ConstantFolding.cpp revision 2ca5c8644e6c35b3a7910a576ed89cddb7b82c3b
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), 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); 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 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 344 if (CE->getOpcode() == Instruction::IntToPtr && 345 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext())) 346 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, 347 BytesLeft, TD); 348 } 349 350 // Otherwise, unknown initializer type. 351 return false; 352} 353 354static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 355 const TargetData &TD) { 356 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType(); 357 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 358 359 // If this isn't an integer load we can't fold it directly. 360 if (!IntType) { 361 // If this is a float/double load, we can try folding it as an int32/64 load 362 // and then bitcast the result. This can be useful for union cases. Note 363 // that address spaces don't matter here since we're not going to result in 364 // an actual new load. 365 const Type *MapTy; 366 if (LoadTy->isFloatTy()) 367 MapTy = Type::getInt32PtrTy(C->getContext()); 368 else if (LoadTy->isDoubleTy()) 369 MapTy = Type::getInt64PtrTy(C->getContext()); 370 else if (LoadTy->isVectorTy()) { 371 MapTy = IntegerType::get(C->getContext(), 372 TD.getTypeAllocSizeInBits(LoadTy)); 373 MapTy = PointerType::getUnqual(MapTy); 374 } else 375 return 0; 376 377 C = FoldBitCast(C, MapTy, TD); 378 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 379 return FoldBitCast(Res, LoadTy, TD); 380 return 0; 381 } 382 383 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 384 if (BytesLoaded > 32 || BytesLoaded == 0) return 0; 385 386 GlobalValue *GVal; 387 int64_t Offset; 388 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 389 return 0; 390 391 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 392 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 393 !GV->getInitializer()->getType()->isSized()) 394 return 0; 395 396 // If we're loading off the beginning of the global, some bytes may be valid, 397 // but we don't try to handle this. 398 if (Offset < 0) return 0; 399 400 // If we're not accessing anything in this constant, the result is undefined. 401 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) 402 return UndefValue::get(IntType); 403 404 unsigned char RawBytes[32] = {0}; 405 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, 406 BytesLoaded, TD)) 407 return 0; 408 409 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); 410 for (unsigned i = 1; i != BytesLoaded; ++i) { 411 ResultVal <<= 8; 412 ResultVal |= RawBytes[BytesLoaded-1-i]; 413 } 414 415 return ConstantInt::get(IntType->getContext(), ResultVal); 416} 417 418/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 419/// produce if it is constant and determinable. If this is not determinable, 420/// return null. 421Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 422 const TargetData *TD) { 423 // First, try the easy cases: 424 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 425 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 426 return GV->getInitializer(); 427 428 // If the loaded value isn't a constant expr, we can't handle it. 429 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 430 if (!CE) return 0; 431 432 if (CE->getOpcode() == Instruction::GetElementPtr) { 433 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 434 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 435 if (Constant *V = 436 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 437 return V; 438 } 439 440 // Instead of loading constant c string, use corresponding integer value 441 // directly if string length is small enough. 442 std::string Str; 443 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) { 444 unsigned StrLen = Str.length(); 445 const Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 446 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 447 // Replace load with immediate integer if the result is an integer or fp 448 // value. 449 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 450 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 451 APInt StrVal(NumBits, 0); 452 APInt SingleChar(NumBits, 0); 453 if (TD->isLittleEndian()) { 454 for (signed i = StrLen-1; i >= 0; i--) { 455 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 456 StrVal = (StrVal << 8) | SingleChar; 457 } 458 } else { 459 for (unsigned i = 0; i < StrLen; i++) { 460 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 461 StrVal = (StrVal << 8) | SingleChar; 462 } 463 // Append NULL at the end. 464 SingleChar = 0; 465 StrVal = (StrVal << 8) | SingleChar; 466 } 467 468 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 469 if (Ty->isFloatingPointTy()) 470 Res = ConstantExpr::getBitCast(Res, Ty); 471 return Res; 472 } 473 } 474 475 // If this load comes from anywhere in a constant global, and if the global 476 // is all undef or zero, we know what it loads. 