//===- AggressiveInstCombine.cpp ------------------------------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This file implements the aggressive expression pattern combiner classes. // Currently, it handles expression patterns for: // * Truncate instruction // //===----------------------------------------------------------------------===// #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h" #include "AggressiveInstCombineInternal.h" #include "llvm-c/Initialization.h" #include "llvm-c/Transforms/AggressiveInstCombine.h" #include "llvm/ADT/Statistic.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/BasicAliasAnalysis.h" #include "llvm/Analysis/GlobalsModRef.h" #include "llvm/Analysis/TargetLibraryInfo.h" #include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/IR/Dominators.h" #include "llvm/IR/Function.h" #include "llvm/IR/IRBuilder.h" #include "llvm/IR/LegacyPassManager.h" #include "llvm/IR/PatternMatch.h" #include "llvm/InitializePasses.h" #include "llvm/Pass.h" #include "llvm/Transforms/Utils/BuildLibCalls.h" #include "llvm/Transforms/Utils/Local.h" using namespace llvm; using namespace PatternMatch; namespace llvm { class DataLayout; } #define DEBUG_TYPE "aggressive-instcombine" STATISTIC(NumAnyOrAllBitsSet, "Number of any/all-bits-set patterns folded"); STATISTIC(NumGuardedRotates, "Number of guarded rotates transformed into funnel shifts"); STATISTIC(NumGuardedFunnelShifts, "Number of guarded funnel shifts transformed into funnel shifts"); STATISTIC(NumPopCountRecognized, "Number of popcount idioms recognized"); namespace { /// Contains expression pattern combiner logic. /// This class provides both the logic to combine expression patterns and /// combine them. It differs from InstCombiner class in that each pattern /// combiner runs only once as opposed to InstCombine's multi-iteration, /// which allows pattern combiner to have higher complexity than the O(1) /// required by the instruction combiner. class AggressiveInstCombinerLegacyPass : public FunctionPass { public: static char ID; // Pass identification, replacement for typeid AggressiveInstCombinerLegacyPass() : FunctionPass(ID) { initializeAggressiveInstCombinerLegacyPassPass( *PassRegistry::getPassRegistry()); } void getAnalysisUsage(AnalysisUsage &AU) const override; /// Run all expression pattern optimizations on the given /p F function. /// /// \param F function to optimize. /// \returns true if the IR is changed. bool runOnFunction(Function &F) override; }; } // namespace /// Match a pattern for a bitwise funnel/rotate operation that partially guards /// against undefined behavior by branching around the funnel-shift/rotation /// when the shift amount is 0. static bool foldGuardedFunnelShift(Instruction &I, const DominatorTree &DT) { if (I.getOpcode() != Instruction::PHI || I.getNumOperands() != 2) return false; // As with the one-use checks below, this is not strictly necessary, but we // are being cautious to avoid potential perf regressions on targets that // do not actually have a funnel/rotate instruction (where the funnel shift // would be expanded back into math/shift/logic ops). if (!isPowerOf2_32(I.getType()->getScalarSizeInBits())) return false; // Match V to funnel shift left/right and capture the source operands and // shift amount. auto matchFunnelShift = [](Value *V, Value *&ShVal0, Value *&ShVal1, Value *&ShAmt) { Value *SubAmt; unsigned Width = V->getType()->getScalarSizeInBits(); // fshl(ShVal0, ShVal1, ShAmt) // == (ShVal0 << ShAmt) | (ShVal1 >> (Width -ShAmt)) if (match(V, m_OneUse(m_c_Or( m_Shl(m_Value(ShVal0), m_Value(ShAmt)), m_LShr(m_Value(ShVal1), m_Sub(m_SpecificInt(Width), m_Value(SubAmt))))))) { if (ShAmt == SubAmt) // TODO: Use m_Specific return Intrinsic::fshl; } // fshr(ShVal0, ShVal1, ShAmt) // == (ShVal0 >> ShAmt) | (ShVal1 << (Width - ShAmt)) if (match(V, m_OneUse(m_c_Or(m_Shl(m_Value(ShVal0), m_Sub(m_SpecificInt(Width), m_Value(SubAmt))), m_LShr(m_Value(ShVal1), m_Value(ShAmt)))))) { if (ShAmt == SubAmt) // TODO: Use m_Specific return Intrinsic::fshr; } return Intrinsic::not_intrinsic; }; // One phi operand must be a funnel/rotate operation, and the other phi // operand must be the source value of that funnel/rotate operation: // phi [ rotate(RotSrc, ShAmt), FunnelBB ], [ RotSrc, GuardBB ] // phi [ fshl(ShVal0, ShVal1, ShAmt), FunnelBB ], [ ShVal0, GuardBB ] // phi [ fshr(ShVal0, ShVal1, ShAmt), FunnelBB ], [ ShVal1, GuardBB ] PHINode &Phi = cast(I); unsigned FunnelOp = 0, GuardOp = 1; Value *P0 = Phi.getOperand(0), *P1 = Phi.getOperand(1); Value *ShVal0, *ShVal1, *ShAmt; Intrinsic::ID IID = matchFunnelShift(P0, ShVal0, ShVal1, ShAmt); if (IID == Intrinsic::not_intrinsic || (IID == Intrinsic::fshl && ShVal0 != P1) || (IID == Intrinsic::fshr && ShVal1 != P1)) { IID = matchFunnelShift(P1, ShVal0, ShVal1, ShAmt); if (IID == Intrinsic::not_intrinsic || (IID == Intrinsic::fshl && ShVal0 != P0) || (IID == Intrinsic::fshr && ShVal1 != P0)) return false; assert((IID == Intrinsic::fshl || IID == Intrinsic::fshr) && "Pattern must match funnel shift left or right"); std::swap(FunnelOp, GuardOp); } // The incoming block with our source operand must be the "guard" block. // That must contain a cmp+branch to avoid the funnel/rotate when the shift // amount is equal to 0. The other incoming block is the block with the // funnel/rotate. BasicBlock *GuardBB = Phi.getIncomingBlock(GuardOp); BasicBlock *FunnelBB = Phi.getIncomingBlock(FunnelOp); Instruction *TermI = GuardBB->getTerminator(); // Ensure that the shift values dominate each block. if (!DT.dominates(ShVal0, TermI) || !DT.dominates(ShVal1, TermI)) return false; ICmpInst::Predicate Pred; BasicBlock *PhiBB = Phi.getParent(); if (!match(TermI, m_Br(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()), m_SpecificBB(PhiBB), m_SpecificBB(FunnelBB)))) return false; if (Pred != CmpInst::ICMP_EQ) return false; IRBuilder<> Builder(PhiBB, PhiBB->getFirstInsertionPt()); if (ShVal0 == ShVal1) ++NumGuardedRotates; else ++NumGuardedFunnelShifts; // If this is not a rotate then the select was blocking poison from the // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. bool IsFshl = IID == Intrinsic::fshl; if (ShVal0 != ShVal1) { if (IsFshl && !llvm::isGuaranteedNotToBePoison(ShVal1)) ShVal1 = Builder.CreateFreeze(ShVal1); else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(ShVal0)) ShVal0 = Builder.CreateFreeze(ShVal0); } // We matched a variation of this IR pattern: // GuardBB: // %cmp = icmp eq i32 %ShAmt, 0 // br i1 %cmp, label %PhiBB, label %FunnelBB // FunnelBB: // %sub = sub i32 32, %ShAmt // %shr = lshr i32 %ShVal1, %sub // %shl = shl i32 %ShVal0, %ShAmt // %fsh = or i32 %shr, %shl // br label %PhiBB // PhiBB: // %cond = phi i32 [ %fsh, %FunnelBB ], [ %ShVal0, %GuardBB ] // --> // llvm.fshl.i32(i32 %ShVal0, i32 %ShVal1, i32 %ShAmt) Function *F = Intrinsic::getDeclaration(Phi.getModule(), IID, Phi.getType()); Phi.replaceAllUsesWith(Builder.CreateCall(F, {ShVal0, ShVal1, ShAmt})); return true; } /// This is used by foldAnyOrAllBitsSet() to capture a source value (Root) and /// the bit indexes (Mask) needed by a masked compare. If we're matching a chain /// of 'and' ops, then we also need to capture the fact that we saw an /// "and X, 1", so that's an extra return value for that case. struct MaskOps { Value *Root = nullptr; APInt Mask; bool MatchAndChain; bool FoundAnd1 = false; MaskOps(unsigned BitWidth, bool MatchAnds) : Mask(APInt::getZero(BitWidth)), MatchAndChain(MatchAnds) {} }; /// This is a recursive helper for foldAnyOrAllBitsSet() that walks through a /// chain of 'and' or 'or' instructions looking for shift ops of a common source /// value. Examples: /// or (or (or X, (X >> 3)), (X >> 5)), (X >> 8) /// returns { X, 0x129 } /// and (and (X >> 1), 1), (X >> 4) /// returns { X, 0x12 } static bool matchAndOrChain(Value *V, MaskOps &MOps) { Value *Op0, *Op1; if (MOps.MatchAndChain) { // Recurse through a chain of 'and' operands. This requires an extra check // vs. the 'or' matcher: we must find an "and X, 1" instruction somewhere // in the chain to know that all of the high bits are cleared. if (match(V, m_And(m_Value(Op0), m_One()))) { MOps.FoundAnd1 = true; return matchAndOrChain(Op0, MOps); } if (match(V, m_And(m_Value(Op0), m_Value(Op1)))) return matchAndOrChain(Op0, MOps) && matchAndOrChain(Op1, MOps); } else { // Recurse through a chain of 'or' operands. if (match(V, m_Or(m_Value(Op0), m_Value(Op1)))) return matchAndOrChain(Op0, MOps) && matchAndOrChain(Op1, MOps); } // We need a shift-right or a bare value representing a compare of bit 0 of // the original source operand. Value *Candidate; const APInt *BitIndex = nullptr; if (!match(V, m_LShr(m_Value(Candidate), m_APInt(BitIndex)))) Candidate = V; // Initialize result source operand. if (!MOps.Root) MOps.Root = Candidate; // The shift constant is out-of-range? This code hasn't been simplified. if (BitIndex && BitIndex->uge(MOps.Mask.getBitWidth())) return false; // Fill in the mask bit derived from the shift constant. MOps.Mask.setBit(BitIndex ? BitIndex->getZExtValue() : 0); return MOps.Root == Candidate; } /// Match patterns that correspond to "any-bits-set" and "all-bits-set". /// These will include a chain of 'or' or 'and'-shifted bits from a /// common source value: /// and (or (lshr X, C), ...), 1 --> (X & CMask) != 0 /// and (and (lshr X, C), ...), 1 --> (X & CMask) == CMask /// Note: "any-bits-clear" and "all-bits-clear" are variations of these patterns /// that differ only with a final 'not' of the result. We expect that final /// 'not' to be folded with the compare that we create here (invert predicate). static bool foldAnyOrAllBitsSet(Instruction &I) { // The 'any-bits-set' ('or' chain) pattern is simpler to match because the // final "and X, 1" instruction must be the final op in the sequence. bool MatchAllBitsSet; if (match(&I, m_c_And(m_OneUse(m_And(m_Value(), m_Value())), m_Value()))) MatchAllBitsSet = true; else if (match(&I, m_And(m_OneUse(m_Or(m_Value(), m_Value())), m_One()))) MatchAllBitsSet = false; else return false; MaskOps MOps(I.getType()->getScalarSizeInBits(), MatchAllBitsSet); if (MatchAllBitsSet) { if (!matchAndOrChain(cast(&I), MOps) || !MOps.FoundAnd1) return false; } else { if (!matchAndOrChain(cast(&I)->getOperand(0), MOps)) return false; } // The pattern was found. Create a masked compare that replaces all of the // shift and logic ops. IRBuilder<> Builder(&I); Constant *Mask = ConstantInt::get(I.getType(), MOps.Mask); Value *And = Builder.CreateAnd(MOps.Root, Mask); Value *Cmp = MatchAllBitsSet ? Builder.CreateICmpEQ(And, Mask) : Builder.CreateIsNotNull(And); Value *Zext = Builder.CreateZExt(Cmp, I.getType()); I.replaceAllUsesWith(Zext); ++NumAnyOrAllBitsSet; return true; } // Try to recognize below function as popcount intrinsic. // This is the "best" algorithm from // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel // Also used in TargetLowering::expandCTPOP(). // // int popcount(unsigned int i) { // i = i - ((i >> 1) & 0x55555555); // i = (i & 0x33333333) + ((i >> 2) & 0x33333333); // i = ((i + (i >> 4)) & 0x0F0F0F0F); // return (i * 0x01010101) >> 24; // } static bool tryToRecognizePopCount(Instruction &I) { if (I.getOpcode() != Instruction::LShr) return false; Type *Ty = I.getType(); if (!Ty->isIntOrIntVectorTy()) return false; unsigned Len = Ty->getScalarSizeInBits(); // FIXME: fix Len == 8 and other irregular type lengths. if (!(Len <= 128 && Len > 8 && Len % 8 == 0)) return false; APInt Mask55 = APInt::getSplat(Len, APInt(8, 0x55)); APInt Mask33 = APInt::getSplat(Len, APInt(8, 0x33)); APInt Mask0F = APInt::getSplat(Len, APInt(8, 0x0F)); APInt Mask01 = APInt::getSplat(Len, APInt(8, 0x01)); APInt MaskShift = APInt(Len, Len - 8); Value *Op0 = I.getOperand(0); Value *Op1 = I.getOperand(1); Value *MulOp0; // Matching "(i * 0x01010101...) >> 24". if ((match(Op0, m_Mul(m_Value(MulOp0), m_SpecificInt(Mask01)))) && match(Op1, m_SpecificInt(MaskShift))) { Value *ShiftOp0; // Matching "((i + (i >> 4)) & 0x0F0F0F0F...)". if (match(MulOp0, m_And(m_c_Add(m_LShr(m_Value(ShiftOp0), m_SpecificInt(4)), m_Deferred(ShiftOp0)), m_SpecificInt(Mask0F)))) { Value *AndOp0; // Matching "(i & 0x33333333...) + ((i >> 2) & 0x33333333...)". if (match(ShiftOp0, m_c_Add(m_And(m_Value(AndOp0), m_SpecificInt(Mask33)), m_And(m_LShr(m_Deferred(AndOp0), m_SpecificInt(2)), m_SpecificInt(Mask33))))) { Value *Root, *SubOp1; // Matching "i - ((i >> 1) & 0x55555555...)". if (match(AndOp0, m_Sub(m_Value(Root), m_Value(SubOp1))) && match(SubOp1, m_And(m_LShr(m_Specific(Root), m_SpecificInt(1)), m_SpecificInt(Mask55)))) { LLVM_DEBUG(dbgs() << "Recognized popcount intrinsic\n"); IRBuilder<> Builder(&I); Function *Func = Intrinsic::getDeclaration( I.getModule(), Intrinsic::ctpop, I.getType()); I.replaceAllUsesWith(Builder.CreateCall(Func, {Root})); ++NumPopCountRecognized; return true; } } } } return false; } /// Fold smin(smax(fptosi(x), C1), C2) to llvm.fptosi.sat(x), providing C1 and /// C2 saturate the value of the fp conversion. The transform is not reversable /// as the fptosi.sat is more defined than the input - all values produce a /// valid value for the fptosi.sat, where as some produce poison for original /// that were out of range of the integer conversion. The reversed pattern may /// use fmax and fmin instead. As we cannot directly reverse the transform, and /// it is not always profitable, we make it conditional on the cost being /// reported as lower by TTI. static bool tryToFPToSat(Instruction &I, TargetTransformInfo &TTI) { // Look for min(max(fptosi, converting to fptosi_sat. Value *In; const APInt *MinC, *MaxC; if (!match(&I, m_SMax(m_OneUse(m_SMin(m_OneUse(m_FPToSI(m_Value(In))), m_APInt(MinC))), m_APInt(MaxC))) && !match(&I, m_SMin(m_OneUse(m_SMax(m_OneUse(m_FPToSI(m_Value(In))), m_APInt(MaxC))), m_APInt(MinC)))) return false; // Check that the constants clamp a saturate. if (!