ConstantFolding.cpp revision 156eb0a569a1ebac86ad8438645f690d8a3894c4
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 "llvm/Support/FEnv.h" 34#include <cerrno> 35#include <cmath> 36using namespace llvm; 37 38//===----------------------------------------------------------------------===// 39// Constant Folding internal helper functions 40//===----------------------------------------------------------------------===// 41 42/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with 43/// TargetData. This always returns a non-null constant, but it may be a 44/// ConstantExpr if unfoldable. 45static Constant *FoldBitCast(Constant *C, const Type *DestTy, 46 const TargetData &TD) { 47 48 // This only handles casts to vectors currently. 49 const VectorType *DestVTy = dyn_cast<VectorType>(DestTy); 50 if (DestVTy == 0) 51 return ConstantExpr::getBitCast(C, DestTy); 52 53 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 54 // vector so the code below can handle it uniformly. 55 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 56 Constant *Ops = C; // don't take the address of C! 57 return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD); 58 } 59 60 // If this is a bitcast from constant vector -> vector, fold it. 61 ConstantVector *CV = dyn_cast<ConstantVector>(C); 62 if (CV == 0) 63 return ConstantExpr::getBitCast(C, DestTy); 64 65 // If the element types match, VMCore can fold it. 66 unsigned NumDstElt = DestVTy->getNumElements(); 67 unsigned NumSrcElt = CV->getNumOperands(); 68 if (NumDstElt == NumSrcElt) 69 return ConstantExpr::getBitCast(C, DestTy); 70 71 const Type *SrcEltTy = CV->getType()->getElementType(); 72 const Type *DstEltTy = DestVTy->getElementType(); 73 74 // Otherwise, we're changing the number of elements in a vector, which 75 // requires endianness information to do the right thing. For example, 76 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 77 // folds to (little endian): 78 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 79 // and to (big endian): 80 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 81 82 // First thing is first. We only want to think about integer here, so if 83 // we have something in FP form, recast it as integer. 84 if (DstEltTy->isFloatingPointTy()) { 85 // Fold to an vector of integers with same size as our FP type. 86 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 87 const Type *DestIVTy = 88 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 89 // Recursively handle this integer conversion, if possible. 90 C = FoldBitCast(C, DestIVTy, TD); 91 if (!C) return ConstantExpr::getBitCast(C, DestTy); 92 93 // Finally, VMCore can handle this now that #elts line up. 94 return ConstantExpr::getBitCast(C, DestTy); 95 } 96 97 // Okay, we know the destination is integer, if the input is FP, convert 98 // it to integer first. 99 if (SrcEltTy->isFloatingPointTy()) { 100 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 101 const Type *SrcIVTy = 102 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 103 // Ask VMCore to do the conversion now that #elts line up. 104 C = ConstantExpr::getBitCast(C, SrcIVTy); 105 CV = dyn_cast<ConstantVector>(C); 106 if (!CV) // If VMCore wasn't able to fold it, bail out. 107 return C; 108 } 109 110 // Now we know that the input and output vectors are both integer vectors 111 // of the same size, and that their #elements is not the same. Do the 112 // conversion here, which depends on whether the input or output has 113 // more elements. 114 bool isLittleEndian = TD.isLittleEndian(); 115 116 SmallVector<Constant*, 32> Result; 117 if (NumDstElt < NumSrcElt) { 118 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 119 Constant *Zero = Constant::getNullValue(DstEltTy); 120 unsigned Ratio = NumSrcElt/NumDstElt; 121 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 122 unsigned SrcElt = 0; 123 for (unsigned i = 0; i != NumDstElt; ++i) { 124 // Build each element of the result. 125 Constant *Elt = Zero; 126 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 127 for (unsigned j = 0; j != Ratio; ++j) { 128 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++)); 129 if (!Src) // Reject constantexpr elements. 130 return ConstantExpr::getBitCast(C, DestTy); 131 132 // Zero extend the element to the right size. 133 Src = ConstantExpr::getZExt(Src, Elt->getType()); 134 135 // Shift it to the right place, depending on endianness. 136 Src = ConstantExpr::getShl(Src, 137 ConstantInt::get(Src->getType(), ShiftAmt)); 138 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 139 140 // Mix it in. 141 Elt = ConstantExpr::getOr(Elt, Src); 142 } 143 Result.push_back(Elt); 144 } 145 } else { 146 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 147 unsigned Ratio = NumDstElt/NumSrcElt; 148 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); 149 150 // Loop over each source value, expanding into multiple results. 151 for (unsigned i = 0; i != NumSrcElt; ++i) { 152 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i)); 153 if (!Src) // Reject constantexpr elements. 154 return ConstantExpr::getBitCast(C, DestTy); 155 156 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 157 for (unsigned j = 0; j != Ratio; ++j) { 158 // Shift the piece of the value into the right place, depending on 159 // endianness. 160 Constant *Elt = ConstantExpr::getLShr(Src, 161 ConstantInt::get(Src->getType(), ShiftAmt)); 162 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 163 164 // Truncate and remember this piece. 165 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 166 } 167 } 168 } 169 170 return ConstantVector::get(Result.data(), Result.size()); 171} 172 173 174/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset 175/// from a global, return the global and the constant. Because of 176/// constantexprs, this function is recursive. 177static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 178 int64_t &Offset, const TargetData &TD) { 179 // Trivial case, constant is the global. 180 if ((GV = dyn_cast<GlobalValue>(C))) { 181 Offset = 0; 182 return true; 183 } 184 185 // Otherwise, if this isn't a constant expr, bail out. 186 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 187 if (!CE) return false; 188 189 // Look through ptr->int and ptr->ptr casts. 