1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements folding of constants for LLVM. This implements the 11 // (internal) ConstantFold.h interface, which is used by the 12 // ConstantExpr::get* methods to automatically fold constants when possible. 13 // 14 // The current constant folding implementation is implemented in two pieces: the 15 // pieces that don't need DataLayout, and the pieces that do. This is to avoid 16 // a dependence in IR on Target. 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "ConstantFold.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/Operator.h" 30 #include "llvm/IR/PatternMatch.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/ManagedStatic.h" 33 #include "llvm/Support/MathExtras.h" 34 using namespace llvm; 35 using namespace llvm::PatternMatch; 36 37 //===----------------------------------------------------------------------===// 38 // ConstantFold*Instruction Implementations 39 //===----------------------------------------------------------------------===// 40 41 /// Convert the specified vector Constant node to the specified vector type. 42 /// At this point, we know that the elements of the input vector constant are 43 /// all simple integer or FP values. 44 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 45 46 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 47 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 48 49 // If this cast changes element count then we can't handle it here: 50 // doing so requires endianness information. This should be handled by 51 // Analysis/ConstantFolding.cpp 52 unsigned NumElts = DstTy->getNumElements(); 53 if (NumElts != CV->getType()->getVectorNumElements()) 54 return nullptr; 55 56 Type *DstEltTy = DstTy->getElementType(); 57 58 SmallVector<Constant*, 16> Result; 59 Type *Ty = IntegerType::get(CV->getContext(), 32); 60 for (unsigned i = 0; i != NumElts; ++i) { 61 Constant *C = 62 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 63 C = ConstantExpr::getBitCast(C, DstEltTy); 64 Result.push_back(C); 65 } 66 67 return ConstantVector::get(Result); 68 } 69 70 /// This function determines which opcode to use to fold two constant cast 71 /// expressions together. It uses CastInst::isEliminableCastPair to determine 72 /// the opcode. Consequently its just a wrapper around that function. 73 /// @brief Determine if it is valid to fold a cast of a cast 74 static unsigned 75 foldConstantCastPair( 76 unsigned opc, ///< opcode of the second cast constant expression 77 ConstantExpr *Op, ///< the first cast constant expression 78 Type *DstTy ///< destination type of the first cast 79 ) { 80 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 81 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 82 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 83 84 // The types and opcodes for the two Cast constant expressions 85 Type *SrcTy = Op->getOperand(0)->getType(); 86 Type *MidTy = Op->getType(); 87 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 88 Instruction::CastOps secondOp = Instruction::CastOps(opc); 89 90 // Assume that pointers are never more than 64 bits wide, and only use this 91 // for the middle type. Otherwise we could end up folding away illegal 92 // bitcasts between address spaces with different sizes. 93 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 94 95 // Let CastInst::isEliminableCastPair do the heavy lifting. 96 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 97 nullptr, FakeIntPtrTy, nullptr); 98 } 99 100 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 101 Type *SrcTy = V->getType(); 102 if (SrcTy == DestTy) 103 return V; // no-op cast 104 105 // Check to see if we are casting a pointer to an aggregate to a pointer to 106 // the first element. If so, return the appropriate GEP instruction. 107 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 108 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 109 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 110 && PTy->getElementType()->isSized()) { 111 SmallVector<Value*, 8> IdxList; 112 Value *Zero = 113 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 114 IdxList.push_back(Zero); 115 Type *ElTy = PTy->getElementType(); 116 while (ElTy != DPTy->getElementType()) { 117 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 118 if (STy->getNumElements() == 0) break; 119 ElTy = STy->getElementType(0); 120 IdxList.push_back(Zero); 121 } else if (SequentialType *STy = 122 dyn_cast<SequentialType>(ElTy)) { 123 if (ElTy->isPointerTy()) break; // Can't index into pointers! 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 = V.lshr(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->isAllOnesValue()) 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 unsigned ShAmt = Amt->getZExtValue(); 272 // Cannot analyze non-byte shifts. 273 if ((ShAmt & 7) != 0) 274 return nullptr; 275 ShAmt >>= 3; 276 277 // If the extract is known to be all zeros, return zero. 278 if (ByteStart >= CSize-ShAmt) 279 return Constant::getNullValue(IntegerType::get(CE->getContext(), 280 ByteSize*8)); 281 // If the extract is known to be fully in the input, extract it. 282 if (ByteStart+ByteSize+ShAmt <= CSize) 283 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 284 285 // TODO: Handle the 'partially zero' case. 286 return nullptr; 287 } 288 289 case Instruction::Shl: { 290 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 291 if (!Amt) 292 return nullptr; 293 unsigned ShAmt = Amt->getZExtValue(); 294 // Cannot analyze non-byte shifts. 295 if ((ShAmt & 7) != 0) 296 return nullptr; 297 ShAmt >>= 3; 298 299 // If the extract is known to be all zeros, return zero. 300 if (ByteStart+ByteSize <= ShAmt) 301 return Constant::getNullValue(IntegerType::get(CE->getContext(), 302 ByteSize*8)); 303 // If the extract is known to be fully in the input, extract it. 304 if (ByteStart >= ShAmt) 305 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 306 307 // TODO: Handle the 'partially zero' case. 308 return nullptr; 309 } 310 311 case Instruction::ZExt: { 312 unsigned SrcBitSize = 313 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 314 315 // If extracting something that is completely zero, return 0. 316 if (ByteStart*8 >= SrcBitSize) 317 return Constant::getNullValue(IntegerType::get(CE->getContext(), 318 ByteSize*8)); 319 320 // If exactly extracting the input, return it. 321 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 322 return CE->getOperand(0); 323 324 // If extracting something completely in the input, if if the input is a 325 // multiple of 8 bits, recurse. 326 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 327 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 328 329 // Otherwise, if extracting a subset of the input, which is not multiple of 330 // 8 bits, do a shift and trunc to get the bits. 331 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 332 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 333 Constant *Res = CE->getOperand(0); 334 if (ByteStart) 335 Res = ConstantExpr::getLShr(Res, 336 ConstantInt::get(Res->getType(), ByteStart*8)); 337 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 338 ByteSize*8)); 339 } 340 341 // TODO: Handle the 'partially zero' case. 342 return nullptr; 343 } 344 } 345 } 346 347 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known 348 /// factors factored out. If Folded is false, return null if no factoring was 349 /// possible, to avoid endlessly bouncing an unfoldable expression back into the 350 /// top-level folder. 351 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 352 bool Folded) { 353 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 354 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 355 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 356 return ConstantExpr::getNUWMul(E, N); 357 } 358 359 if (StructType *STy = dyn_cast<StructType>(Ty)) 360 if (!STy->isPacked()) { 361 unsigned NumElems = STy->getNumElements(); 362 // An empty struct has size zero. 363 if (NumElems == 0) 364 return ConstantExpr::getNullValue(DestTy); 365 // Check for a struct with all members having the same size. 366 Constant *MemberSize = 367 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 368 bool AllSame = true; 369 for (unsigned i = 1; i != NumElems; ++i) 370 if (MemberSize != 371 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 372 AllSame = false; 373 break; 374 } 375 if (AllSame) { 376 Constant *N = ConstantInt::get(DestTy, NumElems); 377 return ConstantExpr::getNUWMul(MemberSize, N); 378 } 379 } 380 381 // Pointer size doesn't depend on the pointee type, so canonicalize them 382 // to an arbitrary pointee. 383 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 384 if (!PTy->getElementType()->isIntegerTy(1)) 385 return 386 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 387 PTy->getAddressSpace()), 388 DestTy, true); 389 390 // If there's no interesting folding happening, bail so that we don't create 391 // a constant that looks like it needs folding but really doesn't. 392 if (!Folded) 393 return nullptr; 394 395 // Base case: Get a regular sizeof expression. 396 Constant *C = ConstantExpr::getSizeOf(Ty); 397 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 398 DestTy, false), 399 C, DestTy); 400 return C; 401 } 402 403 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known 404 /// factors factored out. If Folded is false, return null if no factoring was 405 /// possible, to avoid endlessly bouncing an unfoldable expression back into the 406 /// top-level folder. 