InstCombineCasts.cpp revision 26dbe7ec18740f642febcc738e628d921aafd079
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 uint64_t &Offset) { 26 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 27 Offset = CI->getZExtValue(); 28 Scale = 0; 29 return ConstantInt::get(Val->getType(), 0); 30 } 31 32 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) { 33 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 34 if (I->getOpcode() == Instruction::Shl) { 35 // This is a value scaled by '1 << the shift amt'. 36 Scale = UINT64_C(1) << RHS->getZExtValue(); 37 Offset = 0; 38 return I->getOperand(0); 39 } 40 41 if (I->getOpcode() == Instruction::Mul) { 42 // This value is scaled by 'RHS'. 43 Scale = RHS->getZExtValue(); 44 Offset = 0; 45 return I->getOperand(0); 46 } 47 48 if (I->getOpcode() == Instruction::Add) { 49 // We have X+C. Check to see if we really have (X*C2)+C1, 50 // where C1 is divisible by C2. 51 unsigned SubScale; 52 Value *SubVal = 53 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); 54 Offset += RHS->getZExtValue(); 55 Scale = SubScale; 56 return SubVal; 57 } 58 } 59 } 60 61 // Otherwise, we can't look past this. 62 Scale = 1; 63 Offset = 0; 64 return Val; 65} 66 67/// PromoteCastOfAllocation - If we find a cast of an allocation instruction, 68/// try to eliminate the cast by moving the type information into the alloc. 69Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 70 AllocaInst &AI) { 71 // This requires TargetData to get the alloca alignment and size information. 72 if (!TD) return 0; 73 74 const PointerType *PTy = cast<PointerType>(CI.getType()); 75 76 BuilderTy AllocaBuilder(*Builder); 77 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); 78 79 // Get the type really allocated and the type casted to. 80 const Type *AllocElTy = AI.getAllocatedType(); 81 const Type *CastElTy = PTy->getElementType(); 82 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0; 83 84 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy); 85 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy); 86 if (CastElTyAlign < AllocElTyAlign) return 0; 87 88 // If the allocation has multiple uses, only promote it if we are strictly 89 // increasing the alignment of the resultant allocation. If we keep it the 90 // same, we open the door to infinite loops of various kinds. (A reference 91 // from a dbg.declare doesn't count as a use for this purpose.) 92 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) && 93 CastElTyAlign == AllocElTyAlign) return 0; 94 95 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy); 96 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); 97 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 98 99 // See if we can satisfy the modulus by pulling a scale out of the array 100 // size argument. 101 unsigned ArraySizeScale; 102 uint64_t ArrayOffset; 103 Value *NumElements = // See if the array size is a decomposable linear expr. 104 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 105 106 // If we can now satisfy the modulus, by using a non-1 scale, we really can 107 // do the xform. 108 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 109 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 110 111 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 112 Value *Amt = 0; 113 if (Scale == 1) { 114 Amt = NumElements; 115 } else { 116 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 117 // Insert before the alloca, not before the cast. 118 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp"); 119 } 120 121 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 122 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 123 Offset, true); 124 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp"); 125 } 126 127 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 128 New->setAlignment(AI.getAlignment()); 129 New->takeName(&AI); 130 131 // If the allocation has one real use plus a dbg.declare, just remove the 132 // declare. 133 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) { 134 EraseInstFromFunction(*(Instruction*)DI); 135 } 136 // If the allocation has multiple real uses, insert a cast and change all 137 // things that used it to use the new cast. This will also hack on CI, but it 138 // will die soon. 139 else if (!AI.hasOneUse()) { 140 // New is the allocation instruction, pointer typed. AI is the original 141 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 142 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 143 AI.replaceAllUsesWith(NewCast); 144 } 145 return ReplaceInstUsesWith(CI, New); 146} 147 148 149 150/// EvaluateInDifferentType - Given an expression that 151/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 152/// insert the code to evaluate the expression. 153Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty, 154 bool isSigned) { 155 if (Constant *C = dyn_cast<Constant>(V)) { 156 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 157 // If we got a constantexpr back, try to simplify it with TD info. 158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 159 C = ConstantFoldConstantExpression(CE, TD); 160 return C; 161 } 162 163 // Otherwise, it must be an instruction. 164 Instruction *I = cast<Instruction>(V); 165 Instruction *Res = 0; 166 unsigned Opc = I->getOpcode(); 167 switch (Opc) { 168 case Instruction::Add: 169 case Instruction::Sub: 170 case Instruction::Mul: 171 case Instruction::And: 172 case Instruction::Or: 173 case Instruction::Xor: 174 case Instruction::AShr: 175 case Instruction::LShr: 176 case Instruction::Shl: 177 case Instruction::UDiv: 178 case Instruction::URem: { 179 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 180 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 181 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 182 break; 183 } 184 case Instruction::Trunc: 185 case Instruction::ZExt: 186 case Instruction::SExt: 187 // If the source type of the cast is the type we're trying for then we can 188 // just return the source. There's no need to insert it because it is not 189 // new. 190 if (I->getOperand(0)->getType() == Ty) 191 return I->getOperand(0); 192 193 // Otherwise, must be the same type of cast, so just reinsert a new one. 194 // This also handles the case of zext(trunc(x)) -> zext(x). 195 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 196 Opc == Instruction::SExt); 197 break; 198 case Instruction::Select: { 199 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 200 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 201 Res = SelectInst::Create(I->getOperand(0), True, False); 202 break; 203 } 204 case Instruction::PHI: { 205 PHINode *OPN = cast<PHINode>(I); 206 PHINode *NPN = PHINode::Create(Ty); 207 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 208 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 209 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 210 } 211 Res = NPN; 212 break; 213 } 214 default: 215 // TODO: Can handle more cases here. 216 llvm_unreachable("Unreachable!"); 217 break; 218 } 219 220 Res->takeName(I); 221 return InsertNewInstBefore(Res, *I); 222} 223 224 225/// This function is a wrapper around CastInst::isEliminableCastPair. It 226/// simply extracts arguments and returns what that function returns. 