blob: 88de84a4fd7873255a055c801aee38f750354edb [file] [log] [blame]
//===- ARMTargetTransformInfo.cpp - ARM 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 "ARMTargetTransformInfo.h"
#include "ARMSubtarget.h"
#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/CodeGen/CostTable.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsARM.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/Type.h"
#include "llvm/MC/SubtargetFeature.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MachineValueType.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/InstCombine/InstCombiner.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/LoopUtils.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "armtti"
static cl::opt<bool> EnableMaskedLoadStores(
"enable-arm-maskedldst", cl::Hidden, cl::init(true),
cl::desc("Enable the generation of masked loads and stores"));
static cl::opt<bool> DisableLowOverheadLoops(
"disable-arm-loloops", cl::Hidden, cl::init(false),
cl::desc("Disable the generation of low-overhead loops"));
static cl::opt<bool>
AllowWLSLoops("allow-arm-wlsloops", cl::Hidden, cl::init(true),
cl::desc("Enable the generation of WLS loops"));
extern cl::opt<TailPredication::Mode> EnableTailPredication;
extern cl::opt<bool> EnableMaskedGatherScatters;
extern cl::opt<unsigned> MVEMaxSupportedInterleaveFactor;
/// Convert a vector load intrinsic into a simple llvm load instruction.
/// This is beneficial when the underlying object being addressed comes
/// from a constant, since we get constant-folding for free.
static Value *simplifyNeonVld1(const IntrinsicInst &II, unsigned MemAlign,
InstCombiner::BuilderTy &Builder) {
auto *IntrAlign = dyn_cast<ConstantInt>(II.getArgOperand(1));
if (!IntrAlign)
return nullptr;
unsigned Alignment = IntrAlign->getLimitedValue() < MemAlign
? MemAlign
: IntrAlign->getLimitedValue();
if (!isPowerOf2_32(Alignment))
return nullptr;
auto *BCastInst = Builder.CreateBitCast(II.getArgOperand(0),
PointerType::get(II.getType(), 0));
return Builder.CreateAlignedLoad(II.getType(), BCastInst, Align(Alignment));
}
bool ARMTTIImpl::areInlineCompatible(const Function *Caller,
const Function *Callee) const {
const TargetMachine &TM = getTLI()->getTargetMachine();
const FeatureBitset &CallerBits =
TM.getSubtargetImpl(*Caller)->getFeatureBits();
const FeatureBitset &CalleeBits =
TM.getSubtargetImpl(*Callee)->getFeatureBits();
// To inline a callee, all features not in the allowed list must match exactly.
bool MatchExact = (CallerBits & ~InlineFeaturesAllowed) ==
(CalleeBits & ~InlineFeaturesAllowed);
// For features in the allowed list, the callee's features must be a subset of
// the callers'.
bool MatchSubset = ((CallerBits & CalleeBits) & InlineFeaturesAllowed) ==
(CalleeBits & InlineFeaturesAllowed);
return MatchExact && MatchSubset;
}
TTI::AddressingModeKind
ARMTTIImpl::getPreferredAddressingMode(const Loop *L,
ScalarEvolution *SE) const {
if (ST->hasMVEIntegerOps())
return TTI::AMK_PostIndexed;
if (L->getHeader()->getParent()->hasOptSize())
return TTI::AMK_None;
if (ST->isMClass() && ST->isThumb2() &&
L->getNumBlocks() == 1)
return TTI::AMK_PreIndexed;
return TTI::AMK_None;
}
Optional<Instruction *>
ARMTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
using namespace PatternMatch;
Intrinsic::ID IID = II.getIntrinsicID();
switch (IID) {
default:
break;
case Intrinsic::arm_neon_vld1: {
Align MemAlign =
getKnownAlignment(II.getArgOperand(0), IC.getDataLayout(), &II,
&IC.getAssumptionCache(), &IC.getDominatorTree());
if (Value *V = simplifyNeonVld1(II, MemAlign.value(), IC.Builder)) {
return IC.replaceInstUsesWith(II, V);
}
break;
}
case Intrinsic::arm_neon_vld2:
case Intrinsic::arm_neon_vld3:
case Intrinsic::arm_neon_vld4:
case Intrinsic::arm_neon_vld2lane:
case Intrinsic::arm_neon_vld3lane:
case Intrinsic::arm_neon_vld4lane:
case Intrinsic::arm_neon_vst1:
case Intrinsic::arm_neon_vst2:
case Intrinsic::arm_neon_vst3:
case Intrinsic::arm_neon_vst4:
case Intrinsic::arm_neon_vst2lane:
case Intrinsic::arm_neon_vst3lane:
case Intrinsic::arm_neon_vst4lane: {
Align MemAlign =
getKnownAlignment(II.getArgOperand(0), IC.getDataLayout(), &II,
&IC.getAssumptionCache(), &IC.getDominatorTree());
unsigned AlignArg = II.arg_size() - 1;
Value *AlignArgOp = II.getArgOperand(AlignArg);
MaybeAlign Align = cast<ConstantInt>(AlignArgOp)->getMaybeAlignValue();
if (Align && *Align < MemAlign) {
return IC.replaceOperand(
II, AlignArg,
ConstantInt::get(Type::getInt32Ty(II.getContext()), MemAlign.value(),
false));
}
break;
}
case Intrinsic::arm_mve_pred_i2v: {
Value *Arg = II.getArgOperand(0);
Value *ArgArg;
if (match(Arg, PatternMatch::m_Intrinsic<Intrinsic::arm_mve_pred_v2i>(
PatternMatch::m_Value(ArgArg))) &&
II.getType() == ArgArg->getType()) {
return IC.replaceInstUsesWith(II, ArgArg);
}
Constant *XorMask;
if (match(Arg, m_Xor(PatternMatch::m_Intrinsic<Intrinsic::arm_mve_pred_v2i>(
PatternMatch::m_Value(ArgArg)),
PatternMatch::m_Constant(XorMask))) &&
II.getType() == ArgArg->getType()) {
if (auto *CI = dyn_cast<ConstantInt>(XorMask)) {
if (CI->getValue().trunc(16).isAllOnes()) {
auto TrueVector = IC.Builder.CreateVectorSplat(
cast<FixedVectorType>(II.getType())->getNumElements(),
IC.Builder.getTrue());
return BinaryOperator::Create(Instruction::Xor, ArgArg, TrueVector);
}
}
}
KnownBits ScalarKnown(32);
if (IC.SimplifyDemandedBits(&II, 0, APInt::getLowBitsSet(32, 16),
ScalarKnown, 0)) {
return &II;
}
break;
}
case Intrinsic::arm_mve_pred_v2i: {
Value *Arg = II.getArgOperand(0);
Value *ArgArg;
if (match(Arg, PatternMatch::m_Intrinsic<Intrinsic::arm_mve_pred_i2v>(
PatternMatch::m_Value(ArgArg)))) {
return IC.replaceInstUsesWith(II, ArgArg);
}
if (!II.getMetadata(LLVMContext::MD_range)) {
Type *IntTy32 = Type::getInt32Ty(II.getContext());
Metadata *M[] = {
ConstantAsMetadata::get(ConstantInt::get(IntTy32, 0)),
ConstantAsMetadata::get(ConstantInt::get(IntTy32, 0x10000))};
II.setMetadata(LLVMContext::MD_range, MDNode::get(II.getContext(), M));
return &II;
}
break;
}
case Intrinsic::arm_mve_vadc:
case Intrinsic::arm_mve_vadc_predicated: {
unsigned CarryOp =
(II.getIntrinsicID() == Intrinsic::arm_mve_vadc_predicated) ? 3 : 2;
assert(II.getArgOperand(CarryOp)->getType()->getScalarSizeInBits() == 32 &&
"Bad type for intrinsic!");
KnownBits CarryKnown(32);
if (IC.SimplifyDemandedBits(&II, CarryOp, APInt::getOneBitSet(32, 29),
CarryKnown)) {
return &II;
}
break;
}
case Intrinsic::arm_mve_vmldava: {
Instruction *I = cast<Instruction>(&II);
if (I->hasOneUse()) {
auto *User = cast<Instruction>(*I->user_begin());
Value *OpZ;
if (match(User, m_c_Add(m_Specific(I), m_Value(OpZ))) &&
match(I->getOperand(3), m_Zero())) {
Value *OpX = I->getOperand(4);
Value *OpY = I->getOperand(5);
Type *OpTy = OpX->getType();
IC.Builder.SetInsertPoint(User);
Value *V =
IC.Builder.CreateIntrinsic(Intrinsic::arm_mve_vmldava, {OpTy},
{I->getOperand(0), I->getOperand(1),
I->getOperand(2), OpZ, OpX, OpY});
IC.replaceInstUsesWith(*User, V);
return IC.eraseInstFromFunction(*User);
}
}
return None;
}
}
return None;
}
Optional<Value *> ARMTTIImpl::simplifyDemandedVectorEltsIntrinsic(
InstCombiner &IC, IntrinsicInst &II, APInt OrigDemandedElts,
APInt &UndefElts, APInt &UndefElts2, APInt &UndefElts3,
std::function<void(Instruction *, unsigned, APInt, APInt &)>
SimplifyAndSetOp) const {
// Compute the demanded bits for a narrowing MVE intrinsic. The TopOpc is the
// opcode specifying a Top/Bottom instruction, which can change between
// instructions.
auto SimplifyNarrowInstrTopBottom =[&](unsigned TopOpc) {
unsigned NumElts = cast<FixedVectorType>(II.getType())->getNumElements();
unsigned IsTop = cast<ConstantInt>(II.getOperand(TopOpc))->getZExtValue();
// The only odd/even lanes of operand 0 will only be demanded depending
// on whether this is a top/bottom instruction.
APInt DemandedElts =
APInt::getSplat(NumElts, IsTop ? APInt::getLowBitsSet(2, 1)
: APInt::getHighBitsSet(2, 1));
SimplifyAndSetOp(&II, 0, OrigDemandedElts & DemandedElts, UndefElts);
// The other lanes will be defined from the inserted elements.
