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 "InstCombineInternal.h" 15 #include "llvm/Analysis/ConstantFolding.h" 16 #include "llvm/IR/DataLayout.h" 17 #include "llvm/IR/PatternMatch.h" 18 #include "llvm/Analysis/TargetLibraryInfo.h" 19 using namespace llvm; 20 using namespace PatternMatch; 21 22 #define DEBUG_TYPE "instcombine" 23 24 /// Analyze 'Val', seeing if it is a simple linear expression. 25 /// If so, decompose it, returning some value X, such that Val is 26 /// X*Scale+Offset. 27 /// 28 static Value *decomposeSimpleLinearExpr(Value *Val, unsigned &Scale, 29 uint64_t &Offset) { 30 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 31 Offset = CI->getZExtValue(); 32 Scale = 0; 33 return ConstantInt::get(Val->getType(), 0); 34 } 35 36 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) { 37 // Cannot look past anything that might overflow. 38 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val); 39 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) { 40 Scale = 1; 41 Offset = 0; 42 return Val; 43 } 44 45 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 46 if (I->getOpcode() == Instruction::Shl) { 47 // This is a value scaled by '1 << the shift amt'. 48 Scale = UINT64_C(1) << RHS->getZExtValue(); 49 Offset = 0; 50 return I->getOperand(0); 51 } 52 53 if (I->getOpcode() == Instruction::Mul) { 54 // This value is scaled by 'RHS'. 55 Scale = RHS->getZExtValue(); 56 Offset = 0; 57 return I->getOperand(0); 58 } 59 60 if (I->getOpcode() == Instruction::Add) { 61 // We have X+C. Check to see if we really have (X*C2)+C1, 62 // where C1 is divisible by C2. 63 unsigned SubScale; 64 Value *SubVal = 65 decomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); 66 Offset += RHS->getZExtValue(); 67 Scale = SubScale; 68 return SubVal; 69 } 70 } 71 } 72 73 // Otherwise, we can't look past this. 74 Scale = 1; 75 Offset = 0; 76 return Val; 77 } 78 79 /// If we find a cast of an allocation instruction, try to eliminate the cast by 80 /// moving the type information into the alloc. 81 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 82 AllocaInst &AI) { 83 PointerType *PTy = cast<PointerType>(CI.getType()); 84 85 BuilderTy AllocaBuilder(*Builder); 86 AllocaBuilder.SetInsertPoint(&AI); 87 88 // Get the type really allocated and the type casted to. 89 Type *AllocElTy = AI.getAllocatedType(); 90 Type *CastElTy = PTy->getElementType(); 91 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr; 92 93 unsigned AllocElTyAlign = DL.getABITypeAlignment(AllocElTy); 94 unsigned CastElTyAlign = DL.getABITypeAlignment(CastElTy); 95 if (CastElTyAlign < AllocElTyAlign) return nullptr; 96 97 // If the allocation has multiple uses, only promote it if we are strictly 98 // increasing the alignment of the resultant allocation. If we keep it the 99 // same, we open the door to infinite loops of various kinds. 100 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr; 101 102 uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy); 103 uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy); 104 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr; 105 106 // If the allocation has multiple uses, only promote it if we're not 107 // shrinking the amount of memory being allocated. 108 uint64_t AllocElTyStoreSize = DL.getTypeStoreSize(AllocElTy); 109 uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy); 110 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr; 111 112 // See if we can satisfy the modulus by pulling a scale out of the array 113 // size argument. 114 unsigned ArraySizeScale; 115 uint64_t ArrayOffset; 116 Value *NumElements = // See if the array size is a decomposable linear expr. 117 decomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 118 119 // If we can now satisfy the modulus, by using a non-1 scale, we really can 120 // do the xform. 121 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 122 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr; 123 124 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 125 Value *Amt = nullptr; 126 if (Scale == 1) { 127 Amt = NumElements; 128 } else { 129 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 130 // Insert before the alloca, not before the cast. 131 Amt = AllocaBuilder.CreateMul(Amt, NumElements); 132 } 133 134 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 135 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 136 Offset, true); 137 Amt = AllocaBuilder.CreateAdd(Amt, Off); 138 } 139 140 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 141 New->setAlignment(AI.getAlignment()); 142 New->takeName(&AI); 143 New->setUsedWithInAlloca(AI.isUsedWithInAlloca()); 144 145 // If the allocation has multiple real uses, insert a cast and change all 146 // things that used it to use the new cast. This will also hack on CI, but it 147 // will die soon. 148 if (!AI.hasOneUse()) { 149 // New is the allocation instruction, pointer typed. AI is the original 150 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 151 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 152 replaceInstUsesWith(AI, NewCast); 153 } 154 return replaceInstUsesWith(CI, New); 155 } 156 157 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns 158 /// true for, actually insert the code to evaluate the expression. 159 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, 160 bool isSigned) { 161 if (Constant *C = dyn_cast<Constant>(V)) { 162 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 163 // If we got a constantexpr back, try to simplify it with DL info. 164 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 165 C = ConstantFoldConstantExpression(CE, DL, TLI); 166 return C; 167 } 168 169 // Otherwise, it must be an instruction. 170 Instruction *I = cast<Instruction>(V); 171 Instruction *Res = nullptr; 172 unsigned Opc = I->getOpcode(); 173 switch (Opc) { 174 case Instruction::Add: 175 case Instruction::Sub: 176 case Instruction::Mul: 177 case Instruction::And: 178 case Instruction::Or: 179 case Instruction::Xor: 180 case Instruction::AShr: 181 case Instruction::LShr: 182 case Instruction::Shl: 183 case Instruction::UDiv: 184 case Instruction::URem: { 185 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 186 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 187 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 188 break; 189 } 190 case Instruction::Trunc: 191 case Instruction::ZExt: 192 case Instruction::SExt: 193 // If the source type of the cast is the type we're trying for then we can 194 // just return the source. There's no need to insert it because it is not 195 // new. 196 if (I->getOperand(0)->getType() == Ty) 197 return I->getOperand(0); 198 199 // Otherwise, must be the same type of cast, so just reinsert a new one. 200 // This also handles the case of zext(trunc(x)) -> zext(x). 201 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 202 Opc == Instruction::SExt); 203 break; 204 case Instruction::Select: { 205 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 206 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 207 Res = SelectInst::Create(I->getOperand(0), True, False); 208 break; 209 } 210 case Instruction::PHI: { 211 PHINode *OPN = cast<PHINode>(I); 212 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues()); 213 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 214 Value *V = 215 EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 216 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 217 } 218 Res = NPN; 219 break; 220 } 221 default: 222 // TODO: Can handle more cases here. 223 llvm_unreachable("Unreachable!"); 224 } 225 226 Res->takeName(I); 227 return InsertNewInstWith(Res, *I); 228 } 229 230 231 /// This function is a wrapper around CastInst::isEliminableCastPair. It 232 /// simply extracts arguments and returns what that function returns. 233 static Instruction::CastOps 234 isEliminableCastPair(const CastInst *CI, ///< First cast instruction 235 unsigned opcode, ///< Opcode for the second cast 236 Type *DstTy, ///< Target type for the second cast 237 const DataLayout &DL) { 238 Type *SrcTy = CI->getOperand(0)->getType(); // A from above 239 Type *MidTy = CI->getType(); // B from above 240 241 // Get the opcodes of the two Cast instructions 242 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 243 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 244 Type *SrcIntPtrTy = 245 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr; 246 Type *MidIntPtrTy = 247 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr; 248 Type *DstIntPtrTy = 249 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr; 250 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 251 DstTy, SrcIntPtrTy, MidIntPtrTy, 252 DstIntPtrTy); 253 254 // We don't want to form an inttoptr or ptrtoint that converts to an integer 255 // type that differs from the pointer size. 256 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) || 257 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy)) 258 Res = 0; 259 260 return Instruction::CastOps(Res); 261 } 262 263 /// Return true if the cast from "V to Ty" actually results in any code being 264 /// generated and is interesting to optimize out. 265 /// If the cast can be eliminated by some other simple transformation, we prefer 266 /// to do the simplification first. 267 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 268 Type *Ty) { 269 // Noop casts and casts of constants should be eliminated trivially. 270 if (V->getType() == Ty || isa<Constant>(V)) return false; 271 272 // If this is another cast that can be eliminated, we prefer to have it 273 // eliminated. 