InstCombineCasts.cpp revision 202878
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} 148 149 150 151/// 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/// ValueRequiresCast - Return true if the cast from "V to Ty" actually results 259/// in any code being generated. It does not require codegen if V is simple 260/// enough or if the cast can be folded into other casts. 261bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V, 262 const Type *Ty) { 263 if (V->getType() == Ty || isa<Constant>(V)) return false; 264 265 // If this is another cast that can be eliminated, it isn't codegen either. 266 if (const CastInst *CI = dyn_cast<CastInst>(V)) 267 if (isEliminableCastPair(CI, opcode, Ty, TD)) 268 return false; 269 return true; 270} 271 272 273/// @brief Implement the transforms common to all CastInst visitors. 274Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 275 Value *Src = CI.getOperand(0); 276 277 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 278 // eliminate it now. 279 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 280 if (Instruction::CastOps opc = 281 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 282 // The first cast (CSrc) is eliminable so we need to fix up or replace 283 // the second cast (CI). CSrc will then have a good chance of being dead. 284 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 285 } 286 } 287 288 // If we are casting a select then fold the cast into the select 289 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 290 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 291 return NV; 292 293 // If we are casting a PHI then fold the cast into the PHI 294 if (isa<PHINode>(Src)) { 295 // We don't do this if this would create a PHI node with an illegal type if 296 // it is currently legal. 297 if (!isa<IntegerType>(Src->getType()) || 298 !isa<IntegerType>(CI.getType()) || 299 ShouldChangeType(CI.getType(), Src->getType())) 300 if (Instruction *NV = FoldOpIntoPhi(CI)) 301 return NV; 302 } 303 304 return 0; 305} 306 307/// CanEvaluateTruncated - Return true if we can evaluate the specified 308/// expression tree as type Ty instead of its larger type, and arrive with the 309/// same value. This is used by code that tries to eliminate truncates. 310/// 311/// Ty will always be a type smaller than V. We should return true if trunc(V) 312/// can be computed by computing V in the smaller type. If V is an instruction, 313/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 314/// makes sense if x and y can be efficiently truncated. 315/// 316/// This function works on both vectors and scalars. 317/// 318static bool CanEvaluateTruncated(Value *V, const Type *Ty) { 319 // We can always evaluate constants in another type. 320 if (isa<Constant>(V)) 321 return true; 322 323 Instruction *I = dyn_cast<Instruction>(V); 324 if (!I) return false; 325 326 const Type *OrigTy = V->getType(); 327 328 // If this is an extension from the dest type, we can eliminate it, even if it 329 // has multiple uses. 330 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 331 I->getOperand(0)->getType() == Ty) 332 return true; 333 334 // We can't extend or shrink something that has multiple uses: doing so would 335 // require duplicating the instruction in general, which isn't profitable. 336 if (!I->hasOneUse()) return false; 337 338 unsigned Opc = I->getOpcode(); 339 switch (Opc) { 340 case Instruction::Add: 341 case Instruction::Sub: 342 case Instruction::Mul: 343 case Instruction::And: 344 case Instruction::Or: 345 case Instruction::Xor: 346 // These operators can all arbitrarily be extended or truncated. 347 return CanEvaluateTruncated(I->getOperand(0), Ty) && 348 CanEvaluateTruncated(I->getOperand(1), Ty); 349 350 case Instruction::UDiv: 351 case Instruction::URem: { 352 // UDiv and URem can be truncated if all the truncated bits are zero. 353 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 354 uint32_t BitWidth = Ty->getScalarSizeInBits(); 355 if (BitWidth < OrigBitWidth) { 356 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 357 if (MaskedValueIsZero(I->getOperand(0), Mask) && 358 MaskedValueIsZero(I->getOperand(1), Mask)) { 359 return CanEvaluateTruncated(I->getOperand(0), Ty) && 360 CanEvaluateTruncated(I->getOperand(1), Ty); 361 } 362 } 363 break; 364 } 365 case Instruction::Shl: 366 // If we are truncating the result of this SHL, and if it's a shift of a 367 // constant amount, we can always perform a SHL in a smaller type. 368 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 369 uint32_t BitWidth = Ty->getScalarSizeInBits(); 370 if (CI->getLimitedValue(BitWidth) < BitWidth) 371 return CanEvaluateTruncated(I->getOperand(0), Ty); 372 } 373 break; 374 case Instruction::LShr: 375 // If this is a truncate of a logical shr, we can truncate it to a smaller 376 // lshr iff we know that the bits we would otherwise be shifting in are 377 // already zeros. 