ConstantFolding.cpp revision a9203109f4ac95aa7e9624f2838e3d89623ec902
1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file defines routines for folding instructions into constants. 11// 12// Also, to supplement the basic VMCore ConstantExpr simplifications, 13// this file defines some additional folding routines that can make use of 14// TargetData information. These functions cannot go in VMCore due to library 15// dependency issues. 16// 17//===----------------------------------------------------------------------===// 18 19#include "llvm/Analysis/ConstantFolding.h" 20#include "llvm/Constants.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/Function.h" 23#include "llvm/GlobalVariable.h" 24#include "llvm/Instructions.h" 25#include "llvm/Intrinsics.h" 26#include "llvm/Operator.h" 27#include "llvm/Analysis/ValueTracking.h" 28#include "llvm/Target/TargetData.h" 29#include "llvm/ADT/SmallVector.h" 30#include "llvm/ADT/StringMap.h" 31#include "llvm/Support/ErrorHandling.h" 32#include "llvm/Support/GetElementPtrTypeIterator.h" 33#include "llvm/Support/MathExtras.h" 34#include "llvm/Support/FEnv.h" 35#include <cerrno> 36#include <cmath> 37using namespace llvm; 38 39//===----------------------------------------------------------------------===// 40// Constant Folding internal helper functions 41//===----------------------------------------------------------------------===// 42 43/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with 44/// TargetData. This always returns a non-null constant, but it may be a 45/// ConstantExpr if unfoldable. 46static Constant *FoldBitCast(Constant *C, Type *DestTy, 47 const TargetData &TD) { 48 49 // This only handles casts to vectors currently. 50 VectorType *DestVTy = dyn_cast<VectorType>(DestTy); 51 if (DestVTy == 0) 52 return ConstantExpr::getBitCast(C, DestTy); 53 54 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 55 // vector so the code below can handle it uniformly. 56 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 57 Constant *Ops = C; // don't take the address of C! 58 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD); 59 } 60 61 // If this is a bitcast from constant vector -> vector, fold it. 62 ConstantVector *CV = dyn_cast<ConstantVector>(C); 63 if (CV == 0) 64 return ConstantExpr::getBitCast(C, DestTy); 65 66 // If the element types match, VMCore can fold it. 67 unsigned NumDstElt = DestVTy->getNumElements(); 68 unsigned NumSrcElt = CV->getNumOperands(); 69 if (NumDstElt == NumSrcElt) 70 return ConstantExpr::getBitCast(C, DestTy); 71 72 Type *SrcEltTy = CV->getType()->getElementType(); 73 Type *DstEltTy = DestVTy->getElementType(); 74 75 // Otherwise, we're changing the number of elements in a vector, which 76 // requires endianness information to do the right thing. For example, 77 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 78 // folds to (little endian): 79 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 80 // and to (big endian): 81 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 82 83 // First thing is first. We only want to think about integer here, so if 84 // we have something in FP form, recast it as integer. 85 if (DstEltTy->isFloatingPointTy()) { 86 // Fold to an vector of integers with same size as our FP type. 87 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 88 Type *DestIVTy = 89 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 90 // Recursively handle this integer conversion, if possible. 91 C = FoldBitCast(C, DestIVTy, TD); 92 if (!C) return ConstantExpr::getBitCast(C, DestTy); 93 94 // Finally, VMCore can handle this now that #elts line up. 95 return ConstantExpr::getBitCast(C, DestTy); 96 } 97 98 // Okay, we know the destination is integer, if the input is FP, convert 99 // it to integer first. 100 if (SrcEltTy->isFloatingPointTy()) { 101 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 102 Type *SrcIVTy = 103 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 104 // Ask VMCore to do the conversion now that #elts line up. 105 C = ConstantExpr::getBitCast(C, SrcIVTy); 106 CV = dyn_cast<ConstantVector>(C); 107 if (!CV) // If VMCore wasn't able to fold it, bail out. 108 return C; 109 } 110 111 // Now we know that the input and output vectors are both integer vectors 112 // of the same size, and that their #elements is not the same. Do the 113 // conversion here, which depends on whether the input or output has 114 // more elements. 115 bool isLittleEndian = TD.isLittleEndian(); 116 117 SmallVector<Constant*, 32> Result; 118 if (NumDstElt < NumSrcElt) { 119 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 120 Constant *Zero = Constant::getNullValue(DstEltTy); 121 unsigned Ratio = NumSrcElt/NumDstElt; 122 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 123 unsigned SrcElt = 0; 124 for (unsigned i = 0; i != NumDstElt; ++i) { 125 // Build each element of the result. 126 Constant *Elt = Zero; 127 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 128 for (unsigned j = 0; j != Ratio; ++j) { 129 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++)); 130 if (!Src) // Reject constantexpr elements. 131 return ConstantExpr::getBitCast(C, DestTy); 132 133 // Zero extend the element to the right size. 134 Src = ConstantExpr::getZExt(Src, Elt->getType()); 135 136 // Shift it to the right place, depending on endianness. 137 Src = ConstantExpr::getShl(Src, 138 ConstantInt::get(Src->getType(), ShiftAmt)); 139 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 140 141 // Mix it in. 142 Elt = ConstantExpr::getOr(Elt, Src); 143 } 144 Result.push_back(Elt); 145 } 146 } else { 147 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 148 unsigned Ratio = NumDstElt/NumSrcElt; 149 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); 150 151 // Loop over each source value, expanding into multiple results. 152 for (unsigned i = 0; i != NumSrcElt; ++i) { 153 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i)); 154 if (!Src) // Reject constantexpr elements. 155 return ConstantExpr::getBitCast(C, DestTy); 156 157 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 158 for (unsigned j = 0; j != Ratio; ++j) { 159 // Shift the piece of the value into the right place, depending on 160 // endianness. 