X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FInstCombine%2FInstCombineCalls.cpp;h=cc7d4be7a78063cbfab384f282609aa655bc307c;hp=be1b5aa50b187981b4619a9f9614e771e40362e9;hb=5db31461109dbf8e9a12a78e3128849461d1a2e2;hpb=29f94c72014eaa5d0d3b920686e689e79759cacb diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp index be1b5aa50b1..cc7d4be7a78 100644 --- a/lib/Transforms/InstCombine/InstCombineCalls.cpp +++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp @@ -11,14 +11,17 @@ // //===----------------------------------------------------------------------===// -#include "InstCombine.h" +#include "InstCombineInternal.h" #include "llvm/ADT/Statistic.h" +#include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/MemoryBuiltins.h" #include "llvm/IR/CallSite.h" -#include "llvm/IR/DataLayout.h" +#include "llvm/IR/Dominators.h" #include "llvm/IR/PatternMatch.h" +#include "llvm/IR/Statepoint.h" #include "llvm/Transforms/Utils/BuildLibCalls.h" #include "llvm/Transforms/Utils/Local.h" +#include "llvm/Transforms/Utils/SimplifyLibCalls.h" using namespace llvm; using namespace PatternMatch; @@ -58,14 +61,13 @@ static Type *reduceToSingleValueType(Type *T) { } Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { - unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL); - unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL); + unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, MI, AC, DT); + unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, MI, AC, DT); unsigned MinAlign = std::min(DstAlign, SrcAlign); unsigned CopyAlign = MI->getAlignment(); if (CopyAlign < MinAlign) { - MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), - MinAlign, false)); + MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), MinAlign, false)); return MI; } @@ -105,7 +107,7 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { if (StrippedDest != MI->getArgOperand(0)) { Type *SrcETy = cast(StrippedDest->getType()) ->getElementType(); - if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) { + if (SrcETy->isSized() && DL.getTypeStoreSize(SrcETy) == Size) { // The SrcETy might be something like {{{double}}} or [1 x double]. Rip // down through these levels if so. SrcETy = reduceToSingleValueType(SrcETy); @@ -117,15 +119,14 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { // If the memcpy has metadata describing the members, see if we can // get the TBAA tag describing our copy. if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) { - if (M->getNumOperands() == 3 && - M->getOperand(0) && - isa(M->getOperand(0)) && - cast(M->getOperand(0))->isNullValue() && + if (M->getNumOperands() == 3 && M->getOperand(0) && + mdconst::hasa(M->getOperand(0)) && + mdconst::extract(M->getOperand(0))->isNullValue() && M->getOperand(1) && - isa(M->getOperand(1)) && - cast(M->getOperand(1))->getValue() == Size && - M->getOperand(2) && - isa(M->getOperand(2))) + mdconst::hasa(M->getOperand(1)) && + mdconst::extract(M->getOperand(1))->getValue() == + Size && + M->getOperand(2) && isa(M->getOperand(2))) CopyMD = cast(M->getOperand(2)); } } @@ -154,7 +155,7 @@ Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { } Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { - unsigned Alignment = getKnownAlignment(MI->getDest(), DL); + unsigned Alignment = getKnownAlignment(MI->getDest(), DL, MI, AC, DT); if (MI->getAlignment() < Alignment) { MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Alignment, false)); @@ -196,11 +197,265 @@ Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { return nullptr; } +static Value *SimplifyX86immshift(const IntrinsicInst &II, + InstCombiner::BuilderTy &Builder) { + bool LogicalShift = false; + bool ShiftLeft = false; + + switch (II.getIntrinsicID()) { + default: + return nullptr; + case Intrinsic::x86_sse2_psra_d: + case Intrinsic::x86_sse2_psra_w: + case Intrinsic::x86_sse2_psrai_d: + case Intrinsic::x86_sse2_psrai_w: + case Intrinsic::x86_avx2_psra_d: + case Intrinsic::x86_avx2_psra_w: + case Intrinsic::x86_avx2_psrai_d: + case Intrinsic::x86_avx2_psrai_w: + LogicalShift = false; ShiftLeft = false; + break; + case Intrinsic::x86_sse2_psrl_d: + case Intrinsic::x86_sse2_psrl_q: + case Intrinsic::x86_sse2_psrl_w: + case Intrinsic::x86_sse2_psrli_d: + case Intrinsic::x86_sse2_psrli_q: + case Intrinsic::x86_sse2_psrli_w: + case Intrinsic::x86_avx2_psrl_d: + case Intrinsic::x86_avx2_psrl_q: + case Intrinsic::x86_avx2_psrl_w: + case Intrinsic::x86_avx2_psrli_d: + case Intrinsic::x86_avx2_psrli_q: + case Intrinsic::x86_avx2_psrli_w: + LogicalShift = true; ShiftLeft = false; + break; + case Intrinsic::x86_sse2_psll_d: + case Intrinsic::x86_sse2_psll_q: + case Intrinsic::x86_sse2_psll_w: + case Intrinsic::x86_sse2_pslli_d: + case Intrinsic::x86_sse2_pslli_q: + case Intrinsic::x86_sse2_pslli_w: + case Intrinsic::x86_avx2_psll_d: + case Intrinsic::x86_avx2_psll_q: + case Intrinsic::x86_avx2_psll_w: + case Intrinsic::x86_avx2_pslli_d: + case Intrinsic::x86_avx2_pslli_q: + case Intrinsic::x86_avx2_pslli_w: + LogicalShift = true; ShiftLeft = true; + break; + } + assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left"); + + // Simplify if count is constant. + auto Arg1 = II.getArgOperand(1); + auto CAZ = dyn_cast(Arg1); + auto CDV = dyn_cast(Arg1); + auto CInt = dyn_cast(Arg1); + if (!CAZ && !CDV && !CInt) + return nullptr; + + APInt Count(64, 0); + if (CDV) { + // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector + // operand to compute the shift amount. + auto VT = cast(CDV->getType()); + unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits(); + assert((64 % BitWidth) == 0 && "Unexpected packed shift size"); + unsigned NumSubElts = 64 / BitWidth; + + // Concatenate the sub-elements to create the 64-bit value. + for (unsigned i = 0; i != NumSubElts; ++i) { + unsigned SubEltIdx = (NumSubElts - 1) - i; + auto SubElt = cast(CDV->getElementAsConstant(SubEltIdx)); + Count = Count.shl(BitWidth); + Count |= SubElt->getValue().zextOrTrunc(64); + } + } + else if (CInt) + Count = CInt->getValue(); + + auto Vec = II.getArgOperand(0); + auto VT = cast(Vec->getType()); + auto SVT = VT->getElementType(); + unsigned VWidth = VT->getNumElements(); + unsigned BitWidth = SVT->getPrimitiveSizeInBits(); + + // If shift-by-zero then just return the original value. + if (Count == 0) + return Vec; + + // Handle cases when Shift >= BitWidth. + if (Count.uge(BitWidth)) { + // If LogicalShift - just return zero. + if (LogicalShift) + return ConstantAggregateZero::get(VT); + + // If ArithmeticShift - clamp Shift to (BitWidth - 1). + Count = APInt(64, BitWidth - 1); + } + + // Get a constant vector of the same type as the first operand. + auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth)); + auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt); + + if (ShiftLeft) + return Builder.CreateShl(Vec, ShiftVec); + + if (LogicalShift) + return Builder.CreateLShr(Vec, ShiftVec); + + return Builder.CreateAShr(Vec, ShiftVec); +} + +static Value *SimplifyX86extend(const IntrinsicInst &II, + InstCombiner::BuilderTy &Builder, + bool SignExtend) { + VectorType *SrcTy = cast(II.getArgOperand(0)->getType()); + VectorType *DstTy = cast(II.getType()); + unsigned NumDstElts = DstTy->getNumElements(); + + // Extract a subvector of the first NumDstElts lanes and sign/zero extend. + SmallVector ShuffleMask; + for (int i = 0; i != (int)NumDstElts; ++i) + ShuffleMask.push_back(i); + + Value *SV = Builder.CreateShuffleVector(II.getArgOperand(0), + UndefValue::get(SrcTy), ShuffleMask); + return SignExtend ? Builder.CreateSExt(SV, DstTy) + : Builder.CreateZExt(SV, DstTy); +} + +static Value *SimplifyX86insertps(const IntrinsicInst &II, + InstCombiner::BuilderTy &Builder) { + if (auto *CInt = dyn_cast(II.getArgOperand(2))) { + VectorType *VecTy = cast(II.getType()); + assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type"); + + // The immediate permute control byte looks like this: + // [3:0] - zero mask for each 32-bit lane + // [5:4] - select one 32-bit destination lane + // [7:6] - select one 32-bit source lane + + uint8_t Imm = CInt->getZExtValue(); + uint8_t ZMask = Imm & 0xf; + uint8_t DestLane = (Imm >> 4) & 0x3; + uint8_t SourceLane = (Imm >> 6) & 0x3; + + ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy); + + // If all zero mask bits are set, this was just a weird way to + // generate a zero vector. + if (ZMask == 0xf) + return ZeroVector; + + // Initialize by passing all of the first source bits through. + int ShuffleMask[4] = { 0, 1, 2, 3 }; + + // We may replace the second operand with the zero vector. + Value *V1 = II.getArgOperand(1); + + if (ZMask) { + // If the zero mask is being used with a single input or the zero mask + // overrides the destination lane, this is a shuffle with the zero vector. + if ((II.getArgOperand(0) == II.getArgOperand(1)) || + (ZMask & (1 << DestLane))) { + V1 = ZeroVector; + // We may still move 32-bits of the first source vector from one lane + // to another. + ShuffleMask[DestLane] = SourceLane; + // The zero mask may override the previous insert operation. + for (unsigned i = 0; i < 4; ++i) + if ((ZMask >> i) & 0x1) + ShuffleMask[i] = i + 4; + } else { + // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle? + return nullptr; + } + } else { + // Replace the selected destination lane with the selected source lane. + ShuffleMask[DestLane] = SourceLane + 4; + } + + return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask); + } + return nullptr; +} + +/// The shuffle mask for a perm2*128 selects any two halves of two 256-bit +/// source vectors, unless a zero bit is set. If a zero bit is set, +/// then ignore that half of the mask and clear that half of the vector. +static Value *SimplifyX86vperm2(const IntrinsicInst &II, + InstCombiner::BuilderTy &Builder) { + if (auto *CInt = dyn_cast(II.getArgOperand(2))) { + VectorType *VecTy = cast(II.getType()); + ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy); + + // The immediate permute control byte looks like this: + // [1:0] - select 128 bits from sources for low half of destination + // [2] - ignore + // [3] - zero low half of destination + // [5:4] - select 128 bits from sources for high half of destination + // [6] - ignore + // [7] - zero high half of destination + + uint8_t Imm = CInt->getZExtValue(); + + bool LowHalfZero = Imm & 0x08; + bool HighHalfZero = Imm & 0x80; + + // If both zero mask bits are set, this was just a weird way to + // generate a zero vector. + if (LowHalfZero && HighHalfZero) + return ZeroVector; + + // If 0 or 1 zero mask bits are set, this is a simple shuffle. + unsigned NumElts = VecTy->getNumElements(); + unsigned HalfSize = NumElts / 2; + SmallVector ShuffleMask(NumElts); + + // The high bit of the selection field chooses the 1st or 2nd operand. + bool LowInputSelect = Imm & 0x02; + bool HighInputSelect = Imm & 0x20; + + // The low bit of the selection field chooses the low or high half + // of the selected operand. + bool LowHalfSelect = Imm & 0x01; + bool HighHalfSelect = Imm & 0x10; + + // Determine which operand(s) are actually in use for this instruction. + Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0); + Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0); + + // If needed, replace operands based on zero mask. + V0 = LowHalfZero ? ZeroVector : V0; + V1 = HighHalfZero ? ZeroVector : V1; + + // Permute low half of result. + unsigned StartIndex = LowHalfSelect ? HalfSize : 0; + for (unsigned i = 0; i < HalfSize; ++i) + ShuffleMask[i] = StartIndex + i; + + // Permute high half of result. + StartIndex = HighHalfSelect ? HalfSize : 0; + StartIndex += NumElts; + for (unsigned i = 0; i < HalfSize; ++i) + ShuffleMask[i + HalfSize] = StartIndex + i; + + return Builder.CreateShuffleVector(V0, V1, ShuffleMask); + } + return nullptr; +} + /// visitCallInst - CallInst simplification. This mostly only handles folding /// of intrinsic instructions. For normal calls, it allows visitCallSite to do /// the heavy lifting. /// Instruction *InstCombiner::visitCallInst(CallInst &CI) { + auto Args = CI.arg_operands(); + if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL, + TLI, DT, AC)) + return ReplaceInstUsesWith(CI, V); + if (isFreeCall(&CI, TLI)) return visitFree(CI); @@ -322,7 +577,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { uint32_t BitWidth = IT->getBitWidth(); APInt KnownZero(BitWidth, 0); APInt KnownOne(BitWidth, 0); - computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne); + computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II); unsigned TrailingZeros = KnownOne.countTrailingZeros(); APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); if ((Mask & KnownZero) == Mask) @@ -340,7 +595,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { uint32_t BitWidth = IT->getBitWidth(); APInt KnownZero(BitWidth, 0); APInt