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 VMCore ConstantExpr simplifications, 13 // this file defines some additional folding routines that can make use of 14 // TargetData information. These functions cannot go in VMCore due to library 15 // dependency issues. 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/Analysis/ConstantFolding.h" 20 #include "llvm/Constants.h" 21 #include "llvm/DerivedTypes.h" 22 #include "llvm/Function.h" 23 #include "llvm/GlobalVariable.h" 24 #include "llvm/Instructions.h" 25 #include "llvm/Intrinsics.h" 26 #include "llvm/Operator.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/Target/TargetData.h" 29 #include "llvm/Target/TargetLibraryInfo.h" 30 #include "llvm/ADT/SmallVector.h" 31 #include "llvm/ADT/StringMap.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/GetElementPtrTypeIterator.h" 34 #include "llvm/Support/MathExtras.h" 35 #include "llvm/Support/FEnv.h" 36 #include <cerrno> 37 #include <cmath> 38 using namespace llvm; 39 40 //===----------------------------------------------------------------------===// 41 // Constant Folding internal helper functions 42 //===----------------------------------------------------------------------===// 43 44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with 45 /// TargetData. This always returns a non-null constant, but it may be a 46 /// ConstantExpr if unfoldable. 47 static Constant *FoldBitCast(Constant *C, Type *DestTy, 48 const TargetData &TD) { 49 // Catch the obvious splat cases. 50 if (C->isNullValue() && !DestTy->isX86_MMXTy()) 51 return Constant::getNullValue(DestTy); 52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy()) 53 return Constant::getAllOnesValue(DestTy); 54 55 // Handle a vector->integer cast. 56 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) { 57 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C); 58 if (CDV == 0) 59 return ConstantExpr::getBitCast(C, DestTy); 60 61 unsigned NumSrcElts = CDV->getType()->getNumElements(); 62 63 Type *SrcEltTy = CDV->getType()->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 VMCore to do the conversion now that #elts line up. 72 C = ConstantExpr::getBitCast(C, SrcIVTy); 73 CDV = cast<ConstantDataVector>(C); 74 } 75 76 // Now that we know that the input value is a vector of integers, just shift 77 // and insert them into our result. 78 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy); 79 APInt Result(IT->getBitWidth(), 0); 80 for (unsigned i = 0; i != NumSrcElts; ++i) { 81 Result <<= BitShift; 82 if (TD.isLittleEndian()) 83 Result |= CDV->getElementAsInteger(NumSrcElts-i-1); 84 else 85 Result |= CDV->getElementAsInteger(i); 86 } 87 88 return ConstantInt::get(IT, Result); 89 } 90 91 // The code below only handles casts to vectors currently. 92 VectorType *DestVTy = dyn_cast<VectorType>(DestTy); 93 if (DestVTy == 0) 94 return ConstantExpr::getBitCast(C, DestTy); 95 96 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 97 // vector so the code below can handle it uniformly. 98 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 99 Constant *Ops = C; // don't take the address of C! 100 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD); 101 } 102 103 // If this is a bitcast from constant vector -> vector, fold it. 104 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C)) 105 return ConstantExpr::getBitCast(C, DestTy); 106 107 // If the element types match, VMCore can fold it. 108 unsigned NumDstElt = DestVTy->getNumElements(); 109 unsigned NumSrcElt = C->getType()->getVectorNumElements(); 110 if (NumDstElt == NumSrcElt) 111 return ConstantExpr::getBitCast(C, DestTy); 112 113 Type *SrcEltTy = C->getType()->getVectorElementType(); 114 Type *DstEltTy = DestVTy->getElementType(); 115 116 // Otherwise, we're changing the number of elements in a vector, which 117 // requires endianness information to do the right thing. For example, 118 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 119 // folds to (little endian): 120 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 121 // and to (big endian): 122 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 123 124 // First thing is first. We only want to think about integer here, so if 125 // we have something in FP form, recast it as integer. 126 if (DstEltTy->isFloatingPointTy()) { 127 // Fold to an vector of integers with same size as our FP type. 128 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 129 Type *DestIVTy = 130 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 131 // Recursively handle this integer conversion, if possible. 132 C = FoldBitCast(C, DestIVTy, TD); 133 134 // Finally, VMCore can handle this now that #elts line up. 135 return ConstantExpr::getBitCast(C, DestTy); 136 } 137 138 // Okay, we know the destination is integer, if the input is FP, convert 139 // it to integer first. 140 if (SrcEltTy->isFloatingPointTy()) { 141 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 142 Type *SrcIVTy = 143 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 144 // Ask VMCore to do the conversion now that #elts line up. 145 C = ConstantExpr::getBitCast(C, SrcIVTy); 146 // If VMCore wasn't able to fold it, bail out. 147 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector. 148 !isa<ConstantDataVector>(C)) 149 return C; 150 } 151 152 // Now we know that the input and output vectors are both integer vectors 153 // of the same size, and that their #elements is not the same. Do the 154 // conversion here, which depends on whether the input or output has 155 // more elements. 156 bool isLittleEndian = TD.isLittleEndian(); 157 158 SmallVector<Constant*, 32> Result; 159 if (NumDstElt < NumSrcElt) { 160 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 161 Constant *Zero = Constant::getNullValue(DstEltTy); 162 unsigned Ratio = NumSrcElt/NumDstElt; 163 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 164 unsigned SrcElt = 0; 165 for (unsigned i = 0; i != NumDstElt; ++i) { 166 // Build each element of the result. 167 Constant *Elt = Zero; 168 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 169 for (unsigned j = 0; j != Ratio; ++j) { 170 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++)); 171 if (!Src) // Reject constantexpr elements. 172 return ConstantExpr::getBitCast(C, DestTy); 173 174 // Zero extend the element to the right size. 175 Src = ConstantExpr::getZExt(Src, Elt->getType()); 176 177 // Shift it to the right place, depending on endianness. 