1 //===- BasicTargetTransformInfo.cpp - Basic target-independent TTI impl ---===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file 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. 15 /// 16 //===----------------------------------------------------------------------===// 17 18 #define DEBUG_TYPE "basictti" 19 #include "llvm/CodeGen/Passes.h" 20 #include "llvm/Analysis/TargetTransformInfo.h" 21 #include "llvm/Target/TargetLowering.h" 22 #include <utility> 23 24 using namespace llvm; 25 26 namespace { 27 28 class BasicTTI : public ImmutablePass, public TargetTransformInfo { 29 const TargetLoweringBase *TLI; 30 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; 34 35 public: 36 BasicTTI() : ImmutablePass(ID), TLI(0) { 37 llvm_unreachable("This pass cannot be directly constructed"); 38 } 39 40 BasicTTI(const TargetLoweringBase *TLI) : ImmutablePass(ID), TLI(TLI) { 41 initializeBasicTTIPass(*PassRegistry::getPassRegistry()); 42 } 43 44 virtual void initializePass() { 45 pushTTIStack(this); 46 } 47 48 virtual void finalizePass() { 49 popTTIStack(); 50 } 51 52 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 53 TargetTransformInfo::getAnalysisUsage(AU); 54 } 55 56 /// Pass identification. 57 static char ID; 58 59 /// Provide necessary pointer adjustments for the two base classes. 60 virtual void *getAdjustedAnalysisPointer(const void *ID) { 61 if (ID == &TargetTransformInfo::ID) 62 return (TargetTransformInfo*)this; 63 return this; 64 } 65 66 /// \name Scalar TTI Implementations 67 /// @{ 68 69 virtual bool isLegalAddImmediate(int64_t imm) const; 70 virtual bool isLegalICmpImmediate(int64_t imm) const; 71 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 72 int64_t BaseOffset, bool HasBaseReg, 73 int64_t Scale) const; 74 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const; 75 virtual bool isTypeLegal(Type *Ty) const; 76 virtual unsigned getJumpBufAlignment() const; 77 virtual unsigned getJumpBufSize() const; 78 virtual bool shouldBuildLookupTables() const; 79 80 /// @} 81 82 /// \name Vector TTI Implementations 83 /// @{ 84 85 virtual unsigned getNumberOfRegisters(bool Vector) const; 86 virtual unsigned getMaximumUnrollFactor() const; 87 virtual unsigned getRegisterBitWidth(bool Vector) const; 88 virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const; 89 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, 90 int Index, Type *SubTp) const; 91 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, 92 Type *Src) const; 93 virtual unsigned getCFInstrCost(unsigned Opcode) const; 94 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 95 Type *CondTy) const; 96 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val, 97 unsigned Index) const; 98 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src, 99 unsigned Alignment, 100 unsigned AddressSpace) const; 101 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID, Type *RetTy, 102 ArrayRef<Type*> Tys) const; 103 virtual unsigned getNumberOfParts(Type *Tp) const; 104 virtual unsigned getAddressComputationCost(Type *Ty) const; 105 106 /// @} 107 }; 108 109 } 110 111 INITIALIZE_AG_PASS(BasicTTI, TargetTransformInfo, "basictti", 112 "Target independent code generator's TTI", true, true, false) 113 char BasicTTI::ID = 0; 114 115 ImmutablePass * 116 llvm::createBasicTargetTransformInfoPass(const TargetLoweringBase *TLI) { 117 return new BasicTTI(TLI); 118 } 119 120 121 bool BasicTTI::isLegalAddImmediate(int64_t imm) const { 122 return TLI->isLegalAddImmediate(imm); 123 } 124 125 bool BasicTTI::isLegalICmpImmediate(int64_t imm) const { 126 return TLI->isLegalICmpImmediate(imm); 127 } 128 129 bool BasicTTI::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 130 int64_t BaseOffset, bool HasBaseReg, 131 