InstCombineCasts.cpp revision 3ecfc861b4365f341c5c969b40e1afccde676e6f
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. 91 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0; 92 93 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy); 94 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); 95 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 96 97 // See if we can satisfy the modulus by pulling a scale out of the array 98 // size argument. 99 unsigned ArraySizeScale; 100 uint64_t ArrayOffset; 101 Value *NumElements = // See if the array size is a decomposable linear expr. 102 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 103 104 // If we can now satisfy the modulus, by using a non-1 scale, we really can 105 // do the xform. 106 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 107 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 108 109 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 110 Value *Amt = 0; 111 if (Scale == 1) { 112 Amt = NumElements; 113 } else { 114 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 115 // Insert before the alloca, not before the cast. 116 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp"); 117 } 118 119 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 120 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 121 Offset, true); 122 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp"); 123 } 124 125 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 126 New->setAlignment(AI.getAlignment()); 127 New->takeName(&AI); 128 129 // If the allocation has multiple real uses, insert a cast and change all 130 // things that used it to use the new cast. This will also hack on CI, but it 131 // will die soon. 132 if (!AI.hasOneUse()) { 133 // New is the allocation instruction, pointer typed. AI is the original 134 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 135 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 136 AI.replaceAllUsesWith(NewCast); 137 } 138 return ReplaceInstUsesWith(CI, New); 139} 140 141 142 143/// EvaluateInDifferentType - Given an expression that 144/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 145/// insert the code to evaluate the expression. 146Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty, 147 bool isSigned) { 148 if (Constant *C = dyn_cast<Constant>(V)) { 149 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 150 // If we got a constantexpr back, try to simplify it with TD info. 151 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 152 C = ConstantFoldConstantExpression(CE, TD); 153 return C; 154 } 155 156 // Otherwise, it must be an instruction. 157 Instruction *I = cast<Instruction>(V); 158 Instruction *Res = 0; 159 unsigned Opc = I->getOpcode(); 160 switch (Opc) { 161 case Instruction::Add: 162 case Instruction::Sub: 163 case Instruction::Mul: 164 case Instruction::And: 165 case Instruction::Or: 166 case Instruction::Xor: 167 case Instruction::AShr: 168 case Instruction::LShr: 169 case Instruction::Shl: 170 case Instruction::UDiv: 171 case Instruction::URem: { 172 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 173 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 174 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 175 break; 176 } 177 case Instruction::Trunc: 178 case Instruction::ZExt: 179 case Instruction::SExt: 180 // If the source type of the cast is the type we're trying for then we can 181 // just return the source. There's no need to insert it because it is not 182 // new. 183 if (I->getOperand(0)->getType() == Ty) 184 return I->getOperand(0); 185 186 // Otherwise, must be the same type of cast, so just reinsert a new one. 187 // This also handles the case of zext(trunc(x)) -> zext(x). 188 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 189 Opc == Instruction::SExt); 190 break; 191 case Instruction::Select: { 192 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 193 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 194 Res = SelectInst::Create(I->getOperand(0), True, False); 195 break; 196 } 197 case Instruction::PHI: { 198 PHINode *OPN = cast<PHINode>(I); 199 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues()); 200 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 201 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 202 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 203 } 204 Res = NPN; 205 break; 206 } 207 default: 208 // TODO: Can handle more cases here. 209 llvm_unreachable("Unreachable!"); 210 break; 211 } 212 213 Res->takeName(I); 214 return InsertNewInstBefore(Res, *I); 215} 216 217 218/// This function is a wrapper around CastInst::isEliminableCastPair. It 219/// simply extracts arguments and returns what that function returns. 220static Instruction::CastOps 221isEliminableCastPair( 222 const CastInst *CI, ///< The first cast instruction 223 unsigned opcode, ///< The opcode of the second cast instruction 224 const Type *DstTy, ///< The target type for the second cast instruction 225 TargetData *TD ///< The target data for pointer size 226) { 227 228 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above 229 const Type *MidTy = CI->getType(); // B from above 230 231 // Get the opcodes of the two Cast instructions 232 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 233 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 234 235 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 236 DstTy, 237 TD ? TD->getIntPtrType(CI->getContext()) : 0); 238 239 // We don't want to form an inttoptr or ptrtoint that converts to an integer 240 // type that differs from the pointer size. 241 if ((Res == Instruction::IntToPtr && 242 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 243 (Res == Instruction::PtrToInt && 244 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 245 Res = 0; 246 247 return Instruction::CastOps(Res); 248} 249 250/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 251/// results in any code being generated and is interesting to optimize out. If 252/// the cast can be eliminated by some other simple transformation, we prefer 253/// to do the simplification first. 254bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 255 const Type *Ty) { 256 // Noop casts and casts of constants should be eliminated trivially. 257 if (V->getType() == Ty || isa<Constant>(V)) return false; 258 259 // If this is another cast that can be eliminated, we prefer to have it 260 // eliminated. 261 if (const CastInst *CI = dyn_cast<CastInst>(V)) 262 if (isEliminableCastPair(CI, opc, Ty, TD)) 263 return false; 264 265 // If this is a vector sext from a compare, then we don't want to break the 266 // idiom where each element of the extended vector is either zero or all ones. 267 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 268 return false; 269 270 return true; 271} 272 273 274/// @brief Implement the transforms common to all CastInst visitors. 275Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 276 Value *Src = CI.getOperand(0); 277 278 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 279 // eliminate it now. 280 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 281 if (Instruction::CastOps opc = 282 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 283 // The first cast (CSrc) is eliminable so we need to fix up or replace 284 // the second cast (CI). CSrc will then have a good chance of being dead. 285 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 286 } 287 } 288 289 // If we are casting a select then fold the cast into the select 290 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 291 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 292 return NV; 293 294 // If we are casting a PHI then fold the cast into the PHI 295 if (isa<PHINode>(Src)) { 296 // We don't do this if this would create a PHI node with an illegal type if 297 // it is currently legal. 