477 if (GlobalVariable *GV = 478 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) { 479 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 480 const Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 481 if (GV->getInitializer()->isNullValue()) 482 return Constant::getNullValue(ResTy); 483 if (isa<UndefValue>(GV->getInitializer())) 484 return UndefValue::get(ResTy); 485 } 486 } 487 488 // Try hard to fold loads from bitcasted strange and non-type-safe things. We 489 // currently don't do any of this for big endian systems. It can be 490 // generalized in the future if someone is interested. 491 if (TD && TD->isLittleEndian()) 492 return FoldReinterpretLoadFromConstPtr(CE, *TD); 493 return 0; 494} 495 496static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ 497 if (LI->isVolatile()) return 0; 498 499 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 500 return ConstantFoldLoadFromConstPtr(C, TD); 501 502 return 0; 503} 504 505/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 506/// Attempt to symbolically evaluate the result of a binary operator merging 507/// these together. If target data info is available, it is provided as TD, 508/// otherwise TD is null. 509static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 510 Constant *Op1, const TargetData *TD){ 511 // SROA 512 513 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 514 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 515 // bits. 516 517 518 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 519 // constant. This happens frequently when iterating over a global array. 520 if (Opc == Instruction::Sub && TD) { 521 GlobalValue *GV1, *GV2; 522 int64_t Offs1, Offs2; 523 524 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) 525 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && 526 GV1 == GV2) { 527 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 528 return ConstantInt::get(Op0->getType(), Offs1-Offs2); 529 } 530 } 531 532 return 0; 533} 534 535/// CastGEPIndices - If array indices are not pointer-sized integers, 536/// explicitly cast them so that they aren't implicitly casted by the 537/// getelementptr. 538static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, 539 const Type *ResultTy, 540 const TargetData *TD) { 541 if (!TD) return 0; 542 const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); 543 544 bool Any = false; 545 SmallVector<Constant*, 32> NewIdxs; 546 for (unsigned i = 1; i != NumOps; ++i) { 547 if ((i == 1 || 548 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(), 549 reinterpret_cast<Value *const *>(Ops+1), 550 i-1))) && 551 Ops[i]->getType() != IntPtrTy) { 552 Any = true; 553 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 554 true, 555 IntPtrTy, 556 true), 557 Ops[i], IntPtrTy)); 558 } else 559 NewIdxs.push_back(Ops[i]); 560 } 561 if (!Any) return 0; 562 563 Constant *C = 564 ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size()); 565 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 566 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD)) 567 C = Folded; 568 return C; 569} 570 571/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP 572/// constant expression, do so. 573static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, 574 const Type *ResultTy, 575 const TargetData *TD) { 576 Constant *Ptr = Ops[0]; 577 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized()) 578 return 0; 579 580 const Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext()); 581 582 // If this is a constant expr gep that is effectively computing an 583 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' 584 for (unsigned i = 1; i != NumOps; ++i) 585 if (!isa<ConstantInt>(Ops[i])) { 586 587 // If this is "gep i8* Ptr, (sub 0, V)", fold this as: 588 // "inttoptr (sub (ptrtoint Ptr), V)" 589 if (NumOps == 2 && 590 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) { 591 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]); 592 assert((CE == 0 || CE->getType() == IntPtrTy) && 593 "CastGEPIndices didn't canonicalize index types!"); 594 if (CE && CE->getOpcode() == Instruction::Sub && 595 CE->getOperand(0)->isNullValue()) { 596 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType()); 597 Res = ConstantExpr::getSub(Res, CE->getOperand(1)); 598 Res = ConstantExpr::getIntToPtr(Res, ResultTy); 599 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res)) 600 Res = ConstantFoldConstantExpression(ResCE, TD); 601 return Res; 602 } 603 } 604 return 0; 605 } 606 607 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy); 608 APInt Offset = APInt(BitWidth, 609 TD->getIndexedOffset(Ptr->getType(), 610 (Value**)Ops+1, NumOps-1)); 611 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 612 613 // If this is a GEP of a GEP, fold it all into a single GEP. 614 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 615 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end()); 616 617 // Do not try the incorporate the sub-GEP if some index is not a number. 