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/Target/TargetData.h" 17 #include "llvm/Target/TargetLibraryInfo.h" 18 #include "llvm/Support/PatternMatch.h" 19 using namespace llvm; 20 using 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 /// 26 static 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. 79 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 80 AllocaInst &AI) { 81 // This requires TargetData 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. 154 Value *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. 227 static Instruction::CastOps 228 isEliminableCastPair( 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 TargetData *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); 241 242 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. 261 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 262 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. 282 Instruction *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 /// 326 static bool CanEvaluateTruncated(Value *V, 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 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 } 426 427 Instruction *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 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); 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 (Src->hasOneUse() && 466 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 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 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 487 // type isn't non-native. 488 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) && 489 ShouldChangeType(Src->getType(), CI.getType()) && 490 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 491 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr"); 492 return BinaryOperator::CreateAnd(NewTrunc, 493 ConstantExpr::getTrunc(Cst, CI.getType())); 494 } 495 496 return 0; 497 } 498 499 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 500 /// in order to eliminate the icmp. 501 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 502 bool DoXform) { 503 // If we are just checking for a icmp eq of a single bit and zext'ing it 504 // to an integer, then shift the bit to the appropriate place and then 505 // cast to integer to avoid the comparison. 506 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 507 const APInt &Op1CV = Op1C->getValue(); 508 509 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 510 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 511 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 512 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 513 if (!DoXform) return ICI; 514 515 Value *In = ICI->getOperand(0); 516 Value *Sh = ConstantInt::get(In->getType(), 517 In->getType()->getScalarSizeInBits()-1); 518 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 519 if (In->getType() != CI.getType()) 520 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); 521 522 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 523 Constant *One = ConstantInt::get(In->getType(), 1); 524 In = Builder->CreateXor(In, One, In->getName()+".not"); 525 } 526 527 return ReplaceInstUsesWith(CI, In); 528 } 529 530 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 531 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 532 // zext (X == 1) to i32 --> X iff X has only the low bit set. 533 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 534 // zext (X != 0) to i32 --> X iff X has only the low bit set. 535 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 536 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 537 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 538 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 539 // This only works for EQ and NE 540 ICI->isEquality()) { 541 // If Op1C some other power of two, convert: 542 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 543 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 544 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne); 545 546 APInt KnownZeroMask(~KnownZero); 547 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 548 if (!DoXform) return ICI; 549 550 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 551 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 552 // (X&4) == 2 --> false 553 // (X&4) != 2 --> true 554 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 555 isNE); 556 Res = ConstantExpr::getZExt(Res, CI.getType()); 557 return ReplaceInstUsesWith(CI, Res); 558 } 559 560 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 561 Value *In = ICI->getOperand(0); 562 if (ShiftAmt) { 563 // Perform a logical shr by shiftamt. 564 // Insert the shift to put the result in the low bit. 565 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 566 In->getName()+".lobit"); 567 } 568 569 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 570 Constant *One = ConstantInt::get(In->getType(), 1); 571 In = Builder->CreateXor(In, One); 572 } 573 574 if (CI.getType() == In->getType()) 575 return ReplaceInstUsesWith(CI, In); 576 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 577 } 578 } 579 } 580 581 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 582 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 583 // may lead to additional simplifications. 584 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 585 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 586 uint32_t BitWidth = ITy->getBitWidth(); 587 Value *LHS = ICI->getOperand(0); 588 Value *RHS = ICI->getOperand(1); 589 590 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 591 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 592 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS); 593 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS); 594 595 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 596 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 597 APInt UnknownBit = ~KnownBits; 598 if (UnknownBit.countPopulation() == 1) { 599 if (!DoXform) return ICI; 600 601 Value *Result = Builder->CreateXor(LHS, RHS); 602 603 // Mask off any bits that are set and won't be shifted away. 604 if (KnownOneLHS.uge(UnknownBit)) 605 Result = Builder->CreateAnd(Result, 606 ConstantInt::get(ITy, UnknownBit)); 607 608 // Shift the bit we're testing down to the lsb. 609 Result = Builder->CreateLShr( 610 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 611 612 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 613 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 614 Result->takeName(ICI); 615 return ReplaceInstUsesWith(CI, Result); 616 } 617 } 618 } 619 } 620 621 return 0; 622 } 623 624 /// CanEvaluateZExtd - Determine if the specified value can be computed in the 625 /// specified wider type and produce the same low bits. If not, return false. 