298 if (!Src->getType()->isIntegerTy() || 299 !CI.getType()->isIntegerTy() || 300 ShouldChangeType(CI.getType(), Src->getType())) 301 if (Instruction *NV = FoldOpIntoPhi(CI)) 302 return NV; 303 } 304 305 return 0; 306} 307 308/// CanEvaluateTruncated - Return true if we can evaluate the specified 309/// expression tree as type Ty instead of its larger type, and arrive with the 310/// same value. This is used by code that tries to eliminate truncates. 311/// 312/// Ty will always be a type smaller than V. We should return true if trunc(V) 313/// can be computed by computing V in the smaller type. If V is an instruction, 314/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 315/// makes sense if x and y can be efficiently truncated. 316/// 317/// This function works on both vectors and scalars. 318/// 319static bool CanEvaluateTruncated(Value *V, const Type *Ty) { 320 // We can always evaluate constants in another type. 321 if (isa<Constant>(V)) 322 return true; 323 324 Instruction *I = dyn_cast<Instruction>(V); 325 if (!I) return false; 326 327 const Type *OrigTy = V->getType(); 328 329 // If this is an extension from the dest type, we can eliminate it, even if it 330 // has multiple uses. 331 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 332 I->getOperand(0)->getType() == Ty) 333 return true; 334 335 // We can't extend or shrink something that has multiple uses: doing so would 336 // require duplicating the instruction in general, which isn't profitable. 337 if (!I->hasOneUse()) return false; 338 339 unsigned Opc = I->getOpcode(); 340 switch (Opc) { 341 case Instruction::Add: 342 case Instruction::Sub: 343 case Instruction::Mul: 344 case Instruction::And: 345 case Instruction::Or: 346 case Instruction::Xor: 347 // These operators can all arbitrarily be extended or truncated. 348 return CanEvaluateTruncated(I->getOperand(0), Ty) && 349 CanEvaluateTruncated(I->getOperand(1), Ty); 350 351 case Instruction::UDiv: 352 case Instruction::URem: { 353 // UDiv and URem can be truncated if all the truncated bits are zero. 354 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 355 uint32_t BitWidth = Ty->getScalarSizeInBits(); 356 if (BitWidth < OrigBitWidth) { 357 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 358 if (MaskedValueIsZero(I->getOperand(0), Mask) && 359 MaskedValueIsZero(I->getOperand(1), Mask)) { 360 return CanEvaluateTruncated(I->getOperand(0), Ty) && 361 CanEvaluateTruncated(I->getOperand(1), Ty); 362 } 363 } 364 break; 365 } 366 case Instruction::Shl: 367 // If we are truncating the result of this SHL, and if it's a shift of a 368 // constant amount, we can always perform a SHL in a smaller type. 369 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 370 uint32_t BitWidth = Ty->getScalarSizeInBits(); 371 if (CI->getLimitedValue(BitWidth) < BitWidth) 372 return CanEvaluateTruncated(I->getOperand(0), Ty); 373 } 374 break; 375 case Instruction::LShr: 376 // If this is a truncate of a logical shr, we can truncate it to a smaller 377 // lshr iff we know that the bits we would otherwise be shifting in are 378 // already zeros. 379 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 380 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 381 uint32_t BitWidth = Ty->getScalarSizeInBits(); 382 if (MaskedValueIsZero(I->getOperand(0), 383 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 384 CI->getLimitedValue(BitWidth) < BitWidth) { 385 return CanEvaluateTruncated(I->getOperand(0), Ty); 386 } 387 } 388 break; 389 case Instruction::Trunc: 390 // trunc(trunc(x)) -> trunc(x) 391 return true; 392 case Instruction::ZExt: 393 case Instruction::SExt: 394 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 395 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 396 return true; 397 case Instruction::Select: { 398 SelectInst *SI = cast<SelectInst>(I); 399 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 400 CanEvaluateTruncated(SI->getFalseValue(), Ty); 401 } 402 case Instruction::PHI: { 403 // We can change a phi if we can change all operands. Note that we never 404 // get into trouble with cyclic PHIs here because we only consider 405 // instructions with a single use. 406 PHINode *PN = cast<PHINode>(I); 407 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 408 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 409 return false; 410 return true; 411 } 412 default: 413 // TODO: Can handle more cases here. 414 break; 415 } 416 417 return false; 418} 419 420Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 421 if (Instruction *Result = commonCastTransforms(CI)) 422 return Result; 423 424 // See if we can simplify any instructions used by the input whose sole 425 // purpose is to compute bits we don't care about. 426 if (SimplifyDemandedInstructionBits(CI)) 427 return &CI; 428 429 Value *Src = CI.getOperand(0); 430 const Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 431 432 // Attempt to truncate the entire input expression tree to the destination 433 // type. Only do this if the dest type is a simple type, don't convert the 434 // expression tree to something weird like i93 unless the source is also 435 // strange. 436 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 437 CanEvaluateTruncated(Src, DestTy)) { 438 439 // If this cast is a truncate, evaluting in a different type always 440 // eliminates the cast, so it is always a win. 441 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 442 " to avoid cast: " << CI << '\n'); 443 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 444 assert(Res->getType() == DestTy); 445 return ReplaceInstUsesWith(CI, Res); 446 } 447 448 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 449 if (DestTy->getScalarSizeInBits() == 1) { 450 Constant *One = ConstantInt::get(Src->getType(), 1); 451 Src = Builder->CreateAnd(Src, One, "tmp"); 452 Value *Zero = Constant::getNullValue(Src->getType()); 453 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 454 } 455 456 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 457 Value *A = 0; ConstantInt *Cst = 0; 458 if (Src->hasOneUse() && 459 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 460 // We have three types to worry about here, the type of A, the source of 461 // the truncate (MidSize), and the destination of the truncate. We know that 462 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 463 // between ASize and ResultSize. 464 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 465 466 // If the shift amount is larger than the size of A, then the result is 467 // known to be zero because all the input bits got shifted out. 468 if (Cst->getZExtValue() >= ASize) 469 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType())); 470 471 // Since we're doing an lshr and a zero extend, and know that the shift 472 // amount is smaller than ASize, it is always safe to do the shift in A's 473 // type, then zero extend or truncate to the result. 474 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 475 Shift->takeName(Src); 476 return CastInst::CreateIntegerCast(Shift, CI.getType(), false); 477 } 478 479 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 480 // type isn't non-native. 481 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) && 482 ShouldChangeType(Src->getType(), CI.getType()) && 483 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 484 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr"); 485 return BinaryOperator::CreateAnd(NewTrunc, 486 ConstantExpr::getTrunc(Cst, CI.getType())); 487 } 488 489 return 0; 490} 491 492/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 493/// in order to eliminate the icmp. 494Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 495 bool DoXform) { 496 // If we are just checking for a icmp eq of a single bit and zext'ing it 497 // to an integer, then shift the bit to the appropriate place and then 498 // cast to integer to avoid the comparison. 