Google Git
Sign in
llvm / llvm-project / f5c8c1eedb44861cda8885a04813de999ca9d18a / . / llvm / lib / Target / AArch64 / AArch64TargetTransformInfo.cpp
blob: 50a88b944ef57de948655ddcd0b79219d7f59a91 [file] [log] [blame]
//===-- AArch64TargetTransformInfo.cpp - AArch64 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 "AArch64TargetTransformInfo.h"
#include "AArch64ExpandImm.h"
#include "AArch64PerfectShuffle.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
#include "Utils/AArch64SMEAttributes.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/CodeGen/BasicTTIImpl.h"
#include "llvm/CodeGen/CostTable.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsAArch64.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/Debug.h"
#include "llvm/TargetParser/AArch64TargetParser.h"
#include "llvm/Transforms/InstCombine/InstCombiner.h"
#include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"
#include <algorithm>
#include <optional>
using namespace llvm;
using namespace llvm::PatternMatch;
#define DEBUG_TYPE "aarch64tti"
static cl::opt<bool> EnableFalkorHWPFUnrollFix("enable-falkor-hwpf-unroll-fix",
cl::init(true), cl::Hidden);
static cl::opt<bool> SVEPreferFixedOverScalableIfEqualCost(
"sve-prefer-fixed-over-scalable-if-equal", cl::Hidden);
static cl::opt<unsigned> SVEGatherOverhead("sve-gather-overhead", cl::init(10),
cl::Hidden);
static cl::opt<unsigned> SVEScatterOverhead("sve-scatter-overhead",
cl::init(10), cl::Hidden);
static cl::opt<unsigned> SVETailFoldInsnThreshold("sve-tail-folding-insn-threshold",
cl::init(15), cl::Hidden);
static cl::opt<unsigned>
NeonNonConstStrideOverhead("neon-nonconst-stride-overhead", cl::init(10),
cl::Hidden);
static cl::opt<unsigned> CallPenaltyChangeSM(
"call-penalty-sm-change", cl::init(5), cl::Hidden,
cl::desc(
"Penalty of calling a function that requires a change to PSTATE.SM"));
static cl::opt<unsigned> InlineCallPenaltyChangeSM(
"inline-call-penalty-sm-change", cl::init(10), cl::Hidden,
cl::desc("Penalty of inlining a call that requires a change to PSTATE.SM"));
static cl::opt<bool> EnableOrLikeSelectOpt("enable-aarch64-or-like-select",
cl::init(true), cl::Hidden);
static cl::opt<bool> EnableLSRCostOpt("enable-aarch64-lsr-cost-opt",
cl::init(true), cl::Hidden);
// A complete guess as to a reasonable cost.
static cl::opt<unsigned>
BaseHistCntCost("aarch64-base-histcnt-cost", cl::init(8), cl::Hidden,
cl::desc("The cost of a histcnt instruction"));
static cl::opt<unsigned> DMBLookaheadThreshold(
"dmb-lookahead-threshold", cl::init(10), cl::Hidden,
cl::desc("The number of instructions to search for a redundant dmb"));
namespace {
class TailFoldingOption {
// These bitfields will only ever be set to something non-zero in operator=,
// when setting the -sve-tail-folding option. This option should always be of
// the form (default|simple|all|disable)[+(Flag1|Flag2|etc)], where here
// InitialBits is one of (disabled|all|simple). EnableBits represents
// additional flags we're enabling, and DisableBits for those flags we're
// disabling. The default flag is tracked in the variable NeedsDefault, since
// at the time of setting the option we may not know what the default value
// for the CPU is.
TailFoldingOpts InitialBits = TailFoldingOpts::Disabled;
TailFoldingOpts EnableBits = TailFoldingOpts::Disabled;
TailFoldingOpts DisableBits = TailFoldingOpts::Disabled;
// This value needs to be initialised to true in case the user does not
// explicitly set the -sve-tail-folding option.
bool NeedsDefault = true;
void setInitialBits(TailFoldingOpts Bits) { InitialBits = Bits; }
void setNeedsDefault(bool V) { NeedsDefault = V; }
void setEnableBit(TailFoldingOpts Bit) {
EnableBits |= Bit;
DisableBits &= ~Bit;
}
void setDisableBit(TailFoldingOpts Bit) {
EnableBits &= ~Bit;
DisableBits |= Bit;
}
TailFoldingOpts getBits(TailFoldingOpts DefaultBits) const {
TailFoldingOpts Bits = TailFoldingOpts::Disabled;
assert((InitialBits == TailFoldingOpts::Disabled || !NeedsDefault) &&
"Initial bits should only include one of "
"(disabled|all|simple|default)");
Bits = NeedsDefault ? DefaultBits : InitialBits;
Bits |= EnableBits;
Bits &= ~DisableBits;
return Bits;
}
void reportError(std::string Opt) {
errs() << "invalid argument '" << Opt
<< "' to -sve-tail-folding=; the option should be of the form\n"
" (disabled|all|default|simple)[+(reductions|recurrences"
"|reverse|noreductions|norecurrences|noreverse)]\n";
report_fatal_error("Unrecognised tail-folding option");
}
public:
void operator=(const std::string &Val) {
// If the user explicitly sets -sve-tail-folding= then treat as an error.
if (Val.empty()) {
reportError("");
return;
}
// Since the user is explicitly setting the option we don't automatically
// need the default unless they require it.
setNeedsDefault(false);
SmallVector<StringRef, 4> TailFoldTypes;
StringRef(Val).split(TailFoldTypes, '+', -1, false);
unsigned StartIdx = 1;
if (TailFoldTypes[0] == "disabled")
setInitialBits(TailFoldingOpts::Disabled);
else if (TailFoldTypes[0] == "all")
setInitialBits(TailFoldingOpts::All);
else if (TailFoldTypes[0] == "default")
setNeedsDefault(true);
else if (TailFoldTypes[0] == "simple")
setInitialBits(TailFoldingOpts::Simple);
else {
StartIdx = 0;
setInitialBits(TailFoldingOpts::Disabled);
}
for (unsigned I = StartIdx; I < TailFoldTypes.size(); I++) {
if (TailFoldTypes[I] == "reductions")
setEnableBit(TailFoldingOpts::Reductions);
else if (TailFoldTypes[I] == "recurrences")
setEnableBit(TailFoldingOpts::Recurrences);
else if (TailFoldTypes[I] == "reverse")
setEnableBit(TailFoldingOpts::Reverse);
else if (TailFoldTypes[I] == "noreductions")
setDisableBit(TailFoldingOpts::Reductions);
else if (TailFoldTypes[I] == "norecurrences")
setDisableBit(TailFoldingOpts::Recurrences);
else if (TailFoldTypes[I] == "noreverse")
setDisableBit(TailFoldingOpts::Reverse);
else
reportError(Val);
}
}
bool satisfies(TailFoldingOpts DefaultBits, TailFoldingOpts Required) const {
return (getBits(DefaultBits) & Required) == Required;
}
};
} // namespace
TailFoldingOption TailFoldingOptionLoc;
static cl::opt<TailFoldingOption, true, cl::parser<std::string>> SVETailFolding(
"sve-tail-folding",
cl::desc(
"Control the use of vectorisation using tail-folding for SVE where the"
" option is specified in the form (Initial)[+(Flag1|Flag2|...)]:"
"\ndisabled (Initial) No loop types will vectorize using "
"tail-folding"
"\ndefault (Initial) Uses the default tail-folding settings for "
"the target CPU"
"\nall (Initial) All legal loop types will vectorize using "
"tail-folding"
"\nsimple (Initial) Use tail-folding for simple loops (not "
"reductions or recurrences)"
"\nreductions Use tail-folding for loops containing reductions"
"\nnoreductions Inverse of above"
"\nrecurrences Use tail-folding for loops containing fixed order "
"recurrences"
"\nnorecurrences Inverse of above"
"\nreverse Use tail-folding for loops requiring reversed "
"predicates"
"\nnoreverse Inverse of above"),
cl::location(TailFoldingOptionLoc));
// Experimental option that will only be fully functional when the
// code-generator is changed to use SVE instead of NEON for all fixed-width
// operations.
static cl::opt<bool> EnableFixedwidthAutovecInStreamingMode(
"enable-fixedwidth-autovec-in-streaming-mode", cl::init(false), cl::Hidden);
// Experimental option that will only be fully functional when the cost-model
// and code-generator have been changed to avoid using scalable vector
// instructions that are not legal in streaming SVE mode.
static cl::opt<bool> EnableScalableAutovecInStreamingMode(
"enable-scalable-autovec-in-streaming-mode", cl::init(false), cl::Hidden);
static bool isSMEABIRoutineCall(const CallInst &CI) {
const auto *F = CI.getCalledFunction();
return F && StringSwitch<bool>(F->getName())
.Case("__arm_sme_state", true)
.Case("__arm_tpidr2_save", true)
.Case("__arm_tpidr2_restore", true)
.Case("__arm_za_disable", true)
.Default(false);
}
/// Returns true if the function has explicit operations that can only be
/// lowered using incompatible instructions for the selected mode. This also
/// returns true if the function F may use or modify ZA state.
static bool hasPossibleIncompatibleOps(const Function *F) {
for (const BasicBlock &BB : *F) {
for (const Instruction &I : BB) {
// Be conservative for now and assume that any call to inline asm or to
// intrinsics could could result in non-streaming ops (e.g. calls to
// @llvm.aarch64.* or @llvm.gather/scatter intrinsics). We can assume that
// all native LLVM instructions can be lowered to compatible instructions.
if (isa<CallInst>(I) && !I.isDebugOrPseudoInst() &&
(cast<CallInst>(I).isInlineAsm() || isa<IntrinsicInst>(I) ||
isSMEABIRoutineCall(cast<CallInst>(I))))
return true;
}
}
return false;
}
uint64_t AArch64TTIImpl::getFeatureMask(const Function &F) const {
StringRef AttributeStr =
isMultiversionedFunction(F) ? "fmv-features" : "target-features";
StringRef FeatureStr = F.getFnAttribute(AttributeStr).getValueAsString();
SmallVector<StringRef, 8> Features;
FeatureStr.split(Features, ",");
return AArch64::getFMVPriority(Features);
}
bool AArch64TTIImpl::isMultiversionedFunction(const Function &F) const {
return F.hasFnAttribute("fmv-features");
}
const FeatureBitset AArch64TTIImpl::InlineInverseFeatures = {
AArch64::FeatureExecuteOnly,
};
bool AArch64TTIImpl::areInlineCompatible(const Function *Caller,
const Function *Callee) const {
SMEAttrs CallerAttrs(*Caller), CalleeAttrs(*Callee);
// When inlining, we should consider the body of the function, not the
// interface.
if (CalleeAttrs.hasStreamingBody()) {
CalleeAttrs.set(SMEAttrs::SM_Compatible, false);
CalleeAttrs.set(SMEAttrs::SM_Enabled, true);
}
if (CalleeAttrs.isNewZA() || CalleeAttrs.isNewZT0())
return false;
if (CallerAttrs.requiresLazySave(CalleeAttrs) ||
CallerAttrs.requiresSMChange(CalleeAttrs) ||
CallerAttrs.requiresPreservingZT0(CalleeAttrs) ||
CallerAttrs.requiresPreservingAllZAState(CalleeAttrs)) {
if (hasPossibleIncompatibleOps(Callee))
return false;
}
const TargetMachine &TM = getTLI()->getTargetMachine();
const FeatureBitset &CallerBits =
TM.getSubtargetImpl(*Caller)->getFeatureBits();
const FeatureBitset &CalleeBits =
TM.getSubtargetImpl(*Callee)->getFeatureBits();
// Adjust the feature bitsets by inverting some of the bits. This is needed
// for target features that represent restrictions rather than capabilities,
// for example a "+execute-only" callee can be inlined into a caller without
// "+execute-only", but not vice versa.
FeatureBitset EffectiveCallerBits = CallerBits ^ InlineInverseFeatures;
FeatureBitset EffectiveCalleeBits = CalleeBits ^ InlineInverseFeatures;
return (EffectiveCallerBits & EffectiveCalleeBits) == EffectiveCalleeBits;
}
bool AArch64TTIImpl::areTypesABICompatible(
const Function *Caller, const Function *Callee,
const ArrayRef<Type *> &Types) const {
if (!BaseT::areTypesABICompatible(Caller, Callee, Types))
return false;
// We need to ensure that argument promotion does not attempt to promote
// pointers to fixed-length vector types larger than 128 bits like
// <8 x float> (and pointers to aggregate types which have such fixed-length
// vector type members) into the values of the pointees. Such vector types
// are used for SVE VLS but there is no ABI for SVE VLS arguments and the
// backend cannot lower such value arguments. The 128-bit fixed-length SVE
// types can be safely treated as 128-bit NEON types and they cannot be
// distinguished in IR.
if (ST->useSVEForFixedLengthVectors() && llvm::any_of(Types, [](Type *Ty) {
auto FVTy = dyn_cast<FixedVectorType>(Ty);
return FVTy &&
FVTy->getScalarSizeInBits() * FVTy->getNumElements() > 128;
}))
return false;
return true;
}
unsigned
AArch64TTIImpl::getInlineCallPenalty(const Function *F, const CallBase &Call,
unsigned DefaultCallPenalty) const {
// This function calculates a penalty for executing Call in F.
//
// There are two ways this function can be called:
// (1) F:
// call from F -> G (the call here is Call)
//
// For (1), Call.getCaller() == F, so it will always return a high cost if
// a streaming-mode change is required (thus promoting the need to inline the
// function)
//
// (2) F:
// call from F -> G (the call here is not Call)
// G:
// call from G -> H (the call here is Call)
//
// For (2), if after inlining the body of G into F the call to H requires a
// streaming-mode change, and the call to G from F would also require a
// streaming-mode change, then there is benefit to do the streaming-mode
// change only once and avoid inlining of G into F.
SMEAttrs FAttrs(*F);
SMEAttrs CalleeAttrs(Call);
if (FAttrs.requiresSMChange(CalleeAttrs)) {
if (F == Call.getCaller()) // (1)
return CallPenaltyChangeSM * DefaultCallPenalty;
if (FAttrs.requiresSMChange(SMEAttrs(*Call.getCaller()))) // (2)
return InlineCallPenaltyChangeSM * DefaultCallPenalty;
}
return DefaultCallPenalty;
}
bool AArch64TTIImpl::shouldMaximizeVectorBandwidth(
TargetTransformInfo::RegisterKind K) const {
assert(K != TargetTransformInfo::RGK_Scalar);
return (K == TargetTransformInfo::RGK_FixedWidthVector &&
ST->isNeonAvailable());
}
/// Calculate the cost of materializing a 64-bit value. This helper
/// method might only calculate a fraction of a larger immediate. Therefore it
/// is valid to return a cost of ZERO.
InstructionCost AArch64TTIImpl::getIntImmCost(int64_t Val) const {
// Check if the immediate can be encoded within an instruction.
if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, 64))
return 0;
if (Val < 0)
Val = ~Val;
// Calculate how many moves we will need to materialize this constant.
SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn;
AArch64_IMM::expandMOVImm(Val, 64, Insn);
return Insn.size();
}
/// Calculate the cost of materializing the given constant.
InstructionCost
AArch64TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind) const {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
if (BitSize == 0)
return ~0U;
// Sign-extend all constants to a multiple of 64-bit.
APInt ImmVal = Imm;
if (BitSize & 0x3f)
ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
// Split the constant into 64-bit chunks and calculate the cost for each
// chunk.
InstructionCost Cost = 0;
for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
int64_t Val = Tmp.getSExtValue();
Cost += getIntImmCost(Val);
}
// We need at least one instruction to materialze the constant.
return std::max<InstructionCost>(1, Cost);
}
InstructionCost AArch64TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind,
Instruction *Inst) const {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
// There is no cost model for constants with a bit size of 0. Return TCC_Free
// here, so that constant hoisting will ignore this constant.
if (BitSize == 0)
return TTI::TCC_Free;
unsigned ImmIdx = ~0U;
switch (Opcode) {
default:
return TTI::TCC_Free;
case Instruction::GetElementPtr:
// Always hoist the base address of a GetElementPtr.
if (Idx == 0)
return 2 * TTI::TCC_Basic;
return TTI::TCC_Free;
case Instruction::Store:
ImmIdx = 0;
break;
case Instruction::Add:
case Instruction::Sub:
case Instruction::Mul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor:
case Instruction::ICmp:
ImmIdx = 1;
break;
// Always return TCC_Free for the shift value of a shift instruction.
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
if (Idx == 1)
return TTI::TCC_Free;
break;
case Instruction::Trunc:
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::IntToPtr:
case Instruction::PtrToInt:
case Instruction::BitCast:
case Instruction::PHI:
case Instruction::Call:
case Instruction::Select:
case Instruction::Ret:
case Instruction::Load:
break;
}
if (Idx == ImmIdx) {
int NumConstants = (BitSize + 63) / 64;
InstructionCost Cost = AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
return (Cost <= NumConstants * TTI::TCC_Basic)
? static_cast<int>(TTI::TCC_Free)
: Cost;
}
return AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
}
InstructionCost
AArch64TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind) const {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
// There is no cost model for constants with a bit size of 0. Return TCC_Free
// here, so that constant hoisting will ignore this constant.
if (BitSize == 0)
return TTI::TCC_Free;
// Most (all?) AArch64 intrinsics do not support folding immediates into the
// selected instruction, so we compute the materialization cost for the
// immediate directly.
if (IID >= Intrinsic::aarch64_addg && IID <= Intrinsic::aarch64_udiv)
return AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
switch (IID) {
default:
return TTI::TCC_Free;
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow:
if (Idx == 1) {
int NumConstants = (BitSize + 63) / 64;
InstructionCost Cost = AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
return (Cost <= NumConstants * TTI::TCC_Basic)
? static_cast<int>(TTI::TCC_Free)
: Cost;
}
break;
case Intrinsic::experimental_stackmap:
if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
return TTI::TCC_Free;
break;
case Intrinsic::experimental_patchpoint_void:
case Intrinsic::experimental_patchpoint:
if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
return TTI::TCC_Free;
break;
case Intrinsic::experimental_gc_statepoint:
if ((Idx < 5) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
return TTI::TCC_Free;
break;
}
return AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
}
TargetTransformInfo::PopcntSupportKind
AArch64TTIImpl::getPopcntSupport(unsigned TyWidth) const {
assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
if (TyWidth == 32 || TyWidth == 64)
return TTI::PSK_FastHardware;
// TODO: AArch64TargetLowering::LowerCTPOP() supports 128bit popcount.
return TTI::PSK_Software;
}
static bool isUnpackedVectorVT(EVT VecVT) {
return VecVT.isScalableVector() &&
VecVT.getSizeInBits().getKnownMinValue() < AArch64::SVEBitsPerBlock;
}
static InstructionCost getHistogramCost(const IntrinsicCostAttributes &ICA) {
Type *BucketPtrsTy = ICA.getArgTypes()[0]; // Type of vector of pointers
Type *EltTy = ICA.getArgTypes()[1]; // Type of bucket elements
unsigned TotalHistCnts = 1;
unsigned EltSize = EltTy->getScalarSizeInBits();
// Only allow (up to 64b) integers or pointers
if ((!EltTy->isIntegerTy() && !EltTy->isPointerTy()) || EltSize > 64)
return InstructionCost::getInvalid();
// FIXME: We should be able to generate histcnt for fixed-length vectors
// using ptrue with a specific VL.
if (VectorType *VTy = dyn_cast<VectorType>(BucketPtrsTy)) {
unsigned EC = VTy->getElementCount().getKnownMinValue();
if (!isPowerOf2_64(EC) || !VTy->isScalableTy())
return InstructionCost::getInvalid();
// HistCnt only supports 32b and 64b element types
unsigned LegalEltSize = EltSize <= 32 ? 32 : 64;
if (EC == 2 || (LegalEltSize == 32 && EC == 4))
return InstructionCost(BaseHistCntCost);
unsigned NaturalVectorWidth = AArch64::SVEBitsPerBlock / LegalEltSize;
TotalHistCnts = EC / NaturalVectorWidth;
}
return InstructionCost(BaseHistCntCost * TotalHistCnts);
}
InstructionCost
AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
TTI::TargetCostKind CostKind) const {
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
auto *RetTy = ICA.getReturnType();
if (auto *VTy = dyn_cast<ScalableVectorType>(RetTy))
if (VTy->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
switch (ICA.getID()) {
case Intrinsic::experimental_vector_histogram_add:
if (!ST->hasSVE2())
return InstructionCost::getInvalid();
return getHistogramCost(ICA);
case Intrinsic::umin:
case Intrinsic::umax:
case Intrinsic::smin:
case Intrinsic::smax: {
static const auto ValidMinMaxTys = {MVT::v8i8, MVT::v16i8, MVT::v4i16,
MVT::v8i16, MVT::v2i32, MVT::v4i32,
MVT::nxv16i8, MVT::nxv8i16, MVT::nxv4i32,
MVT::nxv2i64};
auto LT = getTypeLegalizationCost(RetTy);
// v2i64 types get converted to cmp+bif hence the cost of 2
if (LT.second == MVT::v2i64)
return LT.first * 2;
if (any_of(ValidMinMaxTys, [&LT](MVT M) { return M == LT.second; }))
return LT.first;
break;
}
case Intrinsic::sadd_sat:
case Intrinsic::ssub_sat:
case Intrinsic::uadd_sat:
case Intrinsic::usub_sat: {
static const auto ValidSatTys = {MVT::v8i8, MVT::v16i8, MVT::v4i16,
MVT::v8i16, MVT::v2i32, MVT::v4i32,
MVT::v2i64};
auto LT = getTypeLegalizationCost(RetTy);
// This is a base cost of 1 for the vadd, plus 3 extract shifts if we
// need to extend the type, as it uses shr(qadd(shl, shl)).
unsigned Instrs =
LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits() ? 1 : 4;
if (any_of(ValidSatTys, [&LT](MVT M) { return M == LT.second; }))
return LT.first * Instrs;
break;
}
case Intrinsic::abs: {
static const auto ValidAbsTys = {MVT::v8i8, MVT::v16i8, MVT::v4i16,
MVT::v8i16, MVT::v2i32, MVT::v4i32,
MVT::v2i64};
auto LT = getTypeLegalizationCost(RetTy);
if (any_of(ValidAbsTys, [&LT](MVT M) { return M == LT.second; }))
return LT.first;
break;
}
case Intrinsic::bswap: {
static const auto ValidAbsTys = {MVT::v4i16, MVT::v8i16, MVT::v2i32,
MVT::v4i32, MVT::v2i64};
auto LT = getTypeLegalizationCost(RetTy);
if (any_of(ValidAbsTys, [&LT](MVT M) { return M == LT.second; }) &&
LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits())
return LT.first;
break;
}
case Intrinsic::stepvector: {
InstructionCost Cost = 1; // Cost of the `index' instruction
auto LT = getTypeLegalizationCost(RetTy);
// Legalisation of illegal vectors involves an `index' instruction plus
// (LT.first - 1) vector adds.
if (LT.first > 1) {
Type *LegalVTy = EVT(LT.second).getTypeForEVT(RetTy->getContext());
InstructionCost AddCost =
getArithmeticInstrCost(Instruction::Add, LegalVTy, CostKind);
Cost += AddCost * (LT.first - 1);
}
return Cost;
}
case Intrinsic::vector_extract:
case Intrinsic::vector_insert: {
// If both the vector and subvector types are legal types and the index
// is 0, then this should be a no-op or simple operation; return a
// relatively low cost.
// If arguments aren't actually supplied, then we cannot determine the
// value of the index. We also want to skip predicate types.
if (ICA.getArgs().size() != ICA.getArgTypes().size() ||
ICA.getReturnType()->getScalarType()->isIntegerTy(1))
break;
LLVMContext &C = RetTy->getContext();
EVT VecVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
bool IsExtract = ICA.getID() == Intrinsic::vector_extract;
EVT SubVecVT = IsExtract ? getTLI()->getValueType(DL, RetTy)
: getTLI()->getValueType(DL, ICA.getArgTypes()[1]);
// Skip this if either the vector or subvector types are unpacked
// SVE types; they may get lowered to stack stores and loads.
if (isUnpackedVectorVT(VecVT) || isUnpackedVectorVT(SubVecVT))
break;
TargetLoweringBase::LegalizeKind SubVecLK =
getTLI()->getTypeConversion(C, SubVecVT);
TargetLoweringBase::LegalizeKind VecLK =
getTLI()->getTypeConversion(C, VecVT);
const Value *Idx = IsExtract ? ICA.getArgs()[1] : ICA.getArgs()[2];
const ConstantInt *CIdx = cast<ConstantInt>(Idx);
if (SubVecLK.first == TargetLoweringBase::TypeLegal &&
VecLK.first == TargetLoweringBase::TypeLegal && CIdx->isZero())
return TTI::TCC_Free;
break;
}
case Intrinsic::bitreverse: {
static const CostTblEntry BitreverseTbl[] = {
{Intrinsic::bitreverse, MVT::i32, 1},
{Intrinsic::bitreverse, MVT::i64, 1},
{Intrinsic::bitreverse, MVT::v8i8, 1},
{Intrinsic::bitreverse, MVT::v16i8, 1},
{Intrinsic::bitreverse, MVT::v4i16, 2},
{Intrinsic::bitreverse, MVT::v8i16, 2},
{Intrinsic::bitreverse, MVT::v2i32, 2},
{Intrinsic::bitreverse, MVT::v4i32, 2},
{Intrinsic::bitreverse, MVT::v1i64, 2},
{Intrinsic::bitreverse, MVT::v2i64, 2},
};
const auto LegalisationCost = getTypeLegalizationCost(RetTy);
const auto *Entry =
CostTableLookup(BitreverseTbl, ICA.getID(), LegalisationCost.second);
if (Entry) {
// Cost Model is using the legal type(i32) that i8 and i16 will be
// converted to +1 so that we match the actual lowering cost
if (TLI->getValueType(DL, RetTy, true) == MVT::i8 ||
TLI->getValueType(DL, RetTy, true) == MVT::i16)
return LegalisationCost.first * Entry->Cost + 1;
return LegalisationCost.first * Entry->Cost;
}
break;
}
case Intrinsic::ctpop: {
if (!ST->hasNEON()) {
// 32-bit or 64-bit ctpop without NEON is 12 instructions.
return getTypeLegalizationCost(RetTy).first * 12;
}
static const CostTblEntry CtpopCostTbl[] = {
{ISD::CTPOP, MVT::v2i64, 4},
{ISD::CTPOP, MVT::v4i32, 3},
{ISD::CTPOP, MVT::v8i16, 2},
{ISD::CTPOP, MVT::v16i8, 1},
{ISD::CTPOP, MVT::i64, 4},
{ISD::CTPOP, MVT::v2i32, 3},
{ISD::CTPOP, MVT::v4i16, 2},
{ISD::CTPOP, MVT::v8i8, 1},
{ISD::CTPOP, MVT::i32, 5},
};
auto LT = getTypeLegalizationCost(RetTy);
MVT MTy = LT.second;
if (const auto *Entry = CostTableLookup(CtpopCostTbl, ISD::CTPOP, MTy)) {
// Extra cost of +1 when illegal vector types are legalized by promoting
// the integer type.
int ExtraCost = MTy.isVector() && MTy.getScalarSizeInBits() !=
RetTy->getScalarSizeInBits()
? 1
: 0;
return LT.first * Entry->Cost + ExtraCost;
}
break;
}
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow: {
static const CostTblEntry WithOverflowCostTbl[] = {
{Intrinsic::sadd_with_overflow, MVT::i8, 3},
{Intrinsic::uadd_with_overflow, MVT::i8, 3},
{Intrinsic::sadd_with_overflow, MVT::i16, 3},
{Intrinsic::uadd_with_overflow, MVT::i16, 3},
{Intrinsic::sadd_with_overflow, MVT::i32, 1},
{Intrinsic::uadd_with_overflow, MVT::i32, 1},
{Intrinsic::sadd_with_overflow, MVT::i64, 1},
{Intrinsic::uadd_with_overflow, MVT::i64, 1},
{Intrinsic::ssub_with_overflow, MVT::i8, 3},
{Intrinsic::usub_with_overflow, MVT::i8, 3},
{Intrinsic::ssub_with_overflow, MVT::i16, 3},
{Intrinsic::usub_with_overflow, MVT::i16, 3},
{Intrinsic::ssub_with_overflow, MVT::i32, 1},
{Intrinsic::usub_with_overflow, MVT::i32, 1},
{Intrinsic::ssub_with_overflow, MVT::i64, 1},
{Intrinsic::usub_with_overflow, MVT::i64, 1},
{Intrinsic::smul_with_overflow, MVT::i8, 5},
{Intrinsic::umul_with_overflow, MVT::i8, 4},
{Intrinsic::smul_with_overflow, MVT::i16, 5},
{Intrinsic::umul_with_overflow, MVT::i16, 4},
{Intrinsic::smul_with_overflow, MVT::i32, 2}, // eg umull;tst
{Intrinsic::umul_with_overflow, MVT::i32, 2}, // eg umull;cmp sxtw
{Intrinsic::smul_with_overflow, MVT::i64, 3}, // eg mul;smulh;cmp
{Intrinsic::umul_with_overflow, MVT::i64, 3}, // eg mul;umulh;cmp asr
};
EVT MTy = TLI->getValueType(DL, RetTy->getContainedType(0), true);
if (MTy.isSimple())
if (const auto *Entry = CostTableLookup(WithOverflowCostTbl, ICA.getID(),
MTy.getSimpleVT()))
return Entry->Cost;
break;
}
case Intrinsic::fptosi_sat:
case Intrinsic::fptoui_sat: {
if (ICA.getArgTypes().empty())
break;
bool IsSigned = ICA.getID() == Intrinsic::fptosi_sat;
auto LT = getTypeLegalizationCost(ICA.getArgTypes()[0]);
EVT MTy = TLI->getValueType(DL, RetTy);
// Check for the legal types, which are where the size of the input and the
// output are the same, or we are using cvt f64->i32 or f32->i64.
if ((LT.second == MVT::f32 || LT.second == MVT::f64 ||
LT.second == MVT::v2f32 || LT.second == MVT::v4f32 ||
LT.second == MVT::v2f64)) {
if ((LT.second.getScalarSizeInBits() == MTy.getScalarSizeInBits() ||
(LT.second == MVT::f64 && MTy == MVT::i32) ||
(LT.second == MVT::f32 && MTy == MVT::i64)))
return LT.first;
// Extending vector types v2f32->v2i64, fcvtl*2 + fcvt*2
if (LT.second.getScalarType() == MVT::f32 && MTy.isFixedLengthVector() &&
MTy.getScalarSizeInBits() == 64)
return LT.first * (MTy.getVectorNumElements() > 2 ? 4 : 2);
}
// Similarly for fp16 sizes. Without FullFP16 we generally need to fcvt to
// f32.
if (LT.second.getScalarType() == MVT::f16 && !ST->hasFullFP16())
return LT.first + getIntrinsicInstrCost(
{ICA.getID(),
RetTy,
{ICA.getArgTypes()[0]->getWithNewType(
Type::getFloatTy(RetTy->getContext()))}},
CostKind);
if ((LT.second == MVT::f16 && MTy == MVT::i32) ||
(LT.second == MVT::f16 && MTy == MVT::i64) ||
((LT.second == MVT::v4f16 || LT.second == MVT::v8f16) &&
(LT.second.getScalarSizeInBits() == MTy.getScalarSizeInBits())))
return LT.first;
// Extending vector types v8f16->v8i32, fcvtl*2 + fcvt*2
if (LT.second.getScalarType() == MVT::f16 && MTy.isFixedLengthVector() &&
MTy.getScalarSizeInBits() == 32)
return LT.first * (MTy.getVectorNumElements() > 4 ? 4 : 2);
// Extending vector types v8f16->v8i32. These current scalarize but the
// codegen could be better.
if (LT.second.getScalarType() == MVT::f16 && MTy.isFixedLengthVector() &&
MTy.getScalarSizeInBits() == 64)
return MTy.getVectorNumElements() * 3;
// If we can we use a legal convert followed by a min+max
if ((LT.second.getScalarType() == MVT::f32 ||
LT.second.getScalarType() == MVT::f64 ||
LT.second.getScalarType() == MVT::f16) &&
LT.second.getScalarSizeInBits() >= MTy.getScalarSizeInBits()) {
Type *LegalTy =
Type::getIntNTy(RetTy->getContext(), LT.second.getScalarSizeInBits());
if (LT.second.isVector())
LegalTy = VectorType::get(LegalTy, LT.second.getVectorElementCount());
InstructionCost Cost = 1;
IntrinsicCostAttributes Attrs1(IsSigned ? Intrinsic::smin : Intrinsic::umin,
LegalTy, {LegalTy, LegalTy});
Cost += getIntrinsicInstrCost(Attrs1, CostKind);
IntrinsicCostAttributes Attrs2(IsSigned ? Intrinsic::smax : Intrinsic::umax,
LegalTy, {LegalTy, LegalTy});
Cost += getIntrinsicInstrCost(Attrs2, CostKind);
return LT.first * Cost +
((LT.second.getScalarType() != MVT::f16 || ST->hasFullFP16()) ? 0
: 1);
}
// Otherwise we need to follow the default expansion that clamps the value
// using a float min/max with a fcmp+sel for nan handling when signed.
