1 //===-- ConstantFolding.cpp - Fold instructions into constants ------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines routines for folding instructions into constants. 10 // 11 // Also, to supplement the basic IR ConstantExpr simplifications, 12 // this file defines some additional folding routines that can make use of 13 // DataLayout information. These functions cannot go in IR due to library 14 // dependency issues. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "llvm/Analysis/ConstantFolding.h" 19 #include "llvm/ADT/APFloat.h" 20 #include "llvm/ADT/APInt.h" 21 #include "llvm/ADT/ArrayRef.h" 22 #include "llvm/ADT/DenseMap.h" 23 #include "llvm/ADT/STLExtras.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/Analysis/TargetLibraryInfo.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/Analysis/VectorUtils.h" 29 #include "llvm/Config/config.h" 30 #include "llvm/IR/Constant.h" 31 #include "llvm/IR/Constants.h" 32 #include "llvm/IR/DataLayout.h" 33 #include "llvm/IR/DerivedTypes.h" 34 #include "llvm/IR/Function.h" 35 #include "llvm/IR/GlobalValue.h" 36 #include "llvm/IR/GlobalVariable.h" 37 #include "llvm/IR/InstrTypes.h" 38 #include "llvm/IR/Instruction.h" 39 #include "llvm/IR/Instructions.h" 40 #include "llvm/IR/Operator.h" 41 #include "llvm/IR/Type.h" 42 #include "llvm/IR/Value.h" 43 #include "llvm/Support/Casting.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/KnownBits.h" 46 #include "llvm/Support/MathExtras.h" 47 #include <cassert> 48 #include <cerrno> 49 #include <cfenv> 50 #include <cmath> 51 #include <cstddef> 52 #include <cstdint> 53 54 using namespace llvm; 55 56 namespace { 57 58 //===----------------------------------------------------------------------===// 59 // Constant Folding internal helper functions 60 //===----------------------------------------------------------------------===// 61 62 static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy, 63 Constant *C, Type *SrcEltTy, 64 unsigned NumSrcElts, 65 const DataLayout &DL) { 66 // Now that we know that the input value is a vector of integers, just shift 67 // and insert them into our result. 68 unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy); 69 for (unsigned i = 0; i != NumSrcElts; ++i) { 70 Constant *Element; 71 if (DL.isLittleEndian()) 72 Element = C->getAggregateElement(NumSrcElts - i - 1); 73 else 74 Element = C->getAggregateElement(i); 75 76 if (Element && isa<UndefValue>(Element)) { 77 Result <<= BitShift; 78 continue; 79 } 80 81 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); 82 if (!ElementCI) 83 return ConstantExpr::getBitCast(C, DestTy); 84 85 Result <<= BitShift; 86 Result |= ElementCI->getValue().zextOrSelf(Result.getBitWidth()); 87 } 88 89 return nullptr; 90 } 91 92 /// Constant fold bitcast, symbolically evaluating it with DataLayout. 93 /// This always returns a non-null constant, but it may be a 94 /// ConstantExpr if unfoldable. 95 Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) { 96 // Catch the obvious splat cases. 97 if (C->isNullValue() && !DestTy->isX86_MMXTy()) 98 return Constant::getNullValue(DestTy); 99 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() && 100 !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types! 101 return Constant::getAllOnesValue(DestTy); 102 103 if (auto *VTy = dyn_cast<VectorType>(C->getType())) { 104 // Handle a vector->scalar integer/fp cast. 105 if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) { 106 unsigned NumSrcElts = VTy->getNumElements(); 107 Type *SrcEltTy = VTy->getElementType(); 108 109 // If the vector is a vector of floating point, convert it to vector of int 110 // to simplify things. 111 if (SrcEltTy->isFloatingPointTy()) { 112 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 113 Type *SrcIVTy = 114 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts); 115 // Ask IR to do the conversion now that #elts line up. 116 C = ConstantExpr::getBitCast(C, SrcIVTy); 117 } 118 119 APInt Result(DL.getTypeSizeInBits(DestTy), 0); 120 if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C, 121 SrcEltTy, NumSrcElts, DL)) 122 return CE; 123 124 if (isa<IntegerType>(DestTy)) 125 return ConstantInt::get(DestTy, Result); 126 127 APFloat FP(DestTy->getFltSemantics(), Result); 128 return ConstantFP::get(DestTy->getContext(), FP); 129 } 130 } 131 132 // The code below only handles casts to vectors currently. 133 auto *DestVTy = dyn_cast<VectorType>(DestTy); 134 if (!DestVTy) 135 return ConstantExpr::getBitCast(C, DestTy); 136 137 // If this is a scalar -> vector cast, convert the input into a <1 x scalar> 138 // vector so the code below can handle it uniformly. 139 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) { 140 Constant *Ops = C; // don't take the address of C! 141 return FoldBitCast(ConstantVector::get(Ops), DestTy, DL); 142 } 143 144 // If this is a bitcast from constant vector -> vector, fold it. 145 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C)) 146 return ConstantExpr::getBitCast(C, DestTy); 147 148 // If the element types match, IR can fold it. 149 unsigned NumDstElt = DestVTy->getNumElements(); 150 unsigned NumSrcElt = C->getType()->getVectorNumElements(); 151 if (NumDstElt == NumSrcElt) 152 return ConstantExpr::getBitCast(C, DestTy); 153 154 Type *SrcEltTy = C->getType()->getVectorElementType(); 155 Type *DstEltTy = DestVTy->getElementType(); 156 157 // Otherwise, we're changing the number of elements in a vector, which 158 // requires endianness information to do the right thing. For example, 159 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 160 // folds to (little endian): 161 // <4 x i32> <i32 0, i32 0, i32 1, i32 0> 162 // and to (big endian): 163 // <4 x i32> <i32 0, i32 0, i32 0, i32 1> 164 165 // First thing is first. We only want to think about integer here, so if 166 // we have something in FP form, recast it as integer. 167 if (DstEltTy->isFloatingPointTy()) { 168 // Fold to an vector of integers with same size as our FP type. 169 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); 170 Type *DestIVTy = 171 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt); 172 // Recursively handle this integer conversion, if possible. 173 C = FoldBitCast(C, DestIVTy, DL); 174 175 // Finally, IR can handle this now that #elts line up. 176 return ConstantExpr::getBitCast(C, DestTy); 177 } 178 179 // Okay, we know the destination is integer, if the input is FP, convert 180 // it to integer first. 181 if (SrcEltTy->isFloatingPointTy()) { 182 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); 183 Type *SrcIVTy = 184 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt); 185 // Ask IR to do the conversion now that #elts line up. 186 C = ConstantExpr::getBitCast(C, SrcIVTy); 187 // If IR wasn't able to fold it, bail out. 188 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector. 189 !isa<ConstantDataVector>(C)) 190 return C; 191 } 192 193 // Now we know that the input and output vectors are both integer vectors 194 // of the same size, and that their #elements is not the same. Do the 195 // conversion here, which depends on whether the input or output has 196 // more elements. 197 bool isLittleEndian = DL.isLittleEndian(); 198 199 SmallVector<Constant*, 32> Result; 200 if (NumDstElt < NumSrcElt) { 201 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) 202 Constant *Zero = Constant::getNullValue(DstEltTy); 203 unsigned Ratio = NumSrcElt/NumDstElt; 204 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); 205 unsigned SrcElt = 0; 206 for (unsigned i = 0; i != NumDstElt; ++i) { 207 // Build each element of the result. 208 Constant *Elt = Zero; 209 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); 210 for (unsigned j = 0; j != Ratio; ++j) { 211 Constant *Src = C->getAggregateElement(SrcElt++); 212 if (Src && isa<UndefValue>(Src)) 213 Src = Constant::getNullValue(C->getType()->getVectorElementType()); 214 else 215 Src = dyn_cast_or_null<ConstantInt>(Src); 216 if (!Src) // Reject constantexpr elements. 217 return ConstantExpr::getBitCast(C, DestTy); 218 219 // Zero extend the element to the right size. 220 Src = ConstantExpr::getZExt(Src, Elt->getType()); 221 222 // Shift it to the right place, depending on endianness. 223 Src = ConstantExpr::getShl(Src, 224 ConstantInt::get(Src->getType(), ShiftAmt)); 225 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; 226 227 // Mix it in. 228 Elt = ConstantExpr::getOr(Elt, Src); 229 } 230 Result.push_back(Elt); 231 } 232 return ConstantVector::get(Result); 233 } 234 235 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) 236 unsigned Ratio = NumDstElt/NumSrcElt; 237 unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy); 238 239 // Loop over each source value, expanding into multiple results. 240 for (unsigned i = 0; i != NumSrcElt; ++i) { 241 auto *Element = C->getAggregateElement(i); 242 243 if (!Element) // Reject constantexpr elements. 244 return ConstantExpr::getBitCast(C, DestTy); 245 246 if (isa<UndefValue>(Element)) { 247 // Correctly Propagate undef values. 248 Result.append(Ratio, UndefValue::get(DstEltTy)); 249 continue; 250 } 251 252 auto *Src = dyn_cast<ConstantInt>(Element); 253 if (!Src) 254 return ConstantExpr::getBitCast(C, DestTy); 255 256 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); 257 for (unsigned j = 0; j != Ratio; ++j) { 258 // Shift the piece of the value into the right place, depending on 259 // endianness. 260 Constant *Elt = ConstantExpr::getLShr(Src, 261 ConstantInt::get(Src->getType(), ShiftAmt)); 262 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; 263 264 // Truncate the element to an integer with the same pointer size and 265 // convert the element back to a pointer using a inttoptr. 266 if (DstEltTy->isPointerTy()) { 267 IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize); 268 Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy); 269 Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy)); 270 continue; 271 } 272 273 // Truncate and remember this piece. 274 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy)); 275 } 276 } 277 278 return ConstantVector::get(Result); 279 } 280 281 } // end anonymous namespace 282 283 /// If this constant is a constant offset from a global, return the global and 284 /// the constant. Because of constantexprs, this function is recursive. 285 bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, 286 APInt &Offset, const DataLayout &DL) { 287 // Trivial case, constant is the global. 288 if ((GV = dyn_cast<GlobalValue>(C))) { 289 unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType()); 290 Offset = APInt(BitWidth, 0); 291 return true; 292 } 293 294 // Otherwise, if this isn't a constant expr, bail out. 295 auto *CE = dyn_cast<ConstantExpr>(C); 296 if (!CE) return false; 297 298 // Look through ptr->int and ptr->ptr casts. 299 if (CE->getOpcode() == Instruction::PtrToInt || 300 CE->getOpcode() == Instruction::BitCast) 301 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL); 302 303 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) 304 auto *GEP = dyn_cast<GEPOperator>(CE); 305 if (!GEP) 306 return false; 307 308 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType()); 309 APInt TmpOffset(BitWidth, 0); 310 311 // If the base isn't a global+constant, we aren't either. 