ConstantFolding.cpp revision a6aac4c5bc22bb10c7adb11eee3f82c703af7002
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/Analysis/ValueTracking.h" 27#include "llvm/Target/TargetData.h" 28#include "llvm/ADT/SmallVector.h" 29#include "llvm/ADT/StringMap.h" 30#include "llvm/Support/ErrorHandling.h" 31#include "llvm/Support/GetElementPtrTypeIterator.h" 32#include "llvm/Support/MathExtras.h" 33#include <cerrno> 34#include <cmath> 35using namespace llvm; 36 37//===----------------------------------------------------------------------===// 38// Constant Folding internal helper functions 39//===----------------------------------------------------------------------===// 40 41/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with 42/// TargetData. This always returns a non-null constant, but it may be a 43/// ConstantExpr if unfoldable. 44static Constant *FoldBitCast(Constant *C, const Type *DestTy, 45 const TargetData &TD) { 46 47 // This only handles casts to vectors currently. 48 const VectorType *DestVTy = dyn_cast<VectorType>(DestTy); 49 if (DestVTy == 0) 50 return ConstantExpr::getBitCast(C, DestTy); 51 52 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 53 // vector so the code below can handle it uniformly. 54 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 55 Constant *Ops = C; // don't take the address of C! 56 return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD); 57 } 58 59 // If this is a bitcast from constant vector -> vector, fold it. 60 ConstantVector *CV = dyn_cast<ConstantVector>(C); 61 if (CV == 0) 62 return ConstantExpr::getBitCast(C, DestTy); 63 64 // If the element types match, VMCore can fold it. 65 unsigned NumDstElt = DestVTy->getNumElements(); 66 unsigned NumSrcElt = CV->getNumOperands(); 67 if (NumDstElt == NumSrcElt) 68 return ConstantExpr::getBitCast(C, DestTy); 69 70 const Type *SrcEltTy = CV->getType()->getElementType(); 71 const Type *DstEltTy = DestVTy->getElementType(); 72 73 // Otherwise, we're changing the number of elements in a vector, which 74 // requires endianness information to do the right thing. For example, 75 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 76 // folds to (little endian): 77 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 78 // and to (big endian): 79 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 80 81 // First thing is first. We only want to think about integer here, so if 82 // we have something in FP form, recast it as integer. 83 if (DstEltTy->isFloatingPointTy()) { 84 // Fold to an vector of integers with same size as our FP type. 85 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 86 const Type *DestIVTy = 87 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 88 // Recursively handle this integer conversion, if possible. 89 C = FoldBitCast(C, DestIVTy, TD); 90 if (!C) return ConstantExpr::getBitCast(C, DestTy); 91 92 // Finally, VMCore can handle this now that #elts line up. 93 return ConstantExpr::getBitCast(C, DestTy); 94 } 95 96 // Okay, we know the destination is integer, if the input is FP, convert 97 // it to integer first. 98 if (SrcEltTy->isFloatingPointTy()) { 99 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 100 const Type *SrcIVTy = 101 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 102 // Ask VMCore to do the conversion now that #elts line up. 103 C = ConstantExpr::getBitCast(C, SrcIVTy); 104 CV = dyn_cast<ConstantVector>(C); 105 if (!CV) // If VMCore wasn't able to fold it, bail out. 106 return C; 107 } 108 109 // Now we know that the input and output vectors are both integer vectors 110 // of the same size, and that their #elements is not the same. Do the 111 // conversion here, which depends on whether the input or output has 112 // more elements. 113 bool isLittleEndian = TD.isLittleEndian(); 114 115 SmallVector<Constant*, 32> Result; 116 if (NumDstElt < NumSrcElt) { 117 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 118 Constant *Zero = Constant::getNullValue(DstEltTy); 119 unsigned Ratio = NumSrcElt/NumDstElt; 120 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 121 unsigned SrcElt = 0; 122 for (unsigned i = 0; i != NumDstElt; ++i) { 123 // Build each element of the result. 124 Constant *Elt = Zero; 125 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 126 for (unsigned j = 0; j != Ratio; ++j) { 127 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++)); 128 if (!Src) // Reject constantexpr elements. 129 return ConstantExpr::getBitCast(C, DestTy); 130 131 // Zero extend the element to the right size. 132 Src = ConstantExpr::getZExt(Src, Elt->getType()); 133 134 // Shift it to the right place, depending on endianness. 135 Src = ConstantExpr::getShl(Src, 136 ConstantInt::get(Src->getType(), ShiftAmt)); 137 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 138 139 // Mix it in. 140 Elt = ConstantExpr::getOr(Elt, Src); 141 } 142 Result.push_back(Elt); 143 } 144 } else { 145 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 146 unsigned Ratio = NumDstElt/NumSrcElt; 147 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); 148 149 // Loop over each source value, expanding into multiple results. 150 for (unsigned i = 0; i != NumSrcElt; ++i) { 151 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i)); 152 if (!Src) // Reject constantexpr elements. 153 return ConstantExpr::getBitCast(C, DestTy); 154 155 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 156 for (unsigned j = 0; j != Ratio; ++j) { 157 // Shift the piece of the value into the right place, depending on 158 // endianness. 159 Constant *Elt = ConstantExpr::getLShr(Src, 160 ConstantInt::get(Src->getType(), ShiftAmt)); 161 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 162 163 // Truncate and remember this piece. 