Type *FPTy = ICA.getArgTypes()[0]->getScalarType();
RetTy = RetTy->getScalarType();
if (LT.second.isVector()) {
FPTy = VectorType::get(FPTy, LT.second.getVectorElementCount());
RetTy = VectorType::get(RetTy, LT.second.getVectorElementCount());
}
IntrinsicCostAttributes Attrs1(Intrinsic::minnum, FPTy, {FPTy, FPTy});
InstructionCost Cost = getIntrinsicInstrCost(Attrs1, CostKind);
IntrinsicCostAttributes Attrs2(Intrinsic::maxnum, FPTy, {FPTy, FPTy});
Cost += getIntrinsicInstrCost(Attrs2, CostKind);
Cost +=
getCastInstrCost(IsSigned ? Instruction::FPToSI : Instruction::FPToUI,
RetTy, FPTy, TTI::CastContextHint::None, CostKind);
if (IsSigned) {
Type *CondTy = RetTy->getWithNewBitWidth(1);
Cost += getCmpSelInstrCost(BinaryOperator::FCmp, FPTy, CondTy,
CmpInst::FCMP_UNO, CostKind);
Cost += getCmpSelInstrCost(BinaryOperator::Select, RetTy, CondTy,
CmpInst::FCMP_UNO, CostKind);
}
return LT.first * Cost;
}
case Intrinsic::fshl:
case Intrinsic::fshr: {
if (ICA.getArgs().empty())
break;
// TODO: Add handling for fshl where third argument is not a constant.
const TTI::OperandValueInfo OpInfoZ = TTI::getOperandInfo(ICA.getArgs()[2]);
if (!OpInfoZ.isConstant())
break;
const auto LegalisationCost = getTypeLegalizationCost(RetTy);
if (OpInfoZ.isUniform()) {
static const CostTblEntry FshlTbl[] = {
{Intrinsic::fshl, MVT::v4i32, 2}, // shl + usra
{Intrinsic::fshl, MVT::v2i64, 2}, {Intrinsic::fshl, MVT::v16i8, 2},
{Intrinsic::fshl, MVT::v8i16, 2}, {Intrinsic::fshl, MVT::v2i32, 2},
{Intrinsic::fshl, MVT::v8i8, 2}, {Intrinsic::fshl, MVT::v4i16, 2}};
// Costs for both fshl & fshr are the same, so just pass Intrinsic::fshl
// to avoid having to duplicate the costs.
const auto *Entry =
CostTableLookup(FshlTbl, Intrinsic::fshl, LegalisationCost.second);
if (Entry)
return LegalisationCost.first * Entry->Cost;
}
auto TyL = getTypeLegalizationCost(RetTy);
if (!RetTy->isIntegerTy())
break;
// Estimate cost manually, as types like i8 and i16 will get promoted to
// i32 and CostTableLookup will ignore the extra conversion cost.
bool HigherCost = (RetTy->getScalarSizeInBits() != 32 &&
RetTy->getScalarSizeInBits() < 64) ||
(RetTy->getScalarSizeInBits() % 64 != 0);
unsigned ExtraCost = HigherCost ? 1 : 0;
if (RetTy->getScalarSizeInBits() == 32 ||
RetTy->getScalarSizeInBits() == 64)
ExtraCost = 0; // fhsl/fshr for i32 and i64 can be lowered to a single
// extr instruction.
else if (HigherCost)
ExtraCost = 1;
else
break;
return TyL.first + ExtraCost;
}
case Intrinsic::get_active_lane_mask: {
auto *RetTy = dyn_cast<FixedVectorType>(ICA.getReturnType());
if (RetTy) {
EVT RetVT = getTLI()->getValueType(DL, RetTy);
EVT OpVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
if (!getTLI()->shouldExpandGetActiveLaneMask(RetVT, OpVT) &&
!getTLI()->isTypeLegal(RetVT)) {
// We don't have enough context at this point to determine if the mask
// is going to be kept live after the block, which will force the vXi1
// type to be expanded to legal vectors of integers, e.g. v4i1->v4i32.
// For now, we just assume the vectorizer created this intrinsic and
// the result will be the input for a PHI. In this case the cost will
// be extremely high for fixed-width vectors.
// NOTE: getScalarizationOverhead returns a cost that's far too
// pessimistic for the actual generated codegen. In reality there are
// two instructions generated per lane.
return RetTy->getNumElements() * 2;
}
}
break;
}
case Intrinsic::experimental_vector_match: {
auto *NeedleTy = cast<FixedVectorType>(ICA.getArgTypes()[1]);
EVT SearchVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
unsigned SearchSize = NeedleTy->getNumElements();
if (!getTLI()->shouldExpandVectorMatch(SearchVT, SearchSize)) {
// Base cost for MATCH instructions. At least on the Neoverse V2 and
// Neoverse V3, these are cheap operations with the same latency as a
// vector ADD. In most cases, however, we also need to do an extra DUP.
// For fixed-length vectors we currently need an extra five--six
// instructions besides the MATCH.
InstructionCost Cost = 4;
if (isa<FixedVectorType>(RetTy))
Cost += 10;
return Cost;
}
break;
}
case Intrinsic::experimental_cttz_elts: {
EVT ArgVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
if (!getTLI()->shouldExpandCttzElements(ArgVT)) {
// This will consist of a SVE brkb and a cntp instruction. These
// typically have the same latency and half the throughput as a vector
// add instruction.
return 4;
}
break;
}
default:
break;
}
return BaseT::getIntrinsicInstrCost(ICA, CostKind);
}
/// The function will remove redundant reinterprets casting in the presence
/// of the control flow
static std::optional<Instruction *> processPhiNode(InstCombiner &IC,
IntrinsicInst &II) {
SmallVector<Instruction *, 32> Worklist;
auto RequiredType = II.getType();
auto *PN = dyn_cast<PHINode>(II.getArgOperand(0));
assert(PN && "Expected Phi Node!");
// Don't create a new Phi unless we can remove the old one.
if (!PN->hasOneUse())
return std::nullopt;
for (Value *IncValPhi : PN->incoming_values()) {
auto *Reinterpret = dyn_cast<IntrinsicInst>(IncValPhi);
if (!Reinterpret ||
Reinterpret->getIntrinsicID() !=
Intrinsic::aarch64_sve_convert_to_svbool ||
RequiredType != Reinterpret->getArgOperand(0)->getType())
return std::nullopt;
}
// Create the new Phi
IC.Builder.SetInsertPoint(PN);
PHINode *NPN = IC.Builder.CreatePHI(RequiredType, PN->getNumIncomingValues());
Worklist.push_back(PN);
for (unsigned I = 0; I < PN->getNumIncomingValues(); I++) {
auto *Reinterpret = cast<Instruction>(PN->getIncomingValue(I));
NPN->addIncoming(Reinterpret->getOperand(0), PN->getIncomingBlock(I));
Worklist.push_back(Reinterpret);
}
// Cleanup Phi Node and reinterprets
return IC.replaceInstUsesWith(II, NPN);
}
// A collection of properties common to SVE intrinsics that allow for combines
// to be written without needing to know the specific intrinsic.
struct SVEIntrinsicInfo {
//
// Helper routines for common intrinsic definitions.
//
// e.g. llvm.aarch64.sve.add pg, op1, op2
// with IID ==> llvm.aarch64.sve.add_u
static SVEIntrinsicInfo
defaultMergingOp(Intrinsic::ID IID = Intrinsic::not_intrinsic) {
return SVEIntrinsicInfo()
.setGoverningPredicateOperandIdx(0)
.setOperandIdxInactiveLanesTakenFrom(1)
.setMatchingUndefIntrinsic(IID);
}
// e.g. llvm.aarch64.sve.neg inactive, pg, op
static SVEIntrinsicInfo defaultMergingUnaryOp() {
return SVEIntrinsicInfo()
.setGoverningPredicateOperandIdx(1)
.setOperandIdxInactiveLanesTakenFrom(0)
.setOperandIdxWithNoActiveLanes(0);
}
// e.g. llvm.aarch64.sve.fcvtnt inactive, pg, op
static SVEIntrinsicInfo defaultMergingUnaryNarrowingTopOp() {
return SVEIntrinsicInfo()
.setGoverningPredicateOperandIdx(1)
.setOperandIdxInactiveLanesTakenFrom(0);
}
// e.g. llvm.aarch64.sve.add_u pg, op1, op2
static SVEIntrinsicInfo defaultUndefOp() {
return SVEIntrinsicInfo()
.setGoverningPredicateOperandIdx(0)
.setInactiveLanesAreNotDefined();
}
// e.g. llvm.aarch64.sve.prf pg, ptr (GPIndex = 0)
// llvm.aarch64.sve.st1 data, pg, ptr (GPIndex = 1)
static SVEIntrinsicInfo defaultVoidOp(unsigned GPIndex) {
return SVEIntrinsicInfo()
.setGoverningPredicateOperandIdx(GPIndex)
.setInactiveLanesAreUnused();
}
// e.g. llvm.aarch64.sve.cmpeq pg, op1, op2
// llvm.aarch64.sve.ld1 pg, ptr
static SVEIntrinsicInfo defaultZeroingOp() {
return SVEIntrinsicInfo()
.setGoverningPredicateOperandIdx(0)
.setInactiveLanesAreUnused()
.setResultIsZeroInitialized();
}
// All properties relate to predication and thus having a general predicate
// is the minimum requirement to say there is intrinsic info to act on.
explicit operator bool() const { return hasGoverningPredicate(); }
//
// Properties relating to the governing predicate.
//
bool hasGoverningPredicate() const {
return GoverningPredicateIdx != std::numeric_limits<unsigned>::max();
}
unsigned getGoverningPredicateOperandIdx() const {
assert(hasGoverningPredicate() && "Propery not set!");
return GoverningPredicateIdx;
}
SVEIntrinsicInfo &setGoverningPredicateOperandIdx(unsigned Index) {
assert(!hasGoverningPredicate() && "Cannot set property twice!");
GoverningPredicateIdx = Index;
return *this;
}
//
// Properties relating to operations the intrinsic could be transformed into.
// NOTE: This does not mean such a transformation is always possible, but the
// knowledge makes it possible to reuse existing optimisations without needing
// to embed specific handling for each intrinsic. For example, instruction
// simplification can be used to optimise an intrinsic's active lanes.
//
bool hasMatchingUndefIntrinsic() const {
return UndefIntrinsic != Intrinsic::not_intrinsic;
}
Intrinsic::ID getMatchingUndefIntrinsic() const {
assert(hasMatchingUndefIntrinsic() && "Propery not set!");
return UndefIntrinsic;
}
SVEIntrinsicInfo &setMatchingUndefIntrinsic(Intrinsic::ID IID) {
assert(!hasMatchingUndefIntrinsic() && "Cannot set property twice!");
UndefIntrinsic = IID;
return *this;
}
bool hasMatchingIROpode() const { return IROpcode != 0; }
unsigned getMatchingIROpode() const {
assert(hasMatchingIROpode() && "Propery not set!");
return IROpcode;
}
SVEIntrinsicInfo &setMatchingIROpcode(unsigned Opcode) {
assert(!hasMatchingIROpode() && "Cannot set property twice!");
IROpcode = Opcode;
return *this;
}
//
// Properties relating to the result of inactive lanes.
//
bool inactiveLanesTakenFromOperand() const {
return ResultLanes == InactiveLanesTakenFromOperand;
}
unsigned getOperandIdxInactiveLanesTakenFrom() const {
assert(inactiveLanesTakenFromOperand() && "Propery not set!");
return OperandIdxForInactiveLanes;
}
SVEIntrinsicInfo &setOperandIdxInactiveLanesTakenFrom(unsigned Index) {
assert(ResultLanes == Uninitialized && "Cannot set property twice!");
ResultLanes = InactiveLanesTakenFromOperand;
OperandIdxForInactiveLanes = Index;
return *this;
}
bool inactiveLanesAreNotDefined() const {
return ResultLanes == InactiveLanesAreNotDefined;
}
SVEIntrinsicInfo &setInactiveLanesAreNotDefined() {
assert(ResultLanes == Uninitialized && "Cannot set property twice!");
ResultLanes = InactiveLanesAreNotDefined;
return *this;
}
bool inactiveLanesAreUnused() const {
return ResultLanes == InactiveLanesAreUnused;
}
SVEIntrinsicInfo &setInactiveLanesAreUnused() {
assert(ResultLanes == Uninitialized && "Cannot set property twice!");
ResultLanes = InactiveLanesAreUnused;
return *this;
}
// NOTE: Whilst not limited to only inactive lanes, the common use case is:
// inactiveLanesAreZerod =
// resultIsZeroInitialized() && inactiveLanesAreUnused()
bool resultIsZeroInitialized() const { return ResultIsZeroInitialized; }
SVEIntrinsicInfo &setResultIsZeroInitialized() {
ResultIsZeroInitialized = true;
return *this;
}
//
// The first operand of unary merging operations is typically only used to
// set the result for inactive lanes. Knowing this allows us to deadcode the
// operand when we can prove there are no inactive lanes.
//
bool hasOperandWithNoActiveLanes() const {
return OperandIdxWithNoActiveLanes != std::numeric_limits<unsigned>::max();
}
unsigned getOperandIdxWithNoActiveLanes() const {
assert(hasOperandWithNoActiveLanes() && "Propery not set!");
return OperandIdxWithNoActiveLanes;
}
SVEIntrinsicInfo &setOperandIdxWithNoActiveLanes(unsigned Index) {
assert(!hasOperandWithNoActiveLanes() && "Cannot set property twice!");
OperandIdxWithNoActiveLanes = Index;
return *this;
}
private:
unsigned GoverningPredicateIdx = std::numeric_limits<unsigned>::max();
Intrinsic::ID UndefIntrinsic = Intrinsic::not_intrinsic;
unsigned IROpcode = 0;
enum PredicationStyle {
Uninitialized,
InactiveLanesTakenFromOperand,
InactiveLanesAreNotDefined,
InactiveLanesAreUnused
} ResultLanes = Uninitialized;
bool ResultIsZeroInitialized = false;
unsigned OperandIdxForInactiveLanes = std::numeric_limits<unsigned>::max();
unsigned OperandIdxWithNoActiveLanes = std::numeric_limits<unsigned>::max();
};
static SVEIntrinsicInfo constructSVEIntrinsicInfo(IntrinsicInst &II) {
// Some SVE intrinsics do not use scalable vector types, but since they are
// not relevant from an SVEIntrinsicInfo perspective, they are also ignored.
if (!isa<ScalableVectorType>(II.getType()) &&
all_of(II.args(), [&](const Value *V) {
return !isa<ScalableVectorType>(V->getType());
}))
return SVEIntrinsicInfo();
Intrinsic::ID IID = II.getIntrinsicID();
switch (IID) {
default:
break;
case Intrinsic::aarch64_sve_fcvt_bf16f32_v2:
case Intrinsic::aarch64_sve_fcvt_f16f32:
case Intrinsic::aarch64_sve_fcvt_f16f64:
case Intrinsic::aarch64_sve_fcvt_f32f16:
case Intrinsic::aarch64_sve_fcvt_f32f64:
case Intrinsic::aarch64_sve_fcvt_f64f16:
case Intrinsic::aarch64_sve_fcvt_f64f32:
case Intrinsic::aarch64_sve_fcvtlt_f32f16:
case Intrinsic::aarch64_sve_fcvtlt_f64f32:
case Intrinsic::aarch64_sve_fcvtx_f32f64:
case Intrinsic::aarch64_sve_fcvtzs:
case Intrinsic::aarch64_sve_fcvtzs_i32f16:
case Intrinsic::aarch64_sve_fcvtzs_i32f64:
case Intrinsic::aarch64_sve_fcvtzs_i64f16:
case Intrinsic::aarch64_sve_fcvtzs_i64f32:
case Intrinsic::aarch64_sve_fcvtzu:
case Intrinsic::aarch64_sve_fcvtzu_i32f16:
case Intrinsic::aarch64_sve_fcvtzu_i32f64:
case Intrinsic::aarch64_sve_fcvtzu_i64f16:
case Intrinsic::aarch64_sve_fcvtzu_i64f32:
case Intrinsic::aarch64_sve_scvtf:
case Intrinsic::aarch64_sve_scvtf_f16i32:
case Intrinsic::aarch64_sve_scvtf_f16i64:
case Intrinsic::aarch64_sve_scvtf_f32i64:
case Intrinsic::aarch64_sve_scvtf_f64i32:
case Intrinsic::aarch64_sve_ucvtf:
case Intrinsic::aarch64_sve_ucvtf_f16i32:
case Intrinsic::aarch64_sve_ucvtf_f16i64:
case Intrinsic::aarch64_sve_ucvtf_f32i64:
case Intrinsic::aarch64_sve_ucvtf_f64i32:
return SVEIntrinsicInfo::defaultMergingUnaryOp();
case Intrinsic::aarch64_sve_fcvtnt_bf16f32_v2:
case Intrinsic::aarch64_sve_fcvtnt_f16f32:
case Intrinsic::aarch64_sve_fcvtnt_f32f64:
case Intrinsic::aarch64_sve_fcvtxnt_f32f64:
return SVEIntrinsicInfo::defaultMergingUnaryNarrowingTopOp();
case Intrinsic::aarch64_sve_fabd:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fabd_u);
case Intrinsic::aarch64_sve_fadd:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fadd_u)
.setMatchingIROpcode(Instruction::FAdd);
case Intrinsic::aarch64_sve_fdiv:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fdiv_u)
.setMatchingIROpcode(Instruction::FDiv);
case Intrinsic::aarch64_sve_fmax:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmax_u);
case Intrinsic::aarch64_sve_fmaxnm:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmaxnm_u);
case Intrinsic::aarch64_sve_fmin:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmin_u);
case Intrinsic::aarch64_sve_fminnm:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fminnm_u);
case Intrinsic::aarch64_sve_fmla:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmla_u);
case Intrinsic::aarch64_sve_fmls:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmls_u);
case Intrinsic::aarch64_sve_fmul:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmul_u)
.setMatchingIROpcode(Instruction::FMul);
case Intrinsic::aarch64_sve_fmulx:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmulx_u);
case Intrinsic::aarch64_sve_fnmla:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fnmla_u);
case Intrinsic::aarch64_sve_fnmls:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fnmls_u);
case Intrinsic::aarch64_sve_fsub:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fsub_u)
.setMatchingIROpcode(Instruction::FSub);
case Intrinsic::aarch64_sve_add:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_add_u)
.setMatchingIROpcode(Instruction::Add);
case Intrinsic::aarch64_sve_mla:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_mla_u);
case Intrinsic::aarch64_sve_mls:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_mls_u);
case Intrinsic::aarch64_sve_mul:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_mul_u)
.setMatchingIROpcode(Instruction::Mul);
case Intrinsic::aarch64_sve_sabd:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sabd_u);
case Intrinsic::aarch64_sve_smax:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_smax_u);
case Intrinsic::aarch64_sve_smin:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_smin_u);
case Intrinsic::aarch64_sve_smulh:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_smulh_u);
case Intrinsic::aarch64_sve_sub:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sub_u)
.setMatchingIROpcode(Instruction::Sub);
case Intrinsic::aarch64_sve_uabd:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uabd_u);
case Intrinsic::aarch64_sve_umax:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_umax_u);
case Intrinsic::aarch64_sve_umin:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_umin_u);
case Intrinsic::aarch64_sve_umulh:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_umulh_u);
case Intrinsic::aarch64_sve_asr:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_asr_u)
.setMatchingIROpcode(Instruction::AShr);
case Intrinsic::aarch64_sve_lsl:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_lsl_u)
.setMatchingIROpcode(Instruction::Shl);
case Intrinsic::aarch64_sve_lsr:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_lsr_u)
.setMatchingIROpcode(Instruction::LShr);
case Intrinsic::aarch64_sve_and:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_and_u)
.setMatchingIROpcode(Instruction::And);
case Intrinsic::aarch64_sve_bic:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_bic_u);
case Intrinsic::aarch64_sve_eor:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_eor_u)
.setMatchingIROpcode(Instruction::Xor);
case Intrinsic::aarch64_sve_orr:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_orr_u)
.setMatchingIROpcode(Instruction::Or);
case Intrinsic::aarch64_sve_sqsub:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sqsub_u);
case Intrinsic::aarch64_sve_uqsub:
return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uqsub_u);
case Intrinsic::aarch64_sve_add_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::Add);
case Intrinsic::aarch64_sve_and_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::And);
case Intrinsic::aarch64_sve_asr_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::AShr);
case Intrinsic::aarch64_sve_eor_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::Xor);
case Intrinsic::aarch64_sve_fadd_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::FAdd);
case Intrinsic::aarch64_sve_fdiv_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::FDiv);
case Intrinsic::aarch64_sve_fmul_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::FMul);
case Intrinsic::aarch64_sve_fsub_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::FSub);
case Intrinsic::aarch64_sve_lsl_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::Shl);
case Intrinsic::aarch64_sve_lsr_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::LShr);
case Intrinsic::aarch64_sve_mul_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::Mul);
case Intrinsic::aarch64_sve_orr_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::Or);
case Intrinsic::aarch64_sve_sub_u:
return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
Instruction::Sub);
case Intrinsic::aarch64_sve_addqv:
case Intrinsic::aarch64_sve_and_z:
case Intrinsic::aarch64_sve_bic_z:
case Intrinsic::aarch64_sve_brka_z:
case Intrinsic::aarch64_sve_brkb_z:
case Intrinsic::aarch64_sve_brkn_z:
case Intrinsic::aarch64_sve_brkpa_z:
case Intrinsic::aarch64_sve_brkpb_z:
case Intrinsic::aarch64_sve_cntp:
case Intrinsic::aarch64_sve_compact:
case Intrinsic::aarch64_sve_eor_z:
case Intrinsic::aarch64_sve_eorv:
case Intrinsic::aarch64_sve_eorqv:
case Intrinsic::aarch64_sve_nand_z:
case Intrinsic::aarch64_sve_nor_z:
case Intrinsic::aarch64_sve_orn_z:
case Intrinsic::aarch64_sve_orr_z:
case Intrinsic::aarch64_sve_orv:
case Intrinsic::aarch64_sve_orqv:
case Intrinsic::aarch64_sve_pnext:
case Intrinsic::aarch64_sve_rdffr_z:
case Intrinsic::aarch64_sve_saddv:
case Intrinsic::aarch64_sve_uaddv:
case Intrinsic::aarch64_sve_umaxv:
case Intrinsic::aarch64_sve_umaxqv:
case Intrinsic::aarch64_sve_cmpeq:
case Intrinsic::aarch64_sve_cmpeq_wide:
case Intrinsic::aarch64_sve_cmpge:
case Intrinsic::aarch64_sve_cmpge_wide:
case Intrinsic::aarch64_sve_cmpgt:
case Intrinsic::aarch64_sve_cmpgt_wide:
case Intrinsic::aarch64_sve_cmphi:
case Intrinsic::aarch64_sve_cmphi_wide:
case Intrinsic::aarch64_sve_cmphs:
case Intrinsic::aarch64_sve_cmphs_wide:
case Intrinsic::aarch64_sve_cmple_wide:
case Intrinsic::aarch64_sve_cmplo_wide:
case Intrinsic::aarch64_sve_cmpls_wide:
case Intrinsic::aarch64_sve_cmplt_wide:
case Intrinsic::aarch64_sve_cmpne:
case Intrinsic::aarch64_sve_cmpne_wide:
case Intrinsic::aarch64_sve_facge:
case Intrinsic::aarch64_sve_facgt:
case Intrinsic::aarch64_sve_fcmpeq:
case Intrinsic::aarch64_sve_fcmpge:
case Intrinsic::aarch64_sve_fcmpgt:
case Intrinsic::aarch64_sve_fcmpne:
case Intrinsic::aarch64_sve_fcmpuo:
case Intrinsic::aarch64_sve_ld1:
case Intrinsic::aarch64_sve_ld1_gather:
case Intrinsic::aarch64_sve_ld1_gather_index:
case Intrinsic::aarch64_sve_ld1_gather_scalar_offset:
case Intrinsic::aarch64_sve_ld1_gather_sxtw:
case Intrinsic::aarch64_sve_ld1_gather_sxtw_index:
case Intrinsic::aarch64_sve_ld1_gather_uxtw:
case Intrinsic::aarch64_sve_ld1_gather_uxtw_index:
case Intrinsic::aarch64_sve_ld1q_gather_index:
case Intrinsic::aarch64_sve_ld1q_gather_scalar_offset:
case Intrinsic::aarch64_sve_ld1q_gather_vector_offset:
case Intrinsic::aarch64_sve_ld1ro:
case Intrinsic::aarch64_sve_ld1rq:
case Intrinsic::aarch64_sve_ld1udq:
case Intrinsic::aarch64_sve_ld1uwq:
case Intrinsic::aarch64_sve_ld2_sret:
case Intrinsic::aarch64_sve_ld2q_sret:
case Intrinsic::aarch64_sve_ld3_sret:
case Intrinsic::aarch64_sve_ld3q_sret:
case Intrinsic::aarch64_sve_ld4_sret:
case Intrinsic::aarch64_sve_ld4q_sret:
case Intrinsic::aarch64_sve_ldff1:
case Intrinsic::aarch64_sve_ldff1_gather:
case Intrinsic::aarch64_sve_ldff1_gather_index:
case Intrinsic::aarch64_sve_ldff1_gather_scalar_offset:
case Intrinsic::aarch64_sve_ldff1_gather_sxtw:
case Intrinsic::aarch64_sve_ldff1_gather_sxtw_index:
case Intrinsic::aarch64_sve_ldff1_gather_uxtw:
case Intrinsic::aarch64_sve_ldff1_gather_uxtw_index:
case Intrinsic::aarch64_sve_ldnf1:
case Intrinsic::aarch64_sve_ldnt1:
case Intrinsic::aarch64_sve_ldnt1_gather:
case Intrinsic::aarch64_sve_ldnt1_gather_index:
case Intrinsic::aarch64_sve_ldnt1_gather_scalar_offset:
case Intrinsic::aarch64_sve_ldnt1_gather_uxtw:
return SVEIntrinsicInfo::defaultZeroingOp();
case Intrinsic::aarch64_sve_prf:
case Intrinsic::aarch64_sve_prfb_gather_index:
case Intrinsic::aarch64_sve_prfb_gather_scalar_offset:
case Intrinsic::aarch64_sve_prfb_gather_sxtw_index:
case Intrinsic::aarch64_sve_prfb_gather_uxtw_index:
case Intrinsic::aarch64_sve_prfd_gather_index:
case Intrinsic::aarch64_sve_prfd_gather_scalar_offset:
case Intrinsic::aarch64_sve_prfd_gather_sxtw_index:
case Intrinsic::aarch64_sve_prfd_gather_uxtw_index:
case Intrinsic::aarch64_sve_prfh_gather_index:
case Intrinsic::aarch64_sve_prfh_gather_scalar_offset:
case Intrinsic::aarch64_sve_prfh_gather_sxtw_index:
case Intrinsic::aarch64_sve_prfh_gather_uxtw_index:
case Intrinsic::aarch64_sve_prfw_gather_index:
case Intrinsic::aarch64_sve_prfw_gather_scalar_offset:
case Intrinsic::aarch64_sve_prfw_gather_sxtw_index:
case Intrinsic::aarch64_sve_prfw_gather_uxtw_index:
return SVEIntrinsicInfo::defaultVoidOp(0);
case Intrinsic::aarch64_sve_st1_scatter:
case Intrinsic::aarch64_sve_st1_scatter_scalar_offset:
case Intrinsic::aarch64_sve_st1_scatter_sxtw:
case Intrinsic::aarch64_sve_st1_scatter_sxtw_index:
case Intrinsic::aarch64_sve_st1_scatter_uxtw:
case Intrinsic::aarch64_sve_st1_scatter_uxtw_index:
case Intrinsic::aarch64_sve_st1dq:
case Intrinsic::aarch64_sve_st1q_scatter_index:
case Intrinsic::aarch64_sve_st1q_scatter_scalar_offset:
case Intrinsic::aarch64_sve_st1q_scatter_vector_offset:
case Intrinsic::aarch64_sve_st1wq:
case Intrinsic::aarch64_sve_stnt1:
case Intrinsic::aarch64_sve_stnt1_scatter:
case Intrinsic::aarch64_sve_stnt1_scatter_index:
case Intrinsic::aarch64_sve_stnt1_scatter_scalar_offset:
case Intrinsic::aarch64_sve_stnt1_scatter_uxtw:
return SVEIntrinsicInfo::defaultVoidOp(1);
case Intrinsic::aarch64_sve_st2:
case Intrinsic::aarch64_sve_st2q:
return SVEIntrinsicInfo::defaultVoidOp(2);
case Intrinsic::aarch64_sve_st3:
case Intrinsic::aarch64_sve_st3q:
return SVEIntrinsicInfo::defaultVoidOp(3);
case Intrinsic::aarch64_sve_st4:
case Intrinsic::aarch64_sve_st4q:
return SVEIntrinsicInfo::defaultVoidOp(4);
}
return SVEIntrinsicInfo();
}
static bool isAllActivePredicate(Value *Pred) {
// Look through convert.from.svbool(convert.to.svbool(...) chain.
Value *UncastedPred;
if (match(Pred, m_Intrinsic<Intrinsic::aarch64_sve_convert_from_svbool>(
m_Intrinsic<Intrinsic::aarch64_sve_convert_to_svbool>(
m_Value(UncastedPred)))))
// If the predicate has the same or less lanes than the uncasted
// predicate then we know the casting has no effect.
if (cast<ScalableVectorType>(Pred->getType())->getMinNumElements() <=
cast<ScalableVectorType>(UncastedPred->getType())->getMinNumElements())
Pred = UncastedPred;
auto *C = dyn_cast<Constant>(Pred);
return (C && C->isAllOnesValue());
}
// Simplify `V` by only considering the operations that affect active lanes.
// This function should only return existing Values or newly created Constants.
static Value *stripInactiveLanes(Value *V, const Value *Pg) {
auto *Dup = dyn_cast<IntrinsicInst>(V);
if (Dup && Dup->getIntrinsicID() == Intrinsic::aarch64_sve_dup &&
Dup->getOperand(1) == Pg && isa<Constant>(Dup->getOperand(2)))
return ConstantVector::getSplat(
cast<VectorType>(V->getType())->getElementCount(),
cast<Constant>(Dup->getOperand(2)));
return V;
}
static std::optional<Instruction *>
simplifySVEIntrinsicBinOp(InstCombiner &IC, IntrinsicInst &II,
const SVEIntrinsicInfo &IInfo) {
const unsigned Opc = IInfo.getMatchingIROpode();
assert(Instruction::isBinaryOp(Opc) && "Expected a binary operation!");
Value *Pg = II.getOperand(0);
Value *Op1 = II.getOperand(1);
Value *Op2 = II.getOperand(2);
const DataLayout &DL = II.getDataLayout();
// Canonicalise constants to the RHS.
if (Instruction::isCommutative(Opc) && IInfo.inactiveLanesAreNotDefined() &&
isa<Constant>(Op1) && !isa<Constant>(Op2)) {
IC.replaceOperand(II, 1, Op2);
IC.replaceOperand(II, 2, Op1);
return &II;
}
// Only active lanes matter when simplifying the operation.
Op1 = stripInactiveLanes(Op1, Pg);
Op2 = stripInactiveLanes(Op2, Pg);
Value *SimpleII;
if (auto FII = dyn_cast<FPMathOperator>(&II))
SimpleII = simplifyBinOp(Opc, Op1, Op2, FII->getFastMathFlags(), DL);
else
SimpleII = simplifyBinOp(Opc, Op1, Op2, DL);
// An SVE intrinsic's result is always defined. However, this is not the case
// for its equivalent IR instruction (e.g. when shifting by an amount more
// than the data's bitwidth). Simplifications to an undefined result must be
// ignored to preserve the intrinsic's expected behaviour.
if (!SimpleII || isa<UndefValue>(SimpleII))
return std::nullopt;
if (IInfo.inactiveLanesAreNotDefined())
return IC.replaceInstUsesWith(II, SimpleII);
Value *Inactive = II.getOperand(IInfo.getOperandIdxInactiveLanesTakenFrom());
// The intrinsic does nothing (e.g. sve.mul(pg, A, 1.0)).
if (SimpleII == Inactive)
return IC.replaceInstUsesWith(II, SimpleII);
// Inactive lanes must be preserved.