(*MinC + 1).isPowerOf2() || -*MaxC != *MinC + 1) return false; Type *IntTy = I.getType(); Type *FpTy = In->getType(); Type *SatTy = IntegerType::get(IntTy->getContext(), (*MinC + 1).exactLogBase2() + 1); if (auto *VecTy = dyn_cast(IntTy)) SatTy = VectorType::get(SatTy, VecTy->getElementCount()); // Get the cost of the intrinsic, and check that against the cost of // fptosi+smin+smax InstructionCost SatCost = TTI.getIntrinsicInstrCost( IntrinsicCostAttributes(Intrinsic::fptosi_sat, SatTy, {In}, {FpTy}), TTI::TCK_RecipThroughput); SatCost += TTI.getCastInstrCost(Instruction::SExt, SatTy, IntTy, TTI::CastContextHint::None, TTI::TCK_RecipThroughput); InstructionCost MinMaxCost = TTI.getCastInstrCost( Instruction::FPToSI, IntTy, FpTy, TTI::CastContextHint::None, TTI::TCK_RecipThroughput); MinMaxCost += TTI.getIntrinsicInstrCost( IntrinsicCostAttributes(Intrinsic::smin, IntTy, {IntTy}), TTI::TCK_RecipThroughput); MinMaxCost += TTI.getIntrinsicInstrCost( IntrinsicCostAttributes(Intrinsic::smax, IntTy, {IntTy}), TTI::TCK_RecipThroughput); if (SatCost >= MinMaxCost) return false; IRBuilder<> Builder(&I); Function *Fn = Intrinsic::getDeclaration(I.getModule(), Intrinsic::fptosi_sat, {SatTy, FpTy}); Value *Sat = Builder.CreateCall(Fn, In); I.replaceAllUsesWith(Builder.CreateSExt(Sat, IntTy)); return true; } /// Try to replace a mathlib call to sqrt with the LLVM intrinsic. This avoids /// pessimistic codegen that has to account for setting errno and can enable /// vectorization. static bool foldSqrt(Instruction &I, TargetTransformInfo &TTI, TargetLibraryInfo &TLI) { // Match a call to sqrt mathlib function. auto *Call = dyn_cast(&I); if (!Call) return false; Module *M = Call->getModule(); LibFunc Func; if (!TLI.getLibFunc(*Call, Func) || !isLibFuncEmittable(M, &TLI, Func)) return false; if (Func != LibFunc_sqrt && Func != LibFunc_sqrtf && Func != LibFunc_sqrtl) return false; // If (1) this is a sqrt libcall, (2) we can assume that NAN is not created, // and (3) we would not end up lowering to a libcall anyway (which could // change the value of errno), then: // (1) the operand arg must not be less than -0.0. // (2) errno won't be set. // (3) it is safe to convert this to an intrinsic call. // TODO: Check if the arg is known non-negative. Type *Ty = Call->getType(); if (TTI.haveFastSqrt(Ty) && Call->hasNoNaNs()) { IRBuilder<> Builder(&I); IRBuilderBase::FastMathFlagGuard Guard(Builder); Builder.setFastMathFlags(Call->getFastMathFlags()); Function *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, Ty); Value *NewSqrt = Builder.CreateCall(Sqrt, Call->getArgOperand(0), "sqrt"); I.replaceAllUsesWith(NewSqrt); // Explicitly erase the old call because a call with side effects is not // trivially dead. I.eraseFromParent(); return true; } return false; } /// This is the entry point for folds that could be implemented in regular /// InstCombine, but they are separated because they are not expected to /// occur