1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 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 implements folding of constants for LLVM. This implements the 10 // (internal) ConstantFold.h interface, which is used by the 11 // ConstantExpr::get* methods to automatically fold constants when possible. 12 // 13 // The current constant folding implementation is implemented in two pieces: the 14 // pieces that don't need DataLayout, and the pieces that do. This is to avoid 15 // a dependence in IR on Target. 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "ConstantFold.h" 20 #include "llvm/ADT/APSInt.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/IR/Constants.h" 23 #include "llvm/IR/DerivedTypes.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/GetElementPtrTypeIterator.h" 26 #include "llvm/IR/GlobalAlias.h" 27 #include "llvm/IR/GlobalVariable.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/Module.h" 30 #include "llvm/IR/Operator.h" 31 #include "llvm/IR/PatternMatch.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/ManagedStatic.h" 34 #include "llvm/Support/MathExtras.h" 35 using namespace llvm; 36 using namespace llvm::PatternMatch; 37 38 //===----------------------------------------------------------------------===// 39 // ConstantFold*Instruction Implementations 40 //===----------------------------------------------------------------------===// 41 42 /// Convert the specified vector Constant node to the specified vector type. 43 /// At this point, we know that the elements of the input vector constant are 44 /// all simple integer or FP values. 45 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 46 47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 48 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 49 50 // If this cast changes element count then we can't handle it here: 51 // doing so requires endianness information. This should be handled by 52 // Analysis/ConstantFolding.cpp 53 unsigned NumElts = DstTy->getNumElements(); 54 if (NumElts != CV->getType()->getVectorNumElements()) 55 return nullptr; 56 57 Type *DstEltTy = DstTy->getElementType(); 58 59 SmallVector<Constant*, 16> Result; 60 Type *Ty = IntegerType::get(CV->getContext(), 32); 61 for (unsigned i = 0; i != NumElts; ++i) { 62 Constant *C = 63 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 64 C = ConstantExpr::getBitCast(C, DstEltTy); 65 Result.push_back(C); 66 } 67 68 return ConstantVector::get(Result); 69 } 70 71 /// This function determines which opcode to use to fold two constant cast 72 /// expressions together. It uses CastInst::isEliminableCastPair to determine 73 /// the opcode. Consequently its just a wrapper around that function. 74 /// Determine if it is valid to fold a cast of a cast 75 static unsigned 76 foldConstantCastPair( 77 unsigned opc, ///< opcode of the second cast constant expression 78 ConstantExpr *Op, ///< the first cast constant expression 79 Type *DstTy ///< destination type of the first cast 80 ) { 81 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 82 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 83 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 84 85 // The types and opcodes for the two Cast constant expressions 86 Type *SrcTy = Op->getOperand(0)->getType(); 87 Type *MidTy = Op->getType(); 88 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 89 Instruction::CastOps secondOp = Instruction::CastOps(opc); 90 91 // Assume that pointers are never more than 64 bits wide, and only use this 92 // for the middle type. Otherwise we could end up folding away illegal 93 // bitcasts between address spaces with different sizes. 94 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 95 96 // Let CastInst::isEliminableCastPair do the heavy lifting. 97 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 98 nullptr, FakeIntPtrTy, nullptr); 99 } 100 101 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 102 Type *SrcTy = V->getType(); 103 if (SrcTy == DestTy) 104 return V; // no-op cast 105 106 // Check to see if we are casting a pointer to an aggregate to a pointer to 107 // the first element. If so, return the appropriate GEP instruction. 108 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 109 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 110 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 111 && PTy->getElementType()->isSized()) { 112 SmallVector<Value*, 8> IdxList; 113 Value *Zero = 114 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 115 IdxList.push_back(Zero); 116 Type *ElTy = PTy->getElementType(); 117 while (ElTy != DPTy->getElementType()) { 118 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 119 if (STy->getNumElements() == 0) break; 120 ElTy = STy->getElementType(0); 121 IdxList.push_back(Zero); 122 } else if (SequentialType *STy = 123 dyn_cast<SequentialType>(ElTy)) { 124 ElTy = STy->getElementType(); 125 IdxList.push_back(Zero); 126 } else { 127 break; 128 } 129 } 130 131 if (ElTy == DPTy->getElementType()) 132 // This GEP is inbounds because all indices are zero. 133 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), 134 V, IdxList); 135 } 136 137 // Handle casts from one vector constant to another. We know that the src 138 // and dest type have the same size (otherwise its an illegal cast). 139 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 140 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 141 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 142 "Not cast between same sized vectors!"); 143 SrcTy = nullptr; 144 // First, check for null. Undef is already handled. 145 if (isa<ConstantAggregateZero>(V)) 146 return Constant::getNullValue(DestTy); 147 148 // Handle ConstantVector and ConstantAggregateVector. 149 return BitCastConstantVector(V, DestPTy); 150 } 151 152 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 153 // This allows for other simplifications (although some of them 154 // can only be handled by Analysis/ConstantFolding.cpp). 155 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 156 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 157 } 158 159 // Finally, implement bitcast folding now. The code below doesn't handle 160 // bitcast right. 161 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 162 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 163 164 // Handle integral constant input. 165 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 166 if (DestTy->isIntegerTy()) 167 // Integral -> Integral. This is a no-op because the bit widths must 168 // be the same. Consequently, we just fold to V. 169 return V; 170 171 // See note below regarding the PPC_FP128 restriction. 172 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 173 return ConstantFP::get(DestTy->getContext(), 174 APFloat(DestTy->getFltSemantics(), 175 CI->getValue())); 176 177 // Otherwise, can't fold this (vector?) 178 return nullptr; 179 } 180 181 // Handle ConstantFP input: FP -> Integral. 182 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 183 // PPC_FP128 is really the sum of two consecutive doubles, where the first 184 // double is always stored first in memory, regardless of the target 185 // endianness. The memory layout of i128, however, depends on the target 186 // endianness, and so we can't fold this without target endianness 187 // information. This should instead be handled by 188 // Analysis/ConstantFolding.cpp 189 if (FP->getType()->isPPC_FP128Ty()) 190 return nullptr; 191 192 // Make sure dest type is compatible with the folded integer constant. 193 if (!DestTy->isIntegerTy()) 194 return nullptr; 195 196 return ConstantInt::get(FP->getContext(), 197 FP->getValueAPF().bitcastToAPInt()); 198 } 199 200 return nullptr; 201 } 202 203 204 /// V is an integer constant which only has a subset of its bytes used. 205 /// The bytes used are indicated by ByteStart (which is the first byte used, 206 /// counting from the least significant byte) and ByteSize, which is the number 207 /// of bytes used. 208 /// 209 /// This function analyzes the specified constant to see if the specified byte 210 /// range can be returned as a simplified constant. If so, the constant is 211 /// returned, otherwise null is returned. 212 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 213 unsigned ByteSize) { 214 assert(C->getType()->isIntegerTy() && 215 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 216 "Non-byte sized integer input"); 217 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 218 assert(ByteSize && "Must be accessing some piece"); 219 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 220 assert(ByteSize != CSize && "Should not extract everything"); 221 222 // Constant Integers are simple. 223 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 224 APInt V = CI->getValue(); 225 if (ByteStart) 226 V.lshrInPlace(ByteStart*8); 227 V = V.trunc(ByteSize*8); 228 return ConstantInt::get(CI->getContext(), V); 229 } 230 231 // In the input is a constant expr, we might be able to recursively simplify. 232 // If not, we definitely can't do anything. 233 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 234 if (!CE) return nullptr; 235 236 switch (CE->getOpcode()) { 237 default: return nullptr; 238 case Instruction::Or: { 239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 240 if (!RHS) 241 return nullptr; 242 243 // X | -1 -> -1. 244 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 245 if (RHSC->isMinusOne()) 246 return RHSC; 247 248 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 249 if (!LHS) 250 return nullptr; 251 return ConstantExpr::getOr(LHS, RHS); 252 } 253 case Instruction::And: { 254 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 255 if (!RHS) 256 return nullptr; 257 258 // X & 0 -> 0. 259 if (RHS->isNullValue()) 260 return RHS; 261 262 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 263 if (!LHS) 264 return nullptr; 265 return ConstantExpr::getAnd(LHS, RHS); 266 } 267 case Instruction::LShr: { 268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 269 if (!Amt) 270 return nullptr; 271 APInt ShAmt = Amt->getValue(); 272 // Cannot analyze non-byte shifts. 273 if ((ShAmt & 7) != 0) 274 return nullptr; 275 ShAmt.lshrInPlace(3); 276 277 // If the extract is known to be all zeros, return zero. 278 if (ShAmt.uge(CSize - ByteStart)) 279 return Constant::getNullValue( 280 IntegerType::get(CE->getContext(), ByteSize * 8)); 281 // If the extract is known to be fully in the input, extract it. 282 if (ShAmt.ule(CSize - (ByteStart + ByteSize))) 283 return ExtractConstantBytes(CE->getOperand(0), 284 ByteStart + ShAmt.getZExtValue(), ByteSize); 285 286 // TODO: Handle the 'partially zero' case. 287 return nullptr; 288 } 289 290 case Instruction::Shl: { 291 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 292 if (!Amt) 293 return nullptr; 294 APInt ShAmt = Amt->getValue(); 295 // Cannot analyze non-byte shifts. 296 if ((ShAmt & 7) != 0) 297 return nullptr; 298 ShAmt.lshrInPlace(3); 299 300 // If the extract is known to be all zeros, return zero. 301 if (ShAmt.uge(ByteStart + ByteSize)) 302 return Constant::getNullValue( 303 IntegerType::get(CE->getContext(), ByteSize * 8)); 304 // If the extract is known to be fully in the input, extract it. 305 if (ShAmt.ule(ByteStart)) 306 return ExtractConstantBytes(CE->getOperand(0), 307 ByteStart - ShAmt.getZExtValue(), ByteSize); 308 309 // TODO: Handle the 'partially zero' case. 310 return nullptr; 311 } 312 313 case Instruction::ZExt: { 314 unsigned SrcBitSize = 315 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 316 317 // If extracting something that is completely zero, return 0. 318 if (ByteStart*8 >= SrcBitSize) 319 return Constant::getNullValue(IntegerType::get(CE->getContext(), 320 ByteSize*8)); 321 322 // If exactly extracting the input, return it. 323 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 324 return CE->getOperand(0); 325 326 // If extracting something completely in the input, if the input is a 327 // multiple of 8 bits, recurse. 