161 Constant *Elt = ConstantExpr::getLShr(Src, 162 ConstantInt::get(Src->getType(), ShiftAmt)); 163 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 164 165 // Truncate and remember this piece. 166 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 167 } 168 } 169 } 170 171 return ConstantVector::get(Result); 172} 173 174 175/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset 176/// from a global, return the global and the constant. Because of 177/// constantexprs, this function is recursive. 178static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 179 int64_t &Offset, const TargetData &TD) { 180 // Trivial case, constant is the global. 181 if ((GV = dyn_cast<GlobalValue>(C))) { 182 Offset = 0; 183 return true; 184 } 185 186 // Otherwise, if this isn't a constant expr, bail out. 187 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 188 if (!CE) return false; 189 190 // Look through ptr->int and ptr->ptr casts. 191 if (CE->getOpcode() == Instruction::PtrToInt || 192 CE->getOpcode() == Instruction::BitCast) 193 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); 194 195 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 196 if (CE->getOpcode() == Instruction::GetElementPtr) { 197 // Cannot compute this if the element type of the pointer is missing size 198 // info. 199 if (!cast<PointerType>(CE->getOperand(0)->getType()) 200 ->getElementType()->isSized()) 201 return false; 202 203 // If the base isn't a global+constant, we aren't either. 204 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) 205 return false; 206 207 // Otherwise, add any offset that our operands provide. 208 gep_type_iterator GTI = gep_type_begin(CE); 209 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end(); 210 i != e; ++i, ++GTI) { 211 ConstantInt *CI = dyn_cast<ConstantInt>(*i); 212 if (!CI) return false; // Index isn't a simple constant? 213 if (CI->isZero()) continue; // Not adding anything. 214 215 if (StructType *ST = dyn_cast<StructType>(*GTI)) { 216 // N = N + Offset 217 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); 218 } else { 219 SequentialType *SQT = cast<SequentialType>(*GTI); 220 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); 221 } 222 } 223 return true; 224 } 225 226 return false; 227} 228 229/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the 230/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the 231/// pointer to copy results into and BytesLeft is the number of bytes left in 232/// the CurPtr buffer. TD is the target data. 233static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, 234 unsigned char *CurPtr, unsigned BytesLeft, 235 const TargetData &TD) { 236 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && 237 "Out of range access"); 238 239 // If this element is zero or undefined, we can just return since *CurPtr is 240 // zero initialized. 241 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 242 return true; 243 244 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 245 if (CI->getBitWidth() > 64 || 246 (CI->getBitWidth() & 7) != 0) 247 return false; 248 249 uint64_t Val = CI->getZExtValue(); 250 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 251 252 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 253 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8)); 254 ++ByteOffset; 255 } 256 return true; 257 } 258 259 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 260 if (CFP->getType()->isDoubleTy()) { 261 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); 262 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 263 } 264 if (CFP->getType()->isFloatTy()){ 265 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); 266 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 267 } 268 return false; 269 } 270 271 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 272 const StructLayout *SL = TD.getStructLayout(CS->getType()); 273 unsigned Index = SL->getElementContainingOffset(ByteOffset); 274 uint64_t CurEltOffset = SL->getElementOffset(Index); 275 ByteOffset -= CurEltOffset; 276 277 while (1) { 278 // If the element access is to the element itself and not to tail padding, 279 // read the bytes from the element. 280 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); 281 282 if (ByteOffset < EltSize && 283 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 284 BytesLeft, TD)) 285 return false; 286 287 ++Index; 288 289 // Check to see if we read from the last struct element, if so we're done. 290 if (Index == CS->getType()->getNumElements()) 291 return true; 292 293 // If we read all of the bytes we needed from this element we're done. 294 uint64_t NextEltOffset = SL->getElementOffset(Index); 295 296 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset) 297 return true; 298 299 // Move to the next element of the struct. 300 CurPtr += NextEltOffset-CurEltOffset-ByteOffset; 301 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset; 302 ByteOffset = 0; 303 CurEltOffset = NextEltOffset; 304 } 305 // not reached. 306 } 307 308 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) { 309 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType()); 310 uint64_t Index = ByteOffset / EltSize; 311 uint64_t Offset = ByteOffset - Index * EltSize; 312 for (; Index != CA->getType()->getNumElements(); ++Index) { 313 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr, 314 BytesLeft, TD)) 315 return false; 316 if (EltSize >= BytesLeft) 317 return true; 318 319 Offset = 0; 320 BytesLeft -= EltSize; 321 CurPtr += EltSize; 322 } 323 return true; 324 } 325 326 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) { 327 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType()); 328 uint64_t Index = ByteOffset / EltSize; 329 uint64_t Offset = ByteOffset - Index * EltSize; 330 for (; Index != CV->getType()->getNumElements(); ++Index) { 331 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr, 332 BytesLeft, TD)) 333 return false; 334 if (EltSize >= BytesLeft) 335 return true; 336 337 Offset = 0; 338 BytesLeft -= EltSize; 339 CurPtr += EltSize; 340 } 341 return true; 342 } 343 344 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 345 if (CE->getOpcode() == Instruction::IntToPtr && 346 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext())) 347 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, 348 BytesLeft, TD); 349 } 350 351 // Otherwise, unknown initializer type. 352 return false; 353} 354 355static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 356 const TargetData &TD) { 357 Type *LoadTy = cast<PointerType>(C->getType())->getElementType(); 358 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 359 360 // If this isn't an integer load we can't fold it directly. 