178 Src = ConstantExpr::getShl(Src, 179 ConstantInt::get(Src->getType(), ShiftAmt)); 180 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 181 182 // Mix it in. 183 Elt = ConstantExpr::getOr(Elt, Src); 184 } 185 Result.push_back(Elt); 186 } 187 return ConstantVector::get(Result); 188 } 189 190 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 191 unsigned Ratio = NumDstElt/NumSrcElt; 192 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); 193 194 // Loop over each source value, expanding into multiple results. 195 for (unsigned i = 0; i != NumSrcElt; ++i) { 196 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i)); 197 if (!Src) // Reject constantexpr elements. 198 return ConstantExpr::getBitCast(C, DestTy); 199 200 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 201 for (unsigned j = 0; j != Ratio; ++j) { 202 // Shift the piece of the value into the right place, depending on 203 // endianness. 204 Constant *Elt = ConstantExpr::getLShr(Src, 205 ConstantInt::get(Src->getType(), ShiftAmt)); 206 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 207 208 // Truncate and remember this piece. 209 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 210 } 211 } 212 213 return ConstantVector::get(Result); 214 } 215 216 217 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset 218 /// from a global, return the global and the constant. Because of 219 /// constantexprs, this function is recursive. 220 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 221 int64_t &Offset, const TargetData &TD) { 222 // Trivial case, constant is the global. 223 if ((GV = dyn_cast<GlobalValue>(C))) { 224 Offset = 0; 225 return true; 226 } 227 228 // Otherwise, if this isn't a constant expr, bail out. 229 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 230 if (!CE) return false; 231 232 // Look through ptr->int and ptr->ptr casts. 233 if (CE->getOpcode() == Instruction::PtrToInt || 234 CE->getOpcode() == Instruction::BitCast) 235 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); 236 237 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 238 if (CE->getOpcode() == Instruction::GetElementPtr) { 239 // Cannot compute this if the element type of the pointer is missing size 240 // info. 241 if (!cast<PointerType>(CE->getOperand(0)->getType()) 242 ->getElementType()->isSized()) 243 return false; 244 245 // If the base isn't a global+constant, we aren't either. 246 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) 247 return false; 248 249 // Otherwise, add any offset that our operands provide. 250 gep_type_iterator GTI = gep_type_begin(CE); 251 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end(); 252 i != e; ++i, ++GTI) { 253 ConstantInt *CI = dyn_cast<ConstantInt>(*i); 254 if (!CI) return false; // Index isn't a simple constant? 255 if (CI->isZero()) continue; // Not adding anything. 256 257 if (StructType *ST = dyn_cast<StructType>(*GTI)) { 258 // N = N + Offset 259 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); 260 } else { 261 SequentialType *SQT = cast<SequentialType>(*GTI); 262 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); 263 } 264 } 265 return true; 266 } 267 268 return false; 269 } 270 271 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the 272 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the 273 /// pointer to copy results into and BytesLeft is the number of bytes left in 274 /// the CurPtr buffer. TD is the target data. 275 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, 276 unsigned char *CurPtr, unsigned BytesLeft, 277 const TargetData &TD) { 278 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && 279 "Out of range access"); 280 281 // If this element is zero or undefined, we can just return since *CurPtr is 282 // zero initialized. 283 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 284 return true; 285 286 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 287 if (CI->getBitWidth() > 64 || 288 (CI->getBitWidth() & 7) != 0) 289 return false; 290 291 uint64_t Val = CI->getZExtValue(); 292 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 293 294 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 295 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8)); 296 ++ByteOffset; 297 } 298 return true; 299 } 300 301 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 302 if (CFP->getType()->isDoubleTy()) { 303 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); 304 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 305 } 306 if (CFP->getType()->isFloatTy()){ 307 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); 308 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 309 } 310 return false; 311 } 312 313 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 314 const StructLayout *SL = TD.getStructLayout(CS->getType()); 315 unsigned Index = SL->getElementContainingOffset(ByteOffset); 316 uint64_t CurEltOffset = SL->getElementOffset(Index); 317 ByteOffset -= CurEltOffset; 318 319 while (1) { 320 // If the element access is to the element itself and not to tail padding, 321 // read the bytes from the element. 322 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); 323 324 if (ByteOffset < EltSize && 325 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 326 BytesLeft, TD)) 327 return false; 328 329 ++Index; 330 331 // Check to see if we read from the last struct element, if so we're done. 332 if (Index == CS->getType()->getNumElements()) 333 return true; 334 335 // If we read all of the bytes we needed from this element we're done. 336 uint64_t NextEltOffset = SL->getElementOffset(Index); 337 338 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset) 339 return true; 340 341 // Move to the next element of the struct. 342 CurPtr += NextEltOffset-CurEltOffset-ByteOffset; 343 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset; 344 ByteOffset = 0; 345 CurEltOffset = NextEltOffset; 346 } 347 // not reached. 348 } 349 350 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) || 351 isa<ConstantDataSequential>(C)) { 352 Type *EltTy = cast<SequentialType>(C->getType())->getElementType(); 353 uint64_t EltSize = TD.getTypeAllocSize(EltTy); 354 uint64_t Index = ByteOffset / EltSize; 355 uint64_t Offset = ByteOffset - Index * EltSize; 356 uint64_t NumElts; 357 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType())) 358 NumElts = AT->getNumElements(); 359 else 360 NumElts = cast<VectorType>(C->getType())->getNumElements(); 361 362 for (; Index != NumElts; ++Index) { 363 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr, 364 BytesLeft, TD)) 365 return false; 366 if (EltSize >= BytesLeft) 367 return true; 368 369 Offset = 0; 370 BytesLeft -= EltSize; 371 CurPtr += EltSize; 372 } 373 return true; 374 } 375 376 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 377 if (CE->getOpcode() == Instruction::IntToPtr && 378 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext())) 379 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, 380 BytesLeft, TD); 381 } 382 383 // Otherwise, unknown initializer type. 384 return false; 385 } 386 387 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 388 const TargetData &TD) { 389 Type *LoadTy = cast<PointerType>(C->getType())->getElementType(); 390 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 391 392 // If this isn't an integer load we can't fold it directly. 