}
X86TTI(const X86TargetMachine *TM)
- : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
- TLI(TM->getTargetLowering()) {
+ : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
+ TLI(TM->getSubtargetImpl()->getTargetLowering()) {
initializeX86TTIPass(*PassRegistry::getPassRegistry());
}
unsigned getNumberOfRegisters(bool Vector) const override;
unsigned getRegisterBitWidth(bool Vector) const override;
- unsigned getMaximumUnrollFactor() const override;
+ unsigned getMaxInterleaveFactor() const override;
unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
- OperandValueKind) const override;
+ OperandValueKind, OperandValueProperties,
+ OperandValueProperties) const override;
unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
int Index, Type *SubTp) const override;
unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
Type *Ty) const override;
unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
Type *Ty) const override;
+ bool isLegalMaskedLoad (Type *DataType, int Consecutive) const override;
+ bool isLegalMaskedStore(Type *DataType, int Consecutive) const override;
/// @}
};
if (Vector && !ST->hasSSE1())
return 0;
- if (ST->is64Bit())
+ if (ST->is64Bit()) {
+ if (Vector && ST->hasAVX512())
+ return 32;
return 16;
+ }
return 8;
}
unsigned X86TTI::getRegisterBitWidth(bool Vector) const {
if (Vector) {
+ if (ST->hasAVX512()) return 512;
if (ST->hasAVX()) return 256;
if (ST->hasSSE1()) return 128;
return 0;
}
-unsigned X86TTI::getMaximumUnrollFactor() const {
+unsigned X86TTI::getMaxInterleaveFactor() const {
if (ST->isAtom())
return 1;
return 2;
}
-unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
- OperandValueKind Op1Info,
- OperandValueKind Op2Info) const {
+unsigned X86TTI::getArithmeticInstrCost(
+ unsigned Opcode, Type *Ty, OperandValueKind Op1Info,
+ OperandValueKind Op2Info, OperandValueProperties Opd1PropInfo,
+ OperandValueProperties Opd2PropInfo) const {
// Legalize the type.
std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
+ if (ISD == ISD::SDIV &&
+ Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
+ Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
+ // On X86, vector signed division by constants power-of-two are
+ // normally expanded to the sequence SRA + SRL + ADD + SRA.
+ // The OperandValue properties many not be same as that of previous
+ // operation;conservatively assume OP_None.
+ unsigned Cost =
+ 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, Op2Info,
+ TargetTransformInfo::OP_None,
+ TargetTransformInfo::OP_None);
+ Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
+ TargetTransformInfo::OP_None,
+ TargetTransformInfo::OP_None);
+ Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
+ TargetTransformInfo::OP_None,
+ TargetTransformInfo::OP_None);
+
+ return Cost;
+ }
+
static const CostTblEntry<MVT::SimpleValueType>
AVX2UniformConstCostTable[] = {
{ ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
return LT.first * AVX2UniformConstCostTable[Idx].Cost;
}
+ static const CostTblEntry<MVT::SimpleValueType> AVX512CostTable[] = {
+ { ISD::SHL, MVT::v16i32, 1 },
+ { ISD::SRL, MVT::v16i32, 1 },
+ { ISD::SRA, MVT::v16i32, 1 },
+ { ISD::SHL, MVT::v8i64, 1 },
+ { ISD::SRL, MVT::v8i64, 1 },
+ { ISD::SRA, MVT::v8i64, 1 },
+ };
+
static const CostTblEntry<MVT::SimpleValueType> AVX2CostTable[] = {
// Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
// customize them to detect the cases where shift amount is a scalar one.
{ ISD::UDIV, MVT::v4i64, 4*20 },
};
+ if (ST->hasAVX512()) {
+ int Idx = CostTableLookup(AVX512CostTable, ISD, LT.second);
+ if (Idx != -1)
+ return LT.first * AVX512CostTable[Idx].Cost;
+ }
// Look for AVX2 lowering tricks.
if (ST->hasAVX2()) {
if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
{ ISD::SHL, MVT::v8i16, 8*10 }, // Scalarized.
{ ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
{ ISD::SHL, MVT::v2i64, 2*10 }, // Scalarized.
- { ISD::SHL, MVT::v4i64, 4*10 }, // Scalarized.
+ { ISD::SHL, MVT::v4i64, 4*10 }, // Scalarized.
{ ISD::SRL, MVT::v16i8, 16*10 }, // Scalarized.
{ ISD::SRL, MVT::v8i16, 8*10 }, // Scalarized.
unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
Type *SubTp) const {
- // We only estimate the cost of reverse shuffles.
