InstCombineCasts.cpp revision 204642
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()->isIntegerTy(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} 148 149 150 151/// EvaluateInDifferentType - Given an expression that 152/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 153/// insert the code to evaluate the expression. 154Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty, 155 bool isSigned) { 156 if (Constant *C = dyn_cast<Constant>(V)) { 157 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 158 // If we got a constantexpr back, try to simplify it with TD info. 159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 160 C = ConstantFoldConstantExpression(CE, TD); 161 return C; 162 } 163 164 // Otherwise, it must be an instruction. 165 Instruction *I = cast<Instruction>(V); 166 Instruction *Res = 0; 167 unsigned Opc = I->getOpcode(); 168 switch (Opc) { 169 case Instruction::Add: 170 case Instruction::Sub: 171 case Instruction::Mul: 172 case Instruction::And: 173 case Instruction::Or: 174 case Instruction::Xor: 175 case Instruction::AShr: 176 case Instruction::LShr: 177 case Instruction::Shl: 178 case Instruction::UDiv: 179 case Instruction::URem: { 180 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 181 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 182 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 183 break; 184 } 185 case Instruction::Trunc: 186 case Instruction::ZExt: 187 case Instruction::SExt: 188 // If the source type of the cast is the type we're trying for then we can 189 // just return the source. There's no need to insert it because it is not 190 // new. 191 if (I->getOperand(0)->getType() == Ty) 192 return I->getOperand(0); 193 194 // Otherwise, must be the same type of cast, so just reinsert a new one. 195 // This also handles the case of zext(trunc(x)) -> zext(x). 196 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 197 Opc == Instruction::SExt); 198 break; 199 case Instruction::Select: { 200 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 201 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 202 Res = SelectInst::Create(I->getOperand(0), True, False); 203 break; 204 } 205 case Instruction::PHI: { 206 PHINode *OPN = cast<PHINode>(I); 207 PHINode *NPN = PHINode::Create(Ty); 208 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 209 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 210 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 211 } 212 Res = NPN; 213 break; 214 } 215 default: 216 // TODO: Can handle more cases here. 217 llvm_unreachable("Unreachable!"); 218 break; 219 } 220 221 Res->takeName(I); 222 return InsertNewInstBefore(Res, *I); 223} 224 225 226/// This function is a wrapper around CastInst::isEliminableCastPair. It 227/// simply extracts arguments and returns what that function returns. 228static Instruction::CastOps 229isEliminableCastPair( 230 const CastInst *CI, ///< The first cast instruction 231 unsigned opcode, ///< The opcode of the second cast instruction 232 const Type *DstTy, ///< The target type for the second cast instruction 233 TargetData *TD ///< The target data for pointer size 234) { 235 236 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above 237 const Type *MidTy = CI->getType(); // B from above 238 239 // Get the opcodes of the two Cast instructions 240 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 241 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 242 243 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 244 DstTy, 245 TD ? TD->getIntPtrType(CI->getContext()) : 0); 246 247 // We don't want to form an inttoptr or ptrtoint that converts to an integer 248 // type that differs from the pointer size. 249 if ((Res == Instruction::IntToPtr && 250 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 251 (Res == Instruction::PtrToInt && 252 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 253 Res = 0; 254 255 return Instruction::CastOps(Res); 256} 257 258/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 259/// results in any code being generated and is interesting to optimize out. If 260/// the cast can be eliminated by some other simple transformation, we prefer 261/// to do the simplification first. 262bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 263 const Type *Ty) { 264 // Noop casts and casts of constants should be eliminated trivially. 265 if (V->getType() == Ty || isa<Constant>(V)) return false; 266 267 // If this is another cast that can be eliminated, we prefer to have it 268 // eliminated. 269 if (const CastInst *CI = dyn_cast<CastInst>(V)) 270 if (isEliminableCastPair(CI, opc, Ty, TD)) 271 return false; 272 273 // If this is a vector sext from a compare, then we don't want to break the 274 // idiom where each element of the extended vector is either zero or all ones. 275 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 276 return false; 277 278 return true; 279} 280 281 282/// @brief Implement the transforms common to all CastInst visitors. 283Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 284 Value *Src = CI.getOperand(0); 285 286 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 287 // eliminate it now. 288 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 289 if (Instruction::CastOps opc = 290 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 291 // The first cast (CSrc) is eliminable so we need to fix up or replace 292 // the second cast (CI). CSrc will then have a good chance of being dead. 293 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 294 } 295 } 296 297 // If we are casting a select then fold the cast into the select 298 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 299 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 300 return NV; 301 302 // If we are casting a PHI then fold the cast into the PHI 303 if (isa<PHINode>(Src)) { 304 // We don't do this if this would create a PHI node with an illegal type if 305 // it is currently legal. 