. - PredHasWeights = false; - SuccHasWeights = false; - // If this is not the default destination from PSI, only the edges // in SI that occur in PSI with a destination of BB will be // activated. std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; + std::map<ConstantInt*, uint64_t> WeightsForHandled; for (unsigned i = 0, e = PredCases.size(); i != e; ++i) if (PredCases[i].Dest == BB) { PTIHandled.insert(PredCases[i].Value); + + if (PredHasWeights || SuccHasWeights) { + WeightsForHandled[PredCases[i].Value] = Weights[i+1]; + std::swap(Weights[i+1], Weights.back()); + Weights.pop_back(); + } + std::swap(PredCases[i], PredCases.back()); PredCases.pop_back(); --i; --e; @@ -920,6 +953,8 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, for (unsigned i = 0, e = BBCases.size(); i != e; ++i) if (PTIHandled.count(BBCases[i].Value)) { // If this is one we are capable of getting... + if (PredHasWeights || SuccHasWeights) + Weights.push_back(WeightsForHandled[BBCases[i].Value]); PredCases.push_back(BBCases[i]); NewSuccessors.push_back(BBCases[i].Dest); PTIHandled.erase(BBCases[i].Value);// This constant is taken care of @@ -930,6 +965,8 @@ bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI, for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = PTIHandled.begin(), E = PTIHandled.end(); I != E; ++I) { + if (PredHasWeights || SuccHasWeights) + Weights.push_back(WeightsForHandled[*I]); PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault)); NewSuccessors.push_back(BBDefault); } @@ -1124,6 +1161,171 @@ HoistTerminator: return true; } +/// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd, +/// check whether BBEnd has only two predecessors and the other predecessor +/// ends with an unconditional branch. If it is true, sink any common code +/// in the two predecessors to BBEnd. +static bool SinkThenElseCodeToEnd(BranchInst *BI1) { + assert(BI1->isUnconditional()); + BasicBlock *BB1 = BI1->getParent(); + BasicBlock *BBEnd = BI1->getSuccessor(0); + + // Check that BBEnd has two predecessors and the other predecessor ends with + // an unconditional branch. + SmallVector<BasicBlock*, 16> Preds(pred_begin(BBEnd), pred_end(BBEnd)); + if (Preds.size() != 2) + return false; + BasicBlock *BB2 = (Preds[0] == BB1) ? Preds[1] : Preds[0]; + BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator()); + if (!BI2 || !BI2->isUnconditional()) + return false; + + // Gather the PHI nodes in BBEnd. + std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2; + Instruction *FirstNonPhiInBBEnd = 0; + for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); + I != E; ++I) { + if (PHINode *PN = dyn_cast<PHINode>(I)) { + Value *BB1V = PN->getIncomingValueForBlock(BB1); + Value *BB2V = PN->getIncomingValueForBlock(BB2); + MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN); + } else { + FirstNonPhiInBBEnd = &*I; + break; + } + } + if (!FirstNonPhiInBBEnd) + return false; + + + // This does very trivial matching, with limited scanning, to find identical + // instructions in the two blocks. We scan backward for obviously identical + // instructions in an identical order. + BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(), + RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(), + RE2 = BB2->getInstList().rend(); + // Skip debug info. + while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; + if (RI1 == RE1) + return false; + while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; + if (RI2 == RE2) + return false; + // Skip the unconditional branches. + ++RI1; + ++RI2; + + bool Changed = false; + while (RI1 != RE1 && RI2 != RE2) { + // Skip debug info. + while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1; + if (RI1 == RE1) + return Changed; + while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2; + if (RI2 == RE2) + return Changed; + + Instruction *I1 = &*RI1, *I2 = &*RI2; + // I1 and I2 should have a single use in the same PHI node, and they + // perform the same operation. + // Cannot move control-flow-involving, volatile loads, vaarg, etc. + if (isa<PHINode>(I1) || isa<PHINode>(I2) || + isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) || + isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) || + isa<AllocaInst>(I1) || isa<AllocaInst>(I2) || + I1->mayHaveSideEffects() || I2->mayHaveSideEffects() || + I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() || + !I1->hasOneUse() || !I2->hasOneUse() || + MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() || + MapValueFromBB1ToBB2[I1].first != I2) + return Changed; + + // Check whether we should swap the operands of ICmpInst. + ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2); + bool SwapOpnds = false; + if (ICmp1 && ICmp2 && + ICmp1->getOperand(0) != ICmp2->getOperand(0) && + ICmp1->getOperand(1) != ICmp2->getOperand(1) && + (ICmp1->getOperand(0) == ICmp2->getOperand(1) || + ICmp1->getOperand(1) == ICmp2->getOperand(0))) { + ICmp2->swapOperands(); + SwapOpnds = true; + } + if (!I1->isSameOperationAs(I2)) { + if (SwapOpnds) + ICmp2->swapOperands(); + return Changed; + } + + // The operands should be either the same or they need to be generated + // with a PHI node after sinking. We only handle the case where there is + // a single pair of different operands. + Value *DifferentOp1 = 0, *DifferentOp2 = 0; + unsigned Op1Idx = 0; + for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) { + if (I1->getOperand(I) == I2->getOperand(I)) + continue; + // Early exit if we have more-than one pair of different operands or + // the different operand is already in MapValueFromBB1ToBB2. + // Early exit if we need a PHI node to replace a constant. + if (DifferentOp1 || + MapValueFromBB1ToBB2.find(I1->getOperand(I)) != + MapValueFromBB1ToBB2.end() || + isa<Constant>(I1->getOperand(I)) || + isa<Constant>(I2->getOperand(I))) { + // If we can't sink the instructions, undo the swapping. + if (SwapOpnds) + ICmp2->swapOperands(); + return Changed; + } + DifferentOp1 = I1->getOperand(I); + Op1Idx = I; + DifferentOp2 = I2->getOperand(I); + } + + // We insert the pair of different operands to MapValueFromBB1ToBB2 and + // remove (I1, I2) from MapValueFromBB1ToBB2. + if (DifferentOp1) { + PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2, + DifferentOp1->getName() + ".sink", + BBEnd->begin()); + MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN); + // I1 should use NewPN instead of DifferentOp1. + I1->setOperand(Op1Idx, NewPN); + NewPN->addIncoming(DifferentOp1, BB1); + NewPN->addIncoming(DifferentOp2, BB2); + DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";); + } + PHINode *OldPN = MapValueFromBB1ToBB2[I1].second; + MapValueFromBB1ToBB2.erase(I1); + + DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";); + DEBUG(dbgs() << " " << *I2 << "\n";); + // We need to update RE1 and RE2 if we are going to sink the first + // instruction in the basic block down. + bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin()); + // Sink the instruction. + BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1); + if (!OldPN->use_empty()) + OldPN->replaceAllUsesWith(I1); + OldPN->eraseFromParent(); + + if (!I2->use_empty()) + I2->replaceAllUsesWith(I1); + I1->intersectOptionalDataWith(I2); + I2->eraseFromParent(); + + if (UpdateRE1) + RE1 = BB1->getInstList().rend(); + if (UpdateRE2) + RE2 = BB2->getInstList().rend(); + FirstNonPhiInBBEnd = I1; + NumSinkCommons++; + Changed = true; + } + return Changed; +} + /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code /// (for now, restricted to a single instruction that's side effect free) from @@ -1626,7 +1828,7 @@ static bool SimplifyCondBranchToTwoReturns(BranchInst *BI, /// parameters and return true, or returns false if no or invalid metadata was /// found. static bool ExtractBranchMetadata(BranchInst *BI, - APInt &ProbTrue, APInt &ProbFalse) { + uint64_t &ProbTrue, uint64_t &ProbFalse) { assert(BI->isConditional() && "Looking for probabilities on unconditional branch?"); MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof); @@ -1634,35 +1836,11 @@ static bool ExtractBranchMetadata(BranchInst *BI, ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1)); ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2)); if (!CITrue || !CIFalse) return false; - ProbTrue = CITrue->getValue(); - ProbFalse = CIFalse->getValue(); - assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 && - "Branch probability metadata must be 32-bit integers"); + ProbTrue = CITrue->getValue().getZExtValue(); + ProbFalse = CIFalse->getValue().getZExtValue(); return true; } -/// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In -/// the event of overflow, logically-shifts all four inputs right until the -/// multiply fits. -static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D, - unsigned &BitsLost) { - BitsLost = 0; - bool Overflow = false; - APInt Result = A.umul_ov(B, Overflow); - if (Overflow) { - APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A); - do { - B = B.lshr(1); - ++BitsLost; - } while (B.ugt(MaxB)); - A = A.lshr(BitsLost); - C = C.lshr(BitsLost); - D = D.lshr(BitsLost); - Result = A * B; - } - return Result; -} - /// checkCSEInPredecessor - Return true if the given instruction is available /// in its predecessor block. If yes, the instruction will be removed. /// @@ -1788,7 +1966,7 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { continue; // Determine if the two branches share a common destination. - Instruction::BinaryOps Opc; + Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd; bool InvertPredCond = false; if (BI->isConditional()) { @@ -1887,14 +2065,53 @@ bool llvm::FoldBranchToCommonDest(BranchInst *BI) { New, "or.cond")); PBI->setCondition(NewCond); + uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight; + bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight, + PredFalseWeight); + bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight, + SuccFalseWeight); + SmallVector<uint64_t, 8> NewWeights; + if (PBI->getSuccessor(0) == BB) { + if (PredHasWeights && SuccHasWeights) { + // PBI: br i1 %x, BB, FalseDest + // BI: br i1 %y, TrueDest, FalseDest + //TrueWeight is TrueWeight for PBI * TrueWeight for BI. + NewWeights.push_back(PredTrueWeight * SuccTrueWeight); + //FalseWeight is FalseWeight for PBI * TotalWeight for BI + + // TrueWeight for PBI * FalseWeight for BI. + // We assume that total weights of a BranchInst can fit into 32 bits. + // Therefore, we will not have overflow using 64-bit arithmetic. + NewWeights.push_back(PredFalseWeight * (SuccFalseWeight + + SuccTrueWeight) + PredTrueWeight * SuccFalseWeight); + } AddPredecessorToBlock(TrueDest, PredBlock, BB); PBI->setSuccessor(0, TrueDest); } if (PBI->getSuccessor(1) == BB) { + if (PredHasWeights && SuccHasWeights) { + // PBI: br i1 %x, TrueDest, BB + // BI: br i1 %y, TrueDest, FalseDest + //TrueWeight is TrueWeight for PBI * TotalWeight for BI + + // FalseWeight for PBI * TrueWeight for BI. + NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + + SuccTrueWeight) + PredFalseWeight * SuccTrueWeight); + //FalseWeight is FalseWeight for PBI * FalseWeight for BI. + NewWeights.push_back(PredFalseWeight * SuccFalseWeight); + } AddPredecessorToBlock(FalseDest, PredBlock, BB); PBI->setSuccessor(1, FalseDest); } + if (NewWeights.size() == 2) { + // Halve the weights if any of them cannot fit in an uint32_t + FitWeights(NewWeights); + + SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end()); + PBI->setMetadata(LLVM |