328 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 329 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 330 331 // Otherwise, if extracting a subset of the input, which is not multiple of 332 // 8 bits, do a shift and trunc to get the bits. 333 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 334 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 335 Constant *Res = CE->getOperand(0); 336 if (ByteStart) 337 Res = ConstantExpr::getLShr(Res, 338 ConstantInt::get(Res->getType(), ByteStart*8)); 339 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 340 ByteSize*8)); 341 } 342 343 // TODO: Handle the 'partially zero' case. 344 return nullptr; 345 } 346 } 347 } 348 349 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known 350 /// factors factored out. If Folded is false, return null if no factoring was 351 /// possible, to avoid endlessly bouncing an unfoldable expression back into the 352 /// top-level folder. 353 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, bool Folded) { 354 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 355 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 356 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 357 return ConstantExpr::getNUWMul(E, N); 358 } 359 360 if (StructType *STy = dyn_cast<StructType>(Ty)) 361 if (!STy->isPacked()) { 362 unsigned NumElems = STy->getNumElements(); 363 // An empty struct has size zero. 364 if (NumElems == 0) 365 return ConstantExpr::getNullValue(DestTy); 366 // Check for a struct with all members having the same size. 367 Constant *MemberSize = 368 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 369 bool AllSame = true; 370 for (unsigned i = 1; i != NumElems; ++i) 371 if (MemberSize != 372 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 373 AllSame = false; 374 break; 375 } 376 if (AllSame) { 377 Constant *N = ConstantInt::get(DestTy, NumElems); 378 return ConstantExpr::getNUWMul(MemberSize, N); 379 } 380 } 381 382 // Pointer size doesn't depend on the pointee type, so canonicalize them 383 // to an arbitrary pointee. 384 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 385 if (!PTy->getElementType()->isIntegerTy(1)) 386 return 387 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 388 PTy->getAddressSpace()), 389 DestTy, true); 390 391 // If there's no interesting folding happening, bail so that we don't create 392 // a constant that looks like it needs folding but really doesn't. 393 if (!Folded) 394 return nullptr; 395 396 // Base case: Get a regular sizeof expression. 397 Constant *C = ConstantExpr::getSizeOf(Ty); 398 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 399 DestTy, false), 400 C, DestTy); 401 return C; 402 } 403 404 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known 405 /// factors factored out. If Folded is false, return null if no factoring was 406 /// possible, to avoid endlessly bouncing an unfoldable expression back into the 407 /// top-level folder. 408 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, bool Folded) { 409 // The alignment of an array is equal to the alignment of the 410 // array element. Note that this is not always true for vectors. 411 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 412 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 413 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 414 DestTy, 415 false), 416 C, DestTy); 417 return C; 418 } 419 420 if (StructType *STy = dyn_cast<StructType>(Ty)) { 421 // Packed structs always have an alignment of 1. 422 if (STy->isPacked()) 423 return ConstantInt::get(DestTy, 1); 424 425 // Otherwise, struct alignment is the maximum alignment of any member. 426 // Without target data, we can't compare much, but we can check to see 427 // if all the members have the same alignment. 428 unsigned NumElems = STy->getNumElements(); 429 // An empty struct has minimal alignment. 430 if (NumElems == 0) 431 return ConstantInt::get(DestTy, 1); 432 // Check for a struct with all members having the same alignment. 433 Constant *MemberAlign = 434 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 435 bool AllSame = true; 436 for (unsigned i = 1; i != NumElems; ++i) 437 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 438 AllSame = false; 439 break; 440 } 441 if (AllSame) 442 return MemberAlign; 443 } 444 445 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 446 // to an arbitrary pointee. 447 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 448 if (!PTy->getElementType()->isIntegerTy(1)) 449 return 450 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 451 1), 452 PTy->getAddressSpace()), 453 DestTy, true); 454 455 // If there's no interesting folding happening, bail so that we don't create 456 // a constant that looks like it needs folding but really doesn't. 457 if (!Folded) 458 return nullptr; 459 460 // Base case: Get a regular alignof expression. 461 Constant *C = ConstantExpr::getAlignOf(Ty); 462 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 463 DestTy, false), 464 C, DestTy); 465 return C; 466 } 467 468 /// Return a ConstantExpr with type DestTy for offsetof on Ty and FieldNo, with 469 /// any known factors factored out. If Folded is false, return null if no 470 /// factoring was possible, to avoid endlessly bouncing an unfoldable expression 471 /// back into the top-level folder. 472 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, Type *DestTy, 473 bool Folded) { 474 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 475 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 476 DestTy, false), 477 FieldNo, DestTy); 478 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 479 return ConstantExpr::getNUWMul(E, N); 480 } 481 482 if (StructType *STy = dyn_cast<StructType>(Ty)) 483 if (!STy->isPacked()) { 484 unsigned NumElems = STy->getNumElements(); 485 // An empty struct has no members. 486 if (NumElems == 0) 487 return nullptr; 488 // Check for a struct with all members having the same size. 489 Constant *MemberSize = 490 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 491 bool AllSame = true; 492 for (unsigned i = 1; i != NumElems; ++i) 493 if (MemberSize != 494 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 495 AllSame = false; 496 break; 497 } 498 if (AllSame) { 499 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 500 false, 501 DestTy, 502 false), 503 FieldNo, DestTy); 504 return ConstantExpr::getNUWMul(MemberSize, N); 505 } 506 } 507 508 // If there's no interesting folding happening, bail so that we don't create 509 // a constant that looks like it needs folding but really doesn't. 510 if (!Folded) 511 return nullptr; 512 513 // Base case: Get a regular offsetof expression. 514 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 515 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 516 DestTy, false), 517 C, DestTy); 518 return C; 519 } 520 521 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 522 Type *DestTy) { 523 if (isa<UndefValue>(V)) { 524 // zext(undef) = 0, because the top bits will be zero. 525 // sext(undef) = 0, because the top bits will all be the same. 526 // [us]itofp(undef) = 0, because the result value is bounded. 527 if (opc == Instruction::ZExt || opc == Instruction::SExt || 528 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 529 return Constant::getNullValue(DestTy); 530 return UndefValue::get(DestTy); 531 } 532 533 if (V->isNullValue() && !DestTy->isX86_MMXTy() && 534 opc != Instruction::AddrSpaceCast) 535 return Constant::getNullValue(DestTy); 536 537 // If the cast operand is a constant expression, there's a few things we can 538 // do to try to simplify it. 539 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 540 if (CE->isCast()) { 541 // Try hard to fold cast of cast because they are often eliminable. 542 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 543 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 544 } else if (CE->getOpcode() == Instruction::GetElementPtr && 545 // Do not fold addrspacecast (gep 0, .., 0). It might make the 546 // addrspacecast uncanonicalized. 547 opc != Instruction::AddrSpaceCast && 548 // Do not fold bitcast (gep) with inrange index, as this loses 549 // information. 550 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() && 551 // Do not fold if the gep type is a vector, as bitcasting 552 // operand 0 of a vector gep will result in a bitcast between 553 // different sizes. 554 !CE->getType()->isVectorTy()) { 555 // If all of the indexes in the GEP are null values, there is no pointer 556 // adjustment going on. We might as well cast the source pointer. 557 bool isAllNull = true; 558 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 559 if (!CE->getOperand(i)->isNullValue()) { 560 isAllNull = false; 561 break; 562 } 563 if (isAllNull) 564 // This is casting one pointer type to another, always BitCast 565 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 566 } 567 } 568 569 // If the cast operand is a constant vector, perform the cast by 570 // operating on each element. In the cast of bitcasts, the element 571 // count may be mismatched; don't attempt to handle that here. 572 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 573 DestTy->isVectorTy() && 574 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 575 SmallVector<Constant*, 16> res; 576 VectorType *DestVecTy = cast<VectorType>(DestTy); 577 Type *DstEltTy = DestVecTy->getElementType(); 578 Type *Ty = IntegerType::get(V->getContext(), 32); 579 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 580 Constant *C = 581 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 582 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 583 } 584 return ConstantVector::get(res); 585 } 586 587 // We actually have to do a cast now. Perform the cast according to the 588 // opcode specified. 589 switch (opc) { 590 default: 591 llvm_unreachable("Failed to cast constant expression"); 592 case Instruction::FPTrunc: 593 case Instruction::FPExt: 594 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 595 bool ignored; 596 APFloat Val = FPC->getValueAPF(); 597 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() : 598 DestTy->isFloatTy() ? APFloat::IEEEsingle() : 599 DestTy->isDoubleTy() ? APFloat::IEEEdouble() : 600 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() : 601 DestTy->isFP128Ty() ? APFloat::IEEEquad() : 602 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() : 603 APFloat::Bogus(), 604 APFloat::rmNearestTiesToEven, &ignored); 605 return ConstantFP::get(V->getContext(), Val); 606 } 607 return nullptr; // Can't fold. 608 case Instruction::FPToUI: 609 case Instruction::FPToSI: 610 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 611 const APFloat &V = FPC->getValueAPF(); 612 bool ignored; 613 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 614 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI); 615 if (APFloat::opInvalidOp == 616 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) { 617 // Undefined behavior invoked - the destination type can't represent 618 // the input constant. 619 return UndefValue::get(DestTy); 620 } 621 return ConstantInt::get(FPC->getContext(), IntVal); 622 } 623 return nullptr; // Can't fold. 624 case Instruction::IntToPtr: //always treated as unsigned 625 if (V->isNullValue()) // Is it an integral null value? 626 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 627 return nullptr; // Other pointer types cannot be casted 628 case Instruction::PtrToInt: // always treated as unsigned 629 // Is it a null pointer value? 630 if (V->isNullValue()) 631 return ConstantInt::get(DestTy, 0); 632 // If this is a sizeof-like expression, pull out multiplications by 633 // known factors to expose them to subsequent folding. If it's an 634 // alignof-like expression, factor out known factors. 