UndefElts &= APInt::getSplat(NumElts, !IsTop ? APInt::getLowBitsSet(2, 1)
: APInt::getHighBitsSet(2, 1));
return None;
};
switch (II.getIntrinsicID()) {
default:
break;
case Intrinsic::arm_mve_vcvt_narrow:
SimplifyNarrowInstrTopBottom(2);
break;
case Intrinsic::arm_mve_vqmovn:
SimplifyNarrowInstrTopBottom(4);
break;
case Intrinsic::arm_mve_vshrn:
SimplifyNarrowInstrTopBottom(7);
break;
}
return None;
}
InstructionCost ARMTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind) {
assert(Ty->isIntegerTy());
unsigned Bits = Ty->getPrimitiveSizeInBits();
if (Bits == 0 || Imm.getActiveBits() >= 64)
return 4;
int64_t SImmVal = Imm.getSExtValue();
uint64_t ZImmVal = Imm.getZExtValue();
if (!ST->isThumb()) {
if ((SImmVal >= 0 && SImmVal < 65536) ||
(ARM_AM::getSOImmVal(ZImmVal) != -1) ||
(ARM_AM::getSOImmVal(~ZImmVal) != -1))
return 1;
return ST->hasV6T2Ops() ? 2 : 3;
}
if (ST->isThumb2()) {
if ((SImmVal >= 0 && SImmVal < 65536) ||
(ARM_AM::getT2SOImmVal(ZImmVal) != -1) ||
(ARM_AM::getT2SOImmVal(~ZImmVal) != -1))
return 1;
return ST->hasV6T2Ops() ? 2 : 3;
}
// Thumb1, any i8 imm cost 1.
if (Bits == 8 || (SImmVal >= 0 && SImmVal < 256))
return 1;
if ((~SImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal))
return 2;
// Load from constantpool.
return 3;
}
// Constants smaller than 256 fit in the immediate field of
// Thumb1 instructions so we return a zero cost and 1 otherwise.
InstructionCost ARMTTIImpl::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx,
const APInt &Imm, Type *Ty) {
if (Imm.isNonNegative() && Imm.getLimitedValue() < 256)
return 0;
return 1;
}
// Checks whether Inst is part of a min(max()) or max(min()) pattern
// that will match to an SSAT instruction
static bool isSSATMinMaxPattern(Instruction *Inst, const APInt &Imm) {
Value *LHS, *RHS;
ConstantInt *C;
SelectPatternFlavor InstSPF = matchSelectPattern(Inst, LHS, RHS).Flavor;
if (InstSPF == SPF_SMAX &&
PatternMatch::match(RHS, PatternMatch::m_ConstantInt(C)) &&
C->getValue() == Imm && Imm.isNegative() && Imm.isNegatedPowerOf2()) {
auto isSSatMin = [&](Value *MinInst) {
if (isa<SelectInst>(MinInst)) {
Value *MinLHS, *MinRHS;
ConstantInt *MinC;
SelectPatternFlavor MinSPF =
matchSelectPattern(MinInst, MinLHS, MinRHS).Flavor;
if (MinSPF == SPF_SMIN &&
PatternMatch::match(MinRHS, PatternMatch::m_ConstantInt(MinC)) &&
MinC->getValue() == ((-Imm) - 1))
return true;
}
return false;
};
if (isSSatMin(Inst->getOperand(1)) ||
(Inst->hasNUses(2) && (isSSatMin(*Inst->user_begin()) ||
isSSatMin(*(++Inst->user_begin())))))
return true;
}
return false;
}
InstructionCost ARMTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind,
Instruction *Inst) {
// Division by a constant can be turned into multiplication, but only if we
// know it's constant. So it's not so much that the immediate is cheap (it's
// not), but that the alternative is worse.
// FIXME: this is probably unneeded with GlobalISel.
if ((Opcode == Instruction::SDiv || Opcode == Instruction::UDiv ||
Opcode == Instruction::SRem || Opcode == Instruction::URem) &&
Idx == 1)
return 0;
// Leave any gep offsets for the CodeGenPrepare, which will do a better job at
// splitting any large offsets.
if (Opcode == Instruction::GetElementPtr && Idx != 0)
return 0;
if (Opcode == Instruction::And) {
// UXTB/UXTH
if (Imm == 255 || Imm == 65535)
return 0;
// Conversion to BIC is free, and means we can use ~Imm instead.
return std::min(getIntImmCost(Imm, Ty, CostKind),
getIntImmCost(~Imm, Ty, CostKind));
}
if (Opcode == Instruction::Add)
// Conversion to SUB is free, and means we can use -Imm instead.
return std::min(getIntImmCost(Imm, Ty, CostKind),
getIntImmCost(-Imm, Ty, CostKind));
if (Opcode == Instruction::ICmp && Imm.isNegative() &&
Ty->getIntegerBitWidth() == 32) {
int64_t NegImm = -Imm.getSExtValue();
if (ST->isThumb2() && NegImm < 1<<12)
// icmp X, #-C -> cmn X, #C
return 0;
if (ST->isThumb() && NegImm < 1<<8)
// icmp X, #-C -> adds X, #C
return 0;
}
// xor a, -1 can always be folded to MVN
if (Opcode == Instruction::Xor && Imm.isAllOnes())
return 0;
// Ensures negative constant of min(max()) or max(min()) patterns that
// match to SSAT instructions don't get hoisted
if (Inst && ((ST->hasV6Ops() && !ST->isThumb()) || ST->isThumb2()) &&
Ty->getIntegerBitWidth() <= 32) {
if (isSSATMinMaxPattern(Inst, Imm) ||
(isa<ICmpInst>(Inst) && Inst->hasOneUse() &&
isSSATMinMaxPattern(cast<Instruction>(*Inst->user_begin()), Imm)))
return 0;
}
// We can convert <= -1 to < 0, which is generally quite cheap.
if (Inst && Opcode == Instruction::ICmp && Idx == 1 && Imm.isAllOnesValue()) {
ICmpInst::Predicate Pred = cast<ICmpInst>(Inst)->getPredicate();
if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE)
return std::min(getIntImmCost(Imm, Ty, CostKind),
getIntImmCost(Imm + 1, Ty, CostKind));
}
return getIntImmCost(Imm, Ty, CostKind);
}
InstructionCost ARMTTIImpl::getCFInstrCost(unsigned Opcode,
TTI::TargetCostKind CostKind,
const Instruction *I) {
if (CostKind == TTI::TCK_RecipThroughput &&
(ST->hasNEON() || ST->hasMVEIntegerOps())) {
// FIXME: The vectorizer is highly sensistive to the cost of these
// instructions, which suggests that it may be using the costs incorrectly.
// But, for now, just make them free to avoid performance regressions for
// vector targets.
return 0;
}
return BaseT::getCFInstrCost(Opcode, CostKind, I);
}
InstructionCost ARMTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
Type *Src,
TTI::CastContextHint CCH,
TTI::TargetCostKind CostKind,
const Instruction *I) {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
// TODO: Allow non-throughput costs that aren't binary.
auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost {
if (CostKind != TTI::TCK_RecipThroughput)
return Cost == 0 ? 0 : 1;
return Cost;
};
auto IsLegalFPType = [this](EVT VT) {
EVT EltVT = VT.getScalarType();
return (EltVT == MVT::f32 && ST->hasVFP2Base()) ||
(EltVT == MVT::f64 && ST->hasFP64()) ||
(EltVT == MVT::f16 && ST->hasFullFP16());
};
EVT SrcTy = TLI->getValueType(DL, Src);
EVT DstTy = TLI->getValueType(DL, Dst);
if (!SrcTy.isSimple() || !DstTy.isSimple())
return AdjustCost(
BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
// Extending masked load/Truncating masked stores is expensive because we
// currently don't split them. This means that we'll likely end up
// loading/storing each element individually (hence the high cost).
if ((ST->hasMVEIntegerOps() &&
(Opcode == Instruction::Trunc || Opcode == Instruction::ZExt ||
Opcode == Instruction::SExt)) ||
(ST->hasMVEFloatOps() &&
(Opcode == Instruction::FPExt || Opcode == Instruction::FPTrunc) &&
IsLegalFPType(SrcTy) && IsLegalFPType(DstTy)))
if (CCH == TTI::CastContextHint::Masked && DstTy.getSizeInBits() > 128)
return 2 * DstTy.getVectorNumElements() *
ST->getMVEVectorCostFactor(CostKind);
// The extend of other kinds of load is free
if (CCH == TTI::CastContextHint::Normal ||
CCH == TTI::CastContextHint::Masked) {
static const TypeConversionCostTblEntry LoadConversionTbl[] = {
{ISD::SIGN_EXTEND, MVT::i32, MVT::i16, 0},
{ISD::ZERO_EXTEND, MVT::i32, MVT::i16, 0},
{ISD::SIGN_EXTEND, MVT::i32, MVT::i8, 0},
{ISD::ZERO_EXTEND, MVT::i32, MVT::i8, 0},
{ISD::SIGN_EXTEND, MVT::i16, MVT::i8, 0},
{ISD::ZERO_EXTEND, MVT::i16, MVT::i8, 0},
{ISD::SIGN_EXTEND, MVT::i64, MVT::i32, 1},
{ISD::ZERO_EXTEND, MVT::i64, MVT::i32, 1},
{ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 1},
{ISD::ZERO_EXTEND, MVT::i64, MVT::i16, 1},
{ISD::SIGN_EXTEND, MVT::i64, MVT::i8, 1},
{ISD::ZERO_EXTEND, MVT::i64, MVT::i8, 1},
};
if (const auto *Entry = ConvertCostTableLookup(
LoadConversionTbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
static const TypeConversionCostTblEntry MVELoadConversionTbl[] = {
{ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0},
{ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0},
{ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 0},
{ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 0},
{ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 0},
{ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 0},
// The following extend from a legal type to an illegal type, so need to
// split the load. This introduced an extra load operation, but the
// extend is still "free".
{ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1},
{ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1},
{ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 3},
{ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 3},
{ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1},
{ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1},
};
if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
if (const auto *Entry =
ConvertCostTableLookup(MVELoadConversionTbl, ISD,
DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
}
static const TypeConversionCostTblEntry MVEFLoadConversionTbl[] = {
// FPExtends are similar but also require the VCVT instructions.