499 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 500 const APInt &Op1CV = Op1C->getValue(); 501 502 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 503 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 504 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 505 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 506 if (!DoXform) return ICI; 507 508 Value *In = ICI->getOperand(0); 509 Value *Sh = ConstantInt::get(In->getType(), 510 In->getType()->getScalarSizeInBits()-1); 511 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 512 if (In->getType() != CI.getType()) 513 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 514 515 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 516 Constant *One = ConstantInt::get(In->getType(), 1); 517 In = Builder->CreateXor(In, One, In->getName()+".not"); 518 } 519 520 return ReplaceInstUsesWith(CI, In); 521 } 522 523 524 525 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 526 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 527 // zext (X == 1) to i32 --> X iff X has only the low bit set. 528 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 529 // zext (X != 0) to i32 --> X iff X has only the low bit set. 530 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 531 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 532 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 533 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 534 // This only works for EQ and NE 535 ICI->isEquality()) { 536 // If Op1C some other power of two, convert: 537 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 538 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 539 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 540 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 541 542 APInt KnownZeroMask(~KnownZero); 543 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 544 if (!DoXform) return ICI; 545 546 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 547 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 548 // (X&4) == 2 --> false 549 // (X&4) != 2 --> true 550 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 551 isNE); 552 Res = ConstantExpr::getZExt(Res, CI.getType()); 553 return ReplaceInstUsesWith(CI, Res); 554 } 555 556 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 557 Value *In = ICI->getOperand(0); 558 if (ShiftAmt) { 559 // Perform a logical shr by shiftamt. 560 // Insert the shift to put the result in the low bit. 561 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 562 In->getName()+".lobit"); 563 } 564 565 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 566 Constant *One = ConstantInt::get(In->getType(), 1); 567 In = Builder->CreateXor(In, One, "tmp"); 568 } 569 570 if (CI.getType() == In->getType()) 571 return ReplaceInstUsesWith(CI, In); 572 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 573 } 574 } 575 } 576 577 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 578 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 579 // may lead to additional simplifications. 580 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 581 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 582 uint32_t BitWidth = ITy->getBitWidth(); 583 Value *LHS = ICI->getOperand(0); 584 Value *RHS = ICI->getOperand(1); 585 586 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 587 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 588 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 589 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 590 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 591 592 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 593 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 594 APInt UnknownBit = ~KnownBits; 595 if (UnknownBit.countPopulation() == 1) { 596 if (!DoXform) return ICI; 597 598 Value *Result = Builder->CreateXor(LHS, RHS); 599 600 // Mask off any bits that are set and won't be shifted away. 601 if (KnownOneLHS.uge(UnknownBit)) 602 Result = Builder->CreateAnd(Result, 603 ConstantInt::get(ITy, UnknownBit)); 604 605 // Shift the bit we're testing down to the lsb. 606 Result = Builder->CreateLShr( 607 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 608 609 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 610 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 611 Result->takeName(ICI); 612 return ReplaceInstUsesWith(CI, Result); 613 } 614 } 615 } 616 } 617 618 return 0; 619} 620 621/// CanEvaluateZExtd - Determine if the specified value can be computed in the 622/// specified wider type and produce the same low bits. If not, return false. 623/// 624/// If this function returns true, it can also return a non-zero number of bits 625/// (in BitsToClear) which indicates that the value it computes is correct for 626/// the zero extend, but that the additional BitsToClear bits need to be zero'd 627/// out. For example, to promote something like: 628/// 629/// %B = trunc i64 %A to i32 630/// %C = lshr i32 %B, 8 631/// %E = zext i32 %C to i64 632/// 633/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 634/// set to 8 to indicate that the promoted value needs to have bits 24-31 635/// cleared in addition to bits 32-63. Since an 'and' will be generated to 636/// clear the top bits anyway, doing this has no extra cost. 637/// 638/// This function works on both vectors and scalars. 639static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) { 640 BitsToClear = 0; 641 if (isa<Constant>(V)) 642 return true; 643 644 Instruction *I = dyn_cast<Instruction>(V); 645 if (!I) return false; 646 647 // If the input is a truncate from the destination type, we can trivially 648 // eliminate it, even if it has multiple uses. 649 // FIXME: This is currently disabled until codegen can handle this without 650 // pessimizing code, PR5997. 651 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 652 return true; 653 654 // We can't extend or shrink something that has multiple uses: doing so would 655 // require duplicating the instruction in general, which isn't profitable. 656 if (!I->hasOneUse()) return false; 657 658 unsigned Opc = I->getOpcode(), Tmp; 659 switch (Opc) { 660 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 661 case Instruction::SExt: // zext(sext(x)) -> sext(x). 662 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 663 return true; 664 case Instruction::And: 665 case Instruction::Or: 666 case Instruction::Xor: 667 case Instruction::Add: 668 case Instruction::Sub: 669 case Instruction::Mul: 670 case Instruction::Shl: 671 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 672 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 673 return false; 674 // These can all be promoted if neither operand has 'bits to clear'. 675 if (BitsToClear == 0 && Tmp == 0) 676 return true; 677 678 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 679 // other side, BitsToClear is ok. 680 if (Tmp == 0 && 681 (Opc == Instruction::And || Opc == Instruction::Or || 682 Opc == Instruction::Xor)) { 683 // We use MaskedValueIsZero here for generality, but the case we care 684 // about the most is constant RHS. 685 unsigned VSize = V->getType()->getScalarSizeInBits(); 686 if (MaskedValueIsZero(I->getOperand(1), 687 APInt::getHighBitsSet(VSize, BitsToClear))) 688 return true; 689 } 690 691 // Otherwise, we don't know how to analyze this BitsToClear case yet. 692 return false; 693 694 case Instruction::LShr: 695 // We can promote lshr(x, cst) if we can promote x. This requires the 696 // ultimate 'and' to clear out the high zero bits we're clearing out though. 