InstCombineCasts.cpp revision 7a34d6c450035dad9d494502c7c742137c42958e
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/// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear 21/// expression. If so, decompose it, returning some value X, such that Val is 22/// X*Scale+Offset. 23/// 24static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, 25 int &Offset) { 26 assert(Val->getType()->isInteger(32) && "Unexpected allocation size type!"); 27 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 28 Offset = CI->getZExtValue(); 29 Scale = 0; 30 return ConstantInt::get(Type::getInt32Ty(Val->getContext()), 0); 31 } 32 33 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) { 34 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 35 if (I->getOpcode() == Instruction::Shl) { 36 // This is a value scaled by '1 << the shift amt'. 37 Scale = 1U << RHS->getZExtValue(); 38 Offset = 0; 39 return I->getOperand(0); 40 } 41 42 if (I->getOpcode() == Instruction::Mul) { 43 // This value is scaled by 'RHS'. 44 Scale = RHS->getZExtValue(); 45 Offset = 0; 46 return I->getOperand(0); 47 } 48 49 if (I->getOpcode() == Instruction::Add) { 50 // We have X+C. Check to see if we really have (X*C2)+C1, 51 // where C1 is divisible by C2. 52 unsigned SubScale; 53 Value *SubVal = 54 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); 55 Offset += RHS->getZExtValue(); 56 Scale = SubScale; 57 return SubVal; 58 } 59 } 60 } 61 62 // Otherwise, we can't look past this. 63 Scale = 1; 64 Offset = 0; 65 return Val; 66} 67 68/// PromoteCastOfAllocation - If we find a cast of an allocation instruction, 69/// try to eliminate the cast by moving the type information into the alloc. 70Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 71 AllocaInst &AI) { 72 // This requires TargetData to get the alloca alignment and size information. 73 if (!TD) return 0; 74 75 const PointerType *PTy = cast<PointerType>(CI.getType()); 76 77 BuilderTy AllocaBuilder(*Builder); 78 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); 79 80 // Get the type really allocated and the type casted to. 81 const Type *AllocElTy = AI.getAllocatedType(); 82 const Type *CastElTy = PTy->getElementType(); 83 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0; 84 85 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy); 86 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy); 87 if (CastElTyAlign < AllocElTyAlign) return 0; 88 89 // If the allocation has multiple uses, only promote it if we are strictly 90 // increasing the alignment of the resultant allocation. If we keep it the 91 // same, we open the door to infinite loops of various kinds. (A reference 92 // from a dbg.declare doesn't count as a use for this purpose.) 93 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) && 94 CastElTyAlign == AllocElTyAlign) return 0; 95 96 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy); 97 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); 98 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 99 100 // See if we can satisfy the modulus by pulling a scale out of the array 101 // size argument. 102 unsigned ArraySizeScale; 103 int ArrayOffset; 104 Value *NumElements = // See if the array size is a decomposable linear expr. 105 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 106 107 // If we can now satisfy the modulus, by using a non-1 scale, we really can 108 // do the xform. 109 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 110 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 111 112 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 113 Value *Amt = 0; 114 if (Scale == 1) { 115 Amt = NumElements; 116 } else { 117 Amt = ConstantInt::get(Type::getInt32Ty(CI.getContext()), Scale); 118 // Insert before the alloca, not before the cast. 119 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp"); 120 } 121 122 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 123 Value *Off = ConstantInt::get(Type::getInt32Ty(CI.getContext()), 124 Offset, true); 125 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp"); 126 } 127 128 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 129 New->setAlignment(AI.getAlignment()); 130 New->takeName(&AI); 131 132 // If the allocation has one real use plus a dbg.declare, just remove the 133 // declare. 134 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) { 135 EraseInstFromFunction(*(Instruction*)DI); 136 } 137 // If the allocation has multiple real uses, insert a cast and change all 138 // things that used it to use the new cast. This will also hack on CI, but it 139 // will die soon. 140 else if (!AI.hasOneUse()) { 141 // New is the allocation instruction, pointer typed. AI is the original 142 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 143 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 144 AI.replaceAllUsesWith(NewCast); 145 } 146 return ReplaceInstUsesWith(CI, New); 147} 148 149 150/// CanEvaluateInDifferentType - Return true if we can take the specified value 151/// and return it as type Ty without inserting any new casts and without 152/// changing the computed value. This is used by code that tries to decide 153/// whether promoting or shrinking integer operations to wider or smaller types 154/// will allow us to eliminate a truncate or extend. 155/// 156/// This is a truncation operation if Ty is smaller than V->getType(), or an 157/// extension operation if Ty is larger. 158/// 159/// If CastOpc is a truncation, then Ty will be a type smaller than V. We 160/// should return true if trunc(V) can be computed by computing V in the smaller 161/// type. If V is an instruction, then trunc(inst(x,y)) can be computed as 162/// inst(trunc(x),trunc(y)), which only makes sense if x and y can be 163/// efficiently truncated. 