407 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 408 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, 473 Type *DestTy, 474 bool Folded) { 475 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 476 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 477 DestTy, false), 478 FieldNo, DestTy); 479 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 480 return ConstantExpr::getNUWMul(E, N); 481 } 482 483 if (StructType *STy = dyn_cast<StructType>(Ty)) 484 if (!STy->isPacked()) { 485 unsigned NumElems = STy->getNumElements(); 486 // An empty struct has no members. 487 if (NumElems == 0) 488 return nullptr; 489 // Check for a struct with all members having the same size. 490 Constant *MemberSize = 491 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 492 bool AllSame = true; 493 for (unsigned i = 1; i != NumElems; ++i) 494 if (MemberSize != 495 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 496 AllSame = false; 497 break; 498 } 499 if (AllSame) { 500 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 501 false, 502 DestTy, 503 false), 504 FieldNo, DestTy); 505 return ConstantExpr::getNUWMul(MemberSize, N); 506 } 507 } 508 509 // If there's no interesting folding happening, bail so that we don't create 510 // a constant that looks like it needs folding but really doesn't. 511 if (!Folded) 512 return nullptr; 513 514 // Base case: Get a regular offsetof expression. 515 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 516 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 517 DestTy, false), 518 C, DestTy); 519 return C; 520 } 521 522 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 523 Type *DestTy) { 524 if (isa<UndefValue>(V)) { 525 // zext(undef) = 0, because the top bits will be zero. 526 // sext(undef) = 0, because the top bits will all be the same. 527 // [us]itofp(undef) = 0, because the result value is bounded. 528 if (opc == Instruction::ZExt || opc == Instruction::SExt || 529 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 530 return Constant::getNullValue(DestTy); 531 return UndefValue::get(DestTy); 532 } 533 534 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 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 // If all of the indexes in the GEP are null values, there is no pointer 549 // adjustment going on. We might as well cast the source pointer. 550 bool isAllNull = true; 551 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 552 if (!CE->getOperand(i)->isNullValue()) { 553 isAllNull = false; 554 break; 555 } 556 if (isAllNull) 557 // This is casting one pointer type to another, always BitCast 558 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 559 } 560 } 561 562 // If the cast operand is a constant vector, perform the cast by 563 // operating on each element. In the cast of bitcasts, the element 564 // count may be mismatched; don't attempt to handle that here. 565 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 566 DestTy->isVectorTy() && 567 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 568 SmallVector<Constant*, 16> res; 569 VectorType *DestVecTy = cast<VectorType>(DestTy); 570 Type *DstEltTy = DestVecTy->getElementType(); 571 Type *Ty = IntegerType::get(V->getContext(), 32); 572 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 573 Constant *C = 574 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 575 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 576 } 577 return ConstantVector::get(res); 578 } 579 580 // We actually have to do a cast now. Perform the cast according to the 581 // opcode specified. 582 switch (opc) { 583 default: 584 llvm_unreachable("Failed to cast constant expression"); 585 case Instruction::FPTrunc: 586 case Instruction::FPExt: 587 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 588 bool ignored; 589 APFloat Val = FPC->getValueAPF(); 590 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 591 DestTy->isFloatTy() ? APFloat::IEEEsingle : 592 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 593 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 594 DestTy->isFP128Ty() ? APFloat::IEEEquad : 595 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble : 596 APFloat::Bogus, 597 APFloat::rmNearestTiesToEven, &ignored); 598 return ConstantFP::get(V->getContext(), Val); 599 } 600 return nullptr; // Can't fold. 601 case Instruction::FPToUI: 602 case Instruction::FPToSI: 603 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 604 const APFloat &V = FPC->getValueAPF(); 605 bool ignored; 606 uint64_t x[2]; 607 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 608 if (APFloat::opInvalidOp == 609 V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 610 APFloat::rmTowardZero, &ignored)) { 611 // Undefined behavior invoked - the destination type can't represent 612 // the input constant. 613 return UndefValue::get(DestTy); 614 } 615 APInt Val(DestBitWidth, x); 616 return ConstantInt::get(FPC->getContext(), Val); 617 } 618 return nullptr; // Can't fold. 619 case Instruction::IntToPtr: //always treated as unsigned 620 if (V->isNullValue()) // Is it an integral null value? 621 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 622 return nullptr; // Other pointer types cannot be casted 623 case Instruction::PtrToInt: // always treated as unsigned 624 // Is it a null pointer value? 625 if (V->isNullValue()) 626 return ConstantInt::get(DestTy, 0); 627 // If this is a sizeof-like expression, pull out multiplications by 628 // known factors to expose them to subsequent folding. If it's an 629 // alignof-like expression, factor out known factors. 630 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 631 if (CE->getOpcode() == Instruction::GetElementPtr && 632 CE->getOperand(0)->isNullValue()) { 633 GEPOperator *GEPO = cast<GEPOperator>(CE); 634 Type *Ty = GEPO->getSourceElementType(); 635 if (CE->getNumOperands() == 2) { 636 // Handle a sizeof-like expression. 637 Constant *Idx = CE->getOperand(1); 638 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 639 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 640 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 641 DestTy, false), 642 Idx, DestTy); 643 return ConstantExpr::getMul(C, Idx); 644 } 645 } else if (CE->getNumOperands() == 3 && 646 CE->getOperand(1)->isNullValue()) { 647 // Handle an alignof-like expression. 648 if (StructType *STy = dyn_cast<StructType>(Ty)) 649 if (!STy->isPacked()) { 650 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 651 if (CI->isOne() && 652 STy->getNumElements() == 2 && 653 STy->getElementType(0)->isIntegerTy(1)) { 654 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 655 } 656 } 657 // Handle an offsetof-like expression. 658 if (Ty->isStructTy() || Ty->isArrayTy()) { 659 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 660 DestTy, false)) 661 return C; 662 } 663 } 664 } 665 // Other pointer types cannot be casted 666 return nullptr; 667 case Instruction::UIToFP: 668 case Instruction::SIToFP: 669 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 670 const APInt &api = CI->getValue(); 671 APFloat apf(DestTy->getFltSemantics(), 672 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 673 if (APFloat::opOverflow & 674 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 675 APFloat::rmNearestTiesToEven)) { 676 // Undefined behavior invoked - the destination type can't represent 677 // the input constant. 678 return UndefValue::get(DestTy); 679 } 680 return ConstantFP::get(V->getContext(), apf); 681 } 682 return nullptr; 683 case Instruction::ZExt: 684 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 685 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 686 return ConstantInt::get(V->getContext(), 687 CI->getValue().zext(BitWidth)); 688 } 689 return nullptr; 690 case Instruction::SExt: 691 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 692 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 693 return ConstantInt::get(V->getContext(), 694 CI->getValue().sext(BitWidth)); 695 } 696 return nullptr; 697 case Instruction::Trunc: { 698 if (V->getType()->isVectorTy()) 699 return nullptr; 700 701 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 702 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 703 return ConstantInt::get(V->getContext(), 704 CI->getValue().trunc(DestBitWidth)); 705 } 706 707 // The input must be a constantexpr. See if we can simplify this based on 708 // the bytes we are demanding. Only do this if the source and dest are an 709 // even multiple of a byte. 710 if ((DestBitWidth & 7) == 0 && 711 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 712 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 713 return Res; 714 715 return nullptr; 716 } 717 case Instruction::BitCast: 718 return FoldBitCast(V, DestTy); 719 case Instruction::AddrSpaceCast: 720 return nullptr; 721 } 722 } 723 724 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 725 Constant *V1, Constant *V2) { 726 // Check for i1 and vector true/false conditions. 727 if (Cond->isNullValue()) return V2; 728 if (Cond->isAllOnesValue()) return V1; 729 730 // If the condition is a vector constant, fold the result elementwise. 