393 if (!IntType) { 394 // If this is a float/double load, we can try folding it as an int32/64 load 395 // and then bitcast the result. This can be useful for union cases. Note 396 // that address spaces don't matter here since we're not going to result in 397 // an actual new load. 398 Type *MapTy; 399 if (LoadTy->isFloatTy()) 400 MapTy = Type::getInt32PtrTy(C->getContext()); 401 else if (LoadTy->isDoubleTy()) 402 MapTy = Type::getInt64PtrTy(C->getContext()); 403 else if (LoadTy->isVectorTy()) { 404 MapTy = IntegerType::get(C->getContext(), 405 TD.getTypeAllocSizeInBits(LoadTy)); 406 MapTy = PointerType::getUnqual(MapTy); 407 } else 408 return 0; 409 410 C = FoldBitCast(C, MapTy, TD); 411 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 412 return FoldBitCast(Res, LoadTy, TD); 413 return 0; 414 } 415 416 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 417 if (BytesLoaded > 32 || BytesLoaded == 0) return 0; 418 419 GlobalValue *GVal; 420 int64_t Offset; 421 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 422 return 0; 423 424 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 425 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 426 !GV->getInitializer()->getType()->isSized()) 427 return 0; 428 429 // If we're loading off the beginning of the global, some bytes may be valid, 430 // but we don't try to handle this. 431 if (Offset < 0) return 0; 432 433 // If we're not accessing anything in this constant, the result is undefined. 434 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) 435 return UndefValue::get(IntType); 436 437 unsigned char RawBytes[32] = {0}; 438 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, 439 BytesLoaded, TD)) 440 return 0; 441 442 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); 443 for (unsigned i = 1; i != BytesLoaded; ++i) { 444 ResultVal <<= 8; 445 ResultVal |= RawBytes[BytesLoaded-1-i]; 446 } 447 448 return ConstantInt::get(IntType->getContext(), ResultVal); 449 } 450 451 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 452 /// produce if it is constant and determinable. If this is not determinable, 453 /// return null. 454 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 455 const TargetData *TD) { 456 // First, try the easy cases: 457 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 458 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 459 return GV->getInitializer(); 460 461 // If the loaded value isn't a constant expr, we can't handle it. 462 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 463 if (!CE) return 0; 464 465 if (CE->getOpcode() == Instruction::GetElementPtr) { 466 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 467 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 468 if (Constant *V = 469 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 470 return V; 471 } 472 473 // Instead of loading constant c string, use corresponding integer value 474 // directly if string length is small enough. 475 StringRef Str; 476 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) { 477 unsigned StrLen = Str.size(); 478 Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 479 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 480 // Replace load with immediate integer if the result is an integer or fp 481 // value. 482 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 483 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 484 APInt StrVal(NumBits, 0); 485 APInt SingleChar(NumBits, 0); 486 if (TD->isLittleEndian()) { 487 for (signed i = StrLen-1; i >= 0; i--) { 488 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 489 StrVal = (StrVal << 8) | SingleChar; 490 } 491 } else { 492 for (unsigned i = 0; i < StrLen; i++) { 493 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 494 StrVal = (StrVal << 8) | SingleChar; 495 } 496 // Append NULL at the end. 497 SingleChar = 0; 498 StrVal = (StrVal << 8) | SingleChar; 499 } 500 501 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 502 if (Ty->isFloatingPointTy()) 503 Res = ConstantExpr::getBitCast(Res, Ty); 504 return Res; 505 } 506 } 507 508 // If this load comes from anywhere in a constant global, and if the global 509 // is all undef or zero, we know what it loads. 510 if (GlobalVariable *GV = 511 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) { 512 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 513 Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 514 if (GV->getInitializer()->isNullValue()) 515 return Constant::getNullValue(ResTy); 516 if (isa<UndefValue>(GV->getInitializer())) 517 return UndefValue::get(ResTy); 518 } 519 } 520 521 // Try hard to fold loads from bitcasted strange and non-type-safe things. We 522 // currently don't do any of this for big endian systems. It can be 523 // generalized in the future if someone is interested. 524 if (TD && TD->isLittleEndian()) 525 return FoldReinterpretLoadFromConstPtr(CE, *TD); 526 return 0; 527 } 528 529 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ 530 if (LI->isVolatile()) return 0; 531 532 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 533 return ConstantFoldLoadFromConstPtr(C, TD); 534 535 return 0; 536 } 537 538 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 539 /// Attempt to symbolically evaluate the result of a binary operator merging 540 /// these together. If target data info is available, it is provided as TD, 541 /// otherwise TD is null. 542 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 543 Constant *Op1, const TargetData *TD){ 544 // SROA 545 546 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 547 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 548 // bits. 549 550 551 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 552 // constant. This happens frequently when iterating over a global array. 