SimpleII = IC.Builder.CreateSelect(Pg, SimpleII, Inactive);
return IC.replaceInstUsesWith(II, SimpleII);
}
// Use SVE intrinsic info to eliminate redundant operands and/or canonicalise
// to operations with less strict inactive lane requirements.
static std::optional<Instruction *>
simplifySVEIntrinsic(InstCombiner &IC, IntrinsicInst &II,
const SVEIntrinsicInfo &IInfo) {
if (!IInfo.hasGoverningPredicate())
return std::nullopt;
auto *OpPredicate = II.getOperand(IInfo.getGoverningPredicateOperandIdx());
// If there are no active lanes.
if (match(OpPredicate, m_ZeroInt())) {
if (IInfo.inactiveLanesTakenFromOperand())
return IC.replaceInstUsesWith(
II, II.getOperand(IInfo.getOperandIdxInactiveLanesTakenFrom()));
if (IInfo.inactiveLanesAreUnused()) {
if (IInfo.resultIsZeroInitialized())
IC.replaceInstUsesWith(II, Constant::getNullValue(II.getType()));
return IC.eraseInstFromFunction(II);
}
}
// If there are no inactive lanes.
if (isAllActivePredicate(OpPredicate)) {
if (IInfo.hasOperandWithNoActiveLanes()) {
unsigned OpIdx = IInfo.getOperandIdxWithNoActiveLanes();
if (!isa<UndefValue>(II.getOperand(OpIdx)))
return IC.replaceOperand(II, OpIdx, UndefValue::get(II.getType()));
}
if (IInfo.hasMatchingUndefIntrinsic()) {
auto *NewDecl = Intrinsic::getOrInsertDeclaration(
II.getModule(), IInfo.getMatchingUndefIntrinsic(), {II.getType()});
II.setCalledFunction(NewDecl);
return &II;
}
}
// Operation specific simplifications.
if (IInfo.hasMatchingIROpode() &&
Instruction::isBinaryOp(IInfo.getMatchingIROpode()))
return simplifySVEIntrinsicBinOp(IC, II, IInfo);
return std::nullopt;
}
// (from_svbool (binop (to_svbool pred) (svbool_t _) (svbool_t _))))
// => (binop (pred) (from_svbool _) (from_svbool _))
//
// The above transformation eliminates a `to_svbool` in the predicate
// operand of bitwise operation `binop` by narrowing the vector width of
// the operation. For example, it would convert a `<vscale x 16 x i1>
// and` into a `<vscale x 4 x i1> and`. This is profitable because
// to_svbool must zero the new lanes during widening, whereas
// from_svbool is free.
static std::optional<Instruction *>
tryCombineFromSVBoolBinOp(InstCombiner &IC, IntrinsicInst &II) {
auto BinOp = dyn_cast<IntrinsicInst>(II.getOperand(0));
if (!BinOp)
return std::nullopt;
auto IntrinsicID = BinOp->getIntrinsicID();
switch (IntrinsicID) {
case Intrinsic::aarch64_sve_and_z:
case Intrinsic::aarch64_sve_bic_z:
case Intrinsic::aarch64_sve_eor_z:
case Intrinsic::aarch64_sve_nand_z:
case Intrinsic::aarch64_sve_nor_z:
case Intrinsic::aarch64_sve_orn_z:
case Intrinsic::aarch64_sve_orr_z:
break;
default:
return std::nullopt;
}
auto BinOpPred = BinOp->getOperand(0);
auto BinOpOp1 = BinOp->getOperand(1);
auto BinOpOp2 = BinOp->getOperand(2);
auto PredIntr = dyn_cast<IntrinsicInst>(BinOpPred);
if (!PredIntr ||
PredIntr->getIntrinsicID() != Intrinsic::aarch64_sve_convert_to_svbool)
return std::nullopt;
auto PredOp = PredIntr->getOperand(0);
auto PredOpTy = cast<VectorType>(PredOp->getType());
if (PredOpTy != II.getType())
return std::nullopt;
SmallVector<Value *> NarrowedBinOpArgs = {PredOp};
auto NarrowBinOpOp1 = IC.Builder.CreateIntrinsic(
Intrinsic::aarch64_sve_convert_from_svbool, {PredOpTy}, {BinOpOp1});
NarrowedBinOpArgs.push_back(NarrowBinOpOp1);
if (BinOpOp1 == BinOpOp2)
NarrowedBinOpArgs.push_back(NarrowBinOpOp1);
else
NarrowedBinOpArgs.push_back(IC.Builder.CreateIntrinsic(
Intrinsic::aarch64_sve_convert_from_svbool, {PredOpTy}, {BinOpOp2}));
auto NarrowedBinOp =
IC.Builder.CreateIntrinsic(IntrinsicID, {PredOpTy}, NarrowedBinOpArgs);
return IC.replaceInstUsesWith(II, NarrowedBinOp);
}
static std::optional<Instruction *>
instCombineConvertFromSVBool(InstCombiner &IC, IntrinsicInst &II) {
// If the reinterpret instruction operand is a PHI Node
if (isa<PHINode>(II.getArgOperand(0)))
return processPhiNode(IC, II);
if (auto BinOpCombine = tryCombineFromSVBoolBinOp(IC, II))
return BinOpCombine;
// Ignore converts to/from svcount_t.
if (isa<TargetExtType>(II.getArgOperand(0)->getType()) ||
isa<TargetExtType>(II.getType()))
return std::nullopt;
SmallVector<Instruction *, 32> CandidatesForRemoval;
Value *Cursor = II.getOperand(0), *EarliestReplacement = nullptr;
const auto *IVTy = cast<VectorType>(II.getType());
// Walk the chain of conversions.
while (Cursor) {
// If the type of the cursor has fewer lanes than the final result, zeroing
// must take place, which breaks the equivalence chain.
const auto *CursorVTy = cast<VectorType>(Cursor->getType());
if (CursorVTy->getElementCount().getKnownMinValue() <
IVTy->getElementCount().getKnownMinValue())
break;
// If the cursor has the same type as I, it is a viable replacement.
if (Cursor->getType() == IVTy)
EarliestReplacement = Cursor;
auto *IntrinsicCursor = dyn_cast<IntrinsicInst>(Cursor);
// If this is not an SVE conversion intrinsic, this is the end of the chain.
if (!IntrinsicCursor || !(IntrinsicCursor->getIntrinsicID() ==
Intrinsic::aarch64_sve_convert_to_svbool ||
IntrinsicCursor->getIntrinsicID() ==
Intrinsic::aarch64_sve_convert_from_svbool))
break;
CandidatesForRemoval.insert(CandidatesForRemoval.begin(), IntrinsicCursor);
Cursor = IntrinsicCursor->getOperand(0);
}
// If no viable replacement in the conversion chain was found, there is
// nothing to do.
if (!EarliestReplacement)
return std::nullopt;
return IC.replaceInstUsesWith(II, EarliestReplacement);
}
static std::optional<Instruction *> instCombineSVESel(InstCombiner &IC,
IntrinsicInst &II) {
// svsel(ptrue, x, y) => x
auto *OpPredicate = II.getOperand(0);
if (isAllActivePredicate(OpPredicate))
return IC.replaceInstUsesWith(II, II.getOperand(1));
auto Select =
IC.Builder.CreateSelect(OpPredicate, II.getOperand(1), II.getOperand(2));
return IC.replaceInstUsesWith(II, Select);
}
static std::optional<Instruction *> instCombineSVEDup(InstCombiner &IC,
IntrinsicInst &II) {
IntrinsicInst *Pg = dyn_cast<IntrinsicInst>(II.getArgOperand(1));
if (!Pg)
return std::nullopt;
if (Pg->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)
return std::nullopt;
const auto PTruePattern =
cast<ConstantInt>(Pg->getOperand(0))->getZExtValue();
if (PTruePattern != AArch64SVEPredPattern::vl1)
return std::nullopt;
// The intrinsic is inserting into lane zero so use an insert instead.
auto *IdxTy = Type::getInt64Ty(II.getContext());
auto *Insert = InsertElementInst::Create(
II.getArgOperand(0), II.getArgOperand(2), ConstantInt::get(IdxTy, 0));
Insert->insertBefore(II.getIterator());
Insert->takeName(&II);
return IC.replaceInstUsesWith(II, Insert);
}
static std::optional<Instruction *> instCombineSVEDupX(InstCombiner &IC,
IntrinsicInst &II) {
// Replace DupX with a regular IR splat.
auto *RetTy = cast<ScalableVectorType>(II.getType());
Value *Splat = IC.Builder.CreateVectorSplat(RetTy->getElementCount(),
II.getArgOperand(0));
Splat->takeName(&II);
return IC.replaceInstUsesWith(II, Splat);
}
static std::optional<Instruction *> instCombineSVECmpNE(InstCombiner &IC,
IntrinsicInst &II) {
LLVMContext &Ctx = II.getContext();
if (!isAllActivePredicate(II.getArgOperand(0)))
return std::nullopt;
// Check that we have a compare of zero..
auto *SplatValue =
dyn_cast_or_null<ConstantInt>(getSplatValue(II.getArgOperand(2)));
if (!SplatValue || !SplatValue->isZero())
return std::nullopt;
// ..against a dupq
auto *DupQLane = dyn_cast<IntrinsicInst>(II.getArgOperand(1));
if (!DupQLane ||
DupQLane->getIntrinsicID() != Intrinsic::aarch64_sve_dupq_lane)
return std::nullopt;
// Where the dupq is a lane 0 replicate of a vector insert
auto *DupQLaneIdx = dyn_cast<ConstantInt>(DupQLane->getArgOperand(1));
if (!DupQLaneIdx || !DupQLaneIdx->isZero())
return std::nullopt;
auto *VecIns = dyn_cast<IntrinsicInst>(DupQLane->getArgOperand(0));
if (!VecIns || VecIns->getIntrinsicID() != Intrinsic::vector_insert)
return std::nullopt;
// Where the vector insert is a fixed constant vector insert into undef at
// index zero
if (!isa<UndefValue>(VecIns->getArgOperand(0)))
return std::nullopt;
if (!cast<ConstantInt>(VecIns->getArgOperand(2))->isZero())
return std::nullopt;
auto *ConstVec = dyn_cast<Constant>(VecIns->getArgOperand(1));
if (!ConstVec)
return std::nullopt;
auto *VecTy = dyn_cast<FixedVectorType>(ConstVec->getType());
auto *OutTy = dyn_cast<ScalableVectorType>(II.getType());
if (!VecTy || !OutTy || VecTy->getNumElements() != OutTy->getMinNumElements())
return std::nullopt;
unsigned NumElts = VecTy->getNumElements();
unsigned PredicateBits = 0;
// Expand intrinsic operands to a 16-bit byte level predicate
for (unsigned I = 0; I < NumElts; ++I) {
auto *Arg = dyn_cast<ConstantInt>(ConstVec->getAggregateElement(I));
if (!Arg)
return std::nullopt;
if (!Arg->isZero())
PredicateBits |= 1 << (I * (16 / NumElts));
}
// If all bits are zero bail early with an empty predicate
if (PredicateBits == 0) {
auto *PFalse = Constant::getNullValue(II.getType());
PFalse->takeName(&II);
return IC.replaceInstUsesWith(II, PFalse);
}
// Calculate largest predicate type used (where byte predicate is largest)
unsigned Mask = 8;
for (unsigned I = 0; I < 16; ++I)
if ((PredicateBits & (1 << I)) != 0)
Mask |= (I % 8);
unsigned PredSize = Mask & -Mask;
auto *PredType = ScalableVectorType::get(
Type::getInt1Ty(Ctx), AArch64::SVEBitsPerBlock / (PredSize * 8));
// Ensure all relevant bits are set
for (unsigned I = 0; I < 16; I += PredSize)
if ((PredicateBits & (1 << I)) == 0)
return std::nullopt;
auto *PTruePat =
ConstantInt::get(Type::getInt32Ty(Ctx), AArch64SVEPredPattern::all);
auto *PTrue = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_ptrue,
{PredType}, {PTruePat});
auto *ConvertToSVBool = IC.Builder.CreateIntrinsic(
Intrinsic::aarch64_sve_convert_to_svbool, {PredType}, {PTrue});
auto *ConvertFromSVBool =
IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_convert_from_svbool,
{II.getType()}, {ConvertToSVBool});
ConvertFromSVBool->takeName(&II);
return IC.replaceInstUsesWith(II, ConvertFromSVBool);
}
static std::optional<Instruction *> instCombineSVELast(InstCombiner &IC,
IntrinsicInst &II) {
Value *Pg = II.getArgOperand(0);
Value *Vec = II.getArgOperand(1);
auto IntrinsicID = II.getIntrinsicID();
bool IsAfter = IntrinsicID == Intrinsic::aarch64_sve_lasta;
// lastX(splat(X)) --> X
if (auto *SplatVal = getSplatValue(Vec))
return IC.replaceInstUsesWith(II, SplatVal);
// If x and/or y is a splat value then:
// lastX (binop (x, y)) --> binop(lastX(x), lastX(y))
Value *LHS, *RHS;
if (match(Vec, m_OneUse(m_BinOp(m_Value(LHS), m_Value(RHS))))) {
if (isSplatValue(LHS) || isSplatValue(RHS)) {
auto *OldBinOp = cast<BinaryOperator>(Vec);
auto OpC = OldBinOp->getOpcode();
auto *NewLHS =
IC.Builder.CreateIntrinsic(IntrinsicID, {Vec->getType()}, {Pg, LHS});
auto *NewRHS =
IC.Builder.CreateIntrinsic(IntrinsicID, {Vec->getType()}, {Pg, RHS});
auto *NewBinOp = BinaryOperator::CreateWithCopiedFlags(
OpC, NewLHS, NewRHS, OldBinOp, OldBinOp->getName(), II.getIterator());
return IC.replaceInstUsesWith(II, NewBinOp);
}
}
auto *C = dyn_cast<Constant>(Pg);
if (IsAfter && C && C->isNullValue()) {
// The intrinsic is extracting lane 0 so use an extract instead.
auto *IdxTy = Type::getInt64Ty(II.getContext());
auto *Extract = ExtractElementInst::Create(Vec, ConstantInt::get(IdxTy, 0));
Extract->insertBefore(II.getIterator());
Extract->takeName(&II);
return IC.replaceInstUsesWith(II, Extract);
}
auto *IntrPG = dyn_cast<IntrinsicInst>(Pg);
if (!IntrPG)
return std::nullopt;
if (IntrPG->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)
return std::nullopt;
const auto PTruePattern =
cast<ConstantInt>(IntrPG->getOperand(0))->getZExtValue();
// Can the intrinsic's predicate be converted to a known constant index?
unsigned MinNumElts = getNumElementsFromSVEPredPattern(PTruePattern);
if (!MinNumElts)
return std::nullopt;
unsigned Idx = MinNumElts - 1;
// Increment the index if extracting the element after the last active
// predicate element.
if (IsAfter)
++Idx;
// Ignore extracts whose index is larger than the known minimum vector
// length. NOTE: This is an artificial constraint where we prefer to
// maintain what the user asked for until an alternative is proven faster.
auto *PgVTy = cast<ScalableVectorType>(Pg->getType());
if (Idx >= PgVTy->getMinNumElements())
return std::nullopt;
// The intrinsic is extracting a fixed lane so use an extract instead.
auto *IdxTy = Type::getInt64Ty(II.getContext());
auto *Extract = ExtractElementInst::Create(Vec, ConstantInt::get(IdxTy, Idx));
Extract->insertBefore(II.getIterator());
Extract->takeName(&II);
return IC.replaceInstUsesWith(II, Extract);
}
static std::optional<Instruction *> instCombineSVECondLast(InstCombiner &IC,
IntrinsicInst &II) {
// The SIMD&FP variant of CLAST[AB] is significantly faster than the scalar
// integer variant across a variety of micro-architectures. Replace scalar
// integer CLAST[AB] intrinsic with optimal SIMD&FP variant. A simple
// bitcast-to-fp + clast[ab] + bitcast-to-int will cost a cycle or two more
// depending on the micro-architecture, but has been observed as generally
// being faster, particularly when the CLAST[AB] op is a loop-carried
// dependency.
Value *Pg = II.getArgOperand(0);
Value *Fallback = II.getArgOperand(1);
Value *Vec = II.getArgOperand(2);
Type *Ty = II.getType();
if (!Ty->isIntegerTy())
return std::nullopt;
Type *FPTy;
switch (cast<IntegerType>(Ty)->getBitWidth()) {
default:
return std::nullopt;
case 16:
FPTy = IC.Builder.getHalfTy();
break;
case 32:
FPTy = IC.Builder.getFloatTy();
break;
case 64:
FPTy = IC.Builder.getDoubleTy();
break;
}
Value *FPFallBack = IC.Builder.CreateBitCast(Fallback, FPTy);
auto *FPVTy = VectorType::get(
FPTy, cast<VectorType>(Vec->getType())->getElementCount());
Value *FPVec = IC.Builder.CreateBitCast(Vec, FPVTy);
auto *FPII = IC.Builder.CreateIntrinsic(
II.getIntrinsicID(), {FPVec->getType()}, {Pg, FPFallBack, FPVec});
Value *FPIItoInt = IC.Builder.CreateBitCast(FPII, II.getType());
return IC.replaceInstUsesWith(II, FPIItoInt);
}
static std::optional<Instruction *> instCombineRDFFR(InstCombiner &IC,
IntrinsicInst &II) {
LLVMContext &Ctx = II.getContext();
// Replace rdffr with predicated rdffr.z intrinsic, so that optimizePTestInstr
// can work with RDFFR_PP for ptest elimination.
auto *AllPat =
ConstantInt::get(Type::getInt32Ty(Ctx), AArch64SVEPredPattern::all);
auto *PTrue = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_ptrue,
{II.getType()}, {AllPat});
auto *RDFFR =
IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_rdffr_z, {PTrue});
RDFFR->takeName(&II);
return IC.replaceInstUsesWith(II, RDFFR);
}
static std::optional<Instruction *>
instCombineSVECntElts(InstCombiner &IC, IntrinsicInst &II, unsigned NumElts) {
const auto Pattern = cast<ConstantInt>(II.getArgOperand(0))->getZExtValue();
if (Pattern == AArch64SVEPredPattern::all) {
Constant *StepVal = ConstantInt::get(II.getType(), NumElts);
auto *VScale = IC.Builder.CreateVScale(StepVal);
VScale->takeName(&II);
return IC.replaceInstUsesWith(II, VScale);
}
unsigned MinNumElts = getNumElementsFromSVEPredPattern(Pattern);
return MinNumElts && NumElts >= MinNumElts
? std::optional<Instruction *>(IC.replaceInstUsesWith(
II, ConstantInt::get(II.getType(), MinNumElts)))
: std::nullopt;
}
static std::optional<Instruction *> instCombineSVEPTest(InstCombiner &IC,
IntrinsicInst &II) {
Value *PgVal = II.getArgOperand(0);
Value *OpVal = II.getArgOperand(1);
// PTEST_<FIRST|LAST>(X, X) is equivalent to PTEST_ANY(X, X).
// Later optimizations prefer this form.
if (PgVal == OpVal &&
(II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_first ||
II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_last)) {
Value *Ops[] = {PgVal, OpVal};
Type *Tys[] = {PgVal->getType()};
auto *PTest =
IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_ptest_any, Tys, Ops);
PTest->takeName(&II);
return IC.replaceInstUsesWith(II, PTest);
}
IntrinsicInst *Pg = dyn_cast<IntrinsicInst>(PgVal);
IntrinsicInst *Op = dyn_cast<IntrinsicInst>(OpVal);
if (!Pg || !Op)
return std::nullopt;
Intrinsic::ID OpIID = Op->getIntrinsicID();
if (Pg->getIntrinsicID() == Intrinsic::aarch64_sve_convert_to_svbool &&
OpIID == Intrinsic::aarch64_sve_convert_to_svbool &&
Pg->getArgOperand(0)->getType() == Op->getArgOperand(0)->getType()) {
Value *Ops[] = {Pg->getArgOperand(0), Op->getArgOperand(0)};
Type *Tys[] = {Pg->getArgOperand(0)->getType()};
auto *PTest = IC.Builder.CreateIntrinsic(II.getIntrinsicID(), Tys, Ops);
PTest->takeName(&II);
return IC.replaceInstUsesWith(II, PTest);
}
// Transform PTEST_ANY(X=OP(PG,...), X) -> PTEST_ANY(PG, X)).
// Later optimizations may rewrite sequence to use the flag-setting variant
// of instruction X to remove PTEST.
if ((Pg == Op) && (II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_any) &&
((OpIID == Intrinsic::aarch64_sve_brka_z) ||
(OpIID == Intrinsic::aarch64_sve_brkb_z) ||
(OpIID == Intrinsic::aarch64_sve_brkpa_z) ||
(OpIID == Intrinsic::aarch64_sve_brkpb_z) ||
(OpIID == Intrinsic::aarch64_sve_rdffr_z) ||
(OpIID == Intrinsic::aarch64_sve_and_z) ||
(OpIID == Intrinsic::aarch64_sve_bic_z) ||
(OpIID == Intrinsic::aarch64_sve_eor_z) ||
(OpIID == Intrinsic::aarch64_sve_nand_z) ||
(OpIID == Intrinsic::aarch64_sve_nor_z) ||
(OpIID == Intrinsic::aarch64_sve_orn_z) ||
(OpIID == Intrinsic::aarch64_sve_orr_z))) {
Value *Ops[] = {Pg->getArgOperand(0), Pg};
Type *Tys[] = {Pg->getType()};
auto *PTest = IC.Builder.CreateIntrinsic(II.getIntrinsicID(), Tys, Ops);
PTest->takeName(&II);
return IC.replaceInstUsesWith(II, PTest);
}
return std::nullopt;
}
template <Intrinsic::ID MulOpc, typename Intrinsic::ID FuseOpc>
static std::optional<Instruction *>
instCombineSVEVectorFuseMulAddSub(InstCombiner &IC, IntrinsicInst &II,
bool MergeIntoAddendOp) {
Value *P = II.getOperand(0);
Value *MulOp0, *MulOp1, *AddendOp, *Mul;
if (MergeIntoAddendOp) {
AddendOp = II.getOperand(1);
Mul = II.getOperand(2);
} else {
AddendOp = II.getOperand(2);
Mul = II.getOperand(1);
}
if (!match(Mul, m_Intrinsic<MulOpc>(m_Specific(P), m_Value(MulOp0),
m_Value(MulOp1))))
return std::nullopt;
if (!Mul->hasOneUse())
return std::nullopt;
Instruction *FMFSource = nullptr;
if (II.getType()->isFPOrFPVectorTy()) {
llvm::FastMathFlags FAddFlags = II.getFastMathFlags();
// Stop the combine when the flags on the inputs differ in case dropping
// flags would lead to us missing out on more beneficial optimizations.
if (FAddFlags != cast<CallInst>(Mul)->getFastMathFlags())
return std::nullopt;
if (!FAddFlags.allowContract())
return std::nullopt;
FMFSource = &II;
}
CallInst *Res;
if (MergeIntoAddendOp)
Res = IC.Builder.CreateIntrinsic(FuseOpc, {II.getType()},
{P, AddendOp, MulOp0, MulOp1}, FMFSource);
else
Res = IC.Builder.CreateIntrinsic(FuseOpc, {II.getType()},
{P, MulOp0, MulOp1, AddendOp}, FMFSource);
return IC.replaceInstUsesWith(II, Res);
}
static std::optional<Instruction *>
instCombineSVELD1(InstCombiner &IC, IntrinsicInst &II, const DataLayout &DL) {
Value *Pred = II.getOperand(0);
Value *PtrOp = II.getOperand(1);
Type *VecTy = II.getType();
if (isAllActivePredicate(Pred)) {
LoadInst *Load = IC.Builder.CreateLoad(VecTy, PtrOp);
Load->copyMetadata(II);
return IC.replaceInstUsesWith(II, Load);
}
CallInst *MaskedLoad =
IC.Builder.CreateMaskedLoad(VecTy, PtrOp, PtrOp->getPointerAlignment(DL),
Pred, ConstantAggregateZero::get(VecTy));
MaskedLoad->copyMetadata(II);
return IC.replaceInstUsesWith(II, MaskedLoad);
}
static std::optional<Instruction *>
instCombineSVEST1(InstCombiner &IC, IntrinsicInst &II, const DataLayout &DL) {
Value *VecOp = II.getOperand(0);
Value *Pred = II.getOperand(1);
Value *PtrOp = II.getOperand(2);
if (isAllActivePredicate(Pred)) {
StoreInst *Store = IC.Builder.CreateStore(VecOp, PtrOp);
Store->copyMetadata(II);
return IC.eraseInstFromFunction(II);
}
CallInst *MaskedStore = IC.Builder.CreateMaskedStore(
VecOp, PtrOp, PtrOp->getPointerAlignment(DL), Pred);
MaskedStore->copyMetadata(II);
return IC.eraseInstFromFunction(II);
}
static Instruction::BinaryOps intrinsicIDToBinOpCode(unsigned Intrinsic) {
switch (Intrinsic) {
case Intrinsic::aarch64_sve_fmul_u:
return Instruction::BinaryOps::FMul;
case Intrinsic::aarch64_sve_fadd_u:
return Instruction::BinaryOps::FAdd;
case Intrinsic::aarch64_sve_fsub_u:
return Instruction::BinaryOps::FSub;
default:
return Instruction::BinaryOpsEnd;
}
}
static std::optional<Instruction *>
instCombineSVEVectorBinOp(InstCombiner &IC, IntrinsicInst &II) {
// Bail due to missing support for ISD::STRICT_ scalable vector operations.
if (II.isStrictFP())
return std::nullopt;
auto *OpPredicate = II.getOperand(0);
auto BinOpCode = intrinsicIDToBinOpCode(II.getIntrinsicID());
if (BinOpCode == Instruction::BinaryOpsEnd ||
!isAllActivePredicate(OpPredicate))
return std::nullopt;
auto BinOp = IC.Builder.CreateBinOpFMF(
BinOpCode, II.getOperand(1), II.getOperand(2), II.getFastMathFlags());
return IC.replaceInstUsesWith(II, BinOp);
}
static std::optional<Instruction *> instCombineSVEVectorAdd(InstCombiner &IC,
IntrinsicInst &II) {
if (auto MLA = instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul,
Intrinsic::aarch64_sve_mla>(
IC, II, true))
return MLA;
if (auto MAD = instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul,
Intrinsic::aarch64_sve_mad>(
IC, II, false))
return MAD;
return std::nullopt;
}
static std::optional<Instruction *>
instCombineSVEVectorFAdd(InstCombiner &IC, IntrinsicInst &II) {
if (auto FMLA =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fmla>(IC, II,
true))
return FMLA;
if (auto FMAD =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fmad>(IC, II,
false))
return FMAD;
if (auto FMLA =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
Intrinsic::aarch64_sve_fmla>(IC, II,
true))
return FMLA;
return std::nullopt;
}
static std::optional<Instruction *>
instCombineSVEVectorFAddU(InstCombiner &IC, IntrinsicInst &II) {
if (auto FMLA =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fmla>(IC, II,
true))
return FMLA;
if (auto FMAD =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fmad>(IC, II,
false))
return FMAD;
if (auto FMLA_U =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
Intrinsic::aarch64_sve_fmla_u>(
IC, II, true))
return FMLA_U;
return instCombineSVEVectorBinOp(IC, II);
}
static std::optional<Instruction *>
instCombineSVEVectorFSub(InstCombiner &IC, IntrinsicInst &II) {
if (auto FMLS =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fmls>(IC, II,
true))
return FMLS;
if (auto FMSB =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fnmsb>(
IC, II, false))
return FMSB;
if (auto FMLS =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
Intrinsic::aarch64_sve_fmls>(IC, II,
true))
return FMLS;
return std::nullopt;
}
static std::optional<Instruction *>
instCombineSVEVectorFSubU(InstCombiner &IC, IntrinsicInst &II) {
if (auto FMLS =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fmls>(IC, II,
true))
return FMLS;
if (auto FMSB =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
Intrinsic::aarch64_sve_fnmsb>(
IC, II, false))
return FMSB;
if (auto FMLS_U =
instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
Intrinsic::aarch64_sve_fmls_u>(
IC, II, true))
return FMLS_U;
return instCombineSVEVectorBinOp(IC, II);
}
static std::optional<Instruction *> instCombineSVEVectorSub(InstCombiner &IC,
IntrinsicInst &II) {
if (auto MLS = instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul,
Intrinsic::aarch64_sve_mls>(
IC, II, true))
return MLS;
return std::nullopt;
}
static std::optional<Instruction *> instCombineSVEUnpack(InstCombiner &IC,
IntrinsicInst &II) {
Value *UnpackArg = II.getArgOperand(0);
auto *RetTy = cast<ScalableVectorType>(II.getType());
bool IsSigned = II.getIntrinsicID() == Intrinsic::aarch64_sve_sunpkhi ||
II.getIntrinsicID() == Intrinsic::aarch64_sve_sunpklo;
// Hi = uunpkhi(splat(X)) --> Hi = splat(extend(X))
// Lo = uunpklo(splat(X)) --> Lo = splat(extend(X))
if (auto *ScalarArg = getSplatValue(UnpackArg)) {
ScalarArg =
IC.Builder.CreateIntCast(ScalarArg, RetTy->getScalarType(), IsSigned);
Value *NewVal =
IC.Builder.CreateVectorSplat(RetTy->getElementCount(), ScalarArg);
NewVal->takeName(&II);
return IC.replaceInstUsesWith(II, NewVal);
}
return std::nullopt;
}
static std::optional<Instruction *> instCombineSVETBL(InstCombiner &IC,
IntrinsicInst &II) {
auto *OpVal = II.getOperand(0);
auto *OpIndices = II.getOperand(1);
VectorType *VTy = cast<VectorType>(II.getType());
// Check whether OpIndices is a constant splat value < minimal element count
// of result.
auto *SplatValue = dyn_cast_or_null<ConstantInt>(getSplatValue(OpIndices));
if (!SplatValue ||
SplatValue->getValue().uge(VTy->getElementCount().getKnownMinValue()))
return std::nullopt;
// Convert sve_tbl(OpVal sve_dup_x(SplatValue)) to
// splat_vector(extractelement(OpVal, SplatValue)) for further optimization.
auto *Extract = IC.Builder.CreateExtractElement(OpVal, SplatValue);
auto *VectorSplat =
IC.Builder.CreateVectorSplat(VTy->getElementCount(), Extract);
VectorSplat->takeName(&II);
return IC.replaceInstUsesWith(II, VectorSplat);
}
static std::optional<Instruction *> instCombineSVEUzp1(InstCombiner &IC,
IntrinsicInst &II) {
Value *A, *B;
Type *RetTy = II.getType();
constexpr Intrinsic::ID FromSVB = Intrinsic::aarch64_sve_convert_from_svbool;
constexpr Intrinsic::ID ToSVB = Intrinsic::aarch64_sve_convert_to_svbool;
// uzp1(to_svbool(A), to_svbool(B)) --> <A, B>
// uzp1(from_svbool(to_svbool(A)), from_svbool(to_svbool(B))) --> <A, B>
if ((match(II.getArgOperand(0),
m_Intrinsic<FromSVB>(m_Intrinsic<ToSVB>(m_Value(A)))) &&
match(II.getArgOperand(1),
m_Intrinsic<FromSVB>(m_Intrinsic<ToSVB>(m_Value(B))))) ||
(match(II.getArgOperand(0), m_Intrinsic<ToSVB>(m_Value(A))) &&
match(II.getArgOperand(1), m_Intrinsic<ToSVB>(m_Value(B))))) {
auto *TyA = cast<ScalableVectorType>(A->getType());
if (TyA == B->getType() &&
RetTy == ScalableVectorType::getDoubleElementsVectorType(TyA)) {
auto *SubVec = IC.Builder.CreateInsertVector(
RetTy, PoisonValue::get(RetTy), A, IC.Builder.getInt64(0));
auto *ConcatVec = IC.Builder.CreateInsertVector(
RetTy, SubVec, B, IC.Builder.getInt64(TyA->getMinNumElements()));
ConcatVec->takeName(&II);
return IC.replaceInstUsesWith(II, ConcatVec);
}
}
return std::nullopt;
}
static std::optional<Instruction *> instCombineSVEZip(InstCombiner &IC,
IntrinsicInst &II) {
// zip1(uzp1(A, B), uzp2(A, B)) --> A
// zip2(uzp1(A, B), uzp2(A, B)) --> B
Value *A, *B;
if (match(II.getArgOperand(0),
m_Intrinsic<Intrinsic::aarch64_sve_uzp1>(m_Value(A), m_Value(B))) &&
match(II.getArgOperand(1), m_Intrinsic<Intrinsic::aarch64_sve_uzp2>(
m_Specific(A), m_Specific(B))))
return IC.replaceInstUsesWith(
II, (II.getIntrinsicID() == Intrinsic::aarch64_sve_zip1 ? A : B));
return std::nullopt;
}
static std::optional<Instruction *>
instCombineLD1GatherIndex(InstCombiner &IC, IntrinsicInst &II) {
Value *Mask = II.getOperand(0);
Value *BasePtr = II.getOperand(1);
Value *Index = II.getOperand(2);
Type *Ty = II.getType();
Value *PassThru = ConstantAggregateZero::get(Ty);
// Contiguous gather => masked load.