164/// 165/// If CastOpc is a sext or zext, we are asking if the low bits of the value can 166/// bit computed in a larger type, which is then and'd or sext_in_reg'd to get 167/// the final result. 168bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty, 169 unsigned CastOpc, 170 int &NumCastsRemoved){ 171 // We can always evaluate constants in another type. 172 if (isa<Constant>(V)) 173 return true; 174 175 Instruction *I = dyn_cast<Instruction>(V); 176 if (!I) return false; 177 178 const Type *OrigTy = V->getType(); 179 180 // If this is an extension or truncate, we can often eliminate it. 181 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) { 182 // If this is a cast from the destination type, we can trivially eliminate 183 // it, and this will remove a cast overall. 184 if (I->getOperand(0)->getType() == Ty) { 185 // If the first operand is itself a cast, and is eliminable, do not count 186 // this as an eliminable cast. We would prefer to eliminate those two 187 // casts first. 188 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse()) 189 ++NumCastsRemoved; 190 return true; 191 } 192 } 193 194 // We can't extend or shrink something that has multiple uses: doing so would 195 // require duplicating the instruction in general, which isn't profitable. 196 if (!I->hasOneUse()) return false; 197 198 unsigned Opc = I->getOpcode(); 199 switch (Opc) { 200 case Instruction::Add: 201 case Instruction::Sub: 202 case Instruction::Mul: 203 case Instruction::And: 204 case Instruction::Or: 205 case Instruction::Xor: 206 // These operators can all arbitrarily be extended or truncated. 207 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 208 NumCastsRemoved) && 209 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc, 210 NumCastsRemoved); 211 212 case Instruction::UDiv: 213 case Instruction::URem: { 214 // UDiv and URem can be truncated if all the truncated bits are zero. 215 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 216 uint32_t BitWidth = Ty->getScalarSizeInBits(); 217 if (BitWidth < OrigBitWidth) { 218 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 219 if (MaskedValueIsZero(I->getOperand(0), Mask) && 220 MaskedValueIsZero(I->getOperand(1), Mask)) { 221 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 222 NumCastsRemoved) && 223 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc, 224 NumCastsRemoved); 225 } 226 } 227 break; 228 } 229 case Instruction::Shl: 230 // If we are truncating the result of this SHL, and if it's a shift of a 231 // constant amount, we can always perform a SHL in a smaller type. 232 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 233 uint32_t BitWidth = Ty->getScalarSizeInBits(); 234 if (BitWidth < OrigTy->getScalarSizeInBits() && 235 CI->getLimitedValue(BitWidth) < BitWidth) 236 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 237 NumCastsRemoved); 238 } 239 break; 240 case Instruction::LShr: 241 // If this is a truncate of a logical shr, we can truncate it to a smaller 242 // lshr iff we know that the bits we would otherwise be shifting in are 243 // already zeros. 244 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 245 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 246 uint32_t BitWidth = Ty->getScalarSizeInBits(); 247 if (BitWidth < OrigBitWidth && 248 MaskedValueIsZero(I->getOperand(0), 249 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 250 CI->getLimitedValue(BitWidth) < BitWidth) { 251 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc, 252 NumCastsRemoved); 253 } 254 } 255 break; 256 case Instruction::ZExt: 257 case Instruction::SExt: 258 case Instruction::Trunc: 259 // If this is the same kind of case as our original (e.g. zext+zext), we 260 // can safely replace it. Note that replacing it does not reduce the number 261 // of casts in the input. 262 if (Opc == CastOpc) 263 return true; 264 265 // sext (zext ty1), ty2 -> zext ty2 266 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt) 267 return true; 268 break; 269 case Instruction::Select: { 270 SelectInst *SI = cast<SelectInst>(I); 271 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc, 272 NumCastsRemoved) && 273 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc, 274 NumCastsRemoved); 275 } 276 case Instruction::PHI: { 277 // We can change a phi if we can change all operands. 278 PHINode *PN = cast<PHINode>(I); 279 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 280 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc, 281 NumCastsRemoved)) 282 return false; 283 return true; 284 } 285 default: 286 // TODO: Can handle more cases here. 287 break; 288 } 289 290 return false; 291} 292 293/// EvaluateInDifferentType - Given an expression that 294/// CanEvaluateInDifferentType returns true for, actually insert the code to 295/// evaluate the expression. 296Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty, 297 bool isSigned) { 298 if (Constant *C = dyn_cast<Constant>(V)) 299 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 300 301 // Otherwise, it must be an instruction. 