ConstantFolding.cpp revision 80f495aab0b1103b880196191af56f1d1c473ea1
1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===// 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 defines routines for folding instructions into constants. 11// 12// Also, to supplement the basic IR ConstantExpr simplifications, 13// this file defines some additional folding routines that can make use of 14// DataLayout information. These functions cannot go in IR due to library 15// dependency issues. 16// 17//===----------------------------------------------------------------------===// 18 19#include "llvm/Analysis/ConstantFolding.h" 20#include "llvm/ADT/SmallPtrSet.h" 21#include "llvm/ADT/SmallVector.h" 22#include "llvm/ADT/StringMap.h" 23#include "llvm/Analysis/ValueTracking.h" 24#include "llvm/IR/Constants.h" 25#include "llvm/IR/DataLayout.h" 26#include "llvm/IR/DerivedTypes.h" 27#include "llvm/IR/Function.h" 28#include "llvm/IR/GlobalVariable.h" 29#include "llvm/IR/Instructions.h" 30#include "llvm/IR/Intrinsics.h" 31#include "llvm/IR/Operator.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/FEnv.h" 34#include "llvm/Support/GetElementPtrTypeIterator.h" 35#include "llvm/Support/MathExtras.h" 36#include "llvm/Target/TargetLibraryInfo.h" 37#include <cerrno> 38#include <cmath> 39using namespace llvm; 40 41//===----------------------------------------------------------------------===// 42// Constant Folding internal helper functions 43//===----------------------------------------------------------------------===// 44 45/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with 46/// DataLayout. This always returns a non-null constant, but it may be a 47/// ConstantExpr if unfoldable. 48static Constant *FoldBitCast(Constant *C, Type *DestTy, 49 const DataLayout &TD) { 50 // Catch the obvious splat cases. 51 if (C->isNullValue() && !DestTy->isX86_MMXTy()) 52 return Constant::getNullValue(DestTy); 53 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy()) 54 return Constant::getAllOnesValue(DestTy); 55 56 // Handle a vector->integer cast. 57 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) { 58 VectorType *VTy = dyn_cast<VectorType>(C->getType()); 59 if (VTy == 0) 60 return ConstantExpr::getBitCast(C, DestTy); 61 62 unsigned NumSrcElts = VTy->getNumElements(); 63 Type *SrcEltTy = VTy->getElementType(); 64 65 // If the vector is a vector of floating point, convert it to vector of int 66 // to simplify things. 67 if (SrcEltTy->isFloatingPointTy()) { 68 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 69 Type *SrcIVTy = 70 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts); 71 // Ask IR to do the conversion now that #elts line up. 72 C = ConstantExpr::getBitCast(C, SrcIVTy); 73 } 74 75 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C); 76 if (CDV == 0) 77 return ConstantExpr::getBitCast(C, DestTy); 78 79 // Now that we know that the input value is a vector of integers, just shift 80 // and insert them into our result. 81 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy); 82 APInt Result(IT->getBitWidth(), 0); 83 for (unsigned i = 0; i != NumSrcElts; ++i) { 84 Result <<= BitShift; 85 if (TD.isLittleEndian()) 86 Result |= CDV->getElementAsInteger(NumSrcElts-i-1); 87 else 88 Result |= CDV->getElementAsInteger(i); 89 } 90 91 return ConstantInt::get(IT, Result); 92 } 93 94 // The code below only handles casts to vectors currently. 95 VectorType *DestVTy = dyn_cast<VectorType>(DestTy); 96 if (DestVTy == 0) 97 return ConstantExpr::getBitCast(C, DestTy); 98 99 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 100 // vector so the code below can handle it uniformly. 101 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 102 Constant *Ops = C; // don't take the address of C! 103 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD); 104 } 105 106 // If this is a bitcast from constant vector -> vector, fold it. 107 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C)) 108 return ConstantExpr::getBitCast(C, DestTy); 109 110 // If the element types match, IR can fold it. 111 unsigned NumDstElt = DestVTy->getNumElements(); 112 unsigned NumSrcElt = C->getType()->getVectorNumElements(); 113 if (NumDstElt == NumSrcElt) 114 return ConstantExpr::getBitCast(C, DestTy); 115 116 Type *SrcEltTy = C->getType()->getVectorElementType(); 117 Type *DstEltTy = DestVTy->getElementType(); 118 119 // Otherwise, we're changing the number of elements in a vector, which 120 // requires endianness information to do the right thing. For example, 121 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 122 // folds to (little endian): 123 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 124 // and to (big endian): 125 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 126 127 // First thing is first. We only want to think about integer here, so if 128 // we have something in FP form, recast it as integer. 129 if (DstEltTy->isFloatingPointTy()) { 130 // Fold to an vector of integers with same size as our FP type. 131 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 132 Type *DestIVTy = 133 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 134 // Recursively handle this integer conversion, if possible. 135 C = FoldBitCast(C, DestIVTy, TD); 136 137 // Finally, IR can handle this now that #elts line up. 138 return ConstantExpr::getBitCast(C, DestTy); 139 } 140 141 // Okay, we know the destination is integer, if the input is FP, convert 142 // it to integer first. 143 if (SrcEltTy->isFloatingPointTy()) { 144 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 145 Type *SrcIVTy = 146 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 147 // Ask IR to do the conversion now that #elts line up. 148 C = ConstantExpr::getBitCast(C, SrcIVTy); 149 // If IR wasn't able to fold it, bail out. 150 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector. 151 !isa<ConstantDataVector>(C)) 152 return C; 153 } 154 155 // Now we know that the input and output vectors are both integer vectors 156 // of the same size, and that their #elements is not the same. Do the 157 // conversion here, which depends on whether the input or output has 158 // more elements. 159 bool isLittleEndian = TD.isLittleEndian(); 160 161 SmallVector<Constant*, 32> Result; 162 if (NumDstElt < NumSrcElt) { 163 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 164 Constant *Zero = Constant::getNullValue(DstEltTy); 165 unsigned Ratio = NumSrcElt/NumDstElt; 166 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 167 unsigned SrcElt = 0; 168 for (unsigned i = 0; i != NumDstElt; ++i) { 169 // Build each element of the result. 170 Constant *Elt = Zero; 171 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 172 for (unsigned j = 0; j != Ratio; ++j) { 173 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++)); 174 if (!Src) // Reject constantexpr elements. 175 return ConstantExpr::getBitCast(C, DestTy); 176 177 // Zero extend the element to the right size. 178 Src = ConstantExpr::getZExt(Src, Elt->getType()); 179 180 // Shift it to the right place, depending on endianness. 181 Src = ConstantExpr::getShl(Src, 182 ConstantInt::get(Src->getType(), ShiftAmt)); 183 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 184 185 // Mix it in. 186 Elt = ConstantExpr::getOr(Elt, Src); 187 } 188 Result.push_back(Elt); 189 } 190 return ConstantVector::get(Result); 191 } 192 193 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 194 unsigned Ratio = NumDstElt/NumSrcElt; 195 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); 196 197 // Loop over each source value, expanding into multiple results. 