InstCombineCasts.cpp revision ce16339930a2b03e53b4e6399ef59c092a7f2cfa
1//===- InstCombineCasts.cpp -----------------------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the visit functions for cast operations. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Analysis/ConstantFolding.h" 16#include "llvm/Target/TargetData.h" 17#include "llvm/Target/TargetLibraryInfo.h" 18#include "llvm/Support/PatternMatch.h" 19using namespace llvm; 20using namespace PatternMatch; 21 22/// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear 23/// expression. If so, decompose it, returning some value X, such that Val is 24/// X*Scale+Offset. 25/// 26static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale, 27 uint64_t &Offset) { 28 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 29 Offset = CI->getZExtValue(); 30 Scale = 0; 31 return ConstantInt::get(Val->getType(), 0); 32 } 33 34 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) { 35 // Cannot look past anything that might overflow. 36 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val); 37 if (OBI && !OBI->hasNoUnsignedWrap()) { 38 Scale = 1; 39 Offset = 0; 40 return Val; 41 } 42 43 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) { 44 if (I->getOpcode() == Instruction::Shl) { 45 // This is a value scaled by '1 << the shift amt'. 46 Scale = UINT64_C(1) << RHS->getZExtValue(); 47 Offset = 0; 48 return I->getOperand(0); 49 } 50 51 if (I->getOpcode() == Instruction::Mul) { 52 // This value is scaled by 'RHS'. 53 Scale = RHS->getZExtValue(); 54 Offset = 0; 55 return I->getOperand(0); 56 } 57 58 if (I->getOpcode() == Instruction::Add) { 59 // We have X+C. Check to see if we really have (X*C2)+C1, 60 // where C1 is divisible by C2. 61 unsigned SubScale; 62 Value *SubVal = 63 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset); 64 Offset += RHS->getZExtValue(); 65 Scale = SubScale; 66 return SubVal; 67 } 68 } 69 } 70 71 // Otherwise, we can't look past this. 72 Scale = 1; 73 Offset = 0; 74 return Val; 75} 76 77/// PromoteCastOfAllocation - If we find a cast of an allocation instruction, 78/// try to eliminate the cast by moving the type information into the alloc. 79Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI, 80 AllocaInst &AI) { 81 // This requires TargetData to get the alloca alignment and size information. 82 if (!TD) return 0; 83 84 PointerType *PTy = cast<PointerType>(CI.getType()); 85 86 BuilderTy AllocaBuilder(*Builder); 87 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI); 88 89 // Get the type really allocated and the type casted to. 90 Type *AllocElTy = AI.getAllocatedType(); 91 Type *CastElTy = PTy->getElementType(); 92 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0; 93 94 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy); 95 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy); 96 if (CastElTyAlign < AllocElTyAlign) return 0; 97 98 // If the allocation has multiple uses, only promote it if we are strictly 99 // increasing the alignment of the resultant allocation. If we keep it the 100 // same, we open the door to infinite loops of various kinds. 101 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0; 102 103 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy); 104 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy); 105 if (CastElTySize == 0 || AllocElTySize == 0) return 0; 106 107 // See if we can satisfy the modulus by pulling a scale out of the array 108 // size argument. 109 unsigned ArraySizeScale; 110 uint64_t ArrayOffset; 111 Value *NumElements = // See if the array size is a decomposable linear expr. 112 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset); 113 114 // If we can now satisfy the modulus, by using a non-1 scale, we really can 115 // do the xform. 116 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 || 117 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0; 118 119 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize; 120 Value *Amt = 0; 121 if (Scale == 1) { 122 Amt = NumElements; 123 } else { 124 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale); 125 // Insert before the alloca, not before the cast. 126 Amt = AllocaBuilder.CreateMul(Amt, NumElements); 127 } 128 129 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) { 130 Value *Off = ConstantInt::get(AI.getArraySize()->getType(), 131 Offset, true); 132 Amt = AllocaBuilder.CreateAdd(Amt, Off); 133 } 134 135 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt); 136 New->setAlignment(AI.getAlignment()); 137 New->takeName(&AI); 138 139 // If the allocation has multiple real uses, insert a cast and change all 140 // things that used it to use the new cast. This will also hack on CI, but it 141 // will die soon. 142 if (!AI.hasOneUse()) { 143 // New is the allocation instruction, pointer typed. AI is the original 144 // allocation instruction, also pointer typed. Thus, cast to use is BitCast. 145 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast"); 146 ReplaceInstUsesWith(AI, NewCast); 147 } 148 return ReplaceInstUsesWith(CI, New); 149} 150 151/// EvaluateInDifferentType - Given an expression that 152/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually 153/// insert the code to evaluate the expression. 154Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty, 155 bool isSigned) { 156 if (Constant *C = dyn_cast<Constant>(V)) { 157 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/); 158 // If we got a constantexpr back, try to simplify it with TD info. 159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 160 C = ConstantFoldConstantExpression(CE, TD, TLI); 161 return C; 162 } 163 164 // Otherwise, it must be an instruction. 165 Instruction *I = cast<Instruction>(V); 166 Instruction *Res = 0; 167 unsigned Opc = I->getOpcode(); 168 switch (Opc) { 169 case Instruction::Add: 170 case Instruction::Sub: 171 case Instruction::Mul: 172 case Instruction::And: 173 case Instruction::Or: 174 case Instruction::Xor: 175 case Instruction::AShr: 176 case Instruction::LShr: 177 case Instruction::Shl: 178 case Instruction::UDiv: 179 case Instruction::URem: { 180 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned); 181 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 182 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS); 183 break; 184 } 185 case Instruction::Trunc: 186 case Instruction::ZExt: 187 case Instruction::SExt: 188 // If the source type of the cast is the type we're trying for then we can 189 // just return the source. There's no need to insert it because it is not 190 // new. 191 if (I->getOperand(0)->getType() == Ty) 192 return I->getOperand(0); 193 194 // Otherwise, must be the same type of cast, so just reinsert a new one. 195 // This also handles the case of zext(trunc(x)) -> zext(x). 196 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty, 197 Opc == Instruction::SExt); 198 break; 199 case Instruction::Select: { 200 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned); 201 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned); 202 Res = SelectInst::Create(I->getOperand(0), True, False); 203 break; 204 } 205 case Instruction::PHI: { 206 PHINode *OPN = cast<PHINode>(I); 207 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues()); 208 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) { 209 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned); 210 NPN->addIncoming(V, OPN->getIncomingBlock(i)); 211 } 212 Res = NPN; 213 break; 214 } 215 default: 216 // TODO: Can handle more cases here. 217 llvm_unreachable("Unreachable!"); 218 break; 219 } 220 221 Res->takeName(I); 222 return InsertNewInstWith(Res, *I); 223} 224 225 226/// This function is a wrapper around CastInst::isEliminableCastPair. It 227/// simply extracts arguments and returns what that function returns. 228static Instruction::CastOps 229isEliminableCastPair( 230 const CastInst *CI, ///< The first cast instruction 231 unsigned opcode, ///< The opcode of the second cast instruction 232 Type *DstTy, ///< The target type for the second cast instruction 233 TargetData *TD ///< The target data for pointer size 234) { 235 236 Type *SrcTy = CI->getOperand(0)->getType(); // A from above 237 Type *MidTy = CI->getType(); // B from above 238 239 // Get the opcodes of the two Cast instructions 240 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode()); 241 Instruction::CastOps secondOp = Instruction::CastOps(opcode); 242 243 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, 244 DstTy, 245 TD ? TD->getIntPtrType(CI->getContext()) : 0); 246 247 // We don't want to form an inttoptr or ptrtoint that converts to an integer 248 // type that differs from the pointer size. 249 if ((Res == Instruction::IntToPtr && 250 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) || 251 (Res == Instruction::PtrToInt && 252 (!TD || DstTy != TD->getIntPtrType(CI->getContext())))) 253 Res = 0; 254 255 return Instruction::CastOps(Res); 256} 257 258/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 259/// results in any code being generated and is interesting to optimize out. If 260/// the cast can be eliminated by some other simple transformation, we prefer 261/// to do the simplification first. 262bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V, 263 Type *Ty) { 264 // Noop casts and casts of constants should be eliminated trivially. 265 if (V->getType() == Ty || isa<Constant>(V)) return false; 266 267 // If this is another cast that can be eliminated, we prefer to have it 268 // eliminated. 269 if (const CastInst *CI = dyn_cast<CastInst>(V)) 270 if (isEliminableCastPair(CI, opc, Ty, TD)) 271 return false; 272 273 // If this is a vector sext from a compare, then we don't want to break the 274 // idiom where each element of the extended vector is either zero or all ones. 275 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy()) 276 return false; 277 278 return true; 279} 280 281 282/// @brief Implement the transforms common to all CastInst visitors. 283Instruction *InstCombiner::commonCastTransforms(CastInst &CI) { 284 Value *Src = CI.getOperand(0); 285 286 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just 287 // eliminate it now. 288 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast 289 if (Instruction::CastOps opc = 290 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) { 291 // The first cast (CSrc) is eliminable so we need to fix up or replace 292 // the second cast (CI). CSrc will then have a good chance of being dead. 293 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType()); 294 } 295 } 296 297 // If we are casting a select then fold the cast into the select 298 if (SelectInst *SI = dyn_cast<SelectInst>(Src)) 299 if (Instruction *NV = FoldOpIntoSelect(CI, SI)) 300 return NV; 301 302 // If we are casting a PHI then fold the cast into the PHI 303 if (isa<PHINode>(Src)) { 304 // We don't do this if this would create a PHI node with an illegal type if 305 // it is currently legal. 306 if (!Src->getType()->isIntegerTy() || 307 !CI.getType()->isIntegerTy() || 308 ShouldChangeType(CI.getType(), Src->getType())) 309 if (Instruction *NV = FoldOpIntoPhi(CI)) 310 return NV; 311 } 312 313 return 0; 314} 315 316/// CanEvaluateTruncated - Return true if we can evaluate the specified 317/// expression tree as type Ty instead of its larger type, and arrive with the 318/// same value. This is used by code that tries to eliminate truncates. 319/// 320/// Ty will always be a type smaller than V. We should return true if trunc(V) 321/// can be computed by computing V in the smaller type. If V is an instruction, 322/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only 323/// makes sense if x and y can be efficiently truncated. 324/// 325/// This function works on both vectors and scalars. 326/// 327static bool CanEvaluateTruncated(Value *V, Type *Ty) { 328 // We can always evaluate constants in another type. 329 if (isa<Constant>(V)) 330 return true; 331 332 Instruction *I = dyn_cast<Instruction>(V); 333 if (!I) return false; 334 335 Type *OrigTy = V->getType(); 336 337 // If this is an extension from the dest type, we can eliminate it, even if it 338 // has multiple uses. 339 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) && 340 I->getOperand(0)->getType() == Ty) 341 return true; 342 343 // We can't extend or shrink something that has multiple uses: doing so would 344 // require duplicating the instruction in general, which isn't profitable. 345 if (!I->hasOneUse()) return false; 346 347 unsigned Opc = I->getOpcode(); 348 switch (Opc) { 349 case Instruction::Add: 350 case Instruction::Sub: 351 case Instruction::Mul: 352 case Instruction::And: 353 case Instruction::Or: 354 case Instruction::Xor: 355 // These operators can all arbitrarily be extended or truncated. 356 return CanEvaluateTruncated(I->getOperand(0), Ty) && 357 CanEvaluateTruncated(I->getOperand(1), Ty); 358 359 case Instruction::UDiv: 360 case Instruction::URem: { 361 // UDiv and URem can be truncated if all the truncated bits are zero. 362 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 363 uint32_t BitWidth = Ty->getScalarSizeInBits(); 364 if (BitWidth < OrigBitWidth) { 365 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth); 366 if (MaskedValueIsZero(I->getOperand(0), Mask) && 367 MaskedValueIsZero(I->getOperand(1), Mask)) { 368 return CanEvaluateTruncated(I->getOperand(0), Ty) && 369 CanEvaluateTruncated(I->getOperand(1), Ty); 370 } 371 } 372 break; 373 } 374 case Instruction::Shl: 375 // If we are truncating the result of this SHL, and if it's a shift of a 376 // constant amount, we can always perform a SHL in a smaller type. 377 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 378 uint32_t BitWidth = Ty->getScalarSizeInBits(); 379 if (CI->getLimitedValue(BitWidth) < BitWidth) 380 return CanEvaluateTruncated(I->getOperand(0), Ty); 381 } 382 break; 383 case Instruction::LShr: 384 // If this is a truncate of a logical shr, we can truncate it to a smaller 385 // lshr iff we know that the bits we would otherwise be shifting in are 386 // already zeros. 387 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 388 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits(); 389 uint32_t BitWidth = Ty->getScalarSizeInBits(); 390 if (MaskedValueIsZero(I->getOperand(0), 391 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) && 392 CI->getLimitedValue(BitWidth) < BitWidth) { 393 return CanEvaluateTruncated(I->getOperand(0), Ty); 394 } 395 } 396 break; 397 case Instruction::Trunc: 398 // trunc(trunc(x)) -> trunc(x) 399 return true; 400 case Instruction::ZExt: 401 case Instruction::SExt: 402 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 403 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest 404 return true; 405 case Instruction::Select: { 406 SelectInst *SI = cast<SelectInst>(I); 407 return CanEvaluateTruncated(SI->getTrueValue(), Ty) && 408 CanEvaluateTruncated(SI->getFalseValue(), Ty); 409 } 410 case Instruction::PHI: { 411 // We can change a phi if we can change all operands. Note that we never 412 // get into trouble with cyclic PHIs here because we only consider 413 // instructions with a single use. 414 PHINode *PN = cast<PHINode>(I); 415 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 416 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty)) 417 return false; 418 return true; 419 } 420 default: 421 // TODO: Can handle more cases here. 422 break; 423 } 424 425 return false; 426} 427 428Instruction *InstCombiner::visitTrunc(TruncInst &CI) { 429 if (Instruction *Result = commonCastTransforms(CI)) 430 return Result; 431 432 // See if we can simplify any instructions used by the input whose sole 433 // purpose is to compute bits we don't care about. 434 if (SimplifyDemandedInstructionBits(CI)) 435 return &CI; 436 437 Value *Src = CI.getOperand(0); 438 Type *DestTy = CI.getType(), *SrcTy = Src->getType(); 439 440 // Attempt to truncate the entire input expression tree to the destination 441 // type. Only do this if the dest type is a simple type, don't convert the 442 // expression tree to something weird like i93 unless the source is also 443 // strange. 444 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 445 CanEvaluateTruncated(Src, DestTy)) { 446 447 // If this cast is a truncate, evaluting in a different type always 448 // eliminates the cast, so it is always a win. 449 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 450 " to avoid cast: " << CI << '\n'); 451 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 452 assert(Res->getType() == DestTy); 453 return ReplaceInstUsesWith(CI, Res); 454 } 455 456 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector. 457 if (DestTy->getScalarSizeInBits() == 1) { 458 Constant *One = ConstantInt::get(Src->getType(), 1); 459 Src = Builder->CreateAnd(Src, One); 460 Value *Zero = Constant::getNullValue(Src->getType()); 461 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero); 462 } 463 464 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion. 465 Value *A = 0; ConstantInt *Cst = 0; 466 if (Src->hasOneUse() && 467 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) { 468 // We have three types to worry about here, the type of A, the source of 469 // the truncate (MidSize), and the destination of the truncate. We know that 470 // ASize < MidSize and MidSize > ResultSize, but don't know the relation 471 // between ASize and ResultSize. 