190 if (CE->getOpcode() == Instruction::PtrToInt || 191 CE->getOpcode() == Instruction::BitCast) 192 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); 193 194 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 195 if (CE->getOpcode() == Instruction::GetElementPtr) { 196 // Cannot compute this if the element type of the pointer is missing size 197 // info. 198 if (!cast<PointerType>(CE->getOperand(0)->getType()) 199 ->getElementType()->isSized()) 200 return false; 201 202 // If the base isn't a global+constant, we aren't either. 203 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) 204 return false; 205 206 // Otherwise, add any offset that our operands provide. 207 gep_type_iterator GTI = gep_type_begin(CE); 208 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end(); 209 i != e; ++i, ++GTI) { 210 ConstantInt *CI = dyn_cast<ConstantInt>(*i); 211 if (!CI) return false; // Index isn't a simple constant? 212 if (CI->isZero()) continue; // Not adding anything. 213 214 if (const StructType *ST = dyn_cast<StructType>(*GTI)) { 215 // N = N + Offset 216 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); 217 } else { 218 const SequentialType *SQT = cast<SequentialType>(*GTI); 219 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); 220 } 221 } 222 return true; 223 } 224 225 return false; 226} 227 228/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the 229/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the 230/// pointer to copy results into and BytesLeft is the number of bytes left in 231/// the CurPtr buffer. TD is the target data. 232static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, 233 unsigned char *CurPtr, unsigned BytesLeft, 234 const TargetData &TD) { 235 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && 236 "Out of range access"); 237 238 // If this element is zero or undefined, we can just return since *CurPtr is 239 // zero initialized. 240 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 241 return true; 242 243 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 244 if (CI->getBitWidth() > 64 || 245 (CI->getBitWidth() & 7) != 0) 246 return false; 247 248 uint64_t Val = CI->getZExtValue(); 249 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 250 251 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 252 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8)); 253 ++ByteOffset; 254 } 255 return true; 256 } 257 258 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 259 if (CFP->getType()->isDoubleTy()) { 260 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); 261 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 262 } 263 if (CFP->getType()->isFloatTy()){ 264 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); 265 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 266 } 267 return false; 268 } 269 270 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 271 const StructLayout *SL = TD.getStructLayout(CS->getType()); 272 unsigned Index = SL->getElementContainingOffset(ByteOffset); 273 uint64_t CurEltOffset = SL->getElementOffset(Index); 274 ByteOffset -= CurEltOffset; 275 276 while (1) { 277 // If the element access is to the element itself and not to tail padding, 278 // read the bytes from the element. 279 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); 280 281 if (ByteOffset < EltSize && 282 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 283 BytesLeft, TD)) 284 return false; 285 286 ++Index; 287 288 // Check to see if we read from the last struct element, if so we're done. 289 if (Index == CS->getType()->getNumElements()) 290 return true; 291 292 // If we read all of the bytes we needed from this element we're done. 293 uint64_t NextEltOffset = SL->getElementOffset(Index); 294 295 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset) 296 return true; 297 298 // Move to the next element of the struct. 299 CurPtr += NextEltOffset-CurEltOffset-ByteOffset; 300 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset; 301 ByteOffset = 0; 302 CurEltOffset = NextEltOffset; 303 } 304 // not reached. 305 } 306 307 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) { 308 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType()); 309 uint64_t Index = ByteOffset / EltSize; 310 uint64_t Offset = ByteOffset - Index * EltSize; 311 for (; Index != CA->getType()->getNumElements(); ++Index) { 312 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr, 313 BytesLeft, TD)) 314 return false; 315 if (EltSize >= BytesLeft) 316 return true; 317 318 Offset = 0; 319 BytesLeft -= EltSize; 320 CurPtr += EltSize; 321 } 322 return true; 323 } 324 325 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) { 326 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType()); 327 uint64_t Index = ByteOffset / EltSize; 328 uint64_t Offset = ByteOffset - Index * EltSize; 329 for (; Index != CV->getType()->getNumElements(); ++Index) { 330 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr, 331 BytesLeft, TD)) 332 return false; 333 if (EltSize >= BytesLeft) 334 return true; 335 336 Offset = 0; 337 BytesLeft -= EltSize; 338 CurPtr += EltSize; 339 } 340 return true; 341 } 342 343 // Otherwise, unknown initializer type. 344 return false; 345} 346 347static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 348 const TargetData &TD) { 349 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType(); 350 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 351 352 // If this isn't an integer load we can't fold it directly. 