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"
28 class BasicTTI : 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(0) {
39 llvm_unreachable("This pass cannot be directly constructed");
42 BasicTTI(const TargetMachine *TM) : ImmutablePass(ID), TM(TM) {
43 initializeBasicTTIPass(*PassRegistry::getPassRegistry());
46 virtual void initializePass() {
50 virtual void finalizePass() {
54 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
55 TargetTransformInfo::getAnalysisUsage(AU);
58 /// Pass identification.
61 /// Provide necessary pointer adjustments for the two base classes.
62 virtual void *getAdjustedAnalysisPointer(const void *ID) {
63 if (ID == &TargetTransformInfo::ID)
64 return (TargetTransformInfo*)this;
68 virtual bool hasBranchDivergence() const;
70 /// \name Scalar TTI Implementations
73 virtual bool isLegalAddImmediate(int64_t imm) const;
74 virtual bool isLegalICmpImmediate(int64_t imm) const;
75 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
76 int64_t BaseOffset, bool HasBaseReg,
78 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
79 int64_t BaseOffset, bool HasBaseReg,
81 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
82 virtual bool isTypeLegal(Type *Ty) const;
83 virtual unsigned getJumpBufAlignment() const;
84 virtual unsigned getJumpBufSize() const;
85 virtual bool shouldBuildLookupTables() const;
86 virtual bool haveFastSqrt(Type *Ty) const;
87 virtual bool getUnrollingPreferences(UnrollingPreferences &UP) const;
91 /// \name Vector TTI Implementations
94 virtual unsigned getNumberOfRegisters(bool Vector) const;
95 virtual unsigned getMaximumUnrollFactor() const;
96 virtual unsigned getRegisterBitWidth(bool Vector) const;
97 virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
99 OperandValueKind) const;
100 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
101 int Index, Type *SubTp) const;
102 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
104 virtual unsigned getCFInstrCost(unsigned Opcode) const;
105 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
107 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
108 unsigned Index) const;
109 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
111 unsigned AddressSpace) const;
112 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID, Type *RetTy,
113 ArrayRef<Type*> Tys) const;
114 virtual unsigned getNumberOfParts(Type *Tp) const;
115 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) const;
122 INITIALIZE_AG_PASS(BasicTTI, TargetTransformInfo, "basictti",
123 "Target independent code generator's TTI", true, true, false)
124 char BasicTTI::ID = 0;
127 llvm::createBasicTargetTransformInfoPass(const TargetMachine *TM) {
128 return new BasicTTI(TM);
131 bool BasicTTI::hasBranchDivergence() const { return false; }
133 bool BasicTTI::isLegalAddImmediate(int64_t imm) const {
134 return getTLI()->isLegalAddImmediate(imm);
137 bool BasicTTI::isLegalICmpImmediate(int64_t imm) const {
138 return getTLI()->isLegalICmpImmediate(imm);
141 bool BasicTTI::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
142 int64_t BaseOffset, bool HasBaseReg,
143 int64_t Scale) const {
144 TargetLoweringBase::AddrMode AM;
146 AM.BaseOffs = BaseOffset;
147 AM.HasBaseReg = HasBaseReg;
149 return getTLI()->isLegalAddressingMode(AM, Ty);
152 int BasicTTI::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
153 int64_t BaseOffset, bool HasBaseReg,
154 int64_t Scale) const {
155 TargetLoweringBase::AddrMode AM;
157 AM.BaseOffs = BaseOffset;
158 AM.HasBaseReg = HasBaseReg;
160 return getTLI()->getScalingFactorCost(AM, Ty);
163 bool BasicTTI::isTruncateFree(Type *Ty1, Type *Ty2) const {
164 return getTLI()->isTruncateFree(Ty1, Ty2);
167 bool BasicTTI::isTypeLegal(Type *Ty) const {
168 EVT T = getTLI()->getValueType(Ty);
169 return getTLI()->isTypeLegal(T);
172 unsigned BasicTTI::getJumpBufAlignment() const {
173 return getTLI()->getJumpBufAlignment();
176 unsigned BasicTTI::getJumpBufSize() const {
177 return getTLI()->getJumpBufSize();
180 bool BasicTTI::shouldBuildLookupTables() const {
181 const TargetLoweringBase *TLI = getTLI();
182 return TLI->supportJumpTables() &&
183 (TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
184 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
187 bool BasicTTI::haveFastSqrt(Type *Ty) const {
188 const TargetLoweringBase *TLI = getTLI();
189 EVT VT = TLI->getValueType(Ty);
190 return TLI->isTypeLegal(VT) && TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
193 bool BasicTTI::getUnrollingPreferences(UnrollingPreferences &) const {
197 //===----------------------------------------------------------------------===//
199 // Calls used by the vectorizers.
