InstCombineCasts.cpp revision 5f0290e0ef6225114a04517744bf20e93040d2e4
1//===- InstCombineCasts.cpp -----------------------------------------------===// 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// 10// This file implements the visit functions for cast operations. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Target/TargetData.h" 16#include "llvm/Support/PatternMatch.h" 17using namespace llvm; 18using namespace PatternMatch; 19 20/// CanEvaluateInDifferentType - Return true if we can take the specified value 21/// and return it as type Ty without inserting any new casts and without 22/// changing the computed value. This is used by code that tries to decide 23/// whether promoting or shrinking integer operations to wider or smaller types 24/// will allow us to eliminate a truncate or extend. 25/// 26/// This is a truncation operation if Ty is smaller than V->getType(), or an 27/// extension operation if Ty is larger. 28/// 29/// If CastOpc is a truncation, then Ty will be a type smaller than V. We 30/// should return true if trunc(V) can be computed by computing V in the smaller 31/// type. If V is an instruction, then trunc(inst(x,y)) can be computed as 32/// inst(trunc(x),trunc(y)), which only makes sense if x and y can be 33/// efficiently truncated. 34/// 35/// If CastOpc is a sext or zext, we are asking if the low bits of the value can 36/// bit computed in a larger type, which is then and'd or sext_in_reg'd to get 37/// the final result. 38bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty, 39 unsigned CastOpc, 40 int &NumCastsRemoved){ 41 // We can always evaluate constants in another type. 42 if (isa<Constant>(V)) 43 return true; 44 45 Instruction *I = dyn_cast<Instruction>(V); 46 if (!I) return false; 47 48 const Type *OrigTy = V->getType(); 49 50 // If this is an extension or truncate, we can often eliminate it. 51 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) { 52 // If this is a cast from the destination type, we can trivially eliminate 53 // it, and this will remove a cast overall. 54 if (I->getOperand(0)->getType() == Ty) { 55 // If the first operand is itself a cast, and is eliminable, do not count 56 // this as an eliminable cast. We would prefer to eliminate those two 57 // casts first. 58 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse()) 59 ++NumCastsRemoved; 60 return true; 61 } 62 } 63 64 // We can't extend or shrink something that has multiple uses: doing so would 65 // require duplicating the instruction in general, which isn't profitable. 66 if (!I->hasOneUse()) return false; 67 68 unsigned Opc = I->getOpcode(); 69 switch (Opc) { 70 case Instruction::Add: 71 case Instruction::Sub: 72 case Instruction::Mul: 73 case Instruction::And: 74 case Instruction::Or: 75 case Instruction::Xor: 76 // These operators can all arbitrarily be extended or truncated. 77 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 78 NumCastsRemoved) && 79 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc, 80 NumCastsRemoved); 81 82 case Instruction::UDiv: 83 case Instruction::URem: { 84 // UDiv and URem can be truncated if all the truncated bits are zero. 85 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 86 uint32_t BitWidth = Ty->getScalarSizeInBits(); 87 if (BitWidth < OrigBitWidth) { 88 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 89 if (MaskedValueIsZero(I->getOperand(0), Mask) && 90 MaskedValueIsZero(I->getOperand(1), Mask)) { 91 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 92 NumCastsRemoved) && 93 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc, 94 NumCastsRemoved); 95 } 96 } 97 break; 98 } 99 case Instruction::Shl: 100 // If we are truncating the result of this SHL, and if it's a shift of a 101 // constant amount, we can always perform a SHL in a smaller type. 102 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 103 uint32_t BitWidth = Ty->getScalarSizeInBits(); 104 if (BitWidth < OrigTy->getScalarSizeInBits() && 105 CI->getLimitedValue(BitWidth) < BitWidth) 106 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 107 NumCastsRemoved); 108 } 109 break; 110 case Instruction::LShr: 111 // If this is a truncate of a logical shr, we can truncate it to a smaller 112 // lshr iff we know that the bits we would otherwise be shifting in are 113 // already zeros. 114 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 115 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 116 uint32_t BitWidth = Ty->getScalarSizeInBits(); 117 if (BitWidth < OrigBitWidth && 118 MaskedValueIsZero(I->getOperand(0), 119 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 120 CI->getLimitedValue(BitWidth) < BitWidth) { 121 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 122 NumCastsRemoved); 123 } 124 } 125 break; 126 case Instruction::ZExt: 127 case Instruction::SExt: 128 case Instruction::Trunc: 129 // If this is the same kind of case as our original (e.g. zext+zext), we 130 // can safely replace it. Note that replacing it does not reduce the number 131 // of casts in the input. 132 if (Opc == CastOpc) 133 return true; 134 135 // sext (zext ty1), ty2 -> zext ty2 136 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt) 137 return true; 138 break; 139 case Instruction::Select: { 140 SelectInst *SI = cast<SelectInst>(I); 141 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc, 142 NumCastsRemoved) && 143 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc, 144 NumCastsRemoved); 145 } 146 case Instruction::PHI: { 147 // We can change a phi if we can change all operands. 