ConstantFolding.cpp revision c1da204c433459713cbf4c62f0addf458e6b4b9f
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/System/FEnv.h" 34#include <cerrno> 35#include <cmath> 36using namespace llvm; 37 38//===----------------------------------------------------------------------===// 39// Constant Folding internal helper functions 40//===----------------------------------------------------------------------===// 41 42/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with 43/// TargetData. This always returns a non-null constant, but it may be a 44/// ConstantExpr if unfoldable. 45static Constant *FoldBitCast(Constant *C, const Type *DestTy, 46 const TargetData &TD) { 47 48 // This only handles casts to vectors currently. 49 const VectorType *DestVTy = dyn_cast<VectorType>(DestTy); 50 if (DestVTy == 0) 51 return ConstantExpr::getBitCast(C, DestTy); 52 53 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 54 // vector so the code below can handle it uniformly. 55 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 56 Constant *Ops = C; // don't take the address of C! 57 return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD); 58 } 59 60 // If this is a bitcast from constant vector -> vector, fold it. 61 ConstantVector *CV = dyn_cast<ConstantVector>(C); 62 if (CV == 0) 63 return ConstantExpr::getBitCast(C, DestTy); 64 65 // If the element types match, VMCore can fold it. 66 unsigned NumDstElt = DestVTy->getNumElements(); 67 unsigned NumSrcElt = CV->getNumOperands(); 68 if (NumDstElt == NumSrcElt) 69 return ConstantExpr::getBitCast(C, DestTy); 70 71 const Type *SrcEltTy = CV->getType()->getElementType(); 72 const Type *DstEltTy = DestVTy->getElementType(); 73 74 // Otherwise, we're changing the number of elements in a vector, which 75 // requires endianness information to do the right thing. For example, 76 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 77 // folds to (little endian): 78 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 79 // and to (big endian): 80 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 81 82 // First thing is first. We only want to think about integer here, so if 83 // we have something in FP form, recast it as integer. 84 if (DstEltTy->isFloatingPointTy()) { 85 // Fold to an vector of integers with same size as our FP type. 86 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 87 const Type *DestIVTy = 88 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 89 // Recursively handle this integer conversion, if possible. 90 C = FoldBitCast(C, DestIVTy, TD); 91 if (!C) return ConstantExpr::getBitCast(C, DestTy); 92 93 // Finally, VMCore can handle this now that #elts line up. 94 return ConstantExpr::getBitCast(C, DestTy); 95 } 96 97 // Okay, we know the destination is integer, if the input is FP, convert 98 // it to integer first. 99 if (SrcEltTy->isFloatingPointTy()) { 100 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 101 const Type *SrcIVTy = 102 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 103 // Ask VMCore to do the conversion now that #elts line up. 104 C = ConstantExpr::getBitCast(C, SrcIVTy); 105 CV = dyn_cast<ConstantVector>(C); 106 if (!CV) // If VMCore wasn't able to fold it, bail out. 107 return C; 108 } 109 110 // Now we know that the input and output vectors are both integer vectors 111 // of the same size, and that their #elements is not the same. Do the 112 // conversion here, which depends on whether the input or output has 113 // more elements. 114 bool isLittleEndian = TD.isLittleEndian(); 115 116 SmallVector<Constant*, 32> Result; 117 if (NumDstElt < NumSrcElt) { 118 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 119 Constant *Zero = Constant::getNullValue(DstEltTy); 120 unsigned Ratio = NumSrcElt/NumDstElt; 121 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 122 unsigned SrcElt = 0; 123 for (unsigned i = 0; i != NumDstElt; ++i) { 124 // Build each element of the result. 125 Constant *Elt = Zero; 126 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 127 for (unsigned j = 0; j != Ratio; ++j) { 128 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++)); 129 if (!Src) // Reject constantexpr elements. 130 return ConstantExpr::getBitCast(C, DestTy); 131 132 // Zero extend the element to the right size. 133 Src = ConstantExpr::getZExt(Src, Elt->getType()); 134 135 // Shift it to the right place, depending on endianness. 136 Src = ConstantExpr::getShl(Src, 137 ConstantInt::get(Src->getType(), ShiftAmt)); 138 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 139 140 // Mix it in. 141 Elt = ConstantExpr::getOr(Elt, Src); 142 } 143 Result.push_back(Elt); 144 } 145 } else { 146 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 147 unsigned Ratio = NumDstElt/NumSrcElt; 148 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits(); 149 150 // Loop over each source value, expanding into multiple results. 151 for (unsigned i = 0; i != NumSrcElt; ++i) { 152 Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i)); 153 if (!Src) // Reject constantexpr elements. 154 return ConstantExpr::getBitCast(C, DestTy); 155 156 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 157 for (unsigned j = 0; j != Ratio; ++j) { 158 // Shift the piece of the value into the right place, depending on 159 // endianness. 160 Constant *Elt = ConstantExpr::getLShr(Src, 161 ConstantInt::get(Src->getType(), ShiftAmt)); 162 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 163 164 // Truncate and remember this piece. 165 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 166 } 167 } 168 } 169 170 return ConstantVector::get(Result.data(), Result.size()); 171} 172 173 174/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset 175/// from a global, return the global and the constant. Because of 176/// constantexprs, this function is recursive. 