//===-- RISCVTargetTransformInfo.cpp - RISC-V specific TTI ----------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// #include "RISCVTargetTransformInfo.h" #include "MCTargetDesc/RISCVMatInt.h" #include "llvm/ADT/STLExtras.h" #include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/CodeGen/BasicTTIImpl.h" #include "llvm/CodeGen/CostTable.h" #include "llvm/CodeGen/TargetLowering.h" #include "llvm/CodeGen/ValueTypes.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/PatternMatch.h" #include #include using namespace llvm; using namespace llvm::PatternMatch; #define DEBUG_TYPE "riscvtti" static cl::opt RVVRegisterWidthLMUL( "riscv-v-register-bit-width-lmul", cl::desc( "The LMUL to use for getRegisterBitWidth queries. Affects LMUL used " "by autovectorized code. Fractional LMULs are not supported."), cl::init(2), cl::Hidden); static cl::opt SLPMaxVF( "riscv-v-slp-max-vf", cl::desc( "Overrides result used for getMaximumVF query which is used " "exclusively by SLP vectorizer."), cl::Hidden); static cl::opt RVVMinTripCount("riscv-v-min-trip-count", cl::desc("Set the lower bound of a trip count to decide on " "vectorization while tail-folding."), cl::init(5), cl::Hidden); InstructionCost RISCVTTIImpl::getRISCVInstructionCost(ArrayRef OpCodes, MVT VT, TTI::TargetCostKind CostKind) const { // Check if the type is valid for all CostKind if (!VT.isVector()) return InstructionCost::getInvalid(); size_t NumInstr = OpCodes.size(); if (CostKind == TTI::TCK_CodeSize) return NumInstr; InstructionCost LMULCost = TLI->getLMULCost(VT); if ((CostKind != TTI::TCK_RecipThroughput) && (CostKind != TTI::TCK_Latency)) return LMULCost * NumInstr; InstructionCost Cost = 0; for (auto Op : OpCodes) { switch (Op) { case RISCV::VRGATHER_VI: Cost += TLI->getVRGatherVICost(VT); break; case RISCV::VRGATHER_VV: Cost += TLI->getVRGatherVVCost(VT); break; case RISCV::VSLIDEUP_VI: case RISCV::VSLIDEDOWN_VI: Cost += TLI->getVSlideVICost(VT); break; case RISCV::VSLIDEUP_VX: case RISCV::VSLIDEDOWN_VX: Cost += TLI->getVSlideVXCost(VT); break; case RISCV::VREDMAX_VS: case RISCV::VREDMIN_VS: case RISCV::VREDMAXU_VS: case RISCV::VREDMINU_VS: case RISCV::VREDSUM_VS: case RISCV::VREDAND_VS: case RISCV::VREDOR_VS: case RISCV::VREDXOR_VS: case RISCV::VFREDMAX_VS: case RISCV::VFREDMIN_VS: case RISCV::VFREDUSUM_VS: { unsigned VL = VT.getVectorMinNumElements(); if (!VT.isFixedLengthVector()) VL *= *getVScaleForTuning(); Cost += Log2_32_Ceil(VL); break; } case RISCV::VFREDOSUM_VS: { unsigned VL = VT.getVectorMinNumElements(); if (!VT.isFixedLengthVector()) VL *= *getVScaleForTuning(); Cost += VL; break; } case RISCV::VMV_X_S: case RISCV::VMV_S_X: case RISCV::VFMV_F_S: case RISCV::VFMV_S_F: case RISCV::VMOR_MM: case RISCV::VMXOR_MM: case RISCV::VMAND_MM: case RISCV::VMANDN_MM: case RISCV::VMNAND_MM: case RISCV::VCPOP_M: case RISCV::VFIRST_M: Cost += 1; break; default: Cost += LMULCost; } } return Cost; } static InstructionCost getIntImmCostImpl(const DataLayout &DL, const RISCVSubtarget *ST, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind, bool FreeZeroes) { assert(Ty->isIntegerTy() && "getIntImmCost can only estimate cost of materialising integers"); // We have a Zero register, so 0 is always free. if (Imm == 0) return TTI::TCC_Free; // Otherwise, we check how many instructions it will take to materialise. return RISCVMatInt::getIntMatCost(Imm, DL.getTypeSizeInBits(Ty), *ST, /*CompressionCost=*/false, FreeZeroes); } InstructionCost RISCVTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind) const { return getIntImmCostImpl(getDataLayout(), getST(), Imm, Ty, CostKind, false); } // Look for patterns of shift followed by AND that can be turned into a pair of // shifts. We won't need to materialize an immediate for the AND so these can // be considered free. static bool canUseShiftPair(Instruction *Inst, const APInt &Imm) { uint64_t Mask = Imm.getZExtValue(); auto *BO = dyn_cast(Inst->getOperand(0)); if (!BO || !BO->hasOneUse()) return false; if (BO->getOpcode() != Instruction::Shl) return false; if (!isa(BO->getOperand(1))) return false; unsigned ShAmt = cast(BO->getOperand(1))->getZExtValue(); // (and (shl x, c2), c1) will be matched to (srli (slli x, c2+c3), c3) if c1 // is a mask shifted by c2 bits with c3 leading zeros. if (isShiftedMask_64(Mask)) { unsigned Trailing = llvm::countr_zero(Mask); if (ShAmt == Trailing) return true; } return false; } InstructionCost RISCVTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind, Instruction *Inst) const { assert(Ty->isIntegerTy() && "getIntImmCost can only estimate cost of materialising integers"); // We have a Zero register, so 0 is always free. if (Imm == 0) return TTI::TCC_Free; // Some instructions in RISC-V can take a 12-bit immediate. Some of these are // commutative, in others the immediate comes from a specific argument index. bool Takes12BitImm = false; unsigned ImmArgIdx = ~0U; switch (Opcode) { case Instruction::GetElementPtr: // Never hoist any arguments to a GetElementPtr. CodeGenPrepare will // split up large offsets in GEP into better parts than ConstantHoisting // can. return TTI::TCC_Free; case Instruction::Store: { // Use the materialization cost regardless of if it's the address or the // value that is constant, except for if the store is misaligned and // misaligned accesses are not legal (experience shows constant hoisting // can sometimes be harmful in such cases). if (Idx == 1 || !Inst) return getIntImmCostImpl(getDataLayout(), getST(), Imm, Ty, CostKind, /*FreeZeroes=*/true); StoreInst *ST = cast(Inst); if (!getTLI()->allowsMemoryAccessForAlignment( Ty->getContext(), DL, getTLI()->getValueType(DL, Ty), ST->getPointerAddressSpace(), ST->getAlign())) return TTI::TCC_Free; return getIntImmCostImpl(getDataLayout(), getST(), Imm, Ty, CostKind, /*FreeZeroes=*/true); } case Instruction::Load: // If the address is a constant, use the materialization cost. return getIntImmCost(Imm, Ty, CostKind); case Instruction::And: // zext.h if (Imm == UINT64_C(0xffff) && ST->hasStdExtZbb()) return TTI::TCC_Free; // zext.w if (Imm == UINT64_C(0xffffffff) && ((ST->hasStdExtZba() && ST->isRV64()) || ST->isRV32())) return TTI::TCC_Free; // bclri if (ST->hasStdExtZbs() && (~Imm).isPowerOf2()) return TTI::TCC_Free; if (Inst && Idx == 1 && Imm.getBitWidth() <= ST->getXLen() && canUseShiftPair(Inst, Imm)) return TTI::TCC_Free; Takes12BitImm = true; break; case Instruction::Add: Takes12BitImm = true; break; case Instruction::Or: case Instruction::Xor: // bseti/binvi if (ST->hasStdExtZbs() && Imm.isPowerOf2()) return TTI::TCC_Free; Takes12BitImm = true; break; case Instruction::Mul: // Power of 2 is a shift. Negated power of 2 is a shift and a negate. if (Imm.isPowerOf2() || Imm.isNegatedPowerOf2()) return TTI::TCC_Free; // One more or less than a power of 2 can use SLLI+ADD/SUB. if ((Imm + 1).isPowerOf2() || (Imm - 1).isPowerOf2()) return TTI::TCC_Free; // FIXME: There is no MULI instruction. Takes12BitImm = true; break; case Instruction::Sub: case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: Takes12BitImm = true; ImmArgIdx = 1; break; default: break; } if (Takes12BitImm) { // Check immediate is the correct argument... if (Instruction::isCommutative(Opcode) || Idx == ImmArgIdx) { // ... and fits into the 12-bit immediate. if (Imm.getSignificantBits() <= 64 && getTLI()->isLegalAddImmediate(Imm.getSExtValue())) { return TTI::TCC_Free; } } // Otherwise, use the full materialisation cost. return getIntImmCost(Imm, Ty, CostKind); } // By default, prevent hoisting. return TTI::TCC_Free; } InstructionCost RISCVTTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind) const { // Prevent hoisting in unknown cases. return TTI::TCC_Free; } bool RISCVTTIImpl::hasActiveVectorLength() const { return ST->hasVInstructions(); } TargetTransformInfo::PopcntSupportKind RISCVTTIImpl::getPopcntSupport(unsigned TyWidth) const { assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); return ST->hasStdExtZbb() || (ST->hasVendorXCVbitmanip() && !ST->is64Bit()) ? TTI::PSK_FastHardware : TTI::PSK_Software; } InstructionCost RISCVTTIImpl::getPartialReductionCost( unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType, ElementCount VF, TTI::PartialReductionExtendKind OpAExtend, TTI::PartialReductionExtendKind OpBExtend, std::optional BinOp, TTI::TargetCostKind CostKind) const { // zve32x is broken for partial_reduce_umla, but let's make sure we // don't generate them. if (!ST->hasStdExtZvqdotq() || ST->getELen() < 64 || Opcode != Instruction::Add || !BinOp || *BinOp != Instruction::Mul || InputTypeA != InputTypeB || !InputTypeA->isIntegerTy(8) || !AccumType->isIntegerTy(32) || !VF.isKnownMultipleOf(4)) return InstructionCost::getInvalid(); Type *Tp = VectorType::get(AccumType, VF.divideCoefficientBy(4)); std::pair LT = getTypeLegalizationCost(Tp); // Note: Asuming all vqdot* variants are equal cost return LT.first * getRISCVInstructionCost(RISCV::VQDOT_VV, LT.second, CostKind); } bool RISCVTTIImpl::shouldExpandReduction(const IntrinsicInst *II) const { // Currently, the ExpandReductions pass can't expand scalable-vector // reductions, but we still request expansion as RVV doesn't support certain // reductions and the SelectionDAG can't legalize them either. switch (II->getIntrinsicID()) { default: return false; // These reductions have no equivalent in RVV case Intrinsic::vector_reduce_mul: case Intrinsic::vector_reduce_fmul: return true; } } std::optional RISCVTTIImpl::getMaxVScale() const { if (ST->hasVInstructions()) return ST->getRealMaxVLen() / RISCV::RVVBitsPerBlock; return BaseT::getMaxVScale(); } std::optional RISCVTTIImpl::getVScaleForTuning() const { if (ST->hasVInstructions()) if (unsigned MinVLen = ST->getRealMinVLen(); MinVLen >= RISCV::RVVBitsPerBlock) return MinVLen / RISCV::RVVBitsPerBlock; return BaseT::getVScaleForTuning(); } TypeSize RISCVTTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const { unsigned LMUL = llvm::bit_floor(std::clamp(RVVRegisterWidthLMUL, 1, 8)); switch (K) { case TargetTransformInfo::RGK_Scalar: return TypeSize::getFixed(ST->getXLen()); case TargetTransformInfo::RGK_FixedWidthVector: return TypeSize::getFixed( ST->useRVVForFixedLengthVectors() ? LMUL * ST->getRealMinVLen() : 0); case TargetTransformInfo::RGK_ScalableVector: return TypeSize::getScalable( (ST->hasVInstructions() && ST->getRealMinVLen() >= RISCV::RVVBitsPerBlock) ? LMUL * RISCV::RVVBitsPerBlock : 0); } llvm_unreachable("Unsupported register kind"); } InstructionCost RISCVTTIImpl::getConstantPoolLoadCost(Type *Ty, TTI::TargetCostKind CostKind) const { // Add a cost of address generation + the cost of the load. The address // is expected to be a PC relative offset to a constant pool entry // using auipc/addi. return 2 + getMemoryOpCost(Instruction::Load, Ty, DL.getABITypeAlign(Ty), /*AddressSpace=*/0, CostKind); } static bool isRepeatedConcatMask(ArrayRef Mask, int &SubVectorSize) { unsigned Size = Mask.size(); if (!isPowerOf2_32(Size)) return false; for (unsigned I = 0; I != Size; ++I) { if (static_cast(Mask[I]) == I) continue; if (Mask[I] != 0) return false; if (Size % I != 0) return false; for (unsigned J = I + 1; J != Size; ++J) // Check the pattern is repeated. if (static_cast(Mask[J]) != J % I) return false; SubVectorSize = I; return true; } // That means Mask is <0, 1, 2, 3>. This is not a concatenation. return false; } static VectorType *getVRGatherIndexType(MVT DataVT, const RISCVSubtarget &ST, LLVMContext &C) { assert((DataVT.getScalarSizeInBits() != 8 || DataVT.getVectorNumElements() <= 256) && "unhandled case in lowering"); MVT IndexVT = DataVT.changeTypeToInteger(); if (IndexVT.getScalarType().bitsGT(ST.getXLenVT())) IndexVT = IndexVT.changeVectorElementType(MVT::i16); return cast(EVT(IndexVT).getTypeForEVT(C)); } /// Attempt to approximate the cost of a shuffle which will require splitting /// during legalization. Note that processShuffleMasks is not an exact proxy /// for the algorithm used in LegalizeVectorTypes, but hopefully it's a /// reasonably close upperbound. static InstructionCost costShuffleViaSplitting(const RISCVTTIImpl &TTI, MVT LegalVT, VectorType *Tp, ArrayRef Mask, TTI::TargetCostKind