302 Instruction *I = cast<Instruction>(V); 303 Instruction *Res = 0; 304 unsigned Opc = I->getOpcode(); 305 switch (Opc) { 306 case Instruction::Add: 307 case Instruction::Sub: 308 case Instruction::Mul: 309 case Instruction::And: 310 case Instruction::Or: 311 case Instruction::Xor: 312 case Instruction::AShr: 313 case Instruction::LShr: 314 case Instruction::Shl: 315 case Instruction::UDiv: 316 case Instruction::URem: { 317 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 318 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 319 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 320 break; 321 } 322 case Instruction::Trunc: 323 case Instruction::ZExt: 324 case Instruction::SExt: 325 // If the source type of the cast is the type we're trying for then we can 326 // just return the source. There's no need to insert it because it is not 327 // new. 328 if (I->getOperand(0)->getType() == Ty) 329 return I->getOperand(0); 330 331 // Otherwise, must be the same type of cast, so just reinsert a new one. 332 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty); 333 break; 334 case Instruction::Select: { 335 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 336 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 337 Res = SelectInst::Create(I->getOperand(0), True, False); 338 break; 339 } 340 case Instruction::PHI: { 341 PHINode *OPN = cast<PHINode>(I); 342 PHINode *NPN = PHINode::Create(Ty); 343 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 344 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 345 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 346 } 347 Res = NPN; 348 break; 349 } 350 default: 351 // TODO: Can handle more cases here. 352 llvm_unreachable("Unreachable!"); 353 break; 354 } 355 356 Res->takeName(I); 357 return InsertNewInstBefore(Res, *I); 358} 359 360 361/// This function is a wrapper around CastInst::isEliminableCastPair. It 362/// simply extracts arguments and returns what that function returns. 363static Instruction::CastOps 364isEliminableCastPair( 365 const CastInst *CI, ///< The first cast instruction 366 unsigned opcode, ///< The opcode of the second cast instruction 367 const Type *DstTy, ///< The target type for the second cast instruction 368 TargetData *TD ///< The target data for pointer size 369) { 370 371 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above 372 const Type *MidTy = CI->getType(); // B from above 373 374 // Get the opcodes of the two Cast instructions 375 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 376 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 377 378 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 379 DstTy, 380 TD ? TD->getIntPtrType(CI->getContext()) : 0); 381 382 // We don't want to form an inttoptr or ptrtoint that converts to an integer 383 // type that differs from the pointer size. 384 if ((Res == Instruction::IntToPtr && 385 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 386 (Res == Instruction::PtrToInt && 387 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 388 Res = 0; 389 390 return Instruction::CastOps(Res); 391} 392 393/// ValueRequiresCast - Return true if the cast from "V to Ty" actually results 394/// in any code being generated. It does not require codegen if V is simple 395/// enough or if the cast can be folded into other casts. 396bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V, 397 const Type *Ty) { 398 if (V->getType() == Ty || isa<Constant>(V)) return false; 399 400 // If this is another cast that can be eliminated, it isn't codegen either. 401 if (const CastInst *CI = dyn_cast<CastInst>(V)) 402 if (isEliminableCastPair(CI, opcode, Ty, TD)) 403 return false; 404 return true; 405} 406 407 408/// @brief Implement the transforms common to all CastInst visitors. 409Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 410 Value *Src = CI.getOperand(0); 411 412 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 413 // eliminate it now. 414 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 415 if (Instruction::CastOps opc = 416 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 417 // The first cast (CSrc) is eliminable so we need to fix up or replace 418 // the second cast (CI). CSrc will then have a good chance of being dead. 419 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 420 } 421 } 422 423 // If we are casting a select then fold the cast into the select 424 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 425 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 426 return NV; 427 428 // If we are casting a PHI then fold the cast into the PHI 429 if (isa<PHINode>(Src)) { 430 // We don't do this if this would create a PHI node with an illegal type if 431 // it is currently legal. 432 if (!isa<IntegerType>(Src->getType()) || 433 !isa<IntegerType>(CI.getType()) || 434 ShouldChangeType(CI.getType(), Src->getType())) 435 if (Instruction *NV = FoldOpIntoPhi(CI)) 436 return NV; 437 } 438 439 return 0; 440} 441 442/// commonIntCastTransforms - This function implements the common transforms 443/// for trunc, zext, and sext. 444Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) { 445 if (Instruction *Result = commonCastTransforms(CI)) 446 return Result; 447 448 // See if we can simplify any instructions used by the LHS whose sole 449 // purpose is to compute bits we don't care about. 450 if (SimplifyDemandedInstructionBits(CI)) 451 return &CI; 452 453 // If the source isn't an instruction or has more than one use then we 454 // can't do anything more. 455 Instruction *Src = dyn_cast<Instruction>(CI.getOperand(0)); 456 if (!Src || !Src->hasOneUse()) 457 return 0; 458 459 const Type *SrcTy = Src->getType(); 460 const Type *DestTy = CI.getType(); 461 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 462 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 463 464 // Attempt to propagate the cast into the instruction for int->int casts. 465 int NumCastsRemoved = 0; 466 // Only do this if the dest type is a simple type, don't convert the 467 // expression tree to something weird like i93 unless the source is also 468 // strange. 