472 unsigned ASize = A->getType()->getPrimitiveSizeInBits(); 473 474 // If the shift amount is larger than the size of A, then the result is 475 // known to be zero because all the input bits got shifted out. 476 if (Cst->getZExtValue() >= ASize) 477 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType())); 478 479 // Since we're doing an lshr and a zero extend, and know that the shift 480 // amount is smaller than ASize, it is always safe to do the shift in A's 481 // type, then zero extend or truncate to the result. 482 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue()); 483 Shift->takeName(Src); 484 return CastInst::CreateIntegerCast(Shift, CI.getType(), false); 485 } 486 487 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest 488 // type isn't non-native. 489 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) && 490 ShouldChangeType(Src->getType(), CI.getType()) && 491 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) { 492 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr"); 493 return BinaryOperator::CreateAnd(NewTrunc, 494 ConstantExpr::getTrunc(Cst, CI.getType())); 495 } 496 497 return 0; 498} 499 500/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations 501/// in order to eliminate the icmp. 502Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI, 503 bool DoXform) { 504 // If we are just checking for a icmp eq of a single bit and zext'ing it 505 // to an integer, then shift the bit to the appropriate place and then 506 // cast to integer to avoid the comparison. 507 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 508 const APInt &Op1CV = Op1C->getValue(); 509 510 // zext (x <s 0) to i32 --> x>>u31 true if signbit set. 511 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear. 512 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) || 513 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) { 514 if (!DoXform) return ICI; 515 516 Value *In = ICI->getOperand(0); 517 Value *Sh = ConstantInt::get(In->getType(), 518 In->getType()->getScalarSizeInBits()-1); 519 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit"); 520 if (In->getType() != CI.getType()) 521 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/); 522 523 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) { 524 Constant *One = ConstantInt::get(In->getType(), 1); 525 In = Builder->CreateXor(In, One, In->getName()+".not"); 526 } 527 528 return ReplaceInstUsesWith(CI, In); 529 } 530 531 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set. 532 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 533 // zext (X == 1) to i32 --> X iff X has only the low bit set. 534 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set. 535 // zext (X != 0) to i32 --> X iff X has only the low bit set. 536 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set. 537 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set. 538 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set. 539 if ((Op1CV == 0 || Op1CV.isPowerOf2()) && 540 // This only works for EQ and NE 541 ICI->isEquality()) { 542 // If Op1C some other power of two, convert: 543 uint32_t BitWidth = Op1C->getType()->getBitWidth(); 544 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 545 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 546 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne); 547 548 APInt KnownZeroMask(~KnownZero); 549 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1? 550 if (!DoXform) return ICI; 551 552 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE; 553 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) { 554 // (X&4) == 2 --> false 555 // (X&4) != 2 --> true 556 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()), 557 isNE); 558 Res = ConstantExpr::getZExt(Res, CI.getType()); 559 return ReplaceInstUsesWith(CI, Res); 560 } 561 562 uint32_t ShiftAmt = KnownZeroMask.logBase2(); 563 Value *In = ICI->getOperand(0); 564 if (ShiftAmt) { 565 // Perform a logical shr by shiftamt. 566 // Insert the shift to put the result in the low bit. 567 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt), 568 In->getName()+".lobit"); 569 } 570 571 if ((Op1CV != 0) == isNE) { // Toggle the low bit. 572 Constant *One = ConstantInt::get(In->getType(), 1); 573 In = Builder->CreateXor(In, One); 574 } 575 576 if (CI.getType() == In->getType()) 577 return ReplaceInstUsesWith(CI, In); 578 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/); 579 } 580 } 581 } 582 583 // icmp ne A, B is equal to xor A, B when A and B only really have one bit. 584 // It is also profitable to transform icmp eq into not(xor(A, B)) because that 585 // may lead to additional simplifications. 586 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) { 587 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) { 588 uint32_t BitWidth = ITy->getBitWidth(); 589 Value *LHS = ICI->getOperand(0); 590 Value *RHS = ICI->getOperand(1); 591 592 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0); 593 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0); 594 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 595 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS); 596 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS); 597 598 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) { 599 APInt KnownBits = KnownZeroLHS | KnownOneLHS; 600 APInt UnknownBit = ~KnownBits; 601 if (UnknownBit.countPopulation() == 1) { 602 if (!DoXform) return ICI; 603 604 Value *Result = Builder->CreateXor(LHS, RHS); 605 606 // Mask off any bits that are set and won't be shifted away. 607 if (KnownOneLHS.uge(UnknownBit)) 608 Result = Builder->CreateAnd(Result, 609 ConstantInt::get(ITy, UnknownBit)); 610 611 // Shift the bit we're testing down to the lsb. 612 Result = Builder->CreateLShr( 613 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros())); 614 615 if (ICI->getPredicate() == ICmpInst::ICMP_EQ) 616 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1)); 617 Result->takeName(ICI); 618 return ReplaceInstUsesWith(CI, Result); 619 } 620 } 621 } 622 } 623 624 return 0; 625} 626 627/// CanEvaluateZExtd - Determine if the specified value can be computed in the 628/// specified wider type and produce the same low bits. If not, return false. 629/// 630/// If this function returns true, it can also return a non-zero number of bits 631/// (in BitsToClear) which indicates that the value it computes is correct for 632/// the zero extend, but that the additional BitsToClear bits need to be zero'd 633/// out. For example, to promote something like: 634/// 635/// %B = trunc i64 %A to i32 636/// %C = lshr i32 %B, 8 637/// %E = zext i32 %C to i64 638/// 639/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be 640/// set to 8 to indicate that the promoted value needs to have bits 24-31 641/// cleared in addition to bits 32-63. Since an 'and' will be generated to 642/// clear the top bits anyway, doing this has no extra cost. 643/// 644/// This function works on both vectors and scalars. 645static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) { 646 BitsToClear = 0; 647 if (isa<Constant>(V)) 648 return true; 649 650 Instruction *I = dyn_cast<Instruction>(V); 651 if (!I) return false; 652 653 // If the input is a truncate from the destination type, we can trivially 654 // eliminate it, even if it has multiple uses. 655 // FIXME: This is currently disabled until codegen can handle this without 656 // pessimizing code, PR5997. 657 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 658 return true; 659 660 // We can't extend or shrink something that has multiple uses: doing so would 661 // require duplicating the instruction in general, which isn't profitable. 662 if (!I->hasOneUse()) return false; 663 664 unsigned Opc = I->getOpcode(), Tmp; 665 switch (Opc) { 666 case Instruction::ZExt: // zext(zext(x)) -> zext(x). 667 case Instruction::SExt: // zext(sext(x)) -> sext(x). 668 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x) 669 return true; 670 case Instruction::And: 671 case Instruction::Or: 672 case Instruction::Xor: 673 case Instruction::Add: 674 case Instruction::Sub: 675 case Instruction::Mul: 676 case Instruction::Shl: 677 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) || 678 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp)) 679 return false; 680 // These can all be promoted if neither operand has 'bits to clear'. 681 if (BitsToClear == 0 && Tmp == 0) 682 return true; 683 684 // If the operation is an AND/OR/XOR and the bits to clear are zero in the 685 // other side, BitsToClear is ok. 686 if (Tmp == 0 && 687 (Opc == Instruction::And || Opc == Instruction::Or || 688 Opc == Instruction::Xor)) { 689 // We use MaskedValueIsZero here for generality, but the case we care 690 // about the most is constant RHS. 691 unsigned VSize = V->getType()->getScalarSizeInBits(); 692 if (MaskedValueIsZero(I->getOperand(1), 693 APInt::getHighBitsSet(VSize, BitsToClear))) 694 return true; 695 } 696 697 // Otherwise, we don't know how to analyze this BitsToClear case yet. 698 return false; 699 700 case Instruction::LShr: 701 // We can promote lshr(x, cst) if we can promote x. This requires the 702 // ultimate 'and' to clear out the high zero bits we're clearing out though. 