KnownOne(BitWidth, 0); - computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne); + computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II); unsigned LeadingZeros = KnownOne.countLeadingZeros(); APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); if ((Mask & KnownZero) == Mask) @@ -349,173 +604,192 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { } break; - case Intrinsic::uadd_with_overflow: { - Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); - IntegerType *IT = cast(II->getArgOperand(0)->getType()); - uint32_t BitWidth = IT->getBitWidth(); - APInt LHSKnownZero(BitWidth, 0); - APInt LHSKnownOne(BitWidth, 0); - computeKnownBits(LHS, LHSKnownZero, LHSKnownOne); - bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; - bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; - - if (LHSKnownNegative || LHSKnownPositive) { - APInt RHSKnownZero(BitWidth, 0); - APInt RHSKnownOne(BitWidth, 0); - computeKnownBits(RHS, RHSKnownZero, RHSKnownOne); - bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; - bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; - if (LHSKnownNegative && RHSKnownNegative) { - // The sign bit is set in both cases: this MUST overflow. - // Create a simple add instruction, and insert it into the struct. - Value *Add = Builder->CreateAdd(LHS, RHS); - Add->takeName(&CI); - Constant *V[] = { - UndefValue::get(LHS->getType()), - ConstantInt::getTrue(II->getContext()) - }; - StructType *ST = cast(II->getType()); - Constant *Struct = ConstantStruct::get(ST, V); - return InsertValueInst::Create(Struct, Add, 0); - } - if (LHSKnownPositive && RHSKnownPositive) { - // The sign bit is clear in both cases: this CANNOT overflow. - // Create a simple add instruction, and insert it into the struct. - Value *Add = Builder->CreateNUWAdd(LHS, RHS); - Add->takeName(&CI); - Constant *V[] = { - UndefValue::get(LHS->getType()), - ConstantInt::getFalse(II->getContext()) - }; - StructType *ST = cast(II->getType()); - Constant *Struct = ConstantStruct::get(ST, V); - return InsertValueInst::Create(Struct, Add, 0); - } - } - } - // FALL THROUGH uadd into sadd + case Intrinsic::uadd_with_overflow: case Intrinsic::sadd_with_overflow: - // Canonicalize constants into the RHS. + case Intrinsic::umul_with_overflow: + case Intrinsic::smul_with_overflow: if (isa(II->getArgOperand(0)) && !isa(II->getArgOperand(1))) { + // Canonicalize constants into the RHS. Value *LHS = II->getArgOperand(0); II->setArgOperand(0, II->getArgOperand(1)); II->setArgOperand(1, LHS); return II; } + // fall through - // X + undef -> undef - if (isa(II->getArgOperand(1))) - return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); - - if (ConstantInt *RHS = dyn_cast(II->getArgOperand(1))) { - // X + 0 -> {X, false} - if (RHS->isZero()) { - Constant *V[] = { - UndefValue::get(II->getArgOperand(0)->getType()), - ConstantInt::getFalse(II->getContext()) - }; - Constant *Struct = - ConstantStruct::get(cast(II->getType()), V); - return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); - } - } - break; case Intrinsic::usub_with_overflow: - case Intrinsic::ssub_with_overflow: - // undef - X -> undef - // X - undef -> undef - if (isa(II->getArgOperand(0)) || - isa(II->getArgOperand(1))) - return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); - - if (ConstantInt *RHS = dyn_cast(II->getArgOperand(1))) { - // X - 0 -> {X, false} - if (RHS->isZero()) { - Constant *V[] = { - UndefValue::get(II->getArgOperand(0)->getType()), - ConstantInt::getFalse(II->getContext()) - }; - Constant *Struct = - ConstantStruct::get(cast(II->getType()), V); - return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); - } - } + case Intrinsic::ssub_with_overflow: { + OverflowCheckFlavor OCF = + IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID()); + assert(OCF != OCF_INVALID && "unexpected!"); + + Value *OperationResult = nullptr; + Constant *OverflowResult = nullptr; + if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1), + *II, OperationResult, OverflowResult)) + return CreateOverflowTuple(II, OperationResult, OverflowResult); + break; - case Intrinsic::umul_with_overflow: { - Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); - unsigned BitWidth = cast(LHS->getType())->getBitWidth(); - - APInt LHSKnownZero(BitWidth, 0); - APInt LHSKnownOne(BitWidth, 0); - computeKnownBits(LHS, LHSKnownZero, LHSKnownOne); - APInt RHSKnownZero(BitWidth, 0); - APInt RHSKnownOne(BitWidth, 0); - computeKnownBits(RHS, RHSKnownZero, RHSKnownOne); - - // Get the largest possible values for each operand. - APInt LHSMax = ~LHSKnownZero; - APInt RHSMax = ~RHSKnownZero; - - // If multiplying the maximum values does not overflow then we can turn - // this into a plain NUW mul. - bool Overflow; - LHSMax.umul_ov(RHSMax, Overflow); - if (!Overflow) { - Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); - Constant *V[] = { - UndefValue::get(LHS->getType()), - Builder->getFalse() - }; - Constant *Struct = ConstantStruct::get(cast(II->getType()),V); - return InsertValueInst::Create(Struct, Mul, 0); - } - } // FALL THROUGH - case Intrinsic::smul_with_overflow: + } + + case Intrinsic::minnum: + case Intrinsic::maxnum: { + Value *Arg0 = II->getArgOperand(0); + Value *Arg1 = II->getArgOperand(1); + + // fmin(x, x) -> x + if (Arg0 == Arg1) + return ReplaceInstUsesWith(CI, Arg0); + + const ConstantFP *C0 = dyn_cast(Arg0); + const ConstantFP *C1 = dyn_cast(Arg1); + // Canonicalize constants into the RHS. - if (isa(II->getArgOperand(0)) && - !isa(II->getArgOperand(1))) { - Value *LHS = II->getArgOperand(0); - II->setArgOperand(0, II->getArgOperand(1)); - II->setArgOperand(1, LHS); + if (C0 && !C1) { + II->setArgOperand(0, Arg1); + II->setArgOperand(1, Arg0); return II; } - // X * undef -> undef - if (isa(II->getArgOperand(1))) - return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); - - if (ConstantInt *RHSI = dyn_cast(II->getArgOperand(1))) { - // X*0 -> {0, false} - if (RHSI->isZero()) - return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); - - // X * 1 -> {X, false} - if (RHSI->equalsInt(1)) { - Constant *V[] = { - UndefValue::get(II->getArgOperand(0)->getType()), - ConstantInt::getFalse(II->getContext()) - }; - Constant *Struct = - ConstantStruct::get(cast(II->getType()), V); - return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); + // fmin(x, nan) -> x + if (C1 && C1->isNaN()) + return ReplaceInstUsesWith(CI, Arg0); + + // This