{ISD::FP_EXTEND, MVT::v4f32, MVT::v4f16, 1},
{ISD::FP_EXTEND, MVT::v8f32, MVT::v8f16, 3},
};
if (SrcTy.isVector() && ST->hasMVEFloatOps()) {
if (const auto *Entry =
ConvertCostTableLookup(MVEFLoadConversionTbl, ISD,
DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
}
// The truncate of a store is free. This is the mirror of extends above.
static const TypeConversionCostTblEntry MVEStoreConversionTbl[] = {
{ISD::TRUNCATE, MVT::v4i32, MVT::v4i16, 0},
{ISD::TRUNCATE, MVT::v4i32, MVT::v4i8, 0},
{ISD::TRUNCATE, MVT::v8i16, MVT::v8i8, 0},
{ISD::TRUNCATE, MVT::v8i32, MVT::v8i16, 1},
{ISD::TRUNCATE, MVT::v8i32, MVT::v8i8, 1},
{ISD::TRUNCATE, MVT::v16i32, MVT::v16i8, 3},
{ISD::TRUNCATE, MVT::v16i16, MVT::v16i8, 1},
};
if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
if (const auto *Entry =
ConvertCostTableLookup(MVEStoreConversionTbl, ISD,
SrcTy.getSimpleVT(), DstTy.getSimpleVT()))
return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
}
static const TypeConversionCostTblEntry MVEFStoreConversionTbl[] = {
{ISD::FP_ROUND, MVT::v4f32, MVT::v4f16, 1},
{ISD::FP_ROUND, MVT::v8f32, MVT::v8f16, 3},
};
if (SrcTy.isVector() && ST->hasMVEFloatOps()) {
if (const auto *Entry =
ConvertCostTableLookup(MVEFStoreConversionTbl, ISD,
SrcTy.getSimpleVT(), DstTy.getSimpleVT()))
return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
}
}
// NEON vector operations that can extend their inputs.
if ((ISD == ISD::SIGN_EXTEND || ISD == ISD::ZERO_EXTEND) &&
I && I->hasOneUse() && ST->hasNEON() && SrcTy.isVector()) {
static const TypeConversionCostTblEntry NEONDoubleWidthTbl[] = {
// vaddl
{ ISD::ADD, MVT::v4i32, MVT::v4i16, 0 },
{ ISD::ADD, MVT::v8i16, MVT::v8i8, 0 },
// vsubl
{ ISD::SUB, MVT::v4i32, MVT::v4i16, 0 },
{ ISD::SUB, MVT::v8i16, MVT::v8i8, 0 },
// vmull
{ ISD::MUL, MVT::v4i32, MVT::v4i16, 0 },
{ ISD::MUL, MVT::v8i16, MVT::v8i8, 0 },
// vshll
{ ISD::SHL, MVT::v4i32, MVT::v4i16, 0 },
{ ISD::SHL, MVT::v8i16, MVT::v8i8, 0 },
};
auto *User = cast<Instruction>(*I->user_begin());
int UserISD = TLI->InstructionOpcodeToISD(User->getOpcode());
if (auto *Entry = ConvertCostTableLookup(NEONDoubleWidthTbl, UserISD,
DstTy.getSimpleVT(),
SrcTy.getSimpleVT())) {
return AdjustCost(Entry->Cost);
}
}
// Single to/from double precision conversions.
if (Src->isVectorTy() && ST->hasNEON() &&
((ISD == ISD::FP_ROUND && SrcTy.getScalarType() == MVT::f64 &&
DstTy.getScalarType() == MVT::f32) ||
(ISD == ISD::FP_EXTEND && SrcTy.getScalarType() == MVT::f32 &&
DstTy.getScalarType() == MVT::f64))) {
static const CostTblEntry NEONFltDblTbl[] = {
// Vector fptrunc/fpext conversions.
{ISD::FP_ROUND, MVT::v2f64, 2},
{ISD::FP_EXTEND, MVT::v2f32, 2},
{ISD::FP_EXTEND, MVT::v4f32, 4}};
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
if (const auto *Entry = CostTableLookup(NEONFltDblTbl, ISD, LT.second))
return AdjustCost(LT.first * Entry->Cost);
}
// Some arithmetic, load and store operations have specific instructions
// to cast up/down their types automatically at no extra cost.
// TODO: Get these tables to know at least what the related operations are.
static const TypeConversionCostTblEntry NEONVectorConversionTbl[] = {
{ ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
{ ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
{ ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
{ ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 0 },
{ ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 },
// The number of vmovl instructions for the extension.
{ ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
{ ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
{ ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
{ ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i8, 3 },
{ ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i8, 3 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
{ ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
{ ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
{ ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
{ ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
{ ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
{ ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
{ ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
{ ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
{ ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
{ ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
{ ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
// Operations that we legalize using splitting.
{ ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 6 },
{ ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 3 },
// Vector float <-> i32 conversions.
{ ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
{ ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i8, 3 },
{ ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i8, 3 },
{ ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i16, 2 },
{ ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i16, 2 },
{ ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
{ ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
{ ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
{ ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
{ ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
{ ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 },
{ ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 },
{ ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 },
{ ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 },
{ ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 8 },
{ ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 8 },
{ ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 4 },
{ ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 4 },
{ ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 },
{ ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 },
{ ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 3 },
{ ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f32, 3 },
{ ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2 },
{ ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 2 },
// Vector double <-> i32 conversions.
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i8, 4 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 4 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i16, 3 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 3 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
{ ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 2 },
{ ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 2 },
{ ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 4 },
{ ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 4 },
{ ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f32, 8 },
{ ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 8 }
};
if (SrcTy.isVector() && ST->hasNEON()) {
if (const auto *Entry = ConvertCostTableLookup(NEONVectorConversionTbl, ISD,
DstTy.getSimpleVT(),
SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
}
// Scalar float to integer conversions.
static const TypeConversionCostTblEntry NEONFloatConversionTbl[] = {
{ ISD::FP_TO_SINT, MVT::i1, MVT::f32, 2 },
{ ISD::FP_TO_UINT, MVT::i1, MVT::f32, 2 },
{ ISD::FP_TO_SINT, MVT::i1, MVT::f64, 2 },
{ ISD::FP_TO_UINT, MVT::i1, MVT::f64, 2 },
{ ISD::FP_TO_SINT, MVT::i8, MVT::f32, 2 },
{ ISD::FP_TO_UINT, MVT::i8, MVT::f32, 2 },
{ ISD::FP_TO_SINT, MVT::i8, MVT::f64, 2 },
{ ISD::FP_TO_UINT, MVT::i8, MVT::f64, 2 },
{ ISD::FP_TO_SINT, MVT::i16, MVT::f32, 2 },
{ ISD::FP_TO_UINT, MVT::i16, MVT::f32, 2 },
{ ISD::FP_TO_SINT, MVT::i16, MVT::f64, 2 },
{ ISD::FP_TO_UINT, MVT::i16, MVT::f64, 2 },
{ ISD::FP_TO_SINT, MVT::i32, MVT::f32, 2 },
{ ISD::FP_TO_UINT, MVT::i32, MVT::f32, 2 },
{ ISD::FP_TO_SINT, MVT::i32, MVT::f64, 2 },
{ ISD::FP_TO_UINT, MVT::i32, MVT::f64, 2 },
{ ISD::FP_TO_SINT, MVT::i64, MVT::f32, 10 },
{ ISD::FP_TO_UINT, MVT::i64, MVT::f32, 10 },
{ ISD::FP_TO_SINT, MVT::i64, MVT::f64, 10 },
{ ISD::FP_TO_UINT, MVT::i64, MVT::f64, 10 }
};
if (SrcTy.isFloatingPoint() && ST->hasNEON()) {
if (const auto *Entry = ConvertCostTableLookup(NEONFloatConversionTbl, ISD,
DstTy.getSimpleVT(),
SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
}
// Scalar integer to float conversions.
static const TypeConversionCostTblEntry NEONIntegerConversionTbl[] = {
{ ISD::SINT_TO_FP, MVT::f32, MVT::i1, 2 },
{ ISD::UINT_TO_FP, MVT::f32, MVT::i1, 2 },
{ ISD::SINT_TO_FP, MVT::f64, MVT::i1, 2 },
{ ISD::UINT_TO_FP, MVT::f64, MVT::i1, 2 },
{ ISD::SINT_TO_FP, MVT::f32, MVT::i8, 2 },
{ ISD::UINT_TO_FP, MVT::f32, MVT::i8, 2 },
{ ISD::SINT_TO_FP, MVT::f64, MVT::i8, 2 },
{ ISD::UINT_TO_FP, MVT::f64, MVT::i8, 2 },
{ ISD::SINT_TO_FP, MVT::f32, MVT::i16, 2 },
{ ISD::UINT_TO_FP, MVT::f32, MVT::i16, 2 },
{ ISD::SINT_TO_FP, MVT::f64, MVT::i16, 2 },
{ ISD::UINT_TO_FP, MVT::f64, MVT::i16, 2 },
{ ISD::SINT_TO_FP, MVT::f32, MVT::i32, 2 },
{ ISD::UINT_TO_FP, MVT::f32, MVT::i32, 2 },
{ ISD::SINT_TO_FP, MVT::f64, MVT::i32, 2 },
{ ISD::UINT_TO_FP, MVT::f64, MVT::i32, 2 },
{ ISD::SINT_TO_FP, MVT::f32, MVT::i64, 10 },
{ ISD::UINT_TO_FP, MVT::f32, MVT::i64, 10 },
{ ISD::SINT_TO_FP, MVT::f64, MVT::i64, 10 },
{ ISD::UINT_TO_FP, MVT::f64, MVT::i64, 10 }
};
if (SrcTy.isInteger() && ST->hasNEON()) {
if (const auto *Entry = ConvertCostTableLookup(NEONIntegerConversionTbl,
ISD, DstTy.getSimpleVT(),
SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
}
// MVE extend costs, taken from codegen tests. i8->i16 or i16->i32 is one
// instruction, i8->i32 is two. i64 zexts are an VAND with a constant, sext
// are linearised so take more.
static const TypeConversionCostTblEntry MVEVectorConversionTbl[] = {
{ ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
{ ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
{ ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
{ ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i8, 10 },
{ ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i8, 2 },
{ ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
{ ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i16, 10 },
{ ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 8 },
{ ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 2 },
};
if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
if (const auto *Entry = ConvertCostTableLookup(MVEVectorConversionTbl,
ISD, DstTy.getSimpleVT(),
SrcTy.getSimpleVT()))
return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
}
if (ISD == ISD::FP_ROUND || ISD == ISD::FP_EXTEND) {
// As general rule, fp converts that were not matched above are scalarized
// and cost 1 vcvt for each lane, so long as the instruction is available.
// If not it will become a series of function calls.
const InstructionCost CallCost =
getCallInstrCost(nullptr, Dst, {Src}, CostKind);
int Lanes = 1;
if (SrcTy.isFixedLengthVector())
Lanes = SrcTy.getVectorNumElements();
if (IsLegalFPType(SrcTy) && IsLegalFPType(DstTy))
return Lanes;
else
return Lanes * CallCost;
}
if (ISD == ISD::TRUNCATE && ST->hasMVEIntegerOps() &&
SrcTy.isFixedLengthVector()) {
// Treat a truncate with larger than legal source (128bits for MVE) as
// expensive, 2 instructions per lane.
if ((SrcTy.getScalarType() == MVT::i8 ||
SrcTy.getScalarType() == MVT::i16 ||
SrcTy.getScalarType() == MVT::i32) &&
SrcTy.getSizeInBits() > 128 &&
SrcTy.getSizeInBits() > DstTy.getSizeInBits())
return SrcTy.getVectorNumElements() * 2;
}
// Scalar integer conversion costs.
static const TypeConversionCostTblEntry ARMIntegerConversionTbl[] = {
// i16 -> i64 requires two dependent operations.
{ ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 },
// Truncates on i64 are assumed to be free.