164 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 165 } 166 } 167 } 168 169 return ConstantVector::get(Result.data(), Result.size()); 170} 171 172 173/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset 174/// from a global, return the global and the constant. Because of 175/// constantexprs, this function is recursive. 176static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 177 int64_t &Offset, const TargetData &TD) { 178 // Trivial case, constant is the global. 179 if ((GV = dyn_cast<GlobalValue>(C))) { 180 Offset = 0; 181 return true; 182 } 183 184 // Otherwise, if this isn't a constant expr, bail out. 185 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 186 if (!CE) return false; 187 188 // Look through ptr->int and ptr->ptr casts. 189 if (CE->getOpcode() == Instruction::PtrToInt || 190 CE->getOpcode() == Instruction::BitCast) 191 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); 192 193 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 194 if (CE->getOpcode() == Instruction::GetElementPtr) { 195 // Cannot compute this if the element type of the pointer is missing size 196 // info. 197 if (!cast<PointerType>(CE->getOperand(0)->getType()) 198 ->getElementType()->isSized()) 199 return false; 200 201 // If the base isn't a global+constant, we aren't either. 202 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) 203 return false; 204 205 // Otherwise, add any offset that our operands provide. 206 gep_type_iterator GTI = gep_type_begin(CE); 207 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end(); 208 i != e; ++i, ++GTI) { 209 ConstantInt *CI = dyn_cast<ConstantInt>(*i); 210 if (!CI) return false; // Index isn't a simple constant? 211 if (CI->isZero()) continue; // Not adding anything. 212 213 if (const StructType *ST = dyn_cast<StructType>(*GTI)) { 214 // N = N + Offset 215 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); 216 } else { 217 const SequentialType *SQT = cast<SequentialType>(*GTI); 218 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); 219 } 220 } 221 return true; 222 } 223 224 return false; 225} 226 227/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the 228/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the 229/// pointer to copy results into and BytesLeft is the number of bytes left in 230/// the CurPtr buffer. TD is the target data. 231static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, 232 unsigned char *CurPtr, unsigned BytesLeft, 233 const TargetData &TD) { 234 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && 235 "Out of range access"); 236 237 // If this element is zero or undefined, we can just return since *CurPtr is 238 // zero initialized. 239 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 240 return true; 241 242 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 243 if (CI->getBitWidth() > 64 || 244 (CI->getBitWidth() & 7) != 0) 245 return false; 246 247 uint64_t Val = CI->getZExtValue(); 248 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 249 250 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 251 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8)); 252 ++ByteOffset; 253 } 254 return true; 255 } 256 257 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 258 if (CFP->getType()->isDoubleTy()) { 259 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); 260 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 261 } 262 if (CFP->getType()->isFloatTy()){ 263 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); 264 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 265 } 266 return false; 267 } 268 269 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 270 const StructLayout *SL = TD.getStructLayout(CS->getType()); 271 unsigned Index = SL->getElementContainingOffset(ByteOffset); 272 uint64_t CurEltOffset = SL->getElementOffset(Index); 273 ByteOffset -= CurEltOffset; 274 275 while (1) { 276 // If the element access is to the element itself and not to tail padding, 277 // read the bytes from the element. 278 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); 279 280 if (ByteOffset < EltSize && 281 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 282 BytesLeft, TD)) 283 return false; 284 285 ++Index; 286 287 // Check to see if we read from the last struct element, if so we're done. 288 if (Index == CS->getType()->getNumElements()) 289 return true; 290 291 // If we read all of the bytes we needed from this element we're done. 292 uint64_t NextEltOffset = SL->getElementOffset(Index); 293 294 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset) 295 return true; 296 297 // Move to the next element of the struct. 298 CurPtr += NextEltOffset-CurEltOffset-ByteOffset; 299 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset; 300 ByteOffset = 0; 301 CurEltOffset = NextEltOffset; 302 } 303 // not reached. 304 } 305 306 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) { 307 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType()); 308 uint64_t Index = ByteOffset / EltSize; 309 uint64_t Offset = ByteOffset - Index * EltSize; 310 for (; Index != CA->getType()->getNumElements(); ++Index) { 311 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr, 312 BytesLeft, TD)) 313 return false; 314 if (EltSize >= BytesLeft) 315 return true; 316 317 Offset = 0; 318 BytesLeft -= EltSize; 319 CurPtr += EltSize; 320 } 321 return true; 322 } 323 324 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) { 325 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType()); 326 uint64_t Index = ByteOffset / EltSize; 327 uint64_t Offset = ByteOffset - Index * EltSize; 328 for (; Index != CV->getType()->getNumElements(); ++Index) { 329 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr, 330 BytesLeft, TD)) 331 return false; 332 if (EltSize >= BytesLeft) 333 return true; 334 335 Offset = 0; 336 BytesLeft -= EltSize; 337 CurPtr += EltSize; 338 } 339 return true; 340 } 341 342 // Otherwise, unknown initializer type. 343 return false; 344} 345 346static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 347 const TargetData &TD) { 348 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType(); 349 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 350 351 // If this isn't an integer load we can't fold it directly. 