diff options
Diffstat (limited to 'llvm/lib')
86 files changed, 4245 insertions, 785 deletions
diff --git a/llvm/lib/Analysis/DependenceAnalysis.cpp b/llvm/lib/Analysis/DependenceAnalysis.cpp index 853bd66..a572eef 100644 --- a/llvm/lib/Analysis/DependenceAnalysis.cpp +++ b/llvm/lib/Analysis/DependenceAnalysis.cpp @@ -1582,6 +1582,23 @@ static const SCEV *minusSCEVNoSignedOverflow(const SCEV *A, const SCEV *B, return nullptr; } +/// Returns the absolute value of \p A. In the context of dependence analysis, +/// we need an absolute value in a mathematical sense. If \p A is the signed +/// minimum value, we cannot represent it unless extending the original type. +/// Thus if we cannot prove that \p A is not the signed minimum value, returns +/// nullptr. +static const SCEV *absSCEVNoSignedOverflow(const SCEV *A, ScalarEvolution &SE) { + IntegerType *Ty = cast<IntegerType>(A->getType()); + if (!Ty) + return nullptr; + + const SCEV *SMin = + SE.getConstant(APInt::getSignedMinValue(Ty->getBitWidth())); + if (!SE.isKnownPredicate(CmpInst::ICMP_NE, A, SMin)) + return nullptr; + return SE.getAbsExpr(A, /*IsNSW=*/true); +} + /// Returns true iff \p Test is enabled. static bool isDependenceTestEnabled(DependenceTestType Test) { if (EnableDependenceTest == DependenceTestType::All) @@ -1669,21 +1686,25 @@ bool DependenceInfo::strongSIVtest(const SCEV *Coeff, const SCEV *SrcConst, LLVM_DEBUG(dbgs() << ", " << *Delta->getType() << "\n"); // check that |Delta| < iteration count - if (const SCEV *UpperBound = - collectUpperBound(CurSrcLoop, Delta->getType())) { + bool IsDeltaLarge = [&] { + const SCEV *UpperBound = collectUpperBound(CurSrcLoop, Delta->getType()); + if (!UpperBound) + return false; + LLVM_DEBUG(dbgs() << "\t UpperBound = " << *UpperBound); LLVM_DEBUG(dbgs() << ", " << *UpperBound->getType() << "\n"); - const SCEV *AbsDelta = - SE->isKnownNonNegative(Delta) ? Delta : SE->getNegativeSCEV(Delta); - const SCEV *AbsCoeff = - SE->isKnownNonNegative(Coeff) ? Coeff : SE->getNegativeSCEV(Coeff); + const SCEV *AbsDelta = absSCEVNoSignedOverflow(Delta, *SE); + const SCEV *AbsCoeff = absSCEVNoSignedOverflow(Coeff, *SE); + if (!AbsDelta || !AbsCoeff) + return false; const SCEV *Product = SE->getMulExpr(UpperBound, AbsCoeff); - if (isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product)) { - // Distance greater than trip count - no dependence - ++StrongSIVindependence; - ++StrongSIVsuccesses; - return true; - } + return isKnownPredicate(CmpInst::ICMP_SGT, AbsDelta, Product); + }(); + if (IsDeltaLarge) { + // Distance greater than trip count - no dependence + ++StrongSIVindependence; + ++StrongSIVsuccesses; + return true; } // Can we compute distance? @@ -2259,6 +2280,9 @@ bool DependenceInfo::weakZeroSrcSIVtest( const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(DstCoeff); if (!ConstCoeff) return false; + + // Since ConstCoeff is constant, !isKnownNegative means it's non-negative. + // TODO: Bail out if it's a signed minimum value. const SCEV *AbsCoeff = SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(ConstCoeff) : ConstCoeff; @@ -2369,6 +2393,9 @@ bool DependenceInfo::weakZeroDstSIVtest( const SCEVConstant *ConstCoeff = dyn_cast<SCEVConstant>(SrcCoeff); if (!ConstCoeff) return false; + + // Since ConstCoeff is constant, !isKnownNegative means it's non-negative. + // TODO: Bail out if it's a signed minimum value. const SCEV *AbsCoeff = SE->isKnownNegative(ConstCoeff) ? SE->getNegativeSCEV(ConstCoeff) : ConstCoeff; diff --git a/llvm/lib/Analysis/MemoryProfileInfo.cpp b/llvm/lib/Analysis/MemoryProfileInfo.cpp index 92a5b6f..b09f4ed 100644 --- a/llvm/lib/Analysis/MemoryProfileInfo.cpp +++ b/llvm/lib/Analysis/MemoryProfileInfo.cpp @@ -241,9 +241,13 @@ static MDNode *createMIBNode(LLVMContext &Ctx, ArrayRef<uint64_t> MIBCallStack, ColdBytes += TotalSize; // If we have the max cold context size from summary information and have // requested identification of contexts above a percentage of the max, see - // if this context qualifies. - if (MaxColdSize > 0 && MinPercentMaxColdSize < 100 && - TotalSize * 100 >= MaxColdSize * MinPercentMaxColdSize) + // if this context qualifies. We should assume this is large if we rebuilt + // the trie from existing metadata (i.e. to update after inlining), in + // which case we don't have a MaxSize from the profile - we assume any + // context size info in existence on the metadata should be propagated. + if (BuiltFromExistingMetadata || + (MaxColdSize > 0 && MinPercentMaxColdSize < 100 && + TotalSize * 100 >= MaxColdSize * MinPercentMaxColdSize)) LargeColdContext = true; } // Only add the context size info as metadata if we need it in the thin diff --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp index 9e78ec9..8ea1326 100644 --- a/llvm/lib/CodeGen/CodeGenPrepare.cpp +++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp @@ -4030,7 +4030,6 @@ bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const { /// if it is simplified. class SimplificationTracker { DenseMap<Value *, Value *> Storage; - const SimplifyQuery &SQ; // Tracks newly created Phi nodes. The elements are iterated by insertion // order. PhiNodeSet AllPhiNodes; @@ -4038,8 +4037,6 @@ class SimplificationTracker { SmallPtrSet<SelectInst *, 32> AllSelectNodes; public: - SimplificationTracker(const SimplifyQuery &sq) : SQ(sq) {} - Value *Get(Value *V) { do { auto SV = Storage.find(V); @@ -4049,30 +4046,6 @@ public: } while (true); } - Value *Simplify(Value *Val) { - SmallVector<Value *, 32> WorkList; - SmallPtrSet<Value *, 32> Visited; - WorkList.push_back(Val); - while (!WorkList.empty()) { - auto *P = WorkList.pop_back_val(); - if (!Visited.insert(P).second) - continue; - if (auto *PI = dyn_cast<Instruction>(P)) - if (Value *V = simplifyInstruction(cast<Instruction>(PI), SQ)) { - for (auto *U : PI->users()) - WorkList.push_back(cast<Value>(U)); - Put(PI, V); - PI->replaceAllUsesWith(V); - if (auto *PHI = dyn_cast<PHINode>(PI)) - AllPhiNodes.erase(PHI); - if (auto *Select = dyn_cast<SelectInst>(PI)) - AllSelectNodes.erase(Select); - PI->eraseFromParent(); - } - } - return Get(Val); - } - void Put(Value *From, Value *To) { Storage.insert({From, To}); } void ReplacePhi(PHINode *From, PHINode *To) { @@ -4133,8 +4106,7 @@ private: /// Common Type for all different fields in addressing modes. Type *CommonType = nullptr; - /// SimplifyQuery for simplifyInstruction utility. - const SimplifyQuery &SQ; + const DataLayout &DL; /// Original Address. Value *Original; @@ -4143,8 +4115,8 @@ private: Value *CommonValue = nullptr; public: - AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue) - : SQ(_SQ), Original(OriginalValue) {} + AddressingModeCombiner(const DataLayout &DL, Value *OriginalValue) + : DL(DL), Original(OriginalValue) {} ~AddressingModeCombiner() { eraseCommonValueIfDead(); } @@ -4256,7 +4228,7 @@ private: // Keep track of keys where the value is null. We will need to replace it // with constant null when we know the common type. SmallVector<Value *, 2> NullValue; - Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType()); + Type *IntPtrTy = DL.getIntPtrType(AddrModes[0].OriginalValue->getType()); for (auto &AM : AddrModes) { Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy); if (DV) { @@ -4306,7 +4278,7 @@ private: // simplification is possible only if original phi/selects were not // simplified yet. // Using this mapping we can find the current value in AddrToBase. - SimplificationTracker ST(SQ); + SimplificationTracker ST; // First step, DFS to create PHI nodes for all intermediate blocks. // Also fill traverse order for the second step. @@ -4465,7 +4437,6 @@ private: PHI->addIncoming(ST.Get(Map[PV]), B); } } - Map[Current] = ST.Simplify(V); } } @@ -5856,8 +5827,7 @@ bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, // the graph are compatible. bool PhiOrSelectSeen = false; SmallVector<Instruction *, 16> AddrModeInsts; - const SimplifyQuery SQ(*DL, TLInfo); - AddressingModeCombiner AddrModes(SQ, Addr); + AddressingModeCombiner AddrModes(*DL, Addr); TypePromotionTransaction TPT(RemovedInsts); TypePromotionTransaction::ConstRestorationPt LastKnownGood = TPT.getRestorationPoint(); diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp index b425b95..1f10478 100644 --- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp +++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp @@ -391,19 +391,6 @@ void CombinerHelper::applyCombineConcatVectors( MI.eraseFromParent(); } -bool CombinerHelper::matchCombineShuffleToBuildVector(MachineInstr &MI) const { - assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR && - "Invalid instruction"); - auto &Shuffle = cast<GShuffleVector>(MI); - - Register SrcVec1 = Shuffle.getSrc1Reg(); - Register SrcVec2 = Shuffle.getSrc2Reg(); - - LLT SrcVec1Type = MRI.getType(SrcVec1); - LLT SrcVec2Type = MRI.getType(SrcVec2); - return SrcVec1Type.isVector() && SrcVec2Type.isVector(); -} - void CombinerHelper::applyCombineShuffleToBuildVector(MachineInstr &MI) const { auto &Shuffle = cast<GShuffleVector>(MI); @@ -535,11 +522,9 @@ bool CombinerHelper::matchCombineShuffleVector( LLT DstType = MRI.getType(MI.getOperand(0).getReg()); Register Src1 = MI.getOperand(1).getReg(); LLT SrcType = MRI.getType(Src1); - // As bizarre as it may look, shuffle vector can actually produce - // scalar! This is because at the IR level a <1 x ty> shuffle - // vector is perfectly valid. - unsigned DstNumElts = DstType.isVector() ? DstType.getNumElements() : 1; - unsigned SrcNumElts = SrcType.isVector() ? SrcType.getNumElements() : 1; + + unsigned DstNumElts = DstType.getNumElements(); + unsigned SrcNumElts = SrcType.getNumElements(); // If the resulting vector is smaller than the size of the source // vectors being concatenated, we won't be able to replace the @@ -556,7 +541,7 @@ bool CombinerHelper::matchCombineShuffleVector( // // TODO: If the size between the source and destination don't match // we could still emit an extract vector element in that case. - if (DstNumElts < 2 * SrcNumElts && DstNumElts != 1) + if (DstNumElts < 2 * SrcNumElts) return false; // Check that the shuffle mask can be broken evenly between the @@ -619,39 +604,6 @@ void CombinerHelper::applyCombineShuffleVector( MI.eraseFromParent(); } -bool CombinerHelper::matchShuffleToExtract(MachineInstr &MI) const { - assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR && - "Invalid instruction kind"); - - ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); - return Mask.size() == 1; -} - -void CombinerHelper::applyShuffleToExtract(MachineInstr &MI) const { - Register DstReg = MI.getOperand(0).getReg(); - Builder.setInsertPt(*MI.getParent(), MI); - - int I = MI.getOperand(3).getShuffleMask()[0]; - Register Src1 = MI.getOperand(1).getReg(); - LLT Src1Ty = MRI.getType(Src1); - int Src1NumElts = Src1Ty.isVector() ? Src1Ty.getNumElements() : 1; - Register SrcReg; - if (I >= Src1NumElts) { - SrcReg = MI.getOperand(2).getReg(); - I -= Src1NumElts; - } else if (I >= 0) - SrcReg = Src1; - - if (I < 0) - Builder.buildUndef(DstReg); - else if (!MRI.getType(SrcReg).isVector()) - Builder.buildCopy(DstReg, SrcReg); - else - Builder.buildExtractVectorElementConstant(DstReg, SrcReg, I); - - MI.eraseFromParent(); -} - namespace { /// Select a preference between two uses. CurrentUse is the current preference @@ -8369,7 +8321,7 @@ bool CombinerHelper::matchShuffleDisjointMask(MachineInstr &MI, return false; ArrayRef<int> Mask = Shuffle.getMask(); - const unsigned NumSrcElems = Src1Ty.isVector() ? Src1Ty.getNumElements() : 1; + const unsigned NumSrcElems = Src1Ty.getNumElements(); bool TouchesSrc1 = false; bool TouchesSrc2 = false; diff --git a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp index 04d9309..d6f23b6 100644 --- a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp +++ b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp @@ -602,6 +602,8 @@ void GISelValueTracking::computeKnownBitsImpl(Register R, KnownBits &Known, Depth + 1); computeKnownBitsImpl(MI.getOperand(3).getReg(), WidthKnown, DemandedElts, Depth + 1); + OffsetKnown = OffsetKnown.sext(BitWidth); + WidthKnown = WidthKnown.sext(BitWidth); Known = extractBits(BitWidth, SrcOpKnown, OffsetKnown, WidthKnown); // Sign extend the extracted value using shift left and arithmetic shift // right. diff --git a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp index b49040b..1fc90d0 100644 --- a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp +++ b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp @@ -3359,6 +3359,54 @@ bool IRTranslator::translateShuffleVector(const User &U, Mask = SVI->getShuffleMask(); else Mask = cast<ConstantExpr>(U).getShuffleMask(); + + // As GISel does not represent <1 x > vectors as a separate type from scalars, + // we transform shuffle_vector with a scalar output to an + // ExtractVectorElement. If the input type is also scalar it becomes a Copy. + unsigned DstElts = cast<FixedVectorType>(U.getType())->getNumElements(); + unsigned SrcElts = + cast<FixedVectorType>(U.getOperand(0)->getType())->getNumElements(); + if (DstElts == 1) { + unsigned M = Mask[0]; + if (SrcElts == 1) { + if (M == 0 || M == 1) + return translateCopy(U, *U.getOperand(M), MIRBuilder); + MIRBuilder.buildUndef(getOrCreateVReg(U)); + } else { + Register Dst = getOrCreateVReg(U); + if (M < SrcElts) { + MIRBuilder.buildExtractVectorElementConstant( + Dst, getOrCreateVReg(*U.getOperand(0)), M); + } else if (M < SrcElts * 2) { + MIRBuilder.buildExtractVectorElementConstant( + Dst, getOrCreateVReg(*U.getOperand(1)), M - SrcElts); + } else { + MIRBuilder.buildUndef(Dst); + } + } + return true; + } + + // A single element src is transformed to a build_vector. + if (SrcElts == 1) { + SmallVector<Register> Ops; + Register Undef; + for (int M : Mask) { + LLT SrcTy = getLLTForType(*U.getOperand(0)->getType(), *DL); + if (M == 0 || M == 1) { + Ops.push_back(getOrCreateVReg(*U.getOperand(M))); + } else { + if (!Undef.isValid()) { + Undef = MRI->createGenericVirtualRegister(SrcTy); + MIRBuilder.buildUndef(Undef); + } + Ops.push_back(Undef); + } + } + MIRBuilder.buildBuildVector(getOrCreateVReg(U), Ops); + return true; + } + ArrayRef<int> MaskAlloc = MF->allocateShuffleMask(Mask); MIRBuilder .buildInstr(TargetOpcode::G_SHUFFLE_VECTOR, {getOrCreateVReg(U)}, diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp index 38ec83f..178529f 100644 --- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp +++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp @@ -4748,6 +4748,9 @@ LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT LowerHintTy) { case G_FMINIMUMNUM: case G_FMAXIMUMNUM: return lowerFMinNumMaxNum(MI); + case G_FMINIMUM: + case G_FMAXIMUM: + return lowerFMinimumMaximum(MI); case G_MERGE_VALUES: return lowerMergeValues(MI); case G_UNMERGE_VALUES: @@ -5819,6 +5822,8 @@ LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorShuffle( } else if (InputUsed[0] == -1U) { // No input vectors were used! The result is undefined. Output = MIRBuilder.buildUndef(NarrowTy).getReg(0); + } else if (NewElts == 1) { + Output = MIRBuilder.buildCopy(NarrowTy, Inputs[InputUsed[0]]).getReg(0); } else { Register Op0 = Inputs[InputUsed[0]]; // If only one input was used, use an undefined vector for the other. @@ -8775,6 +8780,77 @@ LegalizerHelper::lowerFMinNumMaxNum(MachineInstr &MI) { return Legalized; } +LegalizerHelper::LegalizeResult +LegalizerHelper::lowerFMinimumMaximum(MachineInstr &MI) { + unsigned Opc = MI.getOpcode(); + auto [Dst, Src0, Src1] = MI.getFirst3Regs(); + LLT Ty = MRI.getType(Dst); + LLT CmpTy = Ty.changeElementSize(1); + + bool IsMax = (Opc == TargetOpcode::G_FMAXIMUM); + unsigned OpcIeee = + IsMax ? TargetOpcode::G_FMAXNUM_IEEE : TargetOpcode::G_FMINNUM_IEEE; + unsigned OpcNonIeee = + IsMax ? TargetOpcode::G_FMAXNUM : TargetOpcode::G_FMINNUM; + bool MinMaxMustRespectOrderedZero = false; + Register Res; + + // IEEE variants don't need canonicalization + if (LI.isLegalOrCustom({OpcIeee, Ty})) { + Res = MIRBuilder.buildInstr(OpcIeee, {Ty}, {Src0, Src1}).getReg(0); + MinMaxMustRespectOrderedZero = true; + } else if (LI.isLegalOrCustom({OpcNonIeee, Ty})) { + Res = MIRBuilder.buildInstr(OpcNonIeee, {Ty}, {Src0, Src1}).getReg(0); + } else { + auto Compare = MIRBuilder.buildFCmp( + IsMax ? CmpInst::FCMP_OGT : CmpInst::FCMP_OLT, CmpTy, Src0, Src1); + Res = MIRBuilder.buildSelect(Ty, Compare, Src0, Src1).getReg(0); + } + + // Propagate any NaN of both operands + if (!MI.getFlag(MachineInstr::FmNoNans) && + (!isKnownNeverNaN(Src0, MRI) || isKnownNeverNaN(Src1, MRI))) { + auto IsOrdered = MIRBuilder.buildFCmp(CmpInst::FCMP_ORD, CmpTy, Src0, Src1); + + LLT ElementTy = Ty.isScalar() ? Ty : Ty.getElementType(); + APFloat NaNValue = APFloat::getNaN(getFltSemanticForLLT(ElementTy)); + Register NaN = MIRBuilder.buildFConstant(ElementTy, NaNValue).getReg(0); + if (Ty.isVector()) + NaN = MIRBuilder.buildSplatBuildVector(Ty, NaN).getReg(0); + + Res = MIRBuilder.buildSelect(Ty, IsOrdered, Res, NaN).getReg(0); + } + + // fminimum/fmaximum requires -0.0 less than +0.0 + if (!MinMaxMustRespectOrderedZero && !MI.getFlag(MachineInstr::FmNsz)) { + GISelValueTracking VT(MIRBuilder.getMF()); + KnownFPClass Src0Info = VT.computeKnownFPClass(Src0, fcZero); + KnownFPClass Src1Info = VT.computeKnownFPClass(Src1, fcZero); + + if (!Src0Info.isKnownNeverZero() && !Src1Info.isKnownNeverZero()) { + const unsigned Flags = MI.getFlags(); + Register Zero = MIRBuilder.buildFConstant(Ty, 0.0).getReg(0); + auto IsZero = MIRBuilder.buildFCmp(CmpInst::FCMP_OEQ, CmpTy, Res, Zero); + + unsigned TestClass = IsMax ? fcPosZero : fcNegZero; + + auto LHSTestZero = MIRBuilder.buildIsFPClass(CmpTy, Src0, TestClass); + auto LHSSelect = + MIRBuilder.buildSelect(Ty, LHSTestZero, Src0, Res, Flags); + + auto RHSTestZero = MIRBuilder.buildIsFPClass(CmpTy, Src1, TestClass); + auto RHSSelect = + MIRBuilder.buildSelect(Ty, RHSTestZero, Src1, LHSSelect, Flags); + + Res = MIRBuilder.buildSelect(Ty, IsZero, RHSSelect, Res, Flags).getReg(0); + } + } + + MIRBuilder.buildCopy(Dst, Res); + MI.eraseFromParent(); + return Legalized; +} + LegalizerHelper::LegalizeResult LegalizerHelper::lowerFMad(MachineInstr &MI) { // Expand G_FMAD a, b, c -> G_FADD (G_FMUL a, b), c Register DstReg = MI.getOperand(0).getReg(); @@ -9016,22 +9092,18 @@ LegalizerHelper::lowerShuffleVector(MachineInstr &MI) { continue; } - if (Src0Ty.isScalar()) { - BuildVec.push_back(Idx == 0 ? Src0Reg : Src1Reg); - } else { - int NumElts = Src0Ty.getNumElements(); - Register SrcVec = Idx < NumElts ? Src0Reg : Src1Reg; - int ExtractIdx = Idx < NumElts ? Idx : Idx - NumElts; - auto IdxK = MIRBuilder.buildConstant(IdxTy, ExtractIdx); - auto Extract = MIRBuilder.buildExtractVectorElement(EltTy, SrcVec, IdxK); - BuildVec.push_back(Extract.getReg(0)); - } + assert(!Src0Ty.isScalar() && "Unexpected scalar G_SHUFFLE_VECTOR"); + + int NumElts = Src0Ty.getNumElements(); + Register SrcVec = Idx < NumElts ? Src0Reg : Src1Reg; + int ExtractIdx = Idx < NumElts ? Idx : Idx - NumElts; + auto IdxK = MIRBuilder.buildConstant(IdxTy, ExtractIdx); + auto Extract = MIRBuilder.buildExtractVectorElement(EltTy, SrcVec, IdxK); + BuildVec.push_back(Extract.getReg(0)); } - if (DstTy.isVector()) - MIRBuilder.buildBuildVector(DstReg, BuildVec); - else - MIRBuilder.buildCopy(DstReg, BuildVec[0]); + assert(DstTy.isVector() && "Unexpected scalar G_SHUFFLE_VECTOR"); + MIRBuilder.buildBuildVector(DstReg, BuildVec); MI.eraseFromParent(); return Legalized; } diff --git a/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp b/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp index 27df7e3..4b4df98 100644 --- a/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp +++ b/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp @@ -800,10 +800,11 @@ MachineInstrBuilder MachineIRBuilder::buildShuffleVector(const DstOp &Res, LLT DstTy = Res.getLLTTy(*getMRI()); LLT Src1Ty = Src1.getLLTTy(*getMRI()); LLT Src2Ty = Src2.getLLTTy(*getMRI()); - const LLT DstElemTy = DstTy.isVector() ? DstTy.getElementType() : DstTy; - const LLT ElemTy1 = Src1Ty.isVector() ? Src1Ty.getElementType() : Src1Ty; - const LLT ElemTy2 = Src2Ty.isVector() ? Src2Ty.getElementType() : Src2Ty; + const LLT DstElemTy = DstTy.getScalarType(); + const LLT ElemTy1 = Src1Ty.getScalarType(); + const LLT ElemTy2 = Src2Ty.getScalarType(); assert(DstElemTy == ElemTy1 && DstElemTy == ElemTy2); + assert(Mask.size() > 1 && "Scalar G_SHUFFLE_VECTOR are not supported"); (void)DstElemTy; (void)ElemTy1; (void)ElemTy2; diff --git a/llvm/lib/CodeGen/MIRParser/MIParser.cpp b/llvm/lib/CodeGen/MIRParser/MIParser.cpp index 6a464d9..4795d81 100644 --- a/llvm/lib/CodeGen/MIRParser/MIParser.cpp +++ b/llvm/lib/CodeGen/MIRParser/MIParser.cpp @@ -2788,6 +2788,9 @@ bool MIParser::parseShuffleMaskOperand(MachineOperand &Dest) { if (expectAndConsume(MIToken::rparen)) return error("shufflemask should be terminated by ')'."); + if (ShufMask.size() < 2) + return error("shufflemask should have > 1 element"); + ArrayRef<int> MaskAlloc = MF.allocateShuffleMask(ShufMask); Dest = MachineOperand::CreateShuffleMask(MaskAlloc); return false; diff --git a/llvm/lib/CodeGen/MachineVerifier.cpp b/llvm/lib/CodeGen/MachineVerifier.cpp index 1154855..c0710c4 100644 --- a/llvm/lib/CodeGen/MachineVerifier.cpp +++ b/llvm/lib/CodeGen/MachineVerifier.cpp @@ -1924,13 +1924,23 @@ void MachineVerifier::verifyPreISelGenericInstruction(const MachineInstr *MI) { if (Src0Ty != Src1Ty) report("Source operands must be the same type", MI); - if (Src0Ty.getScalarType() != DstTy.getScalarType()) + if (Src0Ty.getScalarType() != DstTy.getScalarType()) { report("G_SHUFFLE_VECTOR cannot change element type", MI); + break; + } + if (!Src0Ty.isVector()) { + report("G_SHUFFLE_VECTOR must have vector src", MI); + break; + } + if (!DstTy.isVector()) { + report("G_SHUFFLE_VECTOR must have vector dst", MI); + break; + } // Don't check that all operands are vector because scalars are used in // place of 1 element vectors. - int SrcNumElts = Src0Ty.isVector() ? Src0Ty.getNumElements() : 1; - int DstNumElts = DstTy.isVector() ? DstTy.getNumElements() : 1; + int SrcNumElts = Src0Ty.getNumElements(); + int DstNumElts = DstTy.getNumElements(); ArrayRef<int> MaskIdxes = MaskOp.getShuffleMask(); diff --git a/llvm/lib/CodeGen/RegAllocFast.cpp b/llvm/lib/CodeGen/RegAllocFast.cpp index 72b364c..697b779 100644 --- a/llvm/lib/CodeGen/RegAllocFast.cpp +++ b/llvm/lib/CodeGen/RegAllocFast.cpp @@ -211,7 +211,7 @@ private: unsigned getSparseSetIndex() const { return VirtReg.virtRegIndex(); } }; - using LiveRegMap = SparseSet<LiveReg, unsigned, identity_cxx20, uint16_t>; + using LiveRegMap = SparseSet<LiveReg, unsigned, identity, uint16_t>; /// This map contains entries for each virtual register that is currently /// available in a physical register. LiveRegMap LiveVirtRegs; diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index d2ea652..8676060 100644 --- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -19993,8 +19993,12 @@ static SDNode *getPostIndexedLoadStoreOp(SDNode *N, bool &IsLoad, // nor a successor of N. Otherwise, if Op is folded that would // create a cycle. unsigned MaxSteps = SelectionDAG::getHasPredecessorMaxSteps(); - for (SDNode *Op : Ptr->users()) { + for (SDUse &U : Ptr->uses()) { + if (U.getResNo() != Ptr.getResNo()) + continue; + // Check for #1. + SDNode *Op = U.getUser(); if (!shouldCombineToPostInc(N, Ptr, Op, BasePtr, Offset, AM, DAG, TLI)) continue; diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp index bfa566a..dee0909 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp @@ -1162,6 +1162,43 @@ SDValue SelectionDAGBuilder::getMemoryRoot() { return updateRoot(PendingLoads); } +SDValue SelectionDAGBuilder::getFPOperationRoot(fp::ExceptionBehavior EB) { + // If the new exception behavior differs from that of the pending + // ones, chain up them and update the root. + switch (EB) { + case fp::ExceptionBehavior::ebMayTrap: + case fp::ExceptionBehavior::ebIgnore: + // Floating-point exceptions produced by such operations are not intended + // to be observed, so the sequence of these operations does not need to be + // preserved. + // + // They however must not be mixed with the instructions that have strict + // exception behavior. Placing an operation with 'ebIgnore' behavior between + // 'ebStrict' operations could distort the observed exception behavior. + if (!PendingConstrainedFPStrict.empty()) { + assert(PendingConstrainedFP.empty()); + updateRoot(PendingConstrainedFPStrict); + } + break; + case fp::ExceptionBehavior::ebStrict: + // Floating-point exception produced by these operations may be observed, so + // they must be correctly chained. If trapping on FP exceptions is + // disabled, the exceptions can be observed only by functions that read + // exception flags, like 'llvm.get_fpenv' or 'fetestexcept'. It means that + // the order of operations is not significant between barriers. + // + // If trapping is enabled, each operation becomes an implicit observation + // point, so the operations must be sequenced according their original + // source order. + if (!PendingConstrainedFP.empty()) { + assert(PendingConstrainedFPStrict.empty()); + updateRoot(PendingConstrainedFP); + } + // TODO: Add support for trapping-enabled scenarios. + } + return DAG.getRoot(); +} + SDValue SelectionDAGBuilder::getRoot() { // Chain up all pending constrained intrinsics together with all // pending loads, by simply appending them to PendingLoads and @@ -8298,6 +8335,30 @@ void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, } } +void SelectionDAGBuilder::pushFPOpOutChain(SDValue Result, + fp::ExceptionBehavior EB) { + assert(Result.getNode()->getNumValues() == 2); + SDValue OutChain = Result.getValue(1); + assert(OutChain.getValueType() == MVT::Other); + + // Instead of updating the root immediately, push the produced chain to the + // appropriate list, deferring the update until the root is requested. In this + // case, the nodes from the lists are chained using TokenFactor, indicating + // that the operations are independent. + // + // In particular, the root is updated before any call that might access the + // floating-point environment, except for constrained intrinsics. + switch (EB) { + case fp::ExceptionBehavior::ebMayTrap: + case fp::ExceptionBehavior::ebIgnore: + PendingConstrainedFP.push_back(OutChain); + break; + case fp::ExceptionBehavior::ebStrict: + PendingConstrainedFPStrict.push_back(OutChain); + break; + } +} + void SelectionDAGBuilder::visitConstrainedFPIntrinsic( const ConstrainedFPIntrinsic &FPI) { SDLoc sdl = getCurSDLoc(); @@ -8305,42 +8366,16 @@ void SelectionDAGBuilder::visitConstrainedFPIntrinsic( // We do not need to serialize constrained FP intrinsics against // each other or against (nonvolatile) loads, so they can be // chained like loads. - SDValue Chain = DAG.getRoot(); + fp::ExceptionBehavior EB = *FPI.getExceptionBehavior(); + SDValue Chain = getFPOperationRoot(EB); SmallVector<SDValue, 4> Opers; Opers.push_back(Chain); for (unsigned I = 0, E = FPI.getNonMetadataArgCount(); I != E; ++I) Opers.push_back(getValue(FPI.getArgOperand(I))); - auto pushOutChain = [this](SDValue Result, fp::ExceptionBehavior EB) { - assert(Result.getNode()->getNumValues() == 2); - - // Push node to the appropriate list so that future instructions can be - // chained up correctly. - SDValue OutChain = Result.getValue(1); - switch (EB) { - case fp::ExceptionBehavior::ebIgnore: - // The only reason why ebIgnore nodes still need to be chained is that - // they might depend on the current rounding mode, and therefore must - // not be moved across instruction that may change that mode. - [[fallthrough]]; - case fp::ExceptionBehavior::ebMayTrap: - // These must not be moved across calls or instructions that may change - // floating-point exception masks. - PendingConstrainedFP.push_back(OutChain); - break; - case fp::ExceptionBehavior::ebStrict: - // These must not be moved across calls or instructions that may change - // floating-point exception masks or read floating-point exception flags. - // In addition, they cannot be optimized out even if unused. - PendingConstrainedFPStrict.push_back(OutChain); - break; - } - }; - const TargetLowering &TLI = DAG.getTargetLoweringInfo(); EVT VT = TLI.getValueType(DAG.getDataLayout(), FPI.getType()); SDVTList VTs = DAG.getVTList(VT, MVT::Other); - fp::ExceptionBehavior EB = *FPI.getExceptionBehavior(); SDNodeFlags Flags; if (EB == fp::ExceptionBehavior::ebIgnore) @@ -8364,7 +8399,7 @@ void SelectionDAGBuilder::visitConstrainedFPIntrinsic( !TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT)) { Opers.pop_back(); SDValue Mul = DAG.getNode(ISD::STRICT_FMUL, sdl, VTs, Opers, Flags); - pushOutChain(Mul, EB); + pushFPOpOutChain(Mul, EB); Opcode = ISD::STRICT_FADD; Opers.clear(); Opers.push_back(Mul.getValue(1)); @@ -8395,7 +8430,7 @@ void SelectionDAGBuilder::visitConstrainedFPIntrinsic( } SDValue Result = DAG.getNode(Opcode, sdl, VTs, Opers, Flags); - pushOutChain(Result, EB); + pushFPOpOutChain(Result, EB); SDValue FPResult = Result.getValue(0); setValue(&FPI, FPResult); diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h index c7577fa..47e19f7 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h @@ -195,6 +195,11 @@ private: /// Update root to include all chains from the Pending list. SDValue updateRoot(SmallVectorImpl<SDValue> &Pending); + /// Given a node representing a floating-point operation and its specified + /// exception behavior, this either updates the root or stores the node in + /// a list to be added to chains latter. + void pushFPOpOutChain(SDValue Result, fp::ExceptionBehavior EB); + /// A unique monotonically increasing number used to order the SDNodes we /// create. unsigned SDNodeOrder; @@ -300,6 +305,13 @@ public: /// memory node that may need to be ordered after any prior load instructions. SDValue getMemoryRoot(); + /// Return the current virtual root of the Selection DAG, flushing + /// PendingConstrainedFP or PendingConstrainedFPStrict items if the new + /// exception behavior (specified by \p EB) differs from that of the pending + /// instructions. This must be done before emitting constrained FP operation + /// call. + SDValue getFPOperationRoot(fp::ExceptionBehavior EB); + /// Similar to getMemoryRoot, but also flushes PendingConstrainedFP(Strict) /// items. This must be done before emitting any call other any other node /// that may need to be ordered after FP instructions due to other side diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp index 060b1dd..59798b3 100644 --- a/llvm/lib/CodeGen/TargetLoweringBase.cpp +++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp @@ -2097,6 +2097,11 @@ Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { } Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { + // MSVC CRT has a function to validate security cookie. + RTLIB::LibcallImpl SecurityCheckCookieLibcall = + getLibcallImpl(RTLIB::SECURITY_CHECK_COOKIE); + if (SecurityCheckCookieLibcall != RTLIB::Unsupported) + return M.getFunction(getLibcallImplName(SecurityCheckCookieLibcall)); return nullptr; } diff --git a/llvm/lib/ExecutionEngine/Orc/Core.cpp b/llvm/lib/ExecutionEngine/Orc/Core.cpp index 8d413a3..d029ac5 100644 --- a/llvm/lib/ExecutionEngine/Orc/Core.cpp +++ b/llvm/lib/ExecutionEngine/Orc/Core.cpp @@ -2901,13 +2901,23 @@ ExecutionSession::IL_emit(MaterializationResponsibility &MR, for (auto &SN : ER.Ready) IL_collectQueries( - EQ.Updated, SN->defs(), + EQ.Completed, SN->defs(), [](JITDylib::SymbolTableEntry &E) { E.setState(SymbolState::Ready); }, [](AsynchronousSymbolQuery &Q, JITDylib &JD, NonOwningSymbolStringPtr Name, JITDylib::SymbolTableEntry &E) { Q.notifySymbolMetRequiredState(SymbolStringPtr(Name), E.getSymbol()); }); + // std::erase_if is not available in C++17, and llvm::erase_if does not work + // here. + for (auto it = EQ.Completed.begin(), end = EQ.Completed.end(); it != end;) { + if ((*it)->isComplete()) { + ++it; + } else { + it = EQ.Completed.erase(it); + } + } + #ifdef EXPENSIVE_CHECKS verifySessionState("exiting ExecutionSession::IL_emit"); #endif @@ -3043,9 +3053,8 @@ Error ExecutionSession::OL_notifyEmitted( } } - for (auto &UQ : EmitQueries->Updated) - if (UQ->isComplete()) - UQ->handleComplete(*this); + for (auto &UQ : EmitQueries->Completed) + UQ->handleComplete(*this); // If there are any bad dependencies then return an error. if (!BadDeps.empty()) { diff --git a/llvm/lib/IR/AsmWriter.cpp b/llvm/lib/IR/AsmWriter.cpp index 488b078..1096e57 100644 --- a/llvm/lib/IR/AsmWriter.cpp +++ b/llvm/lib/IR/AsmWriter.cpp @@ -4082,10 +4082,10 @@ void AssemblyWriter::printTypeIdentities() { /// printFunction - Print all aspects of a function. void AssemblyWriter::printFunction(const Function *F) { - if (AnnotationWriter) AnnotationWriter->emitFunctionAnnot(F, Out); - if (F->isMaterializable()) Out << "; Materializable\n"; + else if (AnnotationWriter) + AnnotationWriter->emitFunctionAnnot(F, Out); const AttributeList &Attrs = F->getAttributes(); if (Attrs.hasFnAttrs()) { diff --git a/llvm/lib/LTO/LTO.cpp b/llvm/lib/LTO/LTO.cpp index 9d0fa11..4bc2a18 100644 --- a/llvm/lib/LTO/LTO.cpp +++ b/llvm/lib/LTO/LTO.cpp @@ -471,16 +471,14 @@ static void thinLTOInternalizeAndPromoteGUID( ValueInfo VI, function_ref<bool(StringRef, ValueInfo)> isExported, function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> isPrevailing) { - auto ExternallyVisibleCopies = - llvm::count_if(VI.getSummaryList(), - [](const std::unique_ptr<GlobalValueSummary> &Summary) { - return !GlobalValue::isLocalLinkage(Summary->linkage()); - }); - // Before performing index-based internalization and promotion for this GUID, // the local flag should be consistent with the summary list linkage types. VI.verifyLocal(); + const bool SingleExternallyVisibleCopy = + VI.getSummaryList().size() == 1 && + !GlobalValue::isLocalLinkage(VI.getSummaryList().front()->linkage()); + for (auto &S : VI.getSummaryList()) { // First see if we need to promote an internal value because it is not // exported. @@ -543,7 +541,9 @@ static void thinLTOInternalizeAndPromoteGUID( GlobalValue::isExternalWeakLinkage(S->linkage())) continue; - if (isPrevailing(VI.getGUID(), S.get()) && ExternallyVisibleCopies == 1) + // We may have a single summary copy that is externally visible but not + // prevailing if the prevailing copy is in a native object. + if (SingleExternallyVisibleCopy && isPrevailing(VI.getGUID(), S.get())) S->setLinkage(GlobalValue::InternalLinkage); } } @@ -1086,15 +1086,15 @@ LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, GlobalValue::getGlobalIdentifier(Sym.getIRName(), GlobalValue::ExternalLinkage, "")); if (R.Prevailing) - ThinLTO.PrevailingModuleForGUID[GUID] = BM.getModuleIdentifier(); + ThinLTO.setPrevailingModuleForGUID(GUID, BM.getModuleIdentifier()); } } if (Error Err = BM.readSummary(ThinLTO.CombinedIndex, BM.getModuleIdentifier(), [&](GlobalValue::GUID GUID) { - return ThinLTO.PrevailingModuleForGUID[GUID] == - BM.getModuleIdentifier(); + return ThinLTO.isPrevailingModuleForGUID( + GUID, BM.getModuleIdentifier()); })) return Err; LLVM_DEBUG(dbgs() << "Module " << BM.getModuleIdentifier() << "\n"); @@ -1108,8 +1108,8 @@ LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms, GlobalValue::getGlobalIdentifier(Sym.getIRName(), GlobalValue::ExternalLinkage, "")); if (R.Prevailing) { - assert(ThinLTO.PrevailingModuleForGUID[GUID] == - BM.getModuleIdentifier()); + assert( + ThinLTO.isPrevailingModuleForGUID(GUID, BM.getModuleIdentifier())); // For linker redefined symbols (via --wrap or --defsym) we want to // switch the linkage to `weak` to prevent IPOs from happening. @@ -1988,7 +1988,7 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache, LocalWPDTargetsMap); auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) { - return ThinLTO.PrevailingModuleForGUID[GUID] == S->modulePath(); + return ThinLTO.isPrevailingModuleForGUID(GUID, S->modulePath()); }; if (EnableMemProfContextDisambiguation) { MemProfContextDisambiguation ContextDisambiguation; diff --git a/llvm/lib/MC/CMakeLists.txt b/llvm/lib/MC/CMakeLists.txt index 1e1d0a6..70c4577 100644 --- a/llvm/lib/MC/CMakeLists.txt +++ b/llvm/lib/MC/CMakeLists.txt @@ -73,9 +73,10 @@ add_llvm_component_library(LLVMMC ${LLVM_MAIN_INCLUDE_DIR}/llvm/MC LINK_COMPONENTS + BinaryFormat + DebugInfoDWARFLowLevel Support TargetParser - BinaryFormat DEPENDS intrinsics_gen diff --git a/llvm/lib/MC/MCSFrame.cpp b/llvm/lib/MC/MCSFrame.cpp index d6fa54c..e0a90df 100644 --- a/llvm/lib/MC/MCSFrame.cpp +++ b/llvm/lib/MC/MCSFrame.cpp @@ -8,6 +8,8 @@ #include "llvm/MC/MCSFrame.h" #include "llvm/BinaryFormat/SFrame.h" +#include "llvm/DebugInfo/DWARF/LowLevel/DWARFCFIProgram.h" +#include "llvm/DebugInfo/DWARF/LowLevel/DWARFDataExtractorSimple.h" #include "llvm/MC/MCAsmInfo.h" #include "llvm/MC/MCContext.h" #include "llvm/MC/MCObjectFileInfo.h" @@ -211,8 +213,152 @@ class SFrameEmitterImpl { return true; } + // Technically, the escape data could be anything, but it is commonly a dwarf + // CFI program. Even then, it could contain an arbitrarily complicated Dwarf + // expression. Following gnu-gas, look for certain common cases that could + // invalidate an FDE, emit a warning for those sequences, and don't generate + // an FDE in those cases. Allow any that are known safe. It is likely that + // more thorough test cases could refine this code, but it handles the most + // important ones compatibly with gas. + // Returns true if the CFI escape sequence is safe for sframes. + bool isCFIEscapeSafe(SFrameFDE &FDE, const SFrameFRE &FRE, + const MCCFIInstruction &CFI) { + const MCAsmInfo *AI = Streamer.getContext().getAsmInfo(); + DWARFDataExtractorSimple data(CFI.getValues(), AI->isLittleEndian(), + AI->getCodePointerSize()); + + // Normally, both alignment factors are extracted from the enclosing Dwarf + // FDE or CIE. We don't have one here. Alignments are used for scaling + // factors for ops like CFA_def_cfa_offset_sf. But this particular function + // is only interested in registers. + dwarf::CFIProgram P(/*CodeAlignmentFactor=*/1, + /*DataAlignmentFactor=*/1, + Streamer.getContext().getTargetTriple().getArch()); + uint64_t Offset = 0; + if (P.parse(data, &Offset, CFI.getValues().size())) { + // Not a parsable dwarf expression. Assume the worst. + Streamer.getContext().reportWarning( + CFI.getLoc(), + "skipping SFrame FDE; .cfi_escape with unknown effects"); + return false; + } + + // This loop deals with dwarf::CFIProgram::Instructions. Everywhere else + // this file deals with MCCFIInstructions. + for (const dwarf::CFIProgram::Instruction &I : P) { + switch (I.Opcode) { + case dwarf::DW_CFA_nop: + break; + case dwarf::DW_CFA_val_offset: { + // First argument is a register. Anything that touches CFA, FP, or RA is + // a problem, but allow others through. As an even more special case, + // allow SP + 0. + auto Reg = I.getOperandAsUnsigned(P, 0); + // The parser should have failed in this case. + assert(Reg && "DW_CFA_val_offset with no register."); + bool SPOk = true; + if (*Reg == SPReg) { + auto Opnd = I.getOperandAsSigned(P, 1); + if (!Opnd || *Opnd != 0) + SPOk = false; + } + if (!SPOk || *Reg == RAReg || *Reg == FPReg) { + StringRef RN = *Reg == SPReg + ? "SP reg " + : (*Reg == FPReg ? "FP reg " : "RA reg "); + Streamer.getContext().reportWarning( + CFI.getLoc(), + Twine( + "skipping SFrame FDE; .cfi_escape DW_CFA_val_offset with ") + + RN + Twine(*Reg)); + return false; + } + } break; + case dwarf::DW_CFA_expression: { + // First argument is a register. Anything that touches CFA, FP, or RA is + // a problem, but allow others through. + auto Reg = I.getOperandAsUnsigned(P, 0); + if (!Reg) { + Streamer.getContext().reportWarning( + CFI.getLoc(), + "skipping SFrame FDE; .cfi_escape with unknown effects"); + return false; + } + if (*Reg == SPReg || *Reg == RAReg || *Reg == FPReg) { + StringRef RN = *Reg == SPReg + ? "SP reg " + : (*Reg == FPReg ? "FP reg " : "RA reg "); + Streamer.getContext().reportWarning( + CFI.getLoc(), + Twine( + "skipping SFrame FDE; .cfi_escape DW_CFA_expression with ") + + RN + Twine(*Reg)); + return false; + } + } break; + case dwarf::DW_CFA_GNU_args_size: { + auto Size = I.getOperandAsSigned(P, 0); + // Zero size doesn't affect the cfa. + if (Size && *Size == 0) + break; + if (FRE.Info.getBaseRegister() != BaseReg::FP) { + Streamer.getContext().reportWarning( + CFI.getLoc(), + Twine("skipping SFrame FDE; .cfi_escape DW_CFA_GNU_args_size " + "with non frame-pointer CFA")); + return false; + } + } break; + // Cases that gas doesn't specially handle. TODO: Some of these could be + // analyzed and handled instead of just punting. But these are uncommon, + // or should be written as normal cfi directives. Some will need fixes to + // the scaling factor. + case dwarf::DW_CFA_advance_loc: + case dwarf::DW_CFA_offset: + case dwarf::DW_CFA_restore: + case dwarf::DW_CFA_set_loc: + case dwarf::DW_CFA_advance_loc1: + case dwarf::DW_CFA_advance_loc2: + case dwarf::DW_CFA_advance_loc4: + case dwarf::DW_CFA_offset_extended: + case dwarf::DW_CFA_restore_extended: + case dwarf::DW_CFA_undefined: + case dwarf::DW_CFA_same_value: + case dwarf::DW_CFA_register: + case dwarf::DW_CFA_remember_state: + case dwarf::DW_CFA_restore_state: + case dwarf::DW_CFA_def_cfa: + case dwarf::DW_CFA_def_cfa_register: + case dwarf::DW_CFA_def_cfa_offset: + case dwarf::DW_CFA_def_cfa_expression: + case dwarf::DW_CFA_offset_extended_sf: + case dwarf::DW_CFA_def_cfa_sf: + case dwarf::DW_CFA_def_cfa_offset_sf: + case dwarf::DW_CFA_val_offset_sf: + case dwarf::DW_CFA_val_expression: + case dwarf::DW_CFA_MIPS_advance_loc8: + case dwarf::DW_CFA_AARCH64_negate_ra_state_with_pc: + case dwarf::DW_CFA_AARCH64_negate_ra_state: + case dwarf::DW_CFA_LLVM_def_aspace_cfa: + case dwarf::DW_CFA_LLVM_def_aspace_cfa_sf: + Streamer.getContext().reportWarning( + CFI.getLoc(), "skipping SFrame FDE; .cfi_escape " + "CFA expression with unknown side effects"); + return false; + default: + // Dwarf expression was only partially valid, and user could have + // written anything. + Streamer.getContext().reportWarning( + CFI.getLoc(), + "skipping SFrame FDE; .cfi_escape with unknown effects"); + return false; + } + } + return true; + } + // Add the effects of CFI to the current FDE, creating a new FRE when - // necessary. + // necessary. Return true if the CFI is representable in the sframe format. bool handleCFI(SFrameFDE &FDE, SFrameFRE &FRE, const MCCFIInstruction &CFI) { switch (CFI.getOperation()) { case MCCFIInstruction::OpDefCfaRegister: @@ -265,10 +411,11 @@ class SFrameEmitterImpl { FRE = FDE.SaveState.pop_back_val(); return true; case MCCFIInstruction::OpEscape: - // TODO: Implement. Will use FDE. - return true; + // This is a string of bytes that contains an arbitrary dwarf-expression + // that may or may not affect unwind info. + return isCFIEscapeSafe(FDE, FRE, CFI); default: - // Instructions that don't affect the CFA, RA, and SP can be safely + // Instructions that don't affect the CFA, RA, and FP can be safely // ignored. return true; } diff --git a/llvm/lib/Support/SpecialCaseList.cpp b/llvm/lib/Support/SpecialCaseList.cpp index f74e52a..c27f627 100644 --- a/llvm/lib/Support/SpecialCaseList.cpp +++ b/llvm/lib/Support/SpecialCaseList.cpp @@ -89,14 +89,32 @@ void SpecialCaseList::GlobMatcher::preprocess(bool BySize) { return A.Name.size() < B.Name.size(); }); } + + for (const auto &G : reverse(Globs)) { + StringRef Prefix = G.Pattern.prefix(); + + auto &V = PrefixToGlob.emplace(Prefix).first->second; + V.emplace_back(&G); + } } void SpecialCaseList::GlobMatcher::match( StringRef Query, llvm::function_ref<void(StringRef Rule, unsigned LineNo)> Cb) const { - for (const auto &G : reverse(Globs)) - if (G.Pattern.match(Query)) - return Cb(G.Name, G.LineNo); + if (!PrefixToGlob.empty()) { + for (const auto &[_, V] : PrefixToGlob.find_prefixes(Query)) { + for (const auto *G : V) { + if (G->Pattern.match(Query)) { + Cb(G->Name, G->LineNo); + // As soon as we find a match in the vector, we can break for this + // vector, since the globs are already sorted by priority within the + // prefix group. However, we continue searching other prefix groups in + // the map, as they may contain a better match overall. + break; + } + } + } + } } SpecialCaseList::Matcher::Matcher(bool UseGlobs, bool RemoveDotSlash) diff --git a/llvm/lib/Support/Timer.cpp b/llvm/lib/Support/Timer.cpp index 67483ba..9d45096 100644 --- a/llvm/lib/Support/Timer.cpp +++ b/llvm/lib/Support/Timer.cpp @@ -240,7 +240,8 @@ private: getGroupEntry(StringRef GroupName, StringRef GroupDescription) { std::pair<TimerGroup *, Name2TimerMap> &GroupEntry = Map[GroupName]; if (!GroupEntry.first) - GroupEntry.first = new TimerGroup(GroupName, GroupDescription); + GroupEntry.first = + new TimerGroup(GroupName, GroupDescription, /*PrintOnExit=*/true); return GroupEntry; } @@ -270,9 +271,10 @@ TimerGroup &NamedRegionTimer::getNamedTimerGroup(StringRef GroupName, static TimerGroup *TimerGroupList = nullptr; TimerGroup::TimerGroup(StringRef Name, StringRef Description, - sys::SmartMutex<true> &lock) + sys::SmartMutex<true> &lock, bool PrintOnExit) : Name(Name.begin(), Name.end()), - Description(Description.begin(), Description.end()) { + Description(Description.begin(), Description.end()), + PrintOnExit(PrintOnExit) { // Add the group to TimerGroupList. sys::SmartScopedLock<true> L(lock); if (TimerGroupList) @@ -282,12 +284,12 @@ TimerGroup::TimerGroup(StringRef Name, StringRef Description, TimerGroupList = this; } -TimerGroup::TimerGroup(StringRef Name, StringRef Description) - : TimerGroup(Name, Description, timerLock()) {} +TimerGroup::TimerGroup(StringRef Name, StringRef Description, bool PrintOnExit) + : TimerGroup(Name, Description, timerLock(), PrintOnExit) {} TimerGroup::TimerGroup(StringRef Name, StringRef Description, - const StringMap<TimeRecord> &Records) - : TimerGroup(Name, Description) { + const StringMap<TimeRecord> &Records, bool PrintOnExit) + : TimerGroup(Name, Description, PrintOnExit) { TimersToPrint.reserve(Records.size()); for (const auto &P : Records) TimersToPrint.emplace_back(P.getValue(), std::string(P.getKey()), @@ -301,7 +303,7 @@ TimerGroup::~TimerGroup() { while (FirstTimer) removeTimer(*FirstTimer); - if (!TimersToPrint.empty()) { + if (!TimersToPrint.empty() && PrintOnExit) { std::unique_ptr<raw_ostream> OutStream = CreateInfoOutputFile(); PrintQueuedTimers(*OutStream); } @@ -530,7 +532,7 @@ public: sys::SmartMutex<true> TimerLock; TimerGroup DefaultTimerGroup{"misc", "Miscellaneous Ungrouped Timers", - TimerLock}; + TimerLock, /*PrintOnExit=*/true}; SignpostEmitter Signposts; // Order of these members and initialization below is important. For example diff --git a/llvm/lib/Target/AArch64/AArch64Combine.td b/llvm/lib/Target/AArch64/AArch64Combine.td index ecaeff7..b3ec65c 100644 --- a/llvm/lib/Target/AArch64/AArch64Combine.td +++ b/llvm/lib/Target/AArch64/AArch64Combine.td @@ -71,7 +71,6 @@ def AArch64PreLegalizerCombiner: GICombiner< "AArch64PreLegalizerCombinerImpl", [all_combines, icmp_redundant_trunc, fold_global_offset, - shuffle_to_extract, ext_addv_to_udot_addv, ext_uaddv_to_uaddlv, push_sub_through_zext, diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp index a81de5c..d16b116 100644 --- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp +++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp @@ -9002,12 +9002,12 @@ static void analyzeCallOperands(const AArch64TargetLowering &TLI, } static SMECallAttrs -getSMECallAttrs(const Function &Caller, const AArch64TargetLowering &TLI, +getSMECallAttrs(const Function &Caller, const RTLIB::RuntimeLibcallsInfo &RTLCI, const TargetLowering::CallLoweringInfo &CLI) { if (CLI.CB) - return SMECallAttrs(*CLI.CB, &TLI); + return SMECallAttrs(*CLI.CB, &RTLCI); if (auto *ES = dyn_cast<ExternalSymbolSDNode>(CLI.Callee)) - return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(ES->getSymbol(), TLI)); + return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(ES->getSymbol(), RTLCI)); return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(SMEAttrs::Normal)); } @@ -9029,7 +9029,8 @@ bool AArch64TargetLowering::isEligibleForTailCallOptimization( // SME Streaming functions are not eligible for TCO as they may require // the streaming mode or ZA to be restored after returning from the call. - SMECallAttrs CallAttrs = getSMECallAttrs(CallerF, *this, CLI); + SMECallAttrs CallAttrs = + getSMECallAttrs(CallerF, getRuntimeLibcallsInfo(), CLI); if (CallAttrs.requiresSMChange() || CallAttrs.requiresLazySave() || CallAttrs.requiresPreservingAllZAState() || CallAttrs.caller().hasStreamingBody()) @@ -9454,7 +9455,8 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI, } // Determine whether we need any streaming mode changes. - SMECallAttrs CallAttrs = getSMECallAttrs(MF.getFunction(), *this, CLI); + SMECallAttrs CallAttrs = + getSMECallAttrs(MF.getFunction(), getRuntimeLibcallsInfo(), CLI); std::optional<unsigned> ZAMarkerNode; bool UseNewSMEABILowering = getTM().useNewSMEABILowering(); @@ -19476,6 +19478,61 @@ static SDValue performMulVectorExtendCombine(SDNode *Mul, SelectionDAG &DAG) { Op1 ? Op1 : Mul->getOperand(1)); } +// Multiplying an RDSVL value by a constant can sometimes be done cheaper by +// folding a power-of-two factor of the constant into the RDSVL immediate and +// compensating with an extra shift. +// +// We rewrite: +// (mul (srl (rdsvl 1), w), x) +// to one of: +// (shl (rdsvl y), z) if z > 0 +// (srl (rdsvl y), abs(z)) if z < 0 +// where integers y, z satisfy x = y * 2^(w + z) and y ∈ [-32, 31]. +static SDValue performMulRdsvlCombine(SDNode *Mul, SelectionDAG &DAG) { + SDLoc DL(Mul); + EVT VT = Mul->getValueType(0); + SDValue MulOp0 = Mul->getOperand(0); + int ConstMultiplier = + cast<ConstantSDNode>(Mul->getOperand(1))->getSExtValue(); + if ((MulOp0->getOpcode() != ISD::SRL) || + (MulOp0->getOperand(0).getOpcode() != AArch64ISD::RDSVL)) + return SDValue(); + + unsigned AbsConstValue = abs(ConstMultiplier); + unsigned OperandShift = + cast<ConstantSDNode>(MulOp0->getOperand(1))->getZExtValue(); + + // z ≤ ctz(|x|) - w (largest extra shift we can take while keeping y + // integral) + int UpperBound = llvm::countr_zero(AbsConstValue) - OperandShift; + + // To keep y in range, with B = 31 for x > 0 and B = 32 for x < 0, we need: + // 2^(w + z) ≥ ceil(x / B) ⇒ z ≥ ceil_log2(ceil(x / B)) - w (LowerBound). + unsigned B = ConstMultiplier < 0 ? 32 : 31; + unsigned CeilAxOverB = (AbsConstValue + (B - 1)) / B; // ceil(|x|/B) + int LowerBound = llvm::Log2_32_Ceil(CeilAxOverB) - OperandShift; + + // No valid solution found. + if (LowerBound > UpperBound) + return SDValue(); + + // Any value of z in [LowerBound, UpperBound] is valid. Prefer no extra + // shift if possible. + int Shift = std::min(std::max(/*prefer*/ 0, LowerBound), UpperBound); + + // y = x / 2^(w + z) + int32_t RdsvlMul = (AbsConstValue >> (OperandShift + Shift)) * + (ConstMultiplier < 0 ? -1 : 1); + auto Rdsvl = DAG.getNode(AArch64ISD::RDSVL, DL, MVT::i64, + DAG.getSignedConstant(RdsvlMul, DL, MVT::i32)); + + if (Shift == 0) + return Rdsvl; + return DAG.getNode(Shift < 0 ? ISD::SRL : ISD::SHL, DL, VT, Rdsvl, + DAG.getConstant(abs(Shift), DL, MVT::i32), + SDNodeFlags::Exact); +} + // Combine v4i32 Mul(And(Srl(X, 15), 0x10001), 0xffff) -> v8i16 CMLTz // Same for other types with equivalent constants. static SDValue performMulVectorCmpZeroCombine(SDNode *N, SelectionDAG &DAG) { @@ -19604,6 +19661,9 @@ static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG, if (!isa<ConstantSDNode>(N1)) return SDValue(); + if (SDValue Ext = performMulRdsvlCombine(N, DAG)) + return Ext; + ConstantSDNode *C = cast<ConstantSDNode>(N1); const APInt &ConstValue = C->getAPIntValue(); @@ -26665,11 +26725,34 @@ static SDValue performDUPCombine(SDNode *N, } if (N->getOpcode() == AArch64ISD::DUP) { + SDValue Op = N->getOperand(0); + + // Optimize DUP(extload/zextload i8/i16/i32) to avoid GPR->FPR transfer. + // For example: + // v4i32 = DUP (i32 (zextloadi8 addr)) + // => + // v4i32 = SCALAR_TO_VECTOR (i32 (zextloadi8 addr)) ; Matches to ldr b0 + // v4i32 = DUPLANE32 (v4i32), 0 + if (auto *LD = dyn_cast<LoadSDNode>(Op)) { + ISD::LoadExtType ExtType = LD->getExtensionType(); + EVT MemVT = LD->getMemoryVT(); + EVT ElemVT = VT.getVectorElementType(); + if ((ExtType == ISD::EXTLOAD || ExtType == ISD::ZEXTLOAD) && + (MemVT == MVT::i8 || MemVT == MVT::i16 || MemVT == MVT::i32) && + ElemVT != MemVT && LD->hasOneUse()) { + EVT Vec128VT = EVT::getVectorVT(*DCI.DAG.getContext(), ElemVT, + 128 / ElemVT.getSizeInBits()); + SDValue ScalarToVec = + DCI.DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, Vec128VT, Op); + return DCI.DAG.getNode(getDUPLANEOp(ElemVT), DL, VT, ScalarToVec, + DCI.DAG.getConstant(0, DL, MVT::i64)); + } + } + // If the instruction is known to produce a scalar in SIMD registers, we can // duplicate it across the vector lanes using DUPLANE instead of moving it // to a GPR first. For example, this allows us to handle: // v4i32 = DUP (i32 (FCMGT (f32, f32))) - SDValue Op = N->getOperand(0); // FIXME: Ideally, we should be able to handle all instructions that // produce a scalar value in FPRs. if (Op.getOpcode() == AArch64ISD::FCMEQ || @@ -29430,15 +29513,6 @@ void AArch64TargetLowering::insertSSPDeclarations(Module &M) const { TargetLowering::insertSSPDeclarations(M); } -Function *AArch64TargetLowering::getSSPStackGuardCheck(const Module &M) const { - // MSVC CRT has a function to validate security cookie. - RTLIB::LibcallImpl SecurityCheckCookieLibcall = - getLibcallImpl(RTLIB::SECURITY_CHECK_COOKIE); - if (SecurityCheckCookieLibcall != RTLIB::Unsupported) - return M.getFunction(getLibcallImplName(SecurityCheckCookieLibcall)); - return TargetLowering::getSSPStackGuardCheck(M); -} - Value * AArch64TargetLowering::getSafeStackPointerLocation(IRBuilderBase &IRB) const { // Android provides a fixed TLS slot for the SafeStack pointer. See the @@ -29447,11 +29521,6 @@ AArch64TargetLowering::getSafeStackPointerLocation(IRBuilderBase &IRB) const { if (Subtarget->isTargetAndroid()) return UseTlsOffset(IRB, 0x48); - // Fuchsia is similar. - // <zircon/tls.h> defines ZX_TLS_UNSAFE_SP_OFFSET with this value. - if (Subtarget->isTargetFuchsia()) - return UseTlsOffset(IRB, -0x8); - return TargetLowering::getSafeStackPointerLocation(IRB); } @@ -29769,7 +29838,7 @@ bool AArch64TargetLowering::fallBackToDAGISel(const Instruction &Inst) const { // Checks to allow the use of SME instructions if (auto *Base = dyn_cast<CallBase>(&Inst)) { - auto CallAttrs = SMECallAttrs(*Base, this); + auto CallAttrs = SMECallAttrs(*Base, &getRuntimeLibcallsInfo()); if (CallAttrs.requiresSMChange() || CallAttrs.requiresLazySave() || CallAttrs.requiresPreservingZT0() || CallAttrs.requiresPreservingAllZAState()) diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.h b/llvm/lib/Target/AArch64/AArch64ISelLowering.h index 9495c9f..2cb8ed2 100644 --- a/llvm/lib/Target/AArch64/AArch64ISelLowering.h +++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.h @@ -366,7 +366,6 @@ public: Value *getIRStackGuard(IRBuilderBase &IRB) const override; void insertSSPDeclarations(Module &M) const override; - Function *getSSPStackGuardCheck(const Module &M) const override; /// If the target has a standard location for the unsafe stack pointer, /// returns the address of that location. Otherwise, returns nullptr. diff --git a/llvm/lib/Target/AArch64/AArch64InstrFormats.td b/llvm/lib/Target/AArch64/AArch64InstrFormats.td index eab1627..58a53af 100644 --- a/llvm/lib/Target/AArch64/AArch64InstrFormats.td +++ b/llvm/lib/Target/AArch64/AArch64InstrFormats.td @@ -5298,7 +5298,7 @@ multiclass FPToIntegerUnscaled<bits<2> rmode, bits<3> opcode, string asm, } multiclass FPToIntegerSIMDScalar<bits<2> rmode, bits<3> opcode, string asm, - SDPatternOperator OpN = null_frag> { + SDPatternOperator OpN> { // double-precision to 32-bit SIMD/FPR def SDr : BaseFPToIntegerUnscaled<0b01, rmode, opcode, FPR64, FPR32, asm, [(set FPR32:$Rd, (i32 (OpN (f64 FPR64:$Rn))))]> { diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp b/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp index d5117da..457e540 100644 --- a/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp +++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp @@ -5151,7 +5151,15 @@ void AArch64InstrInfo::copyPhysReg(MachineBasicBlock &MBB, // GPR32 zeroing if (AArch64::GPR32spRegClass.contains(DestReg) && SrcReg == AArch64::WZR) { - if (Subtarget.hasZeroCycleZeroingGPR32()) { + if (Subtarget.hasZeroCycleZeroingGPR64() && + !Subtarget.hasZeroCycleZeroingGPR32()) { + MCRegister DestRegX = RI.getMatchingSuperReg(DestReg, AArch64::sub_32, + &AArch64::GPR64spRegClass); + assert(DestRegX.isValid() && "Destination super-reg not valid"); + BuildMI(MBB, I, DL, get(AArch64::MOVZXi), DestRegX) + .addImm(0) + .addImm(AArch64_AM::getShifterImm(AArch64_AM::LSL, 0)); + } else if (Subtarget.hasZeroCycleZeroingGPR32()) { BuildMI(MBB, I, DL, get(AArch64::MOVZWi), DestReg) .addImm(0) .addImm(AArch64_AM::getShifterImm(AArch64_AM::LSL, 0)); diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/llvm/lib/Target/AArch64/AArch64InstrInfo.td index b74ca79..b9e299e 100644 --- a/llvm/lib/Target/AArch64/AArch64InstrInfo.td +++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.td @@ -4022,22 +4022,6 @@ defm LDRSW : LoadUI<0b10, 0, 0b10, GPR64, uimm12s4, "ldrsw", def : Pat<(i64 (zextloadi32 (am_indexed32 GPR64sp:$Rn, uimm12s4:$offset))), (SUBREG_TO_REG (i64 0), (LDRWui GPR64sp:$Rn, uimm12s4:$offset), sub_32)>; -// load zero-extended i32, bitcast to f64 -def : Pat<(f64 (bitconvert (i64 (zextloadi32 (am_indexed32 GPR64sp:$Rn, uimm12s4:$offset))))), - (SUBREG_TO_REG (i64 0), (LDRSui GPR64sp:$Rn, uimm12s4:$offset), ssub)>; -// load zero-extended i16, bitcast to f64 -def : Pat<(f64 (bitconvert (i64 (zextloadi16 (am_indexed16 GPR64sp:$Rn, uimm12s2:$offset))))), - (SUBREG_TO_REG (i64 0), (LDRHui GPR64sp:$Rn, uimm12s2:$offset), hsub)>; -// load zero-extended i8, bitcast to f64 -def : Pat<(f64 (bitconvert (i64 (zextloadi8 (am_indexed8 GPR64sp:$Rn, uimm12s1:$offset))))), - (SUBREG_TO_REG (i64 0), (LDRBui GPR64sp:$Rn, uimm12s1:$offset), bsub)>; -// load zero-extended i16, bitcast to f32 -def : Pat<(f32 (bitconvert (i32 (zextloadi16 (am_indexed16 GPR64sp:$Rn, uimm12s2:$offset))))), - (SUBREG_TO_REG (i32 0), (LDRHui GPR64sp:$Rn, uimm12s2:$offset), hsub)>; -// load zero-extended i8, bitcast to f32 -def : Pat<(f32 (bitconvert (i32 (zextloadi8 (am_indexed8 GPR64sp:$Rn, uimm12s1:$offset))))), - (SUBREG_TO_REG (i32 0), (LDRBui GPR64sp:$Rn, uimm12s1:$offset), bsub)>; - // Pre-fetch. def PRFMui : PrefetchUI<0b11, 0, 0b10, "prfm", [(AArch64Prefetch timm:$Rt, @@ -4389,6 +4373,64 @@ def : Pat <(v1i64 (scalar_to_vector (i64 (load (ro64.Xpat GPR64sp:$Rn, GPR64:$Rm, ro64.Xext:$extend))))), (LDRDroX GPR64sp:$Rn, GPR64:$Rm, ro64.Xext:$extend)>; +// Patterns for bitconvert or scalar_to_vector of load operations. +// Enables direct SIMD register loads for small integer types (i8/i16) that are +// naturally zero-extended to i32/i64. +multiclass ExtLoad8_16AllModes<ValueType OutTy, ValueType InnerTy, + SDPatternOperator OuterOp, + PatFrags LoadOp8, PatFrags LoadOp16> { + // 8-bit loads. + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp8 (am_indexed8 GPR64sp:$Rn, uimm12s1:$offset))))), + (SUBREG_TO_REG (i64 0), (LDRBui GPR64sp:$Rn, uimm12s1:$offset), bsub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp8 (am_unscaled8 GPR64sp:$Rn, simm9:$offset))))), + (SUBREG_TO_REG (i64 0), (LDURBi GPR64sp:$Rn, simm9:$offset), bsub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp8 (ro8.Wpat GPR64sp:$Rn, GPR32:$Rm, ro8.Wext:$extend))))), + (SUBREG_TO_REG (i64 0), (LDRBroW GPR64sp:$Rn, GPR32:$Rm, ro8.Wext:$extend), bsub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp8 (ro8.Xpat GPR64sp:$Rn, GPR64:$Rm, ro8.Xext:$extend))))), + (SUBREG_TO_REG (i64 0), (LDRBroX GPR64sp:$Rn, GPR64:$Rm, ro8.Xext:$extend), bsub)>; + + // 16-bit loads. + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp16 (am_indexed16 GPR64sp:$Rn, uimm12s2:$offset))))), + (SUBREG_TO_REG (i64 0), (LDRHui GPR64sp:$Rn, uimm12s2:$offset), hsub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp16 (am_unscaled16 GPR64sp:$Rn, simm9:$offset))))), + (SUBREG_TO_REG (i64 0), (LDURHi GPR64sp:$Rn, simm9:$offset), hsub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp16 (ro16.Wpat GPR64sp:$Rn, GPR32:$Rm, ro16.Wext:$extend))))), + (SUBREG_TO_REG (i64 0), (LDRHroW GPR64sp:$Rn, GPR32:$Rm, ro16.Wext:$extend), hsub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp16 (ro16.Xpat GPR64sp:$Rn, GPR64:$Rm, ro16.Xext:$extend))))), + (SUBREG_TO_REG (i64 0), (LDRHroX GPR64sp:$Rn, GPR64:$Rm, ro16.Xext:$extend), hsub)>; +} + +// Extended multiclass that includes 32-bit loads in addition to 8-bit and 16-bit. +multiclass ExtLoad8_16_32AllModes<ValueType OutTy, ValueType InnerTy, + SDPatternOperator OuterOp, + PatFrags LoadOp8, PatFrags LoadOp16, PatFrags LoadOp32> { + defm : ExtLoad8_16AllModes<OutTy, InnerTy, OuterOp, LoadOp8, LoadOp16>; + + // 32-bit loads. + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp32 (am_indexed32 GPR64sp:$Rn, uimm12s4:$offset))))), + (SUBREG_TO_REG (i64 0), (LDRSui GPR64sp:$Rn, uimm12s4:$offset), ssub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp32 (am_unscaled32 GPR64sp:$Rn, simm9:$offset))))), + (SUBREG_TO_REG (i64 0), (LDURSi GPR64sp:$Rn, simm9:$offset), ssub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp32 (ro32.Wpat GPR64sp:$Rn, GPR32:$Rm, ro32.Wext:$extend))))), + (SUBREG_TO_REG (i64 0), (LDRSroW GPR64sp:$Rn, GPR32:$Rm, ro32.Wext:$extend), ssub)>; + def : Pat<(OutTy (OuterOp (InnerTy (LoadOp32 (ro32.Xpat GPR64sp:$Rn, GPR64:$Rm, ro32.Xext:$extend))))), + (SUBREG_TO_REG (i64 0), (LDRSroX GPR64sp:$Rn, GPR64:$Rm, ro32.Xext:$extend), ssub)>; +} + +// Instantiate bitconvert patterns for floating-point types. +defm : ExtLoad8_16AllModes<f32, i32, bitconvert, zextloadi8, zextloadi16>; +defm : ExtLoad8_16_32AllModes<f64, i64, bitconvert, zextloadi8, zextloadi16, zextloadi32>; + +// Instantiate scalar_to_vector patterns for all vector types. +defm : ExtLoad8_16AllModes<v16i8, i32, scalar_to_vector, zextloadi8, zextloadi16>; +defm : ExtLoad8_16AllModes<v16i8, i32, scalar_to_vector, extloadi8, extloadi16>; +defm : ExtLoad8_16AllModes<v8i16, i32, scalar_to_vector, zextloadi8, zextloadi16>; +defm : ExtLoad8_16AllModes<v8i16, i32, scalar_to_vector, extloadi8, extloadi16>; +defm : ExtLoad8_16AllModes<v4i32, i32, scalar_to_vector, zextloadi8, zextloadi16>; +defm : ExtLoad8_16AllModes<v4i32, i32, scalar_to_vector, extloadi8, extloadi16>; +defm : ExtLoad8_16_32AllModes<v2i64, i64, scalar_to_vector, zextloadi8, zextloadi16, zextloadi32>; +defm : ExtLoad8_16_32AllModes<v2i64, i64, scalar_to_vector, extloadi8, extloadi16, extloadi32>; + // Pre-fetch. defm PRFUM : PrefetchUnscaled<0b11, 0, 0b10, "prfum", [(AArch64Prefetch timm:$Rt, @@ -5253,113 +5295,10 @@ let Predicates = [HasNEON, HasFPRCVT] in{ defm FCVTNU : FPToIntegerSIMDScalar<0b01, 0b011, "fcvtnu", int_aarch64_neon_fcvtnu>; defm FCVTPS : FPToIntegerSIMDScalar<0b10, 0b010, "fcvtps", int_aarch64_neon_fcvtps>; defm FCVTPU : FPToIntegerSIMDScalar<0b10, 0b011, "fcvtpu", int_aarch64_neon_fcvtpu>; - defm FCVTZS : FPToIntegerSIMDScalar<0b10, 0b110, "fcvtzs">; - defm FCVTZU : FPToIntegerSIMDScalar<0b10, 0b111, "fcvtzu">; -} - - -// AArch64's FCVT instructions saturate when out of range. -multiclass FPToIntegerSatPats<SDNode to_int_sat, SDNode to_int_sat_gi, string INST> { - let Predicates = [HasFullFP16] in { - def : Pat<(i32 (to_int_sat f16:$Rn, i32)), - (!cast<Instruction>(INST # UWHr) f16:$Rn)>; - def : Pat<(i64 (to_int_sat f16:$Rn, i64)), - (!cast<Instruction>(INST # UXHr) f16:$Rn)>; - } - def : Pat<(i32 (to_int_sat f32:$Rn, i32)), - (!cast<Instruction>(INST # UWSr) f32:$Rn)>; - def : Pat<(i64 (to_int_sat f32:$Rn, i64)), - (!cast<Instruction>(INST # UXSr) f32:$Rn)>; - def : Pat<(i32 (to_int_sat f64:$Rn, i32)), - (!cast<Instruction>(INST # UWDr) f64:$Rn)>; - def : Pat<(i64 (to_int_sat f64:$Rn, i64)), - (!cast<Instruction>(INST # UXDr) f64:$Rn)>; - - let Predicates = [HasFullFP16] in { - def : Pat<(i32 (to_int_sat_gi f16:$Rn)), - (!cast<Instruction>(INST # UWHr) f16:$Rn)>; - def : Pat<(i64 (to_int_sat_gi f16:$Rn)), - (!cast<Instruction>(INST # UXHr) f16:$Rn)>; - } - def : Pat<(i32 (to_int_sat_gi f32:$Rn)), - (!cast<Instruction>(INST # UWSr) f32:$Rn)>; - def : Pat<(i64 (to_int_sat_gi f32:$Rn)), - (!cast<Instruction>(INST # UXSr) f32:$Rn)>; - def : Pat<(i32 (to_int_sat_gi f64:$Rn)), - (!cast<Instruction>(INST # UWDr) f64:$Rn)>; - def : Pat<(i64 (to_int_sat_gi f64:$Rn)), - (!cast<Instruction>(INST # UXDr) f64:$Rn)>; - - let Predicates = [HasFullFP16] in { - def : Pat<(i32 (to_int_sat (fmul f16:$Rn, fixedpoint_f16_i32:$scale), i32)), - (!cast<Instruction>(INST # SWHri) $Rn, $scale)>; - def : Pat<(i64 (to_int_sat (fmul f16:$Rn, fixedpoint_f16_i64:$scale), i64)), - (!cast<Instruction>(INST # SXHri) $Rn, $scale)>; - } - def : Pat<(i32 (to_int_sat (fmul f32:$Rn, fixedpoint_f32_i32:$scale), i32)), - (!cast<Instruction>(INST # SWSri) $Rn, $scale)>; - def : Pat<(i64 (to_int_sat (fmul f32:$Rn, fixedpoint_f32_i64:$scale), i64)), - (!cast<Instruction>(INST # SXSri) $Rn, $scale)>; - def : Pat<(i32 (to_int_sat (fmul f64:$Rn, fixedpoint_f64_i32:$scale), i32)), - (!cast<Instruction>(INST # SWDri) $Rn, $scale)>; - def : Pat<(i64 (to_int_sat (fmul f64:$Rn, fixedpoint_f64_i64:$scale), i64)), - (!cast<Instruction>(INST # SXDri) $Rn, $scale)>; - - let Predicates = [HasFullFP16] in { - def : Pat<(i32 (to_int_sat_gi (fmul f16:$Rn, fixedpoint_f16_i32:$scale))), - (!cast<Instruction>(INST # SWHri) $Rn, $scale)>; - def : Pat<(i64 (to_int_sat_gi (fmul f16:$Rn, fixedpoint_f16_i64:$scale))), - (!cast<Instruction>(INST # SXHri) $Rn, $scale)>; - } - def : Pat<(i32 (to_int_sat_gi (fmul f32:$Rn, fixedpoint_f32_i32:$scale))), - (!cast<Instruction>(INST # SWSri) $Rn, $scale)>; - def : Pat<(i64 (to_int_sat_gi (fmul f32:$Rn, fixedpoint_f32_i64:$scale))), - (!cast<Instruction>(INST # SXSri) $Rn, $scale)>; - def : Pat<(i32 (to_int_sat_gi (fmul f64:$Rn, fixedpoint_f64_i32:$scale))), - (!cast<Instruction>(INST # SWDri) $Rn, $scale)>; - def : Pat<(i64 (to_int_sat_gi (fmul f64:$Rn, fixedpoint_f64_i64:$scale))), - (!cast<Instruction>(INST # SXDri) $Rn, $scale)>; -} - -defm : FPToIntegerSatPats<fp_to_sint_sat, fp_to_sint_sat_gi, "FCVTZS">; -defm : FPToIntegerSatPats<fp_to_uint_sat, fp_to_uint_sat_gi, "FCVTZU">; - -multiclass FPToIntegerPats<SDNode to_int, SDNode to_int_sat, SDNode round, string INST> { - def : Pat<(i32 (to_int (round f32:$Rn))), - (!cast<Instruction>(INST # UWSr) f32:$Rn)>; - def : Pat<(i64 (to_int (round f32:$Rn))), - (!cast<Instruction>(INST # UXSr) f32:$Rn)>; - def : Pat<(i32 (to_int (round f64:$Rn))), - (!cast<Instruction>(INST # UWDr) f64:$Rn)>; - def : Pat<(i64 (to_int (round f64:$Rn))), - (!cast<Instruction>(INST # UXDr) f64:$Rn)>; - - // These instructions saturate like fp_to_[su]int_sat. - let Predicates = [HasFullFP16] in { - def : Pat<(i32 (to_int_sat (round f16:$Rn), i32)), - (!cast<Instruction>(INST # UWHr) f16:$Rn)>; - def : Pat<(i64 (to_int_sat (round f16:$Rn), i64)), - (!cast<Instruction>(INST # UXHr) f16:$Rn)>; - } - def : Pat<(i32 (to_int_sat (round f32:$Rn), i32)), - (!cast<Instruction>(INST # UWSr) f32:$Rn)>; - def : Pat<(i64 (to_int_sat (round f32:$Rn), i64)), - (!cast<Instruction>(INST # UXSr) f32:$Rn)>; - def : Pat<(i32 (to_int_sat (round f64:$Rn), i32)), - (!cast<Instruction>(INST # UWDr) f64:$Rn)>; - def : Pat<(i64 (to_int_sat (round f64:$Rn), i64)), - (!cast<Instruction>(INST # UXDr) f64:$Rn)>; + defm FCVTZS : FPToIntegerSIMDScalar<0b10, 0b110, "fcvtzs", any_fp_to_sint>; + defm FCVTZU : FPToIntegerSIMDScalar<0b10, 0b111, "fcvtzu", any_fp_to_uint>; } -defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, fceil, "FCVTPS">; -defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, fceil, "FCVTPU">; -defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, ffloor, "FCVTMS">; -defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, ffloor, "FCVTMU">; -defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, ftrunc, "FCVTZS">; -defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, ftrunc, "FCVTZU">; -defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, fround, "FCVTAS">; -defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, fround, "FCVTAU">; - let Predicates = [HasFullFP16] in { @@ -6567,8 +6506,8 @@ defm FCVTNU : SIMDFPTwoScalar< 1, 0, 0b11010, "fcvtnu", int_aarch64_neon_fcvtn defm FCVTPS : SIMDFPTwoScalar< 0, 1, 0b11010, "fcvtps", int_aarch64_neon_fcvtps>; defm FCVTPU : SIMDFPTwoScalar< 1, 1, 0b11010, "fcvtpu", int_aarch64_neon_fcvtpu>; def FCVTXNv1i64 : SIMDInexactCvtTwoScalar<0b10110, "fcvtxn">; -defm FCVTZS : SIMDFPTwoScalar< 0, 1, 0b11011, "fcvtzs">; -defm FCVTZU : SIMDFPTwoScalar< 1, 1, 0b11011, "fcvtzu">; +defm FCVTZS : SIMDFPTwoScalar< 0, 1, 0b11011, "fcvtzs", any_fp_to_sint>; +defm FCVTZU : SIMDFPTwoScalar< 1, 1, 0b11011, "fcvtzu", any_fp_to_uint>; defm FRECPE : SIMDFPTwoScalar< 0, 1, 0b11101, "frecpe">; defm FRECPX : SIMDFPTwoScalar< 0, 1, 0b11111, "frecpx">; defm FRSQRTE : SIMDFPTwoScalar< 1, 1, 0b11101, "frsqrte">; @@ -6588,6 +6527,7 @@ defm USQADD : SIMDTwoScalarBHSDTied< 1, 0b00011, "usqadd", // Floating-point conversion patterns. multiclass FPToIntegerSIMDScalarPatterns<SDPatternOperator OpN, string INST> { + let Predicates = [HasFPRCVT] in { def : Pat<(f32 (bitconvert (i32 (OpN (f64 FPR64:$Rn))))), (!cast<Instruction>(INST # SDr) FPR64:$Rn)>; def : Pat<(f32 (bitconvert (i32 (OpN (f16 FPR16:$Rn))))), @@ -6596,6 +6536,7 @@ multiclass FPToIntegerSIMDScalarPatterns<SDPatternOperator OpN, string INST> { (!cast<Instruction>(INST # DHr) FPR16:$Rn)>; def : Pat<(f64 (bitconvert (i64 (OpN (f32 FPR32:$Rn))))), (!cast<Instruction>(INST # DSr) FPR32:$Rn)>; + } def : Pat<(f32 (bitconvert (i32 (OpN (f32 FPR32:$Rn))))), (!cast<Instruction>(INST # v1i32) FPR32:$Rn)>; def : Pat<(f64 (bitconvert (i64 (OpN (f64 FPR64:$Rn))))), @@ -6610,6 +6551,8 @@ defm: FPToIntegerSIMDScalarPatterns<int_aarch64_neon_fcvtns, "FCVTNS">; defm: FPToIntegerSIMDScalarPatterns<int_aarch64_neon_fcvtnu, "FCVTNU">; defm: FPToIntegerSIMDScalarPatterns<int_aarch64_neon_fcvtps, "FCVTPS">; defm: FPToIntegerSIMDScalarPatterns<int_aarch64_neon_fcvtpu, "FCVTPU">; +defm: FPToIntegerSIMDScalarPatterns<any_fp_to_sint, "FCVTZS">; +defm: FPToIntegerSIMDScalarPatterns<any_fp_to_uint, "FCVTZU">; multiclass FPToIntegerIntPats<Intrinsic round, string INST> { let Predicates = [HasFullFP16] in { @@ -6666,6 +6609,196 @@ multiclass FPToIntegerIntPats<Intrinsic round, string INST> { defm : FPToIntegerIntPats<int_aarch64_neon_fcvtzs, "FCVTZS">; defm : FPToIntegerIntPats<int_aarch64_neon_fcvtzu, "FCVTZU">; +// AArch64's FCVT instructions saturate when out of range. +multiclass FPToIntegerSatPats<SDNode to_int_sat, SDNode to_int_sat_gi, string INST> { + let Predicates = [HasFullFP16] in { + def : Pat<(i32 (to_int_sat f16:$Rn, i32)), + (!cast<Instruction>(INST # UWHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat f16:$Rn, i64)), + (!cast<Instruction>(INST # UXHr) f16:$Rn)>; + } + def : Pat<(i32 (to_int_sat f32:$Rn, i32)), + (!cast<Instruction>(INST # UWSr) f32:$Rn)>; + def : Pat<(i64 (to_int_sat f32:$Rn, i64)), + (!cast<Instruction>(INST # UXSr) f32:$Rn)>; + def : Pat<(i32 (to_int_sat f64:$Rn, i32)), + (!cast<Instruction>(INST # UWDr) f64:$Rn)>; + def : Pat<(i64 (to_int_sat f64:$Rn, i64)), + (!cast<Instruction>(INST # UXDr) f64:$Rn)>; + + let Predicates = [HasFullFP16] in { + def : Pat<(i32 (to_int_sat_gi f16:$Rn)), + (!cast<Instruction>(INST # UWHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat_gi f16:$Rn)), + (!cast<Instruction>(INST # UXHr) f16:$Rn)>; + } + def : Pat<(i32 (to_int_sat_gi f32:$Rn)), + (!cast<Instruction>(INST # UWSr) f32:$Rn)>; + def : Pat<(i64 (to_int_sat_gi f32:$Rn)), + (!cast<Instruction>(INST # UXSr) f32:$Rn)>; + def : Pat<(i32 (to_int_sat_gi f64:$Rn)), + (!cast<Instruction>(INST # UWDr) f64:$Rn)>; + def : Pat<(i64 (to_int_sat_gi f64:$Rn)), + (!cast<Instruction>(INST # UXDr) f64:$Rn)>; + + // For global-isel we can use register classes to determine + // which FCVT instruction to use. + let Predicates = [HasFPRCVT] in { + def : Pat<(i32 (to_int_sat_gi f16:$Rn)), + (!cast<Instruction>(INST # SHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat_gi f16:$Rn)), + (!cast<Instruction>(INST # DHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat_gi f32:$Rn)), + (!cast<Instruction>(INST # DSr) f32:$Rn)>; + def : Pat<(i32 (to_int_sat_gi f64:$Rn)), + (!cast<Instruction>(INST # SDr) f64:$Rn)>; + } + def : Pat<(i32 (to_int_sat_gi f32:$Rn)), + (!cast<Instruction>(INST # v1i32) f32:$Rn)>; + def : Pat<(i64 (to_int_sat_gi f64:$Rn)), + (!cast<Instruction>(INST # v1i64) f64:$Rn)>; + + let Predicates = [HasFPRCVT] in { + def : Pat<(f32 (bitconvert (i32 (to_int_sat f16:$Rn, i32)))), + (!cast<Instruction>(INST # SHr) f16:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int_sat f16:$Rn, i64)))), + (!cast<Instruction>(INST # DHr) f16:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int_sat f32:$Rn, i64)))), + (!cast<Instruction>(INST # DSr) f32:$Rn)>; + def : Pat<(f32 (bitconvert (i32 (to_int_sat f64:$Rn, i32)))), + (!cast<Instruction>(INST # SDr) f64:$Rn)>; + } + def : Pat<(f32 (bitconvert (i32 (to_int_sat f32:$Rn, i32)))), + (!cast<Instruction>(INST # v1i32) f32:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int_sat f64:$Rn, i64)))), + (!cast<Instruction>(INST # v1i64) f64:$Rn)>; + + let Predicates = [HasFullFP16] in { + def : Pat<(i32 (to_int_sat (fmul f16:$Rn, fixedpoint_f16_i32:$scale), i32)), + (!cast<Instruction>(INST # SWHri) $Rn, $scale)>; + def : Pat<(i64 (to_int_sat (fmul f16:$Rn, fixedpoint_f16_i64:$scale), i64)), + (!cast<Instruction>(INST # SXHri) $Rn, $scale)>; + } + def : Pat<(i32 (to_int_sat (fmul f32:$Rn, fixedpoint_f32_i32:$scale), i32)), + (!cast<Instruction>(INST # SWSri) $Rn, $scale)>; + def : Pat<(i64 (to_int_sat (fmul f32:$Rn, fixedpoint_f32_i64:$scale), i64)), + (!cast<Instruction>(INST # SXSri) $Rn, $scale)>; + def : Pat<(i32 (to_int_sat (fmul f64:$Rn, fixedpoint_f64_i32:$scale), i32)), + (!cast<Instruction>(INST # SWDri) $Rn, $scale)>; + def : Pat<(i64 (to_int_sat (fmul f64:$Rn, fixedpoint_f64_i64:$scale), i64)), + (!cast<Instruction>(INST # SXDri) $Rn, $scale)>; + + let Predicates = [HasFullFP16] in { + def : Pat<(i32 (to_int_sat_gi (fmul f16:$Rn, fixedpoint_f16_i32:$scale))), + (!cast<Instruction>(INST # SWHri) $Rn, $scale)>; + def : Pat<(i64 (to_int_sat_gi (fmul f16:$Rn, fixedpoint_f16_i64:$scale))), + (!cast<Instruction>(INST # SXHri) $Rn, $scale)>; + } + def : Pat<(i32 (to_int_sat_gi (fmul f32:$Rn, fixedpoint_f32_i32:$scale))), + (!cast<Instruction>(INST # SWSri) $Rn, $scale)>; + def : Pat<(i64 (to_int_sat_gi (fmul f32:$Rn, fixedpoint_f32_i64:$scale))), + (!cast<Instruction>(INST # SXSri) $Rn, $scale)>; + def : Pat<(i32 (to_int_sat_gi (fmul f64:$Rn, fixedpoint_f64_i32:$scale))), + (!cast<Instruction>(INST # SWDri) $Rn, $scale)>; + def : Pat<(i64 (to_int_sat_gi (fmul f64:$Rn, fixedpoint_f64_i64:$scale))), + (!cast<Instruction>(INST # SXDri) $Rn, $scale)>; +} + +defm : FPToIntegerSatPats<fp_to_sint_sat, fp_to_sint_sat_gi, "FCVTZS">; +defm : FPToIntegerSatPats<fp_to_uint_sat, fp_to_uint_sat_gi, "FCVTZU">; + +multiclass FPToIntegerPats<SDNode to_int, SDNode to_int_sat, SDNode to_int_sat_gi, SDNode round, string INST> { + def : Pat<(i32 (to_int (round f32:$Rn))), + (!cast<Instruction>(INST # UWSr) f32:$Rn)>; + def : Pat<(i64 (to_int (round f32:$Rn))), + (!cast<Instruction>(INST # UXSr) f32:$Rn)>; + def : Pat<(i32 (to_int (round f64:$Rn))), + (!cast<Instruction>(INST # UWDr) f64:$Rn)>; + def : Pat<(i64 (to_int (round f64:$Rn))), + (!cast<Instruction>(INST # UXDr) f64:$Rn)>; + + // For global-isel we can use register classes to determine + // which FCVT instruction to use. + let Predicates = [HasFPRCVT] in { + def : Pat<(i64 (to_int (round f32:$Rn))), + (!cast<Instruction>(INST # DSr) f32:$Rn)>; + def : Pat<(i32 (to_int (round f64:$Rn))), + (!cast<Instruction>(INST # SDr) f64:$Rn)>; + } + def : Pat<(i32 (to_int (round f32:$Rn))), + (!cast<Instruction>(INST # v1i32) f32:$Rn)>; + def : Pat<(i64 (to_int (round f64:$Rn))), + (!cast<Instruction>(INST # v1i64) f64:$Rn)>; + + let Predicates = [HasFPRCVT] in { + def : Pat<(f64 (bitconvert (i64 (to_int (round f32:$Rn))))), + (!cast<Instruction>(INST # DSr) f32:$Rn)>; + def : Pat<(f32 (bitconvert (i32 (to_int (round f64:$Rn))))), + (!cast<Instruction>(INST # SDr) f64:$Rn)>; + } + def : Pat<(f32 (bitconvert (i32 (to_int (round f32:$Rn))))), + (!cast<Instruction>(INST # v1i32) f32:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int (round f64:$Rn))))), + (!cast<Instruction>(INST # v1i64) f64:$Rn)>; + + // These instructions saturate like fp_to_[su]int_sat. + let Predicates = [HasFullFP16] in { + def : Pat<(i32 (to_int_sat (round f16:$Rn), i32)), + (!cast<Instruction>(INST # UWHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat (round f16:$Rn), i64)), + (!cast<Instruction>(INST # UXHr) f16:$Rn)>; + } + def : Pat<(i32 (to_int_sat (round f32:$Rn), i32)), + (!cast<Instruction>(INST # UWSr) f32:$Rn)>; + def : Pat<(i64 (to_int_sat (round f32:$Rn), i64)), + (!cast<Instruction>(INST # UXSr) f32:$Rn)>; + def : Pat<(i32 (to_int_sat (round f64:$Rn), i32)), + (!cast<Instruction>(INST # UWDr) f64:$Rn)>; + def : Pat<(i64 (to_int_sat (round f64:$Rn), i64)), + (!cast<Instruction>(INST # UXDr) f64:$Rn)>; + + // For global-isel we can use register classes to determine + // which FCVT instruction to use. + let Predicates = [HasFPRCVT] in { + def : Pat<(i32 (to_int_sat_gi (round f16:$Rn))), + (!cast<Instruction>(INST # SHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat_gi (round f16:$Rn))), + (!cast<Instruction>(INST # DHr) f16:$Rn)>; + def : Pat<(i64 (to_int_sat_gi (round f32:$Rn))), + (!cast<Instruction>(INST # DSr) f32:$Rn)>; + def : Pat<(i32 (to_int_sat_gi (round f64:$Rn))), + (!cast<Instruction>(INST # SDr) f64:$Rn)>; + } + def : Pat<(i32 (to_int_sat_gi (round f32:$Rn))), + (!cast<Instruction>(INST # v1i32) f32:$Rn)>; + def : Pat<(i64 (to_int_sat_gi (round f64:$Rn))), + (!cast<Instruction>(INST # v1i64) f64:$Rn)>; + + let Predicates = [HasFPRCVT] in { + def : Pat<(f32 (bitconvert (i32 (to_int_sat (round f16:$Rn), i32)))), + (!cast<Instruction>(INST # SHr) f16:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int_sat (round f16:$Rn), i64)))), + (!cast<Instruction>(INST # DHr) f16:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int_sat (round f32:$Rn), i64)))), + (!cast<Instruction>(INST # DSr) f32:$Rn)>; + def : Pat<(f32 (bitconvert (i32 (to_int_sat (round f64:$Rn), i32)))), + (!cast<Instruction>(INST # SDr) f64:$Rn)>; + } + def : Pat<(f32 (bitconvert (i32 (to_int_sat (round f32:$Rn), i32)))), + (!cast<Instruction>(INST # v1i32) f32:$Rn)>; + def : Pat<(f64 (bitconvert (i64 (to_int_sat (round f64:$Rn), i64)))), + (!cast<Instruction>(INST # v1i64) f64:$Rn)>; +} + +defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, fp_to_sint_sat_gi, fceil, "FCVTPS">; +defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, fp_to_uint_sat_gi, fceil, "FCVTPU">; +defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, fp_to_sint_sat_gi, ffloor, "FCVTMS">; +defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, fp_to_uint_sat_gi, ffloor, "FCVTMU">; +defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, fp_to_sint_sat_gi, ftrunc, "FCVTZS">; +defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, fp_to_uint_sat_gi, ftrunc, "FCVTZU">; +defm : FPToIntegerPats<fp_to_sint, fp_to_sint_sat, fp_to_sint_sat_gi, fround, "FCVTAS">; +defm : FPToIntegerPats<fp_to_uint, fp_to_uint_sat, fp_to_uint_sat_gi, fround, "FCVTAU">; + // f16 -> s16 conversions let Predicates = [HasFullFP16] in { def : Pat<(i16(fp_to_sint_sat_gi f16:$Rn)), (FCVTZSv1f16 f16:$Rn)>; diff --git a/llvm/lib/Target/AArch64/AArch64SchedNeoverseV2.td b/llvm/lib/Target/AArch64/AArch64SchedNeoverseV2.td index bdde8e3..2387f17 100644 --- a/llvm/lib/Target/AArch64/AArch64SchedNeoverseV2.td +++ b/llvm/lib/Target/AArch64/AArch64SchedNeoverseV2.td @@ -2762,11 +2762,11 @@ def : InstRW<[V2Write_11c_18L01_18V01], (instregex "^ST4[BHWD]_IMM$")>; def : InstRW<[V2Write_11c_18L01_18S_18V01], (instregex "^ST4[BHWD]$")>; // Non temporal store, scalar + imm -def : InstRW<[V2Write_2c_1L01_1V], (instregex "^STNT1[BHWD]_ZRI$")>; +def : InstRW<[V2Write_2c_1L01_1V01], (instregex "^STNT1[BHWD]_ZRI$")>; // Non temporal store, scalar + scalar -def : InstRW<[V2Write_2c_1L01_1S_1V], (instrs STNT1H_ZRR)>; -def : InstRW<[V2Write_2c_1L01_1V], (instregex "^STNT1[BWD]_ZRR$")>; +def : InstRW<[V2Write_2c_1L01_1S_1V01], (instrs STNT1H_ZRR)>; +def : InstRW<[V2Write_2c_1L01_1V01], (instregex "^STNT1[BWD]_ZRR$")>; // Scatter non temporal store, vector + scalar 32-bit element size def : InstRW<[V2Write_4c_4L01_4V01], (instregex "^STNT1[BHW]_ZZR_S")>; diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp index 2053fc4..fede586 100644 --- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp +++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp @@ -224,7 +224,8 @@ static cl::opt<bool> EnableScalableAutovecInStreamingMode( static bool isSMEABIRoutineCall(const CallInst &CI, const AArch64TargetLowering &TLI) { const auto *F = CI.getCalledFunction(); - return F && SMEAttrs(F->getName(), TLI).isSMEABIRoutine(); + return F && + SMEAttrs(F->getName(), TLI.getRuntimeLibcallsInfo()).isSMEABIRoutine(); } /// Returns true if the function has explicit operations that can only be @@ -355,7 +356,7 @@ AArch64TTIImpl::getInlineCallPenalty(const Function *F, const CallBase &Call, // change only once and avoid inlining of G into F. SMEAttrs FAttrs(*F); - SMECallAttrs CallAttrs(Call, getTLI()); + SMECallAttrs CallAttrs(Call, &getTLI()->getRuntimeLibcallsInfo()); if (SMECallAttrs(FAttrs, CallAttrs.callee()).requiresSMChange()) { if (F == Call.getCaller()) // (1) @@ -957,23 +958,50 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, return TyL.first + ExtraCost; } case Intrinsic::get_active_lane_mask: { - auto *RetTy = dyn_cast<FixedVectorType>(ICA.getReturnType()); - if (RetTy) { - EVT RetVT = getTLI()->getValueType(DL, RetTy); - EVT OpVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]); - if (!getTLI()->shouldExpandGetActiveLaneMask(RetVT, OpVT) && - !getTLI()->isTypeLegal(RetVT)) { - // We don't have enough context at this point to determine if the mask - // is going to be kept live after the block, which will force the vXi1 - // type to be expanded to legal vectors of integers, e.g. v4i1->v4i32. - // For now, we just assume the vectorizer created this intrinsic and - // the result will be the input for a PHI. In this case the cost will - // be extremely high for fixed-width vectors. - // NOTE: getScalarizationOverhead returns a cost that's far too - // pessimistic for the actual generated codegen. In reality there are - // two instructions generated per lane. - return RetTy->getNumElements() * 2; + auto RetTy = cast<VectorType>(ICA.getReturnType()); + EVT RetVT = getTLI()->getValueType(DL, RetTy); + EVT OpVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]); + if (getTLI()->shouldExpandGetActiveLaneMask(RetVT, OpVT)) + break; + + if (RetTy->isScalableTy()) { + if (TLI->getTypeAction(RetTy->getContext(), RetVT) != + TargetLowering::TypeSplitVector) + break; + + auto LT = getTypeLegalizationCost(RetTy); + InstructionCost Cost = LT.first; + // When SVE2p1 or SME2 is available, we can halve getTypeLegalizationCost + // as get_active_lane_mask may lower to the sve_whilelo_x2 intrinsic, e.g. + // nxv32i1 = get_active_lane_mask(base, idx) -> + // {nxv16i1, nxv16i1} = sve_whilelo_x2(base, idx) + if (ST->hasSVE2p1() || ST->hasSME2()) { + Cost /= 2; + if (Cost == 1) + return Cost; } + + // If more than one whilelo intrinsic is required, include the extra cost + // required by the saturating add & select required to increment the + // start value after the first intrinsic call. + Type *OpTy = ICA.getArgTypes()[0]; + IntrinsicCostAttributes AddAttrs(Intrinsic::uadd_sat, OpTy, {OpTy, OpTy}); + InstructionCost SplitCost = getIntrinsicInstrCost(AddAttrs, CostKind); + Type *CondTy = OpTy->getWithNewBitWidth(1); + SplitCost += getCmpSelInstrCost(Instruction::Select, OpTy, CondTy, + CmpInst::ICMP_UGT, CostKind); + return Cost + (SplitCost * (Cost - 1)); + } else if (!getTLI()->isTypeLegal(RetVT)) { + // We don't have enough context at this point to determine if the mask + // is going to be kept live after the block, which will force the vXi1 + // type to be expanded to legal vectors of integers, e.g. v4i1->v4i32. + // For now, we just assume the vectorizer created this intrinsic and + // the result will be the input for a PHI. In this case the cost will + // be extremely high for fixed-width vectors. + // NOTE: getScalarizationOverhead returns a cost that's far too + // pessimistic for the actual generated codegen. In reality there are + // two instructions generated per lane. + return cast<FixedVectorType>(RetTy)->getNumElements() * 2; } break; } diff --git a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp index 3e55b76..14b0f9a 100644 --- a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp +++ b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp @@ -5126,23 +5126,13 @@ bool AArch64InstructionSelector::selectShuffleVector( MachineInstr &I, MachineRegisterInfo &MRI) { const LLT DstTy = MRI.getType(I.getOperand(0).getReg()); Register Src1Reg = I.getOperand(1).getReg(); - const LLT Src1Ty = MRI.getType(Src1Reg); Register Src2Reg = I.getOperand(2).getReg(); - const LLT Src2Ty = MRI.getType(Src2Reg); ArrayRef<int> Mask = I.getOperand(3).getShuffleMask(); MachineBasicBlock &MBB = *I.getParent(); MachineFunction &MF = *MBB.getParent(); LLVMContext &Ctx = MF.getFunction().getContext(); - // G_SHUFFLE_VECTOR is weird in that the source operands can be scalars, if - // it's originated from a <1 x T> type. Those should have been lowered into - // G_BUILD_VECTOR earlier. - if (!Src1Ty.isVector() || !Src2Ty.isVector()) { - LLVM_DEBUG(dbgs() << "Could not select a \"scalar\" G_SHUFFLE_VECTOR\n"); - return false; - } - unsigned BytesPerElt = DstTy.getElementType().getSizeInBits() / 8; SmallVector<Constant *, 64> CstIdxs; diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp index 05a4313..5f93847 100644 --- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp +++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp @@ -1201,25 +1201,17 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST) return llvm::is_contained( {v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64}, DstTy); }) - // G_SHUFFLE_VECTOR can have scalar sources (from 1 x s vectors) or scalar - // destinations, we just want those lowered into G_BUILD_VECTOR or - // G_EXTRACT_ELEMENT. - .lowerIf([=](const LegalityQuery &Query) { - return !Query.Types[0].isVector() || !Query.Types[1].isVector(); - }) .moreElementsIf( [](const LegalityQuery &Query) { - return Query.Types[0].isVector() && Query.Types[1].isVector() && - Query.Types[0].getNumElements() > - Query.Types[1].getNumElements(); + return Query.Types[0].getNumElements() > + Query.Types[1].getNumElements(); }, changeTo(1, 0)) .moreElementsToNextPow2(0) .moreElementsIf( [](const LegalityQuery &Query) { - return Query.Types[0].isVector() && Query.Types[1].isVector() && - Query.Types[0].getNumElements() < - Query.Types[1].getNumElements(); + return Query.Types[0].getNumElements() < + Query.Types[1].getNumElements(); }, changeTo(0, 1)) .widenScalarOrEltToNextPow2OrMinSize(0, 8) diff --git a/llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp index 830a35bb..6d2d705 100644 --- a/llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp +++ b/llvm/lib/Target/AArch64/GISel/AArch64RegisterBankInfo.cpp @@ -856,7 +856,9 @@ AArch64RegisterBankInfo::getInstrMapping(const MachineInstr &MI) const { break; } case TargetOpcode::G_FPTOSI_SAT: - case TargetOpcode::G_FPTOUI_SAT: { + case TargetOpcode::G_FPTOUI_SAT: + case TargetOpcode::G_FPTOSI: + case TargetOpcode::G_FPTOUI: { LLT DstType = MRI.getType(MI.getOperand(0).getReg()); if (DstType.isVector()) break; @@ -864,11 +866,19 @@ AArch64RegisterBankInfo::getInstrMapping(const MachineInstr &MI) const { OpRegBankIdx = {PMI_FirstFPR, PMI_FirstFPR}; break; } - OpRegBankIdx = {PMI_FirstGPR, PMI_FirstFPR}; + TypeSize DstSize = getSizeInBits(MI.getOperand(0).getReg(), MRI, TRI); + TypeSize SrcSize = getSizeInBits(MI.getOperand(1).getReg(), MRI, TRI); + if (((DstSize == SrcSize) || STI.hasFeature(AArch64::FeatureFPRCVT)) && + all_of(MRI.use_nodbg_instructions(MI.getOperand(0).getReg()), + [&](const MachineInstr &UseMI) { + return onlyUsesFP(UseMI, MRI, TRI) || + prefersFPUse(UseMI, MRI, TRI); + })) + OpRegBankIdx = {PMI_FirstFPR, PMI_FirstFPR}; + else + OpRegBankIdx = {PMI_FirstGPR, PMI_FirstFPR}; break; } - case TargetOpcode::G_FPTOSI: - case TargetOpcode::G_FPTOUI: case TargetOpcode::G_INTRINSIC_LRINT: case TargetOpcode::G_INTRINSIC_LLRINT: if (MRI.getType(MI.getOperand(0).getReg()).isVector()) diff --git a/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.cpp b/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.cpp index d71f728..085c8588 100644 --- a/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.cpp +++ b/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.cpp @@ -75,8 +75,8 @@ SMEAttrs::SMEAttrs(const AttributeList &Attrs) { } void SMEAttrs::addKnownFunctionAttrs(StringRef FuncName, - const AArch64TargetLowering &TLI) { - RTLIB::LibcallImpl Impl = TLI.getSupportedLibcallImpl(FuncName); + const RTLIB::RuntimeLibcallsInfo &RTLCI) { + RTLIB::LibcallImpl Impl = RTLCI.getSupportedLibcallImpl(FuncName); if (Impl == RTLIB::Unsupported) return; unsigned KnownAttrs = SMEAttrs::Normal; @@ -124,21 +124,22 @@ bool SMECallAttrs::requiresSMChange() const { return true; } -SMECallAttrs::SMECallAttrs(const CallBase &CB, const AArch64TargetLowering *TLI) +SMECallAttrs::SMECallAttrs(const CallBase &CB, + const RTLIB::RuntimeLibcallsInfo *RTLCI) : CallerFn(*CB.getFunction()), CalledFn(SMEAttrs::Normal), Callsite(CB.getAttributes()), IsIndirect(CB.isIndirectCall()) { if (auto *CalledFunction = CB.getCalledFunction()) - CalledFn = SMEAttrs(*CalledFunction, TLI); - - // An `invoke` of an agnostic ZA function may not return normally (it may - // resume in an exception block). In this case, it acts like a private ZA - // callee and may require a ZA save to be set up before it is called. - if (isa<InvokeInst>(CB)) - CalledFn.set(SMEAttrs::ZA_State_Agnostic, /*Enable=*/false); + CalledFn = SMEAttrs(*CalledFunction, RTLCI); // FIXME: We probably should not allow SME attributes on direct calls but // clang duplicates streaming mode attributes at each callsite. assert((IsIndirect || ((Callsite.withoutPerCallsiteFlags() | CalledFn) == CalledFn)) && "SME attributes at callsite do not match declaration"); + + // An `invoke` of an agnostic ZA function may not return normally (it may + // resume in an exception block). In this case, it acts like a private ZA + // callee and may require a ZA save to be set up before it is called. + if (isa<InvokeInst>(CB)) + CalledFn.set(SMEAttrs::ZA_State_Agnostic, /*Enable=*/false); } diff --git a/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.h b/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.h index d26e3cd..28c397e 100644 --- a/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.h +++ b/llvm/lib/Target/AArch64/Utils/AArch64SMEAttributes.h @@ -12,8 +12,9 @@ #include "llvm/IR/Function.h" namespace llvm { - -class AArch64TargetLowering; +namespace RTLIB { +struct RuntimeLibcallsInfo; +} class Function; class CallBase; @@ -52,14 +53,14 @@ public: SMEAttrs() = default; SMEAttrs(unsigned Mask) { set(Mask); } - SMEAttrs(const Function &F, const AArch64TargetLowering *TLI = nullptr) + SMEAttrs(const Function &F, const RTLIB::RuntimeLibcallsInfo *RTLCI = nullptr) : SMEAttrs(F.getAttributes()) { - if (TLI) - addKnownFunctionAttrs(F.getName(), *TLI); + if (RTLCI) + addKnownFunctionAttrs(F.getName(), *RTLCI); } SMEAttrs(const AttributeList &L); - SMEAttrs(StringRef FuncName, const AArch64TargetLowering &TLI) { - addKnownFunctionAttrs(FuncName, TLI); + SMEAttrs(StringRef FuncName, const RTLIB::RuntimeLibcallsInfo &RTLCI) { + addKnownFunctionAttrs(FuncName, RTLCI); }; void set(unsigned M, bool Enable = true) { @@ -157,7 +158,7 @@ public: private: void addKnownFunctionAttrs(StringRef FuncName, - const AArch64TargetLowering &TLI); + const RTLIB::RuntimeLibcallsInfo &RTLCI); void validate() const; }; @@ -175,7 +176,7 @@ public: SMEAttrs Callsite = SMEAttrs::Normal) : CallerFn(Caller), CalledFn(Callee), Callsite(Callsite) {} - SMECallAttrs(const CallBase &CB, const AArch64TargetLowering *TLI); + SMECallAttrs(const CallBase &CB, const RTLIB::RuntimeLibcallsInfo *RTLCI); SMEAttrs &caller() { return CallerFn; } SMEAttrs &callee() { return IsIndirect ? Callsite : CalledFn; } diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCombine.td b/llvm/lib/Target/AMDGPU/AMDGPUCombine.td index e8b211f..7f00ead 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUCombine.td +++ b/llvm/lib/Target/AMDGPU/AMDGPUCombine.td @@ -176,6 +176,19 @@ def binop_s64_with_s32_mask_combines : GICombineGroup<[ combine_or_s64_with_s32_mask, combine_and_s64_with_s32_mask ]>; +// (or i64:x, (zext i32:y)) -> i64:(merge (or lo_32(x), i32:y), hi_32(x)) +// (or (zext i32:y), i64:x) -> i64:(merge (or lo_32(x), i32:y), hi_32(x)) +def or_s64_zext_s32_frag : GICombinePatFrag<(outs root:$dst), (ins $src_s64, $src_s32), + [(pattern (G_OR $dst, i64:$src_s64, i64:$zext_val), (G_ZEXT i64:$zext_val, i32:$src_s32)), + (pattern (G_OR $dst, i64:$zext_val, i64:$src_s64), (G_ZEXT i64:$zext_val, i32:$src_s32))]>; + +def combine_or_s64_s32 : GICombineRule< + (defs root:$dst), + (match (or_s64_zext_s32_frag $dst, i64:$x, i32:$y):$dst), + (apply (G_UNMERGE_VALUES $x_lo, $x_hi, $x), + (G_OR $or, $x_lo, $y), + (G_MERGE_VALUES $dst, $or, $x_hi))>; + let Predicates = [Has16BitInsts, NotHasMed3_16] in { // For gfx8, expand