{ ISD::TRUNCATE, MVT::i32, MVT::i64, 0 },
{ ISD::TRUNCATE, MVT::i16, MVT::i64, 0 },
{ ISD::TRUNCATE, MVT::i8, MVT::i64, 0 },
{ ISD::TRUNCATE, MVT::i1, MVT::i64, 0 }
};
if (SrcTy.isInteger()) {
if (const auto *Entry = ConvertCostTableLookup(ARMIntegerConversionTbl, ISD,
DstTy.getSimpleVT(),
SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
}
int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
? ST->getMVEVectorCostFactor(CostKind)
: 1;
return AdjustCost(
BaseCost * BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
}
InstructionCost ARMTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
unsigned Index) {
// Penalize inserting into an D-subregister. We end up with a three times
// lower estimated throughput on swift.
if (ST->hasSlowLoadDSubregister() && Opcode == Instruction::InsertElement &&
ValTy->isVectorTy() && ValTy->getScalarSizeInBits() <= 32)
return 3;
if (ST->hasNEON() && (Opcode == Instruction::InsertElement ||
Opcode == Instruction::ExtractElement)) {
// Cross-class copies are expensive on many microarchitectures,
// so assume they are expensive by default.
if (cast<VectorType>(ValTy)->getElementType()->isIntegerTy())
return 3;
// Even if it's not a cross class copy, this likely leads to mixing
// of NEON and VFP code and should be therefore penalized.
if (ValTy->isVectorTy() &&
ValTy->getScalarSizeInBits() <= 32)
return std::max<InstructionCost>(
BaseT::getVectorInstrCost(Opcode, ValTy, Index), 2U);
}
if (ST->hasMVEIntegerOps() && (Opcode == Instruction::InsertElement ||
Opcode == Instruction::ExtractElement)) {
// Integer cross-lane moves are more expensive than float, which can
// sometimes just be vmovs. Integer involve being passes to GPR registers,
// causing more of a delay.
std::pair<InstructionCost, MVT> LT =
getTLI()->getTypeLegalizationCost(DL, ValTy->getScalarType());
return LT.first * (ValTy->getScalarType()->isIntegerTy() ? 4 : 1);
}
return BaseT::getVectorInstrCost(Opcode, ValTy, Index);
}
InstructionCost ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
Type *CondTy,
CmpInst::Predicate VecPred,
TTI::TargetCostKind CostKind,
const Instruction *I) {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
// Thumb scalar code size cost for select.
if (CostKind == TTI::TCK_CodeSize && ISD == ISD::SELECT &&
ST->isThumb() && !ValTy->isVectorTy()) {
// Assume expensive structs.
if (TLI->getValueType(DL, ValTy, true) == MVT::Other)
return TTI::TCC_Expensive;
// Select costs can vary because they:
// - may require one or more conditional mov (including an IT),
// - can't operate directly on immediates,
// - require live flags, which we can't copy around easily.
InstructionCost Cost = TLI->getTypeLegalizationCost(DL, ValTy).first;
// Possible IT instruction for Thumb2, or more for Thumb1.
++Cost;
// i1 values may need rematerialising by using mov immediates and/or
// flag setting instructions.
if (ValTy->isIntegerTy(1))
++Cost;
return Cost;
}
// If this is a vector min/max/abs, use the cost of that intrinsic directly
// instead. Hopefully when min/max intrinsics are more prevalent this code
// will not be needed.
const Instruction *Sel = I;
if ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && Sel &&
Sel->hasOneUse())
Sel = cast<Instruction>(Sel->user_back());
if (Sel && ValTy->isVectorTy() &&
(ValTy->isIntOrIntVectorTy() || ValTy->isFPOrFPVectorTy())) {
const Value *LHS, *RHS;
SelectPatternFlavor SPF = matchSelectPattern(Sel, LHS, RHS).Flavor;
unsigned IID = 0;
switch (SPF) {
case SPF_ABS:
IID = Intrinsic::abs;
break;
case SPF_SMIN:
IID = Intrinsic::smin;
break;
case SPF_SMAX:
IID = Intrinsic::smax;
break;
case SPF_UMIN:
IID = Intrinsic::umin;
break;
case SPF_UMAX:
IID = Intrinsic::umax;
break;
case SPF_FMINNUM:
IID = Intrinsic::minnum;
break;
case SPF_FMAXNUM:
IID = Intrinsic::maxnum;
break;
default:
break;
}
if (IID) {
// The ICmp is free, the select gets the cost of the min/max/etc
if (Sel != I)
return 0;
IntrinsicCostAttributes CostAttrs(IID, ValTy, {ValTy, ValTy});
return getIntrinsicInstrCost(CostAttrs, CostKind);
}
}
// On NEON a vector select gets lowered to vbsl.
if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT && CondTy) {
// Lowering of some vector selects is currently far from perfect.
static const TypeConversionCostTblEntry NEONVectorSelectTbl[] = {
{ ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
};
EVT SelCondTy = TLI->getValueType(DL, CondTy);
EVT SelValTy = TLI->getValueType(DL, ValTy);
if (SelCondTy.isSimple() && SelValTy.isSimple()) {
if (const auto *Entry = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
SelCondTy.getSimpleVT(),
SelValTy.getSimpleVT()))
return Entry->Cost;
}
std::pair<InstructionCost, MVT> LT =
TLI->getTypeLegalizationCost(DL, ValTy);
return LT.first;
}
if (ST->hasMVEIntegerOps() && ValTy->isVectorTy() &&
(Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
cast<FixedVectorType>(ValTy)->getNumElements() > 1) {
FixedVectorType *VecValTy = cast<FixedVectorType>(ValTy);
FixedVectorType *VecCondTy = dyn_cast_or_null<FixedVectorType>(CondTy);
if (!VecCondTy)
VecCondTy = cast<FixedVectorType>(CmpInst::makeCmpResultType(VecValTy));
// If we don't have mve.fp any fp operations will need to be scalarized.
if (Opcode == Instruction::FCmp && !ST->hasMVEFloatOps()) {
// One scalaization insert, one scalarization extract and the cost of the
// fcmps.
return BaseT::getScalarizationOverhead(VecValTy, false, true) +
BaseT::getScalarizationOverhead(VecCondTy, true, false) +
VecValTy->getNumElements() *
getCmpSelInstrCost(Opcode, ValTy->getScalarType(),
VecCondTy->getScalarType(), VecPred, CostKind,
I);
}
std::pair<InstructionCost, MVT> LT =
TLI->getTypeLegalizationCost(DL, ValTy);
int BaseCost = ST->getMVEVectorCostFactor(CostKind);
// There are two types - the input that specifies the type of the compare
// and the output vXi1 type. Because we don't know how the output will be
// split, we may need an expensive shuffle to get two in sync. This has the
// effect of making larger than legal compares (v8i32 for example)
// expensive.
if (LT.second.getVectorNumElements() > 2) {
if (LT.first > 1)
return LT.first * BaseCost +
BaseT::getScalarizationOverhead(VecCondTy, true, false);
return BaseCost;
}
}
// Default to cheap (throughput/size of 1 instruction) but adjust throughput
// for "multiple beats" potentially needed by MVE instructions.
int BaseCost = 1;
if (ST->hasMVEIntegerOps() && ValTy->isVectorTy())
BaseCost = ST->getMVEVectorCostFactor(CostKind);
return BaseCost *
BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
}
InstructionCost ARMTTIImpl::getAddressComputationCost(Type *Ty,
ScalarEvolution *SE,
const SCEV *Ptr) {
// Address computations in vectorized code with non-consecutive addresses will
// likely result in more instructions compared to scalar code where the
// computation can more often be merged into the index mode. The resulting
// extra micro-ops can significantly decrease throughput.
unsigned NumVectorInstToHideOverhead = 10;
int MaxMergeDistance = 64;
if (ST->hasNEON()) {
if (Ty->isVectorTy() && SE &&
!BaseT::isConstantStridedAccessLessThan(SE, Ptr, MaxMergeDistance + 1))
return NumVectorInstToHideOverhead;
// In many cases the address computation is not merged into the instruction
// addressing mode.
return 1;
}
return BaseT::getAddressComputationCost(Ty, SE, Ptr);
}
bool ARMTTIImpl::isProfitableLSRChainElement(Instruction *I) {
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
// If a VCTP is part of a chain, it's already profitable and shouldn't be
// optimized, else LSR may block tail-predication.
switch (II->getIntrinsicID()) {
case Intrinsic::arm_mve_vctp8:
case Intrinsic::arm_mve_vctp16:
case Intrinsic::arm_mve_vctp32:
case Intrinsic::arm_mve_vctp64:
return true;
default:
break;
}
}
return false;
}
bool ARMTTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
if (!EnableMaskedLoadStores || !ST->hasMVEIntegerOps())
return false;
if (auto *VecTy = dyn_cast<FixedVectorType>(DataTy)) {
// Don't support v2i1 yet.
if (VecTy->getNumElements() == 2)
return false;
// We don't support extending fp types.
unsigned VecWidth = DataTy->getPrimitiveSizeInBits();
if (VecWidth != 128 && VecTy->getElementType()->isFloatingPointTy())
return false;
}
unsigned EltWidth = DataTy->getScalarSizeInBits();
return (EltWidth == 32 && Alignment >= 4) ||
(EltWidth == 16 && Alignment >= 2) || (EltWidth == 8);
}
bool ARMTTIImpl::isLegalMaskedGather(Type *Ty, Align Alignment) {
if (!EnableMaskedGatherScatters || !ST->hasMVEIntegerOps())
return false;
// This method is called in 2 places:
// - from the vectorizer with a scalar type, in which case we need to get
// this as good as we can with the limited info we have (and rely on the cost
// model for the rest).
// - from the masked intrinsic lowering pass with the actual vector type.