697 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 698 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 699 return false; 700 BitsToClear += Amt->getZExtValue(); 701 if (BitsToClear > V->getType()->getScalarSizeInBits()) 702 BitsToClear = V->getType()->getScalarSizeInBits(); 703 return true; 704 } 705 // Cannot promote variable LSHR. 706 return false; 707 case Instruction::Select: 708 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 709 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 710 // TODO: If important, we could handle the case when the BitsToClear are 711 // known zero in the disagreeing side. 712 Tmp != BitsToClear) 713 return false; 714 return true; 715 716 case Instruction::PHI: { 717 // We can change a phi if we can change all operands. Note that we never 718 // get into trouble with cyclic PHIs here because we only consider 719 // instructions with a single use. 720 PHINode *PN = cast<PHINode>(I); 721 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 722 return false; 723 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 724 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 725 // TODO: If important, we could handle the case when the BitsToClear 726 // are known zero in the disagreeing input. 727 Tmp != BitsToClear) 728 return false; 729 return true; 730 } 731 default: 732 // TODO: Can handle more cases here. 733 return false; 734 } 735} 736 737Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 738 // If this zero extend is only used by a truncate, let the truncate by 739 // eliminated before we try to optimize this zext. 740 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 741 return 0; 742 743 // If one of the common conversion will work, do it. 744 if (Instruction *Result = commonCastTransforms(CI)) 745 return Result; 746 747 // See if we can simplify any instructions used by the input whose sole 748 // purpose is to compute bits we don't care about. 749 if (SimplifyDemandedInstructionBits(CI)) 750 return &CI; 751 752 Value *Src = CI.getOperand(0); 753 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 754 755 // Attempt to extend the entire input expression tree to the destination 756 // type. Only do this if the dest type is a simple type, don't convert the 757 // expression tree to something weird like i93 unless the source is also 758 // strange. 759 unsigned BitsToClear; 760 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 761 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 762 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 763 "Unreasonable BitsToClear"); 764 765 // Okay, we can transform this! Insert the new expression now. 766 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 767 " to avoid zero extend: " << CI); 768 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 769 assert(Res->getType() == DestTy); 770 771 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 772 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 773 774 // If the high bits are already filled with zeros, just replace this 775 // cast with the result. 776 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 777 DestBitSize-SrcBitsKept))) 778 return ReplaceInstUsesWith(CI, Res); 779 780 // We need to emit an AND to clear the high bits. 781 Constant *C = ConstantInt::get(Res->getType(), 782 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 783 return BinaryOperator::CreateAnd(Res, C); 784 } 785 786 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 787 // types and if the sizes are just right we can convert this into a logical 788 // 'and' which will be much cheaper than the pair of casts. 789 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 790 // TODO: Subsume this into EvaluateInDifferentType. 791 792 // Get the sizes of the types involved. We know that the intermediate type 793 // will be smaller than A or C, but don't know the relation between A and C. 794 Value *A = CSrc->getOperand(0); 795 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 796 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 797 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 798 // If we're actually extending zero bits, then if 799 // SrcSize < DstSize: zext(a & mask) 800 // SrcSize == DstSize: a & mask 801 // SrcSize > DstSize: trunc(a) & mask 802 if (SrcSize < DstSize) { 803 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 804 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 805 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 806 return new ZExtInst(And, CI.getType()); 807 } 808 809 if (SrcSize == DstSize) { 810 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 811 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 812 AndValue)); 813 } 814 if (SrcSize > DstSize) { 815 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 816 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 817 return BinaryOperator::CreateAnd(Trunc, 818 ConstantInt::get(Trunc->getType(), 819 AndValue)); 820 } 821 } 822 823 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 824 return transformZExtICmp(ICI, CI); 825 826 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 827 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 828 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 829 // of the (zext icmp) will be transformed. 830 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 831 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 832 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 833 (transformZExtICmp(LHS, CI, false) || 834 transformZExtICmp(RHS, CI, false))) { 835 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 836 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 837 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 838 } 839 } 840 841 // zext(trunc(t) & C) -> (t & zext(C)). 842 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 843 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 844 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 845 Value *TI0 = TI->getOperand(0); 846 if (TI0->getType() == CI.getType()) 847 return 848 BinaryOperator::CreateAnd(TI0, 849 ConstantExpr::getZExt(C, CI.getType())); 850 } 851 852 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 853 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 854 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 855 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 856 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 857 And->getOperand(1) == C) 858 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 859 Value *TI0 = TI->getOperand(0); 860 if (TI0->getType() == CI.getType()) { 861 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 862 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 863 return BinaryOperator::CreateXor(NewAnd, ZC); 864 } 865 } 866 867 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 868 Value *X; 869 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 870 match(SrcI, m_Not(m_Value(X))) && 871 (!X->hasOneUse() || !isa<CmpInst>(X))) { 872 Value *New = Builder->CreateZExt(X, CI.getType()); 873 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 874 } 875 876 return 0; 877} 878 879/// CanEvaluateSExtd - Return true if we can take the specified value 880/// and return it as type Ty without inserting any new casts and without 881/// changing the value of the common low bits. This is used by code that tries 882/// to promote integer operations to a wider types will allow us to eliminate 883/// the extension. 884/// 885/// This function works on both vectors and scalars. 886/// 887static bool CanEvaluateSExtd(Value *V, const Type *Ty) { 888 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 889 "Can't sign extend type to a smaller type"); 890 // If this is a constant, it can be trivially promoted. 891 if (isa<Constant>(V)) 892 return true; 893 894 Instruction *I = dyn_cast<Instruction>(V); 895 if (!I) return false; 896 897 // If this is a truncate from the dest type, we can trivially eliminate it, 898 // even if it has multiple uses. 899 // FIXME: This is currently disabled until codegen can handle this without 900 // pessimizing code, PR5997. 901 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 902 return true; 903 904 // We can't extend or shrink something that has multiple uses: doing so would 905 // require duplicating the instruction in general, which isn't profitable. 906 if (!I->hasOneUse()) return false; 907 908 switch (I->getOpcode()) { 909 case Instruction::SExt: // sext(sext(x)) -> sext(x) 910 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 911 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 912 return true; 913 case Instruction::And: 914 case Instruction::Or: 915 case Instruction::Xor: 916 case Instruction::Add: 917 case Instruction::Sub: 918 case Instruction::Mul: 919 // These operators can all arbitrarily be extended if their inputs can. 920 return CanEvaluateSExtd(I->getOperand(0), Ty) && 921 CanEvaluateSExtd(I->getOperand(1), Ty); 922 923 //case Instruction::Shl: TODO 924 //case Instruction::LShr: TODO 925 926 case Instruction::Select: 927 return CanEvaluateSExtd(I->getOperand(1), Ty) && 928 CanEvaluateSExtd(I->getOperand(2), Ty); 929 930 case Instruction::PHI: { 931 // We can change a phi if we can change all operands. Note that we never 932 // get into trouble with cyclic PHIs here because we only consider 933 // instructions with a single use. 934 PHINode *PN = cast<PHINode>(I); 935 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 936 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 937 return true; 938 } 939 default: 940 // TODO: Can handle more cases here. 941 break; 942 } 943 944 return false; 945} 946 947Instruction *InstCombiner::visitSExt(SExtInst &CI) { 948 // If this sign extend is only used by a truncate, let the truncate by 949 // eliminated before we try to optimize this zext. 950 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 951 return 0; 952 953 if (Instruction *I = commonCastTransforms(CI)) 954 return I; 955 956 // See if we can simplify any instructions used by the input whose sole 957 // purpose is to compute bits we don't care about. 958 if (SimplifyDemandedInstructionBits(CI)) 959 return &CI; 960 961 Value *Src = CI.getOperand(0); 962 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 963 964 // Attempt to extend the entire input expression tree to the destination 965 // type. Only do this if the dest type is a simple type, don't convert the 966 // expression tree to something weird like i93 unless the source is also 967 // strange. 968 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 969 CanEvaluateSExtd(Src, DestTy)) { 970 // Okay, we can transform this! Insert the new expression now. 971 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 972 " to avoid sign extend: " << CI); 973 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 974 assert(Res->getType() == DestTy); 975 976 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 977 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 978 979 // If the high bits are already filled with sign bit, just replace this 980 // cast with the result. 981 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 982 return ReplaceInstUsesWith(CI, Res); 983 984 // We need to emit a shl + ashr to do the sign extend. 985 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 986 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 987 ShAmt); 988 } 989 990 // If this input is a trunc from our destination, then turn sext(trunc(x)) 991 // into shifts. 992 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 993 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 994 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 995 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 996 997 // We need to emit a shl + ashr to do the sign extend. 998 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 999 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1000 return BinaryOperator::CreateAShr(Res, ShAmt); 1001 } 1002 1003 1004 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed 1005 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed 1006 { 1007 ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS; 1008 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) { 1009 // sext (x <s 0) to i32 --> x>>s31 true if signbit set. 1010 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear. 1011 if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) || 1012 (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) { 1013 Value *Sh = ConstantInt::get(CmpLHS->getType(), 1014 CmpLHS->getType()->getScalarSizeInBits()-1); 1015 Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit"); 1016 if (In->getType() != CI.getType()) 1017 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp"); 1018 1019 if (Pred == ICmpInst::ICMP_SGT) 1020 In = Builder->CreateNot(In, In->getName()+".not"); 1021 return ReplaceInstUsesWith(CI, In); 1022 } 1023 } 1024 } 1025 1026 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed. 1027 if (const VectorType *VTy = dyn_cast<VectorType>(DestTy)) { 1028 ICmpInst::Predicate Pred; Value *CmpLHS; 1029 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_Zero()))) { 1030 if (Pred == ICmpInst::ICMP_SLT && CmpLHS->getType() == DestTy) { 1031 const Type *EltTy = VTy->getElementType(); 1032 1033 // splat the shift constant to a constant vector. 1034 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1); 1035 Value *In = Builder->CreateAShr(CmpLHS, VSh,CmpLHS->getName()+".lobit"); 1036 return ReplaceInstUsesWith(CI, In); 1037 } 1038 } 1039 } 1040 1041 // If the input is a shl/ashr pair of a same constant, then this is a sign 1042 // extension from a smaller value. If we could trust arbitrary bitwidth 1043 // integers, we could turn this into a truncate to the smaller bit and then 1044 // use a sext for the whole extension. Since we don't, look deeper and check 1045 // for a truncate. If the source and dest are the same type, eliminate the 1046 // trunc and extend and just do shifts. For example, turn: 1047 // %a = trunc i32 %i to i8 1048 // %b = shl i8 %a, 6 1049 // %c = ashr i8 %b, 6 1050 // %d = sext i8 %c to i32 1051 // into: 1052 // %a = shl i32 %i, 30 1053 // %d = ashr i32 %a, 30 1054 Value *A = 0; 1055 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1056 ConstantInt *BA = 0, *CA = 0; 1057 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1058 m_ConstantInt(CA))) && 1059 BA == CA && A->getType() == CI.getType()) { 1060 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1061 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1062 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1063 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1064 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1065 return BinaryOperator::CreateAShr(A, ShAmtV); 1066 } 1067 1068 return 0; 1069} 1070 1071 1072/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1073/// in the specified FP type without changing its value. 1074static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1075 bool losesInfo; 1076 APFloat F = CFP->getValueAPF(); 1077 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1078 if (!losesInfo) 1079 return ConstantFP::get(CFP->getContext(), F); 1080 return 0; 1081} 1082 1083/// LookThroughFPExtensions - If this is an fp extension instruction, look 1084/// through it until we get the source value. 1085static Value *LookThroughFPExtensions(Value *V) { 1086 if (Instruction *I = dyn_cast<Instruction>(V)) 1087 if (I->getOpcode() == Instruction::FPExt) 1088 return LookThroughFPExtensions(I->getOperand(0)); 1089 1090 // If this value is a constant, return the constant in the smallest FP type 1091 // that can accurately represent it. This allows us to turn 1092 // (float)((double)X+2.0) into x+2.0f. 1093 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1094 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1095 return V; // No constant folding of this. 1096 // See if the value can be truncated to float and then reextended. 1097 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1098 return V; 1099 if (CFP->getType()->isDoubleTy()) 1100 return V; // Won't shrink. 1101 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1102 return V; 1103 // Don't try to shrink to various long double types. 1104 } 1105 1106 return V; 1107} 1108 1109Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1110 if (Instruction *I = commonCastTransforms(CI)) 1111 return I; 1112 1113 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1114 // smaller than the destination type, we can eliminate the truncate by doing 1115 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1116 // as many builtins (sqrt, etc). 1117 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1118 if (OpI && OpI->hasOneUse()) { 1119 switch (OpI->getOpcode()) { 1120 default: break; 1121 case Instruction::FAdd: 1122 case Instruction::FSub: 1123 case Instruction::FMul: 1124 case Instruction::FDiv: 1125 case Instruction::FRem: 1126 const Type *SrcTy = OpI->getType(); 1127 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1128 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1129 if (LHSTrunc->getType() != SrcTy && 1130 RHSTrunc->getType() != SrcTy) { 1131 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1132 // If the source types were both smaller than the destination type of 1133 // the cast, do this xform. 1134 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1135 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1136 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1137 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1138 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1139 } 1140 } 1141 break; 1142 } 1143 } 1144 1145 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1146 // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it. 1147 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1148 if (Call && Call->getCalledFunction() && 1149 Call->getCalledFunction()->getName() == "sqrt" && 1150 Call->getNumArgOperands() == 1) { 1151 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1152 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1153 CI.getType()->isFloatTy() && 1154 Call->getType()->isDoubleTy() && 1155 Arg->getType()->isDoubleTy() && 1156 Arg->getOperand(0)->getType()->isFloatTy()) { 1157 Function *Callee = Call->getCalledFunction(); 1158 Module *M = CI.getParent()->getParent()->getParent(); 1159 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf", 1160 Callee->getAttributes(), 1161 Builder->getFloatTy(), 1162 Builder->getFloatTy(), 1163 NULL); 1164 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1165 "sqrtfcall"); 1166 ret->setAttributes(Callee->getAttributes()); 1167 1168 1169 // Remove the old Call. With -fmath-errno, it won't get marked readnone. 1170 Call->replaceAllUsesWith(UndefValue::get(Call->getType())); 1171 EraseInstFromFunction(*Call); 1172 return ret; 1173 } 1174 } 1175 1176 return 0; 1177} 1178 1179Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1180 return commonCastTransforms(CI); 1181} 1182 1183Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1184 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1185 if (OpI == 0) 1186 return commonCastTransforms(FI); 1187 1188 // fptoui(uitofp(X)) --> X 1189 // fptoui(sitofp(X)) --> X 1190 // This is safe if the intermediate type has enough bits in its mantissa to 1191 // accurately represent all values of X. For example, do not do this with 1192 // i64->float->i64. This is also safe for sitofp case, because any negative 1193 // 'X' value would cause an undefined result for the fptoui. 1194 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1195 OpI->getOperand(0)->getType() == FI.getType() && 1196 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1197 OpI->getType()->getFPMantissaWidth()) 1198 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1199 1200 return commonCastTransforms(FI); 1201} 1202 1203Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1204 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1205 if (OpI == 0) 1206 return commonCastTransforms(FI); 1207 1208 // fptosi(sitofp(X)) --> X 1209 // fptosi(uitofp(X)) --> X 1210 // This is safe if the intermediate type has enough bits in its mantissa to 1211 // accurately represent all values of X. For example, do not do this with 1212 // i64->float->i64. This is also safe for sitofp case, because any negative 1213 // 'X' value would cause an undefined result for the fptoui. 1214 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1215 OpI->getOperand(0)->getType() == FI.getType() && 1216 (int)FI.getType()->getScalarSizeInBits() <= 1217 OpI->getType()->getFPMantissaWidth()) 1218 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1219 1220 return commonCastTransforms(FI); 1221} 1222 1223Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1224 return commonCastTransforms(CI); 1225} 1226 1227Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1228 return commonCastTransforms(CI); 1229} 1230 1231Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1232 // If the source integer type is not the intptr_t type for this target, do a 1233 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1234 // cast to be exposed to other transforms. 1235 if (TD) { 1236 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1237 TD->getPointerSizeInBits()) { 1238 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1239 TD->getIntPtrType(CI.getContext()), "tmp"); 1240 return new IntToPtrInst(P, CI.getType()); 1241 } 1242 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1243 TD->getPointerSizeInBits()) { 1244 Value *P = Builder->CreateZExt(CI.getOperand(0), 1245 TD->getIntPtrType(CI.getContext()), "tmp"); 1246 return new IntToPtrInst(P, CI.getType()); 1247 } 1248 } 1249 1250 if (Instruction *I = commonCastTransforms(CI)) 1251 return I; 1252 1253 return 0; 1254} 1255 1256/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1257Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1258 Value *Src = CI.getOperand(0); 1259 1260 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1261 // If casting the result of a getelementptr instruction with no offset, turn 1262 // this into a cast of the original pointer! 1263 if (GEP->hasAllZeroIndices()) { 1264 // Changing the cast operand is usually not a good idea but it is safe 1265 // here because the pointer operand is being replaced with another 1266 // pointer operand so the opcode doesn't need to change. 1267 Worklist.Add(GEP); 1268 CI.setOperand(0, GEP->getOperand(0)); 1269 return &CI; 1270 } 1271 1272 // If the GEP has a single use, and the base pointer is a bitcast, and the 1273 // GEP computes a constant offset, see if we can convert these three 1274 // instructions into fewer. This typically happens with unions and other 1275 // non-type-safe code. 1276 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1277 GEP->hasAllConstantIndices()) { 1278 // We are guaranteed to get a constant from EmitGEPOffset. 1279 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1280 int64_t Offset = OffsetV->getSExtValue(); 1281 1282 // Get the base pointer input of the bitcast, and the type it points to. 1283 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1284 const Type *GEPIdxTy = 1285 cast<PointerType>(OrigBase->getType())->getElementType(); 1286 SmallVector<Value*, 8> NewIndices; 1287 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1288 // If we were able to index down into an element, create the GEP 1289 // and bitcast the result. This eliminates one bitcast, potentially 1290 // two. 1291 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1292 Builder->CreateInBoundsGEP(OrigBase, 1293 NewIndices.begin(), NewIndices.end()) : 1294 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 1295 NGEP->takeName(GEP); 1296 1297 if (isa<BitCastInst>(CI)) 1298 return new BitCastInst(NGEP, CI.getType()); 1299 assert(isa<PtrToIntInst>(CI)); 1300 return new PtrToIntInst(NGEP, CI.getType()); 1301 } 1302 } 1303 } 1304 1305 return commonCastTransforms(CI); 1306} 1307 1308Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1309 // If the destination integer type is not the intptr_t type for this target, 1310 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1311 // to be exposed to other transforms. 