378 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 379 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 380 uint32_t BitWidth = Ty->getScalarSizeInBits(); 381 if (MaskedValueIsZero(I->getOperand(0), 382 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 383 CI->getLimitedValue(BitWidth) < BitWidth) { 384 return CanEvaluateTruncated(I->getOperand(0), Ty); 385 } 386 } 387 break; 388 case Instruction::Trunc: 389 // trunc(trunc(x)) -> trunc(x) 390 return true; 391 case Instruction::Select: { 392 SelectInst *SI = cast<SelectInst>(I); 393 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 394 CanEvaluateTruncated(SI->getFalseValue(), Ty); 395 } 396 case Instruction::PHI: { 397 // We can change a phi if we can change all operands. Note that we never 398 // get into trouble with cyclic PHIs here because we only consider 399 // instructions with a single use. 400 PHINode *PN = cast<PHINode>(I); 401 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 402 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 403 return false; 404 return true; 405 } 406 default: 407 // TODO: Can handle more cases here. 408 break; 409 } 410 411 return false; 412} 413 414Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 415 if (Instruction *Result = commonCastTransforms(CI)) 416 return Result; 417 418 // See if we can simplify any instructions used by the input whose sole 419 // purpose is to compute bits we don't care about. 420 if (SimplifyDemandedInstructionBits(CI)) 421 return &CI; 422 423 Value *Src = CI.getOperand(0); 424 const Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 425 426 // Attempt to truncate the entire input expression tree to the destination 427 // type. Only do this if the dest type is a simple type, don't convert the 428 // expression tree to something weird like i93 unless the source is also 429 // strange. 430 if ((isa<VectorType>(DestTy) || ShouldChangeType(SrcTy, DestTy)) && 431 CanEvaluateTruncated(Src, DestTy)) { 432 433 // If this cast is a truncate, evaluting in a different type always 434 // eliminates the cast, so it is always a win. 435 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 436 " to avoid cast: " << CI); 437 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 438 assert(Res->getType() == DestTy); 439 return ReplaceInstUsesWith(CI, Res); 440 } 441 442 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 443 if (DestTy->getScalarSizeInBits() == 1) { 444 Constant *One = ConstantInt::get(Src->getType(), 1); 445 Src = Builder->CreateAnd(Src, One, "tmp"); 446 Value *Zero = Constant::getNullValue(Src->getType()); 447 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 448 } 449 450 return 0; 451} 452 453/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 454/// in order to eliminate the icmp. 455Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 456 bool DoXform) { 457 // If we are just checking for a icmp eq of a single bit and zext'ing it 458 // to an integer, then shift the bit to the appropriate place and then 459 // cast to integer to avoid the comparison. 460 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 461 const APInt &Op1CV = Op1C->getValue(); 462 463 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 464 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 465 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 466 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 467 if (!DoXform) return ICI; 468 469 Value *In = ICI->getOperand(0); 470 Value *Sh = ConstantInt::get(In->getType(), 471 In->getType()->getScalarSizeInBits()-1); 472 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 473 if (In->getType() != CI.getType()) 474 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 475 476 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 477 Constant *One = ConstantInt::get(In->getType(), 1); 478 In = Builder->CreateXor(In, One, In->getName()+".not"); 479 } 480 481 return ReplaceInstUsesWith(CI, In); 482 } 483 484 485 486 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 487 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 488 // zext (X == 1) to i32 --> X iff X has only the low bit set. 489 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 490 // zext (X != 0) to i32 --> X iff X has only the low bit set. 491 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 492 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 493 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 494 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 495 // This only works for EQ and NE 496 ICI->isEquality()) { 497 // If Op1C some other power of two, convert: 498 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 499 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 500 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 501 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 502 503 APInt KnownZeroMask(~KnownZero); 504 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 505 if (!DoXform) return ICI; 506 507 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 508 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 509 // (X&4) == 2 --> false 510 // (X&4) != 2 --> true 511 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 512 isNE); 513 Res = ConstantExpr::getZExt(Res, CI.getType()); 514 return ReplaceInstUsesWith(CI, Res); 515 } 516 517 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 518 Value *In = ICI->getOperand(0); 519 if (ShiftAmt) { 520 // Perform a logical shr by shiftamt. 