553 if (Opc == Instruction::Sub && TD) { 554 GlobalValue *GV1, *GV2; 555 int64_t Offs1, Offs2; 556 557 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) 558 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && 559 GV1 == GV2) { 560 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 561 return ConstantInt::get(Op0->getType(), Offs1-Offs2); 562 } 563 } 564 565 return 0; 566 } 567 568 /// CastGEPIndices - If array indices are not pointer-sized integers, 569 /// explicitly cast them so that they aren't implicitly casted by the 570 /// getelementptr. 571 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops, 572 Type *ResultTy, const TargetData *TD, 573 const TargetLibraryInfo *TLI) { 574 if (!TD) return 0; 575 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); 576 577 bool Any = false; 578 SmallVector<Constant*, 32> NewIdxs; 579 for (unsigned i = 1, e = Ops.size(); i != e; ++i) { 580 if ((i == 1 || 581 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(), 582 Ops.slice(1, i-1)))) && 583 Ops[i]->getType() != IntPtrTy) { 584 Any = true; 585 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 586 true, 587 IntPtrTy, 588 true), 589 Ops[i], IntPtrTy)); 590 } else 591 NewIdxs.push_back(Ops[i]); 592 } 593 if (!Any) return 0; 594 595 Constant *C = 596 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs); 597 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 598 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) 599 C = Folded; 600 return C; 601 } 602 603 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP 604 /// constant expression, do so. 605 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops, 606 Type *ResultTy, const TargetData *TD, 607 const TargetLibraryInfo *TLI) { 608 Constant *Ptr = Ops[0]; 609 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() || 610 !Ptr->getType()->isPointerTy()) 611 return 0; 612 613 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext()); 614 615 // If this is a constant expr gep that is effectively computing an 616 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' 617 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 618 if (!isa<ConstantInt>(Ops[i])) { 619 620 // If this is "gep i8* Ptr, (sub 0, V)", fold this as: 621 // "inttoptr (sub (ptrtoint Ptr), V)" 622 if (Ops.size() == 2 && 623 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) { 624 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]); 625 assert((CE == 0 || CE->getType() == IntPtrTy) && 626 "CastGEPIndices didn't canonicalize index types!"); 627 if (CE && CE->getOpcode() == Instruction::Sub && 628 CE->getOperand(0)->isNullValue()) { 629 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType()); 630 Res = ConstantExpr::getSub(Res, CE->getOperand(1)); 631 Res = ConstantExpr::getIntToPtr(Res, ResultTy); 632 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res)) 633 Res = ConstantFoldConstantExpression(ResCE, TD, TLI); 634 return Res; 635 } 636 } 637 return 0; 638 } 639 640 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy); 641 APInt Offset = 642 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(), 643 makeArrayRef((Value **)Ops.data() + 1, 644 Ops.size() - 1))); 645 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 646 647 // If this is a GEP of a GEP, fold it all into a single GEP. 648 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 649 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end()); 650 651 // Do not try the incorporate the sub-GEP if some index is not a number. 652 bool AllConstantInt = true; 653 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 654 if (!isa<ConstantInt>(NestedOps[i])) { 655 AllConstantInt = false; 656 break; 657 } 658 if (!AllConstantInt) 659 break; 660 661 Ptr = cast<Constant>(GEP->getOperand(0)); 662 Offset += APInt(BitWidth, 663 TD->getIndexedOffset(Ptr->getType(), NestedOps)); 664 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 665 } 666 667 // If the base value for this address is a literal integer value, fold the 668 // getelementptr to the resulting integer value casted to the pointer type. 669 APInt BasePtr(BitWidth, 0); 670 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 671 if (CE->getOpcode() == Instruction::IntToPtr) 672 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) 673 BasePtr = Base->getValue().zextOrTrunc(BitWidth); 674 if (Ptr->isNullValue() || BasePtr != 0) { 675 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); 676 return ConstantExpr::getIntToPtr(C, ResultTy); 677 } 678 679 // Otherwise form a regular getelementptr. Recompute the indices so that 680 // we eliminate over-indexing of the notional static type array bounds. 681 // This makes it easy to determine if the getelementptr is "inbounds". 682 // Also, this helps GlobalOpt do SROA on GlobalVariables. 683 Type *Ty = Ptr->getType(); 684 SmallVector<Constant*, 32> NewIdxs; 685 do { 686 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 687 if (ATy->isPointerTy()) { 688 // The only pointer indexing we'll do is on the first index of the GEP. 689 if (!NewIdxs.empty()) 690 break; 691 692 // Only handle pointers to sized types, not pointers to functions. 693 if (!ATy->getElementType()->isSized()) 694 return 0; 695 } 696 697 // Determine which element of the array the offset points into. 698 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 699 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 700 if (ElemSize == 0) 701 // The element size is 0. This may be [0 x Ty]*, so just use a zero 702 // index for this level and proceed to the next level to see if it can 703 // accommodate the offset. 704 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 705 else { 706 // The element size is non-zero divide the offset by the element 707 // size (rounding down), to compute the index at this level. 708 APInt NewIdx = Offset.udiv(ElemSize); 709 Offset -= NewIdx * ElemSize; 710 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 711 } 712 Ty = ATy->getElementType(); 713 } else if (StructType *STy = dyn_cast<StructType>(Ty)) { 714 // Determine which field of the struct the offset points into. The 715 // getZExtValue is at least as safe as the StructLayout API because we 716 // know the offset is within the struct at this point. 