#include "ARMTargetMachine.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/CostTable.h"
+#include "llvm/Target/TargetLowering.h"
using namespace llvm;
// Declare the pass initialization routine locally as target-specific passes
namespace {
-class ARMTTI : public ImmutablePass, public TargetTransformInfo {
+class ARMTTI final : public ImmutablePass, public TargetTransformInfo {
const ARMBaseTargetMachine *TM;
const ARMSubtarget *ST;
const ARMTargetLowering *TLI;
initializeARMTTIPass(*PassRegistry::getPassRegistry());
}
- virtual void initializePass() {
+ virtual void initializePass() override {
pushTTIStack(this);
}
popTTIStack();
}
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const override {
TargetTransformInfo::getAnalysisUsage(AU);
}
static char ID;
/// Provide necessary pointer adjustments for the two base classes.
- virtual void *getAdjustedAnalysisPointer(const void *ID) {
+ virtual void *getAdjustedAnalysisPointer(const void *ID) override {
if (ID == &TargetTransformInfo::ID)
return (TargetTransformInfo*)this;
return this;
/// \name Scalar TTI Implementations
/// @{
-
- virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
+ using TargetTransformInfo::getIntImmCost;
+ virtual unsigned
+ getIntImmCost(const APInt &Imm, Type *Ty) const override;
/// @}
if (ST->isThumb1Only())
return 8;
- return 16;
+ return 13;
}
unsigned getRegisterBitWidth(bool Vector) const {
unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
OperandValueKind Op1Info = OK_AnyValue,
OperandValueKind Op2Info = OK_AnyValue) const;
+
+ unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
+ unsigned AddressSpace) const;
/// @}
};
// Load from constantpool.
return 3;
}
- return 2;
}
unsigned ARMTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
- Type *Src) const {
+ Type *Src) const {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
// Single to/from double precision conversions.
- static const CostTblEntry<MVT> NEONFltDblTbl[] = {
+ static const CostTblEntry<MVT::SimpleValueType> NEONFltDblTbl[] = {
// Vector fptrunc/fpext conversions.
{ ISD::FP_ROUND, MVT::v2f64, 2 },
{ ISD::FP_EXTEND, MVT::v2f32, 2 },
if (Src->isVectorTy() && ST->hasNEON() && (ISD == ISD::FP_ROUND ||
ISD == ISD::FP_EXTEND)) {
std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
- int Idx = CostTableLookup<MVT>(NEONFltDblTbl, array_lengthof(NEONFltDblTbl),
- ISD, LT.second);
+ int Idx = CostTableLookup(NEONFltDblTbl, ISD, LT.second);
if (Idx != -1)
return LT.first * NEONFltDblTbl[Idx].Cost;
}
// Some arithmetic, load and store operations have specific instructions
// to cast up/down their types automatically at no extra cost.
// TODO: Get these tables to know at least what the related operations are.
- static const TypeConversionCostTblEntry<MVT> NEONVectorConversionTbl[] = {
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ NEONVectorConversionTbl[] = {
{ ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0 },
{ ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0 },
{ ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
};
if (SrcTy.isVector() && ST->hasNEON()) {
- int Idx = ConvertCostTableLookup<MVT>(NEONVectorConversionTbl,
- array_lengthof(NEONVectorConversionTbl),
- ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT());
+ int Idx = ConvertCostTableLookup(NEONVectorConversionTbl, ISD,
+ DstTy.getSimpleVT(), SrcTy.getSimpleVT());
if (Idx != -1)
return NEONVectorConversionTbl[Idx].Cost;
}
// Scalar float to integer conversions.
- static const TypeConversionCostTblEntry<MVT> NEONFloatConversionTbl[] = {
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ NEONFloatConversionTbl[] = {
{ ISD::FP_TO_SINT, MVT::i1, MVT::f32, 2 },
{ ISD::FP_TO_UINT, MVT::i1, MVT::f32, 2 },
{ ISD::FP_TO_SINT, MVT::i1, MVT::f64, 2 },
{ ISD::FP_TO_UINT, MVT::i64, MVT::f64, 10 }
};
if (SrcTy.isFloatingPoint() && ST->hasNEON()) {
- int Idx = ConvertCostTableLookup<MVT>(NEONFloatConversionTbl,
- array_lengthof(NEONFloatConversionTbl),
- ISD, DstTy.getSimpleVT(),
- SrcTy.getSimpleVT());
+ int Idx = ConvertCostTableLookup(NEONFloatConversionTbl, ISD,
+ DstTy.getSimpleVT(), SrcTy.getSimpleVT());
if (Idx != -1)
return NEONFloatConversionTbl[Idx].Cost;
}
// Scalar integer to float conversions.
- static const TypeConversionCostTblEntry<MVT> NEONIntegerConversionTbl[] = {
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ NEONIntegerConversionTbl[] = {
{ ISD::SINT_TO_FP, MVT::f32, MVT::i1, 2 },
{ ISD::UINT_TO_FP, MVT::f32, MVT::i1, 2 },
{ ISD::SINT_TO_FP, MVT::f64, MVT::i1, 2 },
};
if (SrcTy.isInteger() && ST->hasNEON()) {
- int Idx = ConvertCostTableLookup<MVT>(NEONIntegerConversionTbl,
- array_lengthof(NEONIntegerConversionTbl),
- ISD, DstTy.getSimpleVT(),
- SrcTy.getSimpleVT());
+ int Idx = ConvertCostTableLookup(NEONIntegerConversionTbl, ISD,
+ DstTy.getSimpleVT(), SrcTy.getSimpleVT());
if (Idx != -1)
return NEONIntegerConversionTbl[Idx].Cost;
}
// Scalar integer conversion costs.