353 if (!IntType) { 354 // If this is a float/double load, we can try folding it as an int32/64 load 355 // and then bitcast the result. This can be useful for union cases. Note 356 // that address spaces don't matter here since we're not going to result in 357 // an actual new load. 358 const Type *MapTy; 359 if (LoadTy->isFloatTy()) 360 MapTy = Type::getInt32PtrTy(C->getContext()); 361 else if (LoadTy->isDoubleTy()) 362 MapTy = Type::getInt64PtrTy(C->getContext()); 363 else if (LoadTy->isVectorTy()) { 364 MapTy = IntegerType::get(C->getContext(), 365 TD.getTypeAllocSizeInBits(LoadTy)); 366 MapTy = PointerType::getUnqual(MapTy); 367 } else 368 return 0; 369 370 C = FoldBitCast(C, MapTy, TD); 371 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 372 return FoldBitCast(Res, LoadTy, TD); 373 return 0; 374 } 375 376 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 377 if (BytesLoaded > 32 || BytesLoaded == 0) return 0; 378 379 GlobalValue *GVal; 380 int64_t Offset; 381 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 382 return 0; 383 384 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 385 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 386 !GV->getInitializer()->getType()->isSized()) 387 return 0; 388 389 // If we're loading off the beginning of the global, some bytes may be valid, 390 // but we don't try to handle this. 391 if (Offset < 0) return 0; 392 393 // If we're not accessing anything in this constant, the result is undefined. 394 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) 395 return UndefValue::get(IntType); 396 397 unsigned char RawBytes[32] = {0}; 398 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, 399 BytesLoaded, TD)) 400 return 0; 401 402 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); 403 for (unsigned i = 1; i != BytesLoaded; ++i) { 404 ResultVal <<= 8; 405 ResultVal |= RawBytes[BytesLoaded-1-i]; 406 } 407 408 return ConstantInt::get(IntType->getContext(), ResultVal); 409} 410 411/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 412/// produce if it is constant and determinable. If this is not determinable, 413/// return null. 414Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 415 const TargetData *TD) { 416 // First, try the easy cases: 417 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 418 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 419 return GV->getInitializer(); 420 421 // If the loaded value isn't a constant expr, we can't handle it. 422 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 423 if (!CE) return 0; 424 425 if (CE->getOpcode() == Instruction::GetElementPtr) { 426 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 427 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 428 if (Constant *V = 429 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 430 return V; 431 } 432 433 // Instead of loading constant c string, use corresponding integer value 434 // directly if string length is small enough. 435 std::string Str; 436 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) { 437 unsigned StrLen = Str.length(); 438 const Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 439 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 440 // Replace load with immediate integer if the result is an integer or fp 441 // value. 442 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 443 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 444 APInt StrVal(NumBits, 0); 445 APInt SingleChar(NumBits, 0); 446 if (TD->isLittleEndian()) { 447 for (signed i = StrLen-1; i >= 0; i--) { 448 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 449 StrVal = (StrVal << 8) | SingleChar; 450 } 451 } else { 452 for (unsigned i = 0; i < StrLen; i++) { 453 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 454 StrVal = (StrVal << 8) | SingleChar; 455 } 456 // Append NULL at the end. 457 SingleChar = 0; 458 StrVal = (StrVal << 8) | SingleChar; 459 } 460 461 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 462 if (Ty->isFloatingPointTy()) 463 Res = ConstantExpr::getBitCast(Res, Ty); 464 return Res; 465 } 466 } 467 468 // If this load comes from anywhere in a constant global, and if the global 469 // is all undef or zero, we know what it loads. 470 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE))){ 471 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 472 const Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 473 if (GV->getInitializer()->isNullValue()) 474 return Constant::getNullValue(ResTy); 475 if (isa<UndefValue>(GV->getInitializer())) 476 return UndefValue::get(ResTy); 477 } 478 } 479 480 // Try hard to fold loads from bitcasted strange and non-type-safe things. We 481 // currently don't do any of this for big endian systems. It can be 482 // generalized in the future if someone is interested. 483 if (TD && TD->isLittleEndian()) 484 return FoldReinterpretLoadFromConstPtr(CE, *TD); 485 return 0; 486} 487 488static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ 489 if (LI->isVolatile()) return 0; 490 491 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 492 return ConstantFoldLoadFromConstPtr(C, TD); 493 494 return 0; 495} 496 497/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 498/// Attempt to symbolically evaluate the result of a binary operator merging 499/// these together. If target data info is available, it is provided as TD, 500/// otherwise TD is null. 501static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 502 Constant *Op1, const TargetData *TD){ 503 // SROA 504 505 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 506 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 507 // bits. 508 509 510 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 511 // constant. This happens frequently when iterating over a global array. 512 if (Opc == Instruction::Sub && TD) { 513 GlobalValue *GV1, *GV2; 514 int64_t Offs1, Offs2; 515 516 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) 517 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && 518 GV1 == GV2) { 519 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 520 return ConstantInt::get(Op0->getType(), Offs1-Offs2); 521 } 522 } 523 524 return 0; 525} 526 527/// CastGEPIndices - If array indices are not pointer-sized integers, 528/// explicitly cast them so that they aren't implicitly casted by the 529/// getelementptr. 