// For MVE, we have a custom lowering pass that will already have custom
// legalised any gathers that we can to MVE intrinsics, and want to expand all
// the rest. The pass runs before the masked intrinsic lowering pass, so if we
// are here, we know we want to expand.
if (isa<VectorType>(Ty))
return false;
unsigned EltWidth = Ty->getScalarSizeInBits();
return ((EltWidth == 32 && Alignment >= 4) ||
(EltWidth == 16 && Alignment >= 2) || EltWidth == 8);
}
/// Given a memcpy/memset/memmove instruction, return the number of memory
/// operations performed, via querying findOptimalMemOpLowering. Returns -1 if a
/// call is used.
int ARMTTIImpl::getNumMemOps(const IntrinsicInst *I) const {
MemOp MOp;
unsigned DstAddrSpace = ~0u;
unsigned SrcAddrSpace = ~0u;
const Function *F = I->getParent()->getParent();
if (const auto *MC = dyn_cast<MemTransferInst>(I)) {
ConstantInt *C = dyn_cast<ConstantInt>(MC->getLength());
// If 'size' is not a constant, a library call will be generated.
if (!C)
return -1;
const unsigned Size = C->getValue().getZExtValue();
const Align DstAlign = *MC->getDestAlign();
const Align SrcAlign = *MC->getSourceAlign();
MOp = MemOp::Copy(Size, /*DstAlignCanChange*/ false, DstAlign, SrcAlign,
/*IsVolatile*/ false);
DstAddrSpace = MC->getDestAddressSpace();
SrcAddrSpace = MC->getSourceAddressSpace();
}
else if (const auto *MS = dyn_cast<MemSetInst>(I)) {
ConstantInt *C = dyn_cast<ConstantInt>(MS->getLength());
// If 'size' is not a constant, a library call will be generated.
if (!C)
return -1;
const unsigned Size = C->getValue().getZExtValue();
const Align DstAlign = *MS->getDestAlign();
MOp = MemOp::Set(Size, /*DstAlignCanChange*/ false, DstAlign,
/*IsZeroMemset*/ false, /*IsVolatile*/ false);
DstAddrSpace = MS->getDestAddressSpace();
}
else
llvm_unreachable("Expected a memcpy/move or memset!");
unsigned Limit, Factor = 2;
switch(I->getIntrinsicID()) {
case Intrinsic::memcpy:
Limit = TLI->getMaxStoresPerMemcpy(F->hasMinSize());
break;
case Intrinsic::memmove:
Limit = TLI->getMaxStoresPerMemmove(F->hasMinSize());
break;
case Intrinsic::memset:
Limit = TLI->getMaxStoresPerMemset(F->hasMinSize());
Factor = 1;
break;
default:
llvm_unreachable("Expected a memcpy/move or memset!");
}
// MemOps will be poplulated with a list of data types that needs to be
// loaded and stored. That's why we multiply the number of elements by 2 to
// get the cost for this memcpy.
std::vector<EVT> MemOps;
if (getTLI()->findOptimalMemOpLowering(
MemOps, Limit, MOp, DstAddrSpace,
SrcAddrSpace, F->getAttributes()))
return MemOps.size() * Factor;
// If we can't find an optimal memop lowering, return the default cost
return -1;
}
InstructionCost ARMTTIImpl::getMemcpyCost(const Instruction *I) {
int NumOps = getNumMemOps(cast<IntrinsicInst>(I));
// To model the cost of a library call, we assume 1 for the call, and
// 3 for the argument setup.
if (NumOps == -1)
return 4;
return NumOps;
}
InstructionCost ARMTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
VectorType *Tp, ArrayRef<int> Mask,
int Index, VectorType *SubTp) {
Kind = improveShuffleKindFromMask(Kind, Mask);
if (ST->hasNEON()) {
if (Kind == TTI::SK_Broadcast) {
static const CostTblEntry NEONDupTbl[] = {
// VDUP handles these cases.
{ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4i16, 1},
{ISD::VECTOR_SHUFFLE, MVT::v8i8, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
{ISD::VECTOR_SHUFFLE, MVT::v16i8, 1}};
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
if (const auto *Entry =
CostTableLookup(NEONDupTbl, ISD::VECTOR_SHUFFLE, LT.second))
return LT.first * Entry->Cost;
}
if (Kind == TTI::SK_Reverse) {
static const CostTblEntry NEONShuffleTbl[] = {
// Reverse shuffle cost one instruction if we are shuffling within a
// double word (vrev) or two if we shuffle a quad word (vrev, vext).
{ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4i16, 1},
{ISD::VECTOR_SHUFFLE, MVT::v8i8, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
{ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
{ISD::VECTOR_SHUFFLE, MVT::v8i16, 2},
{ISD::VECTOR_SHUFFLE, MVT::v16i8, 2}};
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
if (const auto *Entry =
CostTableLookup(NEONShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second))
return LT.first * Entry->Cost;
}
if (Kind == TTI::SK_Select) {
static const CostTblEntry NEONSelShuffleTbl[] = {
// Select shuffle cost table for ARM. Cost is the number of
// instructions
// required to create the shuffled vector.
{ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
{ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
{ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
{ISD::VECTOR_SHUFFLE, MVT::v4i16, 2},
{ISD::VECTOR_SHUFFLE, MVT::v8i16, 16},
{ISD::VECTOR_SHUFFLE, MVT::v16i8, 32}};
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
if (const auto *Entry = CostTableLookup(NEONSelShuffleTbl,
ISD::VECTOR_SHUFFLE, LT.second))
return LT.first * Entry->Cost;
}
}
if (ST->hasMVEIntegerOps()) {
if (Kind == TTI::SK_Broadcast) {
static const CostTblEntry MVEDupTbl[] = {
// VDUP handles these cases.
{ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
{ISD::VECTOR_SHUFFLE, MVT::v16i8, 1},
{ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
{ISD::VECTOR_SHUFFLE, MVT::v8f16, 1}};
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
if (const auto *Entry = CostTableLookup(MVEDupTbl, ISD::VECTOR_SHUFFLE,
LT.second))
return LT.first * Entry->Cost *
ST->getMVEVectorCostFactor(TTI::TCK_RecipThroughput);
}
if (!Mask.empty()) {
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
if (Mask.size() <= LT.second.getVectorNumElements() &&
(isVREVMask(Mask, LT.second, 16) || isVREVMask(Mask, LT.second, 32) ||
isVREVMask(Mask, LT.second, 64)))
return ST->getMVEVectorCostFactor(TTI::TCK_RecipThroughput) * LT.first;
}
}
int BaseCost = ST->hasMVEIntegerOps() && Tp->isVectorTy()
? ST->getMVEVectorCostFactor(TTI::TCK_RecipThroughput)
: 1;
return BaseCost * BaseT::getShuffleCost(Kind, Tp, Mask, Index, SubTp);
}
InstructionCost ARMTTIImpl::getArithmeticInstrCost(
unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
TTI::OperandValueProperties Opd1PropInfo,
TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
const Instruction *CxtI) {
int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
if (ST->isThumb() && CostKind == TTI::TCK_CodeSize && Ty->isIntegerTy(1)) {
// Make operations on i1 relatively expensive as this often involves
// combining predicates. AND and XOR should be easier to handle with IT
// blocks.
switch (ISDOpcode) {
default:
break;
case ISD::AND:
case ISD::XOR:
return 2;
case ISD::OR:
return 3;
}
}
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
if (ST->hasNEON()) {
const unsigned FunctionCallDivCost = 20;
const unsigned ReciprocalDivCost = 10;
static const CostTblEntry CostTbl[] = {
// Division.
// These costs are somewhat random. Choose a cost of 20 to indicate that
// vectorizing devision (added function call) is going to be very expensive.
// Double registers types.
{ ISD::SDIV, MVT::v1i64, 1 * FunctionCallDivCost},
{ ISD::UDIV, MVT::v1i64, 1 * FunctionCallDivCost},
{ ISD::SREM, MVT::v1i64, 1 * FunctionCallDivCost},
{ ISD::UREM, MVT::v1i64, 1 * FunctionCallDivCost},
{ ISD::SDIV, MVT::v2i32, 2 * FunctionCallDivCost},
{ ISD::UDIV, MVT::v2i32, 2 * FunctionCallDivCost},
{ ISD::SREM, MVT::v2i32, 2 * FunctionCallDivCost},
{ ISD::UREM, MVT::v2i32, 2 * FunctionCallDivCost},
{ ISD::SDIV, MVT::v4i16, ReciprocalDivCost},
{ ISD::UDIV, MVT::v4i16, ReciprocalDivCost},
{ ISD::SREM, MVT::v4i16, 4 * FunctionCallDivCost},
{ ISD::UREM, MVT::v4i16, 4 * FunctionCallDivCost},
{ ISD::SDIV, MVT::v8i8, ReciprocalDivCost},
{ ISD::UDIV, MVT::v8i8, ReciprocalDivCost},
{ ISD::SREM, MVT::v8i8, 8 * FunctionCallDivCost},
{ ISD::UREM, MVT::v8i8, 8 * FunctionCallDivCost},
// Quad register types.
{ ISD::SDIV, MVT::v2i64, 2 * FunctionCallDivCost},
{ ISD::UDIV, MVT::v2i64, 2 * FunctionCallDivCost},
{ ISD::SREM, MVT::v2i64, 2 * FunctionCallDivCost},
{ ISD::UREM, MVT::v2i64, 2 * FunctionCallDivCost},
{ ISD::SDIV, MVT::v4i32, 4 * FunctionCallDivCost},
{ ISD::UDIV, MVT::v4i32, 4 * FunctionCallDivCost},
{ ISD::SREM, MVT::v4i32, 4 * FunctionCallDivCost},
{ ISD::UREM, MVT::v4i32, 4 * FunctionCallDivCost},
{ ISD::SDIV, MVT::v8i16, 8 * FunctionCallDivCost},
{ ISD::UDIV, MVT::v8i16, 8 * FunctionCallDivCost},
{ ISD::SREM, MVT::v8i16, 8 * FunctionCallDivCost},
{ ISD::UREM, MVT::v8i16, 8 * FunctionCallDivCost},
{ ISD::SDIV, MVT::v16i8, 16 * FunctionCallDivCost},
{ ISD::UDIV, MVT::v16i8, 16 * FunctionCallDivCost},
{ ISD::SREM, MVT::v16i8, 16 * FunctionCallDivCost},
{ ISD::UREM, MVT::v16i8, 16 * FunctionCallDivCost},
// Multiplication.
};
if (const auto *Entry = CostTableLookup(CostTbl, ISDOpcode, LT.second))
return LT.first * Entry->Cost;
InstructionCost Cost = BaseT::getArithmeticInstrCost(
Opcode, Ty, CostKind, Op1Info, Op2Info, Opd1PropInfo, Opd2PropInfo);
// This is somewhat of a hack. The problem that we are facing is that SROA
// creates a sequence of shift, and, or instructions to construct values.
// These sequences are recognized by the ISel and have zero-cost. Not so for
// the vectorized code. Because we have support for v2i64 but not i64 those
// sequences look particularly beneficial to vectorize.