227static Instruction::CastOps 228isEliminableCastPair( 229 const CastInst *CI, ///< The first cast instruction 230 unsigned opcode, ///< The opcode of the second cast instruction 231 const Type *DstTy, ///< The target type for the second cast instruction 232 TargetData *TD ///< The target data for pointer size 233) { 234 235 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above 236 const Type *MidTy = CI->getType(); // B from above 237 238 // Get the opcodes of the two Cast instructions 239 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 240 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 241 242 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 243 DstTy, 244 TD ? TD->getIntPtrType(CI->getContext()) : 0); 245 246 // We don't want to form an inttoptr or ptrtoint that converts to an integer 247 // type that differs from the pointer size. 248 if ((Res == Instruction::IntToPtr && 249 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 250 (Res == Instruction::PtrToInt && 251 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 252 Res = 0; 253 254 return Instruction::CastOps(Res); 255} 256 257/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 258/// results in any code being generated and is interesting to optimize out. If 259/// the cast can be eliminated by some other simple transformation, we prefer 260/// to do the simplification first. 261bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 262 const Type *Ty) { 263 // Noop casts and casts of constants should be eliminated trivially. 264 if (V->getType() == Ty || isa<Constant>(V)) return false; 265 266 // If this is another cast that can be eliminated, we prefer to have it 267 // eliminated. 268 if (const CastInst *CI = dyn_cast<CastInst>(V)) 269 if (isEliminableCastPair(CI, opc, Ty, TD)) 270 return false; 271 272 // If this is a vector sext from a compare, then we don't want to break the 273 // idiom where each element of the extended vector is either zero or all ones. 274 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 275 return false; 276 277 return true; 278} 279 280 281/// @brief Implement the transforms common to all CastInst visitors. 282Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 283 Value *Src = CI.getOperand(0); 284 285 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 286 // eliminate it now. 287 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 288 if (Instruction::CastOps opc = 289 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 290 // The first cast (CSrc) is eliminable so we need to fix up or replace 291 // the second cast (CI). CSrc will then have a good chance of being dead. 292 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 293 } 294 } 295 296 // If we are casting a select then fold the cast into the select 297 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 298 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 299 return NV; 300 301 // If we are casting a PHI then fold the cast into the PHI 302 if (isa<PHINode>(Src)) { 303 // We don't do this if this would create a PHI node with an illegal type if 304 // it is currently legal. 305 if (!Src->getType()->isIntegerTy() || 306 !CI.getType()->isIntegerTy() || 307 ShouldChangeType(CI.getType(), Src->getType())) 308 if (Instruction *NV = FoldOpIntoPhi(CI)) 309 return NV; 310 } 311 312 return 0; 313} 314 315/// CanEvaluateTruncated - Return true if we can evaluate the specified 316/// expression tree as type Ty instead of its larger type, and arrive with the 317/// same value. This is used by code that tries to eliminate truncates. 318/// 319/// Ty will always be a type smaller than V. We should return true if trunc(V) 320/// can be computed by computing V in the smaller type. If V is an instruction, 321/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 322/// makes sense if x and y can be efficiently truncated. 323/// 324/// This function works on both vectors and scalars. 325/// 326static bool CanEvaluateTruncated(Value *V, const Type *Ty) { 327 // We can always evaluate constants in another type. 328 if (isa<Constant>(V)) 329 return true; 330 331 Instruction *I = dyn_cast<Instruction>(V); 332 if (!I) return false; 333 334 const Type *OrigTy = V->getType(); 335 336 // If this is an extension from the dest type, we can eliminate it, even if it 337 // has multiple uses. 338 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 339 I->getOperand(0)->getType() == Ty) 340 return true; 341 342 // We can't extend or shrink something that has multiple uses: doing so would 343 // require duplicating the instruction in general, which isn't profitable. 344 if (!I->hasOneUse()) return false; 345 346 unsigned Opc = I->getOpcode(); 347 switch (Opc) { 348 case Instruction::Add: 349 case Instruction::Sub: 350 case Instruction::Mul: 351 case Instruction::And: 352 case Instruction::Or: 353 case Instruction::Xor: 354 // These operators can all arbitrarily be extended or truncated. 355 return CanEvaluateTruncated(I->getOperand(0), Ty) && 356 CanEvaluateTruncated(I->getOperand(1), Ty); 357 358 case Instruction::UDiv: 359 case Instruction::URem: { 360 // UDiv and URem can be truncated if all the truncated bits are zero. 361 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 362 uint32_t BitWidth = Ty->getScalarSizeInBits(); 363 if (BitWidth < OrigBitWidth) { 364 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 365 if (MaskedValueIsZero(I->getOperand(0), Mask) && 366 MaskedValueIsZero(I->getOperand(1), Mask)) { 367 return CanEvaluateTruncated(I->getOperand(0), Ty) && 368 CanEvaluateTruncated(I->getOperand(1), Ty); 369 } 370 } 371 break; 372 } 373 case Instruction::Shl: 374 // If we are truncating the result of this SHL, and if it's a shift of a 375 // constant amount, we can always perform a SHL in a smaller type. 376 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 377 uint32_t BitWidth = Ty->getScalarSizeInBits(); 378 if (CI->getLimitedValue(BitWidth) < BitWidth) 379 return CanEvaluateTruncated(I->getOperand(0), Ty); 380 } 381 break; 382 case Instruction::LShr: 383 // If this is a truncate of a logical shr, we can truncate it to a smaller 384 // lshr iff we know that the bits we would otherwise be shifting in are 385 // already zeros. 386 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 387 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 388 uint32_t BitWidth = Ty->getScalarSizeInBits(); 389 if (MaskedValueIsZero(I->getOperand(0), 390 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 391 CI->getLimitedValue(BitWidth) < BitWidth) { 392 return CanEvaluateTruncated(I->getOperand(0), Ty); 393 } 394 } 395 break; 396 case Instruction::Trunc: 397 // trunc(trunc(x)) -> trunc(x) 398 return true; 399 case Instruction::Select: { 400 SelectInst *SI = cast<SelectInst>(I); 401 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 402 CanEvaluateTruncated(SI->getFalseValue(), Ty); 403 } 404 case Instruction::PHI: { 405 // We can change a phi if we can change all operands. Note that we never 406 // get into trouble with cyclic PHIs here because we only consider 407 // instructions with a single use. 408 PHINode *PN = cast<PHINode>(I); 409 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 410 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 411 return false; 412 return true; 413 } 414 default: 415 // TODO: Can handle more cases here. 416 break; 417 } 418 419 return false; 420} 421 422Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 423 if (Instruction *Result = commonCastTransforms(CI)) 424 return Result; 425 426 // See if we can simplify any instructions used by the input whose sole 427 // purpose is to compute bits we don't care about. 