frequently and/or have more than a constant-length pattern match. static bool foldUnusualPatterns(Function &F, DominatorTree &DT, TargetTransformInfo &TTI, TargetLibraryInfo &TLI) { bool MadeChange = false; for (BasicBlock &BB : F) { // Ignore unreachable basic blocks. if (!DT.isReachableFromEntry(&BB)) continue; // Walk the block backwards for efficiency. We're matching a chain of // use->defs, so we're more likely to succeed by starting from the bottom. // Also, we want to avoid matching partial patterns. // TODO: It would be more efficient if we removed dead instructions // iteratively in this loop rather than waiting until the end. for (Instruction &I : make_early_inc_range(llvm::reverse(BB))) { MadeChange |= foldAnyOrAllBitsSet(I); MadeChange |= foldGuardedFunnelShift(I, DT); MadeChange |= tryToRecognizePopCount(I); MadeChange |= tryToFPToSat(I, TTI); MadeChange |= foldSqrt(I, TTI, TLI); } } // We're done with transforms, so remove dead instructions. if (MadeChange) for (BasicBlock &BB : F) SimplifyInstructionsInBlock(&BB); return MadeChange; } /// This is the entry point for all transforms. Pass manager differences are /// handled in the callers of this function. static bool runImpl(Function &F, AssumptionCache &AC, TargetTransformInfo &TTI, TargetLibraryInfo &TLI, DominatorTree &DT) { bool MadeChange = false; const DataLayout &DL = F.getParent()->getDataLayout(); TruncInstCombine TIC(AC, TLI, DL, DT); MadeChange |= TIC.run(F); MadeChange |= foldUnusualPatterns(F, DT, TTI, TLI); return MadeChange; } void AggressiveInstCombinerLegacyPass::getAnalysisUsage( AnalysisUsage &AU) const { AU.setPreservesCFG(); AU.addRequired(); AU.addRequired(); AU.addRequired(); AU.addRequired(); AU.addPreserved(); AU.addPreserved(); AU.addPreserved(); AU.addPreserved(); } bool AggressiveInstCombinerLegacyPass::runOnFunction(Function &F) { auto &AC = getAnalysis().getAssumptionCache(F); auto &TLI = getAnalysis().getTLI(F); auto &DT = getAnalysis().getDomTree(); auto &TTI = getAnalysis().getTTI(F); return runImpl(F, AC, TTI, TLI, DT); } PreservedAnalyses AggressiveInstCombinePass::run(Function &F, FunctionAnalysisManager &AM) { auto &AC = AM.getResult(F); auto &TLI = AM.getResult(F); auto &DT = AM.getResult(F); auto &TTI = AM.getResult(F); if (!runImpl(F, AC, TTI, TLI, DT)) { // No changes, all analyses are preserved. return PreservedAnalyses::all(); } // Mark all the analyses that instcombine updates as preserved. PreservedAnalyses PA; PA.preserveSet(); return PA; } char AggressiveInstCombinerLegacyPass::ID = 0; INITIALIZE_PASS_BEGIN(AggressiveInstCombinerLegacyPass, "aggressive-instcombine", "Combine pattern based expressions", false, false) INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) INITIALIZE_PASS_END(AggressiveInstCombinerLegacyPass, "aggressive-instcombine", "Combine pattern based expressions", false, false) // Initialization Routines void llvm::initializeAggressiveInstCombine(PassRegistry &Registry) { initializeAggressiveInstCombinerLegacyPassPass(Registry); } void LLVMInitializeAggressiveInstCombiner(LLVMPassRegistryRef R) { initializeAggressiveInstCombinerLegacyPassPass(*unwrap(R)); } FunctionPass *llvm::createAggressiveInstCombinerPass() { return new AggressiveInstCombinerLegacyPass(); } void LLVMAddAggressiveInstCombinerPass(LLVMPassManagerRef PM) { unwrap(PM)->add(createAggressiveInstCombinerPass()); }