201 //===----------------------------------------------------------------------===//
203 unsigned BasicTTI::getScalarizationOverhead(Type *Ty, bool Insert,
204 bool Extract) const {
205 assert (Ty->isVectorTy() && "Can only scalarize vectors");
208 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
210 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
212 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
218 unsigned BasicTTI::getNumberOfRegisters(bool Vector) const {
222 unsigned BasicTTI::getRegisterBitWidth(bool Vector) const {
226 unsigned BasicTTI::getMaximumUnrollFactor() const {
230 unsigned BasicTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
232 OperandValueKind) const {
233 // Check if any of the operands are vector operands.
234 const TargetLoweringBase *TLI = getTLI();
235 int ISD = TLI->InstructionOpcodeToISD(Opcode);
236 assert(ISD && "Invalid opcode");
238 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
240 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
241 // Assume that floating point arithmetic operations cost twice as much as
242 // integer operations.
243 unsigned OpCost = (IsFloat ? 2 : 1);
245 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
246 // The operation is legal. Assume it costs 1.
247 // If the type is split to multiple registers, assume that there is some
249 // TODO: Once we have extract/insert subvector cost we need to use them.
251 return LT.first * 2 * OpCost;
252 return LT.first * 1 * OpCost;
255 if (!TLI->isOperationExpand(ISD, LT.second)) {
256 // If the operation is custom lowered then assume
257 // thare the code is twice as expensive.
258 return LT.first * 2 * OpCost;
261 // Else, assume that we need to scalarize this op.
262 if (Ty->isVectorTy()) {
263 unsigned Num = Ty->getVectorNumElements();
264 unsigned Cost = TopTTI->getArithmeticInstrCost(Opcode, Ty->getScalarType());
265 // return the cost of multiple scalar invocation plus the cost of inserting
266 // and extracting the values.
267 return getScalarizationOverhead(Ty, true, true) + Num * Cost;
270 // We don't know anything about this scalar instruction.
274 unsigned BasicTTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
279 unsigned BasicTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
281 const TargetLoweringBase *TLI = getTLI();
282 int ISD = TLI->InstructionOpcodeToISD(Opcode);
283 assert(ISD && "Invalid opcode");
285 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(Src);
286 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(Dst);
288 // Check for NOOP conversions.
289 if (SrcLT.first == DstLT.first &&
290 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
292 // Bitcast between types that are legalized to the same type are free.
293 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
297 if (Opcode == Instruction::Trunc &&
298 TLI->isTruncateFree(SrcLT.second, DstLT.second))
301 if (Opcode == Instruction::ZExt &&
302 TLI->isZExtFree(SrcLT.second, DstLT.second))
305 // If the cast is marked as legal (or promote) then assume low cost.
306 if (TLI->isOperationLegalOrPromote(ISD, DstLT.second))
309 // Handle scalar conversions.
310 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
312 // Scalar bitcasts are usually free.
313 if (Opcode == Instruction::BitCast)
316 // Just check the op cost. If the operation is legal then assume it costs 1.
317 if (!TLI->isOperationExpand(ISD, DstLT.second))
320 // Assume that illegal scalar instruction are expensive.
324 // Check vector-to-vector casts.
325 if (Dst->isVectorTy() && Src->isVectorTy()) {
327 // If the cast is between same-sized registers, then the check is simple.
328 if (SrcLT.first == DstLT.first &&
329 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
331 // Assume that Zext is done using AND.
332 if (Opcode == Instruction::ZExt)
335 // Assume that sext is done using SHL and SRA.
336 if (Opcode == Instruction::SExt)
339 // Just check the op cost. If the operation is legal then assume it costs
340 // 1 and multiply by the type-legalization overhead.
341 if (!TLI->isOperationExpand(ISD, DstLT.second))
342 return SrcLT.first * 1;
345 // If we are converting vectors and the operation is illegal, or
346 // if the vectors are legalized to different types, estimate the
347 // scalarization costs.
348 unsigned Num = Dst->getVectorNumElements();
349 unsigned Cost = TopTTI->getCastInstrCost(Opcode, Dst->getScalarType(),
350 Src->getScalarType());
352 // Return the cost of multiple scalar invocation plus the cost of
353 // inserting and extracting the values.
354 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
357 // We already handled vector-to-vector and scalar-to-scalar conversions. This
358 // is where we handle bitcast between vectors and scalars. We need to assume
359 // that the conversion is scalarized in one way or another.
360 if (Opcode == Instruction::BitCast)
361 // Illegal bitcasts are done by storing and loading from a stack slot.
362 return (Src->isVectorTy()? getScalarizationOverhead(Src, false, true):0) +
363 (Dst->isVectorTy()? getScalarizationOverhead(Dst, true, false):0);
365 llvm_unreachable("Unhandled cast");
368 unsigned BasicTTI::getCFInstrCost(unsigned Opcode) const {
369 // Branches are assumed to be predicted.