626 /// 627 /// If this function returns true, it can also return a non-zero number of bits 628 /// (in BitsToClear) which indicates that the value it computes is correct for 629 /// the zero extend, but that the additional BitsToClear bits need to be zero'd 630 /// out. For example, to promote something like: 631 /// 632 /// %B = trunc i64 %A to i32 633 /// %C = lshr i32 %B, 8 634 /// %E = zext i32 %C to i64 635 /// 636 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 637 /// set to 8 to indicate that the promoted value needs to have bits 24-31 638 /// cleared in addition to bits 32-63. Since an 'and' will be generated to 639 /// clear the top bits anyway, doing this has no extra cost. 640 /// 641 /// This function works on both vectors and scalars. 642 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { 643 BitsToClear = 0; 644 if (isa<Constant>(V)) 645 return true; 646 647 Instruction *I = dyn_cast<Instruction>(V); 648 if (!I) return false; 649 650 // If the input is a truncate from the destination type, we can trivially 651 // eliminate it. 652 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 653 return true; 654 655 // We can't extend or shrink something that has multiple uses: doing so would 656 // require duplicating the instruction in general, which isn't profitable. 657 if (!I->hasOneUse()) return false; 658 659 unsigned Opc = I->getOpcode(), Tmp; 660 switch (Opc) { 661 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 662 case Instruction::SExt: // zext(sext(x)) -> sext(x). 663 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 664 return true; 665 case Instruction::And: 666 case Instruction::Or: 667 case Instruction::Xor: 668 case Instruction::Add: 669 case Instruction::Sub: 670 case Instruction::Mul: 671 case Instruction::Shl: 672 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 673 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 674 return false; 675 // These can all be promoted if neither operand has 'bits to clear'. 676 if (BitsToClear == 0 && Tmp == 0) 677 return true; 678 679 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 680 // other side, BitsToClear is ok. 681 if (Tmp == 0 && 682 (Opc == Instruction::And || Opc == Instruction::Or || 683 Opc == Instruction::Xor)) { 684 // We use MaskedValueIsZero here for generality, but the case we care 685 // about the most is constant RHS. 686 unsigned VSize = V->getType()->getScalarSizeInBits(); 687 if (MaskedValueIsZero(I->getOperand(1), 688 APInt::getHighBitsSet(VSize, BitsToClear))) 689 return true; 690 } 691 692 // Otherwise, we don't know how to analyze this BitsToClear case yet. 693 return false; 694 695 case Instruction::LShr: 696 // We can promote lshr(x, cst) if we can promote x. This requires the 697 // ultimate 'and' to clear out the high zero bits we're clearing out though. 698 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 699 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 700 return false; 701 BitsToClear += Amt->getZExtValue(); 702 if (BitsToClear > V->getType()->getScalarSizeInBits()) 703 BitsToClear = V->getType()->getScalarSizeInBits(); 704 return true; 705 } 706 // Cannot promote variable LSHR. 707 return false; 708 case Instruction::Select: 709 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 710 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 711 // TODO: If important, we could handle the case when the BitsToClear are 712 // known zero in the disagreeing side. 713 Tmp != BitsToClear) 714 return false; 715 return true; 716 717 case Instruction::PHI: { 718 // We can change a phi if we can change all operands. Note that we never 719 // get into trouble with cyclic PHIs here because we only consider 720 // instructions with a single use. 721 PHINode *PN = cast<PHINode>(I); 722 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 723 return false; 724 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 725 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 726 // TODO: If important, we could handle the case when the BitsToClear 727 // are known zero in the disagreeing input. 728 Tmp != BitsToClear) 729 return false; 730 return true; 731 } 732 default: 733 // TODO: Can handle more cases here. 734 return false; 735 } 736 } 737 738 Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 739 // If this zero extend is only used by a truncate, let the truncate by 740 // eliminated before we try to optimize this zext. 741 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 742 return 0; 743 744 // If one of the common conversion will work, do it. 745 if (Instruction *Result = commonCastTransforms(CI)) 746 return Result; 747 748 // See if we can simplify any instructions used by the input whose sole 749 // purpose is to compute bits we don't care about. 750 if (SimplifyDemandedInstructionBits(CI)) 751 return &CI; 752 753 Value *Src = CI.getOperand(0); 754 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 755 756 // Attempt to extend the entire input expression tree to the destination 757 // type. Only do this if the dest type is a simple type, don't convert the 758 // expression tree to something weird like i93 unless the source is also 759 // strange. 760 unsigned BitsToClear; 761 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 762 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 763 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 764 "Unreasonable BitsToClear"); 765 766 // Okay, we can transform this! Insert the new expression now. 767 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 768 " to avoid zero extend: " << CI); 769 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 770 assert(Res->getType() == DestTy); 771 772 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 773 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 774 775 // If the high bits are already filled with zeros, just replace this 776 // cast with the result. 