InstCombineCasts.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===- InstCombineCasts.cpp -----------------------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the visit functions for cast operations. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Analysis/ConstantFolding.h" 16#include "llvm/IR/DataLayout.h" 17#include "llvm/IR/PatternMatch.h" 18#include "llvm/Target/TargetLibraryInfo.h" 19using namespace llvm; 20using namespace PatternMatch; 21 22#define DEBUG_TYPE "instcombine" 23 24/// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear 25/// expression. If so, decompose it, returning some value X, such that Val is 26/// X*Scale+Offset. 27/// 28static 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/// PromoteCastOfAllocation - If we find a cast of an allocation instruction, 80/// try to eliminate the cast by moving the type information into the alloc. 81Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 82 AllocaInst &AI) { 83 // This requires DataLayout to get the alloca alignment and size information. 84 if (!DL) return nullptr; 85 86 PointerType *PTy = cast<PointerType>(CI.getType()); 87 88 BuilderTy AllocaBuilder(*Builder); 89 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); 90 91 // Get the type really allocated and the type casted to. 92 Type *AllocElTy = AI.getAllocatedType(); 93 Type *CastElTy = PTy->getElementType(); 94 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr; 95 96 unsigned AllocElTyAlign = DL->getABITypeAlignment(AllocElTy); 97 unsigned CastElTyAlign = DL->getABITypeAlignment(CastElTy); 98 if (CastElTyAlign < AllocElTyAlign) return nullptr; 99 100 // If the allocation has multiple uses, only promote it if we are strictly 101 // increasing the alignment of the resultant allocation. If we keep it the 102 // same, we open the door to infinite loops of various kinds. 103 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr; 104 105 uint64_t AllocElTySize = DL->getTypeAllocSize(AllocElTy); 106 uint64_t CastElTySize = DL->getTypeAllocSize(CastElTy); 107 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr; 108 109 // If the allocation has multiple uses, only promote it if we're not 110 // shrinking the amount of memory being allocated. 111 uint64_t AllocElTyStoreSize = DL->getTypeStoreSize(AllocElTy); 112 uint64_t CastElTyStoreSize = DL->getTypeStoreSize(CastElTy); 113 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr; 114 115 // See if we can satisfy the modulus by pulling a scale out of the array 116 // size argument. 117 unsigned ArraySizeScale; 118 uint64_t ArrayOffset; 119 Value *NumElements = // See if the array size is a decomposable linear expr. 120 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 121 122 // If we can now satisfy the modulus, by using a non-1 scale, we really can 123 // do the xform. 124 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 125 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr; 126 127 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 128 Value *Amt = nullptr; 129 if (Scale == 1) { 130 Amt = NumElements; 131 } else { 132 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 133 // Insert before the alloca, not before the cast. 134 Amt = AllocaBuilder.CreateMul(Amt, NumElements); 135 } 136 137 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 138 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 139 Offset, true); 140 Amt = AllocaBuilder.CreateAdd(Amt, Off); 141 } 142 143 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 144 New->setAlignment(AI.getAlignment()); 145 New->takeName(&AI); 146 New->setUsedWithInAlloca(AI.isUsedWithInAlloca()); 147 148 // If the allocation has multiple real uses, insert a cast and change all 149 // things that used it to use the new cast. This will also hack on CI, but it 150 // will die soon. 151 if (!AI.hasOneUse()) { 152 // New is the allocation instruction, pointer typed. AI is the original 153 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 154 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 155 ReplaceInstUsesWith(AI, NewCast); 156 } 157 return ReplaceInstUsesWith(CI, New); 158} 159 160/// EvaluateInDifferentType - Given an expression that 161/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 162/// insert the code to evaluate the expression. 163Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, 164 bool isSigned) { 165 if (Constant *C = dyn_cast<Constant>(V)) { 166 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 167 // If we got a constantexpr back, try to simplify it with DL info. 168 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 169 C = ConstantFoldConstantExpression(CE, DL, TLI); 170 return C; 171 } 172 173 // Otherwise, it must be an instruction. 174 Instruction *I = cast<Instruction>(V); 175 Instruction *Res = nullptr; 176 unsigned Opc = I->getOpcode(); 177 switch (Opc) { 178 case Instruction::Add: 179 case Instruction::Sub: 180 case Instruction::Mul: 181 case Instruction::And: 182 case Instruction::Or: 183 case Instruction::Xor: 184 case Instruction::AShr: 185 case Instruction::LShr: 186 case Instruction::Shl: 187 case Instruction::UDiv: 188 case Instruction::URem: { 189 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 190 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 191 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 192 break; 193 } 194 case Instruction::Trunc: 195 case Instruction::ZExt: 196 case Instruction::SExt: 197 // If the source type of the cast is the type we're trying for then we can 198 // just return the source. There's no need to insert it because it is not 199 // new. 200 if (I->getOperand(0)->getType() == Ty) 201 return I->getOperand(0); 202 203 // Otherwise, must be the same type of cast, so just reinsert a new one. 204 // This also handles the case of zext(trunc(x)) -> zext(x). 205 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 206 Opc == Instruction::SExt); 207 break; 208 case Instruction::Select: { 209 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 210 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 211 Res = SelectInst::Create(I->getOperand(0), True, False); 212 break; 213 } 214 case Instruction::PHI: { 215 PHINode *OPN = cast<PHINode>(I); 216 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues()); 217 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 218 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 219 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 220 } 221 Res = NPN; 222 break; 223 } 224 default: 225 // TODO: Can handle more cases here. 226 llvm_unreachable("Unreachable!"); 227 } 228 229 Res->takeName(I); 230 return InsertNewInstWith(Res, *I); 231} 232 233 234/// This function is a wrapper around CastInst::isEliminableCastPair. It 235/// simply extracts arguments and returns what that function returns. 236static Instruction::CastOps 237isEliminableCastPair( 238 const CastInst *CI, ///< The first cast instruction 239 unsigned opcode, ///< The opcode of the second cast instruction 240 Type *DstTy, ///< The target type for the second cast instruction 241 const DataLayout *DL ///< The target data for pointer size 242) { 243 244 Type *SrcTy = CI->getOperand(0)->getType(); // A from above 245 Type *MidTy = CI->getType(); // B from above 246 247 // Get the opcodes of the two Cast instructions 248 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 249 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 250 Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ? 251 DL->getIntPtrType(SrcTy) : nullptr; 252 Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ? 253 DL->getIntPtrType(MidTy) : nullptr; 254 Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ? 255 DL->getIntPtrType(DstTy) : nullptr; 256 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 257 DstTy, SrcIntPtrTy, MidIntPtrTy, 258 DstIntPtrTy); 259 260 // We don't want to form an inttoptr or ptrtoint that converts to an integer 261 // type that differs from the pointer size. 262 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) || 263 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy)) 264 Res = 0; 265 266 return Instruction::CastOps(Res); 267} 268 269/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 270/// results in any code being generated and is interesting to optimize out. If 271/// the cast can be eliminated by some other simple transformation, we prefer 272/// to do the simplification first. 273bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 274 Type *Ty) { 275 // Noop casts and casts of constants should be eliminated trivially. 276 if (V->getType() == Ty || isa<Constant>(V)) return false; 277 278 // If this is another cast that can be eliminated, we prefer to have it 279 // eliminated. 280 if (const CastInst *CI = dyn_cast<CastInst>(V)) 281 if (isEliminableCastPair(CI, opc, Ty, DL)) 282 return false; 283 284 // If this is a vector sext from a compare, then we don't want to break the 285 // idiom where each element of the extended vector is either zero or all ones. 286 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 287 return false; 288 289 return true; 290} 291 292 293/// @brief Implement the transforms common to all CastInst visitors. 294Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 295 Value *Src = CI.getOperand(0); 296 297 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 298 // eliminate it now. 299 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 300 if (Instruction::CastOps opc = 301 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) { 302 // The first cast (CSrc) is eliminable so we need to fix up or replace 303 // the second cast (CI). CSrc will then have a good chance of being dead. 304 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 305 } 306 } 307 308 // If we are casting a select then fold the cast into the select 309 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 310 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 311 return NV; 312 313 // If we are casting a PHI then fold the cast into the PHI 314 if (isa<PHINode>(Src)) { 315 // We don't do this if this would create a PHI node with an illegal type if 316 // it is currently legal. 317 if (!Src->getType()->isIntegerTy() || 318 !CI.getType()->isIntegerTy() || 319 ShouldChangeType(CI.getType(), Src->getType())) 320 if (Instruction *NV = FoldOpIntoPhi(CI)) 321 return NV; 322 } 323 324 return nullptr; 325} 326 327/// CanEvaluateTruncated - Return true if we can evaluate the specified 328/// expression tree as type Ty instead of its larger type, and arrive with the 329/// same value. This is used by code that tries to eliminate truncates. 330/// 331/// Ty will always be a type smaller than V. We should return true if trunc(V) 332/// can be computed by computing V in the smaller type. If V is an instruction, 333/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 334/// makes sense if x and y can be efficiently truncated. 335/// 336/// This function works on both vectors and scalars. 337/// 338static bool CanEvaluateTruncated(Value *V, Type *Ty) { 339 // We can always evaluate constants in another type. 340 if (isa<Constant>(V)) 341 return true; 342 343 Instruction *I = dyn_cast<Instruction>(V); 344 if (!I) return false; 345 346 Type *OrigTy = V->getType(); 347 348 // If this is an extension from the dest type, we can eliminate it, even if it 349 // has multiple uses. 350 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 351 I->getOperand(0)->getType() == Ty) 352 return true; 353 354 // We can't extend or shrink something that has multiple uses: doing so would 355 // require duplicating the instruction in general, which isn't profitable. 356 if (!I->hasOneUse()) return false; 357 358 unsigned Opc = I->getOpcode(); 359 switch (Opc) { 360 case Instruction::Add: 361 case Instruction::Sub: 362 case Instruction::Mul: 363 case Instruction::And: 364 case Instruction::Or: 365 case Instruction::Xor: 366 // These operators can all arbitrarily be extended or truncated. 367 return CanEvaluateTruncated(I->getOperand(0), Ty) && 368 CanEvaluateTruncated(I->getOperand(1), Ty); 369 370 case Instruction::UDiv: 371 case Instruction::URem: { 372 // UDiv and URem can be truncated if all the truncated bits are zero. 373 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 374 uint32_t BitWidth = Ty->getScalarSizeInBits(); 375 if (BitWidth < OrigBitWidth) { 376 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 377 if (MaskedValueIsZero(I->getOperand(0), Mask) && 378 MaskedValueIsZero(I->getOperand(1), Mask)) { 379 return CanEvaluateTruncated(I->getOperand(0), Ty) && 380 CanEvaluateTruncated(I->getOperand(1), Ty); 381 } 382 } 383 break; 384 } 385 case Instruction::Shl: 386 // If we are truncating the result of this SHL, and if it's a shift of a 387 // constant amount, we can always perform a SHL in a smaller type. 388 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 389 uint32_t BitWidth = Ty->getScalarSizeInBits(); 390 if (CI->getLimitedValue(BitWidth) < BitWidth) 391 return CanEvaluateTruncated(I->getOperand(0), Ty); 392 } 393 break; 394 case Instruction::LShr: 395 // If this is a truncate of a logical shr, we can truncate it to a smaller 396 // lshr iff we know that the bits we would otherwise be shifting in are 397 // already zeros. 398 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 399 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 400 uint32_t BitWidth = Ty->getScalarSizeInBits(); 401 if (MaskedValueIsZero(I->getOperand(0), 402 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 403 CI->getLimitedValue(BitWidth) < BitWidth) { 404 return CanEvaluateTruncated(I->getOperand(0), Ty); 405 } 406 } 407 break; 408 case Instruction::Trunc: 409 // trunc(trunc(x)) -> trunc(x) 410 return true; 411 case Instruction::ZExt: 412 case Instruction::SExt: 413 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 414 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 415 return true; 416 case Instruction::Select: { 417 SelectInst *SI = cast<SelectInst>(I); 418 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 419 CanEvaluateTruncated(SI->getFalseValue(), Ty); 420 } 421 case Instruction::PHI: { 422 // We can change a phi if we can change all operands. Note that we never 423 // get into trouble with cyclic PHIs here because we only consider 424 // instructions with a single use. 425 PHINode *PN = cast<PHINode>(I); 426 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 427 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 428 return false; 429 return true; 430 } 431 default: 432 // TODO: Can handle more cases here. 433 break; 434 } 435 436 return false; 437} 438 439Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 440 if (Instruction *Result = commonCastTransforms(CI)) 441 return Result; 442 443 // See if we can simplify any instructions used by the input whose sole 444 // purpose is to compute bits we don't care about. 445 if (SimplifyDemandedInstructionBits(CI)) 446 return &CI; 447 448 Value *Src = CI.getOperand(0); 449 Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 450 451 // Attempt to truncate the entire input expression tree to the destination 452 // type. Only do this if the dest type is a simple type, don't convert the 453 // expression tree to something weird like i93 unless the source is also 454 // strange. 455 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 456 CanEvaluateTruncated(Src, DestTy)) { 457 458 // If this cast is a truncate, evaluting in a different type always 459 // eliminates the cast, so it is always a win. 460 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 461 " to avoid cast: " << CI << '\n'); 462 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 463 assert(Res->getType() == DestTy); 464 return ReplaceInstUsesWith(CI, Res); 465 } 466 467 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 468 if (DestTy->getScalarSizeInBits() == 1) { 469 Constant *One = ConstantInt::get(Src->getType(), 1); 470 Src = Builder->CreateAnd(Src, One); 471 Value *Zero = Constant::getNullValue(Src->getType()); 472 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 473 } 474 475 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 476 Value *A = nullptr; ConstantInt *Cst = nullptr; 477 if (Src->hasOneUse() && 478 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 479 // We have three types to worry about here, the type of A, the source of 480 // the truncate (MidSize), and the destination of the truncate. We know that 481 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 482 // between ASize and ResultSize. 483 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 484 485 // If the shift amount is larger than the size of A, then the result is 486 // known to be zero because all the input bits got shifted out. 487 if (Cst->getZExtValue() >= ASize) 488 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType())); 489 490 // Since we're doing an lshr and a zero extend, and know that the shift 491 // amount is smaller than ASize, it is always safe to do the shift in A's 492 // type, then zero extend or truncate to the result. 493 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 494 Shift->takeName(Src); 495 return CastInst::CreateIntegerCast(Shift, CI.getType(), false); 496 } 497 498 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 499 // type isn't non-native. 500 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) && 501 ShouldChangeType(Src->getType(), CI.getType()) && 502 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 503 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr"); 504 return BinaryOperator::CreateAnd(NewTrunc, 505 ConstantExpr::getTrunc(Cst, CI.getType())); 506 } 507 508 return nullptr; 509} 510 511/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 512/// in order to eliminate the icmp. 513Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 514 bool DoXform) { 515 // If we are just checking for a icmp eq of a single bit and zext'ing it 516 // to an integer, then shift the bit to the appropriate place and then 517 // cast to integer to avoid the comparison. 