428 if (SimplifyDemandedInstructionBits(CI)) 429 return &CI; 430 431 Value *Src = CI.getOperand(0); 432 const Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 433 434 // Attempt to truncate the entire input expression tree to the destination 435 // type. Only do this if the dest type is a simple type, don't convert the 436 // expression tree to something weird like i93 unless the source is also 437 // strange. 438 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 439 CanEvaluateTruncated(Src, DestTy)) { 440 441 // If this cast is a truncate, evaluting in a different type always 442 // eliminates the cast, so it is always a win. 443 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 444 " to avoid cast: " << CI << '\n'); 445 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 446 assert(Res->getType() == DestTy); 447 return ReplaceInstUsesWith(CI, Res); 448 } 449 450 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 451 if (DestTy->getScalarSizeInBits() == 1) { 452 Constant *One = ConstantInt::get(Src->getType(), 1); 453 Src = Builder->CreateAnd(Src, One, "tmp"); 454 Value *Zero = Constant::getNullValue(Src->getType()); 455 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 456 } 457 458 return 0; 459} 460 461/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 462/// in order to eliminate the icmp. 463Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 464 bool DoXform) { 465 // If we are just checking for a icmp eq of a single bit and zext'ing it 466 // to an integer, then shift the bit to the appropriate place and then 467 // cast to integer to avoid the comparison. 468 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 469 const APInt &Op1CV = Op1C->getValue(); 470 471 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 472 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 473 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 474 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 475 if (!DoXform) return ICI; 476 477 Value *In = ICI->getOperand(0); 478 Value *Sh = ConstantInt::get(In->getType(), 479 In->getType()->getScalarSizeInBits()-1); 480 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 481 if (In->getType() != CI.getType()) 482 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 483 484 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 485 Constant *One = ConstantInt::get(In->getType(), 1); 486 In = Builder->CreateXor(In, One, In->getName()+".not"); 487 } 488 489 return ReplaceInstUsesWith(CI, In); 490 } 491 492 493 494 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 495 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 496 // zext (X == 1) to i32 --> X iff X has only the low bit set. 497 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 498 // zext (X != 0) to i32 --> X iff X has only the low bit set. 499 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 500 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 501 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 502 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 503 // This only works for EQ and NE 504 ICI->isEquality()) { 505 // If Op1C some other power of two, convert: 506 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 507 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 508 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 509 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 510 511 APInt KnownZeroMask(~KnownZero); 512 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 513 if (!DoXform) return ICI; 514 515 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 516 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 517 // (X&4) == 2 --> false 518 // (X&4) != 2 --> true 519 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 520 isNE); 521 Res = ConstantExpr::getZExt(Res, CI.getType()); 522 return ReplaceInstUsesWith(CI, Res); 523 } 524 525 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 526 Value *In = ICI->getOperand(0); 527 if (ShiftAmt) { 528 // Perform a logical shr by shiftamt. 529 // Insert the shift to put the result in the low bit. 530 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 531 In->getName()+".lobit"); 532 } 533 534 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 535 Constant *One = ConstantInt::get(In->getType(), 1); 536 In = Builder->CreateXor(In, One, "tmp"); 537 } 538 539 if (CI.getType() == In->getType()) 540 return ReplaceInstUsesWith(CI, In); 541 else 542 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 543 } 544 } 545 } 546 547 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 548 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 549 // may lead to additional simplifications. 550 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 551 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 552 uint32_t BitWidth = ITy->getBitWidth(); 553 Value *LHS = ICI->getOperand(0); 554 Value *RHS = ICI->getOperand(1); 555 556 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 557 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 558 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 559 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 560 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 561 562 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 563 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 564 APInt UnknownBit = ~KnownBits; 565 if (UnknownBit.countPopulation() == 1) { 566 if (!DoXform) return ICI; 567 568 Value *Result = Builder->CreateXor(LHS, RHS); 569 570 // Mask off any bits that are set and won't be shifted away. 571 if (KnownOneLHS.uge(UnknownBit)) 572 Result = Builder->CreateAnd(Result, 573 ConstantInt::get(ITy, UnknownBit)); 574 575 // Shift the bit we're testing down to the lsb. 576 Result = Builder->CreateLShr( 577 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 578 579 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 580 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 581 Result->takeName(ICI); 582 return ReplaceInstUsesWith(CI, Result); 583 } 584 } 585 } 586 } 587 588 return 0; 589} 590 591/// CanEvaluateZExtd - Determine if the specified value can be computed in the 592/// specified wider type and produce the same low bits. If not, return false. 593/// 594/// If this function returns true, it can also return a non-zero number of bits 595/// (in BitsToClear) which indicates that the value it computes is correct for 596/// the zero extend, but that the additional BitsToClear bits need to be zero'd 597/// out. For example, to promote something like: 598/// 599/// %B = trunc i64 %A to i32 600/// %C = lshr i32 %B, 8 601/// %E = zext i32 %C to i64 602/// 603/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 604/// set to 8 to indicate that the promoted value needs to have bits 24-31 605/// cleared in addition to bits 32-63. Since an 'and' will be generated to 606/// clear the top bits anyway, doing this has no extra cost. 607/// 608/// This function works on both vectors and scalars. 