306 if (!Src->getType()->isIntegerTy() || 307 !CI.getType()->isIntegerTy() || 308 ShouldChangeType(CI.getType(), Src->getType())) 309 if (Instruction *NV = FoldOpIntoPhi(CI)) 310 return NV; 311 } 312 313 return 0; 314} 315 316/// CanEvaluateTruncated - Return true if we can evaluate the specified 317/// expression tree as type Ty instead of its larger type, and arrive with the 318/// same value. This is used by code that tries to eliminate truncates. 319/// 320/// Ty will always be a type smaller than V. We should return true if trunc(V) 321/// can be computed by computing V in the smaller type. If V is an instruction, 322/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 323/// makes sense if x and y can be efficiently truncated. 324/// 325/// This function works on both vectors and scalars. 326/// 327static bool CanEvaluateTruncated(Value *V, const Type *Ty) { 328 // We can always evaluate constants in another type. 329 if (isa<Constant>(V)) 330 return true; 331 332 Instruction *I = dyn_cast<Instruction>(V); 333 if (!I) return false; 334 335 const Type *OrigTy = V->getType(); 336 337 // If this is an extension from the dest type, we can eliminate it, even if it 338 // has multiple uses. 339 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 340 I->getOperand(0)->getType() == Ty) 341 return true; 342 343 // We can't extend or shrink something that has multiple uses: doing so would 344 // require duplicating the instruction in general, which isn't profitable. 345 if (!I->hasOneUse()) return false; 346 347 unsigned Opc = I->getOpcode(); 348 switch (Opc) { 349 case Instruction::Add: 350 case Instruction::Sub: 351 case Instruction::Mul: 352 case Instruction::And: 353 case Instruction::Or: 354 case Instruction::Xor: 355 // These operators can all arbitrarily be extended or truncated. 356 return CanEvaluateTruncated(I->getOperand(0), Ty) && 357 CanEvaluateTruncated(I->getOperand(1), Ty); 358 359 case Instruction::UDiv: 360 case Instruction::URem: { 361 // UDiv and URem can be truncated if all the truncated bits are zero. 362 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 363 uint32_t BitWidth = Ty->getScalarSizeInBits(); 364 if (BitWidth < OrigBitWidth) { 365 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 366 if (MaskedValueIsZero(I->getOperand(0), Mask) && 367 MaskedValueIsZero(I->getOperand(1), Mask)) { 368 return CanEvaluateTruncated(I->getOperand(0), Ty) && 369 CanEvaluateTruncated(I->getOperand(1), Ty); 370 } 371 } 372 break; 373 } 374 case Instruction::Shl: 375 // If we are truncating the result of this SHL, and if it's a shift of a 376 // constant amount, we can always perform a SHL in a smaller type. 377 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 378 uint32_t BitWidth = Ty->getScalarSizeInBits(); 379 if (CI->getLimitedValue(BitWidth) < BitWidth) 380 return CanEvaluateTruncated(I->getOperand(0), Ty); 381 } 382 break; 383 case Instruction::LShr: 384 // If this is a truncate of a logical shr, we can truncate it to a smaller 385 // lshr iff we know that the bits we would otherwise be shifting in are 386 // already zeros. 387 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 388 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 389 uint32_t BitWidth = Ty->getScalarSizeInBits(); 390 if (MaskedValueIsZero(I->getOperand(0), 391 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 392 CI->getLimitedValue(BitWidth) < BitWidth) { 393 return CanEvaluateTruncated(I->getOperand(0), Ty); 394 } 395 } 396 break; 397 case Instruction::Trunc: 398 // trunc(trunc(x)) -> trunc(x) 399 return true; 400 case Instruction::Select: { 401 SelectInst *SI = cast<SelectInst>(I); 402 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 403 CanEvaluateTruncated(SI->getFalseValue(), Ty); 404 } 405 case Instruction::PHI: { 406 // We can change a phi if we can change all operands. Note that we never 407 // get into trouble with cyclic PHIs here because we only consider 408 // instructions with a single use. 409 PHINode *PN = cast<PHINode>(I); 410 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 411 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 412 return false; 413 return true; 414 } 415 default: 416 // TODO: Can handle more cases here. 417 break; 418 } 419 420 return false; 421} 422 423Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 424 if (Instruction *Result = commonCastTransforms(CI)) 425 return Result; 426 427 // See if we can simplify any instructions used by the input whose sole 428 // purpose is to compute bits we don't care about. 429 if (SimplifyDemandedInstructionBits(CI)) 430 return &CI; 431 432 Value *Src = CI.getOperand(0); 433 const Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 434 435 // Attempt to truncate the entire input expression tree to the destination 436 // type. Only do this if the dest type is a simple type, don't convert the 437 // expression tree to something weird like i93 unless the source is also 438 // strange. 439 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 440 CanEvaluateTruncated(Src, DestTy)) { 441 442 // If this cast is a truncate, evaluting in a different type always 443 // eliminates the cast, so it is always a win. 444 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 445 " to avoid cast: " << CI); 446 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 447 assert(Res->getType() == DestTy); 448 return ReplaceInstUsesWith(CI, Res); 449 } 450 451 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 452 if (DestTy->getScalarSizeInBits() == 1) { 453 Constant *One = ConstantInt::get(Src->getType(), 1); 454 Src = Builder->CreateAnd(Src, One, "tmp"); 455 Value *Zero = Constant::getNullValue(Src->getType()); 456 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 457 } 458 459 return 0; 460} 461 462/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 463/// in order to eliminate the icmp. 464Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 465 bool DoXform) { 466 // If we are just checking for a icmp eq of a single bit and zext'ing it 467 // to an integer, then shift the bit to the appropriate place and then 468 // cast to integer to avoid the comparison. 