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 #define DEBUG_TYPE "basictti"
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/Target/TargetLowering.h"
27 class BasicTTI final : public ImmutablePass, public TargetTransformInfo {
28 const TargetMachine *TM;
30 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
31 /// are set if the result needs to be inserted and/or extracted from vectors.
32 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
34 const TargetLoweringBase *getTLI() const { return TM->getTargetLowering(); }
37 BasicTTI() : ImmutablePass(ID), TM(0) {
38 llvm_unreachable("This pass cannot be directly constructed");
41 BasicTTI(const TargetMachine *TM) : ImmutablePass(ID), TM(TM) {
42 initializeBasicTTIPass(*PassRegistry::getPassRegistry());
45 void initializePass() override {
49 void getAnalysisUsage(AnalysisUsage &AU) const override {
50 TargetTransformInfo::getAnalysisUsage(AU);
53 /// Pass identification.
56 /// Provide necessary pointer adjustments for the two base classes.
57 void *getAdjustedAnalysisPointer(const void *ID) override {
58 if (ID == &TargetTransformInfo::ID)
59 return (TargetTransformInfo*)this;
63 bool hasBranchDivergence() const override;
65 /// \name Scalar TTI Implementations
68 bool isLegalAddImmediate(int64_t imm) const override;
69 bool isLegalICmpImmediate(int64_t imm) const override;
70 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
71 int64_t BaseOffset, bool HasBaseReg,
72 int64_t Scale) const override;
73 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
74 int64_t BaseOffset, bool HasBaseReg,
75 int64_t Scale) const override;
76 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
77 bool isTypeLegal(Type *Ty) const override;
78 unsigned getJumpBufAlignment() const override;
79 unsigned getJumpBufSize() const override;
80 bool shouldBuildLookupTables() const override;
81 bool haveFastSqrt(Type *Ty) const override;
82 void getUnrollingPreferences(Loop *L,
83 UnrollingPreferences &UP) const override;
87 /// \name Vector TTI Implementations
90 unsigned getNumberOfRegisters(bool Vector) const override;
91 unsigned getMaximumUnrollFactor() const override;
92 unsigned getRegisterBitWidth(bool Vector) const override;
93 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
94 OperandValueKind) const override;
95 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
96 int Index, Type *SubTp) const override;
97 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
98 Type *Src) const override;
99 unsigned getCFInstrCost(unsigned Opcode) const override;
100 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
101 Type *CondTy) const override;
102 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
103 unsigned Index) const override;
104 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
105 unsigned AddressSpace) const override;
106 unsigned getIntrinsicInstrCost(Intrinsic::ID, Type *RetTy,
107 ArrayRef<Type*> Tys) const override;
108 unsigned getNumberOfParts(Type *Tp) const override;
109 unsigned getAddressComputationCost( Type *Ty, bool IsComplex) const override;
110 unsigned getReductionCost(unsigned Opcode, Type *Ty,
111 bool IsPairwise) const override;
118 INITIALIZE_AG_PASS(BasicTTI, TargetTransformInfo, "basictti",
119 "Target independent code generator's TTI", true, true, false)
120 char BasicTTI::ID = 0;
123 llvm::createBasicTargetTransformInfoPass(const TargetMachine *TM) {
124 return new BasicTTI(TM);
127 bool BasicTTI::hasBranchDivergence() const { return false; }
129 bool BasicTTI::isLegalAddImmediate(int64_t imm) const {
130 return getTLI()->isLegalAddImmediate(imm);
133 bool BasicTTI::isLegalICmpImmediate(int64_t imm) const {
134 return getTLI()->isLegalICmpImmediate(imm);
137 bool BasicTTI::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
138 int64_t BaseOffset, bool HasBaseReg,
139 int64_t Scale) const {
140 TargetLoweringBase::AddrMode AM;
142 AM.BaseOffs = BaseOffset;
143 AM.HasBaseReg = HasBaseReg;
145 return getTLI()->isLegalAddressingMode(AM, Ty);
148 int BasicTTI::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
149 int64_t BaseOffset, bool HasBaseReg,
150 int64_t Scale) const {
151 TargetLoweringBase::AddrMode AM;
153 AM.BaseOffs = BaseOffset;