352 if (!IntType) { 353 // If this is a float/double load, we can try folding it as an int32/64 load 354 // and then bitcast the result. This can be useful for union cases. Note 355 // that address spaces don't matter here since we're not going to result in 356 // an actual new load. 357 const Type *MapTy; 358 if (LoadTy->isFloatTy()) 359 MapTy = Type::getInt32PtrTy(C->getContext()); 360 else if (LoadTy->isDoubleTy()) 361 MapTy = Type::getInt64PtrTy(C->getContext()); 362 else if (LoadTy->isVectorTy()) { 363 MapTy = IntegerType::get(C->getContext(), 364 TD.getTypeAllocSizeInBits(LoadTy)); 365 MapTy = PointerType::getUnqual(MapTy); 366 } else 367 return 0; 368 369 C = FoldBitCast(C, MapTy, TD); 370 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 371 return FoldBitCast(Res, LoadTy, TD); 372 return 0; 373 } 374 375 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 376 if (BytesLoaded > 32 || BytesLoaded == 0) return 0; 377 378 GlobalValue *GVal; 379 int64_t Offset; 380 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 381 return 0; 382 383 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 384 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 385 !GV->getInitializer()->getType()->isSized()) 386 return 0; 387 388 // If we're loading off the beginning of the global, some bytes may be valid, 389 // but we don't try to handle this. 390 if (Offset < 0) return 0; 391 392 // If we're not accessing anything in this constant, the result is undefined. 393 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) 394 return UndefValue::get(IntType); 395 396 unsigned char RawBytes[32] = {0}; 397 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, 398 BytesLoaded, TD)) 399 return 0; 400 401 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); 402 for (unsigned i = 1; i != BytesLoaded; ++i) { 403 ResultVal <<= 8; 404 ResultVal |= RawBytes[BytesLoaded-1-i]; 405 } 406 407 return ConstantInt::get(IntType->getContext(), ResultVal); 408} 409 410/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 411/// produce if it is constant and determinable. If this is not determinable, 412/// return null. 413Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 414 const TargetData *TD) { 415 // First, try the easy cases: 416 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 417 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 418 return GV->getInitializer(); 419 420 // If the loaded value isn't a constant expr, we can't handle it. 421 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 422 if (!CE) return 0; 423 424 if (CE->getOpcode() == Instruction::GetElementPtr) { 425 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 426 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 427 if (Constant *V = 428 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 429 return V; 430 } 431 432 // Instead of loading constant c string, use corresponding integer value 433 // directly if string length is small enough. 434 std::string Str; 435 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) { 436 unsigned StrLen = Str.length(); 437 const Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 438 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 439 // Replace load with immediate integer if the result is an integer or fp 440 // value. 441 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 442 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 443 APInt StrVal(NumBits, 0); 444 APInt SingleChar(NumBits, 0); 445 if (TD->isLittleEndian()) { 446 for (signed i = StrLen-1; i >= 0; i--) { 447 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 448 StrVal = (StrVal << 8) | SingleChar; 449 } 450 } else { 451 for (unsigned i = 0; i < StrLen; i++) { 452 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 453 StrVal = (StrVal << 8) | SingleChar; 454 } 455 // Append NULL at the end. 456 SingleChar = 0; 457 StrVal = (StrVal << 8) | SingleChar; 458 } 459 460 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 461 if (Ty->isFloatingPointTy()) 462 Res = ConstantExpr::getBitCast(Res, Ty); 463 return Res; 464 } 465 } 466 467 // If this load comes from anywhere in a constant global, and if the global 468 // is all undef or zero, we know what it loads. 469 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){ 470 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 471 const Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 472 if (GV->getInitializer()->isNullValue()) 473 return Constant::getNullValue(ResTy); 474 if (isa<UndefValue>(GV->getInitializer())) 475 return UndefValue::get(ResTy); 476 } 477 } 478 479 // Try hard to fold loads from bitcasted strange and non-type-safe things. We 480 // currently don't do any of this for big endian systems. It can be 481 // generalized in the future if someone is interested. 