777 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 778 DestBitSize-SrcBitsKept))) 779 return ReplaceInstUsesWith(CI, Res); 780 781 // We need to emit an AND to clear the high bits. 782 Constant *C = ConstantInt::get(Res->getType(), 783 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 784 return BinaryOperator::CreateAnd(Res, C); 785 } 786 787 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 788 // types and if the sizes are just right we can convert this into a logical 789 // 'and' which will be much cheaper than the pair of casts. 790 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 791 // TODO: Subsume this into EvaluateInDifferentType. 792 793 // Get the sizes of the types involved. We know that the intermediate type 794 // will be smaller than A or C, but don't know the relation between A and C. 795 Value *A = CSrc->getOperand(0); 796 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 797 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 798 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 799 // If we're actually extending zero bits, then if 800 // SrcSize < DstSize: zext(a & mask) 801 // SrcSize == DstSize: a & mask 802 // SrcSize > DstSize: trunc(a) & mask 803 if (SrcSize < DstSize) { 804 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 805 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 806 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 807 return new ZExtInst(And, CI.getType()); 808 } 809 810 if (SrcSize == DstSize) { 811 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 812 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 813 AndValue)); 814 } 815 if (SrcSize > DstSize) { 816 Value *Trunc = Builder->CreateTrunc(A, CI.getType()); 817 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 818 return BinaryOperator::CreateAnd(Trunc, 819 ConstantInt::get(Trunc->getType(), 820 AndValue)); 821 } 822 } 823 824 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 825 return transformZExtICmp(ICI, CI); 826 827 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 828 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 829 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 830 // of the (zext icmp) will be transformed. 831 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 832 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 833 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 834 (transformZExtICmp(LHS, CI, false) || 835 transformZExtICmp(RHS, CI, false))) { 836 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 837 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 838 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 839 } 840 } 841 842 // zext(trunc(t) & C) -> (t & zext(C)). 843 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 844 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 845 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 846 Value *TI0 = TI->getOperand(0); 847 if (TI0->getType() == CI.getType()) 848 return 849 BinaryOperator::CreateAnd(TI0, 850 ConstantExpr::getZExt(C, CI.getType())); 851 } 852 853 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 854 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 855 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 856 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 857 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 858 And->getOperand(1) == C) 859 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 860 Value *TI0 = TI->getOperand(0); 861 if (TI0->getType() == CI.getType()) { 862 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 863 Value *NewAnd = Builder->CreateAnd(TI0, ZC); 864 return BinaryOperator::CreateXor(NewAnd, ZC); 865 } 866 } 867 868 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 869 Value *X; 870 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 871 match(SrcI, m_Not(m_Value(X))) && 872 (!X->hasOneUse() || !isa<CmpInst>(X))) { 873 Value *New = Builder->CreateZExt(X, CI.getType()); 874 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 875 } 876 877 return 0; 878 } 879 880 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations 881 /// in order to eliminate the icmp. 882 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { 883 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); 884 ICmpInst::Predicate Pred = ICI->getPredicate(); 885 886 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 887 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative 888 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive 889 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) || 890 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { 891 892 Value *Sh = ConstantInt::get(Op0->getType(), 893 Op0->getType()->getScalarSizeInBits()-1); 894 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); 895 if (In->getType() != CI.getType()) 896 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); 897 898 if (Pred == ICmpInst::ICMP_SGT) 899 In = Builder->CreateNot(In, In->getName()+".not"); 900 return ReplaceInstUsesWith(CI, In); 901 } 902 903 // If we know that only one bit of the LHS of the icmp can be set and we 904 // have an equality comparison with zero or a power of 2, we can transform 905 // the icmp and sext into bitwise/integer operations. 906 if (ICI->hasOneUse() && 907 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ 908 unsigned BitWidth = Op1C->getType()->getBitWidth(); 909 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 910 ComputeMaskedBits(Op0, KnownZero, KnownOne); 911 912 APInt KnownZeroMask(~KnownZero); 913 if (KnownZeroMask.isPowerOf2()) { 914 Value *In = ICI->getOperand(0); 915 916 // If the icmp tests for a known zero bit we can constant fold it. 917 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) { 918 Value *V = Pred == ICmpInst::ICMP_NE ? 919 ConstantInt::getAllOnesValue(CI.getType()) : 920 ConstantInt::getNullValue(CI.getType()); 921 return ReplaceInstUsesWith(CI, V); 922 } 923 924 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) { 925 // sext ((x & 2^n) == 0) -> (x >> n) - 1 926 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1 927 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros(); 928 // Perform a right shift to place the desired bit in the LSB. 929 if (ShiftAmt) 930 In = Builder->CreateLShr(In, 931 ConstantInt::get(In->getType(), ShiftAmt)); 932 933 // At this point "In" is either 1 or 0. Subtract 1 to turn 934 // {1, 0} -> {0, -1}. 935 In = Builder->CreateAdd(In, 936 ConstantInt::getAllOnesValue(In->getType()), 937 "sext"); 938 } else { 939 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1 940 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1 941 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros(); 942 // Perform a left shift to place the desired bit in the MSB. 943 if (ShiftAmt) 944 In = Builder->CreateShl(In, 945 ConstantInt::get(In->getType(), ShiftAmt)); 946 947 // Distribute the bit over the whole bit width. 