InstCombineCasts.cpp revision 3dd08734c1812e47ae5f6aceba15f28865f75943
1//===- InstCombineCasts.cpp -----------------------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the visit functions for cast operations. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Target/TargetData.h" 16#include "llvm/Support/PatternMatch.h" 17using namespace llvm; 18using namespace PatternMatch; 19 20/// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear 21/// expression. If so, decompose it, returning some value X, such that Val is 22/// X*Scale+Offset. 23/// 24static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, 25 uint64_t &Offset) { 26 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 27 Offset = CI->getZExtValue(); 28 Scale = 0; 29 return ConstantInt::get(Val->getType(), 0); 30 } 31 32 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) { 33 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 34 if (I->getOpcode() == Instruction::Shl) { 35 // This is a value scaled by '1 << the shift amt'. 36 Scale = UINT64_C(1) << RHS->getZExtValue(); 37 Offset = 0; 38 return I->getOperand(0); 39 } 40 41 if (I->getOpcode() == Instruction::Mul) { 42 // This value is scaled by 'RHS'. 43 Scale = RHS->getZExtValue(); 44 Offset = 0; 45 return I->getOperand(0); 46 } 47 48 if (I->getOpcode() == Instruction::Add) { 49 // We have X+C. Check to see if we really have (X*C2)+C1, 50 // where C1 is divisible by C2. 51 unsigned SubScale; 52 Value *SubVal = 53 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); 54 Offset += RHS->getZExtValue(); 55 Scale = SubScale; 56 return SubVal; 57 } 58 } 59 } 60 61 // Otherwise, we can't look past this. 62 Scale = 1; 63 Offset = 0; 64 return Val; 65} 66 67/// PromoteCastOfAllocation - If we find a cast of an allocation instruction, 68/// try to eliminate the cast by moving the type information into the alloc. 69Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 70 AllocaInst &AI) { 71 // This requires TargetData to get the alloca alignment and size information. 72 if (!TD) return 0; 73 74 const PointerType *PTy = cast<PointerType>(CI.getType()); 75 76 BuilderTy AllocaBuilder(*Builder); 77 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); 78 79 // Get the type really allocated and the type casted to. 80 const Type *AllocElTy = AI.getAllocatedType(); 81 const Type *CastElTy = PTy->getElementType(); 82 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0; 83 84 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy); 85 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy); 86 if (CastElTyAlign < AllocElTyAlign) return 0; 87 88 // If the allocation has multiple uses, only promote it if we are strictly 89 // increasing the alignment of the resultant allocation. If we keep it the 90 // same, we open the door to infinite loops of various kinds. (A reference 91 // from a dbg.declare doesn't count as a use for this purpose.) 92 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) && 93 CastElTyAlign == AllocElTyAlign) return 0; 94 95 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy); 96 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); 97 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 98 99 // See if we can satisfy the modulus by pulling a scale out of the array 100 // size argument. 101 unsigned ArraySizeScale; 102 uint64_t ArrayOffset; 103 Value *NumElements = // See if the array size is a decomposable linear expr. 104 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 105 106 // If we can now satisfy the modulus, by using a non-1 scale, we really can 107 // do the xform. 108 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 109 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 110 111 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 112 Value *Amt = 0; 113 if (Scale == 1) { 114 Amt = NumElements; 115 } else { 116 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 117 // Insert before the alloca, not before the cast. 118 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp"); 119 } 120 121 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 122 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 123 Offset, true); 124 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp"); 125 } 126 127 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 128 New->setAlignment(AI.getAlignment()); 129 New->takeName(&AI); 130 131 // If the allocation has one real use plus a dbg.declare, just remove the 132 // declare. 133 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) { 134 EraseInstFromFunction(*(Instruction*)DI); 135 } 136 // If the allocation has multiple real uses, insert a cast and change all 137 // things that used it to use the new cast. This will also hack on CI, but it 138 // will die soon. 139 else if (!AI.hasOneUse()) { 140 // New is the allocation instruction, pointer typed. AI is the original 141 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 142 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 143 AI.replaceAllUsesWith(NewCast); 144 } 145 return ReplaceInstUsesWith(CI, New); 146} 147 148 149 150/// EvaluateInDifferentType - Given an expression that 151/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 152/// insert the code to evaluate the expression. 153Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty, 154 bool isSigned) { 155 if (Constant *C = dyn_cast<Constant>(V)) { 156 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 157 // If we got a constantexpr back, try to simplify it with TD info. 158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 159 C = ConstantFoldConstantExpression(CE, TD); 160 return C; 161 } 162 163 // Otherwise, it must be an instruction. 164 Instruction *I = cast<Instruction>(V); 165 Instruction *Res = 0; 166 unsigned Opc = I->getOpcode(); 167 switch (Opc) { 168 case Instruction::Add: 169 case Instruction::Sub: 170 case Instruction::Mul: 171 case Instruction::And: 172 case Instruction::Or: 173 case Instruction::Xor: 174 case Instruction::AShr: 175 case Instruction::LShr: 176 case Instruction::Shl: 177 case Instruction::UDiv: 178 case Instruction::URem: { 179 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 180 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 181 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 182 break; 183 } 184 case Instruction::Trunc: 185 case Instruction::ZExt: 186 case Instruction::SExt: 187 // If the source type of the cast is the type we're trying for then we can 188 // just return the source. There's no need to insert it because it is not 189 // new. 190 if (I->getOperand(0)->getType() == Ty) 191 return I->getOperand(0); 192 193 // Otherwise, must be the same type of cast, so just reinsert a new one. 194 // This also handles the case of zext(trunc(x)) -> zext(x). 195 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 196 Opc == Instruction::SExt); 197 break; 198 case Instruction::Select: { 199 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 200 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 201 Res = SelectInst::Create(I->getOperand(0), True, False); 202 break; 203 } 204 case Instruction::PHI: { 205 PHINode *OPN = cast<PHINode>(I); 206 PHINode *NPN = PHINode::Create(Ty); 207 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 208 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 209 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 210 } 211 Res = NPN; 212 break; 213 } 214 default: 215 // TODO: Can handle more cases here. 216 llvm_unreachable("Unreachable!"); 217 break; 218 } 219 220 Res->takeName(I); 221 return InsertNewInstBefore(Res, *I); 222} 223 224 225/// This function is a wrapper around CastInst::isEliminableCastPair. It 226/// simply extracts arguments and returns what that function returns. 