// (sve.ld1.gather.index Mask BasePtr (sve.index IndexBase 1))
// => (masked.load (gep BasePtr IndexBase) Align Mask zeroinitializer)
Value *IndexBase;
if (match(Index, m_Intrinsic<Intrinsic::aarch64_sve_index>(
m_Value(IndexBase), m_SpecificInt(1)))) {
Align Alignment =
BasePtr->getPointerAlignment(II.getDataLayout());
Value *Ptr = IC.Builder.CreateGEP(cast<VectorType>(Ty)->getElementType(),
BasePtr, IndexBase);
CallInst *MaskedLoad =
IC.Builder.CreateMaskedLoad(Ty, Ptr, Alignment, Mask, PassThru);
MaskedLoad->takeName(&II);
return IC.replaceInstUsesWith(II, MaskedLoad);
}
return std::nullopt;
}
static std::optional<Instruction *>
instCombineST1ScatterIndex(InstCombiner &IC, IntrinsicInst &II) {
Value *Val = II.getOperand(0);
Value *Mask = II.getOperand(1);
Value *BasePtr = II.getOperand(2);
Value *Index = II.getOperand(3);
Type *Ty = Val->getType();
// Contiguous scatter => masked store.
// (sve.st1.scatter.index Value Mask BasePtr (sve.index IndexBase 1))
// => (masked.store Value (gep BasePtr IndexBase) Align Mask)
Value *IndexBase;
if (match(Index, m_Intrinsic<Intrinsic::aarch64_sve_index>(
m_Value(IndexBase), m_SpecificInt(1)))) {
Align Alignment =
BasePtr->getPointerAlignment(II.getDataLayout());
Value *Ptr = IC.Builder.CreateGEP(cast<VectorType>(Ty)->getElementType(),
BasePtr, IndexBase);
(void)IC.Builder.CreateMaskedStore(Val, Ptr, Alignment, Mask);
return IC.eraseInstFromFunction(II);
}
return std::nullopt;
}
static std::optional<Instruction *> instCombineSVESDIV(InstCombiner &IC,
IntrinsicInst &II) {
Type *Int32Ty = IC.Builder.getInt32Ty();
Value *Pred = II.getOperand(0);
Value *Vec = II.getOperand(1);
Value *DivVec = II.getOperand(2);
Value *SplatValue = getSplatValue(DivVec);
ConstantInt *SplatConstantInt = dyn_cast_or_null<ConstantInt>(SplatValue);
if (!SplatConstantInt)
return std::nullopt;
APInt Divisor = SplatConstantInt->getValue();
const int64_t DivisorValue = Divisor.getSExtValue();
if (DivisorValue == -1)
return std::nullopt;
if (DivisorValue == 1)
IC.replaceInstUsesWith(II, Vec);
if (Divisor.isPowerOf2()) {
Constant *DivisorLog2 = ConstantInt::get(Int32Ty, Divisor.logBase2());
auto ASRD = IC.Builder.CreateIntrinsic(
Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, DivisorLog2});
return IC.replaceInstUsesWith(II, ASRD);
}
if (Divisor.isNegatedPowerOf2()) {
Divisor.negate();
Constant *DivisorLog2 = ConstantInt::get(Int32Ty, Divisor.logBase2());
auto ASRD = IC.Builder.CreateIntrinsic(
Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, DivisorLog2});
auto NEG = IC.Builder.CreateIntrinsic(
Intrinsic::aarch64_sve_neg, {ASRD->getType()}, {ASRD, Pred, ASRD});
return IC.replaceInstUsesWith(II, NEG);
}
return std::nullopt;
}
bool SimplifyValuePattern(SmallVector<Value *> &Vec, bool AllowPoison) {
size_t VecSize = Vec.size();
if (VecSize == 1)
return true;
if (!isPowerOf2_64(VecSize))
return false;
size_t HalfVecSize = VecSize / 2;
for (auto LHS = Vec.begin(), RHS = Vec.begin() + HalfVecSize;
RHS != Vec.end(); LHS++, RHS++) {
if (*LHS != nullptr && *RHS != nullptr) {
if (*LHS == *RHS)
continue;
else
return false;
}
if (!AllowPoison)
return false;
if (*LHS == nullptr && *RHS != nullptr)
*LHS = *RHS;
}
Vec.resize(HalfVecSize);
SimplifyValuePattern(Vec, AllowPoison);
return true;
}
// Try to simplify dupqlane patterns like dupqlane(f32 A, f32 B, f32 A, f32 B)
// to dupqlane(f64(C)) where C is A concatenated with B
static std::optional<Instruction *> instCombineSVEDupqLane(InstCombiner &IC,
IntrinsicInst &II) {
Value *CurrentInsertElt = nullptr, *Default = nullptr;
if (!match(II.getOperand(0),
m_Intrinsic<Intrinsic::vector_insert>(
m_Value(Default), m_Value(CurrentInsertElt), m_Value())) ||
!isa<FixedVectorType>(CurrentInsertElt->getType()))
return std::nullopt;
auto IIScalableTy = cast<ScalableVectorType>(II.getType());
// Insert the scalars into a container ordered by InsertElement index
SmallVector<Value *> Elts(IIScalableTy->getMinNumElements(), nullptr);
while (auto InsertElt = dyn_cast<InsertElementInst>(CurrentInsertElt)) {
auto Idx = cast<ConstantInt>(InsertElt->getOperand(2));
Elts[Idx->getValue().getZExtValue()] = InsertElt->getOperand(1);
CurrentInsertElt = InsertElt->getOperand(0);
}
bool AllowPoison =
isa<PoisonValue>(CurrentInsertElt) && isa<PoisonValue>(Default);
if (!SimplifyValuePattern(Elts, AllowPoison))
return std::nullopt;
// Rebuild the simplified chain of InsertElements. e.g. (a, b, a, b) as (a, b)
Value *InsertEltChain = PoisonValue::get(CurrentInsertElt->getType());
for (size_t I = 0; I < Elts.size(); I++) {
if (Elts[I] == nullptr)
continue;
InsertEltChain = IC.Builder.CreateInsertElement(InsertEltChain, Elts[I],
IC.Builder.getInt64(I));
}
if (InsertEltChain == nullptr)
return std::nullopt;
// Splat the simplified sequence, e.g. (f16 a, f16 b, f16 c, f16 d) as one i64
// value or (f16 a, f16 b) as one i32 value. This requires an InsertSubvector
// be bitcast to a type wide enough to fit the sequence, be splatted, and then
// be narrowed back to the original type.
unsigned PatternWidth = IIScalableTy->getScalarSizeInBits() * Elts.size();
unsigned PatternElementCount = IIScalableTy->getScalarSizeInBits() *
IIScalableTy->getMinNumElements() /
PatternWidth;
IntegerType *WideTy = IC.Builder.getIntNTy(PatternWidth);
auto *WideScalableTy = ScalableVectorType::get(WideTy, PatternElementCount);
auto *WideShuffleMaskTy =
ScalableVectorType::get(IC.Builder.getInt32Ty(), PatternElementCount);
auto ZeroIdx = ConstantInt::get(IC.Builder.getInt64Ty(), APInt(64, 0));
auto InsertSubvector = IC.Builder.CreateInsertVector(
II.getType(), PoisonValue::get(II.getType()), InsertEltChain, ZeroIdx);
auto WideBitcast =
IC.Builder.CreateBitOrPointerCast(InsertSubvector, WideScalableTy);
auto WideShuffleMask = ConstantAggregateZero::get(WideShuffleMaskTy);
auto WideShuffle = IC.Builder.CreateShuffleVector(
WideBitcast, PoisonValue::get(WideScalableTy), WideShuffleMask);
auto NarrowBitcast =
IC.Builder.CreateBitOrPointerCast(WideShuffle, II.getType());
return IC.replaceInstUsesWith(II, NarrowBitcast);
}
static std::optional<Instruction *> instCombineMaxMinNM(InstCombiner &IC,
IntrinsicInst &II) {
Value *A = II.getArgOperand(0);
Value *B = II.getArgOperand(1);
if (A == B)
return IC.replaceInstUsesWith(II, A);
return std::nullopt;
}
static std::optional<Instruction *> instCombineSVESrshl(InstCombiner &IC,
IntrinsicInst &II) {
Value *Pred = II.getOperand(0);
Value *Vec = II.getOperand(1);
Value *Shift = II.getOperand(2);
// Convert SRSHL into the simpler LSL intrinsic when fed by an ABS intrinsic.
Value *AbsPred, *MergedValue;
if (!match(Vec, m_Intrinsic<Intrinsic::aarch64_sve_sqabs>(
m_Value(MergedValue), m_Value(AbsPred), m_Value())) &&
!match(Vec, m_Intrinsic<Intrinsic::aarch64_sve_abs>(
m_Value(MergedValue), m_Value(AbsPred), m_Value())))
return std::nullopt;
// Transform is valid if any of the following are true:
// * The ABS merge value is an undef or non-negative
// * The ABS predicate is all active
// * The ABS predicate and the SRSHL predicates are the same
if (!isa<UndefValue>(MergedValue) && !match(MergedValue, m_NonNegative()) &&
AbsPred != Pred && !isAllActivePredicate(AbsPred))
return std::nullopt;
// Only valid when the shift amount is non-negative, otherwise the rounding
// behaviour of SRSHL cannot be ignored.
if (!match(Shift, m_NonNegative()))
return std::nullopt;
auto LSL = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_lsl,
{II.getType()}, {Pred, Vec, Shift});
return IC.replaceInstUsesWith(II, LSL);
}
static std::optional<Instruction *> instCombineSVEInsr(InstCombiner &IC,
IntrinsicInst &II) {
Value *Vec = II.getOperand(0);
if (getSplatValue(Vec) == II.getOperand(1))
return IC.replaceInstUsesWith(II, Vec);
return std::nullopt;
}
static std::optional<Instruction *> instCombineDMB(InstCombiner &IC,
IntrinsicInst &II) {
// If this barrier is post-dominated by identical one we can remove it
auto *NI = II.getNextNonDebugInstruction();
unsigned LookaheadThreshold = DMBLookaheadThreshold;
auto CanSkipOver = [](Instruction *I) {
return !I->mayReadOrWriteMemory() && !I->mayHaveSideEffects();
};
while (LookaheadThreshold-- && CanSkipOver(NI)) {
auto *NIBB = NI->getParent();
NI = NI->getNextNonDebugInstruction();
if (!NI) {
if (auto *SuccBB = NIBB->getUniqueSuccessor())
NI = &*SuccBB->getFirstNonPHIOrDbgOrLifetime();
else
break;
}
}
auto *NextII = dyn_cast_or_null<IntrinsicInst>(NI);
if (NextII && II.isIdenticalTo(NextII))
return IC.eraseInstFromFunction(II);
return std::nullopt;
}
static std::optional<Instruction *> instCombinePTrue(InstCombiner &IC,
IntrinsicInst &II) {
if (match(II.getOperand(0), m_ConstantInt<AArch64SVEPredPattern::all>()))
return IC.replaceInstUsesWith(II, Constant::getAllOnesValue(II.getType()));
return std::nullopt;
}
std::optional<Instruction *>
AArch64TTIImpl::instCombineIntrinsic(InstCombiner &IC,
IntrinsicInst &II) const {
const SVEIntrinsicInfo &IInfo = constructSVEIntrinsicInfo(II);
if (std::optional<Instruction *> I = simplifySVEIntrinsic(IC, II, IInfo))
return I;
Intrinsic::ID IID = II.getIntrinsicID();
switch (IID) {
default:
break;
case Intrinsic::aarch64_dmb:
return instCombineDMB(IC, II);
case Intrinsic::aarch64_neon_fmaxnm:
case Intrinsic::aarch64_neon_fminnm:
return instCombineMaxMinNM(IC, II);
case Intrinsic::aarch64_sve_convert_from_svbool:
return instCombineConvertFromSVBool(IC, II);
case Intrinsic::aarch64_sve_dup:
return instCombineSVEDup(IC, II);
case Intrinsic::aarch64_sve_dup_x:
return instCombineSVEDupX(IC, II);
case Intrinsic::aarch64_sve_cmpne:
case Intrinsic::aarch64_sve_cmpne_wide:
return instCombineSVECmpNE(IC, II);
case Intrinsic::aarch64_sve_rdffr:
return instCombineRDFFR(IC, II);
case Intrinsic::aarch64_sve_lasta:
case Intrinsic::aarch64_sve_lastb:
return instCombineSVELast(IC, II);
case Intrinsic::aarch64_sve_clasta_n:
case Intrinsic::aarch64_sve_clastb_n:
return instCombineSVECondLast(IC, II);
case Intrinsic::aarch64_sve_cntd:
return instCombineSVECntElts(IC, II, 2);
case Intrinsic::aarch64_sve_cntw:
return instCombineSVECntElts(IC, II, 4);
case Intrinsic::aarch64_sve_cnth:
return instCombineSVECntElts(IC, II, 8);
case Intrinsic::aarch64_sve_cntb:
return instCombineSVECntElts(IC, II, 16);
case Intrinsic::aarch64_sve_ptest_any:
case Intrinsic::aarch64_sve_ptest_first:
case Intrinsic::aarch64_sve_ptest_last:
return instCombineSVEPTest(IC, II);
case Intrinsic::aarch64_sve_fadd:
return instCombineSVEVectorFAdd(IC, II);
case Intrinsic::aarch64_sve_fadd_u:
return instCombineSVEVectorFAddU(IC, II);
case Intrinsic::aarch64_sve_fmul_u:
return instCombineSVEVectorBinOp(IC, II);
case Intrinsic::aarch64_sve_fsub:
return instCombineSVEVectorFSub(IC, II);
case Intrinsic::aarch64_sve_fsub_u:
return instCombineSVEVectorFSubU(IC, II);
case Intrinsic::aarch64_sve_add:
return instCombineSVEVectorAdd(IC, II);
case Intrinsic::aarch64_sve_add_u:
return instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul_u,
Intrinsic::aarch64_sve_mla_u>(
IC, II, true);
case Intrinsic::aarch64_sve_sub:
return instCombineSVEVectorSub(IC, II);
case Intrinsic::aarch64_sve_sub_u:
return instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul_u,
Intrinsic::aarch64_sve_mls_u>(
IC, II, true);
case Intrinsic::aarch64_sve_tbl:
return instCombineSVETBL(IC, II);
case Intrinsic::aarch64_sve_uunpkhi:
case Intrinsic::aarch64_sve_uunpklo:
case Intrinsic::aarch64_sve_sunpkhi:
case Intrinsic::aarch64_sve_sunpklo:
return instCombineSVEUnpack(IC, II);
case Intrinsic::aarch64_sve_uzp1:
return instCombineSVEUzp1(IC, II);
case Intrinsic::aarch64_sve_zip1:
case Intrinsic::aarch64_sve_zip2:
return instCombineSVEZip(IC, II);
case Intrinsic::aarch64_sve_ld1_gather_index:
return instCombineLD1GatherIndex(IC, II);
case Intrinsic::aarch64_sve_st1_scatter_index:
return instCombineST1ScatterIndex(IC, II);
case Intrinsic::aarch64_sve_ld1:
return instCombineSVELD1(IC, II, DL);
case Intrinsic::aarch64_sve_st1:
return instCombineSVEST1(IC, II, DL);
case Intrinsic::aarch64_sve_sdiv:
return instCombineSVESDIV(IC, II);
case Intrinsic::aarch64_sve_sel:
return instCombineSVESel(IC, II);
case Intrinsic::aarch64_sve_srshl:
return instCombineSVESrshl(IC, II);
case Intrinsic::aarch64_sve_dupq_lane:
return instCombineSVEDupqLane(IC, II);
case Intrinsic::aarch64_sve_insr:
return instCombineSVEInsr(IC, II);
case Intrinsic::aarch64_sve_ptrue:
return instCombinePTrue(IC, II);
}
return std::nullopt;
}
std::optional<Value *> AArch64TTIImpl::simplifyDemandedVectorEltsIntrinsic(
InstCombiner &IC, IntrinsicInst &II, APInt OrigDemandedElts,
APInt &UndefElts, APInt &UndefElts2, APInt &UndefElts3,
std::function<void(Instruction *, unsigned, APInt, APInt &)>
SimplifyAndSetOp) const {
switch (II.getIntrinsicID()) {
default:
break;
case Intrinsic::aarch64_neon_fcvtxn:
case Intrinsic::aarch64_neon_rshrn:
case Intrinsic::aarch64_neon_sqrshrn:
case Intrinsic::aarch64_neon_sqrshrun:
case Intrinsic::aarch64_neon_sqshrn:
case Intrinsic::aarch64_neon_sqshrun:
case Intrinsic::aarch64_neon_sqxtn:
case Intrinsic::aarch64_neon_sqxtun:
case Intrinsic::aarch64_neon_uqrshrn:
case Intrinsic::aarch64_neon_uqshrn:
case Intrinsic::aarch64_neon_uqxtn:
SimplifyAndSetOp(&II, 0, OrigDemandedElts, UndefElts);
break;
}
return std::nullopt;
}
bool AArch64TTIImpl::enableScalableVectorization() const {
return ST->isSVEAvailable() || (ST->isSVEorStreamingSVEAvailable() &&
EnableScalableAutovecInStreamingMode);
}
TypeSize
AArch64TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
switch (K) {
case TargetTransformInfo::RGK_Scalar:
return TypeSize::getFixed(64);
case TargetTransformInfo::RGK_FixedWidthVector:
if (ST->useSVEForFixedLengthVectors() &&
(ST->isSVEAvailable() || EnableFixedwidthAutovecInStreamingMode))
return TypeSize::getFixed(
std::max(ST->getMinSVEVectorSizeInBits(), 128u));
else if (ST->isNeonAvailable())
return TypeSize::getFixed(128);
else
return TypeSize::getFixed(0);
case TargetTransformInfo::RGK_ScalableVector:
if (ST->isSVEAvailable() || (ST->isSVEorStreamingSVEAvailable() &&
EnableScalableAutovecInStreamingMode))
return TypeSize::getScalable(128);
else
return TypeSize::getScalable(0);
}
llvm_unreachable("Unsupported register kind");
}
bool AArch64TTIImpl::isWideningInstruction(Type *DstTy, unsigned Opcode,
ArrayRef<const Value *> Args,
Type *SrcOverrideTy) const {
// A helper that returns a vector type from the given type. The number of
// elements in type Ty determines the vector width.
auto toVectorTy = [&](Type *ArgTy) {
return VectorType::get(ArgTy->getScalarType(),
cast<VectorType>(DstTy)->getElementCount());
};
// Exit early if DstTy is not a vector type whose elements are one of [i16,
// i32, i64]. SVE doesn't generally have the same set of instructions to
// perform an extend with the add/sub/mul. There are SMULLB style
// instructions, but they operate on top/bottom, requiring some sort of lane
// interleaving to be used with zext/sext.
unsigned DstEltSize = DstTy->getScalarSizeInBits();
if (!useNeonVector(DstTy) || Args.size() != 2 ||
(DstEltSize != 16 && DstEltSize != 32 && DstEltSize != 64))
return false;
// Determine if the operation has a widening variant. We consider both the
// "long" (e.g., usubl) and "wide" (e.g., usubw) versions of the
// instructions.
//
// TODO: Add additional widening operations (e.g., shl, etc.) once we
// verify that their extending operands are eliminated during code
// generation.
Type *SrcTy = SrcOverrideTy;
switch (Opcode) {
case Instruction::Add: // UADDL(2), SADDL(2), UADDW(2), SADDW(2).
case Instruction::Sub: // USUBL(2), SSUBL(2), USUBW(2), SSUBW(2).
// The second operand needs to be an extend
if (isa<SExtInst>(Args[1]) || isa<ZExtInst>(Args[1])) {
if (!SrcTy)
SrcTy =
toVectorTy(cast<Instruction>(Args[1])->getOperand(0)->getType());
} else
return false;
break;
case Instruction::Mul: { // SMULL(2), UMULL(2)
// Both operands need to be extends of the same type.
if ((isa<SExtInst>(Args[0]) && isa<SExtInst>(Args[1])) ||
(isa<ZExtInst>(Args[0]) && isa<ZExtInst>(Args[1]))) {
if (!SrcTy)
SrcTy =
toVectorTy(cast<Instruction>(Args[0])->getOperand(0)->getType());
} else if (isa<ZExtInst>(Args[0]) || isa<ZExtInst>(Args[1])) {
// If one of the operands is a Zext and the other has enough zero bits to
// be treated as unsigned, we can still general a umull, meaning the zext
// is free.
KnownBits Known =
computeKnownBits(isa<ZExtInst>(Args[0]) ? Args[1] : Args[0], DL);
if (Args[0]->getType()->getScalarSizeInBits() -
Known.Zero.countLeadingOnes() >
DstTy->getScalarSizeInBits() / 2)
return false;
if (!SrcTy)
SrcTy = toVectorTy(Type::getIntNTy(DstTy->getContext(),
DstTy->getScalarSizeInBits() / 2));
} else
return false;
break;
}
default:
return false;
}
// Legalize the destination type and ensure it can be used in a widening
// operation.
auto DstTyL = getTypeLegalizationCost(DstTy);
if (!DstTyL.second.isVector() || DstEltSize != DstTy->getScalarSizeInBits())
return false;
// Legalize the source type and ensure it can be used in a widening
// operation.
assert(SrcTy && "Expected some SrcTy");
auto SrcTyL = getTypeLegalizationCost(SrcTy);
unsigned SrcElTySize = SrcTyL.second.getScalarSizeInBits();
if (!SrcTyL.second.isVector() || SrcElTySize != SrcTy->getScalarSizeInBits())
return false;
// Get the total number of vector elements in the legalized types.
InstructionCost NumDstEls =
DstTyL.first * DstTyL.second.getVectorMinNumElements();
InstructionCost NumSrcEls =
SrcTyL.first * SrcTyL.second.getVectorMinNumElements();
// Return true if the legalized types have the same number of vector elements
// and the destination element type size is twice that of the source type.
return NumDstEls == NumSrcEls && 2 * SrcElTySize == DstEltSize;
}
// s/urhadd instructions implement the following pattern, making the
// extends free:
// %x = add ((zext i8 -> i16), 1)
// %y = (zext i8 -> i16)
// trunc i16 (lshr (add %x, %y), 1) -> i8
//
bool AArch64TTIImpl::isExtPartOfAvgExpr(const Instruction *ExtUser, Type *Dst,
Type *Src) const {
// The source should be a legal vector type.
if (!Src->isVectorTy() || !TLI->isTypeLegal(TLI->getValueType(DL, Src)) ||
(Src->isScalableTy() && !ST->hasSVE2()))
return false;
if (ExtUser->getOpcode() != Instruction::Add || !ExtUser->hasOneUse())
return false;
// Look for trunc/shl/add before trying to match the pattern.
const Instruction *Add = ExtUser;
auto *AddUser =
dyn_cast_or_null<Instruction>(Add->getUniqueUndroppableUser());
if (AddUser && AddUser->getOpcode() == Instruction::Add)
Add = AddUser;
auto *Shr = dyn_cast_or_null<Instruction>(Add->getUniqueUndroppableUser());
if (!Shr || Shr->getOpcode() != Instruction::LShr)
return false;
auto *Trunc = dyn_cast_or_null<Instruction>(Shr->getUniqueUndroppableUser());
if (!Trunc || Trunc->getOpcode() != Instruction::Trunc ||
Src->getScalarSizeInBits() !=
cast<CastInst>(Trunc)->getDestTy()->getScalarSizeInBits())
return false;
// Try to match the whole pattern. Ext could be either the first or second
// m_ZExtOrSExt matched.
Instruction *Ex1, *Ex2;
if (!(match(Add, m_c_Add(m_Instruction(Ex1),
m_c_Add(m_Instruction(Ex2), m_SpecificInt(1))))))
return false;
// Ensure both extends are of the same type
if (match(Ex1, m_ZExtOrSExt(m_Value())) &&
Ex1->getOpcode() == Ex2->getOpcode())
return true;
return false;
}
InstructionCost AArch64TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
Type *Src,
TTI::CastContextHint CCH,
TTI::TargetCostKind CostKind,
const Instruction *I) const {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
// If the cast is observable, and it is used by a widening instruction (e.g.,
// uaddl, saddw, etc.), it may be free.
if (I && I->hasOneUser()) {
auto *SingleUser = cast<Instruction>(*I->user_begin());
SmallVector<const Value *, 4> Operands(SingleUser->operand_values());
if (isWideningInstruction(Dst, SingleUser->getOpcode(), Operands, Src)) {
// For adds only count the second operand as free if both operands are
// extends but not the same operation. (i.e both operands are not free in
// add(sext, zext)).
if (SingleUser->getOpcode() == Instruction::Add) {
if (I == SingleUser->getOperand(1) ||
(isa<CastInst>(SingleUser->getOperand(1)) &&
cast<CastInst>(SingleUser->getOperand(1))->getOpcode() == Opcode))
return 0;
} else // Others are free so long as isWideningInstruction returned true.
return 0;
}
// The cast will be free for the s/urhadd instructions
if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
isExtPartOfAvgExpr(SingleUser, Dst, Src))
return 0;
}
// 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;
};
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));
static const TypeConversionCostTblEntry BF16Tbl[] = {
{ISD::FP_ROUND, MVT::bf16, MVT::f32, 1}, // bfcvt
{ISD::FP_ROUND, MVT::bf16, MVT::f64, 1}, // bfcvt
{ISD::FP_ROUND, MVT::v4bf16, MVT::v4f32, 1}, // bfcvtn
{ISD::FP_ROUND, MVT::v8bf16, MVT::v8f32, 2}, // bfcvtn+bfcvtn2
{ISD::FP_ROUND, MVT::v2bf16, MVT::v2f64, 2}, // bfcvtn+fcvtn
{ISD::FP_ROUND, MVT::v4bf16, MVT::v4f64, 3}, // fcvtn+fcvtl2+bfcvtn
{ISD::FP_ROUND, MVT::v8bf16, MVT::v8f64, 6}, // 2 * fcvtn+fcvtn2+bfcvtn
};
if (ST->hasBF16())
if (const auto *Entry = ConvertCostTableLookup(
BF16Tbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
// Symbolic constants for the SVE sitofp/uitofp entries in the table below
// The cost of unpacking twice is artificially increased for now in order
// to avoid regressions against NEON, which will use tbl instructions directly
// instead of multiple layers of [s|u]unpk[lo|hi].
// We use the unpacks in cases where the destination type is illegal and
// requires splitting of the input, even if the input type itself is legal.
const unsigned int SVE_EXT_COST = 1;
const unsigned int SVE_FCVT_COST = 1;
const unsigned int SVE_UNPACK_ONCE = 4;
const unsigned int SVE_UNPACK_TWICE = 16;
static const TypeConversionCostTblEntry ConversionTbl[] = {
{ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 1}, // xtn
{ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 1}, // xtn
{ISD::TRUNCATE, MVT::v2i32, MVT::v2i64, 1}, // xtn
{ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 1}, // xtn
{ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 3}, // 2 xtn + 1 uzp1
{ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1}, // xtn
{ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2}, // 1 uzp1 + 1 xtn
{ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1}, // 1 uzp1
{ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 1}, // 1 xtn
{ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2}, // 1 uzp1 + 1 xtn
{ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 4}, // 3 x uzp1 + xtn
{ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1}, // 1 uzp1
{ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 3}, // 3 x uzp1
{ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 2}, // 2 x uzp1
{ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 1}, // uzp1
{ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 3}, // (2 + 1) x uzp1
{ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 7}, // (4 + 2 + 1) x uzp1
{ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2}, // 2 x uzp1
{ISD::TRUNCATE, MVT::v16i16, MVT::v16i64, 6}, // (4 + 2) x uzp1
{ISD::TRUNCATE, MVT::v16i32, MVT::v16i64, 4}, // 4 x uzp1
// Truncations on nxvmiN
{ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i8, 2},
{ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i16, 2},
{ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i32, 2},
{ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i64, 2},
{ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i8, 2},
{ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i16, 2},
{ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i32, 2},
{ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i64, 5},
{ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i8, 2},
{ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i16, 2},
{ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i32, 5},
{ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i64, 11},
{ISD::TRUNCATE, MVT::nxv16i1, MVT::nxv16i8, 2},
{ISD::TRUNCATE, MVT::nxv2i8, MVT::nxv2i16, 0},
{ISD::TRUNCATE, MVT::nxv2i8, MVT::nxv2i32, 0},
{ISD::TRUNCATE, MVT::nxv2i8, MVT::nxv2i64, 0},
{ISD::TRUNCATE, MVT::nxv2i16, MVT::nxv2i32, 0},
{ISD::TRUNCATE, MVT::nxv2i16, MVT::nxv2i64, 0},
{ISD::TRUNCATE, MVT::nxv2i32, MVT::nxv2i64, 0},
{ISD::TRUNCATE, MVT::nxv4i8, MVT::nxv4i16, 0},
{ISD::TRUNCATE, MVT::nxv4i8, MVT::nxv4i32, 0},
{ISD::TRUNCATE, MVT::nxv4i8, MVT::nxv4i64, 1},
{ISD::TRUNCATE, MVT::nxv4i16, MVT::nxv4i32, 0},
{ISD::TRUNCATE, MVT::nxv4i16, MVT::nxv4i64, 1},
{ISD::TRUNCATE, MVT::nxv4i32, MVT::nxv4i64, 1},
{ISD::TRUNCATE, MVT::nxv8i8, MVT::nxv8i16, 0},
{ISD::TRUNCATE, MVT::nxv8i8, MVT::nxv8i32, 1},
{ISD::TRUNCATE, MVT::nxv8i8, MVT::nxv8i64, 3},
{ISD::TRUNCATE, MVT::nxv8i16, MVT::nxv8i32, 1},
{ISD::TRUNCATE, MVT::nxv8i16, MVT::nxv8i64, 3},
{ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i16, 1},
{ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i32, 3},
{ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i64, 7},
// The number of shll instructions for the extension.
{ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3},
{ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3},
{ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2},
{ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2},
{ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3},
{ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3},
{ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2},
{ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2},
{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::v16i16, MVT::v16i8, 2},
{ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2},
{ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6},
{ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6},
// FP Ext and trunc
{ISD::FP_EXTEND, MVT::f64, MVT::f32, 1}, // fcvt
{ISD::FP_EXTEND, MVT::v2f64, MVT::v2f32, 1}, // fcvtl
{ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 2}, // fcvtl+fcvtl2
// FP16
{ISD::FP_EXTEND, MVT::f32, MVT::f16, 1}, // fcvt
{ISD::FP_EXTEND, MVT::f64, MVT::f16, 1}, // fcvt
{ISD::FP_EXTEND, MVT::v4f32, MVT::v4f16, 1}, // fcvtl
{ISD::FP_EXTEND, MVT::v8f32, MVT::v8f16, 2}, // fcvtl+fcvtl2
{ISD::FP_EXTEND, MVT::v2f64, MVT::v2f16, 2}, // fcvtl+fcvtl
{ISD::FP_EXTEND, MVT::v4f64, MVT::v4f16, 3}, // fcvtl+fcvtl2+fcvtl
{ISD::FP_EXTEND, MVT::v8f64, MVT::v8f16, 6}, // 2 * fcvtl+fcvtl2+fcvtl
// BF16 (uses shift)
{ISD::FP_EXTEND, MVT::f32, MVT::bf16, 1}, // shl
{ISD::FP_EXTEND, MVT::f64, MVT::bf16, 2}, // shl+fcvt
{ISD::FP_EXTEND, MVT::v4f32, MVT::v4bf16, 1}, // shll
{ISD::FP_EXTEND, MVT::v8f32, MVT::v8bf16, 2}, // shll+shll2
{ISD::FP_EXTEND, MVT::v2f64, MVT::v2bf16, 2}, // shll+fcvtl
{ISD::FP_EXTEND, MVT::v4f64, MVT::v4bf16, 3}, // shll+fcvtl+fcvtl2
{ISD::FP_EXTEND, MVT::v8f64, MVT::v8bf16, 6}, // 2 * shll+fcvtl+fcvtl2
// FP Ext and trunc
{ISD::FP_ROUND, MVT::f32, MVT::f64, 1}, // fcvt
{ISD::FP_ROUND, MVT::v2f32, MVT::v2f64, 1}, // fcvtn
{ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 2}, // fcvtn+fcvtn2
// FP16
{ISD::FP_ROUND, MVT::f16, MVT::f32, 1}, // fcvt
{ISD::FP_ROUND, MVT::f16, MVT::f64, 1}, // fcvt
{ISD::FP_ROUND, MVT::v4f16, MVT::v4f32, 1}, // fcvtn
{ISD::FP_ROUND, MVT::v8f16, MVT::v8f32, 2}, // fcvtn+fcvtn2
{ISD::FP_ROUND, MVT::v2f16, MVT::v2f64, 2}, // fcvtn+fcvtn
{ISD::FP_ROUND, MVT::v4f16, MVT::v4f64, 3}, // fcvtn+fcvtn2+fcvtn
{ISD::FP_ROUND, MVT::v8f16, MVT::v8f64, 6}, // 2 * fcvtn+fcvtn2+fcvtn
// BF16 (more complex, with +bf16 is handled above)
{ISD::FP_ROUND, MVT::bf16, MVT::f32, 8}, // Expansion is ~8 insns
{ISD::FP_ROUND, MVT::bf16, MVT::f64, 9}, // fcvtn + above
{ISD::FP_ROUND, MVT::v2bf16, MVT::v2f32, 8},
{ISD::FP_ROUND, MVT::v4bf16, MVT::v4f32, 8},
{ISD::FP_ROUND, MVT::v8bf16, MVT::v8f32, 15},
{ISD::FP_ROUND, MVT::v2bf16, MVT::v2f64, 9},
{ISD::FP_ROUND, MVT::v4bf16, MVT::v4f64, 10},
{ISD::FP_ROUND, MVT::v8bf16, MVT::v8f64, 19},
// LowerVectorINT_TO_FP:
{ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i32, 1},
{ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1},
{ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1},
{ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1},
{ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1},
{ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1},
// SVE: to nxv2f16
{ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i16, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i32, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i64, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i16, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i32, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i64, SVE_FCVT_COST},
// SVE: to nxv4f16
{ISD::SINT_TO_FP, MVT::nxv4f16, MVT::nxv4i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f16, MVT::nxv4i16, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f16, MVT::nxv4i32, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f16, MVT::nxv4i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f16, MVT::nxv4i16, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f16, MVT::nxv4i32, SVE_FCVT_COST},
// SVE: to nxv8f16
{ISD::SINT_TO_FP, MVT::nxv8f16, MVT::nxv8i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv8f16, MVT::nxv8i16, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv8f16, MVT::nxv8i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv8f16, MVT::nxv8i16, SVE_FCVT_COST},
// SVE: to nxv16f16
{ISD::SINT_TO_FP, MVT::nxv16f16, MVT::nxv16i8,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv16f16, MVT::nxv16i8,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
// Complex: to v2f32
{ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i8, 3},
{ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i16, 3},
{ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i8, 3},
{ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i16, 3},
// SVE: to nxv2f32
{ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i16, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i32, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i64, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i16, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i32, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i64, SVE_FCVT_COST},
// Complex: to v4f32
{ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 4},
{ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 2},
{ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 3},
{ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2},
// SVE: to nxv4f32
{ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i16, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i32, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f32, MVT::nxv4i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f32, MVT::nxv4i16, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i32, SVE_FCVT_COST},
// Complex: to v8f32
{ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 10},
{ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4},
{ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 10},
{ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4},
// SVE: to nxv8f32
{ISD::SINT_TO_FP, MVT::nxv8f32, MVT::nxv8i8,
SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv8f32, MVT::nxv8i16,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv8f32, MVT::nxv8i8,
SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv8f32, MVT::nxv8i16,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
// SVE: to nxv16f32
{ISD::SINT_TO_FP, MVT::nxv16f32, MVT::nxv16i8,
SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv16f32, MVT::nxv16i8,
SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
// Complex: to v16f32
{ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 21},
{ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 21},
// Complex: to v2f64
{ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i8, 4},
{ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i16, 4},
{ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2},
{ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 4},
{ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 4},
{ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2},
// SVE: to nxv2f64
{ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i16, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i32, SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i64, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i8,
SVE_EXT_COST + SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i16, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i32, SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i64, SVE_FCVT_COST},
// Complex: to v4f64
{ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 4},
{ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 4},
// SVE: to nxv4f64
{ISD::SINT_TO_FP, MVT::nxv4f64, MVT::nxv4i8,
SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f64, MVT::nxv4i16,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv4f64, MVT::nxv4i32,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f64, MVT::nxv4i8,
SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f64, MVT::nxv4i16,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv4f64, MVT::nxv4i32,
SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
// SVE: to nxv8f64
{ISD::SINT_TO_FP, MVT::nxv8f64, MVT::nxv8i8,
SVE_EXT_COST + SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
{ISD::SINT_TO_FP, MVT::nxv8f64, MVT::nxv8i16,
SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv8f64, MVT::nxv8i8,
SVE_EXT_COST + SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
{ISD::UINT_TO_FP, MVT::nxv8f64, MVT::nxv8i16,
SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
// LowerVectorFP_TO_INT
{ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f32, 1},
{ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1},
{ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1},
{ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f32, 1},
{ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1},
{ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1},
// Complex, from v2f32: legal type is v2i32 (no cost) or v2i64 (1 ext).
{ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f32, 2},
{ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f32, 1},
{ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f32, 1},
{ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f32, 2},
{ISD::FP_TO_UINT, MVT::v2i16, MVT::v2f32, 1},
{ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f32, 1},
// Complex, from v4f32: legal type is v4i16, 1 narrowing => ~2
{ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2},
{ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 2},
{ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 2},
{ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f32, 2},
// Complex, from nxv2f32.
{ISD::FP_TO_SINT, MVT::nxv2i64, MVT::nxv2f32, 1},
{ISD::FP_TO_SINT, MVT::nxv2i32, MVT::nxv2f32, 1},
{ISD::FP_TO_SINT, MVT::nxv2i16, MVT::nxv2f32, 1},
{ISD::FP_TO_SINT, MVT::nxv2i8, MVT::nxv2f32, 1},
{ISD::FP_TO_UINT, MVT::nxv2i64, MVT::nxv2f32, 1},
{ISD::FP_TO_UINT, MVT::nxv2i32, MVT::nxv2f32, 1},
{ISD::FP_TO_UINT, MVT::nxv2i16, MVT::nxv2f32, 1},
{ISD::FP_TO_UINT, MVT::nxv2i8, MVT::nxv2f32, 1},
// Complex, from v2f64: legal type is v2i32, 1 narrowing => ~2.
{ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 2},
{ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f64, 2},
{ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f64, 2},
{ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 2},
{ISD::FP_TO_UINT, MVT::v2i16, MVT::v2f64, 2},
{ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f64, 2},
// Complex, from nxv2f64.
{ISD::FP_TO_SINT, MVT::nxv2i64, MVT::nxv2f64, 1},
{ISD::FP_TO_SINT, MVT::nxv2i32, MVT::nxv2f64, 1},
{ISD::FP_TO_SINT, MVT::nxv2i16, MVT::nxv2f64, 1},
{ISD::FP_TO_SINT, MVT::nxv2i8, MVT::nxv2f64, 1},
{ISD::FP_TO_SINT, MVT::nxv2i1, MVT::nxv2f64, 1},
{ISD::FP_TO_UINT, MVT::nxv2i64, MVT::nxv2f64, 1},
{ISD::FP_TO_UINT, MVT::nxv2i32, MVT::nxv2f64, 1},
{ISD::FP_TO_UINT, MVT::nxv2i16, MVT::nxv2f64, 1},
{ISD::FP_TO_UINT, MVT::nxv2i8, MVT::nxv2f64, 1},
{ISD::FP_TO_UINT, MVT::nxv2i1, MVT::nxv2f64, 1},
// Complex, from nxv4f32.
{ISD::FP_TO_SINT, MVT::nxv4i64, MVT::nxv4f32, 4},
{ISD::FP_TO_SINT, MVT::nxv4i32, MVT::nxv4f32, 1},
{ISD::FP_TO_SINT, MVT::nxv4i16, MVT::nxv4f32, 1},
{ISD::FP_TO_SINT, MVT::nxv4i8, MVT::nxv4f32, 1},
{ISD::FP_TO_SINT, MVT::nxv4i1, MVT::nxv4f32, 1},
{ISD::FP_TO_UINT, MVT::nxv4i64, MVT::nxv4f32, 4},
{ISD::FP_TO_UINT, MVT::nxv4i32, MVT::nxv4f32, 1},
{ISD::FP_TO_UINT, MVT::nxv4i16, MVT::nxv4f32, 1},
{ISD::FP_TO_UINT, MVT::nxv4i8, MVT::nxv4f32, 1},
{ISD::FP_TO_UINT, MVT::nxv4i1, MVT::nxv4f32, 1},
// Complex, from nxv8f64. Illegal -> illegal conversions not required.
{ISD::FP_TO_SINT, MVT::nxv8i16, MVT::nxv8f64, 7},
{ISD::FP_TO_SINT, MVT::nxv8i8, MVT::nxv8f64, 7},
{ISD::FP_TO_UINT, MVT::nxv8i16, MVT::nxv8f64, 7},
{ISD::FP_TO_UINT, MVT::nxv8i8, MVT::nxv8f64, 7},
// Complex, from nxv4f64. Illegal -> illegal conversions not required.
{ISD::FP_TO_SINT, MVT::nxv4i32, MVT::nxv4f64, 3},
{ISD::FP_TO_SINT, MVT::nxv4i16, MVT::nxv4f64, 3},
{ISD::FP_TO_SINT, MVT::nxv4i8, MVT::nxv4f64, 3},
{ISD::FP_TO_UINT, MVT::nxv4i32, MVT::nxv4f64, 3},
{ISD::FP_TO_UINT, MVT::nxv4i16, MVT::nxv4f64, 3},
{ISD::FP_TO_UINT, MVT::nxv4i8, MVT::nxv4f64, 3},
// Complex, from nxv8f32. Illegal -> illegal conversions not required.
{ISD::FP_TO_SINT, MVT::nxv8i16, MVT::nxv8f32, 3},
{ISD::FP_TO_SINT, MVT::nxv8i8, MVT::nxv8f32, 3},
{ISD::FP_TO_UINT, MVT::nxv8i16, MVT::nxv8f32, 3},
{ISD::FP_TO_UINT, MVT::nxv8i8, MVT::nxv8f32, 3},
// Complex, from nxv8f16.
{ISD::FP_TO_SINT, MVT::nxv8i64, MVT::nxv8f16, 10},
{ISD::FP_TO_SINT, MVT::nxv8i32, MVT::nxv8f16, 4},
{ISD::FP_TO_SINT, MVT::nxv8i16, MVT::nxv8f16, 1},
{ISD::FP_TO_SINT, MVT::nxv8i8, MVT::nxv8f16, 1},
{ISD::FP_TO_SINT, MVT::nxv8i1, MVT::nxv8f16, 1},
{ISD::FP_TO_UINT, MVT::nxv8i64, MVT::nxv8f16, 10},
{ISD::FP_TO_UINT, MVT::nxv8i32, MVT::nxv8f16, 4},
{ISD::FP_TO_UINT, MVT::nxv8i16, MVT::nxv8f16, 1},
{ISD::FP_TO_UINT, MVT::nxv8i8, MVT::nxv8f16, 1},
{ISD::FP_TO_UINT, MVT::nxv8i1, MVT::nxv8f16, 1},
// Complex, from nxv4f16.
{ISD::FP_TO_SINT, MVT::nxv4i64, MVT::nxv4f16, 4},
{ISD::FP_TO_SINT, MVT::nxv4i32, MVT::nxv4f16, 1},
{ISD::FP_TO_SINT, MVT::nxv4i16, MVT::nxv4f16, 1},
{ISD::FP_TO_SINT, MVT::nxv4i8, MVT::nxv4f16, 1},
{ISD::FP_TO_UINT, MVT::nxv4i64, MVT::nxv4f16, 4},
{ISD::FP_TO_UINT, MVT::nxv4i32, MVT::nxv4f16, 1},
{ISD::FP_TO_UINT, MVT::nxv4i16, MVT::nxv4f16, 1},
{ISD::FP_TO_UINT, MVT::nxv4i8, MVT::nxv4f16, 1},
// Complex, from nxv2f16.
{ISD::FP_TO_SINT, MVT::nxv2i64, MVT::nxv2f16, 1},
{ISD::FP_TO_SINT, MVT::nxv2i32, MVT::nxv2f16, 1},
{ISD::FP_TO_SINT, MVT::nxv2i16, MVT::nxv2f16, 1},
{ISD::FP_TO_SINT, MVT::nxv2i8, MVT::nxv2f16, 1},
{ISD::FP_TO_UINT, MVT::nxv2i64, MVT::nxv2f16, 1},
{ISD::FP_TO_UINT, MVT::nxv2i32, MVT::nxv2f16, 1},
{ISD::FP_TO_UINT, MVT::nxv2i16, MVT::nxv2f16, 1},
{ISD::FP_TO_UINT, MVT::nxv2i8, MVT::nxv2f16, 1},
// Truncate from nxvmf32 to nxvmf16.
{ISD::FP_ROUND, MVT::nxv2f16, MVT::nxv2f32, 1},
{ISD::FP_ROUND, MVT::nxv4f16, MVT::nxv4f32, 1},
{ISD::FP_ROUND, MVT::nxv8f16, MVT::nxv8f32, 3},
// Truncate from nxvmf64 to nxvmf16.
{ISD::FP_ROUND, MVT::nxv2f16, MVT::nxv2f64, 1},
{ISD::FP_ROUND, MVT::nxv4f16, MVT::nxv4f64, 3},
{ISD::FP_ROUND, MVT::nxv8f16, MVT::nxv8f64, 7},
// Truncate from nxvmf64 to nxvmf32.
{ISD::FP_ROUND, MVT::nxv2f32, MVT::nxv2f64, 1},
{ISD::FP_ROUND, MVT::nxv4f32, MVT::nxv4f64, 3},
{ISD::FP_ROUND, MVT::nxv8f32, MVT::nxv8f64, 6},
// Extend from nxvmf16 to nxvmf32.
{ISD::FP_EXTEND, MVT::nxv2f32, MVT::nxv2f16, 1},
{ISD::FP_EXTEND, MVT::nxv4f32, MVT::nxv4f16, 1},
{ISD::FP_EXTEND, MVT::nxv8f32, MVT::nxv8f16, 2},
// Extend from nxvmf16 to nxvmf64.
{ISD::FP_EXTEND, MVT::nxv2f64, MVT::nxv2f16, 1},
{ISD::FP_EXTEND, MVT::nxv4f64, MVT::nxv4f16, 2},
{ISD::FP_EXTEND, MVT::nxv8f64, MVT::nxv8f16, 4},
// Extend from nxvmf32 to nxvmf64.
{ISD::FP_EXTEND, MVT::nxv2f64, MVT::nxv2f32, 1},
{ISD::FP_EXTEND, MVT::nxv4f64, MVT::nxv4f32, 2},
{ISD::FP_EXTEND, MVT::nxv8f64, MVT::nxv8f32, 6},
// Bitcasts from float to integer
{ISD::BITCAST, MVT::nxv2f16, MVT::nxv2i16, 0},
{ISD::BITCAST, MVT::nxv4f16, MVT::nxv4i16, 0},
{ISD::BITCAST, MVT::nxv2f32, MVT::nxv2i32, 0},
// Bitcasts from integer to float
{ISD::BITCAST, MVT::nxv2i16, MVT::nxv2f16, 0},
{ISD::BITCAST, MVT::nxv4i16, MVT::nxv4f16, 0},
{ISD::BITCAST, MVT::nxv2i32, MVT::nxv2f32, 0},
// Add cost for extending to illegal -too wide- scalable vectors.
// zero/sign extend are implemented by multiple unpack operations,
// where each operation has a cost of 1.
{ISD::ZERO_EXTEND, MVT::nxv16i16, MVT::nxv16i8, 2},
{ISD::ZERO_EXTEND, MVT::nxv16i32, MVT::nxv16i8, 6},
{ISD::ZERO_EXTEND, MVT::nxv16i64, MVT::nxv16i8, 14},
{ISD::ZERO_EXTEND, MVT::nxv8i32, MVT::nxv8i16, 2},
{ISD::ZERO_EXTEND, MVT::nxv8i64, MVT::nxv8i16, 6},
{ISD::ZERO_EXTEND, MVT::nxv4i64, MVT::nxv4i32, 2},
{ISD::SIGN_EXTEND, MVT::nxv16i16, MVT::nxv16i8, 2},
{ISD::SIGN_EXTEND, MVT::nxv16i32, MVT::nxv16i8, 6},
{ISD::SIGN_EXTEND, MVT::nxv16i64, MVT::nxv16i8, 14},
{ISD::SIGN_EXTEND, MVT::nxv8i32, MVT::nxv8i16, 2},
{ISD::SIGN_EXTEND, MVT::nxv8i64, MVT::nxv8i16, 6},
{ISD::SIGN_EXTEND, MVT::nxv4i64, MVT::nxv4i32, 2},
};
// We have to estimate a cost of fixed length operation upon
// SVE registers(operations) with the number of registers required
// for a fixed type to be represented upon SVE registers.
EVT WiderTy = SrcTy.bitsGT(DstTy) ? SrcTy : DstTy;
if (SrcTy.isFixedLengthVector() && DstTy.isFixedLengthVector() &&
SrcTy.getVectorNumElements() == DstTy.getVectorNumElements() &&
ST->useSVEForFixedLengthVectors(WiderTy)) {
std::pair<InstructionCost, MVT> LT =
getTypeLegalizationCost(WiderTy.getTypeForEVT(Dst->getContext()));
unsigned NumElements =
AArch64::SVEBitsPerBlock / LT.second.getScalarSizeInBits();
return AdjustCost(
LT.first *
getCastInstrCost(
Opcode, ScalableVectorType::get(Dst->getScalarType(), NumElements),
ScalableVectorType::get(Src->getScalarType(), NumElements), CCH,
CostKind, I));
}
if (const auto *Entry = ConvertCostTableLookup(
ConversionTbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
static const TypeConversionCostTblEntry FP16Tbl[] = {
{ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f16, 1}, // fcvtzs
{ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f16, 1},
{ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f16, 1}, // fcvtzs
{ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f16, 1},
{ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f16, 2}, // fcvtl+fcvtzs
{ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f16, 2},
{ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f16, 2}, // fcvtzs+xtn
{ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f16, 2},
{ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f16, 1}, // fcvtzs
{ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f16, 1},
{ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f16, 4}, // 2*fcvtl+2*fcvtzs
{ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f16, 4},
{ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f16, 3}, // 2*fcvtzs+xtn
{ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f16, 3},
{ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f16, 2}, // 2*fcvtzs
{ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f16, 2},
{ISD::FP_TO_SINT, MVT::v16i32, MVT::v16f16, 8}, // 4*fcvtl+4*fcvtzs
{ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f16, 8},
{ISD::UINT_TO_FP, MVT::v8f16, MVT::v8i8, 2}, // ushll + ucvtf
{ISD::SINT_TO_FP, MVT::v8f16, MVT::v8i8, 2}, // sshll + scvtf
{ISD::UINT_TO_FP, MVT::v16f16, MVT::v16i8, 4}, // 2 * ushl(2) + 2 * ucvtf
{ISD::SINT_TO_FP, MVT::v16f16, MVT::v16i8, 4}, // 2 * sshl(2) + 2 * scvtf
};
if (ST->hasFullFP16())
if (const auto *Entry = ConvertCostTableLookup(
FP16Tbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
return AdjustCost(Entry->Cost);
// INT_TO_FP of i64->f32 will scalarize, which is required to avoid
// double-rounding issues.
if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
DstTy.getScalarType() == MVT::f32 && SrcTy.getScalarSizeInBits() > 32 &&
isa<FixedVectorType>(Dst) && isa<FixedVectorType>(Src))
return AdjustCost(
cast<FixedVectorType>(Dst)->getNumElements() *
getCastInstrCost(Opcode, Dst->getScalarType(), Src->getScalarType(),
CCH, CostKind) +
BaseT::getScalarizationOverhead(cast<FixedVectorType>(Src), false, true,
CostKind) +
BaseT::getScalarizationOverhead(cast<FixedVectorType>(Dst), true, false,
CostKind));
if ((ISD == ISD::ZERO_EXTEND || ISD == ISD::SIGN_EXTEND) &&
CCH == TTI::CastContextHint::Masked &&
ST->isSVEorStreamingSVEAvailable() &&
TLI->getTypeAction(Src->getContext(), SrcTy) ==
TargetLowering::TypePromoteInteger &&
TLI->getTypeAction(Dst->getContext(), DstTy) ==
TargetLowering::TypeSplitVector) {
// The standard behaviour in the backend for these cases is to split the
// extend up into two parts:
// 1. Perform an extending load or masked load up to the legal type.
// 2. Extend the loaded data to the final type.
std::pair<InstructionCost, MVT> SrcLT = getTypeLegalizationCost(Src);
Type *LegalTy = EVT(SrcLT.second).getTypeForEVT(Src->getContext());
InstructionCost Part1 = AArch64TTIImpl::getCastInstrCost(
Opcode, LegalTy, Src, CCH, CostKind, I);
InstructionCost Part2 = AArch64TTIImpl::getCastInstrCost(
Opcode, Dst, LegalTy, TTI::CastContextHint::None, CostKind, I);
return Part1 + Part2;
}
// The BasicTTIImpl version only deals with CCH==TTI::CastContextHint::Normal,
// but we also want to include the TTI::CastContextHint::Masked case too.
if ((ISD == ISD::ZERO_EXTEND || ISD == ISD::SIGN_EXTEND) &&
CCH == TTI::CastContextHint::Masked &&
ST->isSVEorStreamingSVEAvailable() && TLI->isTypeLegal(DstTy))
CCH = TTI::CastContextHint::Normal;
return AdjustCost(
BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
}
InstructionCost
AArch64TTIImpl::getExtractWithExtendCost(unsigned Opcode, Type *Dst,
VectorType *VecTy, unsigned Index,
TTI::TargetCostKind CostKind) const {
// Make sure we were given a valid extend opcode.
assert((Opcode == Instruction::SExt || Opcode == Instruction::ZExt) &&
"Invalid opcode");
// We are extending an element we extract from a vector, so the source type
// of the extend is the element type of the vector.
auto *Src = VecTy->getElementType();
// Sign- and zero-extends are for integer types only.
assert(isa<IntegerType>(Dst) && isa<IntegerType>(Src) && "Invalid type");
// Get the cost for the extract. We compute the cost (if any) for the extend
// below.
InstructionCost Cost = getVectorInstrCost(Instruction::ExtractElement, VecTy,
CostKind, Index, nullptr, nullptr);
// Legalize the types.
auto VecLT = getTypeLegalizationCost(VecTy);
auto DstVT = TLI->getValueType(DL, Dst);
auto SrcVT = TLI->getValueType(DL, Src);
// If the resulting type is still a vector and the destination type is legal,
// we may get the extension for free. If not, get the default cost for the
// extend.
if (!VecLT.second.isVector() || !TLI->isTypeLegal(DstVT))
return Cost + getCastInstrCost(Opcode, Dst, Src, TTI::CastContextHint::None,
CostKind);
// The destination type should be larger than the element type. If not, get
// the default cost for the extend.
if (DstVT.getFixedSizeInBits() < SrcVT.getFixedSizeInBits())
return Cost + getCastInstrCost(Opcode, Dst, Src, TTI::CastContextHint::None,
CostKind);
switch (Opcode) {
default:
llvm_unreachable("Opcode should be either SExt or ZExt");
// For sign-extends, we only need a smov, which performs the extension
// automatically.
case Instruction::SExt:
return Cost;
// For zero-extends, the extend is performed automatically by a umov unless
// the destination type is i64 and the element type is i8 or i16.
case Instruction::ZExt:
if (DstVT.getSizeInBits() != 64u || SrcVT.getSizeInBits() == 32u)
return Cost;
}
// If we are unable to perform the extend for free, get the default cost.
return Cost + getCastInstrCost(Opcode, Dst, Src, TTI::CastContextHint::None,
CostKind);
}
InstructionCost AArch64TTIImpl::getCFInstrCost(unsigned Opcode,
TTI::TargetCostKind CostKind,
const Instruction *I) const {
if (CostKind != TTI::TCK_RecipThroughput)
return Opcode == Instruction::PHI ? 0 : 1;
assert(CostKind == TTI::TCK_RecipThroughput && "unexpected CostKind");
// Branches are assumed to be predicted.
return 0;
}
InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(
unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
bool HasRealUse, const Instruction *I, Value *Scalar,
ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) const {
assert(Val->isVectorTy() && "This must be a vector type");
if (Index != -1U) {
// Legalize the type.
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Val);
// This type is legalized to a scalar type.
if (!LT.second.isVector())
return 0;
// 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;
}
// The element at index zero is already inside the vector.
// - For a physical (HasRealUse==true) insert-element or extract-element
// instruction that extracts integers, an explicit FPR -> GPR move is
// needed. So it has non-zero cost.
// - For the rest of cases (virtual instruction or element type is float),
// consider the instruction free.
if (Index == 0 && (!HasRealUse || !Val->getScalarType()->isIntegerTy()))
return 0;
// This is recognising a LD1 single-element structure to one lane of one
// register instruction. I.e., if this is an `insertelement` instruction,
// and its second operand is a load, then we will generate a LD1, which
// are expensive instructions.
if (I && dyn_cast<LoadInst>(I->getOperand(1)))
return CostKind == TTI::TCK_CodeSize
? 0
: ST->getVectorInsertExtractBaseCost() + 1;
// i1 inserts and extract will include an extra cset or cmp of the vector
// value. Increase the cost by 1 to account.
if (Val->getScalarSizeInBits() == 1)
return CostKind == TTI::TCK_CodeSize
? 2
: ST->getVectorInsertExtractBaseCost() + 1;
// FIXME:
// If the extract-element and insert-element instructions could be
// simplified away (e.g., could be combined into users by looking at use-def
// context), they have no cost. This is not done in the first place for
// compile-time considerations.
}
// In case of Neon, if there exists extractelement from lane != 0 such that
// 1. extractelement does not necessitate a move from vector_reg -> GPR.
// 2. extractelement result feeds into fmul.
// 3. Other operand of fmul is an extractelement from lane 0 or lane
// equivalent to 0.
// then the extractelement can be merged with fmul in the backend and it
// incurs no cost.
// e.g.
// define double @foo(<2 x double> %a) {
// %1 = extractelement <2 x double> %a, i32 0
// %2 = extractelement <2 x double> %a, i32 1
// %res = fmul double %1, %2
// ret double %res
// }
// %2 and %res can be merged in the backend to generate fmul d0, d0, v1.d[1]
auto ExtractCanFuseWithFmul = [&]() {
// We bail out if the extract is from lane 0.
if (Index == 0)
return false;
// Check if the scalar element type of the vector operand of ExtractElement
// instruction is one of the allowed types.
auto IsAllowedScalarTy = [&](const Type *T) {
return T->isFloatTy() || T->isDoubleTy() ||
(T->isHalfTy() && ST->hasFullFP16());
};
// Check if the extractelement user is scalar fmul.
auto IsUserFMulScalarTy = [](const Value *EEUser) {
// Check if the user is scalar fmul.
const auto *BO = dyn_cast<BinaryOperator>(EEUser);
return BO && BO->getOpcode() == BinaryOperator::FMul &&
!BO->getType()->isVectorTy();
};
// Check if the extract index is from lane 0 or lane equivalent to 0 for a
// certain scalar type and a certain vector register width.
auto IsExtractLaneEquivalentToZero = [&](unsigned Idx, unsigned EltSz) {
auto RegWidth =
getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
.getFixedValue();
return Idx == 0 || (RegWidth != 0 && (Idx * EltSz) % RegWidth == 0);
};
// Check if the type constraints on input vector type and result scalar type
// of extractelement instruction are satisfied.
if (!isa<FixedVectorType>(Val) || !IsAllowedScalarTy(Val->getScalarType()))
return false;
if (Scalar) {
DenseMap<User *, unsigned> UserToExtractIdx;
for (auto *U : Scalar->users()) {
if (!IsUserFMulScalarTy(U))
return false;
// Recording entry for the user is important. Index value is not
// important.
UserToExtractIdx[U];
}
if (UserToExtractIdx.empty())
return false;
for (auto &[S, U, L] : ScalarUserAndIdx) {
for (auto *U : S->users()) {
if (UserToExtractIdx.find(U) != UserToExtractIdx.end()) {
auto *FMul = cast<BinaryOperator>(U);
auto *Op0 = FMul->getOperand(0);
auto *Op1 = FMul->getOperand(1);
if ((Op0 == S && Op1 == S) || Op0 != S || Op1 != S) {
UserToExtractIdx[U] = L;
break;
}
}
}
}
for (auto &[U, L] : UserToExtractIdx) {
if (!IsExtractLaneEquivalentToZero(Index, Val->getScalarSizeInBits()) &&
!IsExtractLaneEquivalentToZero(L, Val->getScalarSizeInBits()))
return false;
}
} else {
const auto *EE = cast<ExtractElementInst>(I);
const auto *IdxOp = dyn_cast<ConstantInt>(EE->getIndexOperand());
if (!IdxOp)
return false;
return !EE->users().empty() && all_of(EE->users(), [&](const User *U) {
if (!IsUserFMulScalarTy(U))
return false;
// Check if the other operand of extractelement is also extractelement
// from lane equivalent to 0.
const auto *BO = cast<BinaryOperator>(U);
const auto *OtherEE = dyn_cast<ExtractElementInst>(
BO->getOperand(0) == EE ? BO->getOperand(1) : BO->getOperand(0));
if (OtherEE) {
const auto *IdxOp = dyn_cast<ConstantInt>(OtherEE->getIndexOperand());
if (!IdxOp)
return false;
return IsExtractLaneEquivalentToZero(
cast<ConstantInt>(OtherEE->getIndexOperand())
->getValue()
.getZExtValue(),
OtherEE->getType()->getScalarSizeInBits());
}
return true;
});
}
return true;
};
if (Opcode == Instruction::ExtractElement && (I || Scalar) &&
ExtractCanFuseWithFmul())
return 0;
// All other insert/extracts cost this much.
return CostKind == TTI::TCK_CodeSize ? 1
: ST->getVectorInsertExtractBaseCost();
}
InstructionCost AArch64TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
TTI::TargetCostKind CostKind,
unsigned Index, Value *Op0,
Value *Op1) const {
bool HasRealUse =
Opcode == Instruction::InsertElement && Op0 && !isa<UndefValue>(Op0);
return getVectorInstrCostHelper(Opcode, Val, CostKind, Index, HasRealUse);
}
InstructionCost AArch64TTIImpl::getVectorInstrCost(
unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
Value *Scalar,
ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) const {
return getVectorInstrCostHelper(Opcode, Val, CostKind, Index, false, nullptr,
Scalar, ScalarUserAndIdx);
}
InstructionCost AArch64TTIImpl::getVectorInstrCost(const Instruction &I,
Type *Val,
TTI::TargetCostKind CostKind,
unsigned Index) const {
return getVectorInstrCostHelper(I.getOpcode(), Val, CostKind, Index,
true /* HasRealUse */, &I);
}
InstructionCost AArch64TTIImpl::getScalarizationOverhead(
VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract,
TTI::TargetCostKind CostKind, bool ForPoisonSrc,
ArrayRef<Value *> VL) const {
if (isa<ScalableVectorType>(Ty))
return InstructionCost::getInvalid();
if (Ty->getElementType()->isFloatingPointTy())
return BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, Extract,
CostKind);
unsigned VecInstCost =
CostKind == TTI::TCK_CodeSize ? 1 : ST->getVectorInsertExtractBaseCost();
return DemandedElts.popcount() * (Insert + Extract) * VecInstCost;
}
InstructionCost AArch64TTIImpl::getArithmeticInstrCost(
unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info,
ArrayRef<const Value *> Args, const Instruction *CxtI) const {
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
if (auto *VTy = dyn_cast<ScalableVectorType>(Ty))
if (VTy->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
// TODO: Handle more cost kinds.
if (CostKind != TTI::TCK_RecipThroughput)
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
Op2Info, Args, CxtI);
// Legalize the type.