618 bool AllConstantInt = true; 619 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 620 if (!isa<ConstantInt>(NestedOps[i])) { 621 AllConstantInt = false; 622 break; 623 } 624 if (!AllConstantInt) 625 break; 626 627 Ptr = cast<Constant>(GEP->getOperand(0)); 628 Offset += APInt(BitWidth, 629 TD->getIndexedOffset(Ptr->getType(), 630 (Value**)NestedOps.data(), 631 NestedOps.size())); 632 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 633 } 634 635 // If the base value for this address is a literal integer value, fold the 636 // getelementptr to the resulting integer value casted to the pointer type. 637 APInt BasePtr(BitWidth, 0); 638 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 639 if (CE->getOpcode() == Instruction::IntToPtr) 640 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) 641 BasePtr = Base->getValue().zextOrTrunc(BitWidth); 642 if (Ptr->isNullValue() || BasePtr != 0) { 643 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); 644 return ConstantExpr::getIntToPtr(C, ResultTy); 645 } 646 647 // Otherwise form a regular getelementptr. Recompute the indices so that 648 // we eliminate over-indexing of the notional static type array bounds. 649 // This makes it easy to determine if the getelementptr is "inbounds". 650 // Also, this helps GlobalOpt do SROA on GlobalVariables. 651 const Type *Ty = Ptr->getType(); 652 SmallVector<Constant*, 32> NewIdxs; 653 do { 654 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 655 if (ATy->isPointerTy()) { 656 // The only pointer indexing we'll do is on the first index of the GEP. 657 if (!NewIdxs.empty()) 658 break; 659 660 // Only handle pointers to sized types, not pointers to functions. 661 if (!ATy->getElementType()->isSized()) 662 return 0; 663 } 664 665 // Determine which element of the array the offset points into. 666 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 667 const IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 668 if (ElemSize == 0) 669 // The element size is 0. This may be [0 x Ty]*, so just use a zero 670 // index for this level and proceed to the next level to see if it can 671 // accommodate the offset. 672 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 673 else { 674 // The element size is non-zero divide the offset by the element 675 // size (rounding down), to compute the index at this level. 676 APInt NewIdx = Offset.udiv(ElemSize); 677 Offset -= NewIdx * ElemSize; 678 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 679 } 680 Ty = ATy->getElementType(); 681 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) { 682 // Determine which field of the struct the offset points into. The 683 // getZExtValue is at least as safe as the StructLayout API because we 684 // know the offset is within the struct at this point. 685 const StructLayout &SL = *TD->getStructLayout(STy); 686 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 687 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 688 ElIdx)); 689 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 690 Ty = STy->getTypeAtIndex(ElIdx); 691 } else { 692 // We've reached some non-indexable type. 693 break; 694 } 695 } while (Ty != cast<PointerType>(ResultTy)->getElementType()); 696 697 // If we haven't used up the entire offset by descending the static 698 // type, then the offset is pointing into the middle of an indivisible 699 // member, so we can't simplify it. 700 if (Offset != 0) 701 return 0; 702 703 // Create a GEP. 704 Constant *C = 705 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()); 706 assert(cast<PointerType>(C->getType())->getElementType() == Ty && 707 "Computed GetElementPtr has unexpected type!"); 708 709 // If we ended up indexing a member with a type that doesn't match 710 // the type of what the original indices indexed, add a cast. 711 if (Ty != cast<PointerType>(ResultTy)->getElementType()) 712 C = FoldBitCast(C, ResultTy, *TD); 713 714 return C; 715} 716 717 718 719//===----------------------------------------------------------------------===// 720// Constant Folding public APIs 721//===----------------------------------------------------------------------===// 722 723/// ConstantFoldInstruction - Try to constant fold the specified instruction. 724/// If successful, the constant result is returned, if not, null is returned. 