469 if ((isa<VectorType>(DestTy) || 470 ShouldChangeType(Src->getType(), DestTy)) && 471 CanEvaluateInDifferentType(Src, DestTy, 472 CI.getOpcode(), NumCastsRemoved)) { 473 // If this cast is a truncate, evaluting in a different type always 474 // eliminates the cast, so it is always a win. If this is a zero-extension, 475 // we need to do an AND to maintain the clear top-part of the computation, 476 // so we require that the input have eliminated at least one cast. If this 477 // is a sign extension, we insert two new casts (to do the extension) so we 478 // require that two casts have been eliminated. 479 bool DoXForm = false; 480 bool JustReplace = false; 481 switch (CI.getOpcode()) { 482 default: 483 // All the others use floating point so we shouldn't actually 484 // get here because of the check above. 485 llvm_unreachable("Unknown cast type"); 486 case Instruction::Trunc: 487 DoXForm = true; 488 break; 489 case Instruction::ZExt: { 490 DoXForm = NumCastsRemoved >= 1; 491 492 if (!DoXForm && 0) { 493 // If it's unnecessary to issue an AND to clear the high bits, it's 494 // always profitable to do this xform. 495 Value *TryRes = EvaluateInDifferentType(Src, DestTy, false); 496 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize)); 497 if (MaskedValueIsZero(TryRes, Mask)) 498 return ReplaceInstUsesWith(CI, TryRes); 499 500 if (Instruction *TryI = dyn_cast<Instruction>(TryRes)) 501 if (TryI->use_empty()) 502 EraseInstFromFunction(*TryI); 503 } 504 break; 505 } 506 case Instruction::SExt: { 507 DoXForm = NumCastsRemoved >= 2; 508 if (!DoXForm && !isa<TruncInst>(Src) && 0) { 509 // If we do not have to emit the truncate + sext pair, then it's always 510 // profitable to do this xform. 511 // 512 // It's not safe to eliminate the trunc + sext pair if one of the 513 // eliminated cast is a truncate. e.g. 514 // t2 = trunc i32 t1 to i16 515 // t3 = sext i16 t2 to i32 516 // != 517 // i32 t1 518 Value *TryRes = EvaluateInDifferentType(Src, DestTy, true); 519 unsigned NumSignBits = ComputeNumSignBits(TryRes); 520 if (NumSignBits > (DestBitSize - SrcBitSize)) 521 return ReplaceInstUsesWith(CI, TryRes); 522 523 if (Instruction *TryI = dyn_cast<Instruction>(TryRes)) 524 if (TryI->use_empty()) 525 EraseInstFromFunction(*TryI); 526 } 527 break; 528 } 529 } 530 531 if (DoXForm) { 532 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type" 533 " to avoid cast: " << CI); 534 Value *Res = EvaluateInDifferentType(Src, DestTy, 535 CI.getOpcode() == Instruction::SExt); 536 if (JustReplace) 537 // Just replace this cast with the result. 538 return ReplaceInstUsesWith(CI, Res); 539 540 assert(Res->getType() == DestTy); 541 switch (CI.getOpcode()) { 542 default: llvm_unreachable("Unknown cast type!"); 543 case Instruction::Trunc: 544 // Just replace this cast with the result. 545 return ReplaceInstUsesWith(CI, Res); 546 case Instruction::ZExt: { 547 assert(SrcBitSize < DestBitSize && "Not a zext?"); 548 549 // If the high bits are already zero, just replace this cast with the 550 // result. 551 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize)); 552 if (MaskedValueIsZero(Res, Mask)) 553 return ReplaceInstUsesWith(CI, Res); 554 555 // We need to emit an AND to clear the high bits. 556 Constant *C = ConstantInt::get(CI.getContext(), 557 APInt::getLowBitsSet(DestBitSize, SrcBitSize)); 558 return BinaryOperator::CreateAnd(Res, C); 559 } 560 case Instruction::SExt: { 561 // If the high bits are already filled with sign bit, just replace this 562 // cast with the result. 563 unsigned NumSignBits = ComputeNumSignBits(Res); 564 if (NumSignBits > (DestBitSize - SrcBitSize)) 565 return ReplaceInstUsesWith(CI, Res); 566 567 // We need to emit a cast to truncate, then a cast to sext. 568 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy); 569 } 570 } 571 } 572 } 573 574 return 0; 575} 576 577Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 578 if (Instruction *Result = commonIntCastTransforms(CI)) 579 return Result; 580 581 Value *Src = CI.getOperand(0); 582 const Type *DestTy = CI.getType(); 583 584 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 585 if (DestTy->getScalarSizeInBits() == 1) { 586 Constant *One = ConstantInt::get(Src->getType(), 1); 587 Src = Builder->CreateAnd(Src, One, "tmp"); 588 Value *Zero = Constant::getNullValue(Src->getType()); 589 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 590 } 591 592 return 0; 593} 594 595/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 596/// in order to eliminate the icmp. 597Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 598 bool DoXform) { 599 // If we are just checking for a icmp eq of a single bit and zext'ing it 600 // to an integer, then shift the bit to the appropriate place and then 601 // cast to integer to avoid the comparison. 602 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 603 const APInt &Op1CV = Op1C->getValue(); 604 605 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 606 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 607 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 608 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 609 if (!DoXform) return ICI; 610 611 Value *In = ICI->getOperand(0); 612 Value *Sh = ConstantInt::get(In->getType(), 613 In->getType()->getScalarSizeInBits()-1); 614 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 615 if (In->getType() != CI.getType()) 616 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 617 618 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 619 Constant *One = ConstantInt::get(In->getType(), 1); 620 In = Builder->CreateXor(In, One, In->getName()+".not"); 621 } 622 623 return ReplaceInstUsesWith(CI, In); 624 } 625 626 627 628 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 629 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 