312 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL)) 313 return false; 314 315 // Otherwise, add any offset that our operands provide. 316 if (!GEP->accumulateConstantOffset(DL, TmpOffset)) 317 return false; 318 319 Offset = TmpOffset; 320 return true; 321 } 322 323 Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy, 324 const DataLayout &DL) { 325 do { 326 Type *SrcTy = C->getType(); 327 328 // If the type sizes are the same and a cast is legal, just directly 329 // cast the constant. 330 if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) { 331 Instruction::CastOps Cast = Instruction::BitCast; 332 // If we are going from a pointer to int or vice versa, we spell the cast 333 // differently. 334 if (SrcTy->isIntegerTy() && DestTy->isPointerTy()) 335 Cast = Instruction::IntToPtr; 336 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy()) 337 Cast = Instruction::PtrToInt; 338 339 if (CastInst::castIsValid(Cast, C, DestTy)) 340 return ConstantExpr::getCast(Cast, C, DestTy); 341 } 342 343 // If this isn't an aggregate type, there is nothing we can do to drill down 344 // and find a bitcastable constant. 345 if (!SrcTy->isAggregateType()) 346 return nullptr; 347 348 // We're simulating a load through a pointer that was bitcast to point to 349 // a different type, so we can try to walk down through the initial 350 // elements of an aggregate to see if some part of the aggregate is 351 // castable to implement the "load" semantic model. 352 if (SrcTy->isStructTy()) { 353 // Struct types might have leading zero-length elements like [0 x i32], 354 // which are certainly not what we are looking for, so skip them. 355 unsigned Elem = 0; 356 Constant *ElemC; 357 do { 358 ElemC = C->getAggregateElement(Elem++); 359 } while (ElemC && DL.getTypeSizeInBits(ElemC->getType()) == 0); 360 C = ElemC; 361 } else { 362 C = C->getAggregateElement(0u); 363 } 364 } while (C); 365 366 return nullptr; 367 } 368 369 namespace { 370 371 /// Recursive helper to read bits out of global. C is the constant being copied 372 /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy 373 /// results into and BytesLeft is the number of bytes left in 374 /// the CurPtr buffer. DL is the DataLayout. 375 bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr, 376 unsigned BytesLeft, const DataLayout &DL) { 377 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) && 378 "Out of range access"); 379 380 // If this element is zero or undefined, we can just return since *CurPtr is 381 // zero initialized. 382 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) 383 return true; 384 385 if (auto *CI = dyn_cast<ConstantInt>(C)) { 386 if (CI->getBitWidth() > 64 || 387 (CI->getBitWidth() & 7) != 0) 388 return false; 389 390 uint64_t Val = CI->getZExtValue(); 391 unsigned IntBytes = unsigned(CI->getBitWidth()/8); 392 393 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { 394 int n = ByteOffset; 395 if (!DL.isLittleEndian()) 396 n = IntBytes - n - 1; 397 CurPtr[i] = (unsigned char)(Val >> (n * 8)); 398 ++ByteOffset; 399 } 400 return true; 401 } 402 403 if (auto *CFP = dyn_cast<ConstantFP>(C)) { 404 if (CFP->getType()->isDoubleTy()) { 405 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL); 406 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); 407 } 408 if (CFP->getType()->isFloatTy()){ 409 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL); 410 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); 411 } 412 if (CFP->getType()->isHalfTy()){ 413 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL); 414 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); 415 } 416 return false; 417 } 418 419 if (auto *CS = dyn_cast<ConstantStruct>(C)) { 420 const StructLayout *SL = DL.getStructLayout(CS->getType()); 421 unsigned Index = SL->getElementContainingOffset(ByteOffset); 422 uint64_t CurEltOffset = SL->getElementOffset(Index); 423 ByteOffset -= CurEltOffset; 424 425 while (true) { 426 // If the element access is to the element itself and not to tail padding, 427 // read the bytes from the element. 428 uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType()); 429 430 if (ByteOffset < EltSize && 431 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr, 432 BytesLeft, DL)) 433 return false; 434 435 ++Index; 436 437 // Check to see if we read from the last struct element, if so we're done. 438 if (Index == CS->getType()->getNumElements()) 439 return true; 440 441 // If we read all of the bytes we needed from this element we're done. 442 uint64_t NextEltOffset = SL->getElementOffset(Index); 443 444 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset) 445 return true; 446 447 // Move to the next element of the struct. 448 CurPtr += NextEltOffset - CurEltOffset - ByteOffset; 449 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset; 450 ByteOffset = 0; 451 CurEltOffset = NextEltOffset; 452 } 453 // not reached. 454 } 455 456 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) || 457 isa<ConstantDataSequential>(C)) { 458 Type *EltTy = C->getType()->getSequentialElementType(); 459 uint64_t EltSize = DL.getTypeAllocSize(EltTy); 460 uint64_t Index = ByteOffset / EltSize; 461 uint64_t Offset = ByteOffset - Index * EltSize; 462 uint64_t NumElts; 463 if (auto *AT = dyn_cast<ArrayType>(C->getType())) 464 NumElts = AT->getNumElements(); 465 else 466 NumElts = C->getType()->getVectorNumElements(); 467 468 for (; Index != NumElts; ++Index) { 469 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr, 470 BytesLeft, DL)) 471 return false; 472 473 uint64_t BytesWritten = EltSize - Offset; 474 assert(BytesWritten <= EltSize && "Not indexing into this element?"); 475 if (BytesWritten >= BytesLeft) 476 return true; 477 478 Offset = 0; 479 BytesLeft -= BytesWritten; 480 CurPtr += BytesWritten; 481 } 482 return true; 483 } 484 485 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 486 if (CE->getOpcode() == Instruction::IntToPtr && 487 CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) { 488 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr, 489 BytesLeft, DL); 490 } 491 } 492 493 // Otherwise, unknown initializer type. 494 return false; 495 } 496 497 Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy, 498 const DataLayout &DL) { 499 auto *PTy = cast<PointerType>(C->getType()); 500 auto *IntType = dyn_cast<IntegerType>(LoadTy); 501 502 // If this isn't an integer load we can't fold it directly. 503 if (!IntType) { 504 unsigned AS = PTy->getAddressSpace(); 505 506 // If this is a float/double load, we can try folding it as an int32/64 load 507 // and then bitcast the result. This can be useful for union cases. Note 508 // that address spaces don't matter here since we're not going to result in 509 // an actual new load. 510 Type *MapTy; 511 if (LoadTy->isHalfTy()) 512 MapTy = Type::getInt16Ty(C->getContext()); 513 else if (LoadTy->isFloatTy()) 514 MapTy = Type::getInt32Ty(C->getContext()); 515 else if (LoadTy->isDoubleTy()) 516 MapTy = Type::getInt64Ty(C->getContext()); 517 else if (LoadTy->isVectorTy()) { 518 MapTy = PointerType::getIntNTy(C->getContext(), 519 DL.getTypeSizeInBits(LoadTy)); 520 } else 521 return nullptr; 522 523 C = FoldBitCast(C, MapTy->getPointerTo(AS), DL); 524 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL)) 525 return FoldBitCast(Res, LoadTy, DL); 526 return nullptr; 527 } 528 529 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; 530 if (BytesLoaded > 32 || BytesLoaded == 0) 531 return nullptr; 532 533 GlobalValue *GVal; 534 APInt OffsetAI; 535 if (!IsConstantOffsetFromGlobal(C, GVal, OffsetAI, DL)) 536 return nullptr; 537 538 auto *GV = dyn_cast<GlobalVariable>(GVal); 539 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 540 !GV->getInitializer()->getType()->isSized()) 541 return nullptr; 542 543 int64_t Offset = OffsetAI.getSExtValue(); 544 int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType()); 545 546 // If we're not accessing anything in this constant, the result is undefined. 547 if (Offset + BytesLoaded <= 0) 548 return UndefValue::get(IntType); 549 550 // If we're not accessing anything in this constant, the result is undefined. 551 if (Offset >= InitializerSize) 552 return UndefValue::get(IntType); 553 554 unsigned char RawBytes[32] = {0}; 555 unsigned char *CurPtr = RawBytes; 556 unsigned BytesLeft = BytesLoaded; 557 558 // If we're loading off the beginning of the global, some bytes may be valid. 559 if (Offset < 0) { 560 CurPtr += -Offset; 561 BytesLeft += Offset; 562 Offset = 0; 563 } 564 565 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, CurPtr, BytesLeft, DL)) 566 return nullptr; 567 568 APInt ResultVal = APInt(IntType->getBitWidth(), 0); 569 if (DL.isLittleEndian()) { 570 ResultVal = RawBytes[BytesLoaded - 1]; 571 for (unsigned i = 1; i != BytesLoaded; ++i) { 572 ResultVal <<= 8; 573 ResultVal |= RawBytes[BytesLoaded - 1 - i]; 574 } 575 } else { 576 ResultVal = RawBytes[0]; 577 for (unsigned i = 1; i != BytesLoaded; ++i) { 578 ResultVal <<= 8; 579 ResultVal |= RawBytes[i]; 580 } 581 } 582 583 return ConstantInt::get(IntType->getContext(), ResultVal); 584 } 585 586 Constant *ConstantFoldLoadThroughBitcastExpr(ConstantExpr *CE, Type *DestTy, 587 const DataLayout &DL) { 588 auto *SrcPtr = CE->getOperand(0); 589 auto *SrcPtrTy = dyn_cast<PointerType>(SrcPtr->getType()); 590 if (!SrcPtrTy) 591 return nullptr; 592 Type *SrcTy = SrcPtrTy->getPointerElementType(); 593 594 Constant *C = ConstantFoldLoadFromConstPtr(SrcPtr, SrcTy, DL); 595 if (!C) 596 return nullptr; 597 598 return llvm::ConstantFoldLoadThroughBitcast(C, DestTy, DL); 599 } 600 601 } // end anonymous namespace 602 603 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, 604 const DataLayout &DL) { 605 // First, try the easy cases: 606 if (auto *GV = dyn_cast<GlobalVariable>(C)) 607 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 608 return GV->getInitializer(); 609 610 if (auto *GA = dyn_cast<GlobalAlias>(C)) 611 if (GA->getAliasee() && !GA->isInterposable()) 612 return ConstantFoldLoadFromConstPtr(GA->getAliasee(), Ty, DL); 613 614 // If the loaded value isn't a constant expr, we can't handle it. 615 auto *CE = dyn_cast<ConstantExpr>(C); 616 if (!CE) 617 return nullptr; 618 619 if (CE->getOpcode() == Instruction::GetElementPtr) { 620 if (auto *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) { 621 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 622 if (Constant *V = 623 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) 624 return V; 625 } 626 } 627 } 628 629 if (CE->getOpcode() == Instruction::BitCast) 630 if (Constant *LoadedC = ConstantFoldLoadThroughBitcastExpr(CE, Ty, DL)) 631 return LoadedC; 632 633 // Instead of loading constant c string, use corresponding integer value 634 // directly if string length is small enough. 635 StringRef Str; 636 if (getConstantStringInfo(CE, Str) && !Str.empty()) { 637 size_t StrLen = Str.size(); 638 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 639 // Replace load with immediate integer if the result is an integer or fp 640 // value. 