CostKind) { assert(LegalVT.isFixedLengthVector() && !Mask.empty() && "Expected fixed vector type and non-empty mask"); unsigned LegalNumElts = LegalVT.getVectorNumElements(); // Number of destination vectors after legalization: unsigned NumOfDests = divideCeil(Mask.size(), LegalNumElts); // We are going to permute multiple sources and the result will be in // multiple destinations. Providing an accurate cost only for splits where // the element type remains the same. if (NumOfDests <= 1 || LegalVT.getVectorElementType().getSizeInBits() != Tp->getElementType()->getPrimitiveSizeInBits() || LegalNumElts >= Tp->getElementCount().getFixedValue()) return InstructionCost::getInvalid(); unsigned VecTySize = TTI.getDataLayout().getTypeStoreSize(Tp); unsigned LegalVTSize = LegalVT.getStoreSize(); // Number of source vectors after legalization: unsigned NumOfSrcs = divideCeil(VecTySize, LegalVTSize); auto *SingleOpTy = FixedVectorType::get(Tp->getElementType(), LegalNumElts); unsigned NormalizedVF = LegalNumElts * std::max(NumOfSrcs, NumOfDests); unsigned NumOfSrcRegs = NormalizedVF / LegalNumElts; unsigned NumOfDestRegs = NormalizedVF / LegalNumElts; SmallVector NormalizedMask(NormalizedVF, PoisonMaskElem); assert(NormalizedVF >= Mask.size() && "Normalized mask expected to be not shorter than original mask."); copy(Mask, NormalizedMask.begin()); InstructionCost Cost = 0; SmallDenseSet, unsigned>> ReusedSingleSrcShuffles; processShuffleMasks( NormalizedMask, NumOfSrcRegs, NumOfDestRegs, NumOfDestRegs, []() {}, [&](ArrayRef RegMask, unsigned SrcReg, unsigned DestReg) { if (ShuffleVectorInst::isIdentityMask(RegMask, RegMask.size())) return; if (!ReusedSingleSrcShuffles.insert(std::make_pair(RegMask, SrcReg)) .second) return; Cost += TTI.getShuffleCost( TTI::SK_PermuteSingleSrc, FixedVectorType::get(SingleOpTy->getElementType(), RegMask.size()), SingleOpTy, RegMask, CostKind, 0, nullptr); }, [&](ArrayRef RegMask, unsigned Idx1, unsigned Idx2, bool NewReg) { Cost += TTI.getShuffleCost( TTI::SK_PermuteTwoSrc, FixedVectorType::get(SingleOpTy->getElementType(), RegMask.size()), SingleOpTy, RegMask, CostKind, 0, nullptr); }); return Cost; } /// Try to perform better estimation of the permutation. /// 1. Split the source/destination vectors into real registers. /// 2. Do the mask analysis to identify which real registers are /// permuted. If more than 1 source registers are used for the /// destination register building, the cost for this destination register /// is (Number_of_source_register - 1) * Cost_PermuteTwoSrc. If only one /// source register is used, build mask and calculate the cost as a cost /// of PermuteSingleSrc. /// Also, for the single register permute we try to identify if the /// destination register is just a copy of the source register or the /// copy of the previous destination register (the cost is /// TTI::TCC_Basic). If the source register is just reused, the cost for /// this operation is 0. static InstructionCost costShuffleViaVRegSplitting(const RISCVTTIImpl &TTI, MVT LegalVT, std::optional VLen, VectorType *Tp, ArrayRef Mask, TTI::TargetCostKind CostKind) { assert(LegalVT.isFixedLengthVector()); if (!VLen || Mask.empty()) return InstructionCost::getInvalid(); MVT ElemVT = LegalVT.getVectorElementType(); unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits(); LegalVT = TTI.getTypeLegalizationCost( FixedVectorType::get(Tp->getElementType(), ElemsPerVReg)) .second; // Number of destination vectors after legalization: InstructionCost NumOfDests = divideCeil(Mask.size(), LegalVT.getVectorNumElements()); if (NumOfDests <= 1 || LegalVT.getVectorElementType().getSizeInBits() != Tp->getElementType()->getPrimitiveSizeInBits() || LegalVT.getVectorNumElements() >= Tp->getElementCount().getFixedValue()) return InstructionCost::getInvalid(); unsigned VecTySize = TTI.getDataLayout().getTypeStoreSize(Tp); unsigned LegalVTSize = LegalVT.getStoreSize(); // Number of source vectors after legalization: unsigned NumOfSrcs = divideCeil(VecTySize, LegalVTSize); auto *SingleOpTy = FixedVectorType::get(Tp->getElementType(), LegalVT.getVectorNumElements()); unsigned E = NumOfDests.getValue(); unsigned NormalizedVF = LegalVT.getVectorNumElements() * std::max(NumOfSrcs, E); unsigned NumOfSrcRegs = NormalizedVF / LegalVT.getVectorNumElements(); unsigned NumOfDestRegs = NormalizedVF / LegalVT.getVectorNumElements(); SmallVector NormalizedMask(NormalizedVF, PoisonMaskElem); assert(NormalizedVF >= Mask.size() && "Normalized mask expected to be not shorter than original mask."); copy(Mask, NormalizedMask.begin()); InstructionCost Cost = 0; int NumShuffles = 0; SmallDenseSet, unsigned>> ReusedSingleSrcShuffles; processShuffleMasks( NormalizedMask, NumOfSrcRegs, NumOfDestRegs, NumOfDestRegs, []() {}, [&](ArrayRef RegMask, unsigned SrcReg, unsigned DestReg) { if (ShuffleVectorInst::isIdentityMask(RegMask, RegMask.size())) return; if (!ReusedSingleSrcShuffles.insert(std::make_pair(RegMask, SrcReg)) .second) return; ++NumShuffles; Cost += TTI.getShuffleCost(TTI::SK_PermuteSingleSrc, SingleOpTy, SingleOpTy, RegMask, CostKind, 0, nullptr); }, [&](ArrayRef RegMask, unsigned Idx1, unsigned Idx2, bool NewReg) { Cost += TTI.getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, SingleOpTy, RegMask, CostKind, 0, nullptr); NumShuffles += 2; }); // Note: check that we do not emit too many shuffles here to prevent code // size explosion. // TODO: investigate, if it can be improved by extra analysis of the masks // to check if the code is more profitable. if ((NumOfDestRegs > 2 && NumShuffles <= static_cast(NumOfDestRegs)) || (NumOfDestRegs <= 2 && NumShuffles < 4)) return Cost; return InstructionCost::getInvalid(); } InstructionCost RISCVTTIImpl::getSlideCost(FixedVectorType *Tp, ArrayRef Mask, TTI::TargetCostKind CostKind) const { // Avoid missing masks and length changing shuffles if (Mask.size() <= 2 || Mask.size() != Tp->getNumElements()) return InstructionCost::getInvalid(); int NumElts = Tp->getNumElements(); std::pair LT = getTypeLegalizationCost(Tp); // Avoid scalarization cases if (!LT.second.isFixedLengthVector()) return InstructionCost::getInvalid(); // Requires moving elements between parts, which requires additional // unmodeled instructions. if (LT.first != 1) return InstructionCost::getInvalid(); auto GetSlideOpcode = [&](int SlideAmt) { assert(SlideAmt != 0); bool IsVI = isUInt<5>(std::abs(SlideAmt)); if (SlideAmt < 0) return IsVI ? RISCV::VSLIDEDOWN_VI : RISCV::VSLIDEDOWN_VX; return IsVI ? RISCV::VSLIDEUP_VI : RISCV::VSLIDEUP_VX; }; std::array, 2> SrcInfo; if (!isMaskedSlidePair(Mask, NumElts, SrcInfo)) return InstructionCost::getInvalid(); if (SrcInfo[1].second == 0) std::swap(SrcInfo[0], SrcInfo[1]); InstructionCost FirstSlideCost = 0; if (SrcInfo[0].second != 0) { unsigned Opcode = GetSlideOpcode(SrcInfo[0].second); FirstSlideCost = getRISCVInstructionCost(Opcode, LT.second, CostKind); } if (SrcInfo[1].first == -1) return FirstSlideCost; InstructionCost SecondSlideCost = 0; if (SrcInfo[1].second != 0) { unsigned Opcode = GetSlideOpcode(SrcInfo[1].second); SecondSlideCost = getRISCVInstructionCost(Opcode, LT.second, CostKind); } else { SecondSlideCost = getRISCVInstructionCost(RISCV::VMERGE_VVM, LT.second, CostKind); } auto EC = Tp->getElementCount(); VectorType *MaskTy = VectorType::get(IntegerType::getInt1Ty(Tp->getContext()), EC); InstructionCost MaskCost = getConstantPoolLoadCost(MaskTy, CostKind); return FirstSlideCost + SecondSlideCost + MaskCost; } InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, VectorType *DstTy, VectorType *SrcTy, ArrayRef Mask, TTI::TargetCostKind CostKind, int Index, VectorType *SubTp, ArrayRef Args, const Instruction *CxtI) const { assert((Mask.empty() || DstTy->isScalableTy() || Mask.size() == DstTy->getElementCount().getKnownMinValue()) && "Expected the Mask to match the return size if given"); assert(SrcTy->getScalarType() == DstTy->getScalarType() && "Expected the same scalar types"); Kind = improveShuffleKindFromMask(Kind, Mask, SrcTy, Index, SubTp); std::pair LT = getTypeLegalizationCost(SrcTy); // First, handle cases where having a fixed length vector enables us to // give a more accurate cost than falling back to generic scalable codegen. // TODO: Each of these cases hints at a modeling gap around scalable vectors. if (auto *FVTp = dyn_cast(SrcTy); FVTp && ST->hasVInstructions() && LT.second.isFixedLengthVector()) { InstructionCost VRegSplittingCost = costShuffleViaVRegSplitting( *this, LT.second, ST->getRealVLen(), Kind == TTI::SK_InsertSubvector ? DstTy : SrcTy, Mask, CostKind); if (VRegSplittingCost.isValid()) return VRegSplittingCost; switch (Kind) { default: break; case TTI::SK_PermuteSingleSrc: { if (Mask.size() >= 2) { MVT EltTp = LT.second.getVectorElementType(); // If the size of the element is < ELEN then shuffles of interleaves and // deinterleaves of 2 vectors can be lowered into the following // sequences if (EltTp.getScalarSizeInBits() < ST->getELen()) { // Example sequence: // vsetivli zero, 4, e8, mf4, ta, ma (ignored) // vwaddu.vv v10, v8, v9 // li a0, -1 (ignored) // vwmaccu.vx v10, a0, v9 if (ShuffleVectorInst::isInterleaveMask(Mask, 2, Mask.size())) return 2 * LT.first * TLI->getLMULCost(LT.second); if (Mask[0] == 0 || Mask[0] == 1) { auto DeinterleaveMask = createStrideMask(Mask[0], 2, Mask.size()); // Example sequence: // vnsrl.wi v10, v8, 0 if (equal(DeinterleaveMask, Mask)) return LT.first * getRISCVInstructionCost(RISCV::VNSRL_WI, LT.second, CostKind); } } int SubVectorSize; if (LT.second.getScalarSizeInBits() != 1 && isRepeatedConcatMask(Mask, SubVectorSize)) { InstructionCost Cost = 0; unsigned NumSlides = Log2_32(Mask.size() / SubVectorSize); // The cost of extraction from a subvector is 0 if the index is 0. for (unsigned I = 0; I != NumSlides; ++I) { unsigned InsertIndex = SubVectorSize * (1 << I); FixedVectorType *SubTp = FixedVectorType::get(SrcTy->getElementType(), InsertIndex); FixedVectorType *DestTp = FixedVectorType::getDoubleElementsVectorType(SubTp); std::pair DestLT = getTypeLegalizationCost(DestTp); // Add the cost of whole vector register move because the // destination vector register group for vslideup cannot overlap the // source. Cost += DestLT.first * TLI->getLMULCost(DestLT.second); Cost += getShuffleCost(TTI::SK_InsertSubvector, DestTp, DestTp, {}, CostKind, InsertIndex, SubTp); } return Cost; } } if (InstructionCost SlideCost = getSlideCost(FVTp, Mask, CostKind); SlideCost.isValid()) return SlideCost; // vrgather + cost of generating the mask constant. // We model this for an unknown mask with a single vrgather. if (LT.first == 1 && (LT.second.getScalarSizeInBits() != 8 || LT.second.getVectorNumElements() <= 256)) { VectorType *IdxTy = getVRGatherIndexType(LT.second, *ST, SrcTy->getContext()); InstructionCost IndexCost = getConstantPoolLoadCost(IdxTy, CostKind); return IndexCost + getRISCVInstructionCost(RISCV::VRGATHER_VV, LT.second, CostKind); } break; } case TTI::SK_Transpose: case TTI::SK_PermuteTwoSrc: { if (InstructionCost SlideCost = getSlideCost(FVTp, Mask, CostKind); SlideCost.isValid()) return SlideCost; // 2 x (vrgather + cost of generating the mask constant) + cost of mask // register for the second vrgather. We model this for an unknown // (shuffle) mask. if (LT.first == 1 && (LT.second.getScalarSizeInBits() != 8 || LT.second.getVectorNumElements() <= 256)) { auto &C = SrcTy->getContext(); auto EC = SrcTy->getElementCount(); VectorType *IdxTy = getVRGatherIndexType(LT.second, *ST, C); VectorType *MaskTy = VectorType::get(IntegerType::getInt1Ty(C), EC); InstructionCost IndexCost = getConstantPoolLoadCost(IdxTy, CostKind); InstructionCost MaskCost = getConstantPoolLoadCost(MaskTy, CostKind); return 2 * IndexCost + getRISCVInstructionCost({RISCV::VRGATHER_VV, RISCV::VRGATHER_VV}, LT.second, CostKind) + MaskCost; } break; } } auto shouldSplit = [](TTI::ShuffleKind Kind) { switch (Kind) { default: return false; case TTI::SK_PermuteSingleSrc: case TTI::SK_Transpose: case TTI::SK_PermuteTwoSrc: return true; } }; if (!Mask.empty() && LT.first.isValid() && LT.first != 1 && shouldSplit(Kind)) { InstructionCost SplitCost = costShuffleViaSplitting(*this, LT.second, FVTp, Mask, CostKind); if (SplitCost.isValid()) return SplitCost; } } // Handle scalable vectors (and fixed vectors legalized to scalable vectors). switch (Kind) { default: // Fallthrough to generic handling. // TODO: Most of these cases will return getInvalid in generic code, and // must be implemented here. break; case TTI::SK_ExtractSubvector: // Extract at zero is always a subregister extract if (Index == 0) return TTI::TCC_Free; // If we're extracting a subvector of at most m1 size at a sub-register // boundary - which unfortunately we need exact vlen to identify - this is // a subregister extract at worst and thus won't require a vslidedown. // TODO: Extend for aligned m2, m4 subvector extracts // TODO: Extend for misalgined (but contained) extracts // TODO: Extend for scalable subvector types if (std::pair SubLT = getTypeLegalizationCost(SubTp); SubLT.second.isValid() && SubLT.second.isFixedLengthVector()) { if (std::optional VLen = ST->getRealVLen(); VLen && SubLT.second.getScalarSizeInBits() * Index % *VLen == 0 && SubLT.second.getSizeInBits() <= *VLen) return TTI::TCC_Free; } // Example sequence: // vsetivli zero, 4, e8, mf2, tu, ma (ignored) // vslidedown.vi v8, v9, 2 return LT.first * getRISCVInstructionCost(RISCV::VSLIDEDOWN_VI, LT.second, CostKind); case TTI::SK_InsertSubvector: // Example sequence: // vsetivli zero, 4, e8, mf2, tu, ma (ignored) // vslideup.vi v8, v9, 2 LT = getTypeLegalizationCost(DstTy); return LT.first * getRISCVInstructionCost(RISCV::VSLIDEUP_VI, LT.second, CostKind); case TTI::SK_Select: { // Example sequence: // li a0, 90 // vsetivli zero, 8, e8, mf2, ta, ma (ignored) // vmv.s.x v0, a0 // vmerge.vvm v8, v9, v8, v0 // We use 2 for the cost of the mask materialization as this is the true // cost for small masks and most shuffles are small. At worst, this cost // should be a very small constant for the constant pool load. As such, // we may bias towards large selects slightly more than truly warranted. return LT.first * (1 + getRISCVInstructionCost({RISCV::VMV_S_X, RISCV::VMERGE_VVM}, LT.second, CostKind)); } case TTI::SK_Broadcast: { bool HasScalar = (Args.size() > 0) && (Operator::getOpcode(Args[0]) == Instruction::InsertElement); if (LT.second.getScalarSizeInBits() == 1) { if (HasScalar) { // Example sequence: // andi a0, a0, 1 // vsetivli zero, 2, e8, mf8, ta, ma (ignored) // vmv.v.x v8, a0 // vmsne.vi v0, v8, 0 return LT.first * (1 + getRISCVInstructionCost({RISCV::VMV_V_X, RISCV::VMSNE_VI}, LT.second, CostKind)); } // Example sequence: // vsetivli zero, 2, e8, mf8, ta, mu (ignored) // vmv.v.i v8, 0 // vmerge.vim v8, v8, 1, v0 // vmv.x.s a0, v8 // andi a0, a0, 1 // vmv.v.x v8, a0 // vmsne.vi v0, v8, 0 return LT.first * (1 + getRISCVInstructionCost({RISCV::VMV_V_I, RISCV::VMERGE_VIM, RISCV::VMV_X_S, RISCV::VMV_V_X, RISCV::VMSNE_VI}, LT.second, CostKind)); } if (HasScalar) { // Example sequence: // vmv.v.x v8, a0 return LT.first * getRISCVInstructionCost(RISCV::VMV_V_X, LT.second, CostKind); } // Example sequence: // vrgather.vi v9, v8, 0 return LT.first * getRISCVInstructionCost(RISCV::VRGATHER_VI, LT.second, CostKind); } case TTI::SK_Splice: { // vslidedown+vslideup. // TODO: Multiplying by LT.first implies this legalizes into multiple copies // of similar code, but I think we expand through memory. unsigned Opcodes[2] = {RISCV::VSLIDEDOWN_VX, RISCV::VSLIDEUP_VX}; if (Index >= 0 && Index < 32) Opcodes[0] = RISCV::VSLIDEDOWN_VI; else if (Index < 0 && Index > -32) Opcodes[1] = RISCV::VSLIDEUP_VI; return LT.first * getRISCVInstructionCost(Opcodes, LT.second, CostKind); } case TTI::SK_Reverse: { if (!LT.second.isVector()) return InstructionCost::getInvalid(); // TODO: Cases to improve here: // * Illegal vector types // * i64 on RV32 if (SrcTy->getElementType()->isIntegerTy(1)) { VectorType *WideTy = VectorType::get(IntegerType::get(SrcTy->getContext(), 8), cast(SrcTy)->getElementCount()); return getCastInstrCost(Instruction::ZExt, WideTy, SrcTy, TTI::CastContextHint::None, CostKind) + getShuffleCost(TTI::SK_Reverse, WideTy, WideTy, {}, CostKind, 0, nullptr) + getCastInstrCost(Instruction::Trunc, SrcTy, WideTy, TTI::CastContextHint::None, CostKind); } MVT ContainerVT = LT.second; if (LT.second.isFixedLengthVector()) ContainerVT = TLI->getContainerForFixedLengthVector(LT.second); MVT M1VT = RISCVTargetLowering::getM1VT(ContainerVT); if (ContainerVT.bitsLE(M1VT)) { // Example sequence: // csrr a0, vlenb // srli a0, a0, 3 // addi a0, a0, -1 // vsetvli a1, zero, e8, mf8, ta, mu (ignored) // vid.v v9 // vrsub.vx v10, v9, a0 // vrgather.vv v9, v8, v10 InstructionCost LenCost = 3; if (LT.second.isFixedLengthVector()) // vrsub.vi has a 5 bit immediate field, otherwise an li suffices LenCost = isInt<5>(LT.second.getVectorNumElements() - 1) ? 0 : 1; unsigned Opcodes[] = {RISCV::VID_V, RISCV::VRSUB_VX, RISCV::VRGATHER_VV}; if (LT.second.isFixedLengthVector() && isInt<5>(LT.second.getVectorNumElements() - 1)) Opcodes[1] = RISCV::VRSUB_VI; InstructionCost GatherCost = getRISCVInstructionCost(Opcodes, LT.second, CostKind); return LT.first * (LenCost + GatherCost); } // At high LMUL, we split into a series of M1 reverses (see // lowerVECTOR_REVERSE) and then do a single slide at the end to eliminate // the resulting gap at the bottom (for fixed vectors only). The important // bit is that the cost scales linearly, not quadratically with LMUL. unsigned M1Opcodes[] = {RISCV::VID_V, RISCV::VRSUB_VX}; InstructionCost FixedCost = getRISCVInstructionCost(M1Opcodes, M1VT, CostKind) + 3; unsigned Ratio = ContainerVT.getVectorMinNumElements() / M1VT.getVectorMinNumElements(); InstructionCost GatherCost = getRISCVInstructionCost({RISCV::VRGATHER_VV}, M1VT, CostKind) * Ratio; InstructionCost SlideCost = !LT.second.isFixedLengthVector() ? 0 : getRISCVInstructionCost({RISCV::VSLIDEDOWN_VX}, LT.second, CostKind); return FixedCost + LT.first * (GatherCost + SlideCost); } } return BaseT::getShuffleCost(Kind, DstTy, SrcTy, Mask, CostKind, Index, SubTp); } static unsigned isM1OrSmaller(MVT VT) { RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); return (LMUL == RISCVVType::VLMUL::LMUL_F8 || LMUL == RISCVVType::VLMUL::LMUL_F4 || LMUL == RISCVVType::VLMUL::LMUL_F2 || LMUL == RISCVVType::VLMUL::LMUL_1); } InstructionCost RISCVTTIImpl::getScalarizationOverhead( VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract, TTI::TargetCostKind CostKind, bool ForPoisonSrc, ArrayRef VL) const { if (isa(Ty)) return InstructionCost::getInvalid(); // A build_vector (which is m1 sized or smaller) can be done in no // worse than one vslide1down.vx per element in the type. We could // in theory do an explode_vector in the inverse manner, but our // lowering today does not have a first class node for this pattern. InstructionCost Cost = BaseT::getScalarizationOverhead( Ty, DemandedElts, Insert, Extract, CostKind); std::pair LT = getTypeLegalizationCost(Ty); if (Insert && !Extract && LT.first.isValid() && LT.second.isVector()) { if (Ty->getScalarSizeInBits() == 1) { auto *WideVecTy = cast(Ty->getWithNewBitWidth(8)); // Note: Implicit scalar anyextend is assumed to be free since the i1 // must be stored in a GPR. return getScalarizationOverhead(WideVecTy, DemandedElts, Insert, Extract, CostKind) + getCastInstrCost(Instruction::Trunc, Ty, WideVecTy, TTI::CastContextHint::None, CostKind, nullptr); } assert(LT.second.isFixedLengthVector()); MVT ContainerVT = TLI->getContainerForFixedLengthVector(LT.second); if (isM1OrSmaller(ContainerVT)) { InstructionCost BV = cast(Ty)->getNumElements() * getRISCVInstructionCost(RISCV::VSLIDE1DOWN_VX, LT.second, CostKind); if (BV < Cost) Cost = BV; } } return Cost; } InstructionCost RISCVTTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind) const { if (!isLegalMaskedLoadStore(Src, Alignment) || CostKind != TTI::TCK_RecipThroughput) return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind); return getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind); } InstructionCost RISCVTTIImpl::getInterleavedMemoryOpCost( unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef Indices, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) const { // The interleaved memory access pass will lower interleaved memory ops (i.e // a load and store followed by a specific shuffle) to vlseg/vsseg // intrinsics. if (!UseMaskForCond && !UseMaskForGaps && Factor <= TLI->getMaxSupportedInterleaveFactor()) { auto *VTy = cast(VecTy); std::pair LT = getTypeLegalizationCost(VTy); // Need to make sure type has't been scalarized if (LT.second.isVector()) { auto *SubVecTy = VectorType::get(VTy->getElementType(), VTy->getElementCount().divideCoefficientBy(Factor)); if (VTy->getElementCount().isKnownMultipleOf(Factor) && TLI->isLegalInterleavedAccessType(SubVecTy, Factor, Alignment, AddressSpace, DL)) { // Some processors optimize segment loads/stores as one wide memory op + // Factor * LMUL shuffle ops. if (ST->hasOptimizedSegmentLoadStore(Factor)) { InstructionCost Cost = getMemoryOpCost(Opcode, VTy, Alignment, AddressSpace, CostKind); MVT SubVecVT = getTLI()->getValueType(DL, SubVecTy).getSimpleVT(); Cost += Factor * TLI->getLMULCost(SubVecVT); return LT.first * Cost; } // Otherwise, the cost is proportional to the number of elements (VL * // Factor ops). InstructionCost MemOpCost = getMemoryOpCost(Opcode, VTy->getElementType(), Alignment, 0, CostKind, {TTI::OK_AnyValue, TTI::OP_None}); unsigned NumLoads = getEstimatedVLFor(VTy); return NumLoads * MemOpCost; } } } // TODO: Return the cost of interleaved accesses for scalable vector when // unable to convert to segment accesses instructions. if (isa(VecTy)) return InstructionCost::getInvalid(); auto *FVTy = cast(VecTy); InstructionCost MemCost = getMemoryOpCost(Opcode, VecTy, Alignment, AddressSpace, CostKind); unsigned VF = FVTy->getNumElements() / Factor; // An interleaved load will look like this for Factor=3: // %wide.vec = load <12 x i32>, ptr %3, align 4 // %strided.vec = shufflevector %wide.vec, poison, <4 x i32> // %strided.vec1 = shufflevector %wide.vec, poison, <4 x i32> // %strided.vec2 = shufflevector %wide.vec, poison, <4 x i32> if (Opcode == Instruction::Load) { InstructionCost Cost = MemCost; for (unsigned Index : Indices) { FixedVectorType *VecTy = FixedVectorType::get(FVTy->getElementType(), VF * Factor); auto Mask = createStrideMask(Index, Factor, VF); Mask.resize(VF * Factor, -1); InstructionCost ShuffleCost = getShuffleCost(TTI::ShuffleKind::SK_PermuteSingleSrc, VecTy, VecTy, Mask, CostKind, 0, nullptr, {}); Cost += ShuffleCost; } return Cost; } // TODO: Model for NF > 2 // We'll need to enhance getShuffleCost to model shuffles that are just // inserts and extracts into subvectors, since they won't have the full cost // of a vrgather. // An interleaved store for 3 vectors of 4 lanes will look like // %11 = shufflevector <4 x i32> %4, <4 x i32> %6, <8 x i32> <0...7> // %12 = shufflevector <4 x i32> %9, <4 x i32> poison, <8 x i32> <0...3> // %13 = shufflevector <8 x i32> %11, <8 x i32> %12, <12 x i32> <0...11> // %interleaved.vec = shufflevector %13, poison, <12 x i32> // store <12 x i32> %interleaved.vec, ptr %10, align 4 if (Factor != 2) return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, Alignment, AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); assert(Opcode == Instruction::Store && "Opcode must be a store"); // For an interleaving store of 2 vectors, we perform one large interleaving // shuffle that goes into the wide store auto Mask = createInterleaveMask(VF, Factor); InstructionCost ShuffleCost = getShuffleCost(TTI::ShuffleKind::SK_PermuteSingleSrc, FVTy, FVTy, Mask, CostKind, 0, nullptr, {}); return MemCost + ShuffleCost; } InstructionCost RISCVTTIImpl::getGatherScatterOpCost( unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask, Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const { if (CostKind != TTI::TCK_RecipThroughput) return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, Alignment, CostKind, I); if ((Opcode == Instruction::Load && !isLegalMaskedGather(DataTy, Align(Alignment))) || (Opcode == Instruction::Store && !isLegalMaskedScatter(DataTy, Align(Alignment)))) return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, Alignment, CostKind, I); // Cost