609static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) { 610 BitsToClear = 0; 611 if (isa<Constant>(V)) 612 return true; 613 614 Instruction *I = dyn_cast<Instruction>(V); 615 if (!I) return false; 616 617 // If the input is a truncate from the destination type, we can trivially 618 // eliminate it, even if it has multiple uses. 619 // FIXME: This is currently disabled until codegen can handle this without 620 // pessimizing code, PR5997. 621 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 622 return true; 623 624 // We can't extend or shrink something that has multiple uses: doing so would 625 // require duplicating the instruction in general, which isn't profitable. 626 if (!I->hasOneUse()) return false; 627 628 unsigned Opc = I->getOpcode(), Tmp; 629 switch (Opc) { 630 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 631 case Instruction::SExt: // zext(sext(x)) -> sext(x). 632 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 633 return true; 634 case Instruction::And: 635 case Instruction::Or: 636 case Instruction::Xor: 637 case Instruction::Add: 638 case Instruction::Sub: 639 case Instruction::Mul: 640 case Instruction::Shl: 641 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 642 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 643 return false; 644 // These can all be promoted if neither operand has 'bits to clear'. 645 if (BitsToClear == 0 && Tmp == 0) 646 return true; 647 648 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 649 // other side, BitsToClear is ok. 650 if (Tmp == 0 && 651 (Opc == Instruction::And || Opc == Instruction::Or || 652 Opc == Instruction::Xor)) { 653 // We use MaskedValueIsZero here for generality, but the case we care 654 // about the most is constant RHS. 655 unsigned VSize = V->getType()->getScalarSizeInBits(); 656 if (MaskedValueIsZero(I->getOperand(1), 657 APInt::getHighBitsSet(VSize, BitsToClear))) 658 return true; 659 } 660 661 // Otherwise, we don't know how to analyze this BitsToClear case yet. 662 return false; 663 664 case Instruction::LShr: 665 // We can promote lshr(x, cst) if we can promote x. This requires the 666 // ultimate 'and' to clear out the high zero bits we're clearing out though. 667 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 668 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 669 return false; 670 BitsToClear += Amt->getZExtValue(); 671 if (BitsToClear > V->getType()->getScalarSizeInBits()) 672 BitsToClear = V->getType()->getScalarSizeInBits(); 673 return true; 674 } 675 // Cannot promote variable LSHR. 676 return false; 677 case Instruction::Select: 678 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 679 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 680 // TODO: If important, we could handle the case when the BitsToClear are 681 // known zero in the disagreeing side. 682 Tmp != BitsToClear) 683 return false; 684 return true; 685 686 case Instruction::PHI: { 687 // We can change a phi if we can change all operands. Note that we never 688 // get into trouble with cyclic PHIs here because we only consider 689 // instructions with a single use. 690 PHINode *PN = cast<PHINode>(I); 691 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 692 return false; 693 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 694 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 695 // TODO: If important, we could handle the case when the BitsToClear 696 // are known zero in the disagreeing input. 697 Tmp != BitsToClear) 698 return false; 699 return true; 700 } 701 default: 702 // TODO: Can handle more cases here. 703 return false; 704 } 705} 706 707Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 708 // If this zero extend is only used by a truncate, let the truncate by 709 // eliminated before we try to optimize this zext. 710 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 711 return 0; 712 713 // If one of the common conversion will work, do it. 714 if (Instruction *Result = commonCastTransforms(CI)) 715 return Result; 716 717 // See if we can simplify any instructions used by the input whose sole 718 // purpose is to compute bits we don't care about. 719 if (SimplifyDemandedInstructionBits(CI)) 720 return &CI; 721 722 Value *Src = CI.getOperand(0); 723 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 724 725 // Attempt to extend the entire input expression tree to the destination 726 // type. Only do this if the dest type is a simple type, don't convert the 727 // expression tree to something weird like i93 unless the source is also 728 // strange. 729 unsigned BitsToClear; 730 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 731 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 732 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 733 "Unreasonable BitsToClear"); 734 735 // Okay, we can transform this! Insert the new expression now. 736 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 737 " to avoid zero extend: " << CI); 738 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 739 assert(Res->getType() == DestTy); 740 741 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 742 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 743 744 // If the high bits are already filled with zeros, just replace this 745 // cast with the result. 746 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 747 DestBitSize-SrcBitsKept))) 748 return ReplaceInstUsesWith(CI, Res); 749 750 // We need to emit an AND to clear the high bits. 751 Constant *C = ConstantInt::get(Res->getType(), 752 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 753 return BinaryOperator::CreateAnd(Res, C); 754 } 755 756 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 757 // types and if the sizes are just right we can convert this into a logical 758 // 'and' which will be much cheaper than the pair of casts. 759 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 760 // TODO: Subsume this into EvaluateInDifferentType. 761 762 // Get the sizes of the types involved. We know that the intermediate type 763 // will be smaller than A or C, but don't know the relation between A and C. 764 Value *A = CSrc->getOperand(0); 765 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 766 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 767 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 768 // If we're actually extending zero bits, then if 769 // SrcSize < DstSize: zext(a & mask) 770 // SrcSize == DstSize: a & mask 771 // SrcSize > DstSize: trunc(a) & mask 772 if (SrcSize < DstSize) { 773 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 774 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 775 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 776 return new ZExtInst(And, CI.getType()); 777 } 778 779 if (SrcSize == DstSize) { 780 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 781 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 782 AndValue)); 783 } 784 if (SrcSize > DstSize) { 785 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 786 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 787 return BinaryOperator::CreateAnd(Trunc, 788 ConstantInt::get(Trunc->getType(), 789 AndValue)); 790 } 791 } 792 793 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 794 return transformZExtICmp(ICI, CI); 795 796 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 797 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 798 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 799 // of the (zext icmp) will be transformed. 