703 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) { 704 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear)) 705 return false; 706 BitsToClear += Amt->getZExtValue(); 707 if (BitsToClear > V->getType()->getScalarSizeInBits()) 708 BitsToClear = V->getType()->getScalarSizeInBits(); 709 return true; 710 } 711 // Cannot promote variable LSHR. 712 return false; 713 case Instruction::Select: 714 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) || 715 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) || 716 // TODO: If important, we could handle the case when the BitsToClear are 717 // known zero in the disagreeing side. 718 Tmp != BitsToClear) 719 return false; 720 return true; 721 722 case Instruction::PHI: { 723 // We can change a phi if we can change all operands. Note that we never 724 // get into trouble with cyclic PHIs here because we only consider 725 // instructions with a single use. 726 PHINode *PN = cast<PHINode>(I); 727 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear)) 728 return false; 729 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 730 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) || 731 // TODO: If important, we could handle the case when the BitsToClear 732 // are known zero in the disagreeing input. 733 Tmp != BitsToClear) 734 return false; 735 return true; 736 } 737 default: 738 // TODO: Can handle more cases here. 739 return false; 740 } 741} 742 743Instruction *InstCombiner::visitZExt(ZExtInst &CI) { 744 // If this zero extend is only used by a truncate, let the truncate by 745 // eliminated before we try to optimize this zext. 746 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 747 return 0; 748 749 // If one of the common conversion will work, do it. 750 if (Instruction *Result = commonCastTransforms(CI)) 751 return Result; 752 753 // See if we can simplify any instructions used by the input whose sole 754 // purpose is to compute bits we don't care about. 755 if (SimplifyDemandedInstructionBits(CI)) 756 return &CI; 757 758 Value *Src = CI.getOperand(0); 759 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 760 761 // Attempt to extend the entire input expression tree to the destination 762 // type. Only do this if the dest type is a simple type, don't convert the 763 // expression tree to something weird like i93 unless the source is also 764 // strange. 765 unsigned BitsToClear; 766 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 767 CanEvaluateZExtd(Src, DestTy, BitsToClear)) { 768 assert(BitsToClear < SrcTy->getScalarSizeInBits() && 769 "Unreasonable BitsToClear"); 770 771 // Okay, we can transform this! Insert the new expression now. 772 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 773 " to avoid zero extend: " << CI); 774 Value *Res = EvaluateInDifferentType(Src, DestTy, false); 775 assert(Res->getType() == DestTy); 776 777 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear; 778 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 779 780 // If the high bits are already filled with zeros, just replace this 781 // cast with the result. 782 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize, 783 DestBitSize-SrcBitsKept))) 784 return ReplaceInstUsesWith(CI, Res); 785 786 // We need to emit an AND to clear the high bits. 787 Constant *C = ConstantInt::get(Res->getType(), 788 APInt::getLowBitsSet(DestBitSize, SrcBitsKept)); 789 return BinaryOperator::CreateAnd(Res, C); 790 } 791 792 // If this is a TRUNC followed by a ZEXT then we are dealing with integral 793 // types and if the sizes are just right we can convert this into a logical 794 // 'and' which will be much cheaper than the pair of casts. 795 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast 796 // TODO: Subsume this into EvaluateInDifferentType. 797 798 // Get the sizes of the types involved. We know that the intermediate type 799 // will be smaller than A or C, but don't know the relation between A and C. 800 Value *A = CSrc->getOperand(0); 801 unsigned SrcSize = A->getType()->getScalarSizeInBits(); 802 unsigned MidSize = CSrc->getType()->getScalarSizeInBits(); 803 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 804 // If we're actually extending zero bits, then if 805 // SrcSize < DstSize: zext(a & mask) 806 // SrcSize == DstSize: a & mask 807 // SrcSize > DstSize: trunc(a) & mask 808 if (SrcSize < DstSize) { 809 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 810 Constant *AndConst = ConstantInt::get(A->getType(), AndValue); 811 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask"); 812 return new ZExtInst(And, CI.getType()); 813 } 814 815 if (SrcSize == DstSize) { 816 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize)); 817 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(), 818 AndValue)); 819 } 820 if (SrcSize > DstSize) { 821 Value *Trunc = Builder->CreateTrunc(A, CI.getType()); 822 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize)); 823 return BinaryOperator::CreateAnd(Trunc, 824 ConstantInt::get(Trunc->getType(), 825 AndValue)); 826 } 827 } 828 829 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 830 return transformZExtICmp(ICI, CI); 831 832 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src); 833 if (SrcI && SrcI->getOpcode() == Instruction::Or) { 834 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one 835 // of the (zext icmp) will be transformed. 836 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0)); 837 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1)); 838 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() && 839 (transformZExtICmp(LHS, CI, false) || 840 transformZExtICmp(RHS, CI, false))) { 841 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName()); 842 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName()); 843 return BinaryOperator::Create(Instruction::Or, LCast, RCast); 844 } 845 } 846 847 // zext(trunc(t) & C) -> (t & zext(C)). 848 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse()) 849 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 850 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) { 851 Value *TI0 = TI->getOperand(0); 852 if (TI0->getType() == CI.getType()) 853 return 854 BinaryOperator::CreateAnd(TI0, 855 ConstantExpr::getZExt(C, CI.getType())); 856 } 857 858 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)). 859 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse()) 860 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1))) 861 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0))) 862 if (And->getOpcode() == Instruction::And && And->hasOneUse() && 863 And->getOperand(1) == C) 864 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) { 865 Value *TI0 = TI->getOperand(0); 866 if (TI0->getType() == CI.getType()) { 867 Constant *ZC = ConstantExpr::getZExt(C, CI.getType()); 868 Value *NewAnd = Builder->CreateAnd(TI0, ZC); 869 return BinaryOperator::CreateXor(NewAnd, ZC); 870 } 871 } 872 873 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1 874 Value *X; 875 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) && 876 match(SrcI, m_Not(m_Value(X))) && 877 (!X->hasOneUse() || !isa<CmpInst>(X))) { 878 Value *New = Builder->CreateZExt(X, CI.getType()); 879 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1)); 880 } 881 882 return 0; 883} 884 885/// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations 886/// in order to eliminate the icmp. 887Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) { 888 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1); 889 ICmpInst::Predicate Pred = ICI->getPredicate(); 890 891 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 892 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative 893 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive 894 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) || 895 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) { 896 897 Value *Sh = ConstantInt::get(Op0->getType(), 898 Op0->getType()->getScalarSizeInBits()-1); 899 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit"); 900 if (In->getType() != CI.getType()) 901 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/); 902 903 if (Pred == ICmpInst::ICMP_SGT) 904 In = Builder->CreateNot(In, In->getName()+".not"); 905 return ReplaceInstUsesWith(CI, In); 906 } 907 908 // If we know that only one bit of the LHS of the icmp can be set and we 909 // have an equality comparison with zero or a power of 2, we can transform 910 // the icmp and sext into bitwise/integer operations. 