198 for (unsigned i = 0; i != NumSrcElt; ++i) { 199 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i)); 200 if (!Src) // Reject constantexpr elements. 201 return ConstantExpr::getBitCast(C, DestTy); 202 203 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 204 for (unsigned j = 0; j != Ratio; ++j) { 205 // Shift the piece of the value into the right place, depending on 206 // endianness. 207 Constant *Elt = ConstantExpr::getLShr(Src, 208 ConstantInt::get(Src->getType(), ShiftAmt)); 209 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 210 211 // Truncate and remember this piece. 212 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 213 } 214 } 215 216 return ConstantVector::get(Result); 217} 218 219 220/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset 221/// from a global, return the global and the constant. Because of 222/// constantexprs, this function is recursive. 223static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 224 APInt &Offset, const DataLayout &TD) { 225 // Trivial case, constant is the global. 226 if ((GV = dyn_cast<GlobalValue>(C))) { 227 Offset.clearAllBits(); 228 return true; 229 } 230 231 // Otherwise, if this isn't a constant expr, bail out. 232 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 233 if (!CE) return false; 234 235 // Look through ptr->int and ptr->ptr casts. 236 if (CE->getOpcode() == Instruction::PtrToInt || 237 CE->getOpcode() == Instruction::BitCast) 238 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); 239 240 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 241 if (GEPOperator *GEP = dyn_cast<GEPOperator>(CE)) { 242 // If the base isn't a global+constant, we aren't either. 243 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) 244 return false; 245 246 // Otherwise, add any offset that our operands provide. 247 return GEP->accumulateConstantOffset(TD, Offset); 248 } 249 250 return false; 251} 252 253/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the 254/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the 255/// pointer to copy results into and BytesLeft is the number of bytes left in 256/// the CurPtr buffer. TD is the target data. 257static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, 258 unsigned char *CurPtr, unsigned BytesLeft, 259 const DataLayout &TD) { 260 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && 261 "Out of range access"); 262 263 // If this element is zero or undefined, we can just return since *CurPtr is 264 // zero initialized. 265 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 266 return true; 267 268 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 269 if (CI->getBitWidth() > 64 || 270 (CI->getBitWidth() & 7) != 0) 271 return false; 272 273 uint64_t Val = CI->getZExtValue(); 274 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 275 276 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 277 int n = ByteOffset; 278 if (!TD.isLittleEndian()) 279 n = IntBytes - n - 1; 280 CurPtr[i] = (unsigned char)(Val >> (n * 8)); 281 ++ByteOffset; 282 } 283 return true; 284 } 285 286 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 287 if (CFP->getType()->isDoubleTy()) { 288 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); 289 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 290 } 291 if (CFP->getType()->isFloatTy()){ 292 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); 293 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 294 } 295 if (CFP->getType()->isHalfTy()){ 296 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), TD); 297 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 298 } 299 return false; 300 } 301 302 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 303 const StructLayout *SL = TD.getStructLayout(CS->getType()); 304 unsigned Index = SL->getElementContainingOffset(ByteOffset); 305 uint64_t CurEltOffset = SL->getElementOffset(Index); 306 ByteOffset -= CurEltOffset; 307 308 while (1) { 309 // If the element access is to the element itself and not to tail padding, 310 // read the bytes from the element. 311 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); 312 313 if (ByteOffset < EltSize && 314 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 315 BytesLeft, TD)) 316 return false; 317 318 ++Index; 319 320 // Check to see if we read from the last struct element, if so we're done. 321 if (Index == CS->getType()->getNumElements()) 322 return true; 323 324 // If we read all of the bytes we needed from this element we're done. 325 uint64_t NextEltOffset = SL->getElementOffset(Index); 326 327 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset) 328 return true; 329 330 // Move to the next element of the struct. 331 CurPtr += NextEltOffset - CurEltOffset - ByteOffset; 332 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset; 333 ByteOffset = 0; 334 CurEltOffset = NextEltOffset; 335 } 336 // not reached. 337 } 338 339 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) || 340 isa<ConstantDataSequential>(C)) { 341 Type *EltTy = C->getType()->getSequentialElementType(); 342 uint64_t EltSize = TD.getTypeAllocSize(EltTy); 343 uint64_t Index = ByteOffset / EltSize; 344 uint64_t Offset = ByteOffset - Index * EltSize; 345 uint64_t NumElts; 346 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType())) 347 NumElts = AT->getNumElements(); 348 else 349 NumElts = C->getType()->getVectorNumElements(); 350 351 for (; Index != NumElts; ++Index) { 352 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr, 353 BytesLeft, TD)) 354 return false; 355 356 uint64_t BytesWritten = EltSize - Offset; 357 assert(BytesWritten <= EltSize && "Not indexing into this element?"); 358 if (BytesWritten >= BytesLeft) 359 return true; 360 361 Offset = 0; 362 BytesLeft -= BytesWritten; 363 CurPtr += BytesWritten; 364 } 365 return true; 366 } 367 368 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 369 if (CE->getOpcode() == Instruction::IntToPtr && 370 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getType())) { 371 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, 372 BytesLeft, TD); 373 } 374 } 375 376 // Otherwise, unknown initializer type. 377 return false; 378} 379 380static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 381 const DataLayout &TD) { 382 PointerType *PTy = cast<PointerType>(C->getType()); 383 Type *LoadTy = PTy->getElementType(); 384 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 385 386 // If this isn't an integer load we can't fold it directly. 