518 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 519 const APInt &Op1CV = Op1C->getValue(); 520 521 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 522 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 523 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 524 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 525 if (!DoXform) return ICI; 526 527 Value *In = ICI->getOperand(0); 528 Value *Sh = ConstantInt::get(In->getType(), 529 In->getType()->getScalarSizeInBits()-1); 530 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 531 if (In->getType() != CI.getType()) 532 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); 533 534 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 535 Constant *One = ConstantInt::get(In->getType(), 1); 536 In = Builder->CreateXor(In, One, In->getName()+".not"); 537 } 538 539 return ReplaceInstUsesWith(CI, In); 540 } 541 542 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 543 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 544 // zext (X == 1) to i32 --> X iff X has only the low bit set. 545 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 546 // zext (X != 0) to i32 --> X iff X has only the low bit set. 547 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 548 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 549 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 550 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 551 // This only works for EQ and NE 552 ICI->isEquality()) { 553 // If Op1C some other power of two, convert: 554 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 555 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 556 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne); 557 558 APInt KnownZeroMask(~KnownZero); 559 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 560 if (!DoXform) return ICI; 561 562 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 563 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 564 // (X&4) == 2 --> false 565 // (X&4) != 2 --> true 566 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 567 isNE); 568 Res = ConstantExpr::getZExt(Res, CI.getType()); 569 return ReplaceInstUsesWith(CI, Res); 570 } 571 572 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 573 Value *In = ICI->getOperand(0); 574 if (ShiftAmt) { 575 // Perform a logical shr by shiftamt. 576 // Insert the shift to put the result in the low bit. 577 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 578 In->getName()+".lobit"); 579 } 580 581 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 582 Constant *One = ConstantInt::get(In->getType(), 1); 583 In = Builder->CreateXor(In, One); 584 } 585 586 if (CI.getType() == In->getType()) 587 return ReplaceInstUsesWith(CI, In); 588 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 589 } 590 } 591 } 592 593 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 594 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 595 // may lead to additional simplifications. 596 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 597 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 598 uint32_t BitWidth = ITy->getBitWidth(); 599 Value *LHS = ICI->getOperand(0); 600 Value *RHS = ICI->getOperand(1); 601 602 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 603 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 604 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS); 605 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS); 606 607 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 608 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 609 APInt UnknownBit = ~KnownBits; 610 if (UnknownBit.countPopulation() == 1) { 611 if (!DoXform) return ICI; 612 613 Value *Result = Builder->CreateXor(LHS, RHS); 614 615 // Mask off any bits that are set and won't be shifted away. 616 if (KnownOneLHS.uge(UnknownBit)) 617 Result = Builder->CreateAnd(Result, 618 ConstantInt::get(ITy, UnknownBit)); 619 620 // Shift the bit we're testing down to the lsb. 621 Result = Builder->CreateLShr( 622 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 623 624 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 625 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 626 Result->takeName(ICI); 627 return ReplaceInstUsesWith(CI, Result); 628 } 629 } 630 } 631 } 632 633 return nullptr; 634} 635 636/// CanEvaluateZExtd - Determine if the specified value can be computed in the 637/// specified wider type and produce the same low bits. If not, return false. 638/// 639/// If this function returns true, it can also return a non-zero number of bits 640/// (in BitsToClear) which indicates that the value it computes is correct for 641/// the zero extend, but that the additional BitsToClear bits need to be zero'd 642/// out. For example, to promote something like: 643/// 644/// %B = trunc i64 %A to i32 645/// %C = lshr i32 %B, 8 646/// %E = zext i32 %C to i64 647/// 648/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 649/// set to 8 to indicate that the promoted value needs to have bits 24-31 650/// cleared in addition to bits 32-63. Since an 'and' will be generated to 651/// clear the top bits anyway, doing this has no extra cost. 652/// 653/// This function works on both vectors and scalars. 654static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { 655 BitsToClear = 0; 656 if (isa<Constant>(V)) 657 return true; 658 659 Instruction *I = dyn_cast<Instruction>(V); 660 if (!I) return false; 661 662 // If the input is a truncate from the destination type, we can trivially 663 // eliminate it. 664 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 665 return true; 666 667 // We can't extend or shrink something that has multiple uses: doing so would 668 // require duplicating the instruction in general, which isn't profitable. 669 if (!I->hasOneUse()) return false; 670 671 unsigned Opc = I->getOpcode(), Tmp; 672 switch (Opc) { 673 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 674 case Instruction::SExt: // zext(sext(x)) -> sext(x). 675 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 676 return true; 677 case Instruction::And: 678 case Instruction::Or: 679 case Instruction::Xor: 680 case Instruction::Add: 681 case Instruction::Sub: 682 case Instruction::Mul: 683 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 684 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 685 return false; 686 // These can all be promoted if neither operand has 'bits to clear'. 687 if (BitsToClear == 0 && Tmp == 0) 688 return true; 689 690 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 691 // other side, BitsToClear is ok. 692 if (Tmp == 0 && 693 (Opc == Instruction::And || Opc == Instruction::Or || 694 Opc == Instruction::Xor)) { 695 // We use MaskedValueIsZero here for generality, but the case we care 696 // about the most is constant RHS. 697 unsigned VSize = V->getType()->getScalarSizeInBits(); 698 if (MaskedValueIsZero(I->getOperand(1), 699 APInt::getHighBitsSet(VSize, BitsToClear))) 700 return true; 701 } 702 703 // Otherwise, we don't know how to analyze this BitsToClear case yet. 704 return false; 705 706 case Instruction::Shl: 707 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the 708 // upper bits we can reduce BitsToClear by the shift amount. 709 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 710 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 711 return false; 712 uint64_t ShiftAmt = Amt->getZExtValue(); 713 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0; 714 return true; 715 } 716 return false; 717 case Instruction::LShr: 718 // We can promote lshr(x, cst) if we can promote x. This requires the 719 // ultimate 'and' to clear out the high zero bits we're clearing out though. 720 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 721 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 722 return false; 723 BitsToClear += Amt->getZExtValue(); 724 if (BitsToClear > V->getType()->getScalarSizeInBits()) 725 BitsToClear = V->getType()->getScalarSizeInBits(); 726 return true; 727 } 728 // Cannot promote variable LSHR. 729 return false; 730 case Instruction::Select: 731 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 732 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 733 // TODO: If important, we could handle the case when the BitsToClear are 734 // known zero in the disagreeing side. 735 Tmp != BitsToClear) 736 return false; 737 return true; 738 739 case Instruction::PHI: { 740 // We can change a phi if we can change all operands. Note that we never 741 // get into trouble with cyclic PHIs here because we only consider 742 // instructions with a single use. 743 PHINode *PN = cast<PHINode>(I); 744 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 745 return false; 746 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 747 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 748 // TODO: If important, we could handle the case when the BitsToClear 749 // are known zero in the disagreeing input. 750 Tmp != BitsToClear) 751 return false; 752 return true; 753 } 754 default: 755 // TODO: Can handle more cases here. 756 return false; 757 } 758} 759 760Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 761 // If this zero extend is only used by a truncate, let the truncate be 762 // eliminated before we try to optimize this zext. 763 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back())) 764 return nullptr; 765 766 // If one of the common conversion will work, do it. 767 if (Instruction *Result = commonCastTransforms(CI)) 768 return Result; 769 770 // See if we can simplify any instructions used by the input whose sole 771 // purpose is to compute bits we don't care about. 