148 PHINode *PN = cast<PHINode>(I); 149 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 150 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc, 151 NumCastsRemoved)) 152 return false; 153 return true; 154 } 155 default: 156 // TODO: Can handle more cases here. 157 break; 158 } 159 160 return false; 161} 162 163/// EvaluateInDifferentType - Given an expression that 164/// CanEvaluateInDifferentType returns true for, actually insert the code to 165/// evaluate the expression. 166Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty, 167 bool isSigned) { 168 if (Constant *C = dyn_cast<Constant>(V)) 169 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 170 171 // Otherwise, it must be an instruction. 172 Instruction *I = cast<Instruction>(V); 173 Instruction *Res = 0; 174 unsigned Opc = I->getOpcode(); 175 switch (Opc) { 176 case Instruction::Add: 177 case Instruction::Sub: 178 case Instruction::Mul: 179 case Instruction::And: 180 case Instruction::Or: 181 case Instruction::Xor: 182 case Instruction::AShr: 183 case Instruction::LShr: 184 case Instruction::Shl: 185 case Instruction::UDiv: 186 case Instruction::URem: { 187 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 188 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 189 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 190 break; 191 } 192 case Instruction::Trunc: 193 case Instruction::ZExt: 194 case Instruction::SExt: 195 // If the source type of the cast is the type we're trying for then we can 196 // just return the source. There's no need to insert it because it is not 197 // new. 198 if (I->getOperand(0)->getType() == Ty) 199 return I->getOperand(0); 200 201 // Otherwise, must be the same type of cast, so just reinsert a new one. 202 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty); 203 break; 204 case Instruction::Select: { 205 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 206 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 207 Res = SelectInst::Create(I->getOperand(0), True, False); 208 break; 209 } 210 case Instruction::PHI: { 211 PHINode *OPN = cast<PHINode>(I); 212 PHINode *NPN = PHINode::Create(Ty); 213 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 214 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 215 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 216 } 217 Res = NPN; 218 break; 219 } 220 default: 221 // TODO: Can handle more cases here. 222 llvm_unreachable("Unreachable!"); 223 break; 224 } 225 226 Res->takeName(I); 227 return InsertNewInstBefore(Res, *I); 228} 229 230 231/// This function is a wrapper around CastInst::isEliminableCastPair. It 232/// simply extracts arguments and returns what that function returns. 233static Instruction::CastOps 234isEliminableCastPair( 235 const CastInst *CI, ///< The first cast instruction 236 unsigned opcode, ///< The opcode of the second cast instruction 237 const Type *DstTy, ///< The target type for the second cast instruction 238 TargetData *TD ///< The target data for pointer size 239) { 240 241 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above 242 const Type *MidTy = CI->getType(); // B from above 243 244 // Get the opcodes of the two Cast instructions 245 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 246 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 247 248 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 249 DstTy, 250 TD ? TD->getIntPtrType(CI->getContext()) : 0); 251 252 // We don't want to form an inttoptr or ptrtoint that converts to an integer 253 // type that differs from the pointer size. 254 if ((Res == Instruction::IntToPtr && 255 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 256 (Res == Instruction::PtrToInt && 257 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 258 Res = 0; 259 260 return Instruction::CastOps(Res); 261} 262 263/// ValueRequiresCast - Return true if the cast from "V to Ty" actually results 264/// in any code being generated. It does not require codegen if V is simple 265/// enough or if the cast can be folded into other casts. 266bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V, 267 const Type *Ty) { 268 if (V->getType() == Ty || isa<Constant>(V)) return false; 269 270 // If this is another cast that can be eliminated, it isn't codegen either. 271 if (const CastInst *CI = dyn_cast<CastInst>(V)) 272 if (isEliminableCastPair(CI, opcode, Ty, TD)) 273 return false; 274 return true; 275} 276 277 278/// @brief Implement the transforms common to all CastInst visitors. 279Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 280 Value *Src = CI.getOperand(0); 281 282 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 283 // eliminate it now. 284 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 285 if (Instruction::CastOps opc = 286 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 287 // The first cast (CSrc) is eliminable so we need to fix up or replace 288 // the second cast (CI). CSrc will then have a good chance of being dead. 289 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 290 } 291 } 292 293 // If we are casting a select then fold the cast into the select 294 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 295 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 296 return NV; 297 298 // If we are casting a PHI then fold the cast into the PHI 299 if (isa<PHINode>(Src)) { 300 // We don't do this if this would create a PHI node with an illegal type if 301 // it is currently legal. 