521 // Insert the shift to put the result in the low bit. 522 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 523 In->getName()+".lobit"); 524 } 525 526 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 527 Constant *One = ConstantInt::get(In->getType(), 1); 528 In = Builder->CreateXor(In, One, "tmp"); 529 } 530 531 if (CI.getType() == In->getType()) 532 return ReplaceInstUsesWith(CI, In); 533 else 534 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 535 } 536 } 537 } 538 539 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 540 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 541 // may lead to additional simplifications. 542 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 543 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 544 uint32_t BitWidth = ITy->getBitWidth(); 545 Value *LHS = ICI->getOperand(0); 546 Value *RHS = ICI->getOperand(1); 547 548 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 549 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 550 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 551 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 552 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 553 554 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 555 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 556 APInt UnknownBit = ~KnownBits; 557 if (UnknownBit.countPopulation() == 1) { 558 if (!DoXform) return ICI; 559 560 Value *Result = Builder->CreateXor(LHS, RHS); 561 562 // Mask off any bits that are set and won't be shifted away. 563 if (KnownOneLHS.uge(UnknownBit)) 564 Result = Builder->CreateAnd(Result, 565 ConstantInt::get(ITy, UnknownBit)); 566 567 // Shift the bit we're testing down to the lsb. 568 Result = Builder->CreateLShr( 569 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 570 571 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 572 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 573 Result->takeName(ICI); 574 return ReplaceInstUsesWith(CI, Result); 575 } 576 } 577 } 578 } 579 580 return 0; 581} 582 583/// CanEvaluateZExtd - Determine if the specified value can be computed in the 584/// specified wider type and produce the same low bits. If not, return false. 585/// 586/// If this function returns true, it can also return a non-zero number of bits 587/// (in BitsToClear) which indicates that the value it computes is correct for 588/// the zero extend, but that the additional BitsToClear bits need to be zero'd 589/// out. For example, to promote something like: 590/// 591/// %B = trunc i64 %A to i32 592/// %C = lshr i32 %B, 8 593/// %E = zext i32 %C to i64 594/// 595/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 596/// set to 8 to indicate that the promoted value needs to have bits 24-31 597/// cleared in addition to bits 32-63. Since an 'and' will be generated to 598/// clear the top bits anyway, doing this has no extra cost. 599/// 600/// This function works on both vectors and scalars. 601static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) { 602 BitsToClear = 0; 603 if (isa<Constant>(V)) 604 return true; 605 606 Instruction *I = dyn_cast<Instruction>(V); 607 if (!I) return false; 608 609 // If the input is a truncate from the destination type, we can trivially 610 // eliminate it, even if it has multiple uses. 611 // FIXME: This is currently disabled until codegen can handle this without 612 // pessimizing code, PR5997. 613 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 614 return true; 615 616 // We can't extend or shrink something that has multiple uses: doing so would 617 // require duplicating the instruction in general, which isn't profitable. 618 if (!I->hasOneUse()) return false; 619 620 unsigned Opc = I->getOpcode(), Tmp; 621 switch (Opc) { 622 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 623 case Instruction::SExt: // zext(sext(x)) -> sext(x). 624 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 625 return true; 626 case Instruction::And: 627 case Instruction::Or: 628 case Instruction::Xor: 629 case Instruction::Add: 630 case Instruction::Sub: 631 case Instruction::Mul: 632 case Instruction::Shl: 633 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 634 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 635 return false; 636 // These can all be promoted if neither operand has 'bits to clear'. 637 if (BitsToClear == 0 && Tmp == 0) 638 return true; 639 640 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 641 // other side, BitsToClear is ok. 642 if (Tmp == 0 && 643 (Opc == Instruction::And || Opc == Instruction::Or || 644 Opc == Instruction::Xor)) { 645 // We use MaskedValueIsZero here for generality, but the case we care 646 // about the most is constant RHS. 647 unsigned VSize = V->getType()->getScalarSizeInBits(); 648 if (MaskedValueIsZero(I->getOperand(1), 649 APInt::getHighBitsSet(VSize, BitsToClear))) 650 return true; 651 } 652 653 // Otherwise, we don't know how to analyze this BitsToClear case yet. 654 return false; 655 656 case Instruction::LShr: 657 // We can promote lshr(x, cst) if we can promote x. This requires the 658 // ultimate 'and' to clear out the high zero bits we're clearing out though. 