641 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 && 642 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) { 643 APInt StrVal(NumBits, 0); 644 APInt SingleChar(NumBits, 0); 645 if (DL.isLittleEndian()) { 646 for (unsigned char C : reverse(Str.bytes())) { 647 SingleChar = static_cast<uint64_t>(C); 648 StrVal = (StrVal << 8) | SingleChar; 649 } 650 } else { 651 for (unsigned char C : Str.bytes()) { 652 SingleChar = static_cast<uint64_t>(C); 653 StrVal = (StrVal << 8) | SingleChar; 654 } 655 // Append NULL at the end. 656 SingleChar = 0; 657 StrVal = (StrVal << 8) | SingleChar; 658 } 659 660 Constant *Res = ConstantInt::get(CE->getContext(), StrVal); 661 if (Ty->isFloatingPointTy()) 662 Res = ConstantExpr::getBitCast(Res, Ty); 663 return Res; 664 } 665 } 666 667 // If this load comes from anywhere in a constant global, and if the global 668 // is all undef or zero, we know what it loads. 669 if (auto *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) { 670 if (GV->isConstant() && GV->hasDefinitiveInitializer()) { 671 if (GV->getInitializer()->isNullValue()) 672 return Constant::getNullValue(Ty); 673 if (isa<UndefValue>(GV->getInitializer())) 674 return UndefValue::get(Ty); 675 } 676 } 677 678 // Try hard to fold loads from bitcasted strange and non-type-safe things. 679 return FoldReinterpretLoadFromConstPtr(CE, Ty, DL); 680 } 681 682 namespace { 683 684 Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout &DL) { 685 if (LI->isVolatile()) return nullptr; 686 687 if (auto *C = dyn_cast<Constant>(LI->getOperand(0))) 688 return ConstantFoldLoadFromConstPtr(C, LI->getType(), DL); 689 690 return nullptr; 691 } 692 693 /// One of Op0/Op1 is a constant expression. 694 /// Attempt to symbolically evaluate the result of a binary operator merging 695 /// these together. If target data info is available, it is provided as DL, 696 /// otherwise DL is null. 697 Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1, 698 const DataLayout &DL) { 699 // SROA 700 701 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. 702 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute 703 // bits. 704 705 if (Opc == Instruction::And) { 706 KnownBits Known0 = computeKnownBits(Op0, DL); 707 KnownBits Known1 = computeKnownBits(Op1, DL); 708 if ((Known1.One | Known0.Zero).isAllOnesValue()) { 709 // All the bits of Op0 that the 'and' could be masking are already zero. 710 return Op0; 711 } 712 if ((Known0.One | Known1.Zero).isAllOnesValue()) { 713 // All the bits of Op1 that the 'and' could be masking are already zero. 714 return Op1; 715 } 716 717 Known0.Zero |= Known1.Zero; 718 Known0.One &= Known1.One; 719 if (Known0.isConstant()) 720 return ConstantInt::get(Op0->getType(), Known0.getConstant()); 721 } 722 723 // If the constant expr is something like &A[123] - &A[4].f, fold this into a 724 // constant. This happens frequently when iterating over a global array. 725 if (Opc == Instruction::Sub) { 726 GlobalValue *GV1, *GV2; 727 APInt Offs1, Offs2; 728 729 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL)) 730 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) { 731 unsigned OpSize = DL.getTypeSizeInBits(Op0->getType()); 732 733 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. 734 // PtrToInt may change the bitwidth so we have convert to the right size 735 // first. 736 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) - 737 Offs2.zextOrTrunc(OpSize)); 738 } 739 } 740 741 return nullptr; 742 } 743 744 /// If array indices are not pointer-sized integers, explicitly cast them so 745 /// that they aren't implicitly casted by the getelementptr. 746 Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops, 747 Type *ResultTy, Optional<unsigned> InRangeIndex, 748 const DataLayout &DL, const TargetLibraryInfo *TLI) { 749 Type *IntPtrTy = DL.getIntPtrType(ResultTy); 750 Type *IntPtrScalarTy = IntPtrTy->getScalarType(); 751 752 bool Any = false; 753 SmallVector<Constant*, 32> NewIdxs; 754 for (unsigned i = 1, e = Ops.size(); i != e; ++i) { 755 if ((i == 1 || 756 !isa<StructType>(GetElementPtrInst::getIndexedType( 757 SrcElemTy, Ops.slice(1, i - 1)))) && 758 Ops[i]->getType()->getScalarType() != IntPtrScalarTy) { 759 Any = true; 760 Type *NewType = Ops[i]->getType()->isVectorTy() 761 ? IntPtrTy 762 : IntPtrTy->getScalarType(); 763 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i], 764 true, 765 NewType, 766 true), 767 Ops[i], NewType)); 768 } else 769 NewIdxs.push_back(Ops[i]); 770 } 771 772 if (!Any) 773 return nullptr; 774 775 Constant *C = ConstantExpr::getGetElementPtr( 776 SrcElemTy, Ops[0], NewIdxs, /*InBounds=*/false, InRangeIndex); 777 if (Constant *Folded = ConstantFoldConstant(C, DL, TLI)) 778 C = Folded; 779 780 return C; 781 } 782 783 /// Strip the pointer casts, but preserve the address space information. 784 Constant* StripPtrCastKeepAS(Constant* Ptr, Type *&ElemTy) { 785 assert(Ptr->getType()->isPointerTy() && "Not a pointer type"); 786 auto *OldPtrTy = cast<PointerType>(Ptr->getType()); 787 Ptr = Ptr->stripPointerCasts(); 788 auto *NewPtrTy = cast<PointerType>(Ptr->getType()); 789 790 ElemTy = NewPtrTy->getPointerElementType(); 791 792 // Preserve the address space number of the pointer. 793 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) { 794 NewPtrTy = ElemTy->getPointerTo(OldPtrTy->getAddressSpace()); 795 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy); 796 } 797 return Ptr; 798 } 799 800 /// If we can symbolically evaluate the GEP constant expression, do so. 801 Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP, 802 ArrayRef<Constant *> Ops, 803 const DataLayout &DL, 804 const TargetLibraryInfo *TLI) { 805 const GEPOperator *InnermostGEP = GEP; 806 bool InBounds = GEP->isInBounds(); 807 808 Type *SrcElemTy = GEP->getSourceElementType(); 809 Type *ResElemTy = GEP->getResultElementType(); 810 Type *ResTy = GEP->getType(); 811 if (!SrcElemTy->isSized()) 812 return nullptr; 813 814 if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy, 815 GEP->getInRangeIndex(), DL, TLI)) 816 return C; 817 818 Constant *Ptr = Ops[0]; 819 if (!Ptr->getType()->isPointerTy()) 820 return nullptr; 821 822 Type *IntPtrTy = DL.getIntPtrType(Ptr->getType()); 823 824 // If this is a constant expr gep that is effectively computing an 825 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12' 826 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 827 if (!isa<ConstantInt>(Ops[i])) { 828 829 // If this is "gep i8* Ptr, (sub 0, V)", fold this as: 830 // "inttoptr (sub (ptrtoint Ptr), V)" 831 if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) { 832 auto *CE = dyn_cast<ConstantExpr>(Ops[1]); 833 assert((!CE || CE->getType() == IntPtrTy) && 834 "CastGEPIndices didn't canonicalize index types!"); 835 if (CE && CE->getOpcode() == Instruction::Sub && 836 CE->getOperand(0)->isNullValue()) { 837 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType()); 838 Res = ConstantExpr::getSub(Res, CE->getOperand(1)); 839 Res = ConstantExpr::getIntToPtr(Res, ResTy); 840 if (auto *FoldedRes = ConstantFoldConstant(Res, DL, TLI)) 841 Res = FoldedRes; 842 return Res; 843 } 844 } 845 return nullptr; 846 } 847 848 unsigned BitWidth = DL.getTypeSizeInBits(IntPtrTy); 849 APInt Offset = 850 APInt(BitWidth, 851 DL.getIndexedOffsetInType( 852 SrcElemTy, 853 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1))); 854 Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy); 855 856 // If this is a GEP of a GEP, fold it all into a single GEP. 857 while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { 858 InnermostGEP = GEP; 859 InBounds &= GEP->isInBounds(); 860 861 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end()); 862 863 // Do not try the incorporate the sub-GEP if some index is not a number. 864 bool AllConstantInt = true; 865 for (Value *NestedOp : NestedOps) 866 if (!isa<ConstantInt>(NestedOp)) { 867 AllConstantInt = false; 868 break; 869 } 870 if (!AllConstantInt) 871 break; 872 873 Ptr = cast<Constant>(GEP->getOperand(0)); 874 SrcElemTy = GEP->getSourceElementType(); 875 Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps)); 876 Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy); 877 } 878 879 // If the base value for this address is a literal integer value, fold the 880 // getelementptr to the resulting integer value casted to the pointer type. 881 APInt BasePtr(BitWidth, 0); 882 if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) { 883 if (CE->getOpcode() == Instruction::IntToPtr) { 884 if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) 885 BasePtr = Base->getValue().zextOrTrunc(BitWidth); 886 } 887 } 888 889 auto *PTy = cast<PointerType>(Ptr->getType()); 890 if ((Ptr->isNullValue() || BasePtr != 0) && 891 !DL.isNonIntegralPointerType(PTy)) { 892 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr); 893 return ConstantExpr::getIntToPtr(C, ResTy); 894 } 895 896 // Otherwise form a regular getelementptr. Recompute the indices so that 897 // we eliminate over-indexing of the notional static type array bounds. 898 // This makes it easy to determine if the getelementptr is "inbounds". 899 // Also, this helps GlobalOpt do SROA on GlobalVariables. 900 Type *Ty = PTy; 901 SmallVector<Constant *, 32> NewIdxs; 902 903 do { 904 if (!Ty->isStructTy()) { 905 if (Ty->isPointerTy()) { 906 // The only pointer indexing we'll do is on the first index of the GEP. 907 if (!NewIdxs.empty()) 908 break; 909 910 Ty = SrcElemTy; 911 912 // Only handle pointers to sized types, not pointers to functions. 913 if (!Ty->isSized()) 914 return nullptr; 915 } else if (auto *ATy = dyn_cast<SequentialType>(Ty)) { 916 Ty = ATy->getElementType(); 917 } else { 918 // We've reached some non-indexable type. 919 break; 920 } 921 922 // Determine which element of the array the offset points into. 923 APInt ElemSize(BitWidth, DL.getTypeAllocSize(Ty)); 924 if (ElemSize == 0) { 925 // The element size is 0. This may be [0 x Ty]*, so just use a zero 926 // index for this level and proceed to the next level to see if it can 927 // accommodate the offset. 928 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0)); 929 } else { 930 // The element size is non-zero divide the offset by the element 931 // size (rounding down), to compute the index at this level. 932 bool Overflow; 933 APInt NewIdx = Offset.sdiv_ov(ElemSize, Overflow); 934 if (Overflow) 935 break; 936 Offset -= NewIdx * ElemSize; 937 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx)); 938 } 939 } else { 940 auto *STy = cast<StructType>(Ty); 941 // If we end up with an offset that isn't valid for this struct type, we 942 // can't re-form this GEP in a regular form, so bail out. The pointer 943 // operand likely went through casts that are necessary to make the GEP 944 // sensible. 945 const StructLayout &SL = *DL.getStructLayout(STy); 946 if (Offset.isNegative() || Offset.uge(SL.getSizeInBytes())) 947 break; 948 949 // Determine which field of the struct the offset points into. The 950 // getZExtValue is fine as we've already ensured that the offset is 951 // within the range representable by the StructLayout API. 952 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue()); 953 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()), 954 ElIdx)); 955 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx)); 956 Ty = STy->getTypeAtIndex(ElIdx); 957 } 958 } while (Ty != ResElemTy); 959 960 // If we haven't used up the entire offset by descending the static 961 // type, then the offset is pointing into the middle of an indivisible 962 // member, so we can't simplify it. 963 if (Offset != 0) 964 return nullptr; 965 966 // Preserve the inrange index from the innermost GEP if possible. We must 967 // have calculated the same indices up to and including the inrange index. 