OSchip
/
llvm-project
mirror of https://github.com/THU-DSP-LAB/llvm-project.git
13
0
Fork 6
Code Issues Packages Projects Releases Wiki Activity
llvm-project/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

809 lines
29 KiB
C++
Raw Blame History

//===-- 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/Analysis/TargetTransformInfo.h"
#include "llvm/CodeGen/BasicTTIImpl.h"
#include "llvm/CodeGen/CostTable.h"
#include "llvm/CodeGen/TargetLowering.h"
#include <cmath>
#include <optional>
using namespace llvm;
#define DEBUG_TYPE "riscvtti"
static cl::opt<unsigned> 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(1), cl::Hidden);
static cl::opt<unsigned> SLPMaxVF(
"riscv-v-slp-max-vf",
cl::desc(
"Result used for getMaximumVF query which is used exclusively by "
"SLP vectorizer. Defaults to 1 which disables SLP."),
cl::init(1), cl::Hidden);
InstructionCost RISCVTTIImpl::getLMULCost(MVT VT) {
// TODO: Here assume reciprocal throughput is 1 for LMUL_1, it is
// implementation-defined.
if (!VT.isVector())
return InstructionCost::getInvalid();
unsigned Cost;
if (VT.isScalableVector()) {
unsigned LMul;
bool Fractional;
std::tie(LMul, Fractional) =
RISCVVType::decodeVLMUL(RISCVTargetLowering::getLMUL(VT));
if (Fractional)
Cost = 1;
else
Cost = LMul;
} else {
Cost = VT.getSizeInBits() / ST->getRealMinVLen();
}
return std::max<unsigned>(Cost, 1);
}
InstructionCost RISCVTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind) {
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.
const DataLayout &DL = getDataLayout();
return RISCVMatInt::getIntMatCost(Imm, DL.getTypeSizeInBits(Ty),
getST()->getFeatureBits());
}
// 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<BinaryOperator>(Inst->getOperand(0));
if (!BO || !BO->hasOneUse())
return false;
if (BO->getOpcode() != Instruction::Shl)
return false;
if (!isa<ConstantInt>(BO->getOperand(1)))
return false;
unsigned ShAmt = cast<ConstantInt>(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 = countTrailingZeros(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) {
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::And:
// zext.h
if (Imm == UINT64_C(0xffff) && ST->hasStdExtZbb())
return TTI::TCC_Free;
// zext.w
if (Imm == UINT64_C(0xffffffff) && ST->hasStdExtZba())
return TTI::TCC_Free;
if (Inst && Idx == 1 && Imm.getBitWidth() <= ST->getXLen() &&
canUseShiftPair(Inst, Imm))
return TTI::TCC_Free;
[[fallthrough]];
case Instruction::Add:
case Instruction::Or:
case Instruction::Xor:
Takes12BitImm = true;
break;
case Instruction::Mul:
// Negated power of 2 is a shift and a negate.
if (Imm.isNegatedPowerOf2())
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.getMinSignedBits() <= 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) {
// Prevent hoisting in unknown cases.
return TTI::TCC_Free;
}
TargetTransformInfo::PopcntSupportKind
RISCVTTIImpl::getPopcntSupport(unsigned TyWidth) {
assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
return ST->hasStdExtZbb() ? TTI::PSK_FastHardware : TTI::PSK_Software;
}
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<unsigned> RISCVTTIImpl::getMaxVScale() const {
if (ST->hasVInstructions())
return ST->getRealMaxVLen() / RISCV::RVVBitsPerBlock;
return BaseT::getMaxVScale();
}
std::optional<unsigned> 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 = PowerOf2Floor(
std::max<unsigned>(std::min<unsigned>(RVVRegisterWidthLMUL, 8), 1));
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::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
unsigned AddressSpace,
TTI::TargetCostKind CostKind) {
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::getCastInstrCost(unsigned Opcode, Type *Dst,
Type *Src,
TTI::CastContextHint CCH,
TTI::TargetCostKind CostKind,
const Instruction *I) {
if (isa<VectorType>(Dst) && isa<VectorType>(Src)) {
// FIXME: Need to compute legalizing cost for illegal types.
if (!isTypeLegal(Src) || !isTypeLegal(Dst))
return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
// Skip if element size of Dst or Src is bigger than ELEN.
if (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");
// FIXME: Need to consider vsetvli and lmul.
int PowDiff = (int)Log2_32(Dst->getScalarSizeInBits()) -
(int)Log2_32(Src->getScalarSizeInBits());
switch (ISD) {
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
return 2;
}
return 1;
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 2;
}
[[fallthrough]];
case ISD::FP_EXTEND:
case ISD::FP_ROUND:
// Counts of narrow/widen instructions.
return std::abs(PowDiff);
case ISD::FP_TO_SINT:
case ISD::FP_TO_UINT:
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP:
if (Src->getScalarSizeInBits() == 1 || Dst->getScalarSizeInBits() == 1) {
// The cost of convert from or to mask vector is different from other
// cases. We could not use PowDiff to calculate it.
