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/Compiler.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/ManagedStatic.h" 34 #include "llvm/Support/MathExtras.h" 35 #include <limits> 36 using namespace llvm; 37 using namespace llvm::PatternMatch; 38 39 //===----------------------------------------------------------------------===// 40 // ConstantFold*Instruction Implementations 41 //===----------------------------------------------------------------------===// 42 43 /// BitCastConstantVector - Convert the specified vector Constant node to the 44 /// specified vector type. At this point, we know that the elements of the 45 /// input vector constant are all simple integer or FP values. 46 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 47 48 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 49 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 50 51 // If this cast changes element count then we can't handle it here: 52 // doing so requires endianness information. This should be handled by 53 // Analysis/ConstantFolding.cpp 54 unsigned NumElts = DstTy->getNumElements(); 55 if (NumElts != CV->getType()->getVectorNumElements()) 56 return nullptr; 57 58 Type *DstEltTy = DstTy->getElementType(); 59 60 SmallVector<Constant*, 16> Result; 61 Type *Ty = IntegerType::get(CV->getContext(), 32); 62 for (unsigned i = 0; i != NumElts; ++i) { 63 Constant *C = 64 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 65 C = ConstantExpr::getBitCast(C, DstEltTy); 66 Result.push_back(C); 67 } 68 69 return ConstantVector::get(Result); 70 } 71 72 /// This function determines which opcode to use to fold two constant cast 73 /// expressions together. It uses CastInst::isEliminableCastPair to determine 74 /// the opcode. Consequently its just a wrapper around that function. 75 /// @brief Determine if it is valid to fold a cast of a cast 76 static unsigned 77 foldConstantCastPair( 78 unsigned opc, ///< opcode of the second cast constant expression 79 ConstantExpr *Op, ///< the first cast constant expression 80 Type *DstTy ///< destination type of the first cast 81 ) { 82 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 83 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 84 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 85 86 // The types and opcodes for the two Cast constant expressions 87 Type *SrcTy = Op->getOperand(0)->getType(); 88 Type *MidTy = Op->getType(); 89 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 90 Instruction::CastOps secondOp = Instruction::CastOps(opc); 91 92 // Assume that pointers are never more than 64 bits wide, and only use this 93 // for the middle type. Otherwise we could end up folding away illegal 94 // bitcasts between address spaces with different sizes. 95 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 96 97 // Let CastInst::isEliminableCastPair do the heavy lifting. 98 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 99 nullptr, FakeIntPtrTy, nullptr); 100 } 101 102 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 103 Type *SrcTy = V->getType(); 104 if (SrcTy == DestTy) 105 return V; // no-op cast 106 107 // Check to see if we are casting a pointer to an aggregate to a pointer to 108 // the first element. If so, return the appropriate GEP instruction. 109 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 110 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 111 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 112 && PTy->getElementType()->isSized()) { 113 SmallVector<Value*, 8> IdxList; 114 Value *Zero = 115 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 116 IdxList.push_back(Zero); 117 Type *ElTy = PTy->getElementType(); 118 while (ElTy != DPTy->getElementType()) { 119 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 120 if (STy->getNumElements() == 0) break; 121 ElTy = STy->getElementType(0); 122 IdxList.push_back(Zero); 123 } else if (SequentialType *STy = 124 dyn_cast<SequentialType>(ElTy)) { 125 if (ElTy->isPointerTy()) break; // Can't index into pointers! 126 ElTy = STy->getElementType(); 127 IdxList.push_back(Zero); 128 } else { 129 break; 130 } 131 } 132 133 if (ElTy == DPTy->getElementType()) 134 // This GEP is inbounds because all indices are zero. 135 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), 136 V, IdxList); 137 } 138 139 // Handle casts from one vector constant to another. We know that the src 140 // and dest type have the same size (otherwise its an illegal cast). 141 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 142 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 143 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 144 "Not cast between same sized vectors!"); 145 SrcTy = nullptr; 146 // First, check for null. Undef is already handled. 147 if (isa<ConstantAggregateZero>(V)) 148 return Constant::getNullValue(DestTy); 149 150 // Handle ConstantVector and ConstantAggregateVector. 151 return BitCastConstantVector(V, DestPTy); 152 } 153 154 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 155 // This allows for other simplifications (although some of them 156 // can only be handled by Analysis/ConstantFolding.cpp). 157 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 158 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 159 } 160 161 // Finally, implement bitcast folding now. The code below doesn't handle 162 // bitcast right. 163 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 164 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 165 166 // Handle integral constant input. 167 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 168 if (DestTy->isIntegerTy()) 169 // Integral -> Integral. This is a no-op because the bit widths must 170 // be the same. Consequently, we just fold to V. 171 return V; 172 173 // See note below regarding the PPC_FP128 restriction. 174 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 175 return ConstantFP::get(DestTy->getContext(), 176 APFloat(DestTy->getFltSemantics(), 177 CI->getValue())); 178 179 // Otherwise, can't fold this (vector?) 180 return nullptr; 181 } 182 183 // Handle ConstantFP input: FP -> Integral. 184 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 185 // PPC_FP128 is really the sum of two consecutive doubles, where the first 186 // double is always stored first in memory, regardless of the target 187 // endianness. The memory layout of i128, however, depends on the target 188 // endianness, and so we can't fold this without target endianness 189 // information. This should instead be handled by 190 // Analysis/ConstantFolding.cpp 191 if (FP->getType()->isPPC_FP128Ty()) 192 return nullptr; 193 194 return ConstantInt::get(FP->getContext(), 195 FP->getValueAPF().bitcastToAPInt()); 196 } 197 198 return nullptr; 199 } 200 201 202 /// ExtractConstantBytes - V is an integer constant which only has a subset of 203 /// its bytes used. The bytes used are indicated by ByteStart (which is the 204 /// first byte used, counting from the least significant byte) and ByteSize, 205 /// which is the number of bytes used. 206 /// 207 /// This function analyzes the specified constant to see if the specified byte 208 /// range can be returned as a simplified constant. If so, the constant is 209 /// returned, otherwise null is returned. 210 /// 211 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 212 unsigned ByteSize) { 213 assert(C->getType()->isIntegerTy() && 214 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 215 "Non-byte sized integer input"); 216 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 217 assert(ByteSize && "Must be accessing some piece"); 218 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 219 assert(ByteSize != CSize && "Should not extract everything"); 220 221 // Constant Integers are simple. 