387 if (!IntType) { 388 unsigned AS = PTy->getAddressSpace(); 389 390 // If this is a float/double load, we can try folding it as an int32/64 load 391 // and then bitcast the result. This can be useful for union cases. Note 392 // that address spaces don't matter here since we're not going to result in 393 // an actual new load. 394 Type *MapTy; 395 if (LoadTy->isHalfTy()) 396 MapTy = Type::getInt16PtrTy(C->getContext(), AS); 397 else if (LoadTy->isFloatTy()) 398 MapTy = Type::getInt32PtrTy(C->getContext(), AS); 399 else if (LoadTy->isDoubleTy()) 400 MapTy = Type::getInt64PtrTy(C->getContext(), AS); 401 else if (LoadTy->isVectorTy()) { 402 MapTy = PointerType::getIntNPtrTy(C->getContext(), 403 TD.getTypeAllocSizeInBits(LoadTy), 404 AS); 405 } else 406 return 0; 407 408 C = FoldBitCast(C, MapTy, TD); 409 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 410 return FoldBitCast(Res, LoadTy, TD); 411 return 0; 412 } 413 414 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 415 if (BytesLoaded > 32 || BytesLoaded == 0) 416 return 0; 417 418 GlobalValue *GVal; 419 APInt Offset(TD.getPointerTypeSizeInBits(PTy), 0); 420 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 421 return 0; 422 423 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 424 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 425 !GV->getInitializer()->getType()->isSized()) 426 return 0; 427 428 // If we're loading off the beginning of the global, some bytes may be valid, 429 // but we don't try to handle this. 430 if (Offset.isNegative()) 431 return 0; 432 433 // If we're not accessing anything in this constant, the result is undefined. 434 if (Offset.getZExtValue() >= 435 TD.getTypeAllocSize(GV->getInitializer()->getType())) 436 return UndefValue::get(IntType); 437 438 unsigned char RawBytes[32] = {0}; 439 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes, 440 BytesLoaded, TD)) 441 return 0; 442 443 APInt ResultVal = APInt(IntType->getBitWidth(), 0); 444 if (TD.isLittleEndian()) { 445 ResultVal = RawBytes[BytesLoaded - 1]; 446 for (unsigned i = 1; i != BytesLoaded; ++i) { 447 ResultVal <<= 8; 448 ResultVal |= RawBytes[BytesLoaded - 1 - i]; 449 } 450 } else { 451 ResultVal = RawBytes[0]; 452 for (unsigned i = 1; i != BytesLoaded; ++i) { 453 ResultVal <<= 8; 454 ResultVal |= RawBytes[i]; 455 } 456 } 457 458 return ConstantInt::get(IntType->getContext(), ResultVal); 459} 460 461/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 462/// produce if it is constant and determinable. If this is not determinable, 463/// return null. 464Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 465 const DataLayout *TD) { 466 // First, try the easy cases: 467 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 468 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 469 return GV->getInitializer(); 470 471 // If the loaded value isn't a constant expr, we can't handle it. 472 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 473 if (!CE) 474 return 0; 475 476 if (CE->getOpcode() == Instruction::GetElementPtr) { 477 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) { 478 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 479 if (Constant *V = 480 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 481 return V; 482 } 483 } 484 } 485 486 // Instead of loading constant c string, use corresponding integer value 487 // directly if string length is small enough. 488 StringRef Str; 489 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) { 490 unsigned StrLen = Str.size(); 491 Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 492 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 493 // Replace load with immediate integer if the result is an integer or fp 494 // value. 495 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 496 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 497 APInt StrVal(NumBits, 0); 498 APInt SingleChar(NumBits, 0); 499 if (TD->isLittleEndian()) { 500 for (signed i = StrLen-1; i >= 0; i--) { 501 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 502 StrVal = (StrVal << 8) | SingleChar; 503 } 504 } else { 505 for (unsigned i = 0; i < StrLen; i++) { 506 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 507 StrVal = (StrVal << 8) | SingleChar; 508 } 509 // Append NULL at the end. 510 SingleChar = 0; 511 StrVal = (StrVal << 8) | SingleChar; 512 } 513 514 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 515 if (Ty->isFloatingPointTy()) 516 Res = ConstantExpr::getBitCast(Res, Ty); 517 return Res; 518 } 519 } 520 521 // If this load comes from anywhere in a constant global, and if the global 522 // is all undef or zero, we know what it loads. 523 if (GlobalVariable *GV = 524 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) { 525 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 526 Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 527 if (GV->getInitializer()->isNullValue()) 528 return Constant::getNullValue(ResTy); 529 if (isa<UndefValue>(GV->getInitializer())) 530 return UndefValue::get(ResTy); 531 } 532 } 533 534 // Try hard to fold loads from bitcasted strange and non-type-safe things. 535 if (TD) 536 return FoldReinterpretLoadFromConstPtr(CE, *TD); 537 return 0; 538} 539 540static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){ 541 if (LI->isVolatile()) return 0; 542 543 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 544 return ConstantFoldLoadFromConstPtr(C, TD); 545 546 return 0; 547} 548 549/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 550/// Attempt to symbolically evaluate the result of a binary operator merging 551/// these together. If target data info is available, it is provided as DL, 552/// otherwise DL is null. 553static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 554 Constant *Op1, const DataLayout *DL){ 555 // SROA 556 557 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 558 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 559 // bits. 560 561 562 if (Opc == Instruction::And && DL) { 563 unsigned BitWidth = DL->getTypeSizeInBits(Op0->getType()->getScalarType()); 564 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0); 565 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0); 566 ComputeMaskedBits(Op0, KnownZero0, KnownOne0, DL); 567 ComputeMaskedBits(Op1, KnownZero1, KnownOne1, DL); 568 if ((KnownOne1 | KnownZero0).isAllOnesValue()) { 569 // All the bits of Op0 that the 'and' could be masking are already zero. 570 return Op0; 571 } 572 if ((KnownOne0 | KnownZero1).isAllOnesValue()) { 573 // All the bits of Op1 that the 'and' could be masking are already zero. 574 return Op1; 575 } 576 577 APInt KnownZero = KnownZero0 | KnownZero1; 578 APInt KnownOne = KnownOne0 & KnownOne1; 579 if ((KnownZero | KnownOne).isAllOnesValue()) { 580 return ConstantInt::get(Op0->getType(), KnownOne); 581 } 582 } 583 584 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 585 // constant. This happens frequently when iterating over a global array. 586 if (Opc == Instruction::Sub && DL) { 587 GlobalValue *GV1, *GV2; 588 unsigned PtrSize = DL->getPointerSizeInBits(); 589 unsigned OpSize = DL->getTypeSizeInBits(Op0->getType()); 590 APInt Offs1(PtrSize, 0), Offs2(PtrSize, 0); 591 592 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *DL)) 593 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *DL) && 594 GV1 == GV2) { 595 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 596 // PtrToInt may change the bitwidth so we have convert to the right size 597 // first. 598 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) - 599 Offs2.zextOrTrunc(OpSize)); 600 } 601 } 602 603 return 0; 604} 605 606/// CastGEPIndices - If array indices are not pointer-sized integers, 607/// explicitly cast them so that they aren't implicitly casted by the 608/// getelementptr. 