361 if (!IntType) { 362 // If this is a float/double load, we can try folding it as an int32/64 load 363 // and then bitcast the result. This can be useful for union cases. Note 364 // that address spaces don't matter here since we're not going to result in 365 // an actual new load. 366 Type *MapTy; 367 if (LoadTy->isFloatTy()) 368 MapTy = Type::getInt32PtrTy(C->getContext()); 369 else if (LoadTy->isDoubleTy()) 370 MapTy = Type::getInt64PtrTy(C->getContext()); 371 else if (LoadTy->isVectorTy()) { 372 MapTy = IntegerType::get(C->getContext(), 373 TD.getTypeAllocSizeInBits(LoadTy)); 374 MapTy = PointerType::getUnqual(MapTy); 375 } else 376 return 0; 377 378 C = FoldBitCast(C, MapTy, TD); 379 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 380 return FoldBitCast(Res, LoadTy, TD); 381 return 0; 382 } 383 384 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 385 if (BytesLoaded > 32 || BytesLoaded == 0) return 0; 386 387 GlobalValue *GVal; 388 int64_t Offset; 389 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 390 return 0; 391 392 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 393 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 394 !GV->getInitializer()->getType()->isSized()) 395 return 0; 396 397 // If we're loading off the beginning of the global, some bytes may be valid, 398 // but we don't try to handle this. 399 if (Offset < 0) return 0; 400 401 // If we're not accessing anything in this constant, the result is undefined. 402 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) 403 return UndefValue::get(IntType); 404 405 unsigned char RawBytes[32] = {0}; 406 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, 407 BytesLoaded, TD)) 408 return 0; 409 410 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); 411 for (unsigned i = 1; i != BytesLoaded; ++i) { 412 ResultVal <<= 8; 413 ResultVal |= RawBytes[BytesLoaded-1-i]; 414 } 415 416 return ConstantInt::get(IntType->getContext(), ResultVal); 417} 418 419/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 420/// produce if it is constant and determinable. If this is not determinable, 421/// return null. 422Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 423 const TargetData *TD) { 424 // First, try the easy cases: 425 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 426 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 427 return GV->getInitializer(); 428 429 // If the loaded value isn't a constant expr, we can't handle it. 430 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 431 if (!CE) return 0; 432 433 if (CE->getOpcode() == Instruction::GetElementPtr) { 434 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 435 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 436 if (Constant *V = 437 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 438 return V; 439 } 440 441 // Instead of loading constant c string, use corresponding integer value 442 // directly if string length is small enough. 443 std::string Str; 444 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) { 445 unsigned StrLen = Str.length(); 446 Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 447 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 448 // Replace load with immediate integer if the result is an integer or fp 449 // value. 450 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 451 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 452 APInt StrVal(NumBits, 0); 453 APInt SingleChar(NumBits, 0); 454 if (TD->isLittleEndian()) { 455 for (signed i = StrLen-1; i >= 0; i--) { 456 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 457 StrVal = (StrVal << 8) | SingleChar; 458 } 459 } else { 460 for (unsigned i = 0; i < StrLen; i++) { 461 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 462 StrVal = (StrVal << 8) | SingleChar; 463 } 464 // Append NULL at the end. 465 SingleChar = 0; 466 StrVal = (StrVal << 8) | SingleChar; 467 } 468 469 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 470 if (Ty->isFloatingPointTy()) 471 Res = ConstantExpr::getBitCast(Res, Ty); 472 return Res; 473 } 474 } 475 476 // If this load comes from anywhere in a constant global, and if the global 477 // is all undef or zero, we know what it loads. 478 if (GlobalVariable *GV = 479 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) { 480 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 481 Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 482 if (GV->getInitializer()->isNullValue()) 483 return Constant::getNullValue(ResTy); 484 if (isa<UndefValue>(GV->getInitializer())) 485 return UndefValue::get(ResTy); 486 } 487 } 488 489 // Try hard to fold loads from bitcasted strange and non-type-safe things. We 490 // currently don't do any of this for big endian systems. It can be 491 // generalized in the future if someone is interested. 492 if (TD && TD->isLittleEndian()) 493 return FoldReinterpretLoadFromConstPtr(CE, *TD); 494 return 0; 495} 496 497static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ 498 if (LI->isVolatile()) return 0; 499 500 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 501 return ConstantFoldLoadFromConstPtr(C, TD); 502 503 return 0; 504} 505 506/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 507/// Attempt to symbolically evaluate the result of a binary operator merging 508/// these together. If target data info is available, it is provided as TD, 509/// otherwise TD is null. 510static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 511 Constant *Op1, const TargetData *TD){ 512 // SROA 513 514 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 515 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 516 // bits. 517 518 519 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 520 // constant. This happens frequently when iterating over a global array. 