274 if (const CastInst *CI = dyn_cast<CastInst>(V)) 275 if (isEliminableCastPair(CI, opc, Ty, DL)) 276 return false; 277 278 // If this is a vector sext from a compare, then we don't want to break the 279 // idiom where each element of the extended vector is either zero or all ones. 280 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 281 return false; 282 283 return true; 284 } 285 286 287 /// @brief Implement the transforms common to all CastInst visitors. 288 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 289 Value *Src = CI.getOperand(0); 290 291 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 292 // eliminate it now. 293 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 294 if (Instruction::CastOps opc = 295 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) { 296 // The first cast (CSrc) is eliminable so we need to fix up or replace 297 // the second cast (CI). CSrc will then have a good chance of being dead. 298 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 299 } 300 } 301 302 // If we are casting a select then fold the cast into the select 303 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 304 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 305 return NV; 306 307 // If we are casting a PHI then fold the cast into the PHI 308 if (isa<PHINode>(Src)) { 309 // We don't do this if this would create a PHI node with an illegal type if 310 // it is currently legal. 311 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() || 312 ShouldChangeType(CI.getType(), Src->getType())) 313 if (Instruction *NV = FoldOpIntoPhi(CI)) 314 return NV; 315 } 316 317 return nullptr; 318 } 319 320 /// Return true if we can evaluate the specified expression tree as type Ty 321 /// instead of its larger type, and arrive with the same value. 322 /// This is used by code that tries to eliminate truncates. 323 /// 324 /// Ty will always be a type smaller than V. We should return true if trunc(V) 325 /// can be computed by computing V in the smaller type. If V is an instruction, 326 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 327 /// makes sense if x and y can be efficiently truncated. 328 /// 329 /// This function works on both vectors and scalars. 330 /// 331 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC, 332 Instruction *CxtI) { 333 // We can always evaluate constants in another type. 334 if (isa<Constant>(V)) 335 return true; 336 337 Instruction *I = dyn_cast<Instruction>(V); 338 if (!I) return false; 339 340 Type *OrigTy = V->getType(); 341 342 // If this is an extension from the dest type, we can eliminate it, even if it 343 // has multiple uses. 344 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 345 I->getOperand(0)->getType() == Ty) 346 return true; 347 348 // We can't extend or shrink something that has multiple uses: doing so would 349 // require duplicating the instruction in general, which isn't profitable. 350 if (!I->hasOneUse()) return false; 351 352 unsigned Opc = I->getOpcode(); 353 switch (Opc) { 354 case Instruction::Add: 355 case Instruction::Sub: 356 case Instruction::Mul: 357 case Instruction::And: 358 case Instruction::Or: 359 case Instruction::Xor: 360 // These operators can all arbitrarily be extended or truncated. 361 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) && 362 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI); 363 364 case Instruction::UDiv: 365 case Instruction::URem: { 366 // UDiv and URem can be truncated if all the truncated bits are zero. 367 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 368 uint32_t BitWidth = Ty->getScalarSizeInBits(); 369 if (BitWidth < OrigBitWidth) { 370 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 371 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) && 372 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) { 373 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) && 374 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI); 375 } 376 } 377 break; 378 } 379 case Instruction::Shl: 380 // If we are truncating the result of this SHL, and if it's a shift of a 381 // constant amount, we can always perform a SHL in a smaller type. 382 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 383 uint32_t BitWidth = Ty->getScalarSizeInBits(); 384 if (CI->getLimitedValue(BitWidth) < BitWidth) 385 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI); 386 } 387 break; 388 case Instruction::LShr: 389 // If this is a truncate of a logical shr, we can truncate it to a smaller 390 // lshr iff we know that the bits we would otherwise be shifting in are 391 // already zeros. 392 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 393 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 394 uint32_t BitWidth = Ty->getScalarSizeInBits(); 395 if (IC.MaskedValueIsZero(I->getOperand(0), 396 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) && 397 CI->getLimitedValue(BitWidth) < BitWidth) { 398 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI); 399 } 400 } 401 break; 402 case Instruction::Trunc: 403 // trunc(trunc(x)) -> trunc(x) 404 return true; 405 case Instruction::ZExt: 406 case Instruction::SExt: 407 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 408 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 409 return true; 410 case Instruction::Select: { 411 SelectInst *SI = cast<SelectInst>(I); 412 return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) && 413 canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI); 414 } 415 case Instruction::PHI: { 416 // We can change a phi if we can change all operands. Note that we never 417 // get into trouble with cyclic PHIs here because we only consider 418 // instructions with a single use. 419 PHINode *PN = cast<PHINode>(I); 420 for (Value *IncValue : PN->incoming_values()) 421 if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI)) 422 return false; 423 return true; 424 } 425 default: 426 // TODO: Can handle more cases here. 427 break; 428 } 429 430 return false; 431 } 432 433 /// Given a vector that is bitcast to an integer, optionally logically 434 /// right-shifted, and truncated, convert it to an extractelement. 435 /// Example (big endian): 436 /// trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32 437 /// ---> 438 /// extractelement <4 x i32> %X, 1 439 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc, InstCombiner &IC, 440 const DataLayout &DL) { 441 Value *TruncOp = Trunc.getOperand(0); 442 Type *DestType = Trunc.getType(); 443 if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType)) 444 return nullptr; 445 446 Value *VecInput = nullptr; 447 ConstantInt *ShiftVal = nullptr; 448 if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)), 449 m_LShr(m_BitCast(m_Value(VecInput)), 450 m_ConstantInt(ShiftVal)))) || 451 !isa<VectorType>(VecInput->getType())) 452 return nullptr; 453 454 VectorType *VecType = cast<VectorType>(VecInput->getType()); 455 unsigned VecWidth = VecType->getPrimitiveSizeInBits(); 456 unsigned DestWidth = DestType->getPrimitiveSizeInBits(); 457 unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0; 458 459 if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0)) 460 return nullptr; 461 462 // If the element type of the vector doesn't match the result type, 463 // bitcast it to a vector type that we can extract from. 464 unsigned NumVecElts = VecWidth / DestWidth; 465 if (VecType->getElementType() != DestType) { 466 VecType = VectorType::get(DestType, NumVecElts); 467 VecInput = IC.Builder->CreateBitCast(VecInput, VecType, "bc"); 468 } 469 470 unsigned Elt = ShiftAmount / DestWidth; 471 if (DL.isBigEndian()) 472 Elt = NumVecElts - 1 - Elt; 473 474 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 475 } 476 477 Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 478 if (Instruction *Result = commonCastTransforms(CI)) 479 return Result; 480 481 // Test if the trunc is the user of a select which is part of a 482 // minimum or maximum operation. If so, don't do any more simplification. 483 // Even simplifying demanded bits can break the canonical form of a 484 // min/max. 485 Value *LHS, *RHS; 486 if (SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0))) 487 if (matchSelectPattern(SI, LHS, RHS).Flavor != SPF_UNKNOWN) 488 return nullptr; 489 490 // See if we can simplify any instructions used by the input whose sole 491 // purpose is to compute bits we don't care about. 492 if (SimplifyDemandedInstructionBits(CI)) 493 return &CI; 494 495 Value *Src = CI.getOperand(0); 496 Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 497 498 // Attempt to truncate the entire input expression tree to the destination 499 // type. Only do this if the dest type is a simple type, don't convert the 500 // expression tree to something weird like i93 unless the source is also 501 // strange. 502 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 503 canEvaluateTruncated(Src, DestTy, *this, &CI)) { 504 505 // If this cast is a truncate, evaluting in a different type always 506 // eliminates the cast, so it is always a win. 507 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 508 " to avoid cast: " << CI << '\n'); 509 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 510 assert(Res->getType() == DestTy); 511 return replaceInstUsesWith(CI, Res); 512 } 513 514 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 515 if (DestTy->getScalarSizeInBits() == 1) { 516 Constant *One = ConstantInt::get(SrcTy, 1); 517 Src = Builder->CreateAnd(Src, One); 518 Value *Zero = Constant::getNullValue(Src->getType()); 519 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 520 } 521 522 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 523 Value *A = nullptr; ConstantInt *Cst = nullptr; 524 if (Src->hasOneUse() && 525 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 526 // We have three types to worry about here, the type of A, the source of 527 // the truncate (MidSize), and the destination of the truncate. We know that 528 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 529 // between ASize and ResultSize. 