InstCombineCasts.cpp revision 243830
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/DataLayout.h" 17#include "llvm/Target/TargetLibraryInfo.h" 18#include "llvm/Support/PatternMatch.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 (!TD) 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 = TD->getABITypeAlignment(AllocElTy); 95 unsigned CastElTyAlign = TD->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 = TD->getTypeAllocSize(AllocElTy); 104 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); 105 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 106 107 // See if we can satisfy the modulus by pulling a scale out of the array 108 // size argument. 109 unsigned ArraySizeScale; 110 uint64_t ArrayOffset; 111 Value *NumElements = // See if the array size is a decomposable linear expr. 112 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 113 114 // If we can now satisfy the modulus, by using a non-1 scale, we really can 115 // do the xform. 116 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 117 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 118 119 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 120 Value *Amt = 0; 121 if (Scale == 1) { 122 Amt = NumElements; 123 } else { 124 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 125 // Insert before the alloca, not before the cast. 126 Amt = AllocaBuilder.CreateMul(Amt, NumElements); 127 } 128 129 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 130 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 131 Offset, true); 132 Amt = AllocaBuilder.CreateAdd(Amt, Off); 133 } 134 135 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 136 New->setAlignment(AI.getAlignment()); 137 New->takeName(&AI); 138 139 // If the allocation has multiple real uses, insert a cast and change all 140 // things that used it to use the new cast. This will also hack on CI, but it 141 // will die soon. 142 if (!AI.hasOneUse()) { 143 // New is the allocation instruction, pointer typed. AI is the original 144 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 145 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 146 ReplaceInstUsesWith(AI, NewCast); 147 } 148 return ReplaceInstUsesWith(CI, New); 149} 150 151/// EvaluateInDifferentType - Given an expression that 152/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 153/// insert the code to evaluate the expression. 154Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, 155 bool isSigned) { 156 if (Constant *C = dyn_cast<Constant>(V)) { 157 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 158 // If we got a constantexpr back, try to simplify it with TD info. 159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 160 C = ConstantFoldConstantExpression(CE, TD, TLI); 161 return C; 162 } 163 164 // Otherwise, it must be an instruction. 165 Instruction *I = cast<Instruction>(V); 166 Instruction *Res = 0; 167 unsigned Opc = I->getOpcode(); 168 switch (Opc) { 169 case Instruction::Add: 170 case Instruction::Sub: 171 case Instruction::Mul: 172 case Instruction::And: 173 case Instruction::Or: 174 case Instruction::Xor: 175 case Instruction::AShr: 176 case Instruction::LShr: 177 case Instruction::Shl: 178 case Instruction::UDiv: 179 case Instruction::URem: { 180 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 181 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 182 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 183 break; 184 } 185 case Instruction::Trunc: 186 case Instruction::ZExt: 187 case Instruction::SExt: 188 // If the source type of the cast is the type we're trying for then we can 189 // just return the source. There's no need to insert it because it is not 190 // new. 191 if (I->getOperand(0)->getType() == Ty) 192 return I->getOperand(0); 193 194 // Otherwise, must be the same type of cast, so just reinsert a new one. 195 // This also handles the case of zext(trunc(x)) -> zext(x). 196 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 197 Opc == Instruction::SExt); 198 break; 199 case Instruction::Select: { 200 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 201 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 202 Res = SelectInst::Create(I->getOperand(0), True, False); 203 break; 204 } 205 case Instruction::PHI: { 206 PHINode *OPN = cast<PHINode>(I); 207 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues()); 208 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 209 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 210 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 211 } 212 Res = NPN; 213 break; 214 } 215 default: 216 // TODO: Can handle more cases here. 217 llvm_unreachable("Unreachable!"); 218 } 219 220 Res->takeName(I); 221 return InsertNewInstWith(Res, *I); 222} 223 224 225/// This function is a wrapper around CastInst::isEliminableCastPair. It 226/// simply extracts arguments and returns what that function returns. 227static Instruction::CastOps 228isEliminableCastPair( 229 const CastInst *CI, ///< The first cast instruction 230 unsigned opcode, ///< The opcode of the second cast instruction 231 Type *DstTy, ///< The target type for the second cast instruction 232 DataLayout *TD ///< The target data for pointer size 233) { 234 235 Type *SrcTy = CI->getOperand(0)->getType(); // A from above 236 Type *MidTy = CI->getType(); // B from above 237 238 // Get the opcodes of the two Cast instructions 239 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 240 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 241 Type *SrcIntPtrTy = TD && SrcTy->isPtrOrPtrVectorTy() ? 242 TD->getIntPtrType(SrcTy) : 0; 243 Type *MidIntPtrTy = TD && MidTy->isPtrOrPtrVectorTy() ? 244 TD->getIntPtrType(MidTy) : 0; 245 Type *DstIntPtrTy = TD && DstTy->isPtrOrPtrVectorTy() ? 246 TD->getIntPtrType(DstTy) : 0; 247 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 248 DstTy, SrcIntPtrTy, MidIntPtrTy, 249 DstIntPtrTy); 250 251 // We don't want to form an inttoptr or ptrtoint that converts to an integer 252 // type that differs from the pointer size. 253 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) || 254 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy)) 255 Res = 0; 256 257 return Instruction::CastOps(Res); 258} 259 260/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 261/// results in any code being generated and is interesting to optimize out. If 262/// the cast can be eliminated by some other simple transformation, we prefer 263/// to do the simplification first. 264bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 265 Type *Ty) { 266 // Noop casts and casts of constants should be eliminated trivially. 267 if (V->getType() == Ty || isa<Constant>(V)) return false; 268 269 // If this is another cast that can be eliminated, we prefer to have it 270 // eliminated. 271 if (const CastInst *CI = dyn_cast<CastInst>(V)) 272 if (isEliminableCastPair(CI, opc, Ty, TD)) 273 return false; 274 275 // If this is a vector sext from a compare, then we don't want to break the 276 // idiom where each element of the extended vector is either zero or all ones. 277 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 278 return false; 279 280 return true; 281} 282 283 284/// @brief Implement the transforms common to all CastInst visitors. 285Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 286 Value *Src = CI.getOperand(0); 287 288 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 289 // eliminate it now. 290 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 291 if (Instruction::CastOps opc = 292 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 293 // The first cast (CSrc) is eliminable so we need to fix up or replace 294 // the second cast (CI). CSrc will then have a good chance of being dead. 295 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 296 } 297 } 298 299 // If we are casting a select then fold the cast into the select 300 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 301 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 302 return NV; 303 304 // If we are casting a PHI then fold the cast into the PHI 305 if (isa<PHINode>(Src)) { 306 // We don't do this if this would create a PHI node with an illegal type if 307 // it is currently legal. 