f16-fmed3-as-f32 into a min/max f16 sequence. This // saves one instruction compared to the promotion. @@ -206,7 +219,7 @@ def AMDGPUPreLegalizerCombiner: GICombiner< "AMDGPUPreLegalizerCombinerImpl", [all_combines, combine_fmul_with_select_to_fldexp, clamp_i64_to_i16, foldable_fneg, combine_shuffle_vector_to_build_vector, - binop_s64_with_s32_mask_combines]> { + binop_s64_with_s32_mask_combines, combine_or_s64_s32]> { let CombineAllMethodName = "tryCombineAllImpl"; } @@ -215,7 +228,7 @@ def AMDGPUPostLegalizerCombiner: GICombiner< [all_combines, gfx6gfx7_combines, gfx8_combines, combine_fmul_with_select_to_fldexp, uchar_to_float, cvt_f32_ubyteN, remove_fcanonicalize, foldable_fneg, rcp_sqrt_to_rsq, fdiv_by_sqrt_to_rsq_f16, sign_extension_in_reg, smulu64, - binop_s64_with_s32_mask_combines]> { + binop_s64_with_s32_mask_combines, combine_or_s64_s32]> { let CombineAllMethodName = "tryCombineAllImpl"; } diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp index 596a895..1a13b22 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp @@ -976,9 +976,25 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_, FPOpActions.clampMaxNumElementsStrict(0, S32, 2); } + auto &MinNumMaxNumIeee = + getActionDefinitionsBuilder({G_FMINNUM_IEEE, G_FMAXNUM_IEEE}); + + if (ST.hasVOP3PInsts()) { + MinNumMaxNumIeee.legalFor(FPTypesPK16) + .moreElementsIf(isSmallOddVector(0), oneMoreElement(0)) + .clampMaxNumElements(0, S16, 2) + .clampScalar(0, S16, S64) + .scalarize(0); + } else if (ST.has16BitInsts()) { + MinNumMaxNumIeee.legalFor(FPTypes16).clampScalar(0, S16, S64).scalarize(0); + } else { + MinNumMaxNumIeee.legalFor(FPTypesBase) + .clampScalar(0, S32, S64) + .scalarize(0); + } + auto &MinNumMaxNum = getActionDefinitionsBuilder( - {G_FMINNUM, G_FMAXNUM, G_FMINIMUMNUM, G_FMAXIMUMNUM, G_FMINNUM_IEEE, - G_FMAXNUM_IEEE}); + {G_FMINNUM, G_FMAXNUM, G_FMINIMUMNUM, G_FMAXIMUMNUM}); if (ST.hasVOP3PInsts()) { MinNumMaxNum.customFor(FPTypesPK16) @@ -2136,9 +2152,17 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_, .legalFor(FPTypesPK16) .clampMaxNumElements(0, S16, 2) .scalarize(0); + } else if (ST.hasVOP3PInsts()) { + getActionDefinitionsBuilder({G_FMINIMUM, G_FMAXIMUM}) + .lowerFor({V2S16}) + .clampMaxNumElementsStrict(0, S16, 2) + .scalarize(0) + .lower(); } else { - // TODO: Implement - getActionDefinitionsBuilder({G_FMINIMUM, G_FMAXIMUM}).lower(); + getActionDefinitionsBuilder({G_FMINIMUM, G_FMAXIMUM}) + .scalarize(0) + .clampScalar(0, S32, S64) + .lower(); } getActionDefinitionsBuilder({G_MEMCPY, G_MEMCPY_INLINE, G_MEMMOVE, G_MEMSET}) @@ -2195,8 +2219,6 @@ bool AMDGPULegalizerInfo::legalizeCustom( case TargetOpcode::G_FMAXNUM: case TargetOpcode::G_FMINIMUMNUM: case TargetOpcode::G_FMAXIMUMNUM: - case TargetOpcode::G_FMINNUM_IEEE: - case TargetOpcode::G_FMAXNUM_IEEE: return legalizeMinNumMaxNum(Helper, MI); case TargetOpcode::G_EXTRACT_VECTOR_ELT: return legalizeExtractVectorElt(MI, MRI, B); @@ -2817,23 +2839,8 @@ bool AMDGPULegalizerInfo::legalizeMinNumMaxNum(LegalizerHelper &Helper, MachineFunction &MF = Helper.MIRBuilder.getMF(); const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); - const bool IsIEEEOp = MI.getOpcode() == AMDGPU::G_FMINNUM_IEEE || - MI.getOpcode() == AMDGPU::G_FMAXNUM_IEEE; - - // With ieee_mode disabled, the instructions have the correct behavior - // already for G_FMINIMUMNUM/G_FMAXIMUMNUM. - // - // FIXME: G_FMINNUM/G_FMAXNUM should match the behavior with ieee_mode - // enabled. - if (!MFI->getMode().IEEE) { - if (MI.getOpcode() == AMDGPU::G_FMINIMUMNUM || - MI.getOpcode() == AMDGPU::G_FMAXIMUMNUM) - return true; - - return !IsIEEEOp; - } - - if (IsIEEEOp) + // With ieee_mode disabled, the instructions have the correct behavior. + if (!MFI->getMode().IEEE) return true; return Helper.lowerFMinNumMaxNum(MI) == LegalizerHelper::Legalized; diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp index 8122db2..313ae3d 100644 --- a/llvm/lib/Target/ARM/ARMISelLowering.cpp +++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp @@ -21381,15 +21381,6 @@ void ARMTargetLowering::insertSSPDeclarations(Module &M) const { TargetLowering::insertSSPDeclarations(M); } -Function *ARMTargetLowering::getSSPStackGuardCheck(const Module &M) const { - // MSVC CRT has a function to validate security cookie. - RTLIB::LibcallImpl SecurityCheckCookie = - getLibcallImpl(RTLIB::SECURITY_CHECK_COOKIE); - if (SecurityCheckCookie != RTLIB::Unsupported) - return M.getFunction(getLibcallImplName(SecurityCheckCookie)); - return TargetLowering::getSSPStackGuardCheck(M); -} - bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx, unsigned &Cost) const { // If we do not have NEON, vector types are not natively supported. diff --git a/llvm/lib/Target/ARM/ARMISelLowering.h b/llvm/lib/Target/ARM/ARMISelLowering.h index 8c5e0cf..357d2c5 100644 --- a/llvm/lib/Target/ARM/ARMISelLowering.h +++ b/llvm/lib/Target/ARM/ARMISelLowering.h @@ -708,7 +708,6 @@ class VectorType; bool useLoadStackGuardNode(const Module &M) const override; void insertSSPDeclarations(Module &M) const override; - Function *getSSPStackGuardCheck(const Module &M) const override; bool canCombineStoreAndExtract(Type *VectorTy, Value *Idx, unsigned &Cost) const override; diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td index 53be167..10d4cd5 100644 --- a/llvm/lib/Target/ARM/ARMInstrInfo.td +++ b/llvm/lib/Target/ARM/ARMInstrInfo.td @@ -6546,23 +6546,25 @@ def KCFI_CHECK_ARM : PseudoInst<(outs), (ins GPR:$ptr, i32imm:$type), NoItinerary, []>, Sched<[]>, Requires<[IsARM]> { - let Size = 28; // 7 instructions (bic, ldr, 4x eor, beq, udf) + let Size = 40; // worst-case 10 instructions @ 4 bytes each + // (push, bic, ldr, 4x eor, pop, beq, udf) } def KCFI_CHECK_Thumb2 : PseudoInst<(outs), (ins GPR:$ptr, i32imm:$type), NoItinerary, []>, Sched<[]>, Requires<[IsThumb2]> { - let Size = - 32; // worst-case 9 instructions (push, bic, ldr, 4x eor, pop, beq.w, udf) + let Size = 34; // worst-case (push.w[2], bic[4], ldr[4], 4x eor[16], pop.w[2], + // beq.w[4], udf[2]) } def KCFI_CHECK_Thumb1 : PseudoInst<(outs), (ins GPR:$ptr, i32imm:$type), NoItinerary, []>, Sched<[]>, Requires<[IsThumb1Only]> { - let Size = 50; // worst-case 25 instructions (pushes, bic helper, type - // building, cmp, pops) + let Size = 38; // worst-case 19 instructions @ 2 bytes each + // (2x push, 3x bic-helper, subs+ldr, 13x type-building, cmp, + // 2x pop, beq, bkpt) } //===----------------------------------------------------------------------===// diff --git a/llvm/lib/Target/DirectX/DXILPrepare.cpp b/llvm/lib/Target/DirectX/DXILPrepare.cpp index 42e90f0..d6fa65f 100644 --- a/llvm/lib/Target/DirectX/DXILPrepare.cpp +++ b/llvm/lib/Target/DirectX/DXILPrepare.cpp @@ -6,7 +6,7 @@ // //===----------------------------------------------------------------------===// /// -/// \file This file contains pases and utilities to convert a modern LLVM +/// \file This file contains passes and utilities to convert a modern LLVM /// module into a module compatible with the LLVM 3.7-based DirectX Intermediate /// Language (DXIL). //===----------------------------------------------------------------------===// @@ -16,7 +16,6 @@ #include "DirectX.h" #include "DirectXIRPasses/PointerTypeAnalysis.h" #include "llvm/ADT/STLExtras.h" -#include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringSet.h" #include "llvm/Analysis/DXILMetadataAnalysis.h" #include "llvm/Analysis/DXILResource.h" @@ -27,7 +26,6 @@ #include "llvm/IR/Module.h" #include "llvm/InitializePasses.h" #include "llvm/Pass.h" -#include "llvm/Support/Compiler.h" #include "llvm/Support/VersionTuple.h" #define DEBUG_TYPE "dxil-prepare" @@ -116,31 +114,6 @@ static void removeStringFunctionAttributes(Function &F, F.removeRetAttrs(DeadAttrs); } -static void cleanModuleFlags(Module &M) { - NamedMDNode *MDFlags = M.getModuleFlagsMetadata(); - if (!MDFlags) - return; - - SmallVector<llvm::Module::ModuleFlagEntry> FlagEntries; - M.getModuleFlagsMetadata(FlagEntries); - bool Updated = false; - for (auto &Flag : FlagEntries) { - // llvm 3.7 only supports behavior up to AppendUnique. - if (Flag.Behavior <= Module::ModFlagBehavior::AppendUnique) - continue; - Flag.Behavior = Module::ModFlagBehavior::Warning; - Updated = true; - } - - if (!Updated) - return; - - MDFlags->eraseFromParent(); - - for (auto &Flag : FlagEntries) - M.addModuleFlag(Flag.Behavior, Flag.Key->getString(), Flag.Val); -} - class DXILPrepareModule : public ModulePass { static Value *maybeGenerateBitcast(IRBuilder<> &Builder, @@ -202,15 +175,6 @@ class DXILPrepareModule : public ModulePass { Builder.getPtrTy(PtrTy->getAddressSpace()))); } - static std::array<unsigned, 6> getCompatibleInstructionMDs(llvm::Module &M) { - return {M.getMDKindID("dx.nonuniform"), - M.getMDKindID("dx.controlflow.hints"), - M.getMDKindID("dx.precise"), - llvm::LLVMContext::MD_range, - llvm::LLVMContext::MD_alias_scope, - llvm::LLVMContext::MD_noalias}; - } - public: bool runOnModule(Module &M) override { PointerTypeMap PointerTypes = PointerTypeAnalysis::run(M); @@ -224,10 +188,7 @@ public: const dxil::ModuleMetadataInfo MetadataInfo = getAnalysis<DXILMetadataAnalysisWrapperPass>().getModuleMetadata(); VersionTuple ValVer = MetadataInfo.ValidatorVersion; - bool SkipValidation = ValVer.getMajor() == 0 && ValVer.getMinor() == 0; - - // construct allowlist of valid metadata node kinds - std::array<unsigned, 6> DXILCompatibleMDs = getCompatibleInstructionMDs(M); + bool AllowExperimental = ValVer.getMajor() == 0 && ValVer.getMinor() == 0; for (auto &F : M.functions()) { F.removeFnAttrs(AttrMask); @@ -235,7 +196,7 @@ public: // Only remove string attributes if we are not skipping validation. // This will reserve the experimental attributes when validation version // is 0.0 for experiment mode. - removeStringFunctionAttributes(F, SkipValidation); + removeStringFunctionAttributes(F, AllowExperimental); for (size_t Idx = 0, End = F.arg_size(); Idx < End; ++Idx) F.removeParamAttrs(Idx, AttrMask); @@ -243,11 +204,17 @@ public: IRBuilder<> Builder(&BB); for (auto &I : make_early_inc_range(BB)) { - I.dropUnknownNonDebugMetadata(DXILCompatibleMDs); + if (auto *CB = dyn_cast<CallBase>(&I)) { + CB->removeFnAttrs(AttrMask); + CB->removeRetAttrs(AttrMask); + for (size_t Idx = 0, End = CB->arg_size(); Idx < End; ++Idx) + CB->removeParamAttrs(Idx, AttrMask); + continue; + } // Emtting NoOp bitcast instructions allows the ValueEnumerator to be // unmodified as it reserves instruction IDs during contruction. - if (auto LI = dyn_cast<LoadInst>(&I)) { + if (auto *LI = dyn_cast<LoadInst>(&I)) { if (Value *NoOpBitcast = maybeGenerateBitcast( Builder, PointerTypes, I, LI->getPointerOperand(), LI->getType())) { @@ -257,7 +224,7 @@ public: } continue; } - if (auto SI = dyn_cast<StoreInst>(&I)) { + if (auto *SI = dyn_cast<StoreInst>(&I)) { if (Value *NoOpBitcast = maybeGenerateBitcast( Builder, PointerTypes, I, SI->getPointerOperand(), SI->getValueOperand()->getType())) { @@ -268,39 +235,16 @@ public: } continue; } - if (auto GEP = dyn_cast<GetElementPtrInst>(&I)) { + if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { if (Value *NoOpBitcast = maybeGenerateBitcast( Builder, PointerTypes, I, GEP->getPointerOperand(), GEP->getSourceElementType())) GEP->setOperand(0, NoOpBitcast); continue; } - if (auto *CB = dyn_cast<CallBase>(&I)) { - CB->removeFnAttrs(AttrMask); - CB->removeRetAttrs(AttrMask); - for (size_t Idx = 0, End = CB->arg_size(); Idx < End; ++Idx) - CB->removeParamAttrs(Idx, AttrMask); - continue; - } } } } - // Remove flags not for DXIL. - cleanModuleFlags(M); - - // dx.rootsignatures will have been parsed from its metadata form as its - // binary form as part of the RootSignatureAnalysisWrapper, so safely - // remove it as it is not recognized in DXIL - if (NamedMDNode *RootSignature = M.getNamedMetadata("dx.rootsignatures")) - RootSignature->eraseFromParent(); - - // llvm.errno.tbaa was recently added but is not supported in LLVM 3.7 and - // causes all tests using the DXIL Validator to fail. - // - // This is a temporary fix and should be replaced with a whitelist once - // we have determined all metadata that the DXIL Validator allows - if (NamedMDNode *ErrNo = M.getNamedMetadata("llvm.errno.tbaa")) - ErrNo->eraseFromParent(); return true; } @@ -308,11 +252,11 @@ public: DXILPrepareModule() : ModulePass(ID) {} void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired<DXILMetadataAnalysisWrapperPass>(); - AU.addRequired<RootSignatureAnalysisWrapper>(); - AU.addPreserved<RootSignatureAnalysisWrapper>(); - AU.addPreserved<ShaderFlagsAnalysisWrapper>(); + AU.addPreserved<DXILMetadataAnalysisWrapperPass>(); AU.addPreserved<DXILResourceWrapperPass>(); + AU.addPreserved<RootSignatureAnalysisWrapper>(); + AU.addPreserved<ShaderFlagsAnalysisWrapper>(); } static char ID; // Pass identification. }; @@ -323,7 +267,6 @@ char DXILPrepareModule::ID = 0; INITIALIZE_PASS_BEGIN(DXILPrepareModule, DEBUG_TYPE, "DXIL Prepare Module", false, false) INITIALIZE_PASS_DEPENDENCY(DXILMetadataAnalysisWrapperPass) -INITIALIZE_PASS_DEPENDENCY(RootSignatureAnalysisWrapper) INITIALIZE_PASS_END(DXILPrepareModule, DEBUG_TYPE, "DXIL Prepare Module", false, false) diff --git a/llvm/lib/Target/DirectX/DXILTranslateMetadata.cpp b/llvm/lib/Target/DirectX/DXILTranslateMetadata.cpp index 9eebcc9..1e4797b 100644 --- a/llvm/lib/Target/DirectX/DXILTranslateMetadata.cpp +++ b/llvm/lib/Target/DirectX/DXILTranslateMetadata.cpp @@ -7,8 +7,10 @@ //===----------------------------------------------------------------------===// #include "DXILTranslateMetadata.h" +#include "DXILRootSignature.h" #include "DXILShaderFlags.h" #include "DirectX.h" +#include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Twine.h" #include "llvm/Analysis/DXILMetadataAnalysis.h" @@ -204,9 +206,9 @@ getEntryPropAsMetadata(const EntryProperties &EP, uint64_t EntryShaderFlags, return MDNode::get(Ctx, MDVals); } -MDTuple *constructEntryMetadata(const Function *EntryFn, MDTuple *Signatures, - MDNode *Resources, MDTuple *Properties, - LLVMContext &Ctx) { +static MDTuple *constructEntryMetadata(const Function *EntryFn, + MDTuple *Signatures, MDNode *Resources, + MDTuple *Properties, LLVMContext &Ctx) { // Each entry point metadata record specifies: // * reference to the entry point function global symbol // * unmangled name @@ -290,42 +292,82 @@ static MDTuple *emitTopLevelLibraryNode(Module &M, MDNode *RMD, return constructEntryMetadata(nullptr, nullptr, RMD, Properties, Ctx); } -// TODO: We might need to refactor this to be more generic, -// in case we need more metadata to be replaced. -static void translateBranchMetadata(Module &M) { - for (Function &F : M) { - for (BasicBlock &BB : F) { - Instruction *BBTerminatorInst = BB.getTerminator(); +static void translateBranchMetadata(Module &M, Instruction *BBTerminatorInst) { + MDNode *HlslControlFlowMD = + BBTerminatorInst->getMetadata("hlsl.controlflow.hint"); + + if (!HlslControlFlowMD) + return; - MDNode *HlslControlFlowMD = - BBTerminatorInst->getMetadata("hlsl.controlflow.hint"); + assert(HlslControlFlowMD->getNumOperands() == 2 && + "invalid operands for hlsl.controlflow.hint"); - if (!HlslControlFlowMD) - continue; + MDBuilder MDHelper(M.getContext()); - assert(HlslControlFlowMD->getNumOperands() == 2 && - "invalid operands for hlsl.controlflow.hint"); + llvm::Metadata *HintsStr = MDHelper.createString("dx.controlflow.hints"); + llvm::Metadata *HintsValue = MDHelper.createConstant( + mdconst::extract<ConstantInt>(HlslControlFlowMD->getOperand(1))); - MDBuilder MDHelper(M.getContext()); - ConstantInt *Op1 = - mdconst::extract<ConstantInt>(HlslControlFlowMD->getOperand(1)); + MDNode *MDNode = llvm::MDNode::get(M.getContext(), {HintsStr, HintsValue}); - SmallVector<llvm::Metadata *, 2> Vals( - ArrayRef<Metadata *>{MDHelper.createString("dx.controlflow.hints"), - MDHelper.createConstant(Op1)}); + BBTerminatorInst->setMetadata("dx.controlflow.hints", MDNode); + BBTerminatorInst->setMetadata("hlsl.controlflow.hint", nullptr); +} + +static std::array<unsigned, 6> getCompatibleInstructionMDs(llvm::Module &M) { + return { + M.getMDKindID("dx.nonuniform"), M.getMDKindID("dx.controlflow.hints"), + M.getMDKindID("dx.precise"), llvm::LLVMContext::MD_range, + llvm::LLVMContext::MD_alias_scope, llvm::LLVMContext::MD_noalias}; +} - MDNode *MDNode = llvm::MDNode::get(M.getContext(), Vals); +static void translateInstructionMetadata(Module &M) { + // construct allowlist of valid metadata node kinds + std::array<unsigned, 6> DXILCompatibleMDs = getCompatibleInstructionMDs(M); - BBTerminatorInst->setMetadata("dx.controlflow.hints", MDNode); - BBTerminatorInst->setMetadata("hlsl.controlflow.hint", nullptr); + for (Function &F : M) { + for (BasicBlock &BB : F) { + // This needs to be done first so that "hlsl.controlflow.hints" isn't + // removed in the whitelist below + if (auto *I = BB.getTerminator()) + translateBranchMetadata(M, I); + + for (auto &I : make_early_inc_range(BB)) { + I.dropUnknownNonDebugMetadata(DXILCompatibleMDs); + } } } } -static void translateMetadata(Module &M, DXILResourceMap &DRM, - DXILResourceTypeMap &DRTM, - const ModuleShaderFlags &ShaderFlags, - const ModuleMetadataInfo &MMDI) { +static void cleanModuleFlags(Module &M) { + NamedMDNode *MDFlags = M.getModuleFlagsMetadata(); + if (!MDFlags) + return; + + SmallVector<llvm::Module::ModuleFlagEntry> FlagEntries; + M.getModuleFlagsMetadata(FlagEntries); + bool Updated = false; + for (auto &Flag : FlagEntries) { + // llvm 3.7 only supports behavior up to AppendUnique. + if (Flag.Behavior <= Module::ModFlagBehavior::AppendUnique) + continue; + Flag.Behavior = Module::ModFlagBehavior::Warning; + Updated = true; + } + + if (!Updated) + return; + + MDFlags->eraseFromParent(); + + for (auto &Flag : FlagEntries) + M.addModuleFlag(Flag.Behavior, Flag.Key->getString(), Flag.Val); +} + +static void translateGlobalMetadata(Module &M, DXILResourceMap &DRM, + DXILResourceTypeMap &DRTM, + const ModuleShaderFlags &ShaderFlags, + const ModuleMetadataInfo &MMDI) { LLVMContext &Ctx = M.getContext(); IRBuilder<> IRB(Ctx); SmallVector<MDNode *> EntryFnMDNodes; @@ -381,6 +423,22 @@ static void translateMetadata(Module &M, DXILResourceMap &DRM, M.getOrInsertNamedMetadata("dx.entryPoints"); for (auto *Entry : EntryFnMDNodes) EntryPointsNamedMD->addOperand(Entry); + + cleanModuleFlags(M); + + // dx.rootsignatures will have been parsed from its metadata form as its + // binary form as part of the RootSignatureAnalysisWrapper, so safely + // remove it as it is not recognized in DXIL + if (NamedMDNode *RootSignature = M.getNamedMetadata("dx.rootsignatures")) + RootSignature->eraseFromParent(); + + // llvm.errno.tbaa was recently added but is not supported in LLVM 3.7 and + // causes all tests using the DXIL Validator to fail. + // + // This is a temporary fix and should be replaced with a allowlist once + // we have determined all metadata that the DXIL Validator allows + if (NamedMDNode *ErrNo = M.getNamedMetadata("llvm.errno.tbaa")) + ErrNo->eraseFromParent(); } PreservedAnalyses DXILTranslateMetadata::run(Module &M, @@ -390,8 +448,8 @@ PreservedAnalyses DXILTranslateMetadata::run(Module &M, const ModuleShaderFlags &ShaderFlags = MAM.getResult<ShaderFlagsAnalysis>(M); const dxil::ModuleMetadataInfo MMDI = MAM.getResult<DXILMetadataAnalysis>(M); - translateMetadata(M, DRM, DRTM, ShaderFlags, MMDI); - translateBranchMetadata(M); + translateGlobalMetadata(M, DRM, DRTM, ShaderFlags, MMDI); + translateInstructionMetadata(M); return PreservedAnalyses::all(); } @@ -409,10 +467,13 @@ public: AU.addRequired<DXILResourceWrapperPass>(); AU.addRequired<ShaderFlagsAnalysisWrapper>(); AU.addRequired<DXILMetadataAnalysisWrapperPass>(); - AU.addPreserved<DXILResourceWrapperPass>(); + AU.addRequired<RootSignatureAnalysisWrapper>(); + AU.addPreserved<DXILMetadataAnalysisWrapperPass>(); - AU.addPreserved<ShaderFlagsAnalysisWrapper>(); AU.addPreserved<DXILResourceBindingWrapperPass>(); + AU.addPreserved<DXILResourceWrapperPass>(); + AU.addPreserved<RootSignatureAnalysisWrapper>(); + AU.addPreserved<ShaderFlagsAnalysisWrapper>(); } bool runOnModule(Module &M) override { @@ -425,8 +486,8 @@ public: dxil::ModuleMetadataInfo MMDI = getAnalysis<DXILMetadataAnalysisWrapperPass>().getModuleMetadata(); - translateMetadata(M, DRM, DRTM, ShaderFlags, MMDI); - translateBranchMetadata(M); + translateGlobalMetadata(M, DRM, DRTM, ShaderFlags, MMDI); + translateInstructionMetadata(M); return true; } }; @@ -443,6 +504,7 @@ INITIALIZE_PASS_BEGIN(DXILTranslateMetadataLegacy, "dxil-translate-metadata", "DXIL Translate Metadata", false, false) INITIALIZE_PASS_DEPENDENCY(DXILResourceWrapperPass) INITIALIZE_PASS_DEPENDENCY(ShaderFlagsAnalysisWrapper) +INITIALIZE_PASS_DEPENDENCY(RootSignatureAnalysisWrapper) INITIALIZE_PASS_DEPENDENCY(DXILMetadataAnalysisWrapperPass) INITIALIZE_PASS_END(DXILTranslateMetadataLegacy, "dxil-translate-metadata", "DXIL Translate Metadata", false, false) diff --git a/llvm/lib/Target/DirectX/DXILTranslateMetadata.h b/llvm/lib/Target/DirectX/DXILTranslateMetadata.h index f3f5eb1..4c1ffac 100644 --- a/llvm/lib/Target/DirectX/DXILTranslateMetadata.h +++ b/llvm/lib/Target/DirectX/DXILTranslateMetadata.h @@ -13,7 +13,8 @@ namespace llvm { -/// A pass that transforms DXIL Intrinsics that don't have DXIL opCodes +/// A pass that transforms LLVM Metadata in the module to it's DXIL equivalent, +/// then emits all recognized DXIL Metadata class DXILTranslateMetadata : public PassInfoMixin<DXILTranslateMetadata> { public: PreservedAnalyses run(Module &M, ModuleAnalysisManager &); diff --git a/llvm/lib/Target/Hexagon/Hexagon.td b/llvm/lib/Target/Hexagon/Hexagon.td index fb0928b8..ede8463 100644 --- a/llvm/lib/Target/Hexagon/Hexagon.td +++ b/llvm/lib/Target/Hexagon/Hexagon.td @@ -79,6 +79,12 @@ def ExtensionHVXV79: SubtargetFeature<"hvxv79", "HexagonHVXVersion", ExtensionHVXV67, ExtensionHVXV68, ExtensionHVXV69, ExtensionHVXV71, ExtensionHVXV73, ExtensionHVXV75]>; +def ExtensionHVXV81: SubtargetFeature<"hvxv81", "HexagonHVXVersion", + "Hexagon::ArchEnum::V81", "Hexagon HVX instructions", + [ExtensionHVXV65, ExtensionHVXV66, ExtensionHVXV67, + ExtensionHVXV68, ExtensionHVXV69, ExtensionHVXV71, + ExtensionHVXV73, ExtensionHVXV75, ExtensionHVXV79]>; + def ExtensionHVX64B: SubtargetFeature<"hvx-length64b", "UseHVX64BOps", "true", "Hexagon HVX 64B instructions", [ExtensionHVX]>; def ExtensionHVX128B: SubtargetFeature<"hvx-length128b", "UseHVX128BOps", @@ -151,6 +157,8 @@ def UseHVXV75 : Predicate<"HST->useHVXV75Ops()">, AssemblerPredicate<(all_of ExtensionHVXV75)>; def UseHVXV79 : Predicate<"HST->useHVXV79Ops()">, AssemblerPredicate<(all_of ExtensionHVXV79)>; +def UseHVXV81 : Predicate<"HST->useHVXV81Ops()">, + AssemblerPredicate<(all_of ExtensionHVXV81)>; def UseAudio : Predicate<"HST->useAudioOps()">, AssemblerPredicate<(all_of ExtensionAudio)>; def UseZReg : Predicate<"HST->useZRegOps()">, @@ -488,6 +496,11 @@ def : Proc<"hexagonv79", HexagonModelV79, ArchV68, ArchV69, ArchV71, ArchV73, ArchV75, ArchV79, FeatureCompound, FeatureDuplex, FeatureMemNoShuf, FeatureMemops, FeatureNVJ, FeatureNVS, FeaturePackets, FeatureSmallData]>; +def : Proc<"hexagonv81", HexagonModelV81, + [ArchV65, ArchV66, ArchV67, ArchV68, ArchV69, ArchV71, ArchV73, + ArchV75, ArchV79, ArchV81, + FeatureCompound, FeatureDuplex, FeatureMemNoShuf, FeatureMemops, + FeatureNVJ, FeatureNVS, FeaturePackets, FeatureSmallData]>; // Need to update the correct features for tiny core. // Disable NewValueJumps since the packetizer is unable to handle a packet with diff --git a/llvm/lib/Target/Hexagon/HexagonDepArch.h b/llvm/lib/Target/Hexagon/HexagonDepArch.h index 8984534..9bf4034 100644 --- a/llvm/lib/Target/Hexagon/HexagonDepArch.h +++ b/llvm/lib/Target/Hexagon/HexagonDepArch.h @@ -29,7 +29,8 @@ enum class ArchEnum { V71, V73, V75, - V79 + V79, + V81 }; inline std::optional<Hexagon::ArchEnum> getCpu(StringRef CPU) { @@ -50,6 +51,7 @@ inline std::optional<Hexagon::ArchEnum> getCpu(StringRef CPU) { .Case("hexagonv73", Hexagon::ArchEnum::V73) .Case("hexagonv75", Hexagon::ArchEnum::V75) .Case("hexagonv79", Hexagon::ArchEnum::V79) + .Case("hexagonv81", Hexagon::ArchEnum::V81) .Default(std::nullopt); } } // namespace Hexagon diff --git a/llvm/lib/Target/Hexagon/HexagonDepArch.td b/llvm/lib/Target/Hexagon/HexagonDepArch.td index 8ec1d93..f623fd0 100644 --- a/llvm/lib/Target/Hexagon/HexagonDepArch.td +++ b/llvm/lib/Target/Hexagon/HexagonDepArch.td @@ -34,3 +34,5 @@ def ArchV75: SubtargetFeature<"v75", "HexagonArchVersion", "Hexagon::ArchEnum::V def HasV75 : Predicate<"HST->hasV75Ops()">, AssemblerPredicate<(all_of ArchV75)>; def ArchV79: SubtargetFeature<"v79", "HexagonArchVersion", "Hexagon::ArchEnum::V79", "Enable Hexagon V79 architecture">; def HasV79 : Predicate<"HST->hasV79Ops()">, AssemblerPredicate<(all_of ArchV79)>; +def ArchV81: SubtargetFeature<"v81", "HexagonArchVersion", "Hexagon::ArchEnum::V81", "Enable Hexagon V81 architecture">; +def HasV81 : Predicate<"HST->hasV81Ops()">, AssemblerPredicate<(all_of ArchV81)>; diff --git a/llvm/lib/Target/Hexagon/HexagonDepIICHVX.td b/llvm/lib/Target/Hexagon/HexagonDepIICHVX.td index 93696e0..f4e36fa7 100644 --- a/llvm/lib/Target/Hexagon/HexagonDepIICHVX.td +++ b/llvm/lib/Target/Hexagon/HexagonDepIICHVX.td @@ -7222,3 +7222,595 @@ class DepHVXItinV79 { [Hex_FWD, Hex_FWD, HVX_FWD]> ]; } + +class DepHVXItinV81 { + list<InstrItinData> DepHVXItinV81_list = [ + InstrItinData <tc_0390c1ca, /*SLOT01,LOAD,VA,VX_DV*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_04da405a, /*SLOT0123,VP_VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLSHF]>], [9, 5], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_05ca8cfd, /*SLOT0123,VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_SHIFT]>], [9, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_08a4f1b6, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 7, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_0afc8be9, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_0b04c6c7, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_0ec46cf9, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 7], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_131f1c81, /*SLOT0,NOSLOT1,STORE,VP*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [SLOT1], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_XLANE]>], [2, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_1381a97c, /*SLOT0123,4SLOT*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL]>], [], + []>, + + InstrItinData <tc_15fdf750, /*SLOT23,VS_VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1], 0>, + InstrStage<1, [CVI_SHIFT, CVI_XLANE]>], [9, 7, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_16ff9ef8, /*SLOT0123,VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_SHIFT]>], [9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_191381c1, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [3, 7, 1, 2, 7], + [Hex_FWD, HVX_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_1ad8a370, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5, 2, 2], + [HVX_FWD, HVX_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_1ba8a0cd, /*SLOT01,LOAD,VA*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 3, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_20a4bbec, /*SLOT0,STORE*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST]>], [3, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_227864f7, /*SLOT0,STORE,VA,VX_DV*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE], 0>, + InstrStage<1, [CVI_MPY01]>], [3, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_257f6f7c, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 7, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_26a377fe, /*SLOT23,4SLOT_MPY*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL_NOMEM]>], [9, 3, 5, 2], + [HVX_FWD, Hex_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_2b4c548e, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_2c745bb8, /*SLOT0123,VP_VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLSHF]>], [9, 7, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_2d4051cd, /*SLOT23,4SLOT_MPY*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL_NOMEM]>], [9, 3, 7, 5, 2], + [HVX_FWD, Hex_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_2e8f5f6e, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 7, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_309dbb4f, /*SLOT0123,VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_SHIFT]>], [9, 7, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_37820f4c, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_3904b926, /*SLOT01,LOAD*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD]>], [9, 2, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3aacf4a8, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 2, 7], + [HVX_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_3ad719fb, /*SLOT01,ZW*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_ZW]>], [3, 2, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3c56e5ce, /*SLOT0,NOSLOT1,LOAD,VP*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 3, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3c8c15d0, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 5], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_3ce09744, /*SLOT0,STORE*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST]>], [1, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3e2aaafc, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [3, 1, 2, 7], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_447d9895, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [7, 1, 2, 7], + [HVX_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_453fe68d, /*SLOT01,LOAD,VA*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 3, 2, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_46d6c3e0, /*SLOT0123,VP*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_4942646a, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_51d0ecc3, /*SLOT0123,VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_SHIFT]>], [9, 5], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_52447ecc, /*SLOT01,LOAD*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD]>], [9, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_531b383c, /*SLOT0123*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [9, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_540c3da3, /*SLOT0,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [4, 7, 1], + [Hex_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_54a0dc47, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [3, 2, 1, 2, 7], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_561aaa58, /*SLOT0123,VP_VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLSHF]>], [9, 9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_56c4f9fe, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_56e64202, /*SLOT0123,VP*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_58d21193, /*SLOT0,STORE,VA_DV*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [7, 1, 2, 7, 7], + [HVX_FWD, Hex_FWD, Hex_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_5bf8afbb, /*SLOT0123,VP*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 