302 if (!isa<IntegerType>(Src->getType()) || 303 !isa<IntegerType>(CI.getType()) || 304 ShouldChangeType(CI.getType(), Src->getType())) 305 if (Instruction *NV = FoldOpIntoPhi(CI)) 306 return NV; 307 } 308 309 return 0; 310} 311 312/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 313Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 314 Value *Src = CI.getOperand(0); 315 316 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 317 // If casting the result of a getelementptr instruction with no offset, turn 318 // this into a cast of the original pointer! 319 if (GEP->hasAllZeroIndices()) { 320 // Changing the cast operand is usually not a good idea but it is safe 321 // here because the pointer operand is being replaced with another 322 // pointer operand so the opcode doesn't need to change. 323 Worklist.Add(GEP); 324 CI.setOperand(0, GEP->getOperand(0)); 325 return &CI; 326 } 327 328 // If the GEP has a single use, and the base pointer is a bitcast, and the 329 // GEP computes a constant offset, see if we can convert these three 330 // instructions into fewer. This typically happens with unions and other 331 // non-type-safe code. 332 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) { 333 if (GEP->hasAllConstantIndices()) { 334 // We are guaranteed to get a constant from EmitGEPOffset. 335 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 336 int64_t Offset = OffsetV->getSExtValue(); 337 338 // Get the base pointer input of the bitcast, and the type it points to. 339 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 340 const Type *GEPIdxTy = 341 cast<PointerType>(OrigBase->getType())->getElementType(); 342 SmallVector<Value*, 8> NewIndices; 343 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 344 // If we were able to index down into an element, create the GEP 345 // and bitcast the result. This eliminates one bitcast, potentially 346 // two. 347 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 348 Builder->CreateInBoundsGEP(OrigBase, 349 NewIndices.begin(), NewIndices.end()) : 350 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 351 NGEP->takeName(GEP); 352 353 if (isa<BitCastInst>(CI)) 354 return new BitCastInst(NGEP, CI.getType()); 355 assert(isa<PtrToIntInst>(CI)); 356 return new PtrToIntInst(NGEP, CI.getType()); 357 } 358 } 359 } 360 } 361 362 return commonCastTransforms(CI); 363} 364 365/// commonIntCastTransforms - This function implements the common transforms 366/// for trunc, zext, and sext. 367Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) { 368 if (Instruction *Result = commonCastTransforms(CI)) 369 return Result; 370 371 Value *Src = CI.getOperand(0); 372 const Type *SrcTy = Src->getType(); 373 const Type *DestTy = CI.getType(); 374 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 375 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 376 377 // See if we can simplify any instructions used by the LHS whose sole 378 // purpose is to compute bits we don't care about. 379 if (SimplifyDemandedInstructionBits(CI)) 380 return &CI; 381 382 // If the source isn't an instruction or has more than one use then we 383 // can't do anything more. 384 Instruction *SrcI = dyn_cast<Instruction>(Src); 385 if (!SrcI || !Src->hasOneUse()) 386 return 0; 387 388 // Attempt to propagate the cast into the instruction for int->int casts. 389 int NumCastsRemoved = 0; 390 // Only do this if the dest type is a simple type, don't convert the 391 // expression tree to something weird like i93 unless the source is also 392 // strange. 393 if ((isa<VectorType>(DestTy) || 394 ShouldChangeType(SrcI->getType(), DestTy)) && 395 CanEvaluateInDifferentType(SrcI, DestTy, 396 CI.getOpcode(), NumCastsRemoved)) { 397 // If this cast is a truncate, evaluting in a different type always 398 // eliminates the cast, so it is always a win. If this is a zero-extension, 399 // we need to do an AND to maintain the clear top-part of the computation, 400 // so we require that the input have eliminated at least one cast. If this 401 // is a sign extension, we insert two new casts (to do the extension) so we 402 // require that two casts have been eliminated. 403 bool DoXForm = false; 404 bool JustReplace = false; 405 switch (CI.getOpcode()) { 406 default: 407 // All the others use floating point so we shouldn't actually 408 // get here because of the check above. 409 llvm_unreachable("Unknown cast type"); 410 case Instruction::Trunc: 411 DoXForm = true; 412 break; 413 case Instruction::ZExt: { 414 DoXForm = NumCastsRemoved >= 1; 415 416 if (!DoXForm && 0) { 417 // If it's unnecessary to issue an AND to clear the high bits, it's 418 // always profitable to do this xform. 419 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false); 420 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize)); 421 if (MaskedValueIsZero(TryRes, Mask)) 422 return ReplaceInstUsesWith(CI, TryRes); 423 424 if (Instruction *TryI = dyn_cast<Instruction>(TryRes)) 425 if (TryI->use_empty()) 426 EraseInstFromFunction(*TryI); 427 } 428 break; 429 } 430 case Instruction::SExt: { 431 DoXForm = NumCastsRemoved >= 2; 432 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) { 433 // If we do not have to emit the truncate + sext pair, then it's always 434 // profitable to do this xform. 