717 const StructLayout &SL = *TD->getStructLayout(STy); 718 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 719 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 720 ElIdx)); 721 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 722 Ty = STy->getTypeAtIndex(ElIdx); 723 } else { 724 // We've reached some non-indexable type. 725 break; 726 } 727 } while (Ty != cast<PointerType>(ResultTy)->getElementType()); 728 729 // If we haven't used up the entire offset by descending the static 730 // type, then the offset is pointing into the middle of an indivisible 731 // member, so we can't simplify it. 732 if (Offset != 0) 733 return 0; 734 735 // Create a GEP. 736 Constant *C = 737 ConstantExpr::getGetElementPtr(Ptr, NewIdxs); 738 assert(cast<PointerType>(C->getType())->getElementType() == Ty && 739 "Computed GetElementPtr has unexpected type!"); 740 741 // If we ended up indexing a member with a type that doesn't match 742 // the type of what the original indices indexed, add a cast. 743 if (Ty != cast<PointerType>(ResultTy)->getElementType()) 744 C = FoldBitCast(C, ResultTy, *TD); 745 746 return C; 747 } 748 749 750 751 //===----------------------------------------------------------------------===// 752 // Constant Folding public APIs 753 //===----------------------------------------------------------------------===// 754 755 /// ConstantFoldInstruction - Try to constant fold the specified instruction. 756 /// If successful, the constant result is returned, if not, null is returned. 757 /// Note that this fails if not all of the operands are constant. Otherwise, 758 /// this function can only fail when attempting to fold instructions like loads 759 /// and stores, which have no constant expression form. 760 Constant *llvm::ConstantFoldInstruction(Instruction *I, 761 const TargetData *TD, 762 const TargetLibraryInfo *TLI) { 763 // Handle PHI nodes quickly here... 764 if (PHINode *PN = dyn_cast<PHINode>(I)) { 765 Constant *CommonValue = 0; 766 767 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 768 Value *Incoming = PN->getIncomingValue(i); 769 // If the incoming value is undef then skip it. Note that while we could 770 // skip the value if it is equal to the phi node itself we choose not to 771 // because that would break the rule that constant folding only applies if 772 // all operands are constants. 773 if (isa<UndefValue>(Incoming)) 774 continue; 775 // If the incoming value is not a constant, or is a different constant to 776 // the one we saw previously, then give up. 777 Constant *C = dyn_cast<Constant>(Incoming); 778 if (!C || (CommonValue && C != CommonValue)) 779 return 0; 780 CommonValue = C; 781 } 782 783 // If we reach here, all incoming values are the same constant or undef. 784 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 785 } 786 787 // Scan the operand list, checking to see if they are all constants, if so, 788 // hand off to ConstantFoldInstOperands. 789 SmallVector<Constant*, 8> Ops; 790 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 791 if (Constant *Op = dyn_cast<Constant>(*i)) 792 Ops.push_back(Op); 793 else 794 return 0; // All operands not constant! 795 796 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 797 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 798 TD, TLI); 799 800 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 801 return ConstantFoldLoadInst(LI, TD); 802 803 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) 804 return ConstantExpr::getInsertValue( 805 cast<Constant>(IVI->getAggregateOperand()), 806 cast<Constant>(IVI->getInsertedValueOperand()), 807 IVI->getIndices()); 808 809 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) 810 return ConstantExpr::getExtractValue( 811 cast<Constant>(EVI->getAggregateOperand()), 812 EVI->getIndices()); 813 814 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI); 815 } 816 817 /// ConstantFoldConstantExpression - Attempt to fold the constant expression 818 /// using the specified TargetData. If successful, the constant result is 819 /// result is returned, if not, null is returned. 820 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 821 const TargetData *TD, 822 const TargetLibraryInfo *TLI) { 823 SmallVector<Constant*, 8> Ops; 824 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); 825 i != e; ++i) { 826 Constant *NewC = cast<Constant>(*i); 827 // Recursively fold the ConstantExpr's operands. 828 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) 829 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI); 830 Ops.push_back(NewC); 831 } 832 833 if (CE->isCompare()) 834 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 835 TD, TLI); 836 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI); 837 } 838 839 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 840 /// specified opcode and operands. If successful, the constant result is 841 /// returned, if not, null is returned. Note that this function can fail when 842 /// attempting to fold instructions like loads and stores, which have no 843 /// constant expression form. 844 /// 845 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 846 /// information, due to only being passed an opcode and operands. Constant 847 /// folding using this function strips this information. 848 /// 849 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, 850 ArrayRef<Constant *> Ops, 851 const TargetData *TD, 852 const TargetLibraryInfo *TLI) { 853 // Handle easy binops first. 854 if (Instruction::isBinaryOp(Opcode)) { 855 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) 856 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 857 return C; 858 859 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 860 } 861 862 switch (Opcode) { 863 default: return 0; 864 case Instruction::ICmp: 865 case Instruction::FCmp: llvm_unreachable("Invalid for compares"); 866 case Instruction::Call: 867 if (Function *F = dyn_cast<Function>(Ops.back())) 868 if (canConstantFoldCallTo(F)) 869 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI); 870 return 0; 871 case Instruction::PtrToInt: 872 // If the input is a inttoptr, eliminate the pair. This requires knowing 873 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 874 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 875 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 876 Constant *Input = CE->getOperand(0); 877 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 878 if (TD->getPointerSizeInBits() < InWidth) { 879 Constant *Mask = 880 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 881 TD->getPointerSizeInBits())); 882 Input = ConstantExpr::getAnd(Input, Mask); 883 } 884 // Do a zext or trunc to get to the dest size. 885 return ConstantExpr::getIntegerCast(Input, DestTy, false); 886 } 887 } 888 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 889 case Instruction::IntToPtr: 890 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 891 // the int size is >= the ptr size. This requires knowing the width of a 892 // pointer, so it can't be done in ConstantExpr::getCast. 