is proportional to the number of memory operations implied. For // scalable vectors, we use an estimate on that number since we don't // know exactly what VL will be. auto &VTy = *cast(DataTy); InstructionCost MemOpCost = getMemoryOpCost(Opcode, VTy.getElementType(), Alignment, 0, CostKind, {TTI::OK_AnyValue, TTI::OP_None}, I); unsigned NumLoads = getEstimatedVLFor(&VTy); return NumLoads * MemOpCost; } InstructionCost RISCVTTIImpl::getExpandCompressMemoryOpCost( unsigned Opcode, Type *DataTy, bool VariableMask, Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const { bool IsLegal = (Opcode == Instruction::Store && isLegalMaskedCompressStore(DataTy, Alignment)) || (Opcode == Instruction::Load && isLegalMaskedExpandLoad(DataTy, Alignment)); if (!IsLegal || CostKind != TTI::TCK_RecipThroughput) return BaseT::getExpandCompressMemoryOpCost(Opcode, DataTy, VariableMask, Alignment, CostKind, I); // Example compressstore sequence: // vsetivli zero, 8, e32, m2, ta, ma (ignored) // vcompress.vm v10, v8, v0 // vcpop.m a1, v0 // vsetvli zero, a1, e32, m2, ta, ma // vse32.v v10, (a0) // Example expandload sequence: // vsetivli zero, 8, e8, mf2, ta, ma (ignored) // vcpop.m a1, v0 // vsetvli zero, a1, e32, m2, ta, ma // vle32.v v10, (a0) // vsetivli zero, 8, e32, m2, ta, ma // viota.m v12, v0 // vrgather.vv v8, v10, v12, v0.t auto MemOpCost = getMemoryOpCost(Opcode, DataTy, Alignment, /*AddressSpace*/ 0, CostKind); auto LT = getTypeLegalizationCost(DataTy); SmallVector Opcodes{RISCV::VSETVLI}; if (VariableMask) Opcodes.push_back(RISCV::VCPOP_M); if (Opcode == Instruction::Store) Opcodes.append({RISCV::VCOMPRESS_VM}); else Opcodes.append({RISCV::VSETIVLI, RISCV::VIOTA_M, RISCV::VRGATHER_VV}); return MemOpCost + LT.first * getRISCVInstructionCost(Opcodes, LT.second, CostKind); } InstructionCost RISCVTTIImpl::getStridedMemoryOpCost( unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask, Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const { if (((Opcode == Instruction::Load || Opcode == Instruction::Store) && !isLegalStridedLoadStore(DataTy, Alignment)) || (Opcode != Instruction::Load && Opcode != Instruction::Store)) return BaseT::getStridedMemoryOpCost(Opcode, DataTy, Ptr, VariableMask, Alignment, CostKind, I); if (CostKind == TTI::TCK_CodeSize) return TTI::TCC_Basic; // Cost is proportional to the number of memory operations implied. For // scalable vectors, we use an estimate on that number since we don't // know exactly what VL will be. auto &VTy = *cast(DataTy); InstructionCost MemOpCost = getMemoryOpCost(Opcode, VTy.getElementType(), Alignment, 0, CostKind, {TTI::OK_AnyValue, TTI::OP_None}, I); unsigned NumLoads = getEstimatedVLFor(&VTy); return NumLoads * MemOpCost; } InstructionCost RISCVTTIImpl::getCostOfKeepingLiveOverCall(ArrayRef Tys) const { // FIXME: This is a property of the default vector convention, not // all possible calling conventions. Fixing that will require // some TTI API and SLP rework. InstructionCost Cost = 0; TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; for (auto *Ty : Tys) { if (!Ty->isVectorTy()) continue; Align A = DL.getPrefTypeAlign(Ty); Cost += getMemoryOpCost(Instruction::Store, Ty, A, 0, CostKind) + getMemoryOpCost(Instruction::Load, Ty, A, 0, CostKind); } return Cost; } // Currently, these represent both throughput and codesize costs // for the respective intrinsics. The costs in this table are simply // instruction counts with the following adjustments made: // * One vsetvli is considered free. static const CostTblEntry VectorIntrinsicCostTable[]{ {Intrinsic::floor, MVT::f32, 9}, {Intrinsic::floor, MVT::f64, 9}, {Intrinsic::ceil, MVT::f32, 9}, {Intrinsic::ceil, MVT::f64, 9}, {Intrinsic::trunc, MVT::f32, 7}, {Intrinsic::trunc, MVT::f64, 7}, {Intrinsic::round, MVT::f32, 9}, {Intrinsic::round, MVT::f64, 9}, {Intrinsic::roundeven, MVT::f32, 9}, {Intrinsic::roundeven, MVT::f64, 9}, {Intrinsic::rint, MVT::f32, 7}, {Intrinsic::rint, MVT::f64, 7}, {Intrinsic::lrint, MVT::i32, 1}, {Intrinsic::lrint, MVT::i64, 1}, {Intrinsic::llrint, MVT::i64, 1}, {Intrinsic::nearbyint, MVT::f32, 9}, {Intrinsic::nearbyint, MVT::f64, 9}, {Intrinsic::bswap, MVT::i16, 3}, {Intrinsic::bswap, MVT::i32, 12}, {Intrinsic::bswap, MVT::i64, 31}, {Intrinsic::vp_bswap, MVT::i16, 3}, {Intrinsic::vp_bswap, MVT::i32, 12}, {Intrinsic::vp_bswap, MVT::i64, 31}, {Intrinsic::vp_fshl, MVT::i8, 7}, {Intrinsic::vp_fshl, MVT::i16, 7}, {Intrinsic::vp_fshl, MVT::i32, 7}, {Intrinsic::vp_fshl, MVT::i64, 7}, {Intrinsic::vp_fshr, MVT::i8, 7}, {Intrinsic::vp_fshr, MVT::i16, 7}, {Intrinsic::vp_fshr, MVT::i32, 7}, {Intrinsic::vp_fshr, MVT::i64, 7}, {Intrinsic::bitreverse, MVT::i8, 17}, {Intrinsic::bitreverse, MVT::i16, 24}, {Intrinsic::bitreverse, MVT::i32, 33}, {Intrinsic::bitreverse, MVT::i64, 52}, {Intrinsic::vp_bitreverse, MVT::i8, 17}, {Intrinsic::vp_bitreverse, MVT::i16, 24}, {Intrinsic::vp_bitreverse, MVT::i32, 33}, {Intrinsic::vp_bitreverse, MVT::i64, 52}, {Intrinsic::ctpop, MVT::i8, 12}, {Intrinsic::ctpop, MVT::i16, 19}, {Intrinsic::ctpop, MVT::i32, 20}, {Intrinsic::ctpop, MVT::i64, 21}, {Intrinsic::ctlz, MVT::i8, 19}, {Intrinsic::ctlz, MVT::i16, 28}, {Intrinsic::ctlz, MVT::i32, 31}, {Intrinsic::ctlz, MVT::i64, 35}, {Intrinsic::cttz, MVT::i8, 16}, {Intrinsic::cttz, MVT::i16, 23}, {Intrinsic::cttz, MVT::i32, 24}, {Intrinsic::cttz, MVT::i64, 25}, {Intrinsic::vp_ctpop, MVT::i8, 12}, {Intrinsic::vp_ctpop, MVT::i16, 19}, {Intrinsic::vp_ctpop, MVT::i32, 20}, {Intrinsic::vp_ctpop, MVT::i64, 21}, {Intrinsic::vp_ctlz, MVT::i8, 19}, {Intrinsic::vp_ctlz, MVT::i16, 28}, {Intrinsic::vp_ctlz, MVT::i32, 31}, {Intrinsic::vp_ctlz, MVT::i64, 35}, {Intrinsic::vp_cttz, MVT::i8, 16}, {Intrinsic::vp_cttz, MVT::i16, 23}, {Intrinsic::vp_cttz, MVT::i32, 24}, {Intrinsic::vp_cttz, MVT::i64, 25}, }; InstructionCost RISCVTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind) const { auto *RetTy = ICA.getReturnType(); switch (ICA.getID()) { case Intrinsic::lrint: case Intrinsic::llrint: // We can't currently lower half or bfloat vector lrint/llrint. if (auto *VecTy = dyn_cast(ICA.getArgTypes()[0]); VecTy && VecTy->getElementType()->is16bitFPTy()) return InstructionCost::getInvalid(); [[fallthrough]]; case Intrinsic::ceil: case Intrinsic::floor: case Intrinsic::trunc: case Intrinsic::rint: case Intrinsic::round: case Intrinsic::roundeven: { // These all use the same code. auto LT = getTypeLegalizationCost(RetTy); if (!LT.second.isVector() && TLI->isOperationCustom(ISD::FCEIL, LT.second)) return LT.first * 8; break; } case Intrinsic::umin: case Intrinsic::umax: case Intrinsic::smin: case Intrinsic::smax: { auto LT = getTypeLegalizationCost(RetTy); if (LT.second.isScalarInteger() && ST->hasStdExtZbb()) return LT.first; if (ST->hasVInstructions() && LT.second.isVector()) { unsigned Op; switch (ICA.getID()) { case Intrinsic::umin: Op = RISCV::VMINU_VV; break; case Intrinsic::umax: Op = RISCV::VMAXU_VV; break; case Intrinsic::smin: Op = RISCV::VMIN_VV; break; case Intrinsic::smax: Op = RISCV::VMAX_VV; break; } return LT.first * getRISCVInstructionCost(Op, LT.second, CostKind); } break; } case Intrinsic::sadd_sat: case Intrinsic::ssub_sat: case Intrinsic::uadd_sat: case Intrinsic::usub_sat: { auto LT = getTypeLegalizationCost(RetTy); if (ST->hasVInstructions() && LT.second.isVector()) { unsigned Op; switch (ICA.getID()) { case Intrinsic::sadd_sat: Op = RISCV::VSADD_VV; break; case Intrinsic::ssub_sat: Op = RISCV::VSSUBU_VV; break; case Intrinsic::uadd_sat: Op = RISCV::VSADDU_VV; break; case Intrinsic::usub_sat: Op = RISCV::VSSUBU_VV; break; } return LT.first * getRISCVInstructionCost(Op, LT.second, CostKind); } break; } case Intrinsic::fma: case Intrinsic::fmuladd: { // TODO: handle promotion with f16/bf16 with zvfhmin/zvfbfmin auto LT = getTypeLegalizationCost(RetTy); if (ST->hasVInstructions() && LT.second.isVector()) return LT.first * getRISCVInstructionCost(RISCV::VFMADD_VV, LT.second, CostKind); break; } case Intrinsic::fabs: { auto LT = getTypeLegalizationCost(RetTy); if (ST->hasVInstructions() && LT.second.isVector()) { // lui a0, 8 // addi a0, a0, -1 // vsetvli a1, zero, e16, m1, ta, ma // vand.vx v8, v8, a0 // f16 with zvfhmin and bf16 with zvfhbmin if (LT.second.getVectorElementType() == MVT::bf16 || (LT.second.getVectorElementType() == MVT::f16 && !ST->hasVInstructionsF16())) return LT.first * getRISCVInstructionCost(RISCV::VAND_VX, LT.second, CostKind) + 2; else return LT.first * getRISCVInstructionCost(RISCV::VFSGNJX_VV, LT.second, CostKind); } break; } case Intrinsic::sqrt: { auto LT = getTypeLegalizationCost(RetTy); if (ST->hasVInstructions() && LT.second.isVector()) { SmallVector ConvOp; SmallVector FsqrtOp; MVT ConvType = LT.second; MVT FsqrtType = LT.second; // f16 with zvfhmin and bf16 with zvfbfmin and the type of nxv32[b]f16 // will be spilt. if (LT.second.getVectorElementType() == MVT::bf16) { if (LT.second == MVT::nxv32bf16) { ConvOp = {RISCV::VFWCVTBF16_F_F_V, RISCV::VFWCVTBF16_F_F_V, RISCV::VFNCVTBF16_F_F_W, RISCV::VFNCVTBF16_F_F_W}; FsqrtOp = {RISCV::VFSQRT_V, RISCV::VFSQRT_V}; ConvType = MVT::nxv16f16; FsqrtType = MVT::nxv16f32; } else { ConvOp = {RISCV::VFWCVTBF16_F_F_V, RISCV::VFNCVTBF16_F_F_W}; FsqrtOp = {RISCV::VFSQRT_V}; FsqrtType = TLI->getTypeToPromoteTo(ISD::FSQRT, FsqrtType); } } else if (LT.second.getVectorElementType() == MVT::f16 && !ST->hasVInstructionsF16()) { if (LT.second == MVT::nxv32f16) { ConvOp = {RISCV::VFWCVT_F_F_V, RISCV::VFWCVT_F_F_V, RISCV::VFNCVT_F_F_W, RISCV::VFNCVT_F_F_W}; FsqrtOp = {RISCV::VFSQRT_V, RISCV::VFSQRT_V}; ConvType = MVT::nxv16f16; FsqrtType = MVT::nxv16f32; } else { ConvOp = {RISCV::VFWCVT_F_F_V, RISCV::VFNCVT_F_F_W}; FsqrtOp = {RISCV::VFSQRT_V}; FsqrtType = TLI->getTypeToPromoteTo(ISD::FSQRT, FsqrtType); } } else { FsqrtOp = {RISCV::VFSQRT_V}; } return LT.first * (getRISCVInstructionCost(FsqrtOp, FsqrtType, CostKind) + getRISCVInstructionCost(ConvOp, ConvType, CostKind)); } break; } case Intrinsic::cttz: case Intrinsic::ctlz: case Intrinsic::ctpop: { auto LT = getTypeLegalizationCost(RetTy); if (ST->hasVInstructions() && ST->hasStdExtZvbb() && LT.second.isVector()) { unsigned Op; switch (ICA.getID()) { case Intrinsic::cttz: Op = RISCV::VCTZ_V; break; case Intrinsic::ctlz: Op = RISCV::VCLZ_V; break; case Intrinsic::ctpop: Op = RISCV::VCPOP_V; break; } return LT.first * getRISCVInstructionCost(Op, LT.second, CostKind); } break; } case Intrinsic::abs: { auto LT = getTypeLegalizationCost(RetTy); if (ST->hasVInstructions() && LT.second.isVector()) { // vrsub.vi v10, v8, 0 // vmax.vv v8, v8, v10 return LT.first * getRISCVInstructionCost({RISCV::VRSUB_VI, RISCV::VMAX_VV}, LT.second, CostKind); } break; } case Intrinsic::get_active_lane_mask: { if (ST->hasVInstructions()) { Type *ExpRetTy = VectorType::get( ICA.getArgTypes()[0], cast(RetTy)->getElementCount()); auto LT = getTypeLegalizationCost(ExpRetTy); // vid.v v8 // considered hoisted // vsaddu.vx v8, v8, a0 // vmsltu.vx v0, v8, a1 return LT.first * getRISCVInstructionCost({RISCV::VSADDU_VX, RISCV::VMSLTU_VX}, LT.second, CostKind); } break; } // TODO: add more intrinsic case Intrinsic::stepvector: { auto LT = getTypeLegalizationCost(RetTy); // Legalisation of illegal types involves an `index' instruction plus // (LT.first - 1) vector adds. if (ST->hasVInstructions()) return getRISCVInstructionCost(RISCV::VID_V, LT.second, CostKind) + (LT.first - 1) * getRISCVInstructionCost(RISCV::VADD_VX, LT.second, CostKind); return 1 + (LT.first - 1); } case Intrinsic::experimental_cttz_elts: { Type *ArgTy = ICA.getArgTypes()[0]; EVT ArgType = TLI->getValueType(DL, ArgTy, true); if (getTLI()->shouldExpandCttzElements(ArgType)) break; InstructionCost Cost = getRISCVInstructionCost( RISCV::VFIRST_M, getTypeLegalizationCost(ArgTy).second, CostKind); // If zero_is_poison is false, then we will generate additional // cmp + select instructions to convert -1 to EVL. Type *BoolTy = Type::getInt1Ty(RetTy->getContext()); if (ICA.getArgs().size() > 1 && cast(ICA.getArgs()[1])->isZero()) Cost += getCmpSelInstrCost(Instruction::ICmp, BoolTy, RetTy, CmpInst::ICMP_SLT, CostKind) + getCmpSelInstrCost(Instruction::Select, RetTy, BoolTy, CmpInst::BAD_ICMP_PREDICATE, CostKind); return