435 // 436 // It's not safe to eliminate the trunc + sext pair if one of the 437 // eliminated cast is a truncate. e.g. 438 // t2 = trunc i32 t1 to i16 439 // t3 = sext i16 t2 to i32 440 // != 441 // i32 t1 442 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true); 443 unsigned NumSignBits = ComputeNumSignBits(TryRes); 444 if (NumSignBits > (DestBitSize - SrcBitSize)) 445 return ReplaceInstUsesWith(CI, TryRes); 446 447 if (Instruction *TryI = dyn_cast<Instruction>(TryRes)) 448 if (TryI->use_empty()) 449 EraseInstFromFunction(*TryI); 450 } 451 break; 452 } 453 } 454 455 if (DoXForm) { 456 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type" 457 " to avoid cast: " << CI); 458 Value *Res = EvaluateInDifferentType(SrcI, DestTy, 459 CI.getOpcode() == Instruction::SExt); 460 if (JustReplace) 461 // Just replace this cast with the result. 462 return ReplaceInstUsesWith(CI, Res); 463 464 assert(Res->getType() == DestTy); 465 switch (CI.getOpcode()) { 466 default: llvm_unreachable("Unknown cast type!"); 467 case Instruction::Trunc: 468 // Just replace this cast with the result. 469 return ReplaceInstUsesWith(CI, Res); 470 case Instruction::ZExt: { 471 assert(SrcBitSize < DestBitSize && "Not a zext?"); 472 473 // If the high bits are already zero, just replace this cast with the 474 // result. 475 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize)); 476 if (MaskedValueIsZero(Res, Mask)) 477 return ReplaceInstUsesWith(CI, Res); 478 479 // We need to emit an AND to clear the high bits. 480 Constant *C = ConstantInt::get(CI.getContext(), 481 APInt::getLowBitsSet(DestBitSize, SrcBitSize)); 482 return BinaryOperator::CreateAnd(Res, C); 483 } 484 case Instruction::SExt: { 485 // If the high bits are already filled with sign bit, just replace this 486 // cast with the result. 487 unsigned NumSignBits = ComputeNumSignBits(Res); 488 if (NumSignBits > (DestBitSize - SrcBitSize)) 489 return ReplaceInstUsesWith(CI, Res); 490 491 // We need to emit a cast to truncate, then a cast to sext. 492 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy); 493 } 494 } 495 } 496 } 497 498 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0; 499 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0; 500 501 switch (SrcI->getOpcode()) { 502 case Instruction::Add: 503 case Instruction::Mul: 504 case Instruction::And: 505 case Instruction::Or: 506 case Instruction::Xor: 507 // If we are discarding information, rewrite. 508 if (DestBitSize < SrcBitSize && DestBitSize != 1) { 509 // Don't insert two casts unless at least one can be eliminated. 510 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy) || 511 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy)) { 512 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName()); 513 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName()); 514 return BinaryOperator::Create( 515 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c); 516 } 517 } 518 519 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1 520 if (isa<ZExtInst>(CI) && SrcBitSize == 1 && 521 SrcI->getOpcode() == Instruction::Xor && 522 Op1 == ConstantInt::getTrue(CI.getContext()) && 523 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) { 524 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName()); 525 return BinaryOperator::CreateXor(New, 526 ConstantInt::get(CI.getType(), 1)); 527 } 528 break; 529 530 case Instruction::Shl: { 531 // Canonicalize trunc inside shl, if we can. 532 ConstantInt *CI = dyn_cast<ConstantInt>(Op1); 533 if (CI && DestBitSize < SrcBitSize && 534 CI->getLimitedValue(DestBitSize) < DestBitSize) { 535 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName()); 536 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName()); 537 return BinaryOperator::CreateShl(Op0c, Op1c); 538 } 539 break; 540 } 541 } 542 return 0; 543} 544 545 546Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 547 if (Instruction *Result = commonIntCastTransforms(CI)) 548 return Result; 549 550 Value *Src = CI.getOperand(0); 551 const Type *Ty = CI.getType(); 552 uint32_t DestBitWidth = Ty->getScalarSizeInBits(); 553 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits(); 554 555 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0) 556 if (DestBitWidth == 1) { 557 Constant *One = ConstantInt::get(Src->getType(), 1); 558 Src = Builder->CreateAnd(Src, One, "tmp"); 559 Value *Zero = Constant::getNullValue(Src->getType()); 560 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 561 } 562 563 // Optimize trunc(lshr(), c) to pull the shift through the truncate. 564 ConstantInt *ShAmtV = 0; 565 Value *ShiftOp = 0; 566 if (Src->hasOneUse() && 567 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) { 568 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth); 569 570 // Get a mask for the bits shifting in. 571 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth)); 572 if (MaskedValueIsZero(ShiftOp, Mask)) { 573 if (ShAmt >= DestBitWidth) // All zeros. 574 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty)); 575 576 // Okay, we can shrink this. Truncate the input, then return a new 577 // shift. 578 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName()); 579 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty); 580 return BinaryOperator::CreateLShr(V1, V2); 581 } 582 } 583 584 return 0; 585} 586 587/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 588/// in order to eliminate the icmp. 589Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 590 bool DoXform) { 591 // If we are just checking for a icmp eq of a single bit and zext'ing it 592 // to an integer, then shift the bit to the appropriate place and then 593 // cast to integer to avoid the comparison. 