469 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 470 const APInt &Op1CV = Op1C->getValue(); 471 472 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 473 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 474 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 475 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 476 if (!DoXform) return ICI; 477 478 Value *In = ICI->getOperand(0); 479 Value *Sh = ConstantInt::get(In->getType(), 480 In->getType()->getScalarSizeInBits()-1); 481 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 482 if (In->getType() != CI.getType()) 483 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 484 485 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 486 Constant *One = ConstantInt::get(In->getType(), 1); 487 In = Builder->CreateXor(In, One, In->getName()+".not"); 488 } 489 490 return ReplaceInstUsesWith(CI, In); 491 } 492 493 494 495 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 496 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 497 // zext (X == 1) to i32 --> X iff X has only the low bit set. 498 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 499 // zext (X != 0) to i32 --> X iff X has only the low bit set. 500 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 501 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 502 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 503 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 504 // This only works for EQ and NE 505 ICI->isEquality()) { 506 // If Op1C some other power of two, convert: 507 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 508 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 509 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 510 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 511 512 APInt KnownZeroMask(~KnownZero); 513 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 514 if (!DoXform) return ICI; 515 516 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 517 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 518 // (X&4) == 2 --> false 519 // (X&4) != 2 --> true 520 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 521 isNE); 522 Res = ConstantExpr::getZExt(Res, CI.getType()); 523 return ReplaceInstUsesWith(CI, Res); 524 } 525 526 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 527 Value *In = ICI->getOperand(0); 528 if (ShiftAmt) { 529 // Perform a logical shr by shiftamt. 530 // Insert the shift to put the result in the low bit. 531 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 532 In->getName()+".lobit"); 533 } 534 535 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 536 Constant *One = ConstantInt::get(In->getType(), 1); 537 In = Builder->CreateXor(In, One, "tmp"); 538 } 539 540 if (CI.getType() == In->getType()) 541 return ReplaceInstUsesWith(CI, In); 542 else 543 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 544 } 545 } 546 } 547 548 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 549 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 550 // may lead to additional simplifications. 551 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 552 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 553 uint32_t BitWidth = ITy->getBitWidth(); 554 Value *LHS = ICI->getOperand(0); 555 Value *RHS = ICI->getOperand(1); 556 557 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 558 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 559 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 560 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 561 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 562 563 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 564 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 565 APInt UnknownBit = ~KnownBits; 566 if (UnknownBit.countPopulation() == 1) { 567 if (!DoXform) return ICI; 568 569 Value *Result = Builder->CreateXor(LHS, RHS); 570 571 // Mask off any bits that are set and won't be shifted away. 572 if (KnownOneLHS.uge(UnknownBit)) 573 Result = Builder->CreateAnd(Result, 574 ConstantInt::get(ITy, UnknownBit)); 575 576 // Shift the bit we're testing down to the lsb. 577 Result = Builder->CreateLShr( 578 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 579 580 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 581 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 582 Result->takeName(ICI); 583 return ReplaceInstUsesWith(CI, Result); 584 } 585 } 586 } 587 } 588 589 return 0; 590} 591 592/// CanEvaluateZExtd - Determine if the specified value can be computed in the 593/// specified wider type and produce the same low bits. If not, return false. 594/// 595/// If this function returns true, it can also return a non-zero number of bits 596/// (in BitsToClear) which indicates that the value it computes is correct for 597/// the zero extend, but that the additional BitsToClear bits need to be zero'd 598/// out. For example, to promote something like: 599/// 600/// %B = trunc i64 %A to i32 601/// %C = lshr i32 %B, 8 602/// %E = zext i32 %C to i64 603/// 604/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 605/// set to 8 to indicate that the promoted value needs to have bits 24-31 606/// cleared in addition to bits 32-63. Since an 'and' will be generated to 607/// clear the top bits anyway, doing this has no extra cost. 608/// 609/// This function works on both vectors and scalars. 