635 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 636 if (CE->getOpcode() == Instruction::GetElementPtr && 637 CE->getOperand(0)->isNullValue()) { 638 // FIXME: Looks like getFoldedSizeOf(), getFoldedOffsetOf() and 639 // getFoldedAlignOf() don't handle the case when DestTy is a vector of 640 // pointers yet. We end up in asserts in CastInst::getCastOpcode (see 641 // test/Analysis/ConstantFolding/cast-vector.ll). I've only seen this 642 // happen in one "real" C-code test case, so it does not seem to be an 643 // important optimization to handle vectors here. For now, simply bail 644 // out. 645 if (DestTy->isVectorTy()) 646 return nullptr; 647 GEPOperator *GEPO = cast<GEPOperator>(CE); 648 Type *Ty = GEPO->getSourceElementType(); 649 if (CE->getNumOperands() == 2) { 650 // Handle a sizeof-like expression. 651 Constant *Idx = CE->getOperand(1); 652 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 653 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 654 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 655 DestTy, false), 656 Idx, DestTy); 657 return ConstantExpr::getMul(C, Idx); 658 } 659 } else if (CE->getNumOperands() == 3 && 660 CE->getOperand(1)->isNullValue()) { 661 // Handle an alignof-like expression. 662 if (StructType *STy = dyn_cast<StructType>(Ty)) 663 if (!STy->isPacked()) { 664 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 665 if (CI->isOne() && 666 STy->getNumElements() == 2 && 667 STy->getElementType(0)->isIntegerTy(1)) { 668 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 669 } 670 } 671 // Handle an offsetof-like expression. 672 if (Ty->isStructTy() || Ty->isArrayTy()) { 673 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 674 DestTy, false)) 675 return C; 676 } 677 } 678 } 679 // Other pointer types cannot be casted 680 return nullptr; 681 case Instruction::UIToFP: 682 case Instruction::SIToFP: 683 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 684 const APInt &api = CI->getValue(); 685 APFloat apf(DestTy->getFltSemantics(), 686 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 687 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 688 APFloat::rmNearestTiesToEven); 689 return ConstantFP::get(V->getContext(), apf); 690 } 691 return nullptr; 692 case Instruction::ZExt: 693 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 694 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 695 return ConstantInt::get(V->getContext(), 696 CI->getValue().zext(BitWidth)); 697 } 698 return nullptr; 699 case Instruction::SExt: 700 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 701 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 702 return ConstantInt::get(V->getContext(), 703 CI->getValue().sext(BitWidth)); 704 } 705 return nullptr; 706 case Instruction::Trunc: { 707 if (V->getType()->isVectorTy()) 708 return nullptr; 709 710 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 711 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 712 return ConstantInt::get(V->getContext(), 713 CI->getValue().trunc(DestBitWidth)); 714 } 715 716 // The input must be a constantexpr. See if we can simplify this based on 717 // the bytes we are demanding. Only do this if the source and dest are an 718 // even multiple of a byte. 719 if ((DestBitWidth & 7) == 0 && 720 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 721 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 722 return Res; 723 724 return nullptr; 725 } 726 case Instruction::BitCast: 727 return FoldBitCast(V, DestTy); 728 case Instruction::AddrSpaceCast: 729 return nullptr; 730 } 731 } 732 733 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 734 Constant *V1, Constant *V2) { 735 // Check for i1 and vector true/false conditions. 736 if (Cond->isNullValue()) return V2; 737 if (Cond->isAllOnesValue()) return V1; 738 739 // If the condition is a vector constant, fold the result elementwise. 740 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 741 SmallVector<Constant*, 16> Result; 742 Type *Ty = IntegerType::get(CondV->getContext(), 32); 743 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 744 Constant *V; 745 Constant *V1Element = ConstantExpr::getExtractElement(V1, 746 ConstantInt::get(Ty, i)); 747 Constant *V2Element = ConstantExpr::getExtractElement(V2, 748 ConstantInt::get(Ty, i)); 749 auto *Cond = cast<Constant>(CondV->getOperand(i)); 750 if (V1Element == V2Element) { 751 V = V1Element; 752 } else if (isa<UndefValue>(Cond)) { 753 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 754 } else { 755 if (!isa<ConstantInt>(Cond)) break; 756 V = Cond->isNullValue() ? V2Element : V1Element; 757 } 758 Result.push_back(V); 759 } 760 761 // If we were able to build the vector, return it. 762 if (Result.size() == V1->getType()->getVectorNumElements()) 763 return ConstantVector::get(Result); 764 } 765 766 if (isa<UndefValue>(Cond)) { 767 if (isa<UndefValue>(V1)) return V1; 768 return V2; 769 } 770 if (isa<UndefValue>(V1)) return V2; 771 if (isa<UndefValue>(V2)) return V1; 772 if (V1 == V2) return V1; 773 774 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 775 if (TrueVal->getOpcode() == Instruction::Select) 776 if (TrueVal->getOperand(0) == Cond) 777 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 778 } 779 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 780 if (FalseVal->getOpcode() == Instruction::Select) 781 if (FalseVal->getOperand(0) == Cond) 782 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 783 } 784 785 return nullptr; 786 } 787 788 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 789 Constant *Idx) { 790 // extractelt undef, C -> undef 791 // extractelt C, undef -> undef 792 if (isa<UndefValue>(Val) || isa<UndefValue>(Idx)) 793 return UndefValue::get(Val->getType()->getVectorElementType()); 794 795 auto *CIdx = dyn_cast<ConstantInt>(Idx); 796 if (!CIdx) 797 return nullptr; 798 799 // ee({w,x,y,z}, wrong_value) -> undef 800 if (CIdx->uge(Val->getType()->getVectorNumElements())) 801 return UndefValue::get(Val->getType()->getVectorElementType()); 802 803 // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...) 804 if (auto *CE = dyn_cast<ConstantExpr>(Val)) { 805 if (CE->getOpcode() == Instruction::GetElementPtr) { 806 SmallVector<Constant *, 8> Ops; 807 Ops.reserve(CE->getNumOperands()); 808 for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) { 809 Constant *Op = CE->getOperand(i); 810 if (Op->getType()->isVectorTy()) { 811 Constant *ScalarOp = ConstantExpr::getExtractElement(Op, Idx); 812 if (!ScalarOp) 813 return nullptr; 814 Ops.push_back(ScalarOp); 815 } else 816 Ops.push_back(Op); 817 } 818 return CE->getWithOperands(Ops, CE->getType()->getVectorElementType(), 819 false, 820 Ops[0]->getType()->getPointerElementType()); 821 } 822 } 823 824 return Val->getAggregateElement(CIdx); 825 } 826 827 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 828 Constant *Elt, 829 Constant *Idx) { 830 if (isa<UndefValue>(Idx)) 831 return UndefValue::get(Val->getType()); 832 833 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 834 if (!CIdx) return nullptr; 835 836 // Do not iterate on scalable vector. The num of elements is unknown at 837 // compile-time. 838 VectorType *ValTy = cast<VectorType>(Val->getType()); 839 if (ValTy->isScalable()) 840 return nullptr; 841 842 unsigned NumElts = Val->getType()->getVectorNumElements(); 843 if (CIdx->uge(NumElts)) 844 return UndefValue::get(Val->getType()); 845 846 SmallVector<Constant*, 16> Result; 847 Result.reserve(NumElts); 848 auto *Ty = Type::getInt32Ty(Val->getContext()); 849 uint64_t IdxVal = CIdx->getZExtValue(); 850 for (unsigned i = 0; i != NumElts; ++i) { 851 if (i == IdxVal) { 852 Result.push_back(Elt); 853 continue; 854 } 855 856 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 857 Result.push_back(C); 858 } 859 860 return ConstantVector::get(Result); 861 } 862 863 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 864 Constant *V2, 865 Constant *Mask) { 866 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 867 Type *EltTy = V1->getType()->getVectorElementType(); 868 869 // Undefined shuffle mask -> undefined value. 870 if (isa<UndefValue>(Mask)) 871 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 872 873 // Don't break the bitcode reader hack. 874 if (isa<ConstantExpr>(Mask)) return nullptr; 875 876 // Do not iterate on scalable vector. The num of elements is unknown at 877 // compile-time. 878 VectorType *ValTy = cast<VectorType>(V1->getType()); 879 if (ValTy->isScalable()) 880 return nullptr; 881 882 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 883 884 // Loop over the shuffle mask, evaluating each element. 885 SmallVector<Constant*, 32> Result; 886 for (unsigned i = 0; i != MaskNumElts; ++i) { 887 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 888 if (Elt == -1) { 889 Result.push_back(UndefValue::get(EltTy)); 890 continue; 891 } 892 Constant *InElt; 893 if (unsigned(Elt) >= SrcNumElts*2) 894 InElt = UndefValue::get(EltTy); 895 else if (unsigned(Elt) >= SrcNumElts) { 896 Type *Ty = IntegerType::get(V2->getContext(), 32); 897 InElt = 898 ConstantExpr::getExtractElement(V2, 899 ConstantInt::get(Ty, Elt - SrcNumElts)); 900 } else { 901 Type *Ty = IntegerType::get(V1->getContext(), 32); 902 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 903 } 904 Result.push_back(InElt); 905 } 906 907 return ConstantVector::get(Result); 908 } 909 910 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 911 ArrayRef<unsigned> Idxs) { 912 // Base case: no indices, so return the entire value. 913 if (Idxs.empty()) 914 return Agg; 915 916 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 917 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 918 919 return nullptr; 920 } 921 922 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 923 Constant *Val, 924 ArrayRef<unsigned> Idxs) { 925 // Base case: no indices, so replace the entire value. 926 if (Idxs.empty()) 927 return Val; 928 929 unsigned NumElts; 930 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 931 NumElts = ST->getNumElements(); 932 else 933 NumElts = cast<SequentialType>(Agg->getType())->getNumElements(); 934 935 SmallVector<Constant*, 32> Result; 936 for (unsigned i = 0; i != NumElts; ++i) { 937 Constant *C = Agg->getAggregateElement(i); 938 if (!C) return nullptr; 939 940 if (Idxs[0] == i) 941 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 942 943 Result.push_back(C); 944 } 945 946 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 947 return ConstantStruct::get(ST, Result); 948 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 949 return ConstantArray::get(AT, Result); 950 return ConstantVector::get(Result); 951 } 952 953 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) { 954 assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected"); 955 956 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length 957 // vectors are always evaluated per element. 958 bool IsScalableVector = 959 C->getType()->isVectorTy() && C->getType()->getVectorIsScalable(); 960 bool HasScalarUndefOrScalableVectorUndef = 961 (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(C); 962 963 if (HasScalarUndefOrScalableVectorUndef) { 964 switch (static_cast<Instruction::UnaryOps>(Opcode)) { 965 case Instruction::FNeg: 966 return C; // -undef -> undef 967 case Instruction::UnaryOpsEnd: 968 llvm_unreachable("Invalid UnaryOp"); 969 } 970 } 971 972 // Constant should not be UndefValue, unless these are vector constants. 973 assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue"); 974 // We only have FP UnaryOps right now. 975 assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp"); 976 977 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 978 const APFloat &CV = CFP->getValueAPF(); 979 switch (Opcode) { 980 default: 981 break; 982 case Instruction::FNeg: 983 return ConstantFP::get(C->getContext(), neg(CV)); 984 } 985 } else if (VectorType *VTy = dyn_cast<VectorType>(C->getType())) { 986 // Do not iterate on scalable vector. The number of elements is unknown at 987 // compile-time. 988 if (IsScalableVector) 989 return nullptr; 990 991 // Fold each element and create a vector constant from those constants. 