948 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(), 949 BitWidth - 1), "sext"); 950 } 951 952 if (CI.getType() == In->getType()) 953 return ReplaceInstUsesWith(CI, In); 954 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/); 955 } 956 } 957 } 958 959 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed. 960 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) { 961 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) && 962 Op0->getType() == CI.getType()) { 963 Type *EltTy = VTy->getElementType(); 964 965 // splat the shift constant to a constant vector. 966 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1); 967 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit"); 968 return ReplaceInstUsesWith(CI, In); 969 } 970 } 971 972 return 0; 973 } 974 975 /// CanEvaluateSExtd - Return true if we can take the specified value 976 /// and return it as type Ty without inserting any new casts and without 977 /// changing the value of the common low bits. This is used by code that tries 978 /// to promote integer operations to a wider types will allow us to eliminate 979 /// the extension. 980 /// 981 /// This function works on both vectors and scalars. 982 /// 983 static bool CanEvaluateSExtd(Value *V, Type *Ty) { 984 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 985 "Can't sign extend type to a smaller type"); 986 // If this is a constant, it can be trivially promoted. 987 if (isa<Constant>(V)) 988 return true; 989 990 Instruction *I = dyn_cast<Instruction>(V); 991 if (!I) return false; 992 993 // If this is a truncate from the dest type, we can trivially eliminate it. 994 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 995 return true; 996 997 // We can't extend or shrink something that has multiple uses: doing so would 998 // require duplicating the instruction in general, which isn't profitable. 999 if (!I->hasOneUse()) return false; 1000 1001 switch (I->getOpcode()) { 1002 case Instruction::SExt: // sext(sext(x)) -> sext(x) 1003 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 1004 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 1005 return true; 1006 case Instruction::And: 1007 case Instruction::Or: 1008 case Instruction::Xor: 1009 case Instruction::Add: 1010 case Instruction::Sub: 1011 case Instruction::Mul: 1012 // These operators can all arbitrarily be extended if their inputs can. 1013 return CanEvaluateSExtd(I->getOperand(0), Ty) && 1014 CanEvaluateSExtd(I->getOperand(1), Ty); 1015 1016 //case Instruction::Shl: TODO 1017 //case Instruction::LShr: TODO 1018 1019 case Instruction::Select: 1020 return CanEvaluateSExtd(I->getOperand(1), Ty) && 1021 CanEvaluateSExtd(I->getOperand(2), Ty); 1022 1023 case Instruction::PHI: { 1024 // We can change a phi if we can change all operands. Note that we never 1025 // get into trouble with cyclic PHIs here because we only consider 1026 // instructions with a single use. 1027 PHINode *PN = cast<PHINode>(I); 1028 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1029 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 1030 return true; 1031 } 1032 default: 1033 // TODO: Can handle more cases here. 1034 break; 1035 } 1036 1037 return false; 1038 } 1039 1040 Instruction *InstCombiner::visitSExt(SExtInst &CI) { 1041 // If this sign extend is only used by a truncate, let the truncate by 1042 // eliminated before we try to optimize this zext. 1043 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 1044 return 0; 1045 1046 if (Instruction *I = commonCastTransforms(CI)) 1047 return I; 1048 1049 // See if we can simplify any instructions used by the input whose sole 1050 // purpose is to compute bits we don't care about. 1051 if (SimplifyDemandedInstructionBits(CI)) 1052 return &CI; 1053 1054 Value *Src = CI.getOperand(0); 1055 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 1056 1057 // Attempt to extend the entire input expression tree to the destination 1058 // type. Only do this if the dest type is a simple type, don't convert the 1059 // expression tree to something weird like i93 unless the source is also 1060 // strange. 1061 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 1062 CanEvaluateSExtd(Src, DestTy)) { 1063 // Okay, we can transform this! Insert the new expression now. 1064 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 1065 " to avoid sign extend: " << CI); 1066 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 1067 assert(Res->getType() == DestTy); 1068 1069 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1070 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1071 1072 // If the high bits are already filled with sign bit, just replace this 1073 // cast with the result. 1074 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 1075 return ReplaceInstUsesWith(CI, Res); 1076 1077 // We need to emit a shl + ashr to do the sign extend. 1078 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1079 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 1080 ShAmt); 1081 } 1082 1083 // If this input is a trunc from our destination, then turn sext(trunc(x)) 1084 // into shifts. 1085 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 1086 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 1087 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1088 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1089 1090 // We need to emit a shl + ashr to do the sign extend. 1091 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1092 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1093 return BinaryOperator::CreateAShr(Res, ShAmt); 1094 } 1095 1096 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 1097 return transformSExtICmp(ICI, CI); 1098 1099 // If the input is a shl/ashr pair of a same constant, then this is a sign 1100 // extension from a smaller value. If we could trust arbitrary bitwidth 1101 // integers, we could turn this into a truncate to the smaller bit and then 1102 // use a sext for the whole extension. Since we don't, look deeper and check 1103 // for a truncate. If the source and dest are the same type, eliminate the 1104 // trunc and extend and just do shifts. For example, turn: 1105 // %a = trunc i32 %i to i8 1106 // %b = shl i8 %a, 6 1107 // %c = ashr i8 %b, 6 1108 // %d = sext i8 %c to i32 1109 // into: 1110 // %a = shl i32 %i, 30 1111 // %d = ashr i32 %a, 30 1112 Value *A = 0; 1113 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1114 ConstantInt *BA = 0, *CA = 0; 1115 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1116 m_ConstantInt(CA))) && 1117 BA == CA && A->getType() == CI.getType()) { 1118 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1119 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1120 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1121 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1122 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1123 return BinaryOperator::CreateAShr(A, ShAmtV); 1124 } 1125 1126 return 0; 1127 } 1128 1129 1130 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1131 /// in the specified FP type without changing its value. 