// For mask vector to fp, we should use the following instructions:
// vmv.v.i v8, 0
// vmerge.vim v8, v8, -1, v0
// vfcvt.f.x.v v8, v8
// And for fp vector to mask, we use:
// vfncvt.rtz.x.f.w v9, v8
// vand.vi v8, v9, 1
// vmsne.vi v0, v8, 0
return 3;
}
if (std::abs(PowDiff) <= 1)
return 1;
// Backend could lower (v[sz]ext i8 to double) to vfcvt(v[sz]ext.f8 i8),
// so it only need two conversion.
if (Src->isIntOrIntVectorTy())
return 2;
// Counts of narrow/widen instructions.
return std::abs(PowDiff);
}
}
return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
}
unsigned RISCVTTIImpl::getEstimatedVLFor(VectorType *Ty) {
if (isa<ScalableVectorType>(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<FixedVectorType>(Ty)->getNumElements();
}
InstructionCost
RISCVTTIImpl::getMinMaxReductionCost(VectorType *Ty, VectorType *CondTy,
bool IsUnsigned,
TTI::TargetCostKind CostKind) {
if (isa<FixedVectorType>(Ty) && !ST->useRVVForFixedLengthVectors())
return BaseT::getMinMaxReductionCost(Ty, CondTy, IsUnsigned, CostKind);
// Skip if scalar size of Ty is bigger than ELEN.
if (Ty->getScalarSizeInBits() > ST->getELEN())
return BaseT::getMinMaxReductionCost(Ty, CondTy, IsUnsigned, CostKind);
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
if (Ty->getElementType()->isIntegerTy(1))
// vcpop sequences, see vreduction-mask.ll. umax, smin actually only
// cost 2, but we don't have enough info here so we slightly over cost.
return (LT.first - 1) + 3;
// IR Reduction is composed by two vmv and one rvv reduction instruction.
InstructionCost BaseCost = 2;
unsigned VL = getEstimatedVLFor(Ty);
return (LT.first - 1) + BaseCost + Log2_32_Ceil(VL);
}
InstructionCost
RISCVTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *Ty,
std::optional<FastMathFlags> FMF,
TTI::TargetCostKind CostKind) {
if (isa<FixedVectorType>(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<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
if (Ty->getElementType()->isIntegerTy(1))
// vcpop sequences, see vreduction-mask.ll
return (LT.first - 1) + (ISD == ISD::AND ? 3 : 2);
// IR Reduction is composed by two vmv and one rvv reduction instruction.
InstructionCost BaseCost = 2;
unsigned VL = getEstimatedVLFor(Ty);
if (TTI::requiresOrderedReduction(FMF))
return (LT.first - 1) + BaseCost + VL;
return (LT.first - 1) + BaseCost + Log2_32_Ceil(VL);
}
InstructionCost RISCVTTIImpl::getExtendedReductionCost(
unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *ValTy,
std::optional<FastMathFlags> FMF, TTI::TargetCostKind CostKind) {
if (isa<FixedVectorType>(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<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
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) {
assert(OpInfo.isConstant() && "non constant operand?");
if (!isa<VectorType>(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.x.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;
// Add a cost of address generation + the cost of the vector 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);
}
InstructionCost RISCVTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
MaybeAlign Alignment,
unsigned AddressSpace,
TTI::TargetCostKind CostKind,
TTI::OperandValueInfo OpInfo,
const Instruction *I) {
InstructionCost Cost = 0;
if (Opcode == Instruction::Store && OpInfo.isConstant())
Cost += getStoreImmCost(Src, OpInfo, CostKind);
return Cost + BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
CostKind, OpInfo, I);
}
InstructionCost RISCVTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
Type *CondTy,
CmpInst::Predicate VecPred,
TTI::TargetCostKind CostKind,
const Instruction *I) {
if (CostKind != TTI::TCK_RecipThroughput)
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
I);
if (isa<FixedVectorType>(ValTy) && !ST->useRVVForFixedLengthVectors())
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
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,
I);
if (Opcode == Instruction::Select && ValTy->isVectorTy()) {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
if (CondTy->isVectorTy()) {
if (ValTy->getScalarSizeInBits() == 1) {
// vmandn.mm v8, v8, v9
// vmand.mm v9, v0, v9
// vmor.mm v0, v9, v8
return LT.first * 3;
}
// vselect and max/min are supported natively.
return LT.first * 1;
}
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
return LT.first * 5;
}
// vmv.v.x v10, a0
// vmsne.vi v0, v10, 0
// vmerge.vvm v8, v9, v8, v0
return LT.first * 3;
}
if ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
ValTy->isVectorTy()) {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
// Support natively.
if (CmpInst::isIntPredicate(VecPred))
return LT.first * 1;
// 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,
I);
switch (VecPred) {
// Support natively.
case CmpInst::FCMP_OEQ:
case CmpInst::FCMP_OGT:
case CmpInst::FCMP_OGE:
case CmpInst::FCMP_OLT:
case CmpInst::FCMP_OLE:
case CmpInst::FCMP_UNE:
return LT.first * 1;
// TODO: Other comparisons?
default:
break;
}
}
// TODO: Add cost for scalar type.