177static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 178 int64_t &Offset, const TargetData &TD) { 179 // Trivial case, constant is the global. 180 if ((GV = dyn_cast<GlobalValue>(C))) { 181 Offset = 0; 182 return true; 183 } 184 185 // Otherwise, if this isn't a constant expr, bail out. 186 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 187 if (!CE) return false; 188 189 // Look through ptr->int and ptr->ptr casts. 190 if (CE->getOpcode() == Instruction::PtrToInt || 191 CE->getOpcode() == Instruction::BitCast) 192 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD); 193 194 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 195 if (CE->getOpcode() == Instruction::GetElementPtr) { 196 // Cannot compute this if the element type of the pointer is missing size 197 // info. 198 if (!cast<PointerType>(CE->getOperand(0)->getType()) 199 ->getElementType()->isSized()) 200 return false; 201 202 // If the base isn't a global+constant, we aren't either. 203 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD)) 204 return false; 205 206 // Otherwise, add any offset that our operands provide. 207 gep_type_iterator GTI = gep_type_begin(CE); 208 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end(); 209 i != e; ++i, ++GTI) { 210 ConstantInt *CI = dyn_cast<ConstantInt>(*i); 211 if (!CI) return false; // Index isn't a simple constant? 212 if (CI->isZero()) continue; // Not adding anything. 213 214 if (const StructType *ST = dyn_cast<StructType>(*GTI)) { 215 // N = N + Offset 216 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue()); 217 } else { 218 const SequentialType *SQT = cast<SequentialType>(*GTI); 219 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue(); 220 } 221 } 222 return true; 223 } 224 225 return false; 226} 227 228/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the 229/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the 230/// pointer to copy results into and BytesLeft is the number of bytes left in 231/// the CurPtr buffer. TD is the target data. 232static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, 233 unsigned char *CurPtr, unsigned BytesLeft, 234 const TargetData &TD) { 235 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) && 236 "Out of range access"); 237 238 // If this element is zero or undefined, we can just return since *CurPtr is 239 // zero initialized. 240 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 241 return true; 242 243 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 244 if (CI->getBitWidth() > 64 || 245 (CI->getBitWidth() & 7) != 0) 246 return false; 247 248 uint64_t Val = CI->getZExtValue(); 249 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 250 251 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 252 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8)); 253 ++ByteOffset; 254 } 255 return true; 256 } 257 258 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 259 if (CFP->getType()->isDoubleTy()) { 260 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD); 261 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 262 } 263 if (CFP->getType()->isFloatTy()){ 264 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD); 265 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD); 266 } 267 return false; 268 } 269 270 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 271 const StructLayout *SL = TD.getStructLayout(CS->getType()); 272 unsigned Index = SL->getElementContainingOffset(ByteOffset); 273 uint64_t CurEltOffset = SL->getElementOffset(Index); 274 ByteOffset -= CurEltOffset; 275 276 while (1) { 277 // If the element access is to the element itself and not to tail padding, 278 // read the bytes from the element. 279 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType()); 280 281 if (ByteOffset < EltSize && 282 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 283 BytesLeft, TD)) 284 return false; 285 286 ++Index; 287 288 // Check to see if we read from the last struct element, if so we're done. 289 if (Index == CS->getType()->getNumElements()) 290 return true; 291 292 // If we read all of the bytes we needed from this element we're done. 293 uint64_t NextEltOffset = SL->getElementOffset(Index); 294 295 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset) 296 return true; 297 298 // Move to the next element of the struct. 299 CurPtr += NextEltOffset-CurEltOffset-ByteOffset; 300 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset; 301 ByteOffset = 0; 302 CurEltOffset = NextEltOffset; 303 } 304 // not reached. 305 } 306 307 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) { 308 uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType()); 309 uint64_t Index = ByteOffset / EltSize; 310 uint64_t Offset = ByteOffset - Index * EltSize; 311 for (; Index != CA->getType()->getNumElements(); ++Index) { 312 if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr, 313 BytesLeft, TD)) 314 return false; 315 if (EltSize >= BytesLeft) 316 return true; 317 318 Offset = 0; 319 BytesLeft -= EltSize; 320 CurPtr += EltSize; 321 } 322 return true; 323 } 324 325 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) { 326 uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType()); 327 uint64_t Index = ByteOffset / EltSize; 328 uint64_t Offset = ByteOffset - Index * EltSize; 329 for (; Index != CV->getType()->getNumElements(); ++Index) { 330 if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr, 331 BytesLeft, TD)) 332 return false; 333 if (EltSize >= BytesLeft) 334 return true; 335 336 Offset = 0; 337 BytesLeft -= EltSize; 338 CurPtr += EltSize; 339 } 340 return true; 341 } 342 343 // Otherwise, unknown initializer type. 344 return false; 345} 346 347static Constant *FoldReinterpretLoadFromConstPtr(Constant *C, 348 const TargetData &TD) { 349 const Type *LoadTy = cast<PointerType>(C->getType())->getElementType(); 350 const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy); 351 352 // If this isn't an integer load we can't fold it directly. 