308 if (!Src->getType()->isIntegerTy() || 309 !CI.getType()->isIntegerTy() || 310 ShouldChangeType(CI.getType(), Src->getType())) 311 if (Instruction *NV = FoldOpIntoPhi(CI)) 312 return NV; 313 } 314 315 return 0; 316} 317 318/// CanEvaluateTruncated - Return true if we can evaluate the specified 319/// expression tree as type Ty instead of its larger type, and arrive with the 320/// same value. This is used by code that tries to eliminate truncates. 321/// 322/// Ty will always be a type smaller than V. We should return true if trunc(V) 323/// can be computed by computing V in the smaller type. If V is an instruction, 324/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 325/// makes sense if x and y can be efficiently truncated. 326/// 327/// This function works on both vectors and scalars. 328/// 329static bool CanEvaluateTruncated(Value *V, Type *Ty) { 330 // We can always evaluate constants in another type. 331 if (isa<Constant>(V)) 332 return true; 333 334 Instruction *I = dyn_cast<Instruction>(V); 335 if (!I) return false; 336 337 Type *OrigTy = V->getType(); 338 339 // If this is an extension from the dest type, we can eliminate it, even if it 340 // has multiple uses. 341 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 342 I->getOperand(0)->getType() == Ty) 343 return true; 344 345 // We can't extend or shrink something that has multiple uses: doing so would 346 // require duplicating the instruction in general, which isn't profitable. 347 if (!I->hasOneUse()) return false; 348 349 unsigned Opc = I->getOpcode(); 350 switch (Opc) { 351 case Instruction::Add: 352 case Instruction::Sub: 353 case Instruction::Mul: 354 case Instruction::And: 355 case Instruction::Or: 356 case Instruction::Xor: 357 // These operators can all arbitrarily be extended or truncated. 358 return CanEvaluateTruncated(I->getOperand(0), Ty) && 359 CanEvaluateTruncated(I->getOperand(1), Ty); 360 361 case Instruction::UDiv: 362 case Instruction::URem: { 363 // UDiv and URem can be truncated if all the truncated bits are zero. 364 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 365 uint32_t BitWidth = Ty->getScalarSizeInBits(); 366 if (BitWidth < OrigBitWidth) { 367 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 368 if (MaskedValueIsZero(I->getOperand(0), Mask) && 369 MaskedValueIsZero(I->getOperand(1), Mask)) { 370 return CanEvaluateTruncated(I->getOperand(0), Ty) && 371 CanEvaluateTruncated(I->getOperand(1), Ty); 372 } 373 } 374 break; 375 } 376 case Instruction::Shl: 377 // If we are truncating the result of this SHL, and if it's a shift of a 378 // constant amount, we can always perform a SHL in a smaller type. 379 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 380 uint32_t BitWidth = Ty->getScalarSizeInBits(); 381 if (CI->getLimitedValue(BitWidth) < BitWidth) 382 return CanEvaluateTruncated(I->getOperand(0), Ty); 383 } 384 break; 385 case Instruction::LShr: 386 // If this is a truncate of a logical shr, we can truncate it to a smaller 387 // lshr iff we know that the bits we would otherwise be shifting in are 388 // already zeros. 389 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 390 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 391 uint32_t BitWidth = Ty->getScalarSizeInBits(); 392 if (MaskedValueIsZero(I->getOperand(0), 393 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 394 CI->getLimitedValue(BitWidth) < BitWidth) { 395 return CanEvaluateTruncated(I->getOperand(0), Ty); 396 } 397 } 398 break; 399 case Instruction::Trunc: 400 // trunc(trunc(x)) -> trunc(x) 401 return true; 402 case Instruction::ZExt: 403 case Instruction::SExt: 404 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 405 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 406 return true; 407 case Instruction::Select: { 408 SelectInst *SI = cast<SelectInst>(I); 409 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 410 CanEvaluateTruncated(SI->getFalseValue(), Ty); 411 } 412 case Instruction::PHI: { 413 // We can change a phi if we can change all operands. Note that we never 414 // get into trouble with cyclic PHIs here because we only consider 415 // instructions with a single use. 416 PHINode *PN = cast<PHINode>(I); 417 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 418 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 419 return false; 420 return true; 421 } 422 default: 423 // TODO: Can handle more cases here. 424 break; 425 } 426 427 return false; 428} 429 430Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 431 if (Instruction *Result = commonCastTransforms(CI)) 432 return Result; 433 434 // See if we can simplify any instructions used by the input whose sole 435 // purpose is to compute bits we don't care about. 436 if (SimplifyDemandedInstructionBits(CI)) 437 return &CI; 438 439 Value *Src = CI.getOperand(0); 440 Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 441 442 // Attempt to truncate the entire input expression tree to the destination 443 // type. Only do this if the dest type is a simple type, don't convert the 444 // expression tree to something weird like i93 unless the source is also 445 // strange. 446 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 447 CanEvaluateTruncated(Src, DestTy)) { 448 449 // If this cast is a truncate, evaluting in a different type always 450 // eliminates the cast, so it is always a win. 451 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 452 " to avoid cast: " << CI << '\n'); 453 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 454 assert(Res->getType() == DestTy); 455 return ReplaceInstUsesWith(CI, Res); 456 } 457 458 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 459 if (DestTy->getScalarSizeInBits() == 1) { 460 Constant *One = ConstantInt::get(Src->getType(), 1); 461 Src = Builder->CreateAnd(Src, One); 462 Value *Zero = Constant::getNullValue(Src->getType()); 463 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 464 } 465 466 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 467 Value *A = 0; ConstantInt *Cst = 0; 468 if (Src->hasOneUse() && 469 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 470 // We have three types to worry about here, the type of A, the source of 471 // the truncate (MidSize), and the destination of the truncate. We know that 472 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 473 // between ASize and ResultSize. 474 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 475 476 // If the shift amount is larger than the size of A, then the result is 477 // known to be zero because all the input bits got shifted out. 478 if (Cst->getZExtValue() >= ASize) 479 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType())); 480 481 // Since we're doing an lshr and a zero extend, and know that the shift 482 // amount is smaller than ASize, it is always safe to do the shift in A's 483 // type, then zero extend or truncate to the result. 484 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 485 Shift->takeName(Src); 486 return CastInst::CreateIntegerCast(Shift, CI.getType(), false); 487 } 488 489 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 490 // type isn't non-native. 491 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) && 492 ShouldChangeType(Src->getType(), CI.getType()) && 493 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 494 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr"); 495 return BinaryOperator::CreateAnd(NewTrunc, 496 ConstantExpr::getTrunc(Cst, CI.getType())); 497 } 498 499 return 0; 500} 501 502/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 503/// in order to eliminate the icmp. 504Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 505 bool DoXform) { 506 // If we are just checking for a icmp eq of a single bit and zext'ing it 507 // to an integer, then shift the bit to the appropriate place and then 508 // cast to integer to avoid the comparison. 