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
}
InstructionCost RISCVTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
unsigned Index) {
assert(Val->isVectorTy() && "This must be a vector type");
if (Opcode != Instruction::ExtractElement &&
Opcode != Instruction::InsertElement)
return BaseT::getVectorInstrCost(Opcode, Val, Index);
// Legalize the type.
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Val);
// This type is legalized to a scalar type.
if (!LT.second.isVector())
return 0;
// For unsupported scalable vector.
if (LT.second.isScalableVector() && !LT.first.isValid())
return LT.first;
if (!isTypeLegal(Val))
return BaseT::getVectorInstrCost(Opcode, Val, Index);
// 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;
}
// We could extract/insert the first element without vslidedown/vslideup.
if (Index == 0)
SlideCost = 0;
else if (Opcode == Instruction::InsertElement)
SlideCost = 1; // With a constant index, we do not need to use addi.
}
// Mask vector extract/insert element is different from normal case.
if (Val->getScalarSizeInBits() == 1) {
// For extractelement, we need the following instructions:
// vmv.v.i v8, 0
// vmerge.vim v8, v8, 1, v0
// vsetivli zero, 1, e8, m2, ta, mu (not count)
// vslidedown.vx v8, v8, a0
// vmv.x.s a0, v8
// For insertelement, we need the following instructions:
// vsetvli a2, zero, e8, m1, ta, mu (not count)
// vmv.s.x v8, a0
// vmv.v.i v9, 0
// vmerge.vim v9, v9, 1, v0
// addi a0, a1, 1
// vsetvli zero, a0, e8, m1, tu, mu (not count)
// vslideup.vx v9, v8, a1
// vsetvli a0, zero, e8, m1, ta, mu (not count)
// vand.vi v8, v9, 1
// vmsne.vi v0, v8, 0
// TODO: should we count these special vsetvlis?
BaseCost = Opcode == Instruction::InsertElement ? 5 : 3;
}
// 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<const Value *> Args, const Instruction *CxtI) {
// TODO: Handle more cost kinds.
if (CostKind != TTI::TCK_RecipThroughput)
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
Args, CxtI);
if (isa<FixedVectorType>(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<VectorType>(Ty) && Ty->getScalarSizeInBits() > ST->getELEN())
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
Args, CxtI);
// Legalize the type.
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
// TODO: Handle scalar type.
if (!LT.second.isVector())
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
Args, CxtI);
auto getConstantMatCost =
[&](unsigned Operand, TTI::OperandValueInfo OpInfo) -> InstructionCost {
if (OpInfo.isUniform() && TLI->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;
// Add a cost of address generation + the cost of the vector 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);
};
// 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);
switch (TLI->InstructionOpcodeToISD(Opcode)) {
case ISD::ADD:
case ISD::SUB:
case ISD::AND:
case ISD::OR:
case ISD::XOR:
case ISD::SHL:
case ISD::SRL:
case ISD::SRA:
case ISD::MUL:
case ISD::MULHS:
case ISD::MULHU:
case ISD::FADD:
case ISD::FSUB:
case ISD::FMUL:
case ISD::FNEG: {
return ConstantMatCost + getLMULCost(LT.second) * LT.first * 1;
}
default:
return ConstantMatCost +
BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
Args, CxtI);
}
}
void RISCVTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
TTI::UnrollingPreferences &UP,
OptimizationRemarkEmitter *ORE) {
// 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 dependant 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<BasicBlock *, 4> 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<CallInst>(I) || isa<InvokeInst>(I)) {
if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
if (!isLoweredToCall(F))
continue;
}
return;
}
SmallVector<const Value *> 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;
UP.UnrollAndJamInnerLoopThreshold = 60;
// 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) {
BaseT::getPeelingPreferences(L, SE, PP);
}
unsigned RISCVTTIImpl::getRegUsageForType(Type *Ty) {
TypeSize Size = DL.getTypeSizeInBits(Ty);
if (Ty->isVectorTy()) {
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 {
// This interface is currently only used by SLP. Returning 1 (which is the
// default value for SLPMaxVF) disables SLP. We currently have a cost modeling
// problem w/ constant materialization which causes SLP to perform majorly
// unprofitable transformations.
// TODO: Figure out constant materialization cost modeling and remove.
return SLPMaxVF;
}
Reference in New Issue View Git Blame Copy Permalink