911 if (ICI->hasOneUse() && 912 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){ 913 unsigned BitWidth = Op1C->getType()->getBitWidth(); 914 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 915 APInt TypeMask(APInt::getAllOnesValue(BitWidth)); 916 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne); 917 918 APInt KnownZeroMask(~KnownZero); 919 if (KnownZeroMask.isPowerOf2()) { 920 Value *In = ICI->getOperand(0); 921 922 // If the icmp tests for a known zero bit we can constant fold it. 923 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) { 924 Value *V = Pred == ICmpInst::ICMP_NE ? 925 ConstantInt::getAllOnesValue(CI.getType()) : 926 ConstantInt::getNullValue(CI.getType()); 927 return ReplaceInstUsesWith(CI, V); 928 } 929 930 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) { 931 // sext ((x & 2^n) == 0) -> (x >> n) - 1 932 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1 933 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros(); 934 // Perform a right shift to place the desired bit in the LSB. 935 if (ShiftAmt) 936 In = Builder->CreateLShr(In, 937 ConstantInt::get(In->getType(), ShiftAmt)); 938 939 // At this point "In" is either 1 or 0. Subtract 1 to turn 940 // {1, 0} -> {0, -1}. 941 In = Builder->CreateAdd(In, 942 ConstantInt::getAllOnesValue(In->getType()), 943 "sext"); 944 } else { 945 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1 946 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1 947 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros(); 948 // Perform a left shift to place the desired bit in the MSB. 949 if (ShiftAmt) 950 In = Builder->CreateShl(In, 951 ConstantInt::get(In->getType(), ShiftAmt)); 952 953 // Distribute the bit over the whole bit width. 954 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(), 955 BitWidth - 1), "sext"); 956 } 957 958 if (CI.getType() == In->getType()) 959 return ReplaceInstUsesWith(CI, In); 960 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/); 961 } 962 } 963 } 964 965 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed. 966 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) { 967 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) && 968 Op0->getType() == CI.getType()) { 969 Type *EltTy = VTy->getElementType(); 970 971 // splat the shift constant to a constant vector. 972 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1); 973 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit"); 974 return ReplaceInstUsesWith(CI, In); 975 } 976 } 977 978 return 0; 979} 980 981/// CanEvaluateSExtd - Return true if we can take the specified value 982/// and return it as type Ty without inserting any new casts and without 983/// changing the value of the common low bits. This is used by code that tries 984/// to promote integer operations to a wider types will allow us to eliminate 985/// the extension. 986/// 987/// This function works on both vectors and scalars. 988/// 989static bool CanEvaluateSExtd(Value *V, Type *Ty) { 990 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() && 991 "Can't sign extend type to a smaller type"); 992 // If this is a constant, it can be trivially promoted. 993 if (isa<Constant>(V)) 994 return true; 995 996 Instruction *I = dyn_cast<Instruction>(V); 997 if (!I) return false; 998 999 // If this is a truncate from the dest type, we can trivially eliminate it, 1000 // even if it has multiple uses. 1001 // FIXME: This is currently disabled until codegen can handle this without 1002 // pessimizing code, PR5997. 1003 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) 1004 return true; 1005 1006 // We can't extend or shrink something that has multiple uses: doing so would 1007 // require duplicating the instruction in general, which isn't profitable. 1008 if (!I->hasOneUse()) return false; 1009 1010 switch (I->getOpcode()) { 1011 case Instruction::SExt: // sext(sext(x)) -> sext(x) 1012 case Instruction::ZExt: // sext(zext(x)) -> zext(x) 1013 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x) 1014 return true; 1015 case Instruction::And: 1016 case Instruction::Or: 1017 case Instruction::Xor: 1018 case Instruction::Add: 1019 case Instruction::Sub: 1020 case Instruction::Mul: 1021 // These operators can all arbitrarily be extended if their inputs can. 1022 return CanEvaluateSExtd(I->getOperand(0), Ty) && 1023 CanEvaluateSExtd(I->getOperand(1), Ty); 1024 1025 //case Instruction::Shl: TODO 1026 //case Instruction::LShr: TODO 1027 1028 case Instruction::Select: 1029 return CanEvaluateSExtd(I->getOperand(1), Ty) && 1030 CanEvaluateSExtd(I->getOperand(2), Ty); 1031 1032 case Instruction::PHI: { 1033 // We can change a phi if we can change all operands. Note that we never 1034 // get into trouble with cyclic PHIs here because we only consider 1035 // instructions with a single use. 1036 PHINode *PN = cast<PHINode>(I); 1037 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1038 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false; 1039 return true; 1040 } 1041 default: 1042 // TODO: Can handle more cases here. 1043 break; 1044 } 1045 1046 return false; 1047} 1048 1049Instruction *InstCombiner::visitSExt(SExtInst &CI) { 1050 // If this sign extend is only used by a truncate, let the truncate by 1051 // eliminated before we try to optimize this zext. 1052 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back())) 1053 return 0; 1054 1055 if (Instruction *I = commonCastTransforms(CI)) 1056 return I; 1057 1058 // See if we can simplify any instructions used by the input whose sole 1059 // purpose is to compute bits we don't care about. 1060 if (SimplifyDemandedInstructionBits(CI)) 1061 return &CI; 1062 1063 Value *Src = CI.getOperand(0); 1064 Type *SrcTy = Src->getType(), *DestTy = CI.getType(); 1065 1066 // Attempt to extend the entire input expression tree to the destination 1067 // type. Only do this if the dest type is a simple type, don't convert the 1068 // expression tree to something weird like i93 unless the source is also 1069 // strange. 1070 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) && 1071 CanEvaluateSExtd(Src, DestTy)) { 1072 // Okay, we can transform this! Insert the new expression now. 1073 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type" 1074 " to avoid sign extend: " << CI); 1075 Value *Res = EvaluateInDifferentType(Src, DestTy, true); 1076 assert(Res->getType() == DestTy); 1077 1078 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1079 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1080 1081 // If the high bits are already filled with sign bit, just replace this 1082 // cast with the result. 1083 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize) 1084 return ReplaceInstUsesWith(CI, Res); 1085 1086 // We need to emit a shl + ashr to do the sign extend. 1087 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1088 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"), 1089 ShAmt); 1090 } 1091 1092 // If this input is a trunc from our destination, then turn sext(trunc(x)) 1093 // into shifts. 1094 if (TruncInst *TI = dyn_cast<TruncInst>(Src)) 1095 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) { 1096 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits(); 1097 uint32_t DestBitSize = DestTy->getScalarSizeInBits(); 1098 1099 // We need to emit a shl + ashr to do the sign extend. 1100 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize); 1101 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext"); 1102 return BinaryOperator::CreateAShr(Res, ShAmt); 1103 } 1104 1105 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) 1106 return transformSExtICmp(ICI, CI); 1107 1108 // If the input is a shl/ashr pair of a same constant, then this is a sign 1109 // extension from a smaller value. If we could trust arbitrary bitwidth 1110 // integers, we could turn this into a truncate to the smaller bit and then 1111 // use a sext for the whole extension. Since we don't, look deeper and check 1112 // for a truncate. If the source and dest are the same type, eliminate the 1113 // trunc and extend and just do shifts. For example, turn: 1114 // %a = trunc i32 %i to i8 1115 // %b = shl i8 %a, 6 1116 // %c = ashr i8 %b, 6 1117 // %d = sext i8 %c to i32 1118 // into: 1119 // %a = shl i32 %i, 30 1120 // %d = ashr i32 %a, 30 1121 Value *A = 0; 1122 // TODO: Eventually this could be subsumed by EvaluateInDifferentType. 