530static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, 531 const Type *ResultTy, 532 const TargetData *TD) { 533 if (!TD) return 0; 534 const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); 535 536 bool Any = false; 537 SmallVector<Constant*, 32> NewIdxs; 538 for (unsigned i = 1; i != NumOps; ++i) { 539 if ((i == 1 || 540 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(), 541 reinterpret_cast<Value *const *>(Ops+1), 542 i-1))) && 543 Ops[i]->getType() != IntPtrTy) { 544 Any = true; 545 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 546 true, 547 IntPtrTy, 548 true), 549 Ops[i], IntPtrTy)); 550 } else 551 NewIdxs.push_back(Ops[i]); 552 } 553 if (!Any) return 0; 554 555 Constant *C = 556 ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size()); 557 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 558 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD)) 559 C = Folded; 560 return C; 561} 562 563/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP 564/// constant expression, do so. 565static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, 566 const Type *ResultTy, 567 const TargetData *TD) { 568 Constant *Ptr = Ops[0]; 569 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized()) 570 return 0; 571 572 const Type *IntPtrTy = 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])) { 578 579 // If this is "gep i8* Ptr, (sub 0, V)", fold this as: 580 // "inttoptr (sub (ptrtoint Ptr), V)" 581 if (NumOps == 2 && 582 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) { 583 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]); 584 assert((CE == 0 || CE->getType() == IntPtrTy) && 585 "CastGEPIndices didn't canonicalize index types!"); 586 if (CE && CE->getOpcode() == Instruction::Sub && 587 CE->getOperand(0)->isNullValue()) { 588 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType()); 589 Res = ConstantExpr::getSub(Res, CE->getOperand(1)); 590 Res = ConstantExpr::getIntToPtr(Res, ResultTy); 591 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res)) 592 Res = ConstantFoldConstantExpression(ResCE, TD); 593 return Res; 594 } 595 } 596 return 0; 597 } 598 599 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy); 600 APInt Offset = APInt(BitWidth, 601 TD->getIndexedOffset(Ptr->getType(), 602 (Value**)Ops+1, NumOps-1)); 603 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 604 605 // If this is a GEP of a GEP, fold it all into a single GEP. 606 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 607 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end()); 608 609 // Do not try the incorporate the sub-GEP if some index is not a number. 610 bool AllConstantInt = true; 611 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 612 if (!isa<ConstantInt>(NestedOps[i])) { 613 AllConstantInt = false; 614 break; 615 } 616 if (!AllConstantInt) 617 break; 618 619 Ptr = cast<Constant>(GEP->getOperand(0)); 620 Offset += APInt(BitWidth, 621 TD->getIndexedOffset(Ptr->getType(), 622 (Value**)NestedOps.data(), 623 NestedOps.size())); 624 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 625 } 626 627 // If the base value for this address is a literal integer value, fold the 628 // getelementptr to the resulting integer value casted to the pointer type. 629 APInt BasePtr(BitWidth, 0); 630 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 631 if (CE->getOpcode() == Instruction::IntToPtr) 632 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) 633 BasePtr = Base->getValue().zextOrTrunc(BitWidth); 634 if (Ptr->isNullValue() || BasePtr != 0) { 635 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); 636 return ConstantExpr::getIntToPtr(C, ResultTy); 637 } 638 639 // Otherwise form a regular getelementptr. Recompute the indices so that 640 // we eliminate over-indexing of the notional static type array bounds. 641 // This makes it easy to determine if the getelementptr is "inbounds". 642 // Also, this helps GlobalOpt do SROA on GlobalVariables. 643 const Type *Ty = Ptr->getType(); 644 SmallVector<Constant*, 32> NewIdxs; 645 do { 646 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 647 if (ATy->isPointerTy()) { 648 // The only pointer indexing we'll do is on the first index of the GEP. 649 if (!NewIdxs.empty()) 650 break; 651 652 // Only handle pointers to sized types, not pointers to functions. 653 if (!ATy->getElementType()->isSized()) 654 return 0; 655 } 656 657 // Determine which element of the array the offset points into. 658 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 659 const IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 660 if (ElemSize == 0) 661 // The element size is 0. This may be [0 x Ty]*, so just use a zero 662 // index for this level and proceed to the next level to see if it can 663 // accommodate the offset. 664 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 665 else { 666 // The element size is non-zero divide the offset by the element 667 // size (rounding down), to compute the index at this level. 668 APInt NewIdx = Offset.udiv(ElemSize); 669 Offset -= NewIdx * ElemSize; 670 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 671 } 672 Ty = ATy->getElementType(); 673 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) { 674 // Determine which field of the struct the offset points into. The 675 // getZExtValue is at least as safe as the StructLayout API because we 676 // know the offset is within the struct at this point. 677 const StructLayout &SL = *TD->getStructLayout(STy); 678 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 679 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 680 ElIdx)); 681 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 682 Ty = STy->getTypeAtIndex(ElIdx); 683 } else { 684 // We've reached some non-indexable type. 685 break; 686 } 687 } while (Ty != cast<PointerType>(ResultTy)->getElementType()); 688 689 // If we haven't used up the entire offset by descending the static 690 // type, then the offset is pointing into the middle of an indivisible 691 // member, so we can't simplify it. 692 if (Offset != 0) 693 return 0; 694 695 // Create a GEP. 696 Constant *C = 697 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()); 698 assert(cast<PointerType>(C->getType())->getElementType() == Ty && 699 "Computed GetElementPtr has unexpected type!"); 700 701 // If we ended up indexing a member with a type that doesn't match 702 // the type of what the original indices indexed, add a cast. 703 if (Ty != cast<PointerType>(ResultTy)->getElementType()) 704 C = FoldBitCast(C, ResultTy, *TD); 705 706 return C; 707} 708 709 710 711//===----------------------------------------------------------------------===// 712// Constant Folding public APIs 713//===----------------------------------------------------------------------===// 714 715/// ConstantFoldInstruction - Try to constant fold the specified instruction. 