968 Optional<unsigned> InRangeIndex; 969 if (Optional<unsigned> LastIRIndex = InnermostGEP->getInRangeIndex()) 970 if (SrcElemTy == InnermostGEP->getSourceElementType() && 971 NewIdxs.size() > *LastIRIndex) { 972 InRangeIndex = LastIRIndex; 973 for (unsigned I = 0; I <= *LastIRIndex; ++I) 974 if (NewIdxs[I] != InnermostGEP->getOperand(I + 1)) 975 return nullptr; 976 } 977 978 // Create a GEP. 979 Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs, 980 InBounds, InRangeIndex); 981 assert(C->getType()->getPointerElementType() == Ty && 982 "Computed GetElementPtr has unexpected type!"); 983 984 // If we ended up indexing a member with a type that doesn't match 985 // the type of what the original indices indexed, add a cast. 986 if (Ty != ResElemTy) 987 C = FoldBitCast(C, ResTy, DL); 988 989 return C; 990 } 991 992 /// Attempt to constant fold an instruction with the 993 /// specified opcode and operands. If successful, the constant result is 994 /// returned, if not, null is returned. Note that this function can fail when 995 /// attempting to fold instructions like loads and stores, which have no 996 /// constant expression form. 997 Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode, 998 ArrayRef<Constant *> Ops, 999 const DataLayout &DL, 1000 const TargetLibraryInfo *TLI) { 1001 Type *DestTy = InstOrCE->getType(); 1002 1003 if (Instruction::isUnaryOp(Opcode)) 1004 return ConstantFoldUnaryOpOperand(Opcode, Ops[0], DL); 1005 1006 if (Instruction::isBinaryOp(Opcode)) 1007 return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL); 1008 1009 if (Instruction::isCast(Opcode)) 1010 return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL); 1011 1012 if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) { 1013 if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI)) 1014 return C; 1015 1016 return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), Ops[0], 1017 Ops.slice(1), GEP->isInBounds(), 1018 GEP->getInRangeIndex()); 1019 } 1020 1021 if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE)) 1022 return CE->getWithOperands(Ops); 1023 1024 switch (Opcode) { 1025 default: return nullptr; 1026 case Instruction::ICmp: 1027 case Instruction::FCmp: llvm_unreachable("Invalid for compares"); 1028 case Instruction::Call: 1029 if (auto *F = dyn_cast<Function>(Ops.back())) { 1030 const auto *Call = cast<CallBase>(InstOrCE); 1031 if (canConstantFoldCallTo(Call, F)) 1032 return ConstantFoldCall(Call, F, Ops.slice(0, Ops.size() - 1), TLI); 1033 } 1034 return nullptr; 1035 case Instruction::Select: 1036 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]); 1037 case Instruction::ExtractElement: 1038 return ConstantExpr::getExtractElement(Ops[0], Ops[1]); 1039 case Instruction::ExtractValue: 1040 return ConstantExpr::getExtractValue( 1041 Ops[0], dyn_cast<ExtractValueInst>(InstOrCE)->getIndices()); 1042 case Instruction::InsertElement: 1043 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]); 1044 case Instruction::ShuffleVector: 1045 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]); 1046 } 1047 } 1048 1049 } // end anonymous namespace 1050 1051 //===----------------------------------------------------------------------===// 1052 // Constant Folding public APIs 1053 //===----------------------------------------------------------------------===// 1054 1055 namespace { 1056 1057 Constant * 1058 ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL, 1059 const TargetLibraryInfo *TLI, 1060 SmallDenseMap<Constant *, Constant *> &FoldedOps) { 1061 if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C)) 1062 return nullptr; 1063 1064 SmallVector<Constant *, 8> Ops; 1065 for (const Use &NewU : C->operands()) { 1066 auto *NewC = cast<Constant>(&NewU); 1067 // Recursively fold the ConstantExpr's operands. If we have already folded 1068 // a ConstantExpr, we don't have to process it again. 1069 if (isa<ConstantVector>(NewC) || isa<ConstantExpr>(NewC)) { 1070 auto It = FoldedOps.find(NewC); 1071 if (It == FoldedOps.end()) { 1072 if (auto *FoldedC = 1073 ConstantFoldConstantImpl(NewC, DL, TLI, FoldedOps)) { 1074 FoldedOps.insert({NewC, FoldedC}); 1075 NewC = FoldedC; 1076 } else { 1077 FoldedOps.insert({NewC, NewC}); 1078 } 1079 } else { 1080 NewC = It->second; 1081 } 1082 } 1083 Ops.push_back(NewC); 1084 } 1085 1086 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 1087 if (CE->isCompare()) 1088 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1], 1089 DL, TLI); 1090 1091 return ConstantFoldInstOperandsImpl(CE, CE->getOpcode(), Ops, DL, TLI); 1092 } 1093 1094 assert(isa<ConstantVector>(C)); 1095 return ConstantVector::get(Ops); 1096 } 1097 1098 } // end anonymous namespace 1099 1100 Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL, 1101 const TargetLibraryInfo *TLI) { 1102 // Handle PHI nodes quickly here... 1103 if (auto *PN = dyn_cast<PHINode>(I)) { 1104 Constant *CommonValue = nullptr; 1105 1106 SmallDenseMap<Constant *, Constant *> FoldedOps; 1107 for (Value *Incoming : PN->incoming_values()) { 1108 // If the incoming value is undef then skip it. Note that while we could 1109 // skip the value if it is equal to the phi node itself we choose not to 1110 // because that would break the rule that constant folding only applies if 1111 // all operands are constants. 1112 if (isa<UndefValue>(Incoming)) 1113 continue; 1114 // If the incoming value is not a constant, then give up. 1115 auto *C = dyn_cast<Constant>(Incoming); 1116 if (!C) 1117 return nullptr; 1118 // Fold the PHI's operands. 1119 if (auto *FoldedC = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps)) 1120 C = FoldedC; 1121 // If the incoming value is a different constant to 1122 // the one we saw previously, then give up. 1123 if (CommonValue && C != CommonValue) 1124 return nullptr; 1125 CommonValue = C; 1126 } 1127 1128 // If we reach here, all incoming values are the same constant or undef. 1129 return CommonValue ? CommonValue : UndefValue::get(PN->getType()); 1130 } 1131 1132 // Scan the operand list, checking to see if they are all constants, if so, 1133 // hand off to ConstantFoldInstOperandsImpl. 1134 if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); })) 1135 return nullptr; 1136 1137 SmallDenseMap<Constant *, Constant *> FoldedOps; 1138 SmallVector<Constant *, 8> Ops; 1139 for (const Use &OpU : I->operands()) { 1140 auto *Op = cast<Constant>(&OpU); 1141 // Fold the Instruction's operands. 1142 if (auto *FoldedOp = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps)) 1143 Op = FoldedOp; 1144 1145 Ops.push_back(Op); 1146 } 1147 1148 if (const auto *CI = dyn_cast<CmpInst>(I)) 1149 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1], 1150 DL, TLI); 1151 1152 if (const auto *LI = dyn_cast<LoadInst>(I)) 1153 return ConstantFoldLoadInst(LI, DL); 1154 1155 if (auto *IVI = dyn_cast<InsertValueInst>(I)) { 1156 return ConstantExpr::getInsertValue( 1157 cast<Constant>(IVI->getAggregateOperand()), 1158 cast<Constant>(IVI->getInsertedValueOperand()), 1159 IVI->getIndices()); 1160 } 1161 1162 if (auto *EVI = dyn_cast<ExtractValueInst>(I)) { 1163 return ConstantExpr::getExtractValue( 1164 cast<Constant>(EVI->getAggregateOperand()), 1165 EVI->getIndices()); 1166 } 1167 1168 return ConstantFoldInstOperands(I, Ops, DL, TLI); 1169 } 1170 1171 Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL, 1172 const TargetLibraryInfo *TLI) { 1173 SmallDenseMap<Constant *, Constant *> FoldedOps; 1174 return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps); 1175 } 1176 1177 Constant *llvm::ConstantFoldInstOperands(Instruction *I, 1178 ArrayRef<Constant *> Ops, 1179 const DataLayout &DL, 1180 const TargetLibraryInfo *TLI) { 1181 return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI); 1182 } 1183 1184 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate, 1185 Constant *Ops0, Constant *Ops1, 1186 const DataLayout &DL, 1187 const TargetLibraryInfo *TLI) { 1188 // fold: icmp (inttoptr x), null -> icmp x, 0 1189 // fold: icmp null, (inttoptr x) -> icmp 0, x 1190 // fold: icmp (ptrtoint x), 0 -> icmp x, null 1191 // fold: icmp 0, (ptrtoint x) -> icmp null, x 1192 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y 1193 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y 1194 // 1195 // FIXME: The following comment is out of data and the DataLayout is here now. 1196 // ConstantExpr::getCompare cannot do this, because it doesn't have DL 1197 // around to know if bit truncation is happening. 1198 if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) { 1199 if (Ops1->isNullValue()) { 1200 if (CE0->getOpcode() == Instruction::IntToPtr) { 1201 Type *IntPtrTy = DL.getIntPtrType(CE0->getType()); 1202 // Convert the integer value to the right size to ensure we get the 1203 // proper extension or truncation. 1204 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0), 1205 IntPtrTy, false); 1206 Constant *Null = Constant::getNullValue(C->getType()); 1207 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI); 1208 } 1209 1210 // Only do this transformation if the int is intptrty in size, otherwise 1211 // there is a truncation or extension that we aren't modeling. 1212 if (CE0->getOpcode() == Instruction::PtrToInt) { 1213 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType()); 1214 if (CE0->getType() == IntPtrTy) { 1215 Constant *C = CE0->getOperand(0); 1216 Constant *Null = Constant::getNullValue(C->getType()); 1217 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI); 1218 } 1219 } 1220 } 1221 1222 if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) { 1223 if (CE0->getOpcode() == CE1->getOpcode()) { 1224 if (CE0->getOpcode() == Instruction::IntToPtr) { 1225 Type *IntPtrTy = DL.getIntPtrType(CE0->getType()); 1226 1227 // Convert the integer value to the right size to ensure we get the 1228 // proper extension or truncation. 1229 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0), 1230 IntPtrTy, false); 1231 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0), 1232 IntPtrTy, false); 1233 return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI); 1234 } 1235 1236 // Only do this transformation if the int is intptrty in size, otherwise 1237 // there is a truncation or extension that we aren't modeling. 1238 if (CE0->getOpcode() == Instruction::PtrToInt) { 1239 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType()); 1240 if (CE0->getType() == IntPtrTy && 1241 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) { 1242 return ConstantFoldCompareInstOperands( 1243 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI); 1244 } 1245 } 1246 } 1247 } 1248 1249 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0) 1250 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0) 1251 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) && 1252 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) { 1253 Constant *LHS = ConstantFoldCompareInstOperands( 1254 Predicate, CE0->getOperand(0), Ops1, DL, TLI); 1255 Constant *RHS = ConstantFoldCompareInstOperands( 1256 Predicate, CE0->getOperand(1), Ops1, DL, TLI); 1257 unsigned OpC = 1258 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 1259 return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL); 1260 } 1261 } else if (isa<ConstantExpr>(Ops1)) { 1262 // If RHS is a constant expression, but the left side isn't, swap the 1263 // operands and try again. 