521 if (Opc == Instruction::Sub && TD) { 522 GlobalValue *GV1, *GV2; 523 int64_t Offs1, Offs2; 524 525 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) 526 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && 527 GV1 == GV2) { 528 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 529 return ConstantInt::get(Op0->getType(), Offs1-Offs2); 530 } 531 } 532 533 return 0; 534} 535 536/// CastGEPIndices - If array indices are not pointer-sized integers, 537/// explicitly cast them so that they aren't implicitly casted by the 538/// getelementptr. 539static Constant *CastGEPIndices(ArrayRef<Constant *> Ops, 540 Type *ResultTy, 541 const TargetData *TD) { 542 if (!TD) return 0; 543 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); 544 545 bool Any = false; 546 SmallVector<Constant*, 32> NewIdxs; 547 for (unsigned i = 1, e = Ops.size(); i != e; ++i) { 548 if ((i == 1 || 549 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(), 550 Ops.slice(1, 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); 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(ArrayRef<Constant *> Ops, 574 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 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, e = Ops.size(); i != e; ++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 (Ops.size() == 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 = 609 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(), 610 makeArrayRef((Value **)Ops.data() + 1, 611 Ops.size() - 1))); 612 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 613 614 // If this is a GEP of a GEP, fold it all into a single GEP. 615 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 616 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end()); 617 618 // Do not try the incorporate the sub-GEP if some index is not a number. 619 bool AllConstantInt = true; 620 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 621 if (!isa<ConstantInt>(NestedOps[i])) { 622 AllConstantInt = false; 623 break; 624 } 625 if (!AllConstantInt) 626 break; 627 628 Ptr = cast<Constant>(GEP->getOperand(0)); 629 Offset += APInt(BitWidth, 630 TD->getIndexedOffset(Ptr->getType(), NestedOps)); 631 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 632 } 633 634 // If the base value for this address is a literal integer value, fold the 635 // getelementptr to the resulting integer value casted to the pointer type. 636 APInt BasePtr(BitWidth, 0); 637 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 638 if (CE->getOpcode() == Instruction::IntToPtr) 639 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) 640 BasePtr = Base->getValue().zextOrTrunc(BitWidth); 641 if (Ptr->isNullValue() || BasePtr != 0) { 642 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); 643 return ConstantExpr::getIntToPtr(C, ResultTy); 644 } 645 646 // Otherwise form a regular getelementptr. Recompute the indices so that 647 // we eliminate over-indexing of the notional static type array bounds. 648 // This makes it easy to determine if the getelementptr is "inbounds". 649 // Also, this helps GlobalOpt do SROA on GlobalVariables. 650 Type *Ty = Ptr->getType(); 651 SmallVector<Constant*, 32> NewIdxs; 652 do { 653 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 654 if (ATy->isPointerTy()) { 655 // The only pointer indexing we'll do is on the first index of the GEP. 656 if (!NewIdxs.empty()) 657 break; 658 659 // Only handle pointers to sized types, not pointers to functions. 660 if (!ATy->getElementType()->isSized()) 661 return 0; 662 } 663 664 // Determine which element of the array the offset points into. 665 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 666 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 667 if (ElemSize == 0) 668 // The element size is 0. This may be [0 x Ty]*, so just use a zero 669 // index for this level and proceed to the next level to see if it can 670 // accommodate the offset. 671 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 672 else { 673 // The element size is non-zero divide the offset by the element 674 // size (rounding down), to compute the index at this level. 675 APInt NewIdx = Offset.udiv(ElemSize); 676 Offset -= NewIdx * ElemSize; 677 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 678 } 679 Ty = ATy->getElementType(); 680 } else if (StructType *STy = dyn_cast<StructType>(Ty)) { 681 // Determine which field of the struct the offset points into. The 682 // getZExtValue is at least as safe as the StructLayout API because we 683 // know the offset is within the struct at this point. 684 const StructLayout &SL = *TD->getStructLayout(STy); 685 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 686 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 687 ElIdx)); 688 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 689 Ty = STy->getTypeAtIndex(ElIdx); 690 } else { 691 // We've reached some non-indexable type. 692 break; 693 } 694 } while (Ty != cast<PointerType>(ResultTy)->getElementType()); 695 696 // If we haven't used up the entire offset by descending the static 697 // type, then the offset is pointing into the middle of an indivisible 698 // member, so we can't simplify it. 699 if (Offset != 0) 700 return 0; 701 702 // Create a GEP. 703 Constant *C = 704 ConstantExpr::getGetElementPtr(Ptr, NewIdxs); 705 assert(cast<PointerType>(C->getType())->getElementType() == Ty && 706 "Computed GetElementPtr has unexpected type!"); 707 708 // If we ended up indexing a member with a type that doesn't match 709 // the type of what the original indices indexed, add a cast. 710 if (Ty != cast<PointerType>(ResultTy)->getElementType()) 711 C = FoldBitCast(C, ResultTy, *TD); 712 713 return C; 714} 715 716 717 718//===----------------------------------------------------------------------===// 719// Constant Folding public APIs 720//===----------------------------------------------------------------------===// 721 722/// ConstantFoldInstruction - Try to constant fold the specified instruction. 723/// If successful, the constant result is returned, if not, null is returned. 