716/// If successful, the constant result is returned, if not, null is returned. 717/// Note that this fails if not all of the operands are constant. Otherwise, 718/// this function can only fail when attempting to fold instructions like loads 719/// and stores, which have no constant expression form. 720Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { 721 // Handle PHI nodes quickly here... 722 if (PHINode *PN = dyn_cast<PHINode>(I)) { 723 Constant *CommonValue = 0; 724 725 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 726 Value *Incoming = PN->getIncomingValue(i); 727 // If the incoming value is undef then skip it. Note that while we could 728 // skip the value if it is equal to the phi node itself we choose not to 729 // because that would break the rule that constant folding only applies if 730 // all operands are constants. 731 if (isa<UndefValue>(Incoming)) 732 continue; 733 // If the incoming value is not a constant, or is a different constant to 734 // the one we saw previously, then give up. 735 Constant *C = dyn_cast<Constant>(Incoming); 736 if (!C || (CommonValue && C != CommonValue)) 737 return 0; 738 CommonValue = C; 739 } 740 741 // If we reach here, all incoming values are the same constant or undef. 742 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 743 } 744 745 // Scan the operand list, checking to see if they are all constants, if so, 746 // hand off to ConstantFoldInstOperands. 747 SmallVector<Constant*, 8> Ops; 748 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 749 if (Constant *Op = dyn_cast<Constant>(*i)) 750 Ops.push_back(Op); 751 else 752 return 0; // All operands not constant! 753 754 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 755 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 756 TD); 757 758 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 759 return ConstantFoldLoadInst(LI, TD); 760 761 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) 762 return ConstantExpr::getInsertValue( 763 cast<Constant>(IVI->getAggregateOperand()), 764 cast<Constant>(IVI->getInsertedValueOperand()), 765 IVI->idx_begin(), IVI->getNumIndices()); 766 767 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) 768 return ConstantExpr::getExtractValue( 769 cast<Constant>(EVI->getAggregateOperand()), 770 EVI->idx_begin(), EVI->getNumIndices()); 771 772 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), 773 Ops.data(), Ops.size(), TD); 774} 775 776/// ConstantFoldConstantExpression - Attempt to fold the constant expression 777/// using the specified TargetData. If successful, the constant result is 778/// result is returned, if not, null is returned. 779Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 780 const TargetData *TD) { 781 SmallVector<Constant*, 8> Ops; 782 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); 783 i != e; ++i) { 784 Constant *NewC = cast<Constant>(*i); 785 // Recursively fold the ConstantExpr's operands. 786 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) 787 NewC = ConstantFoldConstantExpression(NewCE, TD); 788 Ops.push_back(NewC); 789 } 790 791 if (CE->isCompare()) 792 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 793 TD); 794 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), 795 Ops.data(), Ops.size(), TD); 796} 797 798/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 799/// specified opcode and operands. If successful, the constant result is 800/// returned, if not, null is returned. Note that this function can fail when 801/// attempting to fold instructions like loads and stores, which have no 802/// constant expression form. 803/// 804/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 805/// information, due to only being passed an opcode and operands. Constant 806/// folding using this function strips this information. 807/// 808Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, 809 Constant* const* Ops, unsigned NumOps, 810 const TargetData *TD) { 811 // Handle easy binops first. 812 if (Instruction::isBinaryOp(Opcode)) { 813 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) 814 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 815 return C; 816 817 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 818 } 819 820 switch (Opcode) { 821 default: return 0; 822 case Instruction::ICmp: 823 case Instruction::FCmp: assert(0 && "Invalid for compares"); 824 case Instruction::Call: 825 if (Function *F = dyn_cast<Function>(Ops[NumOps - 1])) 826 if (canConstantFoldCallTo(F)) 827 return ConstantFoldCall(F, Ops, NumOps - 1); 828 return 0; 829 case Instruction::PtrToInt: 830 // If the input is a inttoptr, eliminate the pair. This requires knowing 831 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 832 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 833 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 834 Constant *Input = CE->getOperand(0); 835 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 836 if (TD->getPointerSizeInBits() < InWidth) { 837 Constant *Mask = 838 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 839 TD->getPointerSizeInBits())); 840 Input = ConstantExpr::getAnd(Input, Mask); 841 } 842 // Do a zext or trunc to get to the dest size. 843 return ConstantExpr::getIntegerCast(Input, DestTy, false); 844 } 845 } 846 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 847 case Instruction::IntToPtr: 848 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 849 // the int size is >= the ptr size. This requires knowing the width of a 850 // pointer, so it can't be done in ConstantExpr::getCast. 