Cost; } case Intrinsic::experimental_vp_splat: { auto LT = getTypeLegalizationCost(RetTy); // TODO: Lower i1 experimental_vp_splat if (!ST->hasVInstructions() || LT.second.getScalarType() == MVT::i1) return InstructionCost::getInvalid(); return LT.first * getRISCVInstructionCost(LT.second.isFloatingPoint() ? RISCV::VFMV_V_F : RISCV::VMV_V_X, LT.second, CostKind); } case Intrinsic::experimental_vp_splice: { // To support type-based query from vectorizer, set the index to 0. // Note that index only change the cost from vslide.vx to vslide.vi and in // current implementations they have same costs. return getShuffleCost(TTI::SK_Splice, cast(ICA.getReturnType()), cast(ICA.getArgTypes()[0]), {}, CostKind, 0, cast(ICA.getReturnType())); } } if (ST->hasVInstructions() && RetTy->isVectorTy()) { if (auto LT = getTypeLegalizationCost(RetTy); LT.second.isVector()) { MVT EltTy = LT.second.getVectorElementType(); if (const auto *Entry = CostTableLookup(VectorIntrinsicCostTable, ICA.getID(), EltTy)) return LT.first * Entry->Cost; } } return BaseT::getIntrinsicInstrCost(ICA, CostKind); } InstructionCost RISCVTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, TTI::CastContextHint CCH, TTI::TargetCostKind CostKind, const Instruction *I) const { bool IsVectorType = isa(Dst) && isa(Src); if (!IsVectorType) return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); // FIXME: Need to compute legalizing cost for illegal types. The current // code handles only legal types and those which can be trivially // promoted to legal. if (!ST->hasVInstructions() || Src->getScalarSizeInBits() > ST->getELen() || Dst->getScalarSizeInBits() > ST->getELen()) return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); int ISD = TLI->InstructionOpcodeToISD(Opcode); assert(ISD && "Invalid opcode"); std::pair SrcLT = getTypeLegalizationCost(Src); std::pair DstLT = getTypeLegalizationCost(Dst); // Handle i1 source and dest cases *before* calling logic in BasicTTI. // The shared implementation doesn't model vector widening during legalization // and instead assumes scalarization. In order to scalarize an // vector, we need to extend/trunc to/from i8. If we don't special case // this, we can get an infinite recursion cycle. switch (ISD) { default: break; case ISD::SIGN_EXTEND: case ISD::ZERO_EXTEND: if (Src->getScalarSizeInBits() == 1) { // We do not use vsext/vzext to extend from mask vector. // Instead we use the following instructions to extend from mask vector: // vmv.v.i v8, 0 // vmerge.vim v8, v8, -1, v0 (repeated per split) return getRISCVInstructionCost(RISCV::VMV_V_I, DstLT.second, CostKind) + DstLT.first * getRISCVInstructionCost(RISCV::VMERGE_VIM, DstLT.second, CostKind) + DstLT.first - 1; } break; case ISD::TRUNCATE: if (Dst->getScalarSizeInBits() == 1) { // We do not use several vncvt to truncate to mask vector. So we could // not use PowDiff to calculate it. // Instead we use the following instructions to truncate to mask vector: // vand.vi v8, v8, 1 // vmsne.vi v0, v8, 0 return SrcLT.first * getRISCVInstructionCost({RISCV::VAND_VI, RISCV::VMSNE_VI}, SrcLT.second, CostKind) + SrcLT.first - 1; } break; }; // Our actual lowering for the case where a wider legal type is available // uses promotion to the wider type. This is reflected in the result of // getTypeLegalizationCost, but BasicTTI assumes the widened cases are // scalarized if the legalized Src and Dst are not equal sized. const DataLayout &DL = this->getDataLayout(); if (!SrcLT.second.isVector() || !DstLT.second.isVector() || !TypeSize::isKnownLE(DL.getTypeSizeInBits(Src), SrcLT.second.getSizeInBits()) || !TypeSize::isKnownLE(DL.getTypeSizeInBits(Dst), DstLT.second.getSizeInBits())) return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); // The split cost is handled by the base getCastInstrCost assert((SrcLT.first == 1) && (DstLT.first == 1) && "Illegal type"); int PowDiff = (int)Log2_32(DstLT.second.getScalarSizeInBits()) - (int)Log2_32(SrcLT.second.getScalarSizeInBits()); switch (ISD) { case ISD::SIGN_EXTEND: case ISD::ZERO_EXTEND: { if ((PowDiff < 1) || (PowDiff > 3)) return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); unsigned SExtOp[] = {RISCV::VSEXT_VF2, RISCV::VSEXT_VF4, RISCV::VSEXT_VF8}; unsigned ZExtOp[] = {RISCV::VZEXT_VF2, RISCV::VZEXT_VF4, RISCV::VZEXT_VF8}; unsigned Op = (ISD == ISD::SIGN_EXTEND) ? SExtOp[PowDiff - 1] : ZExtOp[PowDiff - 1]; return getRISCVInstructionCost(Op, DstLT.second, CostKind); } case ISD::TRUNCATE: case ISD::FP_EXTEND: case ISD::FP_ROUND: { // Counts of narrow/widen instructions. unsigned SrcEltSize = SrcLT.second.getScalarSizeInBits(); unsigned DstEltSize = DstLT.second.getScalarSizeInBits(); unsigned Op = (ISD == ISD::TRUNCATE) ? RISCV::VNSRL_WI : (ISD == ISD::FP_EXTEND) ? RISCV::VFWCVT_F_F_V : RISCV::VFNCVT_F_F_W; InstructionCost Cost = 0; for (; SrcEltSize != DstEltSize;) { MVT ElementMVT = (ISD == ISD::TRUNCATE) ? MVT::getIntegerVT(DstEltSize) : MVT::getFloatingPointVT(DstEltSize); MVT DstMVT = DstLT.second.changeVectorElementType(ElementMVT); DstEltSize = (DstEltSize > SrcEltSize) ? DstEltSize >> 1 : DstEltSize << 1; Cost += getRISCVInstructionCost(Op, DstMVT, CostKind); } return Cost; } case ISD::FP_TO_SINT: case ISD::FP_TO_UINT: { unsigned IsSigned = ISD == ISD::FP_TO_SINT; unsigned FCVT = IsSigned ? RISCV::VFCVT_RTZ_X_F_V : RISCV::VFCVT_RTZ_XU_F_V; unsigned FWCVT = IsSigned ? RISCV::VFWCVT_RTZ_X_F_V : RISCV::VFWCVT_RTZ_XU_F_V; unsigned FNCVT = IsSigned ? RISCV::VFNCVT_RTZ_X_F_W : RISCV::VFNCVT_RTZ_XU_F_W; unsigned SrcEltSize = Src->getScalarSizeInBits(); unsigned DstEltSize = Dst->getScalarSizeInBits(); InstructionCost Cost = 0; if ((SrcEltSize == 16) && (!ST->hasVInstructionsF16() || ((DstEltSize / 2) > SrcEltSize))) { // If the target only supports zvfhmin or it is fp16-to-i64 conversion // pre-widening to f32 and then convert f32 to integer VectorType *VecF32Ty = VectorType::get(Type::getFloatTy(Dst->getContext()), cast(Dst)->getElementCount()); std::pair VecF32LT = getTypeLegalizationCost(VecF32Ty); Cost += VecF32LT.first * getRISCVInstructionCost(RISCV::VFWCVT_F_F_V, VecF32LT.second, CostKind); Cost += getCastInstrCost(Opcode, Dst, VecF32Ty, CCH, CostKind, I); return Cost; } if (DstEltSize == SrcEltSize) Cost += getRISCVInstructionCost(FCVT, DstLT.second, CostKind); else if (DstEltSize > SrcEltSize) Cost += getRISCVInstructionCost(FWCVT, DstLT.second, CostKind); else { // (SrcEltSize > DstEltSize) // First do a narrowing conversion to an integer half the size, then // truncate if needed. MVT ElementVT = MVT::getIntegerVT(SrcEltSize / 2); MVT VecVT = DstLT.second.changeVectorElementType(ElementVT); Cost += getRISCVInstructionCost(FNCVT, VecVT, CostKind); if ((SrcEltSize / 2) > DstEltSize) { Type *VecTy = EVT(VecVT).getTypeForEVT(Dst->getContext()); Cost += getCastInstrCost(Instruction::Trunc, Dst, VecTy, CCH, CostKind, I); } } return Cost; } case ISD::SINT_TO_FP: case ISD::UINT_TO_FP: { unsigned IsSigned = ISD == ISD::SINT_TO_FP; unsigned FCVT = IsSigned ? RISCV::VFCVT_F_X_V : RISCV::VFCVT_F_XU_V; unsigned FWCVT = IsSigned ? RISCV::VFWCVT_F_X_V : RISCV::VFWCVT_F_XU_V; unsigned FNCVT = IsSigned ? RISCV::VFNCVT_F_X_W : RISCV::VFNCVT_F_XU_W; unsigned SrcEltSize = Src->getScalarSizeInBits(); unsigned DstEltSize = Dst->getScalarSizeInBits(); InstructionCost Cost = 0; if ((DstEltSize == 16) && (!ST->hasVInstructionsF16() || ((SrcEltSize / 2) > DstEltSize))) { // If the target only supports zvfhmin or it is i64-to-fp16 conversion // it is converted to f32 and then converted to f16 VectorType *VecF32Ty = VectorType::get(Type::getFloatTy(Dst->getContext()), cast(Dst)->getElementCount()); std::pair VecF32LT = getTypeLegalizationCost(VecF32Ty); Cost += getCastInstrCost(Opcode, VecF32Ty, Src, CCH, CostKind, I); Cost += VecF32LT.first * getRISCVInstructionCost(RISCV::VFNCVT_F_F_W, DstLT.second, CostKind); return Cost; } if (DstEltSize == SrcEltSize) Cost += getRISCVInstructionCost(FCVT, DstLT.second, CostKind); else if (DstEltSize > SrcEltSize) { if ((DstEltSize / 2) > SrcEltSize) { VectorType *VecTy = VectorType::get(IntegerType::get(Dst->getContext(), DstEltSize / 2), cast(Dst)->getElementCount()); unsigned Op = IsSigned ? Instruction::SExt : Instruction::ZExt; Cost += getCastInstrCost(Op, VecTy, Src, CCH, CostKind, I); } Cost += getRISCVInstructionCost(FWCVT, DstLT.second, CostKind); } else Cost += getRISCVInstructionCost(FNCVT, DstLT.second, CostKind); return Cost; } } return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); } unsigned RISCVTTIImpl::getEstimatedVLFor(VectorType *Ty) const { if (isa(Ty)) { const unsigned EltSize = DL.getTypeSizeInBits(Ty->getElementType()); const unsigned MinSize = DL.getTypeSizeInBits(Ty).getKnownMinValue(); const unsigned VectorBits = *getVScaleForTuning() * RISCV::RVVBitsPerBlock; return RISCVTargetLowering::computeVLMAX(VectorBits, EltSize, MinSize); } return cast(Ty)->getNumElements(); } InstructionCost RISCVTTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF, TTI::TargetCostKind CostKind) const { if (isa(Ty) && !ST->useRVVForFixedLengthVectors()) return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind); // Skip if scalar size of Ty is bigger than ELEN. if (Ty->getScalarSizeInBits() > ST->getELen()) return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind); std::pair LT = getTypeLegalizationCost(Ty); if (Ty->getElementType()->isIntegerTy(1)) { // SelectionDAGBuilder does following transforms: // vector_reduce_{smin,umax}() --> vector_reduce_or() // vector_reduce_{smax,umin}() --> vector_reduce_and() if (IID == Intrinsic::umax || IID == Intrinsic::smin) return getArithmeticReductionCost(Instruction::Or, Ty, FMF, CostKind); else return getArithmeticReductionCost(Instruction::And, Ty, FMF, CostKind); } if (IID == Intrinsic::maximum || IID == Intrinsic::minimum) { SmallVector Opcodes; InstructionCost ExtraCost = 0; switch (IID) { case Intrinsic::maximum: if (FMF.noNaNs()) { Opcodes = {RISCV::VFREDMAX_VS, RISCV::VFMV_F_S}; } else { Opcodes = {RISCV::VMFNE_VV, RISCV::VCPOP_M, RISCV::VFREDMAX_VS, RISCV::VFMV_F_S}; // Cost of Canonical Nan + branch // lui a0, 523264 // fmv.w.x fa0, a0 Type *DstTy = Ty->getScalarType(); const unsigned EltTyBits = DstTy->getScalarSizeInBits(); Type *SrcTy = IntegerType::getIntNTy(DstTy->getContext(), EltTyBits); ExtraCost = 1 + getCastInstrCost(Instruction::UIToFP, DstTy, SrcTy, TTI::CastContextHint::None, CostKind) + getCFInstrCost(Instruction::Br, CostKind); } break; case Intrinsic::minimum: if (FMF.noNaNs()) { Opcodes = {RISCV::VFREDMIN_VS, RISCV::VFMV_F_S}; } else { Opcodes = {RISCV::VMFNE_VV, RISCV::VCPOP_M, RISCV::VFREDMIN_VS, RISCV::VFMV_F_S}; // Cost of Canonical Nan + branch // lui a0, 523264 // fmv.w.x fa0, a0 Type *DstTy = Ty->getScalarType(); const unsigned EltTyBits = DL.getTypeSizeInBits(DstTy); Type *SrcTy = IntegerType::getIntNTy(DstTy->getContext(), EltTyBits); ExtraCost = 1 + getCastInstrCost(Instruction::UIToFP, DstTy, SrcTy, TTI::CastContextHint::None, CostKind) + getCFInstrCost(Instruction::Br, CostKind); } break; } return ExtraCost + getRISCVInstructionCost(Opcodes, LT.second, CostKind); } // IR Reduction is composed by one rvv reduction instruction and vmv unsigned SplitOp; SmallVector Opcodes; switch (IID) { default: llvm_unreachable("Unsupported intrinsic"); case Intrinsic::smax: SplitOp = RISCV::VMAX_VV; Opcodes = {RISCV::VREDMAX_VS, RISCV::VMV_X_S}; break; case Intrinsic::smin: SplitOp = RISCV::VMIN_VV; Opcodes = {RISCV::VREDMIN_VS, RISCV::VMV_X_S}; break; case Intrinsic::umax: SplitOp = RISCV::VMAXU_VV; Opcodes = {RISCV::VREDMAXU_VS, RISCV::VMV_X_S}; break; case Intrinsic::umin: SplitOp = RISCV::VMINU_VV; Opcodes = {RISCV::VREDMINU_VS, RISCV::VMV_X_S}; break; case Intrinsic::maxnum: SplitOp = RISCV::VFMAX_VV; Opcodes = {RISCV::VFREDMAX_VS, RISCV::VFMV_F_S}; break; case Intrinsic::minnum: SplitOp = RISCV::VFMIN_VV; Opcodes = {RISCV::VFREDMIN_VS, RISCV::VFMV_F_S}; break; } // Add a cost for data larger than LMUL8 InstructionCost SplitCost = (LT.first > 1) ? (LT.first - 1) * getRISCVInstructionCost(SplitOp, LT.second, CostKind) : 0; return SplitCost + getRISCVInstructionCost(Opcodes, LT.second, CostKind); } InstructionCost RISCVTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *Ty, std::optional FMF, TTI::TargetCostKind CostKind) const { if (isa(Ty) && !ST->useRVVForFixedLengthVectors()) return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind); // Skip if scalar size of Ty is bigger than ELEN. if (Ty->getScalarSizeInBits() > ST->getELen()) return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind); int ISD = TLI->InstructionOpcodeToISD(Opcode); assert(ISD && "Invalid opcode"); if (ISD != ISD::ADD && ISD != ISD::OR && ISD != ISD::XOR && ISD != ISD::AND && ISD != ISD::FADD) return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind); std::pair LT = getTypeLegalizationCost(Ty); Type *ElementTy = Ty->getElementType(); if (ElementTy->isIntegerTy(1)) { // Example sequences: // vfirst.m a0, v0 // seqz a0, a0 if (LT.second == MVT::v1i1) return getRISCVInstructionCost(RISCV::VFIRST_M, LT.second, CostKind) + getCmpSelInstrCost(Instruction::ICmp, ElementTy, ElementTy, CmpInst::ICMP_EQ, CostKind); if (ISD == ISD::AND) { // Example sequences: // vmand.mm v8, v9, v8 ; needed every time type is split // vmnot.m v8, v0 ; alias for vmnand // vcpop.m a0, v8 // seqz a0, a0 // See the discussion: https://github.com/llvm/llvm-project/pull/119160 // For LMUL <= 8, there is no splitting, // the sequences are vmnot, vcpop and seqz. // When LMUL > 8 and split = 1, // the sequences are vmnand, vcpop and seqz. // When LMUL > 8 and split > 1, // the sequences are (LT.first-2) * vmand, vmnand, vcpop and seqz. return ((LT.first > 2) ? (LT.first - 2) : 0) * getRISCVInstructionCost(RISCV::VMAND_MM, LT.second, CostKind) + getRISCVInstructionCost(RISCV::VMNAND_MM, LT.second, CostKind) + getRISCVInstructionCost(RISCV::VCPOP_M, LT.second, CostKind) + getCmpSelInstrCost(Instruction::ICmp, ElementTy, ElementTy, CmpInst::ICMP_EQ, CostKind); } else if (ISD == ISD::XOR || ISD == ISD::ADD) { // Example sequences: // vsetvli a0, zero, e8, mf8, ta, ma // vmxor.mm v8, v0, v8 ; needed every time type is split // vcpop.m a0, v8 // andi a0, a0, 1 return (LT.first - 1) * getRISCVInstructionCost(RISCV::VMXOR_MM, LT.second, CostKind) + getRISCVInstructionCost(RISCV::VCPOP_M, LT.second, CostKind) + 1; } else { assert(ISD == ISD::OR); // Example sequences: // vsetvli a0, zero, e8, mf8, ta, ma // vmor.mm v8, v9, v8 ; needed every time type is split // vcpop.m a0, v0 // snez a0, a0 return (LT.first - 1) * getRISCVInstructionCost(RISCV::VMOR_MM, LT.second, CostKind) + getRISCVInstructionCost(RISCV::VCPOP_M, LT.second, CostKind) + getCmpSelInstrCost(Instruction::ICmp, ElementTy, ElementTy, CmpInst::ICMP_NE, CostKind); } } // IR Reduction of or/and is composed by one vmv and one rvv reduction // instruction, and others is composed by two vmv and one rvv reduction // instruction unsigned SplitOp; SmallVector Opcodes; switch (ISD) { case ISD::ADD: SplitOp = RISCV::VADD_VV; Opcodes = {RISCV::VMV_S_X, RISCV::VREDSUM_VS, RISCV::VMV_X_S}; break; case ISD::OR: SplitOp = RISCV::VOR_VV; Opcodes = {RISCV::VREDOR_VS, RISCV::VMV_X_S}; break; case ISD::XOR: SplitOp = RISCV::VXOR_VV; Opcodes = {RISCV::VMV_S_X, RISCV::VREDXOR_VS, RISCV::VMV_X_S}; break; case ISD::AND: SplitOp = RISCV::VAND_VV; Opcodes = {RISCV::VREDAND_VS, RISCV::VMV_X_S}; break; case ISD::FADD: // We can't promote f16/bf16 fadd reductions. if ((LT.second.getScalarType() == MVT::f16 && !ST->hasVInstructionsF16()) || LT.second.getScalarType() == MVT::bf16) return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind); if (TTI::requiresOrderedReduction(FMF)) { Opcodes.push_back(RISCV::VFMV_S_F); for (unsigned i = 0; i < LT.first.getValue(); i++) Opcodes.push_back(RISCV::VFREDOSUM_VS); Opcodes.push_back(RISCV::VFMV_F_S); return getRISCVInstructionCost(Opcodes, LT.second, CostKind); } SplitOp = RISCV::VFADD_VV; Opcodes = {RISCV::VFMV_S_F, RISCV::VFREDUSUM_VS, RISCV::VFMV_F_S}; break; } // Add a cost for data larger than LMUL8 InstructionCost SplitCost = (LT.first > 1) ? (LT.first - 1) * getRISCVInstructionCost(SplitOp, LT.second, CostKind) : 0; return SplitCost + getRISCVInstructionCost(Opcodes, LT.second, CostKind); } InstructionCost RISCVTTIImpl::getExtendedReductionCost( unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *ValTy, std::optional FMF, TTI::TargetCostKind CostKind) const { if (isa(ValTy) && !ST->useRVVForFixedLengthVectors()) return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy, FMF, CostKind); // Skip if scalar size of ResTy is bigger than ELEN. if (ResTy->getScalarSizeInBits() > ST->getELen()) return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy, FMF, CostKind); if (Opcode != Instruction::Add && Opcode != Instruction::FAdd) return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy, FMF, CostKind); std::pair LT = getTypeLegalizationCost(ValTy); if (IsUnsigned && Opcode == Instruction::Add && LT.second.isFixedLengthVector() && LT.second.getScalarType() == MVT::i1) { // Represent vector_reduce_add(ZExt()) as // ZExtOrTrunc(ctpop(bitcast to in)). return LT.first * getRISCVInstructionCost(RISCV::VCPOP_M, LT.second, CostKind); } if (ResTy->getScalarSizeInBits() != 2 * LT.second.getScalarSizeInBits()) return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy, FMF, CostKind); return (LT.first - 1) + getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind); } InstructionCost RISCVTTIImpl::getStoreImmCost(Type *Ty, TTI::OperandValueInfo OpInfo, TTI::TargetCostKind CostKind) const { assert(OpInfo.isConstant() && "non constant operand?"); if (!isa(Ty)) // FIXME: We need to account for immediate materialization here, but doing // a decent job requires more knowledge about the immediate than we // currently have here. return 0; if (OpInfo.isUniform()) // vmv.v.i, vmv.v.x, or vfmv.v.f // We ignore the cost of the scalar constant materialization to be consistent // with how we treat scalar constants themselves just above. return 1; return getConstantPoolLoadCost(Ty, CostKind); } InstructionCost RISCVTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpInfo, const Instruction *I) const { EVT VT = TLI->getValueType(DL, Src, true); // Type legalization can't handle structs if (VT == MVT::Other) return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind, OpInfo, I); InstructionCost Cost = 0; if (Opcode == Instruction::Store && OpInfo.isConstant()) Cost += getStoreImmCost(Src, OpInfo, CostKind); std::pair LT = getTypeLegalizationCost(Src); InstructionCost BaseCost = [&]() { InstructionCost Cost = LT.first; if (CostKind != TTI::TCK_RecipThroughput) return Cost; // Our actual lowering for the case where a wider legal type is available // uses the a VL predicated load on the wider type. This is reflected in // the result of getTypeLegalizationCost, but BasicTTI assumes the // widened cases are scalarized. const DataLayout &DL = this->getDataLayout(); if (Src->isVectorTy() && LT.second.isVector() && TypeSize::isKnownLT(DL.getTypeStoreSizeInBits(Src), LT.second.getSizeInBits())) return Cost; return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind, OpInfo, I); }(); // Assume memory ops cost scale with the number of vector registers // possible accessed by the instruction. Note that BasicTTI already // handles the LT.first term for us. if (LT.second.isVector() && CostKind != TTI::TCK_CodeSize) BaseCost *= TLI->getLMULCost(LT.second); return Cost + BaseCost; } InstructionCost RISCVTTIImpl::getCmpSelInstrCost( unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info, const Instruction *I) const { if (CostKind != TTI::TCK_RecipThroughput) return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I); if (isa(ValTy) && !ST->useRVVForFixedLengthVectors()) return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I); // Skip if scalar size of ValTy is bigger than ELEN. if (ValTy->isVectorTy() && ValTy->getScalarSizeInBits() > ST->getELen()) return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I); auto GetConstantMatCost = [&](TTI::OperandValueInfo OpInfo) -> InstructionCost { if (OpInfo.isUniform()) // We return 0 we currently ignore the cost of materializing scalar // constants in GPRs. return 0; return getConstantPoolLoadCost(ValTy, CostKind); }; InstructionCost ConstantMatCost; if (Op1Info.isConstant()) ConstantMatCost += GetConstantMatCost(Op1Info); if (Op2Info.isConstant()) ConstantMatCost += GetConstantMatCost(Op2Info); std::pair LT = getTypeLegalizationCost(ValTy); if (Opcode == Instruction::Select && ValTy->isVectorTy()) { if (CondTy->isVectorTy()) { if (ValTy->getScalarSizeInBits() == 1) { // vmandn.mm v8, v8, v9 // vmand.mm v9, v0, v9 // vmor.mm v0, v9, v8 return ConstantMatCost + LT.first * getRISCVInstructionCost( {RISCV::VMANDN_MM, RISCV::VMAND_MM, RISCV::VMOR_MM}, LT.second, CostKind); } // vselect and max/min are supported natively. return ConstantMatCost + LT.first * getRISCVInstructionCost(RISCV::VMERGE_VVM, LT.second, CostKind); } if (ValTy->getScalarSizeInBits() == 1) { // vmv.v.x v9, a0 // vmsne.vi v9, v9, 0 // vmandn.mm v8, v8, v9 // vmand.mm v9, v0, v9 // vmor.mm v0, v9, v8 MVT InterimVT = LT.second.changeVectorElementType(MVT::i8); return ConstantMatCost + LT.first * getRISCVInstructionCost({RISCV::VMV_V_X, RISCV::VMSNE_VI}, InterimVT, CostKind) + LT.first * getRISCVInstructionCost( {RISCV::VMANDN_MM, RISCV::VMAND_MM, RISCV::VMOR_MM}, LT.second, CostKind); } // vmv.v.x v10, a0 // vmsne.vi v0, v10, 0 // vmerge.vvm v8, v9, v8, v0 return ConstantMatCost + LT.first * getRISCVInstructionCost( {RISCV::VMV_V_X, RISCV::VMSNE_VI, RISCV::VMERGE_VVM}, LT.second, CostKind); } if ((Opcode == Instruction::ICmp) && ValTy->isVectorTy() && CmpInst::isIntPredicate(VecPred)) { // Use VMSLT_VV to represent VMSEQ, VMSNE, VMSLTU, VMSLEU, VMSLT, VMSLE // provided they incur the same cost across all implementations return ConstantMatCost + LT.first * getRISCVInstructionCost(RISCV::VMSLT_VV, LT.second, CostKind); } if ((Opcode == Instruction::FCmp) && ValTy->isVectorTy() && CmpInst::isFPPredicate(VecPred)) { // Use VMXOR_MM and VMXNOR_MM to generate all true/false mask if ((VecPred == CmpInst::FCMP_FALSE) || (VecPred == CmpInst::FCMP_TRUE)) return ConstantMatCost + getRISCVInstructionCost(RISCV::VMXOR_MM, LT.second, CostKind); // If we do not support the input floating point vector type, use the base // one which will calculate as: // ScalarizeCost + Num * Cost for fixed vector, // InvalidCost for scalable vector. if ((ValTy->getScalarSizeInBits() == 16 && !ST->hasVInstructionsF16()) || (ValTy->getScalarSizeInBits() == 32 && !ST->hasVInstructionsF32()) || (ValTy->getScalarSizeInBits() == 64 && !ST->hasVInstructionsF64())) return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I); // Assuming vector fp compare and mask instructions are all the same cost // until a need arises to differentiate them. switch (VecPred) { case CmpInst::FCMP_ONE: // vmflt.vv + vmflt.vv + vmor.mm case CmpInst::FCMP_ORD: // vmfeq.vv + vmfeq.vv + vmand.mm case CmpInst::FCMP_UNO: // vmfne.vv + vmfne.vv + vmor.mm case CmpInst::FCMP_UEQ: // vmflt.vv + vmflt.vv + vmnor.mm return ConstantMatCost + LT.first * getRISCVInstructionCost( {RISCV::VMFLT_VV, RISCV::VMFLT_VV, RISCV::VMOR_MM}, LT.second, CostKind); case CmpInst::FCMP_UGT: // vmfle.vv + vmnot.m case CmpInst::FCMP_UGE: // vmflt.vv + vmnot.m case CmpInst::FCMP_ULT: // vmfle.vv + vmnot.m case CmpInst::FCMP_ULE: // vmflt.vv + vmnot.m return ConstantMatCost + LT.first * getRISCVInstructionCost({RISCV::VMFLT_VV, RISCV::VMNAND_MM}, LT.second, CostKind); case CmpInst::FCMP_OEQ: // vmfeq.vv case CmpInst::FCMP_OGT: // vmflt.vv case CmpInst::FCMP_OGE: // vmfle.vv case CmpInst::FCMP_OLT: // vmflt.vv case CmpInst::FCMP_OLE: // vmfle.vv case CmpInst::FCMP_UNE: // vmfne.vv return ConstantMatCost + LT.first * getRISCVInstructionCost(RISCV::VMFLT_VV, LT.second, CostKind); default: break; } } // With ShortForwardBranchOpt or ConditionalMoveFusion, scalar icmp + select // instructions will lower to SELECT_CC and lower to PseudoCCMOVGPR which will // generate a conditional branch + mv. The cost of scalar (icmp + select) will // be (0 + select instr cost). if (ST->hasConditionalMoveFusion() && I && isa(I) && ValTy->isIntegerTy() && !I->user_empty()) { if (all_of(I->users(), [&](const User *U) { return match(U, m_Select(m_Specific(I), m_Value(), m_Value())) && U->getType()->isIntegerTy() && !isa(U->getOperand(1)) && !isa(U->getOperand(2)); })) return 0; } // TODO: Add cost for scalar type. return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I); } InstructionCost RISCVTTIImpl::getCFInstrCost(unsigned Opcode, TTI::TargetCostKind CostKind, const Instruction *I) const { if (CostKind != TTI::TCK_RecipThroughput) return Opcode == Instruction::PHI ? 