- static const TypeConversionCostTblEntry<MVT> ARMIntegerConversionTbl[] = {
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ ARMIntegerConversionTbl[] = {
// i16 -> i64 requires two dependent operations.
{ ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 },
};
if (SrcTy.isInteger()) {
- int Idx =
- ConvertCostTableLookup<MVT>(ARMIntegerConversionTbl,
- array_lengthof(ARMIntegerConversionTbl),
- ISD, DstTy.getSimpleVT(),
- SrcTy.getSimpleVT());
+ int Idx = ConvertCostTableLookup(ARMIntegerConversionTbl, ISD,
+ DstTy.getSimpleVT(), SrcTy.getSimpleVT());
if (Idx != -1)
return ARMIntegerConversionTbl[Idx].Cost;
}
// On NEON a a vector select gets lowered to vbsl.
if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
// Lowering of some vector selects is currently far from perfect.
- static const TypeConversionCostTblEntry<MVT> NEONVectorSelectTbl[] = {
+ static const TypeConversionCostTblEntry<MVT::SimpleValueType>
+ NEONVectorSelectTbl[] = {
{ ISD::SELECT, MVT::v16i1, MVT::v16i16, 2*16 + 1 + 3*1 + 4*1 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i32, 4*8 + 1*3 + 1*4 + 1*2 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i32, 4*16 + 1*6 + 1*8 + 1*4 },
EVT SelCondTy = TLI->getValueType(CondTy);
EVT SelValTy = TLI->getValueType(ValTy);
- int Idx = ConvertCostTableLookup<MVT>(NEONVectorSelectTbl,
- array_lengthof(NEONVectorSelectTbl),
- ISD, SelCondTy.getSimpleVT(),
- SelValTy.getSimpleVT());
- if (Idx != -1)
- return NEONVectorSelectTbl[Idx].Cost;
+ if (SelCondTy.isSimple() && SelValTy.isSimple()) {
+ int Idx = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
+ SelCondTy.getSimpleVT(),
+ SelValTy.getSimpleVT());
+ if (Idx != -1)
+ return NEONVectorSelectTbl[Idx].Cost;
+ }
std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
return LT.first;
}
unsigned ARMTTI::getAddressComputationCost(Type *Ty, bool IsComplex) 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 = 10;
+
+ if (Ty->isVectorTy() && IsComplex)
+ return NumVectorInstToHideOverhead;
+
// In many cases the address computation is not merged into the instruction
// addressing mode.
return 1;
if (Kind != SK_Reverse)
return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
- static const CostTblEntry<MVT> NEONShuffleTbl[] = {
+ static const CostTblEntry<MVT::SimpleValueType> NEONShuffleTbl[] = {
// Reverse shuffle cost one instruction if we are shuffling within a double
// word (vrev) or two if we shuffle a quad word (vrev, vext).
{ ISD::VECTOR_SHUFFLE, MVT::v2i32, 1 },
std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
- int Idx = CostTableLookup<MVT>(NEONShuffleTbl, array_lengthof(NEONShuffleTbl),
- ISD::VECTOR_SHUFFLE, LT.second);
+ int Idx = CostTableLookup(NEONShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
if (Idx == -1)
return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
return LT.first * NEONShuffleTbl[Idx].Cost;
}
-unsigned ARMTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Op1Info,
+unsigned ARMTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
+ OperandValueKind Op1Info,
OperandValueKind Op2Info) const {
int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
const unsigned FunctionCallDivCost = 20;
const unsigned ReciprocalDivCost = 10;
- static const CostTblEntry<MVT> CostTbl[] = {
+ static const CostTblEntry<MVT::SimpleValueType> CostTbl[] = {
// Division.
// These costs are somewhat random. Choose a cost of 20 to indicate that
// vectorizing devision (added function call) is going to be very expensive.
int Idx = -1;
if (ST->hasNEON())
- Idx = CostTableLookup<MVT>(CostTbl, array_lengthof(CostTbl), ISDOpcode,
- LT.second);
+ Idx = CostTableLookup(CostTbl, ISDOpcode, LT.second);
if (Idx != -1)
return LT.first * CostTbl[Idx].Cost;
-
- return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
- Op2Info);
+ unsigned Cost =
+ TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
+
+ // This is somewhat of a hack. The problem that we are facing is that SROA
+ // creates a sequence of shift, and, or instructions to construct values.
+ // These sequences are recognized by the ISel and have zero-cost. Not so for
+ // the vectorized code. Because we have support for v2i64 but not i64 those
+ // sequences look particularly beneficial to vectorize.
+ // To work around this we increase the cost of v2i64 operations to make them
+ // seem less beneficial.
+ if (LT.second == MVT::v2i64 &&
+ Op2Info == TargetTransformInfo::OK_UniformConstantValue)
+ Cost += 4;
+
+ return Cost;
}
+unsigned ARMTTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
+ unsigned AddressSpace) const {
+ std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
+
+ if (Src->isVectorTy() && Alignment != 16 &&
+ Src->getVectorElementType()->isDoubleTy()) {
+ // Unaligned loads/stores are extremely inefficient.
+ // We need 4 uops for vst.1/vld.1 vs 1uop for vldr/vstr.
+ return LT.first * 4;
+ }
+ return LT.first;
+}