609static Constant *CastGEPIndices(ArrayRef<Constant *> Ops, 610 Type *ResultTy, const DataLayout *TD, 611 const TargetLibraryInfo *TLI) { 612 if (!TD) 613 return 0; 614 615 Type *IntPtrTy = TD->getIntPtrType(ResultTy); 616 617 bool Any = false; 618 SmallVector<Constant*, 32> NewIdxs; 619 for (unsigned i = 1, e = Ops.size(); i != e; ++i) { 620 if ((i == 1 || 621 !isa<StructType>(GetElementPtrInst::getIndexedType( 622 Ops[0]->getType(), 623 Ops.slice(1, i - 1)))) && 624 Ops[i]->getType() != IntPtrTy) { 625 Any = true; 626 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 627 true, 628 IntPtrTy, 629 true), 630 Ops[i], IntPtrTy)); 631 } else 632 NewIdxs.push_back(Ops[i]); 633 } 634 635 if (!Any) 636 return 0; 637 638 Constant *C = ConstantExpr::getGetElementPtr(Ops[0], NewIdxs); 639 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 640 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) 641 C = Folded; 642 } 643 644 return C; 645} 646 647/// Strip the pointer casts, but preserve the address space information. 648static Constant* StripPtrCastKeepAS(Constant* Ptr) { 649 assert(Ptr->getType()->isPointerTy() && "Not a pointer type"); 650 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType()); 651 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 652 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType()); 653 654 // Preserve the address space number of the pointer. 655 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) { 656 NewPtrTy = NewPtrTy->getElementType()->getPointerTo( 657 OldPtrTy->getAddressSpace()); 658 Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy); 659 } 660 return Ptr; 661} 662 663/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP 664/// constant expression, do so. 665static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops, 666 Type *ResultTy, const DataLayout *TD, 667 const TargetLibraryInfo *TLI) { 668 Constant *Ptr = Ops[0]; 669 if (!TD || !Ptr->getType()->getPointerElementType()->isSized() || 670 !Ptr->getType()->isPointerTy()) 671 return 0; 672 673 Type *IntPtrTy = TD->getIntPtrType(Ptr->getType()); 674 Type *ResultElementTy = ResultTy->getPointerElementType(); 675 676 // If this is a constant expr gep that is effectively computing an 677 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' 678 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 679 if (!isa<ConstantInt>(Ops[i])) { 680 681 // If this is "gep i8* Ptr, (sub 0, V)", fold this as: 682 // "inttoptr (sub (ptrtoint Ptr), V)" 683 if (Ops.size() == 2 && ResultElementTy->isIntegerTy(8)) { 684 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]); 685 assert((CE == 0 || CE->getType() == IntPtrTy) && 686 "CastGEPIndices didn't canonicalize index types!"); 687 if (CE && CE->getOpcode() == Instruction::Sub && 688 CE->getOperand(0)->isNullValue()) { 689 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType()); 690 Res = ConstantExpr::getSub(Res, CE->getOperand(1)); 691 Res = ConstantExpr::getIntToPtr(Res, ResultTy); 692 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res)) 693 Res = ConstantFoldConstantExpression(ResCE, TD, TLI); 694 return Res; 695 } 696 } 697 return 0; 698 } 699 700 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy); 701 APInt Offset = 702 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(), 703 makeArrayRef((Value *const*) 704 Ops.data() + 1, 705 Ops.size() - 1))); 706 Ptr = StripPtrCastKeepAS(Ptr); 707 708 // If this is a GEP of a GEP, fold it all into a single GEP. 709 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 710 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end()); 711 712 // Do not try the incorporate the sub-GEP if some index is not a number. 713 bool AllConstantInt = true; 714 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 715 if (!isa<ConstantInt>(NestedOps[i])) { 716 AllConstantInt = false; 717 break; 718 } 719 if (!AllConstantInt) 720 break; 721 722 Ptr = cast<Constant>(GEP->getOperand(0)); 723 Offset += APInt(BitWidth, 724 TD->getIndexedOffset(Ptr->getType(), NestedOps)); 725 Ptr = StripPtrCastKeepAS(Ptr); 726 } 727 728 // If the base value for this address is a literal integer value, fold the 729 // getelementptr to the resulting integer value casted to the pointer type. 730 APInt BasePtr(BitWidth, 0); 731 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) { 732 if (CE->getOpcode() == Instruction::IntToPtr) { 733 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) 734 BasePtr = Base->getValue().zextOrTrunc(BitWidth); 735 } 736 } 737 738 if (Ptr->isNullValue() || BasePtr != 0) { 739 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr); 740 return ConstantExpr::getIntToPtr(C, ResultTy); 741 } 742 743 // Otherwise form a regular getelementptr. Recompute the indices so that 744 // we eliminate over-indexing of the notional static type array bounds. 745 // This makes it easy to determine if the getelementptr is "inbounds". 746 // Also, this helps GlobalOpt do SROA on GlobalVariables. 747 Type *Ty = Ptr->getType(); 748 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type"); 749 SmallVector<Constant *, 32> NewIdxs; 750 751 do { 752 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 753 if (ATy->isPointerTy()) { 754 // The only pointer indexing we'll do is on the first index of the GEP. 755 if (!NewIdxs.empty()) 756 break; 757 758 // Only handle pointers to sized types, not pointers to functions. 759 if (!ATy->getElementType()->isSized()) 760 return 0; 761 } 762 763 // Determine which element of the array the offset points into. 764 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 765 if (ElemSize == 0) 766 // The element size is 0. This may be [0 x Ty]*, so just use a zero 767 // index for this level and proceed to the next level to see if it can 768 // accommodate the offset. 769 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 770 else { 771 // The element size is non-zero divide the offset by the element 772 // size (rounding down), to compute the index at this level. 773 APInt NewIdx = Offset.udiv(ElemSize); 774 Offset -= NewIdx * ElemSize; 775 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 776 } 777 Ty = ATy->getElementType(); 778 } else if (StructType *STy = dyn_cast<StructType>(Ty)) { 779 // If we end up with an offset that isn't valid for this struct type, we 780 // can't re-form this GEP in a regular form, so bail out. The pointer 781 // operand likely went through casts that are necessary to make the GEP 782 // sensible. 783 const StructLayout &SL = *TD->getStructLayout(STy); 784 if (Offset.uge(SL.getSizeInBytes())) 785 break; 786 787 // Determine which field of the struct the offset points into. The 788 // getZExtValue is fine as we've already ensured that the offset is 789 // within the range representable by the StructLayout API. 790 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 791 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 792 ElIdx)); 793 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 794 Ty = STy->getTypeAtIndex(ElIdx); 795 } else { 796 // We've reached some non-indexable type. 797 break; 798 } 799 } while (Ty != ResultElementTy); 800 801 // If we haven't used up the entire offset by descending the static 802 // type, then the offset is pointing into the middle of an indivisible 803 // member, so we can't simplify it. 804 if (Offset != 0) 805 return 0; 806 807 // Create a GEP. 808 Constant *C = ConstantExpr::getGetElementPtr(Ptr, NewIdxs); 809 assert(C->getType()->getPointerElementType() == Ty && 810 "Computed GetElementPtr has unexpected type!"); 811 812 // If we ended up indexing a member with a type that doesn't match 813 // the type of what the original indices indexed, add a cast. 814 if (Ty != ResultElementTy) 815 C = FoldBitCast(C, ResultTy, *TD); 816 817 return C; 818} 819 820 821 822//===----------------------------------------------------------------------===// 823// Constant Folding public APIs 824//===----------------------------------------------------------------------===// 825 826/// ConstantFoldInstruction - Try to constant fold the specified instruction. 