482 if (TD && TD->isLittleEndian()) 483 return FoldReinterpretLoadFromConstPtr(CE, *TD); 484 return 0; 485} 486 487static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ 488 if (LI->isVolatile()) return 0; 489 490 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 491 return ConstantFoldLoadFromConstPtr(C, TD); 492 493 return 0; 494} 495 496/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 497/// Attempt to symbolically evaluate the result of a binary operator merging 498/// these together. If target data info is available, it is provided as TD, 499/// otherwise TD is null. 500static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 501 Constant *Op1, const TargetData *TD){ 502 // SROA 503 504 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 505 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 506 // bits. 507 508 509 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 510 // constant. This happens frequently when iterating over a global array. 511 if (Opc == Instruction::Sub && TD) { 512 GlobalValue *GV1, *GV2; 513 int64_t Offs1, Offs2; 514 515 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) 516 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && 517 GV1 == GV2) { 518 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 519 return ConstantInt::get(Op0->getType(), Offs1-Offs2); 520 } 521 } 522 523 return 0; 524} 525 526/// CastGEPIndices - If array indices are not pointer-sized integers, 527/// explicitly cast them so that they aren't implicitly casted by the 528/// getelementptr. 529static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, 530 const Type *ResultTy, 531 const TargetData *TD) { 532 if (!TD) return 0; 533 const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); 534 535 bool Any = false; 536 SmallVector<Constant*, 32> NewIdxs; 537 for (unsigned i = 1; i != NumOps; ++i) { 538 if ((i == 1 || 539 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(), 540 reinterpret_cast<Value *const *>(Ops+1), 541 i-1))) && 542 Ops[i]->getType() != IntPtrTy) { 543 Any = true; 544 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 545 true, 546 IntPtrTy, 547 true), 548 Ops[i], IntPtrTy)); 549 } else 550 NewIdxs.push_back(Ops[i]); 551 } 552 if (!Any) return 0; 553 554 Constant *C = 555 ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size()); 556 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 557 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD)) 558 C = Folded; 559 return C; 560} 561 562/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP 563/// constant expression, do so. 564static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, 565 const Type *ResultTy, 566 const TargetData *TD) { 567 Constant *Ptr = Ops[0]; 568 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized()) 569 return 0; 570 571 unsigned BitWidth = 572 TD->getTypeSizeInBits(TD->getIntPtrType(Ptr->getContext())); 573 574 // If this is a constant expr gep that is effectively computing an 575 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' 576 for (unsigned i = 1; i != NumOps; ++i) 577 if (!isa<ConstantInt>(Ops[i])) 578 return 0; 579 580 APInt Offset = APInt(BitWidth, 581 TD->getIndexedOffset(Ptr->getType(), 582 (Value**)Ops+1, NumOps-1)); 583 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 584 585 // If this is a GEP of a GEP, fold it all into a single GEP. 586 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 587 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end()); 588 589 // Do not try the incorporate the sub-GEP if some index is not a number. 590 bool AllConstantInt = true; 591 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 592 if (!isa<ConstantInt>(NestedOps[i])) { 593 AllConstantInt = false; 594 break; 595 } 596 if (!AllConstantInt) 597 break; 598 599 Ptr = cast<Constant>(GEP->getOperand(0)); 600 Offset += APInt(BitWidth, 601 TD->getIndexedOffset(Ptr->getType(), 602 (Value**)NestedOps.data(), 603 NestedOps.size())); 604 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 605 } 606 607 // If the base value for this address is a literal integer value, fold the 608 // getelementptr to the resulting integer value casted to the pointer type. 609 APInt BasePtr(BitWidth, 0); 610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 611 if (CE->getOpcode() == Instruction::IntToPtr) 612 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) { 613 BasePtr = Base->getValue(); 614 BasePtr.zextOrTrunc(BitWidth); 615 } 616 if (Ptr->isNullValue() || BasePtr != 0) { 617 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); 618 return ConstantExpr::getIntToPtr(C, ResultTy); 619 } 620 621 // Otherwise form a regular getelementptr. Recompute the indices so that 622 // we eliminate over-indexing of the notional static type array bounds. 623 // This makes it easy to determine if the getelementptr is "inbounds". 624 // Also, this helps GlobalOpt do SROA on GlobalVariables. 625 const Type *Ty = Ptr->getType(); 626 SmallVector<Constant*, 32> NewIdxs; 627 do { 628 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 629 if (ATy->isPointerTy()) { 630 // The only pointer indexing we'll do is on the first index of the GEP. 631 if (!NewIdxs.empty()) 632 break; 633 634 // Only handle pointers to sized types, not pointers to functions. 635 if (!ATy->getElementType()->isSized()) 636 return 0; 637 } 638 639 // Determine which element of the array the offset points into. 640 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 641 if (ElemSize == 0) 642 return 0; 643 APInt NewIdx = Offset.udiv(ElemSize); 644 Offset -= NewIdx * ElemSize; 645 NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Ty->getContext()), 646 NewIdx)); 647 Ty = ATy->getElementType(); 648 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) { 649 // Determine which field of the struct the offset points into. The 650 // getZExtValue is at least as safe as the StructLayout API because we 651 // know the offset is within the struct at this point. 