#include "llvm/Support/Debug.h"
#include "llvm/Target/CostTable.h"
#include "llvm/Target/TargetLowering.h"
+
using namespace llvm;
#define DEBUG_TYPE "x86tti"
if (ST->is64Bit())
return 64;
- return 32;
+ return 32;
}
unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
return 2;
}
-unsigned X86TTIImpl::getArithmeticInstrCost(
+int X86TTIImpl::getArithmeticInstrCost(
unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info,
TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo,
TTI::OperandValueProperties Opd2PropInfo) {
// Legalize the type.
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
// 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);
+ int 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);
static const CostTblEntry<MVT::SimpleValueType>
AVX2UniformConstCostTable[] = {
+ { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle.
+
{ ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
{ ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
{ ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
{ ISD::SRA, MVT::v8i64, 1 },
};
+ if (ST->hasAVX512()) {
+ int Idx = CostTableLookup(AVX512CostTable, ISD, LT.second);
+ if (Idx != -1)
+ return LT.first * AVX512CostTable[Idx].Cost;
+ }
+
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::SRL, MVT::v2i64, 1 },
{ ISD::SHL, MVT::v4i64, 1 },
{ ISD::SRL, MVT::v4i64, 1 },
+ };
+
+ // Look for AVX2 lowering tricks.
+ if (ST->hasAVX2()) {
+ if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
+ (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
+ Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
+ // On AVX2, a packed v16i16 shift left by a constant build_vector
+ // is lowered into a vector multiply (vpmullw).
+ return LT.first;
+
+ int Idx = CostTableLookup(AVX2CostTable, ISD, LT.second);
+ if (Idx != -1)
+ return LT.first * AVX2CostTable[Idx].Cost;
+ }
- { ISD::SHL, MVT::v32i8, 42 }, // cmpeqb sequence.
+ static const CostTblEntry<MVT::SimpleValueType> XOPCostTable[] = {
+ // 128bit shifts take 1cy, but right shifts require negation beforehand.
+ { ISD::SHL, MVT::v16i8, 1 },
+ { ISD::SRL, MVT::v16i8, 2 },
+ { ISD::SRA, MVT::v16i8, 2 },
+ { ISD::SHL, MVT::v8i16, 1 },
+ { ISD::SRL, MVT::v8i16, 2 },
+ { ISD::SRA, MVT::v8i16, 2 },
+ { ISD::SHL, MVT::v4i32, 1 },
+ { ISD::SRL, MVT::v4i32, 2 },
+ { ISD::SRA, MVT::v4i32, 2 },
+ { ISD::SHL, MVT::v2i64, 1 },
+ { ISD::SRL, MVT::v2i64, 2 },
+ { ISD::SRA, MVT::v2i64, 2 },
+ // 256bit shifts require splitting if AVX2 didn't catch them above.
+ { ISD::SHL, MVT::v32i8, 2 },
+ { ISD::SRL, MVT::v32i8, 4 },
+ { ISD::SRA, MVT::v32i8, 4 },
+ { ISD::SHL, MVT::v16i16, 2 },
+ { ISD::SRL, MVT::v16i16, 4 },
+ { ISD::SRA, MVT::v16i16, 4 },
+ { ISD::SHL, MVT::v8i32, 2 },
+ { ISD::SRL, MVT::v8i32, 4 },
+ { ISD::SRA, MVT::v8i32, 4 },
+ { ISD::SHL, MVT::v4i64, 2 },
+ { ISD::SRL, MVT::v4i64, 4 },
+ { ISD::SRA, MVT::v4i64, 4 },
+ };
+
+ // Look for XOP lowering tricks.
+ if (ST->hasXOP()) {
+ int Idx = CostTableLookup(XOPCostTable, ISD, LT.second);
+ if (Idx != -1)
+ return LT.first * XOPCostTable[Idx].Cost;
+ }
+
+ static const CostTblEntry<MVT::SimpleValueType> AVX2CustomCostTable[] = {
+ { ISD::SHL, MVT::v32i8, 11 }, // vpblendvb sequence.
{ ISD::SHL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence.
- { ISD::SRL, MVT::v32i8, 32*10 }, // Scalarized.
+ { ISD::SRL, MVT::v32i8, 11 }, // vpblendvb sequence.
{ ISD::SRL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence.
- { ISD::SRA, MVT::v32i8, 32*10 }, // Scalarized.
+ { ISD::SRA, MVT::v32i8, 24 }, // vpblendvb sequence.
{ ISD::SRA, MVT::v16i16, 10 }, // extend/vpsravd/pack sequence.