353 if (!IntType) { 354 // If this is a float/double load, we can try folding it as an int32/64 load 355 // and then bitcast the result. This can be useful for union cases. Note 356 // that address spaces don't matter here since we're not going to result in 357 // an actual new load. 358 const Type *MapTy; 359 if (LoadTy->isFloatTy()) 360 MapTy = Type::getInt32PtrTy(C->getContext()); 361 else if (LoadTy->isDoubleTy()) 362 MapTy = Type::getInt64PtrTy(C->getContext()); 363 else if (LoadTy->isVectorTy()) { 364 MapTy = IntegerType::get(C->getContext(), 365 TD.getTypeAllocSizeInBits(LoadTy)); 366 MapTy = PointerType::getUnqual(MapTy); 367 } else 368 return 0; 369 370 C = FoldBitCast(C, MapTy, TD); 371 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD)) 372 return FoldBitCast(Res, LoadTy, TD); 373 return 0; 374 } 375 376 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 377 if (BytesLoaded > 32 || BytesLoaded == 0) return 0; 378 379 GlobalValue *GVal; 380 int64_t Offset; 381 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD)) 382 return 0; 383 384 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal); 385 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 386 !GV->getInitializer()->getType()->isSized()) 387 return 0; 388 389 // If we're loading off the beginning of the global, some bytes may be valid, 390 // but we don't try to handle this. 391 if (Offset < 0) return 0; 392 393 // If we're not accessing anything in this constant, the result is undefined. 394 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType())) 395 return UndefValue::get(IntType); 396 397 unsigned char RawBytes[32] = {0}; 398 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes, 399 BytesLoaded, TD)) 400 return 0; 401 402 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]); 403 for (unsigned i = 1; i != BytesLoaded; ++i) { 404 ResultVal <<= 8; 405 ResultVal |= RawBytes[BytesLoaded-1-i]; 406 } 407 408 return ConstantInt::get(IntType->getContext(), ResultVal); 409} 410 411/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would 412/// produce if it is constant and determinable. If this is not determinable, 413/// return null. 414Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, 415 const TargetData *TD) { 416 // First, try the easy cases: 417 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 418 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 419 return GV->getInitializer(); 420 421 // If the loaded value isn't a constant expr, we can't handle it. 422 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 423 if (!CE) return 0; 424 425 if (CE->getOpcode() == Instruction::GetElementPtr) { 426 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) 427 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 428 if (Constant *V = 429 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 430 return V; 431 } 432 433 // Instead of loading constant c string, use corresponding integer value 434 // directly if string length is small enough. 435 std::string Str; 436 if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) { 437 unsigned StrLen = Str.length(); 438 const Type *Ty = cast<PointerType>(CE->getType())->getElementType(); 439 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 440 // Replace load with immediate integer if the result is an integer or fp 441 // value. 442 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 443 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 444 APInt StrVal(NumBits, 0); 445 APInt SingleChar(NumBits, 0); 446 if (TD->isLittleEndian()) { 447 for (signed i = StrLen-1; i >= 0; i--) { 448 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 449 StrVal = (StrVal << 8) | SingleChar; 450 } 451 } else { 452 for (unsigned i = 0; i < StrLen; i++) { 453 SingleChar = (uint64_t) Str[i] & UCHAR_MAX; 454 StrVal = (StrVal << 8) | SingleChar; 455 } 456 // Append NULL at the end. 457 SingleChar = 0; 458 StrVal = (StrVal << 8) | SingleChar; 459 } 460 461 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 462 if (Ty->isFloatingPointTy()) 463 Res = ConstantExpr::getBitCast(Res, Ty); 464 return Res; 465 } 466 } 467 468 // If this load comes from anywhere in a constant global, and if the global 469 // is all undef or zero, we know what it loads. 470 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){ 471 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 472 const Type *ResTy = cast<PointerType>(C->getType())->getElementType(); 473 if (GV->getInitializer()->isNullValue()) 474 return Constant::getNullValue(ResTy); 475 if (isa<UndefValue>(GV->getInitializer())) 476 return UndefValue::get(ResTy); 477 } 478 } 479 480 // Try hard to fold loads from bitcasted strange and non-type-safe things. We 481 // currently don't do any of this for big endian systems. It can be 482 // generalized in the future if someone is interested. 483 if (TD && TD->isLittleEndian()) 484 return FoldReinterpretLoadFromConstPtr(CE, *TD); 485 return 0; 486} 487 488static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){ 489 if (LI->isVolatile()) return 0; 490 491 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0))) 492 return ConstantFoldLoadFromConstPtr(C, TD); 493 494 return 0; 495} 496 497/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression. 