530 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 531 532 // If the shift amount is larger than the size of A, then the result is 533 // known to be zero because all the input bits got shifted out. 534 if (Cst->getZExtValue() >= ASize) 535 return replaceInstUsesWith(CI, Constant::getNullValue(DestTy)); 536 537 // Since we're doing an lshr and a zero extend, and know that the shift 538 // amount is smaller than ASize, it is always safe to do the shift in A's 539 // type, then zero extend or truncate to the result. 540 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 541 Shift->takeName(Src); 542 return CastInst::CreateIntegerCast(Shift, DestTy, false); 543 } 544 545 // Transform trunc(lshr (sext A), Cst) to ashr A, Cst to eliminate type 546 // conversion. 547 // It works because bits coming from sign extension have the same value as 548 // the sign bit of the original value; performing ashr instead of lshr 549 // generates bits of the same value as the sign bit. 550 if (Src->hasOneUse() && 551 match(Src, m_LShr(m_SExt(m_Value(A)), m_ConstantInt(Cst))) && 552 cast<Instruction>(Src)->getOperand(0)->hasOneUse()) { 553 const unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 554 // This optimization can be only performed when zero bits generated by 555 // the original lshr aren't pulled into the value after truncation, so we 556 // can only shift by values smaller than the size of destination type (in 557 // bits). 558 if (Cst->getValue().ult(ASize)) { 559 Value *Shift = Builder->CreateAShr(A, Cst->getZExtValue()); 560 Shift->takeName(Src); 561 return CastInst::CreateIntegerCast(Shift, CI.getType(), true); 562 } 563 } 564 565 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 566 // type isn't non-native. 567 if (Src->hasOneUse() && isa<IntegerType>(SrcTy) && 568 ShouldChangeType(SrcTy, DestTy) && 569 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 570 Value *NewTrunc = Builder->CreateTrunc(A, DestTy, A->getName() + ".tr"); 571 return BinaryOperator::CreateAnd(NewTrunc, 572 ConstantExpr::getTrunc(Cst, DestTy)); 573 } 574 575 if (Instruction *I = foldVecTruncToExtElt(CI, *this, DL)) 576 return I; 577 578 return nullptr; 579 } 580 581 /// Transform (zext icmp) to bitwise / integer operations in order to eliminate 582 /// the icmp. 583 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 584 bool DoXform) { 585 // If we are just checking for a icmp eq of a single bit and zext'ing it 586 // to an integer, then shift the bit to the appropriate place and then 587 // cast to integer to avoid the comparison. 588 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 589 const APInt &Op1CV = Op1C->getValue(); 590 591 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 592 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 593 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 594 (ICI->getPredicate() == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) { 595 if (!DoXform) return ICI; 596 597 Value *In = ICI->getOperand(0); 598 Value *Sh = ConstantInt::get(In->getType(), 599 In->getType()->getScalarSizeInBits() - 1); 600 In = Builder->CreateLShr(In, Sh, In->getName() + ".lobit"); 601 if (In->getType() != CI.getType()) 602 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); 603 604 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 605 Constant *One = ConstantInt::get(In->getType(), 1); 606 In = Builder->CreateXor(In, One, In->getName() + ".not"); 607 } 608 609 return replaceInstUsesWith(CI, In); 610 } 611 612 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 613 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 614 // zext (X == 1) to i32 --> X iff X has only the low bit set. 615 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 616 // zext (X != 0) to i32 --> X iff X has only the low bit set. 617 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 618 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 619 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 620 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 621 // This only works for EQ and NE 622 ICI->isEquality()) { 623 // If Op1C some other power of two, convert: 624 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 625 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 626 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI); 627 628 APInt KnownZeroMask(~KnownZero); 629 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 630 if (!DoXform) return ICI; 631 632 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 633 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 634 // (X&4) == 2 --> false 635 // (X&4) != 2 --> true 636 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 637 isNE); 638 Res = ConstantExpr::getZExt(Res, CI.getType()); 639 return replaceInstUsesWith(CI, Res); 640 } 641 642 uint32_t ShAmt = KnownZeroMask.logBase2(); 643 Value *In = ICI->getOperand(0); 644 if (ShAmt) { 645 // Perform a logical shr by shiftamt. 646 // Insert the shift to put the result in the low bit. 647 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(), ShAmt), 648 In->getName() + ".lobit"); 649 } 650 651 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 652 Constant *One = ConstantInt::get(In->getType(), 1); 653 In = Builder->CreateXor(In, One); 654 } 655 656 if (CI.getType() == In->getType()) 657 return replaceInstUsesWith(CI, In); 658 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 659 } 660 } 661 } 662 663 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 664 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 665 // may lead to additional simplifications. 666 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 667 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 668 uint32_t BitWidth = ITy->getBitWidth(); 669 Value *LHS = ICI->getOperand(0); 670 Value *RHS = ICI->getOperand(1); 671 672 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 673 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 674 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI); 675 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI); 676 677 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 678 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 679 APInt UnknownBit = ~KnownBits; 680 if (UnknownBit.countPopulation() == 1) { 681 if (!DoXform) return ICI; 682 683 Value *Result = Builder->CreateXor(LHS, RHS); 684 685 // Mask off any bits that are set and won't be shifted away. 686 if (KnownOneLHS.uge(UnknownBit)) 687 Result = Builder->CreateAnd(Result, 688 ConstantInt::get(ITy, UnknownBit)); 689 690 // Shift the bit we're testing down to the lsb. 691 Result = Builder->CreateLShr( 692 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 693 694 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 695 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 696 Result->takeName(ICI); 697 return replaceInstUsesWith(CI, Result); 698 } 699 } 700 } 701 } 702 703 return nullptr; 704 } 705 706 /// Determine if the specified value can be computed in the specified wider type 707 /// and produce the same low bits. If not, return false. 708 /// 709 /// If this function returns true, it can also return a non-zero number of bits 710 /// (in BitsToClear) which indicates that the value it computes is correct for 711 /// the zero extend, but that the additional BitsToClear bits need to be zero'd 712 /// out. For example, to promote something like: 713 /// 714 /// %B = trunc i64 %A to i32 715 /// %C = lshr i32 %B, 8 716 /// %E = zext i32 %C to i64 717 /// 718 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 719 /// set to 8 to indicate that the promoted value needs to have bits 24-31 720 /// cleared in addition to bits 32-63. Since an 'and' will be generated to 721 /// clear the top bits anyway, doing this has no extra cost. 722 /// 723 /// This function works on both vectors and scalars. 724 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear, 725 InstCombiner &IC, Instruction *CxtI) { 726 BitsToClear = 0; 727 if (isa<Constant>(V)) 728 return true; 729 730 Instruction *I = dyn_cast<Instruction>(V); 731 if (!I) return false; 732 733 // If the input is a truncate from the destination type, we can trivially 734 // eliminate it. 735 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 736 return true; 737 738 // We can't extend or shrink something that has multiple uses: doing so would 739 // require duplicating the instruction in general, which isn't profitable. 740 if (!I->hasOneUse()) return false; 741 742 unsigned Opc = I->getOpcode(), Tmp; 743 switch (Opc) { 744 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 745 case Instruction::SExt: // zext(sext(x)) -> sext(x). 746 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 747 return true; 748 case Instruction::And: 749 case Instruction::Or: 750 case Instruction::Xor: 751 case Instruction::Add: 752 case Instruction::Sub: 753 case Instruction::Mul: 754 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) || 755 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI)) 756 return false; 757 // These can all be promoted if neither operand has 'bits to clear'. 