- if (Kind != SK_Reverse)
+ // We only estimate the cost of reverse and alternate shuffles.
+ if (Kind != SK_Reverse && Kind != SK_Alternate)
return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
- unsigned Cost = 1;
- if (LT.second.getSizeInBits() > 128)
- Cost = 3; // Extract + insert + copy.
+ if (Kind == SK_Reverse) {
+ std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
+ unsigned Cost = 1;
+ if (LT.second.getSizeInBits() > 128)
+ Cost = 3; // Extract + insert + copy.
+
+ // Multiple by the number of parts.
+ return Cost * LT.first;
+ }
+
+ if (Kind == SK_Alternate) {
+ // 64-bit packed float vectors (v2f32) are widened to type v4f32.
+ // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
+ std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
+
+ // The backend knows how to generate a single VEX.256 version of
+ // instruction VPBLENDW if the target supports AVX2.
+ if (ST->hasAVX2() && LT.second == MVT::v16i16)
+ return LT.first;
+
+ static const CostTblEntry<MVT::SimpleValueType> AVXAltShuffleTbl[] = {
+ {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1}, // vblendpd
+ {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1}, // vblendpd
+
+ {ISD::VECTOR_SHUFFLE, MVT::v8i32, 1}, // vblendps
+ {ISD::VECTOR_SHUFFLE, MVT::v8f32, 1}, // vblendps
+
+ // This shuffle is custom lowered into a sequence of:
+ // 2x vextractf128 , 2x vpblendw , 1x vinsertf128
+ {ISD::VECTOR_SHUFFLE, MVT::v16i16, 5},
+
+ // This shuffle is custom lowered into a long sequence of:
+ // 2x vextractf128 , 4x vpshufb , 2x vpor , 1x vinsertf128
+ {ISD::VECTOR_SHUFFLE, MVT::v32i8, 9}
+ };
+
+ if (ST->hasAVX()) {
+ int Idx = CostTableLookup(AVXAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
+ if (Idx != -1)
+ return LT.first * AVXAltShuffleTbl[Idx].Cost;
+ }
+
+ static const CostTblEntry<MVT::SimpleValueType> SSE41AltShuffleTbl[] = {
+ // These are lowered into movsd.
+ {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
+ {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
+
+ // packed float vectors with four elements are lowered into BLENDI dag
+ // nodes. A v4i32/v4f32 BLENDI generates a single 'blendps'/'blendpd'.
+ {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
+ {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
+
+ // This shuffle generates a single pshufw.
+ {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
+
+ // There is no instruction that matches a v16i8 alternate shuffle.
+ // The backend will expand it into the sequence 'pshufb + pshufb + or'.
+ {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}
+ };
+
+ if (ST->hasSSE41()) {
+ int Idx = CostTableLookup(SSE41AltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
+ if (Idx != -1)
+ return LT.first * SSE41AltShuffleTbl[Idx].Cost;
+ }
+
+ static const CostTblEntry<MVT::SimpleValueType> SSSE3AltShuffleTbl[] = {
+ {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd
+ {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd
+
+ // SSE3 doesn't have 'blendps'. The following shuffles are expanded into
+ // the sequence 'shufps + pshufd'
+ {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
+ {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
+
+ {ISD::VECTOR_SHUFFLE, MVT::v8i16, 3}, // pshufb + pshufb + or
+ {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3} // pshufb + pshufb + or
+ };
+
+ if (ST->hasSSSE3()) {
+ int Idx = CostTableLookup(SSSE3AltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
+ if (Idx != -1)
+ return LT.first * SSSE3AltShuffleTbl[Idx].Cost;
+ }
+
+ static const CostTblEntry<MVT::SimpleValueType> SSEAltShuffleTbl[] = {
+ {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd
+ {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd
+
+ {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, // shufps + pshufd
+ {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, // shufps + pshufd
+
+ // This is expanded into a long sequence of four extract + four insert.
+ {ISD::VECTOR_SHUFFLE, MVT::v8i16, 8}, // 4 x pextrw + 4 pinsrw.
+
+ // 8 x (pinsrw + pextrw + and + movb + movzb + or)
+ {ISD::VECTOR_SHUFFLE, MVT::v16i8, 48}
+ };
- // Multiple by the number of parts.
- return Cost * LT.first;
+ // Fall-back (SSE3 and SSE2).
+ int Idx = CostTableLookup(SSEAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
+ if (Idx != -1)
+ return LT.first * SSEAltShuffleTbl[Idx].Cost;
+ return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
+ }
+
+ return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
}
unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
// There are faster sequences for float conversions.