630 // zext (X == 1) to i32 --> X iff X has only the low bit set. 631 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 632 // zext (X != 0) to i32 --> X iff X has only the low bit set. 633 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 634 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 635 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 636 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 637 // This only works for EQ and NE 638 ICI->isEquality()) { 639 // If Op1C some other power of two, convert: 640 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 641 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 642 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 643 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 644 645 APInt KnownZeroMask(~KnownZero); 646 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 647 if (!DoXform) return ICI; 648 649 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 650 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 651 // (X&4) == 2 --> false 652 // (X&4) != 2 --> true 653 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 654 isNE); 655 Res = ConstantExpr::getZExt(Res, CI.getType()); 656 return ReplaceInstUsesWith(CI, Res); 657 } 658 659 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 660 Value *In = ICI->getOperand(0); 661 if (ShiftAmt) { 662 // Perform a logical shr by shiftamt. 663 // Insert the shift to put the result in the low bit. 664 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 665 In->getName()+".lobit"); 666 } 667 668 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 669 Constant *One = ConstantInt::get(In->getType(), 1); 670 In = Builder->CreateXor(In, One, "tmp"); 671 } 672 673 if (CI.getType() == In->getType()) 674 return ReplaceInstUsesWith(CI, In); 675 else 676 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 677 } 678 } 679 } 680 681 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 682 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 683 // may lead to additional simplifications. 684 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 685 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 686 uint32_t BitWidth = ITy->getBitWidth(); 687 Value *LHS = ICI->getOperand(0); 688 Value *RHS = ICI->getOperand(1); 689 690 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 691 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 692 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 693 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 694 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 695 696 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 697 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 698 APInt UnknownBit = ~KnownBits; 699 if (UnknownBit.countPopulation() == 1) { 700 if (!DoXform) return ICI; 701 702 Value *Result = Builder->CreateXor(LHS, RHS); 703 704 // Mask off any bits that are set and won't be shifted away. 705 if (KnownOneLHS.uge(UnknownBit)) 706 Result = Builder->CreateAnd(Result, 707 ConstantInt::get(ITy, UnknownBit)); 708 709 // Shift the bit we're testing down to the lsb. 710 Result = Builder->CreateLShr( 711 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 712 713 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 714 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 715 Result->takeName(ICI); 716 return ReplaceInstUsesWith(CI, Result); 717 } 718 } 719 } 720 } 721 722 return 0; 723} 724 725Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 726 // If one of the common conversion will work, do it. 727 if (Instruction *Result = commonIntCastTransforms(CI)) 728 return Result; 729 730 Value *Src = CI.getOperand(0); 731 732 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 733 // types and if the sizes are just right we can convert this into a logical 734 // 'and' which will be much cheaper than the pair of casts. 735 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 736 // Get the sizes of the types involved. We know that the intermediate type 737 // will be smaller than A or C, but don't know the relation between A and C. 738 Value *A = CSrc->getOperand(0); 739 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 740 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 741 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 742 // If we're actually extending zero bits, then if 743 // SrcSize < DstSize: zext(a & mask) 744 // SrcSize == DstSize: a & mask 745 // SrcSize > DstSize: trunc(a) & mask 746 if (SrcSize < DstSize) { 747 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 748 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 749 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 750 return new ZExtInst(And, CI.getType()); 751 } 752 753 if (SrcSize == DstSize) { 754 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 755 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 756 AndValue)); 757 } 758 if (SrcSize > DstSize) { 759 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 760 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 761 return BinaryOperator::CreateAnd(Trunc, 762 ConstantInt::get(Trunc->getType(), 763 AndValue)); 764 } 765 } 766 767 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 768 return transformZExtICmp(ICI, CI); 769 770 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 771 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 772 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 773 // of the (zext icmp) will be transformed. 