594 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 595 const APInt &Op1CV = Op1C->getValue(); 596 597 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 598 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 599 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 600 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 601 if (!DoXform) return ICI; 602 603 Value *In = ICI->getOperand(0); 604 Value *Sh = ConstantInt::get(In->getType(), 605 In->getType()->getScalarSizeInBits()-1); 606 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 607 if (In->getType() != CI.getType()) 608 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 609 610 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 611 Constant *One = ConstantInt::get(In->getType(), 1); 612 In = Builder->CreateXor(In, One, In->getName()+".not"); 613 } 614 615 return ReplaceInstUsesWith(CI, In); 616 } 617 618 619 620 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 621 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 622 // zext (X == 1) to i32 --> X iff X has only the low bit set. 623 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 624 // zext (X != 0) to i32 --> X iff X has only the low bit set. 625 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 626 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 627 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 628 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 629 // This only works for EQ and NE 630 ICI->isEquality()) { 631 // If Op1C some other power of two, convert: 632 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 633 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 634 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 635 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 636 637 APInt KnownZeroMask(~KnownZero); 638 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 639 if (!DoXform) return ICI; 640 641 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 642 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 643 // (X&4) == 2 --> false 644 // (X&4) != 2 --> true 645 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 646 isNE); 647 Res = ConstantExpr::getZExt(Res, CI.getType()); 648 return ReplaceInstUsesWith(CI, Res); 649 } 650 651 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 652 Value *In = ICI->getOperand(0); 653 if (ShiftAmt) { 654 // Perform a logical shr by shiftamt. 655 // Insert the shift to put the result in the low bit. 656 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 657 In->getName()+".lobit"); 658 } 659 660 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 661 Constant *One = ConstantInt::get(In->getType(), 1); 662 In = Builder->CreateXor(In, One, "tmp"); 663 } 664 665 if (CI.getType() == In->getType()) 666 return ReplaceInstUsesWith(CI, In); 667 else 668 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 669 } 670 } 671 } 672 673 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 674 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 675 // may lead to additional simplifications. 676 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 677 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 678 uint32_t BitWidth = ITy->getBitWidth(); 679 Value *LHS = ICI->getOperand(0); 680 Value *RHS = ICI->getOperand(1); 681 682 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 683 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 684 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 685 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 686 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 687 688 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 689 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 690 APInt UnknownBit = ~KnownBits; 691 if (UnknownBit.countPopulation() == 1) { 692 if (!DoXform) return ICI; 693 694 Value *Result = Builder->CreateXor(LHS, RHS); 695 696 // Mask off any bits that are set and won't be shifted away. 697 if (KnownOneLHS.uge(UnknownBit)) 698 Result = Builder->CreateAnd(Result, 699 ConstantInt::get(ITy, UnknownBit)); 700 701 // Shift the bit we're testing down to the lsb. 702 Result = Builder->CreateLShr( 703 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 704 705 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 706 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 707 Result->takeName(ICI); 708 return ReplaceInstUsesWith(CI, Result); 709 } 710 } 711 } 712 } 713 714 return 0; 715} 716 717Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 718 // If one of the common conversion will work, do it. 719 if (Instruction *Result = commonIntCastTransforms(CI)) 720 return Result; 721 722 Value *Src = CI.getOperand(0); 723 724 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 725 // types and if the sizes are just right we can convert this into a logical 726 // 'and' which will be much cheaper than the pair of casts. 