509 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 510 const APInt &Op1CV = Op1C->getValue(); 511 512 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 513 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 514 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 515 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 516 if (!DoXform) return ICI; 517 518 Value *In = ICI->getOperand(0); 519 Value *Sh = ConstantInt::get(In->getType(), 520 In->getType()->getScalarSizeInBits()-1); 521 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 522 if (In->getType() != CI.getType()) 523 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); 524 525 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 526 Constant *One = ConstantInt::get(In->getType(), 1); 527 In = Builder->CreateXor(In, One, In->getName()+".not"); 528 } 529 530 return ReplaceInstUsesWith(CI, In); 531 } 532 533 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 534 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 535 // zext (X == 1) to i32 --> X iff X has only the low bit set. 536 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 537 // zext (X != 0) to i32 --> X iff X has only the low bit set. 538 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 539 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 540 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 541 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 542 // This only works for EQ and NE 543 ICI->isEquality()) { 544 // If Op1C some other power of two, convert: 545 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 546 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 547 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne); 548 549 APInt KnownZeroMask(~KnownZero); 550 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 551 if (!DoXform) return ICI; 552 553 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 554 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 555 // (X&4) == 2 --> false 556 // (X&4) != 2 --> true 557 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 558 isNE); 559 Res = ConstantExpr::getZExt(Res, CI.getType()); 560 return ReplaceInstUsesWith(CI, Res); 561 } 562 563 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 564 Value *In = ICI->getOperand(0); 565 if (ShiftAmt) { 566 // Perform a logical shr by shiftamt. 567 // Insert the shift to put the result in the low bit. 568 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 569 In->getName()+".lobit"); 570 } 571 572 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 573 Constant *One = ConstantInt::get(In->getType(), 1); 574 In = Builder->CreateXor(In, One); 575 } 576 577 if (CI.getType() == In->getType()) 578 return ReplaceInstUsesWith(CI, In); 579 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 580 } 581 } 582 } 583 584 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 585 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 586 // may lead to additional simplifications. 587 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 588 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 589 uint32_t BitWidth = ITy->getBitWidth(); 590 Value *LHS = ICI->getOperand(0); 591 Value *RHS = ICI->getOperand(1); 592 593 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 594 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 595 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS); 596 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS); 597 598 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 599 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 600 APInt UnknownBit = ~KnownBits; 601 if (UnknownBit.countPopulation() == 1) { 602 if (!DoXform) return ICI; 603 604 Value *Result = Builder->CreateXor(LHS, RHS); 605 606 // Mask off any bits that are set and won't be shifted away. 607 if (KnownOneLHS.uge(UnknownBit)) 608 Result = Builder->CreateAnd(Result, 609 ConstantInt::get(ITy, UnknownBit)); 610 611 // Shift the bit we're testing down to the lsb. 612 Result = Builder->CreateLShr( 613 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 614 615 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 616 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 617 Result->takeName(ICI); 618 return ReplaceInstUsesWith(CI, Result); 619 } 620 } 621 } 622 } 623 624 return 0; 625} 626 627/// CanEvaluateZExtd - Determine if the specified value can be computed in the 628/// specified wider type and produce the same low bits. If not, return false. 629/// 630/// If this function returns true, it can also return a non-zero number of bits 631/// (in BitsToClear) which indicates that the value it computes is correct for 632/// the zero extend, but that the additional BitsToClear bits need to be zero'd 633/// out. For example, to promote something like: 634/// 635/// %B = trunc i64 %A to i32 636/// %C = lshr i32 %B, 8 637/// %E = zext i32 %C to i64 638/// 639/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 640/// set to 8 to indicate that the promoted value needs to have bits 24-31 641/// cleared in addition to bits 32-63. Since an 'and' will be generated to 642/// clear the top bits anyway, doing this has no extra cost. 643/// 644/// This function works on both vectors and scalars. 645static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { 646 BitsToClear = 0; 647 if (isa<Constant>(V)) 648 return true; 649 650 Instruction *I = dyn_cast<Instruction>(V); 651 if (!I) return false; 652 653 // If the input is a truncate from the destination type, we can trivially 654 // eliminate it. 655 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 656 return true; 657 658 // We can't extend or shrink something that has multiple uses: doing so would 659 // require duplicating the instruction in general, which isn't profitable. 660 if (!I->hasOneUse()) return false; 661 662 unsigned Opc = I->getOpcode(), Tmp; 663 switch (Opc) { 664 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 665 case Instruction::SExt: // zext(sext(x)) -> sext(x). 666 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 667 return true; 668 case Instruction::And: 669 case Instruction::Or: 670 case Instruction::Xor: 671 case Instruction::Add: 672 case Instruction::Sub: 673 case Instruction::Mul: 674 case Instruction::Shl: 675 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 676 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 677 return false; 678 // These can all be promoted if neither operand has 'bits to clear'. 679 if (BitsToClear == 0 && Tmp == 0) 680 return true; 681 682 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 683 // other side, BitsToClear is ok. 684 if (Tmp == 0 && 685 (Opc == Instruction::And || Opc == Instruction::Or || 686 Opc == Instruction::Xor)) { 687 // We use MaskedValueIsZero here for generality, but the case we care 688 // about the most is constant RHS. 689 unsigned VSize = V->getType()->getScalarSizeInBits(); 690 if (MaskedValueIsZero(I->getOperand(1), 691 APInt::getHighBitsSet(VSize, BitsToClear))) 692 return true; 693 } 694 695 // Otherwise, we don't know how to analyze this BitsToClear case yet. 696 return false; 697 698 case Instruction::LShr: 699 // We can promote lshr(x, cst) if we can promote x. This requires the 700 // ultimate 'and' to clear out the high zero bits we're clearing out though. 701 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 702 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 703 return false; 704 BitsToClear += Amt->getZExtValue(); 705 if (BitsToClear > V->getType()->getScalarSizeInBits()) 706 BitsToClear = V->getType()->getScalarSizeInBits(); 707 return true; 708 } 709 // Cannot promote variable LSHR. 710 return false; 711 case Instruction::Select: 712 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 713 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 714 // TODO: If important, we could handle the case when the BitsToClear are 715 // known zero in the disagreeing side. 716 Tmp != BitsToClear) 717 return false; 718 return true; 719 720 case Instruction::PHI: { 721 // We can change a phi if we can change all operands. Note that we never 722 // get into trouble with cyclic PHIs here because we only consider 723 // instructions with a single use. 724 PHINode *PN = cast<PHINode>(I); 725 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 726 return false; 727 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 728 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 729 // TODO: If important, we could handle the case when the BitsToClear 730 // are known zero in the disagreeing input. 