1264 Predicate = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)Predicate); 1265 return ConstantFoldCompareInstOperands(Predicate, Ops1, Ops0, DL, TLI); 1266 } 1267 1268 return ConstantExpr::getCompare(Predicate, Ops0, Ops1); 1269 } 1270 1271 Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op, 1272 const DataLayout &DL) { 1273 assert(Instruction::isUnaryOp(Opcode)); 1274 1275 return ConstantExpr::get(Opcode, Op); 1276 } 1277 1278 Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, 1279 Constant *RHS, 1280 const DataLayout &DL) { 1281 assert(Instruction::isBinaryOp(Opcode)); 1282 if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS)) 1283 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL)) 1284 return C; 1285 1286 return ConstantExpr::get(Opcode, LHS, RHS); 1287 } 1288 1289 Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C, 1290 Type *DestTy, const DataLayout &DL) { 1291 assert(Instruction::isCast(Opcode)); 1292 switch (Opcode) { 1293 default: 1294 llvm_unreachable("Missing case"); 1295 case Instruction::PtrToInt: 1296 // If the input is a inttoptr, eliminate the pair. This requires knowing 1297 // the width of a pointer, so it can't be done in ConstantExpr::getCast. 1298 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 1299 if (CE->getOpcode() == Instruction::IntToPtr) { 1300 Constant *Input = CE->getOperand(0); 1301 unsigned InWidth = Input->getType()->getScalarSizeInBits(); 1302 unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType()); 1303 if (PtrWidth < InWidth) { 1304 Constant *Mask = 1305 ConstantInt::get(CE->getContext(), 1306 APInt::getLowBitsSet(InWidth, PtrWidth)); 1307 Input = ConstantExpr::getAnd(Input, Mask); 1308 } 1309 // Do a zext or trunc to get to the dest size. 1310 return ConstantExpr::getIntegerCast(Input, DestTy, false); 1311 } 1312 } 1313 return ConstantExpr::getCast(Opcode, C, DestTy); 1314 case Instruction::IntToPtr: 1315 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if 1316 // the int size is >= the ptr size and the address spaces are the same. 1317 // This requires knowing the width of a pointer, so it can't be done in 1318 // ConstantExpr::getCast. 1319 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 1320 if (CE->getOpcode() == Instruction::PtrToInt) { 1321 Constant *SrcPtr = CE->getOperand(0); 1322 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType()); 1323 unsigned MidIntSize = CE->getType()->getScalarSizeInBits(); 1324 1325 if (MidIntSize >= SrcPtrSize) { 1326 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace(); 1327 if (SrcAS == DestTy->getPointerAddressSpace()) 1328 return FoldBitCast(CE->getOperand(0), DestTy, DL); 1329 } 1330 } 1331 } 1332 1333 return ConstantExpr::getCast(Opcode, C, DestTy); 1334 case Instruction::Trunc: 1335 case Instruction::ZExt: 1336 case Instruction::SExt: 1337 case Instruction::FPTrunc: 1338 case Instruction::FPExt: 1339 case Instruction::UIToFP: 1340 case Instruction::SIToFP: 1341 case Instruction::FPToUI: 1342 case Instruction::FPToSI: 1343 case Instruction::AddrSpaceCast: 1344 return ConstantExpr::getCast(Opcode, C, DestTy); 1345 case Instruction::BitCast: 1346 return FoldBitCast(C, DestTy, DL); 1347 } 1348 } 1349 1350 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C, 1351 ConstantExpr *CE) { 1352 if (!CE->getOperand(1)->isNullValue()) 1353 return nullptr; // Do not allow stepping over the value! 1354 1355 // Loop over all of the operands, tracking down which value we are 1356 // addressing. 1357 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) { 1358 C = C->getAggregateElement(CE->getOperand(i)); 1359 if (!C) 1360 return nullptr; 1361 } 1362 return C; 1363 } 1364 1365 Constant * 1366 llvm::ConstantFoldLoadThroughGEPIndices(Constant *C, 1367 ArrayRef<Constant *> Indices) { 1368 // Loop over all of the operands, tracking down which value we are 1369 // addressing. 1370 for (Constant *Index : Indices) { 1371 C = C->getAggregateElement(Index); 1372 if (!C) 1373 return nullptr; 1374 } 1375 return C; 1376 } 1377 1378 //===----------------------------------------------------------------------===// 1379 // Constant Folding for Calls 1380 // 1381 1382 bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) { 1383 if (Call->isNoBuiltin() || Call->isStrictFP()) 1384 return false; 1385 switch (F->getIntrinsicID()) { 1386 case Intrinsic::fabs: 1387 case Intrinsic::minnum: 1388 case Intrinsic::maxnum: 1389 case Intrinsic::minimum: 1390 case Intrinsic::maximum: 1391 case Intrinsic::log: 1392 case Intrinsic::log2: 1393 case Intrinsic::log10: 1394 case Intrinsic::exp: 1395 case Intrinsic::exp2: 1396 case Intrinsic::floor: 1397 case Intrinsic::ceil: 1398 case Intrinsic::sqrt: 1399 case Intrinsic::sin: 1400 case Intrinsic::cos: 1401 case Intrinsic::trunc: 1402 case Intrinsic::rint: 1403 case Intrinsic::nearbyint: 1404 case Intrinsic::pow: 1405 case Intrinsic::powi: 1406 case Intrinsic::bswap: 1407 case Intrinsic::ctpop: 1408 case Intrinsic::ctlz: 1409 case Intrinsic::cttz: 1410 case Intrinsic::fshl: 1411 case Intrinsic::fshr: 1412 case Intrinsic::fma: 1413 case Intrinsic::fmuladd: 1414 case Intrinsic::copysign: 1415 case Intrinsic::launder_invariant_group: 1416 case Intrinsic::strip_invariant_group: 1417 case Intrinsic::round: 1418 case Intrinsic::masked_load: 1419 case Intrinsic::sadd_with_overflow: 1420 case Intrinsic::uadd_with_overflow: 1421 case Intrinsic::ssub_with_overflow: 1422 case Intrinsic::usub_with_overflow: 1423 case Intrinsic::smul_with_overflow: 1424 case Intrinsic::umul_with_overflow: 1425 case Intrinsic::sadd_sat: 1426 case Intrinsic::uadd_sat: 1427 case Intrinsic::ssub_sat: 1428 case Intrinsic::usub_sat: 1429 case Intrinsic::smul_fix: 1430 case Intrinsic::smul_fix_sat: 1431 case Intrinsic::convert_from_fp16: 1432 case Intrinsic::convert_to_fp16: 1433 case Intrinsic::bitreverse: 1434 case Intrinsic::x86_sse_cvtss2si: 1435 case Intrinsic::x86_sse_cvtss2si64: 1436 case Intrinsic::x86_sse_cvttss2si: 1437 case Intrinsic::x86_sse_cvttss2si64: 1438 case Intrinsic::x86_sse2_cvtsd2si: 1439 case Intrinsic::x86_sse2_cvtsd2si64: 1440 case Intrinsic::x86_sse2_cvttsd2si: 1441 case Intrinsic::x86_sse2_cvttsd2si64: 1442 case Intrinsic::x86_avx512_vcvtss2si32: 1443 case Intrinsic::x86_avx512_vcvtss2si64: 1444 case Intrinsic::x86_avx512_cvttss2si: 1445 case Intrinsic::x86_avx512_cvttss2si64: 1446 case Intrinsic::x86_avx512_vcvtsd2si32: 1447 case Intrinsic::x86_avx512_vcvtsd2si64: 1448 case Intrinsic::x86_avx512_cvttsd2si: 1449 case Intrinsic::x86_avx512_cvttsd2si64: 1450 case Intrinsic::x86_avx512_vcvtss2usi32: 1451 case Intrinsic::x86_avx512_vcvtss2usi64: 1452 case Intrinsic::x86_avx512_cvttss2usi: 1453 case Intrinsic::x86_avx512_cvttss2usi64: 1454 case Intrinsic::x86_avx512_vcvtsd2usi32: 1455 case Intrinsic::x86_avx512_vcvtsd2usi64: 1456 case Intrinsic::x86_avx512_cvttsd2usi: 1457 case Intrinsic::x86_avx512_cvttsd2usi64: 1458 case Intrinsic::is_constant: 1459 return true; 1460 default: 1461 return false; 1462 case Intrinsic::not_intrinsic: break; 1463 } 1464 1465 if (!F->hasName()) 1466 return false; 1467 StringRef Name = F->getName(); 1468 1469 // In these cases, the check of the length is required. We don't want to 1470 // return true for a name like "cos\0blah" which strcmp would return equal to 1471 // "cos", but has length 8. 1472 switch (Name[0]) { 1473 default: 1474 return false; 1475 case 'a': 1476 return Name == "acos" || Name == "asin" || Name == "atan" || 1477 Name == "atan2" || Name == "acosf" || Name == "asinf" || 1478 Name == "atanf" || Name == "atan2f"; 1479 case 'c': 1480 return Name == "ceil" || Name == "cos" || Name == "cosh" || 1481 Name == "ceilf" || Name == "cosf" || Name == "coshf"; 1482 case 'e': 1483 return Name == "exp" || Name == "exp2" || Name == "expf" || Name == "exp2f"; 1484 case 'f': 1485 return Name == "fabs" || Name == "floor" || Name == "fmod" || 1486 Name == "fabsf" || Name == "floorf" || Name == "fmodf"; 1487 case 'l': 1488 return Name == "log" || Name == "log10" || Name == "logf" || 1489 Name == "log10f"; 1490 case 'p': 1491 return Name == "pow" || Name == "powf"; 1492 case 'r': 1493 return Name == "round" || Name == "roundf"; 1494 case 's': 1495 return Name == "sin" || Name == "sinh" || Name == "sqrt" || 1496 Name == "sinf" || Name == "sinhf" || Name == "sqrtf"; 1497 case 't': 1498 return Name == "tan" || Name == "tanh" || Name == "tanf" || Name == "tanhf"; 1499 case '_': 1500 1501 // Check for various function names that get used for the math functions 1502 // when the header files are preprocessed with the macro 1503 // __FINITE_MATH_ONLY__ enabled. 1504 // The '12' here is the length of the shortest name that can match. 1505 // We need to check the size before looking at Name[1] and Name[2] 1506 // so we may as well check a limit that will eliminate mismatches. 1507 if (Name.size() < 12 || Name[1] != '_') 1508 return false; 1509 switch (Name[2]) { 1510 default: 1511 return false; 1512 case 'a': 1513 return Name == "__acos_finite" || Name == "__acosf_finite" || 1514 Name == "__asin_finite" || Name == "__asinf_finite" || 1515 Name == "__atan2_finite" || Name == "__atan2f_finite"; 1516 case 'c': 1517 return Name == "__cosh_finite" || Name == "__coshf_finite"; 1518 case 'e': 1519 return Name == "__exp_finite" || Name == "__expf_finite" || 1520 Name == "__exp2_finite" || Name == "__exp2f_finite"; 1521 case 'l': 1522 return Name == "__log_finite" || Name == "__logf_finite" || 1523 Name == "__log10_finite" || Name == "__log10f_finite"; 1524 case 'p': 1525 return Name == "__pow_finite" || Name == "__powf_finite"; 1526 case 's': 1527 return Name == "__sinh_finite" || Name == "__sinhf_finite"; 1528 } 1529 } 1530 } 1531 1532 namespace { 1533 1534 Constant *GetConstantFoldFPValue(double V, Type *Ty) { 1535 if (Ty->isHalfTy() || Ty->isFloatTy()) { 1536 APFloat APF(V); 1537 bool unused; 1538 APF.convert(Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &unused); 1539 return ConstantFP::get(Ty->getContext(), APF); 1540 } 1541 if (Ty->isDoubleTy()) 1542 return ConstantFP::get(Ty->getContext(), APFloat(V)); 1543 llvm_unreachable("Can only constant fold half/float/double"); 1544 } 1545 1546 /// Clear the floating-point exception state. 1547 inline void llvm_fenv_clearexcept() { 1548 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT 1549 feclearexcept(FE_ALL_EXCEPT); 1550 #endif 1551 errno = 0; 1552 } 1553 1554 /// Test if a floating-point exception was raised. 