758 if (BitsToClear == 0 && Tmp == 0) 759 return true; 760 761 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 762 // other side, BitsToClear is ok. 763 if (Tmp == 0 && 764 (Opc == Instruction::And || Opc == Instruction::Or || 765 Opc == Instruction::Xor)) { 766 // We use MaskedValueIsZero here for generality, but the case we care 767 // about the most is constant RHS. 768 unsigned VSize = V->getType()->getScalarSizeInBits(); 769 if (IC.MaskedValueIsZero(I->getOperand(1), 770 APInt::getHighBitsSet(VSize, BitsToClear), 771 0, CxtI)) 772 return true; 773 } 774 775 // Otherwise, we don't know how to analyze this BitsToClear case yet. 776 return false; 777 778 case Instruction::Shl: 779 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the 780 // upper bits we can reduce BitsToClear by the shift amount. 781 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 782 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI)) 783 return false; 784 uint64_t ShiftAmt = Amt->getZExtValue(); 785 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0; 786 return true; 787 } 788 return false; 789 case Instruction::LShr: 790 // We can promote lshr(x, cst) if we can promote x. This requires the 791 // ultimate 'and' to clear out the high zero bits we're clearing out though. 792 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 793 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI)) 794 return false; 795 BitsToClear += Amt->getZExtValue(); 796 if (BitsToClear > V->getType()->getScalarSizeInBits()) 797 BitsToClear = V->getType()->getScalarSizeInBits(); 798 return true; 799 } 800 // Cannot promote variable LSHR. 801 return false; 802 case Instruction::Select: 803 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) || 804 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) || 805 // TODO: If important, we could handle the case when the BitsToClear are 806 // known zero in the disagreeing side. 807 Tmp != BitsToClear) 808 return false; 809 return true; 810 811 case Instruction::PHI: { 812 // We can change a phi if we can change all operands. Note that we never 813 // get into trouble with cyclic PHIs here because we only consider 814 // instructions with a single use. 815 PHINode *PN = cast<PHINode>(I); 816 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI)) 817 return false; 818 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 819 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) || 820 // TODO: If important, we could handle the case when the BitsToClear 821 // are known zero in the disagreeing input. 822 Tmp != BitsToClear) 823 return false; 824 return true; 825 } 826 default: 827 // TODO: Can handle more cases here. 828 return false; 829 } 830 } 831 832 Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 833 // If this zero extend is only used by a truncate, let the truncate be 834 // eliminated before we try to optimize this zext. 835 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back())) 836 return nullptr; 837 838 // If one of the common conversion will work, do it. 839 if (Instruction *Result = commonCastTransforms(CI)) 840 return Result; 841 842 // See if we can simplify any instructions used by the input whose sole 843 // purpose is to compute bits we don't care about. 844 if (SimplifyDemandedInstructionBits(CI)) 845 return &CI; 846 847 Value *Src = CI.getOperand(0); 848 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 849 850 // Attempt to extend the entire input expression tree to the destination 851 // type. Only do this if the dest type is a simple type, don't convert the 852 // expression tree to something weird like i93 unless the source is also 853 // strange. 854 unsigned BitsToClear; 855 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 856 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) { 857 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 858 "Unreasonable BitsToClear"); 859 860 // Okay, we can transform this! Insert the new expression now. 861 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 862 " to avoid zero extend: " << CI << '\n'); 863 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 864 assert(Res->getType() == DestTy); 865 866 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 867 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 868 869 // If the high bits are already filled with zeros, just replace this 870 // cast with the result. 871 if (MaskedValueIsZero(Res, 872 APInt::getHighBitsSet(DestBitSize, 873 DestBitSize-SrcBitsKept), 874 0, &CI)) 875 return replaceInstUsesWith(CI, Res); 876 877 // We need to emit an AND to clear the high bits. 878 Constant *C = ConstantInt::get(Res->getType(), 879 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 880 return BinaryOperator::CreateAnd(Res, C); 881 } 882 883 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 884 // types and if the sizes are just right we can convert this into a logical 885 // 'and' which will be much cheaper than the pair of casts. 886 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 887 // TODO: Subsume this into EvaluateInDifferentType. 888 889 // Get the sizes of the types involved. We know that the intermediate type 890 // will be smaller than A or C, but don't know the relation between A and C. 891 Value *A = CSrc->getOperand(0); 892 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 893 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 894 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 895 // If we're actually extending zero bits, then if 896 // SrcSize < DstSize: zext(a & mask) 897 // SrcSize == DstSize: a & mask 898 // SrcSize > DstSize: trunc(a) & mask 899 if (SrcSize < DstSize) { 900 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 901 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 902 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 903 return new ZExtInst(And, CI.getType()); 904 } 905 906 if (SrcSize == DstSize) { 907 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 908 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 909 AndValue)); 910 } 911 if (SrcSize > DstSize) { 912 Value *Trunc = Builder->CreateTrunc(A, CI.getType()); 913 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 914 return BinaryOperator::CreateAnd(Trunc, 915 ConstantInt::get(Trunc->getType(), 916 AndValue)); 917 } 918 } 919 920 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 921 return transformZExtICmp(ICI, CI); 922 923 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 924 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 925 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 926 // of the (zext icmp) will be transformed. 927 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 928 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 929 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 930 (transformZExtICmp(LHS, CI, false) || 931 transformZExtICmp(RHS, CI, false))) { 932 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 933 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 934 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 935 } 936 } 937 938 // zext(trunc(X) & C) -> (X & zext(C)). 939 Constant *C; 940 Value *X; 941 if (SrcI && 942 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) && 943 X->getType() == CI.getType()) 944 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType())); 945 946 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)). 947 Value *And; 948 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) && 949 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) && 950 X->getType() == CI.getType()) { 951 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 952 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC); 953 } 954 955 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 956 if (SrcI && SrcI->hasOneUse() && 957 SrcI->getType()->getScalarType()->isIntegerTy(1) && 958 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) { 959 Value *New = Builder->CreateZExt(X, CI.getType()); 960 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 961 } 962 963 return nullptr; 964 } 965 966 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp. 967 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { 968 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); 969 ICmpInst::Predicate Pred = ICI->getPredicate(); 970 971 // Don't bother if Op1 isn't of vector or integer type. 972 if (!Op1->getType()->isIntOrIntVectorTy()) 973 return nullptr; 974 975 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 976 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative 977 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive 978 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) || 979 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { 980 981 Value *Sh = ConstantInt::get(Op0->getType(), 982 Op0->getType()->getScalarSizeInBits()-1); 983 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); 984 if (In->getType() != CI.getType()) 985 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); 986 987 if (Pred == ICmpInst::ICMP_SGT) 988 In = Builder->CreateNot(In, In->getName()+".not"); 989 return replaceInstUsesWith(CI, In); 990 } 991 } 992 993 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 994 // If we know that only one bit of the LHS of the icmp can be set and we 995 // have an equality comparison with zero or a power of 2, we can transform 996 // the icmp and sext into bitwise/integer operations. 