725/// Note that this fails if not all of the operands are constant. Otherwise, 726/// this function can only fail when attempting to fold instructions like loads 727/// and stores, which have no constant expression form. 728Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { 729 // Handle PHI nodes quickly here... 730 if (PHINode *PN = dyn_cast<PHINode>(I)) { 731 Constant *CommonValue = 0; 732 733 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 734 Value *Incoming = PN->getIncomingValue(i); 735 // If the incoming value is undef then skip it. Note that while we could 736 // skip the value if it is equal to the phi node itself we choose not to 737 // because that would break the rule that constant folding only applies if 738 // all operands are constants. 739 if (isa<UndefValue>(Incoming)) 740 continue; 741 // If the incoming value is not a constant, or is a different constant to 742 // the one we saw previously, then give up. 743 Constant *C = dyn_cast<Constant>(Incoming); 744 if (!C || (CommonValue && C != CommonValue)) 745 return 0; 746 CommonValue = C; 747 } 748 749 // If we reach here, all incoming values are the same constant or undef. 750 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 751 } 752 753 // Scan the operand list, checking to see if they are all constants, if so, 754 // hand off to ConstantFoldInstOperands. 755 SmallVector<Constant*, 8> Ops; 756 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 757 if (Constant *Op = dyn_cast<Constant>(*i)) 758 Ops.push_back(Op); 759 else 760 return 0; // All operands not constant! 761 762 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 763 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 764 TD); 765 766 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 767 return ConstantFoldLoadInst(LI, TD); 768 769 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) 770 return ConstantExpr::getInsertValue( 771 cast<Constant>(IVI->getAggregateOperand()), 772 cast<Constant>(IVI->getInsertedValueOperand()), 773 IVI->idx_begin(), IVI->getNumIndices()); 774 775 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) 776 return ConstantExpr::getExtractValue( 777 cast<Constant>(EVI->getAggregateOperand()), 778 EVI->idx_begin(), EVI->getNumIndices()); 779 780 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), 781 Ops.data(), Ops.size(), TD); 782} 783 784/// ConstantFoldConstantExpression - Attempt to fold the constant expression 785/// using the specified TargetData. If successful, the constant result is 786/// result is returned, if not, null is returned. 787Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 788 const TargetData *TD) { 789 SmallVector<Constant*, 8> Ops; 790 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); 791 i != e; ++i) { 792 Constant *NewC = cast<Constant>(*i); 793 // Recursively fold the ConstantExpr's operands. 794 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) 795 NewC = ConstantFoldConstantExpression(NewCE, TD); 796 Ops.push_back(NewC); 797 } 798 799 if (CE->isCompare()) 800 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 801 TD); 802 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), 803 Ops.data(), Ops.size(), TD); 804} 805 806/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 807/// specified opcode and operands. If successful, the constant result is 808/// returned, if not, null is returned. Note that this function can fail when 809/// attempting to fold instructions like loads and stores, which have no 810/// constant expression form. 811/// 812/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 813/// information, due to only being passed an opcode and operands. Constant 814/// folding using this function strips this information. 815/// 816Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, 817 Constant* const* Ops, unsigned NumOps, 818 const TargetData *TD) { 819 // Handle easy binops first. 820 if (Instruction::isBinaryOp(Opcode)) { 821 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) 822 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 823 return C; 824 825 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 826 } 827 828 switch (Opcode) { 829 default: return 0; 830 case Instruction::ICmp: 831 case Instruction::FCmp: assert(0 && "Invalid for compares"); 832 case Instruction::Call: 833 if (Function *F = dyn_cast<Function>(Ops[NumOps - 1])) 834 if (canConstantFoldCallTo(F)) 835 return ConstantFoldCall(F, Ops, NumOps - 1); 836 return 0; 837 case Instruction::PtrToInt: 838 // If the input is a inttoptr, eliminate the pair. This requires knowing 839 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 840 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 841 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 842 Constant *Input = CE->getOperand(0); 843 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 844 if (TD->getPointerSizeInBits() < InWidth) { 845 Constant *Mask = 846 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 847 TD->getPointerSizeInBits())); 848 Input = ConstantExpr::getAnd(Input, Mask); 849 } 850 // Do a zext or trunc to get to the dest size. 851 return ConstantExpr::getIntegerCast(Input, DestTy, false); 852 } 853 } 854 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 855 case Instruction::IntToPtr: 856 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 857 // the int size is >= the ptr size. This requires knowing the width of a 858 // pointer, so it can't be done in ConstantExpr::getCast. 