774 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 775 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 776 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 777 (transformZExtICmp(LHS, CI, false) || 778 transformZExtICmp(RHS, CI, false))) { 779 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 780 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 781 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 782 } 783 } 784 785 // zext(trunc(t) & C) -> (t & zext(C)). 786 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 787 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 788 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 789 Value *TI0 = TI->getOperand(0); 790 if (TI0->getType() == CI.getType()) 791 return 792 BinaryOperator::CreateAnd(TI0, 793 ConstantExpr::getZExt(C, CI.getType())); 794 } 795 796 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 797 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 798 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 799 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 800 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 801 And->getOperand(1) == C) 802 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 803 Value *TI0 = TI->getOperand(0); 804 if (TI0->getType() == CI.getType()) { 805 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 806 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 807 return BinaryOperator::CreateXor(NewAnd, ZC); 808 } 809 } 810 811 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 812 Value *X; 813 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isInteger(1) && 814 match(SrcI, m_Not(m_Value(X))) && 815 (!X->hasOneUse() || !isa<CmpInst>(X))) { 816 Value *New = Builder->CreateZExt(X, CI.getType()); 817 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 818 } 819 820 return 0; 821} 822 823Instruction *InstCombiner::visitSExt(SExtInst &CI) { 824 if (Instruction *I = commonIntCastTransforms(CI)) 825 return I; 826 827 Value *Src = CI.getOperand(0); 828 829 // Canonicalize sign-extend from i1 to a select. 830 if (Src->getType()->isInteger(1)) 831 return SelectInst::Create(Src, 832 Constant::getAllOnesValue(CI.getType()), 833 Constant::getNullValue(CI.getType())); 834 835 // See if the value being truncated is already sign extended. If so, just 836 // eliminate the trunc/sext pair. 837 if (Operator::getOpcode(Src) == Instruction::Trunc) { 838 Value *Op = cast<User>(Src)->getOperand(0); 839 unsigned OpBits = Op->getType()->getScalarSizeInBits(); 840 unsigned MidBits = Src->getType()->getScalarSizeInBits(); 841 unsigned DestBits = CI.getType()->getScalarSizeInBits(); 842 unsigned NumSignBits = ComputeNumSignBits(Op); 843 844 if (OpBits == DestBits) { 845 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign 846 // bits, it is already ready. 847 if (NumSignBits > DestBits-MidBits) 848 return ReplaceInstUsesWith(CI, Op); 849 } else if (OpBits < DestBits) { 850 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign 851 // bits, just sext from i32. 852 if (NumSignBits > OpBits-MidBits) 853 return new SExtInst(Op, CI.getType(), "tmp"); 854 } else { 855 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign 856 // bits, just truncate to i32. 857 if (NumSignBits > OpBits-MidBits) 858 return new TruncInst(Op, CI.getType(), "tmp"); 859 } 860 } 861 862 // If the input is a shl/ashr pair of a same constant, then this is a sign 863 // extension from a smaller value. If we could trust arbitrary bitwidth 864 // integers, we could turn this into a truncate to the smaller bit and then 865 // use a sext for the whole extension. Since we don't, look deeper and check 866 // for a truncate. If the source and dest are the same type, eliminate the 867 // trunc and extend and just do shifts. For example, turn: 868 // %a = trunc i32 %i to i8 869 // %b = shl i8 %a, 6 870 // %c = ashr i8 %b, 6 871 // %d = sext i8 %c to i32 872 // into: 873 // %a = shl i32 %i, 30 874 // %d = ashr i32 %a, 30 875 Value *A = 0; 876 ConstantInt *BA = 0, *CA = 0; 877 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)), 878 m_ConstantInt(CA))) && 879 BA == CA && isa<TruncInst>(A)) { 880 Value *I = cast<TruncInst>(A)->getOperand(0); 881 if (I->getType() == CI.getType()) { 882 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 883 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 884 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 885 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 886 I = Builder->CreateShl(I, ShAmtV, CI.getName()); 887 return BinaryOperator::CreateAShr(I, ShAmtV); 888 } 889 } 890 891 return 0; 892} 893 894 895/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 896/// in the specified FP type without changing its value. 897static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 898 bool losesInfo; 899 APFloat F = CFP->getValueAPF(); 900 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 901 if (!losesInfo) 902 return ConstantFP::get(CFP->getContext(), F); 903 return 0; 904} 905 906/// LookThroughFPExtensions - If this is an fp extension instruction, look 907/// through it until we get the source value. 908static Value *LookThroughFPExtensions(Value *V) { 909 if (Instruction *I = dyn_cast<Instruction>(V)) 910 if (I->getOpcode() == Instruction::FPExt) 911 return LookThroughFPExtensions(I->getOperand(0)); 912 913 // If this value is a constant, return the constant in the smallest FP type 914 // that can accurately represent it. This allows us to turn 915 // (float)((double)X+2.0) into x+2.0f. 916 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 917 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 918 return V; // No constant folding of this. 919 // See if the value can be truncated to float and then reextended. 920 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 921 return V; 922 if (CFP->getType()->isDoubleTy()) 923 return V; // Won't shrink. 