498/// Attempt to symbolically evaluate the result of a binary operator merging 499/// these together. If target data info is available, it is provided as TD, 500/// otherwise TD is null. 501static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, 502 Constant *Op1, const TargetData *TD){ 503 // SROA 504 505 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 506 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 507 // bits. 508 509 510 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 511 // constant. This happens frequently when iterating over a global array. 512 if (Opc == Instruction::Sub && TD) { 513 GlobalValue *GV1, *GV2; 514 int64_t Offs1, Offs2; 515 516 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD)) 517 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) && 518 GV1 == GV2) { 519 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 520 return ConstantInt::get(Op0->getType(), Offs1-Offs2); 521 } 522 } 523 524 return 0; 525} 526 527/// CastGEPIndices - If array indices are not pointer-sized integers, 528/// explicitly cast them so that they aren't implicitly casted by the 529/// getelementptr. 530static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps, 531 const Type *ResultTy, 532 const TargetData *TD) { 533 if (!TD) return 0; 534 const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext()); 535 536 bool Any = false; 537 SmallVector<Constant*, 32> NewIdxs; 538 for (unsigned i = 1; i != NumOps; ++i) { 539 if ((i == 1 || 540 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(), 541 reinterpret_cast<Value *const *>(Ops+1), 542 i-1))) && 543 Ops[i]->getType() != IntPtrTy) { 544 Any = true; 545 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 546 true, 547 IntPtrTy, 548 true), 549 Ops[i], IntPtrTy)); 550 } else 551 NewIdxs.push_back(Ops[i]); 552 } 553 if (!Any) return 0; 554 555 Constant *C = 556 ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size()); 557 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 558 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD)) 559 C = Folded; 560 return C; 561} 562 563/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP 564/// constant expression, do so. 565static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps, 566 const Type *ResultTy, 567 const TargetData *TD) { 568 Constant *Ptr = Ops[0]; 569 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized()) 570 return 0; 571 572 unsigned BitWidth = 573 TD->getTypeSizeInBits(TD->getIntPtrType(Ptr->getContext())); 574 575 // If this is a constant expr gep that is effectively computing an 576 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' 577 for (unsigned i = 1; i != NumOps; ++i) 578 if (!isa<ConstantInt>(Ops[i])) 579 return 0; 580 581 APInt Offset = APInt(BitWidth, 582 TD->getIndexedOffset(Ptr->getType(), 583 (Value**)Ops+1, NumOps-1)); 584 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 585 586 // If this is a GEP of a GEP, fold it all into a single GEP. 587 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 588 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end()); 589 590 // Do not try the incorporate the sub-GEP if some index is not a number. 591 bool AllConstantInt = true; 592 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i) 593 if (!isa<ConstantInt>(NestedOps[i])) { 594 AllConstantInt = false; 595 break; 596 } 597 if (!AllConstantInt) 598 break; 599 600 Ptr = cast<Constant>(GEP->getOperand(0)); 601 Offset += APInt(BitWidth, 602 TD->getIndexedOffset(Ptr->getType(), 603 (Value**)NestedOps.data(), 604 NestedOps.size())); 605 Ptr = cast<Constant>(Ptr->stripPointerCasts()); 606 } 607 608 // If the base value for this address is a literal integer value, fold the 609 // getelementptr to the resulting integer value casted to the pointer type. 610 APInt BasePtr(BitWidth, 0); 611 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 612 if (CE->getOpcode() == Instruction::IntToPtr) 613 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) { 614 BasePtr = Base->getValue(); 615 BasePtr.zextOrTrunc(BitWidth); 616 } 617 if (Ptr->isNullValue() || BasePtr != 0) { 618 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr); 619 return ConstantExpr::getIntToPtr(C, ResultTy); 620 } 621 622 // Otherwise form a regular getelementptr. Recompute the indices so that 623 // we eliminate over-indexing of the notional static type array bounds. 624 // This makes it easy to determine if the getelementptr is "inbounds". 625 // Also, this helps GlobalOpt do SROA on GlobalVariables. 626 const Type *Ty = Ptr->getType(); 627 SmallVector<Constant*, 32> NewIdxs; 628 do { 629 if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) { 630 if (ATy->isPointerTy()) { 631 // The only pointer indexing we'll do is on the first index of the GEP. 632 if (!NewIdxs.empty()) 633 break; 634 635 // Only handle pointers to sized types, not pointers to functions. 636 if (!ATy->getElementType()->isSized()) 637 return 0; 638 } 639 640 // Determine which element of the array the offset points into. 641 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType())); 642 const IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext()); 643 if (ElemSize == 0) 644 // The element size is 0. This may be [0 x Ty]*, so just use a zero 645 // index for this level and procede to the next level to see if it can 646 // accomodate the offset. 647 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 648 else { 649 // The element size is non-zero divide the offset by the element 650 // size (rounding down), to compute the index at this level. 651 APInt NewIdx = Offset.udiv(ElemSize); 652 Offset -= NewIdx * ElemSize; 653 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 654 } 655 Ty = ATy->getElementType(); 656 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) { 657 // Determine which field of the struct the offset points into. The 658 // getZExtValue is at least as safe as the StructLayout API because we 659 // know the offset is within the struct at this point. 