992 SmallVector<Constant*, 16> Result; 993 Type *Ty = IntegerType::get(VTy->getContext(), 32); 994 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 995 Constant *ExtractIdx = ConstantInt::get(Ty, i); 996 Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx); 997 998 Result.push_back(ConstantExpr::get(Opcode, Elt)); 999 } 1000 1001 return ConstantVector::get(Result); 1002 } 1003 1004 // We don't know how to fold this. 1005 return nullptr; 1006 } 1007 1008 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1, 1009 Constant *C2) { 1010 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 1011 1012 // Simplify BinOps with their identity values first. They are no-ops and we 1013 // can always return the other value, including undef or poison values. 1014 // FIXME: remove unnecessary duplicated identity patterns below. 1015 // FIXME: Use AllowRHSConstant with getBinOpIdentity to handle additional ops, 1016 // like X << 0 = X. 1017 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, C1->getType()); 1018 if (Identity) { 1019 if (C1 == Identity) 1020 return C2; 1021 if (C2 == Identity) 1022 return C1; 1023 } 1024 1025 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length 1026 // vectors are always evaluated per element. 1027 bool IsScalableVector = 1028 C1->getType()->isVectorTy() && C1->getType()->getVectorIsScalable(); 1029 bool HasScalarUndefOrScalableVectorUndef = 1030 (!C1->getType()->isVectorTy() || IsScalableVector) && 1031 (isa<UndefValue>(C1) || isa<UndefValue>(C2)); 1032 if (HasScalarUndefOrScalableVectorUndef) { 1033 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 1034 case Instruction::Xor: 1035 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1036 // Handle undef ^ undef -> 0 special case. This is a common 1037 // idiom (misuse). 1038 return Constant::getNullValue(C1->getType()); 1039 LLVM_FALLTHROUGH; 1040 case Instruction::Add: 1041 case Instruction::Sub: 1042 return UndefValue::get(C1->getType()); 1043 case Instruction::And: 1044 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 1045 return C1; 1046 return Constant::getNullValue(C1->getType()); // undef & X -> 0 1047 case Instruction::Mul: { 1048 // undef * undef -> undef 1049 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1050 return C1; 1051 const APInt *CV; 1052 // X * undef -> undef if X is odd 1053 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 1054 if ((*CV)[0]) 1055 return UndefValue::get(C1->getType()); 1056 1057 // X * undef -> 0 otherwise 1058 return Constant::getNullValue(C1->getType()); 1059 } 1060 case Instruction::SDiv: 1061 case Instruction::UDiv: 1062 // X / undef -> undef 1063 if (isa<UndefValue>(C2)) 1064 return C2; 1065 // undef / 0 -> undef 1066 // undef / 1 -> undef 1067 if (match(C2, m_Zero()) || match(C2, m_One())) 1068 return C1; 1069 // undef / X -> 0 otherwise 1070 return Constant::getNullValue(C1->getType()); 1071 case Instruction::URem: 1072 case Instruction::SRem: 1073 // X % undef -> undef 1074 if (match(C2, m_Undef())) 1075 return C2; 1076 // undef % 0 -> undef 1077 if (match(C2, m_Zero())) 1078 return C1; 1079 // undef % X -> 0 otherwise 1080 return Constant::getNullValue(C1->getType()); 1081 case Instruction::Or: // X | undef -> -1 1082 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 1083 return C1; 1084 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 1085 case Instruction::LShr: 1086 // X >>l undef -> undef 1087 if (isa<UndefValue>(C2)) 1088 return C2; 1089 // undef >>l 0 -> undef 1090 if (match(C2, m_Zero())) 1091 return C1; 1092 // undef >>l X -> 0 1093 return Constant::getNullValue(C1->getType()); 1094 case Instruction::AShr: 1095 // X >>a undef -> undef 1096 if (isa<UndefValue>(C2)) 1097 return C2; 1098 // undef >>a 0 -> undef 1099 if (match(C2, m_Zero())) 1100 return C1; 1101 // TODO: undef >>a X -> undef if the shift is exact 1102 // undef >>a X -> 0 1103 return Constant::getNullValue(C1->getType()); 1104 case Instruction::Shl: 1105 // X << undef -> undef 1106 if (isa<UndefValue>(C2)) 1107 return C2; 1108 // undef << 0 -> undef 1109 if (match(C2, m_Zero())) 1110 return C1; 1111 // undef << X -> 0 1112 return Constant::getNullValue(C1->getType()); 1113 case Instruction::FAdd: 1114 case Instruction::FSub: 1115 case Instruction::FMul: 1116 case Instruction::FDiv: 1117 case Instruction::FRem: 1118 // [any flop] undef, undef -> undef 1119 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1120 return C1; 1121 // [any flop] C, undef -> NaN 1122 // [any flop] undef, C -> NaN 1123 // We could potentially specialize NaN/Inf constants vs. 'normal' 1124 // constants (possibly differently depending on opcode and operand). This 1125 // would allow returning undef sometimes. But it is always safe to fold to 1126 // NaN because we can choose the undef operand as NaN, and any FP opcode 1127 // with a NaN operand will propagate NaN. 1128 return ConstantFP::getNaN(C1->getType()); 1129 case Instruction::BinaryOpsEnd: 1130 llvm_unreachable("Invalid BinaryOp"); 1131 } 1132 } 1133 1134 // Neither constant should be UndefValue, unless these are vector constants. 1135 assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue"); 1136 1137 // Handle simplifications when the RHS is a constant int. 1138 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1139 switch (Opcode) { 1140 case Instruction::Add: 1141 if (CI2->isZero()) return C1; // X + 0 == X 1142 break; 1143 case Instruction::Sub: 1144 if (CI2->isZero()) return C1; // X - 0 == X 1145 break; 1146 case Instruction::Mul: 1147 if (CI2->isZero()) return C2; // X * 0 == 0 1148 if (CI2->isOne()) 1149 return C1; // X * 1 == X 1150 break; 1151 case Instruction::UDiv: 1152 case Instruction::SDiv: 1153 if (CI2->isOne()) 1154 return C1; // X / 1 == X 1155 if (CI2->isZero()) 1156 return UndefValue::get(CI2->getType()); // X / 0 == undef 1157 break; 1158 case Instruction::URem: 1159 case Instruction::SRem: 1160 if (CI2->isOne()) 1161 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1162 if (CI2->isZero()) 1163 return UndefValue::get(CI2->getType()); // X % 0 == undef 1164 break; 1165 case Instruction::And: 1166 if (CI2->isZero()) return C2; // X & 0 == 0 1167 if (CI2->isMinusOne()) 1168 return C1; // X & -1 == X 1169 1170 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1171 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1172 if (CE1->getOpcode() == Instruction::ZExt) { 1173 unsigned DstWidth = CI2->getType()->getBitWidth(); 1174 unsigned SrcWidth = 1175 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1176 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1177 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1178 return C1; 1179 } 1180 1181 // If and'ing the address of a global with a constant, fold it. 1182 if (CE1->getOpcode() == Instruction::PtrToInt && 1183 isa<GlobalValue>(CE1->getOperand(0))) { 1184 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1185 1186 MaybeAlign GVAlign; 1187 1188 if (Module *TheModule = GV->getParent()) { 1189 GVAlign = GV->getPointerAlignment(TheModule->getDataLayout()); 1190 1191 // If the function alignment is not specified then assume that it 1192 // is 4. 1193 // This is dangerous; on x86, the alignment of the pointer 1194 // corresponds to the alignment of the function, but might be less 1195 // than 4 if it isn't explicitly specified. 1196 // However, a fix for this behaviour was reverted because it 1197 // increased code size (see https://reviews.llvm.org/D55115) 1198 // FIXME: This code should be deleted once existing targets have 1199 // appropriate defaults 1200 if (!GVAlign && isa<Function>(GV)) 1201 GVAlign = Align(4); 1202 } else if (isa<Function>(GV)) { 1203 // Without a datalayout we have to assume the worst case: that the 1204 // function pointer isn't aligned at all. 1205 GVAlign = llvm::None; 1206 } else { 1207 GVAlign = MaybeAlign(GV->getAlignment()); 1208 } 1209 1210 if (GVAlign && *GVAlign > 1) { 1211 unsigned DstWidth = CI2->getType()->getBitWidth(); 1212 unsigned SrcWidth = std::min(DstWidth, Log2(*GVAlign)); 1213 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1214 1215 // If checking bits we know are clear, return zero. 1216 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1217 return Constant::getNullValue(CI2->getType()); 1218 } 1219 } 1220 } 1221 break; 1222 case Instruction::Or: 1223 if (CI2->isZero()) return C1; // X | 0 == X 1224 if (CI2->isMinusOne()) 1225 return C2; // X | -1 == -1 1226 break; 1227 case Instruction::Xor: 1228 if (CI2->isZero()) return C1; // X ^ 0 == X 1229 1230 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1231 switch (CE1->getOpcode()) { 1232 default: break; 1233 case Instruction::ICmp: 1234 case Instruction::FCmp: 1235 // cmp pred ^ true -> cmp !pred 1236 assert(CI2->isOne()); 1237 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1238 pred = CmpInst::getInversePredicate(pred); 1239 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1240 CE1->getOperand(1)); 1241 } 1242 } 1243 break; 1244 case Instruction::AShr: 1245 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1246 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1247 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1248 return ConstantExpr::getLShr(C1, C2); 1249 break; 1250 } 1251 } else if (isa<ConstantInt>(C1)) { 1252 // If C1 is a ConstantInt and C2 is not, swap the operands. 1253 if (Instruction::isCommutative(Opcode)) 1254 return ConstantExpr::get(Opcode, C2, C1); 1255 } 1256 1257 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1258 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1259 const APInt &C1V = CI1->getValue(); 1260 const APInt &C2V = CI2->getValue(); 1261 switch (Opcode) { 1262 default: 1263 break; 1264 case Instruction::Add: 1265 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1266 case Instruction::Sub: 1267 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1268 case Instruction::Mul: 1269 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1270 case Instruction::UDiv: 1271 assert(!CI2->isZero() && "Div by zero handled above"); 1272 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1273 case Instruction::SDiv: 1274 assert(!CI2->isZero() && "Div by zero handled above"); 1275 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1276 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1277 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1278 case Instruction::URem: 1279 assert(!CI2->isZero() && "Div by zero handled above"); 1280 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1281 case Instruction::SRem: 1282 assert(!CI2->isZero() && "Div by zero handled above"); 1283 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1284 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1285 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1286 case Instruction::And: 1287 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1288 case Instruction::Or: 1289 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1290 case Instruction::Xor: 1291 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1292 case Instruction::Shl: 1293 if (C2V.ult(C1V.getBitWidth())) 1294 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1295 return UndefValue::get(C1->getType()); // too big shift is undef 1296 case Instruction::LShr: 1297 if (C2V.ult(C1V.getBitWidth())) 1298 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1299 return UndefValue::get(C1->getType()); // too big shift is undef 1300 case Instruction::AShr: 1301 if (C2V.ult(C1V.getBitWidth())) 1302 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1303 return UndefValue::get(C1->getType()); // too big shift is undef 1304 } 1305 } 1306 1307 switch (Opcode) { 1308 case Instruction::SDiv: 1309 case Instruction::UDiv: 1310 case Instruction::URem: 1311 case Instruction::SRem: 1312 case Instruction::LShr: 1313 case Instruction::AShr: 1314 case Instruction::Shl: 1315 if (CI1->isZero()) return C1; 1316 break; 1317 default: 1318 break; 1319 } 1320 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1321 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1322 const APFloat &C1V = CFP1->getValueAPF(); 1323 const APFloat &C2V = CFP2->getValueAPF(); 1324 APFloat C3V = C1V; // copy for modification 1325 switch (Opcode) { 1326 default: 1327 break; 1328 case Instruction::FAdd: 1329 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1330 return ConstantFP::get(C1->getContext(), C3V); 1331 case Instruction::FSub: 1332 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1333 return ConstantFP::get(C1->getContext(), C3V); 1334 case Instruction::FMul: 1335 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1336 return ConstantFP::get(C1->getContext(), C3V); 1337 case Instruction::FDiv: 1338 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1339 return ConstantFP::get(C1->getContext(), C3V); 1340 case Instruction::FRem: 1341 (void)C3V.mod(C2V); 1342 return ConstantFP::get(C1->getContext(), C3V); 1343 } 1344 } 1345 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1346 // Do not iterate on scalable vector. The number of elements is unknown at 1347 // compile-time. 1348 if (IsScalableVector) 1349 return nullptr; 1350 1351 // Fold each element and create a vector constant from those constants. 