924 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 925 return V; 926 // Don't try to shrink to various long double types. 927 } 928 929 return V; 930} 931 932Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 933 if (Instruction *I = commonCastTransforms(CI)) 934 return I; 935 936 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 937 // smaller than the destination type, we can eliminate the truncate by doing 938 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 939 // as many builtins (sqrt, etc). 940 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 941 if (OpI && OpI->hasOneUse()) { 942 switch (OpI->getOpcode()) { 943 default: break; 944 case Instruction::FAdd: 945 case Instruction::FSub: 946 case Instruction::FMul: 947 case Instruction::FDiv: 948 case Instruction::FRem: 949 const Type *SrcTy = OpI->getType(); 950 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 951 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 952 if (LHSTrunc->getType() != SrcTy && 953 RHSTrunc->getType() != SrcTy) { 954 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 955 // If the source types were both smaller than the destination type of 956 // the cast, do this xform. 957 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 958 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 959 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 960 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 961 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 962 } 963 } 964 break; 965 } 966 } 967 return 0; 968} 969 970Instruction *InstCombiner::visitFPExt(CastInst &CI) { 971 return commonCastTransforms(CI); 972} 973 974Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 975 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 976 if (OpI == 0) 977 return commonCastTransforms(FI); 978 979 // fptoui(uitofp(X)) --> X 980 // fptoui(sitofp(X)) --> X 981 // This is safe if the intermediate type has enough bits in its mantissa to 982 // accurately represent all values of X. For example, do not do this with 983 // i64->float->i64. This is also safe for sitofp case, because any negative 984 // 'X' value would cause an undefined result for the fptoui. 985 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 986 OpI->getOperand(0)->getType() == FI.getType() && 987 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 988 OpI->getType()->getFPMantissaWidth()) 989 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 990 991 return commonCastTransforms(FI); 992} 993 994Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 995 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 996 if (OpI == 0) 997 return commonCastTransforms(FI); 998 999 // fptosi(sitofp(X)) --> X 1000 // fptosi(uitofp(X)) --> X 1001 // This is safe if the intermediate type has enough bits in its mantissa to 1002 // accurately represent all values of X. For example, do not do this with 1003 // i64->float->i64. This is also safe for sitofp case, because any negative 1004 // 'X' value would cause an undefined result for the fptoui. 1005 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1006 OpI->getOperand(0)->getType() == FI.getType() && 1007 (int)FI.getType()->getScalarSizeInBits() <= 1008 OpI->getType()->getFPMantissaWidth()) 1009 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1010 1011 return commonCastTransforms(FI); 1012} 1013 1014Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1015 return commonCastTransforms(CI); 1016} 1017 1018Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1019 return commonCastTransforms(CI); 1020} 1021 1022Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1023 // If the source integer type is larger than the intptr_t type for 1024 // this target, do a trunc to the intptr_t type, then inttoptr of it. This 1025 // allows the trunc to be exposed to other transforms. Don't do this for 1026 // extending inttoptr's, because we don't know if the target sign or zero 1027 // extends to pointers. 1028 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() > 1029 TD->getPointerSizeInBits()) { 1030 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1031 TD->getIntPtrType(CI.getContext()), "tmp"); 1032 return new IntToPtrInst(P, CI.getType()); 1033 } 1034 1035 if (Instruction *I = commonCastTransforms(CI)) 1036 return I; 1037 1038 return 0; 1039} 1040 1041/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1042Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1043 Value *Src = CI.getOperand(0); 1044 1045 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1046 // If casting the result of a getelementptr instruction with no offset, turn 1047 // this into a cast of the original pointer! 1048 if (GEP->hasAllZeroIndices()) { 1049 // Changing the cast operand is usually not a good idea but it is safe 1050 // here because the pointer operand is being replaced with another 1051 // pointer operand so the opcode doesn't need to change. 1052 Worklist.Add(GEP); 1053 CI.setOperand(0, GEP->getOperand(0)); 1054 return &CI; 1055 } 1056 1057 // If the GEP has a single use, and the base pointer is a bitcast, and the 1058 // GEP computes a constant offset, see if we can convert these three 1059 // instructions into fewer. This typically happens with unions and other 1060 // non-type-safe code. 1061 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1062 GEP->hasAllConstantIndices()) { 1063 // We are guaranteed to get a constant from EmitGEPOffset. 