893 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) 894 if (TD && 895 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && 896 CE->getOpcode() == Instruction::PtrToInt) 897 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 898 899 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 900 case Instruction::Trunc: 901 case Instruction::ZExt: 902 case Instruction::SExt: 903 case Instruction::FPTrunc: 904 case Instruction::FPExt: 905 case Instruction::UIToFP: 906 case Instruction::SIToFP: 907 case Instruction::FPToUI: 908 case Instruction::FPToSI: 909 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 910 case Instruction::BitCast: 911 if (TD) 912 return FoldBitCast(Ops[0], DestTy, *TD); 913 return ConstantExpr::getBitCast(Ops[0], DestTy); 914 case Instruction::Select: 915 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 916 case Instruction::ExtractElement: 917 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 918 case Instruction::InsertElement: 919 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 920 case Instruction::ShuffleVector: 921 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 922 case Instruction::GetElementPtr: 923 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI)) 924 return C; 925 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI)) 926 return C; 927 928 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1)); 929 } 930 } 931 932 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 933 /// instruction (icmp/fcmp) with the specified operands. If it fails, it 934 /// returns a constant expression of the specified operands. 935 /// 936 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 937 Constant *Ops0, Constant *Ops1, 938 const TargetData *TD, 939 const TargetLibraryInfo *TLI) { 940 // fold: icmp (inttoptr x), null -> icmp x, 0 941 // fold: icmp (ptrtoint x), 0 -> icmp x, null 942 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 943 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 944 // 945 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 946 // around to know if bit truncation is happening. 947 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 948 if (TD && Ops1->isNullValue()) { 949 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 950 if (CE0->getOpcode() == Instruction::IntToPtr) { 951 // Convert the integer value to the right size to ensure we get the 952 // proper extension or truncation. 953 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 954 IntPtrTy, false); 955 Constant *Null = Constant::getNullValue(C->getType()); 956 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); 957 } 958 959 // Only do this transformation if the int is intptrty in size, otherwise 960 // there is a truncation or extension that we aren't modeling. 961 if (CE0->getOpcode() == Instruction::PtrToInt && 962 CE0->getType() == IntPtrTy) { 963 Constant *C = CE0->getOperand(0); 964 Constant *Null = Constant::getNullValue(C->getType()); 965 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI); 966 } 967 } 968 969 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 970 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 971 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 972 973 if (CE0->getOpcode() == Instruction::IntToPtr) { 974 // Convert the integer value to the right size to ensure we get the 975 // proper extension or truncation. 976 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 977 IntPtrTy, false); 978 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 979 IntPtrTy, false); 980 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI); 981 } 982 983 // Only do this transformation if the int is intptrty in size, otherwise 984 // there is a truncation or extension that we aren't modeling. 985 if ((CE0->getOpcode() == Instruction::PtrToInt && 986 CE0->getType() == IntPtrTy && 987 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) 988 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), 989 CE1->getOperand(0), TD, TLI); 990 } 991 } 992 993 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 994 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 995 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 996 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 997 Constant *LHS = 998 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1, 999 TD, TLI); 1000 Constant *RHS = 1001 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1, 1002 TD, TLI); 1003 unsigned OpC = 1004 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 1005 Constant *Ops[] = { LHS, RHS }; 1006 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI); 1007 } 1008 } 1009 1010 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 1011 } 1012 1013 1014 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 1015 /// getelementptr constantexpr, return the constant value being addressed by the 1016 /// constant expression, or null if something is funny and we can't decide. 1017 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 1018 ConstantExpr *CE) { 1019 if (!CE->getOperand(1)->isNullValue()) 1020 return 0; // Do not allow stepping over the value! 1021 1022 // Loop over all of the operands, tracking down which value we are 1023 // addressing. 1024 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) { 1025 C = C->getAggregateElement(CE->getOperand(i)); 1026 if (C == 0) return 0; 1027 } 1028 return C; 1029 } 1030 1031 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr 1032 /// indices (with an *implied* zero pointer index that is not in the list), 1033 /// return the constant value being addressed by a virtual load, or null if 1034 /// something is funny and we can't decide. 1035 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C, 1036 ArrayRef<Constant*> Indices) { 1037 // Loop over all of the operands, tracking down which value we are 1038 // addressing. 1039 for (unsigned i = 0, e = Indices.size(); i != e; ++i) { 1040 C = C->getAggregateElement(Indices[i]); 1041 if (C == 0) return 0; 1042 } 1043 return C; 1044 } 1045 1046 1047 //===----------------------------------------------------------------------===// 1048 // Constant Folding for Calls 1049 // 1050 1051 /// canConstantFoldCallTo - Return true if its even possible to fold a call to 1052 /// the specified function. 