652 const StructLayout &SL = *TD->getStructLayout(STy); 653 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 654 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 655 ElIdx)); 656 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 657 Ty = STy->getTypeAtIndex(ElIdx); 658 } else { 659 // We've reached some non-indexable type. 660 break; 661 } 662 } while (Ty != cast<PointerType>(ResultTy)->getElementType()); 663 664 // If we haven't used up the entire offset by descending the static 665 // type, then the offset is pointing into the middle of an indivisible 666 // member, so we can't simplify it. 667 if (Offset != 0) 668 return 0; 669 670 // Create a GEP. 671 Constant *C = 672 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()); 673 assert(cast<PointerType>(C->getType())->getElementType() == Ty && 674 "Computed GetElementPtr has unexpected type!"); 675 676 // If we ended up indexing a member with a type that doesn't match 677 // the type of what the original indices indexed, add a cast. 678 if (Ty != cast<PointerType>(ResultTy)->getElementType()) 679 C = FoldBitCast(C, ResultTy, *TD); 680 681 return C; 682} 683 684 685 686//===----------------------------------------------------------------------===// 687// Constant Folding public APIs 688//===----------------------------------------------------------------------===// 689 690 691/// ConstantFoldInstruction - Attempt to constant fold the specified 692/// instruction. If successful, the constant result is returned, if not, null 693/// is returned. Note that this function can only fail when attempting to fold 694/// instructions like loads and stores, which have no constant expression form. 695/// 696Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { 697 if (PHINode *PN = dyn_cast<PHINode>(I)) { 698 if (PN->getNumIncomingValues() == 0) 699 return UndefValue::get(PN->getType()); 700 701 Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0)); 702 if (Result == 0) return 0; 703 704 // Handle PHI nodes specially here... 705 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) 706 if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN) 707 return 0; // Not all the same incoming constants... 708 709 // If we reach here, all incoming values are the same constant. 710 return Result; 711 } 712 713 // Scan the operand list, checking to see if they are all constants, if so, 714 // hand off to ConstantFoldInstOperands. 715 SmallVector<Constant*, 8> Ops; 716 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 717 if (Constant *Op = dyn_cast<Constant>(*i)) 718 Ops.push_back(Op); 719 else 720 return 0; // All operands not constant! 721 722 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 723 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 724 TD); 725 726 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 727 return ConstantFoldLoadInst(LI, TD); 728 729 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), 730 Ops.data(), Ops.size(), TD); 731} 732 733/// ConstantFoldConstantExpression - Attempt to fold the constant expression 734/// using the specified TargetData. If successful, the constant result is 735/// result is returned, if not, null is returned. 736Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 737 const TargetData *TD) { 738 SmallVector<Constant*, 8> Ops; 739 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i) { 740 Constant *NewC = cast<Constant>(*i); 741 // Recursively fold the ConstantExpr's operands. 742 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) 743 NewC = ConstantFoldConstantExpression(NewCE, TD); 744 Ops.push_back(NewC); 745 } 746 747 if (CE->isCompare()) 748 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 749 TD); 750 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), 751 Ops.data(), Ops.size(), TD); 752} 753 754/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 755/// specified opcode and operands. If successful, the constant result is 756/// returned, if not, null is returned. Note that this function can fail when 757/// attempting to fold instructions like loads and stores, which have no 758/// constant expression form. 759/// 760/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 761/// information, due to only being passed an opcode and operands. Constant 762/// folding using this function strips this information. 763/// 764Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, 765 Constant* const* Ops, unsigned NumOps, 766 const TargetData *TD) { 767 // Handle easy binops first. 768 if (Instruction::isBinaryOp(Opcode)) { 769 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) 770 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 771 return C; 772 773 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 774 } 775 776 switch (Opcode) { 777 default: return 0; 778 case Instruction::ICmp: 779 case Instruction::FCmp: assert(0 && "Invalid for compares"); 780 case Instruction::Call: 781 if (Function *F = dyn_cast<Function>(Ops[NumOps - 1])) 782 if (canConstantFoldCallTo(F)) 783 return ConstantFoldCall(F, Ops, NumOps - 1); 784 return 0; 785 case Instruction::PtrToInt: 786 // If the input is a inttoptr, eliminate the pair. This requires knowing 787 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 788 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 789 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 790 Constant *Input = CE->getOperand(0); 791 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 792 if (TD->getPointerSizeInBits() < InWidth) { 793 Constant *Mask = 794 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 795 TD->getPointerSizeInBits())); 796 Input = ConstantExpr::getAnd(Input, Mask); 797 } 798 // Do a zext or trunc to get to the dest size. 799 return ConstantExpr::getIntegerCast(Input, DestTy, false); 800 } 801 } 802 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 803 case Instruction::IntToPtr: 804 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 805 // the int size is >= the ptr size. This requires knowing the width of a 806 // pointer, so it can't be done in ConstantExpr::getCast. 