1 //===- BasicTargetTransformInfo.cpp - Basic target-independent TTI impl ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file provides the implementation of a basic TargetTransformInfo pass
11 /// predicated on the target abstractions present in the target independent
12 /// code generator. It uses these (primarily TargetLowering) to model as much
13 /// of the TTI query interface as possible. It is included by most targets so
14 /// that they can specialize only a small subset of the query space.
16 //===----------------------------------------------------------------------===//
18 #include "llvm/CodeGen/Passes.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/Target/TargetLowering.h"
24 #define DEBUG_TYPE "basictti"
28 class BasicTTI final : public ImmutablePass, public TargetTransformInfo {
29 const TargetMachine *TM;
31 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
32 /// are set if the result needs to be inserted and/or extracted from vectors.
33 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
35 const TargetLoweringBase *getTLI() const { return TM->getTargetLowering(); }
38 BasicTTI() : ImmutablePass(ID), TM(nullptr) {
39 llvm_unreachable("This pass cannot be directly constructed");
42 BasicTTI(const TargetMachine *TM) : ImmutablePass(ID), TM(TM) {
43 initializeBasicTTIPass(*PassRegistry::getPassRegistry());
46 void initializePass() override {
50 void getAnalysisUsage(AnalysisUsage &AU) const override {
51 TargetTransformInfo::getAnalysisUsage(AU);
54 /// Pass identification.
57 /// Provide necessary pointer adjustments for the two base classes.
58 void *getAdjustedAnalysisPointer(const void *ID) override {
59 if (ID == &TargetTransformInfo::ID)
60 return (TargetTransformInfo*)this;
64 bool hasBranchDivergence() const override;
66 /// \name Scalar TTI Implementations
69 bool isLegalAddImmediate(int64_t imm) const override;
70 bool isLegalICmpImmediate(int64_t imm) const override;
71 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
72 int64_t BaseOffset, bool HasBaseReg,
73 int64_t Scale) const override;
74 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
75 int64_t BaseOffset, bool HasBaseReg,
76 int64_t Scale) const override;
77 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
78 bool isTypeLegal(Type *Ty) const override;
79 unsigned getJumpBufAlignment() const override;
80 unsigned getJumpBufSize() const override;
81 bool shouldBuildLookupTables() const override;
82 bool haveFastSqrt(Type *Ty) const override;
83 void getUnrollingPreferences(Loop *L,
84 UnrollingPreferences &UP) const override;
88 /// \name Vector TTI Implementations
91 unsigned getNumberOfRegisters(bool Vector) const override;
92 unsigned getMaximumUnrollFactor() const override;
93 unsigned getRegisterBitWidth(bool Vector) const override;
94 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
95 OperandValueKind) const override;
96 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
97 int Index, Type *SubTp) const override;
98 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
99 Type *Src) const override;
100 unsigned getCFInstrCost(unsigned Opcode) const override;
101 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
102 Type *CondTy) const override;
103 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
104 unsigned Index) const override;
105 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
106 unsigned AddressSpace) const override;
107 unsigned getIntrinsicInstrCost(Intrinsic::ID, Type *RetTy,
108 ArrayRef<Type*> Tys) const override;
109 unsigned getNumberOfParts(Type *Tp) const override;
110 unsigned getAddressComputationCost( Type *Ty, bool IsComplex) const override;
111 unsigned getReductionCost(unsigned Opcode, Type *Ty,
112 bool IsPairwise) const override;
119 INITIALIZE_AG_PASS(BasicTTI, TargetTransformInfo, "basictti",
120 "Target independent code generator's TTI", true, true, false)
121 char BasicTTI::ID = 0;
124 llvm::createBasicTargetTransformInfoPass(const TargetMachine *TM) {
125 return new BasicTTI(TM);
128 bool BasicTTI::hasBranchDivergence() const { return false; }
130 bool BasicTTI::isLegalAddImmediate(int64_t imm) const {
131 return getTLI()->isLegalAddImmediate(imm);
134 bool BasicTTI::isLegalICmpImmediate(int64_t imm) const {
135 return getTLI()->isLegalICmpImmediate(imm);
138 bool BasicTTI::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