731 Tmp != BitsToClear) 732 return false; 733 return true; 734 } 735 default: 736 // TODO: Can handle more cases here. 737 return false; 738 } 739} 740 741Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 742 // If this zero extend is only used by a truncate, let the truncate by 743 // eliminated before we try to optimize this zext. 744 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 745 return 0; 746 747 // If one of the common conversion will work, do it. 748 if (Instruction *Result = commonCastTransforms(CI)) 749 return Result; 750 751 // See if we can simplify any instructions used by the input whose sole 752 // purpose is to compute bits we don't care about. 753 if (SimplifyDemandedInstructionBits(CI)) 754 return &CI; 755 756 Value *Src = CI.getOperand(0); 757 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 758 759 // Attempt to extend the entire input expression tree to the destination 760 // type. Only do this if the dest type is a simple type, don't convert the 761 // expression tree to something weird like i93 unless the source is also 762 // strange. 763 unsigned BitsToClear; 764 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 765 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 766 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 767 "Unreasonable BitsToClear"); 768 769 // Okay, we can transform this! Insert the new expression now. 770 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 771 " to avoid zero extend: " << CI); 772 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 773 assert(Res->getType() == DestTy); 774 775 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 776 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 777 778 // If the high bits are already filled with zeros, just replace this 779 // cast with the result. 780 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 781 DestBitSize-SrcBitsKept))) 782 return ReplaceInstUsesWith(CI, Res); 783 784 // We need to emit an AND to clear the high bits. 785 Constant *C = ConstantInt::get(Res->getType(), 786 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 787 return BinaryOperator::CreateAnd(Res, C); 788 } 789 790 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 791 // types and if the sizes are just right we can convert this into a logical 792 // 'and' which will be much cheaper than the pair of casts. 793 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 794 // TODO: Subsume this into EvaluateInDifferentType. 795 796 // Get the sizes of the types involved. We know that the intermediate type 797 // will be smaller than A or C, but don't know the relation between A and C. 798 Value *A = CSrc->getOperand(0); 799 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 800 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 801 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 802 // If we're actually extending zero bits, then if 803 // SrcSize < DstSize: zext(a & mask) 804 // SrcSize == DstSize: a & mask 805 // SrcSize > DstSize: trunc(a) & mask 806 if (SrcSize < DstSize) { 807 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 808 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 809 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 810 return new ZExtInst(And, CI.getType()); 811 } 812 813 if (SrcSize == DstSize) { 814 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 815 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 816 AndValue)); 817 } 818 if (SrcSize > DstSize) { 819 Value *Trunc = Builder->CreateTrunc(A, CI.getType()); 820 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 821 return BinaryOperator::CreateAnd(Trunc, 822 ConstantInt::get(Trunc->getType(), 823 AndValue)); 824 } 825 } 826 827 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 828 return transformZExtICmp(ICI, CI); 829 830 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 831 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 832 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 833 // of the (zext icmp) will be transformed. 834 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 835 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 836 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 837 (transformZExtICmp(LHS, CI, false) || 838 transformZExtICmp(RHS, CI, false))) { 839 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 840 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 841 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 842 } 843 } 844 845 // zext(trunc(t) & C) -> (t & zext(C)). 846 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 847 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 848 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 849 Value *TI0 = TI->getOperand(0); 850 if (TI0->getType() == CI.getType()) 851 return 852 BinaryOperator::CreateAnd(TI0, 853 ConstantExpr::getZExt(C, CI.getType())); 854 } 855 856 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 857 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 858 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 859 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 860 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 861 And->getOperand(1) == C) 862 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 863 Value *TI0 = TI->getOperand(0); 864 if (TI0->getType() == CI.getType()) { 865 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 866 Value *NewAnd = Builder->CreateAnd(TI0, ZC); 867 return BinaryOperator::CreateXor(NewAnd, ZC); 868 } 869 } 870 871 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 872 Value *X; 873 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 874 match(SrcI, m_Not(m_Value(X))) && 875 (!X->hasOneUse() || !isa<CmpInst>(X))) { 876 Value *New = Builder->CreateZExt(X, CI.getType()); 877 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 878 } 879 880 return 0; 881} 882 883/// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations 884/// in order to eliminate the icmp. 885Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { 886 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); 887 ICmpInst::Predicate Pred = ICI->getPredicate(); 888 889 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 890 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative 891 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive 892 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) || 893 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { 894 895 Value *Sh = ConstantInt::get(Op0->getType(), 896 Op0->getType()->getScalarSizeInBits()-1); 897 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); 898 if (In->getType() != CI.getType()) 899 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); 900 901 if (Pred == ICmpInst::ICMP_SGT) 902 In = Builder->CreateNot(In, In->getName()+".not"); 903 return ReplaceInstUsesWith(CI, In); 904 } 905 906 // If we know that only one bit of the LHS of the icmp can be set and we 907 // have an equality comparison with zero or a power of 2, we can transform 908 // the icmp and sext into bitwise/integer operations. 909 if (ICI->hasOneUse() && 910 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ 911 unsigned BitWidth = Op1C->getType()->getBitWidth(); 912 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 913 ComputeMaskedBits(Op0, KnownZero, KnownOne); 914 915 APInt KnownZeroMask(~KnownZero); 916 if (KnownZeroMask.isPowerOf2()) { 917 Value *In = ICI->getOperand(0); 918 919 // If the icmp tests for a known zero bit we can constant fold it. 920 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) { 921 Value *V = Pred == ICmpInst::ICMP_NE ? 922 ConstantInt::getAllOnesValue(CI.getType()) : 923 ConstantInt::getNullValue(CI.getType()); 924 return ReplaceInstUsesWith(CI, V); 925 } 926 927 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) { 928 // sext ((x & 2^n) == 0) -> (x >> n) - 1 929 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1 930 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros(); 931 // Perform a right shift to place the desired bit in the LSB. 932 if (ShiftAmt) 933 In = Builder->CreateLShr(In, 934 ConstantInt::get(In->getType(), ShiftAmt)); 935 936 // At this point "In" is either 1 or 0. Subtract 1 to turn 937 // {1, 0} -> {0, -1}. 