859 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) 860 if (TD && 861 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && 862 CE->getOpcode() == Instruction::PtrToInt) 863 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 864 865 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 866 case Instruction::Trunc: 867 case Instruction::ZExt: 868 case Instruction::SExt: 869 case Instruction::FPTrunc: 870 case Instruction::FPExt: 871 case Instruction::UIToFP: 872 case Instruction::SIToFP: 873 case Instruction::FPToUI: 874 case Instruction::FPToSI: 875 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 876 case Instruction::BitCast: 877 if (TD) 878 return FoldBitCast(Ops[0], DestTy, *TD); 879 return ConstantExpr::getBitCast(Ops[0], DestTy); 880 case Instruction::Select: 881 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 882 case Instruction::ExtractElement: 883 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 884 case Instruction::InsertElement: 885 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 886 case Instruction::ShuffleVector: 887 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 888 case Instruction::GetElementPtr: 889 if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD)) 890 return C; 891 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD)) 892 return C; 893 894 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1); 895 } 896} 897 898/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 899/// instruction (icmp/fcmp) with the specified operands. If it fails, it 900/// returns a constant expression of the specified operands. 901/// 902Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 903 Constant *Ops0, Constant *Ops1, 904 const TargetData *TD) { 905 // fold: icmp (inttoptr x), null -> icmp x, 0 906 // fold: icmp (ptrtoint x), 0 -> icmp x, null 907 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 908 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 909 // 910 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 911 // around to know if bit truncation is happening. 912 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 913 if (TD && Ops1->isNullValue()) { 914 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 915 if (CE0->getOpcode() == Instruction::IntToPtr) { 916 // Convert the integer value to the right size to ensure we get the 917 // proper extension or truncation. 918 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 919 IntPtrTy, false); 920 Constant *Null = Constant::getNullValue(C->getType()); 921 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 922 } 923 924 // Only do this transformation if the int is intptrty in size, otherwise 925 // there is a truncation or extension that we aren't modeling. 926 if (CE0->getOpcode() == Instruction::PtrToInt && 927 CE0->getType() == IntPtrTy) { 928 Constant *C = CE0->getOperand(0); 929 Constant *Null = Constant::getNullValue(C->getType()); 930 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 931 } 932 } 933 934 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 935 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 936 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 937 938 if (CE0->getOpcode() == Instruction::IntToPtr) { 939 // Convert the integer value to the right size to ensure we get the 940 // proper extension or truncation. 941 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 942 IntPtrTy, false); 943 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 944 IntPtrTy, false); 945 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); 946 } 947 948 // Only do this transformation if the int is intptrty in size, otherwise 949 // there is a truncation or extension that we aren't modeling. 950 if ((CE0->getOpcode() == Instruction::PtrToInt && 951 CE0->getType() == IntPtrTy && 952 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) 953 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), 954 CE1->getOperand(0), TD); 955 } 956 } 957 958 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 959 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 960 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 961 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 962 Constant *LHS = 963 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); 964 Constant *RHS = 965 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); 966 unsigned OpC = 967 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 968 Constant *Ops[] = { LHS, RHS }; 969 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD); 970 } 971 } 972 973 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 974} 975 976 977/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 978/// getelementptr constantexpr, return the constant value being addressed by the 979/// constant expression, or null if something is funny and we can't decide. 980Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 981 ConstantExpr *CE) { 982 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) 983 return 0; // Do not allow stepping over the value! 984 985 // Loop over all of the operands, tracking down which value we are 986 // addressing... 987 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 988 for (++I; I != E; ++I) 989 if (const StructType *STy = dyn_cast<StructType>(*I)) { 990 ConstantInt *CU = cast<ConstantInt>(I.getOperand()); 991 assert(CU->getZExtValue() < STy->getNumElements() && 992 "Struct index out of range!"); 993 unsigned El = (unsigned)CU->getZExtValue(); 994 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 995 C = CS->getOperand(El); 996 } else if (isa<ConstantAggregateZero>(C)) { 997 C = Constant::getNullValue(STy->getElementType(El)); 998 } else if (isa<UndefValue>(C)) { 999 C = UndefValue::get(STy->getElementType(El)); 1000 } else { 1001 return 0; 1002 } 1003 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 1004 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) { 1005 if (CI->getZExtValue() >= ATy->getNumElements()) 1006 return 0; 1007 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) 1008 C = CA->getOperand(CI->getZExtValue()); 1009 else if (isa<ConstantAggregateZero>(C)) 1010 C = Constant::getNullValue(ATy->getElementType()); 1011 else if (isa<UndefValue>(C)) 1012 C = UndefValue::get(ATy->getElementType()); 1013 else 1014 return 0; 1015 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) { 1016 if (CI->getZExtValue() >= VTy->getNumElements()) 1017 return 0; 1018 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) 1019 C = CP->getOperand(CI->getZExtValue()); 1020 else if (isa<ConstantAggregateZero>(C)) 1021 C = Constant::getNullValue(VTy->getElementType()); 1022 else if (isa<UndefValue>(C)) 1023 C = UndefValue::get(VTy->getElementType()); 1024 else 1025 return 0; 1026 } else { 1027 return 0; 1028 } 1029 } else { 1030 return 0; 1031 } 1032 return C; 1033} 1034 1035 1036//===----------------------------------------------------------------------===// 1037// Constant Folding for Calls 1038// 1039 1040/// canConstantFoldCallTo - Return true if its even possible to fold a call to 1041/// the specified function. 1042bool 1043llvm::canConstantFoldCallTo(const Function *F) { 1044 switch (F->getIntrinsicID()) { 1045 case Intrinsic::sqrt: 1046 case Intrinsic::powi: 1047 case Intrinsic::bswap: 1048 case Intrinsic::ctpop: 1049 case Intrinsic::ctlz: 1050 case Intrinsic::cttz: 1051 case Intrinsic::uadd_with_overflow: 1052 case Intrinsic::usub_with_overflow: 1053 case Intrinsic::sadd_with_overflow: 1054 case Intrinsic::ssub_with_overflow: 1055 case Intrinsic::smul_with_overflow: 1056 case Intrinsic::convert_from_fp16: 1057 case Intrinsic::convert_to_fp16: 1058 case Intrinsic::x86_sse_cvtss2si: 1059 case Intrinsic::x86_sse_cvtss2si64: 1060 case Intrinsic::x86_sse_cvttss2si: 1061 case Intrinsic::x86_sse_cvttss2si64: 1062 case Intrinsic::x86_sse2_cvtsd2si: 1063 case Intrinsic::x86_sse2_cvtsd2si64: 1064 case Intrinsic::x86_sse2_cvttsd2si: 1065 case Intrinsic::x86_sse2_cvttsd2si64: 1066 return true; 1067 default: 1068 return false; 1069 case 0: break; 1070 } 1071 1072 if (!F->hasName()) return false; 1073 StringRef Name = F->getName(); 1074 1075 // In these cases, the check of the length is required. We don't want to 1076 // return true for a name like "cos\0blah" which strcmp would return equal to 1077 // "cos", but has length 8. 1078 switch (Name[0]) { 1079 default: return false; 1080 case 'a': 1081 return Name == "acos" || Name == "asin" || 1082 Name == "atan" || Name == "atan2"; 1083 case 'c': 1084 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1085 case 'e': 1086 return Name == "exp"; 1087 case 'f': 1088 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1089 case 'l': 1090 return Name == "log" || Name == "log10"; 1091 case 'p': 1092 return Name == "pow"; 1093 case 's': 1094 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1095 Name == "sinf" || Name == "sqrtf"; 1096 case 't': 1097 return Name == "tan" || Name == "tanh"; 1098 } 1099} 1100 1101static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1102 const Type *Ty) { 1103 sys::llvm_fenv_clearexcept(); 1104 V = NativeFP(V); 1105 if (sys::llvm_fenv_testexcept()) { 1106 sys::llvm_fenv_clearexcept(); 1107 return 0; 1108 } 1109 1110 if (Ty->isFloatTy()) 1111 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1112 if (Ty->isDoubleTy()) 1113 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1114 llvm_unreachable("Can only constant fold float/double"); 