724/// Note that this fails if not all of the operands are constant. Otherwise, 725/// this function can only fail when attempting to fold instructions like loads 726/// and stores, which have no constant expression form. 727Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { 728 // Handle PHI nodes quickly here... 729 if (PHINode *PN = dyn_cast<PHINode>(I)) { 730 Constant *CommonValue = 0; 731 732 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 733 Value *Incoming = PN->getIncomingValue(i); 734 // If the incoming value is undef then skip it. Note that while we could 735 // skip the value if it is equal to the phi node itself we choose not to 736 // because that would break the rule that constant folding only applies if 737 // all operands are constants. 738 if (isa<UndefValue>(Incoming)) 739 continue; 740 // If the incoming value is not a constant, or is a different constant to 741 // the one we saw previously, then give up. 742 Constant *C = dyn_cast<Constant>(Incoming); 743 if (!C || (CommonValue && C != CommonValue)) 744 return 0; 745 CommonValue = C; 746 } 747 748 // If we reach here, all incoming values are the same constant or undef. 749 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 750 } 751 752 // Scan the operand list, checking to see if they are all constants, if so, 753 // hand off to ConstantFoldInstOperands. 754 SmallVector<Constant*, 8> Ops; 755 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 756 if (Constant *Op = dyn_cast<Constant>(*i)) 757 Ops.push_back(Op); 758 else 759 return 0; // All operands not constant! 760 761 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 762 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 763 TD); 764 765 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 766 return ConstantFoldLoadInst(LI, TD); 767 768 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) 769 return ConstantExpr::getInsertValue( 770 cast<Constant>(IVI->getAggregateOperand()), 771 cast<Constant>(IVI->getInsertedValueOperand()), 772 IVI->getIndices()); 773 774 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) 775 return ConstantExpr::getExtractValue( 776 cast<Constant>(EVI->getAggregateOperand()), 777 EVI->getIndices()); 778 779 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD); 780} 781 782/// ConstantFoldConstantExpression - Attempt to fold the constant expression 783/// using the specified TargetData. If successful, the constant result is 784/// result is returned, if not, null is returned. 785Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 786 const TargetData *TD) { 787 SmallVector<Constant*, 8> Ops; 788 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); 789 i != e; ++i) { 790 Constant *NewC = cast<Constant>(*i); 791 // Recursively fold the ConstantExpr's operands. 792 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) 793 NewC = ConstantFoldConstantExpression(NewCE, TD); 794 Ops.push_back(NewC); 795 } 796 797 if (CE->isCompare()) 798 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 799 TD); 800 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD); 801} 802 803/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 804/// specified opcode and operands. If successful, the constant result is 805/// returned, if not, null is returned. Note that this function can fail when 806/// attempting to fold instructions like loads and stores, which have no 807/// constant expression form. 808/// 809/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 810/// information, due to only being passed an opcode and operands. Constant 811/// folding using this function strips this information. 812/// 813Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy, 814 ArrayRef<Constant *> Ops, 815 const TargetData *TD) { 816 // Handle easy binops first. 817 if (Instruction::isBinaryOp(Opcode)) { 818 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) 819 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 820 return C; 821 822 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 823 } 824 825 switch (Opcode) { 826 default: return 0; 827 case Instruction::ICmp: 828 case Instruction::FCmp: assert(0 && "Invalid for compares"); 829 case Instruction::Call: 830 if (Function *F = dyn_cast<Function>(Ops.back())) 831 if (canConstantFoldCallTo(F)) 832 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1)); 833 return 0; 834 case Instruction::PtrToInt: 835 // If the input is a inttoptr, eliminate the pair. This requires knowing 836 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 837 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 838 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 839 Constant *Input = CE->getOperand(0); 840 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 841 if (TD->getPointerSizeInBits() < InWidth) { 842 Constant *Mask = 843 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 844 TD->getPointerSizeInBits())); 845 Input = ConstantExpr::getAnd(Input, Mask); 846 } 847 // Do a zext or trunc to get to the dest size. 848 return ConstantExpr::getIntegerCast(Input, DestTy, false); 849 } 850 } 851 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 852 case Instruction::IntToPtr: 853 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 854 // the int size is >= the ptr size. This requires knowing the width of a 855 // pointer, so it can't be done in ConstantExpr::getCast. 856 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) 857 if (TD && 858 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && 859 CE->getOpcode() == Instruction::PtrToInt) 860 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 861 862 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 863 case Instruction::Trunc: 864 case Instruction::ZExt: 865 case Instruction::SExt: 866 case Instruction::FPTrunc: 867 case Instruction::FPExt: 868 case Instruction::UIToFP: 869 case Instruction::SIToFP: 870 case Instruction::FPToUI: 871 case Instruction::FPToSI: 872 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 873 case Instruction::BitCast: 874 if (TD) 875 return FoldBitCast(Ops[0], DestTy, *TD); 876 return ConstantExpr::getBitCast(Ops[0], DestTy); 877 case Instruction::Select: 878 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 879 case Instruction::ExtractElement: 880 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 881 case Instruction::InsertElement: 882 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 883 case Instruction::ShuffleVector: 884 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 885 case Instruction::GetElementPtr: 886 if (Constant *C = CastGEPIndices(Ops, DestTy, TD)) 887 return C; 888 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD)) 889 return C; 890 891 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1)); 892 } 893} 894 895/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 896/// instruction (icmp/fcmp) with the specified operands. If it fails, it 897/// returns a constant expression of the specified operands. 