227static Instruction::CastOps 228isEliminableCastPair( 229 const CastInst *CI, ///< The first cast instruction 230 unsigned opcode, ///< The opcode of the second cast instruction 231 const Type *DstTy, ///< The target type for the second cast instruction 232 TargetData *TD ///< The target data for pointer size 233) { 234 235 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above 236 const Type *MidTy = CI->getType(); // B from above 237 238 // Get the opcodes of the two Cast instructions 239 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 240 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 241 242 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 243 DstTy, 244 TD ? TD->getIntPtrType(CI->getContext()) : 0); 245 246 // We don't want to form an inttoptr or ptrtoint that converts to an integer 247 // type that differs from the pointer size. 248 if ((Res == Instruction::IntToPtr && 249 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 250 (Res == Instruction::PtrToInt && 251 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 252 Res = 0; 253 254 return Instruction::CastOps(Res); 255} 256 257/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 258/// results in any code being generated and is interesting to optimize out. If 259/// the cast can be eliminated by some other simple transformation, we prefer 260/// to do the simplification first. 261bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 262 const Type *Ty) { 263 // Noop casts and casts of constants should be eliminated trivially. 264 if (V->getType() == Ty || isa<Constant>(V)) return false; 265 266 // If this is another cast that can be eliminated, we prefer to have it 267 // eliminated. 268 if (const CastInst *CI = dyn_cast<CastInst>(V)) 269 if (isEliminableCastPair(CI, opc, Ty, TD)) 270 return false; 271 272 // If this is a vector sext from a compare, then we don't want to break the 273 // idiom where each element of the extended vector is either zero or all ones. 274 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 275 return false; 276 277 return true; 278} 279 280 281/// @brief Implement the transforms common to all CastInst visitors. 282Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 283 Value *Src = CI.getOperand(0); 284 285 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 286 // eliminate it now. 287 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 288 if (Instruction::CastOps opc = 289 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 290 // The first cast (CSrc) is eliminable so we need to fix up or replace 291 // the second cast (CI). CSrc will then have a good chance of being dead. 292 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 293 } 294 } 295 296 // If we are casting a select then fold the cast into the select 297 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 298 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 299 return NV; 300 301 // If we are casting a PHI then fold the cast into the PHI 302 if (isa<PHINode>(Src)) { 303 // We don't do this if this would create a PHI node with an illegal type if 304 // it is currently legal. 305 if (!Src->getType()->isIntegerTy() || 306 !CI.getType()->isIntegerTy() || 307 ShouldChangeType(CI.getType(), Src->getType())) 308 if (Instruction *NV = FoldOpIntoPhi(CI)) 309 return NV; 310 } 311 312 return 0; 313} 314 315/// CanEvaluateTruncated - Return true if we can evaluate the specified 316/// expression tree as type Ty instead of its larger type, and arrive with the 317/// same value. This is used by code that tries to eliminate truncates. 318/// 319/// Ty will always be a type smaller than V. We should return true if trunc(V) 320/// can be computed by computing V in the smaller type. If V is an instruction, 321/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 322/// makes sense if x and y can be efficiently truncated. 323/// 324/// This function works on both vectors and scalars. 325/// 326static bool CanEvaluateTruncated(Value *V, const Type *Ty) { 327 // We can always evaluate constants in another type. 328 if (isa<Constant>(V)) 329 return true; 330 331 Instruction *I = dyn_cast<Instruction>(V); 332 if (!I) return false; 333 334 const Type *OrigTy = V->getType(); 335 336 // If this is an extension from the dest type, we can eliminate it, even if it 337 // has multiple uses. 338 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 339 I->getOperand(0)->getType() == Ty) 340 return true; 341 342 // We can't extend or shrink something that has multiple uses: doing so would 343 // require duplicating the instruction in general, which isn't profitable. 344 if (!I->hasOneUse()) return false; 345 346 unsigned Opc = I->getOpcode(); 347 switch (Opc) { 348 case Instruction::Add: 349 case Instruction::Sub: 350 case Instruction::Mul: 351 case Instruction::And: 352 case Instruction::Or: 353 case Instruction::Xor: 354 // These operators can all arbitrarily be extended or truncated. 355 return CanEvaluateTruncated(I->getOperand(0), Ty) && 356 CanEvaluateTruncated(I->getOperand(1), Ty); 357 358 case Instruction::UDiv: 359 case Instruction::URem: { 360 // UDiv and URem can be truncated if all the truncated bits are zero. 361 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 362 uint32_t BitWidth = Ty->getScalarSizeInBits(); 363 if (BitWidth < OrigBitWidth) { 364 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 365 if (MaskedValueIsZero(I->getOperand(0), Mask) && 366 MaskedValueIsZero(I->getOperand(1), Mask)) { 367 return CanEvaluateTruncated(I->getOperand(0), Ty) && 368 CanEvaluateTruncated(I->getOperand(1), Ty); 369 } 370 } 371 break; 372 } 373 case Instruction::Shl: 374 // If we are truncating the result of this SHL, and if it's a shift of a 375 // constant amount, we can always perform a SHL in a smaller type. 376 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 377 uint32_t BitWidth = Ty->getScalarSizeInBits(); 378 if (CI->getLimitedValue(BitWidth) < BitWidth) 379 return CanEvaluateTruncated(I->getOperand(0), Ty); 380 } 381 break; 382 case Instruction::LShr: 383 // If this is a truncate of a logical shr, we can truncate it to a smaller 384 // lshr iff we know that the bits we would otherwise be shifting in are 385 // already zeros. 386 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 387 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 388 uint32_t BitWidth = Ty->getScalarSizeInBits(); 389 if (MaskedValueIsZero(I->getOperand(0), 390 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 391 CI->getLimitedValue(BitWidth) < BitWidth) { 392 return CanEvaluateTruncated(I->getOperand(0), Ty); 393 } 394 } 395 break; 396 case Instruction::Trunc: 397 // trunc(trunc(x)) -> trunc(x) 398 return true; 399 case Instruction::ZExt: 400 case Instruction::SExt: 401 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 402 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 403 return true; 404 case Instruction::Select: { 405 SelectInst *SI = cast<SelectInst>(I); 406 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 407 CanEvaluateTruncated(SI->getFalseValue(), Ty); 408 } 409 case Instruction::PHI: { 410 // We can change a phi if we can change all operands. Note that we never 411 // get into trouble with cyclic PHIs here because we only consider 412 // instructions with a single use. 413 PHINode *PN = cast<PHINode>(I); 414 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 415 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 416 return false; 417 return true; 418 } 419 default: 420 // TODO: Can handle more cases here. 421 break; 422 } 423 424 return false; 425} 426 427Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 428 if (Instruction *Result = commonCastTransforms(CI)) 429 return Result; 430 431 // See if we can simplify any instructions used by the input whose sole 432 // purpose is to compute bits we don't care about. 433 if (SimplifyDemandedInstructionBits(CI)) 434 return &CI; 435 436 Value *Src = CI.getOperand(0); 437 const Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 438 439 // Attempt to truncate the entire input expression tree to the destination 440 // type. Only do this if the dest type is a simple type, don't convert the 441 // expression tree to something weird like i93 unless the source is also 442 // strange. 