1115 return 0; // dummy return to suppress warning 1116} 1117 1118static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1119 double V, double W, const Type *Ty) { 1120 sys::llvm_fenv_clearexcept(); 1121 V = NativeFP(V, W); 1122 if (sys::llvm_fenv_testexcept()) { 1123 sys::llvm_fenv_clearexcept(); 1124 return 0; 1125 } 1126 1127 if (Ty->isFloatTy()) 1128 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1129 if (Ty->isDoubleTy()) 1130 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1131 llvm_unreachable("Can only constant fold float/double"); 1132 return 0; // dummy return to suppress warning 1133} 1134 1135/// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer 1136/// conversion of a constant floating point. If roundTowardZero is false, the 1137/// default IEEE rounding is used (toward nearest, ties to even). This matches 1138/// the behavior of the non-truncating SSE instructions in the default rounding 1139/// mode. The desired integer type Ty is used to select how many bits are 1140/// available for the result. Returns null if the conversion cannot be 1141/// performed, otherwise returns the Constant value resulting from the 1142/// conversion. 1143static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero, 1144 const Type *Ty) { 1145 assert(Op && "Called with NULL operand"); 1146 APFloat Val(Op->getValueAPF()); 1147 1148 // All of these conversion intrinsics form an integer of at most 64bits. 1149 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth(); 1150 assert(ResultWidth <= 64 && 1151 "Can only constant fold conversions to 64 and 32 bit ints"); 1152 1153 uint64_t UIntVal; 1154 bool isExact = false; 1155 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero 1156 : APFloat::rmNearestTiesToEven; 1157 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth, 1158 /*isSigned=*/true, mode, 1159 &isExact); 1160 if (status != APFloat::opOK && status != APFloat::opInexact) 1161 return 0; 1162 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true); 1163} 1164 1165/// ConstantFoldCall - Attempt to constant fold a call to the specified function 1166/// with the specified arguments, returning null if unsuccessful. 1167Constant * 1168llvm::ConstantFoldCall(Function *F, 1169 Constant *const *Operands, unsigned NumOperands) { 1170 if (!F->hasName()) return 0; 1171 StringRef Name = F->getName(); 1172 1173 const Type *Ty = F->getReturnType(); 1174 if (NumOperands == 1) { 1175 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1176 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { 1177 APFloat Val(Op->getValueAPF()); 1178 1179 bool lost = false; 1180 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1181 1182 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1183 } 1184 1185 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1186 return 0; 1187 1188 /// We only fold functions with finite arguments. Folding NaN and inf is 1189 /// likely to be aborted with an exception anyway, and some host libms 1190 /// have known errors raising exceptions. 1191 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1192 return 0; 1193 1194 /// Currently APFloat versions of these functions do not exist, so we use 1195 /// the host native double versions. Float versions are not called 1196 /// directly but for all these it is true (float)(f((double)arg)) == 1197 /// f(arg). Long double not supported yet. 1198 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : 1199 Op->getValueAPF().convertToDouble(); 1200 switch (Name[0]) { 1201 case 'a': 1202 if (Name == "acos") 1203 return ConstantFoldFP(acos, V, Ty); 1204 else if (Name == "asin") 1205 return ConstantFoldFP(asin, V, Ty); 1206 else if (Name == "atan") 1207 return ConstantFoldFP(atan, V, Ty); 1208 break; 1209 case 'c': 1210 if (Name == "ceil") 1211 return ConstantFoldFP(ceil, V, Ty); 1212 else if (Name == "cos") 1213 return ConstantFoldFP(cos, V, Ty); 1214 else if (Name == "cosh") 1215 return ConstantFoldFP(cosh, V, Ty); 1216 else if (Name == "cosf") 1217 return ConstantFoldFP(cos, V, Ty); 1218 break; 1219 case 'e': 1220 if (Name == "exp") 1221 return ConstantFoldFP(exp, V, Ty); 1222 break; 1223 case 'f': 1224 if (Name == "fabs") 1225 return ConstantFoldFP(fabs, V, Ty); 1226 else if (Name == "floor") 1227 return ConstantFoldFP(floor, V, Ty); 1228 break; 1229 case 'l': 1230 if (Name == "log" && V > 0) 1231 return ConstantFoldFP(log, V, Ty); 1232 else if (Name == "log10" && V > 0) 1233 return ConstantFoldFP(log10, V, Ty); 1234 else if (F->getIntrinsicID() == Intrinsic::sqrt && 1235 (Ty->isFloatTy() || Ty->isDoubleTy())) { 1236 if (V >= -0.0) 1237 return ConstantFoldFP(sqrt, V, Ty); 1238 else // Undefined 1239 return Constant::getNullValue(Ty); 1240 } 1241 break; 1242 case 's': 1243 if (Name == "sin") 1244 return ConstantFoldFP(sin, V, Ty); 1245 else if (Name == "sinh") 1246 return ConstantFoldFP(sinh, V, Ty); 1247 