731 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 732 SmallVector<Constant*, 16> Result; 733 Type *Ty = IntegerType::get(CondV->getContext(), 32); 734 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 735 Constant *V; 736 Constant *V1Element = ConstantExpr::getExtractElement(V1, 737 ConstantInt::get(Ty, i)); 738 Constant *V2Element = ConstantExpr::getExtractElement(V2, 739 ConstantInt::get(Ty, i)); 740 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 741 if (V1Element == V2Element) { 742 V = V1Element; 743 } else if (isa<UndefValue>(Cond)) { 744 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 745 } else { 746 if (!isa<ConstantInt>(Cond)) break; 747 V = Cond->isNullValue() ? V2Element : V1Element; 748 } 749 Result.push_back(V); 750 } 751 752 // If we were able to build the vector, return it. 753 if (Result.size() == V1->getType()->getVectorNumElements()) 754 return ConstantVector::get(Result); 755 } 756 757 if (isa<UndefValue>(Cond)) { 758 if (isa<UndefValue>(V1)) return V1; 759 return V2; 760 } 761 if (isa<UndefValue>(V1)) return V2; 762 if (isa<UndefValue>(V2)) return V1; 763 if (V1 == V2) return V1; 764 765 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 766 if (TrueVal->getOpcode() == Instruction::Select) 767 if (TrueVal->getOperand(0) == Cond) 768 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 769 } 770 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 771 if (FalseVal->getOpcode() == Instruction::Select) 772 if (FalseVal->getOperand(0) == Cond) 773 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 774 } 775 776 return nullptr; 777 } 778 779 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 780 Constant *Idx) { 781 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 782 return UndefValue::get(Val->getType()->getVectorElementType()); 783 if (Val->isNullValue()) // ee(zero, x) -> zero 784 return Constant::getNullValue(Val->getType()->getVectorElementType()); 785 // ee({w,x,y,z}, undef) -> undef 786 if (isa<UndefValue>(Idx)) 787 return UndefValue::get(Val->getType()->getVectorElementType()); 788 789 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 790 // ee({w,x,y,z}, wrong_value) -> undef 791 if (CIdx->uge(Val->getType()->getVectorNumElements())) 792 return UndefValue::get(Val->getType()->getVectorElementType()); 793 return Val->getAggregateElement(CIdx->getZExtValue()); 794 } 795 return nullptr; 796 } 797 798 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 799 Constant *Elt, 800 Constant *Idx) { 801 if (isa<UndefValue>(Idx)) 802 return UndefValue::get(Val->getType()); 803 804 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 805 if (!CIdx) return nullptr; 806 807 unsigned NumElts = Val->getType()->getVectorNumElements(); 808 if (CIdx->uge(NumElts)) 809 return UndefValue::get(Val->getType()); 810 811 SmallVector<Constant*, 16> Result; 812 Result.reserve(NumElts); 813 auto *Ty = Type::getInt32Ty(Val->getContext()); 814 uint64_t IdxVal = CIdx->getZExtValue(); 815 for (unsigned i = 0; i != NumElts; ++i) { 816 if (i == IdxVal) { 817 Result.push_back(Elt); 818 continue; 819 } 820 821 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 822 Result.push_back(C); 823 } 824 825 return ConstantVector::get(Result); 826 } 827 828 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 829 Constant *V2, 830 Constant *Mask) { 831 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 832 Type *EltTy = V1->getType()->getVectorElementType(); 833 834 // Undefined shuffle mask -> undefined value. 835 if (isa<UndefValue>(Mask)) 836 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 837 838 // Don't break the bitcode reader hack. 839 if (isa<ConstantExpr>(Mask)) return nullptr; 840 841 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 842 843 // Loop over the shuffle mask, evaluating each element. 844 SmallVector<Constant*, 32> Result; 845 for (unsigned i = 0; i != MaskNumElts; ++i) { 846 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 847 if (Elt == -1) { 848 Result.push_back(UndefValue::get(EltTy)); 849 continue; 850 } 851 Constant *InElt; 852 if (unsigned(Elt) >= SrcNumElts*2) 853 InElt = UndefValue::get(EltTy); 854 else if (unsigned(Elt) >= SrcNumElts) { 855 Type *Ty = IntegerType::get(V2->getContext(), 32); 856 InElt = 857 ConstantExpr::getExtractElement(V2, 858 ConstantInt::get(Ty, Elt - SrcNumElts)); 859 } else { 860 Type *Ty = IntegerType::get(V1->getContext(), 32); 861 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 862 } 863 Result.push_back(InElt); 864 } 865 866 return ConstantVector::get(Result); 867 } 868 869 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 870 ArrayRef<unsigned> Idxs) { 871 // Base case: no indices, so return the entire value. 872 if (Idxs.empty()) 873 return Agg; 874 875 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 876 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 877 878 return nullptr; 879 } 880 881 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 882 Constant *Val, 883 ArrayRef<unsigned> Idxs) { 884 // Base case: no indices, so replace the entire value. 885 if (Idxs.empty()) 886 return Val; 887 888 unsigned NumElts; 889 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 890 NumElts = ST->getNumElements(); 891 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 892 NumElts = AT->getNumElements(); 893 else 894 NumElts = Agg->getType()->getVectorNumElements(); 895 896 SmallVector<Constant*, 32> Result; 897 for (unsigned i = 0; i != NumElts; ++i) { 898 Constant *C = Agg->getAggregateElement(i); 899 if (!C) return nullptr; 900 901 if (Idxs[0] == i) 902 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 903 904 Result.push_back(C); 905 } 906 907 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 908 return ConstantStruct::get(ST, Result); 909 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 910 return ConstantArray::get(AT, Result); 911 return ConstantVector::get(Result); 912 } 913 914 915 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 916 Constant *C1, Constant *C2) { 917 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 918 919 // Handle UndefValue up front. 920 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 921 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 922 case Instruction::Xor: 923 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 924 // Handle undef ^ undef -> 0 special case. This is a common 925 // idiom (misuse). 926 return Constant::getNullValue(C1->getType()); 927 // Fallthrough 928 case Instruction::Add: 929 case Instruction::Sub: 930 return UndefValue::get(C1->getType()); 931 case Instruction::And: 932 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 933 return C1; 934 return Constant::getNullValue(C1->getType()); // undef & X -> 0 935 case Instruction::Mul: { 936 // undef * undef -> undef 937 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 938 return C1; 939 const APInt *CV; 940 // X * undef -> undef if X is odd 941 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 942 if ((*CV)[0]) 943 return UndefValue::get(C1->getType()); 944 945 // X * undef -> 0 otherwise 946 return Constant::getNullValue(C1->getType()); 947 } 948 case Instruction::SDiv: 949 case Instruction::UDiv: 950 // X / undef -> undef 951 if (isa<UndefValue>(C2)) 952 return C2; 953 // undef / 0 -> undef 954 // undef / 1 -> undef 955 if (match(C2, m_Zero()) || match(C2, m_One())) 956 return C1; 957 // undef / X -> 0 otherwise 958 return Constant::getNullValue(C1->getType()); 959 case Instruction::URem: 960 case Instruction::SRem: 961 // X % undef -> undef 962 if (match(C2, m_Undef())) 963 return C2; 964 // undef % 0 -> undef 965 if (match(C2, m_Zero())) 966 return C1; 967 // undef % X -> 0 otherwise 968 return Constant::getNullValue(C1->getType()); 969 case Instruction::Or: // X | undef -> -1 970 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 971 return C1; 972 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 973 case Instruction::LShr: 974 // X >>l undef -> undef 975 if (isa<UndefValue>(C2)) 976 return C2; 977 // undef >>l 0 -> undef 978 if (match(C2, m_Zero())) 979 return C1; 980 // undef >>l X -> 0 981 return Constant::getNullValue(C1->getType()); 982 case Instruction::AShr: 983 // X >>a undef -> undef 984 if (isa<UndefValue>(C2)) 985 return C2; 986 // undef >>a 0 -> undef 987 if (match(C2, m_Zero())) 988 return C1; 989 // TODO: undef >>a X -> undef if the shift is exact 990 // undef >>a X -> 0 991 return Constant::getNullValue(C1->getType()); 992 case Instruction::Shl: 993 // X << undef -> undef 994 if (isa<UndefValue>(C2)) 995 return C2; 996 // undef << 0 -> undef 997 if (match(C2, m_Zero())) 998 return C1; 999 // undef << X -> 0 1000 return Constant::getNullValue(C1->getType()); 1001 case Instruction::FAdd: 1002 case Instruction::FSub: 1003 case Instruction::FMul: 1004 case Instruction::FDiv: 1005 case Instruction::FRem: 1006 // TODO: UNDEF handling for binary float instructions. 1007 return nullptr; 1008 case Instruction::BinaryOpsEnd: 1009 llvm_unreachable("Invalid BinaryOp"); 1010 } 1011 } 1012 1013 // At this point neither constant should be an UndefValue. 1014 assert(!isa<UndefValue>(C1) && !isa<UndefValue>(C2) && 1015 "Unexpected UndefValue"); 1016 1017 // Handle simplifications when the RHS is a constant int. 1018 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1019 switch (Opcode) { 1020 case Instruction::Add: 1021 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1022 break; 1023 case Instruction::Sub: 1024 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1025 break; 1026 case Instruction::Mul: 1027 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1028 if (CI2->equalsInt(1)) 1029 return C1; // X * 1 == X 1030 break; 1031 case Instruction::UDiv: 1032 case Instruction::SDiv: 1033 if (CI2->equalsInt(1)) 1034 return C1; // X / 1 == X 1035 if (CI2->equalsInt(0)) 1036 return UndefValue::get(CI2->getType()); // X / 0 == undef 1037 break; 1038 case Instruction::URem: 1039 case Instruction::SRem: 1040 if (CI2->equalsInt(1)) 1041 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1042 if (CI2->equalsInt(0)) 1043 return UndefValue::get(CI2->getType()); // X % 0 == undef 1044 break; 1045 case Instruction::And: 1046 if (CI2->isZero()) return C2; // X & 0 == 0 1047 if (CI2->isAllOnesValue()) 1048 return C1; // X & -1 == X 1049 1050 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1051 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1052 if (CE1->getOpcode() == Instruction::ZExt) { 1053 unsigned DstWidth = CI2->getType()->getBitWidth(); 1054 unsigned SrcWidth = 1055 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1056 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1057 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1058 return C1; 1059 } 1060 1061 // If and'ing the address of a global with a constant, fold it. 1062 if (CE1->getOpcode() == Instruction::PtrToInt && 1063 isa<GlobalValue>(CE1->getOperand(0))) { 1064 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1065 1066 // Functions are at least 4-byte aligned. 