807 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) 808 if (TD && 809 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && 810 CE->getOpcode() == Instruction::PtrToInt) 811 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 812 813 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 814 case Instruction::Trunc: 815 case Instruction::ZExt: 816 case Instruction::SExt: 817 case Instruction::FPTrunc: 818 case Instruction::FPExt: 819 case Instruction::UIToFP: 820 case Instruction::SIToFP: 821 case Instruction::FPToUI: 822 case Instruction::FPToSI: 823 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 824 case Instruction::BitCast: 825 if (TD) 826 return FoldBitCast(Ops[0], DestTy, *TD); 827 return ConstantExpr::getBitCast(Ops[0], DestTy); 828 case Instruction::Select: 829 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 830 case Instruction::ExtractElement: 831 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 832 case Instruction::InsertElement: 833 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 834 case Instruction::ShuffleVector: 835 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 836 case Instruction::GetElementPtr: 837 if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD)) 838 return C; 839 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD)) 840 return C; 841 842 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1); 843 } 844} 845 846/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 847/// instruction (icmp/fcmp) with the specified operands. If it fails, it 848/// returns a constant expression of the specified operands. 849/// 850Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 851 Constant *Ops0, Constant *Ops1, 852 const TargetData *TD) { 853 // fold: icmp (inttoptr x), null -> icmp x, 0 854 // fold: icmp (ptrtoint x), 0 -> icmp x, null 855 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 856 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 857 // 858 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 859 // around to know if bit truncation is happening. 860 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 861 if (TD && Ops1->isNullValue()) { 862 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 863 if (CE0->getOpcode() == Instruction::IntToPtr) { 864 // Convert the integer value to the right size to ensure we get the 865 // proper extension or truncation. 866 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 867 IntPtrTy, false); 868 Constant *Null = Constant::getNullValue(C->getType()); 869 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 870 } 871 872 // Only do this transformation if the int is intptrty in size, otherwise 873 // there is a truncation or extension that we aren't modeling. 874 if (CE0->getOpcode() == Instruction::PtrToInt && 875 CE0->getType() == IntPtrTy) { 876 Constant *C = CE0->getOperand(0); 877 Constant *Null = Constant::getNullValue(C->getType()); 878 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 879 } 880 } 881 882 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 883 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 884 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 885 886 if (CE0->getOpcode() == Instruction::IntToPtr) { 887 // Convert the integer value to the right size to ensure we get the 888 // proper extension or truncation. 889 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 890 IntPtrTy, false); 891 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 892 IntPtrTy, false); 893 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); 894 } 895 896 // Only do this transformation if the int is intptrty in size, otherwise 897 // there is a truncation or extension that we aren't modeling. 898 if ((CE0->getOpcode() == Instruction::PtrToInt && 899 CE0->getType() == IntPtrTy && 900 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) 901 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), 902 CE1->getOperand(0), TD); 903 } 904 } 905 906 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 907 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 908 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 909 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 910 Constant *LHS = 911 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); 912 Constant *RHS = 913 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); 914 unsigned OpC = 915 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 916 Constant *Ops[] = { LHS, RHS }; 917 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD); 918 } 919 } 920 921 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 922} 923 924 925/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 926/// getelementptr constantexpr, return the constant value being addressed by the 927/// constant expression, or null if something is funny and we can't decide. 928Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 929 ConstantExpr *CE) { 930 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) 931 return 0; // Do not allow stepping over the value! 932 933 // Loop over all of the operands, tracking down which value we are 934 // addressing... 