898/// 899Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 900 Constant *Ops0, Constant *Ops1, 901 const TargetData *TD) { 902 // fold: icmp (inttoptr x), null -> icmp x, 0 903 // fold: icmp (ptrtoint x), 0 -> icmp x, null 904 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 905 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 906 // 907 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 908 // around to know if bit truncation is happening. 909 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 910 if (TD && Ops1->isNullValue()) { 911 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 912 if (CE0->getOpcode() == Instruction::IntToPtr) { 913 // Convert the integer value to the right size to ensure we get the 914 // proper extension or truncation. 915 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 916 IntPtrTy, false); 917 Constant *Null = Constant::getNullValue(C->getType()); 918 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 919 } 920 921 // Only do this transformation if the int is intptrty in size, otherwise 922 // there is a truncation or extension that we aren't modeling. 923 if (CE0->getOpcode() == Instruction::PtrToInt && 924 CE0->getType() == IntPtrTy) { 925 Constant *C = CE0->getOperand(0); 926 Constant *Null = Constant::getNullValue(C->getType()); 927 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 928 } 929 } 930 931 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 932 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 933 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 934 935 if (CE0->getOpcode() == Instruction::IntToPtr) { 936 // Convert the integer value to the right size to ensure we get the 937 // proper extension or truncation. 938 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 939 IntPtrTy, false); 940 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 941 IntPtrTy, false); 942 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); 943 } 944 945 // Only do this transformation if the int is intptrty in size, otherwise 946 // there is a truncation or extension that we aren't modeling. 947 if ((CE0->getOpcode() == Instruction::PtrToInt && 948 CE0->getType() == IntPtrTy && 949 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) 950 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), 951 CE1->getOperand(0), TD); 952 } 953 } 954 955 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 956 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 957 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 958 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 959 Constant *LHS = 960 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); 961 Constant *RHS = 962 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); 963 unsigned OpC = 964 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 965 Constant *Ops[] = { LHS, RHS }; 966 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD); 967 } 968 } 969 970 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 971} 972 973 974/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 975/// getelementptr constantexpr, return the constant value being addressed by the 976/// constant expression, or null if something is funny and we can't decide. 977Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 978 ConstantExpr *CE) { 979 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) 980 return 0; // Do not allow stepping over the value! 981 982 // Loop over all of the operands, tracking down which value we are 983 // addressing... 984 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 985 for (++I; I != E; ++I) 986 if (StructType *STy = dyn_cast<StructType>(*I)) { 987 ConstantInt *CU = cast<ConstantInt>(I.getOperand()); 988 assert(CU->getZExtValue() < STy->getNumElements() && 989 "Struct index out of range!"); 990 unsigned El = (unsigned)CU->getZExtValue(); 991 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 992 C = CS->getOperand(El); 993 } else if (isa<ConstantAggregateZero>(C)) { 994 C = Constant::getNullValue(STy->getElementType(El)); 995 } else if (isa<UndefValue>(C)) { 996 C = UndefValue::get(STy->getElementType(El)); 997 } else { 998 return 0; 999 } 1000 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 1001 if (ArrayType *ATy = dyn_cast<ArrayType>(*I)) { 1002 if (CI->getZExtValue() >= ATy->getNumElements()) 1003 return 0; 1004 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) 1005 C = CA->getOperand(CI->getZExtValue()); 1006 else if (isa<ConstantAggregateZero>(C)) 1007 C = Constant::getNullValue(ATy->getElementType()); 1008 else if (isa<UndefValue>(C)) 1009 C = UndefValue::get(ATy->getElementType()); 1010 else 1011 return 0; 1012 } else if (VectorType *VTy = dyn_cast<VectorType>(*I)) { 1013 if (CI->getZExtValue() >= VTy->getNumElements()) 1014 return 0; 1015 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) 1016 C = CP->getOperand(CI->getZExtValue()); 1017 else if (isa<ConstantAggregateZero>(C)) 1018 C = Constant::getNullValue(VTy->getElementType()); 1019 else if (isa<UndefValue>(C)) 1020 C = UndefValue::get(VTy->getElementType()); 1021 else 1022 return 0; 1023 } else { 1024 return 0; 1025 } 1026 } else { 1027 return 0; 1028 } 1029 return C; 1030} 1031 1032 1033//===----------------------------------------------------------------------===// 1034// Constant Folding for Calls 1035// 1036 1037/// canConstantFoldCallTo - Return true if its even possible to fold a call to 1038/// the specified function. 