- { ISD::SRA, MVT::v4i64, 4*10 }, // Scalarized.
+ { ISD::SRA, MVT::v2i64, 4 }, // srl/xor/sub sequence.
+ { ISD::SRA, MVT::v4i64, 4 }, // srl/xor/sub sequence.
// Vectorizing division is a bad idea. See the SSE2 table for more comments.
{ ISD::SDIV, MVT::v32i8, 32*20 },
{ 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.
+ // Look for AVX2 lowering tricks for custom cases.
if (ST->hasAVX2()) {
- if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
- (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
- Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
- // On AVX2, a packed v16i16 shift left by a constant build_vector
- // is lowered into a vector multiply (vpmullw).
- return LT.first;
-
- int Idx = CostTableLookup(AVX2CostTable, ISD, LT.second);
+ int Idx = CostTableLookup(AVX2CustomCostTable, ISD, LT.second);
if (Idx != -1)
- return LT.first * AVX2CostTable[Idx].Cost;
+ return LT.first * AVX2CustomCostTable[Idx].Cost;
}
static const CostTblEntry<MVT::SimpleValueType>
{ ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
{ ISD::SRA, MVT::v8i16, 1 }, // psraw.
{ ISD::SRA, MVT::v4i32, 1 }, // psrad.
+ { ISD::SRA, MVT::v2i64, 4 }, // 2 x psrad + shuffle.
{ ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
{ ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
// to ISel. The cost model must return worst case assumptions because it is
// used for vectorization and we don't want to make vectorized code worse
// than scalar code.
- { ISD::SHL, MVT::v16i8, 30 }, // cmpeqb sequence.
- { 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::SRL, MVT::v16i8, 16*10 }, // Scalarized.
- { ISD::SRL, MVT::v8i16, 8*10 }, // Scalarized.
- { ISD::SRL, MVT::v4i32, 4*10 }, // Scalarized.
- { ISD::SRL, MVT::v2i64, 2*10 }, // Scalarized.
-
- { ISD::SRA, MVT::v16i8, 16*10 }, // Scalarized.
- { ISD::SRA, MVT::v8i16, 8*10 }, // Scalarized.
- { ISD::SRA, MVT::v4i32, 4*10 }, // Scalarized.
- { ISD::SRA, MVT::v2i64, 2*10 }, // Scalarized.
+ { ISD::SHL, MVT::v16i8, 26 }, // cmpgtb sequence.
+ { ISD::SHL, MVT::v8i16, 32 }, // cmpgtb sequence.
+ { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
+ { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence.
+ { ISD::SHL, MVT::v4i64, 8 }, // splat+shuffle sequence.
+
+ { ISD::SRL, MVT::v16i8, 26 }, // cmpgtb sequence.
+ { ISD::SRL, MVT::v8i16, 32 }, // cmpgtb sequence.
+ { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend.
+ { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence.
+
+ { ISD::SRA, MVT::v16i8, 54 }, // unpacked cmpgtb sequence.
+ { ISD::SRA, MVT::v8i16, 32 }, // cmpgtb sequence.
+ { ISD::SRA, MVT::v4i32, 16 }, // Shift each lane + blend.
+ { ISD::SRA, MVT::v2i64, 12 }, // srl/xor/sub sequence.
// It is not a good idea to vectorize division. We have to scalarize it and
// in the process we will often end up having to spilling regular
return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
}
-unsigned X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
- Type *SubTp) {
+int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
+ Type *SubTp) {
// We only estimate the cost of reverse and alternate shuffles.
if (Kind != TTI::SK_Reverse && Kind != TTI::SK_Alternate)
return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
if (Kind == TTI::SK_Reverse) {
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
- unsigned Cost = 1;
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
+ int Cost = 1;
if (LT.second.getSizeInBits() > 128)
Cost = 3; // Extract + insert + copy.
if (Kind == TTI::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);
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
// The backend knows how to generate a single VEX.256 version of
// instruction VPBLENDW if the target supports AVX2.
return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
}
-unsigned X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) {
+int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
- std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
- std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
-
- static const TypeConversionCostTblEntry<MVT::SimpleValueType>
- SSE2ConvTbl[] = {
- // These are somewhat magic numbers justified by looking at the output of
- // Intel's IACA, running some kernels and making sure when we take
- // legalization into account the throughput will be overestimated.
- { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
- { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
- { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
- { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
- { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
- { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
- { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
- { 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, 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 },
- { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
- { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
- { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
- };
-
- if (ST->hasSSE2() && !ST->hasAVX()) {
- int Idx =
- ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
- if (Idx != -1)
- return LTSrc.first * SSE2ConvTbl[Idx].Cost;
- }
-
static const TypeConversionCostTblEntry<MVT::SimpleValueType>
AVX512ConversionTbl[] = {
{ ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 },
{ 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);
-
- // The function getSimpleVT only handles simple value types.
- if (!SrcTy.isSimple() || !DstTy.isSimple())
- return BaseT::getCastInstrCost(Opcode, Dst, Src);
-
static const TypeConversionCostTblEntry<MVT::SimpleValueType>
AVX2ConversionTbl[] = {
{ ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
{ ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 },
};
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ SSE2ConvTbl[] = {
+ // These are somewhat magic numbers justified by looking at the output of
+ // Intel's IACA, running some kernels and making sure when we take
+ // legalization into account the throughput will be overestimated.
+ { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
+ { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
+ { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
+ { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
+ { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
+ { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
+ { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
+ { 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, 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 },
+ { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
+ { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
+ { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
+ };
+
+ std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
+ std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
+
+ if (ST->hasSSE2() && !ST->hasAVX()) {
+ int Idx =
+ ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
+ if (Idx != -1)
+ return LTSrc.first * SSE2ConvTbl[Idx].Cost;
+ }
+
+ if (ST->hasAVX512()) {
+ int Idx = ConvertCostTableLookup(AVX512ConversionTbl, ISD, LTDest.second,
+ LTSrc.second);
+ if (Idx != -1)
+ return AVX512ConversionTbl[Idx].Cost;
+ }
+
+ EVT SrcTy = TLI->getValueType(DL, Src);
+ EVT DstTy = TLI->getValueType(DL, Dst);
+
+ // The function getSimpleVT only handles simple value types.
+ if (!SrcTy.isSimple() || !DstTy.isSimple())
+ return BaseT::getCastInstrCost(Opcode, Dst, Src);
+
if (ST->hasAVX2()) {
int Idx = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
DstTy.getSimpleVT(), SrcTy.getSimpleVT());
return BaseT::getCastInstrCost(Opcode, Dst, Src);
}
-unsigned X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
- Type *CondTy) {
+int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
// Legalize the type.
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
MVT MTy = LT.second;
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy);
}
-unsigned X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
- unsigned Index) {
+int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
assert(Val->isVectorTy() && "This must be a vector type");
if (Index != -1U) {
// Legalize the type.
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val);
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
// This type is legalized to a scalar type.
if (!LT.second.isVector())
return BaseT::getVectorInstrCost(Opcode, Val, Index);
}
-unsigned X86TTIImpl::getScalarizationOverhead(Type *Ty, bool Insert,
- bool Extract) {
+int X86TTIImpl::getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
assert (Ty->isVectorTy() && "Can only scalarize vectors");
- unsigned Cost = 0;
+ int Cost = 0;
for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
if (Insert)
return Cost;
}
-unsigned X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
- unsigned Alignment,
- unsigned AddressSpace) {
+int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
+ unsigned AddressSpace) {
// Handle non-power-of-two vectors such as <3 x float>
if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
unsigned NumElem = VTy->getVectorNumElements();
// Assume that all other non-power-of-two numbers are scalarized.
if (!isPowerOf2_32(NumElem)) {
- unsigned Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(),
- Alignment, AddressSpace);
- unsigned SplitCost = getScalarizationOverhead(Src,
- Opcode == Instruction::Load,
- Opcode==Instruction::Store);
+ int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
+ AddressSpace);
+ int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
+ Opcode == Instruction::Store);
return NumElem * Cost + SplitCost;
}
}
// Legalize the type.
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
"Invalid Opcode");
// Each load/store unit costs 1.
- unsigned Cost = LT.first * 1;
+ int Cost = LT.first * 1;
// On Sandybridge 256bit load/stores are double pumped
// (but not on Haswell).
return Cost;
}
-unsigned X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
- unsigned Alignment,
- unsigned AddressSpace) {
+int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
+ unsigned Alignment,
+ unsigned AddressSpace) {
VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
if (!SrcVTy)
// To calculate scalar take the regular cost, without mask
(Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy, 1)) ||
!isPowerOf2_32(NumElem)) {
// Scalarization
- unsigned MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
- unsigned ScalarCompareCost =
- getCmpSelInstrCost(Instruction::ICmp,
- Type::getInt8Ty(getGlobalContext()), NULL);
- unsigned BranchCost = getCFInstrCost(Instruction::Br);
- unsigned MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
-
- unsigned ValueSplitCost =
- getScalarizationOverhead(SrcVTy, Opcode == Instruction::Load,
- Opcode == Instruction::Store);
- unsigned MemopCost =
+ int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
+ int ScalarCompareCost = getCmpSelInstrCost(
+ Instruction::ICmp, Type::getInt8Ty(getGlobalContext()), nullptr);
+ int BranchCost = getCFInstrCost(Instruction::Br);
+ int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
+
+ int ValueSplitCost = getScalarizationOverhead(
+ SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store);
+ int MemopCost =
NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
Alignment, AddressSpace);
return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
}
// Legalize the type.