727 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 728 // Get the sizes of the types involved. We know that the intermediate type 729 // will be smaller than A or C, but don't know the relation between A and C. 730 Value *A = CSrc->getOperand(0); 731 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 732 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 733 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 734 // If we're actually extending zero bits, then if 735 // SrcSize < DstSize: zext(a & mask) 736 // SrcSize == DstSize: a & mask 737 // SrcSize > DstSize: trunc(a) & mask 738 if (SrcSize < DstSize) { 739 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 740 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 741 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 742 return new ZExtInst(And, CI.getType()); 743 } 744 745 if (SrcSize == DstSize) { 746 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 747 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 748 AndValue)); 749 } 750 if (SrcSize > DstSize) { 751 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 752 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 753 return BinaryOperator::CreateAnd(Trunc, 754 ConstantInt::get(Trunc->getType(), 755 AndValue)); 756 } 757 } 758 759 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 760 return transformZExtICmp(ICI, CI); 761 762 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 763 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 764 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 765 // of the (zext icmp) will be transformed. 766 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 767 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 768 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 769 (transformZExtICmp(LHS, CI, false) || 770 transformZExtICmp(RHS, CI, false))) { 771 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 772 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 773 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 774 } 775 } 776 777 // zext(trunc(t) & C) -> (t & zext(C)). 778 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 779 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 780 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 781 Value *TI0 = TI->getOperand(0); 782 if (TI0->getType() == CI.getType()) 783 return 784 BinaryOperator::CreateAnd(TI0, 785 ConstantExpr::getZExt(C, CI.getType())); 786 } 787 788 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 789 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 790 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 791 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 792 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 793 And->getOperand(1) == C) 794 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 795 Value *TI0 = TI->getOperand(0); 796 if (TI0->getType() == CI.getType()) { 797 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 798 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 799 return BinaryOperator::CreateXor(NewAnd, ZC); 800 } 801 } 802 803 return 0; 804} 805 806Instruction *InstCombiner::visitSExt(SExtInst &CI) { 807 if (Instruction *I = commonIntCastTransforms(CI)) 808 return I; 809 810 Value *Src = CI.getOperand(0); 811 812 // Canonicalize sign-extend from i1 to a select. 813 if (Src->getType() == Type::getInt1Ty(CI.getContext())) 814 return SelectInst::Create(Src, 815 Constant::getAllOnesValue(CI.getType()), 816 Constant::getNullValue(CI.getType())); 817 818 // See if the value being truncated is already sign extended. If so, just 819 // eliminate the trunc/sext pair. 820 if (Operator::getOpcode(Src) == Instruction::Trunc) { 821 Value *Op = cast<User>(Src)->getOperand(0); 822 unsigned OpBits = Op->getType()->getScalarSizeInBits(); 823 unsigned MidBits = Src->getType()->getScalarSizeInBits(); 824 unsigned DestBits = CI.getType()->getScalarSizeInBits(); 825 unsigned NumSignBits = ComputeNumSignBits(Op); 826 827 if (OpBits == DestBits) { 828 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign 829 // bits, it is already ready. 830 if (NumSignBits > DestBits-MidBits) 831 return ReplaceInstUsesWith(CI, Op); 832 } else if (OpBits < DestBits) { 833 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign 834 // bits, just sext from i32. 835 if (NumSignBits > OpBits-MidBits) 836 return new SExtInst(Op, CI.getType(), "tmp"); 837 } else { 838 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign 839 // bits, just truncate to i32. 840 if (NumSignBits > OpBits-MidBits) 841 return new TruncInst(Op, CI.getType(), "tmp"); 842 } 843 } 844 845 // If the input is a shl/ashr pair of a same constant, then this is a sign 846 // extension from a smaller value. If we could trust arbitrary bitwidth 847 // integers, we could turn this into a truncate to the smaller bit and then 848 // use a sext for the whole extension. Since we don't, look deeper and check 849 // for a truncate. If the source and dest are the same type, eliminate the 850 // trunc and extend and just do shifts. For example, turn: 851 // %a = trunc i32 %i to i8 852 // %b = shl i8 %a, 6 853 // %c = ashr i8 %b, 6 854 // %d = sext i8 %c to i32 855 // into: 856 // %a = shl i32 %i, 30 857 // %d = ashr i32 %a, 30 858 Value *A = 0; 859 ConstantInt *BA = 0, *CA = 0; 860 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)), 861 m_ConstantInt(CA))) && 862 BA == CA && isa<TruncInst>(A)) { 863 Value *I = cast<TruncInst>(A)->getOperand(0); 864 if (I->getType() == CI.getType()) { 865 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 866 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 867 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 868 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 869 I = Builder->CreateShl(I, ShAmtV, CI.getName()); 870 return BinaryOperator::CreateAShr(I, ShAmtV); 871 } 872 } 873 874 return 0; 875} 876 877 878/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 879/// in the specified FP type without changing its value. 