1352 SmallVector<Constant*, 16> Result; 1353 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1354 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1355 Constant *ExtractIdx = ConstantInt::get(Ty, i); 1356 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx); 1357 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx); 1358 1359 // If any element of a divisor vector is zero, the whole op is undef. 1360 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue()) 1361 return UndefValue::get(VTy); 1362 1363 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1364 } 1365 1366 return ConstantVector::get(Result); 1367 } 1368 1369 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1370 // There are many possible foldings we could do here. We should probably 1371 // at least fold add of a pointer with an integer into the appropriate 1372 // getelementptr. This will improve alias analysis a bit. 1373 1374 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1375 // (a + (b + c)). 1376 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1377 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1378 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1379 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1380 } 1381 } else if (isa<ConstantExpr>(C2)) { 1382 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1383 // other way if possible. 1384 if (Instruction::isCommutative(Opcode)) 1385 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1386 } 1387 1388 // i1 can be simplified in many cases. 1389 if (C1->getType()->isIntegerTy(1)) { 1390 switch (Opcode) { 1391 case Instruction::Add: 1392 case Instruction::Sub: 1393 return ConstantExpr::getXor(C1, C2); 1394 case Instruction::Mul: 1395 return ConstantExpr::getAnd(C1, C2); 1396 case Instruction::Shl: 1397 case Instruction::LShr: 1398 case Instruction::AShr: 1399 // We can assume that C2 == 0. If it were one the result would be 1400 // undefined because the shift value is as large as the bitwidth. 1401 return C1; 1402 case Instruction::SDiv: 1403 case Instruction::UDiv: 1404 // We can assume that C2 == 1. If it were zero the result would be 1405 // undefined through division by zero. 1406 return C1; 1407 case Instruction::URem: 1408 case Instruction::SRem: 1409 // We can assume that C2 == 1. If it were zero the result would be 1410 // undefined through division by zero. 1411 return ConstantInt::getFalse(C1->getContext()); 1412 default: 1413 break; 1414 } 1415 } 1416 1417 // We don't know how to fold this. 1418 return nullptr; 1419 } 1420 1421 /// This type is zero-sized if it's an array or structure of zero-sized types. 1422 /// The only leaf zero-sized type is an empty structure. 1423 static bool isMaybeZeroSizedType(Type *Ty) { 1424 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1425 if (STy->isOpaque()) return true; // Can't say. 1426 1427 // If all of elements have zero size, this does too. 1428 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1429 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1430 return true; 1431 1432 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1433 return isMaybeZeroSizedType(ATy->getElementType()); 1434 } 1435 return false; 1436 } 1437 1438 /// Compare the two constants as though they were getelementptr indices. 1439 /// This allows coercion of the types to be the same thing. 1440 /// 1441 /// If the two constants are the "same" (after coercion), return 0. If the 1442 /// first is less than the second, return -1, if the second is less than the 1443 /// first, return 1. If the constants are not integral, return -2. 1444 /// 1445 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1446 if (C1 == C2) return 0; 1447 1448 // Ok, we found a different index. If they are not ConstantInt, we can't do 1449 // anything with them. 1450 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1451 return -2; // don't know! 1452 1453 // We cannot compare the indices if they don't fit in an int64_t. 1454 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 || 1455 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64) 1456 return -2; // don't know! 1457 1458 // Ok, we have two differing integer indices. Sign extend them to be the same 1459 // type. 1460 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue(); 1461 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue(); 1462 1463 if (C1Val == C2Val) return 0; // They are equal 1464 1465 // If the type being indexed over is really just a zero sized type, there is 1466 // no pointer difference being made here. 1467 if (isMaybeZeroSizedType(ElTy)) 1468 return -2; // dunno. 1469 1470 // If they are really different, now that they are the same type, then we 1471 // found a difference! 1472 if (C1Val < C2Val) 1473 return -1; 1474 else 1475 return 1; 1476 } 1477 1478 /// This function determines if there is anything we can decide about the two 1479 /// constants provided. This doesn't need to handle simple things like 1480 /// ConstantFP comparisons, but should instead handle ConstantExprs. 1481 /// If we can determine that the two constants have a particular relation to 1482 /// each other, we should return the corresponding FCmpInst predicate, 1483 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1484 /// ConstantFoldCompareInstruction. 1485 /// 1486 /// To simplify this code we canonicalize the relation so that the first 1487 /// operand is always the most "complex" of the two. We consider ConstantFP 1488 /// to be the simplest, and ConstantExprs to be the most complex. 1489 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1490 assert(V1->getType() == V2->getType() && 1491 "Cannot compare values of different types!"); 1492 1493 // We do not know if a constant expression will evaluate to a number or NaN. 1494 // Therefore, we can only say that the relation is unordered or equal. 1495 if (V1 == V2) return FCmpInst::FCMP_UEQ; 1496 1497 if (!isa<ConstantExpr>(V1)) { 1498 if (!isa<ConstantExpr>(V2)) { 1499 // Simple case, use the standard constant folder. 1500 ConstantInt *R = nullptr; 1501 R = dyn_cast<ConstantInt>( 1502 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1503 if (R && !R->isZero()) 1504 return FCmpInst::FCMP_OEQ; 1505 R = dyn_cast<ConstantInt>( 1506 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1507 if (R && !R->isZero()) 1508 return FCmpInst::FCMP_OLT; 1509 R = dyn_cast<ConstantInt>( 1510 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1511 if (R && !R->isZero()) 1512 return FCmpInst::FCMP_OGT; 1513 1514 // Nothing more we can do 1515 return FCmpInst::BAD_FCMP_PREDICATE; 1516 } 1517 1518 // If the first operand is simple and second is ConstantExpr, swap operands. 1519 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1520 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1521 return FCmpInst::getSwappedPredicate(SwappedRelation); 1522 } else { 1523 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1524 // constantexpr or a simple constant. 1525 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1526 switch (CE1->getOpcode()) { 1527 case Instruction::FPTrunc: 1528 case Instruction::FPExt: 1529 case Instruction::UIToFP: 1530 case Instruction::SIToFP: 1531 // We might be able to do something with these but we don't right now. 1532 break; 1533 default: 1534 break; 1535 } 1536 } 1537 // There are MANY other foldings that we could perform here. They will 1538 // probably be added on demand, as they seem needed. 1539 return FCmpInst::BAD_FCMP_PREDICATE; 1540 } 1541 1542 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1543 const GlobalValue *GV2) { 1544 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1545 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1546 return true; 1547 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1548 Type *Ty = GVar->getValueType(); 1549 // A global with opaque type might end up being zero sized. 1550 if (!Ty->isSized()) 1551 return true; 1552 // A global with an empty type might lie at the address of any other 1553 // global. 1554 if (Ty->isEmptyTy()) 1555 return true; 1556 } 1557 return false; 1558 }; 1559 // Don't try to decide equality of aliases. 1560 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1561 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1562 return ICmpInst::ICMP_NE; 1563 return ICmpInst::BAD_ICMP_PREDICATE; 1564 } 1565 1566 /// This function determines if there is anything we can decide about the two 1567 /// constants provided. This doesn't need to handle simple things like integer 1568 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1569 /// If we can determine that the two constants have a particular relation to 1570 /// each other, we should return the corresponding ICmp predicate, otherwise 1571 /// return ICmpInst::BAD_ICMP_PREDICATE. 1572 /// 1573 /// To simplify this code we canonicalize the relation so that the first 1574 /// operand is always the most "complex" of the two. We consider simple 1575 /// constants (like ConstantInt) to be the simplest, followed by 1576 /// GlobalValues, followed by ConstantExpr's (the most complex). 1577 /// 1578 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1579 bool isSigned) { 1580 assert(V1->getType() == V2->getType() && 1581 "Cannot compare different types of values!"); 1582 if (V1 == V2) return ICmpInst::ICMP_EQ; 1583 1584 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1585 !isa<BlockAddress>(V1)) { 1586 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1587 !isa<BlockAddress>(V2)) { 1588 // We distilled this down to a simple case, use the standard constant 1589 // folder. 1590 ConstantInt *R = nullptr; 1591 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1592 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1593 if (R && !R->isZero()) 1594 return pred; 1595 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1596 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1597 if (R && !R->isZero()) 1598 return pred; 1599 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1600 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1601 if (R && !R->isZero()) 1602 return pred; 1603 1604 // If we couldn't figure it out, bail. 1605 return ICmpInst::BAD_ICMP_PREDICATE; 1606 } 1607 1608 // If the first operand is simple, swap operands. 