222 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 223 APInt V = CI->getValue(); 224 if (ByteStart) 225 V = V.lshr(ByteStart*8); 226 V = V.trunc(ByteSize*8); 227 return ConstantInt::get(CI->getContext(), V); 228 } 229 230 // In the input is a constant expr, we might be able to recursively simplify. 231 // If not, we definitely can't do anything. 232 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 233 if (!CE) return nullptr; 234 235 switch (CE->getOpcode()) { 236 default: return nullptr; 237 case Instruction::Or: { 238 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 239 if (!RHS) 240 return nullptr; 241 242 // X | -1 -> -1. 243 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 244 if (RHSC->isAllOnesValue()) 245 return RHSC; 246 247 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 248 if (!LHS) 249 return nullptr; 250 return ConstantExpr::getOr(LHS, RHS); 251 } 252 case Instruction::And: { 253 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 254 if (!RHS) 255 return nullptr; 256 257 // X & 0 -> 0. 258 if (RHS->isNullValue()) 259 return RHS; 260 261 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 262 if (!LHS) 263 return nullptr; 264 return ConstantExpr::getAnd(LHS, RHS); 265 } 266 case Instruction::LShr: { 267 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 268 if (!Amt) 269 return nullptr; 270 unsigned ShAmt = Amt->getZExtValue(); 271 // Cannot analyze non-byte shifts. 272 if ((ShAmt & 7) != 0) 273 return nullptr; 274 ShAmt >>= 3; 275 276 // If the extract is known to be all zeros, return zero. 277 if (ByteStart >= CSize-ShAmt) 278 return Constant::getNullValue(IntegerType::get(CE->getContext(), 279 ByteSize*8)); 280 // If the extract is known to be fully in the input, extract it. 281 if (ByteStart+ByteSize+ShAmt <= CSize) 282 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 283 284 // TODO: Handle the 'partially zero' case. 285 return nullptr; 286 } 287 288 case Instruction::Shl: { 289 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 290 if (!Amt) 291 return nullptr; 292 unsigned ShAmt = Amt->getZExtValue(); 293 // Cannot analyze non-byte shifts. 294 if ((ShAmt & 7) != 0) 295 return nullptr; 296 ShAmt >>= 3; 297 298 // If the extract is known to be all zeros, return zero. 299 if (ByteStart+ByteSize <= ShAmt) 300 return Constant::getNullValue(IntegerType::get(CE->getContext(), 301 ByteSize*8)); 302 // If the extract is known to be fully in the input, extract it. 303 if (ByteStart >= ShAmt) 304 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 305 306 // TODO: Handle the 'partially zero' case. 307 return nullptr; 308 } 309 310 case Instruction::ZExt: { 311 unsigned SrcBitSize = 312 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 313 314 // If extracting something that is completely zero, return 0. 315 if (ByteStart*8 >= SrcBitSize) 316 return Constant::getNullValue(IntegerType::get(CE->getContext(), 317 ByteSize*8)); 318 319 // If exactly extracting the input, return it. 320 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 321 return CE->getOperand(0); 322 323 // If extracting something completely in the input, if if the input is a 324 // multiple of 8 bits, recurse. 325 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 326 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 327 328 // Otherwise, if extracting a subset of the input, which is not multiple of 329 // 8 bits, do a shift and trunc to get the bits. 330 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 331 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 332 Constant *Res = CE->getOperand(0); 333 if (ByteStart) 334 Res = ConstantExpr::getLShr(Res, 335 ConstantInt::get(Res->getType(), ByteStart*8)); 336 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 337 ByteSize*8)); 338 } 339 340 // TODO: Handle the 'partially zero' case. 341 return nullptr; 342 } 343 } 344 } 345 346 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 347 /// on Ty, with any known factors factored out. If Folded is false, 348 /// return null if no factoring was possible, to avoid endlessly 349 /// bouncing an unfoldable expression back into the top-level folder. 350 /// 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 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 404 /// on Ty, with any known factors factored out. If Folded is false, 405 /// return null if no factoring was possible, to avoid endlessly 406 /// bouncing an unfoldable expression back into the top-level folder. 407 /// 408 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 409 bool Folded) { 410 // The alignment of an array is equal to the alignment of the 411 // array element. Note that this is not always true for vectors. 412 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 413 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 414 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 415 DestTy, 416 false), 417 C, DestTy); 418 return C; 419 } 420 421 if (StructType *STy = dyn_cast<StructType>(Ty)) { 422 // Packed structs always have an alignment of 1. 423 if (STy->isPacked()) 424 return ConstantInt::get(DestTy, 1); 425 426 // Otherwise, struct alignment is the maximum alignment of any member. 427 // Without target data, we can't compare much, but we can check to see 428 // if all the members have the same alignment. 429 unsigned NumElems = STy->getNumElements(); 430 // An empty struct has minimal alignment. 431 if (NumElems == 0) 432 return ConstantInt::get(DestTy, 1); 433 // Check for a struct with all members having the same alignment. 434 Constant *MemberAlign = 435 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 436 bool AllSame = true; 437 for (unsigned i = 1; i != NumElems; ++i) 438 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 439 AllSame = false; 440 break; 441 } 442 if (AllSame) 443 return MemberAlign; 444 } 445 446 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 447 // to an arbitrary pointee. 448 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 449 if (!PTy->getElementType()->isIntegerTy(1)) 450 return 451 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 452 1), 453 PTy->getAddressSpace()), 454 DestTy, true); 455 456 // If there's no interesting folding happening, bail so that we don't create 457 // a constant that looks like it needs folding but really doesn't. 458 if (!Folded) 459 return nullptr; 460 461 // Base case: Get a regular alignof expression. 462 Constant *C = ConstantExpr::getAlignOf(Ty); 463 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 464 DestTy, false), 465 C, DestTy); 466 return C; 467 } 468 469 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 470 /// on Ty and FieldNo, with any known factors factored out. If Folded is false, 471 /// return null if no factoring was possible, to avoid endlessly 472 /// bouncing an unfoldable expression back into the top-level folder. 473 /// 474 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 475 Type *DestTy, 476 bool Folded) { 477 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 478 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 479 DestTy, false), 480 FieldNo, DestTy); 481 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 482 return ConstantExpr::getNUWMul(E, N); 483 } 484 485 if (StructType *STy = dyn_cast<StructType>(Ty)) 486 if (!STy->isPacked()) { 487 unsigned NumElems = STy->getNumElements(); 488 // An empty struct has no members. 