0 : 1; // Branches are assumed to be predicted. return 0; } InstructionCost RISCVTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index, const Value *Op0, const Value *Op1) const { assert(Val->isVectorTy() && "This must be a vector type"); if (Opcode != Instruction::ExtractElement && Opcode != Instruction::InsertElement) return BaseT::getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1); // Legalize the type. std::pair LT = getTypeLegalizationCost(Val); // This type is legalized to a scalar type. if (!LT.second.isVector()) { auto *FixedVecTy = cast(Val); // If Index is a known constant, cost is zero. if (Index != -1U) return 0; // Extract/InsertElement with non-constant index is very costly when // scalarized; estimate cost of loads/stores sequence via the stack: // ExtractElement cost: store vector to stack, load scalar; // InsertElement cost: store vector to stack, store scalar, load vector. Type *ElemTy = FixedVecTy->getElementType(); auto NumElems = FixedVecTy->getNumElements(); auto Align = DL.getPrefTypeAlign(ElemTy); InstructionCost LoadCost = getMemoryOpCost(Instruction::Load, ElemTy, Align, 0, CostKind); InstructionCost StoreCost = getMemoryOpCost(Instruction::Store, ElemTy, Align, 0, CostKind); return Opcode == Instruction::ExtractElement ? StoreCost * NumElems + LoadCost : (StoreCost + LoadCost) * NumElems + StoreCost; } // For unsupported scalable vector. if (LT.second.isScalableVector() && !LT.first.isValid()) return LT.first; // Mask vector extract/insert is expanded via e8. if (Val->getScalarSizeInBits() == 1) { VectorType *WideTy = VectorType::get(IntegerType::get(Val->getContext(), 8), cast(Val)->getElementCount()); if (Opcode == Instruction::ExtractElement) { InstructionCost ExtendCost = getCastInstrCost(Instruction::ZExt, WideTy, Val, TTI::CastContextHint::None, CostKind); InstructionCost ExtractCost = getVectorInstrCost(Opcode, WideTy, CostKind, Index, nullptr, nullptr); return ExtendCost + ExtractCost; } InstructionCost ExtendCost = getCastInstrCost(Instruction::ZExt, WideTy, Val, TTI::CastContextHint::None, CostKind); InstructionCost InsertCost = getVectorInstrCost(Opcode, WideTy, CostKind, Index, nullptr, nullptr); InstructionCost TruncCost = getCastInstrCost(Instruction::Trunc, Val, WideTy, TTI::CastContextHint::None, CostKind); return ExtendCost + InsertCost + TruncCost; } // In RVV, we could use vslidedown + vmv.x.s to extract element from vector // and vslideup + vmv.s.x to insert element to vector. unsigned BaseCost = 1; // When insertelement we should add the index with 1 as the input of vslideup. unsigned SlideCost = Opcode == Instruction::InsertElement ? 2 : 1; if (Index != -1U) { // The type may be split. For fixed-width vectors we can normalize the // index to the new type. if (LT.second.isFixedLengthVector()) { unsigned Width = LT.second.getVectorNumElements(); Index = Index % Width; } // If exact VLEN is known, we will insert/extract into the appropriate // subvector with no additional subvector insert/extract cost. if (auto VLEN = ST->getRealVLen()) { unsigned EltSize = LT.second.getScalarSizeInBits(); unsigned M1Max = *VLEN / EltSize; Index = Index % M1Max; } if (Index == 0) // We can extract/insert the first element without vslidedown/vslideup. SlideCost = 0; else if (ST->hasVendorXRivosVisni() && isUInt<5>(Index) && Val->getScalarType()->isIntegerTy()) SlideCost = 0; // With ri.vinsert/ri.vextract there is no slide needed else if (Opcode == Instruction::InsertElement) SlideCost = 1; // With a constant index, we do not need to use addi. } // When the vector needs to split into multiple register groups and the index // exceeds single vector register group, we need to insert/extract the element // via stack. if (LT.first > 1 && ((Index == -1U) || (Index >= LT.second.getVectorMinNumElements() && LT.second.isScalableVector()))) { Type *ScalarType = Val->getScalarType(); Align VecAlign = DL.getPrefTypeAlign(Val); Align SclAlign = DL.getPrefTypeAlign(ScalarType); // Extra addi for unknown index. InstructionCost IdxCost = Index == -1U ? 1 : 0; // Store all split vectors into stack and load the target element. if (Opcode == Instruction::ExtractElement) return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0, CostKind) + getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0, CostKind) + IdxCost; // Store all split vectors into stack and store the target element and load // vectors back. return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0, CostKind) + getMemoryOpCost(Instruction::Load, Val, VecAlign, 0, CostKind) + getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0, CostKind) + IdxCost; } // Extract i64 in the target that has XLEN=32 need more instruction. if (Val->getScalarType()->isIntegerTy() && ST->getXLen() < Val->getScalarSizeInBits()) { // For extractelement, we need the following instructions: // vsetivli zero, 1, e64, m1, ta, mu (not count) // vslidedown.vx v8, v8, a0 // vmv.x.s a0, v8 // li a1, 32 // vsrl.vx v8, v8, a1 // vmv.x.s a1, v8 // For insertelement, we need the following instructions: // vsetivli zero, 2, e32, m4, ta, mu (not count) // vmv.v.i v12, 0 // vslide1up.vx v16, v12, a1 // vslide1up.vx v12, v16, a0 // addi a0, a2, 1 // vsetvli zero, a0, e64, m4, tu, mu (not count) // vslideup.vx v8, v12, a2 // TODO: should we count these special vsetvlis? BaseCost = Opcode == Instruction::InsertElement ? 3 : 4; } return BaseCost + SlideCost; } InstructionCost RISCVTTIImpl::getArithmeticInstrCost( unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info, ArrayRef Args, const Instruction *CxtI) const { // TODO: Handle more cost kinds. if (CostKind != TTI::TCK_RecipThroughput) return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info, Args, CxtI); if (isa(Ty) && !ST->useRVVForFixedLengthVectors()) return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info, Args, CxtI); // Skip if scalar size of Ty is bigger than ELEN. if (isa(Ty) && Ty->getScalarSizeInBits() > ST->getELen()) return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info, Args, CxtI); // Legalize the type. std::pair LT = getTypeLegalizationCost(Ty); // TODO: Handle scalar type. if (!LT.second.isVector()) return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info, Args, CxtI); // f16 with zvfhmin and bf16 will be promoted to f32. // FIXME: nxv32[b]f16 will be custom lowered and split. unsigned ISDOpcode = TLI->InstructionOpcodeToISD(Opcode); InstructionCost CastCost = 0; if ((LT.second.getVectorElementType() == MVT::f16 || LT.second.getVectorElementType() == MVT::bf16) && TLI->getOperationAction(ISDOpcode, LT.second) == TargetLoweringBase::LegalizeAction::Promote) { MVT PromotedVT = TLI->getTypeToPromoteTo(ISDOpcode, LT.second); Type *PromotedTy = EVT(PromotedVT).getTypeForEVT(Ty->getContext()); Type *LegalTy = EVT(LT.second).getTypeForEVT(Ty->getContext()); // Add cost of extending arguments CastCost += LT.first * Args.size() * getCastInstrCost(Instruction::FPExt, PromotedTy, LegalTy, TTI::CastContextHint::None, CostKind); // Add cost of truncating result CastCost += LT.first * getCastInstrCost(Instruction::FPTrunc, LegalTy, PromotedTy, TTI::CastContextHint::None, CostKind); // Compute cost of op in promoted type LT.second = PromotedVT; } auto getConstantMatCost = [&](unsigned Operand, TTI::OperandValueInfo OpInfo) -> InstructionCost { if (OpInfo.isUniform() && canSplatOperand(Opcode, Operand)) // Two sub-cases: // * Has a 5 bit immediate operand which can be splatted. // * Has a larger immediate which must be materialized in scalar register // We return 0 for both as we currently ignore the cost of materializing // scalar constants in GPRs. return 0; return getConstantPoolLoadCost(Ty, CostKind); }; // Add the cost of materializing any constant vectors required. InstructionCost ConstantMatCost = 0; if (Op1Info.isConstant()) ConstantMatCost += getConstantMatCost(0, Op1Info); if (Op2Info.isConstant()) ConstantMatCost += getConstantMatCost(1, Op2Info); unsigned Op; switch (ISDOpcode) { case ISD::ADD: case ISD::SUB: Op = RISCV::VADD_VV; break; case ISD::SHL: case ISD::SRL: case ISD::SRA: Op = RISCV::VSLL_VV; break; case ISD::AND: case ISD::OR: case ISD::XOR: Op = (Ty->getScalarSizeInBits() == 1) ? RISCV::VMAND_MM : RISCV::VAND_VV; break; case ISD::MUL: case ISD::MULHS: case ISD::MULHU: Op = RISCV::VMUL_VV; break; case ISD::SDIV: case ISD::UDIV: Op = RISCV::VDIV_VV; break; case ISD::SREM: case ISD::UREM: Op = RISCV::VREM_VV; break; case ISD::FADD: case ISD::FSUB: Op = RISCV::VFADD_VV; break; case ISD::FMUL: Op = RISCV::VFMUL_VV; break; case ISD::FDIV: Op = RISCV::VFDIV_VV; break; case ISD::FNEG: Op = RISCV::VFSGNJN_VV; break; default: // Assuming all other instructions have the same cost until a need arises to // differentiate them. return CastCost + ConstantMatCost + BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info, Args, CxtI); } InstructionCost InstrCost = getRISCVInstructionCost(Op, LT.second, CostKind); // We use BasicTTIImpl to calculate scalar costs, which assumes floating point // ops are twice as expensive as integer ops. Do the same for vectors so // scalar floating point ops aren't cheaper than their vector equivalents. if (Ty->isFPOrFPVectorTy()) InstrCost *= 2; return CastCost + ConstantMatCost + LT.first * InstrCost; } // TODO: Deduplicate from TargetTransformInfoImplCRTPBase. InstructionCost RISCVTTIImpl::getPointersChainCost( ArrayRef Ptrs, const Value *Base, const TTI::PointersChainInfo &Info, Type *AccessTy, TTI::TargetCostKind CostKind) const { InstructionCost Cost = TTI::TCC_Free; // In the basic model we take into account GEP instructions only // (although here can come alloca instruction, a value, constants and/or // constant expressions, PHIs, bitcasts ... whatever allowed to be used as a // pointer). Typically, if Base is a not a GEP-instruction and all the // pointers are relative to the same base address, all the rest are // either GEP instructions, PHIs, bitcasts or constants. When we have same // base, we just calculate cost of each non-Base GEP as an ADD operation if // any their index is a non-const. // If no known dependencies between the pointers cost is calculated as a sum // of costs of GEP instructions. for (auto [I, V] : enumerate(Ptrs)) { const auto *GEP = dyn_cast(V); if (!GEP) continue; if (Info.isSameBase() && V != Base) { if (GEP->hasAllConstantIndices()) continue; // If the chain is unit-stride and BaseReg + stride*i is a legal // addressing mode, then presume the base GEP is sitting around in a // register somewhere and check if we can fold the offset relative to // it. unsigned Stride = DL.getTypeStoreSize(AccessTy); if (Info.isUnitStride() && isLegalAddressingMode(AccessTy, /* BaseGV */ nullptr, /* BaseOffset */ Stride * I, /* HasBaseReg */ true, /* Scale */ 0, GEP->getType()->getPointerAddressSpace())) continue; Cost += getArithmeticInstrCost(Instruction::Add, GEP->getType(), CostKind, {TTI::OK_AnyValue, TTI::OP_None}, {TTI::OK_AnyValue, TTI::OP_None}, {}); } else { SmallVector Indices(GEP->indices()); Cost += getGEPCost(GEP->getSourceElementType(), GEP->getPointerOperand(), Indices, AccessTy, CostKind); } } return Cost; } void RISCVTTIImpl::getUnrollingPreferences( Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP, OptimizationRemarkEmitter *ORE) const { // TODO: More tuning on benchmarks and metrics with changes as needed // would apply to all settings below to enable performance. if (ST->enableDefaultUnroll()) return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP, ORE); // Enable Upper bound unrolling universally, not dependent upon the conditions // below. UP.UpperBound = true; // Disable loop unrolling for Oz and Os. UP.OptSizeThreshold = 0; UP.PartialOptSizeThreshold = 0; if (L->getHeader()->getParent()->hasOptSize()) return; SmallVector ExitingBlocks; L->getExitingBlocks(ExitingBlocks); LLVM_DEBUG(dbgs() << "Loop has:\n" << "Blocks: " << L->getNumBlocks() << "\n" << "Exit blocks: " << ExitingBlocks.size() << "\n"); // Only allow another exit other than the latch. This acts as an early exit // as it mirrors the profitability calculation of the runtime unroller. if (ExitingBlocks.size() > 2) return; // Limit the CFG of the loop body for targets with a branch predictor. // Allowing 4 blocks permits if-then-else diamonds in the body. if (L->getNumBlocks() > 4) return; // Don't unroll vectorized loops, including the remainder loop if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized")) return; // Scan the loop: don't unroll loops with calls as this could prevent // inlining. InstructionCost Cost = 0; for (auto *BB : L->getBlocks()) { for (auto &I : *BB) { // Initial setting - Don't unroll loops containing vectorized // instructions. if (I.getType()->isVectorTy()) return; if (isa(I) || isa(I)) { if (const Function *F = cast(I).getCalledFunction()) { if (!isLoweredToCall(F)) continue; } return; } SmallVector Operands(I.operand_values()); Cost += getInstructionCost(&I, Operands, TargetTransformInfo::TCK_SizeAndLatency); } } LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n"); UP.Partial = true; UP.Runtime = true; UP.UnrollRemainder = true; UP.UnrollAndJam = true; // Force unrolling small loops can be very useful because of the branch // taken cost of the backedge. if (Cost < 12) UP.Force = true; } void RISCVTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE, TTI::PeelingPreferences &PP) const { BaseT::getPeelingPreferences(L, SE, PP); } unsigned RISCVTTIImpl::getRegUsageForType(Type *Ty) const { if (Ty->isVectorTy()) { // f16 with only zvfhmin and bf16 will be promoted to f32 Type *EltTy = cast(Ty)->getElementType(); if ((EltTy->isHalfTy() && !ST->hasVInstructionsF16()) || EltTy->isBFloatTy()) Ty = VectorType::get(Type::getFloatTy(Ty->getContext()), cast(Ty)); TypeSize Size = DL.getTypeSizeInBits(Ty); if (Size.isScalable() && ST->hasVInstructions()) return divideCeil(Size.getKnownMinValue(), RISCV::RVVBitsPerBlock); if (ST->useRVVForFixedLengthVectors()) return divideCeil(Size, ST->getRealMinVLen()); } return BaseT::getRegUsageForType(Ty); } unsigned RISCVTTIImpl::getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { if (SLPMaxVF.getNumOccurrences()) return SLPMaxVF; // Return how many elements can fit in getRegisterBitwidth. This is the // same routine as used in LoopVectorizer. We should probably be // accounting for whether we actually have instructions with the right // lane type, but we don't have enough information to do that without // some additional plumbing which hasn't been justified yet. TypeSize RegWidth = getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector); // If no vector registers, or absurd element widths, disable // vectorization by returning 1. return std::max(1U, RegWidth.getFixedValue() / ElemWidth); } unsigned RISCVTTIImpl::getMinTripCountTailFoldingThreshold() const { return RVVMinTripCount; } TTI::AddressingModeKind RISCVTTIImpl::getPreferredAddressingMode(const Loop *L, ScalarEvolution *SE) const { if (ST->hasVendorXCVmem() && !ST->is64Bit()) return TTI::AMK_PostIndexed; return BasicTTIImplBase::getPreferredAddressingMode(L, SE); } bool RISCVTTIImpl::isLSRCostLess(const TargetTransformInfo::LSRCost &C1, const TargetTransformInfo::LSRCost &C2) const { // RISC-V specific here are "instruction number 1st priority". // If we need to emit adds inside the loop to add up base registers, then // we need at least one extra temporary register. unsigned C1NumRegs = C1.NumRegs + (C1.NumBaseAdds != 0); unsigned C2NumRegs = C2.NumRegs + (C2.NumBaseAdds != 0); return std::tie(C1.Insns, C1NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds, C1.ScaleCost, C1.ImmCost, C1.SetupCost) < std::tie(C2.Insns, C2NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds, C2.ScaleCost, C2.ImmCost, C2.SetupCost); } bool RISCVTTIImpl::isLegalMaskedExpandLoad(Type *DataTy, Align Alignment) const { auto *VTy = dyn_cast(DataTy); if (!VTy || VTy->isScalableTy()) return false; if (!isLegalMaskedLoadStore(DataTy, Alignment)) return false; // FIXME: If it is an i8 vector and the element count exceeds 256, we should // scalarize these types with LMUL >= maximum fixed-length LMUL. if (VTy->getElementType()->isIntegerTy(8)) if (VTy->getElementCount().getFixedValue() > 256) return VTy->getPrimitiveSizeInBits() / ST->getRealMinVLen() < ST->getMaxLMULForFixedLengthVectors(); return true; } bool RISCVTTIImpl::isLegalMaskedCompressStore(Type *DataTy, Align Alignment) const { auto *VTy = dyn_cast(DataTy); if (!VTy || VTy->isScalableTy()) return false; if (!isLegalMaskedLoadStore(DataTy, Alignment)) return false; return true; } /// See if \p I should be considered for address type promotion. We check if \p /// I is a sext with right type and used in memory accesses. If it used in a /// "complex" getelementptr, we allow it to be promoted without finding other /// sext instructions that sign extended the same initial value. A getelementptr /// is considered as "complex" if it has more than 2 operands. bool RISCVTTIImpl::shouldConsiderAddressTypePromotion( const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { bool Considerable = false; AllowPromotionWithoutCommonHeader = false; if (!isa(&I)) return false; Type *ConsideredSExtType = Type::getInt64Ty(I.getParent()->getParent()->getContext()); if (I.getType() != ConsideredSExtType) return false; // See if the sext is the one with the right type and used in at least one // GetElementPtrInst. for (const User *U : I.users()) { if (const GetElementPtrInst *GEPInst = dyn_cast(U)) { Considerable = true; // A getelementptr is considered as "complex" if it has more than 2 // operands. We will promote a SExt used in such complex GEP as we // expect some computation to be merged if they are done on 64 bits. if (GEPInst->getNumOperands() > 2) { AllowPromotionWithoutCommonHeader = true; break; } } } return Considerable; } bool RISCVTTIImpl::canSplatOperand(unsigned Opcode, int Operand) const { switch (Opcode) { case Instruction::Add: case Instruction::Sub: case Instruction::Mul: case Instruction::And: case Instruction::Or: case Instruction::Xor: case Instruction::FAdd: case Instruction::FSub: case Instruction::FMul: case Instruction::FDiv: case Instruction::ICmp: case Instruction::FCmp: return true; case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: case Instruction::UDiv: case Instruction::SDiv: case Instruction::URem: case Instruction::SRem: case Instruction::Select: return Operand == 1; default: return false; } } bool RISCVTTIImpl::canSplatOperand(Instruction *I, int Operand) const { if (!I->getType()->isVectorTy() || !ST->hasVInstructions()) return false; if (canSplatOperand(I->getOpcode(), Operand)) return true; auto *II = dyn_cast(I); if (!II) return false; switch (II->getIntrinsicID()) { case Intrinsic::fma: case Intrinsic::vp_fma: case Intrinsic::fmuladd: case Intrinsic::vp_fmuladd: return Operand == 0 || Operand == 1; case Intrinsic::vp_shl: case Intrinsic::vp_lshr: case Intrinsic::vp_ashr: case Intrinsic::vp_udiv: case Intrinsic::vp_sdiv: case Intrinsic::vp_urem: case Intrinsic::vp_srem: case Intrinsic::ssub_sat: case Intrinsic::vp_ssub_sat: case Intrinsic::usub_sat: case Intrinsic::vp_usub_sat: case Intrinsic::vp_select: return Operand == 1; // These intrinsics are commutative. case Intrinsic::vp_add: case Intrinsic::vp_mul: case Intrinsic::vp_and: case Intrinsic::vp_or: case Intrinsic::vp_xor: case Intrinsic::vp_fadd: case Intrinsic::vp_fmul: case Intrinsic::vp_icmp: case Intrinsic::vp_fcmp: case Intrinsic::smin: case Intrinsic::vp_smin: case Intrinsic::umin: case Intrinsic::vp_umin: case Intrinsic::smax: case Intrinsic::vp_smax: case Intrinsic::umax: case Intrinsic::vp_umax: case Intrinsic::sadd_sat: case Intrinsic::vp_sadd_sat: case Intrinsic::uadd_sat: case Intrinsic::vp_uadd_sat: // These intrinsics have 'vr' versions. case Intrinsic::vp_sub: case Intrinsic::vp_fsub: case Intrinsic::vp_fdiv: return Operand == 0 || Operand == 1; default: return false; } } /// Check if sinking \p I's operands to I's basic block is profitable, because /// the operands can be folded into a target instruction, e.g. /// splats of scalars can fold into vector instructions. bool RISCVTTIImpl::isProfitableToSinkOperands( Instruction *I, SmallVectorImpl &Ops) const { using namespace llvm::PatternMatch; if (I->isBitwiseLogicOp()) { if (!I->getType()->isVectorTy()) { if (ST->hasStdExtZbb() || ST->hasStdExtZbkb()) { for (auto &Op : I->operands()) { // (and/or/xor X, (not Y)) -> (andn/orn/xnor X, Y) if (match(Op.get(), m_Not(m_Value()))) { Ops.push_back(&Op); return true; } } } } else if (I->getOpcode() == Instruction::And && ST->hasStdExtZvkb()) { for (auto &Op : I->operands()) { // (and X, (not Y)) -> (vandn.vv X, Y) if (match(Op.get(), m_Not(m_Value()))) { Ops.push_back(&Op); return true; } // (and X, (splat (not Y))) -> (vandn.vx X, Y) if (match(Op.get(), m_Shuffle(m_InsertElt(m_Value(), m_Not(m_Value()), m_ZeroInt()), m_Value(), m_ZeroMask()))) { Use &InsertElt = cast(Op)->getOperandUse(0); Use &Not = cast(InsertElt)->getOperandUse(1); Ops.push_back(&Not); Ops.push_back(&InsertElt); Ops.push_back(&Op); return true; } } } } if (!I->getType()->isVectorTy() || !ST->hasVInstructions()) return false; // Don't sink splat operands if the target prefers it. Some targets requires // S2V transfer buffers and we can run out of them copying the same value // repeatedly. // FIXME: It could still be worth doing if it would improve vector register // pressure and prevent a vector spill. if (!ST->sinkSplatOperands()) return false; for (auto OpIdx : enumerate(I->operands())) { if (!canSplatOperand(I, OpIdx.index())) continue; Instruction *Op = dyn_cast(OpIdx.value().get()); // Make sure we are not already sinking this operand if (!Op || any_of(Ops, [&](Use *U) { return U->get() == Op; })) continue; // We are looking for a splat/vp.splat that can be sunk. bool IsVPSplat = match(Op, m_Intrinsic( m_Value(), m_Value(), m_Value())); if (!IsVPSplat && !match(Op, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()), m_Undef(), m_ZeroMask()))) continue; // Don't sink i1 splats. if (cast(Op->getType())->getElementType()->isIntegerTy(1)) continue; // All uses of the shuffle should be sunk to avoid duplicating it across gpr // and vector registers for (Use &U : Op->uses()) { Instruction *Insn = cast(U.getUser()); if (!canSplatOperand(Insn, U.getOperandNo())) return false; } // Sink any fpexts since they might be used in a widening fp pattern. if (IsVPSplat) { if (isa(Op->getOperand(0))) Ops.push_back(&Op->getOperandUse(0)); } else { Use *InsertEltUse = &Op->getOperandUse(0); auto *InsertElt = cast(InsertEltUse); if (isa(InsertElt->getOperand(1))) Ops.push_back(&InsertElt->getOperandUse(1)); Ops.push_back(InsertEltUse); } Ops.push_back(&OpIdx.value()); } return true; } RISCVTTIImpl::TTI::MemCmpExpansionOptions RISCVTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { TTI::MemCmpExpansionOptions Options; // TODO: Enable expansion when unaligned access is not supported after we fix // issues in ExpandMemcmp. if (!ST->enableUnalignedScalarMem()) return Options; if (!ST->hasStdExtZbb() && !ST->hasStdExtZbkb() && !IsZeroCmp) return Options; Options.AllowOverlappingLoads = true; Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); Options.NumLoadsPerBlock = Options.MaxNumLoads; if (ST->is64Bit()) { Options.LoadSizes = {8, 4, 2, 1}; Options.AllowedTailExpansions = {3, 5, 6}; } else { Options.LoadSizes = {4, 2, 1}; Options.AllowedTailExpansions = {3}; } if (IsZeroCmp && ST->hasVInstructions()) { unsigned VLenB = ST->getRealMinVLen() / 8; // The minimum size should be `XLen / 8 + 1`, and the maxinum size should be // `VLenB * MaxLMUL` so that it fits in a single register group. unsigned MinSize = ST->getXLen() / 8 + 1; unsigned MaxSize = VLenB * ST->getMaxLMULForFixedLengthVectors(); for (unsigned Size = MinSize; Size <= MaxSize; Size++) Options.LoadSizes.insert(Options.LoadSizes.begin(), Size); } return Options; }