938 In = Builder->CreateAdd(In, 939 ConstantInt::getAllOnesValue(In->getType()), 940 "sext"); 941 } else { 942 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1 943 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1 944 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros(); 945 // Perform a left shift to place the desired bit in the MSB. 946 if (ShiftAmt) 947 In = Builder->CreateShl(In, 948 ConstantInt::get(In->getType(), ShiftAmt)); 949 950 // Distribute the bit over the whole bit width. 951 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(), 952 BitWidth - 1), "sext"); 953 } 954 955 if (CI.getType() == In->getType()) 956 return ReplaceInstUsesWith(CI, In); 957 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/); 958 } 959 } 960 } 961 962 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed. 963 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) { 964 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) && 965 Op0->getType() == CI.getType()) { 966 Type *EltTy = VTy->getElementType(); 967 968 // splat the shift constant to a constant vector. 969 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1); 970 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit"); 971 return ReplaceInstUsesWith(CI, In); 972 } 973 } 974 975 return 0; 976} 977 978/// CanEvaluateSExtd - Return true if we can take the specified value 979/// and return it as type Ty without inserting any new casts and without 980/// changing the value of the common low bits. This is used by code that tries 981/// to promote integer operations to a wider types will allow us to eliminate 982/// the extension. 983/// 984/// This function works on both vectors and scalars. 985/// 986static bool CanEvaluateSExtd(Value *V, Type *Ty) { 987 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 988 "Can't sign extend type to a smaller type"); 989 // If this is a constant, it can be trivially promoted. 990 if (isa<Constant>(V)) 991 return true; 992 993 Instruction *I = dyn_cast<Instruction>(V); 994 if (!I) return false; 995 996 // If this is a truncate from the dest type, we can trivially eliminate it. 997 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 998 return true; 999 1000 // We can't extend or shrink something that has multiple uses: doing so would 1001 // require duplicating the instruction in general, which isn't profitable. 1002 if (!I->hasOneUse()) return false; 1003 1004 switch (I->getOpcode()) { 1005 case Instruction::SExt: // sext(sext(x)) -> sext(x) 1006 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 1007 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 1008 return true; 1009 case Instruction::And: 1010 case Instruction::Or: 1011 case Instruction::Xor: 1012 case Instruction::Add: 1013 case Instruction::Sub: 1014 case Instruction::Mul: 1015 // These operators can all arbitrarily be extended if their inputs can. 1016 return CanEvaluateSExtd(I->getOperand(0), Ty) && 1017 CanEvaluateSExtd(I->getOperand(1), Ty); 1018 1019 //case Instruction::Shl: TODO 1020 //case Instruction::LShr: TODO 1021 1022 case Instruction::Select: 1023 return CanEvaluateSExtd(I->getOperand(1), Ty) && 1024 CanEvaluateSExtd(I->getOperand(2), Ty); 1025 1026 case Instruction::PHI: { 1027 // We can change a phi if we can change all operands. Note that we never 1028 // get into trouble with cyclic PHIs here because we only consider 1029 // instructions with a single use. 1030 PHINode *PN = cast<PHINode>(I); 1031 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1032 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 1033 return true; 1034 } 1035 default: 1036 // TODO: Can handle more cases here. 1037 break; 1038 } 1039 1040 return false; 1041} 1042 1043Instruction *InstCombiner::visitSExt(SExtInst &CI) { 1044 // If this sign extend is only used by a truncate, let the truncate by 1045 // eliminated before we try to optimize this zext. 1046 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 1047 return 0; 1048 1049 if (Instruction *I = commonCastTransforms(CI)) 1050 return I; 1051 1052 // See if we can simplify any instructions used by the input whose sole 1053 // purpose is to compute bits we don't care about. 1054 if (SimplifyDemandedInstructionBits(CI)) 1055 return &CI; 1056 1057 Value *Src = CI.getOperand(0); 1058 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 1059 1060 // Attempt to extend the entire input expression tree to the destination 1061 // type. Only do this if the dest type is a simple type, don't convert the 1062 // expression tree to something weird like i93 unless the source is also 1063 // strange. 1064 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 1065 CanEvaluateSExtd(Src, DestTy)) { 1066 // Okay, we can transform this! Insert the new expression now. 1067 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 1068 " to avoid sign extend: " << CI); 1069 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 1070 assert(Res->getType() == DestTy); 1071 1072 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1073 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1074 1075 // If the high bits are already filled with sign bit, just replace this 1076 // cast with the result. 1077 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 1078 return ReplaceInstUsesWith(CI, Res); 1079 1080 // We need to emit a shl + ashr to do the sign extend. 1081 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1082 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 1083 ShAmt); 1084 } 1085 1086 // If this input is a trunc from our destination, then turn sext(trunc(x)) 1087 // into shifts. 1088 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 1089 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 1090 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1091 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1092 1093 // We need to emit a shl + ashr to do the sign extend. 1094 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1095 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1096 return BinaryOperator::CreateAShr(Res, ShAmt); 1097 } 1098 1099 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 1100 return transformSExtICmp(ICI, CI); 1101 1102 // If the input is a shl/ashr pair of a same constant, then this is a sign 1103 // extension from a smaller value. If we could trust arbitrary bitwidth 1104 // integers, we could turn this into a truncate to the smaller bit and then 1105 // use a sext for the whole extension. Since we don't, look deeper and check 1106 // for a truncate. If the source and dest are the same type, eliminate the 1107 // trunc and extend and just do shifts. For example, turn: 1108 // %a = trunc i32 %i to i8 1109 // %b = shl i8 %a, 6 1110 // %c = ashr i8 %b, 6 1111 // %d = sext i8 %c to i32 1112 // into: 1113 // %a = shl i32 %i, 30 1114 // %d = ashr i32 %a, 30 1115 Value *A = 0; 1116 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1117 ConstantInt *BA = 0, *CA = 0; 1118 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1119 m_ConstantInt(CA))) && 1120 BA == CA && A->getType() == CI.getType()) { 1121 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1122 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1123 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1124 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1125 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1126 return BinaryOperator::CreateAShr(A, ShAmtV); 1127 } 1128 1129 return 0; 1130} 1131 1132 1133/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1134/// in the specified FP type without changing its value. 1135static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1136 bool losesInfo; 1137 APFloat F = CFP->getValueAPF(); 1138 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1139 if (!losesInfo) 1140 return ConstantFP::get(CFP->getContext(), F); 1141 return 0; 1142} 1143 1144/// LookThroughFPExtensions - If this is an fp extension instruction, look 1145/// through it until we get the source value. 