659 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 660 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 661 return false; 662 BitsToClear += Amt->getZExtValue(); 663 if (BitsToClear > V->getType()->getScalarSizeInBits()) 664 BitsToClear = V->getType()->getScalarSizeInBits(); 665 return true; 666 } 667 // Cannot promote variable LSHR. 668 return false; 669 case Instruction::Select: 670 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 671 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 672 // TODO: If important, we could handle the case when the BitsToClear are 673 // known zero in the disagreeing side. 674 Tmp != BitsToClear) 675 return false; 676 return true; 677 678 case Instruction::PHI: { 679 // We can change a phi if we can change all operands. Note that we never 680 // get into trouble with cyclic PHIs here because we only consider 681 // instructions with a single use. 682 PHINode *PN = cast<PHINode>(I); 683 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 684 return false; 685 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 686 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 687 // TODO: If important, we could handle the case when the BitsToClear 688 // are known zero in the disagreeing input. 689 Tmp != BitsToClear) 690 return false; 691 return true; 692 } 693 default: 694 // TODO: Can handle more cases here. 695 return false; 696 } 697} 698 699Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 700 // If this zero extend is only used by a truncate, let the truncate by 701 // eliminated before we try to optimize this zext. 702 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 703 return 0; 704 705 // If one of the common conversion will work, do it. 706 if (Instruction *Result = commonCastTransforms(CI)) 707 return Result; 708 709 // See if we can simplify any instructions used by the input whose sole 710 // purpose is to compute bits we don't care about. 711 if (SimplifyDemandedInstructionBits(CI)) 712 return &CI; 713 714 Value *Src = CI.getOperand(0); 715 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 716 717 // Attempt to extend the entire input expression tree to the destination 718 // type. Only do this if the dest type is a simple type, don't convert the 719 // expression tree to something weird like i93 unless the source is also 720 // strange. 721 unsigned BitsToClear; 722 if ((isa<VectorType>(DestTy) || ShouldChangeType(SrcTy, DestTy)) && 723 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 724 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 725 "Unreasonable BitsToClear"); 726 727 // Okay, we can transform this! Insert the new expression now. 728 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 729 " to avoid zero extend: " << CI); 730 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 731 assert(Res->getType() == DestTy); 732 733 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 734 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 735 736 // If the high bits are already filled with zeros, just replace this 737 // cast with the result. 738 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 739 DestBitSize-SrcBitsKept))) 740 return ReplaceInstUsesWith(CI, Res); 741 742 // We need to emit an AND to clear the high bits. 743 Constant *C = ConstantInt::get(Res->getType(), 744 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 745 return BinaryOperator::CreateAnd(Res, C); 746 } 747 748 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 749 // types and if the sizes are just right we can convert this into a logical 750 // 'and' which will be much cheaper than the pair of casts. 751 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 752 // TODO: Subsume this into EvaluateInDifferentType. 753 754 // Get the sizes of the types involved. We know that the intermediate type 755 // will be smaller than A or C, but don't know the relation between A and C. 756 Value *A = CSrc->getOperand(0); 757 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 758 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 759 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 760 // If we're actually extending zero bits, then if 761 // SrcSize < DstSize: zext(a & mask) 762 // SrcSize == DstSize: a & mask 763 // SrcSize > DstSize: trunc(a) & mask 764 if (SrcSize < DstSize) { 765 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 766 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 767 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 768 return new ZExtInst(And, CI.getType()); 769 } 770 771 if (SrcSize == DstSize) { 772 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 773 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 774 AndValue)); 775 } 776 if (SrcSize > DstSize) { 777 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 778 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 779 return BinaryOperator::CreateAnd(Trunc, 780 ConstantInt::get(Trunc->getType(), 781 AndValue)); 782 } 783 } 784 785 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 786 return transformZExtICmp(ICI, CI); 787 788 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 789 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 790 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 791 // of the (zext icmp) will be transformed. 