// To work around this we increase the cost of v2i64 operations to make them
// seem less beneficial.
if (LT.second == MVT::v2i64 &&
Op2Info == TargetTransformInfo::OK_UniformConstantValue)
Cost += 4;
return Cost;
}
// If this operation is a shift on arm/thumb2, it might well be folded into
// the following instruction, hence having a cost of 0.
auto LooksLikeAFreeShift = [&]() {
if (ST->isThumb1Only() || Ty->isVectorTy())
return false;
if (!CxtI || !CxtI->hasOneUse() || !CxtI->isShift())
return false;
if (Op2Info != TargetTransformInfo::OK_UniformConstantValue)
return false;
// Folded into a ADC/ADD/AND/BIC/CMP/EOR/MVN/ORR/ORN/RSB/SBC/SUB
switch (cast<Instruction>(CxtI->user_back())->getOpcode()) {
case Instruction::Add:
case Instruction::Sub:
case Instruction::And:
case Instruction::Xor:
case Instruction::Or:
case Instruction::ICmp:
return true;
default:
return false;
}
};
if (LooksLikeAFreeShift())
return 0;
// Default to cheap (throughput/size of 1 instruction) but adjust throughput
// for "multiple beats" potentially needed by MVE instructions.
int BaseCost = 1;
if (ST->hasMVEIntegerOps() && Ty->isVectorTy())
BaseCost = ST->getMVEVectorCostFactor(CostKind);
// The rest of this mostly follows what is done in BaseT::getArithmeticInstrCost,
// without treating floats as more expensive that scalars or increasing the
// costs for custom operations. The results is also multiplied by the
// MVEVectorCostFactor where appropriate.
if (TLI->isOperationLegalOrCustomOrPromote(ISDOpcode, LT.second))
return LT.first * BaseCost;
// Else this is expand, assume that we need to scalarize this op.
if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
unsigned Num = VTy->getNumElements();
InstructionCost Cost =
getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind);
// Return the cost of multiple scalar invocation plus the cost of
// inserting and extracting the values.
SmallVector<Type *> Tys(Args.size(), Ty);
return BaseT::getScalarizationOverhead(VTy, Args, Tys) + Num * Cost;
}
return BaseCost;
}
InstructionCost ARMTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
MaybeAlign Alignment,
unsigned AddressSpace,
TTI::TargetCostKind CostKind,
const Instruction *I) {
// TODO: Handle other cost kinds.
if (CostKind != TTI::TCK_RecipThroughput)
return 1;
// Type legalization can't handle structs
if (TLI->getValueType(DL, Src, true) == MVT::Other)
return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
CostKind);
if (ST->hasNEON() && Src->isVectorTy() &&
(Alignment && *Alignment != Align(16)) &&
cast<VectorType>(Src)->getElementType()->isDoubleTy()) {
// Unaligned loads/stores are extremely inefficient.
// We need 4 uops for vst.1/vld.1 vs 1uop for vldr/vstr.
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
return LT.first * 4;
}
// MVE can optimize a fpext(load(4xhalf)) using an extending integer load.
// Same for stores.
if (ST->hasMVEFloatOps() && isa<FixedVectorType>(Src) && I &&
((Opcode == Instruction::Load && I->hasOneUse() &&
isa<FPExtInst>(*I->user_begin())) ||
(Opcode == Instruction::Store && isa<FPTruncInst>(I->getOperand(0))))) {
FixedVectorType *SrcVTy = cast<FixedVectorType>(Src);
Type *DstTy =
Opcode == Instruction::Load
? (*I->user_begin())->getType()
: cast<Instruction>(I->getOperand(0))->getOperand(0)->getType();
if (SrcVTy->getNumElements() == 4 && SrcVTy->getScalarType()->isHalfTy() &&
DstTy->getScalarType()->isFloatTy())
return ST->getMVEVectorCostFactor(CostKind);
}
int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
? ST->getMVEVectorCostFactor(CostKind)
: 1;
return BaseCost * BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
CostKind, I);
}
InstructionCost
ARMTTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
unsigned AddressSpace,
TTI::TargetCostKind CostKind) {
if (ST->hasMVEIntegerOps()) {
if (Opcode == Instruction::Load && isLegalMaskedLoad(Src, Alignment))
return ST->getMVEVectorCostFactor(CostKind);
if (Opcode == Instruction::Store && isLegalMaskedStore(Src, Alignment))
return ST->getMVEVectorCostFactor(CostKind);
}
if (!isa<FixedVectorType>(Src))
return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
CostKind);
// Scalar cost, which is currently very high due to the efficiency of the
// generated code.
return cast<FixedVectorType>(Src)->getNumElements() * 8;
}
InstructionCost ARMTTIImpl::getInterleavedMemoryOpCost(
unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
bool UseMaskForCond, bool UseMaskForGaps) {
assert(Factor >= 2 && "Invalid interleave factor");
assert(isa<VectorType>(VecTy) && "Expect a vector type");
// vldN/vstN doesn't support vector types of i64/f64 element.
bool EltIs64Bits = DL.getTypeSizeInBits(VecTy->getScalarType()) == 64;
if (Factor <= TLI->getMaxSupportedInterleaveFactor() && !EltIs64Bits &&
!UseMaskForCond && !UseMaskForGaps) {
unsigned NumElts = cast<FixedVectorType>(VecTy)->getNumElements();
auto *SubVecTy =
FixedVectorType::get(VecTy->getScalarType(), NumElts / Factor);
// vldN/vstN only support legal vector types of size 64 or 128 in bits.
// Accesses having vector types that are a multiple of 128 bits can be
// matched to more than one vldN/vstN instruction.
int BaseCost =
ST->hasMVEIntegerOps() ? ST->getMVEVectorCostFactor(CostKind) : 1;
if (NumElts % Factor == 0 &&
TLI->isLegalInterleavedAccessType(Factor, SubVecTy, Alignment, DL))
return Factor * BaseCost * TLI->getNumInterleavedAccesses(SubVecTy, DL);
// Some smaller than legal interleaved patterns are cheap as we can make
// use of the vmovn or vrev patterns to interleave a standard load. This is
// true for v4i8, v8i8 and v4i16 at least (but not for v4f16 as it is
// promoted differently). The cost of 2 here is then a load and vrev or
// vmovn.
if (ST->hasMVEIntegerOps() && Factor == 2 && NumElts / Factor > 2 &&
VecTy->isIntOrIntVectorTy() &&
DL.getTypeSizeInBits(SubVecTy).getFixedSize() <= 64)
return 2 * BaseCost;
}
return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
Alignment, AddressSpace, CostKind,
UseMaskForCond, UseMaskForGaps);
}
InstructionCost ARMTTIImpl::getGatherScatterOpCost(
unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) {
using namespace PatternMatch;
if (!ST->hasMVEIntegerOps() || !EnableMaskedGatherScatters)
return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
Alignment, CostKind, I);
assert(DataTy->isVectorTy() && "Can't do gather/scatters on scalar!");
auto *VTy = cast<FixedVectorType>(DataTy);
// TODO: Splitting, once we do that.
unsigned NumElems = VTy->getNumElements();
unsigned EltSize = VTy->getScalarSizeInBits();
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, DataTy);
// For now, it is assumed that for the MVE gather instructions the loads are
// all effectively serialised. This means the cost is the scalar cost
// multiplied by the number of elements being loaded. This is possibly very
// conservative, but even so we still end up vectorising loops because the
// cost per iteration for many loops is lower than for scalar loops.
InstructionCost VectorCost =
NumElems * LT.first * ST->getMVEVectorCostFactor(CostKind);
// The scalarization cost should be a lot higher. We use the number of vector
// elements plus the scalarization overhead.
InstructionCost ScalarCost =
NumElems * LT.first + BaseT::getScalarizationOverhead(VTy, true, false) +
BaseT::getScalarizationOverhead(VTy, false, true);
if (EltSize < 8 || Alignment < EltSize / 8)
return ScalarCost;
unsigned ExtSize = EltSize;
// Check whether there's a single user that asks for an extended type
if (I != nullptr) {
// Dependent of the caller of this function, a gather instruction will
// either have opcode Instruction::Load or be a call to the masked_gather
// intrinsic
if ((I->getOpcode() == Instruction::Load ||
match(I, m_Intrinsic<Intrinsic::masked_gather>())) &&
I->hasOneUse()) {
const User *Us = *I->users().begin();
if (isa<ZExtInst>(Us) || isa<SExtInst>(Us)) {
// only allow valid type combinations
unsigned TypeSize =
cast<Instruction>(Us)->getType()->getScalarSizeInBits();
if (((TypeSize == 32 && (EltSize == 8 || EltSize == 16)) ||
(TypeSize == 16 && EltSize == 8)) &&
TypeSize * NumElems == 128) {
ExtSize = TypeSize;
}
}
}
// Check whether the input data needs to be truncated
TruncInst *T;
if ((I->getOpcode() == Instruction::Store ||
match(I, m_Intrinsic<Intrinsic::masked_scatter>())) &&
(T = dyn_cast<TruncInst>(I->getOperand(0)))) {
// Only allow valid type combinations
unsigned TypeSize = T->getOperand(0)->getType()->getScalarSizeInBits();
if (((EltSize == 16 && TypeSize == 32) ||
(EltSize == 8 && (TypeSize == 32 || TypeSize == 16))) &&
TypeSize * NumElems == 128)
ExtSize = TypeSize;
}
}
if (ExtSize * NumElems != 128 || NumElems < 4)
return ScalarCost;
// Any (aligned) i32 gather will not need to be scalarised.
if (ExtSize == 32)
return VectorCost;
// For smaller types, we need to ensure that the gep's inputs are correctly
// extended from a small enough value. Other sizes (including i64) are
// scalarized for now.
if (ExtSize != 8 && ExtSize != 16)
return ScalarCost;
if (const auto *BC = dyn_cast<BitCastInst>(Ptr))
Ptr = BC->getOperand(0);
if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
if (GEP->getNumOperands() != 2)
return ScalarCost;
unsigned Scale = DL.getTypeAllocSize(GEP->getResultElementType());
// Scale needs to be correct (which is only relevant for i16s).
if (Scale != 1 && Scale * 8 != ExtSize)
return ScalarCost;
// And we need to zext (not sext) the indexes from a small enough type.
if (const auto *ZExt = dyn_cast<ZExtInst>(GEP->getOperand(1))) {
if (ZExt->getOperand(0)->getType()->getScalarSizeInBits() <= ExtSize)
return VectorCost;
}
return ScalarCost;
}
return ScalarCost;
}
InstructionCost
ARMTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
Optional<FastMathFlags> FMF,
TTI::TargetCostKind CostKind) {
if (TTI::requiresOrderedReduction(FMF))
return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
EVT ValVT = TLI->getValueType(DL, ValTy);
int ISD = TLI->InstructionOpcodeToISD(Opcode);
if (!ST->hasMVEIntegerOps() || !ValVT.isSimple() || ISD != ISD::ADD)
return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
static const CostTblEntry CostTblAdd[]{
{ISD::ADD, MVT::v16i8, 1},
{ISD::ADD, MVT::v8i16, 1},
{ISD::ADD, MVT::v4i32, 1},
};
if (const auto *Entry = CostTableLookup(CostTblAdd, ISD, LT.second))
return Entry->Cost * ST->getMVEVectorCostFactor(CostKind) * LT.first;
return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
}
InstructionCost
ARMTTIImpl::getExtendedAddReductionCost(bool IsMLA, bool IsUnsigned,
Type *ResTy, VectorType *ValTy,
TTI::TargetCostKind CostKind) {
EVT ValVT = TLI->getValueType(DL, ValTy);
EVT ResVT = TLI->getValueType(DL, ResTy);
if (ST->hasMVEIntegerOps() && ValVT.isSimple() && ResVT.isSimple()) {
std::pair<InstructionCost, MVT> LT =
TLI->getTypeLegalizationCost(DL, ValTy);
// The legal cases are:
// VADDV u/s 8/16/32
// VMLAV u/s 8/16/32
// VADDLV u/s 32
// VMLALV u/s 16/32
// Codegen currently cannot always handle larger than legal vectors very
// well, especially for predicated reductions where the mask needs to be
// split, so restrict to 128bit or smaller input types.