1039bool 1040llvm::canConstantFoldCallTo(const Function *F) { 1041 switch (F->getIntrinsicID()) { 1042 case Intrinsic::sqrt: 1043 case Intrinsic::powi: 1044 case Intrinsic::bswap: 1045 case Intrinsic::ctpop: 1046 case Intrinsic::ctlz: 1047 case Intrinsic::cttz: 1048 case Intrinsic::sadd_with_overflow: 1049 case Intrinsic::uadd_with_overflow: 1050 case Intrinsic::ssub_with_overflow: 1051 case Intrinsic::usub_with_overflow: 1052 case Intrinsic::smul_with_overflow: 1053 case Intrinsic::umul_with_overflow: 1054 case Intrinsic::convert_from_fp16: 1055 case Intrinsic::convert_to_fp16: 1056 case Intrinsic::x86_sse_cvtss2si: 1057 case Intrinsic::x86_sse_cvtss2si64: 1058 case Intrinsic::x86_sse_cvttss2si: 1059 case Intrinsic::x86_sse_cvttss2si64: 1060 case Intrinsic::x86_sse2_cvtsd2si: 1061 case Intrinsic::x86_sse2_cvtsd2si64: 1062 case Intrinsic::x86_sse2_cvttsd2si: 1063 case Intrinsic::x86_sse2_cvttsd2si64: 1064 return true; 1065 default: 1066 return false; 1067 case 0: break; 1068 } 1069 1070 if (!F->hasName()) return false; 1071 StringRef Name = F->getName(); 1072 1073 // In these cases, the check of the length is required. We don't want to 1074 // return true for a name like "cos\0blah" which strcmp would return equal to 1075 // "cos", but has length 8. 1076 switch (Name[0]) { 1077 default: return false; 1078 case 'a': 1079 return Name == "acos" || Name == "asin" || 1080 Name == "atan" || Name == "atan2"; 1081 case 'c': 1082 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1083 case 'e': 1084 return Name == "exp" || Name == "exp2"; 1085 case 'f': 1086 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1087 case 'l': 1088 return Name == "log" || Name == "log10"; 1089 case 'p': 1090 return Name == "pow"; 1091 case 's': 1092 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1093 Name == "sinf" || Name == "sqrtf"; 1094 case 't': 1095 return Name == "tan" || Name == "tanh"; 1096 } 1097} 1098 1099static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1100 Type *Ty) { 1101 sys::llvm_fenv_clearexcept(); 1102 V = NativeFP(V); 1103 if (sys::llvm_fenv_testexcept()) { 1104 sys::llvm_fenv_clearexcept(); 1105 return 0; 1106 } 1107 1108 if (Ty->isFloatTy()) 1109 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1110 if (Ty->isDoubleTy()) 1111 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1112 llvm_unreachable("Can only constant fold float/double"); 1113 return 0; // dummy return to suppress warning 1114} 1115 1116static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1117 double V, double W, Type *Ty) { 1118 sys::llvm_fenv_clearexcept(); 1119 V = NativeFP(V, W); 1120 if (sys::llvm_fenv_testexcept()) { 1121 sys::llvm_fenv_clearexcept(); 1122 return 0; 1123 } 1124 1125 if (Ty->isFloatTy()) 1126 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1127 if (Ty->isDoubleTy()) 1128 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1129 llvm_unreachable("Can only constant fold float/double"); 1130 return 0; // dummy return to suppress warning 1131} 1132 1133/// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer 1134/// conversion of a constant floating point. If roundTowardZero is false, the 1135/// default IEEE rounding is used (toward nearest, ties to even). This matches 1136/// the behavior of the non-truncating SSE instructions in the default rounding 1137/// mode. The desired integer type Ty is used to select how many bits are 1138/// available for the result. Returns null if the conversion cannot be 1139/// performed, otherwise returns the Constant value resulting from the 1140/// conversion. 1141static Constant *ConstantFoldConvertToInt(ConstantFP *Op, bool roundTowardZero, 1142 Type *Ty) { 1143 assert(Op && "Called with NULL operand"); 1144 APFloat Val(Op->getValueAPF()); 1145 1146 // All of these conversion intrinsics form an integer of at most 64bits. 1147 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth(); 1148 assert(ResultWidth <= 64 && 1149 "Can only constant fold conversions to 64 and 32 bit ints"); 1150 1151 uint64_t UIntVal; 1152 bool isExact = false; 1153 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero 1154 : APFloat::rmNearestTiesToEven; 1155 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth, 1156 /*isSigned=*/true, mode, 1157 &isExact); 1158 if (status != APFloat::opOK && status != APFloat::opInexact) 1159 return 0; 1160 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true); 1161} 1162 1163/// ConstantFoldCall - Attempt to constant fold a call to the specified function 1164/// with the specified arguments, returning null if unsuccessful. 1165Constant * 1166llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands) { 1167 if (!F->hasName()) return 0; 1168 StringRef Name = F->getName(); 1169 1170 Type *Ty = F->getReturnType(); 1171 if (Operands.size() == 1) { 1172 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1173 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) { 1174 APFloat Val(Op->getValueAPF()); 1175 1176 bool lost = false; 1177 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1178 1179 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1180 } 1181 1182 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1183 return 0; 1184 1185 /// We only fold functions with finite arguments. Folding NaN and inf is 1186 /// likely to be aborted with an exception anyway, and some host libms 1187 /// have known errors raising exceptions. 1188 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1189 return 0; 1190 1191 /// Currently APFloat versions of these functions do not exist, so we use 1192 /// the host native double versions. Float versions are not called 1193 /// directly but for all these it is true (float)(f((double)arg)) == 1194 /// f(arg). Long double not supported yet. 1195 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : 1196 Op->getValueAPF().convertToDouble(); 1197 switch (Name[0]) { 1198 case 'a': 1199 if (Name == "acos") 1200 return ConstantFoldFP(acos, V, Ty); 1201 else if (Name == "asin") 1202 return ConstantFoldFP(asin, V, Ty); 1203 else if (Name == "atan") 1204 return ConstantFoldFP(atan, V, Ty); 1205 break; 1206 case 'c': 1207 if (Name == "ceil") 1208 return ConstantFoldFP(ceil, V, Ty); 1209 else if (Name == "cos") 1210 return ConstantFoldFP(cos, V, Ty); 1211 else if (Name == "cosh") 1212 return ConstantFoldFP(cosh, V, Ty); 1213 else if (Name == "cosf") 1214 return ConstantFoldFP(cos, V, Ty); 1215 break; 1216 case 'e': 1217 if (Name == "exp") 1218 return ConstantFoldFP(exp, V, Ty); 1219 1220 if (Name == "exp2") { 1221 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a 1222 // C99 library. 