489 if (NumElems == 0) 490 return nullptr; 491 // Check for a struct with all members having the same size. 492 Constant *MemberSize = 493 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 494 bool AllSame = true; 495 for (unsigned i = 1; i != NumElems; ++i) 496 if (MemberSize != 497 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 498 AllSame = false; 499 break; 500 } 501 if (AllSame) { 502 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 503 false, 504 DestTy, 505 false), 506 FieldNo, DestTy); 507 return ConstantExpr::getNUWMul(MemberSize, N); 508 } 509 } 510 511 // If there's no interesting folding happening, bail so that we don't create 512 // a constant that looks like it needs folding but really doesn't. 513 if (!Folded) 514 return nullptr; 515 516 // Base case: Get a regular offsetof expression. 517 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 518 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 519 DestTy, false), 520 C, DestTy); 521 return C; 522 } 523 524 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 525 Type *DestTy) { 526 if (isa<UndefValue>(V)) { 527 // zext(undef) = 0, because the top bits will be zero. 528 // sext(undef) = 0, because the top bits will all be the same. 529 // [us]itofp(undef) = 0, because the result value is bounded. 530 if (opc == Instruction::ZExt || opc == Instruction::SExt || 531 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 532 return Constant::getNullValue(DestTy); 533 return UndefValue::get(DestTy); 534 } 535 536 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 537 return Constant::getNullValue(DestTy); 538 539 // If the cast operand is a constant expression, there's a few things we can 540 // do to try to simplify it. 541 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 542 if (CE->isCast()) { 543 // Try hard to fold cast of cast because they are often eliminable. 544 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 545 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 546 } else if (CE->getOpcode() == Instruction::GetElementPtr && 547 // Do not fold addrspacecast (gep 0, .., 0). It might make the 548 // addrspacecast uncanonicalized. 549 opc != Instruction::AddrSpaceCast) { 550 // If all of the indexes in the GEP are null values, there is no pointer 551 // adjustment going on. We might as well cast the source pointer. 552 bool isAllNull = true; 553 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 554 if (!CE->getOperand(i)->isNullValue()) { 555 isAllNull = false; 556 break; 557 } 558 if (isAllNull) 559 // This is casting one pointer type to another, always BitCast 560 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 561 } 562 } 563 564 // If the cast operand is a constant vector, perform the cast by 565 // operating on each element. In the cast of bitcasts, the element 566 // count may be mismatched; don't attempt to handle that here. 567 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 568 DestTy->isVectorTy() && 569 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 570 SmallVector<Constant*, 16> res; 571 VectorType *DestVecTy = cast<VectorType>(DestTy); 572 Type *DstEltTy = DestVecTy->getElementType(); 573 Type *Ty = IntegerType::get(V->getContext(), 32); 574 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 575 Constant *C = 576 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 577 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 578 } 579 return ConstantVector::get(res); 580 } 581 582 // We actually have to do a cast now. Perform the cast according to the 583 // opcode specified. 584 switch (opc) { 585 default: 586 llvm_unreachable("Failed to cast constant expression"); 587 case Instruction::FPTrunc: 588 case Instruction::FPExt: 589 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 590 bool ignored; 591 APFloat Val = FPC->getValueAPF(); 592 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 593 DestTy->isFloatTy() ? APFloat::IEEEsingle : 594 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 595 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 596 DestTy->isFP128Ty() ? APFloat::IEEEquad : 597 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble : 598 APFloat::Bogus, 599 APFloat::rmNearestTiesToEven, &ignored); 600 return ConstantFP::get(V->getContext(), Val); 601 } 602 return nullptr; // Can't fold. 603 case Instruction::FPToUI: 604 case Instruction::FPToSI: 605 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 606 const APFloat &V = FPC->getValueAPF(); 607 bool ignored; 608 uint64_t x[2]; 609 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 610 if (APFloat::opInvalidOp == 611 V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 612 APFloat::rmTowardZero, &ignored)) { 613 // Undefined behavior invoked - the destination type can't represent 614 // the input constant. 615 return UndefValue::get(DestTy); 616 } 617 APInt Val(DestBitWidth, x); 618 return ConstantInt::get(FPC->getContext(), Val); 619 } 620 return nullptr; // Can't fold. 621 case Instruction::IntToPtr: //always treated as unsigned 622 if (V->isNullValue()) // Is it an integral null value? 623 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 624 return nullptr; // Other pointer types cannot be casted 625 case Instruction::PtrToInt: // always treated as unsigned 626 // Is it a null pointer value? 627 if (V->isNullValue()) 628 return ConstantInt::get(DestTy, 0); 629 // If this is a sizeof-like expression, pull out multiplications by 630 // known factors to expose them to subsequent folding. If it's an 631 // alignof-like expression, factor out known factors. 632 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 633 if (CE->getOpcode() == Instruction::GetElementPtr && 634 CE->getOperand(0)->isNullValue()) { 635 GEPOperator *GEPO = cast<GEPOperator>(CE); 636 Type *Ty = GEPO->getSourceElementType(); 637 if (CE->getNumOperands() == 2) { 638 // Handle a sizeof-like expression. 639 Constant *Idx = CE->getOperand(1); 640 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 641 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 642 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 643 DestTy, false), 644 Idx, DestTy); 645 return ConstantExpr::getMul(C, Idx); 646 } 647 } else if (CE->getNumOperands() == 3 && 648 CE->getOperand(1)->isNullValue()) { 649 // Handle an alignof-like expression. 650 if (StructType *STy = dyn_cast<StructType>(Ty)) 651 if (!STy->isPacked()) { 652 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 653 if (CI->isOne() && 654 STy->getNumElements() == 2 && 655 STy->getElementType(0)->isIntegerTy(1)) { 656 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 657 } 658 } 659 // Handle an offsetof-like expression. 660 if (Ty->isStructTy() || Ty->isArrayTy()) { 661 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 662 DestTy, false)) 663 return C; 664 } 665 } 666 } 667 // Other pointer types cannot be casted 668 return nullptr; 669 case Instruction::UIToFP: 670 case Instruction::SIToFP: 671 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 672 const APInt &api = CI->getValue(); 673 APFloat apf(DestTy->getFltSemantics(), 674 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 675 if (APFloat::opOverflow & 676 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 677 APFloat::rmNearestTiesToEven)) { 678 // Undefined behavior invoked - the destination type can't represent 679 // the input constant. 