1053 bool 1054 llvm::canConstantFoldCallTo(const Function *F) { 1055 switch (F->getIntrinsicID()) { 1056 case Intrinsic::sqrt: 1057 case Intrinsic::pow: 1058 case Intrinsic::powi: 1059 case Intrinsic::bswap: 1060 case Intrinsic::ctpop: 1061 case Intrinsic::ctlz: 1062 case Intrinsic::cttz: 1063 case Intrinsic::sadd_with_overflow: 1064 case Intrinsic::uadd_with_overflow: 1065 case Intrinsic::ssub_with_overflow: 1066 case Intrinsic::usub_with_overflow: 1067 case Intrinsic::smul_with_overflow: 1068 case Intrinsic::umul_with_overflow: 1069 case Intrinsic::convert_from_fp16: 1070 case Intrinsic::convert_to_fp16: 1071 case Intrinsic::x86_sse_cvtss2si: 1072 case Intrinsic::x86_sse_cvtss2si64: 1073 case Intrinsic::x86_sse_cvttss2si: 1074 case Intrinsic::x86_sse_cvttss2si64: 1075 case Intrinsic::x86_sse2_cvtsd2si: 1076 case Intrinsic::x86_sse2_cvtsd2si64: 1077 case Intrinsic::x86_sse2_cvttsd2si: 1078 case Intrinsic::x86_sse2_cvttsd2si64: 1079 return true; 1080 default: 1081 return false; 1082 case 0: break; 1083 } 1084 1085 if (!F->hasName()) return false; 1086 StringRef Name = F->getName(); 1087 1088 // In these cases, the check of the length is required. We don't want to 1089 // return true for a name like "cos\0blah" which strcmp would return equal to 1090 // "cos", but has length 8. 1091 switch (Name[0]) { 1092 default: return false; 1093 case 'a': 1094 return Name == "acos" || Name == "asin" || 1095 Name == "atan" || Name == "atan2"; 1096 case 'c': 1097 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1098 case 'e': 1099 return Name == "exp" || Name == "exp2"; 1100 case 'f': 1101 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1102 case 'l': 1103 return Name == "log" || Name == "log10"; 1104 case 'p': 1105 return Name == "pow"; 1106 case 's': 1107 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1108 Name == "sinf" || Name == "sqrtf"; 1109 case 't': 1110 return Name == "tan" || Name == "tanh"; 1111 } 1112 } 1113 1114 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1115 Type *Ty) { 1116 sys::llvm_fenv_clearexcept(); 1117 V = NativeFP(V); 1118 if (sys::llvm_fenv_testexcept()) { 1119 sys::llvm_fenv_clearexcept(); 1120 return 0; 1121 } 1122 1123 if (Ty->isFloatTy()) 1124 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1125 if (Ty->isDoubleTy()) 1126 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1127 llvm_unreachable("Can only constant fold float/double"); 1128 } 1129 1130 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1131 double V, double W, Type *Ty) { 1132 sys::llvm_fenv_clearexcept(); 1133 V = NativeFP(V, W); 1134 if (sys::llvm_fenv_testexcept()) { 1135 sys::llvm_fenv_clearexcept(); 1136 return 0; 1137 } 1138 1139 if (Ty->isFloatTy()) 1140 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1141 if (Ty->isDoubleTy()) 1142 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1143 llvm_unreachable("Can only constant fold float/double"); 1144 } 1145 1146 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer 1147 /// conversion of a constant floating point. If roundTowardZero is false, the 1148 /// default IEEE rounding is used (toward nearest, ties to even). This matches 1149 /// the behavior of the non-truncating SSE instructions in the default rounding 1150 /// mode. The desired integer type Ty is used to select how many bits are 1151 /// available for the result. Returns null if the conversion cannot be 1152 /// performed, otherwise returns the Constant value resulting from the 1153 /// conversion. 1154 static Constant *ConstantFoldConvertToInt(const APFloat &Val, 1155 bool roundTowardZero, Type *Ty) { 1156 // All of these conversion intrinsics form an integer of at most 64bits. 1157 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth(); 1158 assert(ResultWidth <= 64 && 1159 "Can only constant fold conversions to 64 and 32 bit ints"); 1160 1161 uint64_t UIntVal; 1162 bool isExact = false; 1163 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero 1164 : APFloat::rmNearestTiesToEven; 1165 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth, 1166 /*isSigned=*/true, mode, 1167 &isExact); 1168 if (status != APFloat::opOK && status != APFloat::opInexact) 1169 return 0; 1170 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true); 1171 } 1172 1173 /// ConstantFoldCall - Attempt to constant fold a call to the specified function 1174 /// with the specified arguments, returning null if unsuccessful. 1175 Constant * 1176 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands, 1177 const TargetLibraryInfo *TLI) { 1178 if (!F->hasName()) return 0; 1179 StringRef Name = F->getName(); 1180 1181 Type *Ty = F->getReturnType(); 1182 if (Operands.size() == 1) { 1183 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1184 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { 1185 APFloat Val(Op->getValueAPF()); 1186 1187 bool lost = false; 1188 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1189 1190 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1191 } 1192 if (!TLI) 1193 return 0; 1194 1195 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1196 return 0; 1197 1198 /// We only fold functions with finite arguments. Folding NaN and inf is 1199 /// likely to be aborted with an exception anyway, and some host libms 1200 /// have known errors raising exceptions. 1201 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1202 return 0; 1203 1204 /// Currently APFloat versions of these functions do not exist, so we use 1205 /// the host native double versions. Float versions are not called 1206 /// directly but for all these it is true (float)(f((double)arg)) == 1207 /// f(arg). Long double not supported yet. 1208 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : 1209 Op->getValueAPF().convertToDouble(); 1210 switch (Name[0]) { 1211 case 'a': 1212 if (Name == "acos" && TLI->has(LibFunc::acos)) 1213 return ConstantFoldFP(acos, V, Ty); 1214 else if (Name == "asin" && TLI->has(LibFunc::asin)) 1215 return ConstantFoldFP(asin, V, Ty); 1216 else if (Name == "atan" && TLI->has(LibFunc::atan)) 1217 return ConstantFoldFP(atan, V, Ty); 1218 break; 1219 case 'c': 1220 if (Name == "ceil" && TLI->has(LibFunc::ceil)) 1221 return ConstantFoldFP(ceil, V, Ty); 1222 else if (Name == "cos" && TLI->has(LibFunc::cos)) 1223 return ConstantFoldFP(cos, V, Ty); 1224 else if (Name == "cosh" && TLI->has(LibFunc::cosh)) 1225 return ConstantFoldFP(cosh, V, Ty); 1226 else if (Name == "cosf" && TLI->has(LibFunc::cosf)) 1227 return ConstantFoldFP(cos, V, Ty); 1228 break; 1229 case 'e': 1230 if (Name == "exp" && TLI->has(LibFunc::exp)) 1231 return ConstantFoldFP(exp, V, Ty); 1232 1233 if (Name == "exp2" && TLI->has(LibFunc::exp2)) { 1234 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a 1235 // C99 library. 