is the value because if undef were NaN, we would return the other + // value and cannot return a NaN unless both operands are. + // + // fmin(undef, x) -> x + if (isa(Arg0)) + return ReplaceInstUsesWith(CI, Arg1); + + // fmin(x, undef) -> x + if (isa(Arg1)) + return ReplaceInstUsesWith(CI, Arg0); + + Value *X = nullptr; + Value *Y = nullptr; + if (II->getIntrinsicID() == Intrinsic::minnum) { + // fmin(x, fmin(x, y)) -> fmin(x, y) + // fmin(y, fmin(x, y)) -> fmin(x, y) + if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) { + if (Arg0 == X || Arg0 == Y) + return ReplaceInstUsesWith(CI, Arg1); + } + + // fmin(fmin(x, y), x) -> fmin(x, y) + // fmin(fmin(x, y), y) -> fmin(x, y) + if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) { + if (Arg1 == X || Arg1 == Y) + return ReplaceInstUsesWith(CI, Arg0); + } + + // TODO: fmin(nnan x, inf) -> x + // TODO: fmin(nnan ninf x, flt_max) -> x + if (C1 && C1->isInfinity()) { + // fmin(x, -inf) -> -inf + if (C1->isNegative()) + return ReplaceInstUsesWith(CI, Arg1); + } + } else { + assert(II->getIntrinsicID() == Intrinsic::maxnum); + // fmax(x, fmax(x, y)) -> fmax(x, y) + // fmax(y, fmax(x, y)) -> fmax(x, y) + if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) { + if (Arg0 == X || Arg0 == Y) + return ReplaceInstUsesWith(CI, Arg1); + } + + // fmax(fmax(x, y), x) -> fmax(x, y) + // fmax(fmax(x, y), y) -> fmax(x, y) + if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) { + if (Arg1 == X || Arg1 == Y) + return ReplaceInstUsesWith(CI, Arg0); + } + + // TODO: fmax(nnan x, -inf) -> x + // TODO: fmax(nnan ninf x, -flt_max) -> x + if (C1 && C1->isInfinity()) { + // fmax(x, inf) -> inf + if (!C1->isNegative()) + return ReplaceInstUsesWith(CI, Arg1); } } break; + } case Intrinsic::ppc_altivec_lvx: case Intrinsic::ppc_altivec_lvxl: // Turn PPC lvx -> load if the pointer is known aligned. - if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) { + if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >= + 16) { Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), PointerType::getUnqual(II->getType())); return new LoadInst(Ptr); } break; + case Intrinsic::ppc_vsx_lxvw4x: + case Intrinsic::ppc_vsx_lxvd2x: { + // Turn PPC VSX loads into normal loads. + Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), + PointerType::getUnqual(II->getType())); + return new LoadInst(Ptr, Twine(""), false, 1); + } case Intrinsic::ppc_altivec_stvx: case Intrinsic::ppc_altivec_stvxl: // Turn stvx -> store if the pointer is known aligned. - if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL) >= 16) { + if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >= + 16) { + Type *OpPtrTy = + PointerType::getUnqual(II->getArgOperand(0)->getType()); + Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); + return new StoreInst(II->getArgOperand(0), Ptr); + } + break; + case Intrinsic::ppc_vsx_stxvw4x: + case Intrinsic::ppc_vsx_stxvd2x: { + // Turn PPC VSX stores into normal stores. + Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType()); + Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); + return new StoreInst(II->getArgOperand(0), Ptr, false, 1); + } + case Intrinsic::ppc_qpx_qvlfs: + // Turn PPC QPX qvlfs -> load if the pointer is known aligned. + if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >= + 16) { + Type *VTy = VectorType::get(Builder->getFloatTy(), + II->getType()->getVectorNumElements()); + Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), + PointerType::getUnqual(VTy)); + Value *Load = Builder->CreateLoad(Ptr); + return new FPExtInst(Load, II->getType()); + } + break; + case Intrinsic::ppc_qpx_qvlfd: + // Turn PPC QPX qvlfd -> load if the pointer is known aligned. + if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >= + 32) { + Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), + PointerType::getUnqual(II->getType())); + return new LoadInst(Ptr); + } + break; + case Intrinsic::ppc_qpx_qvstfs: + // Turn PPC QPX qvstfs -> store if the pointer is known aligned. + if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >= + 16) { + Type *VTy = VectorType::get(Builder->getFloatTy(), + II->getArgOperand(0)->getType()->getVectorNumElements()); + Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy); + Type *OpPtrTy = PointerType::getUnqual(VTy); + Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); + return new StoreInst(TOp, Ptr); + } + break; + case Intrinsic::ppc_qpx_qvstfd: + // Turn PPC QPX qvstfd -> store if the pointer is known aligned. + if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >= + 32) { Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType()); Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); @@ -526,7 +800,8 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { case Intrinsic::x86_sse2_storeu_pd: case Intrinsic::x86_sse2_storeu_dq: // Turn X86 storeu -> store if the pointer is known aligned. - if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) { + if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >= + 16) { Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(1)->getType()); Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); @@ -556,108 +831,126 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { break; } - // Constant fold << Ci. - // FIXME: We don't handle _dq because it's a shift of an i128, but is - // represented in the IR as <2 x i64>. A per element shift is wrong. - case Intrinsic::x86_sse2_psll_d: - case Intrinsic::x86_sse2_psll_q: - case Intrinsic::x86_sse2_psll_w: + // Constant fold ashr( , Ci ). + // Constant fold lshr( , Ci ). + // Constant fold shl( , Ci ). + case Intrinsic::x86_sse2_psrai_d: + case Intrinsic::x86_sse2_psrai_w: + case Intrinsic::x86_avx2_psrai_d: + case Intrinsic::x86_avx2_psrai_w: + case Intrinsic::x86_sse2_psrli_d: + case Intrinsic::x86_sse2_psrli_q: + case Intrinsic::x86_sse2_psrli_w: + case Intrinsic::x86_avx2_psrli_d: + case Intrinsic::x86_avx2_psrli_q: + case Intrinsic::x86_avx2_psrli_w: case Intrinsic::x86_sse2_pslli_d: case Intrinsic::x86_sse2_pslli_q: case Intrinsic::x86_sse2_pslli_w: - case Intrinsic::x86_avx2_psll_d: - case Intrinsic::x86_avx2_psll_q: - case Intrinsic::x86_avx2_psll_w: case Intrinsic::x86_avx2_pslli_d: case Intrinsic::x86_avx2_pslli_q: case Intrinsic::x86_avx2_pslli_w: + if (Value *V = SimplifyX86immshift(*II, *Builder)) + return ReplaceInstUsesWith(*II, V); + break; + + case Intrinsic::x86_sse2_psra_d: + case Intrinsic::x86_sse2_psra_w: + case Intrinsic::x86_avx2_psra_d: + case Intrinsic::x86_avx2_psra_w: case Intrinsic::x86_sse2_psrl_d: case Intrinsic::x86_sse2_psrl_q: case Intrinsic::x86_sse2_psrl_w: - case Intrinsic::x86_sse2_psrli_d: - case Intrinsic::x86_sse2_psrli_q: - case Intrinsic::x86_sse2_psrli_w: case Intrinsic::x86_avx2_psrl_d: case Intrinsic::x86_avx2_psrl_q: case Intrinsic::x86_avx2_psrl_w: - case Intrinsic::x86_avx2_psrli_d: - case Intrinsic::x86_avx2_psrli_q: - case Intrinsic::x86_avx2_psrli_w: { - // Simplify if count is constant. To 0 if >= BitWidth, - // otherwise to shl/lshr. - auto CDV = dyn_cast(II->getArgOperand(1)); - auto CInt = dyn_cast(II->getArgOperand(1)); - if (!CDV && !CInt) - break; - ConstantInt *Count; - if (CDV) - Count = cast(CDV->getElementAsConstant(0)); - else - Count = CInt; - - auto Vec = II->getArgOperand(0); - auto VT = cast(Vec->getType()); - if (Count->getZExtValue() > - VT->getElementType()->getPrimitiveSizeInBits() - 1) - return ReplaceInstUsesWith( - CI, ConstantAggregateZero::get(Vec->getType())); - - bool isPackedShiftLeft = true; - switch (II->getIntrinsicID()) { - default : break; - case Intrinsic::x86_sse2_psrl_d: - case Intrinsic::x86_sse2_psrl_q: - case Intrinsic::x86_sse2_psrl_w: - case Intrinsic::x86_sse2_psrli_d: - case Intrinsic::x86_sse2_psrli_q: - case Intrinsic::x86_sse2_psrli_w: - case Intrinsic::x86_avx2_psrl_d: - case Intrinsic::x86_avx2_psrl_q: - case Intrinsic::x86_avx2_psrl_w: - case Intrinsic::x86_avx2_psrli_d: - case Intrinsic::x86_avx2_psrli_q: - case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break; + case Intrinsic::x86_sse2_psll_d: + case Intrinsic::x86_sse2_psll_q: + case Intrinsic::x86_sse2_psll_w: + case Intrinsic::x86_avx2_psll_d: + case Intrinsic::x86_avx2_psll_q: + case Intrinsic::x86_avx2_psll_w: { + if (Value *V = SimplifyX86immshift(*II, *Builder)) + return ReplaceInstUsesWith(*II, V); + + // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector + // operand to compute the shift amount. + auto ShiftAmt = II->getArgOperand(1); + auto ShiftType = cast(ShiftAmt->getType()); + assert(ShiftType->getPrimitiveSizeInBits() == 128 && + "Unexpected packed shift size"); + unsigned VWidth = ShiftType->getNumElements(); + + APInt DemandedElts = APInt::getLowBitsSet(VWidth, VWidth / 2); + APInt UndefElts(VWidth, 0); + if (Value *V = + SimplifyDemandedVectorElts(ShiftAmt, DemandedElts, UndefElts)) { + II->setArgOperand(1, V); + return II; } - - unsigned VWidth = VT->getNumElements(); - // Get a constant vector of the same type as the first operand. - auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue()); - if (isPackedShiftLeft) - return BinaryOperator::CreateShl(Vec, - Builder->CreateVectorSplat(VWidth, VTCI)); - - return BinaryOperator::CreateLShr(Vec, - Builder->CreateVectorSplat(VWidth, VTCI)); + break; } + case Intrinsic::x86_sse41_pmovsxbd: + case Intrinsic::x86_sse41_pmovsxbq: case Intrinsic::x86_sse41_pmovsxbw: - case Intrinsic::x86_sse41_pmovsxwd: case Intrinsic::x86_sse41_pmovsxdq: + case Intrinsic::x86_sse41_pmovsxwd: + case Intrinsic::x86_sse41_pmovsxwq: + case Intrinsic::x86_avx2_pmovsxbd: + case Intrinsic::x86_avx2_pmovsxbq: + case Intrinsic::x86_avx2_pmovsxbw: + case Intrinsic::x86_avx2_pmovsxdq: + case Intrinsic::x86_avx2_pmovsxwd: + case Intrinsic::x86_avx2_pmovsxwq: + if (Value *V = SimplifyX86extend(*II, *Builder, true)) + return ReplaceInstUsesWith(*II, V); + break; + + case Intrinsic::x86_sse41_pmovzxbd: + case Intrinsic::x86_sse41_pmovzxbq: case Intrinsic::x86_sse41_pmovzxbw: + case Intrinsic::x86_sse41_pmovzxdq: case Intrinsic::x86_sse41_pmovzxwd: - case Intrinsic::x86_sse41_pmovzxdq: { - // pmov{s|z}x ignores the upper half of their input vectors. - unsigned VWidth = - cast(II->getArgOperand(0)->getType())->getNumElements(); - unsigned LowHalfElts = VWidth / 2; - APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); - APInt UndefElts(VWidth, 0); - if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), - InputDemandedElts, - UndefElts)) { - II->setArgOperand(0, TmpV); - return II; - } + case Intrinsic::x86_sse41_pmovzxwq: + case Intrinsic::x86_avx2_pmovzxbd: + case Intrinsic::x86_avx2_pmovzxbq: + case Intrinsic::x86_avx2_pmovzxbw: + case Intrinsic::x86_avx2_pmovzxdq: + case Intrinsic::x86_avx2_pmovzxwd: + case Intrinsic::x86_avx2_pmovzxwq: + if (Value *V = SimplifyX86extend(*II, *Builder, false)) + return ReplaceInstUsesWith(*II, V); + break; + + case Intrinsic::x86_sse41_insertps: + if (Value *V = SimplifyX86insertps(*II, *Builder)) + return ReplaceInstUsesWith(*II, V); break; - } case Intrinsic::x86_sse4a_insertqi: { // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top // ones undef // TODO: eventually we should lower this intrinsic to IR - if (auto CIWidth = dyn_cast(II->getArgOperand(2))) { - if (auto CIStart = dyn_cast(II->getArgOperand(3))) { - if (CIWidth->equalsInt(64) && CIStart->isZero()) { + if (auto CILength = dyn_cast(II->getArgOperand(2))) { + if (auto CIIndex = dyn_cast(II->getArgOperand(3))) { + unsigned Index = CIIndex->getZExtValue(); + // From AMD documentation: "a value of zero in the field length is + // defined as length of 64". + unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue(); + + // From AMD documentation: "If the sum of the bit index + length field + // is greater than 64, the results are undefined". + unsigned End = Index + Length; + + // Note that both field index and field length are 8-bit quantities. + // Since variables 'Index' and 'Length' are unsigned values + // obtained from zero-extending field index and field length + // respectively, their sum should never wrap around. + if (End > 64) + return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); + + if (Length == 64 && Index == 0) { Value *Vec = II->getArgOperand(1); Value *Undef = UndefValue::get(Vec->getType()); const uint32_t Mask[] = { 0, 2 }; @@ -665,8 +958,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { CI, Builder->CreateShuffleVector( Vec, Undef, ConstantDataVector::get( - II->getContext(), ArrayRef(Mask)))); - + II->getContext(), makeArrayRef(Mask)))); } else if (auto Source = dyn_cast(II->getArgOperand(0))) { if (Source->hasOneUse() && @@ -674,37 +966,34 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { // If the source of the insert has only one use and it's another // insert (and they're both inserting from the same vector), try to // bundle both