1555 inline bool llvm_fenv_testexcept() { 1556 int errno_val = errno; 1557 if (errno_val == ERANGE || errno_val == EDOM) 1558 return true; 1559 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT 1560 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT)) 1561 return true; 1562 #endif 1563 return false; 1564 } 1565 1566 Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) { 1567 llvm_fenv_clearexcept(); 1568 V = NativeFP(V); 1569 if (llvm_fenv_testexcept()) { 1570 llvm_fenv_clearexcept(); 1571 return nullptr; 1572 } 1573 1574 return GetConstantFoldFPValue(V, Ty); 1575 } 1576 1577 Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V, 1578 double W, Type *Ty) { 1579 llvm_fenv_clearexcept(); 1580 V = NativeFP(V, W); 1581 if (llvm_fenv_testexcept()) { 1582 llvm_fenv_clearexcept(); 1583 return nullptr; 1584 } 1585 1586 return GetConstantFoldFPValue(V, Ty); 1587 } 1588 1589 /// Attempt to fold an SSE floating point to integer conversion of a constant 1590 /// floating point. If roundTowardZero is false, the default IEEE rounding is 1591 /// used (toward nearest, ties to even). This matches the behavior of the 1592 /// non-truncating SSE instructions in the default rounding mode. The desired 1593 /// integer type Ty is used to select how many bits are available for the 1594 /// result. Returns null if the conversion cannot be performed, otherwise 1595 /// returns the Constant value resulting from the conversion. 1596 Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero, 1597 Type *Ty, bool IsSigned) { 1598 // All of these conversion intrinsics form an integer of at most 64bits. 1599 unsigned ResultWidth = Ty->getIntegerBitWidth(); 1600 assert(ResultWidth <= 64 && 1601 "Can only constant fold conversions to 64 and 32 bit ints"); 1602 1603 uint64_t UIntVal; 1604 bool isExact = false; 1605 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero 1606 : APFloat::rmNearestTiesToEven; 1607 APFloat::opStatus status = 1608 Val.convertToInteger(makeMutableArrayRef(UIntVal), ResultWidth, 1609 IsSigned, mode, &isExact); 1610 if (status != APFloat::opOK && 1611 (!roundTowardZero || status != APFloat::opInexact)) 1612 return nullptr; 1613 return ConstantInt::get(Ty, UIntVal, IsSigned); 1614 } 1615 1616 double getValueAsDouble(ConstantFP *Op) { 1617 Type *Ty = Op->getType(); 1618 1619 if (Ty->isFloatTy()) 1620 return Op->getValueAPF().convertToFloat(); 1621 1622 if (Ty->isDoubleTy()) 1623 return Op->getValueAPF().convertToDouble(); 1624 1625 bool unused; 1626 APFloat APF = Op->getValueAPF(); 1627 APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused); 1628 return APF.convertToDouble(); 1629 } 1630 1631 static bool isManifestConstant(const Constant *c) { 1632 if (isa<ConstantData>(c)) { 1633 return true; 1634 } else if (isa<ConstantAggregate>(c) || isa<ConstantExpr>(c)) { 1635 for (const Value *subc : c->operand_values()) { 1636 if (!isManifestConstant(cast<Constant>(subc))) 1637 return false; 1638 } 1639 return true; 1640 } 1641 return false; 1642 } 1643 1644 static bool getConstIntOrUndef(Value *Op, const APInt *&C) { 1645 if (auto *CI = dyn_cast<ConstantInt>(Op)) { 1646 C = &CI->getValue(); 1647 return true; 1648 } 1649 if (isa<UndefValue>(Op)) { 1650 C = nullptr; 1651 return true; 1652 } 1653 return false; 1654 } 1655 1656 static Constant *ConstantFoldScalarCall1(StringRef Name, 1657 Intrinsic::ID IntrinsicID, 1658 Type *Ty, 1659 ArrayRef<Constant *> Operands, 1660 const TargetLibraryInfo *TLI, 1661 const CallBase *Call) { 1662 assert(Operands.size() == 1 && "Wrong number of operands."); 1663 1664 if (IntrinsicID == Intrinsic::is_constant) { 1665 // We know we have a "Constant" argument. But we want to only 1666 // return true for manifest constants, not those that depend on 1667 // constants with unknowable values, e.g. GlobalValue or BlockAddress. 1668 if (isManifestConstant(Operands[0])) 1669 return ConstantInt::getTrue(Ty->getContext()); 1670 return nullptr; 1671 } 1672 if (isa<UndefValue>(Operands[0])) { 1673 // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN. 1674 // ctpop() is between 0 and bitwidth, pick 0 for undef. 1675 if (IntrinsicID == Intrinsic::cos || 1676 IntrinsicID == Intrinsic::ctpop) 1677 return Constant::getNullValue(Ty); 1678 if (IntrinsicID == Intrinsic::bswap || 1679 IntrinsicID == Intrinsic::bitreverse || 1680 IntrinsicID == Intrinsic::launder_invariant_group || 1681 IntrinsicID == Intrinsic::strip_invariant_group) 1682 return Operands[0]; 1683 } 1684 1685 if (isa<ConstantPointerNull>(Operands[0])) { 1686 // launder(null) == null == strip(null) iff in addrspace 0 1687 if (IntrinsicID == Intrinsic::launder_invariant_group || 1688 IntrinsicID == Intrinsic::strip_invariant_group) { 1689 // If instruction is not yet put in a basic block (e.g. when cloning 1690 // a function during inlining), Call's caller may not be available. 1691 // So check Call's BB first before querying Call->getCaller. 1692 const Function *Caller = 1693 Call->getParent() ? Call->getCaller() : nullptr; 1694 if (Caller && 1695 !NullPointerIsDefined( 1696 Caller, Operands[0]->getType()->getPointerAddressSpace())) { 1697 return Operands[0]; 1698 } 1699 return nullptr; 1700 } 1701 } 1702 1703 if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) { 1704 if (IntrinsicID == Intrinsic::convert_to_fp16) { 1705 APFloat Val(Op->getValueAPF()); 1706 1707 bool lost = false; 1708 Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost); 1709 1710 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt()); 1711 } 1712 1713 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) 1714 return nullptr; 1715 1716 if (IntrinsicID == Intrinsic::round) { 1717 APFloat V = Op->getValueAPF(); 1718 V.roundToIntegral(APFloat::rmNearestTiesToAway); 1719 return ConstantFP::get(Ty->getContext(), V); 1720 } 1721 1722 if (IntrinsicID == Intrinsic::floor) { 1723 APFloat V = Op->getValueAPF(); 1724 V.roundToIntegral(APFloat::rmTowardNegative); 1725 return ConstantFP::get(Ty->getContext(), V); 1726 } 1727 1728 if (IntrinsicID == Intrinsic::ceil) { 1729 APFloat V = Op->getValueAPF(); 1730 V.roundToIntegral(APFloat::rmTowardPositive); 1731 return ConstantFP::get(Ty->getContext(), V); 1732 } 1733 1734 if (IntrinsicID == Intrinsic::trunc) { 1735 APFloat V = Op->getValueAPF(); 1736 V.roundToIntegral(APFloat::rmTowardZero); 1737 return ConstantFP::get(Ty->getContext(), V); 1738 } 1739 1740 if (IntrinsicID == Intrinsic::rint) { 1741 APFloat V = Op->getValueAPF(); 1742 V.roundToIntegral(APFloat::rmNearestTiesToEven); 1743 return ConstantFP::get(Ty->getContext(), V); 1744 } 1745 1746 if (IntrinsicID == Intrinsic::nearbyint) { 1747 APFloat V = Op->getValueAPF(); 1748 V.roundToIntegral(APFloat::rmNearestTiesToEven); 1749 return ConstantFP::get(Ty->getContext(), V); 1750 } 1751 1752 /// We only fold functions with finite arguments. Folding NaN and inf is 1753 /// likely to be aborted with an exception anyway, and some host libms 1754 /// have known errors raising exceptions. 1755 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity()) 1756 return nullptr; 1757 1758 /// Currently APFloat versions of these functions do not exist, so we use 1759 /// the host native double versions. Float versions are not called 1760 /// directly but for all these it is true (float)(f((double)arg)) == 1761 /// f(arg). Long double not supported yet. 1762 double V = getValueAsDouble(Op); 1763 1764 switch (IntrinsicID) { 1765 default: break; 1766 case Intrinsic::fabs: 1767 return ConstantFoldFP(fabs, V, Ty); 1768 case Intrinsic::log2: 1769 return ConstantFoldFP(Log2, V, Ty); 1770 case Intrinsic::log: 1771 return ConstantFoldFP(log, V, Ty); 1772 case Intrinsic::log10: 1773 return ConstantFoldFP(log10, V, Ty); 1774 case Intrinsic::exp: 1775 return ConstantFoldFP(exp, V, Ty); 1776 case Intrinsic::exp2: 1777 return ConstantFoldFP(exp2, V, Ty); 1778 case Intrinsic::sin: 1779 return ConstantFoldFP(sin, V, Ty); 1780 case Intrinsic::cos: 1781 return ConstantFoldFP(cos, V, Ty); 1782 case Intrinsic::sqrt: 1783 return ConstantFoldFP(sqrt, V, Ty); 1784 } 1785 1786 if (!TLI) 1787 return nullptr; 1788 1789 char NameKeyChar = Name[0]; 1790 if (Name[0] == '_' && Name.size() > 2 && Name[1] == '_') 1791 NameKeyChar = Name[2]; 1792 1793 switch (NameKeyChar) { 1794 case 'a': 1795 if ((Name == "acos" && TLI->has(LibFunc_acos)) || 1796 (Name == "acosf" && TLI->has(LibFunc_acosf)) || 1797 (Name == "__acos_finite" && TLI->has(LibFunc_acos_finite)) || 1798 (Name == "__acosf_finite" && TLI->has(LibFunc_acosf_finite))) 1799 return ConstantFoldFP(acos, V, Ty); 1800 else if ((Name == "asin" && TLI->has(LibFunc_asin)) || 1801 (Name == "asinf" && TLI->has(LibFunc_asinf)) || 1802 (Name == "__asin_finite" && TLI->has(LibFunc_asin_finite)) || 1803 (Name == "__asinf_finite" && TLI->has(LibFunc_asinf_finite))) 1804 return ConstantFoldFP(asin, V, Ty); 1805 else if ((Name == "atan" && TLI->has(LibFunc_atan)) || 1806 (Name == "atanf" && TLI->has(LibFunc_atanf))) 1807 return ConstantFoldFP(atan, V, Ty); 1808 break; 1809 case 'c': 1810 if ((Name == "ceil" && TLI->has(LibFunc_ceil)) || 1811 (Name == "ceilf" && TLI->has(LibFunc_ceilf))) 1812 return ConstantFoldFP(ceil, V, Ty); 1813 else if ((Name == "cos" && TLI->has(LibFunc_cos)) || 1814 (Name == "cosf" && TLI->has(LibFunc_cosf))) 1815 return ConstantFoldFP(cos, V, Ty); 1816 else if ((Name == "cosh" && TLI->has(LibFunc_cosh)) || 1817 (Name == "coshf" && TLI->has(LibFunc_coshf)) || 1818 (Name == "__cosh_finite" && TLI->has(LibFunc_cosh_finite)) || 1819 (Name == "__coshf_finite" && TLI->has(LibFunc_coshf_finite))) 1820 return ConstantFoldFP(cosh, V, Ty); 1821 break; 1822 case 'e': 1823 if ((Name == "exp" && TLI->has(LibFunc_exp)) || 1824 (Name == "expf" && TLI->has(LibFunc_expf)) || 1825 (Name == "__exp_finite" && TLI->has(LibFunc_exp_finite)) || 1826 (Name == "__expf_finite" && TLI->has(LibFunc_expf_finite))) 1827 return ConstantFoldFP(exp, V, Ty); 1828 if ((Name == "exp2" && TLI->has(LibFunc_exp2)) || 1829 (Name == "exp2f" && TLI->has(LibFunc_exp2f)) || 1830 (Name == "__exp2_finite" && TLI->has(LibFunc_exp2_finite)) || 1831 (Name == "__exp2f_finite" && TLI->has(LibFunc_exp2f_finite))) 1832 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a 1833 // C99 library. 1834 return ConstantFoldBinaryFP(pow, 2.0, V, Ty); 1835 break; 1836 case 'f': 1837 if ((Name == "fabs" && TLI->has(LibFunc_fabs)) || 1838 (Name == "fabsf" && TLI->has(LibFunc_fabsf))) 1839 return ConstantFoldFP(fabs, V, Ty); 1840 else if ((Name == "floor" && TLI->has(LibFunc_floor)) || 1841 (Name == "floorf" && TLI->has(LibFunc_floorf))) 1842 return ConstantFoldFP(floor, V, Ty); 1843 break; 1844 case 'l': 1845 if ((Name == "log" && V > 0 && TLI->has(LibFunc_log)) || 1846 (Name == "logf" && V > 0 && TLI->has(LibFunc_logf)) || 1847 (Name == "__log_finite" && V > 0 && 1848 TLI->has(LibFunc_log_finite)) || 1849 (Name == "__logf_finite" && V > 0 && 1850 TLI->has(LibFunc_logf_finite))) 1851 return ConstantFoldFP(log, V, Ty); 1852 else if ((Name == "log10" && V > 0 && TLI->has(LibFunc_log10)) || 1853 (Name == "log10f" && V > 0 && TLI->has(LibFunc_log10f)) || 1854 (Name == "__log10_finite" && V > 0 && 1855 TLI->has(LibFunc_log10_finite)) || 1856 (Name == "__log10f_finite" && V > 0 && 1857 TLI->has(LibFunc_log10f_finite))) 1858 return ConstantFoldFP(log10, V, Ty); 1859 break; 1860 case 'r': 1861 if ((Name == "round" && TLI->has(LibFunc_round)) || 1862 (Name == "roundf" && TLI->has(LibFunc_roundf))) 1863 return ConstantFoldFP(round, V, Ty); 1864 break; 1865 case 's': 1866 if ((Name == "sin" && TLI->has(LibFunc_sin)) || 1867 (Name == "sinf" && TLI->has(LibFunc_sinf))) 1868 return ConstantFoldFP(sin, V, Ty); 1869 else if ((Name == "sinh" && TLI->has(LibFunc_sinh)) || 1870 (Name == "sinhf" && TLI->has(LibFunc_sinhf)) || 1871 (Name == "__sinh_finite" && TLI->has(LibFunc_sinh_finite)) || 1872 (Name == "__sinhf_finite" && TLI->has(LibFunc_sinhf_finite))) 1873 return ConstantFoldFP(sinh, V, Ty); 1874 else if ((Name == "sqrt" && V >= 0 && TLI->has(LibFunc_sqrt)) || 1875 (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc_sqrtf))) 1876 return ConstantFoldFP(sqrt, V, Ty); 1877 break; 1878 case 't': 1879 if ((Name == "tan" && TLI->has(LibFunc_tan)) || 1880 (Name == "tanf" && TLI->has(LibFunc_tanf))) 1881 return ConstantFoldFP(tan, V, Ty); 1882 else if ((Name == "tanh" && TLI->has(LibFunc_tanh)) || 1883 (Name == "tanhf" && TLI->has(LibFunc_tanhf))) 1884 return ConstantFoldFP(tanh, V, Ty); 1885 break; 1886 default: 1887 break; 1888 } 1889 return nullptr; 1890 } 1891 1892 if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) { 1893 switch (IntrinsicID) { 1894 case Intrinsic::bswap: 1895 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap()); 1896 case Intrinsic::ctpop: 1897 return ConstantInt::get(Ty, Op->getValue().countPopulation()); 1898 case Intrinsic::bitreverse: 1899 return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits()); 1900 case Intrinsic::convert_from_fp16: { 1901 APFloat Val(APFloat::IEEEhalf(), Op->getValue()); 1902 1903 bool lost = false; 1904 APFloat::opStatus status = Val.convert( 1905 Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost); 1906 1907 // Conversion is always precise. 