1132 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1133 bool losesInfo; 1134 APFloat F = CFP->getValueAPF(); 1135 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1136 if (!losesInfo) 1137 return ConstantFP::get(CFP->getContext(), F); 1138 return 0; 1139 } 1140 1141 /// LookThroughFPExtensions - If this is an fp extension instruction, look 1142 /// through it until we get the source value. 1143 static Value *LookThroughFPExtensions(Value *V) { 1144 if (Instruction *I = dyn_cast<Instruction>(V)) 1145 if (I->getOpcode() == Instruction::FPExt) 1146 return LookThroughFPExtensions(I->getOperand(0)); 1147 1148 // If this value is a constant, return the constant in the smallest FP type 1149 // that can accurately represent it. This allows us to turn 1150 // (float)((double)X+2.0) into x+2.0f. 1151 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1152 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1153 return V; // No constant folding of this. 1154 // See if the value can be truncated to half and then reextended. 1155 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf)) 1156 return V; 1157 // See if the value can be truncated to float and then reextended. 1158 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1159 return V; 1160 if (CFP->getType()->isDoubleTy()) 1161 return V; // Won't shrink. 1162 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1163 return V; 1164 // Don't try to shrink to various long double types. 1165 } 1166 1167 return V; 1168 } 1169 1170 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1171 if (Instruction *I = commonCastTransforms(CI)) 1172 return I; 1173 1174 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1175 // smaller than the destination type, we can eliminate the truncate by doing 1176 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1177 // as many builtins (sqrt, etc). 1178 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1179 if (OpI && OpI->hasOneUse()) { 1180 switch (OpI->getOpcode()) { 1181 default: break; 1182 case Instruction::FAdd: 1183 case Instruction::FSub: 1184 case Instruction::FMul: 1185 case Instruction::FDiv: 1186 case Instruction::FRem: 1187 Type *SrcTy = OpI->getType(); 1188 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1189 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1190 if (LHSTrunc->getType() != SrcTy && 1191 RHSTrunc->getType() != SrcTy) { 1192 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1193 // If the source types were both smaller than the destination type of 1194 // the cast, do this xform. 1195 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1196 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1197 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1198 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1199 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1200 } 1201 } 1202 break; 1203 } 1204 } 1205 1206 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1207 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1208 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) && 1209 Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) && 1210 Call->getNumArgOperands() == 1 && 1211 Call->hasOneUse()) { 1212 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1213 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1214 CI.getType()->isFloatTy() && 1215 Call->getType()->isDoubleTy() && 1216 Arg->getType()->isDoubleTy() && 1217 Arg->getOperand(0)->getType()->isFloatTy()) { 1218 Function *Callee = Call->getCalledFunction(); 1219 Module *M = CI.getParent()->getParent()->getParent(); 1220 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf", 1221 Callee->getAttributes(), 1222 Builder->getFloatTy(), 1223 Builder->getFloatTy(), 1224 NULL); 1225 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1226 "sqrtfcall"); 1227 ret->setAttributes(Callee->getAttributes()); 1228 1229 1230 // Remove the old Call. With -fmath-errno, it won't get marked readnone. 1231 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType())); 1232 EraseInstFromFunction(*Call); 1233 return ret; 1234 } 1235 } 1236 1237 return 0; 1238 } 1239 1240 Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1241 return commonCastTransforms(CI); 1242 } 1243 1244 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1245 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1246 if (OpI == 0) 1247 return commonCastTransforms(FI); 1248 1249 // fptoui(uitofp(X)) --> X 1250 // fptoui(sitofp(X)) --> X 1251 // This is safe if the intermediate type has enough bits in its mantissa to 1252 // accurately represent all values of X. For example, do not do this with 1253 // i64->float->i64. This is also safe for sitofp case, because any negative 1254 // 'X' value would cause an undefined result for the fptoui. 1255 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1256 OpI->getOperand(0)->getType() == FI.getType() && 1257 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1258 OpI->getType()->getFPMantissaWidth()) 1259 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1260 1261 return commonCastTransforms(FI); 1262 } 1263 1264 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1265 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1266 if (OpI == 0) 1267 return commonCastTransforms(FI); 1268 1269 // fptosi(sitofp(X)) --> X 1270 // fptosi(uitofp(X)) --> X 1271 // This is safe if the intermediate type has enough bits in its mantissa to 1272 // accurately represent all values of X. For example, do not do this with 1273 // i64->float->i64. This is also safe for sitofp case, because any negative 1274 // 'X' value would cause an undefined result for the fptoui. 