154 AM.HasBaseReg = HasBaseReg;
156 return getTLI()->getScalingFactorCost(AM, Ty);
159 bool BasicTTI::isTruncateFree(Type *Ty1, Type *Ty2) const {
160 return getTLI()->isTruncateFree(Ty1, Ty2);
163 bool BasicTTI::isTypeLegal(Type *Ty) const {
164 EVT T = getTLI()->getValueType(Ty);
165 return getTLI()->isTypeLegal(T);
168 unsigned BasicTTI::getJumpBufAlignment() const {
169 return getTLI()->getJumpBufAlignment();
172 unsigned BasicTTI::getJumpBufSize() const {
173 return getTLI()->getJumpBufSize();
176 bool BasicTTI::shouldBuildLookupTables() const {
177 const TargetLoweringBase *TLI = getTLI();
178 return TLI->supportJumpTables() &&
179 (TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
180 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
183 bool BasicTTI::haveFastSqrt(Type *Ty) const {
184 const TargetLoweringBase *TLI = getTLI();
185 EVT VT = TLI->getValueType(Ty);
186 return TLI->isTypeLegal(VT) && TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
189 void BasicTTI::getUnrollingPreferences(Loop *, UnrollingPreferences &) const { }
191 //===----------------------------------------------------------------------===//
193 // Calls used by the vectorizers.
195 //===----------------------------------------------------------------------===//
197 unsigned BasicTTI::getScalarizationOverhead(Type *Ty, bool Insert,
198 bool Extract) const {
199 assert (Ty->isVectorTy() && "Can only scalarize vectors");
202 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
204 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
206 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
212 unsigned BasicTTI::getNumberOfRegisters(bool Vector) const {
216 unsigned BasicTTI::getRegisterBitWidth(bool Vector) const {
220 unsigned BasicTTI::getMaximumUnrollFactor() const {
224 unsigned BasicTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
226 OperandValueKind) const {
227 // Check if any of the operands are vector operands.
228 const TargetLoweringBase *TLI = getTLI();
229 int ISD = TLI->InstructionOpcodeToISD(Opcode);
230 assert(ISD && "Invalid opcode");
232 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
234 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
235 // Assume that floating point arithmetic operations cost twice as much as
236 // integer operations.
237 unsigned OpCost = (IsFloat ? 2 : 1);
239 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
240 // The operation is legal. Assume it costs 1.
241 // If the type is split to multiple registers, assume that there is some
243 // TODO: Once we have extract/insert subvector cost we need to use them.
245 return LT.first * 2 * OpCost;
246 return LT.first * 1 * OpCost;
249 if (!TLI->isOperationExpand(ISD, LT.second)) {
250 // If the operation is custom lowered then assume
251 // thare the code is twice as expensive.
252 return LT.first * 2 * OpCost;
255 // Else, assume that we need to scalarize this op.
256 if (Ty->isVectorTy()) {
257 unsigned Num = Ty->getVectorNumElements();
258 unsigned Cost = TopTTI->getArithmeticInstrCost(Opcode, Ty->getScalarType());
259 // return the cost of multiple scalar invocation plus the cost of inserting
260 // and extracting the values.
261 return getScalarizationOverhead(Ty, true, true) + Num * Cost;
264 // We don't know anything about this scalar instruction.
268 unsigned BasicTTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
273 unsigned BasicTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
275 const TargetLoweringBase *TLI = getTLI();
276 int ISD = TLI->InstructionOpcodeToISD(Opcode);
277 assert(ISD && "Invalid opcode");
279 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(Src);
280 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(Dst);
282 // Check for NOOP conversions.
283 if (SrcLT.first == DstLT.first &&
284 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
286 // Bitcast between types that are legalized to the same type are free.
287 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
291 if (Opcode == Instruction::Trunc &&
292 TLI->isTruncateFree(SrcLT.second, DstLT.second))
295 if (Opcode == Instruction::ZExt &&
296 TLI->isZExtFree(SrcLT.second, DstLT.second))
299 // If the cast is marked as legal (or promote) then assume low cost.