2], + [HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_5cdf8c84, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_61bf7c03, /*SLOT23,4SLOT_MPY*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL_NOMEM]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_649072c2, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_660769f1, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 7, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_663c80a7, /*SLOT01,LOAD*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD]>], [9, 3, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6942b6e0, /*SLOT0,STORE*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST]>], [3, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_6e7fa133, /*SLOT0123,VP*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_7095ecba, /*SLOT01,LOAD,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [1, 2, 7], + [Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_71646d06, /*SLOT0123,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [9, 7, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_7177e272, /*SLOT0,STORE*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST]>], [2, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_718b5c53, /*SLOT0123,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [9], + [HVX_FWD]>, + + InstrItinData <tc_7273323b, /*SLOT0,STORE,VA_DV*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [1, 2, 7, 7], + [Hex_FWD, Hex_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_72e2b393, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_73efe966, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_7417e785, /*SLOT0123,VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_SHIFT]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_767c4e9d, /*SLOT0123,4SLOT*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL]>], [3, 2], + [HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_7d68d5c2, /*SLOT01,LOAD,VA*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [7, 1, 2, 7], + [HVX_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_7e6a3e89, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 9, 7, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_8772086c, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_87adc037, /*SLOT0123,VP_VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLSHF]>], [9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_8e420e4d, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [7, 1, 2, 7, 7], + [HVX_FWD, Hex_FWD, Hex_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_90bcc1db, /*SLOT2,VX_DV*/ + [InstrStage<1, [SLOT2], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_933f2b39, /*SLOT23,4SLOT_MPY*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL_NOMEM]>], [9, 7, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_946013d8, /*SLOT0123,VP*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 5], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_9a1cab75, /*SLOT01,LOAD,VA,VX_DV*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 3, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9aff7a2a, /*SLOT0,STORE,VA,VX_DV*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE], 0>, + InstrStage<1, [CVI_MPY01]>], [1, 2, 5], + [Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_9d1dc972, /*SLOT0123,VP_VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLSHF]>], [9, 7, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_9f363d21, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [1, 2, 7, 7], + [Hex_FWD, Hex_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_a02a10a8, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [2, 1, 2, 7], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_a0dbea28, /*SLOT01,ZW*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_ZW]>], [3, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a19b9305, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_a28f32b5, /*SLOT01,LOAD,VA*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [1, 2, 7], + [Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_a69eeee1, /*SLOT01,LOAD,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [7, 1, 2, 7], + [HVX_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_a7e6707d, /*SLOT0,NOSLOT1,LOAD,VP*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_XLANE]>], [9, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ab23f776, /*SLOT0,STORE*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST]>], [1, 2, 5], + [Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_abe8c3b2, /*SLOT01,LOAD,VA*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 2, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ac4046bc, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_af25efd9, /*SLOT0123,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [9, 2, 7, 7], + [HVX_FWD, Hex_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_b091f1c6, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_b28e51aa, /*SLOT0123,4SLOT*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_b4416217, /*SLOT0123,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [9, 7], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_b9db8205, /*SLOT01,LOAD*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD]>], [9, 3, 2, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_bb599486, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 7, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_c0749f3c, /*SLOT01,LOAD,VA*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_LD], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 1, 2], + [HVX_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_c127de3a, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_c4edf264, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 2], + [HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_c5dba46e, /*SLOT0,STORE,VA*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [1, 2, 7], + [Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_c7039829, /*SLOT0,NOSLOT1,STORE,VP*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [SLOT1], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_XLANE]>], [3, 2, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_cd94bfe0, /*SLOT23,VS_VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1], 0>, + InstrStage<1, [CVI_SHIFT, CVI_XLANE]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_cda936da, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_d8287c14, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_db5555f3, /*SLOT0123,VA_DV*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01, CVI_XLSHF]>], [9, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_dcca380f, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_dd5b0695, /*SLOT01,ZW*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_ZW]>], [2, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_df80eeb0, /*SLOT0123,VP_VS*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_XLSHF]>], [9, 7, 5, 5], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_e2d2e9e5, /*SLOT0,NOSLOT1,STORE,VP*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [SLOT1], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_XLANE]>], [3, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_e2fdd6e6, /*SLOT0123*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [9, 5], + [HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_e35c1e93, /*SLOT0123,VA*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1, CVI_SHIFT, CVI_XLANE]>], [9, 9, 7, 7], + [HVX_FWD, HVX_FWD, HVX_FWD, HVX_FWD]>, + + InstrItinData <tc_e3f68a46, /*SLOT0123,4SLOT*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_ALL]>], [3], + [HVX_FWD]>, + + InstrItinData <tc_e675c45a, /*SLOT23,VX_DV*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 7, 5, 2, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_e699ae41, /*SLOT01,ZW*/ + [InstrStage<1, [SLOT0, SLOT1], 0>, + InstrStage<1, [CVI_ZW]>], [1, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_e99d4c2e, /*SLOT0,STORE*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [CVI_ST]>], [3, 2, 1, 2, 5], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, HVX_FWD]>, + + InstrItinData <tc_f175e046, /*SLOT23,VX*/ + [InstrStage<1, [SLOT2, SLOT3], 0>, + InstrStage<1, [CVI_MPY0, CVI_MPY1]>], [9, 5, 5, 2], + [HVX_FWD, HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_f1de44ef, /*SLOT2,VX_DV*/ + [InstrStage<1, [SLOT2], 0>, + InstrStage<1, [CVI_MPY01]>], [9, 5, 2], + [HVX_FWD, HVX_FWD, Hex_FWD]>, + + InstrItinData <tc_f21e8abb, /*SLOT0,NOSLOT1,STORE,VP*/ + [InstrStage<1, [SLOT0], 0>, + InstrStage<1, [SLOT1], 0>, + InstrStage<1, [CVI_ST], 0>, + InstrStage<1, [CVI_XLANE]>], [1, 2, 5], + [Hex_FWD, Hex_FWD, HVX_FWD]> + ]; +}
\ No newline at end of file diff --git a/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td b/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td index 7a1ad3e..48b665c 100644 --- a/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td +++ b/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td @@ -13740,3 +13740,891 @@ class DepScalarItinV79 { [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]> ]; } + +class DepScalarItinV81 { + list<InstrItinData> DepScalarItinV81_list = [ + InstrItinData <tc_011e0e9d, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [2, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_01d44cb2, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_01e1be3b, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_02fe1c65, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_0655b949, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [2, 3], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_075c8dd8, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_0a195f2c, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 2, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_0a43be35, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_0a6c20ae, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [2, 1, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_0ba0d5da, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_0dfac0a7, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_0fac1eb8, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_112d30d6, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_1242dc2a, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_1248597c, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_139ef484, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [1, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_14ab4f41, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 3, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_151bf368, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_158aa3f7, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_197dce51, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 2, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_1981450d, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3], + [Hex_FWD]>, + + InstrItinData <tc_1c2c7a4a, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_1c7522a8, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 2, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_1d41f8b7, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 4, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_1fcb8495, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_1fe4ab69, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [2, 1, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_20131976, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_2237d952, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [1, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_23708a21, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [], + []>, + + InstrItinData <tc_2471c1c8, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_24e109c7, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 3, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_24f426ab, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_27106296, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_280f7fe1, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_28e55c6f, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [1, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_2c13e7f5, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_2c3e17fc, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_2f573607, /*tc_1*/ + [InstrStage<1, [SLOT2]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_33e7e673, /*tc_2early*/ + [InstrStage<1, [SLOT2]>], [], + []>, + + InstrItinData <tc_362b0be2, /*tc_3*/ + [InstrStage<1, [SLOT2]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_38382228, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_388f9897, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_38e0bae9, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 4, 2, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3d14a17b, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3edca78f, /*tc_2*/ + [InstrStage<1, [SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_3fbf1042, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3], + [Hex_FWD]>, + + InstrItinData <tc_407e96f9, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_40d64c94, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_4222e6bf, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_42ff66ba, /*tc_1*/ + [InstrStage<1, [SLOT2]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_442395f3, /*tc_2latepred*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_449acf79, /*tc_latepredstaia*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 1, 2, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_44d5a428, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_44fffc58, /*tc_3*/ + [InstrStage<1, [SLOT2, SLOT3]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_45791fb8, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 2, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_45f9d1be, /*tc_2early*/ + [InstrStage<1, [SLOT2]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_46c18ecf, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_49fdfd4b, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_4a55d03c, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_4abdbdc6, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_4ac61d92, /*tc_2latepred*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 3, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_4bf903b0, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3], + [Hex_FWD]>, + + InstrItinData <tc_503ce0f3, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_512b1653, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_53c851ab, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [4, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_54f0cee2, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_5502c366, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_55255f2b, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [], + []>, + + InstrItinData <tc_556f6577, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_55a9a350, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 2, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_55b33fda, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_56a124a7, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_57a55b54, /*tc_1*/ + [InstrStage<1, [SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5944960d, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_59a7822c, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5a222e89, /*tc_2early*/ + [InstrStage<1, [SLOT2]>], [1, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5a4b5e58, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5b347363, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5ceb2f9e, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5da50c4b, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5deb5e47, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5e4cf0e8, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_5f2afaf7, /*tc_latepredldaia*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 4, 3, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_60e324ff, /*tc_1*/ + [InstrStage<1, [SLOT2]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_63567288, /*tc_2latepred*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4], + [Hex_FWD]>, + + InstrItinData <tc_64b00d8a, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_651cbe02, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_65279839, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_65cbd974, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_69bfb303, /*tc_3*/ + [InstrStage<1, [SLOT2, SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6aa823ab, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6ae3426b, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6d861a95, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [2, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6e20402a, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [2, 3], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6f42bc60, /*tc_3stall*/ + [InstrStage<1, [SLOT0]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6fb52018, /*tc_3stall*/ + [InstrStage<1, [SLOT0]>], [1, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_6fc5dbea, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_711c805f, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_713b66bf, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7401744f, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7476d766, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_74a42bda, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_759e57be, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_76bb5435, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 2, 1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7d6a2568, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_77f94a5e, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [], + []>, + + InstrItinData <tc_788b1d09, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_78f87ed3, /*tc_3stall*/ + [InstrStage<1, [SLOT0]>], [], + []>, + + InstrItinData <tc_7af3a37e, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 3], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7b9187d3, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7c28bd7e, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3], + [Hex_FWD]>, + + InstrItinData <tc_7c31e19a, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7c6d32e4, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7dc63b5c, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7f58404a, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [], + []>, + + InstrItinData <tc_7f7f45f5, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 5, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_7f8ae742, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_8035e91f, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [2, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_822c3c68, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_829d8a86, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3, 1, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_838c4d7a, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_84a7500d, /*tc_2*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_86173609, /*tc_2latepred*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 3, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_887d1bb7, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_8a6d0d94, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_8a825db2, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_8b5bd4f5, /*tc_2*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_8e82e8ca, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 1, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_8f36a2fd, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9124c04f, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_92240447, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_934753bb, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [3, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_937dd41c, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [], + []>, + + InstrItinData <tc_9406230a, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [2, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_95a33176, /*tc_2*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_95f43c5e, /*tc_3*/ + [InstrStage<1, [SLOT2]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_96ef76ef, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_975a4e54, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 3, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9783714b, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9b20a062, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9b34f5e0, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [], + []>, + + InstrItinData <tc_9b3c0462, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9bcfb2ee, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9c52f549, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9e27f2f9, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9e72dc89, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 2, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9edb7c77, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 2, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9edefe01, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 2, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_9f6cd987, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a08b630b, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a1297125, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a154b476, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a2b365d2, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a3070909, /*tc_3stall*/ + [InstrStage<1, [SLOT0]>], [1, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a32e03e7, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 2, 1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a38c45dc, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a4e22bbd, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a4ee89db, /*tc_2early*/ + [InstrStage<1, [SLOT0]>], [], + []>, + + InstrItinData <tc_a724463d, /*tc_3stall*/ + [InstrStage<1, [SLOT0]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a7a13fac, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a7bdb22c, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_a9edeffa, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_abfd9a6d, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ac65613f, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_addc37a8, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3, 1, 2, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ae5babd7, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_aee6250c, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_af6af259, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_b1ae5f67, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_b2196a3f, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [1, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_b3d46584, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [], + []>, + + InstrItinData <tc_b4dc7630, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 1, 2, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_b7c4062a, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 3, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_b837298f, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [], + []>, + + InstrItinData <tc_b9bec29e, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [], + []>, + + InstrItinData <tc_ba9255a6, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [2, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_bb07f2c5, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [3, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_bb78483e, /*tc_3stall*/ + [InstrStage<1, [SLOT3]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_bb831a7c, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_bf2ffc0f, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_c20701f0, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_c21d7447, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_c57d9f39, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_c818ff7f, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [], + []>, + + InstrItinData <tc_ce59038e, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [3, 2, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_cfa0e29b, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [2, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d03278fd, /*tc_st*/ + [InstrStage<1, [SLOT0, SLOT1]>], [2, 1, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d234b61a, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_d33e5eee, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d3632d88, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d45ba9cd, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_d57d649c, /*tc_3stall*/ + [InstrStage<1, [SLOT2]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_d61dfdc3, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d68dca5c, /*tc_3stall*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d71ea8fa, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [2, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_d7718fbe, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_db596beb, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_db96aa6b, /*tc_st*/ + [InstrStage<1, [SLOT0]>], [1], + [Hex_FWD]>, + + InstrItinData <tc_dc51281d, /*tc_3*/ + [InstrStage<1, [SLOT2]>], [2, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_decdde8a, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_df5d53f9, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 2, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_e3d699e3, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_e60def48, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_e9170fb7, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ed03645c, /*tc_1*/ + [InstrStage<1, [SLOT2]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ed3f8d2a, /*tc_ld*/ + [InstrStage<1, [SLOT0]>], [4, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_eed07714, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_eeda4109, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_ef921005, /*tc_1*/ + [InstrStage<1, [SLOT2, SLOT3]>], [3, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f098b237, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f0cdeccf, /*tc_3x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 1, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f0e8e832, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f34c1c21, /*tc_2*/ + [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f38f92e1, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_f529831b, /*tc_latepredstaia*/ + [InstrStage<1, [SLOT0]>], [4, 3, 1, 2, 3], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f6e2aff9, /*tc_newvjump*/ + [InstrStage<1, [SLOT0]>], [3, 2, 2], + [Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f7569068, /*tc_4x*/ + [InstrStage<1, [SLOT2, SLOT3]>], [5, 5, 1, 1], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_f97707c1, /*tc_1*/ + [InstrStage<1, [SLOT2]>], [2], + [Hex_FWD]>, + + InstrItinData <tc_f999c66e, /*tc_1*/ + [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [2, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_fae9dfa5, /*tc_3x*/ + [InstrStage<1, [SLOT3]>], [4, 2], + [Hex_FWD, Hex_FWD]>, + + InstrItinData <tc_fedb7e19, /*tc_ld*/ + [InstrStage<1, [SLOT0, SLOT1]>], [4, 2, 1, 2], + [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]> + ]; +}