610static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) { 611 BitsToClear = 0; 612 if (isa<Constant>(V)) 613 return true; 614 615 Instruction *I = dyn_cast<Instruction>(V); 616 if (!I) return false; 617 618 // If the input is a truncate from the destination type, we can trivially 619 // eliminate it, even if it has multiple uses. 620 // FIXME: This is currently disabled until codegen can handle this without 621 // pessimizing code, PR5997. 622 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 623 return true; 624 625 // We can't extend or shrink something that has multiple uses: doing so would 626 // require duplicating the instruction in general, which isn't profitable. 627 if (!I->hasOneUse()) return false; 628 629 unsigned Opc = I->getOpcode(), Tmp; 630 switch (Opc) { 631 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 632 case Instruction::SExt: // zext(sext(x)) -> sext(x). 633 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 634 return true; 635 case Instruction::And: 636 case Instruction::Or: 637 case Instruction::Xor: 638 case Instruction::Add: 639 case Instruction::Sub: 640 case Instruction::Mul: 641 case Instruction::Shl: 642 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 643 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 644 return false; 645 // These can all be promoted if neither operand has 'bits to clear'. 646 if (BitsToClear == 0 && Tmp == 0) 647 return true; 648 649 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 650 // other side, BitsToClear is ok. 651 if (Tmp == 0 && 652 (Opc == Instruction::And || Opc == Instruction::Or || 653 Opc == Instruction::Xor)) { 654 // We use MaskedValueIsZero here for generality, but the case we care 655 // about the most is constant RHS. 656 unsigned VSize = V->getType()->getScalarSizeInBits(); 657 if (MaskedValueIsZero(I->getOperand(1), 658 APInt::getHighBitsSet(VSize, BitsToClear))) 659 return true; 660 } 661 662 // Otherwise, we don't know how to analyze this BitsToClear case yet. 663 return false; 664 665 case Instruction::LShr: 666 // We can promote lshr(x, cst) if we can promote x. This requires the 667 // ultimate 'and' to clear out the high zero bits we're clearing out though. 668 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 669 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 670 return false; 671 BitsToClear += Amt->getZExtValue(); 672 if (BitsToClear > V->getType()->getScalarSizeInBits()) 673 BitsToClear = V->getType()->getScalarSizeInBits(); 674 return true; 675 } 676 // Cannot promote variable LSHR. 677 return false; 678 case Instruction::Select: 679 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 680 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 681 // TODO: If important, we could handle the case when the BitsToClear are 682 // known zero in the disagreeing side. 683 Tmp != BitsToClear) 684 return false; 685 return true; 686 687 case Instruction::PHI: { 688 // We can change a phi if we can change all operands. Note that we never 689 // get into trouble with cyclic PHIs here because we only consider 690 // instructions with a single use. 691 PHINode *PN = cast<PHINode>(I); 692 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 693 return false; 694 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 695 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 696 // TODO: If important, we could handle the case when the BitsToClear 697 // are known zero in the disagreeing input. 698 Tmp != BitsToClear) 699 return false; 700 return true; 701 } 702 default: 703 // TODO: Can handle more cases here. 704 return false; 705 } 706} 707 708Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 709 // If this zero extend is only used by a truncate, let the truncate by 710 // eliminated before we try to optimize this zext. 711 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 712 return 0; 713 714 // If one of the common conversion will work, do it. 715 if (Instruction *Result = commonCastTransforms(CI)) 716 return Result; 717 718 // See if we can simplify any instructions used by the input whose sole 719 // purpose is to compute bits we don't care about. 720 if (SimplifyDemandedInstructionBits(CI)) 721 return &CI; 722 723 Value *Src = CI.getOperand(0); 724 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 725 726 // Attempt to extend the entire input expression tree to the destination 727 // type. Only do this if the dest type is a simple type, don't convert the 728 // expression tree to something weird like i93 unless the source is also 729 // strange. 730 unsigned BitsToClear; 731 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 732 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 733 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 734 "Unreasonable BitsToClear"); 735 736 // Okay, we can transform this! Insert the new expression now. 737 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 738 " to avoid zero extend: " << CI); 739 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 740 assert(Res->getType() == DestTy); 741 742 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 743 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 744 745 // If the high bits are already filled with zeros, just replace this 746 // cast with the result. 747 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 748 DestBitSize-SrcBitsKept))) 749 return ReplaceInstUsesWith(CI, Res); 750 751 // We need to emit an AND to clear the high bits. 752 Constant *C = ConstantInt::get(Res->getType(), 753 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 754 return BinaryOperator::CreateAnd(Res, C); 755 } 756 757 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 758 // types and if the sizes are just right we can convert this into a logical 759 // 'and' which will be much cheaper than the pair of casts. 