443 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 444 CanEvaluateTruncated(Src, DestTy)) { 445 446 // If this cast is a truncate, evaluting in a different type always 447 // eliminates the cast, so it is always a win. 448 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 449 " to avoid cast: " << CI << '\n'); 450 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 451 assert(Res->getType() == DestTy); 452 return ReplaceInstUsesWith(CI, Res); 453 } 454 455 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 456 if (DestTy->getScalarSizeInBits() == 1) { 457 Constant *One = ConstantInt::get(Src->getType(), 1); 458 Src = Builder->CreateAnd(Src, One, "tmp"); 459 Value *Zero = Constant::getNullValue(Src->getType()); 460 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 461 } 462 463 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 464 Value *A = 0; ConstantInt *Cst = 0; 465 if (match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst))) && 466 Src->hasOneUse()) { 467 // We have three types to worry about here, the type of A, the source of 468 // the truncate (MidSize), and the destination of the truncate. We know that 469 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 470 // between ASize and ResultSize. 471 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 472 473 // If the shift amount is larger than the size of A, then the result is 474 // known to be zero because all the input bits got shifted out. 475 if (Cst->getZExtValue() >= ASize) 476 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType())); 477 478 // Since we're doing an lshr and a zero extend, and know that the shift 479 // amount is smaller than ASize, it is always safe to do the shift in A's 480 // type, then zero extend or truncate to the result. 481 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 482 Shift->takeName(Src); 483 return CastInst::CreateIntegerCast(Shift, CI.getType(), false); 484 } 485 486 return 0; 487} 488 489/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 490/// in order to eliminate the icmp. 491Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 492 bool DoXform) { 493 // If we are just checking for a icmp eq of a single bit and zext'ing it 494 // to an integer, then shift the bit to the appropriate place and then 495 // cast to integer to avoid the comparison. 496 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 497 const APInt &Op1CV = Op1C->getValue(); 498 499 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 500 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 501 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 502 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 503 if (!DoXform) return ICI; 504 505 Value *In = ICI->getOperand(0); 506 Value *Sh = ConstantInt::get(In->getType(), 507 In->getType()->getScalarSizeInBits()-1); 508 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 509 if (In->getType() != CI.getType()) 510 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp"); 511 512 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 513 Constant *One = ConstantInt::get(In->getType(), 1); 514 In = Builder->CreateXor(In, One, In->getName()+".not"); 515 } 516 517 return ReplaceInstUsesWith(CI, In); 518 } 519 520 521 522 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 523 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 524 // zext (X == 1) to i32 --> X iff X has only the low bit set. 525 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 526 // zext (X != 0) to i32 --> X iff X has only the low bit set. 527 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 528 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 529 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 530 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 531 // This only works for EQ and NE 532 ICI->isEquality()) { 533 // If Op1C some other power of two, convert: 534 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 535 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 536 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 537 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 538 539 APInt KnownZeroMask(~KnownZero); 540 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 541 if (!DoXform) return ICI; 542 543 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 544 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 545 // (X&4) == 2 --> false 546 // (X&4) != 2 --> true 547 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 548 isNE); 549 Res = ConstantExpr::getZExt(Res, CI.getType()); 550 return ReplaceInstUsesWith(CI, Res); 551 } 552 553 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 554 Value *In = ICI->getOperand(0); 555 if (ShiftAmt) { 556 // Perform a logical shr by shiftamt. 557 // Insert the shift to put the result in the low bit. 558 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 559 In->getName()+".lobit"); 560 } 561 562 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 563 Constant *One = ConstantInt::get(In->getType(), 1); 564 In = Builder->CreateXor(In, One, "tmp"); 565 } 566 567 if (CI.getType() == In->getType()) 568 return ReplaceInstUsesWith(CI, In); 569 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 570 } 571 } 572 } 573 574 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 575 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 576 // may lead to additional simplifications. 577 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 578 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 579 uint32_t BitWidth = ITy->getBitWidth(); 580 Value *LHS = ICI->getOperand(0); 581 Value *RHS = ICI->getOperand(1); 582 583 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 584 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 585 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 586 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 587 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 588 589 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 590 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 591 APInt UnknownBit = ~KnownBits; 592 if (UnknownBit.countPopulation() == 1) { 593 if (!DoXform) return ICI; 594 595 Value *Result = Builder->CreateXor(LHS, RHS); 596 597 // Mask off any bits that are set and won't be shifted away. 598 if (KnownOneLHS.uge(UnknownBit)) 599 Result = Builder->CreateAnd(Result, 600 ConstantInt::get(ITy, UnknownBit)); 601 602 // Shift the bit we're testing down to the lsb. 603 Result = Builder->CreateLShr( 604 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 605 606 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 607 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 608 Result->takeName(ICI); 609 return ReplaceInstUsesWith(CI, Result); 610 } 611 } 612 } 613 } 614 615 return 0; 616} 617 618/// CanEvaluateZExtd - Determine if the specified value can be computed in the 619/// specified wider type and produce the same low bits. If not, return false. 620/// 621/// If this function returns true, it can also return a non-zero number of bits 622/// (in BitsToClear) which indicates that the value it computes is correct for 623/// the zero extend, but that the additional BitsToClear bits need to be zero'd 624/// out. For example, to promote something like: 625/// 626/// %B = trunc i64 %A to i32 627/// %C = lshr i32 %B, 8 628/// %E = zext i32 %C to i64 629/// 630/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 631/// set to 8 to indicate that the promoted value needs to have bits 24-31 632/// cleared in addition to bits 32-63. Since an 'and' will be generated to 633/// clear the top bits anyway, doing this has no extra cost. 634/// 635/// This function works on both vectors and scalars. 636static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) { 637 BitsToClear = 0; 638 if (isa<Constant>(V)) 639 return true; 640 641 Instruction *I = dyn_cast<Instruction>(V); 642 if (!I) return false; 643 644 // If the input is a truncate from the destination type, we can trivially 645 // eliminate it, even if it has multiple uses. 646 // FIXME: This is currently disabled until codegen can handle this without 647 // pessimizing code, PR5997. 