373 unsigned BasicTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
374 Type *CondTy) const {
375 const TargetLoweringBase *TLI = getTLI();
376 int ISD = TLI->InstructionOpcodeToISD(Opcode);
377 assert(ISD && "Invalid opcode");
379 // Selects on vectors are actually vector selects.
380 if (ISD == ISD::SELECT) {
381 assert(CondTy && "CondTy must exist");
382 if (CondTy->isVectorTy())
386 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
388 if (!TLI->isOperationExpand(ISD, LT.second)) {
389 // The operation is legal. Assume it costs 1. Multiply
390 // by the type-legalization overhead.
394 // Otherwise, assume that the cast is scalarized.
395 if (ValTy->isVectorTy()) {
396 unsigned Num = ValTy->getVectorNumElements();
398 CondTy = CondTy->getScalarType();
399 unsigned Cost = TopTTI->getCmpSelInstrCost(Opcode, ValTy->getScalarType(),
402 // Return the cost of multiple scalar invocation plus the cost of inserting
403 // and extracting the values.
404 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
407 // Unknown scalar opcode.
411 unsigned BasicTTI::getVectorInstrCost(unsigned Opcode, Type *Val,
412 unsigned Index) const {
416 unsigned BasicTTI::getMemoryOpCost(unsigned Opcode, Type *Src,
418 unsigned AddressSpace) const {
419 assert(!Src->isVoidTy() && "Invalid type");
420 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Src);
422 // Assume that all loads of legal types cost 1.
426 unsigned BasicTTI::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
427 ArrayRef<Type *> Tys) const {
431 // Assume that we need to scalarize this intrinsic.
432 unsigned ScalarizationCost = 0;
433 unsigned ScalarCalls = 1;
434 if (RetTy->isVectorTy()) {
435 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
436 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
438 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
439 if (Tys[i]->isVectorTy()) {
440 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
441 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
445 return ScalarCalls + ScalarizationCost;
447 // Look for intrinsics that can be lowered directly or turned into a scalar
449 case Intrinsic::sqrt: ISD = ISD::FSQRT; break;
450 case Intrinsic::sin: ISD = ISD::FSIN; break;
451 case Intrinsic::cos: ISD = ISD::FCOS; break;
452 case Intrinsic::exp: ISD = ISD::FEXP; break;
453 case Intrinsic::exp2: ISD = ISD::FEXP2; break;
454 case Intrinsic::log: ISD = ISD::FLOG; break;
455 case Intrinsic::log10: ISD = ISD::FLOG10; break;
456 case Intrinsic::log2: ISD = ISD::FLOG2; break;
457 case Intrinsic::fabs: ISD = ISD::FABS; break;
458 case Intrinsic::copysign: ISD = ISD::FCOPYSIGN; break;
459 case Intrinsic::floor: ISD = ISD::FFLOOR; break;
460 case Intrinsic::ceil: ISD = ISD::FCEIL; break;
461 case Intrinsic::trunc: ISD = ISD::FTRUNC; break;
462 case Intrinsic::nearbyint:
463 ISD = ISD::FNEARBYINT; break;
464 case Intrinsic::rint: ISD = ISD::FRINT; break;
465 case Intrinsic::round: ISD = ISD::FROUND; break;
466 case Intrinsic::pow: ISD = ISD::FPOW; break;
467 case Intrinsic::fma: ISD = ISD::FMA; break;
468 case Intrinsic::fmuladd: ISD = ISD::FMA; break; // FIXME: mul + add?
469 case Intrinsic::lifetime_start:
470 case Intrinsic::lifetime_end:
474 const TargetLoweringBase *TLI = getTLI();
475 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(RetTy);
477 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
478 // The operation is legal. Assume it costs 1.
479 // If the type is split to multiple registers, assume that thre is some
481 // TODO: Once we have extract/insert subvector cost we need to use them.
487 if (!TLI->isOperationExpand(ISD, LT.second)) {
488 // If the operation is custom lowered then assume
489 // thare the code is twice as expensive.
493 // Else, assume that we need to scalarize this intrinsic. For math builtins
494 // this will emit a costly libcall, adding call overhead and spills. Make it
496 if (RetTy->isVectorTy()) {
497 unsigned Num = RetTy->getVectorNumElements();
498 unsigned Cost = TopTTI->getIntrinsicInstrCost(IID, RetTy->getScalarType(),
500 return 10 * Cost * Num;
503 // This is going to be turned into a library call, make it expensive.
507 unsigned BasicTTI::getNumberOfParts(Type *Tp) const {
508 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Tp);
512 unsigned BasicTTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {