InstCombineCasts.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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/Analysis/ConstantFolding.h" 16#include "llvm/IR/DataLayout.h" 17#include "llvm/IR/PatternMatch.h" 18#include "llvm/Target/TargetLibraryInfo.h" 19using namespace llvm; 20using namespace PatternMatch; 21 22/// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear 23/// expression. If so, decompose it, returning some value X, such that Val is 24/// X*Scale+Offset. 25/// 26static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, 27 uint64_t &Offset) { 28 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 29 Offset = CI->getZExtValue(); 30 Scale = 0; 31 return ConstantInt::get(Val->getType(), 0); 32 } 33 34 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) { 35 // Cannot look past anything that might overflow. 36 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val); 37 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) { 38 Scale = 1; 39 Offset = 0; 40 return Val; 41 } 42 43 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 44 if (I->getOpcode() == Instruction::Shl) { 45 // This is a value scaled by '1 << the shift amt'. 46 Scale = UINT64_C(1) << RHS->getZExtValue(); 47 Offset = 0; 48 return I->getOperand(0); 49 } 50 51 if (I->getOpcode() == Instruction::Mul) { 52 // This value is scaled by 'RHS'. 53 Scale = RHS->getZExtValue(); 54 Offset = 0; 55 return I->getOperand(0); 56 } 57 58 if (I->getOpcode() == Instruction::Add) { 59 // We have X+C. Check to see if we really have (X*C2)+C1, 60 // where C1 is divisible by C2. 61 unsigned SubScale; 62 Value *SubVal = 63 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); 64 Offset += RHS->getZExtValue(); 65 Scale = SubScale; 66 return SubVal; 67 } 68 } 69 } 70 71 // Otherwise, we can't look past this. 72 Scale = 1; 73 Offset = 0; 74 return Val; 75} 76 77/// PromoteCastOfAllocation - If we find a cast of an allocation instruction, 78/// try to eliminate the cast by moving the type information into the alloc. 79Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 80 AllocaInst &AI) { 81 // This requires DataLayout to get the alloca alignment and size information. 82 if (!DL) return 0; 83 84 PointerType *PTy = cast<PointerType>(CI.getType()); 85 86 BuilderTy AllocaBuilder(*Builder); 87 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); 88 89 // Get the type really allocated and the type casted to. 90 Type *AllocElTy = AI.getAllocatedType(); 91 Type *CastElTy = PTy->getElementType(); 92 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0; 93 94 unsigned AllocElTyAlign = DL->getABITypeAlignment(AllocElTy); 95 unsigned CastElTyAlign = DL->getABITypeAlignment(CastElTy); 96 if (CastElTyAlign < AllocElTyAlign) return 0; 97 98 // If the allocation has multiple uses, only promote it if we are strictly 99 // increasing the alignment of the resultant allocation. If we keep it the 100 // same, we open the door to infinite loops of various kinds. 101 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0; 102 103 uint64_t AllocElTySize = DL->getTypeAllocSize(AllocElTy); 104 uint64_t CastElTySize = DL->getTypeAllocSize(CastElTy); 105 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 106 107 // If the allocation has multiple uses, only promote it if we're not 108 // shrinking the amount of memory being allocated. 109 uint64_t AllocElTyStoreSize = DL->getTypeStoreSize(AllocElTy); 110 uint64_t CastElTyStoreSize = DL->getTypeStoreSize(CastElTy); 111 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return 0; 112 113 // See if we can satisfy the modulus by pulling a scale out of the array 114 // size argument. 115 unsigned ArraySizeScale; 116 uint64_t ArrayOffset; 117 Value *NumElements = // See if the array size is a decomposable linear expr. 118 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 119 120 // If we can now satisfy the modulus, by using a non-1 scale, we really can 121 // do the xform. 122 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 123 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 124 125 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 126 Value *Amt = 0; 127 if (Scale == 1) { 128 Amt = NumElements; 129 } else { 130 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 131 // Insert before the alloca, not before the cast. 132 Amt = AllocaBuilder.CreateMul(Amt, NumElements); 133 } 134 135 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 136 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 137 Offset, true); 138 Amt = AllocaBuilder.CreateAdd(Amt, Off); 139 } 140 141 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 142 New->setAlignment(AI.getAlignment()); 143 New->takeName(&AI); 144 145 // If the allocation has multiple real uses, insert a cast and change all 146 // things that used it to use the new cast. This will also hack on CI, but it 147 // will die soon. 148 if (!AI.hasOneUse()) { 149 // New is the allocation instruction, pointer typed. AI is the original 150 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 151 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 152 ReplaceInstUsesWith(AI, NewCast); 153 } 154 return ReplaceInstUsesWith(CI, New); 155} 156 157/// EvaluateInDifferentType - Given an expression that 158/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 159/// insert the code to evaluate the expression. 160Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, 161 bool isSigned) { 162 if (Constant *C = dyn_cast<Constant>(V)) { 163 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 164 // If we got a constantexpr back, try to simplify it with DL info. 165 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 166 C = ConstantFoldConstantExpression(CE, DL, TLI); 167 return C; 168 } 169 170 // Otherwise, it must be an instruction. 171 Instruction *I = cast<Instruction>(V); 172 Instruction *Res = 0; 173 unsigned Opc = I->getOpcode(); 174 switch (Opc) { 175 case Instruction::Add: 176 case Instruction::Sub: 177 case Instruction::Mul: 178 case Instruction::And: 179 case Instruction::Or: 180 case Instruction::Xor: 181 case Instruction::AShr: 182 case Instruction::LShr: 183 case Instruction::Shl: 184 case Instruction::UDiv: 185 case Instruction::URem: { 186 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 187 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 188 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 189 break; 190 } 191 case Instruction::Trunc: 192 case Instruction::ZExt: 193 case Instruction::SExt: 194 // If the source type of the cast is the type we're trying for then we can 195 // just return the source. There's no need to insert it because it is not 196 // new. 197 if (I->getOperand(0)->getType() == Ty) 198 return I->getOperand(0); 199 200 // Otherwise, must be the same type of cast, so just reinsert a new one. 201 // This also handles the case of zext(trunc(x)) -> zext(x). 202 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 203 Opc == Instruction::SExt); 204 break; 205 case Instruction::Select: { 206 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 207 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 208 Res = SelectInst::Create(I->getOperand(0), True, False); 209 break; 210 } 211 case Instruction::PHI: { 212 PHINode *OPN = cast<PHINode>(I); 213 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues()); 214 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 215 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 216 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 217 } 218 Res = NPN; 219 break; 220 } 221 default: 222 // TODO: Can handle more cases here. 223 llvm_unreachable("Unreachable!"); 224 } 225 226 Res->takeName(I); 227 return InsertNewInstWith(Res, *I); 228} 229 230 231/// This function is a wrapper around CastInst::isEliminableCastPair. It 232/// simply extracts arguments and returns what that function returns. 233static Instruction::CastOps 234isEliminableCastPair( 235 const CastInst *CI, ///< The first cast instruction 236 unsigned opcode, ///< The opcode of the second cast instruction 237 Type *DstTy, ///< The target type for the second cast instruction 238 const DataLayout *DL ///< The target data for pointer size 239) { 240 241 Type *SrcTy = CI->getOperand(0)->getType(); // A from above 242 Type *MidTy = CI->getType(); // B from above 243 244 // Get the opcodes of the two Cast instructions 245 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 246 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 247 Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ? 248 DL->getIntPtrType(SrcTy) : 0; 249 Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ? 250 DL->getIntPtrType(MidTy) : 0; 251 Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ? 252 DL->getIntPtrType(DstTy) : 0; 253 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 254 DstTy, SrcIntPtrTy, MidIntPtrTy, 255 DstIntPtrTy); 256 257 // We don't want to form an inttoptr or ptrtoint that converts to an integer 258 // type that differs from the pointer size. 259 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) || 260 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy)) 261 Res = 0; 262 263 return Instruction::CastOps(Res); 264} 265 266/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 267/// results in any code being generated and is interesting to optimize out. If 268/// the cast can be eliminated by some other simple transformation, we prefer 269/// to do the simplification first. 270bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 271 Type *Ty) { 272 // Noop casts and casts of constants should be eliminated trivially. 273 if (V->getType() == Ty || isa<Constant>(V)) return false; 274 275 // If this is another cast that can be eliminated, we prefer to have it 276 // eliminated. 277 if (const CastInst *CI = dyn_cast<CastInst>(V)) 278 if (isEliminableCastPair(CI, opc, Ty, DL)) 279 return false; 280 281 // If this is a vector sext from a compare, then we don't want to break the 282 // idiom where each element of the extended vector is either zero or all ones. 283 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 284 return false; 285 286 return true; 287} 288 289 290/// @brief Implement the transforms common to all CastInst visitors. 291Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 292 Value *Src = CI.getOperand(0); 293 294 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 295 // eliminate it now. 296 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 297 if (Instruction::CastOps opc = 298 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) { 299 // The first cast (CSrc) is eliminable so we need to fix up or replace 300 // the second cast (CI). CSrc will then have a good chance of being dead. 301 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 302 } 303 } 304 305 // If we are casting a select then fold the cast into the select 306 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 307 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 308 return NV; 309 310 // If we are casting a PHI then fold the cast into the PHI 311 if (isa<PHINode>(Src)) { 312 // We don't do this if this would create a PHI node with an illegal type if 313 // it is currently legal. 314 if (!Src->getType()->isIntegerTy() || 315 !CI.getType()->isIntegerTy() || 316 ShouldChangeType(CI.getType(), Src->getType())) 317 if (Instruction *NV = FoldOpIntoPhi(CI)) 318 return NV; 319 } 320 321 return 0; 322} 323 324/// CanEvaluateTruncated - Return true if we can evaluate the specified 325/// expression tree as type Ty instead of its larger type, and arrive with the 326/// same value. This is used by code that tries to eliminate truncates. 327/// 328/// Ty will always be a type smaller than V. We should return true if trunc(V) 329/// can be computed by computing V in the smaller type. If V is an instruction, 330/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 331/// makes sense if x and y can be efficiently truncated. 332/// 333/// This function works on both vectors and scalars. 334/// 335static bool CanEvaluateTruncated(Value *V, Type *Ty) { 336 // We can always evaluate constants in another type. 337 if (isa<Constant>(V)) 338 return true; 339 340 Instruction *I = dyn_cast<Instruction>(V); 341 if (!I) return false; 342 343 Type *OrigTy = V->getType(); 344 345 // If this is an extension from the dest type, we can eliminate it, even if it 346 // has multiple uses. 347 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 348 I->getOperand(0)->getType() == Ty) 349 return true; 350 351 // We can't extend or shrink something that has multiple uses: doing so would 352 // require duplicating the instruction in general, which isn't profitable. 353 if (!I->hasOneUse()) return false; 354 355 unsigned Opc = I->getOpcode(); 356 switch (Opc) { 357 case Instruction::Add: 358 case Instruction::Sub: 359 case Instruction::Mul: 360 case Instruction::And: 361 case Instruction::Or: 362 case Instruction::Xor: 363 // These operators can all arbitrarily be extended or truncated. 364 return CanEvaluateTruncated(I->getOperand(0), Ty) && 365 CanEvaluateTruncated(I->getOperand(1), Ty); 366 367 case Instruction::UDiv: 368 case Instruction::URem: { 369 // UDiv and URem can be truncated if all the truncated bits are zero. 370 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 371 uint32_t BitWidth = Ty->getScalarSizeInBits(); 372 if (BitWidth < OrigBitWidth) { 373 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 374 if (MaskedValueIsZero(I->getOperand(0), Mask) && 375 MaskedValueIsZero(I->getOperand(1), Mask)) { 376 return CanEvaluateTruncated(I->getOperand(0), Ty) && 377 CanEvaluateTruncated(I->getOperand(1), Ty); 378 } 379 } 380 break; 381 } 382 case Instruction::Shl: 383 // If we are truncating the result of this SHL, and if it's a shift of a 384 // constant amount, we can always perform a SHL in a smaller type. 385 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 386 uint32_t BitWidth = Ty->getScalarSizeInBits(); 387 if (CI->getLimitedValue(BitWidth) < BitWidth) 388 return CanEvaluateTruncated(I->getOperand(0), Ty); 389 } 390 break; 391 case Instruction::LShr: 392 // If this is a truncate of a logical shr, we can truncate it to a smaller 393 // lshr iff we know that the bits we would otherwise be shifting in are 394 // already zeros. 395 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 396 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 397 uint32_t BitWidth = Ty->getScalarSizeInBits(); 398 if (MaskedValueIsZero(I->getOperand(0), 399 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 400 CI->getLimitedValue(BitWidth) < BitWidth) { 401 return CanEvaluateTruncated(I->getOperand(0), Ty); 402 } 403 } 404 break; 405 case Instruction::Trunc: 406 // trunc(trunc(x)) -> trunc(x) 407 return true; 408 case Instruction::ZExt: 409 case Instruction::SExt: 410 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 411 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 412 return true; 413 case Instruction::Select: { 414 SelectInst *SI = cast<SelectInst>(I); 415 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 416 CanEvaluateTruncated(SI->getFalseValue(), Ty); 417 } 418 case Instruction::PHI: { 419 // We can change a phi if we can change all operands. Note that we never 420 // get into trouble with cyclic PHIs here because we only consider 421 // instructions with a single use. 422 PHINode *PN = cast<PHINode>(I); 423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 424 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 425 return false; 426 return true; 427 } 428 default: 429 // TODO: Can handle more cases here. 430 break; 431 } 432 433 return false; 434} 435 436Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 437 if (Instruction *Result = commonCastTransforms(CI)) 438 return Result; 439 440 // See if we can simplify any instructions used by the input whose sole 441 // purpose is to compute bits we don't care about. 442 if (SimplifyDemandedInstructionBits(CI)) 443 return &CI; 444 445 Value *Src = CI.getOperand(0); 446 Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 447 448 // Attempt to truncate the entire input expression tree to the destination 449 // type. Only do this if the dest type is a simple type, don't convert the 450 // expression tree to something weird like i93 unless the source is also 451 // strange. 452 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 453 CanEvaluateTruncated(Src, DestTy)) { 454 455 // If this cast is a truncate, evaluting in a different type always 456 // eliminates the cast, so it is always a win. 457 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 458 " to avoid cast: " << CI << '\n'); 459 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 460 assert(Res->getType() == DestTy); 461 return ReplaceInstUsesWith(CI, Res); 462 } 463 464 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 465 if (DestTy->getScalarSizeInBits() == 1) { 466 Constant *One = ConstantInt::get(Src->getType(), 1); 467 Src = Builder->CreateAnd(Src, One); 468 Value *Zero = Constant::getNullValue(Src->getType()); 469 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 470 } 471 472 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 473 Value *A = 0; ConstantInt *Cst = 0; 474 if (Src->hasOneUse() && 475 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 476 // We have three types to worry about here, the type of A, the source of 477 // the truncate (MidSize), and the destination of the truncate. We know that 478 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 479 // between ASize and ResultSize. 480 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 481 482 // If the shift amount is larger than the size of A, then the result is 483 // known to be zero because all the input bits got shifted out. 484 if (Cst->getZExtValue() >= ASize) 485 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType())); 486 487 // Since we're doing an lshr and a zero extend, and know that the shift 488 // amount is smaller than ASize, it is always safe to do the shift in A's 489 // type, then zero extend or truncate to the result. 490 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 491 Shift->takeName(Src); 492 return CastInst::CreateIntegerCast(Shift, CI.getType(), false); 493 } 494 495 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 496 // type isn't non-native. 497 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) && 498 ShouldChangeType(Src->getType(), CI.getType()) && 499 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 500 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr"); 501 return BinaryOperator::CreateAnd(NewTrunc, 502 ConstantExpr::getTrunc(Cst, CI.getType())); 503 } 504 505 return 0; 506} 507 508/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 509/// in order to eliminate the icmp. 510Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 511 bool DoXform) { 512 // If we are just checking for a icmp eq of a single bit and zext'ing it 513 // to an integer, then shift the bit to the appropriate place and then 514 // cast to integer to avoid the comparison. 515 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 516 const APInt &Op1CV = Op1C->getValue(); 517 518 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 519 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 520 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 521 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 522 if (!DoXform) return ICI; 523 524 Value *In = ICI->getOperand(0); 525 Value *Sh = ConstantInt::get(In->getType(), 526 In->getType()->getScalarSizeInBits()-1); 527 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 528 if (In->getType() != CI.getType()) 529 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); 530 531 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 532 Constant *One = ConstantInt::get(In->getType(), 1); 533 In = Builder->CreateXor(In, One, In->getName()+".not"); 534 } 535 536 return ReplaceInstUsesWith(CI, In); 537 } 538 539 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 540 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 541 // zext (X == 1) to i32 --> X iff X has only the low bit set. 542 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 543 // zext (X != 0) to i32 --> X iff X has only the low bit set. 544 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 545 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 546 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 547 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 548 // This only works for EQ and NE 549 ICI->isEquality()) { 550 // If Op1C some other power of two, convert: 551 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 552 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 553 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne); 554 555 APInt KnownZeroMask(~KnownZero); 556 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 557 if (!DoXform) return ICI; 558 559 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 560 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 561 // (X&4) == 2 --> false 562 // (X&4) != 2 --> true 563 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 564 isNE); 565 Res = ConstantExpr::getZExt(Res, CI.getType()); 566 return ReplaceInstUsesWith(CI, Res); 567 } 568 569 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 570 Value *In = ICI->getOperand(0); 571 if (ShiftAmt) { 572 // Perform a logical shr by shiftamt. 573 // Insert the shift to put the result in the low bit. 574 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 575 In->getName()+".lobit"); 576 } 577 578 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 579 Constant *One = ConstantInt::get(In->getType(), 1); 580 In = Builder->CreateXor(In, One); 581 } 582 583 if (CI.getType() == In->getType()) 584 return ReplaceInstUsesWith(CI, In); 585 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 586 } 587 } 588 } 589 590 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 591 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 592 // may lead to additional simplifications. 593 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 594 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 595 uint32_t BitWidth = ITy->getBitWidth(); 596 Value *LHS = ICI->getOperand(0); 597 Value *RHS = ICI->getOperand(1); 598 599 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 600 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 601 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS); 602 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS); 603 604 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 605 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 606 APInt UnknownBit = ~KnownBits; 607 if (UnknownBit.countPopulation() == 1) { 608 if (!DoXform) return ICI; 609 610 Value *Result = Builder->CreateXor(LHS, RHS); 611 612 // Mask off any bits that are set and won't be shifted away. 613 if (KnownOneLHS.uge(UnknownBit)) 614 Result = Builder->CreateAnd(Result, 615 ConstantInt::get(ITy, UnknownBit)); 616 617 // Shift the bit we're testing down to the lsb. 618 Result = Builder->CreateLShr( 619 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 620 621 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 622 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 623 Result->takeName(ICI); 624 return ReplaceInstUsesWith(CI, Result); 625 } 626 } 627 } 628 } 629 630 return 0; 631} 632 633/// CanEvaluateZExtd - Determine if the specified value can be computed in the 634/// specified wider type and produce the same low bits. If not, return false. 635/// 636/// If this function returns true, it can also return a non-zero number of bits 637/// (in BitsToClear) which indicates that the value it computes is correct for 638/// the zero extend, but that the additional BitsToClear bits need to be zero'd 639/// out. For example, to promote something like: 640/// 641/// %B = trunc i64 %A to i32 642/// %C = lshr i32 %B, 8 643/// %E = zext i32 %C to i64 644/// 645/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 646/// set to 8 to indicate that the promoted value needs to have bits 24-31 647/// cleared in addition to bits 32-63. Since an 'and' will be generated to 648/// clear the top bits anyway, doing this has no extra cost. 649/// 650/// This function works on both vectors and scalars. 651static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { 652 BitsToClear = 0; 653 if (isa<Constant>(V)) 654 return true; 655 656 Instruction *I = dyn_cast<Instruction>(V); 657 if (!I) return false; 658 659 // If the input is a truncate from the destination type, we can trivially 660 // eliminate it. 661 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 662 return true; 663 664 // We can't extend or shrink something that has multiple uses: doing so would 665 // require duplicating the instruction in general, which isn't profitable. 666 if (!I->hasOneUse()) return false; 667 668 unsigned Opc = I->getOpcode(), Tmp; 669 switch (Opc) { 670 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 671 case Instruction::SExt: // zext(sext(x)) -> sext(x). 672 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 673 return true; 674 case Instruction::And: 675 case Instruction::Or: 676 case Instruction::Xor: 677 case Instruction::Add: 678 case Instruction::Sub: 679 case Instruction::Mul: 680 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 681 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 682 return false; 683 // These can all be promoted if neither operand has 'bits to clear'. 684 if (BitsToClear == 0 && Tmp == 0) 685 return true; 686 687 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 688 // other side, BitsToClear is ok. 689 if (Tmp == 0 && 690 (Opc == Instruction::And || Opc == Instruction::Or || 691 Opc == Instruction::Xor)) { 692 // We use MaskedValueIsZero here for generality, but the case we care 693 // about the most is constant RHS. 694 unsigned VSize = V->getType()->getScalarSizeInBits(); 695 if (MaskedValueIsZero(I->getOperand(1), 696 APInt::getHighBitsSet(VSize, BitsToClear))) 697 return true; 698 } 699 700 // Otherwise, we don't know how to analyze this BitsToClear case yet. 701 return false; 702 703 case Instruction::Shl: 704 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the 705 // upper bits we can reduce BitsToClear by the shift amount. 706 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 707 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 708 return false; 709 uint64_t ShiftAmt = Amt->getZExtValue(); 710 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0; 711 return true; 712 } 713 return false; 714 case Instruction::LShr: 715 // We can promote lshr(x, cst) if we can promote x. This requires the 716 // ultimate 'and' to clear out the high zero bits we're clearing out though. 717 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 718 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 719 return false; 720 BitsToClear += Amt->getZExtValue(); 721 if (BitsToClear > V->getType()->getScalarSizeInBits()) 722 BitsToClear = V->getType()->getScalarSizeInBits(); 723 return true; 724 } 725 // Cannot promote variable LSHR. 726 return false; 727 case Instruction::Select: 728 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 729 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 730 // TODO: If important, we could handle the case when the BitsToClear are 731 // known zero in the disagreeing side. 732 Tmp != BitsToClear) 733 return false; 734 return true; 735 736 case Instruction::PHI: { 737 // We can change a phi if we can change all operands. Note that we never 738 // get into trouble with cyclic PHIs here because we only consider 739 // instructions with a single use. 740 PHINode *PN = cast<PHINode>(I); 741 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 742 return false; 743 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 744 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 745 // TODO: If important, we could handle the case when the BitsToClear 746 // are known zero in the disagreeing input. 747 Tmp != BitsToClear) 748 return false; 749 return true; 750 } 751 default: 752 // TODO: Can handle more cases here. 753 return false; 754 } 755} 756 757Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 758 // If this zero extend is only used by a truncate, let the truncate be 759 // eliminated before we try to optimize this zext. 760 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back())) 761 return 0; 762 763 // If one of the common conversion will work, do it. 764 if (Instruction *Result = commonCastTransforms(CI)) 765 return Result; 766 767 // See if we can simplify any instructions used by the input whose sole 768 // purpose is to compute bits we don't care about. 769 if (SimplifyDemandedInstructionBits(CI)) 770 return &CI; 771 772 Value *Src = CI.getOperand(0); 773 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 774 775 // Attempt to extend the entire input expression tree to the destination 776 // type. Only do this if the dest type is a simple type, don't convert the 777 // expression tree to something weird like i93 unless the source is also 778 // strange. 779 unsigned BitsToClear; 780 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 781 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 782 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 783 "Unreasonable BitsToClear"); 784 785 // Okay, we can transform this! Insert the new expression now. 786 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 787 " to avoid zero extend: " << CI); 788 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 789 assert(Res->getType() == DestTy); 790 791 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 792 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 793 794 // If the high bits are already filled with zeros, just replace this 795 // cast with the result. 796 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 797 DestBitSize-SrcBitsKept))) 798 return ReplaceInstUsesWith(CI, Res); 799 800 // We need to emit an AND to clear the high bits. 801 Constant *C = ConstantInt::get(Res->getType(), 802 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 803 return BinaryOperator::CreateAnd(Res, C); 804 } 805 806 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 807 // types and if the sizes are just right we can convert this into a logical 808 // 'and' which will be much cheaper than the pair of casts. 809 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 810 // TODO: Subsume this into EvaluateInDifferentType. 811 812 // Get the sizes of the types involved. We know that the intermediate type 813 // will be smaller than A or C, but don't know the relation between A and C. 814 Value *A = CSrc->getOperand(0); 815 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 816 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 817 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 818 // If we're actually extending zero bits, then if 819 // SrcSize < DstSize: zext(a & mask) 820 // SrcSize == DstSize: a & mask 821 // SrcSize > DstSize: trunc(a) & mask 822 if (SrcSize < DstSize) { 823 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 824 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 825 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 826 return new ZExtInst(And, CI.getType()); 827 } 828 829 if (SrcSize == DstSize) { 830 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 831 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 832 AndValue)); 833 } 834 if (SrcSize > DstSize) { 835 Value *Trunc = Builder->CreateTrunc(A, CI.getType()); 836 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 837 return BinaryOperator::CreateAnd(Trunc, 838 ConstantInt::get(Trunc->getType(), 839 AndValue)); 840 } 841 } 842 843 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 844 return transformZExtICmp(ICI, CI); 845 846 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 847 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 848 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 849 // of the (zext icmp) will be transformed. 850 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 851 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 852 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 853 (transformZExtICmp(LHS, CI, false) || 854 transformZExtICmp(RHS, CI, false))) { 855 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 856 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 857 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 858 } 859 } 860 861 // zext(trunc(X) & C) -> (X & zext(C)). 862 Constant *C; 863 Value *X; 864 if (SrcI && 865 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) && 866 X->getType() == CI.getType()) 867 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType())); 868 869 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)). 870 Value *And; 871 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) && 872 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) && 873 X->getType() == CI.getType()) { 874 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 875 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC); 876 } 877 878 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 879 if (SrcI && SrcI->hasOneUse() && 880 SrcI->getType()->getScalarType()->isIntegerTy(1) && 881 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) { 882 Value *New = Builder->CreateZExt(X, CI.getType()); 883 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 884 } 885 886 return 0; 887} 888 889/// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations 890/// in order to eliminate the icmp. 891Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { 892 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); 893 ICmpInst::Predicate Pred = ICI->getPredicate(); 894 895 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 896 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative 897 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive 898 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) || 899 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { 900 901 Value *Sh = ConstantInt::get(Op0->getType(), 902 Op0->getType()->getScalarSizeInBits()-1); 903 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); 904 if (In->getType() != CI.getType()) 905 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); 906 907 if (Pred == ICmpInst::ICMP_SGT) 908 In = Builder->CreateNot(In, In->getName()+".not"); 909 return ReplaceInstUsesWith(CI, In); 910 } 911 } 912 913 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 914 // If we know that only one bit of the LHS of the icmp can be set and we 915 // have an equality comparison with zero or a power of 2, we can transform 916 // the icmp and sext into bitwise/integer operations. 917 if (ICI->hasOneUse() && 918 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ 919 unsigned BitWidth = Op1C->getType()->getBitWidth(); 920 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 921 ComputeMaskedBits(Op0, KnownZero, KnownOne); 922 923 APInt KnownZeroMask(~KnownZero); 924 if (KnownZeroMask.isPowerOf2()) { 925 Value *In = ICI->getOperand(0); 926 927 // If the icmp tests for a known zero bit we can constant fold it. 928 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) { 929 Value *V = Pred == ICmpInst::ICMP_NE ? 930 ConstantInt::getAllOnesValue(CI.getType()) : 931 ConstantInt::getNullValue(CI.getType()); 932 return ReplaceInstUsesWith(CI, V); 933 } 934 935 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) { 936 // sext ((x & 2^n) == 0) -> (x >> n) - 1 937 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1 938 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros(); 939 // Perform a right shift to place the desired bit in the LSB. 940 if (ShiftAmt) 941 In = Builder->CreateLShr(In, 942 ConstantInt::get(In->getType(), ShiftAmt)); 943 944 // At this point "In" is either 1 or 0. Subtract 1 to turn 945 // {1, 0} -> {0, -1}. 946 In = Builder->CreateAdd(In, 947 ConstantInt::getAllOnesValue(In->getType()), 948 "sext"); 949 } else { 950 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1 951 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1 952 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros(); 953 // Perform a left shift to place the desired bit in the MSB. 954 if (ShiftAmt) 955 In = Builder->CreateShl(In, 956 ConstantInt::get(In->getType(), ShiftAmt)); 957 958 // Distribute the bit over the whole bit width. 959 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(), 960 BitWidth - 1), "sext"); 961 } 962 963 if (CI.getType() == In->getType()) 964 return ReplaceInstUsesWith(CI, In); 965 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/); 966 } 967 } 968 } 969 970 return 0; 971} 972 973/// CanEvaluateSExtd - Return true if we can take the specified value 974/// and return it as type Ty without inserting any new casts and without 975/// changing the value of the common low bits. This is used by code that tries 976/// to promote integer operations to a wider types will allow us to eliminate 977/// the extension. 978/// 979/// This function works on both vectors and scalars. 980/// 981static bool CanEvaluateSExtd(Value *V, Type *Ty) { 982 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 983 "Can't sign extend type to a smaller type"); 984 // If this is a constant, it can be trivially promoted. 985 if (isa<Constant>(V)) 986 return true; 987 988 Instruction *I = dyn_cast<Instruction>(V); 989 if (!I) return false; 990 991 // If this is a truncate from the dest type, we can trivially eliminate it. 992 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 993 return true; 994 995 // We can't extend or shrink something that has multiple uses: doing so would 996 // require duplicating the instruction in general, which isn't profitable. 997 if (!I->hasOneUse()) return false; 998 999 switch (I->getOpcode()) { 1000 case Instruction::SExt: // sext(sext(x)) -> sext(x) 1001 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 1002 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 1003 return true; 1004 case Instruction::And: 1005 case Instruction::Or: 1006 case Instruction::Xor: 1007 case Instruction::Add: 1008 case Instruction::Sub: 1009 case Instruction::Mul: 1010 // These operators can all arbitrarily be extended if their inputs can. 1011 return CanEvaluateSExtd(I->getOperand(0), Ty) && 1012 CanEvaluateSExtd(I->getOperand(1), Ty); 1013 1014 //case Instruction::Shl: TODO 1015 //case Instruction::LShr: TODO 1016 1017 case Instruction::Select: 1018 return CanEvaluateSExtd(I->getOperand(1), Ty) && 1019 CanEvaluateSExtd(I->getOperand(2), Ty); 1020 1021 case Instruction::PHI: { 1022 // We can change a phi if we can change all operands. Note that we never 1023 // get into trouble with cyclic PHIs here because we only consider 1024 // instructions with a single use. 1025 PHINode *PN = cast<PHINode>(I); 1026 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1027 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 1028 return true; 1029 } 1030 default: 1031 // TODO: Can handle more cases here. 1032 break; 1033 } 1034 1035 return false; 1036} 1037 1038Instruction *InstCombiner::visitSExt(SExtInst &CI) { 1039 // If this sign extend is only used by a truncate, let the truncate be 1040 // eliminated before we try to optimize this sext. 1041 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back())) 1042 return 0; 1043 1044 if (Instruction *I = commonCastTransforms(CI)) 1045 return I; 1046 1047 // See if we can simplify any instructions used by the input whose sole 1048 // purpose is to compute bits we don't care about. 1049 if (SimplifyDemandedInstructionBits(CI)) 1050 return &CI; 1051 1052 Value *Src = CI.getOperand(0); 1053 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 1054 1055 // Attempt to extend the entire input expression tree to the destination 1056 // type. Only do this if the dest type is a simple type, don't convert the 1057 // expression tree to something weird like i93 unless the source is also 1058 // strange. 1059 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 1060 CanEvaluateSExtd(Src, DestTy)) { 1061 // Okay, we can transform this! Insert the new expression now. 1062 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 1063 " to avoid sign extend: " << CI); 1064 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 1065 assert(Res->getType() == DestTy); 1066 1067 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1068 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1069 1070 // If the high bits are already filled with sign bit, just replace this 1071 // cast with the result. 1072 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 1073 return ReplaceInstUsesWith(CI, Res); 1074 1075 // We need to emit a shl + ashr to do the sign extend. 1076 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1077 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 1078 ShAmt); 1079 } 1080 1081 // If this input is a trunc from our destination, then turn sext(trunc(x)) 1082 // into shifts. 1083 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 1084 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 1085 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1086 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1087 1088 // We need to emit a shl + ashr to do the sign extend. 1089 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1090 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1091 return BinaryOperator::CreateAShr(Res, ShAmt); 1092 } 1093 1094 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 1095 return transformSExtICmp(ICI, CI); 1096 1097 // If the input is a shl/ashr pair of a same constant, then this is a sign 1098 // extension from a smaller value. If we could trust arbitrary bitwidth 1099 // integers, we could turn this into a truncate to the smaller bit and then 1100 // use a sext for the whole extension. Since we don't, look deeper and check 1101 // for a truncate. If the source and dest are the same type, eliminate the 1102 // trunc and extend and just do shifts. For example, turn: 1103 // %a = trunc i32 %i to i8 1104 // %b = shl i8 %a, 6 1105 // %c = ashr i8 %b, 6 1106 // %d = sext i8 %c to i32 1107 // into: 1108 // %a = shl i32 %i, 30 1109 // %d = ashr i32 %a, 30 1110 Value *A = 0; 1111 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1112 ConstantInt *BA = 0, *CA = 0; 1113 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1114 m_ConstantInt(CA))) && 1115 BA == CA && A->getType() == CI.getType()) { 1116 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1117 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1118 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1119 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1120 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1121 return BinaryOperator::CreateAShr(A, ShAmtV); 1122 } 1123 1124 return 0; 1125} 1126 1127 1128/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1129/// in the specified FP type without changing its value. 1130static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1131 bool losesInfo; 1132 APFloat F = CFP->getValueAPF(); 1133 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1134 if (!losesInfo) 1135 return ConstantFP::get(CFP->getContext(), F); 1136 return 0; 1137} 1138 1139/// LookThroughFPExtensions - If this is an fp extension instruction, look 1140/// through it until we get the source value. 1141static Value *LookThroughFPExtensions(Value *V) { 1142 if (Instruction *I = dyn_cast<Instruction>(V)) 1143 if (I->getOpcode() == Instruction::FPExt) 1144 return LookThroughFPExtensions(I->getOperand(0)); 1145 1146 // If this value is a constant, return the constant in the smallest FP type 1147 // that can accurately represent it. This allows us to turn 1148 // (float)((double)X+2.0) into x+2.0f. 1149 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1150 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1151 return V; // No constant folding of this. 1152 // See if the value can be truncated to half and then reextended. 1153 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf)) 1154 return V; 1155 // See if the value can be truncated to float and then reextended. 1156 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1157 return V; 1158 if (CFP->getType()->isDoubleTy()) 1159 return V; // Won't shrink. 1160 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1161 return V; 1162 // Don't try to shrink to various long double types. 1163 } 1164 1165 return V; 1166} 1167 1168Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1169 if (Instruction *I = commonCastTransforms(CI)) 1170 return I; 1171 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to 1172 // simpilify this expression to avoid one or more of the trunc/extend 1173 // operations if we can do so without changing the numerical results. 1174 // 1175 // The exact manner in which the widths of the operands interact to limit 1176 // what we can and cannot do safely varies from operation to operation, and 1177 // is explained below in the various case statements. 1178 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1179 if (OpI && OpI->hasOneUse()) { 1180 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0)); 1181 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1)); 1182 unsigned OpWidth = OpI->getType()->getFPMantissaWidth(); 1183 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth(); 1184 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth(); 1185 unsigned SrcWidth = std::max(LHSWidth, RHSWidth); 1186 unsigned DstWidth = CI.getType()->getFPMantissaWidth(); 1187 switch (OpI->getOpcode()) { 1188 default: break; 1189 case Instruction::FAdd: 1190 case Instruction::FSub: 1191 // For addition and subtraction, the infinitely precise result can 1192 // essentially be arbitrarily wide; proving that double rounding 1193 // will not occur because the result of OpI is exact (as we will for 1194 // FMul, for example) is hopeless. However, we *can* nonetheless 1195 // frequently know that double rounding cannot occur (or that it is 1196 // innocuous) by taking advantage of the specific structure of 1197 // infinitely-precise results that admit double rounding. 1198 // 1199 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient 1200 // to represent both sources, we can guarantee that the double 1201 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis, 1202 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..." 1203 // for proof of this fact). 1204 // 1205 // Note: Figueroa does not consider the case where DstFormat != 1206 // SrcFormat. It's possible (likely even!) that this analysis 1207 // could be tightened for those cases, but they are rare (the main 1208 // case of interest here is (float)((double)float + float)). 1209 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) { 1210 if (LHSOrig->getType() != CI.getType()) 1211 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1212 if (RHSOrig->getType() != CI.getType()) 1213 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1214 Instruction *RI = 1215 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig); 1216 RI->copyFastMathFlags(OpI); 1217 return RI; 1218 } 1219 break; 1220 case Instruction::FMul: 1221 // For multiplication, the infinitely precise result has at most 1222 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient 1223 // that such a value can be exactly represented, then no double 1224 // rounding can possibly occur; we can safely perform the operation 1225 // in the destination format if it can represent both sources. 1226 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) { 1227 if (LHSOrig->getType() != CI.getType()) 1228 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1229 if (RHSOrig->getType() != CI.getType()) 1230 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1231 Instruction *RI = 1232 BinaryOperator::CreateFMul(LHSOrig, RHSOrig); 1233 RI->copyFastMathFlags(OpI); 1234 return RI; 1235 } 1236 break; 1237 case Instruction::FDiv: 1238 // For division, we use again use the bound from Figueroa's 1239 // dissertation. I am entirely certain that this bound can be 1240 // tightened in the unbalanced operand case by an analysis based on 1241 // the diophantine rational approximation bound, but the well-known 1242 // condition used here is a good conservative first pass. 1243 // TODO: Tighten bound via rigorous analysis of the unbalanced case. 1244 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) { 1245 if (LHSOrig->getType() != CI.getType()) 1246 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1247 if (RHSOrig->getType() != CI.getType()) 1248 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1249 Instruction *RI = 1250 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig); 1251 RI->copyFastMathFlags(OpI); 1252 return RI; 1253 } 1254 break; 1255 case Instruction::FRem: 1256 // Remainder is straightforward. Remainder is always exact, so the 1257 // type of OpI doesn't enter into things at all. We simply evaluate 1258 // in whichever source type is larger, then convert to the 1259 // destination type. 1260 if (LHSWidth < SrcWidth) 1261 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType()); 1262 else if (RHSWidth <= SrcWidth) 1263 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType()); 1264 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig); 1265 if (Instruction *RI = dyn_cast<Instruction>(ExactResult)) 1266 RI->copyFastMathFlags(OpI); 1267 return CastInst::CreateFPCast(ExactResult, CI.getType()); 1268 } 1269 1270 // (fptrunc (fneg x)) -> (fneg (fptrunc x)) 1271 if (BinaryOperator::isFNeg(OpI)) { 1272 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1), 1273 CI.getType()); 1274 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc); 1275 RI->copyFastMathFlags(OpI); 1276 return RI; 1277 } 1278 } 1279 1280 // (fptrunc (select cond, R1, Cst)) --> 1281 // (select cond, (fptrunc R1), (fptrunc Cst)) 1282 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)); 1283 if (SI && 1284 (isa<ConstantFP>(SI->getOperand(1)) || 1285 isa<ConstantFP>(SI->getOperand(2)))) { 1286 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1), 1287 CI.getType()); 1288 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2), 1289 CI.getType()); 1290 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc); 1291 } 1292 1293 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0)); 1294 if (II) { 1295 switch (II->getIntrinsicID()) { 1296 default: break; 1297 case Intrinsic::fabs: { 1298 // (fptrunc (fabs x)) -> (fabs (fptrunc x)) 1299 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0), 1300 CI.getType()); 1301 Type *IntrinsicType[] = { CI.getType() }; 1302 Function *Overload = 1303 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(), 1304 II->getIntrinsicID(), IntrinsicType); 1305 1306 Value *Args[] = { InnerTrunc }; 1307 return CallInst::Create(Overload, Args, II->getName()); 1308 } 1309 } 1310 } 1311 1312 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1313 // Note that we restrict this transformation based on 1314 // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because 1315 // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the 1316 // single-precision intrinsic can be expanded in the backend. 1317 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1318 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) && 1319 (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) || 1320 Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) && 1321 Call->getNumArgOperands() == 1 && 1322 Call->hasOneUse()) { 1323 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1324 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1325 CI.getType()->isFloatTy() && 1326 Call->getType()->isDoubleTy() && 1327 Arg->getType()->isDoubleTy() && 1328 Arg->getOperand(0)->getType()->isFloatTy()) { 1329 Function *Callee = Call->getCalledFunction(); 1330 Module *M = CI.getParent()->getParent()->getParent(); 1331 Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ? 1332 Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) : 1333 M->getOrInsertFunction("sqrtf", Callee->getAttributes(), 1334 Builder->getFloatTy(), Builder->getFloatTy(), 1335 NULL); 1336 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1337 "sqrtfcall"); 1338 ret->setAttributes(Callee->getAttributes()); 1339 1340 1341 // Remove the old Call. With -fmath-errno, it won't get marked readnone. 1342 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType())); 1343 EraseInstFromFunction(*Call); 1344 return ret; 1345 } 1346 } 1347 1348 return 0; 1349} 1350 1351Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1352 return commonCastTransforms(CI); 1353} 1354 1355Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1356 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1357 if (OpI == 0) 1358 return commonCastTransforms(FI); 1359 1360 // fptoui(uitofp(X)) --> X 1361 // fptoui(sitofp(X)) --> X 1362 // This is safe if the intermediate type has enough bits in its mantissa to 1363 // accurately represent all values of X. For example, do not do this with 1364 // i64->float->i64. This is also safe for sitofp case, because any negative 1365 // 'X' value would cause an undefined result for the fptoui. 1366 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1367 OpI->getOperand(0)->getType() == FI.getType() && 1368 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1369 OpI->getType()->getFPMantissaWidth()) 1370 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1371 1372 return commonCastTransforms(FI); 1373} 1374 1375Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1376 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1377 if (OpI == 0) 1378 return commonCastTransforms(FI); 1379 1380 // fptosi(sitofp(X)) --> X 1381 // fptosi(uitofp(X)) --> X 1382 // This is safe if the intermediate type has enough bits in its mantissa to 1383 // accurately represent all values of X. For example, do not do this with 1384 // i64->float->i64. This is also safe for sitofp case, because any negative 1385 // 'X' value would cause an undefined result for the fptoui. 1386 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1387 OpI->getOperand(0)->getType() == FI.getType() && 1388 (int)FI.getType()->getScalarSizeInBits() <= 1389 OpI->getType()->getFPMantissaWidth()) 1390 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1391 1392 return commonCastTransforms(FI); 1393} 1394 1395Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1396 return commonCastTransforms(CI); 1397} 1398 1399Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1400 return commonCastTransforms(CI); 1401} 1402 1403Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1404 // If the source integer type is not the intptr_t type for this target, do a 1405 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1406 // cast to be exposed to other transforms. 1407 1408 if (DL) { 1409 unsigned AS = CI.getAddressSpace(); 1410 if (CI.getOperand(0)->getType()->getScalarSizeInBits() != 1411 DL->getPointerSizeInBits(AS)) { 1412 Type *Ty = DL->getIntPtrType(CI.getContext(), AS); 1413 if (CI.getType()->isVectorTy()) // Handle vectors of pointers. 1414 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements()); 1415 1416 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty); 1417 return new IntToPtrInst(P, CI.getType()); 1418 } 1419 } 1420 1421 if (Instruction *I = commonCastTransforms(CI)) 1422 return I; 1423 1424 return 0; 1425} 1426 1427/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1428Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1429 Value *Src = CI.getOperand(0); 1430 1431 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1432 // If casting the result of a getelementptr instruction with no offset, turn 1433 // this into a cast of the original pointer! 1434 if (GEP->hasAllZeroIndices()) { 1435 // Changing the cast operand is usually not a good idea but it is safe 1436 // here because the pointer operand is being replaced with another 1437 // pointer operand so the opcode doesn't need to change. 1438 Worklist.Add(GEP); 1439 CI.setOperand(0, GEP->getOperand(0)); 1440 return &CI; 1441 } 1442 1443 if (!DL) 1444 return commonCastTransforms(CI); 1445 1446 // If the GEP has a single use, and the base pointer is a bitcast, and the 1447 // GEP computes a constant offset, see if we can convert these three 1448 // instructions into fewer. This typically happens with unions and other 1449 // non-type-safe code. 1450 unsigned AS = GEP->getPointerAddressSpace(); 1451 unsigned OffsetBits = DL->getPointerSizeInBits(AS); 1452 APInt Offset(OffsetBits, 0); 1453 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0)); 1454 if (GEP->hasOneUse() && 1455 BCI && 1456 GEP->accumulateConstantOffset(*DL, Offset)) { 1457 // Get the base pointer input of the bitcast, and the type it points to. 1458 Value *OrigBase = BCI->getOperand(0); 1459 SmallVector<Value*, 8> NewIndices; 1460 if (FindElementAtOffset(OrigBase->getType(), 1461 Offset.getSExtValue(), 1462 NewIndices)) { 1463 // If we were able to index down into an element, create the GEP 1464 // and bitcast the result. This eliminates one bitcast, potentially 1465 // two. 1466 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1467 Builder->CreateInBoundsGEP(OrigBase, NewIndices) : 1468 Builder->CreateGEP(OrigBase, NewIndices); 1469 NGEP->takeName(GEP); 1470 1471 if (isa<BitCastInst>(CI)) 1472 return new BitCastInst(NGEP, CI.getType()); 1473 assert(isa<PtrToIntInst>(CI)); 1474 return new PtrToIntInst(NGEP, CI.getType()); 1475 } 1476 } 1477 } 1478 1479 return commonCastTransforms(CI); 1480} 1481 1482Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1483 // If the destination integer type is not the intptr_t type for this target, 1484 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1485 // to be exposed to other transforms. 1486 1487 if (!DL) 1488 return commonPointerCastTransforms(CI); 1489 1490 Type *Ty = CI.getType(); 1491 unsigned AS = CI.getPointerAddressSpace(); 1492 1493 if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS)) 1494 return commonPointerCastTransforms(CI); 1495 1496 Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS); 1497 if (Ty->isVectorTy()) // Handle vectors of pointers. 1498 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements()); 1499 1500 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy); 1501 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false); 1502} 1503 1504/// OptimizeVectorResize - This input value (which is known to have vector type) 1505/// is being zero extended or truncated to the specified vector type. Try to 1506/// replace it with a shuffle (and vector/vector bitcast) if possible. 1507/// 1508/// The source and destination vector types may have different element types. 1509static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, 1510 InstCombiner &IC) { 1511 // We can only do this optimization if the output is a multiple of the input 1512 // element size, or the input is a multiple of the output element size. 1513 // Convert the input type to have the same element type as the output. 1514 VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1515 1516 if (SrcTy->getElementType() != DestTy->getElementType()) { 1517 // The input types don't need to be identical, but for now they must be the 1518 // same size. There is no specific reason we couldn't handle things like 1519 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1520 // there yet. 1521 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1522 DestTy->getElementType()->getPrimitiveSizeInBits()) 1523 return 0; 1524 1525 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1526 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1527 } 1528 1529 // Now that the element types match, get the shuffle mask and RHS of the 1530 // shuffle to use, which depends on whether we're increasing or decreasing the 1531 // size of the input. 1532 SmallVector<uint32_t, 16> ShuffleMask; 1533 Value *V2; 1534 1535 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1536 // If we're shrinking the number of elements, just shuffle in the low 1537 // elements from the input and use undef as the second shuffle input. 1538 V2 = UndefValue::get(SrcTy); 1539 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1540 ShuffleMask.push_back(i); 1541 1542 } else { 1543 // If we're increasing the number of elements, shuffle in all of the 1544 // elements from InVal and fill the rest of the result elements with zeros 1545 // from a constant zero. 1546 V2 = Constant::getNullValue(SrcTy); 1547 unsigned SrcElts = SrcTy->getNumElements(); 1548 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1549 ShuffleMask.push_back(i); 1550 1551 // The excess elements reference the first element of the zero input. 1552 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i) 1553 ShuffleMask.push_back(SrcElts); 1554 } 1555 1556 return new ShuffleVectorInst(InVal, V2, 1557 ConstantDataVector::get(V2->getContext(), 1558 ShuffleMask)); 1559} 1560 1561static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { 1562 return Value % Ty->getPrimitiveSizeInBits() == 0; 1563} 1564 1565static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { 1566 return Value / Ty->getPrimitiveSizeInBits(); 1567} 1568 1569/// CollectInsertionElements - V is a value which is inserted into a vector of 1570/// VecEltTy. Look through the value to see if we can decompose it into 1571/// insertions into the vector. See the example in the comment for 1572/// OptimizeIntegerToVectorInsertions for the pattern this handles. 1573/// The type of V is always a non-zero multiple of VecEltTy's size. 1574/// Shift is the number of bits between the lsb of V and the lsb of 1575/// the vector. 1576/// 1577/// This returns false if the pattern can't be matched or true if it can, 1578/// filling in Elements with the elements found here. 1579static bool CollectInsertionElements(Value *V, unsigned Shift, 1580 SmallVectorImpl<Value*> &Elements, 1581 Type *VecEltTy, InstCombiner &IC) { 1582 assert(isMultipleOfTypeSize(Shift, VecEltTy) && 1583 "Shift should be a multiple of the element type size"); 1584 1585 // Undef values never contribute useful bits to the result. 1586 if (isa<UndefValue>(V)) return true; 1587 1588 // If we got down to a value of the right type, we win, try inserting into the 1589 // right element. 1590 if (V->getType() == VecEltTy) { 1591 // Inserting null doesn't actually insert any elements. 1592 if (Constant *C = dyn_cast<Constant>(V)) 1593 if (C->isNullValue()) 1594 return true; 1595 1596 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy); 1597 if (IC.getDataLayout()->isBigEndian()) 1598 ElementIndex = Elements.size() - ElementIndex - 1; 1599 1600 // Fail if multiple elements are inserted into this slot. 1601 if (Elements[ElementIndex] != 0) 1602 return false; 1603 1604 Elements[ElementIndex] = V; 1605 return true; 1606 } 1607 1608 if (Constant *C = dyn_cast<Constant>(V)) { 1609 // Figure out the # elements this provides, and bitcast it or slice it up 1610 // as required. 1611 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1612 VecEltTy); 1613 // If the constant is the size of a vector element, we just need to bitcast 1614 // it to the right type so it gets properly inserted. 1615 if (NumElts == 1) 1616 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1617 Shift, Elements, VecEltTy, IC); 1618 1619 // Okay, this is a constant that covers multiple elements. Slice it up into 1620 // pieces and insert each element-sized piece into the vector. 1621 if (!isa<IntegerType>(C->getType())) 1622 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1623 C->getType()->getPrimitiveSizeInBits())); 1624 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1625 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1626 1627 for (unsigned i = 0; i != NumElts; ++i) { 1628 unsigned ShiftI = Shift+i*ElementSize; 1629 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1630 ShiftI)); 1631 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1632 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC)) 1633 return false; 1634 } 1635 return true; 1636 } 1637 1638 if (!V->hasOneUse()) return false; 1639 1640 Instruction *I = dyn_cast<Instruction>(V); 1641 if (I == 0) return false; 1642 switch (I->getOpcode()) { 1643 default: return false; // Unhandled case. 1644 case Instruction::BitCast: 1645 return CollectInsertionElements(I->getOperand(0), Shift, 1646 Elements, VecEltTy, IC); 1647 case Instruction::ZExt: 1648 if (!isMultipleOfTypeSize( 1649 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1650 VecEltTy)) 1651 return false; 1652 return CollectInsertionElements(I->getOperand(0), Shift, 1653 Elements, VecEltTy, IC); 1654 case Instruction::Or: 1655 return CollectInsertionElements(I->getOperand(0), Shift, 1656 Elements, VecEltTy, IC) && 1657 CollectInsertionElements(I->getOperand(1), Shift, 1658 Elements, VecEltTy, IC); 1659 case Instruction::Shl: { 1660 // Must be shifting by a constant that is a multiple of the element size. 1661 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1662 if (CI == 0) return false; 1663 Shift += CI->getZExtValue(); 1664 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false; 1665 return CollectInsertionElements(I->getOperand(0), Shift, 1666 Elements, VecEltTy, IC); 1667 } 1668 1669 } 1670} 1671 1672 1673/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1674/// may be doing shifts and ors to assemble the elements of the vector manually. 1675/// Try to rip the code out and replace it with insertelements. This is to 1676/// optimize code like this: 1677/// 1678/// %tmp37 = bitcast float %inc to i32 1679/// %tmp38 = zext i32 %tmp37 to i64 1680/// %tmp31 = bitcast float %inc5 to i32 1681/// %tmp32 = zext i32 %tmp31 to i64 1682/// %tmp33 = shl i64 %tmp32, 32 1683/// %ins35 = or i64 %tmp33, %tmp38 1684/// %tmp43 = bitcast i64 %ins35 to <2 x float> 1685/// 1686/// Into two insertelements that do "buildvector{%inc, %inc5}". 1687static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1688 InstCombiner &IC) { 1689 // We need to know the target byte order to perform this optimization. 1690 if (!IC.getDataLayout()) return 0; 1691 1692 VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1693 Value *IntInput = CI.getOperand(0); 1694 1695 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1696 if (!CollectInsertionElements(IntInput, 0, Elements, 1697 DestVecTy->getElementType(), IC)) 1698 return 0; 1699 1700 // If we succeeded, we know that all of the element are specified by Elements 1701 // or are zero if Elements has a null entry. Recast this as a set of 1702 // insertions. 1703 Value *Result = Constant::getNullValue(CI.getType()); 1704 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1705 if (Elements[i] == 0) continue; // Unset element. 1706 1707 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1708 IC.Builder->getInt32(i)); 1709 } 1710 1711 return Result; 1712} 1713 1714 1715/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1716/// bitcast. The various long double bitcasts can't get in here. 1717static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1718 // We need to know the target byte order to perform this optimization. 1719 if (!IC.getDataLayout()) return 0; 1720 1721 Value *Src = CI.getOperand(0); 1722 Type *DestTy = CI.getType(); 1723 1724 // If this is a bitcast from int to float, check to see if the int is an 1725 // extraction from a vector. 1726 Value *VecInput = 0; 1727 // bitcast(trunc(bitcast(somevector))) 1728 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1729 isa<VectorType>(VecInput->getType())) { 1730 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1731 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1732 1733 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1734 // If the element type of the vector doesn't match the result type, 1735 // bitcast it to be a vector type we can extract from. 1736 if (VecTy->getElementType() != DestTy) { 1737 VecTy = VectorType::get(DestTy, 1738 VecTy->getPrimitiveSizeInBits() / DestWidth); 1739 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1740 } 1741 1742 unsigned Elt = 0; 1743 if (IC.getDataLayout()->isBigEndian()) 1744 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1; 1745 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1746 } 1747 } 1748 1749 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1750 ConstantInt *ShAmt = 0; 1751 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1752 m_ConstantInt(ShAmt)))) && 1753 isa<VectorType>(VecInput->getType())) { 1754 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1755 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1756 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1757 ShAmt->getZExtValue() % DestWidth == 0) { 1758 // If the element type of the vector doesn't match the result type, 1759 // bitcast it to be a vector type we can extract from. 1760 if (VecTy->getElementType() != DestTy) { 1761 VecTy = VectorType::get(DestTy, 1762 VecTy->getPrimitiveSizeInBits() / DestWidth); 1763 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1764 } 1765 1766 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1767 if (IC.getDataLayout()->isBigEndian()) 1768 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt; 1769 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1770 } 1771 } 1772 return 0; 1773} 1774 1775Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1776 // If the operands are integer typed then apply the integer transforms, 1777 // otherwise just apply the common ones. 1778 Value *Src = CI.getOperand(0); 1779 Type *SrcTy = Src->getType(); 1780 Type *DestTy = CI.getType(); 1781 1782 // Get rid of casts from one type to the same type. These are useless and can 1783 // be replaced by the operand. 1784 if (DestTy == Src->getType()) 1785 return ReplaceInstUsesWith(CI, Src); 1786 1787 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1788 PointerType *SrcPTy = cast<PointerType>(SrcTy); 1789 Type *DstElTy = DstPTy->getElementType(); 1790 Type *SrcElTy = SrcPTy->getElementType(); 1791 1792 // If we are casting a alloca to a pointer to a type of the same 1793 // size, rewrite the allocation instruction to allocate the "right" type. 1794 // There is no need to modify malloc calls because it is their bitcast that 1795 // needs to be cleaned up. 1796 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1797 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1798 return V; 1799 1800 // If the source and destination are pointers, and this cast is equivalent 1801 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1802 // This can enhance SROA and other transforms that want type-safe pointers. 1803 Constant *ZeroUInt = 1804 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1805 unsigned NumZeros = 0; 1806 while (SrcElTy != DstElTy && 1807 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1808 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1809 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1810 ++NumZeros; 1811 } 1812 1813 // If we found a path from the src to dest, create the getelementptr now. 1814 if (SrcElTy == DstElTy) { 1815 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1816 return GetElementPtrInst::CreateInBounds(Src, Idxs); 1817 } 1818 } 1819 1820 // Try to optimize int -> float bitcasts. 1821 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1822 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1823 return I; 1824 1825 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1826 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1827 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1828 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1829 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1830 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1831 } 1832 1833 if (isa<IntegerType>(SrcTy)) { 1834 // If this is a cast from an integer to vector, check to see if the input 1835 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1836 // the casts with a shuffle and (potentially) a bitcast. 1837 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1838 CastInst *SrcCast = cast<CastInst>(Src); 1839 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1840 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1841 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1842 cast<VectorType>(DestTy), *this)) 1843 return I; 1844 } 1845 1846 // If the input is an 'or' instruction, we may be doing shifts and ors to 1847 // assemble the elements of the vector manually. Try to rip the code out 1848 // and replace it with insertelements. 1849 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1850 return ReplaceInstUsesWith(CI, V); 1851 } 1852 } 1853 1854 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1855 if (SrcVTy->getNumElements() == 1) { 1856 // If our destination is not a vector, then make this a straight 1857 // scalar-scalar cast. 1858 if (!DestTy->isVectorTy()) { 1859 Value *Elem = 1860 Builder->CreateExtractElement(Src, 1861 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1862 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1863 } 1864 1865 // Otherwise, see if our source is an insert. If so, then use the scalar 1866 // component directly. 1867 if (InsertElementInst *IEI = 1868 dyn_cast<InsertElementInst>(CI.getOperand(0))) 1869 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1), 1870 DestTy); 1871 } 1872 } 1873 1874 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1875 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1876 // a bitcast to a vector with the same # elts. 1877 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1878 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() && 1879 SVI->getType()->getNumElements() == 1880 SVI->getOperand(0)->getType()->getVectorNumElements()) { 1881 BitCastInst *Tmp; 1882 // If either of the operands is a cast from CI.getType(), then 1883 // evaluating the shuffle in the casted destination's type will allow 1884 // us to eliminate at least one cast. 1885 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1886 Tmp->getOperand(0)->getType() == DestTy) || 1887 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1888 Tmp->getOperand(0)->getType() == DestTy)) { 1889 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1890 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1891 // Return a new shuffle vector. Use the same element ID's, as we 1892 // know the vector types match #elts. 1893 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1894 } 1895 } 1896 } 1897 1898 if (SrcTy->isPointerTy()) 1899 return commonPointerCastTransforms(CI); 1900 return commonCastTransforms(CI); 1901} 1902 1903Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) { 1904 return commonPointerCastTransforms(CI); 1905} 1906