772 if (SimplifyDemandedInstructionBits(CI)) 773 return &CI; 774 775 Value *Src = CI.getOperand(0); 776 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 777 778 // Attempt to extend the entire input expression tree to the destination 779 // type. Only do this if the dest type is a simple type, don't convert the 780 // expression tree to something weird like i93 unless the source is also 781 // strange. 782 unsigned BitsToClear; 783 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 784 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 785 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 786 "Unreasonable BitsToClear"); 787 788 // Okay, we can transform this! Insert the new expression now. 789 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 790 " to avoid zero extend: " << CI); 791 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 792 assert(Res->getType() == DestTy); 793 794 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 795 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 796 797 // If the high bits are already filled with zeros, just replace this 798 // cast with the result. 799 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 800 DestBitSize-SrcBitsKept))) 801 return ReplaceInstUsesWith(CI, Res); 802 803 // We need to emit an AND to clear the high bits. 804 Constant *C = ConstantInt::get(Res->getType(), 805 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 806 return BinaryOperator::CreateAnd(Res, C); 807 } 808 809 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 810 // types and if the sizes are just right we can convert this into a logical 811 // 'and' which will be much cheaper than the pair of casts. 812 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 813 // TODO: Subsume this into EvaluateInDifferentType. 814 815 // Get the sizes of the types involved. We know that the intermediate type 816 // will be smaller than A or C, but don't know the relation between A and C. 817 Value *A = CSrc->getOperand(0); 818 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 819 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 820 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 821 // If we're actually extending zero bits, then if 822 // SrcSize < DstSize: zext(a & mask) 823 // SrcSize == DstSize: a & mask 824 // SrcSize > DstSize: trunc(a) & mask 825 if (SrcSize < DstSize) { 826 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 827 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 828 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 829 return new ZExtInst(And, CI.getType()); 830 } 831 832 if (SrcSize == DstSize) { 833 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 834 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 835 AndValue)); 836 } 837 if (SrcSize > DstSize) { 838 Value *Trunc = Builder->CreateTrunc(A, CI.getType()); 839 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 840 return BinaryOperator::CreateAnd(Trunc, 841 ConstantInt::get(Trunc->getType(), 842 AndValue)); 843 } 844 } 845 846 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 847 return transformZExtICmp(ICI, CI); 848 849 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 850 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 851 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 852 // of the (zext icmp) will be transformed. 853 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 854 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 855 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 856 (transformZExtICmp(LHS, CI, false) || 857 transformZExtICmp(RHS, CI, false))) { 858 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 859 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 860 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 861 } 862 } 863 864 // zext(trunc(X) & C) -> (X & zext(C)). 865 Constant *C; 866 Value *X; 867 if (SrcI && 868 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) && 869 X->getType() == CI.getType()) 870 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType())); 871 872 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)). 873 Value *And; 874 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) && 875 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) && 876 X->getType() == CI.getType()) { 877 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 878 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC); 879 } 880 881 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 882 if (SrcI && SrcI->hasOneUse() && 883 SrcI->getType()->getScalarType()->isIntegerTy(1) && 884 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) { 885 Value *New = Builder->CreateZExt(X, CI.getType()); 886 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 887 } 888 889 return nullptr; 890} 891 892/// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations 893/// in order to eliminate the icmp. 894Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { 895 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); 896 ICmpInst::Predicate Pred = ICI->getPredicate(); 897 898 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 899 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative 900 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive 901 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) || 902 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { 903 904 Value *Sh = ConstantInt::get(Op0->getType(), 905 Op0->getType()->getScalarSizeInBits()-1); 906 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); 907 if (In->getType() != CI.getType()) 908 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); 909 910 if (Pred == ICmpInst::ICMP_SGT) 911 In = Builder->CreateNot(In, In->getName()+".not"); 912 return ReplaceInstUsesWith(CI, In); 913 } 914 } 915 916 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 917 // If we know that only one bit of the LHS of the icmp can be set and we 918 // have an equality comparison with zero or a power of 2, we can transform 919 // the icmp and sext into bitwise/integer operations. 920 if (ICI->hasOneUse() && 921 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ 922 unsigned BitWidth = Op1C->getType()->getBitWidth(); 923 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 924 computeKnownBits(Op0, KnownZero, KnownOne); 925 926 APInt KnownZeroMask(~KnownZero); 927 if (KnownZeroMask.isPowerOf2()) { 928 Value *In = ICI->getOperand(0); 929 930 // If the icmp tests for a known zero bit we can constant fold it. 931 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) { 932 Value *V = Pred == ICmpInst::ICMP_NE ? 933 ConstantInt::getAllOnesValue(CI.getType()) : 934 ConstantInt::getNullValue(CI.getType()); 935 return ReplaceInstUsesWith(CI, V); 936 } 937 938 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) { 939 // sext ((x & 2^n) == 0) -> (x >> n) - 1 940 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1 941 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros(); 942 // Perform a right shift to place the desired bit in the LSB. 943 if (ShiftAmt) 944 In = Builder->CreateLShr(In, 945 ConstantInt::get(In->getType(), ShiftAmt)); 946 947 // At this point "In" is either 1 or 0. Subtract 1 to turn 948 // {1, 0} -> {0, -1}. 949 In = Builder->CreateAdd(In, 950 ConstantInt::getAllOnesValue(In->getType()), 951 "sext"); 952 } else { 953 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1 954 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1 955 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros(); 956 // Perform a left shift to place the desired bit in the MSB. 957 if (ShiftAmt) 958 In = Builder->CreateShl(In, 959 ConstantInt::get(In->getType(), ShiftAmt)); 960 961 // Distribute the bit over the whole bit width. 962 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(), 963 BitWidth - 1), "sext"); 964 } 965 966 if (CI.getType() == In->getType()) 967 return ReplaceInstUsesWith(CI, In); 968 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/); 969 } 970 } 971 } 972 973 return nullptr; 974} 975 976/// CanEvaluateSExtd - Return true if we can take the specified value 977/// and return it as type Ty without inserting any new casts and without 978/// changing the value of the common low bits. This is used by code that tries 979/// to promote integer operations to a wider types will allow us to eliminate 980/// the extension. 981/// 982/// This function works on both vectors and scalars. 983/// 984static bool CanEvaluateSExtd(Value *V, Type *Ty) { 985 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 986 "Can't sign extend type to a smaller type"); 987 // If this is a constant, it can be trivially promoted. 988 if (isa<Constant>(V)) 989 return true; 990 991 Instruction *I = dyn_cast<Instruction>(V); 992 if (!I) return false; 993 994 // If this is a truncate from the dest type, we can trivially eliminate it. 995 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 996 return true; 997 998 // We can't extend or shrink something that has multiple uses: doing so would 999 // require duplicating the instruction in general, which isn't profitable. 