139 int64_t BaseOffset, bool HasBaseReg,
140 int64_t Scale) const {
141 TargetLoweringBase::AddrMode AM;
143 AM.BaseOffs = BaseOffset;
144 AM.HasBaseReg = HasBaseReg;
146 return getTLI()->isLegalAddressingMode(AM, Ty);
149 int BasicTTI::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
150 int64_t BaseOffset, bool HasBaseReg,
151 int64_t Scale) const {
152 TargetLoweringBase::AddrMode AM;
154 AM.BaseOffs = BaseOffset;
155 AM.HasBaseReg = HasBaseReg;
157 return getTLI()->getScalingFactorCost(AM, Ty);
160 bool BasicTTI::isTruncateFree(Type *Ty1, Type *Ty2) const {
161 return getTLI()->isTruncateFree(Ty1, Ty2);
164 bool BasicTTI::isTypeLegal(Type *Ty) const {
165 EVT T = getTLI()->getValueType(Ty);
166 return getTLI()->isTypeLegal(T);
169 unsigned BasicTTI::getJumpBufAlignment() const {
170 return getTLI()->getJumpBufAlignment();
173 unsigned BasicTTI::getJumpBufSize() const {
174 return getTLI()->getJumpBufSize();
177 bool BasicTTI::shouldBuildLookupTables() const {
178 const TargetLoweringBase *TLI = getTLI();
179 return TLI->supportJumpTables() &&
180 (TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
181 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
184 bool BasicTTI::haveFastSqrt(Type *Ty) const {
185 const TargetLoweringBase *TLI = getTLI();
186 EVT VT = TLI->getValueType(Ty);
187 return TLI->isTypeLegal(VT) && TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
190 void BasicTTI::getUnrollingPreferences(Loop *, UnrollingPreferences &) const { }
192 //===----------------------------------------------------------------------===//
194 // Calls used by the vectorizers.
196 //===----------------------------------------------------------------------===//
198 unsigned BasicTTI::getScalarizationOverhead(Type *Ty, bool Insert,
199 bool Extract) const {
200 assert (Ty->isVectorTy() && "Can only scalarize vectors");
203 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
205 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
207 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
213 unsigned BasicTTI::getNumberOfRegisters(bool Vector) const {
217 unsigned BasicTTI::getRegisterBitWidth(bool Vector) const {
221 unsigned BasicTTI::getMaximumUnrollFactor() const {
225 unsigned BasicTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
227 OperandValueKind) const {
228 // Check if any of the operands are vector operands.
229 const TargetLoweringBase *TLI = getTLI();
230 int ISD = TLI->InstructionOpcodeToISD(Opcode);
231 assert(ISD && "Invalid opcode");
233 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
235 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
236 // Assume that floating point arithmetic operations cost twice as much as
237 // integer operations.
238 unsigned OpCost = (IsFloat ? 2 : 1);
240 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
241 // The operation is legal. Assume it costs 1.
242 // If the type is split to multiple registers, assume that there is some
244 // TODO: Once we have extract/insert subvector cost we need to use them.
246 return LT.first * 2 * OpCost;
247 return LT.first * 1 * OpCost;
250 if (!TLI->isOperationExpand(ISD, LT.second)) {
251 // If the operation is custom lowered then assume
252 // thare the code is twice as expensive.
253 return LT.first * 2 * OpCost;
256 // Else, assume that we need to scalarize this op.
257 if (Ty->isVectorTy()) {
258 unsigned Num = Ty->getVectorNumElements();
259 unsigned Cost = TopTTI->getArithmeticInstrCost(Opcode, Ty->getScalarType());
260 // return the cost of multiple scalar invocation plus the cost of inserting
261 // and extracting the values.
262 return getScalarizationOverhead(Ty, true, true) + Num * Cost;
265 // We don't know anything about this scalar instruction.
269 unsigned BasicTTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
274 unsigned BasicTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
276 const TargetLoweringBase *TLI = getTLI();
277 int ISD = TLI->InstructionOpcodeToISD(Opcode);
278 assert(ISD && "Invalid opcode");
280 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(Src);
281 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(Dst);
283 // Check for NOOP conversions.
284 if (SrcLT.first == DstLT.first &&
285 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
287 // Bitcast between types that are legalized to the same type are free.
