///
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "x86tti"
-#include "X86.h"
-#include "X86TargetMachine.h"
+#include "X86TargetTransformInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/CodeGen/BasicTTIImpl.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/CostTable.h"
#include "llvm/Target/TargetLowering.h"
-using namespace llvm;
-
-// Declare the pass initialization routine locally as target-specific passes
-// don't havve a target-wide initialization entry point, and so we rely on the
-// pass constructor initialization.
-namespace llvm {
-void initializeX86TTIPass(PassRegistry &);
-}
-
-namespace {
-
-class X86TTI final : public ImmutablePass, public TargetTransformInfo {
- const X86Subtarget *ST;
- const X86TargetLowering *TLI;
-
- /// Estimate the overhead of scalarizing an instruction. Insert and Extract
- /// are set if the result needs to be inserted and/or extracted from vectors.
- unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
-
-public:
- X86TTI() : ImmutablePass(ID), ST(0), TLI(0) {
- llvm_unreachable("This pass cannot be directly constructed");
- }
-
- X86TTI(const X86TargetMachine *TM)
- : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
- TLI(TM->getTargetLowering()) {
- initializeX86TTIPass(*PassRegistry::getPassRegistry());
- }
-
- void initializePass() override {
- pushTTIStack(this);
- }
-
- void getAnalysisUsage(AnalysisUsage &AU) const override {
- TargetTransformInfo::getAnalysisUsage(AU);
- }
- /// Pass identification.
- static char ID;
-
- /// Provide necessary pointer adjustments for the two base classes.
- void *getAdjustedAnalysisPointer(const void *ID) override {
- if (ID == &TargetTransformInfo::ID)
- return (TargetTransformInfo*)this;
- return this;
- }
-
- /// \name Scalar TTI Implementations
- /// @{
- PopcntSupportKind getPopcntSupport(unsigned TyWidth) const override;
-
- /// @}
-
- /// \name Vector TTI Implementations
- /// @{
-
- unsigned getNumberOfRegisters(bool Vector) const override;
- unsigned getRegisterBitWidth(bool Vector) const override;
- unsigned getMaximumUnrollFactor() const override;
- unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
- OperandValueKind) const override;
- unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
- int Index, Type *SubTp) const override;
- unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
- Type *Src) const override;
- unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
- Type *CondTy) const override;
- unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
- unsigned Index) const override;
- unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
- unsigned AddressSpace) const override;
-
- unsigned getAddressComputationCost(Type *PtrTy,
- bool IsComplex) const override;
-
- unsigned getReductionCost(unsigned Opcode, Type *Ty,
- bool IsPairwiseForm) const override;
-
- unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override;
-
- unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
- Type *Ty) const override;
- unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
- Type *Ty) const override;
-
- /// @}
-};
-
-} // end anonymous namespace
-
-INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti",
- "X86 Target Transform Info", true, true, false)
-char X86TTI::ID = 0;
-
-ImmutablePass *
-llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) {
- return new X86TTI(TM);
-}
+using namespace llvm;
+#define DEBUG_TYPE "x86tti"
//===----------------------------------------------------------------------===//
//
//
//===----------------------------------------------------------------------===//
-X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const {
+TargetTransformInfo::PopcntSupportKind
+X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
// TODO: Currently the __builtin_popcount() implementation using SSE3
// instructions is inefficient. Once the problem is fixed, we should
// call ST->hasSSE3() instead of ST->hasPOPCNT().
- return ST->hasPOPCNT() ? PSK_FastHardware : PSK_Software;
+ return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
}
-unsigned X86TTI::getNumberOfRegisters(bool Vector) const {
+unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) {
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 {
+unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) {
if (Vector) {
+ if (ST->hasAVX512()) return 512;
if (ST->hasAVX()) return 256;
if (ST->hasSSE1()) return 128;
return 0;
if (ST->is64Bit())
return 64;
- return 32;
+ return 32;
}
-unsigned X86TTI::getMaximumUnrollFactor() const {
+unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
+ // If the loop will not be vectorized, don't interleave the loop.
+ // Let regular unroll to unroll the loop, which saves the overflow
+ // check and memory check cost.
+ if (VF == 1)
+ return 1;
+
if (ST->isAtom())
return 1;
return 2;
}
-unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
- OperandValueKind Op1Info,
- OperandValueKind Op2Info) const {
+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");
+ 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.