997 if (ICI->hasOneUse() && 998 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ 999 unsigned BitWidth = Op1C->getType()->getBitWidth(); 1000 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 1001 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI); 1002 1003 APInt KnownZeroMask(~KnownZero); 1004 if (KnownZeroMask.isPowerOf2()) { 1005 Value *In = ICI->getOperand(0); 1006 1007 // If the icmp tests for a known zero bit we can constant fold it. 1008 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) { 1009 Value *V = Pred == ICmpInst::ICMP_NE ? 1010 ConstantInt::getAllOnesValue(CI.getType()) : 1011 ConstantInt::getNullValue(CI.getType()); 1012 return replaceInstUsesWith(CI, V); 1013 } 1014 1015 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) { 1016 // sext ((x & 2^n) == 0) -> (x >> n) - 1 1017 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1 1018 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros(); 1019 // Perform a right shift to place the desired bit in the LSB. 1020 if (ShiftAmt) 1021 In = Builder->CreateLShr(In, 1022 ConstantInt::get(In->getType(), ShiftAmt)); 1023 1024 // At this point "In" is either 1 or 0. Subtract 1 to turn 1025 // {1, 0} -> {0, -1}. 1026 In = Builder->CreateAdd(In, 1027 ConstantInt::getAllOnesValue(In->getType()), 1028 "sext"); 1029 } else { 1030 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1 1031 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1 1032 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros(); 1033 // Perform a left shift to place the desired bit in the MSB. 1034 if (ShiftAmt) 1035 In = Builder->CreateShl(In, 1036 ConstantInt::get(In->getType(), ShiftAmt)); 1037 1038 // Distribute the bit over the whole bit width. 1039 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(), 1040 BitWidth - 1), "sext"); 1041 } 1042 1043 if (CI.getType() == In->getType()) 1044 return replaceInstUsesWith(CI, In); 1045 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/); 1046 } 1047 } 1048 } 1049 1050 return nullptr; 1051 } 1052 1053 /// Return true if we can take the specified value and return it as type Ty 1054 /// without inserting any new casts and without changing the value of the common 1055 /// low bits. This is used by code that tries to promote integer operations to 1056 /// a wider types will allow us to eliminate the extension. 1057 /// 1058 /// This function works on both vectors and scalars. 1059 /// 1060 static bool canEvaluateSExtd(Value *V, Type *Ty) { 1061 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 1062 "Can't sign extend type to a smaller type"); 1063 // If this is a constant, it can be trivially promoted. 1064 if (isa<Constant>(V)) 1065 return true; 1066 1067 Instruction *I = dyn_cast<Instruction>(V); 1068 if (!I) return false; 1069 1070 // If this is a truncate from the dest type, we can trivially eliminate it. 1071 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 1072 return true; 1073 1074 // We can't extend or shrink something that has multiple uses: doing so would 1075 // require duplicating the instruction in general, which isn't profitable. 1076 if (!I->hasOneUse()) return false; 1077 1078 switch (I->getOpcode()) { 1079 case Instruction::SExt: // sext(sext(x)) -> sext(x) 1080 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 1081 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 1082 return true; 1083 case Instruction::And: 1084 case Instruction::Or: 1085 case Instruction::Xor: 1086 case Instruction::Add: 1087 case Instruction::Sub: 1088 case Instruction::Mul: 1089 // These operators can all arbitrarily be extended if their inputs can. 1090 return canEvaluateSExtd(I->getOperand(0), Ty) && 1091 canEvaluateSExtd(I->getOperand(1), Ty); 1092 1093 //case Instruction::Shl: TODO 1094 //case Instruction::LShr: TODO 1095 1096 case Instruction::Select: 1097 return canEvaluateSExtd(I->getOperand(1), Ty) && 1098 canEvaluateSExtd(I->getOperand(2), Ty); 1099 1100 case Instruction::PHI: { 1101 // We can change a phi if we can change all operands. Note that we never 1102 // get into trouble with cyclic PHIs here because we only consider 1103 // instructions with a single use. 1104 PHINode *PN = cast<PHINode>(I); 1105 for (Value *IncValue : PN->incoming_values()) 1106 if (!canEvaluateSExtd(IncValue, Ty)) return false; 1107 return true; 1108 } 1109 default: 1110 // TODO: Can handle more cases here. 1111 break; 1112 } 1113 1114 return false; 1115 } 1116 1117 Instruction *InstCombiner::visitSExt(SExtInst &CI) { 1118 // If this sign extend is only used by a truncate, let the truncate be 1119 // eliminated before we try to optimize this sext. 1120 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back())) 1121 return nullptr; 1122 1123 if (Instruction *I = commonCastTransforms(CI)) 1124 return I; 1125 1126 // See if we can simplify any instructions used by the input whose sole 1127 // purpose is to compute bits we don't care about. 1128 if (SimplifyDemandedInstructionBits(CI)) 1129 return &CI; 1130 1131 Value *Src = CI.getOperand(0); 1132 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 1133 1134 // If we know that the value being extended is positive, we can use a zext 1135 // instead. 1136 bool KnownZero, KnownOne; 1137 ComputeSignBit(Src, KnownZero, KnownOne, 0, &CI); 1138 if (KnownZero) { 1139 Value *ZExt = Builder->CreateZExt(Src, DestTy); 1140 return replaceInstUsesWith(CI, ZExt); 1141 } 1142 1143 // Attempt to extend the entire input expression tree to the destination 1144 // type. Only do this if the dest type is a simple type, don't convert the 1145 // expression tree to something weird like i93 unless the source is also 1146 // strange. 1147 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 1148 canEvaluateSExtd(Src, DestTy)) { 1149 // Okay, we can transform this! Insert the new expression now. 1150 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 1151 " to avoid sign extend: " << CI << '\n'); 1152 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 1153 assert(Res->getType() == DestTy); 1154 1155 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1156 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1157 1158 // If the high bits are already filled with sign bit, just replace this 1159 // cast with the result. 1160 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize) 1161 return replaceInstUsesWith(CI, Res); 1162 1163 // We need to emit a shl + ashr to do the sign extend. 1164 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1165 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 1166 ShAmt); 1167 } 1168 1169 // If this input is a trunc from our destination, then turn sext(trunc(x)) 1170 // into shifts. 1171 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 1172 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 1173 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1174 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1175 1176 // We need to emit a shl + ashr to do the sign extend. 1177 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1178 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1179 return BinaryOperator::CreateAShr(Res, ShAmt); 1180 } 1181 1182 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 1183 return transformSExtICmp(ICI, CI); 1184 1185 // If the input is a shl/ashr pair of a same constant, then this is a sign 1186 // extension from a smaller value. If we could trust arbitrary bitwidth 1187 // integers, we could turn this into a truncate to the smaller bit and then 1188 // use a sext for the whole extension. Since we don't, look deeper and check 1189 // for a truncate. If the source and dest are the same type, eliminate the 1190 // trunc and extend and just do shifts. For example, turn: 1191 // %a = trunc i32 %i to i8 1192 // %b = shl i8 %a, 6 1193 // %c = ashr i8 %b, 6 1194 // %d = sext i8 %c to i32 1195 // into: 1196 // %a = shl i32 %i, 30 1197 // %d = ashr i32 %a, 30 1198 Value *A = nullptr; 1199 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1200 ConstantInt *BA = nullptr, *CA = nullptr; 1201 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1202 m_ConstantInt(CA))) && 1203 BA == CA && A->getType() == CI.getType()) { 1204 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1205 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1206 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1207 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1208 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1209 return BinaryOperator::CreateAShr(A, ShAmtV); 1210 } 1211 1212 return nullptr; 1213 } 1214 1215 1216 /// Return a Constant* for the specified floating-point constant if it fits 1217 /// in the specified FP type without changing its value. 1218 static Constant *fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1219 bool losesInfo; 1220 APFloat F = CFP->getValueAPF(); 1221 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1222 if (!losesInfo) 1223 return ConstantFP::get(CFP->getContext(), F); 1224 return nullptr; 1225 } 1226 1227 /// If this is a floating-point extension instruction, look 1228 /// through it until we get the source value. 1229 static Value *lookThroughFPExtensions(Value *V) { 1230 if (Instruction *I = dyn_cast<Instruction>(V)) 1231 if (I->getOpcode() == Instruction::FPExt) 1232 return lookThroughFPExtensions(I->getOperand(0)); 1233 1234 // If this value is a constant, return the constant in the smallest FP type 1235 // that can accurately represent it. This allows us to turn 1236 // (float)((double)X+2.0) into x+2.0f. 1237 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1238 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1239 return V; // No constant folding of this. 1240 // See if the value can be truncated to half and then reextended. 