680 return UndefValue::get(DestTy); 681 } 682 return ConstantFP::get(V->getContext(), apf); 683 } 684 return nullptr; 685 case Instruction::ZExt: 686 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 687 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 688 return ConstantInt::get(V->getContext(), 689 CI->getValue().zext(BitWidth)); 690 } 691 return nullptr; 692 case Instruction::SExt: 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().sext(BitWidth)); 697 } 698 return nullptr; 699 case Instruction::Trunc: { 700 if (V->getType()->isVectorTy()) 701 return nullptr; 702 703 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 704 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 705 return ConstantInt::get(V->getContext(), 706 CI->getValue().trunc(DestBitWidth)); 707 } 708 709 // The input must be a constantexpr. See if we can simplify this based on 710 // the bytes we are demanding. Only do this if the source and dest are an 711 // even multiple of a byte. 712 if ((DestBitWidth & 7) == 0 && 713 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 714 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 715 return Res; 716 717 return nullptr; 718 } 719 case Instruction::BitCast: 720 return FoldBitCast(V, DestTy); 721 case Instruction::AddrSpaceCast: 722 return nullptr; 723 } 724 } 725 726 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 727 Constant *V1, Constant *V2) { 728 // Check for i1 and vector true/false conditions. 729 if (Cond->isNullValue()) return V2; 730 if (Cond->isAllOnesValue()) return V1; 731 732 // If the condition is a vector constant, fold the result elementwise. 733 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 734 SmallVector<Constant*, 16> Result; 735 Type *Ty = IntegerType::get(CondV->getContext(), 32); 736 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 737 Constant *V; 738 Constant *V1Element = ConstantExpr::getExtractElement(V1, 739 ConstantInt::get(Ty, i)); 740 Constant *V2Element = ConstantExpr::getExtractElement(V2, 741 ConstantInt::get(Ty, i)); 742 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 743 if (V1Element == V2Element) { 744 V = V1Element; 745 } else if (isa<UndefValue>(Cond)) { 746 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 747 } else { 748 if (!isa<ConstantInt>(Cond)) break; 749 V = Cond->isNullValue() ? V2Element : V1Element; 750 } 751 Result.push_back(V); 752 } 753 754 // If we were able to build the vector, return it. 755 if (Result.size() == V1->getType()->getVectorNumElements()) 756 return ConstantVector::get(Result); 757 } 758 759 if (isa<UndefValue>(Cond)) { 760 if (isa<UndefValue>(V1)) return V1; 761 return V2; 762 } 763 if (isa<UndefValue>(V1)) return V2; 764 if (isa<UndefValue>(V2)) return V1; 765 if (V1 == V2) return V1; 766 767 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 768 if (TrueVal->getOpcode() == Instruction::Select) 769 if (TrueVal->getOperand(0) == Cond) 770 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 771 } 772 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 773 if (FalseVal->getOpcode() == Instruction::Select) 774 if (FalseVal->getOperand(0) == Cond) 775 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 776 } 777 778 return nullptr; 779 } 780 781 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 782 Constant *Idx) { 783 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 784 return UndefValue::get(Val->getType()->getVectorElementType()); 785 if (Val->isNullValue()) // ee(zero, x) -> zero 786 return Constant::getNullValue(Val->getType()->getVectorElementType()); 787 // ee({w,x,y,z}, undef) -> undef 788 if (isa<UndefValue>(Idx)) 789 return UndefValue::get(Val->getType()->getVectorElementType()); 790 791 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 792 // ee({w,x,y,z}, wrong_value) -> undef 793 if (CIdx->uge(Val->getType()->getVectorNumElements())) 794 return UndefValue::get(Val->getType()->getVectorElementType()); 795 return Val->getAggregateElement(CIdx->getZExtValue()); 796 } 797 return nullptr; 798 } 799 800 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 801 Constant *Elt, 802 Constant *Idx) { 803 if (isa<UndefValue>(Idx)) 804 return UndefValue::get(Val->getType()); 805 806 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 807 if (!CIdx) return nullptr; 808 809 unsigned NumElts = Val->getType()->getVectorNumElements(); 810 if (CIdx->uge(NumElts)) 811 return UndefValue::get(Val->getType()); 812 813 SmallVector<Constant*, 16> Result; 814 Result.reserve(NumElts); 815 auto *Ty = Type::getInt32Ty(Val->getContext()); 816 uint64_t IdxVal = CIdx->getZExtValue(); 817 for (unsigned i = 0; i != NumElts; ++i) { 818 if (i == IdxVal) { 819 Result.push_back(Elt); 820 continue; 821 } 822 823 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 824 Result.push_back(C); 825 } 826 827 return ConstantVector::get(Result); 828 } 829 830 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 831 Constant *V2, 832 Constant *Mask) { 833 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 834 Type *EltTy = V1->getType()->getVectorElementType(); 835 836 // Undefined shuffle mask -> undefined value. 837 if (isa<UndefValue>(Mask)) 838 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 839 840 // Don't break the bitcode reader hack. 841 if (isa<ConstantExpr>(Mask)) return nullptr; 842 843 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 844 845 // Loop over the shuffle mask, evaluating each element. 846 SmallVector<Constant*, 32> Result; 847 for (unsigned i = 0; i != MaskNumElts; ++i) { 848 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 849 if (Elt == -1) { 850 Result.push_back(UndefValue::get(EltTy)); 851 continue; 852 } 853 Constant *InElt; 854 if (unsigned(Elt) >= SrcNumElts*2) 855 InElt = UndefValue::get(EltTy); 856 else if (unsigned(Elt) >= SrcNumElts) { 857 Type *Ty = IntegerType::get(V2->getContext(), 32); 858 InElt = 859 ConstantExpr::getExtractElement(V2, 860 ConstantInt::get(Ty, Elt - SrcNumElts)); 861 } else { 862 Type *Ty = IntegerType::get(V1->getContext(), 32); 863 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 864 } 865 Result.push_back(InElt); 866 } 867 868 return ConstantVector::get(Result); 869 } 870 871 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 872 ArrayRef<unsigned> Idxs) { 873 // Base case: no indices, so return the entire value. 874 if (Idxs.empty()) 875 return Agg; 876 877 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 878 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 879 880 return nullptr; 881 } 882 883 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 884 Constant *Val, 885 ArrayRef<unsigned> Idxs) { 886 // Base case: no indices, so replace the entire value. 887 if (Idxs.empty()) 888 return Val; 889 890 unsigned NumElts; 891 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 892 NumElts = ST->getNumElements(); 893 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 894 NumElts = AT->getNumElements(); 895 else 896 NumElts = Agg->getType()->getVectorNumElements(); 897 898 SmallVector<Constant*, 32> Result; 899 for (unsigned i = 0; i != NumElts; ++i) { 900 Constant *C = Agg->getAggregateElement(i); 901 if (!C) return nullptr; 902 903 if (Idxs[0] == i) 904 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 905 906 Result.push_back(C); 907 } 908 909 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 910 return ConstantStruct::get(ST, Result); 911 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 912 return ConstantArray::get(AT, Result); 913 return ConstantVector::get(Result); 914 } 915 916 917 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 918 Constant *C1, Constant *C2) { 919 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 920 921 // Handle UndefValue up front. 922 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 923 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 924 case Instruction::Xor: 925 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 926 // Handle undef ^ undef -> 0 special case. This is a common 927 // idiom (misuse). 