1067 unsigned GVAlign = GV->getAlignment(); 1068 if (isa<Function>(GV)) 1069 GVAlign = std::max(GVAlign, 4U); 1070 1071 if (GVAlign > 1) { 1072 unsigned DstWidth = CI2->getType()->getBitWidth(); 1073 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1074 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1075 1076 // If checking bits we know are clear, return zero. 1077 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1078 return Constant::getNullValue(CI2->getType()); 1079 } 1080 } 1081 } 1082 break; 1083 case Instruction::Or: 1084 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1085 if (CI2->isAllOnesValue()) 1086 return C2; // X | -1 == -1 1087 break; 1088 case Instruction::Xor: 1089 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1090 1091 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1092 switch (CE1->getOpcode()) { 1093 default: break; 1094 case Instruction::ICmp: 1095 case Instruction::FCmp: 1096 // cmp pred ^ true -> cmp !pred 1097 assert(CI2->equalsInt(1)); 1098 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1099 pred = CmpInst::getInversePredicate(pred); 1100 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1101 CE1->getOperand(1)); 1102 } 1103 } 1104 break; 1105 case Instruction::AShr: 1106 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1107 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1108 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1109 return ConstantExpr::getLShr(C1, C2); 1110 break; 1111 } 1112 } else if (isa<ConstantInt>(C1)) { 1113 // If C1 is a ConstantInt and C2 is not, swap the operands. 1114 if (Instruction::isCommutative(Opcode)) 1115 return ConstantExpr::get(Opcode, C2, C1); 1116 } 1117 1118 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1119 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1120 const APInt &C1V = CI1->getValue(); 1121 const APInt &C2V = CI2->getValue(); 1122 switch (Opcode) { 1123 default: 1124 break; 1125 case Instruction::Add: 1126 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1127 case Instruction::Sub: 1128 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1129 case Instruction::Mul: 1130 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1131 case Instruction::UDiv: 1132 assert(!CI2->isNullValue() && "Div by zero handled above"); 1133 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1134 case Instruction::SDiv: 1135 assert(!CI2->isNullValue() && "Div by zero handled above"); 1136 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1137 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1138 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1139 case Instruction::URem: 1140 assert(!CI2->isNullValue() && "Div by zero handled above"); 1141 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1142 case Instruction::SRem: 1143 assert(!CI2->isNullValue() && "Div by zero handled above"); 1144 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1145 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1146 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1147 case Instruction::And: 1148 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1149 case Instruction::Or: 1150 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1151 case Instruction::Xor: 1152 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1153 case Instruction::Shl: 1154 if (C2V.ult(C1V.getBitWidth())) 1155 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1156 return UndefValue::get(C1->getType()); // too big shift is undef 1157 case Instruction::LShr: 1158 if (C2V.ult(C1V.getBitWidth())) 1159 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1160 return UndefValue::get(C1->getType()); // too big shift is undef 1161 case Instruction::AShr: 1162 if (C2V.ult(C1V.getBitWidth())) 1163 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1164 return UndefValue::get(C1->getType()); // too big shift is undef 1165 } 1166 } 1167 1168 switch (Opcode) { 1169 case Instruction::SDiv: 1170 case Instruction::UDiv: 1171 case Instruction::URem: 1172 case Instruction::SRem: 1173 case Instruction::LShr: 1174 case Instruction::AShr: 1175 case Instruction::Shl: 1176 if (CI1->equalsInt(0)) return C1; 1177 break; 1178 default: 1179 break; 1180 } 1181 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1182 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1183 const APFloat &C1V = CFP1->getValueAPF(); 1184 const APFloat &C2V = CFP2->getValueAPF(); 1185 APFloat C3V = C1V; // copy for modification 1186 switch (Opcode) { 1187 default: 1188 break; 1189 case Instruction::FAdd: 1190 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1191 return ConstantFP::get(C1->getContext(), C3V); 1192 case Instruction::FSub: 1193 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1194 return ConstantFP::get(C1->getContext(), C3V); 1195 case Instruction::FMul: 1196 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1197 return ConstantFP::get(C1->getContext(), C3V); 1198 case Instruction::FDiv: 1199 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1200 return ConstantFP::get(C1->getContext(), C3V); 1201 case Instruction::FRem: 1202 (void)C3V.mod(C2V); 1203 return ConstantFP::get(C1->getContext(), C3V); 1204 } 1205 } 1206 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1207 // Perform elementwise folding. 1208 SmallVector<Constant*, 16> Result; 1209 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1210 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1211 Constant *LHS = 1212 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1213 Constant *RHS = 1214 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1215 1216 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1217 } 1218 1219 return ConstantVector::get(Result); 1220 } 1221 1222 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1223 // There are many possible foldings we could do here. We should probably 1224 // at least fold add of a pointer with an integer into the appropriate 1225 // getelementptr. This will improve alias analysis a bit. 1226 1227 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1228 // (a + (b + c)). 1229 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1230 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1231 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1232 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1233 } 1234 } else if (isa<ConstantExpr>(C2)) { 1235 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1236 // other way if possible. 1237 if (Instruction::isCommutative(Opcode)) 1238 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1239 } 1240 1241 // i1 can be simplified in many cases. 1242 if (C1->getType()->isIntegerTy(1)) { 1243 switch (Opcode) { 1244 case Instruction::Add: 1245 case Instruction::Sub: 1246 return ConstantExpr::getXor(C1, C2); 1247 case Instruction::Mul: 1248 return ConstantExpr::getAnd(C1, C2); 1249 case Instruction::Shl: 1250 case Instruction::LShr: 1251 case Instruction::AShr: 1252 // We can assume that C2 == 0. If it were one the result would be 1253 // undefined because the shift value is as large as the bitwidth. 1254 return C1; 1255 case Instruction::SDiv: 1256 case Instruction::UDiv: 1257 // We can assume that C2 == 1. If it were zero the result would be 1258 // undefined through division by zero. 1259 return C1; 1260 case Instruction::URem: 1261 case Instruction::SRem: 1262 // We can assume that C2 == 1. If it were zero the result would be 1263 // undefined through division by zero. 1264 return ConstantInt::getFalse(C1->getContext()); 1265 default: 1266 break; 1267 } 1268 } 1269 1270 // We don't know how to fold this. 1271 return nullptr; 1272 } 1273 1274 /// This type is zero-sized if it's an array or structure of zero-sized types. 1275 /// The only leaf zero-sized type is an empty structure. 1276 static bool isMaybeZeroSizedType(Type *Ty) { 1277 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1278 if (STy->isOpaque()) return true; // Can't say. 1279 1280 // If all of elements have zero size, this does too. 1281 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1282 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1283 return true; 1284 1285 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1286 return isMaybeZeroSizedType(ATy->getElementType()); 1287 } 1288 return false; 1289 } 1290 1291 /// Compare the two constants as though they were getelementptr indices. 1292 /// This allows coercion of the types to be the same thing. 1293 /// 1294 /// If the two constants are the "same" (after coercion), return 0. If the 1295 /// first is less than the second, return -1, if the second is less than the 1296 /// first, return 1. If the constants are not integral, return -2. 