int64_t Scale) const { 132 TargetLoweringBase::AddrMode AM; 133 AM.BaseGV = BaseGV; 134 AM.BaseOffs = BaseOffset; 135 AM.HasBaseReg = HasBaseReg; 136 AM.Scale = Scale; 137 return TLI->isLegalAddressingMode(AM, Ty); 138 } 139 140 bool BasicTTI::isTruncateFree(Type *Ty1, Type *Ty2) const { 141 return TLI->isTruncateFree(Ty1, Ty2); 142 } 143 144 bool BasicTTI::isTypeLegal(Type *Ty) const { 145 EVT T = TLI->getValueType(Ty); 146 return TLI->isTypeLegal(T); 147 } 148 149 unsigned BasicTTI::getJumpBufAlignment() const { 150 return TLI->getJumpBufAlignment(); 151 } 152 153 unsigned BasicTTI::getJumpBufSize() const { 154 return TLI->getJumpBufSize(); 155 } 156 157 bool BasicTTI::shouldBuildLookupTables() const { 158 return TLI->supportJumpTables() && 159 (TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) || 160 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other)); 161 } 162 163 //===----------------------------------------------------------------------===// 164 // 165 // Calls used by the vectorizers. 166 // 167 //===----------------------------------------------------------------------===// 168 169 unsigned BasicTTI::getScalarizationOverhead(Type *Ty, bool Insert, 170 bool Extract) const { 171 assert (Ty->isVectorTy() && "Can only scalarize vectors"); 172 unsigned Cost = 0; 173 174 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) { 175 if (Insert) 176 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i); 177 if (Extract) 178 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i); 179 } 180 181 return Cost; 182 } 183 184 unsigned BasicTTI::getNumberOfRegisters(bool Vector) const { 185 return 1; 186 } 187 188 unsigned BasicTTI::getRegisterBitWidth(bool Vector) const { 189 return 32; 190 } 191 192 unsigned BasicTTI::getMaximumUnrollFactor() const { 193 return 1; 194 } 195 196 unsigned BasicTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty) const { 197 // Check if any of the operands are vector operands. 198 int ISD = TLI->InstructionOpcodeToISD(Opcode); 199 assert(ISD && "Invalid opcode"); 200 201 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty); 202 203 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) { 204 // The operation is legal. Assume it costs 1. 205 // If the type is split to multiple registers, assume that thre is some 206 // overhead to this. 207 // TODO: Once we have extract/insert subvector cost we need to use them. 208 if (LT.first > 1) 209 return LT.first * 2; 210 return LT.first * 1; 211 } 212 213 if (!TLI->isOperationExpand(ISD, LT.second)) { 214 // If the operation is custom lowered then assume 215 // thare the code is twice as expensive. 216 return LT.first * 2; 217 } 218 219 // Else, assume that we need to scalarize this op. 220 if (Ty->isVectorTy()) { 221 unsigned Num = Ty->getVectorNumElements(); 222 unsigned Cost = TopTTI->getArithmeticInstrCost(Opcode, Ty->getScalarType()); 223 // return the cost of multiple scalar invocation plus the cost of inserting 224 // and extracting the values. 225 return getScalarizationOverhead(Ty, true, true) + Num * Cost; 226 } 227 228 // We don't know anything about this scalar instruction. 229 return 1; 230 } 231 232 unsigned BasicTTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index, 233 Type *SubTp) const { 234 return 1; 235 } 236 237 unsigned BasicTTI::getCastInstrCost(unsigned Opcode, Type *Dst, 238 Type *Src) const { 239 int ISD = TLI->InstructionOpcodeToISD(Opcode); 240 assert(ISD && "Invalid opcode"); 241 242 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(Src); 243 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(Dst); 244 245 // Check for NOOP conversions. 