1146static Value *LookThroughFPExtensions(Value *V) { 1147 if (Instruction *I = dyn_cast<Instruction>(V)) 1148 if (I->getOpcode() == Instruction::FPExt) 1149 return LookThroughFPExtensions(I->getOperand(0)); 1150 1151 // If this value is a constant, return the constant in the smallest FP type 1152 // that can accurately represent it. This allows us to turn 1153 // (float)((double)X+2.0) into x+2.0f. 1154 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1155 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1156 return V; // No constant folding of this. 1157 // See if the value can be truncated to half and then reextended. 1158 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf)) 1159 return V; 1160 // See if the value can be truncated to float and then reextended. 1161 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1162 return V; 1163 if (CFP->getType()->isDoubleTy()) 1164 return V; // Won't shrink. 1165 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1166 return V; 1167 // Don't try to shrink to various long double types. 1168 } 1169 1170 return V; 1171} 1172 1173Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1174 if (Instruction *I = commonCastTransforms(CI)) 1175 return I; 1176 1177 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1178 // smaller than the destination type, we can eliminate the truncate by doing 1179 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1180 // as many builtins (sqrt, etc). 1181 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1182 if (OpI && OpI->hasOneUse()) { 1183 switch (OpI->getOpcode()) { 1184 default: break; 1185 case Instruction::FAdd: 1186 case Instruction::FSub: 1187 case Instruction::FMul: 1188 case Instruction::FDiv: 1189 case Instruction::FRem: 1190 Type *SrcTy = OpI->getType(); 1191 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1192 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1193 if (LHSTrunc->getType() != SrcTy && 1194 RHSTrunc->getType() != SrcTy) { 1195 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1196 // If the source types were both smaller than the destination type of 1197 // the cast, do this xform. 1198 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1199 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1200 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1201 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1202 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1203 } 1204 } 1205 break; 1206 } 1207 } 1208 1209 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1210 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1211 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) && 1212 Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) && 1213 Call->getNumArgOperands() == 1 && 1214 Call->hasOneUse()) { 1215 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1216 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1217 CI.getType()->isFloatTy() && 1218 Call->getType()->isDoubleTy() && 1219 Arg->getType()->isDoubleTy() && 1220 Arg->getOperand(0)->getType()->isFloatTy()) { 1221 Function *Callee = Call->getCalledFunction(); 1222 Module *M = CI.getParent()->getParent()->getParent(); 1223 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf", 1224 Callee->getAttributes(), 1225 Builder->getFloatTy(), 1226 Builder->getFloatTy(), 1227 NULL); 1228 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1229 "sqrtfcall"); 1230 ret->setAttributes(Callee->getAttributes()); 1231 1232 1233 // Remove the old Call. With -fmath-errno, it won't get marked readnone. 1234 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType())); 1235 EraseInstFromFunction(*Call); 1236 return ret; 1237 } 1238 } 1239 1240 return 0; 1241} 1242 1243Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1244 return commonCastTransforms(CI); 1245} 1246 1247Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1248 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1249 if (OpI == 0) 1250 return commonCastTransforms(FI); 1251 1252 // fptoui(uitofp(X)) --> X 1253 // fptoui(sitofp(X)) --> X 1254 // This is safe if the intermediate type has enough bits in its mantissa to 1255 // accurately represent all values of X. For example, do not do this with 1256 // i64->float->i64. This is also safe for sitofp case, because any negative 1257 // 'X' value would cause an undefined result for the fptoui. 1258 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1259 OpI->getOperand(0)->getType() == FI.getType() && 1260 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1261 OpI->getType()->getFPMantissaWidth()) 1262 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1263 1264 return commonCastTransforms(FI); 1265} 1266 1267Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1268 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1269 if (OpI == 0) 1270 return commonCastTransforms(FI); 1271 1272 // fptosi(sitofp(X)) --> X 1273 // fptosi(uitofp(X)) --> X 1274 // This is safe if the intermediate type has enough bits in its mantissa to 1275 // accurately represent all values of X. For example, do not do this with 1276 // i64->float->i64. This is also safe for sitofp case, because any negative 1277 // 'X' value would cause an undefined result for the fptoui. 1278 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1279 OpI->getOperand(0)->getType() == FI.getType() && 1280 (int)FI.getType()->getScalarSizeInBits() <= 1281 OpI->getType()->getFPMantissaWidth()) 1282 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1283 1284 return commonCastTransforms(FI); 1285} 1286 1287Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1288 return commonCastTransforms(CI); 1289} 1290 1291Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1292 return commonCastTransforms(CI); 1293} 1294 1295Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1296 // If the source integer type is not the intptr_t type for this target, do a 1297 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1298 // cast to be exposed to other transforms. 1299 if (TD) { 1300 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1301 TD->getPointerSizeInBits()) { 1302 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1303 TD->getIntPtrType(CI.getContext())); 1304 return new IntToPtrInst(P, CI.getType()); 1305 } 1306 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1307 TD->getPointerSizeInBits()) { 1308 Value *P = Builder->CreateZExt(CI.getOperand(0), 1309 TD->getIntPtrType(CI.getContext())); 1310 return new IntToPtrInst(P, CI.getType()); 1311 } 1312 } 1313 1314 if (Instruction *I = commonCastTransforms(CI)) 1315 return I; 1316 1317 return 0; 1318} 1319 1320/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1321Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1322 Value *Src = CI.getOperand(0); 1323 1324 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1325 // If casting the result of a getelementptr instruction with no offset, turn 1326 // this into a cast of the original pointer! 1327 if (GEP->hasAllZeroIndices()) { 1328 // Changing the cast operand is usually not a good idea but it is safe 1329 // here because the pointer operand is being replaced with another 1330 // pointer operand so the opcode doesn't need to change. 1331 Worklist.Add(GEP); 1332 CI.setOperand(0, GEP->getOperand(0)); 1333 return &CI; 1334 } 1335 1336 // If the GEP has a single use, and the base pointer is a bitcast, and the 1337 // GEP computes a constant offset, see if we can convert these three 1338 // instructions into fewer. This typically happens with unions and other 1339 // non-type-safe code. 1340 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1341 GEP->hasAllConstantIndices()) { 1342 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end()); 1343 int64_t Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops); 1344 1345 // Get the base pointer input of the bitcast, and the type it points to. 1346 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1347 Type *GEPIdxTy = 1348 cast<PointerType>(OrigBase->getType())->getElementType(); 1349 SmallVector<Value*, 8> NewIndices; 1350 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1351 // If we were able to index down into an element, create the GEP 1352 // and bitcast the result. This eliminates one bitcast, potentially 1353 // two. 1354 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1355 Builder->CreateInBoundsGEP(OrigBase, NewIndices) : 1356 Builder->CreateGEP(OrigBase, NewIndices); 1357 NGEP->takeName(GEP); 1358 1359 if (isa<BitCastInst>(CI)) 1360 return new BitCastInst(NGEP, CI.getType()); 1361 assert(isa<PtrToIntInst>(CI)); 1362 return new PtrToIntInst(NGEP, CI.getType()); 1363 } 1364 } 1365 } 1366 1367 return commonCastTransforms(CI); 1368} 1369 1370Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1371 // If the destination integer type is not the intptr_t type for this target, 1372 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1373 // to be exposed to other transforms. 