together. - auto CISourceWidth = + auto CISourceLength = dyn_cast(Source->getArgOperand(2)); - auto CISourceStart = + auto CISourceIndex = dyn_cast(Source->getArgOperand(3)); - if (CISourceStart && CISourceWidth) { - unsigned Start = CIStart->getZExtValue(); - unsigned Width = CIWidth->getZExtValue(); - unsigned End = Start + Width; - unsigned SourceStart = CISourceStart->getZExtValue(); - unsigned SourceWidth = CISourceWidth->getZExtValue(); - unsigned SourceEnd = SourceStart + SourceWidth; - unsigned NewStart, NewWidth; + if (CISourceIndex && CISourceLength) { + unsigned SourceIndex = CISourceIndex->getZExtValue(); + unsigned SourceLength = CISourceLength->getZExtValue(); + unsigned SourceEnd = SourceIndex + SourceLength; + unsigned NewIndex, NewLength; bool ShouldReplace = false; - if (Start <= SourceStart && SourceStart <= End) { - NewStart = Start; - NewWidth = std::max(End, SourceEnd) - NewStart; + if (Index <= SourceIndex && SourceIndex <= End) { + NewIndex = Index; + NewLength = std::max(End, SourceEnd) - NewIndex; ShouldReplace = true; - } else if (SourceStart <= Start && Start <= SourceEnd) { - NewStart = SourceStart; - NewWidth = std::max(SourceEnd, End) - NewStart; + } else if (SourceIndex <= Index && Index <= SourceEnd) { + NewIndex = SourceIndex; + NewLength = std::max(SourceEnd, End) - NewIndex; ShouldReplace = true; } if (ShouldReplace) { - Constant *ConstantWidth = ConstantInt::get( - II->getArgOperand(2)->getType(), NewWidth, false); - Constant *ConstantStart = ConstantInt::get( - II->getArgOperand(3)->getType(), NewStart, false); + Constant *ConstantLength = ConstantInt::get( + II->getArgOperand(2)->getType(), NewLength, false); + Constant *ConstantIndex = ConstantInt::get( + II->getArgOperand(3)->getType(), NewIndex, false); Value *Args[4] = { Source->getArgOperand(0), - II->getArgOperand(1), ConstantWidth, - ConstantStart }; + II->getArgOperand(1), ConstantLength, + ConstantIndex }; Module *M = CI.getParent()->getParent()->getParent(); Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi); @@ -718,6 +1007,57 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { break; } + case Intrinsic::x86_sse41_pblendvb: + case Intrinsic::x86_sse41_blendvps: + case Intrinsic::x86_sse41_blendvpd: + case Intrinsic::x86_avx_blendv_ps_256: + case Intrinsic::x86_avx_blendv_pd_256: + case Intrinsic::x86_avx2_pblendvb: { + // Convert blendv* to vector selects if the mask is constant. + // This optimization is convoluted because the intrinsic is defined as + // getting a vector of floats or doubles for the ps and pd versions. + // FIXME: That should be changed. + + Value *Op0 = II->getArgOperand(0); + Value *Op1 = II->getArgOperand(1); + Value *Mask = II->getArgOperand(2); + + // fold (blend A, A, Mask) -> A + if (Op0 == Op1) + return ReplaceInstUsesWith(CI, Op0); + + // Zero Mask - select 1st argument. + if (isa(Mask)) + return ReplaceInstUsesWith(CI, Op0); + + // Constant Mask - select 1st/2nd argument lane based on top bit of mask. + if (auto C = dyn_cast(Mask)) { + auto Tyi1 = Builder->getInt1Ty(); + auto SelectorType = cast(Mask->getType()); + auto EltTy = SelectorType->getElementType(); + unsigned Size = SelectorType->getNumElements(); + unsigned BitWidth = + EltTy->isFloatTy() + ? 32 + : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth()); + assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) && + "Wrong arguments for variable blend intrinsic"); + SmallVector Selectors; + for (unsigned I = 0; I < Size; ++I) { + // The intrinsics only read the top bit + uint64_t Selector; + if (BitWidth == 8) + Selector = C->getElementAsInteger(I); + else + Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue(); + Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1))); + } + auto NewSelector = ConstantVector::get(Selectors); + return SelectInst::Create(NewSelector, Op1, Op0, "blendv"); + } + break; + } + case Intrinsic::x86_avx_vpermilvar_ps: case Intrinsic::x86_avx_vpermilvar_ps_256: case Intrinsic::x86_avx_vpermilvar_pd: @@ -758,8 +1098,21 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { return ReplaceInstUsesWith(CI, Shuffle); } + case Intrinsic::x86_avx_vperm2f128_pd_256: + case Intrinsic::x86_avx_vperm2f128_ps_256: + case Intrinsic::x86_avx_vperm2f128_si_256: + case Intrinsic::x86_avx2_vperm2i128: + if (Value *V = SimplifyX86vperm2(*II, *Builder)) + return ReplaceInstUsesWith(*II, V); + break; + case Intrinsic::ppc_altivec_vperm: // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. + // Note that ppc_altivec_vperm has a big-endian bias, so when creating + // a vectorshuffle for little endian, we must undo the transformation + // performed on vec_perm in altivec.h. That is, we must complement + // the permutation mask with respect to 31 and reverse the order of + // V1 and V2. if (Constant *Mask = dyn_cast(II->getArgOperand(2))) { assert(Mask->getType()->getVectorNumElements() == 16 && "Bad type for intrinsic!"); @@ -792,10 +1145,14 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { unsigned Idx = cast(Mask->getAggregateElement(i))->getZExtValue(); Idx &= 31; // Match the hardware behavior. + if (DL.isLittleEndian()) + Idx = 31 - Idx; if (!ExtractedElts[Idx]) { + Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0; + Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1; ExtractedElts[Idx] = - Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, + Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse, Builder->getInt32(Idx&15)); } @@ -822,7 +1179,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { case Intrinsic::arm_neon_vst2lane: case Intrinsic::arm_neon_vst3lane: case Intrinsic::arm_neon_vst4lane: { - unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL); + unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT); unsigned AlignArg = II->getNumArgOperands() - 1; ConstantInt *IntrAlign = dyn_cast(II->getArgOperand(AlignArg)); if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { @@ -873,6 +1230,20 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { break; } + case Intrinsic::AMDGPU_rcp: { + if (const ConstantFP *C = dyn_cast(II->getArgOperand(0))) { + const APFloat &ArgVal = C->getValueAPF(); + APFloat Val(ArgVal.getSemantics(), 1.0); + APFloat::opStatus Status = Val.divide(ArgVal, + APFloat::rmNearestTiesToEven); + // Only do this if it was exact and therefore not dependent on the + // rounding mode. + if (Status == APFloat::opOK) + return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val)); + } + + break; + } case Intrinsic::stackrestore: { // If the save is right next to the restore, remove the restore. This can // happen when variable allocas are DCE'd. @@ -916,6 +1287,102 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) { return EraseInstFromFunction(CI); break; } + case Intrinsic::assume: { + // Canonicalize assume(a && b) -> assume(a); assume(b); + // Note: New assumption intrinsics created here are registered by + // the InstCombineIRInserter object. + Value *IIOperand = II->getArgOperand(0), *A, *B, + *AssumeIntrinsic = II->getCalledValue(); + if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) { + Builder->CreateCall(AssumeIntrinsic, A, II->getName()); + Builder->CreateCall(AssumeIntrinsic, B, II->getName()); + return EraseInstFromFunction(*II); + } + // assume(!