851 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) 852 if (TD && 853 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && 854 CE->getOpcode() == Instruction::PtrToInt) 855 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 856 857 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 858 case Instruction::Trunc: 859 case Instruction::ZExt: 860 case Instruction::SExt: 861 case Instruction::FPTrunc: 862 case Instruction::FPExt: 863 case Instruction::UIToFP: 864 case Instruction::SIToFP: 865 case Instruction::FPToUI: 866 case Instruction::FPToSI: 867 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 868 case Instruction::BitCast: 869 if (TD) 870 return FoldBitCast(Ops[0], DestTy, *TD); 871 return ConstantExpr::getBitCast(Ops[0], DestTy); 872 case Instruction::Select: 873 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 874 case Instruction::ExtractElement: 875 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 876 case Instruction::InsertElement: 877 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 878 case Instruction::ShuffleVector: 879 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 880 case Instruction::GetElementPtr: 881 if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD)) 882 return C; 883 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD)) 884 return C; 885 886 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1); 887 } 888} 889 890/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 891/// instruction (icmp/fcmp) with the specified operands. If it fails, it 892/// returns a constant expression of the specified operands. 893/// 894Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 895 Constant *Ops0, Constant *Ops1, 896 const TargetData *TD) { 897 // fold: icmp (inttoptr x), null -> icmp x, 0 898 // fold: icmp (ptrtoint x), 0 -> icmp x, null 899 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 900 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 901 // 902 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 903 // around to know if bit truncation is happening. 904 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 905 if (TD && Ops1->isNullValue()) { 906 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 907 if (CE0->getOpcode() == Instruction::IntToPtr) { 908 // Convert the integer value to the right size to ensure we get the 909 // proper extension or truncation. 910 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 911 IntPtrTy, false); 912 Constant *Null = Constant::getNullValue(C->getType()); 913 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 914 } 915 916 // Only do this transformation if the int is intptrty in size, otherwise 917 // there is a truncation or extension that we aren't modeling. 918 if (CE0->getOpcode() == Instruction::PtrToInt && 919 CE0->getType() == IntPtrTy) { 920 Constant *C = CE0->getOperand(0); 921 Constant *Null = Constant::getNullValue(C->getType()); 922 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 923 } 924 } 925 926 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 927 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 928 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 929 930 if (CE0->getOpcode() == Instruction::IntToPtr) { 931 // Convert the integer value to the right size to ensure we get the 932 // proper extension or truncation. 933 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 934 IntPtrTy, false); 935 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 936 IntPtrTy, false); 937 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); 938 } 939 940 // Only do this transformation if the int is intptrty in size, otherwise 941 // there is a truncation or extension that we aren't modeling. 942 if ((CE0->getOpcode() == Instruction::PtrToInt && 943 CE0->getType() == IntPtrTy && 944 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) 945 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), 946 CE1->getOperand(0), TD); 947 } 948 } 949 950 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 951 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 952 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 953 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 954 Constant *LHS = 955 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); 956 Constant *RHS = 957 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); 958 unsigned OpC = 959 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 960 Constant *Ops[] = { LHS, RHS }; 961 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD); 962 } 963 } 964 965 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 966} 967 968 969/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 970/// getelementptr constantexpr, return the constant value being addressed by the 971/// constant expression, or null if something is funny and we can't decide. 972Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 973 ConstantExpr *CE) { 974 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) 975 return 0; // Do not allow stepping over the value! 976 977 // Loop over all of the operands, tracking down which value we are 978 // addressing... 979 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 980 for (++I; I != E; ++I) 981 if (const StructType *STy = dyn_cast<StructType>(*I)) { 982 ConstantInt *CU = cast<ConstantInt>(I.getOperand()); 983 assert(CU->getZExtValue() < STy->getNumElements() && 984 "Struct index out of range!"); 985 unsigned El = (unsigned)CU->getZExtValue(); 986 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 987 C = CS->getOperand(El); 988 } else if (isa<ConstantAggregateZero>(C)) { 989 C = Constant::getNullValue(STy->getElementType(El)); 990 } else if (isa<UndefValue>(C)) { 991 C = UndefValue::get(STy->getElementType(El)); 992 } else { 993 return 0; 994 } 995 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 996 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) { 997 if (CI->getZExtValue() >= ATy->getNumElements()) 998 return 0; 999 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) 1000 C = CA->getOperand(CI->getZExtValue()); 1001 else if (isa<ConstantAggregateZero>(C)) 1002 C = Constant::getNullValue(ATy->getElementType()); 1003 else if (isa<UndefValue>(C)) 1004 C = UndefValue::get(ATy->getElementType()); 1005 else 1006 return 0; 1007 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) { 1008 if (CI->getZExtValue() >= VTy->getNumElements()) 1009 return 0; 1010 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) 1011 C = CP->getOperand(CI->getZExtValue()); 1012 else if (isa<ConstantAggregateZero>(C)) 1013 C = Constant::getNullValue(VTy->getElementType()); 1014 else if (isa<UndefValue>(C)) 1015 C = UndefValue::get(VTy->getElementType()); 1016 else 1017 return 0; 1018 } else { 1019 return 0; 1020 } 1021 } else { 1022 return 0; 1023 } 1024 return C; 1025} 1026 1027 1028//===----------------------------------------------------------------------===// 1029// Constant Folding for Calls 1030// 1031 1032/// canConstantFoldCallTo - Return true if its even possible to fold a call to 1033/// the specified function. 