288 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
292 if (Opcode == Instruction::Trunc &&
293 TLI->isTruncateFree(SrcLT.second, DstLT.second))
296 if (Opcode == Instruction::ZExt &&
297 TLI->isZExtFree(SrcLT.second, DstLT.second))
300 // If the cast is marked as legal (or promote) then assume low cost.
301 if (SrcLT.first == DstLT.first &&
302 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
305 // Handle scalar conversions.
306 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
308 // Scalar bitcasts are usually free.
309 if (Opcode == Instruction::BitCast)
312 // Just check the op cost. If the operation is legal then assume it costs 1.
313 if (!TLI->isOperationExpand(ISD, DstLT.second))
316 // Assume that illegal scalar instruction are expensive.
320 // Check vector-to-vector casts.
321 if (Dst->isVectorTy() && Src->isVectorTy()) {
323 // If the cast is between same-sized registers, then the check is simple.
324 if (SrcLT.first == DstLT.first &&
325 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
327 // Assume that Zext is done using AND.
328 if (Opcode == Instruction::ZExt)
331 // Assume that sext is done using SHL and SRA.
332 if (Opcode == Instruction::SExt)
335 // Just check the op cost. If the operation is legal then assume it costs
336 // 1 and multiply by the type-legalization overhead.
337 if (!TLI->isOperationExpand(ISD, DstLT.second))
338 return SrcLT.first * 1;
341 // If we are converting vectors and the operation is illegal, or
342 // if the vectors are legalized to different types, estimate the
343 // scalarization costs.
344 unsigned Num = Dst->getVectorNumElements();
345 unsigned Cost = TopTTI->getCastInstrCost(Opcode, Dst->getScalarType(),
346 Src->getScalarType());
348 // Return the cost of multiple scalar invocation plus the cost of
349 // inserting and extracting the values.
350 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
353 // We already handled vector-to-vector and scalar-to-scalar conversions. This
354 // is where we handle bitcast between vectors and scalars. We need to assume
355 // that the conversion is scalarized in one way or another.
356 if (Opcode == Instruction::BitCast)
357 // Illegal bitcasts are done by storing and loading from a stack slot.
358 return (Src->isVectorTy()? getScalarizationOverhead(Src, false, true):0) +
359 (Dst->isVectorTy()? getScalarizationOverhead(Dst, true, false):0);
361 llvm_unreachable("Unhandled cast");
364 unsigned BasicTTI::getCFInstrCost(unsigned Opcode) const {
365 // Branches are assumed to be predicted.
369 unsigned BasicTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
370 Type *CondTy) const {
371 const TargetLoweringBase *TLI = getTLI();
372 int ISD = TLI->InstructionOpcodeToISD(Opcode);
373 assert(ISD && "Invalid opcode");
375 // Selects on vectors are actually vector selects.
376 if (ISD == ISD::SELECT) {
377 assert(CondTy && "CondTy must exist");
378 if (CondTy->isVectorTy())
382 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
384 if (!TLI->isOperationExpand(ISD, LT.second)) {
385 // The operation is legal. Assume it costs 1. Multiply
386 // by the type-legalization overhead.
390 // Otherwise, assume that the cast is scalarized.
391 if (ValTy->isVectorTy()) {
392 unsigned Num = ValTy->getVectorNumElements();
394 CondTy = CondTy->getScalarType();
395 unsigned Cost = TopTTI->getCmpSelInstrCost(Opcode, ValTy->getScalarType(),
398 // Return the cost of multiple scalar invocation plus the cost of inserting
399 // and extracting the values.
400 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
403 // Unknown scalar opcode.
407 unsigned BasicTTI::getVectorInstrCost(unsigned Opcode, Type *Val,
408 unsigned Index) const {
409 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Val->getScalarType());
414 unsigned BasicTTI::getMemoryOpCost(unsigned Opcode, Type *Src,
416 unsigned AddressSpace) const {
417 assert(!Src->isVoidTy() && "Invalid type");
418 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Src);
420 // Assuming that all loads of legal types cost 1.
421 unsigned Cost = LT.first;
423 if (Src->isVectorTy() &&
424 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
425 // This is a vector load that legalizes to a larger type than the vector
426 // itself. Unless the corresponding extending load or truncating store is
427 // legal, then this will scalarize.