792 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 793 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 794 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 795 (transformZExtICmp(LHS, CI, false) || 796 transformZExtICmp(RHS, CI, false))) { 797 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 798 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 799 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 800 } 801 } 802 803 // zext(trunc(t) & C) -> (t & zext(C)). 804 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 805 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 806 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 807 Value *TI0 = TI->getOperand(0); 808 if (TI0->getType() == CI.getType()) 809 return 810 BinaryOperator::CreateAnd(TI0, 811 ConstantExpr::getZExt(C, CI.getType())); 812 } 813 814 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 815 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 816 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 817 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 818 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 819 And->getOperand(1) == C) 820 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 821 Value *TI0 = TI->getOperand(0); 822 if (TI0->getType() == CI.getType()) { 823 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 824 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 825 return BinaryOperator::CreateXor(NewAnd, ZC); 826 } 827 } 828 829 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 830 Value *X; 831 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isInteger(1) && 832 match(SrcI, m_Not(m_Value(X))) && 833 (!X->hasOneUse() || !isa<CmpInst>(X))) { 834 Value *New = Builder->CreateZExt(X, CI.getType()); 835 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 836 } 837 838 return 0; 839} 840 841/// CanEvaluateSExtd - Return true if we can take the specified value 842/// and return it as type Ty without inserting any new casts and without 843/// changing the value of the common low bits. This is used by code that tries 844/// to promote integer operations to a wider types will allow us to eliminate 845/// the extension. 846/// 847/// This function works on both vectors and scalars. 848/// 849static bool CanEvaluateSExtd(Value *V, const Type *Ty) { 850 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 851 "Can't sign extend type to a smaller type"); 852 // If this is a constant, it can be trivially promoted. 853 if (isa<Constant>(V)) 854 return true; 855 856 Instruction *I = dyn_cast<Instruction>(V); 857 if (!I) return false; 858 859 // If this is a truncate from the dest type, we can trivially eliminate it, 860 // even if it has multiple uses. 861 // FIXME: This is currently disabled until codegen can handle this without 862 // pessimizing code, PR5997. 863 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 864 return true; 865 866 // We can't extend or shrink something that has multiple uses: doing so would 867 // require duplicating the instruction in general, which isn't profitable. 868 if (!I->hasOneUse()) return false; 869 870 switch (I->getOpcode()) { 871 case Instruction::SExt: // sext(sext(x)) -> sext(x) 872 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 873 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 874 return true; 875 case Instruction::And: 876 case Instruction::Or: 877 case Instruction::Xor: 878 case Instruction::Add: 879 case Instruction::Sub: 880 case Instruction::Mul: 881 // These operators can all arbitrarily be extended if their inputs can. 882 return CanEvaluateSExtd(I->getOperand(0), Ty) && 883 CanEvaluateSExtd(I->getOperand(1), Ty); 884 885 //case Instruction::Shl: TODO 886 //case Instruction::LShr: TODO 887 888 case Instruction::Select: 889 return CanEvaluateSExtd(I->getOperand(1), Ty) && 890 CanEvaluateSExtd(I->getOperand(2), Ty); 891 892 case Instruction::PHI: { 893 // We can change a phi if we can change all operands. Note that we never 894 // get into trouble with cyclic PHIs here because we only consider 895 // instructions with a single use. 896 PHINode *PN = cast<PHINode>(I); 897 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 898 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 899 return true; 900 } 901 default: 902 // TODO: Can handle more cases here. 903 break; 904 } 905 906 return false; 907} 908 909Instruction *InstCombiner::visitSExt(SExtInst &CI) { 910 // If this sign extend is only used by a truncate, let the truncate by 911 // eliminated before we try to optimize this zext. 912 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 913 return 0; 914 915 if (Instruction *I = commonCastTransforms(CI)) 916 return I; 917 918 // See if we can simplify any instructions used by the input whose sole 919 // purpose is to compute bits we don't care about. 920 if (SimplifyDemandedInstructionBits(CI)) 921 return &CI; 922 923 Value *Src = CI.getOperand(0); 924 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 925 926 // Canonicalize sign-extend from i1 to a select. 