1312 if (TD) { 1313 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1314 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1315 TD->getIntPtrType(CI.getContext()), 1316 "tmp"); 1317 return new TruncInst(P, CI.getType()); 1318 } 1319 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1320 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1321 TD->getIntPtrType(CI.getContext()), 1322 "tmp"); 1323 return new ZExtInst(P, CI.getType()); 1324 } 1325 } 1326 1327 return commonPointerCastTransforms(CI); 1328} 1329 1330/// OptimizeVectorResize - This input value (which is known to have vector type) 1331/// is being zero extended or truncated to the specified vector type. Try to 1332/// replace it with a shuffle (and vector/vector bitcast) if possible. 1333/// 1334/// The source and destination vector types may have different element types. 1335static Instruction *OptimizeVectorResize(Value *InVal, const VectorType *DestTy, 1336 InstCombiner &IC) { 1337 // We can only do this optimization if the output is a multiple of the input 1338 // element size, or the input is a multiple of the output element size. 1339 // Convert the input type to have the same element type as the output. 1340 const VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1341 1342 if (SrcTy->getElementType() != DestTy->getElementType()) { 1343 // The input types don't need to be identical, but for now they must be the 1344 // same size. There is no specific reason we couldn't handle things like 1345 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1346 // there yet. 1347 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1348 DestTy->getElementType()->getPrimitiveSizeInBits()) 1349 return 0; 1350 1351 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1352 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1353 } 1354 1355 // Now that the element types match, get the shuffle mask and RHS of the 1356 // shuffle to use, which depends on whether we're increasing or decreasing the 1357 // size of the input. 1358 SmallVector<Constant*, 16> ShuffleMask; 1359 Value *V2; 1360 const IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext()); 1361 1362 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1363 // If we're shrinking the number of elements, just shuffle in the low 1364 // elements from the input and use undef as the second shuffle input. 1365 V2 = UndefValue::get(SrcTy); 1366 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1367 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1368 1369 } else { 1370 // If we're increasing the number of elements, shuffle in all of the 1371 // elements from InVal and fill the rest of the result elements with zeros 1372 // from a constant zero. 1373 V2 = Constant::getNullValue(SrcTy); 1374 unsigned SrcElts = SrcTy->getNumElements(); 1375 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1376 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1377 1378 // The excess elements reference the first element of the zero input. 1379 ShuffleMask.append(DestTy->getNumElements()-SrcElts, 1380 ConstantInt::get(Int32Ty, SrcElts)); 1381 } 1382 1383 return new ShuffleVectorInst(InVal, V2, ConstantVector::get(ShuffleMask)); 1384} 1385 1386static bool isMultipleOfTypeSize(unsigned Value, const Type *Ty) { 1387 return Value % Ty->getPrimitiveSizeInBits() == 0; 1388} 1389 1390static unsigned getTypeSizeIndex(unsigned Value, const Type *Ty) { 1391 return Value / Ty->getPrimitiveSizeInBits(); 1392} 1393 1394/// CollectInsertionElements - V is a value which is inserted into a vector of 1395/// VecEltTy. Look through the value to see if we can decompose it into 1396/// insertions into the vector. See the example in the comment for 1397/// OptimizeIntegerToVectorInsertions for the pattern this handles. 1398/// The type of V is always a non-zero multiple of VecEltTy's size. 1399/// 1400/// This returns false if the pattern can't be matched or true if it can, 1401/// filling in Elements with the elements found here. 1402static bool CollectInsertionElements(Value *V, unsigned ElementIndex, 1403 SmallVectorImpl<Value*> &Elements, 1404 const Type *VecEltTy) { 1405 // Undef values never contribute useful bits to the result. 1406 if (isa<UndefValue>(V)) return true; 1407 1408 // If we got down to a value of the right type, we win, try inserting into the 1409 // right element. 1410 if (V->getType() == VecEltTy) { 1411 // Inserting null doesn't actually insert any elements. 1412 if (Constant *C = dyn_cast<Constant>(V)) 1413 if (C->isNullValue()) 1414 return true; 1415 1416 // Fail if multiple elements are inserted into this slot. 1417 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0) 1418 return false; 1419 1420 Elements[ElementIndex] = V; 1421 return true; 1422 } 1423 1424 if (Constant *C = dyn_cast<Constant>(V)) { 1425 // Figure out the # elements this provides, and bitcast it or slice it up 1426 // as required. 1427 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1428 VecEltTy); 1429 // If the constant is the size of a vector element, we just need to bitcast 1430 // it to the right type so it gets properly inserted. 1431 if (NumElts == 1) 1432 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1433 ElementIndex, Elements, VecEltTy); 1434 1435 // Okay, this is a constant that covers multiple elements. Slice it up into 1436 // pieces and insert each element-sized piece into the vector. 1437 if (!isa<IntegerType>(C->getType())) 1438 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1439 C->getType()->getPrimitiveSizeInBits())); 1440 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1441 const Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1442 1443 for (unsigned i = 0; i != NumElts; ++i) { 1444 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1445 i*ElementSize)); 1446 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1447 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy)) 1448 return false; 1449 } 1450 return true; 1451 } 1452 1453 if (!V->hasOneUse()) return false; 1454 1455 Instruction *I = dyn_cast<Instruction>(V); 1456 if (I == 0) return false; 1457 switch (I->getOpcode()) { 1458 default: return false; // Unhandled case. 1459 case Instruction::BitCast: 1460 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1461 Elements, VecEltTy); 1462 case Instruction::ZExt: 1463 if (!isMultipleOfTypeSize( 1464 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1465 VecEltTy)) 1466 return false; 1467 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1468 Elements, VecEltTy); 1469 case Instruction::Or: 1470 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1471 Elements, VecEltTy) && 1472 CollectInsertionElements(I->getOperand(1), ElementIndex, 1473 Elements, VecEltTy); 1474 case Instruction::Shl: { 1475 // Must be shifting by a constant that is a multiple of the element size. 1476 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1477 if (CI == 0) return false; 1478 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false; 1479 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy); 1480 1481 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift, 1482 Elements, VecEltTy); 1483 } 1484 1485 } 1486} 1487 1488 1489/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1490/// may be doing shifts and ors to assemble the elements of the vector manually. 1491/// Try to rip the code out and replace it with insertelements. This is to 1492/// optimize code like this: 1493/// 1494/// %tmp37 = bitcast float %inc to i32 1495/// %tmp38 = zext i32 %tmp37 to i64 1496/// %tmp31 = bitcast float %inc5 to i32 1497/// %tmp32 = zext i32 %tmp31 to i64 1498/// %tmp33 = shl i64 %tmp32, 32 1499/// %ins35 = or i64 %tmp33, %tmp38 1500/// %tmp43 = bitcast i64 %ins35 to <2 x float> 1501/// 1502/// Into two insertelements that do "buildvector{%inc, %inc5}". 