827/// If successful, the constant result is returned, if not, null is returned. 828/// Note that this fails if not all of the operands are constant. Otherwise, 829/// this function can only fail when attempting to fold instructions like loads 830/// and stores, which have no constant expression form. 831Constant *llvm::ConstantFoldInstruction(Instruction *I, 832 const DataLayout *TD, 833 const TargetLibraryInfo *TLI) { 834 // Handle PHI nodes quickly here... 835 if (PHINode *PN = dyn_cast<PHINode>(I)) { 836 Constant *CommonValue = 0; 837 838 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 839 Value *Incoming = PN->getIncomingValue(i); 840 // If the incoming value is undef then skip it. Note that while we could 841 // skip the value if it is equal to the phi node itself we choose not to 842 // because that would break the rule that constant folding only applies if 843 // all operands are constants. 844 if (isa<UndefValue>(Incoming)) 845 continue; 846 // If the incoming value is not a constant, then give up. 847 Constant *C = dyn_cast<Constant>(Incoming); 848 if (!C) 849 return 0; 850 // Fold the PHI's operands. 851 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C)) 852 C = ConstantFoldConstantExpression(NewC, TD, TLI); 853 // If the incoming value is a different constant to 854 // the one we saw previously, then give up. 855 if (CommonValue && C != CommonValue) 856 return 0; 857 CommonValue = C; 858 } 859 860 861 // If we reach here, all incoming values are the same constant or undef. 862 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 863 } 864 865 // Scan the operand list, checking to see if they are all constants, if so, 866 // hand off to ConstantFoldInstOperands. 867 SmallVector<Constant*, 8> Ops; 868 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) { 869 Constant *Op = dyn_cast<Constant>(*i); 870 if (!Op) 871 return 0; // All operands not constant! 872 873 // Fold the Instruction's operands. 874 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op)) 875 Op = ConstantFoldConstantExpression(NewCE, TD, TLI); 876 877 Ops.push_back(Op); 878 } 879 880 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 881 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 882 TD, TLI); 883 884 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 885 return ConstantFoldLoadInst(LI, TD); 886 887 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) { 888 return ConstantExpr::getInsertValue( 889 cast<Constant>(IVI->getAggregateOperand()), 890 cast<Constant>(IVI->getInsertedValueOperand()), 891 IVI->getIndices()); 892 } 893 894 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) { 895 return ConstantExpr::getExtractValue( 896 cast<Constant>(EVI->getAggregateOperand()), 897 EVI->getIndices()); 898 } 899 900 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI); 901} 902 903static Constant * 904ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout *TD, 905 const TargetLibraryInfo *TLI, 906 SmallPtrSet<ConstantExpr *, 4> &FoldedOps) { 907 SmallVector<Constant *, 8> Ops; 908 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; 909 ++i) { 910 Constant *NewC = cast<Constant>(*i); 911 // Recursively fold the ConstantExpr's operands. If we have already folded 912 // a ConstantExpr, we don't have to process it again. 913 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) { 914 if (FoldedOps.insert(NewCE)) 915 NewC = ConstantFoldConstantExpressionImpl(NewCE, TD, TLI, FoldedOps); 916 } 917 Ops.push_back(NewC); 918 } 919 920 if (CE->isCompare()) 921 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 922 TD, TLI); 923 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI); 924} 925 926/// ConstantFoldConstantExpression - Attempt to fold the constant expression 927/// using the specified DataLayout. If successful, the constant result is 928/// result is returned, if not, null is returned. 929Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 930 const DataLayout *TD, 931 const TargetLibraryInfo *TLI) { 932 SmallPtrSet<ConstantExpr *, 4> FoldedOps; 933 return ConstantFoldConstantExpressionImpl(CE, TD, TLI, FoldedOps); 934} 935 936/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 937/// specified opcode and operands. If successful, the constant result is 938/// returned, if not, null is returned. Note that this function can fail when 939/// attempting to fold instructions like loads and stores, which have no 940/// constant expression form. 941/// 942/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 943/// information, due to only being passed an opcode and operands. Constant 944/// folding using this function strips this information. 945/// 946Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, 947 ArrayRef<Constant *> Ops, 948 const DataLayout *TD, 949 const TargetLibraryInfo *TLI) { 950 // Handle easy binops first. 951 if (Instruction::isBinaryOp(Opcode)) { 952 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) { 953 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 954 return C; 955 } 956 957 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 958 } 959 960 switch (Opcode) { 961 default: return 0; 962 case Instruction::ICmp: 963 case Instruction::FCmp: llvm_unreachable("Invalid for compares"); 964 case Instruction::Call: 965 if (Function *F = dyn_cast<Function>(Ops.back())) 966 if (canConstantFoldCallTo(F)) 967 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI); 968 return 0; 969 case Instruction::PtrToInt: 970 // If the input is a inttoptr, eliminate the pair. This requires knowing 971 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 972 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 973 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 974 Constant *Input = CE->getOperand(0); 975 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 976 if (TD->getPointerTypeSizeInBits(CE->getType()) < InWidth) { 977 Constant *Mask = 978 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 979 TD->getPointerSizeInBits())); 980 Input = ConstantExpr::getAnd(Input, Mask); 981 } 982 // Do a zext or trunc to get to the dest size. 983 return ConstantExpr::getIntegerCast(Input, DestTy, false); 984 } 985 } 986 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 987 case Instruction::IntToPtr: 988 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 989 // the int size is >= the ptr size. This requires knowing the width of a 990 // pointer, so it can't be done in ConstantExpr::getCast. 