300 if (SrcLT.first == DstLT.first &&
301 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
304 // Handle scalar conversions.
305 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
307 // Scalar bitcasts are usually free.
308 if (Opcode == Instruction::BitCast)
311 // Just check the op cost. If the operation is legal then assume it costs 1.
312 if (!TLI->isOperationExpand(ISD, DstLT.second))
315 // Assume that illegal scalar instruction are expensive.
319 // Check vector-to-vector casts.
320 if (Dst->isVectorTy() && Src->isVectorTy()) {
322 // If the cast is between same-sized registers, then the check is simple.
323 if (SrcLT.first == DstLT.first &&
324 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
326 // Assume that Zext is done using AND.
327 if (Opcode == Instruction::ZExt)
330 // Assume that sext is done using SHL and SRA.
331 if (Opcode == Instruction::SExt)
334 // Just check the op cost. If the operation is legal then assume it costs
335 // 1 and multiply by the type-legalization overhead.
336 if (!TLI->isOperationExpand(ISD, DstLT.second))
337 return SrcLT.first * 1;
340 // If we are converting vectors and the operation is illegal, or
341 // if the vectors are legalized to different types, estimate the
342 // scalarization costs.
343 unsigned Num = Dst->getVectorNumElements();
344 unsigned Cost = TopTTI->getCastInstrCost(Opcode, Dst->getScalarType(),
345 Src->getScalarType());
347 // Return the cost of multiple scalar invocation plus the cost of
348 // inserting and extracting the values.
349 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
352 // We already handled vector-to-vector and scalar-to-scalar conversions. This
353 // is where we handle bitcast between vectors and scalars. We need to assume
354 // that the conversion is scalarized in one way or another.
355 if (Opcode == Instruction::BitCast)
356 // Illegal bitcasts are done by storing and loading from a stack slot.
357 return (Src->isVectorTy()? getScalarizationOverhead(Src, false, true):0) +
358 (Dst->isVectorTy()? getScalarizationOverhead(Dst, true, false):0);
360 llvm_unreachable("Unhandled cast");
363 unsigned BasicTTI::getCFInstrCost(unsigned Opcode) const {
364 // Branches are assumed to be predicted.
368 unsigned BasicTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
369 Type *CondTy) const {
370 const TargetLoweringBase *TLI = getTLI();
371 int ISD = TLI->InstructionOpcodeToISD(Opcode);
372 assert(ISD && "Invalid opcode");
374 // Selects on vectors are actually vector selects.
375 if (ISD == ISD::SELECT) {
376 assert(CondTy && "CondTy must exist");
377 if (CondTy->isVectorTy())
381 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
383 if (!TLI->isOperationExpand(ISD, LT.second)) {
384 // The operation is legal. Assume it costs 1. Multiply
385 // by the type-legalization overhead.
389 // Otherwise, assume that the cast is scalarized.
390 if (ValTy->isVectorTy()) {
391 unsigned Num = ValTy->getVectorNumElements();
393 CondTy = CondTy->getScalarType();
394 unsigned Cost = TopTTI->getCmpSelInstrCost(Opcode, ValTy->getScalarType(),
397 // Return the cost of multiple scalar invocation plus the cost of inserting
398 // and extracting the values.
399 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
402 // Unknown scalar opcode.
406 unsigned BasicTTI::getVectorInstrCost(unsigned Opcode, Type *Val,
407 unsigned Index) const {
408 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Val->getScalarType());
413 unsigned BasicTTI::getMemoryOpCost(unsigned Opcode, Type *Src,
415 unsigned AddressSpace) const {
416 assert(!Src->isVoidTy() && "Invalid type");
417 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Src);
419 // Assuming that all loads of legal types cost 1.
420 unsigned Cost = LT.first;
422 if (Src->isVectorTy() &&
423 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
424 // This is a vector load that legalizes to a larger type than the vector
425 // itself. Unless the corresponding extending load or truncating store is
426 // legal, then this will scalarize.