1000 if (!I->hasOneUse()) return false; 1001 1002 switch (I->getOpcode()) { 1003 case Instruction::SExt: // sext(sext(x)) -> sext(x) 1004 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 1005 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 1006 return true; 1007 case Instruction::And: 1008 case Instruction::Or: 1009 case Instruction::Xor: 1010 case Instruction::Add: 1011 case Instruction::Sub: 1012 case Instruction::Mul: 1013 // These operators can all arbitrarily be extended if their inputs can. 1014 return CanEvaluateSExtd(I->getOperand(0), Ty) && 1015 CanEvaluateSExtd(I->getOperand(1), Ty); 1016 1017 //case Instruction::Shl: TODO 1018 //case Instruction::LShr: TODO 1019 1020 case Instruction::Select: 1021 return CanEvaluateSExtd(I->getOperand(1), Ty) && 1022 CanEvaluateSExtd(I->getOperand(2), Ty); 1023 1024 case Instruction::PHI: { 1025 // We can change a phi if we can change all operands. Note that we never 1026 // get into trouble with cyclic PHIs here because we only consider 1027 // instructions with a single use. 1028 PHINode *PN = cast<PHINode>(I); 1029 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1030 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 1031 return true; 1032 } 1033 default: 1034 // TODO: Can handle more cases here. 1035 break; 1036 } 1037 1038 return false; 1039} 1040 1041Instruction *InstCombiner::visitSExt(SExtInst &CI) { 1042 // If this sign extend is only used by a truncate, let the truncate be 1043 // eliminated before we try to optimize this sext. 1044 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back())) 1045 return nullptr; 1046 1047 if (Instruction *I = commonCastTransforms(CI)) 1048 return I; 1049 1050 // See if we can simplify any instructions used by the input whose sole 1051 // purpose is to compute bits we don't care about. 1052 if (SimplifyDemandedInstructionBits(CI)) 1053 return &CI; 1054 1055 Value *Src = CI.getOperand(0); 1056 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 1057 1058 // Attempt to extend the entire input expression tree to the destination 1059 // type. Only do this if the dest type is a simple type, don't convert the 1060 // expression tree to something weird like i93 unless the source is also 1061 // strange. 1062 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 1063 CanEvaluateSExtd(Src, DestTy)) { 1064 // Okay, we can transform this! Insert the new expression now. 1065 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 1066 " to avoid sign extend: " << CI); 1067 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 1068 assert(Res->getType() == DestTy); 1069 1070 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1071 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1072 1073 // If the high bits are already filled with sign bit, just replace this 1074 // cast with the result. 1075 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 1076 return ReplaceInstUsesWith(CI, Res); 1077 1078 // We need to emit a shl + ashr to do the sign extend. 1079 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1080 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 1081 ShAmt); 1082 } 1083 1084 // If this input is a trunc from our destination, then turn sext(trunc(x)) 1085 // into shifts. 1086 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 1087 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 1088 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1089 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1090 1091 // We need to emit a shl + ashr to do the sign extend. 1092 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1093 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1094 return BinaryOperator::CreateAShr(Res, ShAmt); 1095 } 1096 1097 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 1098 return transformSExtICmp(ICI, CI); 1099 1100 // If the input is a shl/ashr pair of a same constant, then this is a sign 1101 // extension from a smaller value. If we could trust arbitrary bitwidth 1102 // integers, we could turn this into a truncate to the smaller bit and then 1103 // use a sext for the whole extension. Since we don't, look deeper and check 1104 // for a truncate. If the source and dest are the same type, eliminate the 1105 // trunc and extend and just do shifts. For example, turn: 1106 // %a = trunc i32 %i to i8 1107 // %b = shl i8 %a, 6 1108 // %c = ashr i8 %b, 6 1109 // %d = sext i8 %c to i32 1110 // into: 1111 // %a = shl i32 %i, 30 1112 // %d = ashr i32 %a, 30 1113 Value *A = nullptr; 1114 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1115 ConstantInt *BA = nullptr, *CA = nullptr; 1116 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1117 m_ConstantInt(CA))) && 1118 BA == CA && A->getType() == CI.getType()) { 1119 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1120 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1121 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1122 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1123 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1124 return BinaryOperator::CreateAShr(A, ShAmtV); 1125 } 1126 1127 return nullptr; 1128} 1129 1130 1131/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1132/// in the specified FP type without changing its value. 1133static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1134 bool losesInfo; 1135 APFloat F = CFP->getValueAPF(); 1136 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1137 if (!losesInfo) 1138 return ConstantFP::get(CFP->getContext(), F); 1139 return nullptr; 1140} 1141 1142/// LookThroughFPExtensions - If this is an fp extension instruction, look 1143/// through it until we get the source value. 1144static Value *LookThroughFPExtensions(Value *V) { 1145 if (Instruction *I = dyn_cast<Instruction>(V)) 1146 if (I->getOpcode() == Instruction::FPExt) 1147 return LookThroughFPExtensions(I->getOperand(0)); 1148 1149 // If this value is a constant, return the constant in the smallest FP type 1150 // that can accurately represent it. This allows us to turn 1151 // (float)((double)X+2.0) into x+2.0f. 1152 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1153 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1154 return V; // No constant folding of this. 1155 // See if the value can be truncated to half and then reextended. 1156 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf)) 1157 return V; 1158 // See if the value can be truncated to float and then reextended. 1159 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1160 return V; 1161 if (CFP->getType()->isDoubleTy()) 1162 return V; // Won't shrink. 1163 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1164 return V; 1165 // Don't try to shrink to various long double types. 1166 } 1167 1168 return V; 1169} 1170 1171Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1172 if (Instruction *I = commonCastTransforms(CI)) 1173 return I; 1174 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to 1175 // simpilify this expression to avoid one or more of the trunc/extend 1176 // operations if we can do so without changing the numerical results. 1177 // 1178 // The exact manner in which the widths of the operands interact to limit 1179 // what we can and cannot do safely varies from operation to operation, and 1180 // is explained below in the various case statements. 1181 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1182 if (OpI && OpI->hasOneUse()) { 1183 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0)); 1184 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1)); 1185 unsigned OpWidth = OpI->getType()->getFPMantissaWidth(); 1186 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth(); 1187 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth(); 1188 unsigned SrcWidth = std::max(LHSWidth, RHSWidth); 1189 unsigned DstWidth = CI.getType()->getFPMantissaWidth(); 1190 switch (OpI->getOpcode()) { 1191 default: break; 1192 case Instruction::FAdd: 1193 case Instruction::FSub: 1194 // For addition and subtraction, the infinitely precise result can 1195 // essentially be arbitrarily wide; proving that double rounding 1196 // will not occur because the result of OpI is exact (as we will for 1197 // FMul, for example) is hopeless. However, we *can* nonetheless 1198 // frequently know that double rounding cannot occur (or that it is 1199 // innocuous) by taking advantage of the specific structure of 1200 // infinitely-precise results that admit double rounding. 1201 // 1202 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient 1203 // to represent both sources, we can guarantee that the double 1204 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis, 1205 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..." 1206 // for proof of this fact). 1207 // 1208 // Note: Figueroa does not consider the case where DstFormat != 1209 // SrcFormat. It's possible (likely even!) that this analysis 1210 // could be tightened for those cases, but they are rare (the main 1211 // case of interest here is (float)((double)float + float)). 