1034bool 1035llvm::canConstantFoldCallTo(const Function *F) { 1036 switch (F->getIntrinsicID()) { 1037 case Intrinsic::sqrt: 1038 case Intrinsic::powi: 1039 case Intrinsic::bswap: 1040 case Intrinsic::ctpop: 1041 case Intrinsic::ctlz: 1042 case Intrinsic::cttz: 1043 case Intrinsic::uadd_with_overflow: 1044 case Intrinsic::usub_with_overflow: 1045 case Intrinsic::sadd_with_overflow: 1046 case Intrinsic::ssub_with_overflow: 1047 case Intrinsic::smul_with_overflow: 1048 case Intrinsic::convert_from_fp16: 1049 case Intrinsic::convert_to_fp16: 1050 case Intrinsic::x86_sse_cvtss2si: 1051 case Intrinsic::x86_sse_cvtss2si64: 1052 case Intrinsic::x86_sse_cvttss2si: 1053 case Intrinsic::x86_sse_cvttss2si64: 1054 case Intrinsic::x86_sse2_cvtsd2si: 1055 case Intrinsic::x86_sse2_cvtsd2si64: 1056 case Intrinsic::x86_sse2_cvttsd2si: 1057 case Intrinsic::x86_sse2_cvttsd2si64: 1058 return true; 1059 default: 1060 return false; 1061 case 0: break; 1062 } 1063 1064 if (!F->hasName()) return false; 1065 StringRef Name = F->getName(); 1066 1067 // In these cases, the check of the length is required. We don't want to 1068 // return true for a name like "cos\0blah" which strcmp would return equal to 1069 // "cos", but has length 8. 1070 switch (Name[0]) { 1071 default: return false; 1072 case 'a': 1073 return Name == "acos" || Name == "asin" || 1074 Name == "atan" || Name == "atan2"; 1075 case 'c': 1076 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1077 case 'e': 1078 return Name == "exp"; 1079 case 'f': 1080 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1081 case 'l': 1082 return Name == "log" || Name == "log10"; 1083 case 'p': 1084 return Name == "pow"; 1085 case 's': 1086 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1087 Name == "sinf" || Name == "sqrtf"; 1088 case 't': 1089 return Name == "tan" || Name == "tanh"; 1090 } 1091} 1092 1093static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1094 const Type *Ty) { 1095 sys::llvm_fenv_clearexcept(); 1096 V = NativeFP(V); 1097 if (sys::llvm_fenv_testexcept()) { 1098 sys::llvm_fenv_clearexcept(); 1099 return 0; 1100 } 1101 1102 if (Ty->isFloatTy()) 1103 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1104 if (Ty->isDoubleTy()) 1105 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1106 llvm_unreachable("Can only constant fold float/double"); 1107 return 0; // dummy return to suppress warning 1108} 1109 1110static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1111 double V, double W, const Type *Ty) { 1112 sys::llvm_fenv_clearexcept(); 1113 V = NativeFP(V, W); 1114 if (sys::llvm_fenv_testexcept()) { 1115 sys::llvm_fenv_clearexcept(); 1116 return 0; 1117 } 1118 1119 if (Ty->isFloatTy()) 1120 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1121 if (Ty->isDoubleTy()) 1122 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1123 llvm_unreachable("Can only constant fold float/double"); 1124 return 0; // dummy return to suppress warning 1125} 1126 1127/// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer 1128/// conversion of a constant floating point. If roundTowardZero is false, the 1129/// default IEEE rounding is used (toward nearest, ties to even). This matches 1130/// the behavior of the non-truncating SSE instructions in the default rounding 1131/// mode. The desired integer type Ty is used to select how many bits are 1132/// available for the result. Returns null if the conversion cannot be 1133/// performed, otherwise returns the Constant value resulting from the 1134/// conversion. 1135static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero, 1136 const Type *Ty) { 1137 assert(Op && "Called with NULL operand"); 1138 APFloat Val(Op->getValueAPF()); 1139 1140 // All of these conversion intrinsics form an integer of at most 64bits. 1141 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth(); 1142 assert(ResultWidth <= 64 && 1143 "Can only constant fold conversions to 64 and 32 bit ints"); 1144 1145 uint64_t UIntVal; 1146 bool isExact = false; 1147 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero 1148 : APFloat::rmNearestTiesToEven; 1149 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth, 1150 /*isSigned=*/true, mode, 1151 &isExact); 1152 if (status != APFloat::opOK && status != APFloat::opInexact) 1153 return 0; 1154 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true); 1155} 1156 1157/// ConstantFoldCall - Attempt to constant fold a call to the specified function 1158/// with the specified arguments, returning null if unsuccessful. 1159Constant * 1160llvm::ConstantFoldCall(Function *F, 1161 Constant *const *Operands, unsigned NumOperands) { 1162 if (!F->hasName()) return 0; 1163 StringRef Name = F->getName(); 1164 1165 const Type *Ty = F->getReturnType(); 1166 if (NumOperands == 1) { 1167 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1168 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { 1169 APFloat Val(Op->getValueAPF()); 1170 1171 bool lost = false; 1172 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1173 1174 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1175 } 1176 1177 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1178 return 0; 1179 1180 /// We only fold functions with finite arguments. Folding NaN and inf is 1181 /// likely to be aborted with an exception anyway, and some host libms 1182 /// have known errors raising exceptions. 