991 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 992 if (TD && CE->getOpcode() == Instruction::PtrToInt) { 993 Constant *SrcPtr = CE->getOperand(0); 994 unsigned SrcPtrSize = TD->getPointerTypeSizeInBits(SrcPtr->getType()); 995 unsigned MidIntSize = CE->getType()->getScalarSizeInBits(); 996 997 if (MidIntSize >= SrcPtrSize) { 998 unsigned DestPtrSize = TD->getPointerTypeSizeInBits(DestTy); 999 if (SrcPtrSize == DestPtrSize) 1000 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 1001 } 1002 } 1003 } 1004 1005 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 1006 case Instruction::Trunc: 1007 case Instruction::ZExt: 1008 case Instruction::SExt: 1009 case Instruction::FPTrunc: 1010 case Instruction::FPExt: 1011 case Instruction::UIToFP: 1012 case Instruction::SIToFP: 1013 case Instruction::FPToUI: 1014 case Instruction::FPToSI: 1015 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 1016 case Instruction::BitCast: 1017 if (TD) 1018 return FoldBitCast(Ops[0], DestTy, *TD); 1019 return ConstantExpr::getBitCast(Ops[0], DestTy); 1020 case Instruction::Select: 1021 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 1022 case Instruction::ExtractElement: 1023 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 1024 case Instruction::InsertElement: 1025 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 1026 case Instruction::ShuffleVector: 1027 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 1028 case Instruction::GetElementPtr: 1029 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI)) 1030 return C; 1031 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI)) 1032 return C; 1033 1034 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1)); 1035 } 1036} 1037 1038/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 1039/// instruction (icmp/fcmp) with the specified operands. If it fails, it 1040/// returns a constant expression of the specified operands. 1041/// 1042Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 1043 Constant *Ops0, Constant *Ops1, 1044 const DataLayout *TD, 1045 const TargetLibraryInfo *TLI) { 1046 // fold: icmp (inttoptr x), null -> icmp x, 0 1047 // fold: icmp (ptrtoint x), 0 -> icmp x, null 1048 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 1049 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 1050 // 1051 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 1052 // around to know if bit truncation is happening. 1053 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 1054 if (TD && Ops1->isNullValue()) { 1055 if (CE0->getOpcode() == Instruction::IntToPtr) { 1056 Type *IntPtrTy = TD->getIntPtrType(CE0->getType()); 1057 // Convert the integer value to the right size to ensure we get the 1058 // proper extension or truncation. 1059 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 1060 IntPtrTy, false); 1061 Constant *Null = Constant::getNullValue(C->getType()); 1062 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); 1063 } 1064 1065 // Only do this transformation if the int is intptrty in size, otherwise 1066 // there is a truncation or extension that we aren't modeling. 1067 if (CE0->getOpcode() == Instruction::PtrToInt) { 1068 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType()); 1069 if (CE0->getType() == IntPtrTy) { 1070 Constant *C = CE0->getOperand(0); 1071 Constant *Null = Constant::getNullValue(C->getType()); 1072 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); 1073 } 1074 } 1075 } 1076 1077 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 1078 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 1079 if (CE0->getOpcode() == Instruction::IntToPtr) { 1080 Type *IntPtrTy = TD->getIntPtrType(CE0->getType()); 1081 1082 // Convert the integer value to the right size to ensure we get the 1083 // proper extension or truncation. 1084 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 1085 IntPtrTy, false); 1086 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 1087 IntPtrTy, false); 1088 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI); 1089 } 1090 1091 // Only do this transformation if the int is intptrty in size, otherwise 1092 // there is a truncation or extension that we aren't modeling. 1093 if (CE0->getOpcode() == Instruction::PtrToInt) { 1094 Type *IntPtrTy = TD->getIntPtrType(CE0->getOperand(0)->getType()); 1095 if (CE0->getType() == IntPtrTy && 1096 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) { 1097 return ConstantFoldCompareInstOperands(Predicate, 1098 CE0->getOperand(0), 1099 CE1->getOperand(0), 1100 TD, 1101 TLI); 1102 } 1103 } 1104 } 1105 } 1106 1107 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 1108 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 1109 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 1110 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 1111 Constant *LHS = 1112 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1, 1113 TD, TLI); 1114 Constant *RHS = 1115 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1, 1116 TD, TLI); 1117 unsigned OpC = 1118 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 1119 Constant *Ops[] = { LHS, RHS }; 1120 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI); 1121 } 1122 } 1123 1124 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 1125} 1126 1127 1128/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 1129/// getelementptr constantexpr, return the constant value being addressed by the 1130/// constant expression, or null if something is funny and we can't decide. 1131Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 1132 ConstantExpr *CE) { 1133 if (!CE->getOperand(1)->isNullValue()) 1134 return 0; // Do not allow stepping over the value! 1135 1136 // Loop over all of the operands, tracking down which value we are 1137 // addressing. 1138 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) { 1139 C = C->getAggregateElement(CE->getOperand(i)); 1140 if (C == 0) 1141 return 0; 1142 } 1143 return C; 1144} 1145 1146/// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr 1147/// indices (with an *implied* zero pointer index that is not in the list), 1148/// return the constant value being addressed by a virtual load, or null if 1149/// something is funny and we can't decide. 1150Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C, 1151 ArrayRef<Constant*> Indices) { 1152 // Loop over all of the operands, tracking down which value we are 1153 // addressing. 1154 for (unsigned i = 0, e = Indices.size(); i != e; ++i) { 1155 C = C->getAggregateElement(Indices[i]); 1156 if (C == 0) 1157 return 0; 1158 } 1159 return C; 1160} 1161 1162 1163//===----------------------------------------------------------------------===// 1164// Constant Folding for Calls 1165// 1166 1167/// canConstantFoldCallTo - Return true if its even possible to fold a call to 1168/// the specified function. 1169bool llvm::canConstantFoldCallTo(const Function *F) { 1170 switch (F->getIntrinsicID()) { 1171 case Intrinsic::fabs: 1172 case Intrinsic::log: 1173 case Intrinsic::log2: 1174 case Intrinsic::log10: 1175 case Intrinsic::exp: 1176 case Intrinsic::exp2: 1177 case Intrinsic::floor: 1178 case Intrinsic::sqrt: 1179 case Intrinsic::pow: 1180 case Intrinsic::powi: 1181 case Intrinsic::bswap: 1182 case Intrinsic::ctpop: 1183 case Intrinsic::ctlz: 1184 case Intrinsic::cttz: 1185 case Intrinsic::sadd_with_overflow: 1186 case Intrinsic::uadd_with_overflow: 1187 case Intrinsic::ssub_with_overflow: 1188 case Intrinsic::usub_with_overflow: 1189 case Intrinsic::smul_with_overflow: 1190 case Intrinsic::umul_with_overflow: 1191 case Intrinsic::convert_from_fp16: 1192 case Intrinsic::convert_to_fp16: 1193 case Intrinsic::x86_sse_cvtss2si: 1194 case Intrinsic::x86_sse_cvtss2si64: 1195 case Intrinsic::x86_sse_cvttss2si: 1196 case Intrinsic::x86_sse_cvttss2si64: 1197 case Intrinsic::x86_sse2_cvtsd2si: 1198 case Intrinsic::x86_sse2_cvtsd2si64: 1199 case Intrinsic::x86_sse2_cvttsd2si: 1200 case Intrinsic::x86_sse2_cvttsd2si64: 1201 return true; 1202 default: 1203 return false; 1204 case 0: break; 1205 } 1206 1207 if (!F->hasName()) 1208 return false; 1209 StringRef Name = F->getName(); 1210 1211 // In these cases, the check of the length is required. We don't want to 1212 // return true for a name like "cos\0blah" which strcmp would return equal to 1213 // "cos", but has length 8. 