1297 /// 1298 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1299 if (C1 == C2) return 0; 1300 1301 // Ok, we found a different index. If they are not ConstantInt, we can't do 1302 // anything with them. 1303 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1304 return -2; // don't know! 1305 1306 // We cannot compare the indices if they don't fit in an int64_t. 1307 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 || 1308 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64) 1309 return -2; // don't know! 1310 1311 // Ok, we have two differing integer indices. Sign extend them to be the same 1312 // type. 1313 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue(); 1314 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue(); 1315 1316 if (C1Val == C2Val) return 0; // They are equal 1317 1318 // If the type being indexed over is really just a zero sized type, there is 1319 // no pointer difference being made here. 1320 if (isMaybeZeroSizedType(ElTy)) 1321 return -2; // dunno. 1322 1323 // If they are really different, now that they are the same type, then we 1324 // found a difference! 1325 if (C1Val < C2Val) 1326 return -1; 1327 else 1328 return 1; 1329 } 1330 1331 /// This function determines if there is anything we can decide about the two 1332 /// constants provided. This doesn't need to handle simple things like 1333 /// ConstantFP comparisons, but should instead handle ConstantExprs. 1334 /// If we can determine that the two constants have a particular relation to 1335 /// each other, we should return the corresponding FCmpInst predicate, 1336 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1337 /// ConstantFoldCompareInstruction. 1338 /// 1339 /// To simplify this code we canonicalize the relation so that the first 1340 /// operand is always the most "complex" of the two. We consider ConstantFP 1341 /// to be the simplest, and ConstantExprs to be the most complex. 1342 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1343 assert(V1->getType() == V2->getType() && 1344 "Cannot compare values of different types!"); 1345 1346 // Handle degenerate case quickly 1347 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1348 1349 if (!isa<ConstantExpr>(V1)) { 1350 if (!isa<ConstantExpr>(V2)) { 1351 // Simple case, use the standard constant folder. 1352 ConstantInt *R = nullptr; 1353 R = dyn_cast<ConstantInt>( 1354 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1355 if (R && !R->isZero()) 1356 return FCmpInst::FCMP_OEQ; 1357 R = dyn_cast<ConstantInt>( 1358 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1359 if (R && !R->isZero()) 1360 return FCmpInst::FCMP_OLT; 1361 R = dyn_cast<ConstantInt>( 1362 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1363 if (R && !R->isZero()) 1364 return FCmpInst::FCMP_OGT; 1365 1366 // Nothing more we can do 1367 return FCmpInst::BAD_FCMP_PREDICATE; 1368 } 1369 1370 // If the first operand is simple and second is ConstantExpr, swap operands. 1371 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1372 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1373 return FCmpInst::getSwappedPredicate(SwappedRelation); 1374 } else { 1375 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1376 // constantexpr or a simple constant. 1377 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1378 switch (CE1->getOpcode()) { 1379 case Instruction::FPTrunc: 1380 case Instruction::FPExt: 1381 case Instruction::UIToFP: 1382 case Instruction::SIToFP: 1383 // We might be able to do something with these but we don't right now. 1384 break; 1385 default: 1386 break; 1387 } 1388 } 1389 // There are MANY other foldings that we could perform here. They will 1390 // probably be added on demand, as they seem needed. 1391 return FCmpInst::BAD_FCMP_PREDICATE; 1392 } 1393 1394 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1395 const GlobalValue *GV2) { 1396 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1397 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1398 return true; 1399 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1400 Type *Ty = GVar->getValueType(); 1401 // A global with opaque type might end up being zero sized. 1402 if (!Ty->isSized()) 1403 return true; 1404 // A global with an empty type might lie at the address of any other 1405 // global. 1406 if (Ty->isEmptyTy()) 1407 return true; 1408 } 1409 return false; 1410 }; 1411 // Don't try to decide equality of aliases. 1412 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1413 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1414 return ICmpInst::ICMP_NE; 1415 return ICmpInst::BAD_ICMP_PREDICATE; 1416 } 1417 1418 /// This function determines if there is anything we can decide about the two 1419 /// constants provided. This doesn't need to handle simple things like integer 1420 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1421 /// If we can determine that the two constants have a particular relation to 1422 /// each other, we should return the corresponding ICmp predicate, otherwise 1423 /// return ICmpInst::BAD_ICMP_PREDICATE. 1424 /// 1425 /// To simplify this code we canonicalize the relation so that the first 1426 /// operand is always the most "complex" of the two. We consider simple 1427 /// constants (like ConstantInt) to be the simplest, followed by 1428 /// GlobalValues, followed by ConstantExpr's (the most complex). 1429 /// 1430 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1431 bool isSigned) { 1432 assert(V1->getType() == V2->getType() && 1433 "Cannot compare different types of values!"); 1434 if (V1 == V2) return ICmpInst::ICMP_EQ; 1435 1436 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1437 !isa<BlockAddress>(V1)) { 1438 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1439 !isa<BlockAddress>(V2)) { 1440 // We distilled this down to a simple case, use the standard constant 1441 // folder. 1442 ConstantInt *R = nullptr; 1443 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1444 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1445 if (R && !R->isZero()) 1446 return pred; 1447 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1448 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1449 if (R && !R->isZero()) 1450 return pred; 1451 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1452 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1453 if (R && !R->isZero()) 1454 return pred; 1455 1456 // If we couldn't figure it out, bail. 1457 return ICmpInst::BAD_ICMP_PREDICATE; 1458 } 1459 1460 // If the first operand is simple, swap operands. 1461 ICmpInst::Predicate SwappedRelation = 1462 evaluateICmpRelation(V2, V1, isSigned); 1463 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1464 return ICmpInst::getSwappedPredicate(SwappedRelation); 1465 1466 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1467 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1468 ICmpInst::Predicate SwappedRelation = 1469 evaluateICmpRelation(V2, V1, isSigned); 1470 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1471 return ICmpInst::getSwappedPredicate(SwappedRelation); 1472 return ICmpInst::BAD_ICMP_PREDICATE; 1473 } 1474 1475 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1476 // constant (which, since the types must match, means that it's a 1477 // ConstantPointerNull). 1478 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1479 return areGlobalsPotentiallyEqual(GV, GV2); 1480 } else if (isa<BlockAddress>(V2)) { 1481 return ICmpInst::ICMP_NE; // Globals never equal labels. 1482 } else { 1483 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1484 // GlobalVals can never be null unless they have external weak linkage. 1485 // We don't try to evaluate aliases here. 1486 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1487 return ICmpInst::ICMP_NE; 1488 } 1489 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1490 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1491 ICmpInst::Predicate SwappedRelation = 1492 evaluateICmpRelation(V2, V1, isSigned); 1493 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1494 return ICmpInst::getSwappedPredicate(SwappedRelation); 1495 return ICmpInst::BAD_ICMP_PREDICATE; 1496 } 1497 1498 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1499 // constant (which, since the types must match, means that it is a 1500 // ConstantPointerNull). 1501 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1502 // Block address in another function can't equal this one, but block 1503 // addresses in the current function might be the same if blocks are 1504 // empty. 1505 if (BA2->getFunction() != BA->getFunction()) 1506 return ICmpInst::ICMP_NE; 1507 } else { 1508 // Block addresses aren't null, don't equal the address of globals. 1509 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1510 "Canonicalization guarantee!"); 1511 return ICmpInst::ICMP_NE; 1512 } 1513 } else { 1514 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1515 // constantexpr, a global, block address, or a simple constant. 1516 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1517 Constant *CE1Op0 = CE1->getOperand(0); 1518 1519 switch (CE1->getOpcode()) { 1520 case Instruction::Trunc: 1521 case Instruction::FPTrunc: 1522 case Instruction::FPExt: 1523 case Instruction::FPToUI: 1524 case Instruction::FPToSI: 1525 break; // We can't evaluate floating point casts or truncations. 1526 1527 case Instruction::UIToFP: 1528 case Instruction::SIToFP: 1529 case Instruction::BitCast: 1530 case Instruction::ZExt: 1531 case Instruction::SExt: 1532 // If the cast is not actually changing bits, and the second operand is a 1533 // null pointer, do the comparison with the pre-casted value. 