928 return Constant::getNullValue(C1->getType()); 929 // Fallthrough 930 case Instruction::Add: 931 case Instruction::Sub: 932 return UndefValue::get(C1->getType()); 933 case Instruction::And: 934 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 935 return C1; 936 return Constant::getNullValue(C1->getType()); // undef & X -> 0 937 case Instruction::Mul: { 938 // undef * undef -> undef 939 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 940 return C1; 941 const APInt *CV; 942 // X * undef -> undef if X is odd 943 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 944 if ((*CV)[0]) 945 return UndefValue::get(C1->getType()); 946 947 // X * undef -> 0 otherwise 948 return Constant::getNullValue(C1->getType()); 949 } 950 case Instruction::SDiv: 951 case Instruction::UDiv: 952 // X / undef -> undef 953 if (isa<UndefValue>(C2)) 954 return C2; 955 // undef / 0 -> undef 956 // undef / 1 -> undef 957 if (match(C2, m_Zero()) || match(C2, m_One())) 958 return C1; 959 // undef / X -> 0 otherwise 960 return Constant::getNullValue(C1->getType()); 961 case Instruction::URem: 962 case Instruction::SRem: 963 // X % undef -> undef 964 if (match(C2, m_Undef())) 965 return C2; 966 // undef % 0 -> undef 967 if (match(C2, m_Zero())) 968 return C1; 969 // undef % X -> 0 otherwise 970 return Constant::getNullValue(C1->getType()); 971 case Instruction::Or: // X | undef -> -1 972 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 973 return C1; 974 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 975 case Instruction::LShr: 976 // X >>l undef -> undef 977 if (isa<UndefValue>(C2)) 978 return C2; 979 // undef >>l 0 -> undef 980 if (match(C2, m_Zero())) 981 return C1; 982 // undef >>l X -> 0 983 return Constant::getNullValue(C1->getType()); 984 case Instruction::AShr: 985 // X >>a undef -> undef 986 if (isa<UndefValue>(C2)) 987 return C2; 988 // undef >>a 0 -> undef 989 if (match(C2, m_Zero())) 990 return C1; 991 // TODO: undef >>a X -> undef if the shift is exact 992 // undef >>a X -> 0 993 return Constant::getNullValue(C1->getType()); 994 case Instruction::Shl: 995 // X << undef -> undef 996 if (isa<UndefValue>(C2)) 997 return C2; 998 // undef << 0 -> undef 999 if (match(C2, m_Zero())) 1000 return C1; 1001 // undef << X -> 0 1002 return Constant::getNullValue(C1->getType()); 1003 case Instruction::FAdd: 1004 case Instruction::FSub: 1005 case Instruction::FMul: 1006 case Instruction::FDiv: 1007 case Instruction::FRem: 1008 // TODO: UNDEF handling for binary float instructions. 1009 return nullptr; 1010 case Instruction::BinaryOpsEnd: 1011 llvm_unreachable("Invalid BinaryOp"); 1012 } 1013 } 1014 1015 // At this point neither constant should be an UndefValue. 1016 assert(!isa<UndefValue>(C1) && !isa<UndefValue>(C2) && 1017 "Unexpected UndefValue"); 1018 1019 // Handle simplifications when the RHS is a constant int. 1020 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1021 switch (Opcode) { 1022 case Instruction::Add: 1023 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1024 break; 1025 case Instruction::Sub: 1026 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1027 break; 1028 case Instruction::Mul: 1029 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1030 if (CI2->equalsInt(1)) 1031 return C1; // X * 1 == X 1032 break; 1033 case Instruction::UDiv: 1034 case Instruction::SDiv: 1035 if (CI2->equalsInt(1)) 1036 return C1; // X / 1 == X 1037 if (CI2->equalsInt(0)) 1038 return UndefValue::get(CI2->getType()); // X / 0 == undef 1039 break; 1040 case Instruction::URem: 1041 case Instruction::SRem: 1042 if (CI2->equalsInt(1)) 1043 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1044 if (CI2->equalsInt(0)) 1045 return UndefValue::get(CI2->getType()); // X % 0 == undef 1046 break; 1047 case Instruction::And: 1048 if (CI2->isZero()) return C2; // X & 0 == 0 1049 if (CI2->isAllOnesValue()) 1050 return C1; // X & -1 == X 1051 1052 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1053 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1054 if (CE1->getOpcode() == Instruction::ZExt) { 1055 unsigned DstWidth = CI2->getType()->getBitWidth(); 1056 unsigned SrcWidth = 1057 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1058 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1059 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1060 return C1; 1061 } 1062 1063 // If and'ing the address of a global with a constant, fold it. 1064 if (CE1->getOpcode() == Instruction::PtrToInt && 1065 isa<GlobalValue>(CE1->getOperand(0))) { 1066 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1067 1068 // Functions are at least 4-byte aligned. 1069 unsigned GVAlign = GV->getAlignment(); 1070 if (isa<Function>(GV)) 1071 GVAlign = std::max(GVAlign, 4U); 1072 1073 if (GVAlign > 1) { 1074 unsigned DstWidth = CI2->getType()->getBitWidth(); 1075 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1076 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1077 1078 // If checking bits we know are clear, return zero. 1079 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1080 return Constant::getNullValue(CI2->getType()); 1081 } 1082 } 1083 } 1084 break; 1085 case Instruction::Or: 1086 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1087 if (CI2->isAllOnesValue()) 1088 return C2; // X | -1 == -1 1089 break; 1090 case Instruction::Xor: 1091 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1092 1093 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1094 switch (CE1->getOpcode()) { 1095 default: break; 1096 case Instruction::ICmp: 1097 case Instruction::FCmp: 1098 // cmp pred ^ true -> cmp !pred 1099 assert(CI2->equalsInt(1)); 1100 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1101 pred = CmpInst::getInversePredicate(pred); 1102 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1103 CE1->getOperand(1)); 1104 } 1105 } 1106 break; 1107 case Instruction::AShr: 1108 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1109 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1110 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1111 return ConstantExpr::getLShr(C1, C2); 1112 break; 1113 } 1114 } else if (isa<ConstantInt>(C1)) { 1115 // If C1 is a ConstantInt and C2 is not, swap the operands. 1116 if (Instruction::isCommutative(Opcode)) 1117 return ConstantExpr::get(Opcode, C2, C1); 1118 } 1119 1120 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1121 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1122 const APInt &C1V = CI1->getValue(); 1123 const APInt &C2V = CI2->getValue(); 1124 switch (Opcode) { 1125 default: 1126 break; 1127 case Instruction::Add: 1128 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1129 case Instruction::Sub: 1130 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1131 case Instruction::Mul: 1132 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1133 case Instruction::UDiv: 1134 assert(!CI2->isNullValue() && "Div by zero handled above"); 1135 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1136 case Instruction::SDiv: 1137 assert(!CI2->isNullValue() && "Div by zero handled above"); 1138 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1139 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1140 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1141 case Instruction::URem: 1142 assert(!CI2->isNullValue() && "Div by zero handled above"); 1143 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1144 case Instruction::SRem: 1145 assert(!CI2->isNullValue() && "Div by zero handled above"); 1146 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1147 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1148 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1149 case Instruction::And: 1150 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1151 case Instruction::Or: 1152 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1153 case Instruction::Xor: 1154 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1155 case Instruction::Shl: 1156 if (C2V.ult(C1V.getBitWidth())) 1157 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1158 