1275 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1276 OpI->getOperand(0)->getType() == FI.getType() && 1277 (int)FI.getType()->getScalarSizeInBits() <= 1278 OpI->getType()->getFPMantissaWidth()) 1279 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1280 1281 return commonCastTransforms(FI); 1282 } 1283 1284 Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1285 return commonCastTransforms(CI); 1286 } 1287 1288 Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1289 return commonCastTransforms(CI); 1290 } 1291 1292 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1293 // If the source integer type is not the intptr_t type for this target, do a 1294 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1295 // cast to be exposed to other transforms. 1296 if (TD) { 1297 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1298 TD->getPointerSizeInBits()) { 1299 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1300 TD->getIntPtrType(CI.getContext())); 1301 return new IntToPtrInst(P, CI.getType()); 1302 } 1303 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1304 TD->getPointerSizeInBits()) { 1305 Value *P = Builder->CreateZExt(CI.getOperand(0), 1306 TD->getIntPtrType(CI.getContext())); 1307 return new IntToPtrInst(P, CI.getType()); 1308 } 1309 } 1310 1311 if (Instruction *I = commonCastTransforms(CI)) 1312 return I; 1313 1314 return 0; 1315 } 1316 1317 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1318 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1319 Value *Src = CI.getOperand(0); 1320 1321 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1322 // If casting the result of a getelementptr instruction with no offset, turn 1323 // this into a cast of the original pointer! 1324 if (GEP->hasAllZeroIndices()) { 1325 // Changing the cast operand is usually not a good idea but it is safe 1326 // here because the pointer operand is being replaced with another 1327 // pointer operand so the opcode doesn't need to change. 1328 Worklist.Add(GEP); 1329 CI.setOperand(0, GEP->getOperand(0)); 1330 return &CI; 1331 } 1332 1333 // If the GEP has a single use, and the base pointer is a bitcast, and the 1334 // GEP computes a constant offset, see if we can convert these three 1335 // instructions into fewer. This typically happens with unions and other 1336 // non-type-safe code. 1337 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1338 GEP->hasAllConstantIndices()) { 1339 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end()); 1340 int64_t Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops); 1341 1342 // Get the base pointer input of the bitcast, and the type it points to. 1343 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1344 Type *GEPIdxTy = 1345 cast<PointerType>(OrigBase->getType())->getElementType(); 1346 SmallVector<Value*, 8> NewIndices; 1347 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1348 // If we were able to index down into an element, create the GEP 1349 // and bitcast the result. This eliminates one bitcast, potentially 1350 // two. 1351 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1352 Builder->CreateInBoundsGEP(OrigBase, NewIndices) : 1353 Builder->CreateGEP(OrigBase, NewIndices); 1354 NGEP->takeName(GEP); 1355 1356 if (isa<BitCastInst>(CI)) 1357 return new BitCastInst(NGEP, CI.getType()); 1358 assert(isa<PtrToIntInst>(CI)); 1359 return new PtrToIntInst(NGEP, CI.getType()); 1360 } 1361 } 1362 } 1363 1364 return commonCastTransforms(CI); 1365 } 1366 1367 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1368 // If the destination integer type is not the intptr_t type for this target, 1369 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1370 // to be exposed to other transforms. 1371 if (TD) { 1372 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1373 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1374 TD->getIntPtrType(CI.getContext())); 1375 return new TruncInst(P, CI.getType()); 1376 } 1377 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1378 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1379 TD->getIntPtrType(CI.getContext())); 1380 return new ZExtInst(P, CI.getType()); 1381 } 1382 } 1383 1384 return commonPointerCastTransforms(CI); 1385 } 1386 1387 /// OptimizeVectorResize - This input value (which is known to have vector type) 1388 /// is being zero extended or truncated to the specified vector type. Try to 1389 /// replace it with a shuffle (and vector/vector bitcast) if possible. 1390 /// 1391 /// The source and destination vector types may have different element types. 1392 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, 1393 InstCombiner &IC) { 1394 // We can only do this optimization if the output is a multiple of the input 1395 // element size, or the input is a multiple of the output element size. 1396 // Convert the input type to have the same element type as the output. 1397 VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1398 1399 if (SrcTy->getElementType() != DestTy->getElementType()) { 1400 // The input types don't need to be identical, but for now they must be the 1401 // same size. There is no specific reason we couldn't handle things like 1402 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1403 // there yet. 1404 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1405 DestTy->getElementType()->getPrimitiveSizeInBits()) 1406 return 0; 1407 1408 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1409 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1410 } 1411 1412 // Now that the element types match, get the shuffle mask and RHS of the 1413 // shuffle to use, which depends on whether we're increasing or decreasing the 1414 // size of the input. 1415 SmallVector<uint32_t, 16> ShuffleMask; 1416 Value *V2; 1417 1418 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1419 // If we're shrinking the number of elements, just shuffle in the low 1420 // elements from the input and use undef as the second shuffle input. 1421 V2 = UndefValue::get(SrcTy); 1422 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1423 ShuffleMask.push_back(i); 1424 1425 } else { 1426 // If we're increasing the number of elements, shuffle in all of the 1427 // elements from InVal and fill the rest of the result elements with zeros 1428 // from a constant zero. 1429 V2 = Constant::getNullValue(SrcTy); 1430 unsigned SrcElts = SrcTy->getNumElements(); 1431 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1432 ShuffleMask.push_back(i); 1433 1434 // The excess elements reference the first element of the zero input. 