1123 ConstantInt *BA = 0, *CA = 0; 1124 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)), 1125 m_ConstantInt(CA))) && 1126 BA == CA && A->getType() == CI.getType()) { 1127 unsigned MidSize = Src->getType()->getScalarSizeInBits(); 1128 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits(); 1129 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize; 1130 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt); 1131 A = Builder->CreateShl(A, ShAmtV, CI.getName()); 1132 return BinaryOperator::CreateAShr(A, ShAmtV); 1133 } 1134 1135 return 0; 1136} 1137 1138 1139/// FitsInFPType - Return a Constant* for the specified FP constant if it fits 1140/// in the specified FP type without changing its value. 1141static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) { 1142 bool losesInfo; 1143 APFloat F = CFP->getValueAPF(); 1144 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo); 1145 if (!losesInfo) 1146 return ConstantFP::get(CFP->getContext(), F); 1147 return 0; 1148} 1149 1150/// LookThroughFPExtensions - If this is an fp extension instruction, look 1151/// through it until we get the source value. 1152static Value *LookThroughFPExtensions(Value *V) { 1153 if (Instruction *I = dyn_cast<Instruction>(V)) 1154 if (I->getOpcode() == Instruction::FPExt) 1155 return LookThroughFPExtensions(I->getOperand(0)); 1156 1157 // If this value is a constant, return the constant in the smallest FP type 1158 // that can accurately represent it. This allows us to turn 1159 // (float)((double)X+2.0) into x+2.0f. 1160 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 1161 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext())) 1162 return V; // No constant folding of this. 1163 // See if the value can be truncated to half and then reextended. 1164 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf)) 1165 return V; 1166 // See if the value can be truncated to float and then reextended. 1167 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle)) 1168 return V; 1169 if (CFP->getType()->isDoubleTy()) 1170 return V; // Won't shrink. 1171 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble)) 1172 return V; 1173 // Don't try to shrink to various long double types. 1174 } 1175 1176 return V; 1177} 1178 1179Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { 1180 if (Instruction *I = commonCastTransforms(CI)) 1181 return I; 1182 1183 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are 1184 // smaller than the destination type, we can eliminate the truncate by doing 1185 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well 1186 // as many builtins (sqrt, etc). 1187 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0)); 1188 if (OpI && OpI->hasOneUse()) { 1189 switch (OpI->getOpcode()) { 1190 default: break; 1191 case Instruction::FAdd: 1192 case Instruction::FSub: 1193 case Instruction::FMul: 1194 case Instruction::FDiv: 1195 case Instruction::FRem: 1196 Type *SrcTy = OpI->getType(); 1197 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0)); 1198 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1)); 1199 if (LHSTrunc->getType() != SrcTy && 1200 RHSTrunc->getType() != SrcTy) { 1201 unsigned DstSize = CI.getType()->getScalarSizeInBits(); 1202 // If the source types were both smaller than the destination type of 1203 // the cast, do this xform. 1204 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize && 1205 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) { 1206 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType()); 1207 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType()); 1208 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc); 1209 } 1210 } 1211 break; 1212 } 1213 } 1214 1215 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) 1216 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); 1217 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) && 1218 Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) && 1219 Call->getNumArgOperands() == 1 && 1220 Call->hasOneUse()) { 1221 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); 1222 if (Arg && Arg->getOpcode() == Instruction::FPExt && 1223 CI.getType()->isFloatTy() && 1224 Call->getType()->isDoubleTy() && 1225 Arg->getType()->isDoubleTy() && 1226 Arg->getOperand(0)->getType()->isFloatTy()) { 1227 Function *Callee = Call->getCalledFunction(); 1228 Module *M = CI.getParent()->getParent()->getParent(); 1229 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf", 1230 Callee->getAttributes(), 1231 Builder->getFloatTy(), 1232 Builder->getFloatTy(), 1233 NULL); 1234 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), 1235 "sqrtfcall"); 1236 ret->setAttributes(Callee->getAttributes()); 1237 1238 1239 // Remove the old Call. With -fmath-errno, it won't get marked readnone. 1240 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType())); 1241 EraseInstFromFunction(*Call); 1242 return ret; 1243 } 1244 } 1245 1246 return 0; 1247} 1248 1249Instruction *InstCombiner::visitFPExt(CastInst &CI) { 1250 return commonCastTransforms(CI); 1251} 1252 1253Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) { 1254 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1255 if (OpI == 0) 1256 return commonCastTransforms(FI); 1257 1258 // fptoui(uitofp(X)) --> X 1259 // fptoui(sitofp(X)) --> X 1260 // This is safe if the intermediate type has enough bits in its mantissa to 1261 // accurately represent all values of X. For example, do not do this with 1262 // i64->float->i64. This is also safe for sitofp case, because any negative 1263 // 'X' value would cause an undefined result for the fptoui. 1264 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1265 OpI->getOperand(0)->getType() == FI.getType() && 1266 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */ 1267 OpI->getType()->getFPMantissaWidth()) 1268 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1269 1270 return commonCastTransforms(FI); 1271} 1272 1273Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) { 1274 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0)); 1275 if (OpI == 0) 1276 return commonCastTransforms(FI); 1277 1278 // fptosi(sitofp(X)) --> X 1279 // fptosi(uitofp(X)) --> X 1280 // This is safe if the intermediate type has enough bits in its mantissa to 1281 // accurately represent all values of X. For example, do not do this with 1282 // i64->float->i64. This is also safe for sitofp case, because any negative 1283 // 'X' value would cause an undefined result for the fptoui. 1284 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) && 1285 OpI->getOperand(0)->getType() == FI.getType() && 1286 (int)FI.getType()->getScalarSizeInBits() <= 1287 OpI->getType()->getFPMantissaWidth()) 1288 return ReplaceInstUsesWith(FI, OpI->getOperand(0)); 1289 1290 return commonCastTransforms(FI); 1291} 1292 1293Instruction *InstCombiner::visitUIToFP(CastInst &CI) { 1294 return commonCastTransforms(CI); 1295} 1296 1297Instruction *InstCombiner::visitSIToFP(CastInst &CI) { 1298 return commonCastTransforms(CI); 1299} 1300 1301Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) { 1302 // If the source integer type is not the intptr_t type for this target, do a 1303 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the 1304 // cast to be exposed to other transforms. 1305 if (TD) { 1306 if (CI.getOperand(0)->getType()->getScalarSizeInBits() > 1307 TD->getPointerSizeInBits()) { 1308 Value *P = Builder->CreateTrunc(CI.getOperand(0), 1309 TD->getIntPtrType(CI.getContext())); 1310 return new IntToPtrInst(P, CI.getType()); 1311 } 1312 if (CI.getOperand(0)->getType()->getScalarSizeInBits() < 1313 TD->getPointerSizeInBits()) { 1314 Value *P = Builder->CreateZExt(CI.getOperand(0), 1315 TD->getIntPtrType(CI.getContext())); 1316 return new IntToPtrInst(P, CI.getType()); 1317 } 1318 } 1319 1320 if (Instruction *I = commonCastTransforms(CI)) 1321 return I; 1322 1323 return 0; 1324} 1325 1326/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint) 1327Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) { 1328 Value *Src = CI.getOperand(0); 1329 1330 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) { 1331 // If casting the result of a getelementptr instruction with no offset, turn 1332 // this into a cast of the original pointer! 1333 if (GEP->hasAllZeroIndices()) { 1334 // Changing the cast operand is usually not a good idea but it is safe 1335 // here because the pointer operand is being replaced with another 1336 // pointer operand so the opcode doesn't need to change. 1337 Worklist.Add(GEP); 1338 CI.setOperand(0, GEP->getOperand(0)); 1339 return &CI; 1340 } 1341 1342 // If the GEP has a single use, and the base pointer is a bitcast, and the 1343 // GEP computes a constant offset, see if we can convert these three 1344 // instructions into fewer. This typically happens with unions and other 1345 // non-type-safe code. 1346 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) && 1347 GEP->hasAllConstantIndices()) { 1348 // We are guaranteed to get a constant from EmitGEPOffset. 