\ No newline at end of file diff --git a/llvm/lib/Target/Hexagon/HexagonDepInstrInfo.td b/llvm/lib/Target/Hexagon/HexagonDepInstrInfo.td index ae96753..f8f1c2a 100644 --- a/llvm/lib/Target/Hexagon/HexagonDepInstrInfo.td +++ b/llvm/lib/Target/Hexagon/HexagonDepInstrInfo.td @@ -39178,6 +39178,19 @@ let opNewValue = 0; let isCVI = 1; let DecoderNamespace = "EXT_mmvec"; } +def V6_vsub_hf_mix : HInst< +(outs HvxVR:$Vd32), +(ins HvxVR:$Vu32, HvxVR:$Vv32), +"$Vd32.qf16 = vsub($Vu32.hf,$Vv32.qf16)", +tc_05ca8cfd, TypeCVI_VS>, Enc_45364e, Requires<[UseHVXV81,UseHVXQFloat]> { +let Inst{7-5} = 0b100; +let Inst{13-13} = 0b1; +let Inst{31-21} = 0b00011010000; +let hasNewValue = 1; +let opNewValue = 0; +let isCVI = 1; +let DecoderNamespace = "EXT_mmvec"; +} def V6_vsub_qf16 : HInst< (outs HvxVR:$Vd32), (ins HvxVR:$Vu32, HvxVR:$Vv32), @@ -39269,6 +39282,19 @@ let opNewValue = 0; let isCVI = 1; let DecoderNamespace = "EXT_mmvec"; } +def V6_vsub_sf_mix : HInst< +(outs HvxVR:$Vd32), +(ins HvxVR:$Vu32, HvxVR:$Vv32), +"$Vd32.qf32 = vsub($Vu32.sf,$Vv32.qf32)", +tc_05ca8cfd, TypeCVI_VS>, Enc_45364e, Requires<[UseHVXV81,UseHVXQFloat]> { +let Inst{7-5} = 0b000; +let Inst{13-13} = 0b1; +let Inst{31-21} = 0b00011010000; +let hasNewValue = 1; +let opNewValue = 0; +let isCVI = 1; +let DecoderNamespace = "EXT_mmvec"; +} def V6_vsub_sf_sf : HInst< (outs HvxVR:$Vd32), (ins HvxVR:$Vu32, HvxVR:$Vv32), @@ -41116,6 +41142,17 @@ let hasNewValue = 1; let opNewValue = 0; let isSolo = 1; } +def Y2_tlbpp : HInst< +(outs IntRegs:$Rd32), +(ins DoubleRegs:$Rss32), +"$Rd32 = tlbp($Rss32)", +tc_6aa823ab, TypeCR>, Enc_90cd8b, Requires<[HasV81]> { +let Inst{13-5} = 0b000000000; +let Inst{31-21} = 0b01101100011; +let hasNewValue = 1; +let opNewValue = 0; +let isSolo = 1; +} def Y2_tlbr : HInst< (outs DoubleRegs:$Rdd32), (ins IntRegs:$Rs32), diff --git a/llvm/lib/Target/Hexagon/HexagonDepMapAsm2Intrin.td b/llvm/lib/Target/Hexagon/HexagonDepMapAsm2Intrin.td index 17cb96c..23f4b3a 100644 --- a/llvm/lib/Target/Hexagon/HexagonDepMapAsm2Intrin.td +++ b/llvm/lib/Target/Hexagon/HexagonDepMapAsm2Intrin.td @@ -3827,3 +3827,14 @@ def: Pat<(int_hexagon_V6_vsub_hf_f8 HvxVR:$src1, HvxVR:$src2), (V6_vsub_hf_f8 HvxVR:$src1, HvxVR:$src2)>, Requires<[HasV79, UseHVX64B]>; def: Pat<(int_hexagon_V6_vsub_hf_f8_128B HvxVR:$src1, HvxVR:$src2), (V6_vsub_hf_f8 HvxVR:$src1, HvxVR:$src2)>, Requires<[HasV79, UseHVX128B]>; + +// V81 HVX Instructions. + +def: Pat<(int_hexagon_V6_vsub_hf_mix HvxVR:$src1, HvxVR:$src2), + (V6_vsub_hf_mix HvxVR:$src1, HvxVR:$src2)>, Requires<[UseHVXV81, UseHVX64B, UseHVXQFloat]>; +def: Pat<(int_hexagon_V6_vsub_hf_mix_128B HvxVR:$src1, HvxVR:$src2), + (V6_vsub_hf_mix HvxVR:$src1, HvxVR:$src2)>, Requires<[UseHVXV81, UseHVX128B, UseHVXQFloat]>; +def: Pat<(int_hexagon_V6_vsub_sf_mix HvxVR:$src1, HvxVR:$src2), + (V6_vsub_sf_mix HvxVR:$src1, HvxVR:$src2)>, Requires<[UseHVXV81, UseHVX64B, UseHVXQFloat]>; +def: Pat<(int_hexagon_V6_vsub_sf_mix_128B HvxVR:$src1, HvxVR:$src2), + (V6_vsub_sf_mix HvxVR:$src1, HvxVR:$src2)>, Requires<[UseHVXV81, UseHVX128B, UseHVXQFloat]>; diff --git a/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp b/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp index e285e04..7ee280d 100644 --- a/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp +++ b/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp @@ -654,7 +654,9 @@ void HexagonDAGToDAGISel::SelectIntrinsicWChain(SDNode *N) { IntNo == Intrinsic::hexagon_V6_vgathermh || IntNo == Intrinsic::hexagon_V6_vgathermh_128B || IntNo == Intrinsic::hexagon_V6_vgathermhw || - IntNo == Intrinsic::hexagon_V6_vgathermhw_128B) { + IntNo == Intrinsic::hexagon_V6_vgathermhw_128B || + IntNo == Intrinsic::hexagon_V6_vgather_vscattermh || + IntNo == Intrinsic::hexagon_V6_vgather_vscattermh_128B) { SelectV65Gather(N); return; } diff --git a/llvm/lib/Target/Hexagon/HexagonISelDAGToDAGHVX.cpp b/llvm/lib/Target/Hexagon/HexagonISelDAGToDAGHVX.cpp index c7a4f68..3cc146b 100644 --- a/llvm/lib/Target/Hexagon/HexagonISelDAGToDAGHVX.cpp +++ b/llvm/lib/Target/Hexagon/HexagonISelDAGToDAGHVX.cpp @@ -2953,6 +2953,10 @@ void HexagonDAGToDAGISel::SelectV65Gather(SDNode *N) { case Intrinsic::hexagon_V6_vgathermhw_128B: Opcode = Hexagon::V6_vgathermhw_pseudo; break; + case Intrinsic::hexagon_V6_vgather_vscattermh: + case Intrinsic::hexagon_V6_vgather_vscattermh_128B: + Opcode = Hexagon::V6_vgather_vscatter_mh_pseudo; + break; } SDVTList VTs = CurDAG->getVTList(MVT::Other); diff --git a/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp b/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp index 9f7f434..526b4de 100644 --- a/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp +++ b/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp @@ -2145,7 +2145,9 @@ bool HexagonTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, case Intrinsic::hexagon_V6_vgathermhq: case Intrinsic::hexagon_V6_vgathermhq_128B: case Intrinsic::hexagon_V6_vgathermhwq: - case Intrinsic::hexagon_V6_vgathermhwq_128B: { + case Intrinsic::hexagon_V6_vgathermhwq_128B: + case Intrinsic::hexagon_V6_vgather_vscattermh: + case Intrinsic::hexagon_V6_vgather_vscattermh_128B: { const Module &M = *I.getParent()->getParent()->getParent(); Info.opc = ISD::INTRINSIC_W_CHAIN; Type *VecTy = I.getArgOperand(1)->getType(); diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp b/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp index 939841a..47726d6 100644 --- a/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp +++ b/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp @@ -1554,80 +1554,93 @@ HexagonInstrInfo::expandVGatherPseudo(MachineInstr &MI) const { MachineBasicBlock::iterator First; switch (Opc) { - case Hexagon::V6_vgathermh_pseudo: - First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermh)) - .add(MI.getOperand(2)) - .add(MI.getOperand(3)) - .add(MI.getOperand(4)); - BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) - .add(MI.getOperand(0)) - .addImm(MI.getOperand(1).getImm()) - .addReg(Hexagon::VTMP); - MBB.erase(MI); - return First.getInstrIterator(); - - case Hexagon::V6_vgathermw_pseudo: - First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermw)) - .add(MI.getOperand(2)) - .add(MI.getOperand(3)) - .add(MI.getOperand(4)); - BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) - .add(MI.getOperand(0)) - .addImm(MI.getOperand(1).getImm()) - .addReg(Hexagon::VTMP); - MBB.erase(MI); - return First.getInstrIterator(); - - case Hexagon::V6_vgathermhw_pseudo: - First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhw)) - .add(MI.getOperand(2)) - .add(MI.getOperand(3)) - .add(MI.getOperand(4)); - BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) - .add(MI.getOperand(0)) - .addImm(MI.getOperand(1).getImm()) - .addReg(Hexagon::VTMP); - MBB.erase(MI); - return First.getInstrIterator(); - - case Hexagon::V6_vgathermhq_pseudo: - First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhq)) - .add(MI.getOperand(2)) - .add(MI.getOperand(3)) - .add(MI.getOperand(4)) - .add(MI.getOperand(5)); - BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) - .add(MI.getOperand(0)) - .addImm(MI.getOperand(1).getImm()) - .addReg(Hexagon::VTMP); - MBB.erase(MI); - return First.getInstrIterator(); - - case Hexagon::V6_vgathermwq_pseudo: - First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermwq)) - .add(MI.getOperand(2)) - .add(MI.getOperand(3)) - .add(MI.getOperand(4)) - .add(MI.getOperand(5)); - BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) - .add(MI.getOperand(0)) - .addImm(MI.getOperand(1).getImm()) - .addReg(Hexagon::VTMP); - MBB.erase(MI); - return First.getInstrIterator(); - - case Hexagon::V6_vgathermhwq_pseudo: - First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhwq)) - .add(MI.getOperand(2)) - .add(MI.getOperand(3)) - .add(MI.getOperand(4)) - .add(MI.getOperand(5)); - BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) - .add(MI.getOperand(0)) - .addImm(MI.getOperand(1).getImm()) - .addReg(Hexagon::VTMP); - MBB.erase(MI); - return First.getInstrIterator(); + case Hexagon::V6_vgather_vscatter_mh_pseudo: + // This is mainly a place holder. It will be extended. + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermh)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vscattermh)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); + case Hexagon::V6_vgathermh_pseudo: + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermh)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) + .add(MI.getOperand(0)) + .addImm(MI.getOperand(1).getImm()) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); + + case Hexagon::V6_vgathermw_pseudo: + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermw)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) + .add(MI.getOperand(0)) + .addImm(MI.getOperand(1).getImm()) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); + + case Hexagon::V6_vgathermhw_pseudo: + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhw)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) + .add(MI.getOperand(0)) + .addImm(MI.getOperand(1).getImm()) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); + + case Hexagon::V6_vgathermhq_pseudo: + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhq)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)) + .add(MI.getOperand(5)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) + .add(MI.getOperand(0)) + .addImm(MI.getOperand(1).getImm()) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); + + case Hexagon::V6_vgathermwq_pseudo: + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermwq)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)) + .add(MI.getOperand(5)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) + .add(MI.getOperand(0)) + .addImm(MI.getOperand(1).getImm()) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); + + case Hexagon::V6_vgathermhwq_pseudo: + First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhwq)) + .add(MI.getOperand(2)) + .add(MI.getOperand(3)) + .add(MI.getOperand(4)) + .add(MI.getOperand(5)); + BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) + .add(MI.getOperand(0)) + .addImm(MI.getOperand(1).getImm()) + .addReg(Hexagon::VTMP); + MBB.erase(MI); + return First.getInstrIterator(); } return MI.getIterator(); @@ -2806,6 +2819,7 @@ bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset, case Hexagon::V6_vL32b_nt_tmp_npred_ai: case Hexagon::V6_vS32Ub_npred_ai: case Hexagon::V6_vgathermh_pseudo: + case Hexagon::V6_vgather_vscatter_mh_pseudo: case Hexagon::V6_vgathermw_pseudo: case Hexagon::V6_vgathermhw_pseudo: case Hexagon::V6_vgathermhq_pseudo: diff --git a/llvm/lib/Target/Hexagon/HexagonPatternsV65.td b/llvm/lib/Target/Hexagon/HexagonPatternsV65.td index f927f9b..42393d0 100644 --- a/llvm/lib/Target/Hexagon/HexagonPatternsV65.td +++ b/llvm/lib/Target/Hexagon/HexagonPatternsV65.td @@ -40,6 +40,19 @@ defm V6_vgathermh_pseudo : vgathermh<HvxVR>; defm V6_vgathermw_pseudo : vgathermw<HvxVR>; defm V6_vgathermhw_pseudo : vgathermhw<HvxWR>; + +multiclass vgather_scatter_mh<RegisterClass RC> { + let isCodeGenOnly = 1, isPseudo = 1, mayLoad = 1, + mayStore = 1, addrMode = BaseImmOffset, accessSize = HalfWordAccess in + def NAME : CVI_GATHER_TMP_LD_Resource_NoOpcode<(outs ), + (ins IntRegs:$_dst_, s4_0Imm:$Ii, + IntRegs:$Rt, ModRegs:$Mu, RC:$Vv), + ".error \"should not emit\" ", + []>; +} + +defm V6_vgather_vscatter_mh_pseudo : vgather_scatter_mh<HvxVR>; + multiclass vgathermhq<RegisterClass RC1, RegisterClass RC2> { let isCodeGenOnly = 1, isPseudo = 1, mayLoad = 1, mayStore = 1, addrMode = BaseImmOffset, accessSize = HalfWordAccess in diff --git a/llvm/lib/Target/Hexagon/HexagonSchedule.td b/llvm/lib/Target/Hexagon/HexagonSchedule.td index b8a9cf3..9bcd4bf 100644 --- a/llvm/lib/Target/Hexagon/HexagonSchedule.td +++ b/llvm/lib/Target/Hexagon/HexagonSchedule.td @@ -75,3 +75,4 @@ include "HexagonScheduleV71T.td" include "HexagonScheduleV73.td" include "HexagonScheduleV75.td" include "HexagonScheduleV79.td" +include "HexagonScheduleV81.td"
\ No newline at end of file diff --git a/llvm/lib/Target/Hexagon/HexagonScheduleV81.td b/llvm/lib/Target/Hexagon/HexagonScheduleV81.td new file mode 100644 index 0000000..dd5f5a0 --- /dev/null +++ b/llvm/lib/Target/Hexagon/HexagonScheduleV81.td @@ -0,0 +1,31 @@ +//=-HexagonScheduleV81.td - HexagonV81 Scheduling Definitions *- tablegen -*-=// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// + +def HexagonV81ItinList : DepScalarItinV81, ScalarItin, + DepHVXItinV81, HVXItin, PseudoItin { + list<InstrItinData> ItinList = + !listconcat(DepScalarItinV81_list, ScalarItin_list, + DepHVXItinV81_list, HVXItin_list, PseudoItin_list); +} + +def HexagonItinerariesV81 : + ProcessorItineraries<[SLOT0, SLOT1, SLOT2, SLOT3, SLOT_ENDLOOP, + CVI_ST, CVI_XLANE, CVI_SHIFT, CVI_MPY0, CVI_MPY1, + CVI_LD, CVI_XLSHF, CVI_MPY01, CVI_ALL, + CVI_ALL_NOMEM, CVI_ZW], + [Hex_FWD, HVX_FWD], + HexagonV81ItinList.ItinList>; + +def HexagonModelV81 : SchedMachineModel { + // Max issue per cycle == bundle width. + let IssueWidth = 4; + let Itineraries = HexagonItinerariesV81; + let LoadLatency = 1; + let CompleteModel = 0; +} diff --git a/llvm/lib/Target/Hexagon/HexagonSubtarget.h b/llvm/lib/Target/Hexagon/HexagonSubtarget.h index 7430567..995f66d 100644 --- a/llvm/lib/Target/Hexagon/HexagonSubtarget.h +++ b/llvm/lib/Target/Hexagon/HexagonSubtarget.h @@ -224,6 +224,15 @@ public: bool useHVXV79Ops() const { return HexagonHVXVersion >= Hexagon::ArchEnum::V79; } + bool hasV81Ops() const { + return getHexagonArchVersion() >= Hexagon::ArchEnum::V81; + } + bool hasV81OpsOnly() const { + return getHexagonArchVersion() == Hexagon::ArchEnum::V81; + } + bool useHVXV81Ops() const { + return HexagonHVXVersion >= Hexagon::ArchEnum::V81; + } bool useAudioOps() const { return UseAudioOps; } bool useCompound() const { return UseCompound; } diff --git a/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp b/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp index 171e294..e925e04 100644 --- a/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp +++ b/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp @@ -31,6 +31,10 @@ using namespace llvm; static cl::opt<bool> HexagonAutoHVX("hexagon-autohvx", cl::init(false), cl::Hidden, cl::desc("Enable loop vectorizer for HVX")); +cl::opt<bool> HexagonAllowScatterGatherHVX( + "hexagon-allow-scatter-gather-hvx", cl::init(false), cl::Hidden, + cl::desc("Allow auto-generation of HVX scatter-gather")); + static cl::opt<bool> EnableV68FloatAutoHVX( "force-hvx-float", cl::Hidden, cl::desc("Enable auto-vectorization of floatint point types on v68.")); @@ -354,6 +358,61 @@ bool HexagonTTIImpl::isLegalMaskedLoad(Type *DataType, Align /*Alignment*/, return HexagonMaskedVMem && ST.isTypeForHVX(DataType); } +bool HexagonTTIImpl::isLegalMaskedGather(Type *Ty, Align Alignment) const { + // For now assume we can not deal with all HVX datatypes. + if (!Ty->isVectorTy() || !ST.isTypeForHVX(Ty) || + !HexagonAllowScatterGatherHVX) + return false; + // This must be in sync with HexagonVectorCombine pass. + switch (Ty->getScalarSizeInBits()) { + case 8: + return (getTypeNumElements(Ty) == 128); + case 16: + if (getTypeNumElements(Ty) == 64 || getTypeNumElements(Ty) == 32) + return (Alignment >= 2); + break; + case 32: + if (getTypeNumElements(Ty) == 32) + return (Alignment >= 4); + break; + default: + break; + } + return false; +} + +bool HexagonTTIImpl::isLegalMaskedScatter(Type *Ty, Align Alignment) const { + if (!Ty->isVectorTy() || !ST.isTypeForHVX(Ty) || + !HexagonAllowScatterGatherHVX) + return false; + // This must be in sync with HexagonVectorCombine pass. + switch (Ty->getScalarSizeInBits()) { + case 8: + return (getTypeNumElements(Ty) == 128); + case 16: + if (getTypeNumElements(Ty) == 64) + return (Alignment >= 2); + break; + case 32: + if (getTypeNumElements(Ty) == 32) + return (Alignment >= 4); + break; + default: + break; + } + return false; +} + +bool HexagonTTIImpl::forceScalarizeMaskedGather(VectorType *VTy, + Align Alignment) const { + return !isLegalMaskedGather(VTy, Alignment); +} + +bool HexagonTTIImpl::forceScalarizeMaskedScatter(VectorType *VTy, + Align Alignment) const { + return !isLegalMaskedScatter(VTy, Alignment); +} + /// --- Vector TTI end --- unsigned HexagonTTIImpl::getPrefetchDistance() const { diff --git a/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.h b/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.h index dbf16c9..cec2bf9 100644 --- a/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.h +++ b/llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.h @@ -169,6 +169,12 @@ public: unsigned AddressSpace) const override; bool isLegalMaskedLoad(Type *DataType, Align Alignment, unsigned AddressSpace) const override; + bool isLegalMaskedGather(Type *Ty, Align Alignment) const override; + bool isLegalMaskedScatter(Type *Ty, Align Alignment) const override; + bool forceScalarizeMaskedGather(VectorType *VTy, + Align Alignment) const override; + bool forceScalarizeMaskedScatter(VectorType *VTy, + Align Alignment) const override; /// @} diff --git a/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp b/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp index 9ab5202..5c50ec2 100644 --- a/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp +++ b/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp @@ -57,6 +57,11 @@ #define DEBUG_TYPE "hexagon-vc" +// This is a const that represents default HVX VTCM page size. +// It is boot time configurable, so we probably want an API to +// read it, but for now assume 128KB +#define DEFAULT_HVX_VTCM_PAGE_SIZE 131072 + using namespace llvm; namespace { @@ -418,6 +423,18 @@ raw_ostream &operator<<(raw_ostream &OS, const AlignVectors::ByteSpan &BS) { class HvxIdioms { public: + enum DstQualifier { + Undefined = 0, + Arithmetic, + LdSt, + LLVM_Gather, + LLVM_Scatter, + HEX_Gather_Scatter, + HEX_Gather, + HEX_Scatter, + Call + }; + HvxIdioms(const HexagonVectorCombine &HVC_) : HVC(HVC_) { auto *Int32Ty = HVC.getIntTy(32); HvxI32Ty = HVC.getHvxTy(Int32Ty, /*Pair=*/false); @@ -473,6 +490,11 @@ private: auto createMulLong(IRBuilderBase &Builder, ArrayRef<Value *> WordX, Signedness SgnX, ArrayRef<Value *> WordY, Signedness SgnY) const -> SmallVector<Value *>; + // Vector manipulations for Ripple + bool matchScatter(Instruction &In) const; + bool matchGather(Instruction &In) const; + Value *processVScatter(Instruction &In) const; + Value *processVGather(Instruction &In) const; VectorType *HvxI32Ty; VectorType *HvxP32Ty; @@ -1545,7 +1567,7 @@ auto AlignVectors::isSectorTy(Type *Ty) const -> bool { } auto AlignVectors::run() -> bool { - LLVM_DEBUG(dbgs() << "Running HVC::AlignVectors on " << HVC.F.getName() + LLVM_DEBUG(dbgs() << "\nRunning HVC::AlignVectors on " << HVC.F.getName() << '\n'); if (!createAddressGroups()) return false; @@ -1797,6 +1819,846 @@ auto HvxIdioms::processFxpMul(Instruction &In, const FxpOp &Op) const return Ext; } +inline bool HvxIdioms::matchScatter(Instruction &In) const { + IntrinsicInst *II = dyn_cast<IntrinsicInst>(&In); + if (!II) + return false; + return (II->getIntrinsicID() == Intrinsic::masked_scatter); +} + +inline bool HvxIdioms::matchGather(Instruction &In) const { + IntrinsicInst *II = dyn_cast<IntrinsicInst>(&In); + if (!II) + return false; + return (II->getIntrinsicID() == Intrinsic::masked_gather); +} + +Instruction *locateDestination(Instruction *In, HvxIdioms::DstQualifier &Qual); + +// Binary instructions we want to handle as users of gather/scatter. +inline bool isArithmetic(unsigned Opc) { + switch (Opc) { + case Instruction::Add: + case Instruction::Sub: + case Instruction::Mul: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + case Instruction::AShr: + case Instruction::LShr: + case Instruction::Shl: + case Instruction::UDiv: + return true; + } + return false; +} + +// TODO: Maybe use MemoryLocation for this. See getLocOrNone above. +inline Value *getPointer(Value *Ptr) { + assert(Ptr && "Unable to extract pointer"); + if (isa<AllocaInst>(Ptr) || isa<Argument>(Ptr) || isa<GlobalValue>(Ptr)) + return Ptr; + if (isa<LoadInst>(Ptr) || isa<StoreInst>(Ptr)) + return getLoadStorePointerOperand(Ptr); + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Ptr)) { + if (II->getIntrinsicID() == Intrinsic::masked_store) + return II->getOperand(1); + } + return nullptr; +} + +static Instruction *selectDestination(Instruction *In, + HvxIdioms::DstQualifier &Qual) { + Instruction *Destination = nullptr; + if (!In) + return Destination; + if (isa<StoreInst>(In)) { + Destination = In; + Qual = HvxIdioms::LdSt; + } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(In)) { + if (II->getIntrinsicID() == Intrinsic::masked_gather) { + Destination = In; + Qual = HvxIdioms::LLVM_Gather; + } else if (II->getIntrinsicID() == Intrinsic::masked_scatter) { + Destination = In; + Qual = HvxIdioms::LLVM_Scatter; + } else if (II->getIntrinsicID() == Intrinsic::masked_store) { + Destination = In; + Qual = HvxIdioms::LdSt; + } else if (II->getIntrinsicID() == + Intrinsic::hexagon_V6_vgather_vscattermh) { + Destination = In; + Qual = HvxIdioms::HEX_Gather_Scatter; + } else if (II->getIntrinsicID() == Intrinsic::hexagon_V6_vscattermh_128B) { + Destination = In; + Qual = HvxIdioms::HEX_Scatter; + } else if (II->getIntrinsicID() == Intrinsic::hexagon_V6_vgathermh_128B) { + Destination = In; + Qual = HvxIdioms::HEX_Gather; + } + } else if (isa<ZExtInst>(In)) { + return locateDestination(In, Qual); + } else if (isa<CastInst>(In)) { + return locateDestination(In, Qual); + } else if (isa<CallInst>(In)) { + Destination = In; + Qual = HvxIdioms::Call; + } else if (isa<GetElementPtrInst>(In)) { + return locateDestination(In, Qual); + } else if (isArithmetic(In->getOpcode())) { + Destination = In; + Qual = HvxIdioms::Arithmetic; + } else { + LLVM_DEBUG(dbgs() << "Unhandled destination : " << *In << "\n"); + } + return Destination; +} + +// This method attempts to find destination (user) for a given intrinsic. +// Given that these are produced only by Ripple, the number of options is +// limited. Simplest case is explicit store which in fact is redundant (since +// HVX gater creates its own store during packetization). Nevertheless we need +// to figure address where we storing. Other cases are more complicated, but +// still few. +Instruction *locateDestination(Instruction *In, HvxIdioms::DstQualifier &Qual) { + Instruction *Destination = nullptr; + if (!In) + return Destination; + // Get all possible destinations + SmallVector<Instruction *> Users; + // Iterate over the uses of the instruction + for (auto &U : In->uses()) { + if (auto *UI = dyn_cast<Instruction>(U.getUser())) { + Destination = selectDestination(UI, Qual); + if (Destination) + Users.push_back(Destination); + } + } + // Now see which of the users (if any) is a memory destination. + for (auto *I : Users) + if (getPointer(I)) + return I; + return Destination; +} + +// The two intrinsics we handle here have GEP in a different position. +inline GetElementPtrInst *locateGepFromIntrinsic(Instruction *In) { + assert(In && "Bad instruction"); + IntrinsicInst *IIn = dyn_cast<IntrinsicInst>(In); + assert((IIn && (IIn->getIntrinsicID() == Intrinsic::masked_gather || + IIn->getIntrinsicID() == Intrinsic::masked_scatter)) && + "Not a gather Intrinsic"); + GetElementPtrInst *GEPIndex = nullptr; + if (IIn->getIntrinsicID() == Intrinsic::masked_gather) + GEPIndex = dyn_cast<GetElementPtrInst>(IIn->getOperand(0)); + else + GEPIndex = dyn_cast<GetElementPtrInst>(IIn->getOperand(1)); + return GEPIndex; +} + +// Given the intrinsic find its GEP argument and extract base address it uses. +// The method relies on the way how Ripple typically forms the GEP for +// scatter/gather. +static Value *locateAddressFromIntrinsic(Instruction *In) { + GetElementPtrInst *GEPIndex = locateGepFromIntrinsic(In); + if (!GEPIndex) { + LLVM_DEBUG(dbgs() << " No GEP in intrinsic\n"); + return nullptr; + } + Value *BaseAddress = GEPIndex->getPointerOperand(); + auto *IndexLoad = dyn_cast<LoadInst>(BaseAddress); + if (IndexLoad) + return IndexLoad; + + auto *IndexZEx = dyn_cast<ZExtInst>(BaseAddress); + if (IndexZEx) { + IndexLoad = dyn_cast<LoadInst>(IndexZEx->getOperand(0)); + if (IndexLoad) + return IndexLoad; + IntrinsicInst *II = dyn_cast<IntrinsicInst>(IndexZEx->getOperand(0)); + if (II && II->getIntrinsicID() == Intrinsic::masked_gather) + return locateAddressFromIntrinsic(II); + } + auto *BaseShuffle = dyn_cast<ShuffleVectorInst>(BaseAddress); + if (BaseShuffle) { + IndexLoad = dyn_cast<LoadInst>(BaseShuffle->getOperand(0)); + if (IndexLoad) + return IndexLoad; + auto *IE = dyn_cast<InsertElementInst>(BaseShuffle->getOperand(0)); + if (IE) { + auto *Src = IE->getOperand(1); + IndexLoad = dyn_cast<LoadInst>(Src); + if (IndexLoad) + return IndexLoad; + auto *Alloca = dyn_cast<AllocaInst>(Src); + if (Alloca) + return Alloca; + if (isa<Argument>(Src)) { + return Src; + } + if (isa<GlobalValue>(Src)) { + return Src; + } + } + } + LLVM_DEBUG(dbgs() << " Unable to locate Address from intrinsic\n"); + return nullptr; +} + +static Type *getIndexType(Value *In) { + if (!In) + return nullptr; + + if (isa<LoadInst>(In) || isa<StoreInst>(In)) + return getLoadStoreType(In); + + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(In)) { + if (II->getIntrinsicID() == Intrinsic::masked_load) + return II->getType(); + if (II->getIntrinsicID() == Intrinsic::masked_store) + return II->getOperand(0)->getType(); + } + return In->getType(); +} + +static Value *locateIndexesFromGEP(Value *In) { + if (!In) + return nullptr; + if (isa<LoadInst>(In)) + return In; + if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(In)) { + if (II->getIntrinsicID() == Intrinsic::masked_load) + return In; + if (II->getIntrinsicID() == Intrinsic::masked_gather) + return In; + } + if (auto *IndexZEx = dyn_cast<ZExtInst>(In)) + return locateIndexesFromGEP(IndexZEx->getOperand(0)); + if (auto *IndexSEx = dyn_cast<SExtInst>(In)) + return locateIndexesFromGEP(IndexSEx->getOperand(0)); + if (auto *BaseShuffle = dyn_cast<ShuffleVectorInst>(In)) + return locateIndexesFromGEP(BaseShuffle->getOperand(0)); + if (auto *IE = dyn_cast<InsertElementInst>(In)) + return locateIndexesFromGEP(IE->getOperand(1)); + if (auto *cstDataVector = dyn_cast<ConstantDataVector>(In)) + return cstDataVector; + if (auto *GEPIndex = dyn_cast<GetElementPtrInst>(In)) + return GEPIndex->getOperand(0); + return nullptr; +} + +// Given the intrinsic find its GEP argument and extract offsetts from the base +// address it uses. +static Value *locateIndexesFromIntrinsic(Instruction *In) { + GetElementPtrInst *GEPIndex = locateGepFromIntrinsic(In); + if (!GEPIndex) { + LLVM_DEBUG(dbgs() << " No GEP in intrinsic\n"); + return nullptr; + } + Value *Indexes = GEPIndex->getOperand(1); + if (auto *IndexLoad = locateIndexesFromGEP(Indexes)) + return IndexLoad; + + LLVM_DEBUG(dbgs() << " Unable to locate Index from intrinsic\n"); + return nullptr; +} + +// Because of aukward definition of many Hex intrinsics we often have to +// reinterprete HVX native <64 x i16> as <32 x i32> which in practice is a NOP +// for all use cases, so this only exist to make IR builder happy. +inline Value *getReinterpretiveCast_i16_to_i32(const HexagonVectorCombine &HVC, + IRBuilderBase &Builder, + LLVMContext &Ctx, Value *I) { + assert(I && "Unable to reinterprete cast"); + Type *NT = HVC.getHvxTy(HVC.getIntTy(32), false); + std::vector<unsigned> shuffleMask; + for (unsigned i = 0; i < 64; ++i) + shuffleMask.push_back(i); + Constant *Mask = llvm::ConstantDataVector::get(Ctx, shuffleMask); + Value *CastShuffle = + Builder.CreateShuffleVector(I, I, Mask, "identity_shuffle"); + return Builder.CreateBitCast(CastShuffle, NT, "cst64_i16_to_32_i32"); +} + +// Recast <128 x i8> as <32 x i32> +inline Value *getReinterpretiveCast_i8_to_i32(const HexagonVectorCombine &HVC, + IRBuilderBase &Builder, + LLVMContext &Ctx, Value *I) { + assert(I && "Unable to reinterprete cast"); + Type *NT = HVC.getHvxTy(HVC.getIntTy(32), false); + std::vector<unsigned> shuffleMask; + for (unsigned i = 0; i < 128; ++i) + shuffleMask.push_back(i); + Constant *Mask = llvm::ConstantDataVector::get(Ctx, shuffleMask); + Value *CastShuffle = + Builder.CreateShuffleVector(I, I, Mask, "identity_shuffle"); + return Builder.CreateBitCast(CastShuffle, NT, "cst128_i8_to_32_i32"); +} + +// Create <32 x i32> mask reinterpreted as <128 x i1> with a given pattern +inline Value *get_i32_Mask(const HexagonVectorCombine &HVC, + IRBuilderBase &Builder, LLVMContext &Ctx, + unsigned int pattern) { + std::vector<unsigned int> byteMask; + for (unsigned i = 0; i < 32; ++i) + byteMask.push_back(pattern); + + return Builder.CreateIntrinsic( + HVC.getBoolTy(128), HVC.HST.getIntrinsicId(Hexagon::V6_vandvrt), + {llvm::ConstantDataVector::get(Ctx, byteMask), HVC.getConstInt(~0)}, + nullptr); +} + +Value *HvxIdioms::processVScatter(Instruction &In) const { + auto *InpTy = dyn_cast<VectorType>(In.getOperand(0)->getType()); + assert(InpTy && "Cannot handle no vector type for llvm.scatter/gather"); + unsigned InpSize = HVC.getSizeOf(InpTy); + auto *F = In.getFunction(); + LLVMContext &Ctx = F->getContext(); + auto *ElemTy = dyn_cast<IntegerType>(InpTy->getElementType()); + assert(ElemTy && "llvm.scatter needs integer type argument"); + unsigned ElemWidth = HVC.DL.getTypeAllocSize(ElemTy); + LLVM_DEBUG({ + unsigned Elements = HVC.length(InpTy); + dbgs() << "\n[Process scatter](" << In << ")\n" << *In.getParent() << "\n"; + dbgs() << " Input type(" << *InpTy << ") elements(" << Elements + << ") VecLen(" << InpSize << ") type(" << *ElemTy << ") ElemWidth(" + << ElemWidth << ")\n"; + }); + + IRBuilder Builder(In.getParent(), In.getIterator(), + InstSimplifyFolder(HVC.DL)); + + auto *ValueToScatter = In.getOperand(0); + LLVM_DEBUG(dbgs() << " ValueToScatter : " << *ValueToScatter << "\n"); + + if (HVC.HST.getVectorLength() != InpSize) { + LLVM_DEBUG(dbgs() << "Unhandled vector size(" << InpSize + << ") for vscatter\n"); + return nullptr; + } + + // Base address of indexes. + auto *IndexLoad = locateAddressFromIntrinsic(&In); + if (!IndexLoad) + return nullptr; + LLVM_DEBUG(dbgs() << " IndexLoad : " << *IndexLoad << "\n"); + + // Address of destination. Must be in VTCM. + auto *Ptr = getPointer(IndexLoad); + if (!Ptr) + return nullptr; + LLVM_DEBUG(dbgs() << " Ptr : " << *Ptr << "\n"); + // Indexes/offsets + auto *Indexes = locateIndexesFromIntrinsic(&In); + if (!Indexes) + return nullptr; + LLVM_DEBUG(dbgs() << " Indexes : " << *Indexes << "\n"); + Value *CastedDst = Builder.CreateBitOrPointerCast(Ptr, Type::getInt32Ty(Ctx), + "cst_ptr_to_i32"); + LLVM_DEBUG(dbgs() << " CastedDst : " << *CastedDst << "\n"); + // Adjust Indexes + auto *cstDataVector = dyn_cast<ConstantDataVector>(Indexes); + Value *CastIndex = nullptr; + if (cstDataVector) { + // Our indexes are represented as a constant. We need it in a reg. + AllocaInst *IndexesAlloca = + Builder.CreateAlloca(HVC.getHvxTy(HVC.getIntTy(32), false)); + [[maybe_unused]] auto *StoreIndexes = + Builder.CreateStore(cstDataVector, IndexesAlloca); + LLVM_DEBUG(dbgs() << " StoreIndexes : " << *StoreIndexes << "\n"); + CastIndex = Builder.CreateLoad(IndexesAlloca->getAllocatedType(), + IndexesAlloca, "reload_index"); + } else { + if (ElemWidth == 2) + CastIndex = getReinterpretiveCast_i16_to_i32(HVC, Builder, Ctx, Indexes); + else + CastIndex = Indexes; + } + LLVM_DEBUG(dbgs() << " Cast index : " << *CastIndex << ")\n"); + + if (ElemWidth == 1) { + // v128i8 There is no native instruction for this. + // Do this as two Hi/Lo gathers with masking. + Type *NT = HVC.getHvxTy(HVC.getIntTy(32), false); + // Extend indexes. We assume that indexes are in 128i8 format - need to + // expand them to Hi/Lo 64i16 + Value *CastIndexes = Builder.CreateBitCast(CastIndex, NT, "cast_to_32i32"); + auto V6_vunpack = HVC.HST.getIntrinsicId(Hexagon::V6_vunpackub); + auto *UnpackedIndexes = Builder.CreateIntrinsic( + HVC.getHvxTy(HVC.getIntTy(32), true), V6_vunpack, CastIndexes, nullptr); + LLVM_DEBUG(dbgs() << " UnpackedIndexes : " << *UnpackedIndexes << ")\n"); + + auto V6_hi = HVC.HST.getIntrinsicId(Hexagon::V6_hi); + auto V6_lo = HVC.HST.getIntrinsicId(Hexagon::V6_lo); + [[maybe_unused]] Value *IndexHi = + HVC.createHvxIntrinsic(Builder, V6_hi, NT, UnpackedIndexes); + [[maybe_unused]] Value *IndexLo = + HVC.createHvxIntrinsic(Builder, V6_lo, NT, UnpackedIndexes); + LLVM_DEBUG(dbgs() << " UnpackedIndHi : " << *IndexHi << ")\n"); + LLVM_DEBUG(dbgs() << " UnpackedIndLo : " << *IndexLo << ")\n"); + // Now unpack values to scatter + Value *CastSrc = + getReinterpretiveCast_i8_to_i32(HVC, Builder, Ctx, ValueToScatter); + LLVM_DEBUG(dbgs() << " CastSrc : " << *CastSrc << ")\n"); + auto *UnpackedValueToScatter = Builder.CreateIntrinsic( + HVC.getHvxTy(HVC.getIntTy(32), true), V6_vunpack, CastSrc, nullptr); + LLVM_DEBUG(dbgs() << " UnpackedValToScat: " << *UnpackedValueToScatter + << ")\n"); + + [[maybe_unused]] Value *UVSHi = + HVC.createHvxIntrinsic(Builder, V6_hi, NT, UnpackedValueToScatter); + [[maybe_unused]] Value *UVSLo = + HVC.createHvxIntrinsic(Builder, V6_lo, NT, UnpackedValueToScatter); + LLVM_DEBUG(dbgs() << " UVSHi : " << *UVSHi << ")\n"); + LLVM_DEBUG(dbgs() << " UVSLo : " << *UVSLo << ")\n"); + + // Create the mask for individual bytes + auto *QByteMask = get_i32_Mask(HVC, Builder, Ctx, 0x00ff00ff); + LLVM_DEBUG(dbgs() << " QByteMask : " << *QByteMask << "\n"); + [[maybe_unused]] auto *ResHi = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vscattermhq_128B, + {QByteMask, CastedDst, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + IndexHi, UVSHi}, + nullptr); + LLVM_DEBUG(dbgs() << " ResHi : " << *ResHi << ")\n"); + return Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vscattermhq_128B, + {QByteMask, CastedDst, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + IndexLo, UVSLo}, + nullptr); + } else if (ElemWidth == 2) { + Value *CastSrc = + getReinterpretiveCast_i16_to_i32(HVC, Builder, Ctx, ValueToScatter); + LLVM_DEBUG(dbgs() << " CastSrc : " << *CastSrc << ")\n"); + return Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vscattermh_128B, + {CastedDst, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), CastIndex, + CastSrc}, + nullptr); + } else if (ElemWidth == 4) { + return Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vscattermw_128B, + {CastedDst, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), CastIndex, + ValueToScatter}, + nullptr); + } else { + LLVM_DEBUG(dbgs() << "Unhandled element type for vscatter\n"); + return nullptr; + } +} + +Value *HvxIdioms::processVGather(Instruction &In) const { + [[maybe_unused]] auto *InpTy = + dyn_cast<VectorType>(In.getOperand(0)->getType()); + assert(InpTy && "Cannot handle no vector type for llvm.gather"); + [[maybe_unused]] auto *ElemTy = + dyn_cast<PointerType>(InpTy->getElementType()); + assert(ElemTy && "llvm.gather needs vector of ptr argument"); + auto *F = In.getFunction(); + LLVMContext &Ctx = F->getContext(); + LLVM_DEBUG(dbgs() << "\n[Process gather](" << In << ")\n" + << *In.getParent() << "\n"); + LLVM_DEBUG(dbgs() << " Input