648 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 649 return true; 650 651 // We can't extend or shrink something that has multiple uses: doing so would 652 // require duplicating the instruction in general, which isn't profitable. 653 if (!I->hasOneUse()) return false; 654 655 unsigned Opc = I->getOpcode(), Tmp; 656 switch (Opc) { 657 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 658 case Instruction::SExt: // zext(sext(x)) -> sext(x). 659 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 660 return true; 661 case Instruction::And: 662 case Instruction::Or: 663 case Instruction::Xor: 664 case Instruction::Add: 665 case Instruction::Sub: 666 case Instruction::Mul: 667 case Instruction::Shl: 668 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 669 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 670 return false; 671 // These can all be promoted if neither operand has 'bits to clear'. 672 if (BitsToClear == 0 && Tmp == 0) 673 return true; 674 675 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 676 // other side, BitsToClear is ok. 677 if (Tmp == 0 && 678 (Opc == Instruction::And || Opc == Instruction::Or || 679 Opc == Instruction::Xor)) { 680 // We use MaskedValueIsZero here for generality, but the case we care 681 // about the most is constant RHS. 682 unsigned VSize = V->getType()->getScalarSizeInBits(); 683 if (MaskedValueIsZero(I->getOperand(1), 684 APInt::getHighBitsSet(VSize, BitsToClear))) 685 return true; 686 } 687 688 // Otherwise, we don't know how to analyze this BitsToClear case yet. 689 return false; 690 691 case Instruction::LShr: 692 // We can promote lshr(x, cst) if we can promote x. This requires the 693 // ultimate 'and' to clear out the high zero bits we're clearing out though. 694 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 695 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 696 return false; 697 BitsToClear += Amt->getZExtValue(); 698 if (BitsToClear > V->getType()->getScalarSizeInBits()) 699 BitsToClear = V->getType()->getScalarSizeInBits(); 700 return true; 701 } 702 // Cannot promote variable LSHR. 703 return false; 704 case Instruction::Select: 705 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 706 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 707 // TODO: If important, we could handle the case when the BitsToClear are 708 // known zero in the disagreeing side. 709 Tmp != BitsToClear) 710 return false; 711 return true; 712 713 case Instruction::PHI: { 714 // We can change a phi if we can change all operands. Note that we never 715 // get into trouble with cyclic PHIs here because we only consider 716 // instructions with a single use. 717 PHINode *PN = cast<PHINode>(I); 718 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 719 return false; 720 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 721 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 722 // TODO: If important, we could handle the case when the BitsToClear 723 // are known zero in the disagreeing input. 724 Tmp != BitsToClear) 725 return false; 726 return true; 727 } 728 default: 729 // TODO: Can handle more cases here. 730 return false; 731 } 732} 733 734Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 735 // If this zero extend is only used by a truncate, let the truncate by 736 // eliminated before we try to optimize this zext. 737 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 738 return 0; 739 740 // If one of the common conversion will work, do it. 741 if (Instruction *Result = commonCastTransforms(CI)) 742 return Result; 743 744 // See if we can simplify any instructions used by the input whose sole 745 // purpose is to compute bits we don't care about. 746 if (SimplifyDemandedInstructionBits(CI)) 747 return &CI; 748 749 Value *Src = CI.getOperand(0); 750 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 751 752 // Attempt to extend the entire input expression tree to the destination 753 // type. Only do this if the dest type is a simple type, don't convert the 754 // expression tree to something weird like i93 unless the source is also 755 // strange. 756 unsigned BitsToClear; 757 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 758 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 759 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 760 "Unreasonable BitsToClear"); 761 762 // Okay, we can transform this! Insert the new expression now. 763 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 764 " to avoid zero extend: " << CI); 765 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 766 assert(Res->getType() == DestTy); 767 768 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 769 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 770 771 // If the high bits are already filled with zeros, just replace this 772 // cast with the result. 773 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 774 DestBitSize-SrcBitsKept))) 775 return ReplaceInstUsesWith(CI, Res); 776 777 // We need to emit an AND to clear the high bits. 778 Constant *C = ConstantInt::get(Res->getType(), 779 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 780 return BinaryOperator::CreateAnd(Res, C); 781 } 782 783 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 784 // types and if the sizes are just right we can convert this into a logical 785 // 'and' which will be much cheaper than the pair of casts. 786 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 787 // TODO: Subsume this into EvaluateInDifferentType. 788 789 // Get the sizes of the types involved. We know that the intermediate type 790 // will be smaller than A or C, but don't know the relation between A and C. 791 Value *A = CSrc->getOperand(0); 792 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 793 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 794 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 795 // If we're actually extending zero bits, then if 796 // SrcSize < DstSize: zext(a & mask) 797 // SrcSize == DstSize: a & mask 798 // SrcSize > DstSize: trunc(a) & mask 799 if (SrcSize < DstSize) { 800 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 801 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 802 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 803 return new ZExtInst(And, CI.getType()); 804 } 805 806 if (SrcSize == DstSize) { 807 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 808 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 809 AndValue)); 810 } 811 if (SrcSize > DstSize) { 812 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp"); 813 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 814 return BinaryOperator::CreateAnd(Trunc, 815 ConstantInt::get(Trunc->getType(), 816 AndValue)); 817 } 818 } 819 820 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 821 return transformZExtICmp(ICI, CI); 822 823 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 824 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 825 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 826 // of the (zext icmp) will be transformed. 827 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 828 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 829 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 830 (transformZExtICmp(LHS, CI, false) || 831 transformZExtICmp(RHS, CI, false))) { 832 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 833 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 834 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 835 } 836 } 837 838 // zext(trunc(t) & C) -> (t & zext(C)). 