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
int ISD = TLI->InstructionOpcodeToISD(Opcode);
switch (ISD) {
default:
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
Op2Info);
case ISD::SREM:
case ISD::SDIV:
/*
Notes for sdiv/srem specific costs:
1. This only considers the cases where the divisor is constant, uniform and
(pow-of-2/non-pow-of-2). Other cases are not important since they either
result in some form of (ldr + adrp), corresponding to constant vectors, or
scalarization of the division operation.
2. Constant divisors, either negative in whole or partially, don't result in
significantly different codegen as compared to positive constant divisors.
So, we don't consider negative divisors seperately.
3. If the codegen is significantly different with SVE, it has been indicated
using comments at appropriate places.
sdiv specific cases:
-----------------------------------------------------------------------
codegen | pow-of-2 | Type
-----------------------------------------------------------------------
add + cmp + csel + asr | Y | i64
add + cmp + csel + asr | Y | i32
-----------------------------------------------------------------------
srem specific cases:
-----------------------------------------------------------------------
codegen | pow-of-2 | Type
-----------------------------------------------------------------------
negs + and + and + csneg | Y | i64
negs + and + and + csneg | Y | i32
-----------------------------------------------------------------------
other sdiv/srem cases:
-------------------------------------------------------------------------
commom codegen | + srem | + sdiv | pow-of-2 | Type
-------------------------------------------------------------------------
smulh + asr + add + add | - | - | N | i64
smull + lsr + add + add | - | - | N | i32
usra | and + sub | sshr | Y | <2 x i64>
2 * (scalar code) | - | - | N | <2 x i64>
usra | bic + sub | sshr + neg | Y | <4 x i32>
smull2 + smull + uzp2 | mls | - | N | <4 x i32>
+ sshr + usra | | | |
-------------------------------------------------------------------------
*/
if (Op2Info.isConstant() && Op2Info.isUniform()) {
InstructionCost AddCost =
getArithmeticInstrCost(Instruction::Add, Ty, CostKind,
Op1Info.getNoProps(), Op2Info.getNoProps());
InstructionCost AsrCost =
getArithmeticInstrCost(Instruction::AShr, Ty, CostKind,
Op1Info.getNoProps(), Op2Info.getNoProps());
InstructionCost MulCost =
getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
Op1Info.getNoProps(), Op2Info.getNoProps());
// add/cmp/csel/csneg should have similar cost while asr/negs/and should
// have similar cost.
auto VT = TLI->getValueType(DL, Ty);
if (VT.isScalarInteger() && VT.getSizeInBits() <= 64) {
if (Op2Info.isPowerOf2()) {
return ISD == ISD::SDIV ? (3 * AddCost + AsrCost)
: (3 * AsrCost + AddCost);
} else {
return MulCost + AsrCost + 2 * AddCost;
}
} else if (VT.isVector()) {
InstructionCost UsraCost = 2 * AsrCost;
if (Op2Info.isPowerOf2()) {
// Division with scalable types corresponds to native 'asrd'
// instruction when SVE is available.
// e.g. %1 = sdiv <vscale x 4 x i32> %a, splat (i32 8)
if (Ty->isScalableTy() && ST->hasSVE())
return 2 * AsrCost;
return UsraCost +
(ISD == ISD::SDIV
? (LT.second.getScalarType() == MVT::i64 ? 1 : 2) *
AsrCost
: 2 * AddCost);
} else if (LT.second == MVT::v2i64) {
return VT.getVectorNumElements() *
getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind,
Op1Info.getNoProps(),
Op2Info.getNoProps());
} else {
// When SVE is available, we get:
// smulh + lsr + add/sub + asr + add/sub.
if (Ty->isScalableTy() && ST->hasSVE())
return MulCost /*smulh cost*/ + 2 * AddCost + 2 * AsrCost;
return 2 * MulCost + AddCost /*uzp2 cost*/ + AsrCost + UsraCost;
}
}
}
if (Op2Info.isConstant() && !Op2Info.isUniform() &&
LT.second.isFixedLengthVector()) {
// FIXME: When the constant vector is non-uniform, this may result in
// loading the vector from constant pool or in some cases, may also result
// in scalarization. For now, we are approximating this with the
// scalarization cost.
auto ExtractCost = 2 * getVectorInstrCost(Instruction::ExtractElement, Ty,
CostKind, -1, nullptr, nullptr);
auto InsertCost = getVectorInstrCost(Instruction::InsertElement, Ty,
CostKind, -1, nullptr, nullptr);
unsigned NElts = cast<FixedVectorType>(Ty)->getNumElements();
return ExtractCost + InsertCost +
NElts * getArithmeticInstrCost(Opcode, Ty->getScalarType(),
CostKind, Op1Info.getNoProps(),
Op2Info.getNoProps());
}
[[fallthrough]];
case ISD::UDIV:
case ISD::UREM: {
auto VT = TLI->getValueType(DL, Ty);
if (Op2Info.isConstant()) {
// If the operand is a power of 2 we can use the shift or and cost.
if (ISD == ISD::UDIV && Op2Info.isPowerOf2())
return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind,
Op1Info.getNoProps(),
Op2Info.getNoProps());
if (ISD == ISD::UREM && Op2Info.isPowerOf2())
return getArithmeticInstrCost(Instruction::And, Ty, CostKind,
Op1Info.getNoProps(),
Op2Info.getNoProps());
if (ISD == ISD::UDIV || ISD == ISD::UREM) {
// Divides by a constant are expanded to MULHU + SUB + SRL + ADD + SRL.
// The MULHU will be expanded to UMULL for the types not listed below,
// and will become a pair of UMULL+MULL2 for 128bit vectors.
bool HasMULH = VT == MVT::i64 || LT.second == MVT::nxv2i64 ||
LT.second == MVT::nxv4i32 || LT.second == MVT::nxv8i16 ||
LT.second == MVT::nxv16i8;
bool Is128bit = LT.second.is128BitVector();
InstructionCost MulCost =
getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
Op1Info.getNoProps(), Op2Info.getNoProps());
InstructionCost AddCost =
getArithmeticInstrCost(Instruction::Add, Ty, CostKind,
Op1Info.getNoProps(), Op2Info.getNoProps());
InstructionCost ShrCost =
getArithmeticInstrCost(Instruction::AShr, Ty, CostKind,
Op1Info.getNoProps(), Op2Info.getNoProps());
InstructionCost DivCost = MulCost * (Is128bit ? 2 : 1) + // UMULL/UMULH
(HasMULH ? 0 : ShrCost) + // UMULL shift
AddCost * 2 + ShrCost;
return DivCost + (ISD == ISD::UREM ? MulCost + AddCost : 0);
}
}
// div i128's are lowered as libcalls. Pass nullptr as (u)divti3 calls are
// emitted by the backend even when those functions are not declared in the
// module.
if (!VT.isVector() && VT.getSizeInBits() > 64)
return getCallInstrCost(/*Function*/ nullptr, Ty, {Ty, Ty}, CostKind);
InstructionCost Cost = BaseT::getArithmeticInstrCost(
Opcode, Ty, CostKind, Op1Info, Op2Info);
if (Ty->isVectorTy() && (ISD == ISD::SDIV || ISD == ISD::UDIV)) {
if (TLI->isOperationLegalOrCustom(ISD, LT.second) && ST->hasSVE()) {
// SDIV/UDIV operations are lowered using SVE, then we can have less
// costs.
if (isa<FixedVectorType>(Ty) && cast<FixedVectorType>(Ty)
->getPrimitiveSizeInBits()
.getFixedValue() < 128) {
EVT VT = TLI->getValueType(DL, Ty);
static const CostTblEntry DivTbl[]{
{ISD::SDIV, MVT::v2i8, 5}, {ISD::SDIV, MVT::v4i8, 8},
{ISD::SDIV, MVT::v8i8, 8}, {ISD::SDIV, MVT::v2i16, 5},
{ISD::SDIV, MVT::v4i16, 5}, {ISD::SDIV, MVT::v2i32, 1},
{ISD::UDIV, MVT::v2i8, 5}, {ISD::UDIV, MVT::v4i8, 8},
{ISD::UDIV, MVT::v8i8, 8}, {ISD::UDIV, MVT::v2i16, 5},
{ISD::UDIV, MVT::v4i16, 5}, {ISD::UDIV, MVT::v2i32, 1}};
const auto *Entry = CostTableLookup(DivTbl, ISD, VT.getSimpleVT());
if (nullptr != Entry)
return Entry->Cost;
}
// For 8/16-bit elements, the cost is higher because the type
// requires promotion and possibly splitting:
if (LT.second.getScalarType() == MVT::i8)
Cost *= 8;
else if (LT.second.getScalarType() == MVT::i16)
Cost *= 4;
return Cost;
} else {
// If one of the operands is a uniform constant then the cost for each
// element is Cost for insertion, extraction and division.
// Insertion cost = 2, Extraction Cost = 2, Division = cost for the
// operation with scalar type
if ((Op1Info.isConstant() && Op1Info.isUniform()) ||
(Op2Info.isConstant() && Op2Info.isUniform())) {
if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
InstructionCost DivCost = BaseT::getArithmeticInstrCost(
Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info);
return (4 + DivCost) * VTy->getNumElements();
}
}
// On AArch64, without SVE, vector divisions are expanded
// into scalar divisions of each pair of elements.
Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, CostKind,
-1, nullptr, nullptr);
Cost += getVectorInstrCost(Instruction::InsertElement, Ty, CostKind, -1,
nullptr, nullptr);
}
// TODO: if one of the arguments is scalar, then it's not necessary to
// double the cost of handling the vector elements.
Cost += Cost;
}
return Cost;
}
case ISD::MUL:
// When SVE is available, then we can lower the v2i64 operation using
// the SVE mul instruction, which has a lower cost.
if (LT.second == MVT::v2i64 && ST->hasSVE())
return LT.first;
// When SVE is not available, there is no MUL.2d instruction,
// which means mul <2 x i64> is expensive as elements are extracted
// from the vectors and the muls scalarized.
// As getScalarizationOverhead is a bit too pessimistic, we
// estimate the cost for a i64 vector directly here, which is:
// - four 2-cost i64 extracts,
// - two 2-cost i64 inserts, and
// - two 1-cost muls.
// So, for a v2i64 with LT.First = 1 the cost is 14, and for a v4i64 with
// LT.first = 2 the cost is 28. If both operands are extensions it will not
// need to scalarize so the cost can be cheaper (smull or umull).
// so the cost can be cheaper (smull or umull).
if (LT.second != MVT::v2i64 || isWideningInstruction(Ty, Opcode, Args))
return LT.first;
return cast<VectorType>(Ty)->getElementCount().getKnownMinValue() *
(getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind) +
getVectorInstrCost(Instruction::ExtractElement, Ty, CostKind, -1,
nullptr, nullptr) *
2 +
getVectorInstrCost(Instruction::InsertElement, Ty, CostKind, -1,
nullptr, nullptr));
case ISD::ADD:
case ISD::XOR:
case ISD::OR:
case ISD::AND:
case ISD::SRL:
case ISD::SRA:
case ISD::SHL:
// These nodes are marked as 'custom' for combining purposes only.
// We know that they are legal. See LowerAdd in ISelLowering.
return LT.first;
case ISD::FNEG:
// Scalar fmul(fneg) or fneg(fmul) can be converted to fnmul
if ((Ty->isFloatTy() || Ty->isDoubleTy() ||
(Ty->isHalfTy() && ST->hasFullFP16())) &&
CxtI &&
((CxtI->hasOneUse() &&
match(*CxtI->user_begin(), m_FMul(m_Value(), m_Value()))) ||
match(CxtI->getOperand(0), m_FMul(m_Value(), m_Value()))))
return 0;
[[fallthrough]];
case ISD::FADD:
case ISD::FSUB:
// Increase the cost for half and bfloat types if not architecturally
// supported.
if ((Ty->getScalarType()->isHalfTy() && !ST->hasFullFP16()) ||
(Ty->getScalarType()->isBFloatTy() && !ST->hasBF16()))
return 2 * LT.first;
if (!Ty->getScalarType()->isFP128Ty())
return LT.first;
[[fallthrough]];
case ISD::FMUL:
case ISD::FDIV:
// These nodes are marked as 'custom' just to lower them to SVE.
// We know said lowering will incur no additional cost.
if (!Ty->getScalarType()->isFP128Ty())
return 2 * LT.first;
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
Op2Info);
case ISD::FREM:
// Pass nullptr as fmod/fmodf calls are emitted by the backend even when
// those functions are not declared in the module.
if (!Ty->isVectorTy())
return getCallInstrCost(/*Function*/ nullptr, Ty, {Ty, Ty}, CostKind);
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
Op2Info);
}
}
InstructionCost
AArch64TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
const SCEV *Ptr) const {
// 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 = NeonNonConstStrideOverhead;
int MaxMergeDistance = 64;
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;
}
InstructionCost AArch64TTIImpl::getCmpSelInstrCost(
unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info,
TTI::OperandValueInfo Op2Info, const Instruction *I) const {
// TODO: Handle other cost kinds.
if (CostKind != TTI::TCK_RecipThroughput)
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
Op1Info, Op2Info, I);
int ISD = TLI->InstructionOpcodeToISD(Opcode);
// We don't lower some vector selects well that are wider than the register
// width.
if (isa<FixedVectorType>(ValTy) && ISD == ISD::SELECT) {
// We would need this many instructions to hide the scalarization happening.
const int AmortizationCost = 20;
// If VecPred is not set, check if we can get a predicate from the context
// instruction, if its type matches the requested ValTy.
if (VecPred == CmpInst::BAD_ICMP_PREDICATE && I && I->getType() == ValTy) {
CmpPredicate CurrentPred;
if (match(I, m_Select(m_Cmp(CurrentPred, m_Value(), m_Value()), m_Value(),
m_Value())))
VecPred = CurrentPred;
}
// Check if we have a compare/select chain that can be lowered using
// a (F)CMxx & BFI pair.
if (CmpInst::isIntPredicate(VecPred) || VecPred == CmpInst::FCMP_OLE ||
VecPred == CmpInst::FCMP_OLT || VecPred == CmpInst::FCMP_OGT ||
VecPred == CmpInst::FCMP_OGE || VecPred == CmpInst::FCMP_OEQ ||
VecPred == CmpInst::FCMP_UNE) {
static const auto ValidMinMaxTys = {
MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16, MVT::v2i32,
MVT::v4i32, MVT::v2i64, MVT::v2f32, MVT::v4f32, MVT::v2f64};
static const auto ValidFP16MinMaxTys = {MVT::v4f16, MVT::v8f16};
auto LT = getTypeLegalizationCost(ValTy);
if (any_of(ValidMinMaxTys, [&LT](MVT M) { return M == LT.second; }) ||
(ST->hasFullFP16() &&
any_of(ValidFP16MinMaxTys, [&LT](MVT M) { return M == LT.second; })))
return LT.first;
}
static const TypeConversionCostTblEntry
VectorSelectTbl[] = {
{ ISD::SELECT, MVT::v2i1, MVT::v2f32, 2 },
{ ISD::SELECT, MVT::v2i1, MVT::v2f64, 2 },
{ ISD::SELECT, MVT::v4i1, MVT::v4f32, 2 },
{ ISD::SELECT, MVT::v4i1, MVT::v4f16, 2 },
{ ISD::SELECT, MVT::v8i1, MVT::v8f16, 2 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i16, 16 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i32, 8 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i32, 16 },
{ ISD::SELECT, MVT::v4i1, MVT::v4i64, 4 * AmortizationCost },
{ ISD::SELECT, MVT::v8i1, MVT::v8i64, 8 * AmortizationCost },
{ ISD::SELECT, MVT::v16i1, MVT::v16i64, 16 * AmortizationCost }
};
EVT SelCondTy = TLI->getValueType(DL, CondTy);
EVT SelValTy = TLI->getValueType(DL, ValTy);
if (SelCondTy.isSimple() && SelValTy.isSimple()) {
if (const auto *Entry = ConvertCostTableLookup(VectorSelectTbl, ISD,
SelCondTy.getSimpleVT(),
SelValTy.getSimpleVT()))
return Entry->Cost;
}
}
if (isa<FixedVectorType>(ValTy) && ISD == ISD::SETCC) {
Type *ValScalarTy = ValTy->getScalarType();
if ((ValScalarTy->isHalfTy() && !ST->hasFullFP16()) ||
ValScalarTy->isBFloatTy()) {
auto *ValVTy = cast<FixedVectorType>(ValTy);
// FIXME: We currently scalarise these.
if (ValVTy->getNumElements() > 4)
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred,
CostKind, Op1Info, Op2Info, I);
// Without dedicated instructions we promote [b]f16 compares to f32.
auto *PromotedTy =
VectorType::get(Type::getFloatTy(ValTy->getContext()), ValVTy);
InstructionCost Cost = 0;
// Promote operands to float vectors.
Cost += 2 * getCastInstrCost(Instruction::FPExt, PromotedTy, ValTy,
TTI::CastContextHint::None, CostKind);
// Compare float vectors.
Cost += getCmpSelInstrCost(Opcode, PromotedTy, CondTy, VecPred, CostKind,
Op1Info, Op2Info);
// During codegen we'll truncate the vector result from i32 to i16.
Cost +=
getCastInstrCost(Instruction::Trunc, VectorType::getInteger(ValVTy),
VectorType::getInteger(PromotedTy),
TTI::CastContextHint::None, CostKind);
return Cost;
}
}
// Treat the icmp in icmp(and, 0) as free, as we can make use of ands.
// FIXME: This can apply to more conditions and add/sub if it can be shown to
// be profitable.
if (ValTy->isIntegerTy() && ISD == ISD::SETCC && I &&
ICmpInst::isEquality(VecPred) &&
TLI->isTypeLegal(TLI->getValueType(DL, ValTy)) &&
match(I->getOperand(1), m_Zero()) &&
match(I->getOperand(0), m_And(m_Value(), m_Value())))
return 0;
// The base case handles scalable vectors fine for now, since it treats the
// cost as 1 * legalization cost.
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
Op1Info, Op2Info, I);
}
AArch64TTIImpl::TTI::MemCmpExpansionOptions
AArch64TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
TTI::MemCmpExpansionOptions Options;
if (ST->requiresStrictAlign()) {
// TODO: Add cost modeling for strict align. Misaligned loads expand to
// a bunch of instructions when strict align is enabled.
return Options;
}
Options.AllowOverlappingLoads = true;
Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
Options.NumLoadsPerBlock = Options.MaxNumLoads;
// TODO: Though vector loads usually perform well on AArch64, in some targets
// they may wake up the FP unit, which raises the power consumption. Perhaps
// they could be used with no holds barred (-O3).
Options.LoadSizes = {8, 4, 2, 1};
Options.AllowedTailExpansions = {3, 5, 6};
return Options;
}
bool AArch64TTIImpl::prefersVectorizedAddressing() const {
return ST->hasSVE();
}
InstructionCost
AArch64TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
Align Alignment, unsigned AddressSpace,
TTI::TargetCostKind CostKind) const {
if (useNeonVector(Src))
return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
CostKind);
auto LT = getTypeLegalizationCost(Src);
if (!LT.first.isValid())
return InstructionCost::getInvalid();
// Return an invalid cost for element types that we are unable to lower.
auto *VT = cast<VectorType>(Src);
if (VT->getElementType()->isIntegerTy(1))
return InstructionCost::getInvalid();
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
if (VT->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
return LT.first;
}
// This function returns gather/scatter overhead either from
// user-provided value or specialized values per-target from \p ST.
static unsigned getSVEGatherScatterOverhead(unsigned Opcode,
const AArch64Subtarget *ST) {
assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
"Should be called on only load or stores.");
switch (Opcode) {
case Instruction::Load:
if (SVEGatherOverhead.getNumOccurrences() > 0)
return SVEGatherOverhead;
return ST->getGatherOverhead();
break;
case Instruction::Store:
if (SVEScatterOverhead.getNumOccurrences() > 0)
return SVEScatterOverhead;
return ST->getScatterOverhead();
break;
default:
llvm_unreachable("Shouldn't have reached here");
}
}
InstructionCost AArch64TTIImpl::getGatherScatterOpCost(
unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const {
if (useNeonVector(DataTy) || !isLegalMaskedGatherScatter(DataTy))
return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
Alignment, CostKind, I);
auto *VT = cast<VectorType>(DataTy);
auto LT = getTypeLegalizationCost(DataTy);
if (!LT.first.isValid())
return InstructionCost::getInvalid();
// Return an invalid cost for element types that we are unable to lower.
if (!LT.second.isVector() ||
!isElementTypeLegalForScalableVector(VT->getElementType()) ||
VT->getElementType()->isIntegerTy(1))
return InstructionCost::getInvalid();
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
if (VT->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
ElementCount LegalVF = LT.second.getVectorElementCount();
InstructionCost MemOpCost =
getMemoryOpCost(Opcode, VT->getElementType(), Alignment, 0, CostKind,
{TTI::OK_AnyValue, TTI::OP_None}, I);
// Add on an overhead cost for using gathers/scatters.
MemOpCost *= getSVEGatherScatterOverhead(Opcode, ST);
return LT.first * MemOpCost * getMaxNumElements(LegalVF);
}
bool AArch64TTIImpl::useNeonVector(const Type *Ty) const {
return isa<FixedVectorType>(Ty) && !ST->useSVEForFixedLengthVectors();
}
InstructionCost AArch64TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Ty,
Align Alignment,
unsigned AddressSpace,
TTI::TargetCostKind CostKind,
TTI::OperandValueInfo OpInfo,
const Instruction *I) const {
EVT VT = TLI->getValueType(DL, Ty, true);
// Type legalization can't handle structs
if (VT == MVT::Other)
return BaseT::getMemoryOpCost(Opcode, Ty, Alignment, AddressSpace,
CostKind);
auto LT = getTypeLegalizationCost(Ty);
if (!LT.first.isValid())
return InstructionCost::getInvalid();
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
// We also only support full register predicate loads and stores.
if (auto *VTy = dyn_cast<ScalableVectorType>(Ty))
if (VTy->getElementCount() == ElementCount::getScalable(1) ||
(VTy->getElementType()->isIntegerTy(1) &&
!VTy->getElementCount().isKnownMultipleOf(
ElementCount::getScalable(16))))
return InstructionCost::getInvalid();
// TODO: consider latency as well for TCK_SizeAndLatency.
if (CostKind == TTI::TCK_CodeSize || CostKind == TTI::TCK_SizeAndLatency)
return LT.first;
if (CostKind != TTI::TCK_RecipThroughput)
return 1;
if (ST->isMisaligned128StoreSlow() && Opcode == Instruction::Store &&
LT.second.is128BitVector() && Alignment < Align(16)) {
// Unaligned stores are extremely inefficient. We don't split all
// unaligned 128-bit stores because the negative impact that has shown in
// practice on inlined block copy code.
// We make such stores expensive so that we will only vectorize if there
// are 6 other instructions getting vectorized.
const int AmortizationCost = 6;
return LT.first * 2 * AmortizationCost;
}
// Opaque ptr or ptr vector types are i64s and can be lowered to STP/LDPs.
if (Ty->isPtrOrPtrVectorTy())
return LT.first;
if (useNeonVector(Ty)) {
// Check truncating stores and extending loads.
if (Ty->getScalarSizeInBits() != LT.second.getScalarSizeInBits()) {
// v4i8 types are lowered to scalar a load/store and sshll/xtn.
if (VT == MVT::v4i8)
return 2;
// Otherwise we need to scalarize.
return cast<FixedVectorType>(Ty)->getNumElements() * 2;
}
EVT EltVT = VT.getVectorElementType();
unsigned EltSize = EltVT.getScalarSizeInBits();
if (!isPowerOf2_32(EltSize) || EltSize < 8 || EltSize > 64 ||
VT.getVectorNumElements() >= (128 / EltSize) || Alignment != Align(1))
return LT.first;
// FIXME: v3i8 lowering currently is very inefficient, due to automatic
// widening to v4i8, which produces suboptimal results.
if (VT.getVectorNumElements() == 3 && EltVT == MVT::i8)
return LT.first;
// Check non-power-of-2 loads/stores for legal vector element types with
// NEON. Non-power-of-2 memory ops will get broken down to a set of
// operations on smaller power-of-2 ops, including ld1/st1.
LLVMContext &C = Ty->getContext();
InstructionCost Cost(0);
SmallVector<EVT> TypeWorklist;
TypeWorklist.push_back(VT);
while (!TypeWorklist.empty()) {
EVT CurrVT = TypeWorklist.pop_back_val();
unsigned CurrNumElements = CurrVT.getVectorNumElements();
if (isPowerOf2_32(CurrNumElements)) {
Cost += 1;
continue;
}
unsigned PrevPow2 = NextPowerOf2(CurrNumElements) / 2;
TypeWorklist.push_back(EVT::getVectorVT(C, EltVT, PrevPow2));
TypeWorklist.push_back(
EVT::getVectorVT(C, EltVT, CurrNumElements - PrevPow2));
}
return Cost;
}
return LT.first;
}
InstructionCost AArch64TTIImpl::getInterleavedMemoryOpCost(
unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
bool UseMaskForCond, bool UseMaskForGaps) const {
assert(Factor >= 2 && "Invalid interleave factor");
auto *VecVTy = cast<VectorType>(VecTy);
if (VecTy->isScalableTy() && !ST->hasSVE())
return InstructionCost::getInvalid();
// Vectorization for masked interleaved accesses is only enabled for scalable
// VF.
if (!VecTy->isScalableTy() && (UseMaskForCond || UseMaskForGaps))
return InstructionCost::getInvalid();
if (!UseMaskForGaps && Factor <= TLI->getMaxSupportedInterleaveFactor()) {
unsigned MinElts = VecVTy->getElementCount().getKnownMinValue();
auto *SubVecTy =
VectorType::get(VecVTy->getElementType(),
VecVTy->getElementCount().divideCoefficientBy(Factor));
// ldN/stN 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 ldN/stN instruction.
bool UseScalable;
if (MinElts % Factor == 0 &&
TLI->isLegalInterleavedAccessType(SubVecTy, DL, UseScalable))
return Factor * TLI->getNumInterleavedAccesses(SubVecTy, DL, UseScalable);
}
return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
Alignment, AddressSpace, CostKind,
UseMaskForCond, UseMaskForGaps);
}
InstructionCost
AArch64TTIImpl::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const {
InstructionCost Cost = 0;
TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
for (auto *I : Tys) {
if (!I->isVectorTy())
continue;
if (I->getScalarSizeInBits() * cast<FixedVectorType>(I)->getNumElements() ==
128)
Cost += getMemoryOpCost(Instruction::Store, I, Align(128), 0, CostKind) +
getMemoryOpCost(Instruction::Load, I, Align(128), 0, CostKind);
}
return Cost;
}
unsigned AArch64TTIImpl::getMaxInterleaveFactor(ElementCount VF) const {
return ST->getMaxInterleaveFactor();
}
// For Falkor, we want to avoid having too many strided loads in a loop since
// that can exhaust the HW prefetcher resources. We adjust the unroller
// MaxCount preference below to attempt to ensure unrolling doesn't create too
// many strided loads.
static void
getFalkorUnrollingPreferences(Loop *L, ScalarEvolution &SE,
TargetTransformInfo::UnrollingPreferences &UP) {
enum { MaxStridedLoads = 7 };
auto countStridedLoads = [](Loop *L, ScalarEvolution &SE) {
int StridedLoads = 0;
// FIXME? We could make this more precise by looking at the CFG and
// e.g. not counting loads in each side of an if-then-else diamond.
for (const auto BB : L->blocks()) {
for (auto &I : *BB) {
LoadInst *LMemI = dyn_cast<LoadInst>(&I);
if (!LMemI)
continue;
Value *PtrValue = LMemI->getPointerOperand();
if (L->isLoopInvariant(PtrValue))
continue;
const SCEV *LSCEV = SE.getSCEV(PtrValue);
const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
if (!LSCEVAddRec || !LSCEVAddRec->isAffine())
continue;
// FIXME? We could take pairing of unrolled load copies into account
// by looking at the AddRec, but we would probably have to limit this
// to loops with no stores or other memory optimization barriers.
++StridedLoads;
// We've seen enough strided loads that seeing more won't make a
// difference.
if (StridedLoads > MaxStridedLoads / 2)
return StridedLoads;
}
}
return StridedLoads;
};
int StridedLoads = countStridedLoads(L, SE);
LLVM_DEBUG(dbgs() << "falkor-hwpf: detected " << StridedLoads
<< " strided loads\n");
// Pick the largest power of 2 unroll count that won't result in too many
// strided loads.
if (StridedLoads) {
UP.MaxCount = 1 << Log2_32(MaxStridedLoads / StridedLoads);
LLVM_DEBUG(dbgs() << "falkor-hwpf: setting unroll MaxCount to "
<< UP.MaxCount << '\n');
}
}
// This function returns true if the loop:
// 1. Has a valid cost, and
// 2. Has a cost within the supplied budget.
// Otherwise it returns false.
static bool isLoopSizeWithinBudget(Loop *L, const AArch64TTIImpl &TTI,
InstructionCost Budget,
unsigned *FinalSize) {
// Estimate the size of the loop.
InstructionCost LoopCost = 0;
for (auto *BB : L->getBlocks()) {
for (auto &I : *BB) {
SmallVector<const Value *, 4> Operands(I.operand_values());
InstructionCost Cost =
TTI.getInstructionCost(&I, Operands, TTI::TCK_CodeSize);
// This can happen with intrinsics that don't currently have a cost model
// or for some operations that require SVE.
if (!Cost.isValid())
return false;
LoopCost += Cost;
if (LoopCost > Budget)
return false;
}
}
if (FinalSize)
*FinalSize = LoopCost.getValue();
return true;
}
static bool shouldUnrollMultiExitLoop(Loop *L, ScalarEvolution &SE,
const AArch64TTIImpl &TTI) {
// Only consider loops with unknown trip counts for which we can determine
// a symbolic expression. Multi-exit loops with small known trip counts will
// likely be unrolled anyway.
const SCEV *BTC = SE.getSymbolicMaxBackedgeTakenCount(L);
if (isa<SCEVConstant>(BTC) || isa<SCEVCouldNotCompute>(BTC))
return false;
// It might not be worth unrolling loops with low max trip counts. Restrict
// this to max trip counts > 32 for now.
unsigned MaxTC = SE.getSmallConstantMaxTripCount(L);
if (MaxTC > 0 && MaxTC <= 32)
return false;
// Make sure the loop size is <= 5.
if (!isLoopSizeWithinBudget(L, TTI, 5, nullptr))
return false;
// Small search loops with multiple exits can be highly beneficial to unroll.
// We only care about loops with exactly two exiting blocks, although each
// block could jump to the same exit block.
ArrayRef<BasicBlock *> Blocks = L->getBlocks();
if (Blocks.size() != 2)
return false;
if (any_of(Blocks, [](BasicBlock *BB) {
return !isa<BranchInst>(BB->getTerminator());
}))
return false;
return true;
}
/// For Apple CPUs, we want to runtime-unroll loops to make better use if the
/// OOO engine's wide instruction window and various predictors.
static void
getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE,
TargetTransformInfo::UnrollingPreferences &UP,
const AArch64TTIImpl &TTI) {
// Limit loops with structure that is highly likely to benefit from runtime
// unrolling; that is we exclude outer loops and loops with many blocks (i.e.
// likely with complex control flow). Note that the heuristics here may be
// overly conservative and we err on the side of avoiding runtime unrolling
// rather than unroll excessively. They are all subject to further refinement.
if (!L->isInnermost() || L->getNumBlocks() > 8)
return;
// Loops with multiple exits are handled by common code.
if (!L->getExitBlock())
return;
const SCEV *BTC = SE.getSymbolicMaxBackedgeTakenCount(L);
if (isa<SCEVConstant>(BTC) || isa<SCEVCouldNotCompute>(BTC) ||
(SE.getSmallConstantMaxTripCount(L) > 0 &&
SE.getSmallConstantMaxTripCount(L) <= 32))
return;
if (findStringMetadataForLoop(L, "llvm.loop.isvectorized"))
return;
if (SE.getSymbolicMaxBackedgeTakenCount(L) != SE.getBackedgeTakenCount(L))
return;
// Limit to loops with trip counts that are cheap to expand.
UP.SCEVExpansionBudget = 1;
// Try to unroll small, single block loops, if they have load/store
// dependencies, to expose more parallel memory access streams.