1183 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1184 return 0; 1185 1186 /// Currently APFloat versions of these functions do not exist, so we use 1187 /// the host native double versions. Float versions are not called 1188 /// directly but for all these it is true (float)(f((double)arg)) == 1189 /// f(arg). Long double not supported yet. 1190 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : 1191 Op->getValueAPF().convertToDouble(); 1192 switch (Name[0]) { 1193 case 'a': 1194 if (Name == "acos") 1195 return ConstantFoldFP(acos, V, Ty); 1196 else if (Name == "asin") 1197 return ConstantFoldFP(asin, V, Ty); 1198 else if (Name == "atan") 1199 return ConstantFoldFP(atan, V, Ty); 1200 break; 1201 case 'c': 1202 if (Name == "ceil") 1203 return ConstantFoldFP(ceil, V, Ty); 1204 else if (Name == "cos") 1205 return ConstantFoldFP(cos, V, Ty); 1206 else if (Name == "cosh") 1207 return ConstantFoldFP(cosh, V, Ty); 1208 else if (Name == "cosf") 1209 return ConstantFoldFP(cos, V, Ty); 1210 break; 1211 case 'e': 1212 if (Name == "exp") 1213 return ConstantFoldFP(exp, V, Ty); 1214 break; 1215 case 'f': 1216 if (Name == "fabs") 1217 return ConstantFoldFP(fabs, V, Ty); 1218 else if (Name == "floor") 1219 return ConstantFoldFP(floor, V, Ty); 1220 break; 1221 case 'l': 1222 if (Name == "log" && V > 0) 1223 return ConstantFoldFP(log, V, Ty); 1224 else if (Name == "log10" && V > 0) 1225 return ConstantFoldFP(log10, V, Ty); 1226 else if (F->getIntrinsicID() == Intrinsic::sqrt && 1227 (Ty->isFloatTy() || Ty->isDoubleTy())) { 1228 if (V >= -0.0) 1229 return ConstantFoldFP(sqrt, V, Ty); 1230 else // Undefined 1231 return Constant::getNullValue(Ty); 1232 } 1233 break; 1234 case 's': 1235 if (Name == "sin") 1236 return ConstantFoldFP(sin, V, Ty); 1237 else if (Name == "sinh") 1238 return ConstantFoldFP(sinh, V, Ty); 1239 else if (Name == "sqrt" && V >= 0) 1240 return ConstantFoldFP(sqrt, V, Ty); 1241 else if (Name == "sqrtf" && V >= 0) 1242 return ConstantFoldFP(sqrt, V, Ty); 1243 else if (Name == "sinf") 1244 return ConstantFoldFP(sin, V, Ty); 1245 break; 1246 case 't': 1247 if (Name == "tan") 1248 return ConstantFoldFP(tan, V, Ty); 1249 else if (Name == "tanh") 1250 return ConstantFoldFP(tanh, V, Ty); 1251 break; 1252 default: 1253 break; 1254 } 1255 return 0; 1256 } 1257 1258 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1259 switch (F->getIntrinsicID()) { 1260 case Intrinsic::bswap: 1261 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1262 case Intrinsic::ctpop: 1263 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1264 case Intrinsic::cttz: 1265 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); 1266 case Intrinsic::ctlz: 1267 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); 1268 case Intrinsic::convert_from_fp16: { 1269 APFloat Val(Op->getValue()); 1270 1271 bool lost = false; 1272 APFloat::opStatus status = 1273 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1274 1275 // Conversion is always precise. 1276 (void)status; 1277 assert(status == APFloat::opOK && !lost && 1278 "Precision lost during fp16 constfolding"); 1279 1280 return ConstantFP::get(F->getContext(), Val); 1281 } 1282 default: 1283 return 0; 1284 } 1285 } 1286 1287 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) { 1288 switch (F->getIntrinsicID()) { 1289 default: break; 1290 case Intrinsic::x86_sse_cvtss2si: 1291 case Intrinsic::x86_sse_cvtss2si64: 1292 case Intrinsic::x86_sse2_cvtsd2si: 1293 case Intrinsic::x86_sse2_cvtsd2si64: 1294 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0))) 1295 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty); 1296 case Intrinsic::x86_sse_cvttss2si: 1297 case Intrinsic::x86_sse_cvttss2si64: 1298 case Intrinsic::x86_sse2_cvttsd2si: 1299 case Intrinsic::x86_sse2_cvttsd2si64: 1300 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0))) 1301 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty); 1302 } 1303 } 1304 1305 if (isa<UndefValue>(Operands[0])) { 1306 if (F->getIntrinsicID() == Intrinsic::bswap) 1307 return Operands[0]; 1308 return 0; 1309 } 1310 1311 return 0; 1312 } 1313 1314 if (NumOperands == 2) { 1315 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1316 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1317 return 0; 1318 double Op1V = Ty->isFloatTy() ? 1319 (double)Op1->getValueAPF().convertToFloat() : 1320 Op1->getValueAPF().convertToDouble(); 1321 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1322 if (Op2->getType() != Op1->getType()) 1323 return 0; 1324 1325 double Op2V = Ty->isFloatTy() ? 1326 (double)Op2->getValueAPF().convertToFloat(): 1327 Op2->getValueAPF().convertToDouble(); 1328 1329 if (Name == "pow") 1330 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1331 if (Name == "fmod") 1332 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1333 if (Name == "atan2") 1334 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1335 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1336 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) 1337 return ConstantFP::get(F->getContext(), 1338 APFloat((float)std::pow((float)Op1V, 1339 (int)Op2C->getZExtValue()))); 1340 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy()) 1341 return ConstantFP::get(F->getContext(), 1342 APFloat((double)std::pow((double)Op1V, 1343 (int)Op2C->getZExtValue()))); 1344 } 1345 return 0; 1346 } 1347 1348 1349 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1350 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1351 switch (F->getIntrinsicID()) { 1352 default: break; 1353 case Intrinsic::sadd_with_overflow: 1354 case Intrinsic::uadd_with_overflow: 1355 case Intrinsic::ssub_with_overflow: 1356 case Intrinsic::usub_with_overflow: 1357 case Intrinsic::smul_with_overflow: { 1358 APInt Res; 1359 bool Overflow; 1360 switch (F->getIntrinsicID()) { 1361 default: assert(0 && "Invalid case"); 1362 case Intrinsic::sadd_with_overflow: 1363 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); 1364 break; 1365 case Intrinsic::uadd_with_overflow: 1366 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); 1367 break; 1368 case Intrinsic::ssub_with_overflow: 1369 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); 1370 break; 1371 case Intrinsic::usub_with_overflow: 1372 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); 1373 break; 1374 case Intrinsic::smul_with_overflow: 1375 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); 1376 break; 1377 } 1378 Constant *Ops[] = { 1379 ConstantInt::get(F->getContext(), Res), 1380 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) 1381 }; 1382 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1383 } 1384 } 1385 } 1386 1387 return 0; 1388 } 1389 return 0; 1390 } 1391 return 0; 1392} 1393