type(" << *InpTy << ") elements(" + << HVC.length(InpTy) << ") VecLen(" << HVC.getSizeOf(InpTy) + << ") type(" << *ElemTy << ") Access alignment(" + << *In.getOperand(1) << ") AddressSpace(" + << ElemTy->getAddressSpace() << ")\n"); + + // TODO: Handle masking of elements. + assert(dyn_cast<VectorType>(In.getOperand(2)->getType()) && + "llvm.gather needs vector for mask"); + IRBuilder Builder(In.getParent(), In.getIterator(), + InstSimplifyFolder(HVC.DL)); + + // See who is using the result. The difference between LLVM and HVX vgather + // Intrinsic makes it impossible to handle all cases with temp storage. Alloca + // in VTCM is not yet supported, so for now we just bail out for those cases. + HvxIdioms::DstQualifier Qual = HvxIdioms::Undefined; + Instruction *Dst = locateDestination(&In, Qual); + if (!Dst) { + LLVM_DEBUG(dbgs() << " Unable to locate vgather destination\n"); + return nullptr; + } + LLVM_DEBUG(dbgs() << " Destination : " << *Dst << " Qual(" << Qual + << ")\n"); + + // Address of destination. Must be in VTCM. + auto *Ptr = getPointer(Dst); + if (!Ptr) { + LLVM_DEBUG(dbgs() << "Could not locate vgather destination ptr\n"); + return nullptr; + } + + // Result type. Assume it is a vector type. + auto *DstType = cast<VectorType>(getIndexType(Dst)); + assert(DstType && "Cannot handle non vector dst type for llvm.gather"); + + // Base address for sources to be loaded + auto *IndexLoad = locateAddressFromIntrinsic(&In); + if (!IndexLoad) + return nullptr; + LLVM_DEBUG(dbgs() << " IndexLoad : " << *IndexLoad << "\n"); + + // Gather indexes/offsets + auto *Indexes = locateIndexesFromIntrinsic(&In); + if (!Indexes) + return nullptr; + LLVM_DEBUG(dbgs() << " Indexes : " << *Indexes << "\n"); + + Instruction *Gather = nullptr; + Type *NT = HVC.getHvxTy(HVC.getIntTy(32), false); + if (Qual == HvxIdioms::LdSt || Qual == HvxIdioms::Arithmetic) { + // We fully assume the address space is in VTCM. We also assume that all + // pointers in Operand(0) have the same base(!). + // This is the most basic case of all the above. + unsigned OutputSize = HVC.getSizeOf(DstType); + auto *DstElemTy = cast<IntegerType>(DstType->getElementType()); + unsigned ElemWidth = HVC.DL.getTypeAllocSize(DstElemTy); + LLVM_DEBUG(dbgs() << " Buffer type : " << *Ptr->getType() + << " Address space (" + << Ptr->getType()->getPointerAddressSpace() << ")\n" + << " Result type : " << *DstType + << "\n Size in bytes : " << OutputSize + << " element type(" << *DstElemTy + << ")\n ElemWidth : " << ElemWidth << " bytes\n"); + + auto *IndexType = cast<VectorType>(getIndexType(Indexes)); + assert(IndexType && "Cannot handle non vector index type for llvm.gather"); + unsigned IndexWidth = HVC.DL.getTypeAllocSize(IndexType->getElementType()); + LLVM_DEBUG(dbgs() << " IndexWidth(" << IndexWidth << ")\n"); + + // Intrinsic takes i32 instead of pointer so cast. + Value *CastedPtr = Builder.CreateBitOrPointerCast( + IndexLoad, Type::getInt32Ty(Ctx), "cst_ptr_to_i32"); + // [llvm_ptr_ty, llvm_i32_ty, llvm_i32_ty, ...] + // int_hexagon_V6_vgathermh [... , llvm_v16i32_ty] + // int_hexagon_V6_vgathermh_128B [... , llvm_v32i32_ty] + // int_hexagon_V6_vgathermhw [... , llvm_v32i32_ty] + // int_hexagon_V6_vgathermhw_128B [... , llvm_v64i32_ty] + // int_hexagon_V6_vgathermw [... , llvm_v16i32_ty] + // int_hexagon_V6_vgathermw_128B [... , llvm_v32i32_ty] + if (HVC.HST.getVectorLength() == OutputSize) { + if (ElemWidth == 1) { + // v128i8 There is no native instruction for this. + // Do this as two Hi/Lo gathers with masking. + // Unpack indexes. We assume that indexes are in 128i8 format - need to + // expand them to Hi/Lo 64i16 + Value *CastIndexes = + Builder.CreateBitCast(Indexes, NT, "cast_to_32i32"); + auto V6_vunpack = HVC.HST.getIntrinsicId(Hexagon::V6_vunpackub); + auto *UnpackedIndexes = + Builder.CreateIntrinsic(HVC.getHvxTy(HVC.getIntTy(32), true), + V6_vunpack, CastIndexes, nullptr); + LLVM_DEBUG(dbgs() << " UnpackedIndexes : " << *UnpackedIndexes + << ")\n"); + + auto V6_hi = HVC.HST.getIntrinsicId(Hexagon::V6_hi); + auto V6_lo = HVC.HST.getIntrinsicId(Hexagon::V6_lo); + [[maybe_unused]] Value *IndexHi = + HVC.createHvxIntrinsic(Builder, V6_hi, NT, UnpackedIndexes); + [[maybe_unused]] Value *IndexLo = + HVC.createHvxIntrinsic(Builder, V6_lo, NT, UnpackedIndexes); + LLVM_DEBUG(dbgs() << " UnpackedIndHi : " << *IndexHi << ")\n"); + LLVM_DEBUG(dbgs() << " UnpackedIndLo : " << *IndexLo << ")\n"); + // Create the mask for individual bytes + auto *QByteMask = get_i32_Mask(HVC, Builder, Ctx, 0x00ff00ff); + LLVM_DEBUG(dbgs() << " QByteMask : " << *QByteMask << "\n"); + // We use our destination allocation as a temp storage + // This is unlikely to work properly for masked gather. + auto V6_vgather = HVC.HST.getIntrinsicId(Hexagon::V6_vgathermhq); + [[maybe_unused]] auto GatherHi = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), V6_vgather, + {Ptr, QByteMask, CastedPtr, + HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), IndexHi}, + nullptr); + LLVM_DEBUG(dbgs() << " GatherHi : " << *GatherHi << ")\n"); + // Rematerialize the result + [[maybe_unused]] Value *LoadedResultHi = Builder.CreateLoad( + HVC.getHvxTy(HVC.getIntTy(32), false), Ptr, "temp_result_hi"); + LLVM_DEBUG(dbgs() << " LoadedResultHi : " << *LoadedResultHi << "\n"); + // Same for the low part. Here we use Gather to return non-NULL result + // from this function and continue to iterate. We also are deleting Dst + // store below. + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), V6_vgather, + {Ptr, QByteMask, CastedPtr, + HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), IndexLo}, + nullptr); + LLVM_DEBUG(dbgs() << " GatherLo : " << *Gather << ")\n"); + Value *LoadedResultLo = Builder.CreateLoad( + HVC.getHvxTy(HVC.getIntTy(32), false), Ptr, "temp_result_lo"); + LLVM_DEBUG(dbgs() << " LoadedResultLo : " << *LoadedResultLo << "\n"); + // Now we have properly sized bytes in every other position + // B b A a c a A b B c f F g G h H is presented as + // B . b . A . a . c . a . A . b . B . c . f . F . g . G . h . H + // Use vpack to gather them + auto V6_vpackeb = HVC.HST.getIntrinsicId(Hexagon::V6_vpackeb); + [[maybe_unused]] auto Res = Builder.CreateIntrinsic( + NT, V6_vpackeb, {LoadedResultHi, LoadedResultLo}, nullptr); + LLVM_DEBUG(dbgs() << " ScaledRes : " << *Res << "\n"); + [[maybe_unused]] auto *StoreRes = Builder.CreateStore(Res, Ptr); + LLVM_DEBUG(dbgs() << " StoreRes : " << *StoreRes << "\n"); + } else if (ElemWidth == 2) { + // v32i16 + if (IndexWidth == 2) { + // Reinterprete 64i16 as 32i32. Only needed for syntactic IR match. + Value *CastIndex = + getReinterpretiveCast_i16_to_i32(HVC, Builder, Ctx, Indexes); + LLVM_DEBUG(dbgs() << " Cast index: " << *CastIndex << ")\n"); + // shift all i16 left by 1 to match short addressing mode instead of + // byte. + auto V6_vaslh = HVC.HST.getIntrinsicId(Hexagon::V6_vaslh); + Value *AdjustedIndex = HVC.createHvxIntrinsic( + Builder, V6_vaslh, NT, {CastIndex, HVC.getConstInt(1)}); + LLVM_DEBUG(dbgs() + << " Shifted half index: " << *AdjustedIndex << ")\n"); + + auto V6_vgather = HVC.HST.getIntrinsicId(Hexagon::V6_vgathermh); + // The 3rd argument is the size of the region to gather from. Probably + // want to set it to max VTCM size. + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), V6_vgather, + {Ptr, CastedPtr, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + AdjustedIndex}, + nullptr); + for (auto &U : Dst->uses()) { + if (auto *UI = dyn_cast<Instruction>(U.getUser())) + dbgs() << " dst used by: " << *UI << "\n"; + } + for (auto &U : In.uses()) { + if (auto *UI = dyn_cast<Instruction>(U.getUser())) + dbgs() << " In used by : " << *UI << "\n"; + } + // Create temp load from result in case the result is used by any + // other instruction. + Value *LoadedResult = Builder.CreateLoad( + HVC.getHvxTy(HVC.getIntTy(16), false), Ptr, "temp_result"); + LLVM_DEBUG(dbgs() << " LoadedResult : " << *LoadedResult << "\n"); + In.replaceAllUsesWith(LoadedResult); + } else { + dbgs() << " Unhandled index type for vgather\n"; + return nullptr; + } + } else if (ElemWidth == 4) { + if (IndexWidth == 4) { + // v32i32 + auto V6_vaslh = HVC.HST.getIntrinsicId(Hexagon::V6_vaslh); + Value *AdjustedIndex = HVC.createHvxIntrinsic( + Builder, V6_vaslh, NT, {Indexes, HVC.getConstInt(2)}); + LLVM_DEBUG(dbgs() + << " Shifted word index: " << *AdjustedIndex << ")\n"); + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vgathermw_128B, + {Ptr, CastedPtr, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + AdjustedIndex}, + nullptr); + } else { + LLVM_DEBUG(dbgs() << " Unhandled index type for vgather\n"); + return nullptr; + } + } else { + LLVM_DEBUG(dbgs() << " Unhandled element type for vgather\n"); + return nullptr; + } + } else if (HVC.HST.getVectorLength() == OutputSize * 2) { + // This is half of the reg width, duplicate low in high + LLVM_DEBUG(dbgs() << " Unhandled half of register size\n"); + return nullptr; + } else if (HVC.HST.getVectorLength() * 2 == OutputSize) { + LLVM_DEBUG(dbgs() << " Unhandle twice the register size\n"); + return nullptr; + } + // Erase the original intrinsic and store that consumes it. + // HVX will create a pseudo for gather that is expanded to gather + store + // during packetization. + Dst->eraseFromParent(); + } else if (Qual == HvxIdioms::LLVM_Scatter) { + // Gather feeds directly into scatter. + LLVM_DEBUG({ + auto *DstInpTy = cast<VectorType>(Dst->getOperand(1)->getType()); + assert(DstInpTy && "Cannot handle no vector type for llvm.scatter"); + unsigned DstInpSize = HVC.getSizeOf(DstInpTy); + unsigned DstElements = HVC.length(DstInpTy); + auto *DstElemTy = cast<PointerType>(DstInpTy->getElementType()); + assert(DstElemTy && "llvm.scatter needs vector of ptr argument"); + dbgs() << " Gather feeds into scatter\n Values to scatter : " + << *Dst->getOperand(0) << "\n"; + dbgs() << " Dst type(" << *DstInpTy << ") elements(" << DstElements + << ") VecLen(" << DstInpSize << ") type(" << *DstElemTy + << ") Access alignment(" << *Dst->getOperand(2) << ")\n"; + }); + // Address of source + auto *Src = getPointer(IndexLoad); + if (!Src) + return nullptr; + LLVM_DEBUG(dbgs() << " Src : " << *Src << "\n"); + + if (!isa<PointerType>(Src->getType())) { + LLVM_DEBUG(dbgs() << " Source is not a pointer type...\n"); + return nullptr; + } + + Value *CastedSrc = Builder.CreateBitOrPointerCast( + Src, Type::getInt32Ty(Ctx), "cst_ptr_to_i32"); + LLVM_DEBUG(dbgs() << " CastedSrc: " << *CastedSrc << "\n"); + + auto *DstLoad = locateAddressFromIntrinsic(Dst); + if (!DstLoad) { + LLVM_DEBUG(dbgs() << " Unable to locate DstLoad\n"); + return nullptr; + } + LLVM_DEBUG(dbgs() << " DstLoad : " << *DstLoad << "\n"); + + Value *Ptr = getPointer(DstLoad); + if (!Ptr) + return nullptr; + LLVM_DEBUG(dbgs() << " Ptr : " << *Ptr << "\n"); + Value *CastIndex = + getReinterpretiveCast_i16_to_i32(HVC, Builder, Ctx, IndexLoad); + LLVM_DEBUG(dbgs() << " Cast index: " << *CastIndex << ")\n"); + // Shift all i16 left by 1 to match short addressing mode instead of + // byte. + auto V6_vaslh = HVC.HST.getIntrinsicId(Hexagon::V6_vaslh); + Value *AdjustedIndex = HVC.createHvxIntrinsic( + Builder, V6_vaslh, NT, {CastIndex, HVC.getConstInt(1)}); + LLVM_DEBUG(dbgs() << " Shifted half index: " << *AdjustedIndex << ")\n"); + + return Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vgathermh_128B, + {Ptr, CastedSrc, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + AdjustedIndex}, + nullptr); + } else if (Qual == HvxIdioms::HEX_Gather_Scatter) { + // Gather feeds into previously inserted pseudo intrinsic. + // These could not be in the same packet, so we need to generate another + // pseudo that is expanded to .tmp + store V6_vgathermh_pseudo + // V6_vgathermh_pseudo (ins IntRegs:$_dst_, s4_0Imm:$Ii, IntRegs:$Rt, + // ModRegs:$Mu, HvxVR:$Vv) + if (isa<AllocaInst>(IndexLoad)) { + auto *cstDataVector = dyn_cast<ConstantDataVector>(Indexes); + if (cstDataVector) { + // Our indexes are represented as a constant. We need THEM in a reg. + // This most likely will not work properly since alloca gives us DDR + // stack location. This will be fixed once we teach compiler about VTCM. + AllocaInst *IndexesAlloca = Builder.CreateAlloca(NT); + [[maybe_unused]] auto *StoreIndexes = + Builder.CreateStore(cstDataVector, IndexesAlloca); + LLVM_DEBUG(dbgs() << " StoreIndexes : " << *StoreIndexes << "\n"); + Value *LoadedIndex = Builder.CreateLoad( + IndexesAlloca->getAllocatedType(), IndexesAlloca, "reload_index"); + AllocaInst *ResultAlloca = Builder.CreateAlloca(NT); + LLVM_DEBUG(dbgs() << " ResultAlloca : " << *ResultAlloca << "\n"); + + Value *CastedSrc = Builder.CreateBitOrPointerCast( + IndexLoad, Type::getInt32Ty(Ctx), "cst_ptr_to_i32"); + LLVM_DEBUG(dbgs() << " CastedSrc : " << *CastedSrc << "\n"); + + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vgathermh_128B, + {ResultAlloca, CastedSrc, + HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), LoadedIndex}, + nullptr); + Value *LoadedResult = Builder.CreateLoad( + HVC.getHvxTy(HVC.getIntTy(16), false), ResultAlloca, "temp_result"); + LLVM_DEBUG(dbgs() << " LoadedResult : " << *LoadedResult << "\n"); + LLVM_DEBUG(dbgs() << " Gather : " << *Gather << "\n"); + In.replaceAllUsesWith(LoadedResult); + } + } else { + // Address of source + auto *Src = getPointer(IndexLoad); + if (!Src) + return nullptr; + LLVM_DEBUG(dbgs() << " Src : " << *Src << "\n"); + + Value *CastedSrc = Builder.CreateBitOrPointerCast( + Src, Type::getInt32Ty(Ctx), "cst_ptr_to_i32"); + LLVM_DEBUG(dbgs() << " CastedSrc: " << *CastedSrc << "\n"); + + auto *DstLoad = locateAddressFromIntrinsic(Dst); + if (!DstLoad) + return nullptr; + LLVM_DEBUG(dbgs() << " DstLoad : " << *DstLoad << "\n"); + auto *Ptr = getPointer(DstLoad); + if (!Ptr) + return nullptr; + LLVM_DEBUG(dbgs() << " Ptr : " << *Ptr << "\n"); + + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vgather_vscattermh, + {Ptr, CastedSrc, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + Indexes}, + nullptr); + } + return Gather; + } else if (Qual == HvxIdioms::HEX_Scatter) { + // This is the case when result of a gather is used as an argument to + // Intrinsic::hexagon_V6_vscattermh_128B. Most likely we just inserted it + // ourselves. We have to create alloca, store to it, and replace all uses + // with that. + AllocaInst *ResultAlloca = Builder.CreateAlloca(NT); + Value *CastedSrc = Builder.CreateBitOrPointerCast( + IndexLoad, Type::getInt32Ty(Ctx), "cst_ptr_to_i32"); + LLVM_DEBUG(dbgs() << " CastedSrc : " << *CastedSrc << "\n"); + Value *CastIndex = + getReinterpretiveCast_i16_to_i32(HVC, Builder, Ctx, Indexes); + LLVM_DEBUG(dbgs() << " Cast index : " << *CastIndex << ")\n"); + + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vgathermh_128B, + {ResultAlloca, CastedSrc, HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), + CastIndex}, + nullptr); + Value *LoadedResult = Builder.CreateLoad( + HVC.getHvxTy(HVC.getIntTy(16), false), ResultAlloca, "temp_result"); + LLVM_DEBUG(dbgs() << " LoadedResult : " << *LoadedResult << "\n"); + In.replaceAllUsesWith(LoadedResult); + } else if (Qual == HvxIdioms::HEX_Gather) { + // Gather feeds to another gather but already replaced with + // hexagon_V6_vgathermh_128B + if (isa<AllocaInst>(IndexLoad)) { + auto *cstDataVector = dyn_cast<ConstantDataVector>(Indexes); + if (cstDataVector) { + // Our indexes are represented as a constant. We need it in a reg. + AllocaInst *IndexesAlloca = Builder.CreateAlloca(NT); + + [[maybe_unused]] auto *StoreIndexes = + Builder.CreateStore(cstDataVector, IndexesAlloca); + LLVM_DEBUG(dbgs() << " StoreIndexes : " << *StoreIndexes << "\n"); + Value *LoadedIndex = Builder.CreateLoad( + IndexesAlloca->getAllocatedType(), IndexesAlloca, "reload_index"); + AllocaInst *ResultAlloca = Builder.CreateAlloca(NT); + LLVM_DEBUG(dbgs() << " ResultAlloca : " << *ResultAlloca + << "\n AddressSpace: " + << ResultAlloca->getAddressSpace() << "\n";); + + Value *CastedSrc = Builder.CreateBitOrPointerCast( + IndexLoad, Type::getInt32Ty(Ctx), "cst_ptr_to_i32"); + LLVM_DEBUG(dbgs() << " CastedSrc : " << *CastedSrc << "\n"); + + Gather = Builder.CreateIntrinsic( + Type::getVoidTy(Ctx), Intrinsic::hexagon_V6_vgathermh_128B, + {ResultAlloca, CastedSrc, + HVC.getConstInt(DEFAULT_HVX_VTCM_PAGE_SIZE), LoadedIndex}, + nullptr); + Value *LoadedResult = Builder.CreateLoad( + HVC.getHvxTy(HVC.getIntTy(16), false), ResultAlloca, "temp_result"); + LLVM_DEBUG(dbgs() << " LoadedResult : " << *LoadedResult << "\n"); + LLVM_DEBUG(dbgs() << " Gather : " << *Gather << "\n"); + In.replaceAllUsesWith(LoadedResult); + } + } + } else if (Qual == HvxIdioms::LLVM_Gather) { + // Gather feeds into another gather + errs() << " Underimplemented vgather to vgather sequence\n"; + return nullptr; + } else + llvm_unreachable("Unhandled Qual enum"); + + return Gather; +} + auto HvxIdioms::processFxpMulChopped(IRBuilderBase &Builder, Instruction &In, const FxpOp &Op) const -> Value * { assert(Op.X.Val->getType() == Op.Y.Val->getType()); @@ -2138,6 +3000,26 @@ auto HvxIdioms::run() -> bool { It = StartOver ? B.rbegin() : cast<Instruction>(New)->getReverseIterator(); Changed = true; + } else if (matchGather(*It)) { + Value *New = processVGather(*It); + if (!New) + continue; + LLVM_DEBUG(dbgs() << " Gather : " << *New << "\n"); + // We replace original intrinsic with a new pseudo call. + It->eraseFromParent(); + It = cast<Instruction>(New)->getReverseIterator(); + RecursivelyDeleteTriviallyDeadInstructions(&*It, &HVC.TLI); + Changed = true; + } else if (matchScatter(*It)) { + Value *New = processVScatter(*It); + if (!New) + continue; + LLVM_DEBUG(dbgs() << " Scatter : " << *New << "\n"); + // We replace original intrinsic with a new pseudo call. + It->eraseFromParent(); + It = cast<Instruction>(New)->getReverseIterator(); + RecursivelyDeleteTriviallyDeadInstructions(&*It, &HVC.TLI); + Changed = true; } } } diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCELFStreamer.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCELFStreamer.cpp index 6455757..2f59b7c 100644 --- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCELFStreamer.cpp +++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCELFStreamer.cpp @@ -186,6 +186,9 @@ static unsigned featureToArchVersion(unsigned Feature) { case Hexagon::ArchV79: case Hexagon::ExtensionHVXV79: return 79; + case Hexagon::ArchV81: + case Hexagon::ExtensionHVXV81: + return 81; } llvm_unreachable("Expected valid arch feature"); return 0; diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp index 6b48a21..b8075bd 100644 --- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp +++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp @@ -96,6 +96,8 @@ cl::opt<bool> MV75("mv75", cl::Hidden, cl::desc("Build for Hexagon V75"), cl::init(false)); cl::opt<bool> MV79("mv79", cl::Hidden, cl::desc("Build for Hexagon V79"), cl::init(false)); +cl::opt<bool> MV81("mv81", cl::Hidden, cl::desc("Build for Hexagon V81"), + cl::init(false)); } // namespace static cl::opt<Hexagon::ArchEnum> EnableHVX( @@ -111,6 +113,7 @@ static cl::opt<Hexagon::ArchEnum> EnableHVX( clEnumValN(Hexagon::ArchEnum::V73, "v73", "Build for HVX v73"), clEnumValN(Hexagon::ArchEnum::V75, "v75", "Build for HVX v75"), clEnumValN(Hexagon::ArchEnum::V79, "v79", "Build for HVX v79"), + clEnumValN(Hexagon::ArchEnum::V81, "v81", "Build for HVX v81"), // Sentinel for no value specified. clEnumValN(Hexagon::ArchEnum::Generic, "", "")), // Sentinel for flag not present. @@ -159,6 +162,8 @@ static StringRef HexagonGetArchVariant() { return "hexagonv75"; if (MV79) return "hexagonv79"; + if (MV81) + return "hexagonv81"; return ""; } @@ -474,6 +479,9 @@ std::string selectHexagonFS(StringRef CPU, StringRef FS) { case Hexagon::ArchEnum::V79: Result.push_back("+hvxv79"); break; + case Hexagon::ArchEnum::V81: + Result.push_back("+hvxv81"); + break; case Hexagon::ArchEnum::Generic: { Result.push_back(StringSwitch<StringRef>(CPU) @@ -489,7 +497,8 @@ std::string selectHexagonFS(StringRef CPU, StringRef FS) { .Case("hexagonv71t", "+hvxv71") .Case("hexagonv73", "+hvxv73") .Case("hexagonv75", "+hvxv75") - .Case("hexagonv79", "+hvxv79")); + .Case("hexagonv79", "+hvxv79") + .Case("hexagonv81", "+hvxv81")); break; } case Hexagon::ArchEnum::NoArch: @@ -538,8 +547,8 @@ FeatureBitset Hexagon_MC::completeHVXFeatures(const FeatureBitset &S) { FeatureBitset FB = S; unsigned CpuArch = ArchV5; for (unsigned F : - {ArchV79, ArchV75, ArchV73, ArchV71, ArchV69, ArchV68, ArchV67, ArchV66, - ArchV65, ArchV62, ArchV60, ArchV55, ArchV5}) { + {ArchV81, ArchV79, ArchV75, ArchV73, ArchV71, ArchV69, ArchV68, ArchV67, + ArchV66, ArchV65, ArchV62, ArchV60, ArchV55, ArchV5}) { if (!FB.test(F)) continue; CpuArch = F; @@ -556,7 +565,7 @@ FeatureBitset Hexagon_MC::completeHVXFeatures(const FeatureBitset &S) { for (unsigned F : {ExtensionHVXV60, ExtensionHVXV62, ExtensionHVXV65, ExtensionHVXV66, ExtensionHVXV67, ExtensionHVXV68, ExtensionHVXV69, ExtensionHVXV71, - ExtensionHVXV73, ExtensionHVXV75, ExtensionHVXV79}) { + ExtensionHVXV73, ExtensionHVXV75, ExtensionHVXV79, ExtensionHVXV81}) { if (!FB.test(F)) continue; HasHvxVer = true; @@ -569,6 +578,9 @@ FeatureBitset Hexagon_MC::completeHVXFeatures(const FeatureBitset &S) { // HasHvxVer is false, and UseHvx is true. switch (CpuArch) { + case ArchV81: + FB.set(ExtensionHVXV81); + [[fallthrough]]; case ArchV79: FB.set(ExtensionHVXV79); [[fallthrough]]; @@ -668,12 +680,12 @@ void Hexagon_MC::addArchSubtarget(MCSubtargetInfo const *STI, StringRef FS) { std::optional<unsigned> Hexagon_MC::getHVXVersion(const FeatureBitset &Features) { - for (auto Arch : {Hexagon::ExtensionHVXV79, Hexagon::ExtensionHVXV75, - Hexagon::ExtensionHVXV73, Hexagon::ExtensionHVXV71, - Hexagon::ExtensionHVXV69, Hexagon::ExtensionHVXV68, - Hexagon::ExtensionHVXV67, Hexagon::ExtensionHVXV66, - Hexagon::ExtensionHVXV65, Hexagon::ExtensionHVXV62, - Hexagon::ExtensionHVXV60}) + for (auto Arch : {Hexagon::ExtensionHVXV81, Hexagon::ExtensionHVXV79, + Hexagon::ExtensionHVXV75, Hexagon::ExtensionHVXV73, + Hexagon::ExtensionHVXV71, Hexagon::ExtensionHVXV69, + Hexagon::ExtensionHVXV68, Hexagon::ExtensionHVXV67, + Hexagon::ExtensionHVXV66, Hexagon::ExtensionHVXV65, + Hexagon::ExtensionHVXV62, Hexagon::ExtensionHVXV60}) if (Features.test(Arch)) return Arch; return {}; @@ -681,13 +693,13 @@ Hexagon_MC::getHVXVersion(const FeatureBitset &Features) { unsigned Hexagon_MC::getArchVersion(const FeatureBitset &Features) { for (auto Arch : - {Hexagon::ArchV79, Hexagon::ArchV75, Hexagon::ArchV73, Hexagon::ArchV71, - Hexagon::ArchV69, Hexagon::ArchV68, Hexagon::ArchV67, Hexagon::ArchV66, - Hexagon::ArchV65, Hexagon::ArchV62, Hexagon::ArchV60, Hexagon::ArchV55, - Hexagon::ArchV5}) + {Hexagon::ArchV81, Hexagon::ArchV79, Hexagon::ArchV75, Hexagon::ArchV73, + Hexagon::ArchV71, Hexagon::ArchV69, Hexagon::ArchV68, Hexagon::ArchV67, + Hexagon::ArchV66, Hexagon::ArchV65, Hexagon::ArchV62, Hexagon::ArchV60, + Hexagon::ArchV55, Hexagon::ArchV5}) if (Features.test(Arch)) return Arch; - llvm_unreachable("Expected arch v5-v79"); + llvm_unreachable("Expected arch v5-v81"); return 0; } @@ -708,7 +720,8 @@ unsigned Hexagon_MC::GetELFFlags(const MCSubtargetInfo &STI) { .Case("hexagonv71t", llvm::ELF::EF_HEXAGON_MACH_V71T) .Case("hexagonv73", llvm::ELF::EF_HEXAGON_MACH_V73) .Case("hexagonv75", llvm::ELF::EF_HEXAGON_MACH_V75) - .Case("hexagonv79", llvm::ELF::EF_HEXAGON_MACH_V79); + .Case("hexagonv79", llvm::ELF::EF_HEXAGON_MACH_V79) + .Case("hexagonv81", llvm::ELF::EF_HEXAGON_MACH_V81); } llvm::ArrayRef<MCPhysReg> Hexagon_MC::GetVectRegRev() { diff --git a/llvm/lib/Target/PowerPC/PPCInstrInfo.td b/llvm/lib/Target/PowerPC/PPCInstrInfo.td index aca7abd..44d1a44 100644 --- a/llvm/lib/Target/PowerPC/PPCInstrInfo.td +++ b/llvm/lib/Target/PowerPC/PPCInstrInfo.td @@ -4578,6 +4578,8 @@ def : InstAlias<"mfamr $Rx", (MFSPR gprc:$Rx, 29)>; def : InstAlias<"mtpid $Rx", (MTSPR 48, gprc:$Rx)>, Requires<[IsBookE]>; def : InstAlias<"mfpid $Rx", (MFSPR gprc:$Rx, 48)>, Requires<[IsBookE]>; +def : InstAlias<"mtpidr $Rx", (MTSPR 48, gprc:$Rx)>, Requires<[IsISA3_0]>; +def : InstAlias<"mfpidr $Rx", (MFSPR gprc:$Rx, 48)>, Requires<[IsISA3_0]>; foreach SPRG = 4-7 in { def : InstAlias<"mfsprg $RT, "#SPRG, (MFSPR gprc:$RT, !add(SPRG, 256))>, diff --git a/llvm/lib/Target/RISCV/RISCVFeatures.td b/llvm/lib/Target/RISCV/RISCVFeatures.td index 9e6b7f0..2754d78 100644 --- a/llvm/lib/Target/RISCV/RISCVFeatures.td +++ b/llvm/lib/Target/RISCV/RISCVFeatures.td @@ -1124,7 +1124,8 @@ def HasStdExtZbkbOrP "'Base P' (Packed-SIMD)">; def HasStdExtZbbOrZbkbOrP - : Predicate<"Subtarget->HasStdExtZbbOrZbkb()|| Subtarget->hasStdExtP()">, + : Predicate<"Subtarget->hasStdExtZbb() || Subtarget->hasStdExtZbkb() || " + "Subtarget->hasStdExtP()">, AssemblerPredicate<(any_of FeatureStdExtZbb, FeatureStdExtZbkb, FeatureStdExtP), "'Zbb' (Basic Bit-Manipulation) or " "'Zbkb' (Bitmanip instructions for Cryptography) or " diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp index 219e3f2..1c930ac 100644 --- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp +++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp @@ -318,8 +318,9 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM, setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); - if (!Subtarget.hasStdExtZbb() && !Subtarget.hasVendorXTHeadBb() && - !Subtarget.hasVendorXqcibm() && !Subtarget.hasVendorXAndesPerf() && + if (!Subtarget.hasStdExtZbb() && !Subtarget.hasStdExtP() && + !Subtarget.hasVendorXTHeadBb() && !Subtarget.hasVendorXqcibm() && + !Subtarget.hasVendorXAndesPerf() && !(Subtarget.hasVendorXCValu() && !Subtarget.is64Bit())) setOperationAction(ISD::SIGN_EXTEND_INREG, {MVT::i8, MVT::i16}, Expand); @@ -392,7 +393,7 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM, setOperationAction(ISD::BITREVERSE, MVT::i8, Custom); } - if (Subtarget.hasStdExtZbb() || + if (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtP() || (Subtarget.hasVendorXCValu() && !Subtarget.is64Bit())) { setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, XLenVT, Legal); @@ -403,6 +404,9 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM, setOperationAction({ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF}, MVT::i32, Custom); } else { setOperationAction(ISD::CTTZ, XLenVT, Expand); + // If have a CLZW, but not CTZW, custom promote i32. + if (Subtarget.hasStdExtP() && Subtarget.is64Bit()) + setOperationAction({ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF}, MVT::i32, Custom); } if (!Subtarget.hasCPOPLike()) { @@ -419,13 +423,15 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM, // We need the custom lowering to make sure that the resulting sequence // for the 32bit case is efficient on 64bit targets. // Use default promotion for i32 without Zbb. - if (Subtarget.is64Bit() && Subtarget.hasStdExtZbb()) + if (Subtarget.is64Bit() && + (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtP())) setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, MVT::i32, Custom); } else { setOperationAction(ISD::CTLZ, XLenVT, Expand); } - if (Subtarget.hasVendorXCValu() && !Subtarget.is64Bit()) { + if (Subtarget.hasStdExtP() || + (Subtarget.hasVendorXCValu() && !Subtarget.is64Bit())) { setOperationAction(ISD::ABS, XLenVT, Legal); } else if (Subtarget.hasShortForwardBranchOpt()) { // We can use PseudoCCSUB to implement ABS. @@ -14669,6 +14675,25 @@ void RISCVTargetLowering::ReplaceNodeResults(SDNode *N, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0)); bool IsCTZ = N->getOpcode() == ISD::CTTZ || N->getOpcode() == ISD::CTTZ_ZERO_UNDEF; + + // Without Zbb, lower as 32 - clzw(~X & (X-1)) + if (IsCTZ && !Subtarget.hasStdExtZbb()) { + assert(Subtarget.hasStdExtP()); + + NewOp0 = DAG.getFreeze(NewOp0); + SDValue Not = DAG.getNOT(DL, NewOp0, MVT::i64); + SDValue Minus1 = DAG.getNode(ISD::SUB, DL, MVT::i64, NewOp0, + DAG.getConstant(1, DL, MVT::i64)); + SDValue And = DAG.getNode(ISD::AND, DL, MVT::i64, Not, Minus1); + SDValue CLZW = DAG.getNode(RISCVISD::CLZW, DL, MVT::i64, And); + SDValue Sub = DAG.getNode(ISD::SUB, DL, MVT::i64, + DAG.getConstant(32, DL, MVT::i64), CLZW); + SDValue Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, Sub, + DAG.getValueType(MVT::i32)); + Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res)); + return; + } + unsigned Opc = IsCTZ ? RISCVISD::CTZW : RISCVISD::CLZW; SDValue Res = DAG.getNode(Opc, DL, MVT::i64, NewOp0); Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Res)); diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td index 7d8a919..cc085bb 100644 --- a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td +++ b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td @@ -1455,3 +1455,11 @@ let Predicates = [HasStdExtP, IsRV32] in { def PMAXU_DW : RVPPairBinaryExchanged_rr<0b1111, 0b01, "pmaxu.dw">; def PMAXU_DB : RVPPairBinaryExchanged_rr<0b1111, 0b10, "pmaxu.db">; } // Predicates = [HasStdExtP, IsRV32] + + +//===----------------------------------------------------------------------===// +// Codegen patterns +//===----------------------------------------------------------------------===// + +let Predicates = [HasStdExtP] in +def : PatGpr<abs, ABS>; diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td b/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td index 4c2f7f6..f7b4914 100644 --- a/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td +++ b/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td @@ -218,11 +218,13 @@ let Predicates = [HasVendorXSfvcp], mayLoad = 0, mayStore = 0, } let Predicates = [HasVendorXSfvfexpAny], DecoderNamespace = "XSfvector" in { - def SF_VFEXP_V : VALUVs2<0b010011, 0b00111, OPFVV, "sf.vfexp.v">; + def SF_VFEXP_V : VALUVs2<0b010011, 0b00111, OPFVV, "sf.vfexp.v">, + SchedUnaryMC<"WriteSF_VFExp", "ReadSF_VFExp">; } let Predicates = [HasVendorXSfvfexpa], DecoderNamespace = "XSfvector" in { - def SF_VFEXPA_V : VALUVs2<0b010011, 0b00110, OPFVV, "sf.vfexpa.v">; + def SF_VFEXPA_V : VALUVs2<0b010011, 0b00110, OPFVV, "sf.vfexpa.v">, + SchedUnaryMC<"WriteSF_VFExpa", "ReadSF_VFExpa">; } let Predicates = [HasVendorXSfvqmaccdod], DecoderNamespace = "XSfvector", @@ -487,6 +489,48 @@ let Predicates = [HasVendorXSfvfnrclipxfqf], AltFmtType = IS_NOT_ALTFMT in { defm SF_VFNRCLIP_X_F_QF : VPseudoSiFiveVFNRCLIP; } +class VFExpSchedSEWSet<string mx, bit IsBF16, bit IsApprox> { + defvar BaseSet = SchedSEWSet<mx, isF=1>.val; + list<int> val = !if(IsBF16, !listremove(BaseSet, [32, 64]), + !if(IsApprox, BaseSet, !listremove(BaseSet, [64]))); +} +multiclass VPseudoVFExp_V<bit IsBF16 = false, bit IsApprox = false> { + defvar SchedSuffix = !if(IsApprox, "VFExpa", "VFExp"); + + foreach m = MxListF in { + defvar mx = m.MX; + foreach e = VFExpSchedSEWSet<mx, IsBF16, IsApprox>.val in { + let VLMul = m.value in { + def "_V_" # mx # "_E" # e + : VPseudoUnaryNoMask<m.vrclass, m.vrclass>, + SchedUnary<"WriteSF_" # SchedSuffix, "ReadSF_" # SchedSuffix, + mx, e, forcePassthruRead=true>; + def "_V_" # mx # "_E" # e # "_MASK" + : VPseudoUnaryMask<m.vrclass, m.vrclass>, + RISCVMaskedPseudo<MaskIdx = 2>, + SchedUnary<"WriteSF_" # SchedSuffix, "ReadSF_" # SchedSuffix, + mx, e, forcePassthruRead=true>; + } + } + } +} + +let Predicates = [HasVendorXSfvfbfexp16e], hasSideEffects = 0 in { + let AltFmtType = IS_ALTFMT in { + defm PseudoSF_VFEXP_ALT : VPseudoVFExp_V<IsBF16=true>; + } +} + +let