1908 (void)status; 1909 assert(status == APFloat::opOK && !lost && 1910 "Precision lost during fp16 constfolding"); 1911 1912 return ConstantFP::get(Ty->getContext(), Val); 1913 } 1914 default: 1915 return nullptr; 1916 } 1917 } 1918 1919 // Support ConstantVector in case we have an Undef in the top. 1920 if (isa<ConstantVector>(Operands[0]) || 1921 isa<ConstantDataVector>(Operands[0])) { 1922 auto *Op = cast<Constant>(Operands[0]); 1923 switch (IntrinsicID) { 1924 default: break; 1925 case Intrinsic::x86_sse_cvtss2si: 1926 case Intrinsic::x86_sse_cvtss2si64: 1927 case Intrinsic::x86_sse2_cvtsd2si: 1928 case Intrinsic::x86_sse2_cvtsd2si64: 1929 if (ConstantFP *FPOp = 1930 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 1931 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(), 1932 /*roundTowardZero=*/false, Ty, 1933 /*IsSigned*/true); 1934 break; 1935 case Intrinsic::x86_sse_cvttss2si: 1936 case Intrinsic::x86_sse_cvttss2si64: 1937 case Intrinsic::x86_sse2_cvttsd2si: 1938 case Intrinsic::x86_sse2_cvttsd2si64: 1939 if (ConstantFP *FPOp = 1940 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 1941 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(), 1942 /*roundTowardZero=*/true, Ty, 1943 /*IsSigned*/true); 1944 break; 1945 } 1946 } 1947 1948 return nullptr; 1949 } 1950 1951 static Constant *ConstantFoldScalarCall2(StringRef Name, 1952 Intrinsic::ID IntrinsicID, 1953 Type *Ty, 1954 ArrayRef<Constant *> Operands, 1955 const TargetLibraryInfo *TLI, 1956 const CallBase *Call) { 1957 assert(Operands.size() == 2 && "Wrong number of operands."); 1958 1959 if (auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 1960 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) 1961 return nullptr; 1962 double Op1V = getValueAsDouble(Op1); 1963 1964 if (auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 1965 if (Op2->getType() != Op1->getType()) 1966 return nullptr; 1967 1968 double Op2V = getValueAsDouble(Op2); 1969 if (IntrinsicID == Intrinsic::pow) { 1970 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 1971 } 1972 if (IntrinsicID == Intrinsic::copysign) { 1973 APFloat V1 = Op1->getValueAPF(); 1974 const APFloat &V2 = Op2->getValueAPF(); 1975 V1.copySign(V2); 1976 return ConstantFP::get(Ty->getContext(), V1); 1977 } 1978 1979 if (IntrinsicID == Intrinsic::minnum) { 1980 const APFloat &C1 = Op1->getValueAPF(); 1981 const APFloat &C2 = Op2->getValueAPF(); 1982 return ConstantFP::get(Ty->getContext(), minnum(C1, C2)); 1983 } 1984 1985 if (IntrinsicID == Intrinsic::maxnum) { 1986 const APFloat &C1 = Op1->getValueAPF(); 1987 const APFloat &C2 = Op2->getValueAPF(); 1988 return ConstantFP::get(Ty->getContext(), maxnum(C1, C2)); 1989 } 1990 1991 if (IntrinsicID == Intrinsic::minimum) { 1992 const APFloat &C1 = Op1->getValueAPF(); 1993 const APFloat &C2 = Op2->getValueAPF(); 1994 return ConstantFP::get(Ty->getContext(), minimum(C1, C2)); 1995 } 1996 1997 if (IntrinsicID == Intrinsic::maximum) { 1998 const APFloat &C1 = Op1->getValueAPF(); 1999 const APFloat &C2 = Op2->getValueAPF(); 2000 return ConstantFP::get(Ty->getContext(), maximum(C1, C2)); 2001 } 2002 2003 if (!TLI) 2004 return nullptr; 2005 if ((Name == "pow" && TLI->has(LibFunc_pow)) || 2006 (Name == "powf" && TLI->has(LibFunc_powf)) || 2007 (Name == "__pow_finite" && TLI->has(LibFunc_pow_finite)) || 2008 (Name == "__powf_finite" && TLI->has(LibFunc_powf_finite))) 2009 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty); 2010 if ((Name == "fmod" && TLI->has(LibFunc_fmod)) || 2011 (Name == "fmodf" && TLI->has(LibFunc_fmodf))) 2012 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty); 2013 if ((Name == "atan2" && TLI->has(LibFunc_atan2)) || 2014 (Name == "atan2f" && TLI->has(LibFunc_atan2f)) || 2015 (Name == "__atan2_finite" && TLI->has(LibFunc_atan2_finite)) || 2016 (Name == "__atan2f_finite" && TLI->has(LibFunc_atan2f_finite))) 2017 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty); 2018 } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) { 2019 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy()) 2020 return ConstantFP::get(Ty->getContext(), 2021 APFloat((float)std::pow((float)Op1V, 2022 (int)Op2C->getZExtValue()))); 2023 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy()) 2024 return ConstantFP::get(Ty->getContext(), 2025 APFloat((float)std::pow((float)Op1V, 2026 (int)Op2C->getZExtValue()))); 2027 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy()) 2028 return ConstantFP::get(Ty->getContext(), 2029 APFloat((double)std::pow((double)Op1V, 2030 (int)Op2C->getZExtValue()))); 2031 } 2032 return nullptr; 2033 } 2034 2035 if (Operands[0]->getType()->isIntegerTy() && 2036 Operands[1]->getType()->isIntegerTy()) { 2037 const APInt *C0, *C1; 2038 if (!getConstIntOrUndef(Operands[0], C0) || 2039 !getConstIntOrUndef(Operands[1], C1)) 2040 return nullptr; 2041 2042 switch (IntrinsicID) { 2043 default: break; 2044 case Intrinsic::smul_with_overflow: 2045 case Intrinsic::umul_with_overflow: 2046 // Even if both operands are undef, we cannot fold muls to undef 2047 // in the general case. For example, on i2 there are no inputs 2048 // that would produce { i2 -1, i1 true } as the result. 2049 if (!C0 || !C1) 2050 return Constant::getNullValue(Ty); 2051 LLVM_FALLTHROUGH; 2052 case Intrinsic::sadd_with_overflow: 2053 case Intrinsic::uadd_with_overflow: 2054 case Intrinsic::ssub_with_overflow: 2055 case Intrinsic::usub_with_overflow: { 2056 if (!C0 || !C1) 2057 return UndefValue::get(Ty); 2058 2059 APInt Res; 2060 bool Overflow; 2061 switch (IntrinsicID) { 2062 default: llvm_unreachable("Invalid case"); 2063 case Intrinsic::sadd_with_overflow: 2064 Res = C0->sadd_ov(*C1, Overflow); 2065 break; 2066 case Intrinsic::uadd_with_overflow: 2067 Res = C0->uadd_ov(*C1, Overflow); 2068 break; 2069 case Intrinsic::ssub_with_overflow: 2070 Res = C0->ssub_ov(*C1, Overflow); 2071 break; 2072 case Intrinsic::usub_with_overflow: 2073 Res = C0->usub_ov(*C1, Overflow); 2074 break; 2075 case Intrinsic::smul_with_overflow: 2076 Res = C0->smul_ov(*C1, Overflow); 2077 break; 2078 case Intrinsic::umul_with_overflow: 2079 Res = C0->umul_ov(*C1, Overflow); 2080 break; 2081 } 2082 Constant *Ops[] = { 2083 ConstantInt::get(Ty->getContext(), Res), 2084 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow) 2085 }; 2086 return ConstantStruct::get(cast<StructType>(Ty), Ops); 2087 } 2088 case Intrinsic::uadd_sat: 2089 case Intrinsic::sadd_sat: 2090 if (!C0 && !C1) 2091 return UndefValue::get(Ty); 2092 if (!C0 || !C1) 2093 return Constant::getAllOnesValue(Ty); 2094 if (IntrinsicID == Intrinsic::uadd_sat) 2095 return ConstantInt::get(Ty, C0->uadd_sat(*C1)); 2096 else 2097 return ConstantInt::get(Ty, C0->sadd_sat(*C1)); 2098 case Intrinsic::usub_sat: 2099 case Intrinsic::ssub_sat: 2100 if (!C0 && !C1) 2101 return UndefValue::get(Ty); 2102 if (!C0 || !C1) 2103 return Constant::getNullValue(Ty); 2104 if (IntrinsicID == Intrinsic::usub_sat) 2105 return ConstantInt::get(Ty, C0->usub_sat(*C1)); 2106 else 2107 return ConstantInt::get(Ty, C0->ssub_sat(*C1)); 2108 case Intrinsic::cttz: 2109 case Intrinsic::ctlz: 2110 assert(C1 && "Must be constant int"); 2111 2112 // cttz(0, 1) and ctlz(0, 1) are undef. 2113 if (C1->isOneValue() && (!C0 || C0->isNullValue())) 2114 return UndefValue::get(Ty); 2115 if (!C0) 2116 return Constant::getNullValue(Ty); 2117 if (IntrinsicID == Intrinsic::cttz) 2118 return ConstantInt::get(Ty, C0->countTrailingZeros()); 2119 else 2120 return ConstantInt::get(Ty, C0->countLeadingZeros()); 2121 } 2122 2123 return nullptr; 2124 } 2125 2126 // Support ConstantVector in case we have an Undef in the top. 2127 if ((isa<ConstantVector>(Operands[0]) || 2128 isa<ConstantDataVector>(Operands[0])) && 2129 // Check for default rounding mode. 2130 // FIXME: Support other rounding modes? 2131 isa<ConstantInt>(Operands[1]) && 2132 cast<ConstantInt>(Operands[1])->getValue() == 4) { 2133 auto *Op = cast<Constant>(Operands[0]); 2134 switch (IntrinsicID) { 2135 default: break; 2136 case Intrinsic::x86_avx512_vcvtss2si32: 2137 case Intrinsic::x86_avx512_vcvtss2si64: 2138 case Intrinsic::x86_avx512_vcvtsd2si32: 2139 case Intrinsic::x86_avx512_vcvtsd2si64: 2140 if (ConstantFP *FPOp = 2141 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 2142 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(), 2143 /*roundTowardZero=*/false, Ty, 2144 /*IsSigned*/true); 2145 break; 2146 case Intrinsic::x86_avx512_vcvtss2usi32: 2147 case Intrinsic::x86_avx512_vcvtss2usi64: 2148 case Intrinsic::x86_avx512_vcvtsd2usi32: 2149 case Intrinsic::x86_avx512_vcvtsd2usi64: 2150 if (ConstantFP *FPOp = 2151 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 2152 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(), 2153 /*roundTowardZero=*/false, Ty, 2154 /*IsSigned*/false); 2155 break; 2156 case Intrinsic::x86_avx512_cvttss2si: 2157 case Intrinsic::x86_avx512_cvttss2si64: 2158 case Intrinsic::x86_avx512_cvttsd2si: 2159 case Intrinsic::x86_avx512_cvttsd2si64: 2160 if (ConstantFP *FPOp = 2161 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 2162 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(), 2163 /*roundTowardZero=*/true, Ty, 2164 /*IsSigned*/true); 2165 break; 2166 case Intrinsic::x86_avx512_cvttss2usi: 2167 case Intrinsic::x86_avx512_cvttss2usi64: 2168 case Intrinsic::x86_avx512_cvttsd2usi: 2169 case Intrinsic::x86_avx512_cvttsd2usi64: 2170 if (ConstantFP *FPOp = 2171 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U))) 2172 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(), 2173 /*roundTowardZero=*/true, Ty, 2174 /*IsSigned*/false); 2175 break; 2176 } 2177 } 2178 return nullptr; 2179 } 2180 2181 static Constant *ConstantFoldScalarCall3(StringRef Name, 2182 Intrinsic::ID IntrinsicID, 2183 Type *Ty, 2184 ArrayRef<Constant *> Operands, 2185 const TargetLibraryInfo *TLI, 2186 const CallBase *Call) { 2187 assert(Operands.size() == 3 && "Wrong number of operands."); 2188 2189 if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) { 2190 if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) { 2191 if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) { 2192 switch (IntrinsicID) { 2193 default: break; 2194 case Intrinsic::fma: 2195 case Intrinsic::fmuladd: { 2196 APFloat V = Op1->getValueAPF(); 2197 APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(), 2198 Op3->getValueAPF(), 2199 APFloat::rmNearestTiesToEven); 2200 if (s != APFloat::opInvalidOp) 2201 return ConstantFP::get(Ty->getContext(), V); 2202 2203 return nullptr; 2204 } 2205 } 2206 } 2207 } 2208 } 2209 2210 if (const auto *Op1 = dyn_cast<ConstantInt>(Operands[0])) { 2211 if (const auto *Op2 = dyn_cast<ConstantInt>(Operands[1])) { 2212 if (const auto *Op3 = dyn_cast<ConstantInt>(Operands[2])) { 2213 switch (IntrinsicID) { 2214 default: break; 2215 case Intrinsic::smul_fix: 2216 case Intrinsic::smul_fix_sat: { 2217 // This code performs rounding towards negative infinity in case the 2218 // result cannot be represented exactly for the given scale. Targets 2219 // that do care about rounding should use a target hook for specifying 2220 // how rounding should be done, and provide their own folding to be 2221 // consistent with rounding. This is the same approach as used by 2222 // DAGTypeLegalizer::ExpandIntRes_MULFIX. 