return UndefValue::get(C1->getType()); // too big shift is undef 1159 case Instruction::LShr: 1160 if (C2V.ult(C1V.getBitWidth())) 1161 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1162 return UndefValue::get(C1->getType()); // too big shift is undef 1163 case Instruction::AShr: 1164 if (C2V.ult(C1V.getBitWidth())) 1165 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1166 return UndefValue::get(C1->getType()); // too big shift is undef 1167 } 1168 } 1169 1170 switch (Opcode) { 1171 case Instruction::SDiv: 1172 case Instruction::UDiv: 1173 case Instruction::URem: 1174 case Instruction::SRem: 1175 case Instruction::LShr: 1176 case Instruction::AShr: 1177 case Instruction::Shl: 1178 if (CI1->equalsInt(0)) return C1; 1179 break; 1180 default: 1181 break; 1182 } 1183 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1184 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1185 const APFloat &C1V = CFP1->getValueAPF(); 1186 const APFloat &C2V = CFP2->getValueAPF(); 1187 APFloat C3V = C1V; // copy for modification 1188 switch (Opcode) { 1189 default: 1190 break; 1191 case Instruction::FAdd: 1192 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1193 return ConstantFP::get(C1->getContext(), C3V); 1194 case Instruction::FSub: 1195 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1196 return ConstantFP::get(C1->getContext(), C3V); 1197 case Instruction::FMul: 1198 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1199 return ConstantFP::get(C1->getContext(), C3V); 1200 case Instruction::FDiv: 1201 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1202 return ConstantFP::get(C1->getContext(), C3V); 1203 case Instruction::FRem: 1204 (void)C3V.mod(C2V); 1205 return ConstantFP::get(C1->getContext(), C3V); 1206 } 1207 } 1208 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1209 // Perform elementwise folding. 1210 SmallVector<Constant*, 16> Result; 1211 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1212 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1213 Constant *LHS = 1214 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1215 Constant *RHS = 1216 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1217 1218 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1219 } 1220 1221 return ConstantVector::get(Result); 1222 } 1223 1224 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1225 // There are many possible foldings we could do here. We should probably 1226 // at least fold add of a pointer with an integer into the appropriate 1227 // getelementptr. This will improve alias analysis a bit. 1228 1229 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1230 // (a + (b + c)). 1231 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1232 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1233 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1234 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1235 } 1236 } else if (isa<ConstantExpr>(C2)) { 1237 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1238 // other way if possible. 1239 if (Instruction::isCommutative(Opcode)) 1240 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1241 } 1242 1243 // i1 can be simplified in many cases. 1244 if (C1->getType()->isIntegerTy(1)) { 1245 switch (Opcode) { 1246 case Instruction::Add: 1247 case Instruction::Sub: 1248 return ConstantExpr::getXor(C1, C2); 1249 case Instruction::Mul: 1250 return ConstantExpr::getAnd(C1, C2); 1251 case Instruction::Shl: 1252 case Instruction::LShr: 1253 case Instruction::AShr: 1254 // We can assume that C2 == 0. If it were one the result would be 1255 // undefined because the shift value is as large as the bitwidth. 1256 return C1; 1257 case Instruction::SDiv: 1258 case Instruction::UDiv: 1259 // We can assume that C2 == 1. If it were zero the result would be 1260 // undefined through division by zero. 1261 return C1; 1262 case Instruction::URem: 1263 case Instruction::SRem: 1264 // We can assume that C2 == 1. If it were zero the result would be 1265 // undefined through division by zero. 1266 return ConstantInt::getFalse(C1->getContext()); 1267 default: 1268 break; 1269 } 1270 } 1271 1272 // We don't know how to fold this. 1273 return nullptr; 1274 } 1275 1276 /// isZeroSizedType - This type is zero sized if its an array or structure of 1277 /// zero sized types. The only leaf zero sized type is an empty structure. 1278 static bool isMaybeZeroSizedType(Type *Ty) { 1279 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1280 if (STy->isOpaque()) return true; // Can't say. 1281 1282 // If all of elements have zero size, this does too. 1283 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1284 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1285 return true; 1286 1287 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1288 return isMaybeZeroSizedType(ATy->getElementType()); 1289 } 1290 return false; 1291 } 1292 1293 /// IdxCompare - Compare the two constants as though they were getelementptr 1294 /// indices. This allows coercion of the types to be the same thing. 1295 /// 1296 /// If the two constants are the "same" (after coercion), return 0. If the 1297 /// first is less than the second, return -1, if the second is less than the 1298 /// first, return 1. If the constants are not integral, return -2. 1299 /// 1300 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1301 if (C1 == C2) return 0; 1302 1303 // Ok, we found a different index. If they are not ConstantInt, we can't do 1304 // anything with them. 1305 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1306 return -2; // don't know! 1307 1308 // We cannot compare the indices if they don't fit in an int64_t. 1309 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 || 1310 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64) 1311 return -2; // don't know! 1312 1313 // Ok, we have two differing integer indices. Sign extend them to be the same 1314 // type. 1315 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue(); 1316 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue(); 1317 1318 if (C1Val == C2Val) return 0; // They are equal 1319 1320 // If the type being indexed over is really just a zero sized type, there is 1321 // no pointer difference being made here. 1322 if (isMaybeZeroSizedType(ElTy)) 1323 return -2; // dunno. 1324 1325 // If they are really different, now that they are the same type, then we 1326 // found a difference! 1327 if (C1Val < C2Val) 1328 return -1; 1329 else 1330 return 1; 1331 } 1332 1333 /// evaluateFCmpRelation - This function determines if there is anything we can 1334 /// decide about the two constants provided. This doesn't need to handle simple 1335 /// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1336 /// If we can determine that the two constants have a particular relation to 1337 /// each other, we should return the corresponding FCmpInst predicate, 1338 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1339 /// ConstantFoldCompareInstruction. 1340 /// 1341 /// To simplify this code we canonicalize the relation so that the first 1342 /// operand is always the most "complex" of the two. We consider ConstantFP 1343 /// to be the simplest, and ConstantExprs to be the most complex. 1344 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1345 assert(V1->getType() == V2->getType() && 1346 "Cannot compare values of different types!"); 1347 1348 // Handle degenerate case quickly 1349 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1350 1351 if (!isa<ConstantExpr>(V1)) { 1352 if (!isa<ConstantExpr>(V2)) { 1353 // Simple case, use the standard constant folder. 