660 const StructLayout &SL = *TD->getStructLayout(STy); 661 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 662 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 663 ElIdx)); 664 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 665 Ty = STy->getTypeAtIndex(ElIdx); 666 } else { 667 // We've reached some non-indexable type. 668 break; 669 } 670 } while (Ty != cast<PointerType>(ResultTy)->getElementType()); 671 672 // If we haven't used up the entire offset by descending the static 673 // type, then the offset is pointing into the middle of an indivisible 674 // member, so we can't simplify it. 675 if (Offset != 0) 676 return 0; 677 678 // Create a GEP. 679 Constant *C = 680 ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size()); 681 assert(cast<PointerType>(C->getType())->getElementType() == Ty && 682 "Computed GetElementPtr has unexpected type!"); 683 684 // If we ended up indexing a member with a type that doesn't match 685 // the type of what the original indices indexed, add a cast. 686 if (Ty != cast<PointerType>(ResultTy)->getElementType()) 687 C = FoldBitCast(C, ResultTy, *TD); 688 689 return C; 690} 691 692 693 694//===----------------------------------------------------------------------===// 695// Constant Folding public APIs 696//===----------------------------------------------------------------------===// 697 698 699/// ConstantFoldInstruction - Attempt to constant fold the specified 700/// instruction. If successful, the constant result is returned, if not, null 701/// is returned. Note that this function can only fail when attempting to fold 702/// instructions like loads and stores, which have no constant expression form. 703/// 704Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) { 705 // Handle PHI nodes specially here... 706 if (PHINode *PN = dyn_cast<PHINode>(I)) { 707 Constant *CommonValue = 0; 708 709 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 710 Value *Incoming = PN->getIncomingValue(i); 711 // If the incoming value is equal to the phi node itself or is undef then 712 // skip it. 713 if (Incoming == PN || isa<UndefValue>(Incoming)) 714 continue; 715 // If the incoming value is not a constant, or is a different constant to 716 // the one we saw previously, then give up. 717 Constant *C = dyn_cast<Constant>(Incoming); 718 if (!C || (CommonValue && C != CommonValue)) 719 return 0; 720 CommonValue = C; 721 } 722 723 // If we reach here, all incoming values are the same constant or undef. 724 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 725 } 726 727 // Scan the operand list, checking to see if they are all constants, if so, 728 // hand off to ConstantFoldInstOperands. 729 SmallVector<Constant*, 8> Ops; 730 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 731 if (Constant *Op = dyn_cast<Constant>(*i)) 732 Ops.push_back(Op); 733 else 734 return 0; // All operands not constant! 735 736 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 737 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 738 TD); 739 740 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 741 return ConstantFoldLoadInst(LI, TD); 742 743 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), 744 Ops.data(), Ops.size(), TD); 745} 746 747/// ConstantFoldConstantExpression - Attempt to fold the constant expression 748/// using the specified TargetData. If successful, the constant result is 749/// result is returned, if not, null is returned. 750Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE, 751 const TargetData *TD) { 752 SmallVector<Constant*, 8> Ops; 753 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); 754 i != e; ++i) { 755 Constant *NewC = cast<Constant>(*i); 756 // Recursively fold the ConstantExpr's operands. 757 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) 758 NewC = ConstantFoldConstantExpression(NewCE, TD); 759 Ops.push_back(NewC); 760 } 761 762 if (CE->isCompare()) 763 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 764 TD); 765 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), 766 Ops.data(), Ops.size(), TD); 767} 768 769/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the 770/// specified opcode and operands. If successful, the constant result is 771/// returned, if not, null is returned. Note that this function can fail when 772/// attempting to fold instructions like loads and stores, which have no 773/// constant expression form. 774/// 775/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc 776/// information, due to only being passed an opcode and operands. Constant 777/// folding using this function strips this information. 778/// 779Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy, 780 Constant* const* Ops, unsigned NumOps, 781 const TargetData *TD) { 782 // Handle easy binops first. 783 if (Instruction::isBinaryOp(Opcode)) { 784 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) 785 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD)) 786 return C; 787 788 return ConstantExpr::get(Opcode, Ops[0], Ops[1]); 789 } 790 791 switch (Opcode) { 792 default: return 0; 793 case Instruction::ICmp: 794 case Instruction::FCmp: assert(0 && "Invalid for compares"); 795 case Instruction::Call: 796 if (Function *F = dyn_cast<Function>(Ops[NumOps - 1])) 797 if (canConstantFoldCallTo(F)) 798 return ConstantFoldCall(F, Ops, NumOps - 1); 799 return 0; 800 case Instruction::PtrToInt: 801 // If the input is a inttoptr, eliminate the pair. This requires knowing 802 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 803 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) { 804 if (TD && CE->getOpcode() == Instruction::IntToPtr) { 805 Constant *Input = CE->getOperand(0); 806 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 807 if (TD->getPointerSizeInBits() < InWidth) { 808 Constant *Mask = 809 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth, 810 TD->getPointerSizeInBits())); 811 Input = ConstantExpr::getAnd(Input, Mask); 812 } 813 // Do a zext or trunc to get to the dest size. 814 return ConstantExpr::getIntegerCast(Input, DestTy, false); 815 } 816 } 817 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 818 case Instruction::IntToPtr: 819 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 820 // the int size is >= the ptr size. This requires knowing the width of a 821 // pointer, so it can't be done in ConstantExpr::getCast. 