+ 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);
+ Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
+ TargetTransformInfo::OP_None,
+ TargetTransformInfo::OP_None);
+
+ return Cost;
+ }
+
+ 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::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
+ };
+
+ if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
+ ST->hasAVX2()) {
+ int Idx = CostTableLookup(AVX2UniformConstCostTable, ISD, LT.second);
+ if (Idx != -1)
+ 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 },
+ };
+
+ 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;
+ }
+
+ 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;
+ }
- { ISD::SHL, MVT::v32i8, 42 }, // cmpeqb sequence.
- { ISD::SHL, MVT::v16i16, 16*10 }, // Scalarized.
+ 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::v16i16, 8*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::v16i16, 16*10 }, // Scalarized.
- { ISD::SRA, MVT::v4i64, 4*10 }, // Scalarized.
+ { ISD::SRA, MVT::v32i8, 24 }, // vpblendvb sequence.
+ { ISD::SRA, MVT::v16i16, 10 }, // extend/vpsravd/pack sequence.
+ { 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 },
};
- // 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
+ { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
+ { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
};
if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
ST->hasSSE2()) {
+ // pmuldq sequence.
+ if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
+ return LT.first * 15;
+
int Idx = CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second);
if (Idx != -1)
return LT.first * SSE2UniformConstCostTable[Idx].Cost;
// 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 LT.first * 6;
// Fallback to the default implementation.
- return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
- Op2Info);
+ return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
}
-unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
- Type *SubTp) const {
- // We only estimate the cost of reverse shuffles.
- if (Kind != SK_Reverse)
- return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, 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);
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
- unsigned Cost = 1;
- if (LT.second.getSizeInBits() > 128)
- Cost = 3; // Extract + insert + copy.
+ if (Kind == TTI::SK_Reverse) {
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
+ int Cost = 1;
+ if (LT.second.getSizeInBits() > 128)
+ Cost = 3; // Extract + insert + copy.
- // Multiple by the number of parts.
- return Cost * LT.first;
-}
+ // Multiple by the number of parts.
+ return Cost * LT.first;
+ }
-unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
- int ISD = TLI->InstructionOpcodeToISD(Opcode);
- assert(ISD && "Invalid opcode");
+ 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<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
- std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
- std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
+ // 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 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, 15 },
- { 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 },
- };
+ static const CostTblEntry<MVT::SimpleValueType> AVXAltShuffleTbl[] = {
+ {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1}, // vblendpd
+ {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1}, // vblendpd
- if (ST->hasSSE2() && !ST->hasAVX()) {
- int Idx =
- ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
+ {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}
+ };
+
+ // Fall-back (SSE3 and SSE2).
+ int Idx = CostTableLookup(SSEAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
if (Idx != -1)
- return LTSrc.first * SSE2ConvTbl[Idx].Cost;
+ return LT.first * SSEAltShuffleTbl[Idx].Cost;
+ return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
}
- EVT SrcTy = TLI->getValueType(Src);
- EVT DstTy = TLI->getValueType(Dst);
+ return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
+}
- // The function getSimpleVT only handles simple value types.
- if (!SrcTy.isSimple() || !DstTy.isSimple())
- return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
+int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) {
+ int ISD = TLI->InstructionOpcodeToISD(Opcode);
+ assert(ISD && "Invalid opcode");
+
+ 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 },
+ };
static const TypeConversionCostTblEntry<MVT::SimpleValueType>
AVX2ConversionTbl[] = {
{ 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>
// problem is that the inserts form a read-modify-write chain so latency
// should be factored in too. Inflating the cost per element by 1.
{ ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 },
+ { 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 AVXConversionTbl[Idx].Cost;
}
- return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
+ return BaseT::getCastInstrCost(Opcode, Dst, Src);
}
-unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
- Type *CondTy) const {
+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;
{ 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)
return LT.first * SSE42CostTbl[Idx].Cost;
}
- return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
+ return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy);
}
-unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val,
- unsigned Index) const {
+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 0;
}
- return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index);
+ return BaseT::getVectorInstrCost(Opcode, Val, Index);
}
-unsigned X86TTI::getScalarizationOverhead(Type *Ty, bool Insert,
- bool Extract) const {
+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)
- Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
+ Cost += getVectorInstrCost(Instruction::InsertElement, Ty, i);
if (Extract)
- Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
+ Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, i);
}
return Cost;
}
-unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
- unsigned AddressSpace) const {
+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 = TargetTransformInfo::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 X86TTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
+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
+ return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace);
+
+ unsigned NumElem = SrcVTy->getVectorNumElements();
+ VectorType *MaskTy =
+ VectorType::get(Type::getInt8Ty(getGlobalContext()), NumElem);
+ if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy, 1)) ||
+ (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy, 1)) ||
+ !isPowerOf2_32(NumElem)) {
+ // Scalarization
+ 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<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, nullptr) +
+ getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr);
+
+ else if (LT.second.getVectorNumElements() > NumElem) {
+ VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
+ LT.second.getVectorNumElements());
+ // Expanding requires fill mask with zeroes
+ Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
+ }
+ if (!ST->hasAVX512())
+ return Cost + LT.first*4; // Each maskmov costs 4
+
+ // AVX-512 masked load/store is cheapper
+ return Cost+LT.first;
+}
+
+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
if (Ty->isVectorTy() && IsComplex)
return NumVectorInstToHideOverhead;
- return TargetTransformInfo::getAddressComputationCost(Ty, IsComplex);
+ return BaseT::getAddressComputationCost(Ty, IsComplex);
}
-unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
- bool IsPairwise) const {
-
- std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
-
+int X86TTIImpl::getReductionCost(unsigned Opcode, Type *ValTy,
+ bool IsPairwise) {
+
+ std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, 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 TargetTransformInfo::getReductionCost(Opcode, ValTy, IsPairwise);
+ return BaseT::getReductionCost(Opcode, ValTy, IsPairwise);
+}
+
+/// \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.
+int X86TTIImpl::getIntImmCost(int64_t Val) {
+ if (Val == 0)
+ return TTI::TCC_Free;
+
+ if (isInt<32>(Val))
+ return TTI::TCC_Basic;
+
+ return 2 * TTI::TCC_Basic;
}
-unsigned X86TTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
+int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
if (BitSize == 0)
return ~0U;
- if (Imm == 0)
- return TCC_Free;
+ // Never hoist constants larger than 128bit, because this might lead to
+ // incorrect code generation or assertions in codegen.
+ // Fixme: Create a cost model for types larger than i128 once the codegen
+ // issues have been fixed.
+ if (BitSize > 128)
+ return TTI::TCC_Free;
- if (Imm.getBitWidth() <= 64 &&
- (isInt<32>(Imm.getSExtValue()) || isUInt<32>(Imm.getZExtValue())))
- return TCC_Basic;
- else
- return 2 * TCC_Basic;
+ if (Imm == 0)
+ return TTI::TCC_Free;
+
+ // 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.
+ 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(1, Cost);
}
-unsigned X86TTI::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
- Type *Ty) const {
+int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
+ Type *Ty) {
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 ~0U;
+ return TTI::TCC_Free;
unsigned ImmIdx = ~0U;
switch (Opcode) {
- default: return TCC_Free;
+ default:
+ return TTI::TCC_Free;
case Instruction::GetElementPtr:
+ // Always hoist the base address of a GetElementPtr. This prevents the
+ // creation of new constants for every base constant that gets constant
+ // folded with the offset.
if (Idx == 0)
- return 2 * TCC_Basic;
- return TCC_Free;
+ return 2 * TTI::TCC_Basic;
+ return TTI::TCC_Free;
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::Shl:
- case Instruction::LShr:
- case Instruction::AShr:
- 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:
break;
}
- if ((Idx == ImmIdx) &&
- Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
- return TCC_Free;
+ if (Idx == ImmIdx) {
+ int NumConstants = (BitSize + 63) / 64;
+ int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
+ return (Cost <= NumConstants * TTI::TCC_Basic)
+ ? static_cast<int>(TTI::TCC_Free)
+ : Cost;
+ }
- return X86TTI::getIntImmCost(Imm, Ty);
+ return X86TTIImpl::getIntImmCost(Imm, Ty);
}
-unsigned X86TTI::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
- const APInt &Imm, Type *Ty) const {
+int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
+ Type *Ty) {
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 ~0U;
+ return TTI::TCC_Free;
switch (IID) {
- default: return TCC_Free;
+ default:
+ return TTI::TCC_Free;
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow:
if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
- return TCC_Free;
+ return TTI::TCC_Free;
break;
case Intrinsic::experimental_stackmap:
if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
- return TCC_Free;
+ return TTI::TCC_Free;
break;
case Intrinsic::experimental_patchpoint_void:
case Intrinsic::experimental_patchpoint_i64:
if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
- return TCC_Free;
+ return TTI::TCC_Free;
break;
}
- return X86TTI::getIntImmCost(Imm, Ty);
+ return X86TTIImpl::getIntImmCost(Imm, Ty);
+}
+
+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())
+ return true;
+ return false;
+}
+
+bool X86TTIImpl::isLegalMaskedStore(Type *DataType, int Consecutive) {
+ 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;
}