1609 ICmpInst::Predicate SwappedRelation = 1610 evaluateICmpRelation(V2, V1, isSigned); 1611 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1612 return ICmpInst::getSwappedPredicate(SwappedRelation); 1613 1614 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1615 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1616 ICmpInst::Predicate SwappedRelation = 1617 evaluateICmpRelation(V2, V1, isSigned); 1618 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1619 return ICmpInst::getSwappedPredicate(SwappedRelation); 1620 return ICmpInst::BAD_ICMP_PREDICATE; 1621 } 1622 1623 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1624 // constant (which, since the types must match, means that it's a 1625 // ConstantPointerNull). 1626 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1627 return areGlobalsPotentiallyEqual(GV, GV2); 1628 } else if (isa<BlockAddress>(V2)) { 1629 return ICmpInst::ICMP_NE; // Globals never equal labels. 1630 } else { 1631 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1632 // GlobalVals can never be null unless they have external weak linkage. 1633 // We don't try to evaluate aliases here. 1634 // NOTE: We should not be doing this constant folding if null pointer 1635 // is considered valid for the function. But currently there is no way to 1636 // query it from the Constant type. 1637 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) && 1638 !NullPointerIsDefined(nullptr /* F */, 1639 GV->getType()->getAddressSpace())) 1640 return ICmpInst::ICMP_NE; 1641 } 1642 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1643 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1644 ICmpInst::Predicate SwappedRelation = 1645 evaluateICmpRelation(V2, V1, isSigned); 1646 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1647 return ICmpInst::getSwappedPredicate(SwappedRelation); 1648 return ICmpInst::BAD_ICMP_PREDICATE; 1649 } 1650 1651 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1652 // constant (which, since the types must match, means that it is a 1653 // ConstantPointerNull). 1654 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1655 // Block address in another function can't equal this one, but block 1656 // addresses in the current function might be the same if blocks are 1657 // empty. 1658 if (BA2->getFunction() != BA->getFunction()) 1659 return ICmpInst::ICMP_NE; 1660 } else { 1661 // Block addresses aren't null, don't equal the address of globals. 1662 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1663 "Canonicalization guarantee!"); 1664 return ICmpInst::ICMP_NE; 1665 } 1666 } else { 1667 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1668 // constantexpr, a global, block address, or a simple constant. 1669 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1670 Constant *CE1Op0 = CE1->getOperand(0); 1671 1672 switch (CE1->getOpcode()) { 1673 case Instruction::Trunc: 1674 case Instruction::FPTrunc: 1675 case Instruction::FPExt: 1676 case Instruction::FPToUI: 1677 case Instruction::FPToSI: 1678 break; // We can't evaluate floating point casts or truncations. 1679 1680 case Instruction::UIToFP: 1681 case Instruction::SIToFP: 1682 case Instruction::BitCast: 1683 case Instruction::ZExt: 1684 case Instruction::SExt: 1685 // We can't evaluate floating point casts or truncations. 1686 if (CE1Op0->getType()->isFPOrFPVectorTy()) 1687 break; 1688 1689 // If the cast is not actually changing bits, and the second operand is a 1690 // null pointer, do the comparison with the pre-casted value. 1691 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) { 1692 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1693 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1694 return evaluateICmpRelation(CE1Op0, 1695 Constant::getNullValue(CE1Op0->getType()), 1696 isSigned); 1697 } 1698 break; 1699 1700 case Instruction::GetElementPtr: { 1701 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1702 // Ok, since this is a getelementptr, we know that the constant has a 1703 // pointer type. Check the various cases. 1704 if (isa<ConstantPointerNull>(V2)) { 1705 // If we are comparing a GEP to a null pointer, check to see if the base 1706 // of the GEP equals the null pointer. 1707 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1708 if (GV->hasExternalWeakLinkage()) 1709 // Weak linkage GVals could be zero or not. We're comparing that 1710 // to null pointer so its greater-or-equal 1711 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1712 else 1713 // If its not weak linkage, the GVal must have a non-zero address 1714 // so the result is greater-than 1715 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1716 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1717 // If we are indexing from a null pointer, check to see if we have any 1718 // non-zero indices. 1719 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1720 if (!CE1->getOperand(i)->isNullValue()) 1721 // Offsetting from null, must not be equal. 1722 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1723 // Only zero indexes from null, must still be zero. 1724 return ICmpInst::ICMP_EQ; 1725 } 1726 // Otherwise, we can't really say if the first operand is null or not. 1727 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1728 if (isa<ConstantPointerNull>(CE1Op0)) { 1729 if (GV2->hasExternalWeakLinkage()) 1730 // Weak linkage GVals could be zero or not. We're comparing it to 1731 // a null pointer, so its less-or-equal 1732 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1733 else 1734 // If its not weak linkage, the GVal must have a non-zero address 1735 // so the result is less-than 1736 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1737 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1738 if (GV == GV2) { 1739 // If this is a getelementptr of the same global, then it must be 1740 // different. Because the types must match, the getelementptr could 1741 // only have at most one index, and because we fold getelementptr's 1742 // with a single zero index, it must be nonzero. 1743 assert(CE1->getNumOperands() == 2 && 1744 !CE1->getOperand(1)->isNullValue() && 1745 "Surprising getelementptr!"); 1746 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1747 } else { 1748 if (CE1GEP->hasAllZeroIndices()) 1749 return areGlobalsPotentiallyEqual(GV, GV2); 1750 return ICmpInst::BAD_ICMP_PREDICATE; 1751 } 1752 } 1753 } else { 1754 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1755 Constant *CE2Op0 = CE2->getOperand(0); 1756 1757 // There are MANY other foldings that we could perform here. They will 1758 // probably be added on demand, as they seem needed. 1759 switch (CE2->getOpcode()) { 1760 default: break; 1761 case Instruction::GetElementPtr: 1762 // By far the most common case to handle is when the base pointers are 1763 // obviously to the same global. 1764 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1765 // Don't know relative ordering, but check for inequality. 1766 if (CE1Op0 != CE2Op0) { 1767 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1768 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1769 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1770 cast<GlobalValue>(CE2Op0)); 1771 return ICmpInst::BAD_ICMP_PREDICATE; 1772 } 1773 // Ok, we know that both getelementptr instructions are based on the 1774 // same global. From this, we can precisely determine the relative 1775 // ordering of the resultant pointers. 1776 unsigned i = 1; 1777 1778 // The logic below assumes that the result of the comparison 1779 // can be determined by finding the first index that differs. 1780 // This doesn't work if there is over-indexing in any 1781 // subsequent indices, so check for that case first. 1782 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1783 !CE2->isGEPWithNoNotionalOverIndexing()) 1784 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1785 1786 // Compare all of the operands the GEP's have in common. 1787 gep_type_iterator GTI = gep_type_begin(CE1); 1788 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1789 ++i, ++GTI) 1790 switch (IdxCompare(CE1->getOperand(i), 1791 CE2->getOperand(i), GTI.getIndexedType())) { 1792 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1793 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1794 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1795 } 1796 1797 // Ok, we ran out of things they have in common. If any leftovers 1798 // are non-zero then we have a difference, otherwise we are equal. 1799 for (; i < CE1->getNumOperands(); ++i) 1800 if (!CE1->getOperand(i)->isNullValue()) { 1801 if (isa<ConstantInt>(CE1->getOperand(i))) 1802 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1803 else 1804 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1805 } 1806 1807 for (; i < CE2->getNumOperands(); ++i) 1808 if (!CE2->getOperand(i)->isNullValue()) { 1809 if (isa<ConstantInt>(CE2->getOperand(i))) 1810 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1811 else 1812 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1813 } 1814 return ICmpInst::ICMP_EQ; 1815 } 1816 } 1817 } 1818 break; 1819 } 1820 default: 1821 break; 1822 } 1823 } 1824 1825 return ICmpInst::BAD_ICMP_PREDICATE; 1826 } 1827 1828 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1829 Constant *C1, Constant *C2) { 1830 Type *ResultTy; 1831 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1832 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1833 VT->getNumElements()); 1834 else 1835 ResultTy = Type::getInt1Ty(C1->getContext()); 1836 1837 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1838 if (pred == FCmpInst::FCMP_FALSE) 1839 return Constant::getNullValue(ResultTy); 1840 1841 if (pred == FCmpInst::FCMP_TRUE) 1842 return Constant::getAllOnesValue(ResultTy); 1843 1844 // Handle some degenerate cases first 1845 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1846 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1847 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1848 // For EQ and NE, we can always pick a value for the undef to make the 1849 // predicate pass or fail, so we can return undef. 1850 // Also, if both operands are undef, we can return undef for int comparison. 1851 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1852 return UndefValue::get(ResultTy); 1853 1854 // Otherwise, for integer compare, pick the same value as the non-undef 1855 // operand, and fold it to true or false. 1856 if (isIntegerPredicate) 1857 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1858 1859 // Choosing NaN for the undef will always make unordered comparison succeed 1860 // and ordered comparison fails. 