(a || b)) -> assume(!a); assume(!b); + if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) { + Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A), + II->getName()); + Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B), + II->getName()); + return EraseInstFromFunction(*II); + } + + // assume( (load addr) != null ) -> add 'nonnull' metadata to load + // (if assume is valid at the load) + if (ICmpInst* ICmp = dyn_cast(IIOperand)) { + Value *LHS = ICmp->getOperand(0); + Value *RHS = ICmp->getOperand(1); + if (ICmpInst::ICMP_NE == ICmp->getPredicate() && + isa(LHS) && + isa(RHS) && + RHS->getType()->isPointerTy() && + cast(RHS)->isNullValue()) { + LoadInst* LI = cast(LHS); + if (isValidAssumeForContext(II, LI, DT)) { + MDNode *MD = MDNode::get(II->getContext(), None); + LI->setMetadata(LLVMContext::MD_nonnull, MD); + return EraseInstFromFunction(*II); + } + } + // TODO: apply nonnull return attributes to calls and invokes + // TODO: apply range metadata for range check patterns? + } + // If there is a dominating assume with the same condition as this one, + // then this one is redundant, and should be removed. + APInt KnownZero(1, 0), KnownOne(1, 0); + computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II); + if (KnownOne.isAllOnesValue()) + return EraseInstFromFunction(*II); + + break; + } + case Intrinsic::experimental_gc_relocate: { + // Translate facts known about a pointer before relocating into + // facts about the relocate value, while being careful to + // preserve relocation semantics. + GCRelocateOperands Operands(II); + Value *DerivedPtr = Operands.getDerivedPtr(); + auto *GCRelocateType = cast(II->getType()); + + // Remove the relocation if unused, note that this check is required + // to prevent the cases below from looping forever. + if (II->use_empty()) + return EraseInstFromFunction(*II); + + // Undef is undef, even after relocation. + // TODO: provide a hook for this in GCStrategy. This is clearly legal for + // most practical collectors, but there was discussion in the review thread + // about whether it was legal for all possible collectors. + if (isa(DerivedPtr)) { + // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it. + return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType)); + } + + // The relocation of null will be null for most any collector. + // TODO: provide a hook for this in GCStrategy. There might be some weird + // collector this property does not hold for. + if (isa(DerivedPtr)) { + // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it. + return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType)); + } + + // isKnownNonNull -> nonnull attribute + if (isKnownNonNullAt(DerivedPtr, II, DT, TLI)) + II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull); + + // isDereferenceablePointer -> deref attribute + if (isDereferenceablePointer(DerivedPtr, DL)) { + if (Argument *A = dyn_cast(DerivedPtr)) { + uint64_t Bytes = A->getDereferenceableBytes(); + II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes); + } + } + + // TODO: bitcast(relocate(p)) -> relocate(bitcast(p)) + // Canonicalize on the type from the uses to the defs + + // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...) + } } return visitCallSite(II); @@ -930,12 +1397,20 @@ Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { /// isSafeToEliminateVarargsCast - If this cast does not affect the value /// passed through the varargs area, we can eliminate the use of the cast. static bool isSafeToEliminateVarargsCast(const CallSite CS, - const CastInst * const CI, - const DataLayout * const DL, + const DataLayout &DL, + const CastInst *const CI, const int ix) { if (!CI->isLosslessCast()) return false; + // If this is a GC intrinsic, avoid munging types. We need types for + // statepoint reconstruction in SelectionDAG. + // TODO: This is probably something which should be expanded to all + // intrinsics since the entire point of intrinsics is that + // they are understandable by the optimizer. + if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS)) + return false; + // The size of ByVal or InAlloca arguments is derived from the type, so we // can't change to a type with a different size. If the size were // passed explicitly we could avoid this check. @@ -947,7 +1422,7 @@ static bool isSafeToEliminateVarargsCast(const CallSite CS, Type* DstTy = cast(CI->getType())->getElementType(); if (!SrcTy->isSized() || !DstTy->isSized()) return false; - if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy)) + if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy)) return false; return true; } @@ -956,10 +1431,14 @@ static bool isSafeToEliminateVarargsCast(const CallSite CS, // Currently we're only working with the checking functions, memcpy_chk, // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, // strcat_chk and strncat_chk. -Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) { +Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) { if (!CI->getCalledFunction()) return nullptr; - if (Value *With = Simplifier->optimizeCall(CI)) { + auto InstCombineRAUW = [this](Instruction *From, Value *With) { + ReplaceInstUsesWith(*From, With); + }; + LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW); + if (Value *With = Simplifier.optimizeCall(CI)) { ++NumSimplified; return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With); } @@ -1046,11 +1525,29 @@ static IntrinsicInst *FindInitTrampoline(Value *Callee) { // visitCallSite - Improvements for call and invoke instructions. // Instruction *InstCombiner::visitCallSite(CallSite