BasicBlock *Header = L->getHeader();
if (Header == L->getLoopLatch()) {
// Estimate the size of the loop.
unsigned Size;
if (!isLoopSizeWithinBudget(L, TTI, 8, &Size))
return;
SmallPtrSet<Value *, 8> LoadedValues;
SmallVector<StoreInst *> Stores;
for (auto *BB : L->blocks()) {
for (auto &I : *BB) {
Value *Ptr = getLoadStorePointerOperand(&I);
if (!Ptr)
continue;
const SCEV *PtrSCEV = SE.getSCEV(Ptr);
if (SE.isLoopInvariant(PtrSCEV, L))
continue;
if (isa<LoadInst>(&I))
LoadedValues.insert(&I);
else
Stores.push_back(cast<StoreInst>(&I));
}
}
// Try to find an unroll count that maximizes the use of the instruction
// window, i.e. trying to fetch as many instructions per cycle as possible.
unsigned MaxInstsPerLine = 16;
unsigned UC = 1;
unsigned BestUC = 1;
unsigned SizeWithBestUC = BestUC * Size;
while (UC <= 8) {
unsigned SizeWithUC = UC * Size;
if (SizeWithUC > 48)
break;
if ((SizeWithUC % MaxInstsPerLine) == 0 ||
(SizeWithBestUC % MaxInstsPerLine) < (SizeWithUC % MaxInstsPerLine)) {
BestUC = UC;
SizeWithBestUC = BestUC * Size;
}
UC++;
}
if (BestUC == 1 || none_of(Stores, [&LoadedValues](StoreInst *SI) {
return LoadedValues.contains(SI->getOperand(0));
}))
return;
UP.Runtime = true;
UP.DefaultUnrollRuntimeCount = BestUC;
return;
}
// Try to runtime-unroll loops with early-continues depending on loop-varying
// loads; this helps with branch-prediction for the early-continues.
auto *Term = dyn_cast<BranchInst>(Header->getTerminator());
auto *Latch = L->getLoopLatch();
SmallVector<BasicBlock *> Preds(predecessors(Latch));
if (!Term || !Term->isConditional() || Preds.size() == 1 ||
none_of(Preds, [Header](BasicBlock *Pred) { return Header == Pred; }) ||
none_of(Preds, [L](BasicBlock *Pred) { return L->contains(Pred); }))
return;
std::function<bool(Instruction *, unsigned)> DependsOnLoopLoad =
[&](Instruction *I, unsigned Depth) -> bool {
if (isa<PHINode>(I) || L->isLoopInvariant(I) || Depth > 8)
return false;
if (isa<LoadInst>(I))
return true;
return any_of(I->operands(), [&](Value *V) {
auto *I = dyn_cast<Instruction>(V);
return I && DependsOnLoopLoad(I, Depth + 1);
});
};
CmpPredicate Pred;
Instruction *I;
if (match(Term, m_Br(m_ICmp(Pred, m_Instruction(I), m_Value()), m_Value(),
m_Value())) &&
DependsOnLoopLoad(I, 0)) {
UP.Runtime = true;
}
}
void AArch64TTIImpl::getUnrollingPreferences(
Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP,
OptimizationRemarkEmitter *ORE) const {
// Enable partial unrolling and runtime unrolling.
BaseT::getUnrollingPreferences(L, SE, UP, ORE);
UP.UpperBound = true;
// For inner loop, it is more likely to be a hot one, and the runtime check
// can be promoted out from LICM pass, so the overhead is less, let's try
// a larger threshold to unroll more loops.
if (L->getLoopDepth() > 1)
UP.PartialThreshold *= 2;
// Disable partial & runtime unrolling on -Os.
UP.PartialOptSizeThreshold = 0;
// Scan the loop: don't unroll loops with calls as this could prevent
// inlining. Don't unroll vector loops either, as they don't benefit much from
// unrolling.
for (auto *BB : L->getBlocks()) {
for (auto &I : *BB) {
// Don't unroll vectorised loop.
if (I.getType()->isVectorTy())
return;
if (isa<CallBase>(I)) {
if (isa<CallInst>(I) || isa<InvokeInst>(I))
if (const Function *F = cast<CallBase>(I).getCalledFunction())
if (!isLoweredToCall(F))
continue;
return;
}
}
}
// Apply subtarget-specific unrolling preferences.
switch (ST->getProcFamily()) {
case AArch64Subtarget::AppleA14:
case AArch64Subtarget::AppleA15:
case AArch64Subtarget::AppleA16:
case AArch64Subtarget::AppleM4:
getAppleRuntimeUnrollPreferences(L, SE, UP, *this);
break;
case AArch64Subtarget::Falkor:
if (EnableFalkorHWPFUnrollFix)
getFalkorUnrollingPreferences(L, SE, UP);
break;
default:
break;
}
// If this is a small, multi-exit loop similar to something like std::find,
// then there is typically a performance improvement achieved by unrolling.
if (!L->getExitBlock() && shouldUnrollMultiExitLoop(L, SE, *this)) {
UP.RuntimeUnrollMultiExit = true;
UP.Runtime = true;
// Limit unroll count.
UP.DefaultUnrollRuntimeCount = 4;
// Allow slightly more costly trip-count expansion to catch search loops
// with pointer inductions.
UP.SCEVExpansionBudget = 5;
return;
}
// Enable runtime unrolling for in-order models
// If mcpu is omitted, getProcFamily() returns AArch64Subtarget::Others, so by
// checking for that case, we can ensure that the default behaviour is
// unchanged
if (ST->getProcFamily() != AArch64Subtarget::Generic &&
!ST->getSchedModel().isOutOfOrder()) {
UP.Runtime = true;
UP.Partial = true;
UP.UnrollRemainder = true;
UP.DefaultUnrollRuntimeCount = 4;
UP.UnrollAndJam = true;
UP.UnrollAndJamInnerLoopThreshold = 60;
}
}
void AArch64TTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
TTI::PeelingPreferences &PP) const {
BaseT::getPeelingPreferences(L, SE, PP);
}
Value *
AArch64TTIImpl::getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
Type *ExpectedType) const {
switch (Inst->getIntrinsicID()) {
default:
return nullptr;
case Intrinsic::aarch64_neon_st2:
case Intrinsic::aarch64_neon_st3:
case Intrinsic::aarch64_neon_st4: {
// Create a struct type
StructType *ST = dyn_cast<StructType>(ExpectedType);
if (!ST)
return nullptr;
unsigned NumElts = Inst->arg_size() - 1;
if (ST->getNumElements() != NumElts)
return nullptr;
for (unsigned i = 0, e = NumElts; i != e; ++i) {
if (Inst->getArgOperand(i)->getType() != ST->getElementType(i))
return nullptr;
}
Value *Res = PoisonValue::get(ExpectedType);
IRBuilder<> Builder(Inst);
for (unsigned i = 0, e = NumElts; i != e; ++i) {
Value *L = Inst->getArgOperand(i);
Res = Builder.CreateInsertValue(Res, L, i);
}
return Res;
}
case Intrinsic::aarch64_neon_ld2:
case Intrinsic::aarch64_neon_ld3:
case Intrinsic::aarch64_neon_ld4:
if (Inst->getType() == ExpectedType)
return Inst;
return nullptr;
}
}
bool AArch64TTIImpl::getTgtMemIntrinsic(IntrinsicInst *Inst,
MemIntrinsicInfo &Info) const {
switch (Inst->getIntrinsicID()) {
default:
break;
case Intrinsic::aarch64_neon_ld2:
case Intrinsic::aarch64_neon_ld3:
case Intrinsic::aarch64_neon_ld4:
Info.ReadMem = true;
Info.WriteMem = false;
Info.PtrVal = Inst->getArgOperand(0);
break;
case Intrinsic::aarch64_neon_st2:
case Intrinsic::aarch64_neon_st3:
case Intrinsic::aarch64_neon_st4:
Info.ReadMem = false;
Info.WriteMem = true;
Info.PtrVal = Inst->getArgOperand(Inst->arg_size() - 1);
break;
}
switch (Inst->getIntrinsicID()) {
default:
return false;
case Intrinsic::aarch64_neon_ld2:
case Intrinsic::aarch64_neon_st2:
Info.MatchingId = VECTOR_LDST_TWO_ELEMENTS;
break;
case Intrinsic::aarch64_neon_ld3:
case Intrinsic::aarch64_neon_st3:
Info.MatchingId = VECTOR_LDST_THREE_ELEMENTS;
break;
case Intrinsic::aarch64_neon_ld4:
case Intrinsic::aarch64_neon_st4:
Info.MatchingId = VECTOR_LDST_FOUR_ELEMENTS;
break;
}
return true;
}
/// See if \p I should be considered for address type promotion. We check if \p
/// I is a sext with right type and used in memory accesses. If it used in a
/// "complex" getelementptr, we allow it to be promoted without finding other
/// sext instructions that sign extended the same initial value. A getelementptr
/// is considered as "complex" if it has more than 2 operands.
bool AArch64TTIImpl::shouldConsiderAddressTypePromotion(
const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const {
bool Considerable = false;
AllowPromotionWithoutCommonHeader = false;
if (!isa<SExtInst>(&I))
return false;
Type *ConsideredSExtType =
Type::getInt64Ty(I.getParent()->getParent()->getContext());
if (I.getType() != ConsideredSExtType)
return false;
// See if the sext is the one with the right type and used in at least one
// GetElementPtrInst.
for (const User *U : I.users()) {
if (const GetElementPtrInst *GEPInst = dyn_cast<GetElementPtrInst>(U)) {
Considerable = true;
// A getelementptr is considered as "complex" if it has more than 2
// operands. We will promote a SExt used in such complex GEP as we
// expect some computation to be merged if they are done on 64 bits.
if (GEPInst->getNumOperands() > 2) {
AllowPromotionWithoutCommonHeader = true;
break;
}
}
}
return Considerable;
}
bool AArch64TTIImpl::isLegalToVectorizeReduction(
const RecurrenceDescriptor &RdxDesc, ElementCount VF) const {
if (!VF.isScalable())
return true;
Type *Ty = RdxDesc.getRecurrenceType();
if (Ty->isBFloatTy() || !isElementTypeLegalForScalableVector(Ty))
return false;
switch (RdxDesc.getRecurrenceKind()) {
case RecurKind::Add:
case RecurKind::FAdd:
case RecurKind::And:
case RecurKind::Or:
case RecurKind::Xor:
case RecurKind::SMin:
case RecurKind::SMax:
case RecurKind::UMin:
case RecurKind::UMax:
case RecurKind::FMin:
case RecurKind::FMax:
case RecurKind::FMulAdd:
case RecurKind::IAnyOf:
case RecurKind::FAnyOf:
return true;
default:
return false;
}
}
InstructionCost
AArch64TTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty,
FastMathFlags FMF,
TTI::TargetCostKind CostKind) const {
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
if (auto *VTy = dyn_cast<ScalableVectorType>(Ty))
if (VTy->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
if (LT.second.getScalarType() == MVT::f16 && !ST->hasFullFP16())
return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind);
InstructionCost LegalizationCost = 0;
if (LT.first > 1) {
Type *LegalVTy = EVT(LT.second).getTypeForEVT(Ty->getContext());
IntrinsicCostAttributes Attrs(IID, LegalVTy, {LegalVTy, LegalVTy}, FMF);
LegalizationCost = getIntrinsicInstrCost(Attrs, CostKind) * (LT.first - 1);
}
return LegalizationCost + /*Cost of horizontal reduction*/ 2;
}
InstructionCost AArch64TTIImpl::getArithmeticReductionCostSVE(
unsigned Opcode, VectorType *ValTy, TTI::TargetCostKind CostKind) const {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
InstructionCost LegalizationCost = 0;
if (LT.first > 1) {
Type *LegalVTy = EVT(LT.second).getTypeForEVT(ValTy->getContext());
LegalizationCost = getArithmeticInstrCost(Opcode, LegalVTy, CostKind);
LegalizationCost *= LT.first - 1;
}
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
// Add the final reduction cost for the legal horizontal reduction
switch (ISD) {
case ISD::ADD:
case ISD::AND:
case ISD::OR:
case ISD::XOR:
case ISD::FADD:
return LegalizationCost + 2;
default:
return InstructionCost::getInvalid();
}
}
InstructionCost
AArch64TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
std::optional<FastMathFlags> FMF,
TTI::TargetCostKind CostKind) const {
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
if (auto *VTy = dyn_cast<ScalableVectorType>(ValTy))
if (VTy->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
if (TTI::requiresOrderedReduction(FMF)) {
if (auto *FixedVTy = dyn_cast<FixedVectorType>(ValTy)) {
InstructionCost BaseCost =
BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
// Add on extra cost to reflect the extra overhead on some CPUs. We still
// end up vectorizing for more computationally intensive loops.
return BaseCost + FixedVTy->getNumElements();
}
if (Opcode != Instruction::FAdd)
return InstructionCost::getInvalid();
auto *VTy = cast<ScalableVectorType>(ValTy);
InstructionCost Cost =
getArithmeticInstrCost(Opcode, VTy->getScalarType(), CostKind);
Cost *= getMaxNumElements(VTy->getElementCount());
return Cost;
}
if (isa<ScalableVectorType>(ValTy))
return getArithmeticReductionCostSVE(Opcode, ValTy, CostKind);
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
MVT MTy = LT.second;
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
// Horizontal adds can use the 'addv' instruction. We model the cost of these
// instructions as twice a normal vector add, plus 1 for each legalization
// step (LT.first). This is the only arithmetic vector reduction operation for
// which we have an instruction.
// OR, XOR and AND costs should match the codegen from:
// OR: llvm/test/CodeGen/AArch64/reduce-or.ll
// XOR: llvm/test/CodeGen/AArch64/reduce-xor.ll
// AND: llvm/test/CodeGen/AArch64/reduce-and.ll
static const CostTblEntry CostTblNoPairwise[]{
{ISD::ADD, MVT::v8i8, 2},
{ISD::ADD, MVT::v16i8, 2},
{ISD::ADD, MVT::v4i16, 2},
{ISD::ADD, MVT::v8i16, 2},
{ISD::ADD, MVT::v4i32, 2},
{ISD::ADD, MVT::v2i64, 2},
{ISD::OR, MVT::v8i8, 15},
{ISD::OR, MVT::v16i8, 17},
{ISD::OR, MVT::v4i16, 7},
{ISD::OR, MVT::v8i16, 9},
{ISD::OR, MVT::v2i32, 3},
{ISD::OR, MVT::v4i32, 5},
{ISD::OR, MVT::v2i64, 3},
{ISD::XOR, MVT::v8i8, 15},
{ISD::XOR, MVT::v16i8, 17},
{ISD::XOR, MVT::v4i16, 7},
{ISD::XOR, MVT::v8i16, 9},
{ISD::XOR, MVT::v2i32, 3},
{ISD::XOR, MVT::v4i32, 5},
{ISD::XOR, MVT::v2i64, 3},
{ISD::AND, MVT::v8i8, 15},
{ISD::AND, MVT::v16i8, 17},
{ISD::AND, MVT::v4i16, 7},
{ISD::AND, MVT::v8i16, 9},
{ISD::AND, MVT::v2i32, 3},
{ISD::AND, MVT::v4i32, 5},
{ISD::AND, MVT::v2i64, 3},
};
switch (ISD) {
default:
break;
case ISD::FADD:
if (Type *EltTy = ValTy->getScalarType();
// FIXME: For half types without fullfp16 support, this could extend and
// use a fp32 faddp reduction but current codegen unrolls.
MTy.isVector() && (EltTy->isFloatTy() || EltTy->isDoubleTy() ||
(EltTy->isHalfTy() && ST->hasFullFP16()))) {
const unsigned NElts = MTy.getVectorNumElements();
if (ValTy->getElementCount().getFixedValue() >= 2 && NElts >= 2 &&
isPowerOf2_32(NElts))
// Reduction corresponding to series of fadd instructions is lowered to
// series of faddp instructions. faddp has latency/throughput that
// matches fadd instruction and hence, every faddp instruction can be
// considered to have a relative cost = 1 with
// CostKind = TCK_RecipThroughput.
// An faddp will pairwise add vector elements, so the size of input
// vector reduces by half every time, requiring
// #(faddp instructions) = log2_32(NElts).
return (LT.first - 1) + /*No of faddp instructions*/ Log2_32(NElts);
}
break;
case ISD::ADD:
if (const auto *Entry = CostTableLookup(CostTblNoPairwise, ISD, MTy))
return (LT.first - 1) + Entry->Cost;
break;
case ISD::XOR:
case ISD::AND:
case ISD::OR:
const auto *Entry = CostTableLookup(CostTblNoPairwise, ISD, MTy);
if (!Entry)
break;
auto *ValVTy = cast<FixedVectorType>(ValTy);
if (MTy.getVectorNumElements() <= ValVTy->getNumElements() &&
isPowerOf2_32(ValVTy->getNumElements())) {
InstructionCost ExtraCost = 0;
if (LT.first != 1) {
// Type needs to be split, so there is an extra cost of LT.first - 1
// arithmetic ops.
auto *Ty = FixedVectorType::get(ValTy->getElementType(),
MTy.getVectorNumElements());
ExtraCost = getArithmeticInstrCost(Opcode, Ty, CostKind);
ExtraCost *= LT.first - 1;
}
// All and/or/xor of i1 will be lowered with maxv/minv/addv + fmov
auto Cost = ValVTy->getElementType()->isIntegerTy(1) ? 2 : Entry->Cost;
return Cost + ExtraCost;
}
break;
}
return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
}
InstructionCost AArch64TTIImpl::getExtendedReductionCost(
unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *VecTy,
std::optional<FastMathFlags> FMF, TTI::TargetCostKind CostKind) const {
EVT VecVT = TLI->getValueType(DL, VecTy);
EVT ResVT = TLI->getValueType(DL, ResTy);
if (Opcode == Instruction::Add && VecVT.isSimple() && ResVT.isSimple() &&
VecVT.getSizeInBits() >= 64) {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(VecTy);
// The legal cases are:
// UADDLV 8/16/32->32
// UADDLP 32->64
unsigned RevVTSize = ResVT.getSizeInBits();
if (((LT.second == MVT::v8i8 || LT.second == MVT::v16i8) &&
RevVTSize <= 32) ||
((LT.second == MVT::v4i16 || LT.second == MVT::v8i16) &&
RevVTSize <= 32) ||
((LT.second == MVT::v2i32 || LT.second == MVT::v4i32) &&
RevVTSize <= 64))
return (LT.first - 1) * 2 + 2;
}
return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, VecTy, FMF,
CostKind);
}
InstructionCost
AArch64TTIImpl::getMulAccReductionCost(bool IsUnsigned, Type *ResTy,
VectorType *VecTy,
TTI::TargetCostKind CostKind) const {
EVT VecVT = TLI->getValueType(DL, VecTy);
EVT ResVT = TLI->getValueType(DL, ResTy);
if (ST->hasDotProd() && VecVT.isSimple() && ResVT.isSimple()) {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(VecTy);
// The legal cases with dotprod are
// UDOT 8->32
// Which requires an additional uaddv to sum the i32 values.
if ((LT.second == MVT::v8i8 || LT.second == MVT::v16i8) &&
ResVT == MVT::i32)
return LT.first + 2;
}
return BaseT::getMulAccReductionCost(IsUnsigned, ResTy, VecTy, CostKind);
}
InstructionCost
AArch64TTIImpl::getSpliceCost(VectorType *Tp, int Index,
TTI::TargetCostKind CostKind) const {
static const CostTblEntry ShuffleTbl[] = {
{ TTI::SK_Splice, MVT::nxv16i8, 1 },
{ TTI::SK_Splice, MVT::nxv8i16, 1 },
{ TTI::SK_Splice, MVT::nxv4i32, 1 },
{ TTI::SK_Splice, MVT::nxv2i64, 1 },
{ TTI::SK_Splice, MVT::nxv2f16, 1 },
{ TTI::SK_Splice, MVT::nxv4f16, 1 },
{ TTI::SK_Splice, MVT::nxv8f16, 1 },
{ TTI::SK_Splice, MVT::nxv2bf16, 1 },
{ TTI::SK_Splice, MVT::nxv4bf16, 1 },
{ TTI::SK_Splice, MVT::nxv8bf16, 1 },
{ TTI::SK_Splice, MVT::nxv2f32, 1 },
{ TTI::SK_Splice, MVT::nxv4f32, 1 },
{ TTI::SK_Splice, MVT::nxv2f64, 1 },
};
// The code-generator is currently not able to handle scalable vectors
// of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
// it. This change will be removed when code-generation for these types is
// sufficiently reliable.
if (Tp->getElementCount() == ElementCount::getScalable(1))
return InstructionCost::getInvalid();
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
Type *LegalVTy = EVT(LT.second).getTypeForEVT(Tp->getContext());
EVT PromotedVT = LT.second.getScalarType() == MVT::i1
? TLI->getPromotedVTForPredicate(EVT(LT.second))
: LT.second;
Type *PromotedVTy = EVT(PromotedVT).getTypeForEVT(Tp->getContext());
InstructionCost LegalizationCost = 0;
if (Index < 0) {
LegalizationCost =
getCmpSelInstrCost(Instruction::ICmp, PromotedVTy, PromotedVTy,
CmpInst::BAD_ICMP_PREDICATE, CostKind) +
getCmpSelInstrCost(Instruction::Select, PromotedVTy, LegalVTy,
CmpInst::BAD_ICMP_PREDICATE, CostKind);
}
// Predicated splice are promoted when lowering. See AArch64ISelLowering.cpp
// Cost performed on a promoted type.
if (LT.second.getScalarType() == MVT::i1) {
LegalizationCost +=
getCastInstrCost(Instruction::ZExt, PromotedVTy, LegalVTy,
TTI::CastContextHint::None, CostKind) +
getCastInstrCost(Instruction::Trunc, LegalVTy, PromotedVTy,
TTI::CastContextHint::None, CostKind);
}
const auto *Entry =
CostTableLookup(ShuffleTbl, TTI::SK_Splice, PromotedVT.getSimpleVT());
assert(Entry && "Illegal Type for Splice");
LegalizationCost += Entry->Cost;
return LegalizationCost * LT.first;
}
InstructionCost AArch64TTIImpl::getPartialReductionCost(
unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,
ElementCount VF, TTI::PartialReductionExtendKind OpAExtend,
TTI::PartialReductionExtendKind OpBExtend,
std::optional<unsigned> BinOp) const {
InstructionCost Invalid = InstructionCost::getInvalid();
InstructionCost Cost(TTI::TCC_Basic);
// Sub opcodes currently only occur in chained cases.
// Independent partial reduction subtractions are still costed as an add
if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
return Invalid;
if (InputTypeA != InputTypeB)
return Invalid;
EVT InputEVT = EVT::getEVT(InputTypeA);
EVT AccumEVT = EVT::getEVT(AccumType);
unsigned VFMinValue = VF.getKnownMinValue();
if (VF.isScalable()) {
if (!ST->isSVEorStreamingSVEAvailable())
return Invalid;
// Don't accept a partial reduction if the scaled accumulator is vscale x 1,
// since we can't lower that type.
unsigned Scale =
AccumEVT.getScalarSizeInBits() / InputEVT.getScalarSizeInBits();
if (VFMinValue == Scale)
return Invalid;
}
if (VF.isFixed() &&
(!ST->isNeonAvailable() || !ST->hasDotProd() || AccumEVT == MVT::i64))
return Invalid;
if (InputEVT == MVT::i8) {
switch (VFMinValue) {
default:
return Invalid;
case 8:
if (AccumEVT == MVT::i32)
Cost *= 2;
else if (AccumEVT != MVT::i64)
return Invalid;
break;
case 16:
if (AccumEVT == MVT::i64)
Cost *= 2;
else if (AccumEVT != MVT::i32)
return Invalid;
break;
}
} else if (InputEVT == MVT::i16) {
// FIXME: Allow i32 accumulator but increase cost, as we would extend
// it to i64.
if (VFMinValue != 8 || AccumEVT != MVT::i64)
return Invalid;
} else
return Invalid;
// AArch64 supports lowering mixed extensions to a usdot but only if the
// i8mm or sve/streaming features are available.
if (OpAExtend == TTI::PR_None || OpBExtend == TTI::PR_None ||
(OpAExtend != OpBExtend && !ST->hasMatMulInt8() &&
!ST->isSVEorStreamingSVEAvailable()))
return Invalid;
if (!BinOp || *BinOp != Instruction::Mul)
return Invalid;
return Cost;
}
InstructionCost AArch64TTIImpl::getShuffleCost(
TTI::ShuffleKind Kind, VectorType *Tp, ArrayRef<int> Mask,
TTI::TargetCostKind CostKind, int Index, VectorType *SubTp,
ArrayRef<const Value *> Args, const Instruction *CxtI) const {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
// If we have a Mask, and the LT is being legalized somehow, split the Mask
// into smaller vectors and sum the cost of each shuffle.
if (!Mask.empty() && isa<FixedVectorType>(Tp) && LT.second.isVector() &&
Tp->getScalarSizeInBits() == LT.second.getScalarSizeInBits() &&
Mask.size() > LT.second.getVectorNumElements() && !Index && !SubTp) {
// Check for LD3/LD4 instructions, which are represented in llvm IR as
// deinterleaving-shuffle(load). The shuffle cost could potentially be free,
// but we model it with a cost of LT.first so that LD3/LD4 have a higher
// cost than just the load.
if (Args.size() >= 1 && isa<LoadInst>(Args[0]) &&
(ShuffleVectorInst::isDeInterleaveMaskOfFactor(Mask, 3) ||
ShuffleVectorInst::isDeInterleaveMaskOfFactor(Mask, 4)))
return std::max<InstructionCost>(1, LT.first / 4);
// Check for ST3/ST4 instructions, which are represented in llvm IR as
// store(interleaving-shuffle). The shuffle cost could potentially be free,
// but we model it with a cost of LT.first so that ST3/ST4 have a higher
// cost than just the store.
if (CxtI && CxtI->hasOneUse() && isa<StoreInst>(*CxtI->user_begin()) &&
(ShuffleVectorInst::isInterleaveMask(
Mask, 4, Tp->getElementCount().getKnownMinValue() * 2) ||
ShuffleVectorInst::isInterleaveMask(
Mask, 3, Tp->getElementCount().getKnownMinValue() * 2)))
return LT.first;
unsigned TpNumElts = Mask.size();
unsigned LTNumElts = LT.second.getVectorNumElements();
unsigned NumVecs = (TpNumElts + LTNumElts - 1) / LTNumElts;
VectorType *NTp =
VectorType::get(Tp->getScalarType(), LT.second.getVectorElementCount());
InstructionCost Cost;
for (unsigned N = 0; N < NumVecs; N++) {
SmallVector<int> NMask;
// Split the existing mask into chunks of size LTNumElts. Track the source
// sub-vectors to ensure the result has at most 2 inputs.
unsigned Source1, Source2;
unsigned NumSources = 0;
for (unsigned E = 0; E < LTNumElts; E++) {
int MaskElt = (N * LTNumElts + E < TpNumElts) ? Mask[N * LTNumElts + E]
: PoisonMaskElem;
if (MaskElt < 0) {
NMask.push_back(PoisonMaskElem);
continue;
}
// Calculate which source from the input this comes from and whether it
// is new to us.
unsigned Source = MaskElt / LTNumElts;
if (NumSources == 0) {
Source1 = Source;
NumSources = 1;
} else if (NumSources == 1 && Source != Source1) {
Source2 = Source;
NumSources = 2;
} else if (NumSources >= 2 && Source != Source1 && Source != Source2) {
NumSources++;
}
// Add to the new mask. For the NumSources>2 case these are not correct,
// but are only used for the modular lane number.
if (Source == Source1)
NMask.push_back(MaskElt % LTNumElts);
else if (Source == Source2)
NMask.push_back(MaskElt % LTNumElts + LTNumElts);
else
NMask.push_back(MaskElt % LTNumElts);
}
// If the sub-mask has at most 2 input sub-vectors then re-cost it using
// getShuffleCost. If not then cost it using the worst case as the number
// of element moves into a new vector.
if (NumSources <= 2)
Cost += getShuffleCost(NumSources <= 1 ? TTI::SK_PermuteSingleSrc
: TTI::SK_PermuteTwoSrc,
NTp, NMask, CostKind, 0, nullptr, Args, CxtI);
else
Cost += LTNumElts;
}
return Cost;
}
Kind = improveShuffleKindFromMask(Kind, Mask, Tp, Index, SubTp);
bool IsExtractSubvector = Kind == TTI::SK_ExtractSubvector;
// A subvector extract can be implemented with an ext (or trivial extract, if
// from lane 0). This currently only handles low or high extracts to prevent
// SLP vectorizer regressions.
if (IsExtractSubvector && LT.second.isFixedLengthVector()) {
if (LT.second.is128BitVector() &&
cast<FixedVectorType>(SubTp)->getNumElements() ==
LT.second.getVectorNumElements() / 2) {
if (Index == 0)
return 0;
if (Index == (int)LT.second.getVectorNumElements() / 2)
return 1;
}
Kind = TTI::SK_PermuteSingleSrc;
}
// Check for broadcast loads, which are supported by the LD1R instruction.
// In terms of code-size, the shuffle vector is free when a load + dup get
// folded into a LD1R. That's what we check and return here. For performance
// and reciprocal throughput, a LD1R is not completely free. In this case, we
// return the cost for the broadcast below (i.e. 1 for most/all types), so
// that we model the load + dup sequence slightly higher because LD1R is a
// high latency instruction.
if (CostKind == TTI::TCK_CodeSize && Kind == TTI::SK_Broadcast) {
bool IsLoad = !Args.empty() && isa<LoadInst>(Args[0]);
if (IsLoad && LT.second.isVector() &&
isLegalBroadcastLoad(Tp->getElementType(),
LT.second.getVectorElementCount()))
return 0;
}
// If we have 4 elements for the shuffle and a Mask, get the cost straight
// from the perfect shuffle tables.
if (Mask.size() == 4 && Tp->getElementCount() == ElementCount::getFixed(4) &&
(Tp->getScalarSizeInBits() == 16 || Tp->getScalarSizeInBits() == 32) &&
all_of(Mask, [](int E) { return E < 8; }))
return getPerfectShuffleCost(Mask);
// Check for identity masks, which we can treat as free.
if (!Mask.empty() && LT.second.isFixedLengthVector() &&
(Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc) &&
all_of(enumerate(Mask), [](const auto &M) {
return M.value() < 0 || M.value() == (int)M.index();
}))
return 0;
// Check for other shuffles that are not SK_ kinds but we have native
// instructions for, for example ZIP and UZP.
unsigned Unused;
if (LT.second.isFixedLengthVector() &&
LT.second.getVectorNumElements() == Mask.size() &&
(Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc) &&
(isZIPMask(Mask, LT.second.getVectorNumElements(), Unused) ||
isUZPMask(Mask, LT.second.getVectorNumElements(), Unused) ||
isREVMask(Mask, LT.second.getScalarSizeInBits(),
LT.second.getVectorNumElements(), 16) ||
isREVMask(Mask, LT.second.getScalarSizeInBits(),
LT.second.getVectorNumElements(), 32) ||
isREVMask(Mask, LT.second.getScalarSizeInBits(),
LT.second.getVectorNumElements(), 64) ||
// Check for non-zero lane splats
all_of(drop_begin(Mask),
[&Mask](int M) { return M < 0 || M == Mask[0]; })))
return 1;
if (Kind == TTI::SK_Broadcast || Kind == TTI::SK_Transpose ||
Kind == TTI::SK_Select || Kind == TTI::SK_PermuteSingleSrc ||
Kind == TTI::SK_Reverse || Kind == TTI::SK_Splice) {
static const CostTblEntry ShuffleTbl[] = {
// Broadcast shuffle kinds can be performed with 'dup'.
{TTI::SK_Broadcast, MVT::v8i8, 1},
{TTI::SK_Broadcast, MVT::v16i8, 1},
{TTI::SK_Broadcast, MVT::v4i16, 1},
{TTI::SK_Broadcast, MVT::v8i16, 1},
{TTI::SK_Broadcast, MVT::v2i32, 1},
{TTI::SK_Broadcast, MVT::v4i32, 1},
{TTI::SK_Broadcast, MVT::v2i64, 1},
{TTI::SK_Broadcast, MVT::v4f16, 1},
{TTI::SK_Broadcast, MVT::v8f16, 1},
{TTI::SK_Broadcast, MVT::v2f32, 1},
{TTI::SK_Broadcast, MVT::v4f32, 1},
{TTI::SK_Broadcast, MVT::v2f64, 1},
// Transpose shuffle kinds can be performed with 'trn1/trn2' and
// 'zip1/zip2' instructions.