1861 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1862 } 1863 1864 // icmp eq/ne(null,GV) -> false/true 1865 if (C1->isNullValue()) { 1866 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1867 // Don't try to evaluate aliases. External weak GV can be null. 1868 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1869 !NullPointerIsDefined(nullptr /* F */, 1870 GV->getType()->getAddressSpace())) { 1871 if (pred == ICmpInst::ICMP_EQ) 1872 return ConstantInt::getFalse(C1->getContext()); 1873 else if (pred == ICmpInst::ICMP_NE) 1874 return ConstantInt::getTrue(C1->getContext()); 1875 } 1876 // icmp eq/ne(GV,null) -> false/true 1877 } else if (C2->isNullValue()) { 1878 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1879 // Don't try to evaluate aliases. External weak GV can be null. 1880 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1881 !NullPointerIsDefined(nullptr /* F */, 1882 GV->getType()->getAddressSpace())) { 1883 if (pred == ICmpInst::ICMP_EQ) 1884 return ConstantInt::getFalse(C1->getContext()); 1885 else if (pred == ICmpInst::ICMP_NE) 1886 return ConstantInt::getTrue(C1->getContext()); 1887 } 1888 } 1889 1890 // If the comparison is a comparison between two i1's, simplify it. 1891 if (C1->getType()->isIntegerTy(1)) { 1892 switch(pred) { 1893 case ICmpInst::ICMP_EQ: 1894 if (isa<ConstantInt>(C2)) 1895 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1896 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1897 case ICmpInst::ICMP_NE: 1898 return ConstantExpr::getXor(C1, C2); 1899 default: 1900 break; 1901 } 1902 } 1903 1904 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1905 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1906 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1907 switch (pred) { 1908 default: llvm_unreachable("Invalid ICmp Predicate"); 1909 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1910 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1911 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1912 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1913 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1914 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1915 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1916 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1917 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1918 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1919 } 1920 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1921 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1922 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1923 APFloat::cmpResult R = C1V.compare(C2V); 1924 switch (pred) { 1925 default: llvm_unreachable("Invalid FCmp Predicate"); 1926 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1927 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1928 case FCmpInst::FCMP_UNO: 1929 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1930 case FCmpInst::FCMP_ORD: 1931 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1932 case FCmpInst::FCMP_UEQ: 1933 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1934 R==APFloat::cmpEqual); 1935 case FCmpInst::FCMP_OEQ: 1936 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1937 case FCmpInst::FCMP_UNE: 1938 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1939 case FCmpInst::FCMP_ONE: 1940 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1941 R==APFloat::cmpGreaterThan); 1942 case FCmpInst::FCMP_ULT: 1943 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1944 R==APFloat::cmpLessThan); 1945 case FCmpInst::FCMP_OLT: 1946 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1947 case FCmpInst::FCMP_UGT: 1948 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1949 R==APFloat::cmpGreaterThan); 1950 case FCmpInst::FCMP_OGT: 1951 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1952 case FCmpInst::FCMP_ULE: 1953 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1954 case FCmpInst::FCMP_OLE: 1955 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1956 R==APFloat::cmpEqual); 1957 case FCmpInst::FCMP_UGE: 1958 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1959 case FCmpInst::FCMP_OGE: 1960 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1961 R==APFloat::cmpEqual); 1962 } 1963 } else if (C1->getType()->isVectorTy()) { 1964 // If we can constant fold the comparison of each element, constant fold 1965 // the whole vector comparison. 1966 SmallVector<Constant*, 4> ResElts; 1967 Type *Ty = IntegerType::get(C1->getContext(), 32); 1968 // Compare the elements, producing an i1 result or constant expr. 1969 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1970 Constant *C1E = 1971 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1972 Constant *C2E = 1973 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1974 1975 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1976 } 1977 1978 return ConstantVector::get(ResElts); 1979 } 1980 1981 if (C1->getType()->isFloatingPointTy() && 1982 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1983 // infinite recursive loop 1984 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1985 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1986 switch (evaluateFCmpRelation(C1, C2)) { 1987 default: llvm_unreachable("Unknown relation!"); 1988 case FCmpInst::FCMP_UNO: 1989 case FCmpInst::FCMP_ORD: 1990 case FCmpInst::FCMP_UNE: 1991 case FCmpInst::FCMP_ULT: 1992 case FCmpInst::FCMP_UGT: 1993 case FCmpInst::FCMP_ULE: 1994 case FCmpInst::FCMP_UGE: 1995 case FCmpInst::FCMP_TRUE: 1996 case FCmpInst::FCMP_FALSE: 1997 case FCmpInst::BAD_FCMP_PREDICATE: 1998 break; // Couldn't determine anything about these constants. 1999 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 2000 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 2001 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 2002 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 2003 break; 2004 case FCmpInst::FCMP_OLT: // We know that C1 < C2 2005 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 2006 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 2007 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 2008 break; 2009 case FCmpInst::FCMP_OGT: // We know that C1 > C2 2010 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 2011 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 2012 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 2013 break; 2014 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 2015 // We can only partially decide this relation. 2016 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 2017 Result = 0; 2018 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 2019 Result = 1; 2020 break; 2021 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 2022 // We can only partially decide this relation. 2023 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 2024 Result = 0; 2025 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 2026 Result = 1; 2027 break; 2028 case FCmpInst::FCMP_ONE: // We know that C1 != C2 2029 // We can only partially decide this relation. 2030 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 2031 Result = 0; 2032 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 2033 Result = 1; 2034 break; 2035 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2). 2036 // We can only partially decide this relation. 2037 if (pred == FCmpInst::FCMP_ONE) 2038 Result = 0; 2039 else if (pred == FCmpInst::FCMP_UEQ) 2040 Result = 1; 2041 break; 2042 } 2043 2044 // If we evaluated the result, return it now. 2045 if (Result != -1) 2046 return ConstantInt::get(ResultTy, Result); 2047 2048 } else { 2049 // Evaluate the relation between the two constants, per the predicate. 2050 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 2051 switch (evaluateICmpRelation(C1, C2, 2052 CmpInst::isSigned((CmpInst::Predicate)pred))) { 2053 default: llvm_unreachable("Unknown relational!"); 2054 case ICmpInst::BAD_ICMP_PREDICATE: 2055 break; // Couldn't determine anything about these constants. 2056 case ICmpInst::ICMP_EQ: // We know the constants are equal! 2057 // If we know the constants are equal, we can decide the result of this 2058 // computation precisely. 2059 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 2060 break; 2061 case ICmpInst::ICMP_ULT: 2062 switch (pred) { 2063 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 2064 Result = 1; break; 2065 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 2066 Result = 0; break; 2067 } 2068 break; 2069 case ICmpInst::ICMP_SLT: 2070 switch (pred) { 2071 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 2072 Result = 1; break; 2073 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 2074 Result = 0; break; 2075 } 2076 break; 2077 case ICmpInst::ICMP_UGT: 2078 switch (pred) { 2079 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 2080 Result = 1; break; 2081 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 2082 Result = 0; break; 2083 } 2084 break; 2085 case ICmpInst::ICMP_SGT: 2086 switch (pred) { 2087 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 2088 Result = 1; break; 2089 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 2090 Result = 0; break; 2091 } 2092 break; 2093 case ICmpInst::ICMP_ULE: 2094 if (pred == ICmpInst::ICMP_UGT) Result = 0; 2095 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 2096 break; 2097 case ICmpInst::ICMP_SLE: 2098 if (pred == ICmpInst::ICMP_SGT) Result = 0; 2099 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 2100 break; 2101 case ICmpInst::ICMP_UGE: 2102 if (pred == ICmpInst::ICMP_ULT) Result = 0; 2103 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 2104 break; 2105 case ICmpInst::ICMP_SGE: 2106 if (pred == ICmpInst::ICMP_SLT) Result = 0; 2107 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 2108 break; 2109 case ICmpInst::ICMP_NE: 2110 if (pred == ICmpInst::ICMP_EQ) Result = 0; 2111 if (pred == ICmpInst::ICMP_NE) Result = 1; 2112 break; 2113 } 2114 2115 // If we evaluated the result, return it now. 2116 if (Result != -1) 2117 return ConstantInt::get(ResultTy, Result); 2118 2119 // If the right hand side is a bitcast, try using its inverse to simplify 2120 // it by moving it to the left hand side. We can't do this if it would turn 2121 // a vector compare into a scalar compare or visa versa, or if it would turn 2122 // the operands into FP values. 2123 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 2124 Constant *CE2Op0 = CE2->getOperand(0); 2125 if (CE2->getOpcode() == Instruction::BitCast && 2126 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() && 2127 !CE2Op0->getType()->isFPOrFPVectorTy()) { 2128 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 2129 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 2130 } 2131 } 2132 2133 // If the left hand side is an extension, try eliminating it. 2134 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 2135 if ((CE1->getOpcode() == Instruction::SExt && 2136 ICmpInst::isSigned((ICmpInst::Predicate)pred)) || 2137 (CE1->getOpcode() == Instruction::ZExt && 2138 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){ 2139 Constant *CE1Op0 = CE1->getOperand(0); 2140 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 2141 if (CE1Inverse == CE1Op0) { 2142 // Check whether we can safely truncate the right hand side. 2143 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 2144 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 2145 C2->getType()) == C2) 2146 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 2147 } 2148 } 2149 } 2150 2151 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 2152 (C1->isNullValue() && !C2->isNullValue())) { 2153 // If C2 is a constant expr and C1 isn't, flip them around and fold the 2154 // other way if possible. 2155 // Also, if C1 is null and C2 isn't, flip them around. 2156 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 2157 return ConstantExpr::getICmp(pred, C2, C1); 2158 } 2159 } 2160 return nullptr; 2161 } 2162 2163 /// Test whether the given sequence of *normalized* indices is "inbounds". 2164 template<typename IndexTy> 2165 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 2166 // No indices means nothing that could be out of bounds. 2167 if (Idxs.empty()) return true; 2168 2169 // If the first index is zero, it's in bounds. 2170 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2171 2172 // If the first index is one and all the rest are zero, it's in bounds, 2173 // by the one-past-the-end rule. 2174 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) { 2175 if (!CI->isOne()) 2176 return false; 2177 } else { 2178 auto *CV = cast<ConstantDataVector>(Idxs[0]); 2179 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue()); 2180 if (!CI || !CI->isOne()) 2181 return false; 2182 } 2183 2184 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 2185 if (!cast<Constant>(Idxs[i])->isNullValue()) 2186 return false; 2187 return true; 2188 } 2189 2190 /// Test whether a given ConstantInt is in-range for a SequentialType. 2191 static bool isIndexInRangeOfArrayType(uint64_t NumElements, 2192 const ConstantInt *CI) { 2193 // We cannot bounds check the index if it doesn't fit in an int64_t. 2194 if (CI->getValue().getMinSignedBits() > 64) 2195 return false; 2196 2197 // A negative index or an index past the end of our sequential type is 2198 // considered out-of-range. 