246 if (SrcLT.first == DstLT.first && 247 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) { 248 249 // Bitcast between types that are legalized to the same type are free. 250 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc) 251 return 0; 252 } 253 254 if (Opcode == Instruction::Trunc && 255 TLI->isTruncateFree(SrcLT.second, DstLT.second)) 256 return 0; 257 258 if (Opcode == Instruction::ZExt && 259 TLI->isZExtFree(SrcLT.second, DstLT.second)) 260 return 0; 261 262 // If the cast is marked as legal (or promote) then assume low cost. 263 if (TLI->isOperationLegalOrPromote(ISD, DstLT.second)) 264 return 1; 265 266 // Handle scalar conversions. 267 if (!Src->isVectorTy() && !Dst->isVectorTy()) { 268 269 // Scalar bitcasts are usually free. 270 if (Opcode == Instruction::BitCast) 271 return 0; 272 273 // Just check the op cost. If the operation is legal then assume it costs 1. 274 if (!TLI->isOperationExpand(ISD, DstLT.second)) 275 return 1; 276 277 // Assume that illegal scalar instruction are expensive. 278 return 4; 279 } 280 281 // Check vector-to-vector casts. 282 if (Dst->isVectorTy() && Src->isVectorTy()) { 283 284 // If the cast is between same-sized registers, then the check is simple. 285 if (SrcLT.first == DstLT.first && 286 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) { 287 288 // Assume that Zext is done using AND. 289 if (Opcode == Instruction::ZExt) 290 return 1; 291 292 // Assume that sext is done using SHL and SRA. 293 if (Opcode == Instruction::SExt) 294 return 2; 295 296 // Just check the op cost. If the operation is legal then assume it costs 297 // 1 and multiply by the type-legalization overhead. 298 if (!TLI->isOperationExpand(ISD, DstLT.second)) 299 return SrcLT.first * 1; 300 } 301 302 // If we are converting vectors and the operation is illegal, or 303 // if the vectors are legalized to different types, estimate the 304 // scalarization costs. 305 unsigned Num = Dst->getVectorNumElements(); 306 unsigned Cost = TopTTI->getCastInstrCost(Opcode, Dst->getScalarType(), 307 Src->getScalarType()); 308 309 // Return the cost of multiple scalar invocation plus the cost of 310 // inserting and extracting the values. 311 return getScalarizationOverhead(Dst, true, true) + Num * Cost; 312 } 313 314 // We already handled vector-to-vector and scalar-to-scalar conversions. This 315 // is where we handle bitcast between vectors and scalars. We need to assume 316 // that the conversion is scalarized in one way or another. 317 if (Opcode == Instruction::BitCast) 318 // Illegal bitcasts are done by storing and loading from a stack slot. 319 return (Src->isVectorTy()? getScalarizationOverhead(Src, false, true):0) + 320 (Dst->isVectorTy()? getScalarizationOverhead(Dst, true, false):0); 321 322 llvm_unreachable("Unhandled cast"); 323 } 324 325 unsigned BasicTTI::getCFInstrCost(unsigned Opcode) const { 326 // Branches are assumed to be predicted. 327 return 0; 328 } 329 330 unsigned BasicTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 331 Type *CondTy) const { 332 int ISD = TLI->InstructionOpcodeToISD(Opcode); 333 assert(ISD && "Invalid opcode"); 334 335 // Selects on vectors are actually vector selects. 336 if (ISD == ISD::SELECT) { 337 assert(CondTy && "CondTy must exist"); 338 if (CondTy->isVectorTy()) 339 ISD = ISD::VSELECT; 340 } 341 342 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy); 343 344 if (!TLI->isOperationExpand(ISD, LT.second)) { 345 // The operation is legal. Assume it costs 1. Multiply 346 // by the type-legalization overhead. 347 return LT.first * 1; 348 } 349 350 // Otherwise, assume that the cast is scalarized. 351 if (ValTy->isVectorTy()) { 352 unsigned Num = ValTy->getVectorNumElements(); 353 if (CondTy) 354 CondTy = CondTy->getScalarType(); 355 unsigned Cost = TopTTI->getCmpSelInstrCost(Opcode, ValTy->getScalarType(), 356 CondTy); 357 358 // Return the cost of multiple scalar invocation plus the cost of inserting 359 // and extracting the values. 