{TTI::SK_Transpose, MVT::v8i8, 1},
{TTI::SK_Transpose, MVT::v16i8, 1},
{TTI::SK_Transpose, MVT::v4i16, 1},
{TTI::SK_Transpose, MVT::v8i16, 1},
{TTI::SK_Transpose, MVT::v2i32, 1},
{TTI::SK_Transpose, MVT::v4i32, 1},
{TTI::SK_Transpose, MVT::v2i64, 1},
{TTI::SK_Transpose, MVT::v4f16, 1},
{TTI::SK_Transpose, MVT::v8f16, 1},
{TTI::SK_Transpose, MVT::v2f32, 1},
{TTI::SK_Transpose, MVT::v4f32, 1},
{TTI::SK_Transpose, MVT::v2f64, 1},
// Select shuffle kinds.
// TODO: handle vXi8/vXi16.
{TTI::SK_Select, MVT::v2i32, 1}, // mov.
{TTI::SK_Select, MVT::v4i32, 2}, // rev+trn (or similar).
{TTI::SK_Select, MVT::v2i64, 1}, // mov.
{TTI::SK_Select, MVT::v2f32, 1}, // mov.
{TTI::SK_Select, MVT::v4f32, 2}, // rev+trn (or similar).
{TTI::SK_Select, MVT::v2f64, 1}, // mov.
// PermuteSingleSrc shuffle kinds.
{TTI::SK_PermuteSingleSrc, MVT::v2i32, 1}, // mov.
{TTI::SK_PermuteSingleSrc, MVT::v4i32, 3}, // perfectshuffle worst case.
{TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // mov.
{TTI::SK_PermuteSingleSrc, MVT::v2f32, 1}, // mov.
{TTI::SK_PermuteSingleSrc, MVT::v4f32, 3}, // perfectshuffle worst case.
{TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // mov.
{TTI::SK_PermuteSingleSrc, MVT::v4i16, 3}, // perfectshuffle worst case.
{TTI::SK_PermuteSingleSrc, MVT::v4f16, 3}, // perfectshuffle worst case.
{TTI::SK_PermuteSingleSrc, MVT::v4bf16, 3}, // same
{TTI::SK_PermuteSingleSrc, MVT::v8i16, 8}, // constpool + load + tbl
{TTI::SK_PermuteSingleSrc, MVT::v8f16, 8}, // constpool + load + tbl
{TTI::SK_PermuteSingleSrc, MVT::v8bf16, 8}, // constpool + load + tbl
{TTI::SK_PermuteSingleSrc, MVT::v8i8, 8}, // constpool + load + tbl
{TTI::SK_PermuteSingleSrc, MVT::v16i8, 8}, // constpool + load + tbl
// Reverse can be lowered with `rev`.
{TTI::SK_Reverse, MVT::v2i32, 1}, // REV64
{TTI::SK_Reverse, MVT::v4i32, 2}, // REV64; EXT
{TTI::SK_Reverse, MVT::v2i64, 1}, // EXT
{TTI::SK_Reverse, MVT::v2f32, 1}, // REV64
{TTI::SK_Reverse, MVT::v4f32, 2}, // REV64; EXT
{TTI::SK_Reverse, MVT::v2f64, 1}, // EXT
{TTI::SK_Reverse, MVT::v8f16, 2}, // REV64; EXT
{TTI::SK_Reverse, MVT::v8bf16, 2}, // REV64; EXT
{TTI::SK_Reverse, MVT::v8i16, 2}, // REV64; EXT
{TTI::SK_Reverse, MVT::v16i8, 2}, // REV64; EXT
{TTI::SK_Reverse, MVT::v4f16, 1}, // REV64
{TTI::SK_Reverse, MVT::v4bf16, 1}, // REV64
{TTI::SK_Reverse, MVT::v4i16, 1}, // REV64
{TTI::SK_Reverse, MVT::v8i8, 1}, // REV64
// Splice can all be lowered as `ext`.
{TTI::SK_Splice, MVT::v2i32, 1},
{TTI::SK_Splice, MVT::v4i32, 1},
{TTI::SK_Splice, MVT::v2i64, 1},
{TTI::SK_Splice, MVT::v2f32, 1},
{TTI::SK_Splice, MVT::v4f32, 1},
{TTI::SK_Splice, MVT::v2f64, 1},
{TTI::SK_Splice, MVT::v8f16, 1},
{TTI::SK_Splice, MVT::v8bf16, 1},
{TTI::SK_Splice, MVT::v8i16, 1},
{TTI::SK_Splice, MVT::v16i8, 1},
{TTI::SK_Splice, MVT::v4f16, 1},
{TTI::SK_Splice, MVT::v4bf16, 1},
{TTI::SK_Splice, MVT::v4i16, 1},
{TTI::SK_Splice, MVT::v8i8, 1},
// Broadcast shuffle kinds for scalable vectors
{TTI::SK_Broadcast, MVT::nxv16i8, 1},
{TTI::SK_Broadcast, MVT::nxv8i16, 1},
{TTI::SK_Broadcast, MVT::nxv4i32, 1},
{TTI::SK_Broadcast, MVT::nxv2i64, 1},
{TTI::SK_Broadcast, MVT::nxv2f16, 1},
{TTI::SK_Broadcast, MVT::nxv4f16, 1},
{TTI::SK_Broadcast, MVT::nxv8f16, 1},
{TTI::SK_Broadcast, MVT::nxv2bf16, 1},
{TTI::SK_Broadcast, MVT::nxv4bf16, 1},
{TTI::SK_Broadcast, MVT::nxv8bf16, 1},
{TTI::SK_Broadcast, MVT::nxv2f32, 1},
{TTI::SK_Broadcast, MVT::nxv4f32, 1},
{TTI::SK_Broadcast, MVT::nxv2f64, 1},
{TTI::SK_Broadcast, MVT::nxv16i1, 1},
{TTI::SK_Broadcast, MVT::nxv8i1, 1},
{TTI::SK_Broadcast, MVT::nxv4i1, 1},
{TTI::SK_Broadcast, MVT::nxv2i1, 1},
// Handle the cases for vector.reverse with scalable vectors
{TTI::SK_Reverse, MVT::nxv16i8, 1},
{TTI::SK_Reverse, MVT::nxv8i16, 1},
{TTI::SK_Reverse, MVT::nxv4i32, 1},
{TTI::SK_Reverse, MVT::nxv2i64, 1},
{TTI::SK_Reverse, MVT::nxv2f16, 1},
{TTI::SK_Reverse, MVT::nxv4f16, 1},
{TTI::SK_Reverse, MVT::nxv8f16, 1},
{TTI::SK_Reverse, MVT::nxv2bf16, 1},
{TTI::SK_Reverse, MVT::nxv4bf16, 1},
{TTI::SK_Reverse, MVT::nxv8bf16, 1},
{TTI::SK_Reverse, MVT::nxv2f32, 1},
{TTI::SK_Reverse, MVT::nxv4f32, 1},
{TTI::SK_Reverse, MVT::nxv2f64, 1},
{TTI::SK_Reverse, MVT::nxv16i1, 1},
{TTI::SK_Reverse, MVT::nxv8i1, 1},
{TTI::SK_Reverse, MVT::nxv4i1, 1},
{TTI::SK_Reverse, MVT::nxv2i1, 1},
};
if (const auto *Entry = CostTableLookup(ShuffleTbl, Kind, LT.second))
return LT.first * Entry->Cost;
}
if (Kind == TTI::SK_Splice && isa<ScalableVectorType>(Tp))
return getSpliceCost(Tp, Index, CostKind);
// Inserting a subvector can often be done with either a D, S or H register
// move, so long as the inserted vector is "aligned".
if (Kind == TTI::SK_InsertSubvector && LT.second.isFixedLengthVector() &&
LT.second.getSizeInBits() <= 128 && SubTp) {
std::pair<InstructionCost, MVT> SubLT = getTypeLegalizationCost(SubTp);
if (SubLT.second.isVector()) {
int NumElts = LT.second.getVectorNumElements();
int NumSubElts = SubLT.second.getVectorNumElements();
if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
return SubLT.first;
}
}
// Restore optimal kind.
if (IsExtractSubvector)
Kind = TTI::SK_ExtractSubvector;
return BaseT::getShuffleCost(Kind, Tp, Mask, CostKind, Index, SubTp, Args,
CxtI);
}
static bool containsDecreasingPointers(Loop *TheLoop,
PredicatedScalarEvolution *PSE) {
const auto &Strides = DenseMap<Value *, const SCEV *>();
for (BasicBlock *BB : TheLoop->blocks()) {
// Scan the instructions in the block and look for addresses that are
// consecutive and decreasing.
for (Instruction &I : *BB) {
if (isa<LoadInst>(&I) || isa<StoreInst>(&I)) {
Value *Ptr = getLoadStorePointerOperand(&I);
Type *AccessTy = getLoadStoreType(&I);
if (getPtrStride(*PSE, AccessTy, Ptr, TheLoop, Strides, /*Assume=*/true,
/*ShouldCheckWrap=*/false)
.value_or(0) < 0)
return true;
}
}
}
return false;
}
bool AArch64TTIImpl::preferFixedOverScalableIfEqualCost() const {
if (SVEPreferFixedOverScalableIfEqualCost.getNumOccurrences())
return SVEPreferFixedOverScalableIfEqualCost;
return ST->useFixedOverScalableIfEqualCost();
}
unsigned AArch64TTIImpl::getEpilogueVectorizationMinVF() const {
return ST->getEpilogueVectorizationMinVF();
}
bool AArch64TTIImpl::preferPredicateOverEpilogue(TailFoldingInfo *TFI) const {
if (!ST->hasSVE())
return false;
// We don't currently support vectorisation with interleaving for SVE - with
// such loops we're better off not using tail-folding. This gives us a chance
// to fall back on fixed-width vectorisation using NEON's ld2/st2/etc.
if (TFI->IAI->hasGroups())
return false;
TailFoldingOpts Required = TailFoldingOpts::Disabled;
if (TFI->LVL->getReductionVars().size())
Required |= TailFoldingOpts::Reductions;
if (TFI->LVL->getFixedOrderRecurrences().size())
Required |= TailFoldingOpts::Recurrences;
// We call this to discover whether any load/store pointers in the loop have
// negative strides. This will require extra work to reverse the loop
// predicate, which may be expensive.
if (containsDecreasingPointers(TFI->LVL->getLoop(),
TFI->LVL->getPredicatedScalarEvolution()))
Required |= TailFoldingOpts::Reverse;
if (Required == TailFoldingOpts::Disabled)
Required |= TailFoldingOpts::Simple;
if (!TailFoldingOptionLoc.satisfies(ST->getSVETailFoldingDefaultOpts(),
Required))
return false;
// Don't tail-fold for tight loops where we would be better off interleaving
// with an unpredicated loop.
unsigned NumInsns = 0;
for (BasicBlock *BB : TFI->LVL->getLoop()->blocks()) {
NumInsns += BB->sizeWithoutDebug();
}
// We expect 4 of these to be a IV PHI, IV add, IV compare and branch.
return NumInsns >= SVETailFoldInsnThreshold;
}
InstructionCost
AArch64TTIImpl::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
StackOffset BaseOffset, bool HasBaseReg,
int64_t Scale, unsigned AddrSpace) const {
// Scaling factors are not free at all.
// Operands | Rt Latency
// -------------------------------------------
// Rt, [Xn, Xm] | 4
// -------------------------------------------
// Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
// Rt, [Xn, Wm, <extend> #imm] |
TargetLoweringBase::AddrMode AM;
AM.BaseGV = BaseGV;
AM.BaseOffs = BaseOffset.getFixed();
AM.HasBaseReg = HasBaseReg;
AM.Scale = Scale;
AM.ScalableOffset = BaseOffset.getScalable();
if (getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace))
// Scale represents reg2 * scale, thus account for 1 if
// it is not equal to 0 or 1.
return AM.Scale != 0 && AM.Scale != 1;
return InstructionCost::getInvalid();
}
bool AArch64TTIImpl::shouldTreatInstructionLikeSelect(
const Instruction *I) const {
if (EnableOrLikeSelectOpt) {
// For the binary operators (e.g. or) we need to be more careful than
// selects, here we only transform them if they are already at a natural
// break point in the code - the end of a block with an unconditional
// terminator.
if (I->getOpcode() == Instruction::Or &&
isa<BranchInst>(I->getNextNode()) &&
cast<BranchInst>(I->getNextNode())->isUnconditional())
return true;
if (I->getOpcode() == Instruction::Add ||
I->getOpcode() == Instruction::Sub)
return true;
}
return BaseT::shouldTreatInstructionLikeSelect(I);
}
bool AArch64TTIImpl::isLSRCostLess(
const TargetTransformInfo::LSRCost &C1,
const TargetTransformInfo::LSRCost &C2) const {
// AArch64 specific here is adding the number of instructions to the
// comparison (though not as the first consideration, as some targets do)
// along with changing the priority of the base additions.
// TODO: Maybe a more nuanced tradeoff between instruction count
// and number of registers? To be investigated at a later date.
if (EnableLSRCostOpt)
return std::tie(C1.NumRegs, C1.Insns, C1.NumBaseAdds, C1.AddRecCost,
C1.NumIVMuls, C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
std::tie(C2.NumRegs, C2.Insns, C2.NumBaseAdds, C2.AddRecCost,
C2.NumIVMuls, C2.ScaleCost, C2.ImmCost, C2.SetupCost);
return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
}
static bool isSplatShuffle(Value *V) {
if (auto *Shuf = dyn_cast<ShuffleVectorInst>(V))
return all_equal(Shuf->getShuffleMask());
return false;
}
/// Check if both Op1 and Op2 are shufflevector extracts of either the lower
/// or upper half of the vector elements.
static bool areExtractShuffleVectors(Value *Op1, Value *Op2,
bool AllowSplat = false) {
// Scalable types can't be extract shuffle vectors.
if (Op1->getType()->isScalableTy() || Op2->getType()->isScalableTy())
return false;
auto areTypesHalfed = [](Value *FullV, Value *HalfV) {
auto *FullTy = FullV->getType();
auto *HalfTy = HalfV->getType();
return FullTy->getPrimitiveSizeInBits().getFixedValue() ==
2 * HalfTy->getPrimitiveSizeInBits().getFixedValue();
};
auto extractHalf = [](Value *FullV, Value *HalfV) {
auto *FullVT = cast<FixedVectorType>(FullV->getType());
auto *HalfVT = cast<FixedVectorType>(HalfV->getType());
return FullVT->getNumElements() == 2 * HalfVT->getNumElements();
};
ArrayRef<int> M1, M2;
Value *S1Op1 = nullptr, *S2Op1 = nullptr;
if (!match(Op1, m_Shuffle(m_Value(S1Op1), m_Undef(), m_Mask(M1))) ||
!match(Op2, m_Shuffle(m_Value(S2Op1), m_Undef(), m_Mask(M2))))
return false;
// If we allow splats, set S1Op1/S2Op1 to nullptr for the relavant arg so that
// it is not checked as an extract below.
if (AllowSplat && isSplatShuffle(Op1))
S1Op1 = nullptr;
if (AllowSplat && isSplatShuffle(Op2))
S2Op1 = nullptr;
// Check that the operands are half as wide as the result and we extract
// half of the elements of the input vectors.
if ((S1Op1 && (!areTypesHalfed(S1Op1, Op1) || !extractHalf(S1Op1, Op1))) ||
(S2Op1 && (!areTypesHalfed(S2Op1, Op2) || !extractHalf(S2Op1, Op2))))
return false;
// Check the mask extracts either the lower or upper half of vector
// elements.
int M1Start = 0;
int M2Start = 0;
int NumElements = cast<FixedVectorType>(Op1->getType())->getNumElements() * 2;
if ((S1Op1 &&
!ShuffleVectorInst::isExtractSubvectorMask(M1, NumElements, M1Start)) ||
(S2Op1 &&
!ShuffleVectorInst::isExtractSubvectorMask(M2, NumElements, M2Start)))
return false;
if ((M1Start != 0 && M1Start != (NumElements / 2)) ||
(M2Start != 0 && M2Start != (NumElements / 2)))
return false;
if (S1Op1 && S2Op1 && M1Start != M2Start)
return false;
return true;
}
/// Check if Ext1 and Ext2 are extends of the same type, doubling the bitwidth
/// of the vector elements.
static bool areExtractExts(Value *Ext1, Value *Ext2) {
auto areExtDoubled = [](Instruction *Ext) {
return Ext->getType()->getScalarSizeInBits() ==
2 * Ext->getOperand(0)->getType()->getScalarSizeInBits();
};
if (!match(Ext1, m_ZExtOrSExt(m_Value())) ||
!match(Ext2, m_ZExtOrSExt(m_Value())) ||
!areExtDoubled(cast<Instruction>(Ext1)) ||
!areExtDoubled(cast<Instruction>(Ext2)))
return false;
return true;
}
/// Check if Op could be used with vmull_high_p64 intrinsic.
static bool isOperandOfVmullHighP64(Value *Op) {
Value *VectorOperand = nullptr;
ConstantInt *ElementIndex = nullptr;
return match(Op, m_ExtractElt(m_Value(VectorOperand),
m_ConstantInt(ElementIndex))) &&
ElementIndex->getValue() == 1 &&
isa<FixedVectorType>(VectorOperand->getType()) &&
cast<FixedVectorType>(VectorOperand->getType())->getNumElements() == 2;
}
/// Check if Op1 and Op2 could be used with vmull_high_p64 intrinsic.
static bool areOperandsOfVmullHighP64(Value *Op1, Value *Op2) {
return isOperandOfVmullHighP64(Op1) && isOperandOfVmullHighP64(Op2);
}
static bool shouldSinkVectorOfPtrs(Value *Ptrs, SmallVectorImpl<Use *> &Ops) {
// Restrict ourselves to the form CodeGenPrepare typically constructs.
auto *GEP = dyn_cast<GetElementPtrInst>(Ptrs);
if (!GEP || GEP->getNumOperands() != 2)
return false;
Value *Base = GEP->getOperand(0);
Value *Offsets = GEP->getOperand(1);
// We only care about scalar_base+vector_offsets.
if (Base->getType()->isVectorTy() || !Offsets->getType()->isVectorTy())
return false;
// Sink extends that would allow us to use 32-bit offset vectors.
if (isa<SExtInst>(Offsets) || isa<ZExtInst>(Offsets)) {
auto *OffsetsInst = cast<Instruction>(Offsets);
if (OffsetsInst->getType()->getScalarSizeInBits() > 32 &&
OffsetsInst->getOperand(0)->getType()->getScalarSizeInBits() <= 32)
Ops.push_back(&GEP->getOperandUse(1));
}
// Sink the GEP.
return true;
}
/// We want to sink following cases:
/// (add|sub|gep) A, ((mul|shl) vscale, imm); (add|sub|gep) A, vscale;
/// (add|sub|gep) A, ((mul|shl) zext(vscale), imm);
static bool shouldSinkVScale(Value *Op, SmallVectorImpl<Use *> &Ops) {
if (match(Op, m_VScale()))
return true;
if (match(Op, m_Shl(m_VScale(), m_ConstantInt())) ||
match(Op, m_Mul(m_VScale(), m_ConstantInt()))) {
Ops.push_back(&cast<Instruction>(Op)->getOperandUse(0));
return true;
}
if (match(Op, m_Shl(m_ZExt(m_VScale()), m_ConstantInt())) ||
match(Op, m_Mul(m_ZExt(m_VScale()), m_ConstantInt()))) {
Value *ZExtOp = cast<Instruction>(Op)->getOperand(0);
Ops.push_back(&cast<Instruction>(ZExtOp)->getOperandUse(0));
Ops.push_back(&cast<Instruction>(Op)->getOperandUse(0));
return true;
}
return false;
}
/// Check if sinking \p I's operands to I's basic block is profitable, because
/// the operands can be folded into a target instruction, e.g.
/// shufflevectors extracts and/or sext/zext can be folded into (u,s)subl(2).
bool AArch64TTIImpl::isProfitableToSinkOperands(
Instruction *I, SmallVectorImpl<Use *> &Ops) const {
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
switch (II->getIntrinsicID()) {
case Intrinsic::aarch64_neon_smull:
case Intrinsic::aarch64_neon_umull:
if (areExtractShuffleVectors(II->getOperand(0), II->getOperand(1),
/*AllowSplat=*/true)) {
Ops.push_back(&II->getOperandUse(0));
Ops.push_back(&II->getOperandUse(1));
return true;
}
[[fallthrough]];
case Intrinsic::fma:
case Intrinsic::fmuladd:
if (isa<VectorType>(I->getType()) &&
cast<VectorType>(I->getType())->getElementType()->isHalfTy() &&
!ST->hasFullFP16())
return false;
[[fallthrough]];
case Intrinsic::aarch64_neon_sqdmull:
case Intrinsic::aarch64_neon_sqdmulh:
case Intrinsic::aarch64_neon_sqrdmulh:
// Sink splats for index lane variants
if (isSplatShuffle(II->getOperand(0)))
Ops.push_back(&II->getOperandUse(0));
if (isSplatShuffle(II->getOperand(1)))
Ops.push_back(&II->getOperandUse(1));
return !Ops.empty();
case Intrinsic::aarch64_neon_fmlal:
case Intrinsic::aarch64_neon_fmlal2:
case Intrinsic::aarch64_neon_fmlsl:
case Intrinsic::aarch64_neon_fmlsl2:
// Sink splats for index lane variants
if (isSplatShuffle(II->getOperand(1)))
Ops.push_back(&II->getOperandUse(1));
if (isSplatShuffle(II->getOperand(2)))
Ops.push_back(&II->getOperandUse(2));
return !Ops.empty();
case Intrinsic::aarch64_sve_ptest_first:
case Intrinsic::aarch64_sve_ptest_last:
if (auto *IIOp = dyn_cast<IntrinsicInst>(II->getOperand(0)))
if (IIOp->getIntrinsicID() == Intrinsic::aarch64_sve_ptrue)
Ops.push_back(&II->getOperandUse(0));
return !Ops.empty();
case Intrinsic::aarch64_sme_write_horiz:
case Intrinsic::aarch64_sme_write_vert:
case Intrinsic::aarch64_sme_writeq_horiz:
case Intrinsic::aarch64_sme_writeq_vert: {
auto *Idx = dyn_cast<Instruction>(II->getOperand(1));
if (!Idx || Idx->getOpcode() != Instruction::Add)
return false;
Ops.push_back(&II->getOperandUse(1));
return true;
}
case Intrinsic::aarch64_sme_read_horiz:
case Intrinsic::aarch64_sme_read_vert:
case Intrinsic::aarch64_sme_readq_horiz:
case Intrinsic::aarch64_sme_readq_vert:
case Intrinsic::aarch64_sme_ld1b_vert:
case Intrinsic::aarch64_sme_ld1h_vert:
case Intrinsic::aarch64_sme_ld1w_vert:
case Intrinsic::aarch64_sme_ld1d_vert:
case Intrinsic::aarch64_sme_ld1q_vert:
case Intrinsic::aarch64_sme_st1b_vert:
case Intrinsic::aarch64_sme_st1h_vert:
case Intrinsic::aarch64_sme_st1w_vert:
case Intrinsic::aarch64_sme_st1d_vert:
case Intrinsic::aarch64_sme_st1q_vert:
case Intrinsic::aarch64_sme_ld1b_horiz:
case Intrinsic::aarch64_sme_ld1h_horiz:
case Intrinsic::aarch64_sme_ld1w_horiz:
case Intrinsic::aarch64_sme_ld1d_horiz:
case Intrinsic::aarch64_sme_ld1q_horiz:
case Intrinsic::aarch64_sme_st1b_horiz:
case Intrinsic::aarch64_sme_st1h_horiz:
case Intrinsic::aarch64_sme_st1w_horiz:
case Intrinsic::aarch64_sme_st1d_horiz:
case Intrinsic::aarch64_sme_st1q_horiz: {
auto *Idx = dyn_cast<Instruction>(II->getOperand(3));
if (!Idx || Idx->getOpcode() != Instruction::Add)
return false;
Ops.push_back(&II->getOperandUse(3));
return true;
}
case Intrinsic::aarch64_neon_pmull:
if (!areExtractShuffleVectors(II->getOperand(0), II->getOperand(1)))
return false;
Ops.push_back(&II->getOperandUse(0));
Ops.push_back(&II->getOperandUse(1));
return true;
case Intrinsic::aarch64_neon_pmull64:
if (!areOperandsOfVmullHighP64(II->getArgOperand(0),
II->getArgOperand(1)))
return false;
Ops.push_back(&II->getArgOperandUse(0));
Ops.push_back(&II->getArgOperandUse(1));
return true;
case Intrinsic::masked_gather:
if (!shouldSinkVectorOfPtrs(II->getArgOperand(0), Ops))
return false;
Ops.push_back(&II->getArgOperandUse(0));
return true;
case Intrinsic::masked_scatter:
if (!shouldSinkVectorOfPtrs(II->getArgOperand(1), Ops))
return false;
Ops.push_back(&II->getArgOperandUse(1));
return true;
default:
return false;
}
}
auto ShouldSinkCondition = [](Value *Cond) -> bool {
auto *II = dyn_cast<IntrinsicInst>(Cond);
return II && II->getIntrinsicID() == Intrinsic::vector_reduce_or &&
isa<ScalableVectorType>(II->getOperand(0)->getType());
};
switch (I->getOpcode()) {
case Instruction::GetElementPtr:
case Instruction::Add:
case Instruction::Sub:
// Sink vscales closer to uses for better isel
for (unsigned Op = 0; Op < I->getNumOperands(); ++Op) {
if (shouldSinkVScale(I->getOperand(Op), Ops)) {
Ops.push_back(&I->getOperandUse(Op));
return true;
}
}
break;
case Instruction::Select: {
if (!ShouldSinkCondition(I->getOperand(0)))
return false;
Ops.push_back(&I->getOperandUse(0));
return true;
}
case Instruction::Br: {
if (cast<BranchInst>(I)->isUnconditional())
return false;
if (!ShouldSinkCondition(cast<BranchInst>(I)->getCondition()))
return false;
Ops.push_back(&I->getOperandUse(0));
return true;
}
default:
break;
}
if (!I->getType()->isVectorTy())
return false;
switch (I->getOpcode()) {
case Instruction::Sub:
case Instruction::Add: {
if (!areExtractExts(I->getOperand(0), I->getOperand(1)))
return false;
// If the exts' operands extract either the lower or upper elements, we
// can sink them too.
auto Ext1 = cast<Instruction>(I->getOperand(0));
auto Ext2 = cast<Instruction>(I->getOperand(1));
if (areExtractShuffleVectors(Ext1->getOperand(0), Ext2->getOperand(0))) {
Ops.push_back(&Ext1->getOperandUse(0));
Ops.push_back(&Ext2->getOperandUse(0));
}
Ops.push_back(&I->getOperandUse(0));
Ops.push_back(&I->getOperandUse(1));
return true;
}
case Instruction::Or: {
// Pattern: Or(And(MaskValue, A), And(Not(MaskValue), B)) ->
// bitselect(MaskValue, A, B) where Not(MaskValue) = Xor(MaskValue, -1)
if (ST->hasNEON()) {
Instruction *OtherAnd, *IA, *IB;
Value *MaskValue;
// MainAnd refers to And instruction that has 'Not' as one of its operands
if (match(I, m_c_Or(m_OneUse(m_Instruction(OtherAnd)),
m_OneUse(m_c_And(m_OneUse(m_Not(m_Value(MaskValue))),
m_Instruction(IA)))))) {
if (match(OtherAnd,
m_c_And(m_Specific(MaskValue), m_Instruction(IB)))) {
Instruction *MainAnd = I->getOperand(0) == OtherAnd
? cast<Instruction>(I->getOperand(1))
: cast<Instruction>(I->getOperand(0));
// Both Ands should be in same basic block as Or
if (I->getParent() != MainAnd->getParent() ||
I->getParent() != OtherAnd->getParent())
return false;
// Non-mask operands of both Ands should also be in same basic block
if (I->getParent() != IA->getParent() ||
I->getParent() != IB->getParent())
return false;
Ops.push_back(
&MainAnd->getOperandUse(MainAnd->getOperand(0) == IA ? 1 : 0));
Ops.push_back(&I->getOperandUse(0));
Ops.push_back(&I->getOperandUse(1));
return true;
}
}
}
return false;
}
case Instruction::Mul: {
auto ShouldSinkSplatForIndexedVariant = [](Value *V) {
auto *Ty = cast<VectorType>(V->getType());
// For SVE the lane-indexing is within 128-bits, so we can't fold splats.
if (Ty->isScalableTy())
return false;
// Indexed variants of Mul exist for i16 and i32 element types only.
return Ty->getScalarSizeInBits() == 16 || Ty->getScalarSizeInBits() == 32;
};
int NumZExts = 0, NumSExts = 0;
for (auto &Op : I->operands()) {
// Make sure we are not already sinking this operand
if (any_of(Ops, [&](Use *U) { return U->get() == Op; }))
continue;
if (match(&Op, m_ZExtOrSExt(m_Value()))) {
auto *Ext = cast<Instruction>(Op);
auto *ExtOp = Ext->getOperand(0);
if (isSplatShuffle(ExtOp) && ShouldSinkSplatForIndexedVariant(ExtOp))
Ops.push_back(&Ext->getOperandUse(0));
Ops.push_back(&Op);
if (isa<SExtInst>(Ext))
NumSExts++;
else
NumZExts++;
continue;
}
ShuffleVectorInst *Shuffle = dyn_cast<ShuffleVectorInst>(Op);
if (!Shuffle)
continue;
// If the Shuffle is a splat and the operand is a zext/sext, sinking the
// operand and the s/zext can help create indexed s/umull. This is
// especially useful to prevent i64 mul being scalarized.
if (isSplatShuffle(Shuffle) &&
match(Shuffle->getOperand(0), m_ZExtOrSExt(m_Value()))) {
Ops.push_back(&Shuffle->getOperandUse(0));
Ops.push_back(&Op);
if (match(Shuffle->getOperand(0), m_SExt(m_Value())))
NumSExts++;
else
NumZExts++;
continue;
}
Value *ShuffleOperand = Shuffle->getOperand(0);
InsertElementInst *Insert = dyn_cast<InsertElementInst>(ShuffleOperand);
if (!Insert)
continue;
Instruction *OperandInstr = dyn_cast<Instruction>(Insert->getOperand(1));
if (!OperandInstr)
continue;
ConstantInt *ElementConstant =
dyn_cast<ConstantInt>(Insert->getOperand(2));
// Check that the insertelement is inserting into element 0
if (!ElementConstant || !ElementConstant->isZero())
continue;
unsigned Opcode = OperandInstr->getOpcode();
if (Opcode == Instruction::SExt)
NumSExts++;
else if (Opcode == Instruction::ZExt)
NumZExts++;
else {
// If we find that the top bits are known 0, then we can sink and allow
// the backend to generate a umull.
unsigned Bitwidth = I->getType()->getScalarSizeInBits();
APInt UpperMask = APInt::getHighBitsSet(Bitwidth, Bitwidth / 2);
const DataLayout &DL = I->getDataLayout();
if (!MaskedValueIsZero(OperandInstr, UpperMask, DL))
continue;
NumZExts++;
}
// And(Load) is excluded to prevent CGP getting stuck in a loop of sinking
// the And, just to hoist it again back to the load.
if (!match(OperandInstr, m_And(m_Load(m_Value()), m_Value())))
Ops.push_back(&Insert->getOperandUse(1));
Ops.push_back(&Shuffle->getOperandUse(0));
Ops.push_back(&Op);
}
// It is profitable to sink if we found two of the same type of extends.
if (!Ops.empty() && (NumSExts == 2 || NumZExts == 2))
return true;
// Otherwise, see if we should sink splats for indexed variants.
if (!ShouldSinkSplatForIndexedVariant(I))
return false;
Ops.clear();
if (isSplatShuffle(I->getOperand(0)))
Ops.push_back(&I->getOperandUse(0));
if (isSplatShuffle(I->getOperand(1)))
Ops.push_back(&I->getOperandUse(1));
return !Ops.empty();
}
case Instruction::FMul: {
// For SVE the lane-indexing is within 128-bits, so we can't fold splats.
if (I->getType()->isScalableTy())
return false;
if (cast<VectorType>(I->getType())->getElementType()->isHalfTy() &&
!ST->hasFullFP16())
return false;
// Sink splats for index lane variants
if (isSplatShuffle(I->getOperand(0)))
Ops.push_back(&I->getOperandUse(0));
if (isSplatShuffle(I->getOperand(1)))
Ops.push_back(&I->getOperandUse(1));
return !Ops.empty();
}
default:
return false;
}
return false;
}