428 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
429 EVT MemVT = getTLI()->getValueType(Src, true);
430 if (MemVT.isSimple() && MemVT != MVT::Other) {
431 if (Opcode == Instruction::Store)
432 LA = getTLI()->getTruncStoreAction(LT.second, MemVT.getSimpleVT());
434 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, MemVT.getSimpleVT());
437 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
438 // This is a vector load/store for some illegal type that is scalarized.
439 // We must account for the cost of building or decomposing the vector.
440 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
441 Opcode == Instruction::Store);
448 unsigned BasicTTI::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
449 ArrayRef<Type *> Tys) const {
453 // Assume that we need to scalarize this intrinsic.
454 unsigned ScalarizationCost = 0;
455 unsigned ScalarCalls = 1;
456 if (RetTy->isVectorTy()) {
457 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
458 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
460 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
461 if (Tys[i]->isVectorTy()) {
462 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
463 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
467 return ScalarCalls + ScalarizationCost;
469 // Look for intrinsics that can be lowered directly or turned into a scalar
471 case Intrinsic::sqrt: ISD = ISD::FSQRT; break;
472 case Intrinsic::sin: ISD = ISD::FSIN; break;
473 case Intrinsic::cos: ISD = ISD::FCOS; break;
474 case Intrinsic::exp: ISD = ISD::FEXP; break;
475 case Intrinsic::exp2: ISD = ISD::FEXP2; break;
476 case Intrinsic::log: ISD = ISD::FLOG; break;
477 case Intrinsic::log10: ISD = ISD::FLOG10; break;
478 case Intrinsic::log2: ISD = ISD::FLOG2; break;
479 case Intrinsic::fabs: ISD = ISD::FABS; break;
480 case Intrinsic::copysign: ISD = ISD::FCOPYSIGN; break;
481 case Intrinsic::floor: ISD = ISD::FFLOOR; break;
482 case Intrinsic::ceil: ISD = ISD::FCEIL; break;
483 case Intrinsic::trunc: ISD = ISD::FTRUNC; break;
484 case Intrinsic::nearbyint:
485 ISD = ISD::FNEARBYINT; break;
486 case Intrinsic::rint: ISD = ISD::FRINT; break;
487 case Intrinsic::round: ISD = ISD::FROUND; break;
488 case Intrinsic::pow: ISD = ISD::FPOW; break;
489 case Intrinsic::fma: ISD = ISD::FMA; break;
490 case Intrinsic::fmuladd: ISD = ISD::FMA; break; // FIXME: mul + add?
491 case Intrinsic::lifetime_start:
492 case Intrinsic::lifetime_end:
496 const TargetLoweringBase *TLI = getTLI();
497 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(RetTy);
499 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
500 // The operation is legal. Assume it costs 1.
501 // If the type is split to multiple registers, assume that thre is some
503 // TODO: Once we have extract/insert subvector cost we need to use them.
509 if (!TLI->isOperationExpand(ISD, LT.second)) {
510 // If the operation is custom lowered then assume
511 // thare the code is twice as expensive.
515 // Else, assume that we need to scalarize this intrinsic. For math builtins
516 // this will emit a costly libcall, adding call overhead and spills. Make it
518 if (RetTy->isVectorTy()) {
519 unsigned Num = RetTy->getVectorNumElements();
520 unsigned Cost = TopTTI->getIntrinsicInstrCost(IID, RetTy->getScalarType(),
522 return 10 * Cost * Num;
525 // This is going to be turned into a library call, make it expensive.
529 unsigned BasicTTI::getNumberOfParts(Type *Tp) const {
530 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Tp);
534 unsigned BasicTTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
538 unsigned BasicTTI::getReductionCost(unsigned Opcode, Type *Ty,
539 bool IsPairwise) const {
540 assert(Ty->isVectorTy() && "Expect a vector type");
541 unsigned NumVecElts = Ty->getVectorNumElements();
542 unsigned NumReduxLevels = Log2_32(NumVecElts);
543 unsigned ArithCost = NumReduxLevels *
544 TopTTI->getArithmeticInstrCost(Opcode, Ty);
545 // Assume the pairwise shuffles add a cost.
546 unsigned ShuffleCost =
547 NumReduxLevels * (IsPairwise + 1) *
548 TopTTI->getShuffleCost(SK_ExtractSubvector, Ty, NumVecElts / 2, Ty);
549 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);