760 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 761 // TODO: Subsume this into EvaluateInDifferentType. 762 763 // Get the sizes of the types involved. We know that the intermediate type 764 // will be smaller than A or C, but don't know the relation between A and C. 765 Value *A = CSrc->getOperand(0); 766 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 767 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 768 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 769 // If we're actually extending zero bits, then if 770 // SrcSize < DstSize: zext(a & mask) 771 // SrcSize == DstSize: a & mask 772 // SrcSize > DstSize: trunc(a) & mask 773 if (SrcSize < DstSize) { 774 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 775 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 776 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 777 return new ZExtInst(And, CI.getType()); 778 } 779 780 if (SrcSize == DstSize) { 781 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 782 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 783 AndValue)); 784 } 785 if (SrcSize > DstSize) { 786 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 787 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 788 return BinaryOperator::CreateAnd(Trunc, 789 ConstantInt::get(Trunc->getType(), 790 AndValue)); 791 } 792 } 793 794 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 795 return transformZExtICmp(ICI, CI); 796 797 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 798 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 799 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 800 // of the (zext icmp) will be transformed. 801 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 802 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 803 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 804 (transformZExtICmp(LHS, CI, false) || 805 transformZExtICmp(RHS, CI, false))) { 806 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 807 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 808 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 809 } 810 } 811 812 // zext(trunc(t) & C) -> (t & zext(C)). 813 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 814 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 815 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 816 Value *TI0 = TI->getOperand(0); 817 if (TI0->getType() == CI.getType()) 818 return 819 BinaryOperator::CreateAnd(TI0, 820 ConstantExpr::getZExt(C, CI.getType())); 821 } 822 823 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 824 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 825 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 826 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 827 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 828 And->getOperand(1) == C) 829 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 830 Value *TI0 = TI->getOperand(0); 831 if (TI0->getType() == CI.getType()) { 832 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 833 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 834 return BinaryOperator::CreateXor(NewAnd, ZC); 835 } 836 } 837 838 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 839 Value *X; 840 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 841 match(SrcI, m_Not(m_Value(X))) && 842 (!X->hasOneUse() || !isa<CmpInst>(X))) { 843 Value *New = Builder->CreateZExt(X, CI.getType()); 844 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 845 } 846 847 return 0; 848} 849 850/// CanEvaluateSExtd - Return true if we can take the specified value 851/// and return it as type Ty without inserting any new casts and without 852/// changing the value of the common low bits. This is used by code that tries 853/// to promote integer operations to a wider types will allow us to eliminate 854/// the extension. 855/// 856/// This function works on both vectors and scalars. 857/// 858static bool CanEvaluateSExtd(Value *V, const Type *Ty) { 859 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 860 "Can't sign extend type to a smaller type"); 861 // If this is a constant, it can be trivially promoted. 862 if (isa<Constant>(V)) 863 return true; 864 865 Instruction *I = dyn_cast<Instruction>(V); 866 if (!I) return false; 867 868 // If this is a truncate from the dest type, we can trivially eliminate it, 869 // even if it has multiple uses. 870 // FIXME: This is currently disabled until codegen can handle this without 871 // pessimizing code, PR5997. 872 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 873 return true; 874 875 // We can't extend or shrink something that has multiple uses: doing so would 876 // require duplicating the instruction in general, which isn't profitable. 877 if (!I->hasOneUse()) return false; 878 879 switch (I->getOpcode()) { 880 case Instruction::SExt: // sext(sext(x)) -> sext(x) 881 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 882 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 883 return true; 884 case Instruction::And: 885 case Instruction::Or: 886 case Instruction::Xor: 887 case Instruction::Add: 888 case Instruction::Sub: 889 case Instruction::Mul: 890 // These operators can all arbitrarily be extended if their inputs can. 891 return CanEvaluateSExtd(I->getOperand(0), Ty) && 892 CanEvaluateSExtd(I->getOperand(1), Ty); 893 894 //case Instruction::Shl: TODO 895 //case Instruction::LShr: TODO 896 897 case Instruction::Select: 898 return CanEvaluateSExtd(I->getOperand(1), Ty) && 899 CanEvaluateSExtd(I->getOperand(2), Ty); 900 901 case Instruction::PHI: { 902 // We can change a phi if we can change all operands. Note that we never 903 // get into trouble with cyclic PHIs here because we only consider 904 // instructions with a single use. 905 PHINode *PN = cast<PHINode>(I); 906 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 907 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 908 return true; 909 } 910 default: 911 // TODO: Can handle more cases here. 912 break; 913 } 914 915 return false; 916} 917 918Instruction *InstCombiner::visitSExt(SExtInst &CI) { 919 // If this sign extend is only used by a truncate, let the truncate by 920 // eliminated before we try to optimize this zext. 921 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 922 return 0; 923 924 if (Instruction *I = commonCastTransforms(CI)) 925 return I; 926 927 // See if we can simplify any instructions used by the input whose sole 928 // purpose is to compute bits we don't care about. 