1214 switch (Name[0]) { 1215 default: return false; 1216 case 'a': 1217 return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2"; 1218 case 'c': 1219 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1220 case 'e': 1221 return Name == "exp" || Name == "exp2"; 1222 case 'f': 1223 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1224 case 'l': 1225 return Name == "log" || Name == "log10"; 1226 case 'p': 1227 return Name == "pow"; 1228 case 's': 1229 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1230 Name == "sinf" || Name == "sqrtf"; 1231 case 't': 1232 return Name == "tan" || Name == "tanh"; 1233 } 1234} 1235 1236static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1237 Type *Ty) { 1238 sys::llvm_fenv_clearexcept(); 1239 V = NativeFP(V); 1240 if (sys::llvm_fenv_testexcept()) { 1241 sys::llvm_fenv_clearexcept(); 1242 return 0; 1243 } 1244 1245 if (Ty->isHalfTy()) { 1246 APFloat APF(V); 1247 bool unused; 1248 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused); 1249 return ConstantFP::get(Ty->getContext(), APF); 1250 } 1251 if (Ty->isFloatTy()) 1252 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1253 if (Ty->isDoubleTy()) 1254 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1255 llvm_unreachable("Can only constant fold half/float/double"); 1256} 1257 1258static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1259 double V, double W, Type *Ty) { 1260 sys::llvm_fenv_clearexcept(); 1261 V = NativeFP(V, W); 1262 if (sys::llvm_fenv_testexcept()) { 1263 sys::llvm_fenv_clearexcept(); 1264 return 0; 1265 } 1266 1267 if (Ty->isHalfTy()) { 1268 APFloat APF(V); 1269 bool unused; 1270 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused); 1271 return ConstantFP::get(Ty->getContext(), APF); 1272 } 1273 if (Ty->isFloatTy()) 1274 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1275 if (Ty->isDoubleTy()) 1276 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1277 llvm_unreachable("Can only constant fold half/float/double"); 1278} 1279 1280/// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer 1281/// conversion of a constant floating point. If roundTowardZero is false, the 1282/// default IEEE rounding is used (toward nearest, ties to even). This matches 1283/// the behavior of the non-truncating SSE instructions in the default rounding 1284/// mode. The desired integer type Ty is used to select how many bits are 1285/// available for the result. Returns null if the conversion cannot be 1286/// performed, otherwise returns the Constant value resulting from the 1287/// conversion. 1288static Constant *ConstantFoldConvertToInt(const APFloat &Val, 1289 bool roundTowardZero, Type *Ty) { 1290 // All of these conversion intrinsics form an integer of at most 64bits. 1291 unsigned ResultWidth = Ty->getIntegerBitWidth(); 1292 assert(ResultWidth <= 64 && 1293 "Can only constant fold conversions to 64 and 32 bit ints"); 1294 1295 uint64_t UIntVal; 1296 bool isExact = false; 1297 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero 1298 : APFloat::rmNearestTiesToEven; 1299 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth, 1300 /*isSigned=*/true, mode, 1301 &isExact); 1302 if (status != APFloat::opOK && status != APFloat::opInexact) 1303 return 0; 1304 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true); 1305} 1306 1307/// ConstantFoldCall - Attempt to constant fold a call to the specified function 1308/// with the specified arguments, returning null if unsuccessful. 1309Constant * 1310llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands, 1311 const TargetLibraryInfo *TLI) { 1312 if (!F->hasName()) 1313 return 0; 1314 StringRef Name = F->getName(); 1315 1316 Type *Ty = F->getReturnType(); 1317 if (Operands.size() == 1) { 1318 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1319 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { 1320 APFloat Val(Op->getValueAPF()); 1321 1322 bool lost = false; 1323 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1324 1325 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1326 } 1327 if (!TLI) 1328 return 0; 1329 1330 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) 1331 return 0; 1332 1333 /// We only fold functions with finite arguments. Folding NaN and inf is 1334 /// likely to be aborted with an exception anyway, and some host libms 1335 /// have known errors raising exceptions. 1336 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1337 return 0; 1338 1339 /// Currently APFloat versions of these functions do not exist, so we use 1340 /// the host native double versions. Float versions are not called 1341 /// directly but for all these it is true (float)(f((double)arg)) == 1342 /// f(arg). Long double not supported yet. 1343 double V; 1344 if (Ty->isFloatTy()) 1345 V = Op->getValueAPF().convertToFloat(); 1346 else if (Ty->isDoubleTy()) 1347 V = Op->getValueAPF().convertToDouble(); 1348 else { 1349 bool unused; 1350 APFloat APF = Op->getValueAPF(); 1351 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused); 1352 V = APF.convertToDouble(); 1353 } 1354 1355 switch (F->getIntrinsicID()) { 1356 default: break; 1357 case Intrinsic::fabs: 1358 return ConstantFoldFP(fabs, V, Ty); 1359#if HAVE_LOG2 1360 case Intrinsic::log2: 1361 return ConstantFoldFP(log2, V, Ty); 1362#endif 1363#if HAVE_LOG 1364 case Intrinsic::log: 1365 return ConstantFoldFP(log, V, Ty); 1366#endif 1367#if HAVE_LOG10 1368 case Intrinsic::log10: 1369 return ConstantFoldFP(log10, V, Ty); 1370#endif 1371#if HAVE_EXP 1372 case Intrinsic::exp: 1373 return ConstantFoldFP(exp, V, Ty); 1374#endif 1375#if HAVE_EXP2 1376 case Intrinsic::exp2: 1377 return ConstantFoldFP(exp2, V, Ty); 1378#endif 1379 case Intrinsic::floor: 1380 return ConstantFoldFP(floor, V, Ty); 1381 } 1382 1383 switch (Name[0]) { 1384 case 'a': 1385 if (Name == "acos" && TLI->has(LibFunc::acos)) 1386 return ConstantFoldFP(acos, V, Ty); 1387 else if (Name == "asin" && TLI->has(LibFunc::asin)) 1388 return ConstantFoldFP(asin, V, Ty); 1389 else if (Name == "atan" && TLI->has(LibFunc::atan)) 1390 return ConstantFoldFP(atan, V, Ty); 1391 break; 1392 case 'c': 1393 if (Name == "ceil" && TLI->has(LibFunc::ceil)) 1394 return ConstantFoldFP(ceil, V, Ty); 1395 else if (Name == "cos" && TLI->has(LibFunc::cos)) 1396 return ConstantFoldFP(cos, V, Ty); 1397 else if (Name == "cosh" && TLI->has(LibFunc::cosh)) 1398 return ConstantFoldFP(cosh, V, Ty); 1399 else if (Name == "cosf" && TLI->has(LibFunc::cosf)) 1400 return ConstantFoldFP(cos, V, Ty); 1401 break; 1402 case 'e': 1403 if (Name == "exp" && TLI->has(LibFunc::exp)) 1404 return ConstantFoldFP(exp, V, Ty); 1405 1406 if (Name == "exp2" && TLI->has(LibFunc::exp2)) { 1407 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a 1408 // C99 library. 