360 return getScalarizationOverhead(ValTy, true, false) + Num * Cost; 361 } 362 363 // Unknown scalar opcode. 364 return 1; 365 } 366 367 unsigned BasicTTI::getVectorInstrCost(unsigned Opcode, Type *Val, 368 unsigned Index) const { 369 return 1; 370 } 371 372 unsigned BasicTTI::getMemoryOpCost(unsigned Opcode, Type *Src, 373 unsigned Alignment, 374 unsigned AddressSpace) const { 375 assert(!Src->isVoidTy() && "Invalid type"); 376 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src); 377 378 // Assume that all loads of legal types cost 1. 379 return LT.first; 380 } 381 382 unsigned BasicTTI::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 383 ArrayRef<Type *> Tys) const { 384 unsigned ISD = 0; 385 switch (IID) { 386 default: { 387 // Assume that we need to scalarize this intrinsic. 388 unsigned ScalarizationCost = 0; 389 unsigned ScalarCalls = 1; 390 if (RetTy->isVectorTy()) { 391 ScalarizationCost = getScalarizationOverhead(RetTy, true, false); 392 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements()); 393 } 394 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) { 395 if (Tys[i]->isVectorTy()) { 396 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true); 397 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements()); 398 } 399 } 400 401 return ScalarCalls + ScalarizationCost; 402 } 403 // Look for intrinsics that can be lowered directly or turned into a scalar 404 // intrinsic call. 405 case Intrinsic::sqrt: ISD = ISD::FSQRT; break; 406 case Intrinsic::sin: ISD = ISD::FSIN; break; 407 case Intrinsic::cos: ISD = ISD::FCOS; break; 408 case Intrinsic::exp: ISD = ISD::FEXP; break; 409 case Intrinsic::exp2: ISD = ISD::FEXP2; break; 410 case Intrinsic::log: ISD = ISD::FLOG; break; 411 case Intrinsic::log10: ISD = ISD::FLOG10; break; 412 case Intrinsic::log2: ISD = ISD::FLOG2; break; 413 case Intrinsic::fabs: ISD = ISD::FABS; break; 414 case Intrinsic::floor: ISD = ISD::FFLOOR; break; 415 case Intrinsic::ceil: ISD = ISD::FCEIL; break; 416 case Intrinsic::trunc: ISD = ISD::FTRUNC; break; 417 case Intrinsic::rint: ISD = ISD::FRINT; break; 418 case Intrinsic::pow: ISD = ISD::FPOW; break; 419 case Intrinsic::fma: ISD = ISD::FMA; break; 420 case Intrinsic::fmuladd: ISD = ISD::FMA; break; // FIXME: mul + add? 421 } 422 423 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(RetTy); 424 425 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) { 426 // The operation is legal. Assume it costs 1. 427 // If the type is split to multiple registers, assume that thre is some 428 // overhead to this. 429 // TODO: Once we have extract/insert subvector cost we need to use them. 430 if (LT.first > 1) 431 return LT.first * 2; 432 return LT.first * 1; 433 } 434 435 if (!TLI->isOperationExpand(ISD, LT.second)) { 436 // If the operation is custom lowered then assume 437 // thare the code is twice as expensive. 438 return LT.first * 2; 439 } 440 441 // Else, assume that we need to scalarize this intrinsic. For math builtins 442 // this will emit a costly libcall, adding call overhead and spills. Make it 443 // very expensive. 444 if (RetTy->isVectorTy()) { 445 unsigned Num = RetTy->getVectorNumElements(); 446 unsigned Cost = TopTTI->getIntrinsicInstrCost(IID, RetTy->getScalarType(), 447 Tys); 448 return 10 * Cost * Num; 449 } 450 451 // This is going to be turned into a library call, make it expensive. 452 return 10; 453 } 454 455 unsigned BasicTTI::getNumberOfParts(Type *Tp) const { 456 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp); 457 return LT.first; 458 } 459 460 unsigned BasicTTI::getAddressComputationCost(Type *Ty) const { 461 return 0; 462 } 463