822 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) 823 if (TD && 824 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() && 825 CE->getOpcode() == Instruction::PtrToInt) 826 return FoldBitCast(CE->getOperand(0), DestTy, *TD); 827 828 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 829 case Instruction::Trunc: 830 case Instruction::ZExt: 831 case Instruction::SExt: 832 case Instruction::FPTrunc: 833 case Instruction::FPExt: 834 case Instruction::UIToFP: 835 case Instruction::SIToFP: 836 case Instruction::FPToUI: 837 case Instruction::FPToSI: 838 return ConstantExpr::getCast(Opcode, Ops[0], DestTy); 839 case Instruction::BitCast: 840 if (TD) 841 return FoldBitCast(Ops[0], DestTy, *TD); 842 return ConstantExpr::getBitCast(Ops[0], DestTy); 843 case Instruction::Select: 844 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 845 case Instruction::ExtractElement: 846 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 847 case Instruction::InsertElement: 848 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 849 case Instruction::ShuffleVector: 850 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 851 case Instruction::GetElementPtr: 852 if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD)) 853 return C; 854 if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD)) 855 return C; 856 857 return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1); 858 } 859} 860 861/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare 862/// instruction (icmp/fcmp) with the specified operands. If it fails, it 863/// returns a constant expression of the specified operands. 864/// 865Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 866 Constant *Ops0, Constant *Ops1, 867 const TargetData *TD) { 868 // fold: icmp (inttoptr x), null -> icmp x, 0 869 // fold: icmp (ptrtoint x), 0 -> icmp x, null 870 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 871 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 872 // 873 // ConstantExpr::getCompare cannot do this, because it doesn't have TD 874 // around to know if bit truncation is happening. 875 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 876 if (TD && Ops1->isNullValue()) { 877 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 878 if (CE0->getOpcode() == Instruction::IntToPtr) { 879 // Convert the integer value to the right size to ensure we get the 880 // proper extension or truncation. 881 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 882 IntPtrTy, false); 883 Constant *Null = Constant::getNullValue(C->getType()); 884 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 885 } 886 887 // Only do this transformation if the int is intptrty in size, otherwise 888 // there is a truncation or extension that we aren't modeling. 889 if (CE0->getOpcode() == Instruction::PtrToInt && 890 CE0->getType() == IntPtrTy) { 891 Constant *C = CE0->getOperand(0); 892 Constant *Null = Constant::getNullValue(C->getType()); 893 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD); 894 } 895 } 896 897 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 898 if (TD && CE0->getOpcode() == CE1->getOpcode()) { 899 const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext()); 900 901 if (CE0->getOpcode() == Instruction::IntToPtr) { 902 // Convert the integer value to the right size to ensure we get the 903 // proper extension or truncation. 904 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 905 IntPtrTy, false); 906 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 907 IntPtrTy, false); 908 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD); 909 } 910 911 // Only do this transformation if the int is intptrty in size, otherwise 912 // there is a truncation or extension that we aren't modeling. 913 if ((CE0->getOpcode() == Instruction::PtrToInt && 914 CE0->getType() == IntPtrTy && 915 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) 916 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), 917 CE1->getOperand(0), TD); 918 } 919 } 920 921 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 922 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 923 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 924 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 925 Constant *LHS = 926 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD); 927 Constant *RHS = 928 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD); 929 unsigned OpC = 930 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 931 Constant *Ops[] = { LHS, RHS }; 932 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD); 933 } 934 } 935 936 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 937} 938 939 940/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a 941/// getelementptr constantexpr, return the constant value being addressed by the 942/// constant expression, or null if something is funny and we can't decide. 943Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 944 ConstantExpr *CE) { 945 if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType())) 946 return 0; // Do not allow stepping over the value! 947 948 // Loop over all of the operands, tracking down which value we are 949 // addressing... 