880static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 881 bool losesInfo; 882 APFloat F = CFP->getValueAPF(); 883 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 884 if (!losesInfo) 885 return ConstantFP::get(CFP->getContext(), F); 886 return 0; 887} 888 889/// LookThroughFPExtensions - If this is an fp extension instruction, look 890/// through it until we get the source value. 891static Value *LookThroughFPExtensions(Value *V) { 892 if (Instruction *I = dyn_cast<Instruction>(V)) 893 if (I->getOpcode() == Instruction::FPExt) 894 return LookThroughFPExtensions(I->getOperand(0)); 895 896 // If this value is a constant, return the constant in the smallest FP type 897 // that can accurately represent it. This allows us to turn 898 // (float)((double)X+2.0) into x+2.0f. 899 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 900 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 901 return V; // No constant folding of this. 902 // See if the value can be truncated to float and then reextended. 903 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 904 return V; 905 if (CFP->getType() == Type::getDoubleTy(V->getContext())) 906 return V; // Won't shrink. 907 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 908 return V; 909 // Don't try to shrink to various long double types. 910 } 911 912 return V; 913} 914 915Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 916 if (Instruction *I = commonCastTransforms(CI)) 917 return I; 918 919 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 920 // smaller than the destination type, we can eliminate the truncate by doing 921 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 922 // as many builtins (sqrt, etc). 923 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 924 if (OpI && OpI->hasOneUse()) { 925 switch (OpI->getOpcode()) { 926 default: break; 927 case Instruction::FAdd: 928 case Instruction::FSub: 929 case Instruction::FMul: 930 case Instruction::FDiv: 931 case Instruction::FRem: 932 const Type *SrcTy = OpI->getType(); 933 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 934 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 935 if (LHSTrunc->getType() != SrcTy && 936 RHSTrunc->getType() != SrcTy) { 937 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 938 // If the source types were both smaller than the destination type of 939 // the cast, do this xform. 940 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 941 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 942 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 943 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 944 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 945 } 946 } 947 break; 948 } 949 } 950 return 0; 951} 952 953Instruction *InstCombiner::visitFPExt(CastInst &CI) { 954 return commonCastTransforms(CI); 955} 956 957Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 958 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 959 if (OpI == 0) 960 return commonCastTransforms(FI); 961 962 // fptoui(uitofp(X)) --> X 963 // fptoui(sitofp(X)) --> X 964 // This is safe if the intermediate type has enough bits in its mantissa to 965 // accurately represent all values of X. For example, do not do this with 966 // i64->float->i64. This is also safe for sitofp case, because any negative 967 // 'X' value would cause an undefined result for the fptoui. 968 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 969 OpI->getOperand(0)->getType() == FI.getType() && 970 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 971 OpI->getType()->getFPMantissaWidth()) 972 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 973 974 return commonCastTransforms(FI); 975} 976 977Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 978 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 979 if (OpI == 0) 980 return commonCastTransforms(FI); 981 982 // fptosi(sitofp(X)) --> X 983 // fptosi(uitofp(X)) --> X 984 // This is safe if the intermediate type has enough bits in its mantissa to 985 // accurately represent all values of X. For example, do not do this with 986 // i64->float->i64. This is also safe for sitofp case, because any negative 987 // 'X' value would cause an undefined result for the fptoui. 988 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 989 OpI->getOperand(0)->getType() == FI.getType() && 990 (int)FI.getType()->getScalarSizeInBits() <= 991 OpI->getType()->getFPMantissaWidth()) 992 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 993 994 return commonCastTransforms(FI); 995} 996 997Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 998 return commonCastTransforms(CI); 999} 1000 1001Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1002 return commonCastTransforms(CI); 1003} 1004 1005Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1006 // If the destination integer type is smaller than the intptr_t type for 1007 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the 1008 // trunc to be exposed to other transforms. Don't do this for extending 1009 // ptrtoint's, because we don't know if the target sign or zero extends its 1010 // pointers. 