1064 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1065 int64_t Offset = OffsetV->getSExtValue(); 1066 1067 // Get the base pointer input of the bitcast, and the type it points to. 1068 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1069 const Type *GEPIdxTy = 1070 cast<PointerType>(OrigBase->getType())->getElementType(); 1071 SmallVector<Value*, 8> NewIndices; 1072 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1073 // If we were able to index down into an element, create the GEP 1074 // and bitcast the result. This eliminates one bitcast, potentially 1075 // two. 1076 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1077 Builder->CreateInBoundsGEP(OrigBase, 1078 NewIndices.begin(), NewIndices.end()) : 1079 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 1080 NGEP->takeName(GEP); 1081 1082 if (isa<BitCastInst>(CI)) 1083 return new BitCastInst(NGEP, CI.getType()); 1084 assert(isa<PtrToIntInst>(CI)); 1085 return new PtrToIntInst(NGEP, CI.getType()); 1086 } 1087 } 1088 } 1089 1090 return commonCastTransforms(CI); 1091} 1092 1093Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1094 // If the destination integer type is smaller than the intptr_t type for 1095 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the 1096 // trunc to be exposed to other transforms. Don't do this for extending 1097 // ptrtoint's, because we don't know if the target sign or zero extends its 1098 // pointers. 1099 if (TD && 1100 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1101 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1102 TD->getIntPtrType(CI.getContext()), 1103 "tmp"); 1104 return new TruncInst(P, CI.getType()); 1105 } 1106 1107 return commonPointerCastTransforms(CI); 1108} 1109 1110Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1111 // If the operands are integer typed then apply the integer transforms, 1112 // otherwise just apply the common ones. 1113 Value *Src = CI.getOperand(0); 1114 const Type *SrcTy = Src->getType(); 1115 const Type *DestTy = CI.getType(); 1116 1117 // Get rid of casts from one type to the same type. These are useless and can 1118 // be replaced by the operand. 1119 if (DestTy == Src->getType()) 1120 return ReplaceInstUsesWith(CI, Src); 1121 1122 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1123 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1124 const Type *DstElTy = DstPTy->getElementType(); 1125 const Type *SrcElTy = SrcPTy->getElementType(); 1126 1127 // If the address spaces don't match, don't eliminate the bitcast, which is 1128 // required for changing types. 1129 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1130 return 0; 1131 1132 // If we are casting a alloca to a pointer to a type of the same 1133 // size, rewrite the allocation instruction to allocate the "right" type. 1134 // There is no need to modify malloc calls because it is their bitcast that 1135 // needs to be cleaned up. 1136 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1137 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1138 return V; 1139 1140 // If the source and destination are pointers, and this cast is equivalent 1141 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1142 // This can enhance SROA and other transforms that want type-safe pointers. 1143 Constant *ZeroUInt = 1144 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1145 unsigned NumZeros = 0; 1146 while (SrcElTy != DstElTy && 1147 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) && 1148 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1149 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1150 ++NumZeros; 1151 } 1152 1153 // If we found a path from the src to dest, create the getelementptr now. 1154 if (SrcElTy == DstElTy) { 1155 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1156 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1157 ((Instruction*)NULL)); 1158 } 1159 } 1160 1161 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1162 if (DestVTy->getNumElements() == 1 && !isa<VectorType>(SrcTy)) { 1163 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1164 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1165 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1166 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1167 } 1168 } 1169 1170 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1171 if (SrcVTy->getNumElements() == 1 && !isa<VectorType>(DestTy)) { 1172 Value *Elem = 1173 Builder->CreateExtractElement(Src, 1174 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1175 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1176 } 1177 } 1178 1179 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1180 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1181 // a bitconvert to a vector with the same # elts. 1182 if (SVI->hasOneUse() && isa<VectorType>(DestTy) && 1183 cast<VectorType>(DestTy)->getNumElements() == 1184 SVI->getType()->getNumElements() && 1185 SVI->getType()->getNumElements() == 1186 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1187 BitCastInst *Tmp; 1188 // If either of the operands is a cast from CI.getType(), then 1189 // evaluating the shuffle in the casted destination's type will allow 1190 // us to eliminate at least one cast. 1191 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1192 Tmp->getOperand(0)->getType() == DestTy) || 1193 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1194 Tmp->getOperand(0)->getType() == DestTy)) { 1195 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1196 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1197 // Return a new shuffle vector. Use the same element ID's, as we 1198 // know the vector types match #elts. 1199 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1200 } 1201 } 1202 } 1203 1204 if (isa<PointerType>(SrcTy)) 1205 return commonPointerCastTransforms(CI); 1206 return commonCastTransforms(CI); 1207} 1208