1236 return ConstantFoldBinaryFP(pow, 2.0, V, Ty); 1237 } 1238 break; 1239 case 'f': 1240 if (Name == "fabs" && TLI->has(LibFunc::fabs)) 1241 return ConstantFoldFP(fabs, V, Ty); 1242 else if (Name == "floor" && TLI->has(LibFunc::floor)) 1243 return ConstantFoldFP(floor, V, Ty); 1244 break; 1245 case 'l': 1246 if (Name == "log" && V > 0 && TLI->has(LibFunc::log)) 1247 return ConstantFoldFP(log, V, Ty); 1248 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10)) 1249 return ConstantFoldFP(log10, V, Ty); 1250 else if (F->getIntrinsicID() == Intrinsic::sqrt && 1251 (Ty->isFloatTy() || Ty->isDoubleTy())) { 1252 if (V >= -0.0) 1253 return ConstantFoldFP(sqrt, V, Ty); 1254 else // Undefined 1255 return Constant::getNullValue(Ty); 1256 } 1257 break; 1258 case 's': 1259 if (Name == "sin" && TLI->has(LibFunc::sin)) 1260 return ConstantFoldFP(sin, V, Ty); 1261 else if (Name == "sinh" && TLI->has(LibFunc::sinh)) 1262 return ConstantFoldFP(sinh, V, Ty); 1263 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt)) 1264 return ConstantFoldFP(sqrt, V, Ty); 1265 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf)) 1266 return ConstantFoldFP(sqrt, V, Ty); 1267 else if (Name == "sinf" && TLI->has(LibFunc::sinf)) 1268 return ConstantFoldFP(sin, V, Ty); 1269 break; 1270 case 't': 1271 if (Name == "tan" && TLI->has(LibFunc::tan)) 1272 return ConstantFoldFP(tan, V, Ty); 1273 else if (Name == "tanh" && TLI->has(LibFunc::tanh)) 1274 return ConstantFoldFP(tanh, V, Ty); 1275 break; 1276 default: 1277 break; 1278 } 1279 return 0; 1280 } 1281 1282 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1283 switch (F->getIntrinsicID()) { 1284 case Intrinsic::bswap: 1285 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1286 case Intrinsic::ctpop: 1287 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1288 case Intrinsic::convert_from_fp16: { 1289 APFloat Val(Op->getValue()); 1290 1291 bool lost = false; 1292 APFloat::opStatus status = 1293 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1294 1295 // Conversion is always precise. 1296 (void)status; 1297 assert(status == APFloat::opOK && !lost && 1298 "Precision lost during fp16 constfolding"); 1299 1300 return ConstantFP::get(F->getContext(), Val); 1301 } 1302 default: 1303 return 0; 1304 } 1305 } 1306 1307 // Support ConstantVector in case we have an Undef in the top. 1308 if (isa<ConstantVector>(Operands[0]) || 1309 isa<ConstantDataVector>(Operands[0])) { 1310 Constant *Op = cast<Constant>(Operands[0]); 1311 switch (F->getIntrinsicID()) { 1312 default: break; 1313 case Intrinsic::x86_sse_cvtss2si: 1314 case Intrinsic::x86_sse_cvtss2si64: 1315 case Intrinsic::x86_sse2_cvtsd2si: 1316 case Intrinsic::x86_sse2_cvtsd2si64: 1317 if (ConstantFP *FPOp = 1318 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 1319 return ConstantFoldConvertToInt(FPOp->getValueAPF(), 1320 /*roundTowardZero=*/false, Ty); 1321 case Intrinsic::x86_sse_cvttss2si: 1322 case Intrinsic::x86_sse_cvttss2si64: 1323 case Intrinsic::x86_sse2_cvttsd2si: 1324 case Intrinsic::x86_sse2_cvttsd2si64: 1325 if (ConstantFP *FPOp = 1326 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 1327 return ConstantFoldConvertToInt(FPOp->getValueAPF(), 1328 /*roundTowardZero=*/true, Ty); 1329 } 1330 } 1331 1332 if (isa<UndefValue>(Operands[0])) { 1333 if (F->getIntrinsicID() == Intrinsic::bswap) 1334 return Operands[0]; 1335 return 0; 1336 } 1337 1338 return 0; 1339 } 1340 1341 if (Operands.size() == 2) { 1342 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1343 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1344 return 0; 1345 double Op1V = Ty->isFloatTy() ? 1346 (double)Op1->getValueAPF().convertToFloat() : 1347 Op1->getValueAPF().convertToDouble(); 1348 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1349 if (Op2->getType() != Op1->getType()) 1350 return 0; 1351 1352 double Op2V = Ty->isFloatTy() ? 1353 (double)Op2->getValueAPF().convertToFloat(): 1354 Op2->getValueAPF().convertToDouble(); 1355 1356 if (F->getIntrinsicID() == Intrinsic::pow) { 1357 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1358 } 1359 if (!TLI) 1360 return 0; 1361 if (Name == "pow" && TLI->has(LibFunc::pow)) 1362 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1363 if (Name == "fmod" && TLI->has(LibFunc::fmod)) 1364 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1365 if (Name == "atan2" && TLI->has(LibFunc::atan2)) 1366 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1367 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1368 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) 1369 return ConstantFP::get(F->getContext(), 1370 APFloat((float)std::pow((float)Op1V, 1371 (int)Op2C->getZExtValue()))); 1372 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy()) 1373 return ConstantFP::get(F->getContext(), 1374 APFloat((double)std::pow((double)Op1V, 1375 (int)Op2C->getZExtValue()))); 1376 } 1377 return 0; 1378 } 1379 1380 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1381 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1382 switch (F->getIntrinsicID()) { 1383 default: break; 1384 case Intrinsic::sadd_with_overflow: 1385 case Intrinsic::uadd_with_overflow: 1386 case Intrinsic::ssub_with_overflow: 1387 case Intrinsic::usub_with_overflow: 1388 case Intrinsic::smul_with_overflow: 1389 case Intrinsic::umul_with_overflow: { 1390 APInt Res; 1391 bool Overflow; 1392 switch (F->getIntrinsicID()) { 1393 default: llvm_unreachable("Invalid case"); 1394 case Intrinsic::sadd_with_overflow: 1395 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); 1396 break; 1397 case Intrinsic::uadd_with_overflow: 1398 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); 1399 break; 1400 case Intrinsic::ssub_with_overflow: 1401 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); 1402 break; 1403 case Intrinsic::usub_with_overflow: 1404 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); 1405 break; 1406 case Intrinsic::smul_with_overflow: 1407 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); 1408 break; 1409 case Intrinsic::umul_with_overflow: 1410 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow); 1411 break; 1412 } 1413 Constant *Ops[] = { 1414 ConstantInt::get(F->getContext(), Res), 1415 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) 1416 }; 1417 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops); 1418 } 1419 case Intrinsic::cttz: 1420 // FIXME: This should check for Op2 == 1, and become unreachable if 1421 // Op1 == 0. 1422 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros()); 1423 case Intrinsic::ctlz: 1424 // FIXME: This should check for Op2 == 1, and become unreachable if 1425 // Op1 == 0. 1426 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros()); 1427 } 1428 } 1429 1430 return 0; 1431 } 1432 return 0; 1433 } 1434 return 0; 1435 } 1436