1241 if (Value *V = fitsInFPType(CFP, APFloat::IEEEhalf)) 1242 return V; 1243 // See if the value can be truncated to float and then reextended. 1244 if (Value *V = fitsInFPType(CFP, APFloat::IEEEsingle)) 1245 return V; 1246 if (CFP->getType()->isDoubleTy()) 1247 return V; // Won't shrink. 1248 if (Value *V = fitsInFPType(CFP, APFloat::IEEEdouble)) 1249 return V; 1250 // Don't try to shrink to various long double types. 1251 } 1252 1253 return V; 1254 } 1255 1256 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1257 if (Instruction *I = commonCastTransforms(CI)) 1258 return I; 1259 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to 1260 // simplify this expression to avoid one or more of the trunc/extend 1261 // operations if we can do so without changing the numerical results. 1262 // 1263 // The exact manner in which the widths of the operands interact to limit 1264 // what we can and cannot do safely varies from operation to operation, and 1265 // is explained below in the various case statements. 1266 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1267 if (OpI && OpI->hasOneUse()) { 1268 Value *LHSOrig = lookThroughFPExtensions(OpI->getOperand(0)); 1269 Value *RHSOrig = lookThroughFPExtensions(OpI->getOperand(1)); 1270 unsigned OpWidth = OpI->getType()->getFPMantissaWidth(); 1271 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth(); 1272 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth(); 1273 unsigned SrcWidth = std::max(LHSWidth, RHSWidth); 1274 unsigned DstWidth = CI.getType()->getFPMantissaWidth(); 1275 switch (OpI->getOpcode()) { 1276 default: break; 1277 case Instruction::FAdd: 1278 case Instruction::FSub: 1279 // For addition and subtraction, the infinitely precise result can 1280 // essentially be arbitrarily wide; proving that double rounding 1281 // will not occur because the result of OpI is exact (as we will for 1282 // FMul, for example) is hopeless. However, we *can* nonetheless 1283 // frequently know that double rounding cannot occur (or that it is 1284 // innocuous) by taking advantage of the specific structure of 1285 // infinitely-precise results that admit double rounding. 1286 // 1287 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient 1288 // to represent both sources, we can guarantee that the double 1289 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis, 1290 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..." 1291 // for proof of this fact). 1292 // 1293 // Note: Figueroa does not consider the case where DstFormat != 1294 // SrcFormat. It's possible (likely even!) that this analysis 1295 // could be tightened for those cases, but they are rare (the main 1296 // case of interest here is (float)((double)float + float)). 1297 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) { 1298 if (LHSOrig->getType() != CI.getType()) 1299 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1300 if (RHSOrig->getType() != CI.getType()) 1301 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1302 Instruction *RI = 1303 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig); 1304 RI->copyFastMathFlags(OpI); 1305 return RI; 1306 } 1307 break; 1308 case Instruction::FMul: 1309 // For multiplication, the infinitely precise result has at most 1310 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient 1311 // that such a value can be exactly represented, then no double 1312 // rounding can possibly occur; we can safely perform the operation 1313 // in the destination format if it can represent both sources. 1314 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) { 1315 if (LHSOrig->getType() != CI.getType()) 1316 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1317 if (RHSOrig->getType() != CI.getType()) 1318 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1319 Instruction *RI = 1320 BinaryOperator::CreateFMul(LHSOrig, RHSOrig); 1321 RI->copyFastMathFlags(OpI); 1322 return RI; 1323 } 1324 break; 1325 case Instruction::FDiv: 1326 // For division, we use again use the bound from Figueroa's 1327 // dissertation. I am entirely certain that this bound can be 1328 // tightened in the unbalanced operand case by an analysis based on 1329 // the diophantine rational approximation bound, but the well-known 1330 // condition used here is a good conservative first pass. 1331 // TODO: Tighten bound via rigorous analysis of the unbalanced case. 1332 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) { 1333 if (LHSOrig->getType() != CI.getType()) 1334 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1335 if (RHSOrig->getType() != CI.getType()) 1336 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1337 Instruction *RI = 1338 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig); 1339 RI->copyFastMathFlags(OpI); 1340 return RI; 1341 } 1342 break; 1343 case Instruction::FRem: 1344 // Remainder is straightforward. Remainder is always exact, so the 1345 // type of OpI doesn't enter into things at all. We simply evaluate 1346 // in whichever source type is larger, then convert to the 1347 // destination type. 1348 if (SrcWidth == OpWidth) 1349 break; 1350 if (LHSWidth < SrcWidth) 1351 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType()); 1352 else if (RHSWidth <= SrcWidth) 1353 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType()); 1354 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) { 1355 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig); 1356 if (Instruction *RI = dyn_cast<Instruction>(ExactResult)) 1357 RI->copyFastMathFlags(OpI); 1358 return CastInst::CreateFPCast(ExactResult, CI.getType()); 1359 } 1360 } 1361 1362 // (fptrunc (fneg x)) -> (fneg (fptrunc x)) 1363 if (BinaryOperator::isFNeg(OpI)) { 1364 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1), 1365 CI.getType()); 1366 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc); 1367 RI->copyFastMathFlags(OpI); 1368 return RI; 1369 } 1370 } 1371 1372 // (fptrunc (select cond, R1, Cst)) --> 1373 // (select cond, (fptrunc R1), (fptrunc Cst)) 1374 // 1375 // - but only if this isn't part of a min/max operation, else we'll 1376 // ruin min/max canonical form which is to have the select and 1377 // compare's operands be of the same type with no casts to look through. 1378 Value *LHS, *RHS; 1379 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)); 1380 if (SI && 1381 (isa<ConstantFP>(SI->getOperand(1)) || 1382 isa<ConstantFP>(SI->getOperand(2))) && 1383 matchSelectPattern(SI, LHS, RHS).Flavor == SPF_UNKNOWN) { 1384 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1), 1385 CI.getType()); 1386 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2), 1387 CI.getType()); 1388 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc); 1389 } 1390 1391 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0)); 1392 if (II) { 1393 switch (II->getIntrinsicID()) { 1394 default: break; 1395 case Intrinsic::fabs: { 1396 // (fptrunc (fabs x)) -> (fabs (fptrunc x)) 1397 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0), 1398 CI.getType()); 1399 Type *IntrinsicType[] = { CI.getType() }; 1400 Function *Overload = Intrinsic::getDeclaration( 1401 CI.getModule(), II->getIntrinsicID(), IntrinsicType); 1402 1403 SmallVector<OperandBundleDef, 1> OpBundles; 1404 II->getOperandBundlesAsDefs(OpBundles); 1405 1406 Value *Args[] = { InnerTrunc }; 1407 return CallInst::Create(Overload, Args, OpBundles, II->getName()); 1408 } 1409 } 1410 } 1411 1412 return nullptr; 1413 } 1414 1415 Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1416 return commonCastTransforms(CI); 1417 } 1418 1419 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X) 1420 // This is safe if the intermediate type has enough bits in its mantissa to 1421 // accurately represent all values of X. For example, this won't work with 1422 // i64 -> float -> i64. 1423 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) { 1424 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0))) 1425 return nullptr; 1426 Instruction *OpI = cast<Instruction>(FI.getOperand(0)); 1427 1428 Value *SrcI = OpI->getOperand(0); 1429 Type *FITy = FI.getType(); 1430 Type *OpITy = OpI->getType(); 1431 Type *SrcTy = SrcI->getType(); 1432 bool IsInputSigned = isa<SIToFPInst>(OpI); 1433 bool IsOutputSigned = isa<FPToSIInst>(FI); 1434 1435 // We can safely assume the conversion won't overflow the output range, 1436 // because (for example) (uint8_t)18293.f is undefined behavior. 1437 1438 // Since we can assume the conversion won't overflow, our decision as to 1439 // whether the input will fit in the float should depend on the minimum 1440 // of the input range and output range. 1441 1442 // This means this is also safe for a signed input and unsigned output, since 1443 // a negative input would lead to undefined behavior. 