1349 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP)); 1350 int64_t Offset = OffsetV->getSExtValue(); 1351 1352 // Get the base pointer input of the bitcast, and the type it points to. 1353 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0); 1354 Type *GEPIdxTy = 1355 cast<PointerType>(OrigBase->getType())->getElementType(); 1356 SmallVector<Value*, 8> NewIndices; 1357 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) { 1358 // If we were able to index down into an element, create the GEP 1359 // and bitcast the result. This eliminates one bitcast, potentially 1360 // two. 1361 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ? 1362 Builder->CreateInBoundsGEP(OrigBase, NewIndices) : 1363 Builder->CreateGEP(OrigBase, NewIndices); 1364 NGEP->takeName(GEP); 1365 1366 if (isa<BitCastInst>(CI)) 1367 return new BitCastInst(NGEP, CI.getType()); 1368 assert(isa<PtrToIntInst>(CI)); 1369 return new PtrToIntInst(NGEP, CI.getType()); 1370 } 1371 } 1372 } 1373 1374 return commonCastTransforms(CI); 1375} 1376 1377Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) { 1378 // If the destination integer type is not the intptr_t type for this target, 1379 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast 1380 // to be exposed to other transforms. 1381 if (TD) { 1382 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) { 1383 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1384 TD->getIntPtrType(CI.getContext())); 1385 return new TruncInst(P, CI.getType()); 1386 } 1387 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) { 1388 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), 1389 TD->getIntPtrType(CI.getContext())); 1390 return new ZExtInst(P, CI.getType()); 1391 } 1392 } 1393 1394 return commonPointerCastTransforms(CI); 1395} 1396 1397/// OptimizeVectorResize - This input value (which is known to have vector type) 1398/// is being zero extended or truncated to the specified vector type. Try to 1399/// replace it with a shuffle (and vector/vector bitcast) if possible. 1400/// 1401/// The source and destination vector types may have different element types. 1402static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy, 1403 InstCombiner &IC) { 1404 // We can only do this optimization if the output is a multiple of the input 1405 // element size, or the input is a multiple of the output element size. 1406 // Convert the input type to have the same element type as the output. 1407 VectorType *SrcTy = cast<VectorType>(InVal->getType()); 1408 1409 if (SrcTy->getElementType() != DestTy->getElementType()) { 1410 // The input types don't need to be identical, but for now they must be the 1411 // same size. There is no specific reason we couldn't handle things like 1412 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten 1413 // there yet. 1414 if (SrcTy->getElementType()->getPrimitiveSizeInBits() != 1415 DestTy->getElementType()->getPrimitiveSizeInBits()) 1416 return 0; 1417 1418 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements()); 1419 InVal = IC.Builder->CreateBitCast(InVal, SrcTy); 1420 } 1421 1422 // Now that the element types match, get the shuffle mask and RHS of the 1423 // shuffle to use, which depends on whether we're increasing or decreasing the 1424 // size of the input. 1425 SmallVector<Constant*, 16> ShuffleMask; 1426 Value *V2; 1427 IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext()); 1428 1429 if (SrcTy->getNumElements() > DestTy->getNumElements()) { 1430 // If we're shrinking the number of elements, just shuffle in the low 1431 // elements from the input and use undef as the second shuffle input. 1432 V2 = UndefValue::get(SrcTy); 1433 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i) 1434 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1435 1436 } else { 1437 // If we're increasing the number of elements, shuffle in all of the 1438 // elements from InVal and fill the rest of the result elements with zeros 1439 // from a constant zero. 1440 V2 = Constant::getNullValue(SrcTy); 1441 unsigned SrcElts = SrcTy->getNumElements(); 1442 for (unsigned i = 0, e = SrcElts; i != e; ++i) 1443 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i)); 1444 1445 // The excess elements reference the first element of the zero input. 1446 ShuffleMask.append(DestTy->getNumElements()-SrcElts, 1447 ConstantInt::get(Int32Ty, SrcElts)); 1448 } 1449 1450 return new ShuffleVectorInst(InVal, V2, ConstantVector::get(ShuffleMask)); 1451} 1452 1453static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) { 1454 return Value % Ty->getPrimitiveSizeInBits() == 0; 1455} 1456 1457static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) { 1458 return Value / Ty->getPrimitiveSizeInBits(); 1459} 1460 1461/// CollectInsertionElements - V is a value which is inserted into a vector of 1462/// VecEltTy. Look through the value to see if we can decompose it into 1463/// insertions into the vector. See the example in the comment for 1464/// OptimizeIntegerToVectorInsertions for the pattern this handles. 1465/// The type of V is always a non-zero multiple of VecEltTy's size. 1466/// 1467/// This returns false if the pattern can't be matched or true if it can, 1468/// filling in Elements with the elements found here. 1469static bool CollectInsertionElements(Value *V, unsigned ElementIndex, 1470 SmallVectorImpl<Value*> &Elements, 1471 Type *VecEltTy) { 1472 // Undef values never contribute useful bits to the result. 1473 if (isa<UndefValue>(V)) return true; 1474 1475 // If we got down to a value of the right type, we win, try inserting into the 1476 // right element. 1477 if (V->getType() == VecEltTy) { 1478 // Inserting null doesn't actually insert any elements. 1479 if (Constant *C = dyn_cast<Constant>(V)) 1480 if (C->isNullValue()) 1481 return true; 1482 1483 // Fail if multiple elements are inserted into this slot. 1484 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0) 1485 return false; 1486 1487 Elements[ElementIndex] = V; 1488 return true; 1489 } 1490 1491 if (Constant *C = dyn_cast<Constant>(V)) { 1492 // Figure out the # elements this provides, and bitcast it or slice it up 1493 // as required. 1494 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(), 1495 VecEltTy); 1496 // If the constant is the size of a vector element, we just need to bitcast 1497 // it to the right type so it gets properly inserted. 1498 if (NumElts == 1) 1499 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy), 1500 ElementIndex, Elements, VecEltTy); 1501 1502 // Okay, this is a constant that covers multiple elements. Slice it up into 1503 // pieces and insert each element-sized piece into the vector. 1504 if (!isa<IntegerType>(C->getType())) 1505 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(), 1506 C->getType()->getPrimitiveSizeInBits())); 1507 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits(); 1508 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize); 1509 1510 for (unsigned i = 0; i != NumElts; ++i) { 1511 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(), 1512 i*ElementSize)); 1513 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy); 1514 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy)) 1515 return false; 1516 } 1517 return true; 1518 } 1519 1520 if (!V->hasOneUse()) return false; 1521 1522 Instruction *I = dyn_cast<Instruction>(V); 1523 if (I == 0) return false; 1524 switch (I->getOpcode()) { 1525 default: return false; // Unhandled case. 1526 case Instruction::BitCast: 1527 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1528 Elements, VecEltTy); 1529 case Instruction::ZExt: 1530 if (!isMultipleOfTypeSize( 1531 I->getOperand(0)->getType()->getPrimitiveSizeInBits(), 1532 VecEltTy)) 1533 return false; 1534 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1535 Elements, VecEltTy); 1536 case Instruction::Or: 1537 return CollectInsertionElements(I->getOperand(0), ElementIndex, 1538 Elements, VecEltTy) && 1539 CollectInsertionElements(I->getOperand(1), ElementIndex, 1540 Elements, VecEltTy); 1541 case Instruction::Shl: { 1542 // Must be shifting by a constant that is a multiple of the element size. 1543 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1544 if (CI == 0) return false; 1545 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false; 1546 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy); 1547 1548 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift, 1549 Elements, VecEltTy); 1550 } 1551 1552 } 1553} 1554 1555 1556/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we 1557/// may be doing shifts and ors to assemble the elements of the vector manually. 