1212 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) { 1213 if (LHSOrig->getType() != CI.getType()) 1214 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1215 if (RHSOrig->getType() != CI.getType()) 1216 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1217 Instruction *RI = 1218 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig); 1219 RI->copyFastMathFlags(OpI); 1220 return RI; 1221 } 1222 break; 1223 case Instruction::FMul: 1224 // For multiplication, the infinitely precise result has at most 1225 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient 1226 // that such a value can be exactly represented, then no double 1227 // rounding can possibly occur; we can safely perform the operation 1228 // in the destination format if it can represent both sources. 1229 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) { 1230 if (LHSOrig->getType() != CI.getType()) 1231 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1232 if (RHSOrig->getType() != CI.getType()) 1233 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1234 Instruction *RI = 1235 BinaryOperator::CreateFMul(LHSOrig, RHSOrig); 1236 RI->copyFastMathFlags(OpI); 1237 return RI; 1238 } 1239 break; 1240 case Instruction::FDiv: 1241 // For division, we use again use the bound from Figueroa's 1242 // dissertation. I am entirely certain that this bound can be 1243 // tightened in the unbalanced operand case by an analysis based on 1244 // the diophantine rational approximation bound, but the well-known 1245 // condition used here is a good conservative first pass. 1246 // TODO: Tighten bound via rigorous analysis of the unbalanced case. 1247 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) { 1248 if (LHSOrig->getType() != CI.getType()) 1249 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType()); 1250 if (RHSOrig->getType() != CI.getType()) 1251 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType()); 1252 Instruction *RI = 1253 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig); 1254 RI->copyFastMathFlags(OpI); 1255 return RI; 1256 } 1257 break; 1258 case Instruction::FRem: 1259 // Remainder is straightforward. Remainder is always exact, so the 1260 // type of OpI doesn't enter into things at all. We simply evaluate 1261 // in whichever source type is larger, then convert to the 1262 // destination type. 1263 if (LHSWidth < SrcWidth) 1264 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType()); 1265 else if (RHSWidth <= SrcWidth) 1266 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType()); 1267 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig); 1268 if (Instruction *RI = dyn_cast<Instruction>(ExactResult)) 1269 RI->copyFastMathFlags(OpI); 1270 return CastInst::CreateFPCast(ExactResult, CI.getType()); 1271 } 1272 1273 // (fptrunc (fneg x)) -> (fneg (fptrunc x)) 1274 if (BinaryOperator::isFNeg(OpI)) { 1275 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1), 1276 CI.getType()); 1277 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc); 1278 RI->copyFastMathFlags(OpI); 1279 return RI; 1280 } 1281 } 1282 1283 // (fptrunc (select cond, R1, Cst)) --> 1284 // (select cond, (fptrunc R1), (fptrunc Cst)) 1285 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)); 1286 if (SI && 1287 (isa<ConstantFP>(SI->getOperand(1)) || 1288 isa<ConstantFP>(SI->getOperand(2)))) { 1289 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1), 1290 CI.getType()); 1291 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2), 1292 CI.getType()); 1293 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc); 1294 } 1295 1296 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0)); 1297 if (II) { 1298 switch (II->getIntrinsicID()) { 1299 default: break; 1300 case Intrinsic::fabs: { 1301 // (fptrunc (fabs x)) -> (fabs (fptrunc x)) 1302 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0), 1303 CI.getType()); 1304 Type *IntrinsicType[] = { CI.getType() }; 1305 Function *Overload = 1306 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(), 1307 II->getIntrinsicID(), IntrinsicType); 1308 1309 Value *Args[] = { InnerTrunc }; 1310 return CallInst::Create(Overload, Args, II->getName()); 1311 } 1312 } 1313 } 1314 1315 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1316 // Note that we restrict this transformation based on 1317 // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because 1318 // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the 1319 // single-precision intrinsic can be expanded in the backend. 1320 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1321 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) && 1322 (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) || 1323 Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) && 1324 Call->getNumArgOperands() == 1 && 1325 Call->hasOneUse()) { 1326 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1327 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1328 CI.getType()->isFloatTy() && 1329 Call->getType()->isDoubleTy() && 1330 Arg->getType()->isDoubleTy() && 1331 Arg->getOperand(0)->getType()->isFloatTy()) { 1332 Function *Callee = Call->getCalledFunction(); 1333 Module *M = CI.getParent()->getParent()->getParent(); 1334 Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ? 1335 Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) : 1336 M->getOrInsertFunction("sqrtf", Callee->getAttributes(), 1337 Builder->getFloatTy(), Builder->getFloatTy(), 1338 NULL); 1339 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1340 "sqrtfcall"); 1341 ret->setAttributes(Callee->getAttributes()); 1342 1343 1344 // Remove the old Call. With -fmath-errno, it won't get marked readnone. 1345 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType())); 1346 EraseInstFromFunction(*Call); 1347 return ret; 1348 } 1349 } 1350 1351 return nullptr; 1352} 1353 1354Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1355 return commonCastTransforms(CI); 1356} 1357 1358Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1359 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1360 if (!OpI) 1361 return commonCastTransforms(FI); 1362 1363 // fptoui(uitofp(X)) --> X 1364 // fptoui(sitofp(X)) --> X 1365 // This is safe if the intermediate type has enough bits in its mantissa to 1366 // accurately represent all values of X. For example, do not do this with 1367 // i64->float->i64. This is also safe for sitofp case, because any negative 1368 // 'X' value would cause an undefined result for the fptoui. 1369 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1370 OpI->getOperand(0)->getType() == FI.getType() && 1371 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1372 OpI->getType()->getFPMantissaWidth()) 1373 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1374 1375 return commonCastTransforms(FI); 1376} 1377 1378Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1379 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1380 if (!OpI) 1381 return commonCastTransforms(FI); 1382 1383 // fptosi(sitofp(X)) --> X 1384 // fptosi(uitofp(X)) --> X 1385 // This is safe if the intermediate type has enough bits in its mantissa to 1386 // accurately represent all values of X. For example, do not do this with 1387 // i64->float->i64. This is also safe for sitofp case, because any negative 1388 // 'X' value would cause an undefined result for the fptoui. 1389 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1390 OpI->getOperand(0)->getType() == FI.getType() && 1391 (int)FI.getType()->getScalarSizeInBits() <= 1392 OpI->getType()->getFPMantissaWidth()) 1393 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1394 1395 return commonCastTransforms(FI); 1396} 1397 1398Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1399 return commonCastTransforms(CI); 1400} 1401 1402Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1403 return commonCastTransforms(CI); 1404} 1405 1406Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1407 // If the source integer type is not the intptr_t type for this target, do a 1408 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1409 // cast to be exposed to other transforms. 1410 1411 if (DL) { 1412 unsigned AS = CI.getAddressSpace(); 1413 if (CI.getOperand(0)->getType()->getScalarSizeInBits() != 1414 DL->getPointerSizeInBits(AS)) { 1415 Type *Ty = DL->getIntPtrType(CI.getContext(), AS); 1416 if (CI.getType()->isVectorTy()) // Handle vectors of pointers. 1417 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements()); 1418 1419 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty); 1420 return new IntToPtrInst(P, CI.getType()); 1421 } 1422 } 1423 1424 if (Instruction *I = commonCastTransforms(CI)) 1425 return I; 1426 1427 return nullptr; 1428} 1429 1430/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1431Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1432 Value *Src = CI.getOperand(0); 1433 1434 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1435 // If casting the result of a getelementptr instruction with no offset, turn 1436 // this into a cast of the original pointer! 1437 if (GEP->hasAllZeroIndices()) { 1438 // Changing the cast operand is usually not a good idea but it is safe 1439 // here because the pointer operand is being replaced with another 1440 // pointer operand so the opcode doesn't need to change. 