1223 return ConstantFoldBinaryFP(pow, 2.0, V, Ty); 1224 } 1225 break; 1226 case 'f': 1227 if (Name == "fabs") 1228 return ConstantFoldFP(fabs, V, Ty); 1229 else if (Name == "floor") 1230 return ConstantFoldFP(floor, V, Ty); 1231 break; 1232 case 'l': 1233 if (Name == "log" && V > 0) 1234 return ConstantFoldFP(log, V, Ty); 1235 else if (Name == "log10" && V > 0) 1236 return ConstantFoldFP(log10, V, Ty); 1237 else if (F->getIntrinsicID() == Intrinsic::sqrt && 1238 (Ty->isFloatTy() || Ty->isDoubleTy())) { 1239 if (V >= -0.0) 1240 return ConstantFoldFP(sqrt, V, Ty); 1241 else // Undefined 1242 return Constant::getNullValue(Ty); 1243 } 1244 break; 1245 case 's': 1246 if (Name == "sin") 1247 return ConstantFoldFP(sin, V, Ty); 1248 else if (Name == "sinh") 1249 return ConstantFoldFP(sinh, V, Ty); 1250 else if (Name == "sqrt" && V >= 0) 1251 return ConstantFoldFP(sqrt, V, Ty); 1252 else if (Name == "sqrtf" && V >= 0) 1253 return ConstantFoldFP(sqrt, V, Ty); 1254 else if (Name == "sinf") 1255 return ConstantFoldFP(sin, V, Ty); 1256 break; 1257 case 't': 1258 if (Name == "tan") 1259 return ConstantFoldFP(tan, V, Ty); 1260 else if (Name == "tanh") 1261 return ConstantFoldFP(tanh, V, Ty); 1262 break; 1263 default: 1264 break; 1265 } 1266 return 0; 1267 } 1268 1269 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1270 switch (F->getIntrinsicID()) { 1271 case Intrinsic::bswap: 1272 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1273 case Intrinsic::ctpop: 1274 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1275 case Intrinsic::cttz: 1276 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); 1277 case Intrinsic::ctlz: 1278 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); 1279 case Intrinsic::convert_from_fp16: { 1280 APFloat Val(Op->getValue()); 1281 1282 bool lost = false; 1283 APFloat::opStatus status = 1284 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1285 1286 // Conversion is always precise. 1287 (void)status; 1288 assert(status == APFloat::opOK && !lost && 1289 "Precision lost during fp16 constfolding"); 1290 1291 return ConstantFP::get(F->getContext(), Val); 1292 } 1293 default: 1294 return 0; 1295 } 1296 } 1297 1298 if (ConstantVector *Op = dyn_cast<ConstantVector>(Operands[0])) { 1299 switch (F->getIntrinsicID()) { 1300 default: break; 1301 case Intrinsic::x86_sse_cvtss2si: 1302 case Intrinsic::x86_sse_cvtss2si64: 1303 case Intrinsic::x86_sse2_cvtsd2si: 1304 case Intrinsic::x86_sse2_cvtsd2si64: 1305 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0))) 1306 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/false, Ty); 1307 case Intrinsic::x86_sse_cvttss2si: 1308 case Intrinsic::x86_sse_cvttss2si64: 1309 case Intrinsic::x86_sse2_cvttsd2si: 1310 case Intrinsic::x86_sse2_cvttsd2si64: 1311 if (ConstantFP *FPOp = dyn_cast<ConstantFP>(Op->getOperand(0))) 1312 return ConstantFoldConvertToInt(FPOp, /*roundTowardZero=*/true, Ty); 1313 } 1314 } 1315 1316 if (isa<UndefValue>(Operands[0])) { 1317 if (F->getIntrinsicID() == Intrinsic::bswap) 1318 return Operands[0]; 1319 return 0; 1320 } 1321 1322 return 0; 1323 } 1324 1325 if (Operands.size() == 2) { 1326 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1327 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1328 return 0; 1329 double Op1V = Ty->isFloatTy() ? 1330 (double)Op1->getValueAPF().convertToFloat() : 1331 Op1->getValueAPF().convertToDouble(); 1332 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1333 if (Op2->getType() != Op1->getType()) 1334 return 0; 1335 1336 double Op2V = Ty->isFloatTy() ? 1337 (double)Op2->getValueAPF().convertToFloat(): 1338 Op2->getValueAPF().convertToDouble(); 1339 1340 if (Name == "pow") 1341 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1342 if (Name == "fmod") 1343 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1344 if (Name == "atan2") 1345 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1346 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1347 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy()) 1348 return ConstantFP::get(F->getContext(), 1349 APFloat((float)std::pow((float)Op1V, 1350 (int)Op2C->getZExtValue()))); 1351 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy()) 1352 return ConstantFP::get(F->getContext(), 1353 APFloat((double)std::pow((double)Op1V, 1354 (int)Op2C->getZExtValue()))); 1355 } 1356 return 0; 1357 } 1358 1359 1360 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1361 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1362 switch (F->getIntrinsicID()) { 1363 default: break; 1364 case Intrinsic::sadd_with_overflow: 1365 case Intrinsic::uadd_with_overflow: 1366 case Intrinsic::ssub_with_overflow: 1367 case Intrinsic::usub_with_overflow: 1368 case Intrinsic::smul_with_overflow: 1369 case Intrinsic::umul_with_overflow: { 1370 APInt Res; 1371 bool Overflow; 1372 switch (F->getIntrinsicID()) { 1373 default: assert(0 && "Invalid case"); 1374 case Intrinsic::sadd_with_overflow: 1375 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); 1376 break; 1377 case Intrinsic::uadd_with_overflow: 1378 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); 1379 break; 1380 case Intrinsic::ssub_with_overflow: 1381 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); 1382 break; 1383 case Intrinsic::usub_with_overflow: 1384 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); 1385 break; 1386 case Intrinsic::smul_with_overflow: 1387 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); 1388 break; 1389 case Intrinsic::umul_with_overflow: 1390 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow); 1391 break; 1392 } 1393 Constant *Ops[] = { 1394 ConstantInt::get(F->getContext(), Res), 1395 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) 1396 }; 1397 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops); 1398 } 1399 } 1400 } 1401 1402 return 0; 1403 } 1404 return 0; 1405 } 1406 return 0; 1407} 1408