Predicates = [HasVendorXSfvfexpAnyFloat], hasSideEffects = 0 in { + let AltFmtType = IS_NOT_ALTFMT in { + defm PseudoSF_VFEXP : VPseudoVFExp_V; + } +} + +let Predicates = [HasVendorXSfvfexpa], AltFmtType = IS_NOT_ALTFMT in { + defm PseudoSF_VFEXPA : VPseudoVFExp_V<IsApprox=true>; +} + // SDNode def SDT_SF_VC_V_X : SDTypeProfile<1, 4, [SDTCisVec<0>, SDTCisVT<1, XLenVT>, @@ -893,3 +937,36 @@ let Predicates = [HasVendorXSfcease] in { let rs2 = 0b00101; } } + +let Predicates = [HasVendorXSfvfbfexp16e] in { + defm : VPatUnaryV_V<"int_riscv_sf_vfexp", "PseudoSF_VFEXP_ALT", + AllBF16Vectors, + isSEWAware=1>; +} + +let Predicates = [HasVendorXSfvfexp16e] in { + defm : VPatUnaryV_V<"int_riscv_sf_vfexp", "PseudoSF_VFEXP", + [VF16MF4, VF16MF2, VF16M1, VF16M2, VF16M4, VF16M8], + isSEWAware=1>; +} + +let Predicates = [HasVendorXSfvfexp32e] in { + defm : VPatUnaryV_V<"int_riscv_sf_vfexp", "PseudoSF_VFEXP", + [VF32MF2, VF32M1, VF32M2, VF32M4, VF32M8], isSEWAware=1>; +} + +let Predicates = [HasVendorXSfvfexpa] in { + defm : VPatUnaryV_V<"int_riscv_sf_vfexpa", "PseudoSF_VFEXPA", + [VF32MF2, VF32M1, VF32M2, VF32M4, VF32M8], isSEWAware=1>; +} + +let Predicates = [HasVendorXSfvfexpa, HasVInstructionsF16] in { + defm : VPatUnaryV_V<"int_riscv_sf_vfexpa", "PseudoSF_VFEXPA", + [VF16MF4, VF16MF2, VF16M1, VF16M2, VF16M4, VF16M8], + isSEWAware=1>; +} + +let Predicates = [HasVendorXSfvfexpa64e] in { + defm : VPatUnaryV_V<"int_riscv_sf_vfexpa", "PseudoSF_VFEXPA", + [VF64M1, VF64M2, VF64M4, VF64M8], isSEWAware=1>; +} diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td index 6b9a75f..5429c2a 100644 --- a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td +++ b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td @@ -599,14 +599,20 @@ def : PatGpr<riscv_zip, ZIP_RV32, i32>; def : PatGpr<riscv_unzip, UNZIP_RV32, i32>; } // Predicates = [HasStdExtZbkb, IsRV32] -let Predicates = [HasStdExtZbb] in { +let Predicates = [HasStdExtZbbOrP] in { def : PatGpr<ctlz, CLZ>; +} + +let Predicates = [HasStdExtZbb] in { def : PatGpr<cttz, CTZ>; def : PatGpr<ctpop, CPOP>; } // Predicates = [HasStdExtZbb] -let Predicates = [HasStdExtZbb, IsRV64] in { +let Predicates = [HasStdExtZbbOrP, IsRV64] in { def : PatGpr<riscv_clzw, CLZW>; +} + +let Predicates = [HasStdExtZbb, IsRV64] in { def : PatGpr<riscv_ctzw, CTZW>; def : Pat<(i64 (ctpop (i64 (zexti32 (i64 GPR:$rs1))))), (CPOPW GPR:$rs1)>; @@ -614,22 +620,22 @@ def : Pat<(i64 (riscv_negw_max GPR:$rs1)), (MAX GPR:$rs1, (XLenVT (SUBW (XLenVT X0), GPR:$rs1)))>; } // Predicates = [HasStdExtZbb, IsRV64] -let Predicates = [HasStdExtZbb] in { +let Predicates = [HasStdExtZbbOrP] in { def : Pat<(XLenVT (sext_inreg GPR:$rs1, i8)), (SEXT_B GPR:$rs1)>; def : Pat<(XLenVT (sext_inreg GPR:$rs1, i16)), (SEXT_H GPR:$rs1)>; } // Predicates = [HasStdExtZbb] -let Predicates = [HasStdExtZbb] in { +let Predicates = [HasStdExtZbbOrP] in { def : PatGprGpr<smin, MIN>; def : PatGprGpr<smax, MAX>; def : PatGprGpr<umin, MINU>; def : PatGprGpr<umax, MAXU>; } // Predicates = [HasStdExtZbb] -let Predicates = [HasStdExtZbbOrZbkb, IsRV32] in +let Predicates = [HasStdExtZbbOrZbkbOrP, IsRV32] in def : PatGpr<bswap, REV8_RV32, i32>; -let Predicates = [HasStdExtZbbOrZbkb, IsRV64] in +let Predicates = [HasStdExtZbbOrZbkbOrP, IsRV64] in def : PatGpr<bswap, REV8_RV64, i64>; let Predicates = [HasStdExtZbkb] in { diff --git a/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td b/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td index 637d61fe..36a2f46 100644 --- a/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td +++ b/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td @@ -1588,6 +1588,10 @@ multiclass SiFive7SchedResources<int vlen, bit dualVALU, //===----------------------------------------------------------------------===// // Unsupported extensions defm : UnsupportedSchedQ; + // TODO: scheduling info of XSfvfexp* and XSfvfexpa* + // for SiFive7 will be added in follow-up patches. + defm : UnsupportedSchedXSfvfexp; + defm : UnsupportedSchedXSfvfexpa; defm : UnsupportedSchedZabha; defm : UnsupportedSchedZbc; defm : UnsupportedSchedZbkb; diff --git a/llvm/lib/Target/RISCV/RISCVSchedule.td b/llvm/lib/Target/RISCV/RISCVSchedule.td index 9ab9636..64ccfd8 100644 --- a/llvm/lib/Target/RISCV/RISCVSchedule.td +++ b/llvm/lib/Target/RISCV/RISCVSchedule.td @@ -523,6 +523,8 @@ include "RISCVScheduleZvk.td" // Vendor Extensions multiclass UnsupportedSchedXsf { defm : UnsupportedSchedXsfvcp; + defm : UnsupportedSchedXSfvfexp; + defm : UnsupportedSchedXSfvfexpa; defm : UnsupportedSchedXSfvfnrclipxfqf; defm : UnsupportedSchedXSfvfwmaccqqq; defm : UnsupportedSchedXSfvqmaccdod; diff --git a/llvm/lib/Target/RISCV/RISCVScheduleXSf.td b/llvm/lib/Target/RISCV/RISCVScheduleXSf.td index 99632e4..1ee6dc1 100644 --- a/llvm/lib/Target/RISCV/RISCVScheduleXSf.td +++ b/llvm/lib/Target/RISCV/RISCVScheduleXSf.td @@ -99,3 +99,23 @@ defm : LMULWriteRes<"WriteSF_VFWMACC_QQQ", []>; defm : LMULReadAdvance<"ReadSF_VFWMACC_QQQ", 0>; } // Unsupported = true } + +defm "" : LMULSEWSchedWritesF<"WriteSF_VFExp">; +defm "" : LMULSEWSchedReadsF<"ReadSF_VFExp">; + +multiclass UnsupportedSchedXSfvfexp { +let Unsupported = true in { +defm : LMULSEWWriteResF<"WriteSF_VFExp", []>; +defm : LMULSEWReadAdvanceF<"ReadSF_VFExp", 0>; +} // Unsupported = true +} + +defm "" : LMULSEWSchedWritesF<"WriteSF_VFExpa">; +defm "" : LMULSEWSchedReadsF<"ReadSF_VFExpa">; + +multiclass UnsupportedSchedXSfvfexpa { +let Unsupported = true in { +defm : LMULSEWWriteResF<"WriteSF_VFExpa", []>; +defm : LMULSEWReadAdvanceF<"ReadSF_VFExpa", 0>; +} // Unsupported = true +} diff --git a/llvm/lib/Target/RISCV/RISCVSubtarget.h b/llvm/lib/Target/RISCV/RISCVSubtarget.h index 334db4b..4b4fc8f 100644 --- a/llvm/lib/Target/RISCV/RISCVSubtarget.h +++ b/llvm/lib/Target/RISCV/RISCVSubtarget.h @@ -187,7 +187,7 @@ public: } bool hasCLZLike() const { - return HasStdExtZbb || HasVendorXTHeadBb || + return HasStdExtZbb || HasStdExtP || HasVendorXTHeadBb || (HasVendorXCVbitmanip && !IsRV64); } bool hasCTZLike() const { @@ -197,7 +197,7 @@ public: return HasStdExtZbb || (HasVendorXCVbitmanip && !IsRV64); } bool hasREV8Like() const { - return HasStdExtZbb || HasStdExtZbkb || HasVendorXTHeadBb; + return HasStdExtZbb || HasStdExtZbkb || HasStdExtP || HasVendorXTHeadBb; } bool hasBEXTILike() const { return HasStdExtZbs || HasVendorXTHeadBs; } diff --git a/llvm/lib/Target/X86/X86.h b/llvm/lib/Target/X86/X86.h index 6261fad..706ab2b 100644 --- a/llvm/lib/Target/X86/X86.h +++ b/llvm/lib/Target/X86/X86.h @@ -160,6 +160,14 @@ FunctionPass *createX86PartialReductionPass(); /// // Analyzes and emits pseudos to support Win x64 Unwind V2. FunctionPass *createX86WinEHUnwindV2Pass(); +/// The pass transforms load/store <256 x i32> to AMX load/store intrinsics +/// or split the data to two <128 x i32>. +FunctionPass *createX86LowerAMXTypePass(); + +/// The pass transforms amx intrinsics to scalar operation if the function has +/// optnone attribute or it is O0. +FunctionPass *createX86LowerAMXIntrinsicsPass(); + InstructionSelector *createX86InstructionSelector(const X86TargetMachine &TM, const X86Subtarget &, const X86RegisterBankInfo &); diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp index 62073ec..4393f6e 100644 --- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp +++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp @@ -4721,9 +4721,6 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) { if (!(Subtarget->hasVLX() || NVT.is512BitVector())) return false; - SDValue N0 = N->getOperand(0); - SDValue N1 = N->getOperand(1); - auto getFoldableLogicOp = [](SDValue Op) { // Peek through single use bitcast. if (Op.getOpcode() == ISD::BITCAST && Op.hasOneUse()) @@ -4740,13 +4737,47 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) { return SDValue(); }; - SDValue A, FoldableOp; - if ((FoldableOp = getFoldableLogicOp(N1))) { - A = N0; - } else if ((FoldableOp = getFoldableLogicOp(N0))) { - A = N1; - } else - return false; + SDValue N0, N1, A, FoldableOp; + + // Identify and (optionally) peel an outer NOT that wraps a pure logic tree + auto tryPeelOuterNotWrappingLogic = [&](SDNode *Op) { + if (Op->getOpcode() == ISD::XOR && Op->hasOneUse() && + ISD::isBuildVectorAllOnes(Op->getOperand(1).getNode())) { + SDValue InnerOp = Op->getOperand(0); + + if (!getFoldableLogicOp(InnerOp)) + return SDValue(); + + N0 = InnerOp.getOperand(0); + N1 = InnerOp.getOperand(1); + if ((FoldableOp = getFoldableLogicOp(N1))) { + A = N0; + return InnerOp; + } + if ((FoldableOp = getFoldableLogicOp(N0))) { + A = N1; + return InnerOp; + } + } + return SDValue(); + }; + + bool PeeledOuterNot = false; + SDNode *OriN = N; + if (SDValue InnerOp = tryPeelOuterNotWrappingLogic(N)) { + PeeledOuterNot = true; + N = InnerOp.getNode(); + } else { + N0 = N->getOperand(0); + N1 = N->getOperand(1); + + if ((FoldableOp = getFoldableLogicOp(N1))) + A = N0; + else if ((FoldableOp = getFoldableLogicOp(N0))) + A = N1; + else + return false; + } SDValue B = FoldableOp.getOperand(0); SDValue C = FoldableOp.getOperand(1); @@ -4798,7 +4829,10 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) { case ISD::XOR: Imm ^= TernlogMagicA; break; } - return matchVPTERNLOG(N, ParentA, ParentB, ParentC, A, B, C, Imm); + if (PeeledOuterNot) + Imm = ~Imm; + + return matchVPTERNLOG(OriN, ParentA, ParentB, ParentC, A, B, C, Imm); } /// If the high bits of an 'and' operand are known zero, try setting the diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 4dfc400..410f20e 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -57617,10 +57617,10 @@ static SDValue combineX86AddSub(SDNode *N, SelectionDAG &DAG, } // Fold any similar generic ADD/SUB opcodes to reuse this node. - auto MatchGeneric = [&](SDValue N0, SDValue N1, bool Negate) { + auto MatchGeneric = [&](unsigned Opc, SDValue N0, SDValue N1, bool Negate) { SDValue Ops[] = {N0, N1}; SDVTList VTs = DAG.getVTList(N->getValueType(0)); - if (SDNode *GenericAddSub = DAG.getNodeIfExists(GenericOpc, VTs, Ops)) { + if (SDNode *GenericAddSub = DAG.getNodeIfExists(Opc, VTs, Ops)) { SDValue Op(N, 0); if (Negate) { // Bail if this is only used by a user of the x86 add/sub. @@ -57632,8 +57632,25 @@ static SDValue combineX86AddSub(SDNode *N, SelectionDAG &DAG, DCI.CombineTo(GenericAddSub, Op); } }; - MatchGeneric(LHS, RHS, false); - MatchGeneric(RHS, LHS, X86ISD::SUB == N->getOpcode()); + MatchGeneric(GenericOpc, LHS, RHS, false); + MatchGeneric(GenericOpc, RHS, LHS, X86ISD::SUB == N->getOpcode()); + + if (auto *Const = dyn_cast<ConstantSDNode>(RHS)) { + SDValue NegC = DAG.getConstant(-Const->getAPIntValue(), DL, VT); + if (X86ISD::SUB == N->getOpcode()) { + // Fold generic add(LHS, -C) to X86ISD::SUB(LHS, C). + MatchGeneric(ISD::ADD, LHS, NegC, false); + } else { + // Negate X86ISD::ADD(LHS, C) and replace generic sub(-C, LHS). + MatchGeneric(ISD::SUB, NegC, LHS, true); + } + } else if (auto *Const = dyn_cast<ConstantSDNode>(LHS)) { + if (X86ISD::SUB == N->getOpcode()) { + SDValue NegC = DAG.getConstant(-Const->getAPIntValue(), DL, VT); + // Negate X86ISD::SUB(C, RHS) and replace generic add(RHS, -C). + MatchGeneric(ISD::ADD, RHS, NegC, true); + } + } // TODO: Can we drop the ZeroSecondOpOnly limit? This is to guarantee that the // EFLAGS result doesn't change. diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h index e28b9c1..b7151f6 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.h +++ b/llvm/lib/Target/X86/X86ISelLowering.h @@ -1592,7 +1592,6 @@ namespace llvm { bool useLoadStackGuardNode(const Module &M) const override; bool useStackGuardXorFP() const override; void insertSSPDeclarations(Module &M) const override; - Function *getSSPStackGuardCheck(const Module &M) const override; SDValue emitStackGuardXorFP(SelectionDAG &DAG, SDValue Val, const SDLoc &DL) const override; diff --git a/llvm/lib/Target/X86/X86ISelLoweringCall.cpp b/llvm/lib/Target/X86/X86ISelLoweringCall.cpp index 37d7772..a61bbe5 100644 --- a/llvm/lib/Target/X86/X86ISelLoweringCall.cpp +++ b/llvm/lib/Target/X86/X86ISelLoweringCall.cpp @@ -640,15 +640,6 @@ void X86TargetLowering::insertSSPDeclarations(Module &M) const { TargetLowering::insertSSPDeclarations(M); } -Function *X86TargetLowering::getSSPStackGuardCheck(const Module &M) const { - // MSVC CRT has a function to validate security cookie. - if (Subtarget.getTargetTriple().isWindowsMSVCEnvironment() || - Subtarget.getTargetTriple().isWindowsItaniumEnvironment()) { - return M.getFunction("__security_check_cookie"); - } - return TargetLowering::getSSPStackGuardCheck(M); -} - Value * X86TargetLowering::getSafeStackPointerLocation(IRBuilderBase &IRB) const { // Android provides a fixed TLS slot for the SafeStack pointer. See the diff --git a/llvm/lib/Target/Xtensa/XtensaInstrInfo.td b/llvm/lib/Target/Xtensa/XtensaInstrInfo.td index edcf247..632c6a2 100644 --- a/llvm/lib/Target/Xtensa/XtensaInstrInfo.td +++ b/llvm/lib/Target/Xtensa/XtensaInstrInfo.td @@ -1407,7 +1407,7 @@ let isBarrier = 1, isTerminator = 1 in { let r = 0x04; } - def BREAK_N : RRRN_Inst<0x0C, (outs), (ins uimm4:$imm), + def BREAK_N : RRRN_Inst<0x0D, (outs), (ins uimm4:$imm), "break.n\t$imm", []>, Requires<[HasDensity, HasDebug]> { bits<4> imm; diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp index 669d4f0..8d9933b 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp @@ -582,6 +582,18 @@ static Instruction *foldCttzCtlz(IntrinsicInst &II, InstCombinerImpl &IC) { IC.Builder.CreateBinaryIntrinsic(Intrinsic::ctlz, C, Op1); return BinaryOperator::CreateSub(ConstCtlz, X); } + + // ctlz(~x & (x - 1)) -> bitwidth - cttz(x, false) + if (Op0->hasOneUse() && + match(Op0, + m_c_And(m_Not(m_Value(X)), m_Add(m_Deferred(X), m_AllOnes())))) { + Type *Ty = II.getType(); + unsigned BitWidth = Ty->getScalarSizeInBits(); + auto *Cttz = IC.Builder.CreateIntrinsic(Intrinsic::cttz, Ty, + {X, IC.Builder.getFalse()}); + auto *Bw = ConstantInt::get(Ty, APInt(BitWidth, BitWidth)); + return IC.replaceInstUsesWith(II, IC.Builder.CreateSub(Bw, Cttz)); + } } // cttz(Pow2) -> Log2(Pow2) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp index 5aa8de3..f5130da 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -4697,5 +4697,31 @@ Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { cast<IntrinsicInst>(TrueVal)->getParamAlign(0).valueOrOne(), CondVal, FalseVal)); + // Canonicalize sign function ashr pattern: select (icmp slt X, 1), ashr X, + // bitwidth-1, 1 -> scmp(X, 0) + // Also handles: select (icmp sgt X, 0), 1, ashr X, bitwidth-1 -> scmp(X, 0) + unsigned BitWidth = SI.getType()->getScalarSizeInBits(); + CmpPredicate Pred; + Value *CmpLHS, *CmpRHS; + + // Canonicalize sign function ashr patterns: + // select (icmp slt X, 1), ashr X, bitwidth-1, 1 -> scmp(X, 0) + // select (icmp sgt X, 0), 1, ashr X, bitwidth-1 -> scmp(X, 0) + if (match(&SI, m_Select(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)), + m_Value(TrueVal), m_Value(FalseVal))) && + ((Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_One()) && + match(TrueVal, + m_AShr(m_Specific(CmpLHS), m_SpecificInt(BitWidth - 1))) && + match(FalseVal, m_One())) || + (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_Zero()) && + match(TrueVal, m_One()) && + match(FalseVal, + m_AShr(m_Specific(CmpLHS), m_SpecificInt(BitWidth - 1)))))) { + + Function *Scmp = Intrinsic::getOrInsertDeclaration( + SI.getModule(), Intrinsic::scmp, {SI.getType(), SI.getType()}); + return CallInst::Create(Scmp, {CmpLHS, ConstantInt::get(SI.getType(), 0)}); + } + return nullptr; } diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp index 67e2aae..9c8de45 100644 --- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp +++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp @@ -2327,6 +2327,18 @@ Constant *InstCombinerImpl::unshuffleConstant(ArrayRef<int> ShMask, Constant *C, return ConstantVector::get(NewVecC); } +// Get the result of `Vector Op Splat` (or Splat Op Vector if \p SplatLHS). +static Constant *constantFoldBinOpWithSplat(unsigned Opcode, Constant *Vector, + Constant *Splat, bool SplatLHS, + const DataLayout &DL) { + ElementCount EC = cast<VectorType>(Vector->getType())->getElementCount(); + Constant *LHS = ConstantVector::getSplat(EC, Splat); + Constant *RHS = Vector; + if (!SplatLHS) + std::swap(LHS, RHS); + return ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL); +} + Instruction *InstCombinerImpl::foldVectorBinop(BinaryOperator &Inst) { if (!isa<VectorType>(Inst.getType())) return nullptr; @@ -2338,6 +2350,37 @@ Instruction *InstCombinerImpl::foldVectorBinop(BinaryOperator &Inst) { assert(cast<VectorType>(RHS->getType())->getElementCount() == cast<VectorType>(Inst.getType())->getElementCount()); + auto foldConstantsThroughSubVectorInsertSplat = + [&](Value *MaybeSubVector, Value *MaybeSplat, + bool SplatLHS) -> Instruction * { + Value *Idx; + Constant *Splat, *SubVector, *Dest; + if (!match(MaybeSplat, m_ConstantSplat(m_Constant(Splat))) || + !match(MaybeSubVector, + m_VectorInsert(m_Constant(Dest), m_Constant(SubVector), + m_Value(Idx)))) + return nullptr; + SubVector = + constantFoldBinOpWithSplat(Opcode, SubVector, Splat, SplatLHS, DL); + Dest = constantFoldBinOpWithSplat(Opcode, Dest, Splat, SplatLHS, DL); + if (!SubVector || !Dest) + return nullptr; + auto *InsertVector = + Builder.CreateInsertVector(Dest->getType(), Dest, SubVector, Idx); + return replaceInstUsesWith(Inst, InsertVector); + }; + + // If one operand is a constant splat and the other operand is a + // `vector.insert` where both the destination and subvector are constant, + // apply the operation to both the destination and subvector, returning a new + // constant `vector.insert`. This helps constant folding for scalable vectors. + if (Instruction *Folded = foldConstantsThroughSubVectorInsertSplat( + /*MaybeSubVector=*/LHS, /*MaybeSplat=*/RHS, /*SplatLHS=*/false)) + return Folded; + if (Instruction *Folded = foldConstantsThroughSubVectorInsertSplat( + /*MaybeSubVector=*/RHS, /*MaybeSplat=*/LHS, /*SplatLHS=*/true)) + return Folded; + // If both operands of the binop are vector concatenations, then perform the // narrow binop on each pair of the source operands followed by concatenation // of the results. diff --git a/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp b/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp index b6cbecb..10b03bb 100644 --- a/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp +++ b/llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp @@ -226,6 +226,7 @@ static const Align kMinOriginAlignment = Align(4); static const Align kShadowTLSAlignment = Align(8); // These constants must be kept in sync with the ones in msan.h. +// TODO: increase size to match SVE/SVE2/SME/SME2 limits static const unsigned kParamTLSSize = 800; static const unsigned kRetvalTLSSize = 800; @@ -1544,6 +1545,22 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> { } } + static bool isAArch64SVCount(Type *Ty) { + if (TargetExtType *TTy = dyn_cast<TargetExtType>(Ty)) + return TTy->getName() == "aarch64.svcount"; + return false; + } + + // This is intended to match the "AArch64 Predicate-as-Counter Type" (aka + // 'target("aarch64.svcount")', but not e.g., <vscale x 4 x i32>. + static bool isScalableNonVectorType(Type *Ty) { + if (!isAArch64SVCount(Ty)) + LLVM_DEBUG(dbgs() << "isScalableNonVectorType: Unexpected type " << *Ty + << "\n"); + + return Ty->isScalableTy() && !isa<VectorType>(Ty); + } + void materializeChecks() { #ifndef NDEBUG // For assert below. @@ -1672,6 +1689,12 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> { LLVM_DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n"); return Res; } + if (isScalableNonVectorType(OrigTy)) { + LLVM_DEBUG(dbgs() << "getShadowTy: Scalable non-vector type: " << *OrigTy + << "\n"); + return OrigTy; + } + uint32_t TypeSize = DL.getTypeSizeInBits(OrigTy); return IntegerType::get(*MS.C, TypeSize); } @@ -2185,8 +2208,14 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> { << *OrigIns << "\n"); return; } -#ifndef NDEBUG + Type *ShadowTy = Shadow->getType(); + if (isScalableNonVectorType(ShadowTy)) { + LLVM_DEBUG(dbgs() << "Skipping check of scalable non-vector " << *Shadow + << " before " << *OrigIns << "\n"); + return; + } +#ifndef NDEBUG assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy) || isa<StructType>(ShadowTy) || isa<ArrayType>(ShadowTy)) && "Can only insert checks for integer, vector, and aggregate shadow " @@ -6972,6 +7001,15 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> { // an extra "select". This results in much more compact IR. // Sa = select Sb, poisoned, (select b, Sc, Sd) Sa1 = getPoisonedShadow(getShadowTy(I.getType())); + } else if (isScalableNonVectorType(I.getType())) { + // This is intended to handle target("aarch64.svcount"), which can't be + // handled in the else branch because of incompatibility with CreateXor + // ("The supported LLVM operations on this type are limited to load, + // store, phi, select and alloca instructions"). + + // TODO: this currently underapproximates. Use Arm SVE EOR in the else + // branch as needed instead. + Sa1 = getCleanShadow(getShadowTy(I.getType())); } else { // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ] // If Sb (condition is poisoned), look for bits in c and d that are equal diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h index a1ad2db..2591df8 100644 --- a/llvm/lib/Transforms/Vectorize/VPlan.h +++ b/llvm/lib/Transforms/Vectorize/VPlan.h @@ -4172,11 +4172,6 @@ class VPlan { /// definitions are VPValues that hold a pointer to their underlying IR. SmallVector<VPValue *, 16> VPLiveIns; - /// Mapping from SCEVs to the VPValues representing their expansions. - /// NOTE: This mapping is temporary and will be removed once all users have - /// been modeled in VPlan directly. - DenseMap<const SCEV *, VPValue *> SCEVToExpansion; - /// Blocks allocated and owned by the VPlan. They will be deleted once the /// VPlan is destroyed. SmallVector<VPBlockBase *> CreatedBlocks; @@ -4424,15 +4419,6 @@ public: LLVM_DUMP_METHOD void dump() const; #endif - VPValue *getSCEVExpansion(const SCEV *S) const { - return SCEVToExpansion.lookup(S); - } - - void addSCEVExpansion(const SCEV *S, VPValue *V) { - assert(!SCEVToExpansion.contains(S) && "SCEV already expanded"); - SCEVToExpansion[S] = V; - } - /// Clone the current VPlan, update all VPValues of the new VPlan and cloned /// recipes to refer to the clones, and return it. VPlan *duplicate(); diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp index c385c36..84817d7 100644 --- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp +++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp @@ -943,12 +943,40 @@ static void recursivelyDeleteDeadRecipes(VPValue *V) { } } +/// Get any instruction opcode or intrinsic ID data embedded in recipe \p R. +/// Returns an optional pair, where the first element indicates whether it is +/// an intrinsic ID. +static std::optional<std::pair<bool, unsigned>> +getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) { + return TypeSwitch<const VPSingleDefRecipe *, + std::optional<std::pair<bool, unsigned>>>(R) + .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe, + VPWidenSelectRecipe, VPWidenGEPRecipe, VPReplicateRecipe>( + [](auto *I) { return std::make_pair(false, I->getOpcode()); }) + .Case<VPWidenIntrinsicRecipe>([](auto *I) { + return std::make_pair(true, I->getVectorIntrinsicID()); + }) + .Case<VPVectorPointerRecipe, VPPredInstPHIRecipe>([](auto *I) { + // For recipes that do not directly map to LLVM IR instructions, + // assign opcodes after the last VPInstruction opcode (which is also + // after the last IR Instruction opcode), based on the VPDefID. + return std::make_pair(false, + VPInstruction::OpsEnd + 1 + I->getVPDefID()); + }) + .Default([](auto *) { return std::nullopt; }); +} + /// Try to fold \p R using InstSimplifyFolder. Will succeed and return a -/// non-nullptr Value for a handled \p Opcode if corresponding \p Operands are -/// foldable live-ins. -static Value *tryToFoldLiveIns(const VPRecipeBase &R, unsigned Opcode, - ArrayRef<VPValue *> Operands, - const DataLayout &DL, VPTypeAnalysis &TypeInfo) { +/// non-nullptr VPValue for a handled opcode or intrinsic ID if corresponding \p +/// Operands are foldable live-ins. +static VPValue *tryToFoldLiveIns(VPSingleDefRecipe &R, + ArrayRef<VPValue *> Operands, + const DataLayout &DL, + VPTypeAnalysis &TypeInfo) { + auto OpcodeOrIID = getOpcodeOrIntrinsicID(&R); + if (!OpcodeOrIID) + return nullptr; + SmallVector<Value *, 4> Ops; for (VPValue *Op : Operands) { if (!Op->isLiveIn() || !Op->getLiveInIRValue()) @@ -956,43 +984,57 @@ static Value *tryToFoldLiveIns(const VPRecipeBase &R, unsigned Opcode, Ops.push_back(Op->getLiveInIRValue()); } - InstSimplifyFolder Folder(DL); - if (Instruction::isBinaryOp(Opcode)) - return Folder.FoldBinOp(static_cast<Instruction::BinaryOps>(Opcode), Ops[0], + auto FoldToIRValue = [&]() -> Value * { + InstSimplifyFolder Folder(DL); + if (OpcodeOrIID->first) { + if (R.getNumOperands() != 2) + return nullptr; + unsigned ID = OpcodeOrIID->second; + return Folder.FoldBinaryIntrinsic(ID, Ops[0], Ops[1], + TypeInfo.inferScalarType(&R)); + } + unsigned Opcode = OpcodeOrIID->second; + if (Instruction::isBinaryOp(Opcode)) + return Folder.FoldBinOp(static_cast<Instruction::BinaryOps>(Opcode), + Ops[0], Ops[1]); + if (Instruction::isCast(Opcode)) + return Folder.FoldCast(static_cast<Instruction::CastOps>(Opcode), Ops[0], + TypeInfo.inferScalarType(R.getVPSingleValue())); + switch (Opcode) { + case VPInstruction::LogicalAnd: + return Folder.FoldSelect(Ops[0], Ops[1], + ConstantInt::getNullValue(Ops[1]->getType())); + case VPInstruction::Not: + return Folder.FoldBinOp(Instruction::BinaryOps::Xor, Ops[0], + Constant::getAllOnesValue(Ops[0]->getType())); + case Instruction::Select: + return Folder.FoldSelect(Ops[0], Ops[1], Ops[2]); + case Instruction::ICmp: + case Instruction::FCmp: + return Folder.FoldCmp(cast<VPRecipeWithIRFlags>(R).getPredicate(), Ops[0], Ops[1]); - if (Instruction::isCast(Opcode)) - return Folder.FoldCast(static_cast<Instruction::CastOps>(Opcode), Ops[0], - TypeInfo.inferScalarType(R.getVPSingleValue())); - switch (Opcode) { - case VPInstruction::LogicalAnd: - return Folder.FoldSelect(Ops[0], Ops[1], - ConstantInt::getNullValue(Ops[1]->getType())); - case VPInstruction::Not: - return Folder.FoldBinOp(Instruction::BinaryOps::Xor, Ops[0], - Constant::getAllOnesValue(Ops[0]->getType())); - case Instruction::Select: - return Folder.FoldSelect(Ops[0], Ops[1], Ops[2]); - case Instruction::ICmp: - case Instruction::FCmp: - return Folder.FoldCmp(cast<VPRecipeWithIRFlags>(R).getPredicate(), Ops[0], - Ops[1]); - case Instruction::GetElementPtr: { - auto &RFlags = cast<VPRecipeWithIRFlags>(R); - auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr()); - return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0], drop_begin(Ops), - RFlags.getGEPNoWrapFlags()); - } - case VPInstruction::PtrAdd: - case VPInstruction::WidePtrAdd: - return Folder.FoldGEP(IntegerType::getInt8Ty(TypeInfo.getContext()), Ops[0], - Ops[1], - cast<VPRecipeWithIRFlags>(R).getGEPNoWrapFlags()); - // An extract of a live-in is an extract of a broadcast, so return the - // broadcasted element. - case Instruction::ExtractElement: - assert(!Ops[0]->getType()->isVectorTy() && "Live-ins should be scalar"); - return Ops[0]; - } + case Instruction::GetElementPtr: { + auto &RFlags = cast<VPRecipeWithIRFlags>(R); + auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr()); + return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0], + drop_begin(Ops), RFlags.getGEPNoWrapFlags()); + } + case VPInstruction::PtrAdd: + case VPInstruction::WidePtrAdd: + return Folder.FoldGEP(IntegerType::getInt8Ty(TypeInfo.getContext()), + Ops[0], Ops[1], + cast<VPRecipeWithIRFlags>(R).getGEPNoWrapFlags()); + // An extract of a live-in is an extract of a broadcast, so return the + // broadcasted element. + case Instruction::ExtractElement: + assert(!Ops[0]->getType()->isVectorTy() && "Live-ins should be scalar"); + return Ops[0]; + } + return nullptr; + }; + + if (Value *V = FoldToIRValue()) + return R.getParent()->getPlan()->getOrAddLiveIn(V); return nullptr; } @@ -1006,19 +1048,10 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) { // Simplification of live-in IR values for SingleDef recipes using // InstSimplifyFolder. - if (TypeSwitch<VPRecipeBase *, bool>(&R) - .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe, - VPReplicateRecipe, VPWidenSelectRecipe>([&](auto *I) { - const DataLayout &DL = - Plan->getScalarHeader()->getIRBasicBlock()->getDataLayout(); - Value *V = tryToFoldLiveIns(*I, I->getOpcode(), I->operands(), DL, - TypeInfo); - if (V) - I->replaceAllUsesWith(Plan->getOrAddLiveIn(V)); - return V; - }) - .Default([](auto *) { return false; })) - return; + const DataLayout &DL = + Plan->getScalarHeader()->getIRBasicBlock()->getDataLayout(); + if (VPValue *V = tryToFoldLiveIns(*Def, Def->operands(), DL, TypeInfo)) + return Def->replaceAllUsesWith(V); // Fold PredPHI LiveIn -> LiveIn. if (auto *PredPHI = dyn_cast<VPPredInstPHIRecipe>(&R)) { @@ -1996,29 +2029,6 @@ struct VPCSEDenseMapInfo : public DenseMapInfo<VPSingleDefRecipe *> { return Def == getEmptyKey() || Def == getTombstoneKey(); } - /// Get any instruction opcode or intrinsic ID data embedded in recipe \p R. - /// Returns an optional pair, where the first element indicates whether it is - /// an intrinsic ID. - static std::optional<std::pair<bool, unsigned>> - getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) { - return TypeSwitch<const VPSingleDefRecipe *, - std::optional<std::pair<bool, unsigned>>>(R) - .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe, - VPWidenSelectRecipe, VPWidenGEPRecipe, VPReplicateRecipe>( - [](auto *I) { return std::make_pair(false, I->getOpcode()); }) - .Case<VPWidenIntrinsicRecipe>([](auto *I) { - return std::make_pair(true, I->getVectorIntrinsicID()); - }) - .Case<VPVectorPointerRecipe, VPPredInstPHIRecipe>([](auto *I) { - // For recipes that do not directly map to LLVM IR instructions, - // assign opcodes after the last VPInstruction opcode (which is also - // after the last IR Instruction opcode), based on the VPDefID. - return std::make_pair(false, - VPInstruction::OpsEnd + 1 + I->getVPDefID()); - }) - .Default([](auto *) { return std::nullopt; }); - } - /// If recipe \p R will lower to a GEP with a non-i8 source element type, /// return that source element type. static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) { diff --git a/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp b/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp index 06c3d75..fe66f13 100644 --- a/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp +++ b/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp @@ -32,8 +32,6 @@ bool vputils::onlyScalarValuesUsed(const VPValue *Def) { } VPValue *vputils::getOrCreateVPValueForSCEVExpr(VPlan &Plan, const SCEV *Expr) { - if (auto *Expanded = Plan.getSCEVExpansion(Expr)) - return Expanded; VPValue *Expanded = nullptr; if (auto *E = dyn_cast<SCEVConstant>(Expr)) Expanded = Plan.getOrAddLiveIn(E->getValue()); @@ -50,7 +48,6 @@ VPValue *vputils::getOrCreateVPValueForSCEVExpr(VPlan &Plan, const SCEV *Expr) { Plan.getEntry()->appendRecipe(Expanded->getDefiningRecipe()); } } - Plan.addSCEVExpansion(Expr, Expanded); return Expanded; } |