927 if (Src->getType()->isInteger(1)) 928 return SelectInst::Create(Src, 929 Constant::getAllOnesValue(CI.getType()), 930 Constant::getNullValue(CI.getType())); 931 932 // Attempt to extend the entire input expression tree to the destination 933 // type. Only do this if the dest type is a simple type, don't convert the 934 // expression tree to something weird like i93 unless the source is also 935 // strange. 936 if ((isa<VectorType>(DestTy) || ShouldChangeType(SrcTy, DestTy)) && 937 CanEvaluateSExtd(Src, DestTy)) { 938 // Okay, we can transform this! Insert the new expression now. 939 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 940 " to avoid sign extend: " << CI); 941 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 942 assert(Res->getType() == DestTy); 943 944 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 945 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 946 947 // If the high bits are already filled with sign bit, just replace this 948 // cast with the result. 949 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 950 return ReplaceInstUsesWith(CI, Res); 951 952 // We need to emit a shl + ashr to do the sign extend. 953 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 954 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 955 ShAmt); 956 } 957 958 // If this input is a trunc from our destination, then turn sext(trunc(x)) 959 // into shifts. 960 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 961 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 962 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 963 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 964 965 // We need to emit a shl + ashr to do the sign extend. 966 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 967 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 968 return BinaryOperator::CreateAShr(Res, ShAmt); 969 } 970 971 // If the input is a shl/ashr pair of a same constant, then this is a sign 972 // extension from a smaller value. If we could trust arbitrary bitwidth 973 // integers, we could turn this into a truncate to the smaller bit and then 974 // use a sext for the whole extension. Since we don't, look deeper and check 975 // for a truncate. If the source and dest are the same type, eliminate the 976 // trunc and extend and just do shifts. For example, turn: 977 // %a = trunc i32 %i to i8 978 // %b = shl i8 %a, 6 979 // %c = ashr i8 %b, 6 980 // %d = sext i8 %c to i32 981 // into: 982 // %a = shl i32 %i, 30 983 // %d = ashr i32 %a, 30 984 Value *A = 0; 985 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 986 ConstantInt *BA = 0, *CA = 0; 987 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 988 m_ConstantInt(CA))) && 989 BA == CA && A->getType() == CI.getType()) { 990 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 991 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 992 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 993 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 994 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 995 return BinaryOperator::CreateAShr(A, ShAmtV); 996 } 997 998 return 0; 999} 1000 1001 1002/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1003/// in the specified FP type without changing its value. 1004static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1005 bool losesInfo; 1006 APFloat F = CFP->getValueAPF(); 1007 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1008 if (!losesInfo) 1009 return ConstantFP::get(CFP->getContext(), F); 1010 return 0; 1011} 1012 1013/// LookThroughFPExtensions - If this is an fp extension instruction, look 1014/// through it until we get the source value. 1015static Value *LookThroughFPExtensions(Value *V) { 1016 if (Instruction *I = dyn_cast<Instruction>(V)) 1017 if (I->getOpcode() == Instruction::FPExt) 1018 return LookThroughFPExtensions(I->getOperand(0)); 1019 1020 // If this value is a constant, return the constant in the smallest FP type 1021 // that can accurately represent it. This allows us to turn 1022 // (float)((double)X+2.0) into x+2.0f. 1023 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1024 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1025 return V; // No constant folding of this. 1026 // See if the value can be truncated to float and then reextended. 1027 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1028 return V; 1029 if (CFP->getType()->isDoubleTy()) 1030 return V; // Won't shrink. 1031 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1032 return V; 1033 // Don't try to shrink to various long double types. 1034 } 1035 1036 return V; 1037} 1038 1039Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1040 if (Instruction *I = commonCastTransforms(CI)) 1041 return I; 1042 1043 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1044 // smaller than the destination type, we can eliminate the truncate by doing 1045 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1046 // as many builtins (sqrt, etc). 1047 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1048 if (OpI && OpI->hasOneUse()) { 1049 switch (OpI->getOpcode()) { 1050 default: break; 1051 case Instruction::FAdd: 1052 case Instruction::FSub: 1053 case Instruction::FMul: 1054 case Instruction::FDiv: 1055 case Instruction::FRem: 1056 const Type *SrcTy = OpI->getType(); 1057 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1058 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1059 if (LHSTrunc->getType() != SrcTy && 1060 RHSTrunc->getType() != SrcTy) { 1061 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1062 // If the source types were both smaller than the destination type of 1063 // the cast, do this xform. 