2223 APInt Lhs = Op1->getValue(); 2224 APInt Rhs = Op2->getValue(); 2225 unsigned Scale = Op3->getValue().getZExtValue(); 2226 unsigned Width = Lhs.getBitWidth(); 2227 assert(Scale < Width && "Illegal scale."); 2228 unsigned ExtendedWidth = Width * 2; 2229 APInt Product = (Lhs.sextOrSelf(ExtendedWidth) * 2230 Rhs.sextOrSelf(ExtendedWidth)).ashr(Scale); 2231 if (IntrinsicID == Intrinsic::smul_fix_sat) { 2232 APInt MaxValue = 2233 APInt::getSignedMaxValue(Width).sextOrSelf(ExtendedWidth); 2234 APInt MinValue = 2235 APInt::getSignedMinValue(Width).sextOrSelf(ExtendedWidth); 2236 Product = APIntOps::smin(Product, MaxValue); 2237 Product = APIntOps::smax(Product, MinValue); 2238 } 2239 return ConstantInt::get(Ty->getContext(), 2240 Product.sextOrTrunc(Width)); 2241 } 2242 } 2243 } 2244 } 2245 } 2246 2247 if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) { 2248 const APInt *C0, *C1, *C2; 2249 if (!getConstIntOrUndef(Operands[0], C0) || 2250 !getConstIntOrUndef(Operands[1], C1) || 2251 !getConstIntOrUndef(Operands[2], C2)) 2252 return nullptr; 2253 2254 bool IsRight = IntrinsicID == Intrinsic::fshr; 2255 if (!C2) 2256 return Operands[IsRight ? 1 : 0]; 2257 if (!C0 && !C1) 2258 return UndefValue::get(Ty); 2259 2260 // The shift amount is interpreted as modulo the bitwidth. If the shift 2261 // amount is effectively 0, avoid UB due to oversized inverse shift below. 2262 unsigned BitWidth = C2->getBitWidth(); 2263 unsigned ShAmt = C2->urem(BitWidth); 2264 if (!ShAmt) 2265 return Operands[IsRight ? 1 : 0]; 2266 2267 // (C0 << ShlAmt) | (C1 >> LshrAmt) 2268 unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt; 2269 unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt; 2270 if (!C0) 2271 return ConstantInt::get(Ty, C1->lshr(LshrAmt)); 2272 if (!C1) 2273 return ConstantInt::get(Ty, C0->shl(ShlAmt)); 2274 return ConstantInt::get(Ty, C0->shl(ShlAmt) | C1->lshr(LshrAmt)); 2275 } 2276 2277 return nullptr; 2278 } 2279 2280 static Constant *ConstantFoldScalarCall(StringRef Name, 2281 Intrinsic::ID IntrinsicID, 2282 Type *Ty, 2283 ArrayRef<Constant *> Operands, 2284 const TargetLibraryInfo *TLI, 2285 const CallBase *Call) { 2286 if (Operands.size() == 1) 2287 return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call); 2288 2289 if (Operands.size() == 2) 2290 return ConstantFoldScalarCall2(Name, IntrinsicID, Ty, Operands, TLI, Call); 2291 2292 if (Operands.size() == 3) 2293 return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call); 2294 2295 return nullptr; 2296 } 2297 2298 static Constant *ConstantFoldVectorCall(StringRef Name, 2299 Intrinsic::ID IntrinsicID, 2300 VectorType *VTy, 2301 ArrayRef<Constant *> Operands, 2302 const DataLayout &DL, 2303 const TargetLibraryInfo *TLI, 2304 const CallBase *Call) { 2305 SmallVector<Constant *, 4> Result(VTy->getNumElements()); 2306 SmallVector<Constant *, 4> Lane(Operands.size()); 2307 Type *Ty = VTy->getElementType(); 2308 2309 if (IntrinsicID == Intrinsic::masked_load) { 2310 auto *SrcPtr = Operands[0]; 2311 auto *Mask = Operands[2]; 2312 auto *Passthru = Operands[3]; 2313 2314 Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, VTy, DL); 2315 2316 SmallVector<Constant *, 32> NewElements; 2317 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) { 2318 auto *MaskElt = Mask->getAggregateElement(I); 2319 if (!MaskElt) 2320 break; 2321 auto *PassthruElt = Passthru->getAggregateElement(I); 2322 auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr; 2323 if (isa<UndefValue>(MaskElt)) { 2324 if (PassthruElt) 2325 NewElements.push_back(PassthruElt); 2326 else if (VecElt) 2327 NewElements.push_back(VecElt); 2328 else 2329 return nullptr; 2330 } 2331 if (MaskElt->isNullValue()) { 2332 if (!PassthruElt) 2333 return nullptr; 2334 NewElements.push_back(PassthruElt); 2335 } else if (MaskElt->isOneValue()) { 2336 if (!VecElt) 2337 return nullptr; 2338 NewElements.push_back(VecElt); 2339 } else { 2340 return nullptr; 2341 } 2342 } 2343 if (NewElements.size() != VTy->getNumElements()) 2344 return nullptr; 2345 return ConstantVector::get(NewElements); 2346 } 2347 2348 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) { 2349 // Gather a column of constants. 2350 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) { 2351 // Some intrinsics use a scalar type for certain arguments. 2352 if (hasVectorInstrinsicScalarOpd(IntrinsicID, J)) { 2353 Lane[J] = Operands[J]; 2354 continue; 2355 } 2356 2357 Constant *Agg = Operands[J]->getAggregateElement(I); 2358 if (!Agg) 2359 return nullptr; 2360 2361 Lane[J] = Agg; 2362 } 2363 2364 // Use the regular scalar folding to simplify this column. 2365 Constant *Folded = 2366 ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI, Call); 2367 if (!Folded) 2368 return nullptr; 2369 Result[I] = Folded; 2370 } 2371 2372 return ConstantVector::get(Result); 2373 } 2374 2375 } // end anonymous namespace 2376 2377 Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F, 2378 ArrayRef<Constant *> Operands, 2379 const TargetLibraryInfo *TLI) { 2380 if (Call->isNoBuiltin() || Call->isStrictFP()) 2381 return nullptr; 2382 if (!F->hasName()) 2383 return nullptr; 2384 StringRef Name = F->getName(); 2385 2386 Type *Ty = F->getReturnType(); 2387 2388 if (auto *VTy = dyn_cast<VectorType>(Ty)) 2389 return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands, 2390 F->getParent()->getDataLayout(), TLI, Call); 2391 2392 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI, 2393 Call); 2394 } 2395 2396 bool llvm::isMathLibCallNoop(const CallBase *Call, 2397 const TargetLibraryInfo *TLI) { 2398 // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap 2399 // (and to some extent ConstantFoldScalarCall). 2400 if (Call->isNoBuiltin() || Call->isStrictFP()) 2401 return false; 2402 Function *F = Call->getCalledFunction(); 2403 if (!F) 2404 return false; 2405 2406 LibFunc Func; 2407 if (!TLI || !TLI->getLibFunc(*F, Func)) 2408 return false; 2409 2410 if (Call->getNumArgOperands() == 1) { 2411 if (ConstantFP *OpC = dyn_cast<ConstantFP>(Call->getArgOperand(0))) { 2412 const APFloat &Op = OpC->getValueAPF(); 2413 switch (Func) { 2414 case LibFunc_logl: 2415 case LibFunc_log: 2416 case LibFunc_logf: 2417 case LibFunc_log2l: 2418 case LibFunc_log2: 2419 case LibFunc_log2f: 2420 case LibFunc_log10l: 2421 case LibFunc_log10: 2422 case LibFunc_log10f: 2423 return Op.isNaN() || (!Op.isZero() && !Op.isNegative()); 2424 2425 case LibFunc_expl: 2426 case LibFunc_exp: 2427 case LibFunc_expf: 2428 // FIXME: These boundaries are slightly conservative. 2429 if (OpC->getType()->isDoubleTy()) 2430 return Op.compare(APFloat(-745.0)) != APFloat::cmpLessThan && 2431 Op.compare(APFloat(709.0)) != APFloat::cmpGreaterThan; 2432 if (OpC->getType()->isFloatTy()) 2433 return Op.compare(APFloat(-103.0f)) != APFloat::cmpLessThan && 2434 Op.compare(APFloat(88.0f)) != APFloat::cmpGreaterThan; 2435 break; 2436 2437 case LibFunc_exp2l: 2438 case LibFunc_exp2: 2439 case LibFunc_exp2f: 2440 // FIXME: These boundaries are slightly conservative. 2441 if (OpC->getType()->isDoubleTy()) 2442 return Op.compare(APFloat(-1074.0)) != APFloat::cmpLessThan && 2443 Op.compare(APFloat(1023.0)) != APFloat::cmpGreaterThan; 2444 if (OpC->getType()->isFloatTy()) 2445 return Op.compare(APFloat(-149.0f)) != APFloat::cmpLessThan && 2446 Op.compare(APFloat(127.0f)) != APFloat::cmpGreaterThan; 2447 break; 2448 2449 case LibFunc_sinl: 2450 case LibFunc_sin: 2451 case LibFunc_sinf: 2452 case LibFunc_cosl: 2453 case LibFunc_cos: 2454 case LibFunc_cosf: 2455 return !Op.isInfinity(); 2456 2457 case LibFunc_tanl: 2458 case LibFunc_tan: 2459 case LibFunc_tanf: { 2460 // FIXME: Stop using the host math library. 2461 // FIXME: The computation isn't done in the right precision. 2462 Type *Ty = OpC->getType(); 2463 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) { 2464 double OpV = getValueAsDouble(OpC); 2465 return ConstantFoldFP(tan, OpV, Ty) != nullptr; 2466 } 2467 break; 2468 } 2469 2470 case LibFunc_asinl: 2471 case LibFunc_asin: 2472 case LibFunc_asinf: 2473 case LibFunc_acosl: 2474 case LibFunc_acos: 2475 case LibFunc_acosf: 2476 return Op.compare(APFloat(Op.getSemantics(), "-1")) != 2477 APFloat::cmpLessThan && 2478 Op.compare(APFloat(Op.getSemantics(), "1")) != 2479 APFloat::cmpGreaterThan; 2480 2481 case LibFunc_sinh: 2482 case LibFunc_cosh: 2483 case LibFunc_sinhf: 2484 case LibFunc_coshf: 2485 case LibFunc_sinhl: 2486 case LibFunc_coshl: 2487 // FIXME: These boundaries are slightly conservative. 2488 if (OpC->getType()->isDoubleTy()) 2489 return Op.compare(APFloat(-710.0)) != APFloat::cmpLessThan && 2490 Op.compare(APFloat(710.0)) != APFloat::cmpGreaterThan; 2491 if (OpC->getType()->isFloatTy()) 2492 return Op.compare(APFloat(-89.0f)) != APFloat::cmpLessThan && 2493 Op.compare(APFloat(89.0f)) != APFloat::cmpGreaterThan; 2494 break; 2495 2496 case LibFunc_sqrtl: 2497 case LibFunc_sqrt: 2498 case LibFunc_sqrtf: 2499 return Op.isNaN() || Op.isZero() || !Op.isNegative(); 2500 2501 // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p, 2502 // maybe others? 2503 default: 2504 break; 2505 } 2506 } 2507 } 2508 2509 if (Call->getNumArgOperands() == 2) { 2510 ConstantFP *Op0C = dyn_cast<ConstantFP>(Call->getArgOperand(0)); 2511 ConstantFP *Op1C = dyn_cast<ConstantFP>(Call->getArgOperand(1)); 2512 if (Op0C && Op1C) { 2513 const APFloat &Op0 = Op0C->getValueAPF(); 2514 const APFloat &Op1 = Op1C->getValueAPF(); 2515 2516 switch (Func) { 2517 case LibFunc_powl: 2518 case LibFunc_pow: 2519 case LibFunc_powf: { 2520 // FIXME: Stop using the host math library. 2521 // FIXME: The computation isn't done in the right precision. 2522 Type *Ty = Op0C->getType(); 2523 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) { 2524 if (Ty == Op1C->getType()) { 2525 double Op0V = getValueAsDouble(Op0C); 2526 double Op1V = getValueAsDouble(Op1C); 2527 return ConstantFoldBinaryFP(pow, Op0V, Op1V, Ty) != nullptr; 2528 } 2529 } 2530 break; 2531 } 2532 2533 case LibFunc_fmodl: 2534 case LibFunc_fmod: 2535 case LibFunc_fmodf: 2536 return Op0.isNaN() || Op1.isNaN() || 2537 (!Op0.isInfinity() && !Op1.isZero()); 2538 2539 default: 2540 break; 2541 } 2542 } 2543 } 2544 2545 return false; 2546 } 2547