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
- { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
+ { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
{ ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
return LTSrc.first * SSE2ConvTbl[Idx].Cost;
}
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ AVX512ConversionTbl[] = {
+ { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 },
+ { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 },
+ { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 },
+ { ISD::FP_ROUND, MVT::v16f32, MVT::v8f64, 3 },
+
+ { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 1 },
+ { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 1 },
+ { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 1 },
+ { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 },
+ { ISD::TRUNCATE, MVT::v16i32, MVT::v8i64, 4 },
+
+ // v16i1 -> v16i32 - load + broadcast
+ { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 2 },
+ { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 },
+
+ { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 },
+ { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 },
+ { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
+ { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
+ { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v16i32, 3 },
+ { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v16i32, 3 },
+
+ { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 },
+ { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 },
+ { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 },
+ { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 },
+ { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 },
+ { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 },
+ { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 },
+ };
+
+ if (ST->hasAVX512()) {
+ int Idx = ConvertCostTableLookup(AVX512ConversionTbl, ISD, LTDest.second,
+ LTSrc.second);
+ if (Idx != -1)
+ return AVX512ConversionTbl[Idx].Cost;
+ }
EVT SrcTy = TLI->getValueType(Src);
EVT DstTy = TLI->getValueType(Dst);
{ ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 },
{ ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 },
{ ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 },
+
+ { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 },
+ { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 },
+
+ { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 },
};
static const TypeConversionCostTblEntry<MVT::SimpleValueType>
{ ISD::SETCC, MVT::v32i8, 1 },
};
+ static const CostTblEntry<MVT::SimpleValueType> AVX512CostTbl[] = {
+ { ISD::SETCC, MVT::v8i64, 1 },
+ { ISD::SETCC, MVT::v16i32, 1 },
+ { ISD::SETCC, MVT::v8f64, 1 },
+ { ISD::SETCC, MVT::v16f32, 1 },
+ };
+
+ if (ST->hasAVX512()) {
+ int Idx = CostTableLookup(AVX512CostTbl, ISD, MTy);
+ if (Idx != -1)
+ return LT.first * AVX512CostTbl[Idx].Cost;
+ }
+
if (ST->hasAVX2()) {
int Idx = CostTableLookup(AVX2CostTbl, ISD, MTy);
if (Idx != -1)
unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
bool IsPairwise) const {
-
+
std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
-
+
MVT MTy = LT.second;
-
+
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
-
- // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
- // and make it as the cost.
-
+
+ // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
+ // and make it as the cost.
+
static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblPairWise[] = {
{ ISD::FADD, MVT::v2f64, 2 },
{ ISD::FADD, MVT::v4f32, 4 },
{ ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
{ ISD::ADD, MVT::v8i16, 5 },
};
-
+
static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblPairWise[] = {
{ ISD::FADD, MVT::v4f32, 4 },
{ ISD::FADD, MVT::v4f64, 5 },
{ ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3".
{ ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3".
};
-
+
static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblNoPairWise[] = {
{ ISD::FADD, MVT::v4f32, 3 },
{ ISD::FADD, MVT::v4f64, 3 },
{ ISD::ADD, MVT::v8i16, 4 },
{ ISD::ADD, MVT::v8i32, 5 },
};
-
+
if (IsPairwise) {
if (ST->hasAVX()) {
int Idx = CostTableLookup(AVX1CostTblPairWise, ISD, MTy);
if (Idx != -1)
return LT.first * AVX1CostTblPairWise[Idx].Cost;
}
-
+
if (ST->hasSSE42()) {
int Idx = CostTableLookup(SSE42CostTblPairWise, ISD, MTy);
if (Idx != -1)
if (Idx != -1)
return LT.first * AVX1CostTblNoPairWise[Idx].Cost;
}
-
+
if (ST->hasSSE42()) {
int Idx = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy);
if (Idx != -1)
}
return X86TTI::getIntImmCost(Imm, Ty);
}
+
+bool X86TTI::isLegalMaskedLoad(Type *DataTy, int Consecutive) const {
+ int DataWidth = DataTy->getPrimitiveSizeInBits();
+
+ // Todo: AVX512 allows gather/scatter, works with strided and random as well
+ if ((DataWidth < 32) || (Consecutive == 0))
+ return false;
+ if (ST->hasAVX512() || ST->hasAVX2())
+ return true;
+ return false;
+}
+
+bool X86TTI::isLegalMaskedStore(Type *DataType, int Consecutive) const {
+ return isLegalMaskedLoad(DataType, Consecutive);
+}
+
projects / oota-llvm.git / blobdiff