1435 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i) 1436 ShuffleMask.push_back(SrcElts); 1437 } 1438 1439 return new ShuffleVectorInst(InVal, V2, 1440 ConstantDataVector::get(V2->getContext(), 1441 ShuffleMask)); 1442 } 1443 1444 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { 1445 return Value % Ty->getPrimitiveSizeInBits() == 0; 1446 } 1447 1448 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { 1449 return Value / Ty->getPrimitiveSizeInBits(); 1450 } 1451 1452 /// CollectInsertionElements - V is a value which is inserted into a vector of 1453 /// VecEltTy. Look through the value to see if we can decompose it into 1454 /// insertions into the vector. See the example in the comment for 1455 /// OptimizeIntegerToVectorInsertions for the pattern this handles. 1456 /// The type of V is always a non-zero multiple of VecEltTy's size. 1457 /// 1458 /// This returns false if the pattern can't be matched or true if it can, 1459 /// filling in Elements with the elements found here. 1460 static bool CollectInsertionElements(Value *V, unsigned ElementIndex, 1461 SmallVectorImpl<Value*> &Elements, 1462 Type *VecEltTy) { 1463 // Undef values never contribute useful bits to the result. 1464 if (isa<UndefValue>(V)) return true; 1465 1466 // If we got down to a value of the right type, we win, try inserting into the 1467 // right element. 1468 if (V->getType() == VecEltTy) { 1469 // Inserting null doesn't actually insert any elements. 1470 if (Constant *C = dyn_cast<Constant>(V)) 1471 if (C->isNullValue()) 1472 return true; 1473 1474 // Fail if multiple elements are inserted into this slot. 1475 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0) 1476 return false; 1477 1478 Elements[ElementIndex] = V; 1479 return true; 1480 } 1481 1482 if (Constant *C = dyn_cast<Constant>(V)) { 1483 // Figure out the # elements this provides, and bitcast it or slice it up 1484 // as required. 1485 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1486 VecEltTy); 1487 // If the constant is the size of a vector element, we just need to bitcast 1488 // it to the right type so it gets properly inserted. 1489 if (NumElts == 1) 1490 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1491 ElementIndex, Elements, VecEltTy); 1492 1493 // Okay, this is a constant that covers multiple elements. Slice it up into 1494 // pieces and insert each element-sized piece into the vector. 1495 if (!isa<IntegerType>(C->getType())) 1496 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1497 C->getType()->getPrimitiveSizeInBits())); 1498 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1499 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1500 1501 for (unsigned i = 0; i != NumElts; ++i) { 1502 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1503 i*ElementSize)); 1504 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1505 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy)) 1506 return false; 1507 } 1508 return true; 1509 } 1510 1511 if (!V->hasOneUse()) return false; 1512 1513 Instruction *I = dyn_cast<Instruction>(V); 1514 if (I == 0) return false; 1515 switch (I->getOpcode()) { 1516 default: return false; // Unhandled case. 1517 case Instruction::BitCast: 1518 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1519 Elements, VecEltTy); 1520 case Instruction::ZExt: 1521 if (!isMultipleOfTypeSize( 1522 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1523 VecEltTy)) 1524 return false; 1525 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1526 Elements, VecEltTy); 1527 case Instruction::Or: 1528 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1529 Elements, VecEltTy) && 1530 CollectInsertionElements(I->getOperand(1), ElementIndex, 1531 Elements, VecEltTy); 1532 case Instruction::Shl: { 1533 // Must be shifting by a constant that is a multiple of the element size. 1534 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1535 if (CI == 0) return false; 1536 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false; 1537 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy); 1538 1539 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift, 1540 Elements, VecEltTy); 1541 } 1542 1543 } 1544 } 1545 1546 1547 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1548 /// may be doing shifts and ors to assemble the elements of the vector manually. 1549 /// Try to rip the code out and replace it with insertelements. This is to 1550 /// optimize code like this: 1551 /// 1552 /// %tmp37 = bitcast float %inc to i32 1553 /// %tmp38 = zext i32 %tmp37 to i64 1554 /// %tmp31 = bitcast float %inc5 to i32 1555 /// %tmp32 = zext i32 %tmp31 to i64 1556 /// %tmp33 = shl i64 %tmp32, 32 1557 /// %ins35 = or i64 %tmp33, %tmp38 1558 /// %tmp43 = bitcast i64 %ins35 to <2 x float> 1559 /// 1560 /// Into two insertelements that do "buildvector{%inc, %inc5}". 1561 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1562 InstCombiner &IC) { 1563 VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1564 Value *IntInput = CI.getOperand(0); 1565 1566 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1567 if (!CollectInsertionElements(IntInput, 0, Elements, 1568 DestVecTy->getElementType())) 1569 return 0; 1570 1571 // If we succeeded, we know that all of the element are specified by Elements 1572 // or are zero if Elements has a null entry. Recast this as a set of 1573 // insertions. 1574 Value *Result = Constant::getNullValue(CI.getType()); 1575 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1576 if (Elements[i] == 0) continue; // Unset element. 1577 1578 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1579 IC.Builder->getInt32(i)); 1580 } 1581 1582 return Result; 1583 } 1584 1585 1586 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1587 /// bitcast. The various long double bitcasts can't get in here. 1588 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1589 Value *Src = CI.getOperand(0); 1590 Type *DestTy = CI.getType(); 1591 1592 // If this is a bitcast from int to float, check to see if the int is an 1593 // extraction from a vector. 