839 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 840 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 841 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 842 Value *TI0 = TI->getOperand(0); 843 if (TI0->getType() == CI.getType()) 844 return 845 BinaryOperator::CreateAnd(TI0, 846 ConstantExpr::getZExt(C, CI.getType())); 847 } 848 849 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 850 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 851 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 852 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 853 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 854 And->getOperand(1) == C) 855 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 856 Value *TI0 = TI->getOperand(0); 857 if (TI0->getType() == CI.getType()) { 858 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 859 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp"); 860 return BinaryOperator::CreateXor(NewAnd, ZC); 861 } 862 } 863 864 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 865 Value *X; 866 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 867 match(SrcI, m_Not(m_Value(X))) && 868 (!X->hasOneUse() || !isa<CmpInst>(X))) { 869 Value *New = Builder->CreateZExt(X, CI.getType()); 870 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 871 } 872 873 return 0; 874} 875 876/// CanEvaluateSExtd - Return true if we can take the specified value 877/// and return it as type Ty without inserting any new casts and without 878/// changing the value of the common low bits. This is used by code that tries 879/// to promote integer operations to a wider types will allow us to eliminate 880/// the extension. 881/// 882/// This function works on both vectors and scalars. 883/// 884static bool CanEvaluateSExtd(Value *V, const Type *Ty) { 885 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 886 "Can't sign extend type to a smaller type"); 887 // If this is a constant, it can be trivially promoted. 888 if (isa<Constant>(V)) 889 return true; 890 891 Instruction *I = dyn_cast<Instruction>(V); 892 if (!I) return false; 893 894 // If this is a truncate from the dest type, we can trivially eliminate it, 895 // even if it has multiple uses. 896 // FIXME: This is currently disabled until codegen can handle this without 897 // pessimizing code, PR5997. 898 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 899 return true; 900 901 // We can't extend or shrink something that has multiple uses: doing so would 902 // require duplicating the instruction in general, which isn't profitable. 903 if (!I->hasOneUse()) return false; 904 905 switch (I->getOpcode()) { 906 case Instruction::SExt: // sext(sext(x)) -> sext(x) 907 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 908 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 909 return true; 910 case Instruction::And: 911 case Instruction::Or: 912 case Instruction::Xor: 913 case Instruction::Add: 914 case Instruction::Sub: 915 case Instruction::Mul: 916 // These operators can all arbitrarily be extended if their inputs can. 917 return CanEvaluateSExtd(I->getOperand(0), Ty) && 918 CanEvaluateSExtd(I->getOperand(1), Ty); 919 920 //case Instruction::Shl: TODO 921 //case Instruction::LShr: TODO 922 923 case Instruction::Select: 924 return CanEvaluateSExtd(I->getOperand(1), Ty) && 925 CanEvaluateSExtd(I->getOperand(2), Ty); 926 927 case Instruction::PHI: { 928 // We can change a phi if we can change all operands. Note that we never 929 // get into trouble with cyclic PHIs here because we only consider 930 // instructions with a single use. 931 PHINode *PN = cast<PHINode>(I); 932 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 933 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 934 return true; 935 } 936 default: 937 // TODO: Can handle more cases here. 938 break; 939 } 940 941 return false; 942} 943 944Instruction *InstCombiner::visitSExt(SExtInst &CI) { 945 // If this sign extend is only used by a truncate, let the truncate by 946 // eliminated before we try to optimize this zext. 947 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 948 return 0; 949 950 if (Instruction *I = commonCastTransforms(CI)) 951 return I; 952 953 // See if we can simplify any instructions used by the input whose sole 954 // purpose is to compute bits we don't care about. 955 if (SimplifyDemandedInstructionBits(CI)) 956 return &CI; 957 958 Value *Src = CI.getOperand(0); 959 const Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 960 961 // Attempt to extend the entire input expression tree to the destination 962 // type. Only do this if the dest type is a simple type, don't convert the 963 // expression tree to something weird like i93 unless the source is also 964 // strange. 965 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 966 CanEvaluateSExtd(Src, DestTy)) { 967 // Okay, we can transform this! Insert the new expression now. 968 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 969 " to avoid sign extend: " << CI); 970 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 971 assert(Res->getType() == DestTy); 972 973 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 974 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 975 976 // If the high bits are already filled with sign bit, just replace this 977 // cast with the result. 978 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 979 return ReplaceInstUsesWith(CI, Res); 980 981 // We need to emit a shl + ashr to do the sign extend. 982 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 983 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 984 ShAmt); 985 } 986 987 // If this input is a trunc from our destination, then turn sext(trunc(x)) 988 // into shifts. 989 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 990 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 991 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 992 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 993 994 // We need to emit a shl + ashr to do the sign extend. 995 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 996 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 997 return BinaryOperator::CreateAShr(Res, ShAmt); 998 } 999 1000 1001 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed 1002 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed 1003 { 1004 ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS; 1005 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) { 1006 // sext (x <s 0) to i32 --> x>>s31 true if signbit set. 1007 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear. 1008 if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) || 1009 (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) { 1010 Value *Sh = ConstantInt::get(CmpLHS->getType(), 1011 CmpLHS->getType()->getScalarSizeInBits()-1); 1012 Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit"); 1013 if (In->getType() != CI.getType()) 1014 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp"); 1015 1016 if (Pred == ICmpInst::ICMP_SGT) 1017 In = Builder->CreateNot(In, In->getName()+".not"); 1018 return ReplaceInstUsesWith(CI, In); 1019 } 1020 } 1021 } 1022 1023 1024 // If the input is a shl/ashr pair of a same constant, then this is a sign 1025 // extension from a smaller value. If we could trust arbitrary bitwidth 1026 // integers, we could turn this into a truncate to the smaller bit and then 1027 // use a sext for the whole extension. Since we don't, look deeper and check 1028 // for a truncate. If the source and dest are the same type, eliminate the 1029 // trunc and extend and just do shifts. For example, turn: 1030 // %a = trunc i32 %i to i8 1031 // %b = shl i8 %a, 6 1032 // %c = ashr i8 %b, 6 1033 // %d = sext i8 %c to i32 1034 // into: 1035 // %a = shl i32 %i, 30 1036 // %d = ashr i32 %a, 30 1037 Value *A = 0; 1038 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1039 ConstantInt *BA = 0, *CA = 0; 1040 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1041 m_ConstantInt(CA))) && 1042 BA == CA && A->getType() == CI.getType()) { 1043 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1044 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1045 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1046 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1047 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1048 return BinaryOperator::CreateAShr(A, ShAmtV); 1049 } 1050 1051 return 0; 1052} 1053 1054 1055/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1056/// in the specified FP type without changing its value. 