929 if (SimplifyDemandedInstructionBits(CI)) 930 return &CI; 931 932 Value *Src = CI.getOperand(0); 933 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 934 935 // Attempt to extend the entire input expression tree to the destination 936 // type. Only do this if the dest type is a simple type, don't convert the 937 // expression tree to something weird like i93 unless the source is also 938 // strange. 939 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 940 CanEvaluateSExtd(Src, DestTy)) { 941 // Okay, we can transform this! Insert the new expression now. 942 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 943 " to avoid sign extend: " << CI); 944 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 945 assert(Res->getType() == DestTy); 946 947 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 948 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 949 950 // If the high bits are already filled with sign bit, just replace this 951 // cast with the result. 952 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 953 return ReplaceInstUsesWith(CI, Res); 954 955 // We need to emit a shl + ashr to do the sign extend. 956 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 957 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 958 ShAmt); 959 } 960 961 // If this input is a trunc from our destination, then turn sext(trunc(x)) 962 // into shifts. 963 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 964 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 965 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 966 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 967 968 // We need to emit a shl + ashr to do the sign extend. 969 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 970 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 971 return BinaryOperator::CreateAShr(Res, ShAmt); 972 } 973 974 975 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed 976 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed 977 { 978 ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS; 979 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) { 980 // sext (x <s 0) to i32 --> x>>s31 true if signbit set. 981 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear. 982 if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) || 983 (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) { 984 Value *Sh = ConstantInt::get(CmpLHS->getType(), 985 CmpLHS->getType()->getScalarSizeInBits()-1); 986 Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit"); 987 if (In->getType() != CI.getType()) 988 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp"); 989 990 if (Pred == ICmpInst::ICMP_SGT) 991 In = Builder->CreateNot(In, In->getName()+".not"); 992 return ReplaceInstUsesWith(CI, In); 993 } 994 } 995 } 996 997 998 // If the input is a shl/ashr pair of a same constant, then this is a sign 999 // extension from a smaller value. If we could trust arbitrary bitwidth 1000 // integers, we could turn this into a truncate to the smaller bit and then 1001 // use a sext for the whole extension. Since we don't, look deeper and check 1002 // for a truncate. If the source and dest are the same type, eliminate the 1003 // trunc and extend and just do shifts. For example, turn: 1004 // %a = trunc i32 %i to i8 1005 // %b = shl i8 %a, 6 1006 // %c = ashr i8 %b, 6 1007 // %d = sext i8 %c to i32 1008 // into: 1009 // %a = shl i32 %i, 30 1010 // %d = ashr i32 %a, 30 1011 Value *A = 0; 1012 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1013 ConstantInt *BA = 0, *CA = 0; 1014 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1015 m_ConstantInt(CA))) && 1016 BA == CA && A->getType() == CI.getType()) { 1017 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1018 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1019 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1020 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1021 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1022 return BinaryOperator::CreateAShr(A, ShAmtV); 1023 } 1024 1025 return 0; 1026} 1027 1028 1029/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1030/// in the specified FP type without changing its value. 1031static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1032 bool losesInfo; 1033 APFloat F = CFP->getValueAPF(); 1034 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1035 if (!losesInfo) 1036 return ConstantFP::get(CFP->getContext(), F); 1037 return 0; 1038} 1039 1040/// LookThroughFPExtensions - If this is an fp extension instruction, look 1041/// through it until we get the source value. 1042static Value *LookThroughFPExtensions(Value *V) { 1043 if (Instruction *I = dyn_cast<Instruction>(V)) 1044 if (I->getOpcode() == Instruction::FPExt) 1045 return LookThroughFPExtensions(I->getOperand(0)); 1046 1047 // If this value is a constant, return the constant in the smallest FP type 1048 // that can accurately represent it. This allows us to turn 1049 // (float)((double)X+2.0) into x+2.0f. 1050 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1051 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1052 return V; // No constant folding of this. 1053 // See if the value can be truncated to float and then reextended. 