1594 Value *VecInput = 0; 1595 // bitcast(trunc(bitcast(somevector))) 1596 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1597 isa<VectorType>(VecInput->getType())) { 1598 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1599 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1600 1601 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1602 // If the element type of the vector doesn't match the result type, 1603 // bitcast it to be a vector type we can extract from. 1604 if (VecTy->getElementType() != DestTy) { 1605 VecTy = VectorType::get(DestTy, 1606 VecTy->getPrimitiveSizeInBits() / DestWidth); 1607 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1608 } 1609 1610 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); 1611 } 1612 } 1613 1614 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1615 ConstantInt *ShAmt = 0; 1616 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1617 m_ConstantInt(ShAmt)))) && 1618 isa<VectorType>(VecInput->getType())) { 1619 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1620 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1621 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1622 ShAmt->getZExtValue() % DestWidth == 0) { 1623 // If the element type of the vector doesn't match the result type, 1624 // bitcast it to be a vector type we can extract from. 1625 if (VecTy->getElementType() != DestTy) { 1626 VecTy = VectorType::get(DestTy, 1627 VecTy->getPrimitiveSizeInBits() / DestWidth); 1628 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1629 } 1630 1631 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1632 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1633 } 1634 } 1635 return 0; 1636 } 1637 1638 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1639 // If the operands are integer typed then apply the integer transforms, 1640 // otherwise just apply the common ones. 1641 Value *Src = CI.getOperand(0); 1642 Type *SrcTy = Src->getType(); 1643 Type *DestTy = CI.getType(); 1644 1645 // Get rid of casts from one type to the same type. These are useless and can 1646 // be replaced by the operand. 1647 if (DestTy == Src->getType()) 1648 return ReplaceInstUsesWith(CI, Src); 1649 1650 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1651 PointerType *SrcPTy = cast<PointerType>(SrcTy); 1652 Type *DstElTy = DstPTy->getElementType(); 1653 Type *SrcElTy = SrcPTy->getElementType(); 1654 1655 // If the address spaces don't match, don't eliminate the bitcast, which is 1656 // required for changing types. 1657 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1658 return 0; 1659 1660 // If we are casting a alloca to a pointer to a type of the same 1661 // size, rewrite the allocation instruction to allocate the "right" type. 1662 // There is no need to modify malloc calls because it is their bitcast that 1663 // needs to be cleaned up. 1664 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1665 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1666 return V; 1667 1668 // If the source and destination are pointers, and this cast is equivalent 1669 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1670 // This can enhance SROA and other transforms that want type-safe pointers. 1671 Constant *ZeroUInt = 1672 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1673 unsigned NumZeros = 0; 1674 while (SrcElTy != DstElTy && 1675 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1676 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1677 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1678 ++NumZeros; 1679 } 1680 1681 // If we found a path from the src to dest, create the getelementptr now. 1682 if (SrcElTy == DstElTy) { 1683 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1684 return GetElementPtrInst::CreateInBounds(Src, Idxs); 1685 } 1686 } 1687 1688 // Try to optimize int -> float bitcasts. 1689 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1690 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1691 return I; 1692 1693 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1694 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1695 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1696 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1697 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1698 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1699 } 1700 1701 if (isa<IntegerType>(SrcTy)) { 1702 // If this is a cast from an integer to vector, check to see if the input 1703 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1704 // the casts with a shuffle and (potentially) a bitcast. 1705 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1706 CastInst *SrcCast = cast<CastInst>(Src); 1707 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1708 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1709 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1710 cast<VectorType>(DestTy), *this)) 1711 return I; 1712 } 1713 1714 // If the input is an 'or' instruction, we may be doing shifts and ors to 1715 // assemble the elements of the vector manually. Try to rip the code out 1716 // and replace it with insertelements. 1717 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1718 return ReplaceInstUsesWith(CI, V); 1719 } 1720 } 1721 1722 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1723 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1724 Value *Elem = 1725 Builder->CreateExtractElement(Src, 1726 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1727 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1728 } 1729 } 1730 1731 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1732 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1733 // a bitcast to a vector with the same # elts. 1734 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1735 cast<VectorType>(DestTy)->getNumElements() == 1736 SVI->getType()->getNumElements() && 1737 SVI->getType()->getNumElements() == 1738 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1739 BitCastInst *Tmp; 1740 // If either of the operands is a cast from CI.getType(), then 1741 // evaluating the shuffle in the casted destination's type will allow 1742 // us to eliminate at least one cast. 1743 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1744 Tmp->getOperand(0)->getType() == DestTy) || 1745 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1746 Tmp->getOperand(0)->getType() == DestTy)) { 1747 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1748 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1749 // Return a new shuffle vector. Use the same element ID's, as we 1750 // know the vector types match #elts. 1751 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1752 } 1753 } 1754 } 1755 1756 if (SrcTy->isPointerTy()) 1757 return commonPointerCastTransforms(CI); 1758 return commonCastTransforms(CI); 1759 } 1760