1534 if (V2->isNullValue() && 1535 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1536 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1537 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1538 return evaluateICmpRelation(CE1Op0, 1539 Constant::getNullValue(CE1Op0->getType()), 1540 isSigned); 1541 } 1542 break; 1543 1544 case Instruction::GetElementPtr: { 1545 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1546 // Ok, since this is a getelementptr, we know that the constant has a 1547 // pointer type. Check the various cases. 1548 if (isa<ConstantPointerNull>(V2)) { 1549 // If we are comparing a GEP to a null pointer, check to see if the base 1550 // of the GEP equals the null pointer. 1551 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1552 if (GV->hasExternalWeakLinkage()) 1553 // Weak linkage GVals could be zero or not. We're comparing that 1554 // to null pointer so its greater-or-equal 1555 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1556 else 1557 // If its not weak linkage, the GVal must have a non-zero address 1558 // so the result is greater-than 1559 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1560 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1561 // If we are indexing from a null pointer, check to see if we have any 1562 // non-zero indices. 1563 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1564 if (!CE1->getOperand(i)->isNullValue()) 1565 // Offsetting from null, must not be equal. 1566 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1567 // Only zero indexes from null, must still be zero. 1568 return ICmpInst::ICMP_EQ; 1569 } 1570 // Otherwise, we can't really say if the first operand is null or not. 1571 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1572 if (isa<ConstantPointerNull>(CE1Op0)) { 1573 if (GV2->hasExternalWeakLinkage()) 1574 // Weak linkage GVals could be zero or not. We're comparing it to 1575 // a null pointer, so its less-or-equal 1576 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1577 else 1578 // If its not weak linkage, the GVal must have a non-zero address 1579 // so the result is less-than 1580 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1581 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1582 if (GV == GV2) { 1583 // If this is a getelementptr of the same global, then it must be 1584 // different. Because the types must match, the getelementptr could 1585 // only have at most one index, and because we fold getelementptr's 1586 // with a single zero index, it must be nonzero. 1587 assert(CE1->getNumOperands() == 2 && 1588 !CE1->getOperand(1)->isNullValue() && 1589 "Surprising getelementptr!"); 1590 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1591 } else { 1592 if (CE1GEP->hasAllZeroIndices()) 1593 return areGlobalsPotentiallyEqual(GV, GV2); 1594 return ICmpInst::BAD_ICMP_PREDICATE; 1595 } 1596 } 1597 } else { 1598 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1599 Constant *CE2Op0 = CE2->getOperand(0); 1600 1601 // There are MANY other foldings that we could perform here. They will 1602 // probably be added on demand, as they seem needed. 1603 switch (CE2->getOpcode()) { 1604 default: break; 1605 case Instruction::GetElementPtr: 1606 // By far the most common case to handle is when the base pointers are 1607 // obviously to the same global. 1608 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1609 // Don't know relative ordering, but check for inequality. 1610 if (CE1Op0 != CE2Op0) { 1611 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1612 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1613 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1614 cast<GlobalValue>(CE2Op0)); 1615 return ICmpInst::BAD_ICMP_PREDICATE; 1616 } 1617 // Ok, we know that both getelementptr instructions are based on the 1618 // same global. From this, we can precisely determine the relative 1619 // ordering of the resultant pointers. 1620 unsigned i = 1; 1621 1622 // The logic below assumes that the result of the comparison 1623 // can be determined by finding the first index that differs. 1624 // This doesn't work if there is over-indexing in any 1625 // subsequent indices, so check for that case first. 1626 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1627 !CE2->isGEPWithNoNotionalOverIndexing()) 1628 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1629 1630 // Compare all of the operands the GEP's have in common. 1631 gep_type_iterator GTI = gep_type_begin(CE1); 1632 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1633 ++i, ++GTI) 1634 switch (IdxCompare(CE1->getOperand(i), 1635 CE2->getOperand(i), GTI.getIndexedType())) { 1636 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1637 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1638 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1639 } 1640 1641 // Ok, we ran out of things they have in common. If any leftovers 1642 // are non-zero then we have a difference, otherwise we are equal. 1643 for (; i < CE1->getNumOperands(); ++i) 1644 if (!CE1->getOperand(i)->isNullValue()) { 1645 if (isa<ConstantInt>(CE1->getOperand(i))) 1646 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1647 else 1648 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1649 } 1650 1651 for (; i < CE2->getNumOperands(); ++i) 1652 if (!CE2->getOperand(i)->isNullValue()) { 1653 if (isa<ConstantInt>(CE2->getOperand(i))) 1654 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1655 else 1656 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1657 } 1658 return ICmpInst::ICMP_EQ; 1659 } 1660 } 1661 } 1662 } 1663 default: 1664 break; 1665 } 1666 } 1667 1668 return ICmpInst::BAD_ICMP_PREDICATE; 1669 } 1670 1671 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1672 Constant *C1, Constant *C2) { 1673 Type *ResultTy; 1674 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1675 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1676 VT->getNumElements()); 1677 else 1678 ResultTy = Type::getInt1Ty(C1->getContext()); 1679 1680 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1681 if (pred == FCmpInst::FCMP_FALSE) 1682 return Constant::getNullValue(ResultTy); 1683 1684 if (pred == FCmpInst::FCMP_TRUE) 1685 return Constant::getAllOnesValue(ResultTy); 1686 1687 // Handle some degenerate cases first 1688 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1689 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1690 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1691 // For EQ and NE, we can always pick a value for the undef to make the 1692 // predicate pass or fail, so we can return undef. 1693 // Also, if both operands are undef, we can return undef for int comparison. 1694 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1695 return UndefValue::get(ResultTy); 1696 1697 // Otherwise, for integer compare, pick the same value as the non-undef 1698 // operand, and fold it to true or false. 1699 if (isIntegerPredicate) 1700 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1701 1702 // Choosing NaN for the undef will always make unordered comparison succeed 1703 // and ordered comparison fails. 1704 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1705 } 1706 1707 // icmp eq/ne(null,GV) -> false/true 1708 if (C1->isNullValue()) { 1709 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1710 // Don't try to evaluate aliases. External weak GV can be null. 1711 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1712 if (pred == ICmpInst::ICMP_EQ) 1713 return ConstantInt::getFalse(C1->getContext()); 1714 else if (pred == ICmpInst::ICMP_NE) 1715 return ConstantInt::getTrue(C1->getContext()); 1716 } 1717 // icmp eq/ne(GV,null) -> false/true 1718 } else if (C2->isNullValue()) { 1719 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1720 // Don't try to evaluate aliases. External weak GV can be null. 1721 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1722 if (pred == ICmpInst::ICMP_EQ) 1723 return ConstantInt::getFalse(C1->getContext()); 1724 else if (pred == ICmpInst::ICMP_NE) 1725 return ConstantInt::getTrue(C1->getContext()); 1726 } 1727 } 1728 1729 // If the comparison is a comparison between two i1's, simplify it. 1730 if (C1->getType()->isIntegerTy(1)) { 1731 switch(pred) { 1732 case ICmpInst::ICMP_EQ: 1733 if (isa<ConstantInt>(C2)) 1734 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1735 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1736 case ICmpInst::ICMP_NE: 1737 return ConstantExpr::getXor(C1, C2); 1738 default: 1739 break; 1740 } 1741 } 1742 1743 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1744 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1745 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1746 switch (pred) { 1747 default: llvm_unreachable("Invalid ICmp Predicate"); 1748 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1749 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1750 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1751 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1752 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1753 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1754 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1755 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1756 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1757 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1758 } 1759 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1760 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1761 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1762 APFloat::cmpResult R = C1V.compare(C2V); 1763 switch (pred) { 1764 default: llvm_unreachable("Invalid FCmp Predicate"); 1765 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1766 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1767 