else if (Name == "sqrt" && V >= 0) 1248 return ConstantFoldFP(sqrt, V, Ty); 1249 else if (Name == "sqrtf" && V >= 0) 1250 return ConstantFoldFP(sqrt, V, Ty); 1251 else if (Name == "sinf") 1252 return ConstantFoldFP(sin, V, Ty); 1253 break; 1254 case 't': 1255 if (Name == "tan") 1256 return ConstantFoldFP(tan, V, Ty); 1257 else if (Name == "tanh") 1258 return ConstantFoldFP(tanh, V, Ty); 1259 break; 1260 default: 1261 break; 1262 } 1263 return 0; 1264 } 1265 1266 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1267 switch (F->getIntrinsicID()) { 1268 case Intrinsic::bswap: 1269 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1270 case Intrinsic::ctpop: 1271 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1272 case Intrinsic::cttz: 1273 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); 1274 case Intrinsic::ctlz: 1275 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); 1276 case Intrinsic::convert_from_fp16: { 1277 APFloat Val(Op->getValue()); 1278 1279 bool lost = false; 1280 APFloat::opStatus status = 1281 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1282 1283 // Conversion is always precise. 1284 (void)status; 1285 assert(status == APFloat::opOK && !lost && 1286 "Precision lost during fp16 constfolding"); 1287 1288 return ConstantFP::get(F->getContext(), Val); 1289 } 1290 default: 1291 return 0; 1292 } 1293 } 1294 1295 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) { 1296 switch (F->getIntrinsicID()) { 1297 default: break; 1298 case Intrinsic::x86_sse_cvtss2si: 1299 case Intrinsic::x86_sse_cvtss2si64: 1300 case Intrinsic::x86_sse2_cvtsd2si: 1301 case Intrinsic::x86_sse2_cvtsd2si64: 1302 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0))) 1303 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty); 1304 case Intrinsic::x86_sse_cvttss2si: 1305 case Intrinsic::x86_sse_cvttss2si64: 1306 case Intrinsic::x86_sse2_cvttsd2si: 1307 case Intrinsic::x86_sse2_cvttsd2si64: 1308 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0))) 1309 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty); 1310 } 1311 } 1312 1313 if (isa<UndefValue>(Operands[0])) { 1314 if (F->getIntrinsicID() == Intrinsic::bswap) 1315 return Operands[0]; 1316 return 0; 1317 } 1318 1319 return 0; 1320 } 1321 1322 if (NumOperands == 2) { 1323 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1324 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1325 return 0; 1326 double Op1V = Ty->isFloatTy() ? 1327 (double)Op1->getValueAPF().convertToFloat() : 1328 Op1->getValueAPF().convertToDouble(); 1329 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1330 if (Op2->getType() != Op1->getType()) 1331 return 0; 1332 1333 double Op2V = Ty->isFloatTy() ? 1334 (double)Op2->getValueAPF().convertToFloat(): 1335 Op2->getValueAPF().convertToDouble(); 1336 1337 if (Name == "pow") 1338 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1339 if (Name == "fmod") 1340 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1341 if (Name == "atan2") 1342 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1343 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1344 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) 1345 return ConstantFP::get(F->getContext(), 1346 APFloat((float)std::pow((float)Op1V, 1347 (int)Op2C->getZExtValue()))); 1348 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy()) 1349 return ConstantFP::get(F->getContext(), 1350 APFloat((double)std::pow((double)Op1V, 1351 (int)Op2C->getZExtValue()))); 1352 } 1353 return 0; 1354 } 1355 1356 1357 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1358 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1359 switch (F->getIntrinsicID()) { 1360 default: break; 1361 case Intrinsic::sadd_with_overflow: 1362 case Intrinsic::uadd_with_overflow: 1363 case Intrinsic::ssub_with_overflow: 1364 case Intrinsic::usub_with_overflow: 1365 case Intrinsic::smul_with_overflow: { 1366 APInt Res; 1367 bool Overflow; 1368 switch (F->getIntrinsicID()) { 1369 default: assert(0 && "Invalid case"); 1370 case Intrinsic::sadd_with_overflow: 1371 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); 1372 break; 1373 case Intrinsic::uadd_with_overflow: 1374 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); 1375 break; 1376 case Intrinsic::ssub_with_overflow: 1377 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); 1378 break; 1379 case Intrinsic::usub_with_overflow: 1380 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); 1381 break; 1382 case Intrinsic::smul_with_overflow: 1383 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); 1384 break; 1385 } 1386 Constant *Ops[] = { 1387 ConstantInt::get(F->getContext(), Res), 1388 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) 1389 }; 1390 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1391 } 1392 } 1393 } 1394 1395 return 0; 1396 } 1397 return 0; 1398 } 1399 return 0; 1400} 1401