1354 ConstantInt *R = nullptr; 1355 R = dyn_cast<ConstantInt>( 1356 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1357 if (R && !R->isZero()) 1358 return FCmpInst::FCMP_OEQ; 1359 R = dyn_cast<ConstantInt>( 1360 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1361 if (R && !R->isZero()) 1362 return FCmpInst::FCMP_OLT; 1363 R = dyn_cast<ConstantInt>( 1364 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1365 if (R && !R->isZero()) 1366 return FCmpInst::FCMP_OGT; 1367 1368 // Nothing more we can do 1369 return FCmpInst::BAD_FCMP_PREDICATE; 1370 } 1371 1372 // If the first operand is simple and second is ConstantExpr, swap operands. 1373 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1374 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1375 return FCmpInst::getSwappedPredicate(SwappedRelation); 1376 } else { 1377 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1378 // constantexpr or a simple constant. 1379 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1380 switch (CE1->getOpcode()) { 1381 case Instruction::FPTrunc: 1382 case Instruction::FPExt: 1383 case Instruction::UIToFP: 1384 case Instruction::SIToFP: 1385 // We might be able to do something with these but we don't right now. 1386 break; 1387 default: 1388 break; 1389 } 1390 } 1391 // There are MANY other foldings that we could perform here. They will 1392 // probably be added on demand, as they seem needed. 1393 return FCmpInst::BAD_FCMP_PREDICATE; 1394 } 1395 1396 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1397 const GlobalValue *GV2) { 1398 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1399 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1400 return true; 1401 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1402 Type *Ty = GVar->getValueType(); 1403 // A global with opaque type might end up being zero sized. 1404 if (!Ty->isSized()) 1405 return true; 1406 // A global with an empty type might lie at the address of any other 1407 // global. 1408 if (Ty->isEmptyTy()) 1409 return true; 1410 } 1411 return false; 1412 }; 1413 // Don't try to decide equality of aliases. 1414 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1415 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1416 return ICmpInst::ICMP_NE; 1417 return ICmpInst::BAD_ICMP_PREDICATE; 1418 } 1419 1420 /// evaluateICmpRelation - This function determines if there is anything we can 1421 /// decide about the two constants provided. This doesn't need to handle simple 1422 /// things like integer comparisons, but should instead handle ConstantExprs 1423 /// and GlobalValues. If we can determine that the two constants have a 1424 /// particular relation to each other, we should return the corresponding ICmp 1425 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1426 /// 1427 /// To simplify this code we canonicalize the relation so that the first 1428 /// operand is always the most "complex" of the two. We consider simple 1429 /// constants (like ConstantInt) to be the simplest, followed by 1430 /// GlobalValues, followed by ConstantExpr's (the most complex). 1431 /// 1432 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1433 bool isSigned) { 1434 assert(V1->getType() == V2->getType() && 1435 "Cannot compare different types of values!"); 1436 if (V1 == V2) return ICmpInst::ICMP_EQ; 1437 1438 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1439 !isa<BlockAddress>(V1)) { 1440 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1441 !isa<BlockAddress>(V2)) { 1442 // We distilled this down to a simple case, use the standard constant 1443 // folder. 1444 ConstantInt *R = nullptr; 1445 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1446 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1447 if (R && !R->isZero()) 1448 return pred; 1449 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1450 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1451 if (R && !R->isZero()) 1452 return pred; 1453 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1454 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1455 if (R && !R->isZero()) 1456 return pred; 1457 1458 // If we couldn't figure it out, bail. 1459 return ICmpInst::BAD_ICMP_PREDICATE; 1460 } 1461 1462 // If the first operand is simple, swap operands. 1463 ICmpInst::Predicate SwappedRelation = 1464 evaluateICmpRelation(V2, V1, isSigned); 1465 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1466 return ICmpInst::getSwappedPredicate(SwappedRelation); 1467 1468 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1469 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1470 ICmpInst::Predicate SwappedRelation = 1471 evaluateICmpRelation(V2, V1, isSigned); 1472 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1473 return ICmpInst::getSwappedPredicate(SwappedRelation); 1474 return ICmpInst::BAD_ICMP_PREDICATE; 1475 } 1476 1477 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1478 // constant (which, since the types must match, means that it's a 1479 // ConstantPointerNull). 1480 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1481 return areGlobalsPotentiallyEqual(GV, GV2); 1482 } else if (isa<BlockAddress>(V2)) { 1483 return ICmpInst::ICMP_NE; // Globals never equal labels. 1484 } else { 1485 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1486 // GlobalVals can never be null unless they have external weak linkage. 1487 // We don't try to evaluate aliases here. 1488 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1489 return ICmpInst::ICMP_NE; 1490 } 1491 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1492 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1493 ICmpInst::Predicate SwappedRelation = 1494 evaluateICmpRelation(V2, V1, isSigned); 1495 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1496 return ICmpInst::getSwappedPredicate(SwappedRelation); 1497 return ICmpInst::BAD_ICMP_PREDICATE; 1498 } 1499 1500 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1501 // constant (which, since the types must match, means that it is a 1502 // ConstantPointerNull). 1503 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1504 // Block address in another function can't equal this one, but block 1505 // addresses in the current function might be the same if blocks are 1506 // empty. 1507 if (BA2->getFunction() != BA->getFunction()) 1508 return ICmpInst::ICMP_NE; 1509 } else { 1510 // Block addresses aren't null, don't equal the address of globals. 1511 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1512 "Canonicalization guarantee!"); 1513 return ICmpInst::ICMP_NE; 1514 } 1515 } else { 1516 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1517 // constantexpr, a global, block address, or a simple constant. 1518 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1519 Constant *CE1Op0 = CE1->getOperand(0); 1520 1521 switch (CE1->getOpcode()) { 1522 case Instruction::Trunc: 1523 case Instruction::FPTrunc: 1524 case Instruction::FPExt: 1525 case Instruction::FPToUI: 1526 case Instruction::FPToSI: 1527 break; // We can't evaluate floating point casts or truncations. 1528 1529 case Instruction::UIToFP: 1530 case Instruction::SIToFP: 1531 case Instruction::BitCast: 1532 case Instruction::ZExt: 1533 case Instruction::SExt: 1534 // If the cast is not actually changing bits, and the second operand is a 1535 // null pointer, do the comparison with the pre-casted value. 1536 if (V2->isNullValue() && 1537 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1538 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1539 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1540 return evaluateICmpRelation(CE1Op0, 1541 Constant::getNullValue(CE1Op0->getType()), 1542 isSigned); 1543 } 1544 break; 1545 1546 case Instruction::GetElementPtr: { 1547 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1548 // Ok, since this is a getelementptr, we know that the constant has a 1549 // pointer type. Check the various cases. 