950 gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 951 for (++I; I != E; ++I) 952 if (const StructType *STy = dyn_cast<StructType>(*I)) { 953 ConstantInt *CU = cast<ConstantInt>(I.getOperand()); 954 assert(CU->getZExtValue() < STy->getNumElements() && 955 "Struct index out of range!"); 956 unsigned El = (unsigned)CU->getZExtValue(); 957 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) { 958 C = CS->getOperand(El); 959 } else if (isa<ConstantAggregateZero>(C)) { 960 C = Constant::getNullValue(STy->getElementType(El)); 961 } else if (isa<UndefValue>(C)) { 962 C = UndefValue::get(STy->getElementType(El)); 963 } else { 964 return 0; 965 } 966 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 967 if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) { 968 if (CI->getZExtValue() >= ATy->getNumElements()) 969 return 0; 970 if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) 971 C = CA->getOperand(CI->getZExtValue()); 972 else if (isa<ConstantAggregateZero>(C)) 973 C = Constant::getNullValue(ATy->getElementType()); 974 else if (isa<UndefValue>(C)) 975 C = UndefValue::get(ATy->getElementType()); 976 else 977 return 0; 978 } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) { 979 if (CI->getZExtValue() >= VTy->getNumElements()) 980 return 0; 981 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) 982 C = CP->getOperand(CI->getZExtValue()); 983 else if (isa<ConstantAggregateZero>(C)) 984 C = Constant::getNullValue(VTy->getElementType()); 985 else if (isa<UndefValue>(C)) 986 C = UndefValue::get(VTy->getElementType()); 987 else 988 return 0; 989 } else { 990 return 0; 991 } 992 } else { 993 return 0; 994 } 995 return C; 996} 997 998 999//===----------------------------------------------------------------------===// 1000// Constant Folding for Calls 1001// 1002 1003/// canConstantFoldCallTo - Return true if its even possible to fold a call to 1004/// the specified function. 1005bool 1006llvm::canConstantFoldCallTo(const Function *F) { 1007 switch (F->getIntrinsicID()) { 1008 case Intrinsic::sqrt: 1009 case Intrinsic::powi: 1010 case Intrinsic::bswap: 1011 case Intrinsic::ctpop: 1012 case Intrinsic::ctlz: 1013 case Intrinsic::cttz: 1014 case Intrinsic::uadd_with_overflow: 1015 case Intrinsic::usub_with_overflow: 1016 case Intrinsic::sadd_with_overflow: 1017 case Intrinsic::ssub_with_overflow: 1018 case Intrinsic::smul_with_overflow: 1019 case Intrinsic::convert_from_fp16: 1020 case Intrinsic::convert_to_fp16: 1021 return true; 1022 default: 1023 return false; 1024 case 0: break; 1025 } 1026 1027 if (!F->hasName()) return false; 1028 StringRef Name = F->getName(); 1029 1030 // In these cases, the check of the length is required. We don't want to 1031 // return true for a name like "cos\0blah" which strcmp would return equal to 1032 // "cos", but has length 8. 1033 switch (Name[0]) { 1034 default: return false; 1035 case 'a': 1036 return Name == "acos" || Name == "asin" || 1037 Name == "atan" || Name == "atan2"; 1038 case 'c': 1039 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh"; 1040 case 'e': 1041 return Name == "exp"; 1042 case 'f': 1043 return Name == "fabs" || Name == "fmod" || Name == "floor"; 1044 case 'l': 1045 return Name == "log" || Name == "log10"; 1046 case 'p': 1047 return Name == "pow"; 1048 case 's': 1049 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1050 Name == "sinf" || Name == "sqrtf"; 1051 case 't': 1052 return Name == "tan" || Name == "tanh"; 1053 } 1054} 1055 1056static Constant *ConstantFoldFP(double (*NativeFP)(double), double V, 1057 const Type *Ty) { 1058 sys::llvm_fenv_clearexcept(); 1059 V = NativeFP(V); 1060 if (sys::llvm_fenv_testexcept()) { 1061 sys::llvm_fenv_clearexcept(); 1062 return 0; 1063 } 1064 1065 if (Ty->isFloatTy()) 1066 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1067 if (Ty->isDoubleTy()) 1068 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1069 llvm_unreachable("Can only constant fold float/double"); 1070 return 0; // dummy return to suppress warning 1071} 1072 1073static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), 1074 double V, double W, const Type *Ty) { 1075 sys::llvm_fenv_clearexcept(); 1076 V = NativeFP(V, W); 1077 if (sys::llvm_fenv_testexcept()) { 1078 sys::llvm_fenv_clearexcept(); 1079 return 0; 1080 } 1081 1082 if (Ty->isFloatTy()) 1083 return ConstantFP::get(Ty->getContext(), APFloat((float)V)); 1084 if (Ty->isDoubleTy()) 1085 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1086 llvm_unreachable("Can only constant fold float/double"); 1087 return 0; // dummy return to suppress warning 1088} 1089 1090/// ConstantFoldCall - Attempt to constant fold a call to the specified function 1091/// with the specified arguments, returning null if unsuccessful. 1092Constant * 1093llvm::ConstantFoldCall(Function *F, 1094 Constant *const *Operands, unsigned NumOperands) { 1095 if (!F->hasName()) return 0; 1096 StringRef Name = F->getName(); 1097 1098 const Type *Ty = F->getReturnType(); 1099 if (NumOperands == 1) { 1100 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) { 1101 if (Name == "llvm.convert.to.fp16") { 1102 APFloat Val(Op->getValueAPF()); 1103 1104 bool lost = false; 1105 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost); 1106 1107 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt()); 1108 } 1109 1110 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1111 return 0; 1112 1113 /// We only fold functions with finite arguments. Folding NaN and inf is 1114 /// likely to be aborted with an exception anyway, and some host libms 1115 /// have known errors raising exceptions. 