1011 if (TD && 1012 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1013 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1014 TD->getIntPtrType(CI.getContext()), 1015 "tmp"); 1016 return new TruncInst(P, CI.getType()); 1017 } 1018 1019 return commonPointerCastTransforms(CI); 1020} 1021 1022 1023Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1024 // If the source integer type is larger than the intptr_t type for 1025 // this target, do a trunc to the intptr_t type, then inttoptr of it. This 1026 // allows the trunc to be exposed to other transforms. Don't do this for 1027 // extending inttoptr's, because we don't know if the target sign or zero 1028 // extends to pointers. 1029 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() > 1030 TD->getPointerSizeInBits()) { 1031 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1032 TD->getIntPtrType(CI.getContext()), "tmp"); 1033 return new IntToPtrInst(P, CI.getType()); 1034 } 1035 1036 if (Instruction *I = commonCastTransforms(CI)) 1037 return I; 1038 1039 return 0; 1040} 1041 1042Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1043 // If the operands are integer typed then apply the integer transforms, 1044 // otherwise just apply the common ones. 1045 Value *Src = CI.getOperand(0); 1046 const Type *SrcTy = Src->getType(); 1047 const Type *DestTy = CI.getType(); 1048 1049 if (isa<PointerType>(SrcTy)) { 1050 if (Instruction *I = commonPointerCastTransforms(CI)) 1051 return I; 1052 } else { 1053 if (Instruction *Result = commonCastTransforms(CI)) 1054 return Result; 1055 } 1056 1057 1058 // Get rid of casts from one type to the same type. These are useless and can 1059 // be replaced by the operand. 1060 if (DestTy == Src->getType()) 1061 return ReplaceInstUsesWith(CI, Src); 1062 1063 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1064 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1065 const Type *DstElTy = DstPTy->getElementType(); 1066 const Type *SrcElTy = SrcPTy->getElementType(); 1067 1068 // If the address spaces don't match, don't eliminate the bitcast, which is 1069 // required for changing types. 1070 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1071 return 0; 1072 1073 // If we are casting a alloca to a pointer to a type of the same 1074 // size, rewrite the allocation instruction to allocate the "right" type. 1075 // There is no need to modify malloc calls because it is their bitcast that 1076 // needs to be cleaned up. 1077 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1078 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1079 return V; 1080 1081 // If the source and destination are pointers, and this cast is equivalent 1082 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1083 // This can enhance SROA and other transforms that want type-safe pointers. 1084 Constant *ZeroUInt = 1085 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1086 unsigned NumZeros = 0; 1087 while (SrcElTy != DstElTy && 1088 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) && 1089 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1090 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1091 ++NumZeros; 1092 } 1093 1094 // If we found a path from the src to dest, create the getelementptr now. 1095 if (SrcElTy == DstElTy) { 1096 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1097 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1098 ((Instruction*) NULL)); 1099 } 1100 } 1101 1102 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1103 if (DestVTy->getNumElements() == 1) { 1104 if (!isa<VectorType>(SrcTy)) { 1105 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1106 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1107 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1108 } 1109 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1110 } 1111 } 1112 1113 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1114 if (SrcVTy->getNumElements() == 1) { 1115 if (!isa<VectorType>(DestTy)) { 1116 Value *Elem = 1117 Builder->CreateExtractElement(Src, 1118 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1119 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1120 } 1121 } 1122 } 1123 1124 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1125 if (SVI->hasOneUse()) { 1126 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is 1127 // a bitconvert to a vector with the same # elts. 1128 if (isa<VectorType>(DestTy) && 1129 cast<VectorType>(DestTy)->getNumElements() == 1130 SVI->getType()->getNumElements() && 1131 SVI->getType()->getNumElements() == 1132 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1133 CastInst *Tmp; 1134 // If either of the operands is a cast from CI.getType(), then 1135 // evaluating the shuffle in the casted destination's type will allow 1136 // us to eliminate at least one cast. 1137 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) && 1138 Tmp->getOperand(0)->getType() == DestTy) || 1139 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) && 1140 Tmp->getOperand(0)->getType() == DestTy)) { 1141 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1142 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1143 // Return a new shuffle vector. Use the same element ID's, as we 1144 // know the vector types match #elts. 1145 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1146 } 1147 } 1148 } 1149 } 1150 return 0; 1151} 1152