1444 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned; 1445 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned; 1446 int ActualSize = std::min(InputSize, OutputSize); 1447 1448 if (ActualSize <= OpITy->getFPMantissaWidth()) { 1449 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) { 1450 if (IsInputSigned && IsOutputSigned) 1451 return new SExtInst(SrcI, FITy); 1452 return new ZExtInst(SrcI, FITy); 1453 } 1454 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits()) 1455 return new TruncInst(SrcI, FITy); 1456 if (SrcTy == FITy) 1457 return replaceInstUsesWith(FI, SrcI); 1458 return new BitCastInst(SrcI, FITy); 1459 } 1460 return nullptr; 1461 } 1462 1463 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1464 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1465 if (!OpI) 1466 return commonCastTransforms(FI); 1467 1468 if (Instruction *I = FoldItoFPtoI(FI)) 1469 return I; 1470 1471 return commonCastTransforms(FI); 1472 } 1473 1474 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1475 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1476 if (!OpI) 1477 return commonCastTransforms(FI); 1478 1479 if (Instruction *I = FoldItoFPtoI(FI)) 1480 return I; 1481 1482 return commonCastTransforms(FI); 1483 } 1484 1485 Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1486 return commonCastTransforms(CI); 1487 } 1488 1489 Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1490 return commonCastTransforms(CI); 1491 } 1492 1493 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1494 // If the source integer type is not the intptr_t type for this target, do a 1495 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1496 // cast to be exposed to other transforms. 1497 unsigned AS = CI.getAddressSpace(); 1498 if (CI.getOperand(0)->getType()->getScalarSizeInBits() != 1499 DL.getPointerSizeInBits(AS)) { 1500 Type *Ty = DL.getIntPtrType(CI.getContext(), AS); 1501 if (CI.getType()->isVectorTy()) // Handle vectors of pointers. 1502 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements()); 1503 1504 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty); 1505 return new IntToPtrInst(P, CI.getType()); 1506 } 1507 1508 if (Instruction *I = commonCastTransforms(CI)) 1509 return I; 1510 1511 return nullptr; 1512 } 1513 1514 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1515 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1516 Value *Src = CI.getOperand(0); 1517 1518 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1519 // If casting the result of a getelementptr instruction with no offset, turn 1520 // this into a cast of the original pointer! 1521 if (GEP->hasAllZeroIndices() && 1522 // If CI is an addrspacecast and GEP changes the poiner type, merging 1523 // GEP into CI would undo canonicalizing addrspacecast with different 1524 // pointer types, causing infinite loops. 1525 (!isa<AddrSpaceCastInst>(CI) || 1526 GEP->getType() == GEP->getPointerOperand()->getType())) { 1527 // Changing the cast operand is usually not a good idea but it is safe 1528 // here because the pointer operand is being replaced with another 1529 // pointer operand so the opcode doesn't need to change. 1530 Worklist.Add(GEP); 1531 CI.setOperand(0, GEP->getOperand(0)); 1532 return &CI; 1533 } 1534 } 1535 1536 return commonCastTransforms(CI); 1537 } 1538 1539 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1540 // If the destination integer type is not the intptr_t type for this target, 1541 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1542 // to be exposed to other transforms. 1543 1544 Type *Ty = CI.getType(); 1545 unsigned AS = CI.getPointerAddressSpace(); 1546 1547 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS)) 1548 return commonPointerCastTransforms(CI); 1549 1550 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS); 1551 if (Ty->isVectorTy()) // Handle vectors of pointers. 1552 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements()); 1553 1554 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy); 1555 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false); 1556 } 1557 1558 /// This input value (which is known to have vector type) is being zero extended 1559 /// or truncated to the specified vector type. 1560 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible. 1561 /// 1562 /// The source and destination vector types may have different element types. 1563 static Instruction *optimizeVectorResize(Value *InVal, VectorType *DestTy, 1564 InstCombiner &IC) { 1565 // We can only do this optimization if the output is a multiple of the input 1566 // element size, or the input is a multiple of the output element size. 1567 // Convert the input type to have the same element type as the output. 1568 VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1569 1570 if (SrcTy->getElementType() != DestTy->getElementType()) { 1571 // The input types don't need to be identical, but for now they must be the 1572 // same size. There is no specific reason we couldn't handle things like 1573 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1574 // there yet. 1575 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1576 DestTy->getElementType()->getPrimitiveSizeInBits()) 1577 return nullptr; 1578 1579 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1580 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1581 } 1582 1583 // Now that the element types match, get the shuffle mask and RHS of the 1584 // shuffle to use, which depends on whether we're increasing or decreasing the 1585 // size of the input. 1586 SmallVector<uint32_t, 16> ShuffleMask; 1587 Value *V2; 1588 1589 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1590 // If we're shrinking the number of elements, just shuffle in the low 1591 // elements from the input and use undef as the second shuffle input. 1592 V2 = UndefValue::get(SrcTy); 1593 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1594 ShuffleMask.push_back(i); 1595 1596 } else { 1597 // If we're increasing the number of elements, shuffle in all of the 1598 // elements from InVal and fill the rest of the result elements with zeros 1599 // from a constant zero. 1600 V2 = Constant::getNullValue(SrcTy); 1601 unsigned SrcElts = SrcTy->getNumElements(); 1602 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1603 ShuffleMask.push_back(i); 1604 1605 // The excess elements reference the first element of the zero input. 1606 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i) 1607 ShuffleMask.push_back(SrcElts); 1608 } 1609 1610 return new ShuffleVectorInst(InVal, V2, 1611 ConstantDataVector::get(V2->getContext(), 1612 ShuffleMask)); 1613 } 1614 1615 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { 1616 return Value % Ty->getPrimitiveSizeInBits() == 0; 1617 } 1618 1619 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { 1620 return Value / Ty->getPrimitiveSizeInBits(); 1621 } 1622 1623 /// V is a value which is inserted into a vector of VecEltTy. 1624 /// Look through the value to see if we can decompose it into 1625 /// insertions into the vector. See the example in the comment for 1626 /// OptimizeIntegerToVectorInsertions for the pattern this handles. 1627 /// The type of V is always a non-zero multiple of VecEltTy's size. 1628 /// Shift is the number of bits between the lsb of V and the lsb of 1629 /// the vector. 1630 /// 1631 /// This returns false if the pattern can't be matched or true if it can, 1632 /// filling in Elements with the elements found here. 1633 static bool collectInsertionElements(Value *V, unsigned Shift, 1634 SmallVectorImpl<Value *> &Elements, 1635 Type *VecEltTy, bool isBigEndian) { 1636 assert(isMultipleOfTypeSize(Shift, VecEltTy) && 1637 "Shift should be a multiple of the element type size"); 1638 1639 // Undef values never contribute useful bits to the result. 1640 if (isa<UndefValue>(V)) return true; 1641 1642 // If we got down to a value of the right type, we win, try inserting into the 1643 // right element. 1644 if (V->getType() == VecEltTy) { 1645 // Inserting null doesn't actually insert any elements. 1646 if (Constant *C = dyn_cast<Constant>(V)) 1647 if (C->isNullValue()) 1648 return true; 1649 1650 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy); 1651 if (isBigEndian) 1652 ElementIndex = Elements.size() - ElementIndex - 1; 1653 1654 // Fail if multiple elements are inserted into this slot. 1655 if (Elements[ElementIndex]) 1656 return false; 1657 1658 Elements[ElementIndex] = V; 1659 return true; 1660 } 1661 1662 if (Constant *C = dyn_cast<Constant>(V)) { 1663 // Figure out the # elements this provides, and bitcast it or slice it up 1664 // as required. 1665 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1666 VecEltTy); 1667 // If the constant is the size of a vector element, we just need to bitcast 1668 // it to the right type so it gets properly inserted. 1669 if (NumElts == 1) 1670 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1671 Shift, Elements, VecEltTy, isBigEndian); 1672 1673 // Okay, this is a constant that covers multiple elements. Slice it up into 1674 // pieces and insert each element-sized piece into the vector. 1675 if (!isa<IntegerType>(C->getType())) 1676 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1677 C->getType()->getPrimitiveSizeInBits())); 1678 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1679 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1680 1681 for (unsigned i = 0; i != NumElts; ++i) { 1682 unsigned ShiftI = Shift+i*ElementSize; 1683 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1684 ShiftI)); 1685 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1686 if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy, 1687 isBigEndian)) 1688 return false; 1689 } 1690 return true; 1691 } 1692 1693 if (!V->hasOneUse()) return false; 1694 1695 Instruction *I = dyn_cast<Instruction>(V); 1696 if (!I) return false; 1697 switch (I->getOpcode()) { 1698 default: return false; // Unhandled case. 1699 case Instruction::BitCast: 1700 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy, 1701 isBigEndian); 1702 case Instruction::ZExt: 1703 if (!isMultipleOfTypeSize( 1704 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1705 VecEltTy)) 1706 return false; 1707 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy, 1708 isBigEndian); 1709 case Instruction::Or: 1710 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy, 1711 isBigEndian) && 1712 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy, 1713 isBigEndian); 1714 case Instruction::Shl: { 1715 // Must be shifting by a constant that is a multiple of the element size. 1716 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1717 if (!CI) return false; 1718 Shift += CI->getZExtValue(); 1719 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false; 1720 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy, 1721 isBigEndian); 1722 } 1723 1724 } 1725 } 1726 1727 1728 /// If the input is an 'or' instruction, we may be doing shifts and ors to 1729 /// assemble the elements of the vector manually. 