2199 int64_t IndexVal = CI->getSExtValue(); 2200 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2201 return false; 2202 2203 // Otherwise, it is in-range. 2204 return true; 2205 } 2206 2207 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, 2208 bool InBounds, 2209 Optional<unsigned> InRangeIndex, 2210 ArrayRef<Value *> Idxs) { 2211 if (Idxs.empty()) return C; 2212 2213 Type *GEPTy = GetElementPtrInst::getGEPReturnType( 2214 PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size())); 2215 2216 if (isa<UndefValue>(C)) 2217 return UndefValue::get(GEPTy); 2218 2219 Constant *Idx0 = cast<Constant>(Idxs[0]); 2220 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0))) 2221 return GEPTy->isVectorTy() && !C->getType()->isVectorTy() 2222 ? ConstantVector::getSplat( 2223 cast<VectorType>(GEPTy)->getNumElements(), C) 2224 : C; 2225 2226 if (C->isNullValue()) { 2227 bool isNull = true; 2228 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2229 if (!isa<UndefValue>(Idxs[i]) && 2230 !cast<Constant>(Idxs[i])->isNullValue()) { 2231 isNull = false; 2232 break; 2233 } 2234 if (isNull) { 2235 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2236 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2237 2238 assert(Ty && "Invalid indices for GEP!"); 2239 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2240 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2241 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2242 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements()); 2243 2244 // The GEP returns a vector of pointers when one of more of 2245 // its arguments is a vector. 2246 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) { 2247 if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) { 2248 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements()); 2249 break; 2250 } 2251 } 2252 2253 return Constant::getNullValue(GEPTy); 2254 } 2255 } 2256 2257 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2258 // Combine Indices - If the source pointer to this getelementptr instruction 2259 // is a getelementptr instruction, combine the indices of the two 2260 // getelementptr instructions into a single instruction. 2261 // 2262 if (CE->getOpcode() == Instruction::GetElementPtr) { 2263 gep_type_iterator LastI = gep_type_end(CE); 2264 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2265 I != E; ++I) 2266 LastI = I; 2267 2268 // We cannot combine indices if doing so would take us outside of an 2269 // array or vector. Doing otherwise could trick us if we evaluated such a 2270 // GEP as part of a load. 2271 // 2272 // e.g. Consider if the original GEP was: 2273 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2274 // i32 0, i32 0, i64 0) 2275 // 2276 // If we then tried to offset it by '8' to get to the third element, 2277 // an i8, we should *not* get: 2278 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2279 // i32 0, i32 0, i64 8) 2280 // 2281 // This GEP tries to index array element '8 which runs out-of-bounds. 2282 // Subsequent evaluation would get confused and produce erroneous results. 2283 // 2284 // The following prohibits such a GEP from being formed by checking to see 2285 // if the index is in-range with respect to an array. 2286 // TODO: This code may be extended to handle vectors as well. 2287 bool PerformFold = false; 2288 if (Idx0->isNullValue()) 2289 PerformFold = true; 2290 else if (LastI.isSequential()) 2291 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2292 PerformFold = (!LastI.isBoundedSequential() || 2293 isIndexInRangeOfArrayType( 2294 LastI.getSequentialNumElements(), CI)) && 2295 !CE->getOperand(CE->getNumOperands() - 1) 2296 ->getType() 2297 ->isVectorTy(); 2298 2299 if (PerformFold) { 2300 SmallVector<Value*, 16> NewIndices; 2301 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2302 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2303 2304 // Add the last index of the source with the first index of the new GEP. 2305 // Make sure to handle the case when they are actually different types. 2306 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2307 // Otherwise it must be an array. 2308 if (!Idx0->isNullValue()) { 2309 Type *IdxTy = Combined->getType(); 2310 if (IdxTy != Idx0->getType()) { 2311 unsigned CommonExtendedWidth = 2312 std::max(IdxTy->getIntegerBitWidth(), 2313 Idx0->getType()->getIntegerBitWidth()); 2314 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2315 2316 Type *CommonTy = 2317 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2318 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2319 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2320 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2321 } else { 2322 Combined = 2323 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2324 } 2325 } 2326 2327 NewIndices.push_back(Combined); 2328 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2329 2330 // The combined GEP normally inherits its index inrange attribute from 2331 // the inner GEP, but if the inner GEP's last index was adjusted by the 2332 // outer GEP, any inbounds attribute on that index is invalidated. 2333 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex(); 2334 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue()) 2335 IRIndex = None; 2336 2337 return ConstantExpr::getGetElementPtr( 2338 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2339 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(), 2340 IRIndex); 2341 } 2342 } 2343 2344 // Attempt to fold casts to the same type away. For example, folding: 2345 // 2346 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2347 // i64 0, i64 0) 2348 // into: 2349 // 2350 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2351 // 2352 // Don't fold if the cast is changing address spaces. 2353 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2354 PointerType *SrcPtrTy = 2355 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2356 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2357 if (SrcPtrTy && DstPtrTy) { 2358 ArrayType *SrcArrayTy = 2359 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2360 ArrayType *DstArrayTy = 2361 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2362 if (SrcArrayTy && DstArrayTy 2363 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2364 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2365 return ConstantExpr::getGetElementPtr(SrcArrayTy, 2366 (Constant *)CE->getOperand(0), 2367 Idxs, InBounds, InRangeIndex); 2368 } 2369 } 2370 } 2371 2372 // Check to see if any array indices are not within the corresponding 2373 // notional array or vector bounds. If so, try to determine if they can be 2374 // factored out into preceding dimensions. 2375 SmallVector<Constant *, 8> NewIdxs; 2376 Type *Ty = PointeeTy; 2377 Type *Prev = C->getType(); 2378 bool Unknown = 2379 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]); 2380 for (unsigned i = 1, e = Idxs.size(); i != e; 2381 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2382 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) { 2383 // We don't know if it's in range or not. 2384 Unknown = true; 2385 continue; 2386 } 2387 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1])) 2388 // Skip if the type of the previous index is not supported. 2389 continue; 2390 if (InRangeIndex && i == *InRangeIndex + 1) { 2391 // If an index is marked inrange, we cannot apply this canonicalization to 2392 // the following index, as that will cause the inrange index to point to 2393 // the wrong element. 2394 continue; 2395 } 2396 if (isa<StructType>(Ty)) { 2397 // The verify makes sure that GEPs into a struct are in range. 2398 continue; 2399 } 2400 auto *STy = cast<SequentialType>(Ty); 2401 if (isa<VectorType>(STy)) { 2402 // There can be awkward padding in after a non-power of two vector. 2403 Unknown = true; 2404 continue; 2405 } 2406 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2407 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI)) 2408 // It's in range, skip to the next index. 2409 continue; 2410 if (CI->getSExtValue() < 0) { 2411 // It's out of range and negative, don't try to factor it. 2412 Unknown = true; 2413 continue; 2414 } 2415 } else { 2416 auto *CV = cast<ConstantDataVector>(Idxs[i]); 2417 bool InRange = true; 2418 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) { 2419 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I)); 2420 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI); 2421 if (CI->getSExtValue() < 0) { 2422 Unknown = true; 2423 break; 2424 } 2425 } 2426 if (InRange || Unknown) 2427 // It's in range, skip to the next index. 2428 // It's out of range and negative, don't try to factor it. 2429 continue; 2430 } 2431 if (isa<StructType>(Prev)) { 2432 // It's out of range, but the prior dimension is a struct 2433 // so we can't do anything about it. 2434 Unknown = true; 2435 continue; 2436 } 2437 // It's out of range, but we can factor it into the prior 2438 // dimension. 2439 NewIdxs.resize(Idxs.size()); 2440 // Determine the number of elements in our sequential type. 2441 uint64_t NumElements = STy->getArrayNumElements(); 2442 2443 // Expand the current index or the previous index to a vector from a scalar 2444 // if necessary. 2445 Constant *CurrIdx = cast<Constant>(Idxs[i]); 2446 auto *PrevIdx = 2447 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]); 2448 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy(); 2449 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy(); 2450 bool UseVector = IsCurrIdxVector || IsPrevIdxVector; 2451 2452 if (!IsCurrIdxVector && IsPrevIdxVector) 2453 CurrIdx = ConstantDataVector::getSplat( 2454 PrevIdx->getType()->getVectorNumElements(), CurrIdx); 2455 2456 if (!IsPrevIdxVector && IsCurrIdxVector) 2457 PrevIdx = ConstantDataVector::getSplat( 2458 CurrIdx->getType()->getVectorNumElements(), PrevIdx); 2459 2460 Constant *Factor = 2461 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements); 2462 if (UseVector) 2463 Factor = ConstantDataVector::getSplat( 2464 IsPrevIdxVector ? PrevIdx->getType()->getVectorNumElements() 2465 : CurrIdx->getType()->getVectorNumElements(), 2466 Factor); 2467 2468 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor); 2469 2470 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor); 2471 2472 unsigned CommonExtendedWidth = 2473 std::max(PrevIdx->getType()->getScalarSizeInBits(), 2474 Div->getType()->getScalarSizeInBits()); 2475 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2476 2477 // Before adding, extend both operands to i64 to avoid 2478 // overflow trouble. 2479 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth); 2480 if (UseVector) 2481 ExtendedTy = VectorType::get( 2482 ExtendedTy, IsPrevIdxVector 2483 ? PrevIdx->getType()->getVectorNumElements() 2484 : CurrIdx->getType()->getVectorNumElements()); 2485 2486 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2487 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy); 2488 2489 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2490 Div = ConstantExpr::getSExt(Div, ExtendedTy); 2491 2492 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div); 2493 } 2494 2495 // If we did any factoring, start over with the adjusted indices. 2496 if (!NewIdxs.empty()) { 2497 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2498 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2499 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds, 2500 InRangeIndex); 2501 } 2502 2503 // If all indices are known integers and normalized, we can do a simple 2504 // check for the "inbounds" property. 2505 if (!Unknown && !InBounds) 2506 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2507 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2508 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs, 2509 /*InBounds=*/true, InRangeIndex); 2510 2511 return nullptr; 2512 } 2513