1409 return ConstantFoldBinaryFP(pow, 2.0, V, Ty); 1410 } 1411 break; 1412 case 'f': 1413 if (Name == "fabs" && TLI->has(LibFunc::fabs)) 1414 return ConstantFoldFP(fabs, V, Ty); 1415 else if (Name == "floor" && TLI->has(LibFunc::floor)) 1416 return ConstantFoldFP(floor, V, Ty); 1417 break; 1418 case 'l': 1419 if (Name == "log" && V > 0 && TLI->has(LibFunc::log)) 1420 return ConstantFoldFP(log, V, Ty); 1421 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10)) 1422 return ConstantFoldFP(log10, V, Ty); 1423 else if (F->getIntrinsicID() == Intrinsic::sqrt && 1424 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) { 1425 if (V >= -0.0) 1426 return ConstantFoldFP(sqrt, V, Ty); 1427 else // Undefined 1428 return Constant::getNullValue(Ty); 1429 } 1430 break; 1431 case 's': 1432 if (Name == "sin" && TLI->has(LibFunc::sin)) 1433 return ConstantFoldFP(sin, V, Ty); 1434 else if (Name == "sinh" && TLI->has(LibFunc::sinh)) 1435 return ConstantFoldFP(sinh, V, Ty); 1436 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt)) 1437 return ConstantFoldFP(sqrt, V, Ty); 1438 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf)) 1439 return ConstantFoldFP(sqrt, V, Ty); 1440 else if (Name == "sinf" && TLI->has(LibFunc::sinf)) 1441 return ConstantFoldFP(sin, V, Ty); 1442 break; 1443 case 't': 1444 if (Name == "tan" && TLI->has(LibFunc::tan)) 1445 return ConstantFoldFP(tan, V, Ty); 1446 else if (Name == "tanh" && TLI->has(LibFunc::tanh)) 1447 return ConstantFoldFP(tanh, V, Ty); 1448 break; 1449 default: 1450 break; 1451 } 1452 return 0; 1453 } 1454 1455 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1456 switch (F->getIntrinsicID()) { 1457 case Intrinsic::bswap: 1458 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1459 case Intrinsic::ctpop: 1460 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1461 case Intrinsic::convert_from_fp16: { 1462 APFloat Val(APFloat::IEEEhalf, Op->getValue()); 1463 1464 bool lost = false; 1465 APFloat::opStatus status = 1466 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1467 1468 // Conversion is always precise. 1469 (void)status; 1470 assert(status == APFloat::opOK && !lost && 1471 "Precision lost during fp16 constfolding"); 1472 1473 return ConstantFP::get(F->getContext(), Val); 1474 } 1475 default: 1476 return 0; 1477 } 1478 } 1479 1480 // Support ConstantVector in case we have an Undef in the top. 1481 if (isa<ConstantVector>(Operands[0]) || 1482 isa<ConstantDataVector>(Operands[0])) { 1483 Constant *Op = cast<Constant>(Operands[0]); 1484 switch (F->getIntrinsicID()) { 1485 default: break; 1486 case Intrinsic::x86_sse_cvtss2si: 1487 case Intrinsic::x86_sse_cvtss2si64: 1488 case Intrinsic::x86_sse2_cvtsd2si: 1489 case Intrinsic::x86_sse2_cvtsd2si64: 1490 if (ConstantFP *FPOp = 1491 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 1492 return ConstantFoldConvertToInt(FPOp->getValueAPF(), 1493 /*roundTowardZero=*/false, Ty); 1494 case Intrinsic::x86_sse_cvttss2si: 1495 case Intrinsic::x86_sse_cvttss2si64: 1496 case Intrinsic::x86_sse2_cvttsd2si: 1497 case Intrinsic::x86_sse2_cvttsd2si64: 1498 if (ConstantFP *FPOp = 1499 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 1500 return ConstantFoldConvertToInt(FPOp->getValueAPF(), 1501 /*roundTowardZero=*/true, Ty); 1502 } 1503 } 1504 1505 if (isa<UndefValue>(Operands[0])) { 1506 if (F->getIntrinsicID() == Intrinsic::bswap) 1507 return Operands[0]; 1508 return 0; 1509 } 1510 1511 return 0; 1512 } 1513 1514 if (Operands.size() == 2) { 1515 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1516 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) 1517 return 0; 1518 double Op1V; 1519 if (Ty->isFloatTy()) 1520 Op1V = Op1->getValueAPF().convertToFloat(); 1521 else if (Ty->isDoubleTy()) 1522 Op1V = Op1->getValueAPF().convertToDouble(); 1523 else { 1524 bool unused; 1525 APFloat APF = Op1->getValueAPF(); 1526 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused); 1527 Op1V = APF.convertToDouble(); 1528 } 1529 1530 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1531 if (Op2->getType() != Op1->getType()) 1532 return 0; 1533 1534 double Op2V; 1535 if (Ty->isFloatTy()) 1536 Op2V = Op2->getValueAPF().convertToFloat(); 1537 else if (Ty->isDoubleTy()) 1538 Op2V = Op2->getValueAPF().convertToDouble(); 1539 else { 1540 bool unused; 1541 APFloat APF = Op2->getValueAPF(); 1542 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused); 1543 Op2V = APF.convertToDouble(); 1544 } 1545 1546 if (F->getIntrinsicID() == Intrinsic::pow) { 1547 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1548 } 1549 if (!TLI) 1550 return 0; 1551 if (Name == "pow" && TLI->has(LibFunc::pow)) 1552 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1553 if (Name == "fmod" && TLI->has(LibFunc::fmod)) 1554 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1555 if (Name == "atan2" && TLI->has(LibFunc::atan2)) 1556 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1557 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1558 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isHalfTy()) 1559 return ConstantFP::get(F->getContext(), 1560 APFloat((float)std::pow((float)Op1V, 1561 (int)Op2C->getZExtValue()))); 1562 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) 1563 return ConstantFP::get(F->getContext(), 1564 APFloat((float)std::pow((float)Op1V, 1565 (int)Op2C->getZExtValue()))); 1566 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy()) 1567 return ConstantFP::get(F->getContext(), 1568 APFloat((double)std::pow((double)Op1V, 1569 (int)Op2C->getZExtValue()))); 1570 } 1571 return 0; 1572 } 1573 1574 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1575 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1576 switch (F->getIntrinsicID()) { 1577 default: break; 1578 case Intrinsic::sadd_with_overflow: 1579 case Intrinsic::uadd_with_overflow: 1580 case Intrinsic::ssub_with_overflow: 1581 case Intrinsic::usub_with_overflow: 1582 case Intrinsic::smul_with_overflow: 1583 case Intrinsic::umul_with_overflow: { 1584 APInt Res; 1585 bool Overflow; 1586 switch (F->getIntrinsicID()) { 1587 default: llvm_unreachable("Invalid case"); 1588 case Intrinsic::sadd_with_overflow: 1589 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); 1590 break; 1591 case Intrinsic::uadd_with_overflow: 1592 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); 1593 break; 1594 case Intrinsic::ssub_with_overflow: 1595 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); 1596 break; 1597 case Intrinsic::usub_with_overflow: 1598 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); 1599 break; 1600 case Intrinsic::smul_with_overflow: 1601 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); 1602 break; 1603 case Intrinsic::umul_with_overflow: 1604 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow); 1605 break; 1606 } 1607 Constant *Ops[] = { 1608 ConstantInt::get(F->getContext(), Res), 1609 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) 1610 }; 1611 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops); 1612 } 1613 case Intrinsic::cttz: 1614 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef. 1615 return UndefValue::get(Ty); 1616 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros()); 1617 case Intrinsic::ctlz: 1618 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef. 1619 return UndefValue::get(Ty); 1620 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros()); 1621 } 1622 } 1623 1624 return 0; 1625 } 1626 return 0; 1627 } 1628 return 0; 1629} 1630