1503static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1504 InstCombiner &IC) { 1505 const VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1506 Value *IntInput = CI.getOperand(0); 1507 1508 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1509 if (!CollectInsertionElements(IntInput, 0, Elements, 1510 DestVecTy->getElementType())) 1511 return 0; 1512 1513 // If we succeeded, we know that all of the element are specified by Elements 1514 // or are zero if Elements has a null entry. Recast this as a set of 1515 // insertions. 1516 Value *Result = Constant::getNullValue(CI.getType()); 1517 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1518 if (Elements[i] == 0) continue; // Unset element. 1519 1520 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1521 IC.Builder->getInt32(i)); 1522 } 1523 1524 return Result; 1525} 1526 1527 1528/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1529/// bitcast. The various long double bitcasts can't get in here. 1530static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1531 Value *Src = CI.getOperand(0); 1532 const Type *DestTy = CI.getType(); 1533 1534 // If this is a bitcast from int to float, check to see if the int is an 1535 // extraction from a vector. 1536 Value *VecInput = 0; 1537 // bitcast(trunc(bitcast(somevector))) 1538 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1539 isa<VectorType>(VecInput->getType())) { 1540 const VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1541 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1542 1543 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1544 // If the element type of the vector doesn't match the result type, 1545 // bitcast it to be a vector type we can extract from. 1546 if (VecTy->getElementType() != DestTy) { 1547 VecTy = VectorType::get(DestTy, 1548 VecTy->getPrimitiveSizeInBits() / DestWidth); 1549 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1550 } 1551 1552 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); 1553 } 1554 } 1555 1556 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1557 ConstantInt *ShAmt = 0; 1558 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1559 m_ConstantInt(ShAmt)))) && 1560 isa<VectorType>(VecInput->getType())) { 1561 const VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1562 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1563 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1564 ShAmt->getZExtValue() % DestWidth == 0) { 1565 // If the element type of the vector doesn't match the result type, 1566 // bitcast it to be a vector type we can extract from. 1567 if (VecTy->getElementType() != DestTy) { 1568 VecTy = VectorType::get(DestTy, 1569 VecTy->getPrimitiveSizeInBits() / DestWidth); 1570 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1571 } 1572 1573 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1574 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1575 } 1576 } 1577 return 0; 1578} 1579 1580Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1581 // If the operands are integer typed then apply the integer transforms, 1582 // otherwise just apply the common ones. 1583 Value *Src = CI.getOperand(0); 1584 const Type *SrcTy = Src->getType(); 1585 const Type *DestTy = CI.getType(); 1586 1587 // Get rid of casts from one type to the same type. These are useless and can 1588 // be replaced by the operand. 1589 if (DestTy == Src->getType()) 1590 return ReplaceInstUsesWith(CI, Src); 1591 1592 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1593 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1594 const Type *DstElTy = DstPTy->getElementType(); 1595 const Type *SrcElTy = SrcPTy->getElementType(); 1596 1597 // If the address spaces don't match, don't eliminate the bitcast, which is 1598 // required for changing types. 1599 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1600 return 0; 1601 1602 // If we are casting a alloca to a pointer to a type of the same 1603 // size, rewrite the allocation instruction to allocate the "right" type. 1604 // There is no need to modify malloc calls because it is their bitcast that 1605 // needs to be cleaned up. 1606 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1607 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1608 return V; 1609 1610 // If the source and destination are pointers, and this cast is equivalent 1611 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1612 // This can enhance SROA and other transforms that want type-safe pointers. 1613 Constant *ZeroUInt = 1614 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1615 unsigned NumZeros = 0; 1616 while (SrcElTy != DstElTy && 1617 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1618 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1619 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1620 ++NumZeros; 1621 } 1622 1623 // If we found a path from the src to dest, create the getelementptr now. 1624 if (SrcElTy == DstElTy) { 1625 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1626 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1627 ((Instruction*)NULL)); 1628 } 1629 } 1630 1631 // Try to optimize int -> float bitcasts. 1632 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1633 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1634 return I; 1635 1636 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1637 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1638 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1639 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1640 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1641 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1642 } 1643 1644 if (isa<IntegerType>(SrcTy)) { 1645 // If this is a cast from an integer to vector, check to see if the input 1646 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1647 // the casts with a shuffle and (potentially) a bitcast. 1648 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1649 CastInst *SrcCast = cast<CastInst>(Src); 1650 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1651 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1652 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1653 cast<VectorType>(DestTy), *this)) 1654 return I; 1655 } 1656 1657 // If the input is an 'or' instruction, we may be doing shifts and ors to 1658 // assemble the elements of the vector manually. Try to rip the code out 1659 // and replace it with insertelements. 1660 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1661 return ReplaceInstUsesWith(CI, V); 1662 } 1663 } 1664 1665 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1666 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1667 Value *Elem = 1668 Builder->CreateExtractElement(Src, 1669 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1670 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1671 } 1672 } 1673 1674 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1675 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1676 // a bitcast to a vector with the same # elts. 1677 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1678 cast<VectorType>(DestTy)->getNumElements() == 1679 SVI->getType()->getNumElements() && 1680 SVI->getType()->getNumElements() == 1681 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1682 BitCastInst *Tmp; 1683 // If either of the operands is a cast from CI.getType(), then 1684 // evaluating the shuffle in the casted destination's type will allow 1685 // us to eliminate at least one cast. 1686 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1687 Tmp->getOperand(0)->getType() == DestTy) || 1688 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1689 Tmp->getOperand(0)->getType() == DestTy)) { 1690 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1691 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1692 // Return a new shuffle vector. Use the same element ID's, as we 1693 // know the vector types match #elts. 1694 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1695 } 1696 } 1697 } 1698 1699 if (SrcTy->isPointerTy()) 1700 return commonPointerCastTransforms(CI); 1701 return commonCastTransforms(CI); 1702} 1703