427 TargetLowering::LegalizeAction LA;
428 MVT MemVT = getTLI()->getSimpleValueType(Src, true);
429 if (Opcode == Instruction::Store)
430 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
432 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, MemVT);
434 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
435 // This is a vector load/store for some illegal type that is scalarized.
436 // We must account for the cost of building or decomposing the vector.
437 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
438 Opcode == Instruction::Store);
445 unsigned BasicTTI::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
446 ArrayRef<Type *> Tys) const {
450 // Assume that we need to scalarize this intrinsic.
451 unsigned ScalarizationCost = 0;
452 unsigned ScalarCalls = 1;
453 if (RetTy->isVectorTy()) {
454 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
455 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
457 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
458 if (Tys[i]->isVectorTy()) {
459 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
460 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
464 return ScalarCalls + ScalarizationCost;
466 // Look for intrinsics that can be lowered directly or turned into a scalar
468 case Intrinsic::sqrt: ISD = ISD::FSQRT; break;
469 case Intrinsic::sin: ISD = ISD::FSIN; break;
470 case Intrinsic::cos: ISD = ISD::FCOS; break;
471 case Intrinsic::exp: ISD = ISD::FEXP; break;
472 case Intrinsic::exp2: ISD = ISD::FEXP2; break;
473 case Intrinsic::log: ISD = ISD::FLOG; break;
474 case Intrinsic::log10: ISD = ISD::FLOG10; break;
475 case Intrinsic::log2: ISD = ISD::FLOG2; break;
476 case Intrinsic::fabs: ISD = ISD::FABS; break;
477 case Intrinsic::copysign: ISD = ISD::FCOPYSIGN; break;
478 case Intrinsic::floor: ISD = ISD::FFLOOR; break;
479 case Intrinsic::ceil: ISD = ISD::FCEIL; break;
480 case Intrinsic::trunc: ISD = ISD::FTRUNC; break;
481 case Intrinsic::nearbyint:
482 ISD = ISD::FNEARBYINT; break;
483 case Intrinsic::rint: ISD = ISD::FRINT; break;
484 case Intrinsic::round: ISD = ISD::FROUND; break;
485 case Intrinsic::pow: ISD = ISD::FPOW; break;
486 case Intrinsic::fma: ISD = ISD::FMA; break;
487 case Intrinsic::fmuladd: ISD = ISD::FMA; break; // FIXME: mul + add?
488 case Intrinsic::lifetime_start:
489 case Intrinsic::lifetime_end:
493 const TargetLoweringBase *TLI = getTLI();
494 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(RetTy);
496 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
497 // The operation is legal. Assume it costs 1.
498 // If the type is split to multiple registers, assume that thre is some
500 // TODO: Once we have extract/insert subvector cost we need to use them.
506 if (!TLI->isOperationExpand(ISD, LT.second)) {
507 // If the operation is custom lowered then assume
508 // thare the code is twice as expensive.
512 // Else, assume that we need to scalarize this intrinsic. For math builtins
513 // this will emit a costly libcall, adding call overhead and spills. Make it
515 if (RetTy->isVectorTy()) {
516 unsigned Num = RetTy->getVectorNumElements();
517 unsigned Cost = TopTTI->getIntrinsicInstrCost(IID, RetTy->getScalarType(),
519 return 10 * Cost * Num;
522 // This is going to be turned into a library call, make it expensive.
526 unsigned BasicTTI::getNumberOfParts(Type *Tp) const {
527 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Tp);
531 unsigned BasicTTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
535 unsigned BasicTTI::getReductionCost(unsigned Opcode, Type *Ty,
536 bool IsPairwise) const {
537 assert(Ty->isVectorTy() && "Expect a vector type");
538 unsigned NumVecElts = Ty->getVectorNumElements();
539 unsigned NumReduxLevels = Log2_32(NumVecElts);
540 unsigned ArithCost = NumReduxLevels *
541 TopTTI->getArithmeticInstrCost(Opcode, Ty);
542 // Assume the pairwise shuffles add a cost.
543 unsigned ShuffleCost =
544 NumReduxLevels * (IsPairwise + 1) *
545 TopTTI->getShuffleCost(SK_ExtractSubvector, Ty, NumVecElts / 2, Ty);
546 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);