1057static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1058 bool losesInfo; 1059 APFloat F = CFP->getValueAPF(); 1060 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1061 if (!losesInfo) 1062 return ConstantFP::get(CFP->getContext(), F); 1063 return 0; 1064} 1065 1066/// LookThroughFPExtensions - If this is an fp extension instruction, look 1067/// through it until we get the source value. 1068static Value *LookThroughFPExtensions(Value *V) { 1069 if (Instruction *I = dyn_cast<Instruction>(V)) 1070 if (I->getOpcode() == Instruction::FPExt) 1071 return LookThroughFPExtensions(I->getOperand(0)); 1072 1073 // If this value is a constant, return the constant in the smallest FP type 1074 // that can accurately represent it. This allows us to turn 1075 // (float)((double)X+2.0) into x+2.0f. 1076 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1077 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1078 return V; // No constant folding of this. 1079 // See if the value can be truncated to float and then reextended. 1080 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1081 return V; 1082 if (CFP->getType()->isDoubleTy()) 1083 return V; // Won't shrink. 1084 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1085 return V; 1086 // Don't try to shrink to various long double types. 1087 } 1088 1089 return V; 1090} 1091 1092Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1093 if (Instruction *I = commonCastTransforms(CI)) 1094 return I; 1095 1096 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1097 // smaller than the destination type, we can eliminate the truncate by doing 1098 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1099 // as many builtins (sqrt, etc). 1100 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1101 if (OpI && OpI->hasOneUse()) { 1102 switch (OpI->getOpcode()) { 1103 default: break; 1104 case Instruction::FAdd: 1105 case Instruction::FSub: 1106 case Instruction::FMul: 1107 case Instruction::FDiv: 1108 case Instruction::FRem: 1109 const Type *SrcTy = OpI->getType(); 1110 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1111 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1112 if (LHSTrunc->getType() != SrcTy && 1113 RHSTrunc->getType() != SrcTy) { 1114 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1115 // If the source types were both smaller than the destination type of 1116 // the cast, do this xform. 1117 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1118 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1119 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1120 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1121 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1122 } 1123 } 1124 break; 1125 } 1126 } 1127 1128 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1129 // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it. 1130 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1131 if (Call && Call->getCalledFunction() && 1132 Call->getCalledFunction()->getName() == "sqrt" && 1133 Call->getNumArgOperands() == 1) { 1134 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1135 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1136 CI.getType()->isFloatTy() && 1137 Call->getType()->isDoubleTy() && 1138 Arg->getType()->isDoubleTy() && 1139 Arg->getOperand(0)->getType()->isFloatTy()) { 1140 Function *Callee = Call->getCalledFunction(); 1141 Module *M = CI.getParent()->getParent()->getParent(); 1142 Constant* SqrtfFunc = M->getOrInsertFunction("sqrtf", 1143 Callee->getAttributes(), 1144 Builder->getFloatTy(), 1145 Builder->getFloatTy(), 1146 NULL); 1147 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1148 "sqrtfcall"); 1149 ret->setAttributes(Callee->getAttributes()); 1150 return ret; 1151 } 1152 } 1153 1154 return 0; 1155} 1156 1157Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1158 return commonCastTransforms(CI); 1159} 1160 1161Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1162 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1163 if (OpI == 0) 1164 return commonCastTransforms(FI); 1165 1166 // fptoui(uitofp(X)) --> X 1167 // fptoui(sitofp(X)) --> X 1168 // This is safe if the intermediate type has enough bits in its mantissa to 1169 // accurately represent all values of X. For example, do not do this with 1170 // i64->float->i64. This is also safe for sitofp case, because any negative 1171 // 'X' value would cause an undefined result for the fptoui. 1172 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1173 OpI->getOperand(0)->getType() == FI.getType() && 1174 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1175 OpI->getType()->getFPMantissaWidth()) 1176 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1177 1178 return commonCastTransforms(FI); 1179} 1180 1181Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1182 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1183 if (OpI == 0) 1184 return commonCastTransforms(FI); 1185 1186 // fptosi(sitofp(X)) --> X 1187 // fptosi(uitofp(X)) --> X 1188 // This is safe if the intermediate type has enough bits in its mantissa to 1189 // accurately represent all values of X. For example, do not do this with 1190 // i64->float->i64. This is also safe for sitofp case, because any negative 1191 // 'X' value would cause an undefined result for the fptoui. 1192 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1193 OpI->getOperand(0)->getType() == FI.getType() && 1194 (int)FI.getType()->getScalarSizeInBits() <= 1195 OpI->getType()->getFPMantissaWidth()) 1196 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1197 1198 return commonCastTransforms(FI); 1199} 1200 1201Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1202 return commonCastTransforms(CI); 1203} 1204 1205Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1206 return commonCastTransforms(CI); 1207} 1208 1209Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1210 // If the source integer type is not the intptr_t type for this target, do a 1211 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1212 // cast to be exposed to other transforms. 1213 if (TD) { 1214 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1215 TD->getPointerSizeInBits()) { 1216 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1217 TD->getIntPtrType(CI.getContext()), "tmp"); 1218 return new IntToPtrInst(P, CI.getType()); 1219 } 1220 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1221 TD->getPointerSizeInBits()) { 1222 Value *P = Builder->CreateZExt(CI.getOperand(0), 1223 TD->getIntPtrType(CI.getContext()), "tmp"); 1224 return new IntToPtrInst(P, CI.getType()); 1225 } 1226 } 1227 1228 if (Instruction *I = commonCastTransforms(CI)) 1229 return I; 1230 1231 return 0; 1232} 1233 1234/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1235Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1236 Value *Src = CI.getOperand(0); 1237 1238 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1239 // If casting the result of a getelementptr instruction with no offset, turn 1240 // this into a cast of the original pointer! 1241 if (GEP->hasAllZeroIndices()) { 1242 // Changing the cast operand is usually not a good idea but it is safe 1243 // here because the pointer operand is being replaced with another 1244 // pointer operand so the opcode doesn't need to change. 1245 Worklist.Add(GEP); 1246 CI.setOperand(0, GEP->getOperand(0)); 1247 return &CI; 1248 } 1249 1250 // If the GEP has a single use, and the base pointer is a bitcast, and the 1251 // GEP computes a constant offset, see if we can convert these three 1252 // instructions into fewer. This typically happens with unions and other 1253 // non-type-safe code. 1254 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1255 GEP->hasAllConstantIndices()) { 1256 // We are guaranteed to get a constant from EmitGEPOffset. 