935 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 936 for (++I; I != E; ++I) 937 if (const StructType *STy = dyn_cast<StructType>(*I)) { 938 ConstantInt *CU = cast<ConstantInt>(I.getOperand()); 939 assert(CU->getZExtValue() < STy->getNumElements() && 940 "Struct index out of range!"); 941 unsigned El = (unsigned)CU->getZExtValue(); 942 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 943 C = CS->getOperand(El); 944 } else if (isa<ConstantAggregateZero>(C)) { 945 C = Constant::getNullValue(STy->getElementType(El)); 946 } else if (isa<UndefValue>(C)) { 947 C = UndefValue::get(STy->getElementType(El)); 948 } else { 949 return 0; 950 } 951 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 952 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) { 953 if (CI->getZExtValue() >= ATy->getNumElements()) 954 return 0; 955 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) 956 C = CA->getOperand(CI->getZExtValue()); 957 else if (isa<ConstantAggregateZero>(C)) 958 C = Constant::getNullValue(ATy->getElementType()); 959 else if (isa<UndefValue>(C)) 960 C = UndefValue::get(ATy->getElementType()); 961 else 962 return 0; 963 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) { 964 if (CI->getZExtValue() >= VTy->getNumElements()) 965 return 0; 966 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) 967 C = CP->getOperand(CI->getZExtValue()); 968 else if (isa<ConstantAggregateZero>(C)) 969 C = Constant::getNullValue(VTy->getElementType()); 970 else if (isa<UndefValue>(C)) 971 C = UndefValue::get(VTy->getElementType()); 972 else 973 return 0; 974 } else { 975 return 0; 976 } 977 } else { 978 return 0; 979 } 980 return C; 981} 982 983 984//===----------------------------------------------------------------------===// 985// Constant Folding for Calls 986// 987 988/// canConstantFoldCallTo - Return true if its even possible to fold a call to 989/// the specified function. 990bool 991llvm::canConstantFoldCallTo(const Function *F) { 992 switch (F->getIntrinsicID()) { 993 case Intrinsic::sqrt: 994 case Intrinsic::powi: 995 case Intrinsic::bswap: 996 case Intrinsic::ctpop: 997 case Intrinsic::ctlz: 998 case Intrinsic::cttz: 999 case Intrinsic::uadd_with_overflow: 1000 case Intrinsic::usub_with_overflow: 1001 case Intrinsic::sadd_with_overflow: 1002 case Intrinsic::ssub_with_overflow: 1003 case Intrinsic::convert_from_fp16: 1004 case Intrinsic::convert_to_fp16: 1005 return true; 1006 default: 1007 return false; 1008 case 0: break; 1009 } 1010 1011 if (!F->hasName()) return false; 1012 StringRef Name = F->getName(); 1013 1014 // In these cases, the check of the length is required. We don't want to 1015 // return true for a name like "cos\0blah" which strcmp would return equal to 1016 // "cos", but has length 8. 1017 switch (Name[0]) { 1018 default: return false; 1019 case 'a': 1020 return Name == "acos" || Name == "asin" || 1021 Name == "atan" || Name == "atan2"; 1022 case 'c': 1023 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1024 case 'e': 1025 return Name == "exp"; 1026 case 'f': 1027 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1028 case 'l': 1029 return Name == "log" || Name == "log10"; 1030 case 'p': 1031 return Name == "pow"; 1032 case 's': 1033 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1034 Name == "sinf" || Name == "sqrtf"; 1035 case 't': 1036 return Name == "tan" || Name == "tanh"; 1037 } 1038} 1039 1040static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1041 const Type *Ty) { 1042 errno = 0; 1043 V = NativeFP(V); 1044 if (errno != 0) { 1045 errno = 0; 1046 return 0; 1047 } 1048 1049 if (Ty->isFloatTy()) 1050 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1051 if (Ty->isDoubleTy()) 1052 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1053 llvm_unreachable("Can only constant fold float/double"); 1054 return 0; // dummy return to suppress warning 1055} 1056 1057static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1058 double V, double W, const Type *Ty) { 1059 errno = 0; 1060 V = NativeFP(V, W); 1061 if (errno != 0) { 1062 errno = 0; 1063 return 0; 1064 } 1065 1066 if (Ty->isFloatTy()) 1067 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1068 if (Ty->isDoubleTy()) 1069 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1070 llvm_unreachable("Can only constant fold float/double"); 1071 return 0; // dummy return to suppress warning 1072} 1073 1074/// ConstantFoldCall - Attempt to constant fold a call to the specified function 1075/// with the specified arguments, returning null if unsuccessful. 1076Constant * 1077llvm::ConstantFoldCall(Function *F, 1078 Constant *const *Operands, unsigned NumOperands) { 1079 if (!F->hasName()) return 0; 1080 StringRef Name = F->getName(); 1081 1082 const Type *Ty = F->getReturnType(); 1083 if (NumOperands == 1) { 1084 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1085 if (Name == "llvm.convert.to.fp16") { 1086 APFloat Val(Op->getValueAPF()); 1087 1088 bool lost = false; 1089 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1090 1091 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1092 } 1093 1094 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1095 return 0; 1096 /// Currently APFloat versions of these functions do not exist, so we use 1097 /// the host native double versions. Float versions are not called 1098 /// directly but for all these it is true (float)(f((double)arg)) == 1099 /// f(arg). Long double not supported yet. 1100 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : 1101 Op->getValueAPF().convertToDouble(); 1102 switch (Name[0]) { 1103 case 'a': 1104 if (Name == "acos") 1105 return ConstantFoldFP(acos, V, Ty); 1106 else if (Name == "asin") 1107 return ConstantFoldFP(asin, V, Ty); 1108 else if (Name == "atan") 1109 return