1064 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1065 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1066 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1067 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1068 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1069 } 1070 } 1071 break; 1072 } 1073 } 1074 return 0; 1075} 1076 1077Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1078 return commonCastTransforms(CI); 1079} 1080 1081Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1082 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1083 if (OpI == 0) 1084 return commonCastTransforms(FI); 1085 1086 // fptoui(uitofp(X)) --> X 1087 // fptoui(sitofp(X)) --> X 1088 // This is safe if the intermediate type has enough bits in its mantissa to 1089 // accurately represent all values of X. For example, do not do this with 1090 // i64->float->i64. This is also safe for sitofp case, because any negative 1091 // 'X' value would cause an undefined result for the fptoui. 1092 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1093 OpI->getOperand(0)->getType() == FI.getType() && 1094 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1095 OpI->getType()->getFPMantissaWidth()) 1096 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1097 1098 return commonCastTransforms(FI); 1099} 1100 1101Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1102 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1103 if (OpI == 0) 1104 return commonCastTransforms(FI); 1105 1106 // fptosi(sitofp(X)) --> X 1107 // fptosi(uitofp(X)) --> X 1108 // This is safe if the intermediate type has enough bits in its mantissa to 1109 // accurately represent all values of X. For example, do not do this with 1110 // i64->float->i64. This is also safe for sitofp case, because any negative 1111 // 'X' value would cause an undefined result for the fptoui. 1112 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1113 OpI->getOperand(0)->getType() == FI.getType() && 1114 (int)FI.getType()->getScalarSizeInBits() <= 1115 OpI->getType()->getFPMantissaWidth()) 1116 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1117 1118 return commonCastTransforms(FI); 1119} 1120 1121Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1122 return commonCastTransforms(CI); 1123} 1124 1125Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1126 return commonCastTransforms(CI); 1127} 1128 1129Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1130 // If the source integer type is larger than the intptr_t type for 1131 // this target, do a trunc to the intptr_t type, then inttoptr of it. This 1132 // allows the trunc to be exposed to other transforms. Don't do this for 1133 // extending inttoptr's, because we don't know if the target sign or zero 1134 // extends to pointers. 1135 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() > 1136 TD->getPointerSizeInBits()) { 1137 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1138 TD->getIntPtrType(CI.getContext()), "tmp"); 1139 return new IntToPtrInst(P, CI.getType()); 1140 } 1141 1142 if (Instruction *I = commonCastTransforms(CI)) 1143 return I; 1144 1145 return 0; 1146} 1147 1148/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1149Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1150 Value *Src = CI.getOperand(0); 1151 1152 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1153 // If casting the result of a getelementptr instruction with no offset, turn 1154 // this into a cast of the original pointer! 1155 if (GEP->hasAllZeroIndices()) { 1156 // Changing the cast operand is usually not a good idea but it is safe 1157 // here because the pointer operand is being replaced with another 1158 // pointer operand so the opcode doesn't need to change. 1159 Worklist.Add(GEP); 1160 CI.setOperand(0, GEP->getOperand(0)); 1161 return &CI; 1162 } 1163 1164 // If the GEP has a single use, and the base pointer is a bitcast, and the 1165 // GEP computes a constant offset, see if we can convert these three 1166 // instructions into fewer. This typically happens with unions and other 1167 // non-type-safe code. 1168 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1169 GEP->hasAllConstantIndices()) { 1170 // We are guaranteed to get a constant from EmitGEPOffset. 