1257 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1258 int64_t Offset = OffsetV->getSExtValue(); 1259 1260 // Get the base pointer input of the bitcast, and the type it points to. 1261 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1262 const Type *GEPIdxTy = 1263 cast<PointerType>(OrigBase->getType())->getElementType(); 1264 SmallVector<Value*, 8> NewIndices; 1265 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1266 // If we were able to index down into an element, create the GEP 1267 // and bitcast the result. This eliminates one bitcast, potentially 1268 // two. 1269 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1270 Builder->CreateInBoundsGEP(OrigBase, 1271 NewIndices.begin(), NewIndices.end()) : 1272 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end()); 1273 NGEP->takeName(GEP); 1274 1275 if (isa<BitCastInst>(CI)) 1276 return new BitCastInst(NGEP, CI.getType()); 1277 assert(isa<PtrToIntInst>(CI)); 1278 return new PtrToIntInst(NGEP, CI.getType()); 1279 } 1280 } 1281 } 1282 1283 return commonCastTransforms(CI); 1284} 1285 1286Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1287 // If the destination integer type is not the intptr_t type for this target, 1288 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1289 // to be exposed to other transforms. 1290 if (TD) { 1291 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1292 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1293 TD->getIntPtrType(CI.getContext()), 1294 "tmp"); 1295 return new TruncInst(P, CI.getType()); 1296 } 1297 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1298 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1299 TD->getIntPtrType(CI.getContext()), 1300 "tmp"); 1301 return new ZExtInst(P, CI.getType()); 1302 } 1303 } 1304 1305 return commonPointerCastTransforms(CI); 1306} 1307 1308/// OptimizeVectorResize - This input value (which is known to have vector type) 1309/// is being zero extended or truncated to the specified vector type. Try to 1310/// replace it with a shuffle (and vector/vector bitcast) if possible. 1311/// 1312/// The source and destination vector types may have different element types. 1313static Instruction *OptimizeVectorResize(Value *InVal, const VectorType *DestTy, 1314 InstCombiner &IC) { 1315 // We can only do this optimization if the output is a multiple of the input 1316 // element size, or the input is a multiple of the output element size. 1317 // Convert the input type to have the same element type as the output. 1318 const VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1319 1320 if (SrcTy->getElementType() != DestTy->getElementType()) { 1321 // The input types don't need to be identical, but for now they must be the 1322 // same size. There is no specific reason we couldn't handle things like 1323 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1324 // there yet. 1325 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1326 DestTy->getElementType()->getPrimitiveSizeInBits()) 1327 return 0; 1328 1329 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1330 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1331 } 1332 1333 // Now that the element types match, get the shuffle mask and RHS of the 1334 // shuffle to use, which depends on whether we're increasing or decreasing the 1335 // size of the input. 1336 SmallVector<Constant*, 16> ShuffleMask; 1337 Value *V2; 1338 const IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext()); 1339 1340 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1341 // If we're shrinking the number of elements, just shuffle in the low 1342 // elements from the input and use undef as the second shuffle input. 1343 V2 = UndefValue::get(SrcTy); 1344 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1345 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1346 1347 } else { 1348 // If we're increasing the number of elements, shuffle in all of the 1349 // elements from InVal and fill the rest of the result elements with zeros 1350 // from a constant zero. 1351 V2 = Constant::getNullValue(SrcTy); 1352 unsigned SrcElts = SrcTy->getNumElements(); 1353 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1354 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1355 1356 // The excess elements reference the first element of the zero input. 1357 ShuffleMask.append(DestTy->getNumElements()-SrcElts, 1358 ConstantInt::get(Int32Ty, SrcElts)); 1359 } 1360 1361 Constant *Mask = ConstantVector::get(ShuffleMask.data(), ShuffleMask.size()); 1362 return new ShuffleVectorInst(InVal, V2, Mask); 1363} 1364 1365static bool isMultipleOfTypeSize(unsigned Value, const Type *Ty) { 1366 return Value % Ty->getPrimitiveSizeInBits() == 0; 1367} 1368 1369static bool getTypeSizeIndex(unsigned Value, const Type *Ty) { 1370 return Value / Ty->getPrimitiveSizeInBits(); 1371} 1372 1373/// CollectInsertionElements - V is a value which is inserted into a vector of 1374/// VecEltTy. Look through the value to see if we can decompose it into 1375/// insertions into the vector. See the example in the comment for 1376/// OptimizeIntegerToVectorInsertions for the pattern this handles. 1377/// The type of V is always a non-zero multiple of VecEltTy's size. 1378/// 1379/// This returns false if the pattern can't be matched or true if it can, 1380/// filling in Elements with the elements found here. 1381static bool CollectInsertionElements(Value *V, unsigned ElementIndex, 1382 SmallVectorImpl<Value*> &Elements, 1383 const Type *VecEltTy) { 1384 // If we got down to a value of the right type, we win, try inserting into the 1385 // right element. 1386 if (V->getType() == VecEltTy) { 1387 // Fail if multiple elements are inserted into this slot. 1388 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0) 1389 return false; 1390 1391 Elements[ElementIndex] = V; 1392 return true; 1393 } 1394 1395 //if (Constant *C = dyn_cast<Constant>(V)) { 1396 // Figure out the # elements this provides, and bitcast it or slice it up 1397 // as required. 1398 //} 1399 1400 if (!V->hasOneUse()) return false; 1401 1402 Instruction *I = dyn_cast<Instruction>(V); 1403 if (I == 0) return false; 1404 switch (I->getOpcode()) { 1405 default: return false; // Unhandled case. 1406 case Instruction::BitCast: 1407 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1408 Elements, VecEltTy); 1409 case Instruction::ZExt: 1410 if (!isMultipleOfTypeSize( 1411 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1412 VecEltTy)) 1413 return false; 1414 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1415 Elements, VecEltTy); 1416 case Instruction::Or: 1417 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1418 Elements, VecEltTy) && 1419 CollectInsertionElements(I->getOperand(1), ElementIndex, 1420 Elements, VecEltTy); 1421 case Instruction::Shl: { 1422 // Must be shifting by a constant that is a multiple of the element size. 1423 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1424 if (CI == 0) return false; 1425 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false; 1426 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy); 1427 1428 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift, 1429 Elements, VecEltTy); 1430 } 1431 1432 } 1433} 1434 1435 1436/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1437/// may be doing shifts and ors to assemble the elements of the vector manually. 1438/// Try to rip the code out and replace it with insertelements. This is to 1439/// optimize code like this: 1440/// 1441/// %tmp37 = bitcast float %inc to i32 1442/// %tmp38 = zext i32 %tmp37 to i64 1443/// %tmp31 = bitcast float %inc5 to i32 1444/// %tmp32 = zext i32 %tmp31 to i64 1445/// %tmp33 = shl i64 %tmp32, 32 1446/// %ins35 = or i64 %tmp33, %tmp38 1447/// %tmp43 = bitcast i64 %ins35 to <2 x float> 1448/// 1449/// Into two insertelements that do "buildvector{%inc, %inc5}". 