1730 /// Try to rip the code out and replace it with insertelements. This is to 1731 /// optimize code like this: 1732 /// 1733 /// %tmp37 = bitcast float %inc to i32 1734 /// %tmp38 = zext i32 %tmp37 to i64 1735 /// %tmp31 = bitcast float %inc5 to i32 1736 /// %tmp32 = zext i32 %tmp31 to i64 1737 /// %tmp33 = shl i64 %tmp32, 32 1738 /// %ins35 = or i64 %tmp33, %tmp38 1739 /// %tmp43 = bitcast i64 %ins35 to <2 x float> 1740 /// 1741 /// Into two insertelements that do "buildvector{%inc, %inc5}". 1742 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI, 1743 InstCombiner &IC) { 1744 VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1745 Value *IntInput = CI.getOperand(0); 1746 1747 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1748 if (!collectInsertionElements(IntInput, 0, Elements, 1749 DestVecTy->getElementType(), 1750 IC.getDataLayout().isBigEndian())) 1751 return nullptr; 1752 1753 // If we succeeded, we know that all of the element are specified by Elements 1754 // or are zero if Elements has a null entry. Recast this as a set of 1755 // insertions. 1756 Value *Result = Constant::getNullValue(CI.getType()); 1757 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1758 if (!Elements[i]) continue; // Unset element. 1759 1760 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1761 IC.Builder->getInt32(i)); 1762 } 1763 1764 return Result; 1765 } 1766 1767 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the 1768 /// vector followed by extract element. The backend tends to handle bitcasts of 1769 /// vectors better than bitcasts of scalars because vector registers are 1770 /// usually not type-specific like scalar integer or scalar floating-point. 1771 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast, 1772 InstCombiner &IC, 1773 const DataLayout &DL) { 1774 // TODO: Create and use a pattern matcher for ExtractElementInst. 1775 auto *ExtElt = dyn_cast<ExtractElementInst>(BitCast.getOperand(0)); 1776 if (!ExtElt || !ExtElt->hasOneUse()) 1777 return nullptr; 1778 1779 // The bitcast must be to a vectorizable type, otherwise we can't make a new 1780 // type to extract from. 1781 Type *DestType = BitCast.getType(); 1782 if (!VectorType::isValidElementType(DestType)) 1783 return nullptr; 1784 1785 unsigned NumElts = ExtElt->getVectorOperandType()->getNumElements(); 1786 auto *NewVecType = VectorType::get(DestType, NumElts); 1787 auto *NewBC = IC.Builder->CreateBitCast(ExtElt->getVectorOperand(), 1788 NewVecType, "bc"); 1789 return ExtractElementInst::Create(NewBC, ExtElt->getIndexOperand()); 1790 } 1791 1792 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1793 // If the operands are integer typed then apply the integer transforms, 1794 // otherwise just apply the common ones. 1795 Value *Src = CI.getOperand(0); 1796 Type *SrcTy = Src->getType(); 1797 Type *DestTy = CI.getType(); 1798 1799 // Get rid of casts from one type to the same type. These are useless and can 1800 // be replaced by the operand. 1801 if (DestTy == Src->getType()) 1802 return replaceInstUsesWith(CI, Src); 1803 1804 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1805 PointerType *SrcPTy = cast<PointerType>(SrcTy); 1806 Type *DstElTy = DstPTy->getElementType(); 1807 Type *SrcElTy = SrcPTy->getElementType(); 1808 1809 // If we are casting a alloca to a pointer to a type of the same 1810 // size, rewrite the allocation instruction to allocate the "right" type. 1811 // There is no need to modify malloc calls because it is their bitcast that 1812 // needs to be cleaned up. 1813 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1814 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1815 return V; 1816 1817 // When the type pointed to is not sized the cast cannot be 1818 // turned into a gep. 1819 Type *PointeeType = 1820 cast<PointerType>(Src->getType()->getScalarType())->getElementType(); 1821 if (!PointeeType->isSized()) 1822 return nullptr; 1823 1824 // If the source and destination are pointers, and this cast is equivalent 1825 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1826 // This can enhance SROA and other transforms that want type-safe pointers. 1827 unsigned NumZeros = 0; 1828 while (SrcElTy != DstElTy && 1829 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1830 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1831 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U); 1832 ++NumZeros; 1833 } 1834 1835 // If we found a path from the src to dest, create the getelementptr now. 1836 if (SrcElTy == DstElTy) { 1837 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder->getInt32(0)); 1838 return GetElementPtrInst::CreateInBounds(Src, Idxs); 1839 } 1840 } 1841 1842 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1843 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1844 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1845 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1846 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1847 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1848 } 1849 1850 if (isa<IntegerType>(SrcTy)) { 1851 // If this is a cast from an integer to vector, check to see if the input 1852 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1853 // the casts with a shuffle and (potentially) a bitcast. 1854 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1855 CastInst *SrcCast = cast<CastInst>(Src); 1856 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1857 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1858 if (Instruction *I = optimizeVectorResize(BCIn->getOperand(0), 1859 cast<VectorType>(DestTy), *this)) 1860 return I; 1861 } 1862 1863 // If the input is an 'or' instruction, we may be doing shifts and ors to 1864 // assemble the elements of the vector manually. Try to rip the code out 1865 // and replace it with insertelements. 1866 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this)) 1867 return replaceInstUsesWith(CI, V); 1868 } 1869 } 1870 1871 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1872 if (SrcVTy->getNumElements() == 1) { 1873 // If our destination is not a vector, then make this a straight 1874 // scalar-scalar cast. 1875 if (!DestTy->isVectorTy()) { 1876 Value *Elem = 1877 Builder->CreateExtractElement(Src, 1878 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1879 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1880 } 1881 1882 // Otherwise, see if our source is an insert. If so, then use the scalar 1883 // component directly. 1884 if (InsertElementInst *IEI = 1885 dyn_cast<InsertElementInst>(CI.getOperand(0))) 1886 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1), 1887 DestTy); 1888 } 1889 } 1890 1891 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1892 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1893 // a bitcast to a vector with the same # elts. 1894 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1895 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() && 1896 SVI->getType()->getNumElements() == 1897 SVI->getOperand(0)->getType()->getVectorNumElements()) { 1898 BitCastInst *Tmp; 1899 // If either of the operands is a cast from CI.getType(), then 1900 // evaluating the shuffle in the casted destination's type will allow 1901 // us to eliminate at least one cast. 1902 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1903 Tmp->getOperand(0)->getType() == DestTy) || 1904 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1905 Tmp->getOperand(0)->getType() == DestTy)) { 1906 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1907 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1908 // Return a new shuffle vector. Use the same element ID's, as we 1909 // know the vector types match #elts. 1910 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1911 } 1912 } 1913 } 1914 1915 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this, DL)) 1916 return I; 1917 1918 if (SrcTy->isPointerTy()) 1919 return commonPointerCastTransforms(CI); 1920 return commonCastTransforms(CI); 1921 } 1922 1923 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) { 1924 // If the destination pointer element type is not the same as the source's 1925 // first do a bitcast to the destination type, and then the addrspacecast. 1926 // This allows the cast to be exposed to other transforms. 1927 Value *Src = CI.getOperand(0); 1928 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType()); 1929 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType()); 1930 1931 Type *DestElemTy = DestTy->getElementType(); 1932 if (SrcTy->getElementType() != DestElemTy) { 1933 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace()); 1934 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) { 1935 // Handle vectors of pointers. 1936 MidTy = VectorType::get(MidTy, VT->getNumElements()); 1937 } 1938 1939 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy); 1940 return new AddrSpaceCastInst(NewBitCast, CI.getType()); 1941 } 1942 1943 return commonPointerCastTransforms(CI); 1944 } 1945