800 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 801 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 802 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 803 (transformZExtICmp(LHS, CI, false) || 804 transformZExtICmp(RHS, CI, false))) { 805 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 806 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 807 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 808 } 809 } 810 811 // zext(trunc(t) & C) -> (t & zext(C)). 812 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 813 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 814 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 815 Value *TI0 = TI->getOperand(0); 816 if (TI0->getType() == CI.getType()) 817 return 818 BinaryOperator::CreateAnd(TI0, 819 ConstantExpr::getZExt(C, CI.getType())); 820 } 821 822 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 823 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 824 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 825 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 826 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 827 And->getOperand(1) == C) 828 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 829 Value *TI0 = TI->getOperand(0); 830 if (TI0->getType() == CI.getType()) { 831 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 832 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 833 return BinaryOperator::CreateXor(NewAnd, ZC); 834 } 835 } 836 837 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 838 Value *X; 839 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 840 match(SrcI, m_Not(m_Value(X))) && 841 (!X->hasOneUse() || !isa<CmpInst>(X))) { 842 Value *New = Builder->CreateZExt(X, CI.getType()); 843 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 844 } 845 846 return 0; 847} 848 849/// CanEvaluateSExtd - Return true if we can take the specified value 850/// and return it as type Ty without inserting any new casts and without 851/// changing the value of the common low bits. This is used by code that tries 852/// to promote integer operations to a wider types will allow us to eliminate 853/// the extension. 854/// 855/// This function works on both vectors and scalars. 856/// 857static bool CanEvaluateSExtd(Value *V, const Type *Ty) { 858 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 859 "Can't sign extend type to a smaller type"); 860 // If this is a constant, it can be trivially promoted. 861 if (isa<Constant>(V)) 862 return true; 863 864 Instruction *I = dyn_cast<Instruction>(V); 865 if (!I) return false; 866 867 // If this is a truncate from the dest type, we can trivially eliminate it, 868 // even if it has multiple uses. 869 // FIXME: This is currently disabled until codegen can handle this without 870 // pessimizing code, PR5997. 871 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 872 return true; 873 874 // We can't extend or shrink something that has multiple uses: doing so would 875 // require duplicating the instruction in general, which isn't profitable. 876 if (!I->hasOneUse()) return false; 877 878 switch (I->getOpcode()) { 879 case Instruction::SExt: // sext(sext(x)) -> sext(x) 880 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 881 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 882 return true; 883 case Instruction::And: 884 case Instruction::Or: 885 case Instruction::Xor: 886 case Instruction::Add: 887 case Instruction::Sub: 888 case Instruction::Mul: 889 // These operators can all arbitrarily be extended if their inputs can. 890 return CanEvaluateSExtd(I->getOperand(0), Ty) && 891 CanEvaluateSExtd(I->getOperand(1), Ty); 892 893 //case Instruction::Shl: TODO 894 //case Instruction::LShr: TODO 895 896 case Instruction::Select: 897 return CanEvaluateSExtd(I->getOperand(1), Ty) && 898 CanEvaluateSExtd(I->getOperand(2), Ty); 899 900 case Instruction::PHI: { 901 // We can change a phi if we can change all operands. Note that we never 902 // get into trouble with cyclic PHIs here because we only consider 903 // instructions with a single use. 904 PHINode *PN = cast<PHINode>(I); 905 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 906 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 907 return true; 908 } 909 default: 910 // TODO: Can handle more cases here. 911 break; 912 } 913 914 return false; 915} 916 917Instruction *InstCombiner::visitSExt(SExtInst &CI) { 918 // If this sign extend is only used by a truncate, let the truncate by 919 // eliminated before we try to optimize this zext. 920 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 921 return 0; 922 923 if (Instruction *I = commonCastTransforms(CI)) 924 return I; 925 926 // See if we can simplify any instructions used by the input whose sole 927 // purpose is to compute bits we don't care about. 928 if (SimplifyDemandedInstructionBits(CI)) 929 return &CI; 930 931 Value *Src = CI.getOperand(0); 932 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 933 934 // Attempt to extend the entire input expression tree to the destination 935 // type. Only do this if the dest type is a simple type, don't convert the 936 // expression tree to something weird like i93 unless the source is also 937 // strange. 938 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 939 CanEvaluateSExtd(Src, DestTy)) { 940 // Okay, we can transform this! Insert the new expression now. 941 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 942 " to avoid sign extend: " << CI); 943 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 944 assert(Res->getType() == DestTy); 945 946 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 947 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 948 949 // If the high bits are already filled with sign bit, just replace this 950 // cast with the result. 951 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 952 return ReplaceInstUsesWith(CI, Res); 953 954 // We need to emit a shl + ashr to do the sign extend. 955 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 956 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 957 ShAmt); 958 } 959 960 // If this input is a trunc from our destination, then turn sext(trunc(x)) 961 // into shifts. 962 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 963 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 964 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 965 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 966 967 // We need to emit a shl + ashr to do the sign extend. 