1441 Worklist.Add(GEP); 1442 CI.setOperand(0, GEP->getOperand(0)); 1443 return &CI; 1444 } 1445 1446 if (!DL) 1447 return commonCastTransforms(CI); 1448 1449 // If the GEP has a single use, and the base pointer is a bitcast, and the 1450 // GEP computes a constant offset, see if we can convert these three 1451 // instructions into fewer. This typically happens with unions and other 1452 // non-type-safe code. 1453 unsigned AS = GEP->getPointerAddressSpace(); 1454 unsigned OffsetBits = DL->getPointerSizeInBits(AS); 1455 APInt Offset(OffsetBits, 0); 1456 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0)); 1457 if (GEP->hasOneUse() && 1458 BCI && 1459 GEP->accumulateConstantOffset(*DL, Offset)) { 1460 // Get the base pointer input of the bitcast, and the type it points to. 1461 Value *OrigBase = BCI->getOperand(0); 1462 SmallVector<Value*, 8> NewIndices; 1463 if (FindElementAtOffset(OrigBase->getType(), 1464 Offset.getSExtValue(), 1465 NewIndices)) { 1466 // If we were able to index down into an element, create the GEP 1467 // and bitcast the result. This eliminates one bitcast, potentially 1468 // two. 1469 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1470 Builder->CreateInBoundsGEP(OrigBase, NewIndices) : 1471 Builder->CreateGEP(OrigBase, NewIndices); 1472 NGEP->takeName(GEP); 1473 1474 if (isa<BitCastInst>(CI)) 1475 return new BitCastInst(NGEP, CI.getType()); 1476 assert(isa<PtrToIntInst>(CI)); 1477 return new PtrToIntInst(NGEP, CI.getType()); 1478 } 1479 } 1480 } 1481 1482 return commonCastTransforms(CI); 1483} 1484 1485Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1486 // If the destination integer type is not the intptr_t type for this target, 1487 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1488 // to be exposed to other transforms. 1489 1490 if (!DL) 1491 return commonPointerCastTransforms(CI); 1492 1493 Type *Ty = CI.getType(); 1494 unsigned AS = CI.getPointerAddressSpace(); 1495 1496 if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS)) 1497 return commonPointerCastTransforms(CI); 1498 1499 Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS); 1500 if (Ty->isVectorTy()) // Handle vectors of pointers. 1501 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements()); 1502 1503 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy); 1504 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false); 1505} 1506 1507/// OptimizeVectorResize - This input value (which is known to have vector type) 1508/// is being zero extended or truncated to the specified vector type. Try to 1509/// replace it with a shuffle (and vector/vector bitcast) if possible. 1510/// 1511/// The source and destination vector types may have different element types. 1512static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, 1513 InstCombiner &IC) { 1514 // We can only do this optimization if the output is a multiple of the input 1515 // element size, or the input is a multiple of the output element size. 1516 // Convert the input type to have the same element type as the output. 1517 VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1518 1519 if (SrcTy->getElementType() != DestTy->getElementType()) { 1520 // The input types don't need to be identical, but for now they must be the 1521 // same size. There is no specific reason we couldn't handle things like 1522 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1523 // there yet. 1524 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1525 DestTy->getElementType()->getPrimitiveSizeInBits()) 1526 return nullptr; 1527 1528 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1529 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1530 } 1531 1532 // Now that the element types match, get the shuffle mask and RHS of the 1533 // shuffle to use, which depends on whether we're increasing or decreasing the 1534 // size of the input. 1535 SmallVector<uint32_t, 16> ShuffleMask; 1536 Value *V2; 1537 1538 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1539 // If we're shrinking the number of elements, just shuffle in the low 1540 // elements from the input and use undef as the second shuffle input. 1541 V2 = UndefValue::get(SrcTy); 1542 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1543 ShuffleMask.push_back(i); 1544 1545 } else { 1546 // If we're increasing the number of elements, shuffle in all of the 1547 // elements from InVal and fill the rest of the result elements with zeros 1548 // from a constant zero. 1549 V2 = Constant::getNullValue(SrcTy); 1550 unsigned SrcElts = SrcTy->getNumElements(); 1551 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1552 ShuffleMask.push_back(i); 1553 1554 // The excess elements reference the first element of the zero input. 1555 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i) 1556 ShuffleMask.push_back(SrcElts); 1557 } 1558 1559 return new ShuffleVectorInst(InVal, V2, 1560 ConstantDataVector::get(V2->getContext(), 1561 ShuffleMask)); 1562} 1563 1564static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { 1565 return Value % Ty->getPrimitiveSizeInBits() == 0; 1566} 1567 1568static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { 1569 return Value / Ty->getPrimitiveSizeInBits(); 1570} 1571 1572/// CollectInsertionElements - V is a value which is inserted into a vector of 1573/// VecEltTy. Look through the value to see if we can decompose it into 1574/// insertions into the vector. See the example in the comment for 1575/// OptimizeIntegerToVectorInsertions for the pattern this handles. 1576/// The type of V is always a non-zero multiple of VecEltTy's size. 1577/// Shift is the number of bits between the lsb of V and the lsb of 1578/// the vector. 1579/// 1580/// This returns false if the pattern can't be matched or true if it can, 1581/// filling in Elements with the elements found here. 1582static bool CollectInsertionElements(Value *V, unsigned Shift, 1583 SmallVectorImpl<Value*> &Elements, 1584 Type *VecEltTy, InstCombiner &IC) { 1585 assert(isMultipleOfTypeSize(Shift, VecEltTy) && 1586 "Shift should be a multiple of the element type size"); 1587 1588 // Undef values never contribute useful bits to the result. 1589 if (isa<UndefValue>(V)) return true; 1590 1591 // If we got down to a value of the right type, we win, try inserting into the 1592 // right element. 1593 if (V->getType() == VecEltTy) { 1594 // Inserting null doesn't actually insert any elements. 1595 if (Constant *C = dyn_cast<Constant>(V)) 1596 if (C->isNullValue()) 1597 return true; 1598 1599 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy); 1600 if (IC.getDataLayout()->isBigEndian()) 1601 ElementIndex = Elements.size() - ElementIndex - 1; 1602 1603 // Fail if multiple elements are inserted into this slot. 1604 if (Elements[ElementIndex]) 1605 return false; 1606 1607 Elements[ElementIndex] = V; 1608 return true; 1609 } 1610 1611 if (Constant *C = dyn_cast<Constant>(V)) { 1612 // Figure out the # elements this provides, and bitcast it or slice it up 1613 // as required. 1614 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1615 VecEltTy); 1616 // If the constant is the size of a vector element, we just need to bitcast 1617 // it to the right type so it gets properly inserted. 1618 if (NumElts == 1) 1619 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1620 Shift, Elements, VecEltTy, IC); 1621 1622 // Okay, this is a constant that covers multiple elements. Slice it up into 1623 // pieces and insert each element-sized piece into the vector. 1624 if (!isa<IntegerType>(C->getType())) 1625 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1626 C->getType()->getPrimitiveSizeInBits())); 1627 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1628 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1629 1630 for (unsigned i = 0; i != NumElts; ++i) { 1631 unsigned ShiftI = Shift+i*ElementSize; 1632 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1633 ShiftI)); 1634 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1635 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC)) 1636 return false; 1637 } 1638 return true; 1639 } 1640 1641 if (!V->hasOneUse()) return false; 1642 1643 Instruction *I = dyn_cast<Instruction>(V); 1644 if (!I) return false; 1645 switch (I->getOpcode()) { 1646 default: return false; // Unhandled case. 1647 case Instruction::BitCast: 1648 return CollectInsertionElements(I->getOperand(0), Shift, 1649 Elements, VecEltTy, IC); 1650 case Instruction::ZExt: 1651 if (!isMultipleOfTypeSize( 1652 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1653 VecEltTy)) 1654 return false; 1655 return CollectInsertionElements(I->getOperand(0), Shift, 1656 Elements, VecEltTy, IC); 1657 case Instruction::Or: 1658 return CollectInsertionElements(I->getOperand(0), Shift, 1659 Elements, VecEltTy, IC) && 1660 CollectInsertionElements(I->getOperand(1), Shift, 1661 Elements, VecEltTy, IC); 1662 case Instruction::Shl: { 1663 // Must be shifting by a constant that is a multiple of the element size. 1664 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1665 if (!CI) return false; 1666 Shift += CI->getZExtValue(); 1667 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false; 1668 return CollectInsertionElements(I->getOperand(0), Shift, 1669 Elements, VecEltTy, IC); 1670 } 1671 1672 } 1673} 1674 1675 1676/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1677/// may be doing shifts and ors to assemble the elements of the vector manually. 