1374 if (TD) { 1375 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1376 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1377 TD->getIntPtrType(CI.getContext())); 1378 return new TruncInst(P, CI.getType()); 1379 } 1380 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1381 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1382 TD->getIntPtrType(CI.getContext())); 1383 return new ZExtInst(P, CI.getType()); 1384 } 1385 } 1386 1387 return commonPointerCastTransforms(CI); 1388} 1389 1390/// OptimizeVectorResize - This input value (which is known to have vector type) 1391/// is being zero extended or truncated to the specified vector type. Try to 1392/// replace it with a shuffle (and vector/vector bitcast) if possible. 1393/// 1394/// The source and destination vector types may have different element types. 1395static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, 1396 InstCombiner &IC) { 1397 // We can only do this optimization if the output is a multiple of the input 1398 // element size, or the input is a multiple of the output element size. 1399 // Convert the input type to have the same element type as the output. 1400 VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1401 1402 if (SrcTy->getElementType() != DestTy->getElementType()) { 1403 // The input types don't need to be identical, but for now they must be the 1404 // same size. There is no specific reason we couldn't handle things like 1405 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1406 // there yet. 1407 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1408 DestTy->getElementType()->getPrimitiveSizeInBits()) 1409 return 0; 1410 1411 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1412 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1413 } 1414 1415 // Now that the element types match, get the shuffle mask and RHS of the 1416 // shuffle to use, which depends on whether we're increasing or decreasing the 1417 // size of the input. 1418 SmallVector<uint32_t, 16> ShuffleMask; 1419 Value *V2; 1420 1421 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1422 // If we're shrinking the number of elements, just shuffle in the low 1423 // elements from the input and use undef as the second shuffle input. 1424 V2 = UndefValue::get(SrcTy); 1425 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1426 ShuffleMask.push_back(i); 1427 1428 } else { 1429 // If we're increasing the number of elements, shuffle in all of the 1430 // elements from InVal and fill the rest of the result elements with zeros 1431 // from a constant zero. 1432 V2 = Constant::getNullValue(SrcTy); 1433 unsigned SrcElts = SrcTy->getNumElements(); 1434 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1435 ShuffleMask.push_back(i); 1436 1437 // The excess elements reference the first element of the zero input. 1438 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i) 1439 ShuffleMask.push_back(SrcElts); 1440 } 1441 1442 return new ShuffleVectorInst(InVal, V2, 1443 ConstantDataVector::get(V2->getContext(), 1444 ShuffleMask)); 1445} 1446 1447static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { 1448 return Value % Ty->getPrimitiveSizeInBits() == 0; 1449} 1450 1451static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { 1452 return Value / Ty->getPrimitiveSizeInBits(); 1453} 1454 1455/// CollectInsertionElements - V is a value which is inserted into a vector of 1456/// VecEltTy. Look through the value to see if we can decompose it into 1457/// insertions into the vector. See the example in the comment for 1458/// OptimizeIntegerToVectorInsertions for the pattern this handles. 1459/// The type of V is always a non-zero multiple of VecEltTy's size. 1460/// 1461/// This returns false if the pattern can't be matched or true if it can, 1462/// filling in Elements with the elements found here. 1463static bool CollectInsertionElements(Value *V, unsigned ElementIndex, 1464 SmallVectorImpl<Value*> &Elements, 1465 Type *VecEltTy) { 1466 // Undef values never contribute useful bits to the result. 1467 if (isa<UndefValue>(V)) return true; 1468 1469 // If we got down to a value of the right type, we win, try inserting into the 1470 // right element. 1471 if (V->getType() == VecEltTy) { 1472 // Inserting null doesn't actually insert any elements. 1473 if (Constant *C = dyn_cast<Constant>(V)) 1474 if (C->isNullValue()) 1475 return true; 1476 1477 // Fail if multiple elements are inserted into this slot. 1478 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0) 1479 return false; 1480 1481 Elements[ElementIndex] = V; 1482 return true; 1483 } 1484 1485 if (Constant *C = dyn_cast<Constant>(V)) { 1486 // Figure out the # elements this provides, and bitcast it or slice it up 1487 // as required. 1488 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1489 VecEltTy); 1490 // If the constant is the size of a vector element, we just need to bitcast 1491 // it to the right type so it gets properly inserted. 1492 if (NumElts == 1) 1493 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1494 ElementIndex, Elements, VecEltTy); 1495 1496 // Okay, this is a constant that covers multiple elements. Slice it up into 1497 // pieces and insert each element-sized piece into the vector. 1498 if (!isa<IntegerType>(C->getType())) 1499 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1500 C->getType()->getPrimitiveSizeInBits())); 1501 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1502 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1503 1504 for (unsigned i = 0; i != NumElts; ++i) { 1505 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1506 i*ElementSize)); 1507 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1508 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy)) 1509 return false; 1510 } 1511 return true; 1512 } 1513 1514 if (!V->hasOneUse()) return false; 1515 1516 Instruction *I = dyn_cast<Instruction>(V); 1517 if (I == 0) return false; 1518 switch (I->getOpcode()) { 1519 default: return false; // Unhandled case. 1520 case Instruction::BitCast: 1521 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1522 Elements, VecEltTy); 1523 case Instruction::ZExt: 1524 if (!isMultipleOfTypeSize( 1525 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1526 VecEltTy)) 1527 return false; 1528 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1529 Elements, VecEltTy); 1530 case Instruction::Or: 1531 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1532 Elements, VecEltTy) && 1533 CollectInsertionElements(I->getOperand(1), ElementIndex, 1534 Elements, VecEltTy); 1535 case Instruction::Shl: { 1536 // Must be shifting by a constant that is a multiple of the element size. 1537 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1538 if (CI == 0) return false; 1539 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false; 1540 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy); 1541 1542 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift, 1543 Elements, VecEltTy); 1544 } 1545 1546 } 1547} 1548 1549 1550/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1551/// may be doing shifts and ors to assemble the elements of the vector manually. 1552/// Try to rip the code out and replace it with insertelements. This is to 1553/// optimize code like this: 1554/// 1555/// %tmp37 = bitcast float %inc to i32 1556/// %tmp38 = zext i32 %tmp37 to i64 1557/// %tmp31 = bitcast float %inc5 to i32 1558/// %tmp32 = zext i32 %tmp31 to i64 1559/// %tmp33 = shl i64 %tmp32, 32 1560/// %ins35 = or i64 %tmp33, %tmp38 1561/// %tmp43 = bitcast i64 %ins35 to <2 x float> 1562/// 1563/// Into two insertelements that do "buildvector{%inc, %inc5}". 1564static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1565 InstCombiner &IC) { 1566 VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1567 Value *IntInput = CI.getOperand(0); 1568 1569 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1570 if (!CollectInsertionElements(IntInput, 0, Elements, 1571 DestVecTy->getElementType())) 1572 return 0; 1573 1574 // If we succeeded, we know that all of the element are specified by Elements 1575 // or are zero if Elements has a null entry. Recast this as a set of 1576 // insertions. 1577 Value *Result = Constant::getNullValue(CI.getType()); 1578 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1579 if (Elements[i] == 0) continue; // Unset element. 1580 1581 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1582 IC.Builder->getInt32(i)); 1583 } 1584 1585 return Result; 1586} 1587 1588 1589/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1590/// bitcast. The various long double bitcasts can't get in here. 1591static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1592 Value *Src = CI.getOperand(0); 1593 Type *DestTy = CI.getType(); 1594 1595 // If this is a bitcast from int to float, check to see if the int is an 1596 // extraction from a vector. 1597 Value *VecInput = 0; 1598 // bitcast(trunc(bitcast(somevector))) 1599 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1600 isa<VectorType>(VecInput->getType())) { 1601 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1602 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1603 1604 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1605 // If the element type of the vector doesn't match the result type, 1606 // bitcast it to be a vector type we can extract from. 1607 if (VecTy->getElementType() != DestTy) { 1608 VecTy = VectorType::get(DestTy, 1609 VecTy->getPrimitiveSizeInBits() / DestWidth); 1610 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1611 } 1612 1613 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); 1614 } 1615 } 1616 1617 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1618 ConstantInt *ShAmt = 0; 1619 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1620 m_ConstantInt(ShAmt)))) && 1621 isa<VectorType>(VecInput->getType())) { 1622 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1623 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1624 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1625 ShAmt->getZExtValue() % DestWidth == 0) { 1626 // If the element type of the vector doesn't match the result type, 1627 // bitcast it to be a vector type we can extract from. 1628 if (VecTy->getElementType() != DestTy) { 1629 VecTy = VectorType::get(DestTy, 1630 VecTy->getPrimitiveSizeInBits() / DestWidth); 1631 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1632 } 1633 1634 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1635 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1636 } 1637 } 1638 return 0; 1639} 1640 1641Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1642 // If the operands are integer typed then apply the integer transforms, 1643 // otherwise just apply the common ones. 1644 Value *Src = CI.getOperand(0); 1645 Type *SrcTy = Src->getType(); 1646 Type *DestTy = CI.getType(); 1647 1648 // Get rid of casts from one type to the same type. These are useless and can 1649 // be replaced by the operand. 1650 if (DestTy == Src->getType()) 1651 return ReplaceInstUsesWith(CI, Src); 1652 1653 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1654 PointerType *SrcPTy = cast<PointerType>(SrcTy); 1655 Type *DstElTy = DstPTy->getElementType(); 1656 Type *SrcElTy = SrcPTy->getElementType(); 1657 1658 // If the address spaces don't match, don't eliminate the bitcast, which is 1659 // required for changing types. 1660 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1661 return 0; 1662 1663 // If we are casting a alloca to a pointer to a type of the same 1664 // size, rewrite the allocation instruction to allocate the "right" type. 1665 // There is no need to modify malloc calls because it is their bitcast that 1666 // needs to be cleaned up. 1667 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1668 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1669 return V; 1670 1671 // If the source and destination are pointers, and this cast is equivalent 1672 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1673 // This can enhance SROA and other transforms that want type-safe pointers. 1674 Constant *ZeroUInt = 1675 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1676 unsigned NumZeros = 0; 1677 while (SrcElTy != DstElTy && 1678 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1679 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1680 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1681 ++NumZeros; 1682 } 1683 1684 // If we found a path from the src to dest, create the getelementptr now. 1685 if (SrcElTy == DstElTy) { 1686 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1687 return GetElementPtrInst::CreateInBounds(Src, Idxs); 1688 } 1689 } 1690 1691 // Try to optimize int -> float bitcasts. 1692 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1693 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1694 return I; 1695 1696 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1697 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1698 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1699 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1700 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1701 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1702 } 1703 1704 if (isa<IntegerType>(SrcTy)) { 1705 // If this is a cast from an integer to vector, check to see if the input 1706 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1707 // the casts with a shuffle and (potentially) a bitcast. 1708 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1709 CastInst *SrcCast = cast<CastInst>(Src); 1710 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1711 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1712 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1713 cast<VectorType>(DestTy), *this)) 1714 return I; 1715 } 1716 1717 // If the input is an 'or' instruction, we may be doing shifts and ors to 1718 // assemble the elements of the vector manually. Try to rip the code out 1719 // and replace it with insertelements. 1720 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1721 return ReplaceInstUsesWith(CI, V); 1722 } 1723 } 1724 1725 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1726 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1727 Value *Elem = 1728 Builder->CreateExtractElement(Src, 1729 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1730 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1731 } 1732 } 1733 1734 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1735 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1736 // a bitcast to a vector with the same # elts. 1737 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1738 cast<VectorType>(DestTy)->getNumElements() == 1739 SVI->getType()->getNumElements() && 1740 SVI->getType()->getNumElements() == 1741 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1742 BitCastInst *Tmp; 1743 // If either of the operands is a cast from CI.getType(), then 1744 // evaluating the shuffle in the casted destination's type will allow 1745 // us to eliminate at least one cast. 1746 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1747 Tmp->getOperand(0)->getType() == DestTy) || 1748 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1749 Tmp->getOperand(0)->getType() == DestTy)) { 1750 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1751 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1752 // Return a new shuffle vector. Use the same element ID's, as we 1753 // know the vector types match #elts. 1754 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1755 } 1756 } 1757 } 1758 1759 if (SrcTy->isPointerTy()) 1760 return commonPointerCastTransforms(CI); 1761 return commonCastTransforms(CI); 1762} 1763