968 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 969 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 970 return BinaryOperator::CreateAShr(Res, ShAmt); 971 } 972 973 974 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed 975 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed 976 { 977 ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS; 978 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) { 979 // sext (x <s 0) to i32 --> x>>s31 true if signbit set. 980 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear. 981 if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) || 982 (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) { 983 Value *Sh = ConstantInt::get(CmpLHS->getType(), 984 CmpLHS->getType()->getScalarSizeInBits()-1); 985 Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit"); 986 if (In->getType() != CI.getType()) 987 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp"); 988 989 if (Pred == ICmpInst::ICMP_SGT) 990 In = Builder->CreateNot(In, In->getName()+".not"); 991 return ReplaceInstUsesWith(CI, In); 992 } 993 } 994 } 995 996 997 // If the input is a shl/ashr pair of a same constant, then this is a sign 998 // extension from a smaller value. If we could trust arbitrary bitwidth 999 // integers, we could turn this into a truncate to the smaller bit and then 1000 // use a sext for the whole extension. Since we don't, look deeper and check 1001 // for a truncate. If the source and dest are the same type, eliminate the 1002 // trunc and extend and just do shifts. For example, turn: 1003 // %a = trunc i32 %i to i8 1004 // %b = shl i8 %a, 6 1005 // %c = ashr i8 %b, 6 1006 // %d = sext i8 %c to i32 1007 // into: 1008 // %a = shl i32 %i, 30 1009 // %d = ashr i32 %a, 30 1010 Value *A = 0; 1011 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1012 ConstantInt *BA = 0, *CA = 0; 1013 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1014 m_ConstantInt(CA))) && 1015 BA == CA && A->getType() == CI.getType()) { 1016 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1017 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1018 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1019 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1020 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1021 return BinaryOperator::CreateAShr(A, ShAmtV); 1022 } 1023 1024 return 0; 1025} 1026 1027 1028/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1029/// in the specified FP type without changing its value. 1030static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1031 bool losesInfo; 1032 APFloat F = CFP->getValueAPF(); 1033 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1034 if (!losesInfo) 1035 return ConstantFP::get(CFP->getContext(), F); 1036 return 0; 1037} 1038 1039/// LookThroughFPExtensions - If this is an fp extension instruction, look 1040/// through it until we get the source value. 1041static Value *LookThroughFPExtensions(Value *V) { 1042 if (Instruction *I = dyn_cast<Instruction>(V)) 1043 if (I->getOpcode() == Instruction::FPExt) 1044 return LookThroughFPExtensions(I->getOperand(0)); 1045 1046 // If this value is a constant, return the constant in the smallest FP type 1047 // that can accurately represent it. This allows us to turn 1048 // (float)((double)X+2.0) into x+2.0f. 1049 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1050 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1051 return V; // No constant folding of this. 1052 // See if the value can be truncated to float and then reextended. 1053 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1054 return V; 1055 if (CFP->getType()->isDoubleTy()) 1056 return V; // Won't shrink. 1057 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1058 return V; 1059 // Don't try to shrink to various long double types. 1060 } 1061 1062 return V; 1063} 1064 1065Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1066 if (Instruction *I = commonCastTransforms(CI)) 1067 return I; 1068 1069 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1070 // smaller than the destination type, we can eliminate the truncate by doing 1071 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1072 // as many builtins (sqrt, etc). 1073 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1074 if (OpI && OpI->hasOneUse()) { 1075 switch (OpI->getOpcode()) { 1076 default: break; 1077 case Instruction::FAdd: 1078 case Instruction::FSub: 1079 case Instruction::FMul: 1080 case Instruction::FDiv: 1081 case Instruction::FRem: 1082 const Type *SrcTy = OpI->getType(); 1083 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1084 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1085 if (LHSTrunc->getType() != SrcTy && 1086 RHSTrunc->getType() != SrcTy) { 1087 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1088 // If the source types were both smaller than the destination type of 1089 // the cast, do this xform. 1090 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1091 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1092 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1093 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1094 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1095 } 1096 } 1097 break; 1098 } 1099 } 1100 1101 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1102 // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it. 1103 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1104 if (Call && Call->getCalledFunction() && 1105 Call->getCalledFunction()->getName() == "sqrt" && 1106 Call->getNumArgOperands() == 1) { 1107 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1108 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1109 CI.getType()->isFloatTy() && 1110 Call->getType()->isDoubleTy() && 1111 Arg->getType()->isDoubleTy() && 1112 Arg->getOperand(0)->getType()->isFloatTy()) { 1113 Function *Callee = Call->getCalledFunction(); 1114 Module *M = CI.getParent()->getParent()->getParent(); 1115 Constant* SqrtfFunc = M->getOrInsertFunction("sqrtf", 1116 Callee->getAttributes(), 1117 Builder->getFloatTy(), 1118 Builder->getFloatTy(), 1119 NULL); 1120 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1121 "sqrtfcall"); 1122 ret->setAttributes(Callee->getAttributes()); 1123 return ret; 1124 } 1125 } 1126 1127 return 0; 1128} 1129 1130Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1131 return commonCastTransforms(CI); 1132} 1133 1134Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1135 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1136 if (OpI == 0) 1137 return commonCastTransforms(FI); 1138 1139 // fptoui(uitofp(X)) --> X 1140 // fptoui(sitofp(X)) --> X 1141 // This is safe if the intermediate type has enough bits in its mantissa to 1142 // accurately represent all values of X. For example, do not do this with 1143 // i64->float->i64. This is also safe for sitofp case, because any negative 1144 // 'X' value would cause an undefined result for the fptoui. 1145 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1146 OpI->getOperand(0)->getType() == FI.getType() && 1147 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1148 OpI->getType()->getFPMantissaWidth()) 1149 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1150 1151 return commonCastTransforms(FI); 1152} 1153 1154Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1155 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1156 if (OpI == 0) 1157 return commonCastTransforms(FI); 1158 1159 // fptosi(sitofp(X)) --> X 1160 // fptosi(uitofp(X)) --> X 1161 // This is safe if the intermediate type has enough bits in its mantissa to 1162 // accurately represent all values of X. For example, do not do this with 1163 // i64->float->i64. This is also safe for sitofp case, because any negative 1164 // 'X' value would cause an undefined result for the fptoui. 