case FCmpInst::FCMP_UNO: 1768 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1769 case FCmpInst::FCMP_ORD: 1770 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1771 case FCmpInst::FCMP_UEQ: 1772 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1773 R==APFloat::cmpEqual); 1774 case FCmpInst::FCMP_OEQ: 1775 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1776 case FCmpInst::FCMP_UNE: 1777 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1778 case FCmpInst::FCMP_ONE: 1779 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1780 R==APFloat::cmpGreaterThan); 1781 case FCmpInst::FCMP_ULT: 1782 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1783 R==APFloat::cmpLessThan); 1784 case FCmpInst::FCMP_OLT: 1785 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1786 case FCmpInst::FCMP_UGT: 1787 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1788 R==APFloat::cmpGreaterThan); 1789 case FCmpInst::FCMP_OGT: 1790 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1791 case FCmpInst::FCMP_ULE: 1792 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1793 case FCmpInst::FCMP_OLE: 1794 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1795 R==APFloat::cmpEqual); 1796 case FCmpInst::FCMP_UGE: 1797 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1798 case FCmpInst::FCMP_OGE: 1799 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1800 R==APFloat::cmpEqual); 1801 } 1802 } else if (C1->getType()->isVectorTy()) { 1803 // If we can constant fold the comparison of each element, constant fold 1804 // the whole vector comparison. 1805 SmallVector<Constant*, 4> ResElts; 1806 Type *Ty = IntegerType::get(C1->getContext(), 32); 1807 // Compare the elements, producing an i1 result or constant expr. 1808 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1809 Constant *C1E = 1810 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1811 Constant *C2E = 1812 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1813 1814 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1815 } 1816 1817 return ConstantVector::get(ResElts); 1818 } 1819 1820 if (C1->getType()->isFloatingPointTy() && 1821 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1822 // infinite recursive loop 1823 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1824 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1825 switch (evaluateFCmpRelation(C1, C2)) { 1826 default: llvm_unreachable("Unknown relation!"); 1827 case FCmpInst::FCMP_UNO: 1828 case FCmpInst::FCMP_ORD: 1829 case FCmpInst::FCMP_UEQ: 1830 case FCmpInst::FCMP_UNE: 1831 case FCmpInst::FCMP_ULT: 1832 case FCmpInst::FCMP_UGT: 1833 case FCmpInst::FCMP_ULE: 1834 case FCmpInst::FCMP_UGE: 1835 case FCmpInst::FCMP_TRUE: 1836 case FCmpInst::FCMP_FALSE: 1837 case FCmpInst::BAD_FCMP_PREDICATE: 1838 break; // Couldn't determine anything about these constants. 1839 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1840 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1841 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1842 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1843 break; 1844 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1845 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1846 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1847 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1848 break; 1849 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1850 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1851 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1852 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1853 break; 1854 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1855 // We can only partially decide this relation. 1856 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1857 Result = 0; 1858 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1859 Result = 1; 1860 break; 1861 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1862 // We can only partially decide this relation. 1863 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1864 Result = 0; 1865 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1866 Result = 1; 1867 break; 1868 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1869 // We can only partially decide this relation. 1870 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1871 Result = 0; 1872 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1873 Result = 1; 1874 break; 1875 } 1876 1877 // If we evaluated the result, return it now. 1878 if (Result != -1) 1879 return ConstantInt::get(ResultTy, Result); 1880 1881 } else { 1882 // Evaluate the relation between the two constants, per the predicate. 1883 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1884 switch (evaluateICmpRelation(C1, C2, 1885 CmpInst::isSigned((CmpInst::Predicate)pred))) { 1886 default: llvm_unreachable("Unknown relational!"); 1887 case ICmpInst::BAD_ICMP_PREDICATE: 1888 break; // Couldn't determine anything about these constants. 1889 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1890 // If we know the constants are equal, we can decide the result of this 1891 // computation precisely. 1892 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1893 break; 1894 case ICmpInst::ICMP_ULT: 1895 switch (pred) { 1896 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1897 Result = 1; break; 1898 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1899 Result = 0; break; 1900 } 1901 break; 1902 case ICmpInst::ICMP_SLT: 1903 switch (pred) { 1904 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1905 Result = 1; break; 1906 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1907 Result = 0; break; 1908 } 1909 break; 1910 case ICmpInst::ICMP_UGT: 1911 switch (pred) { 1912 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1913 Result = 1; break; 1914 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1915 Result = 0; break; 1916 } 1917 break; 1918 case ICmpInst::ICMP_SGT: 1919 switch (pred) { 1920 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1921 Result = 1; break; 1922 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1923 Result = 0; break; 1924 } 1925 break; 1926 case ICmpInst::ICMP_ULE: 1927 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1928 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1929 break; 1930 case ICmpInst::ICMP_SLE: 1931 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1932 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1933 break; 1934 case ICmpInst::ICMP_UGE: 1935 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1936 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1937 break; 1938 case ICmpInst::ICMP_SGE: 1939 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1940 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1941 break; 1942 case ICmpInst::ICMP_NE: 1943 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1944 if (pred == ICmpInst::ICMP_NE) Result = 1; 1945 break; 1946 } 1947 1948 // If we evaluated the result, return it now. 1949 if (Result != -1) 1950 return ConstantInt::get(ResultTy, Result); 1951 1952 // If the right hand side is a bitcast, try using its inverse to simplify 1953 // it by moving it to the left hand side. We can't do this if it would turn 1954 // a vector compare into a scalar compare or visa versa. 1955 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1956 Constant *CE2Op0 = CE2->getOperand(0); 1957 if (CE2->getOpcode() == Instruction::BitCast && 1958 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1959 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1960 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1961 } 1962 } 1963 1964 // If the left hand side is an extension, try eliminating it. 1965 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1966 if ((CE1->getOpcode() == Instruction::SExt && 1967 ICmpInst::isSigned((ICmpInst::Predicate)pred)) || 1968 (CE1->getOpcode() == Instruction::ZExt && 1969 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){ 1970 Constant *CE1Op0 = CE1->getOperand(0); 1971 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1972 if (CE1Inverse == CE1Op0) { 1973 // Check whether we can safely truncate the right hand side. 1974 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1975 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1976 C2->getType()) == C2) 1977 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1978 } 1979 } 1980 } 1981 1982 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1983 (C1->isNullValue() && !C2->isNullValue())) { 1984 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1985 // other way if possible. 1986 // Also, if C1 is null and C2 isn't, flip them around. 1987 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1988 return ConstantExpr::getICmp(pred, C2, C1); 1989 } 1990 } 1991 return nullptr; 1992 } 1993 1994 /// Test whether the given sequence of *normalized* indices is "inbounds". 1995 template<typename IndexTy> 1996 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1997 // No indices means nothing that could be out of bounds. 1998 if (Idxs.empty()) return true; 1999 2000 // If the first index is zero, it's in bounds. 2001 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2002 2003 // If the first index is one and all the rest are zero, it's in bounds, 2004 // by the one-past-the-end rule. 2005 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2006 return false; 2007 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 2008 if (!cast<Constant>(Idxs[i])->isNullValue()) 2009 return false; 2010 return true; 2011 } 2012 2013 /// Test whether a given ConstantInt is in-range for a SequentialType. 