CS) { + if (isAllocLikeFn(CS.getInstruction(), TLI)) return visitAllocSite(*CS.getInstruction()); bool Changed = false; + // Mark any parameters that are known to be non-null with the nonnull + // attribute. This is helpful for inlining calls to functions with null + // checks on their arguments. + unsigned ArgNo = 0; + for (Value *V : CS.args()) { + if (!CS.paramHasAttr(ArgNo+1, Attribute::NonNull) && + isKnownNonNull(V)) { + AttributeSet AS = CS.getAttributes(); + AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1, + Attribute::NonNull); + CS.setAttributes(AS); + Changed = true; + } + ArgNo++; + } + assert(ArgNo == CS.arg_size() && "sanity check"); + // If the callee is a pointer to a function, attempt to move any casts to the // arguments of the call/invoke. Value *Callee = CS.getCalledValue(); @@ -1117,7 +1614,7 @@ Instruction *InstCombiner::visitCallSite(CallSite CS) { for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(), E = CS.arg_end(); I != E; ++I, ++ix) { CastInst *CI = dyn_cast(*I); - if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) { + if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) { *I = CI->getOperand(0); Changed = true; } @@ -1134,7 +1631,7 @@ Instruction *InstCombiner::visitCallSite(CallSite CS) { // this. None of these calls are seen as possibly dead so go ahead and // delete the instruction now. if (CallInst *CI = dyn_cast(CS.getInstruction())) { - Instruction *I = tryOptimizeCall(CI, DL); + Instruction *I = tryOptimizeCall(CI); // If we changed something return the result, etc. Otherwise let // the fallthrough check. if (I) return EraseInstFromFunction(*I); @@ -1151,6 +1648,10 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { dyn_cast(CS.getCalledValue()->stripPointerCasts()); if (!Callee) return false; + // The prototype of thunks are a lie, don't try to directly call such + // functions. + if (Callee->hasFnAttribute("thunk")) + return false; Instruction *Caller = CS.getInstruction(); const AttributeSet &CallerPAL = CS.getAttributes(); @@ -1168,22 +1669,19 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { if (NewRetTy->isStructTy()) return false; // TODO: Handle multiple return values. - if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) { + if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) { if (Callee->isDeclaration()) return false; // Cannot transform this return value. if (!Caller->use_empty() && // void -> non-void is handled specially !NewRetTy->isVoidTy()) - return false; // Cannot transform this return value. + return false; // Cannot transform this return value. } if (!CallerPAL.isEmpty() && !Caller->use_empty()) { AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex); - if (RAttrs. - hasAttributes(AttributeFuncs:: - typeIncompatible(NewRetTy, AttributeSet::ReturnIndex), - AttributeSet::ReturnIndex)) + if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy))) return false; // Attribute not compatible with transformed value. } @@ -1203,17 +1701,28 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { unsigned NumActualArgs = CS.arg_size(); unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); + // Prevent us turning: + // declare void @takes_i32_inalloca(i32* inalloca) + // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0) + // + // into: + // call void @takes_i32_inalloca(i32* null) + // + // Similarly, avoid folding away bitcasts of byval calls. + if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) || + Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal)) + return false; + CallSite::arg_iterator AI = CS.arg_begin(); for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { Type *ParamTy = FT->getParamType(i); Type *ActTy = (*AI)->getType(); - if (!CastInst::isBitCastable(ActTy, ParamTy)) + if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL)) return false; // Cannot transform this parameter value. if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1). - hasAttributes(AttributeFuncs:: - typeIncompatible(ParamTy, i + 1), i + 1)) + overlaps(AttributeFuncs::typeIncompatible(ParamTy))) return false; // Attribute not compatible with transformed value. if (CS.isInAllocaArgument(i)) @@ -1225,12 +1734,12 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1, Attribute::ByVal)) { PointerType *ParamPTy = dyn_cast(ParamTy); - if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL) + if (!ParamPTy || !ParamPTy->getElementType()->isSized()) return false; Type *CurElTy = ActTy->getPointerElementType(); - if (DL->getTypeAllocSize(CurElTy) != - DL->getTypeAllocSize(ParamPTy->getElementType())) + if (DL.getTypeAllocSize(CurElTy) != + DL.getTypeAllocSize(ParamPTy->getElementType())) return false; } } @@ -1286,10 +1795,7 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { // If the return value is not being used, the type may not be compatible // with the existing attributes. Wipe out any problematic attributes. - RAttrs. - removeAttributes(AttributeFuncs:: - typeIncompatible(NewRetTy, AttributeSet::ReturnIndex), - AttributeSet::ReturnIndex); + RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy)); // Add the new return attributes. if (RAttrs.hasAttributes()) @@ -1303,7 +1809,7 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { if ((*AI)->getType() == ParamTy) { Args.push_back(*AI); } else { - Args.push_back(Builder->CreateBitCast(*AI, ParamTy)); + Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy)); } // Add any parameter attributes. @@ -1374,7 +1880,7 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { Value *NV = NC; if (OldRetTy != NV->getType() && !Caller->use_empty()) { if (!NV->getType()->isVoidTy()) { - NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy); + NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy); NC->setDebugLoc(Caller->getDebugLoc()); // If this is an invoke instruction, we should insert it after the first @@ -1394,8 +1900,14 @@ bool InstCombiner::transformConstExprCastCall(CallSite CS) { if (!Caller->use_empty()) ReplaceInstUsesWith(*Caller, NV); - else if (Caller->hasValueHandle()) - ValueHandleBase::ValueIsRAUWd(Caller, NV); + else if (Caller->hasValueHandle()) { + if (OldRetTy == NV->getType()) + ValueHandleBase::ValueIsRAUWd(Caller, NV); + else + // We cannot call ValueIsRAUWd with a different type, and the + // actual tracked value will disappear. + ValueHandleBase::ValueIsDeleted(Caller); + } EraseInstFromFunction(*Caller); return true;