1678/// Try to rip the code out and replace it with insertelements. This is to 1679/// optimize code like this: 1680/// 1681/// %tmp37 = bitcast float %inc to i32 1682/// %tmp38 = zext i32 %tmp37 to i64 1683/// %tmp31 = bitcast float %inc5 to i32 1684/// %tmp32 = zext i32 %tmp31 to i64 1685/// %tmp33 = shl i64 %tmp32, 32 1686/// %ins35 = or i64 %tmp33, %tmp38 1687/// %tmp43 = bitcast i64 %ins35 to <2 x float> 1688/// 1689/// Into two insertelements that do "buildvector{%inc, %inc5}". 1690static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1691 InstCombiner &IC) { 1692 // We need to know the target byte order to perform this optimization. 1693 if (!IC.getDataLayout()) return nullptr; 1694 1695 VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1696 Value *IntInput = CI.getOperand(0); 1697 1698 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1699 if (!CollectInsertionElements(IntInput, 0, Elements, 1700 DestVecTy->getElementType(), IC)) 1701 return nullptr; 1702 1703 // If we succeeded, we know that all of the element are specified by Elements 1704 // or are zero if Elements has a null entry. Recast this as a set of 1705 // insertions. 1706 Value *Result = Constant::getNullValue(CI.getType()); 1707 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1708 if (!Elements[i]) continue; // Unset element. 1709 1710 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1711 IC.Builder->getInt32(i)); 1712 } 1713 1714 return Result; 1715} 1716 1717 1718/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1719/// bitcast. The various long double bitcasts can't get in here. 1720static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1721 // We need to know the target byte order to perform this optimization. 1722 if (!IC.getDataLayout()) return nullptr; 1723 1724 Value *Src = CI.getOperand(0); 1725 Type *DestTy = CI.getType(); 1726 1727 // If this is a bitcast from int to float, check to see if the int is an 1728 // extraction from a vector. 1729 Value *VecInput = nullptr; 1730 // bitcast(trunc(bitcast(somevector))) 1731 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1732 isa<VectorType>(VecInput->getType())) { 1733 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1734 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1735 1736 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1737 // If the element type of the vector doesn't match the result type, 1738 // bitcast it to be a vector type we can extract from. 1739 if (VecTy->getElementType() != DestTy) { 1740 VecTy = VectorType::get(DestTy, 1741 VecTy->getPrimitiveSizeInBits() / DestWidth); 1742 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1743 } 1744 1745 unsigned Elt = 0; 1746 if (IC.getDataLayout()->isBigEndian()) 1747 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1; 1748 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1749 } 1750 } 1751 1752 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1753 ConstantInt *ShAmt = nullptr; 1754 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1755 m_ConstantInt(ShAmt)))) && 1756 isa<VectorType>(VecInput->getType())) { 1757 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1758 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1759 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1760 ShAmt->getZExtValue() % DestWidth == 0) { 1761 // If the element type of the vector doesn't match the result type, 1762 // bitcast it to be a vector type we can extract from. 1763 if (VecTy->getElementType() != DestTy) { 1764 VecTy = VectorType::get(DestTy, 1765 VecTy->getPrimitiveSizeInBits() / DestWidth); 1766 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1767 } 1768 1769 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1770 if (IC.getDataLayout()->isBigEndian()) 1771 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt; 1772 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1773 } 1774 } 1775 return nullptr; 1776} 1777 1778Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1779 // If the operands are integer typed then apply the integer transforms, 1780 // otherwise just apply the common ones. 1781 Value *Src = CI.getOperand(0); 1782 Type *SrcTy = Src->getType(); 1783 Type *DestTy = CI.getType(); 1784 1785 // Get rid of casts from one type to the same type. These are useless and can 1786 // be replaced by the operand. 1787 if (DestTy == Src->getType()) 1788 return ReplaceInstUsesWith(CI, Src); 1789 1790 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1791 PointerType *SrcPTy = cast<PointerType>(SrcTy); 1792 Type *DstElTy = DstPTy->getElementType(); 1793 Type *SrcElTy = SrcPTy->getElementType(); 1794 1795 // If we are casting a alloca to a pointer to a type of the same 1796 // size, rewrite the allocation instruction to allocate the "right" type. 1797 // There is no need to modify malloc calls because it is their bitcast that 1798 // needs to be cleaned up. 1799 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1800 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1801 return V; 1802 1803 // If the source and destination are pointers, and this cast is equivalent 1804 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1805 // This can enhance SROA and other transforms that want type-safe pointers. 1806 Constant *ZeroUInt = 1807 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1808 unsigned NumZeros = 0; 1809 while (SrcElTy != DstElTy && 1810 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1811 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1812 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1813 ++NumZeros; 1814 } 1815 1816 // If we found a path from the src to dest, create the getelementptr now. 1817 if (SrcElTy == DstElTy) { 1818 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1819 return GetElementPtrInst::CreateInBounds(Src, Idxs); 1820 } 1821 } 1822 1823 // Try to optimize int -> float bitcasts. 1824 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1825 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1826 return I; 1827 1828 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1829 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1830 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1831 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1832 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1833 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1834 } 1835 1836 if (isa<IntegerType>(SrcTy)) { 1837 // If this is a cast from an integer to vector, check to see if the input 1838 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1839 // the casts with a shuffle and (potentially) a bitcast. 1840 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1841 CastInst *SrcCast = cast<CastInst>(Src); 1842 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1843 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1844 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1845 cast<VectorType>(DestTy), *this)) 1846 return I; 1847 } 1848 1849 // If the input is an 'or' instruction, we may be doing shifts and ors to 1850 // assemble the elements of the vector manually. Try to rip the code out 1851 // and replace it with insertelements. 1852 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1853 return ReplaceInstUsesWith(CI, V); 1854 } 1855 } 1856 1857 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1858 if (SrcVTy->getNumElements() == 1) { 1859 // If our destination is not a vector, then make this a straight 1860 // scalar-scalar cast. 1861 if (!DestTy->isVectorTy()) { 1862 Value *Elem = 1863 Builder->CreateExtractElement(Src, 1864 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1865 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1866 } 1867 1868 // Otherwise, see if our source is an insert. If so, then use the scalar 1869 // component directly. 1870 if (InsertElementInst *IEI = 1871 dyn_cast<InsertElementInst>(CI.getOperand(0))) 1872 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1), 1873 DestTy); 1874 } 1875 } 1876 1877 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1878 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1879 // a bitcast to a vector with the same # elts. 1880 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1881 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() && 1882 SVI->getType()->getNumElements() == 1883 SVI->getOperand(0)->getType()->getVectorNumElements()) { 1884 BitCastInst *Tmp; 1885 // If either of the operands is a cast from CI.getType(), then 1886 // evaluating the shuffle in the casted destination's type will allow 1887 // us to eliminate at least one cast. 1888 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1889 Tmp->getOperand(0)->getType() == DestTy) || 1890 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1891 Tmp->getOperand(0)->getType() == DestTy)) { 1892 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1893 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1894 // Return a new shuffle vector. Use the same element ID's, as we 1895 // know the vector types match #elts. 1896 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1897 } 1898 } 1899 } 1900 1901 if (SrcTy->isPointerTy()) 1902 return commonPointerCastTransforms(CI); 1903 return commonCastTransforms(CI); 1904} 1905 1906Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) { 1907 return commonPointerCastTransforms(CI); 1908} 1909