1450static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1451 InstCombiner &IC) { 1452 const VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1453 Value *IntInput = CI.getOperand(0); 1454 1455 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1456 if (!CollectInsertionElements(IntInput, 0, Elements, 1457 DestVecTy->getElementType())) 1458 return 0; 1459 1460 // If we succeeded, we know that all of the element are specified by Elements 1461 // or are zero if Elements has a null entry. Recast this as a set of 1462 // insertions. 1463 Value *Result = Constant::getNullValue(CI.getType()); 1464 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1465 if (Elements[i] == 0) continue; // Unset element. 1466 1467 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1468 IC.Builder->getInt32(i)); 1469 } 1470 1471 return Result; 1472} 1473 1474 1475/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1476/// bitcast. The various long double bitcasts can't get in here. 1477static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1478 Value *Src = CI.getOperand(0); 1479 const Type *DestTy = CI.getType(); 1480 1481 // If this is a bitcast from int to float, check to see if the int is an 1482 // extraction from a vector. 1483 Value *VecInput = 0; 1484 // bitcast(trunc(bitcast(somevector))) 1485 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1486 isa<VectorType>(VecInput->getType())) { 1487 const VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1488 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1489 1490 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1491 // If the element type of the vector doesn't match the result type, 1492 // bitcast it to be a vector type we can extract from. 1493 if (VecTy->getElementType() != DestTy) { 1494 VecTy = VectorType::get(DestTy, 1495 VecTy->getPrimitiveSizeInBits() / DestWidth); 1496 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1497 } 1498 1499 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); 1500 } 1501 } 1502 1503 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1504 ConstantInt *ShAmt = 0; 1505 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1506 m_ConstantInt(ShAmt)))) && 1507 isa<VectorType>(VecInput->getType())) { 1508 const VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1509 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1510 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1511 ShAmt->getZExtValue() % DestWidth == 0) { 1512 // If the element type of the vector doesn't match the result type, 1513 // bitcast it to be a vector type we can extract from. 1514 if (VecTy->getElementType() != DestTy) { 1515 VecTy = VectorType::get(DestTy, 1516 VecTy->getPrimitiveSizeInBits() / DestWidth); 1517 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1518 } 1519 1520 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1521 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1522 } 1523 } 1524 return 0; 1525} 1526 1527Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1528 // If the operands are integer typed then apply the integer transforms, 1529 // otherwise just apply the common ones. 1530 Value *Src = CI.getOperand(0); 1531 const Type *SrcTy = Src->getType(); 1532 const Type *DestTy = CI.getType(); 1533 1534 // Get rid of casts from one type to the same type. These are useless and can 1535 // be replaced by the operand. 1536 if (DestTy == Src->getType()) 1537 return ReplaceInstUsesWith(CI, Src); 1538 1539 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1540 const PointerType *SrcPTy = cast<PointerType>(SrcTy); 1541 const Type *DstElTy = DstPTy->getElementType(); 1542 const Type *SrcElTy = SrcPTy->getElementType(); 1543 1544 // If the address spaces don't match, don't eliminate the bitcast, which is 1545 // required for changing types. 1546 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1547 return 0; 1548 1549 // If we are casting a alloca to a pointer to a type of the same 1550 // size, rewrite the allocation instruction to allocate the "right" type. 1551 // There is no need to modify malloc calls because it is their bitcast that 1552 // needs to be cleaned up. 1553 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1554 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1555 return V; 1556 1557 // If the source and destination are pointers, and this cast is equivalent 1558 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1559 // This can enhance SROA and other transforms that want type-safe pointers. 1560 Constant *ZeroUInt = 1561 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1562 unsigned NumZeros = 0; 1563 while (SrcElTy != DstElTy && 1564 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1565 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1566 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1567 ++NumZeros; 1568 } 1569 1570 // If we found a path from the src to dest, create the getelementptr now. 1571 if (SrcElTy == DstElTy) { 1572 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1573 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"", 1574 ((Instruction*)NULL)); 1575 } 1576 } 1577 1578 // Try to optimize int -> float bitcasts. 1579 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1580 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1581 return I; 1582 1583 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1584 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1585 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1586 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1587 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1588 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1589 } 1590 1591 if (isa<IntegerType>(SrcTy)) { 1592 // If this is a cast from an integer to vector, check to see if the input 1593 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1594 // the casts with a shuffle and (potentially) a bitcast. 1595 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1596 CastInst *SrcCast = cast<CastInst>(Src); 1597 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1598 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1599 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1600 cast<VectorType>(DestTy), *this)) 1601 return I; 1602 } 1603 1604 // If the input is an 'or' instruction, we may be doing shifts and ors to 1605 // assemble the elements of the vector manually. Try to rip the code out 1606 // and replace it with insertelements. 1607 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1608 return ReplaceInstUsesWith(CI, V); 1609 } 1610 } 1611 1612 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1613 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1614 Value *Elem = 1615 Builder->CreateExtractElement(Src, 1616 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1617 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1618 } 1619 } 1620 1621 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1622 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1623 // a bitcast to a vector with the same # elts. 1624 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1625 cast<VectorType>(DestTy)->getNumElements() == 1626 SVI->getType()->getNumElements() && 1627 SVI->getType()->getNumElements() == 1628 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1629 BitCastInst *Tmp; 1630 // If either of the operands is a cast from CI.getType(), then 1631 // evaluating the shuffle in the casted destination's type will allow 1632 // us to eliminate at least one cast. 1633 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1634 Tmp->getOperand(0)->getType() == DestTy) || 1635 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1636 Tmp->getOperand(0)->getType() == DestTy)) { 1637 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1638 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1639 // Return a new shuffle vector. Use the same element ID's, as we 1640 // know the vector types match #elts. 1641 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1642 } 1643 } 1644 } 1645 1646 if (SrcTy->isPointerTy()) 1647 return commonPointerCastTransforms(CI); 1648 return commonCastTransforms(CI); 1649} 1650