2014 static bool isIndexInRangeOfSequentialType(SequentialType *STy, 2015 const ConstantInt *CI) { 2016 // And indices are valid when indexing along a pointer 2017 if (isa<PointerType>(STy)) 2018 return true; 2019 2020 uint64_t NumElements = 0; 2021 // Determine the number of elements in our sequential type. 2022 if (auto *ATy = dyn_cast<ArrayType>(STy)) 2023 NumElements = ATy->getNumElements(); 2024 else if (auto *VTy = dyn_cast<VectorType>(STy)) 2025 NumElements = VTy->getNumElements(); 2026 2027 assert((isa<ArrayType>(STy) || NumElements > 0) && 2028 "didn't expect non-array type to have zero elements!"); 2029 2030 // We cannot bounds check the index if it doesn't fit in an int64_t. 2031 if (CI->getValue().getActiveBits() > 64) 2032 return false; 2033 2034 // A negative index or an index past the end of our sequential type is 2035 // considered out-of-range. 2036 int64_t IndexVal = CI->getSExtValue(); 2037 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2038 return false; 2039 2040 // Otherwise, it is in-range. 2041 return true; 2042 } 2043 2044 template<typename IndexTy> 2045 static Constant *ConstantFoldGetElementPtrImpl(Type *PointeeTy, Constant *C, 2046 bool inBounds, 2047 ArrayRef<IndexTy> Idxs) { 2048 if (Idxs.empty()) return C; 2049 Constant *Idx0 = cast<Constant>(Idxs[0]); 2050 if ((Idxs.size() == 1 && Idx0->isNullValue())) 2051 return C; 2052 2053 if (isa<UndefValue>(C)) { 2054 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2055 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2056 assert(Ty && "Invalid indices for GEP!"); 2057 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2058 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2059 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 2060 return UndefValue::get(GEPTy); 2061 } 2062 2063 if (C->isNullValue()) { 2064 bool isNull = true; 2065 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2066 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2067 isNull = false; 2068 break; 2069 } 2070 if (isNull) { 2071 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2072 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2073 2074 assert(Ty && "Invalid indices for GEP!"); 2075 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2076 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2077 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 2078 return Constant::getNullValue(GEPTy); 2079 } 2080 } 2081 2082 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2083 // Combine Indices - If the source pointer to this getelementptr instruction 2084 // is a getelementptr instruction, combine the indices of the two 2085 // getelementptr instructions into a single instruction. 2086 // 2087 if (CE->getOpcode() == Instruction::GetElementPtr) { 2088 Type *LastTy = nullptr; 2089 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2090 I != E; ++I) 2091 LastTy = *I; 2092 2093 // We cannot combine indices if doing so would take us outside of an 2094 // array or vector. Doing otherwise could trick us if we evaluated such a 2095 // GEP as part of a load. 2096 // 2097 // e.g. Consider if the original GEP was: 2098 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2099 // i32 0, i32 0, i64 0) 2100 // 2101 // If we then tried to offset it by '8' to get to the third element, 2102 // an i8, we should *not* get: 2103 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2104 // i32 0, i32 0, i64 8) 2105 // 2106 // This GEP tries to index array element '8 which runs out-of-bounds. 2107 // Subsequent evaluation would get confused and produce erroneous results. 2108 // 2109 // The following prohibits such a GEP from being formed by checking to see 2110 // if the index is in-range with respect to an array or vector. 2111 bool PerformFold = false; 2112 if (Idx0->isNullValue()) 2113 PerformFold = true; 2114 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy)) 2115 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2116 PerformFold = isIndexInRangeOfSequentialType(STy, CI); 2117 2118 if (PerformFold) { 2119 SmallVector<Value*, 16> NewIndices; 2120 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2121 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2122 2123 // Add the last index of the source with the first index of the new GEP. 2124 // Make sure to handle the case when they are actually different types. 2125 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2126 // Otherwise it must be an array. 2127 if (!Idx0->isNullValue()) { 2128 Type *IdxTy = Combined->getType(); 2129 if (IdxTy != Idx0->getType()) { 2130 unsigned CommonExtendedWidth = 2131 std::max(IdxTy->getIntegerBitWidth(), 2132 Idx0->getType()->getIntegerBitWidth()); 2133 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2134 2135 Type *CommonTy = 2136 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2137 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2138 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2139 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2140 } else { 2141 Combined = 2142 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2143 } 2144 } 2145 2146 NewIndices.push_back(Combined); 2147 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2148 return ConstantExpr::getGetElementPtr( 2149 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2150 NewIndices, inBounds && cast<GEPOperator>(CE)->isInBounds()); 2151 } 2152 } 2153 2154 // Attempt to fold casts to the same type away. For example, folding: 2155 // 2156 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2157 // i64 0, i64 0) 2158 // into: 2159 // 2160 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2161 // 2162 // Don't fold if the cast is changing address spaces. 2163 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2164 PointerType *SrcPtrTy = 2165 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2166 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2167 if (SrcPtrTy && DstPtrTy) { 2168 ArrayType *SrcArrayTy = 2169 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2170 ArrayType *DstArrayTy = 2171 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2172 if (SrcArrayTy && DstArrayTy 2173 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2174 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2175 return ConstantExpr::getGetElementPtr( 2176 SrcArrayTy, (Constant *)CE->getOperand(0), Idxs, inBounds); 2177 } 2178 } 2179 } 2180 2181 // Check to see if any array indices are not within the corresponding 2182 // notional array or vector bounds. If so, try to determine if they can be 2183 // factored out into preceding dimensions. 2184 SmallVector<Constant *, 8> NewIdxs; 2185 Type *Ty = PointeeTy; 2186 Type *Prev = C->getType(); 2187 bool Unknown = !isa<ConstantInt>(Idxs[0]); 2188 for (unsigned i = 1, e = Idxs.size(); i != e; 2189 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2190 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2191 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) 2192 if (CI->getSExtValue() > 0 && 2193 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) { 2194 if (isa<SequentialType>(Prev)) { 2195 // It's out of range, but we can factor it into the prior 2196 // dimension. 2197 NewIdxs.resize(Idxs.size()); 2198 uint64_t NumElements = 0; 2199 if (auto *ATy = dyn_cast<ArrayType>(Ty)) 2200 NumElements = ATy->getNumElements(); 2201 else 2202 NumElements = cast<VectorType>(Ty)->getNumElements(); 2203 2204 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2205 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2206 2207 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2208 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2209 2210 unsigned CommonExtendedWidth = 2211 std::max(PrevIdx->getType()->getIntegerBitWidth(), 2212 Div->getType()->getIntegerBitWidth()); 2213 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2214 2215 // Before adding, extend both operands to i64 to avoid 2216 // overflow trouble. 2217 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth)) 2218 PrevIdx = ConstantExpr::getSExt( 2219 PrevIdx, 2220 Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2221 if (!Div->getType()->isIntegerTy(CommonExtendedWidth)) 2222 Div = ConstantExpr::getSExt( 2223 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2224 2225 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2226 } else { 2227 // It's out of range, but the prior dimension is a struct 2228 // so we can't do anything about it. 2229 Unknown = true; 2230 } 2231 } 2232 } else { 2233 // We don't know if it's in range or not. 2234 Unknown = true; 2235 } 2236 } 2237 2238 // If we did any factoring, start over with the adjusted indices. 2239 if (!NewIdxs.empty()) { 2240 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2241 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2242 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, inBounds); 2243 } 2244 2245 // If all indices are known integers and normalized, we can do a simple 2246 // check for the "inbounds" property. 2247 if (!Unknown && !inBounds) 2248 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2249 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2250 return ConstantExpr::getInBoundsGetElementPtr(PointeeTy, C, Idxs); 2251 2252 return nullptr; 2253 } 2254 2255 Constant *llvm::ConstantFoldGetElementPtr(Type *Ty, Constant *C, 2256 bool inBounds, 2257 ArrayRef<Constant *> Idxs) { 2258 return ConstantFoldGetElementPtrImpl(Ty, C, inBounds, Idxs); 2259 } 2260 2261 Constant *llvm::ConstantFoldGetElementPtr(Type *Ty, Constant *C, 2262 bool inBounds, 2263 ArrayRef<Value *> Idxs) { 2264 return ConstantFoldGetElementPtrImpl(Ty, C, inBounds, Idxs); 2265 } 2266