unsigned RevVTSize = ResVT.getSizeInBits();
if (ValVT.getSizeInBits() <= 128 &&
((LT.second == MVT::v16i8 && RevVTSize <= 32) ||
(LT.second == MVT::v8i16 && RevVTSize <= (IsMLA ? 64u : 32u)) ||
(LT.second == MVT::v4i32 && RevVTSize <= 64)))
return ST->getMVEVectorCostFactor(CostKind) * LT.first;
}
return BaseT::getExtendedAddReductionCost(IsMLA, IsUnsigned, ResTy, ValTy,
CostKind);
}
InstructionCost
ARMTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
TTI::TargetCostKind CostKind) {
switch (ICA.getID()) {
case Intrinsic::get_active_lane_mask:
// Currently we make a somewhat optimistic assumption that
// active_lane_mask's are always free. In reality it may be freely folded
// into a tail predicated loop, expanded into a VCPT or expanded into a lot
// of add/icmp code. We may need to improve this in the future, but being
// able to detect if it is free or not involves looking at a lot of other
// code. We currently assume that the vectorizer inserted these, and knew
// what it was doing in adding one.
if (ST->hasMVEIntegerOps())
return 0;
break;
case Intrinsic::sadd_sat:
case Intrinsic::ssub_sat:
case Intrinsic::uadd_sat:
case Intrinsic::usub_sat: {
if (!ST->hasMVEIntegerOps())
break;
Type *VT = ICA.getReturnType();
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, VT);
if (LT.second == MVT::v4i32 || LT.second == MVT::v8i16 ||
LT.second == MVT::v16i8) {
// This is a base cost of 1 for the vqadd, plus 3 extract shifts if we
// need to extend the type, as it uses shr(qadd(shl, shl)).
unsigned Instrs =
LT.second.getScalarSizeInBits() == VT->getScalarSizeInBits() ? 1 : 4;
return LT.first * ST->getMVEVectorCostFactor(CostKind) * Instrs;
}
break;
}
case Intrinsic::abs:
case Intrinsic::smin:
case Intrinsic::smax:
case Intrinsic::umin:
case Intrinsic::umax: {
if (!ST->hasMVEIntegerOps())
break;
Type *VT = ICA.getReturnType();
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, VT);
if (LT.second == MVT::v4i32 || LT.second == MVT::v8i16 ||
LT.second == MVT::v16i8)
return LT.first * ST->getMVEVectorCostFactor(CostKind);
break;
}
case Intrinsic::minnum:
case Intrinsic::maxnum: {
if (!ST->hasMVEFloatOps())
break;
Type *VT = ICA.getReturnType();
std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, VT);
if (LT.second == MVT::v4f32 || LT.second == MVT::v8f16)
return LT.first * ST->getMVEVectorCostFactor(CostKind);
break;
}
}
return BaseT::getIntrinsicInstrCost(ICA, CostKind);
}
bool ARMTTIImpl::isLoweredToCall(const Function *F) {
if (!F->isIntrinsic())
BaseT::isLoweredToCall(F);
// Assume all Arm-specific intrinsics map to an instruction.
if (F->getName().startswith("llvm.arm"))
return false;
switch (F->getIntrinsicID()) {
default: break;
case Intrinsic::powi:
case Intrinsic::sin:
case Intrinsic::cos:
case Intrinsic::pow:
case Intrinsic::log:
case Intrinsic::log10:
case Intrinsic::log2:
case Intrinsic::exp:
case Intrinsic::exp2:
return true;
case Intrinsic::sqrt:
case Intrinsic::fabs:
case Intrinsic::copysign:
case Intrinsic::floor:
case Intrinsic::ceil:
case Intrinsic::trunc:
case Intrinsic::rint:
case Intrinsic::nearbyint:
case Intrinsic::round:
case Intrinsic::canonicalize:
case Intrinsic::lround:
case Intrinsic::llround:
case Intrinsic::lrint:
case Intrinsic::llrint:
if (F->getReturnType()->isDoubleTy() && !ST->hasFP64())
return true;
if (F->getReturnType()->isHalfTy() && !ST->hasFullFP16())
return true;
// Some operations can be handled by vector instructions and assume
// unsupported vectors will be expanded into supported scalar ones.
// TODO Handle scalar operations properly.
return !ST->hasFPARMv8Base() && !ST->hasVFP2Base();
case Intrinsic::masked_store:
case Intrinsic::masked_load:
case Intrinsic::masked_gather:
case Intrinsic::masked_scatter:
return !ST->hasMVEIntegerOps();
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
case Intrinsic::sadd_sat:
case Intrinsic::uadd_sat:
case Intrinsic::ssub_sat:
case Intrinsic::usub_sat:
return false;
}
return BaseT::isLoweredToCall(F);
}
bool ARMTTIImpl::maybeLoweredToCall(Instruction &I) {
unsigned ISD = TLI->InstructionOpcodeToISD(I.getOpcode());
EVT VT = TLI->getValueType(DL, I.getType(), true);
if (TLI->getOperationAction(ISD, VT) == TargetLowering::LibCall)
return true;
// Check if an intrinsic will be lowered to a call and assume that any
// other CallInst will generate a bl.
if (auto *Call = dyn_cast<CallInst>(&I)) {
if (auto *II = dyn_cast<IntrinsicInst>(Call)) {
switch(II->getIntrinsicID()) {
case Intrinsic::memcpy:
case Intrinsic::memset:
case Intrinsic::memmove:
return getNumMemOps(II) == -1;
default:
if (const Function *F = Call->getCalledFunction())
return isLoweredToCall(F);
}
}
return true;
}
// FPv5 provides conversions between integer, double-precision,
// single-precision, and half-precision formats.
switch (I.getOpcode()) {
default:
break;
case Instruction::FPToSI:
case Instruction::FPToUI:
case Instruction::SIToFP:
case Instruction::UIToFP:
case Instruction::FPTrunc:
case Instruction::FPExt:
return !ST->hasFPARMv8Base();
}
// FIXME: Unfortunately the approach of checking the Operation Action does
// not catch all cases of Legalization that use library calls. Our
// Legalization step categorizes some transformations into library calls as
// Custom, Expand or even Legal when doing type legalization. So for now
// we have to special case for instance the SDIV of 64bit integers and the
// use of floating point emulation.
if (VT.isInteger() && VT.getSizeInBits() >= 64) {
switch (ISD) {
default:
break;
case ISD::SDIV:
case ISD::UDIV:
case ISD::SREM:
case ISD::UREM:
case ISD::SDIVREM:
case ISD::UDIVREM:
return true;
}
}
// Assume all other non-float operations are supported.
if (!VT.isFloatingPoint())
return false;
// We'll need a library call to handle most floats when using soft.
if (TLI->useSoftFloat()) {
switch (I.getOpcode()) {
default:
return true;
case Instruction::Alloca:
case Instruction::Load:
case Instruction::Store:
case Instruction::Select:
case Instruction::PHI:
return false;
}
}
// We'll need a libcall to perform double precision operations on a single
// precision only FPU.
if (I.getType()->isDoubleTy() && !ST->hasFP64())
return true;
// Likewise for half precision arithmetic.
if (I.getType()->isHalfTy() && !ST->hasFullFP16())
return true;
return false;
}
bool ARMTTIImpl::isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
AssumptionCache &AC,
TargetLibraryInfo *LibInfo,
HardwareLoopInfo &HWLoopInfo) {
// Low-overhead branches are only supported in the 'low-overhead branch'
// extension of v8.1-m.
if (!ST->hasLOB() || DisableLowOverheadLoops) {
LLVM_DEBUG(dbgs() << "ARMHWLoops: Disabled\n");
return false;
}
if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
LLVM_DEBUG(dbgs() << "ARMHWLoops: No BETC\n");
return false;
}
const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
LLVM_DEBUG(dbgs() << "ARMHWLoops: Uncomputable BETC\n");
return false;
}
const SCEV *TripCountSCEV =
SE.getAddExpr(BackedgeTakenCount,
SE.getOne(BackedgeTakenCount->getType()));
// We need to store the trip count in LR, a 32-bit register.
if (SE.getUnsignedRangeMax(TripCountSCEV).getBitWidth() > 32) {
LLVM_DEBUG(dbgs() << "ARMHWLoops: Trip count does not fit into 32bits\n");
return false;
}
// Making a call will trash LR and clear LO_BRANCH_INFO, so there's little
// point in generating a hardware loop if that's going to happen.
auto IsHardwareLoopIntrinsic = [](Instruction &I) {
if (auto *Call = dyn_cast<IntrinsicInst>(&I)) {
switch (Call->getIntrinsicID()) {
default:
break;
case Intrinsic::start_loop_iterations:
case Intrinsic::test_start_loop_iterations:
case Intrinsic::loop_decrement:
case Intrinsic::loop_decrement_reg:
return true;
}
}
return false;
};
// Scan the instructions to see if there's any that we know will turn into a
// call or if this loop is already a low-overhead loop or will become a tail
// predicated loop.
bool IsTailPredLoop = false;
auto ScanLoop = [&](Loop *L) {
for (auto *BB : L->getBlocks()) {
for (auto &I : *BB) {
if (maybeLoweredToCall(I) || IsHardwareLoopIntrinsic(I) ||
isa<InlineAsm>(I)) {
LLVM_DEBUG(dbgs() << "ARMHWLoops: Bad instruction: " << I << "\n");
return false;
}
if (auto *II = dyn_cast<IntrinsicInst>(&I))
IsTailPredLoop |=
II->getIntrinsicID() == Intrinsic::get_active_lane_mask ||
II->getIntrinsicID() == Intrinsic::arm_mve_vctp8 ||
II->getIntrinsicID() == Intrinsic::arm_mve_vctp16 ||
II->getIntrinsicID() == Intrinsic::arm_mve_vctp32 ||
II->getIntrinsicID() == Intrinsic::arm_mve_vctp64;
}
}
return true;
};
// Visit inner loops.
for (auto Inner : *L)
if (!ScanLoop(Inner))
return false;
if (!ScanLoop(L))
return false;
// TODO: Check whether the trip count calculation is expensive. If L is the
// inner loop but we know it has a low trip count, calculating that trip
// count (in the parent loop) may be detrimental.