1165 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1166 OpI->getOperand(0)->getType() == FI.getType() && 1167 (int)FI.getType()->getScalarSizeInBits() <= 1168 OpI->getType()->getFPMantissaWidth()) 1169 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1170 1171 return commonCastTransforms(FI); 1172} 1173 1174Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1175 return commonCastTransforms(CI); 1176} 1177 1178Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1179 return commonCastTransforms(CI); 1180} 1181 1182Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1183 // If the source integer type is not the intptr_t type for this target, do a 1184 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1185 // cast to be exposed to other transforms. 1186 if (TD) { 1187 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1188 TD->getPointerSizeInBits()) { 1189 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1190 TD->getIntPtrType(CI.getContext()), "tmp"); 1191 return new IntToPtrInst(P, CI.getType()); 1192 } 1193 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1194 TD->getPointerSizeInBits()) { 1195 Value *P = Builder->CreateZExt(CI.getOperand(0), 1196 TD->getIntPtrType(CI.getContext()), "tmp"); 1197 return new IntToPtrInst(P, CI.getType()); 1198 } 1199 } 1200 1201 if (Instruction *I = commonCastTransforms(CI)) 1202 return I; 1203 1204 return 0; 1205} 1206 1207/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1208Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1209 Value *Src = CI.getOperand(0); 1210 1211 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1212 // If casting the result of a getelementptr instruction with no offset, turn 1213 // this into a cast of the original pointer! 1214 if (GEP->hasAllZeroIndices()) { 1215 // Changing the cast operand is usually not a good idea but it is safe 1216 // here because the pointer operand is being replaced with another 1217 // pointer operand so the opcode doesn't need to change. 1218 Worklist.Add(GEP); 1219 CI.setOperand(0, GEP->getOperand(0)); 1220 return &CI; 1221 } 1222 1223 // If the GEP has a single use, and the base pointer is a bitcast, and the 1224 // GEP computes a constant offset, see if we can convert these three 1225 // instructions into fewer. This typically happens with unions and other 1226 // non-type-safe code. 1227 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1228 GEP->hasAllConstantIndices()) { 1229 // We are guaranteed to get a constant from EmitGEPOffset. 1230 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1231 int64_t Offset = OffsetV->getSExtValue(); 1232 1233 // Get the base pointer input of the bitcast, and the type it points to. 1234 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1235 const Type *GEPIdxTy = 1236 cast<PointerType>(OrigBase->getType())->getElementType(); 1237 SmallVector<Value*, 8> NewIndices; 1238 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1239 // If we were able to index down into an element, create the GEP 1240 // and bitcast the result. This eliminates one bitcast, potentially 1241 // two. 1242 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1243 Builder->CreateInBoundsGEP(OrigBase, 1244 NewIndices.begin(), NewIndices.end()) : 1245 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 1246 NGEP->takeName(GEP); 1247 1248 if (isa<BitCastInst>(CI)) 1249 return new BitCastInst(NGEP, CI.getType()); 1250 assert(isa<PtrToIntInst>(CI)); 1251 return new PtrToIntInst(NGEP, CI.getType()); 1252 } 1253 } 1254 } 1255 1256 return commonCastTransforms(CI); 1257} 1258 1259Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1260 // If the destination integer type is not the intptr_t type for this target, 1261 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1262 // to be exposed to other transforms. 1263 if (TD) { 1264 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1265 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1266 TD->getIntPtrType(CI.getContext()), 1267 "tmp"); 1268 return new TruncInst(P, CI.getType()); 1269 } 1270 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1271 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1272 TD->getIntPtrType(CI.getContext()), 1273 "tmp"); 1274 return new ZExtInst(P, CI.getType()); 1275 } 1276 } 1277 1278 return commonPointerCastTransforms(CI); 1279} 1280 1281/// OptimizeVectorResize - This input value (which is known to have vector type) 1282/// is being zero extended or truncated to the specified vector type. Try to 1283/// replace it with a shuffle (and vector/vector bitcast) if possible. 1284/// 1285/// The source and destination vector types may have different element types. 1286static Instruction *OptimizeVectorResize(Value *InVal, const VectorType *DestTy, 1287 InstCombiner &IC) { 1288 // We can only do this optimization if the output is a multiple of the input 1289 // element size, or the input is a multiple of the output element size. 1290 // Convert the input type to have the same element type as the output. 1291 const VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1292 1293 if (SrcTy->getElementType() != DestTy->getElementType()) { 1294 // The input types don't need to be identical, but for now they must be the 1295 // same size. There is no specific reason we couldn't handle things like 1296 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1297 // there yet. 1298 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1299 DestTy->getElementType()->getPrimitiveSizeInBits()) 1300 return 0; 1301 1302 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1303 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1304 } 1305 1306 // Now that the element types match, get the shuffle mask and RHS of the 1307 // shuffle to use, which depends on whether we're increasing or decreasing the 1308 // size of the input. 1309 SmallVector<Constant*, 16> ShuffleMask; 1310 Value *V2; 1311 const IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext()); 1312 1313 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1314 // If we're shrinking the number of elements, just shuffle in the low 1315 // elements from the input and use undef as the second shuffle input. 1316 V2 = UndefValue::get(SrcTy); 1317 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1318 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1319 1320 } else { 1321 // If we're increasing the number of elements, shuffle in all of the 1322 // elements from InVal and fill the rest of the result elements with zeros 1323 // from a constant zero. 1324 V2 = Constant::getNullValue(SrcTy); 1325 unsigned SrcElts = SrcTy->getNumElements(); 1326 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1327 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1328 1329 // The excess elements reference the first element of the zero input. 