1054 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1055 return V; 1056 if (CFP->getType()->isDoubleTy()) 1057 return V; // Won't shrink. 1058 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1059 return V; 1060 // Don't try to shrink to various long double types. 1061 } 1062 1063 return V; 1064} 1065 1066Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1067 if (Instruction *I = commonCastTransforms(CI)) 1068 return I; 1069 1070 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1071 // smaller than the destination type, we can eliminate the truncate by doing 1072 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1073 // as many builtins (sqrt, etc). 1074 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1075 if (OpI && OpI->hasOneUse()) { 1076 switch (OpI->getOpcode()) { 1077 default: break; 1078 case Instruction::FAdd: 1079 case Instruction::FSub: 1080 case Instruction::FMul: 1081 case Instruction::FDiv: 1082 case Instruction::FRem: 1083 const Type *SrcTy = OpI->getType(); 1084 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1085 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1086 if (LHSTrunc->getType() != SrcTy && 1087 RHSTrunc->getType() != SrcTy) { 1088 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1089 // If the source types were both smaller than the destination type of 1090 // the cast, do this xform. 1091 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1092 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1093 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1094 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1095 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1096 } 1097 } 1098 break; 1099 } 1100 } 1101 return 0; 1102} 1103 1104Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1105 return commonCastTransforms(CI); 1106} 1107 1108Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1109 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1110 if (OpI == 0) 1111 return commonCastTransforms(FI); 1112 1113 // fptoui(uitofp(X)) --> X 1114 // fptoui(sitofp(X)) --> X 1115 // This is safe if the intermediate type has enough bits in its mantissa to 1116 // accurately represent all values of X. For example, do not do this with 1117 // i64->float->i64. This is also safe for sitofp case, because any negative 1118 // 'X' value would cause an undefined result for the fptoui. 1119 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1120 OpI->getOperand(0)->getType() == FI.getType() && 1121 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1122 OpI->getType()->getFPMantissaWidth()) 1123 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1124 1125 return commonCastTransforms(FI); 1126} 1127 1128Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1129 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1130 if (OpI == 0) 1131 return commonCastTransforms(FI); 1132 1133 // fptosi(sitofp(X)) --> X 1134 // fptosi(uitofp(X)) --> X 1135 // This is safe if the intermediate type has enough bits in its mantissa to 1136 // accurately represent all values of X. For example, do not do this with 1137 // i64->float->i64. This is also safe for sitofp case, because any negative 1138 // 'X' value would cause an undefined result for the fptoui. 1139 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1140 OpI->getOperand(0)->getType() == FI.getType() && 1141 (int)FI.getType()->getScalarSizeInBits() <= 1142 OpI->getType()->getFPMantissaWidth()) 1143 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1144 1145 return commonCastTransforms(FI); 1146} 1147 1148Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1149 return commonCastTransforms(CI); 1150} 1151 1152Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1153 return commonCastTransforms(CI); 1154} 1155 1156Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1157 // If the source integer type is not the intptr_t type for this target, do a 1158 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1159 // cast to be exposed to other transforms. 1160 if (TD) { 1161 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1162 TD->getPointerSizeInBits()) { 1163 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1164 TD->getIntPtrType(CI.getContext()), "tmp"); 1165 return new IntToPtrInst(P, CI.getType()); 1166 } 1167 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1168 TD->getPointerSizeInBits()) { 1169 Value *P = Builder->CreateZExt(CI.getOperand(0), 1170 TD->getIntPtrType(CI.getContext()), "tmp"); 1171 return new IntToPtrInst(P, CI.getType()); 1172 } 1173 } 1174 1175 if (Instruction *I = commonCastTransforms(CI)) 1176 return I; 1177 1178 return 0; 1179} 1180 1181/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1182Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1183 Value *Src = CI.getOperand(0); 1184 1185 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1186 // If casting the result of a getelementptr instruction with no offset, turn 1187 // this into a cast of the original pointer! 1188 if (GEP->hasAllZeroIndices()) { 1189 // Changing the cast operand is usually not a good idea but it is safe 1190 // here because the pointer operand is being replaced with another 1191 // pointer operand so the opcode doesn't need to change. 1192 Worklist.Add(GEP); 1193 CI.setOperand(0, GEP->getOperand(0)); 1194 return &CI; 1195 } 1196 1197 // If the GEP has a single use, and the base pointer is a bitcast, and the 1198 // GEP computes a constant offset, see if we can convert these three 1199 // instructions into fewer. This typically happens with unions and other 1200 // non-type-safe code. 1201 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1202 GEP->hasAllConstantIndices()) { 1203 // We are guaranteed to get a constant from EmitGEPOffset. 