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(SrcVTy);
- unsigned Cost = 0;
- if (LT.second != TLI->getValueType(SrcVTy).getSimpleVT() &&
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
+ int Cost = 0;
+ if (LT.second != TLI->getValueType(DL, SrcVTy).getSimpleVT() &&
LT.second.getVectorNumElements() == NumElem)
// Promotion requires expand/truncate for data and a shuffle for mask.
- Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, 0) +
- getShuffleCost(TTI::SK_Alternate, MaskTy, 0, 0);
+ Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, nullptr) +
+ getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr);
else if (LT.second.getVectorNumElements() > NumElem) {
VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
return Cost+LT.first;
}
-unsigned X86TTIImpl::getAddressComputationCost(Type *Ty, bool IsComplex) {
+int X86TTIImpl::getAddressComputationCost(Type *Ty, bool IsComplex) {
// 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
return BaseT::getAddressComputationCost(Ty, IsComplex);
}
-unsigned X86TTIImpl::getReductionCost(unsigned Opcode, Type *ValTy,
- bool IsPairwise) {
+int X86TTIImpl::getReductionCost(unsigned Opcode, Type *ValTy,
+ bool IsPairwise) {
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
MVT MTy = LT.second;
/// \brief 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.
-unsigned X86TTIImpl::getIntImmCost(int64_t Val) {
+int X86TTIImpl::getIntImmCost(int64_t Val) {
if (Val == 0)
return TTI::TCC_Free;
return 2 * TTI::TCC_Basic;
}
-unsigned X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
+int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
// Split the constant into 64-bit chunks and calculate the cost for each
// chunk.
- unsigned Cost = 0;
+ int 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(1U, Cost);
+ return std::max(1, Cost);
}
-unsigned X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx,
- const APInt &Imm, Type *Ty) {
+int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
+ Type *Ty) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
case Instruction::Store:
ImmIdx = 0;
break;
+ case Instruction::And:
+ // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
+ // by using a 32-bit operation with implicit zero extension. Detect such
+ // immediates here as the normal path expects bit 31 to be sign extended.
+ if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
+ return TTI::TCC_Free;
+ // Fallthrough
case Instruction::Add:
case Instruction::Sub:
case Instruction::Mul:
case Instruction::SDiv:
case Instruction::URem:
case Instruction::SRem:
- case Instruction::And:
case Instruction::Or:
case Instruction::Xor:
case Instruction::ICmp:
}
if (Idx == ImmIdx) {
- unsigned NumConstants = (BitSize + 63) / 64;
- unsigned Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
+ int NumConstants = (BitSize + 63) / 64;
+ int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
return (Cost <= NumConstants * TTI::TCC_Basic)
- ? static_cast<unsigned>(TTI::TCC_Free)
+ ? static_cast<int>(TTI::TCC_Free)
: Cost;
}
return X86TTIImpl::getIntImmCost(Imm, Ty);
}
-unsigned X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
- const APInt &Imm, Type *Ty) {
+int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
+ Type *Ty) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, int Consecutive) {
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())
+ if (ST->hasAVX512() || ST->hasAVX2())
return true;
return false;
}
return isLegalMaskedLoad(DataType, Consecutive);
}
+bool X86TTIImpl::areInlineCompatible(const Function *Caller,
+ const Function *Callee) const {
+ const TargetMachine &TM = getTLI()->getTargetMachine();
+
+ // Work this as a subsetting of subtarget features.
+ const FeatureBitset &CallerBits =
+ TM.getSubtargetImpl(*Caller)->getFeatureBits();
+ const FeatureBitset &CalleeBits =
+ TM.getSubtargetImpl(*Callee)->getFeatureBits();
+
+ // FIXME: This is likely too limiting as it will include subtarget features
+ // that we might not care about for inlining, but it is conservatively
+ // correct.
+ return (CallerBits & CalleeBits) == CalleeBits;
+}