ConstantFoldFP(atan, V, Ty); 1110 break; 1111 case 'c': 1112 if (Name == "ceil") 1113 return ConstantFoldFP(ceil, V, Ty); 1114 else if (Name == "cos") 1115 return ConstantFoldFP(cos, V, Ty); 1116 else if (Name == "cosh") 1117 return ConstantFoldFP(cosh, V, Ty); 1118 else if (Name == "cosf") 1119 return ConstantFoldFP(cos, V, Ty); 1120 break; 1121 case 'e': 1122 if (Name == "exp") 1123 return ConstantFoldFP(exp, V, Ty); 1124 break; 1125 case 'f': 1126 if (Name == "fabs") 1127 return ConstantFoldFP(fabs, V, Ty); 1128 else if (Name == "floor") 1129 return ConstantFoldFP(floor, V, Ty); 1130 break; 1131 case 'l': 1132 if (Name == "log" && V > 0) 1133 return ConstantFoldFP(log, V, Ty); 1134 else if (Name == "log10" && V > 0) 1135 return ConstantFoldFP(log10, V, Ty); 1136 else if (Name == "llvm.sqrt.f32" || 1137 Name == "llvm.sqrt.f64") { 1138 if (V >= -0.0) 1139 return ConstantFoldFP(sqrt, V, Ty); 1140 else // Undefined 1141 return Constant::getNullValue(Ty); 1142 } 1143 break; 1144 case 's': 1145 if (Name == "sin") 1146 return ConstantFoldFP(sin, V, Ty); 1147 else if (Name == "sinh") 1148 return ConstantFoldFP(sinh, V, Ty); 1149 else if (Name == "sqrt" && V >= 0) 1150 return ConstantFoldFP(sqrt, V, Ty); 1151 else if (Name == "sqrtf" && V >= 0) 1152 return ConstantFoldFP(sqrt, V, Ty); 1153 else if (Name == "sinf") 1154 return ConstantFoldFP(sin, V, Ty); 1155 break; 1156 case 't': 1157 if (Name == "tan") 1158 return ConstantFoldFP(tan, V, Ty); 1159 else if (Name == "tanh") 1160 return ConstantFoldFP(tanh, V, Ty); 1161 break; 1162 default: 1163 break; 1164 } 1165 return 0; 1166 } 1167 1168 1169 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1170 if (Name.startswith("llvm.bswap")) 1171 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1172 else if (Name.startswith("llvm.ctpop")) 1173 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1174 else if (Name.startswith("llvm.cttz")) 1175 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); 1176 else if (Name.startswith("llvm.ctlz")) 1177 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); 1178 else if (Name == "llvm.convert.from.fp16") { 1179 APFloat Val(Op->getValue()); 1180 1181 bool lost = false; 1182 APFloat::opStatus status = 1183 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1184 1185 // Conversion is always precise. 1186 status = status; 1187 assert(status == APFloat::opOK && !lost && 1188 "Precision lost during fp16 constfolding"); 1189 1190 return ConstantFP::get(F->getContext(), Val); 1191 } 1192 return 0; 1193 } 1194 1195 if (isa<UndefValue>(Operands[0])) { 1196 if (Name.startswith("llvm.bswap")) 1197 return Operands[0]; 1198 return 0; 1199 } 1200 1201 return 0; 1202 } 1203 1204 if (NumOperands == 2) { 1205 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1206 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1207 return 0; 1208 double Op1V = Ty->isFloatTy() ? 1209 (double)Op1->getValueAPF().convertToFloat() : 1210 Op1->getValueAPF().convertToDouble(); 1211 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1212 if (Op2->getType() != Op1->getType()) 1213 return 0; 1214 1215 double Op2V = Ty->isFloatTy() ? 1216 (double)Op2->getValueAPF().convertToFloat(): 1217 Op2->getValueAPF().convertToDouble(); 1218 1219 if (Name == "pow") 1220 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1221 if (Name == "fmod") 1222 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1223 if (Name == "atan2") 1224 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1225 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1226 if (Name == "llvm.powi.f32") 1227 return ConstantFP::get(F->getContext(), 1228 APFloat((float)std::pow((float)Op1V, 1229 (int)Op2C->getZExtValue()))); 1230 if (Name == "llvm.powi.f64") 1231 return ConstantFP::get(F->getContext(), 1232 APFloat((double)std::pow((double)Op1V, 1233 (int)Op2C->getZExtValue()))); 1234 } 1235 return 0; 1236 } 1237 1238 1239 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1240 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1241 switch (F->getIntrinsicID()) { 1242 default: break; 1243 case Intrinsic::uadd_with_overflow: { 1244 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result. 1245 Constant *Ops[] = { 1246 Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow. 1247 }; 1248 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1249 } 1250 case Intrinsic::usub_with_overflow: { 1251 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result. 1252 Constant *Ops[] = { 1253 Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow. 1254 }; 1255 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1256 } 1257 case Intrinsic::sadd_with_overflow: { 1258 Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result. 1259 Constant *Overflow = ConstantExpr::getSelect( 1260 ConstantExpr::getICmp(CmpInst::ICMP_SGT, 1261 ConstantInt::get(Op1->getType(), 0), Op1), 1262 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2), 1263 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow. 1264 1265 Constant *Ops[] = { Res, Overflow }; 1266 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1267 } 1268 case Intrinsic::ssub_with_overflow: { 1269 Constant *Res = ConstantExpr::getSub(Op1, Op2); // result. 1270 Constant *Overflow = ConstantExpr::getSelect( 1271 ConstantExpr::getICmp(CmpInst::ICMP_SGT, 1272 ConstantInt::get(Op2->getType(), 0), Op2), 1273 ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1), 1274 ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow. 1275 1276 Constant *Ops[] = { Res, Overflow }; 1277 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1278 } 1279 } 1280 } 1281 1282 return 0; 1283 } 1284 return 0; 1285 } 1286 return 0; 1287} 1288 1289