1204 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1205 int64_t Offset = OffsetV->getSExtValue(); 1206 1207 // Get the base pointer input of the bitcast, and the type it points to. 1208 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1209 const Type *GEPIdxTy = 1210 cast<PointerType>(OrigBase->getType())->getElementType(); 1211 SmallVector<Value*, 8> NewIndices; 1212 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1213 // If we were able to index down into an element, create the GEP 1214 // and bitcast the result. This eliminates one bitcast, potentially 1215 // two. 1216 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1217 Builder->CreateInBoundsGEP(OrigBase, 1218 NewIndices.begin(), NewIndices.end()) : 1219 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 1220 NGEP->takeName(GEP); 1221 1222 if (isa<BitCastInst>(CI)) 1223 return new BitCastInst(NGEP, CI.getType()); 1224 assert(isa<PtrToIntInst>(CI)); 1225 return new PtrToIntInst(NGEP, CI.getType()); 1226 } 1227 } 1228 } 1229 1230 return commonCastTransforms(CI); 1231} 1232 1233Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1234 // If the destination integer type is not the intptr_t type for this target, 1235 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1236 // to be exposed to other transforms. 1237 if (TD) { 1238 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1239 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1240 TD->getIntPtrType(CI.getContext()), 1241 "tmp"); 1242 return new TruncInst(P, CI.getType()); 1243 } 1244 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1245 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1246 TD->getIntPtrType(CI.getContext()), 1247 "tmp"); 1248 return new ZExtInst(P, CI.getType()); 1249 } 1250 } 1251 1252 return commonPointerCastTransforms(CI); 1253} 1254 1255Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1256 // If the operands are integer typed then apply the integer transforms, 1257 // otherwise just apply the common ones. 1258 Value *Src = CI.getOperand(0); 1259 const Type *SrcTy = Src->getType(); 1260 const Type *DestTy = CI.getType(); 1261 1262 // Get rid of casts from one type to the same type. These are useless and can 1263 // be replaced by the operand. 1264 if (DestTy == Src->getType()) 1265 return ReplaceInstUsesWith(CI, Src); 1266 1267 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1268 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1269 const Type *DstElTy = DstPTy->getElementType(); 1270 const Type *SrcElTy = SrcPTy->getElementType(); 1271 1272 // If the address spaces don't match, don't eliminate the bitcast, which is 1273 // required for changing types. 1274 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1275 return 0; 1276 1277 // If we are casting a alloca to a pointer to a type of the same 1278 // size, rewrite the allocation instruction to allocate the "right" type. 1279 // There is no need to modify malloc calls because it is their bitcast that 1280 // needs to be cleaned up. 1281 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1282 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1283 return V; 1284 1285 // If the source and destination are pointers, and this cast is equivalent 1286 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1287 // This can enhance SROA and other transforms that want type-safe pointers. 1288 Constant *ZeroUInt = 1289 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1290 unsigned NumZeros = 0; 1291 while (SrcElTy != DstElTy && 1292 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1293 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1294 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1295 ++NumZeros; 1296 } 1297 1298 // If we found a path from the src to dest, create the getelementptr now. 1299 if (SrcElTy == DstElTy) { 1300 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1301 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1302 ((Instruction*)NULL)); 1303 } 1304 } 1305 1306 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1307 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1308 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1309 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1310 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1311 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1312 } 1313 } 1314 1315 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1316 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1317 Value *Elem = 1318 Builder->CreateExtractElement(Src, 1319 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1320 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1321 } 1322 } 1323 1324 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1325 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1326 // a bitconvert to a vector with the same # elts. 1327 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1328 cast<VectorType>(DestTy)->getNumElements() == 1329 SVI->getType()->getNumElements() && 1330 SVI->getType()->getNumElements() == 1331 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1332 BitCastInst *Tmp; 1333 // If either of the operands is a cast from CI.getType(), then 1334 // evaluating the shuffle in the casted destination's type will allow 1335 // us to eliminate at least one cast. 1336 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1337 Tmp->getOperand(0)->getType() == DestTy) || 1338 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1339 Tmp->getOperand(0)->getType() == DestTy)) { 1340 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1341 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1342 // Return a new shuffle vector. Use the same element ID's, as we 1343 // know the vector types match #elts. 1344 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1345 } 1346 } 1347 } 1348 1349 if (SrcTy->isPointerTy()) 1350 return commonPointerCastTransforms(CI); 1351 return commonCastTransforms(CI); 1352} 1353