1116 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1117 return 0; 1118 1119 /// Currently APFloat versions of these functions do not exist, so we use 1120 /// the host native double versions. Float versions are not called 1121 /// directly but for all these it is true (float)(f((double)arg)) == 1122 /// f(arg). Long double not supported yet. 1123 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() : 1124 Op->getValueAPF().convertToDouble(); 1125 switch (Name[0]) { 1126 case 'a': 1127 if (Name == "acos") 1128 return ConstantFoldFP(acos, V, Ty); 1129 else if (Name == "asin") 1130 return ConstantFoldFP(asin, V, Ty); 1131 else if (Name == "atan") 1132 return ConstantFoldFP(atan, V, Ty); 1133 break; 1134 case 'c': 1135 if (Name == "ceil") 1136 return ConstantFoldFP(ceil, V, Ty); 1137 else if (Name == "cos") 1138 return ConstantFoldFP(cos, V, Ty); 1139 else if (Name == "cosh") 1140 return ConstantFoldFP(cosh, V, Ty); 1141 else if (Name == "cosf") 1142 return ConstantFoldFP(cos, V, Ty); 1143 break; 1144 case 'e': 1145 if (Name == "exp") 1146 return ConstantFoldFP(exp, V, Ty); 1147 break; 1148 case 'f': 1149 if (Name == "fabs") 1150 return ConstantFoldFP(fabs, V, Ty); 1151 else if (Name == "floor") 1152 return ConstantFoldFP(floor, V, Ty); 1153 break; 1154 case 'l': 1155 if (Name == "log" && V > 0) 1156 return ConstantFoldFP(log, V, Ty); 1157 else if (Name == "log10" && V > 0) 1158 return ConstantFoldFP(log10, V, Ty); 1159 else if (Name == "llvm.sqrt.f32" || 1160 Name == "llvm.sqrt.f64") { 1161 if (V >= -0.0) 1162 return ConstantFoldFP(sqrt, V, Ty); 1163 else // Undefined 1164 return Constant::getNullValue(Ty); 1165 } 1166 break; 1167 case 's': 1168 if (Name == "sin") 1169 return ConstantFoldFP(sin, V, Ty); 1170 else if (Name == "sinh") 1171 return ConstantFoldFP(sinh, V, Ty); 1172 else if (Name == "sqrt" && V >= 0) 1173 return ConstantFoldFP(sqrt, V, Ty); 1174 else if (Name == "sqrtf" && V >= 0) 1175 return ConstantFoldFP(sqrt, V, Ty); 1176 else if (Name == "sinf") 1177 return ConstantFoldFP(sin, V, Ty); 1178 break; 1179 case 't': 1180 if (Name == "tan") 1181 return ConstantFoldFP(tan, V, Ty); 1182 else if (Name == "tanh") 1183 return ConstantFoldFP(tanh, V, Ty); 1184 break; 1185 default: 1186 break; 1187 } 1188 return 0; 1189 } 1190 1191 1192 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) { 1193 if (Name.startswith("llvm.bswap")) 1194 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap()); 1195 else if (Name.startswith("llvm.ctpop")) 1196 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1197 else if (Name.startswith("llvm.cttz")) 1198 return ConstantInt::get(Ty, Op->getValue().countTrailingZeros()); 1199 else if (Name.startswith("llvm.ctlz")) 1200 return ConstantInt::get(Ty, Op->getValue().countLeadingZeros()); 1201 else if (Name == "llvm.convert.from.fp16") { 1202 APFloat Val(Op->getValue()); 1203 1204 bool lost = false; 1205 APFloat::opStatus status = 1206 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost); 1207 1208 // Conversion is always precise. 1209 status = status; 1210 assert(status == APFloat::opOK && !lost && 1211 "Precision lost during fp16 constfolding"); 1212 1213 return ConstantFP::get(F->getContext(), Val); 1214 } 1215 return 0; 1216 } 1217 1218 if (isa<UndefValue>(Operands[0])) { 1219 if (Name.startswith("llvm.bswap")) 1220 return Operands[0]; 1221 return 0; 1222 } 1223 1224 return 0; 1225 } 1226 1227 if (NumOperands == 2) { 1228 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1229 if (!Ty->isFloatTy() && !Ty->isDoubleTy()) 1230 return 0; 1231 double Op1V = Ty->isFloatTy() ? 1232 (double)Op1->getValueAPF().convertToFloat() : 1233 Op1->getValueAPF().convertToDouble(); 1234 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1235 if (Op2->getType() != Op1->getType()) 1236 return 0; 1237 1238 double Op2V = Ty->isFloatTy() ? 1239 (double)Op2->getValueAPF().convertToFloat(): 1240 Op2->getValueAPF().convertToDouble(); 1241 1242 if (Name == "pow") 1243 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1244 if (Name == "fmod") 1245 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 1246 if (Name == "atan2") 1247 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 1248 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 1249 if (Name == "llvm.powi.f32") 1250 return ConstantFP::get(F->getContext(), 1251 APFloat((float)std::pow((float)Op1V, 1252 (int)Op2C->getZExtValue()))); 1253 if (Name == "llvm.powi.f64") 1254 return ConstantFP::get(F->getContext(), 1255 APFloat((double)std::pow((double)Op1V, 1256 (int)Op2C->getZExtValue()))); 1257 } 1258 return 0; 1259 } 1260 1261 1262 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 1263 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 1264 switch (F->getIntrinsicID()) { 1265 default: break; 1266 case Intrinsic::sadd_with_overflow: 1267 case Intrinsic::uadd_with_overflow: 1268 case Intrinsic::ssub_with_overflow: 1269 case Intrinsic::usub_with_overflow: 1270 case Intrinsic::smul_with_overflow: { 1271 APInt Res; 1272 bool Overflow; 1273 switch (F->getIntrinsicID()) { 1274 default: assert(0 && "Invalid case"); 1275 case Intrinsic::sadd_with_overflow: 1276 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow); 1277 break; 1278 case Intrinsic::uadd_with_overflow: 1279 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow); 1280 break; 1281 case Intrinsic::ssub_with_overflow: 1282 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow); 1283 break; 1284 case Intrinsic::usub_with_overflow: 1285 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow); 1286 break; 1287 case Intrinsic::smul_with_overflow: 1288 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow); 1289 break; 1290 } 1291 Constant *Ops[] = { 1292 ConstantInt::get(F->getContext(), Res), 1293 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow) 1294 }; 1295 return ConstantStruct::get(F->getContext(), Ops, 2, false); 1296 } 1297 } 1298 } 1299 1300 return 0; 1301 } 1302 return 0; 1303 } 1304 return 0; 1305} 1306 1307