LLVMContext &C = L->getHeader()->getContext();
HWLoopInfo.CounterInReg = true;
HWLoopInfo.IsNestingLegal = false;
HWLoopInfo.PerformEntryTest = AllowWLSLoops && !IsTailPredLoop;
HWLoopInfo.CountType = Type::getInt32Ty(C);
HWLoopInfo.LoopDecrement = ConstantInt::get(HWLoopInfo.CountType, 1);
return true;
}
static bool canTailPredicateInstruction(Instruction &I, int &ICmpCount) {
// We don't allow icmp's, and because we only look at single block loops,
// we simply count the icmps, i.e. there should only be 1 for the backedge.
if (isa<ICmpInst>(&I) && ++ICmpCount > 1)
return false;
// FIXME: This is a workaround for poor cost modelling. Min/Max intrinsics are
// not currently canonical, but soon will be. Code without them uses icmp, and
// so is not tail predicated as per the condition above. In order to get the
// same performance we treat min and max the same as an icmp for tailpred
// purposes for the moment (we often rely on non-tailpred and higher VF's to
// pick more optimial instructions like VQDMULH. They need to be recognized
// directly by the vectorizer).
if (auto *II = dyn_cast<IntrinsicInst>(&I))
if ((II->getIntrinsicID() == Intrinsic::smin ||
II->getIntrinsicID() == Intrinsic::smax ||
II->getIntrinsicID() == Intrinsic::umin ||
II->getIntrinsicID() == Intrinsic::umax) &&
++ICmpCount > 1)
return false;
if (isa<FCmpInst>(&I))
return false;
// We could allow extending/narrowing FP loads/stores, but codegen is
// too inefficient so reject this for now.
if (isa<FPExtInst>(&I) || isa<FPTruncInst>(&I))
return false;
// Extends have to be extending-loads
if (isa<SExtInst>(&I) || isa<ZExtInst>(&I) )
if (!I.getOperand(0)->hasOneUse() || !isa<LoadInst>(I.getOperand(0)))
return false;
// Truncs have to be narrowing-stores
if (isa<TruncInst>(&I) )
if (!I.hasOneUse() || !isa<StoreInst>(*I.user_begin()))
return false;
return true;
}
// To set up a tail-predicated loop, we need to know the total number of
// elements processed by that loop. Thus, we need to determine the element
// size and:
// 1) it should be uniform for all operations in the vector loop, so we
// e.g. don't want any widening/narrowing operations.
// 2) it should be smaller than i64s because we don't have vector operations
// that work on i64s.
// 3) we don't want elements to be reversed or shuffled, to make sure the
// tail-predication masks/predicates the right lanes.
//
static bool canTailPredicateLoop(Loop *L, LoopInfo *LI, ScalarEvolution &SE,
const DataLayout &DL,
const LoopAccessInfo *LAI) {
LLVM_DEBUG(dbgs() << "Tail-predication: checking allowed instructions\n");
// If there are live-out values, it is probably a reduction. We can predicate
// most reduction operations freely under MVE using a combination of
// prefer-predicated-reduction-select and inloop reductions. We limit this to
// floating point and integer reductions, but don't check for operators
// specifically here. If the value ends up not being a reduction (and so the
// vectorizer cannot tailfold the loop), we should fall back to standard
// vectorization automatically.
SmallVector< Instruction *, 8 > LiveOuts;
LiveOuts = llvm::findDefsUsedOutsideOfLoop(L);
bool ReductionsDisabled =
EnableTailPredication == TailPredication::EnabledNoReductions ||
EnableTailPredication == TailPredication::ForceEnabledNoReductions;
for (auto *I : LiveOuts) {
if (!I->getType()->isIntegerTy() && !I->getType()->isFloatTy() &&
!I->getType()->isHalfTy()) {
LLVM_DEBUG(dbgs() << "Don't tail-predicate loop with non-integer/float "
"live-out value\n");
return false;
}
if (ReductionsDisabled) {
LLVM_DEBUG(dbgs() << "Reductions not enabled\n");
return false;
}
}
// Next, check that all instructions can be tail-predicated.
PredicatedScalarEvolution PSE = LAI->getPSE();
SmallVector<Instruction *, 16> LoadStores;
int ICmpCount = 0;
for (BasicBlock *BB : L->blocks()) {
for (Instruction &I : BB->instructionsWithoutDebug()) {
if (isa<PHINode>(&I))
continue;
if (!canTailPredicateInstruction(I, ICmpCount)) {
LLVM_DEBUG(dbgs() << "Instruction not allowed: "; I.dump());
return false;
}
Type *T = I.getType();
if (T->isPointerTy())
T = T->getPointerElementType();
if (T->getScalarSizeInBits() > 32) {
LLVM_DEBUG(dbgs() << "Unsupported Type: "; T->dump());
return false;
}
if (isa<StoreInst>(I) || isa<LoadInst>(I)) {
Value *Ptr = getLoadStorePointerOperand(&I);
Type *AccessTy = getLoadStoreType(&I);
int64_t NextStride = getPtrStride(PSE, AccessTy, Ptr, L);
if (NextStride == 1) {
// TODO: for now only allow consecutive strides of 1. We could support
// other strides as long as it is uniform, but let's keep it simple
// for now.
continue;
} else if (NextStride == -1 ||
(NextStride == 2 && MVEMaxSupportedInterleaveFactor >= 2) ||
(NextStride == 4 && MVEMaxSupportedInterleaveFactor >= 4)) {
LLVM_DEBUG(dbgs()
<< "Consecutive strides of 2 found, vld2/vstr2 can't "
"be tail-predicated\n.");
return false;
// TODO: don't tail predicate if there is a reversed load?
} else if (EnableMaskedGatherScatters) {
// Gather/scatters do allow loading from arbitrary strides, at
// least if they are loop invariant.
// TODO: Loop variant strides should in theory work, too, but
// this requires further testing.
const SCEV *PtrScev = PSE.getSE()->getSCEV(Ptr);
if (auto AR = dyn_cast<SCEVAddRecExpr>(PtrScev)) {
const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
if (PSE.getSE()->isLoopInvariant(Step, L))
continue;
}
}
LLVM_DEBUG(dbgs() << "Bad stride found, can't "
"tail-predicate\n.");
return false;
}
}
}
LLVM_DEBUG(dbgs() << "tail-predication: all instructions allowed!\n");
return true;
}
bool ARMTTIImpl::preferPredicateOverEpilogue(Loop *L, LoopInfo *LI,
ScalarEvolution &SE,
AssumptionCache &AC,
TargetLibraryInfo *TLI,
DominatorTree *DT,
const LoopAccessInfo *LAI) {
if (!EnableTailPredication) {
LLVM_DEBUG(dbgs() << "Tail-predication not enabled.\n");
return false;
}
// Creating a predicated vector loop is the first step for generating a
// tail-predicated hardware loop, for which we need the MVE masked
// load/stores instructions:
if (!ST->hasMVEIntegerOps())
return false;
// For now, restrict this to single block loops.
if (L->getNumBlocks() > 1) {
LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: not a single block "
"loop.\n");
return false;
}
assert(L->isInnermost() && "preferPredicateOverEpilogue: inner-loop expected");
HardwareLoopInfo HWLoopInfo(L);
if (!HWLoopInfo.canAnalyze(*LI)) {
LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
"analyzable.\n");
return false;
}
// This checks if we have the low-overhead branch architecture
// extension, and if we will create a hardware-loop:
if (!isHardwareLoopProfitable(L, SE, AC, TLI, HWLoopInfo)) {
LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
"profitable.\n");
return false;
}
if (!HWLoopInfo.isHardwareLoopCandidate(SE, *LI, *DT)) {
LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
"a candidate.\n");
return false;
}
return canTailPredicateLoop(L, LI, SE, DL, LAI);
}
bool ARMTTIImpl::emitGetActiveLaneMask() const {
if (!ST->hasMVEIntegerOps() || !EnableTailPredication)
return false;
// Intrinsic @llvm.get.active.lane.mask is supported.
// It is used in the MVETailPredication pass, which requires the number of
// elements processed by this vector loop to setup the tail-predicated
// loop.
return true;
}
void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
TTI::UnrollingPreferences &UP,
OptimizationRemarkEmitter *ORE) {
// Enable Upper bound unrolling universally, not dependant upon the conditions
// below.
UP.UpperBound = true;
// Only currently enable these preferences for M-Class cores.
if (!ST->isMClass())
return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP, ORE);
// 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 (ST->hasBranchPredictor() && 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) {
// Don't unroll vectorised loop. MVE does not benefit from it as much as
// scalar code.
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*, 4> Operands(I.operand_values());
Cost +=
getUserCost(&I, Operands, TargetTransformInfo::TCK_SizeAndLatency);
}
}
// On v6m cores, there are very few registers available. We can easily end up
// spilling and reloading more registers in an unrolled loop. Look at the
// number of LCSSA phis as a rough measure of how many registers will need to
// be live out of the loop, reducing the default unroll count if more than 1
// value is needed. In the long run, all of this should be being learnt by a
// machine.
unsigned UnrollCount = 4;
if (ST->isThumb1Only()) {
unsigned ExitingValues = 0;
SmallVector<BasicBlock *, 4> ExitBlocks;
L->getExitBlocks(ExitBlocks);
for (auto *Exit : ExitBlocks) {
// Count the number of LCSSA phis. Exclude values coming from GEP's as
// only the last is expected to be needed for address operands.
unsigned LiveOuts = count_if(Exit->phis(), [](auto &PH) {
return PH.getNumOperands() != 1 ||
!isa<GetElementPtrInst>(PH.getOperand(0));
});
ExitingValues = ExitingValues < LiveOuts ? LiveOuts : ExitingValues;
}
if (ExitingValues)
UnrollCount /= ExitingValues;
if (UnrollCount <= 1)
return;
}
LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n");
LLVM_DEBUG(dbgs() << "Default Runtime Unroll Count: " << UnrollCount << "\n");
UP.Partial = true;
UP.Runtime = true;
UP.UnrollRemainder = true;
UP.DefaultUnrollRuntimeCount = UnrollCount;
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 ARMTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
TTI::PeelingPreferences &PP) {
BaseT::getPeelingPreferences(L, SE, PP);
}
bool ARMTTIImpl::preferInLoopReduction(unsigned Opcode, Type *Ty,
TTI::ReductionFlags Flags) const {
if (!ST->hasMVEIntegerOps())
return false;
unsigned ScalarBits = Ty->getScalarSizeInBits();
switch (Opcode) {
case Instruction::Add:
return ScalarBits <= 64;
default:
return false;
}
}
bool ARMTTIImpl::preferPredicatedReductionSelect(
unsigned Opcode, Type *Ty, TTI::ReductionFlags Flags) const {
if (!ST->hasMVEIntegerOps())
return false;
return true;
}