1171 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1172 int64_t Offset = OffsetV->getSExtValue(); 1173 1174 // Get the base pointer input of the bitcast, and the type it points to. 1175 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1176 const Type *GEPIdxTy = 1177 cast<PointerType>(OrigBase->getType())->getElementType(); 1178 SmallVector<Value*, 8> NewIndices; 1179 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1180 // If we were able to index down into an element, create the GEP 1181 // and bitcast the result. This eliminates one bitcast, potentially 1182 // two. 1183 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1184 Builder->CreateInBoundsGEP(OrigBase, 1185 NewIndices.begin(), NewIndices.end()) : 1186 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 1187 NGEP->takeName(GEP); 1188 1189 if (isa<BitCastInst>(CI)) 1190 return new BitCastInst(NGEP, CI.getType()); 1191 assert(isa<PtrToIntInst>(CI)); 1192 return new PtrToIntInst(NGEP, CI.getType()); 1193 } 1194 } 1195 } 1196 1197 return commonCastTransforms(CI); 1198} 1199 1200Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1201 // If the destination integer type is smaller than the intptr_t type for 1202 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the 1203 // trunc to be exposed to other transforms. Don't do this for extending 1204 // ptrtoint's, because we don't know if the target sign or zero extends its 1205 // pointers. 1206 if (TD && 1207 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1208 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1209 TD->getIntPtrType(CI.getContext()), 1210 "tmp"); 1211 return new TruncInst(P, CI.getType()); 1212 } 1213 1214 return commonPointerCastTransforms(CI); 1215} 1216 1217Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1218 // If the operands are integer typed then apply the integer transforms, 1219 // otherwise just apply the common ones. 1220 Value *Src = CI.getOperand(0); 1221 const Type *SrcTy = Src->getType(); 1222 const Type *DestTy = CI.getType(); 1223 1224 // Get rid of casts from one type to the same type. These are useless and can 1225 // be replaced by the operand. 1226 if (DestTy == Src->getType()) 1227 return ReplaceInstUsesWith(CI, Src); 1228 1229 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1230 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1231 const Type *DstElTy = DstPTy->getElementType(); 1232 const Type *SrcElTy = SrcPTy->getElementType(); 1233 1234 // If the address spaces don't match, don't eliminate the bitcast, which is 1235 // required for changing types. 1236 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1237 return 0; 1238 1239 // If we are casting a alloca to a pointer to a type of the same 1240 // size, rewrite the allocation instruction to allocate the "right" type. 1241 // There is no need to modify malloc calls because it is their bitcast that 1242 // needs to be cleaned up. 1243 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1244 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1245 return V; 1246 1247 // If the source and destination are pointers, and this cast is equivalent 1248 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1249 // This can enhance SROA and other transforms that want type-safe pointers. 1250 Constant *ZeroUInt = 1251 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1252 unsigned NumZeros = 0; 1253 while (SrcElTy != DstElTy && 1254 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) && 1255 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1256 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1257 ++NumZeros; 1258 } 1259 1260 // If we found a path from the src to dest, create the getelementptr now. 1261 if (SrcElTy == DstElTy) { 1262 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1263 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1264 ((Instruction*)NULL)); 1265 } 1266 } 1267 1268 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1269 if (DestVTy->getNumElements() == 1 && !isa<VectorType>(SrcTy)) { 1270 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1271 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1272 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1273 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1274 } 1275 } 1276 1277 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1278 if (SrcVTy->getNumElements() == 1 && !isa<VectorType>(DestTy)) { 1279 Value *Elem = 1280 Builder->CreateExtractElement(Src, 1281 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1282 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1283 } 1284 } 1285 1286 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1287 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1288 // a bitconvert to a vector with the same # elts. 1289 if (SVI->hasOneUse() && isa<VectorType>(DestTy) && 1290 cast<VectorType>(DestTy)->getNumElements() == 1291 SVI->getType()->getNumElements() && 1292 SVI->getType()->getNumElements() == 1293 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1294 BitCastInst *Tmp; 1295 // If either of the operands is a cast from CI.getType(), then 1296 // evaluating the shuffle in the casted destination's type will allow 1297 // us to eliminate at least one cast. 1298 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1299 Tmp->getOperand(0)->getType() == DestTy) || 1300 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1301 Tmp->getOperand(0)->getType() == DestTy)) { 1302 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1303 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1304 // Return a new shuffle vector. Use the same element ID's, as we 1305 // know the vector types match #elts. 1306 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1307 } 1308 } 1309 } 1310 1311 if (isa<PointerType>(SrcTy)) 1312 return commonPointerCastTransforms(CI); 1313 return commonCastTransforms(CI); 1314} 1315