1558/// Try to rip the code out and replace it with insertelements. This is to 1559/// optimize code like this: 1560/// 1561/// %tmp37 = bitcast float %inc to i32 1562/// %tmp38 = zext i32 %tmp37 to i64 1563/// %tmp31 = bitcast float %inc5 to i32 1564/// %tmp32 = zext i32 %tmp31 to i64 1565/// %tmp33 = shl i64 %tmp32, 32 1566/// %ins35 = or i64 %tmp33, %tmp38 1567/// %tmp43 = bitcast i64 %ins35 to <2 x float> 1568/// 1569/// Into two insertelements that do "buildvector{%inc, %inc5}". 1570static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI, 1571 InstCombiner &IC) { 1572 VectorType *DestVecTy = cast<VectorType>(CI.getType()); 1573 Value *IntInput = CI.getOperand(0); 1574 1575 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements()); 1576 if (!CollectInsertionElements(IntInput, 0, Elements, 1577 DestVecTy->getElementType())) 1578 return 0; 1579 1580 // If we succeeded, we know that all of the element are specified by Elements 1581 // or are zero if Elements has a null entry. Recast this as a set of 1582 // insertions. 1583 Value *Result = Constant::getNullValue(CI.getType()); 1584 for (unsigned i = 0, e = Elements.size(); i != e; ++i) { 1585 if (Elements[i] == 0) continue; // Unset element. 1586 1587 Result = IC.Builder->CreateInsertElement(Result, Elements[i], 1588 IC.Builder->getInt32(i)); 1589 } 1590 1591 return Result; 1592} 1593 1594 1595/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double 1596/// bitcast. The various long double bitcasts can't get in here. 1597static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){ 1598 Value *Src = CI.getOperand(0); 1599 Type *DestTy = CI.getType(); 1600 1601 // If this is a bitcast from int to float, check to see if the int is an 1602 // extraction from a vector. 1603 Value *VecInput = 0; 1604 // bitcast(trunc(bitcast(somevector))) 1605 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) && 1606 isa<VectorType>(VecInput->getType())) { 1607 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1608 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1609 1610 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) { 1611 // If the element type of the vector doesn't match the result type, 1612 // bitcast it to be a vector type we can extract from. 1613 if (VecTy->getElementType() != DestTy) { 1614 VecTy = VectorType::get(DestTy, 1615 VecTy->getPrimitiveSizeInBits() / DestWidth); 1616 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1617 } 1618 1619 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0)); 1620 } 1621 } 1622 1623 // bitcast(trunc(lshr(bitcast(somevector), cst)) 1624 ConstantInt *ShAmt = 0; 1625 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)), 1626 m_ConstantInt(ShAmt)))) && 1627 isa<VectorType>(VecInput->getType())) { 1628 VectorType *VecTy = cast<VectorType>(VecInput->getType()); 1629 unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); 1630 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 && 1631 ShAmt->getZExtValue() % DestWidth == 0) { 1632 // If the element type of the vector doesn't match the result type, 1633 // bitcast it to be a vector type we can extract from. 1634 if (VecTy->getElementType() != DestTy) { 1635 VecTy = VectorType::get(DestTy, 1636 VecTy->getPrimitiveSizeInBits() / DestWidth); 1637 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy); 1638 } 1639 1640 unsigned Elt = ShAmt->getZExtValue() / DestWidth; 1641 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt)); 1642 } 1643 } 1644 return 0; 1645} 1646 1647Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { 1648 // If the operands are integer typed then apply the integer transforms, 1649 // otherwise just apply the common ones. 1650 Value *Src = CI.getOperand(0); 1651 Type *SrcTy = Src->getType(); 1652 Type *DestTy = CI.getType(); 1653 1654 // Get rid of casts from one type to the same type. These are useless and can 1655 // be replaced by the operand. 1656 if (DestTy == Src->getType()) 1657 return ReplaceInstUsesWith(CI, Src); 1658 1659 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) { 1660 PointerType *SrcPTy = cast<PointerType>(SrcTy); 1661 Type *DstElTy = DstPTy->getElementType(); 1662 Type *SrcElTy = SrcPTy->getElementType(); 1663 1664 // If the address spaces don't match, don't eliminate the bitcast, which is 1665 // required for changing types. 1666 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace()) 1667 return 0; 1668 1669 // If we are casting a alloca to a pointer to a type of the same 1670 // size, rewrite the allocation instruction to allocate the "right" type. 1671 // There is no need to modify malloc calls because it is their bitcast that 1672 // needs to be cleaned up. 1673 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src)) 1674 if (Instruction *V = PromoteCastOfAllocation(CI, *AI)) 1675 return V; 1676 1677 // If the source and destination are pointers, and this cast is equivalent 1678 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep. 1679 // This can enhance SROA and other transforms that want type-safe pointers. 1680 Constant *ZeroUInt = 1681 Constant::getNullValue(Type::getInt32Ty(CI.getContext())); 1682 unsigned NumZeros = 0; 1683 while (SrcElTy != DstElTy && 1684 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() && 1685 SrcElTy->getNumContainedTypes() /* not "{}" */) { 1686 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt); 1687 ++NumZeros; 1688 } 1689 1690 // If we found a path from the src to dest, create the getelementptr now. 1691 if (SrcElTy == DstElTy) { 1692 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt); 1693 return GetElementPtrInst::CreateInBounds(Src, Idxs); 1694 } 1695 } 1696 1697 // Try to optimize int -> float bitcasts. 1698 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy)) 1699 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this)) 1700 return I; 1701 1702 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) { 1703 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) { 1704 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType()); 1705 return InsertElementInst::Create(UndefValue::get(DestTy), Elem, 1706 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1707 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast) 1708 } 1709 1710 if (isa<IntegerType>(SrcTy)) { 1711 // If this is a cast from an integer to vector, check to see if the input 1712 // is a trunc or zext of a bitcast from vector. If so, we can replace all 1713 // the casts with a shuffle and (potentially) a bitcast. 1714 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) { 1715 CastInst *SrcCast = cast<CastInst>(Src); 1716 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0))) 1717 if (isa<VectorType>(BCIn->getOperand(0)->getType())) 1718 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0), 1719 cast<VectorType>(DestTy), *this)) 1720 return I; 1721 } 1722 1723 // If the input is an 'or' instruction, we may be doing shifts and ors to 1724 // assemble the elements of the vector manually. Try to rip the code out 1725 // and replace it with insertelements. 1726 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this)) 1727 return ReplaceInstUsesWith(CI, V); 1728 } 1729 } 1730 1731 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) { 1732 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) { 1733 Value *Elem = 1734 Builder->CreateExtractElement(Src, 1735 Constant::getNullValue(Type::getInt32Ty(CI.getContext()))); 1736 return CastInst::Create(Instruction::BitCast, Elem, DestTy); 1737 } 1738 } 1739 1740 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) { 1741 // Okay, we have (bitcast (shuffle ..)). Check to see if this is 1742 // a bitcast to a vector with the same # elts. 1743 if (SVI->hasOneUse() && DestTy->isVectorTy() && 1744 cast<VectorType>(DestTy)->getNumElements() == 1745 SVI->getType()->getNumElements() && 1746 SVI->getType()->getNumElements() == 1747 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) { 1748 BitCastInst *Tmp; 1749 // If either of the operands is a cast from CI.getType(), then 1750 // evaluating the shuffle in the casted destination's type will allow 1751 // us to eliminate at least one cast. 1752 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) && 1753 Tmp->getOperand(0)->getType() == DestTy) || 1754 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) && 1755 Tmp->getOperand(0)->getType() == DestTy)) { 1756 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy); 1757 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy); 1758 // Return a new shuffle vector. Use the same element ID's, as we 1759 // know the vector types match #elts. 1760 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2)); 1761 } 1762 } 1763 } 1764 1765 if (SrcTy->isPointerTy()) 1766 return commonPointerCastTransforms(CI); 1767 return commonCastTransforms(CI); 1768} 1769