1330 ShuffleMask.append(DestTy->getNumElements()-SrcElts, 1331 ConstantInt::get(Int32Ty, SrcElts)); 1332 } 1333 1334 Constant *Mask = ConstantVector::get(ShuffleMask.data(), ShuffleMask.size()); 1335 return new ShuffleVectorInst(InVal, V2, Mask); 1336} 1337 1338/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1339/// bitcast. The various long double bitcasts can't get in here. 1340static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1341 Value *Src = CI.getOperand(0); 1342 const Type *DestTy = CI.getType(); 1343 1344 // If this is a bitcast from int to float, check to see if the int is an 1345 // extraction from a vector. 1346 Value *VecInput = 0; 1347 // bitcast(trunc(bitcast(somevector))) 1348 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1349 isa<VectorType>(VecInput->getType())) { 1350 const VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1351 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1352 1353 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1354 // If the element type of the vector doesn't match the result type, 1355 // bitcast it to be a vector type we can extract from. 1356 if (VecTy->getElementType() != DestTy) { 1357 VecTy = VectorType::get(DestTy, 1358 VecTy->getPrimitiveSizeInBits() / DestWidth); 1359 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1360 } 1361 1362 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); 1363 } 1364 } 1365 1366 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1367 ConstantInt *ShAmt = 0; 1368 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1369 m_ConstantInt(ShAmt)))) && 1370 isa<VectorType>(VecInput->getType())) { 1371 const VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1372 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1373 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1374 ShAmt->getZExtValue() % DestWidth == 0) { 1375 // If the element type of the vector doesn't match the result type, 1376 // bitcast it to be a vector type we can extract from. 1377 if (VecTy->getElementType() != DestTy) { 1378 VecTy = VectorType::get(DestTy, 1379 VecTy->getPrimitiveSizeInBits() / DestWidth); 1380 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1381 } 1382 1383 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1384 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1385 } 1386 } 1387 return 0; 1388} 1389 1390Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1391 // If the operands are integer typed then apply the integer transforms, 1392 // otherwise just apply the common ones. 1393 Value *Src = CI.getOperand(0); 1394 const Type *SrcTy = Src->getType(); 1395 const Type *DestTy = CI.getType(); 1396 1397 // Get rid of casts from one type to the same type. These are useless and can 1398 // be replaced by the operand. 1399 if (DestTy == Src->getType()) 1400 return ReplaceInstUsesWith(CI, Src); 1401 1402 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1403 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1404 const Type *DstElTy = DstPTy->getElementType(); 1405 const Type *SrcElTy = SrcPTy->getElementType(); 1406 1407 // If the address spaces don't match, don't eliminate the bitcast, which is 1408 // required for changing types. 1409 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1410 return 0; 1411 1412 // If we are casting a alloca to a pointer to a type of the same 1413 // size, rewrite the allocation instruction to allocate the "right" type. 1414 // There is no need to modify malloc calls because it is their bitcast that 1415 // needs to be cleaned up. 1416 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1417 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1418 return V; 1419 1420 // If the source and destination are pointers, and this cast is equivalent 1421 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1422 // This can enhance SROA and other transforms that want type-safe pointers. 1423 Constant *ZeroUInt = 1424 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1425 unsigned NumZeros = 0; 1426 while (SrcElTy != DstElTy && 1427 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1428 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1429 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1430 ++NumZeros; 1431 } 1432 1433 // If we found a path from the src to dest, create the getelementptr now. 1434 if (SrcElTy == DstElTy) { 1435 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1436 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1437 ((Instruction*)NULL)); 1438 } 1439 } 1440 1441 // Try to optimize int -> float bitcasts. 1442 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1443 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1444 return I; 1445 1446 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1447 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1448 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1449 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1450 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1451 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1452 } 1453 1454 // If this is a cast from an integer to vector, check to see if the input 1455 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1456 // the casts with a shuffle and (potentially) a bitcast. 1457 if (isa<IntegerType>(SrcTy) && (isa<TruncInst>(Src) || isa<ZExtInst>(Src))){ 1458 CastInst *SrcCast = cast<CastInst>(Src); 1459 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1460 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1461 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1462 cast<VectorType>(DestTy), *this)) 1463 return I; 1464 } 1465 } 1466 1467 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1468 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1469 Value *Elem = 1470 Builder->CreateExtractElement(Src, 1471 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1472 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1473 } 1474 } 1475 1476 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1477 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1478 // a bitcast to a vector with the same # elts. 1479 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1480 cast<VectorType>(DestTy)->getNumElements() == 1481 SVI->getType()->getNumElements() && 1482 SVI->getType()->getNumElements() == 1483 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1484 BitCastInst *Tmp; 1485 // If either of the operands is a cast from CI.getType(), then 1486 // evaluating the shuffle in the casted destination's type will allow 1487 // us to eliminate at least one cast. 1488 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1489 Tmp->getOperand(0)->getType() == DestTy) || 1490 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1491 Tmp->getOperand(0)->getType() == DestTy)) { 1492 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1493 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1494 // Return a new shuffle vector. Use the same element ID's, as we 1495 // know the vector types match #elts. 1496 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1497 } 1498 } 1499 } 1500 1501 if (SrcTy->isPointerTy()) 1502 return commonPointerCastTransforms(CI); 1503 return commonCastTransforms(CI); 1504} 1505