1550 if (isa<ConstantPointerNull>(V2)) { 1551 // If we are comparing a GEP to a null pointer, check to see if the base 1552 // of the GEP equals the null pointer. 1553 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1554 if (GV->hasExternalWeakLinkage()) 1555 // Weak linkage GVals could be zero or not. We're comparing that 1556 // to null pointer so its greater-or-equal 1557 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1558 else 1559 // If its not weak linkage, the GVal must have a non-zero address 1560 // so the result is greater-than 1561 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1562 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1563 // If we are indexing from a null pointer, check to see if we have any 1564 // non-zero indices. 1565 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1566 if (!CE1->getOperand(i)->isNullValue()) 1567 // Offsetting from null, must not be equal. 1568 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1569 // Only zero indexes from null, must still be zero. 1570 return ICmpInst::ICMP_EQ; 1571 } 1572 // Otherwise, we can't really say if the first operand is null or not. 1573 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1574 if (isa<ConstantPointerNull>(CE1Op0)) { 1575 if (GV2->hasExternalWeakLinkage()) 1576 // Weak linkage GVals could be zero or not. We're comparing it to 1577 // a null pointer, so its less-or-equal 1578 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1579 else 1580 // If its not weak linkage, the GVal must have a non-zero address 1581 // so the result is less-than 1582 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1583 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1584 if (GV == GV2) { 1585 // If this is a getelementptr of the same global, then it must be 1586 // different. Because the types must match, the getelementptr could 1587 // only have at most one index, and because we fold getelementptr's 1588 // with a single zero index, it must be nonzero. 1589 assert(CE1->getNumOperands() == 2 && 1590 !CE1->getOperand(1)->isNullValue() && 1591 "Surprising getelementptr!"); 1592 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1593 } else { 1594 if (CE1GEP->hasAllZeroIndices()) 1595 return areGlobalsPotentiallyEqual(GV, GV2); 1596 return ICmpInst::BAD_ICMP_PREDICATE; 1597 } 1598 } 1599 } else { 1600 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1601 Constant *CE2Op0 = CE2->getOperand(0); 1602 1603 // There are MANY other foldings that we could perform here. They will 1604 // probably be added on demand, as they seem needed. 1605 switch (CE2->getOpcode()) { 1606 default: break; 1607 case Instruction::GetElementPtr: 1608 // By far the most common case to handle is when the base pointers are 1609 // obviously to the same global. 1610 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1611 // Don't know relative ordering, but check for inequality. 1612 if (CE1Op0 != CE2Op0) { 1613 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1614 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1615 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1616 cast<GlobalValue>(CE2Op0)); 1617 return ICmpInst::BAD_ICMP_PREDICATE; 1618 } 1619 // Ok, we know that both getelementptr instructions are based on the 1620 // same global. From this, we can precisely determine the relative 1621 // ordering of the resultant pointers. 1622 unsigned i = 1; 1623 1624 // The logic below assumes that the result of the comparison 1625 // can be determined by finding the first index that differs. 1626 // This doesn't work if there is over-indexing in any 1627 // subsequent indices, so check for that case first. 1628 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1629 !CE2->isGEPWithNoNotionalOverIndexing()) 1630 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1631 1632 // Compare all of the operands the GEP's have in common. 1633 gep_type_iterator GTI = gep_type_begin(CE1); 1634 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1635 ++i, ++GTI) 1636 switch (IdxCompare(CE1->getOperand(i), 1637 CE2->getOperand(i), GTI.getIndexedType())) { 1638 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1639 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1640 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1641 } 1642 1643 // Ok, we ran out of things they have in common. If any leftovers 1644 // are non-zero then we have a difference, otherwise we are equal. 1645 for (; i < CE1->getNumOperands(); ++i) 1646 if (!CE1->getOperand(i)->isNullValue()) { 1647 if (isa<ConstantInt>(CE1->getOperand(i))) 1648 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1649 else 1650 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1651 } 1652 1653 for (; i < CE2->getNumOperands(); ++i) 1654 if (!CE2->getOperand(i)->isNullValue()) { 1655 if (isa<ConstantInt>(CE2->getOperand(i))) 1656 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1657 else 1658 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1659 } 1660 return ICmpInst::ICMP_EQ; 1661 } 1662 } 1663 } 1664 } 1665 default: 1666 break; 1667 } 1668 } 1669 1670 return ICmpInst::BAD_ICMP_PREDICATE; 1671 } 1672 1673 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1674 Constant *C1, Constant *C2) { 1675 Type *ResultTy; 1676 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1677 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1678 VT->getNumElements()); 1679 else 1680 ResultTy = Type::getInt1Ty(C1->getContext()); 1681 1682 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1683 if (pred == FCmpInst::FCMP_FALSE) 1684 return Constant::getNullValue(ResultTy); 1685 1686 if (pred == FCmpInst::FCMP_TRUE) 1687 return Constant::getAllOnesValue(ResultTy); 1688 1689 // Handle some degenerate cases first 1690 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1691 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1692 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1693 // For EQ and NE, we can always pick a value for the undef to make the 1694 // predicate pass or fail, so we can return undef. 1695 // Also, if both operands are undef, we can return undef for int comparison. 1696 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1697 return UndefValue::get(ResultTy); 1698 1699 // Otherwise, for integer compare, pick the same value as the non-undef 1700 // operand, and fold it to true or false. 1701 if (isIntegerPredicate) 1702 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1703 1704 // Choosing NaN for the undef will always make unordered comparison succeed 1705 // and ordered comparison fails. 1706 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1707 } 1708 1709 // icmp eq/ne(null,GV) -> false/true 1710 if (C1->isNullValue()) { 1711 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1712 // Don't try to evaluate aliases. External weak GV can be null. 1713 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1714 if (pred == ICmpInst::ICMP_EQ) 1715 return ConstantInt::getFalse(C1->getContext()); 1716 else if (pred == ICmpInst::ICMP_NE) 1717 return ConstantInt::getTrue(C1->getContext()); 1718 } 1719 // icmp eq/ne(GV,null) -> false/true 1720 } else if (C2->isNullValue()) { 1721 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1722 // Don't try to evaluate aliases. External weak GV can be null. 1723 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1724 if (pred == ICmpInst::ICMP_EQ) 1725 return ConstantInt::getFalse(C1->getContext()); 1726 else if (pred == ICmpInst::ICMP_NE) 1727 return ConstantInt::getTrue(C1->getContext()); 1728 } 1729 } 1730 1731 // If the comparison is a comparison between two i1's, simplify it. 1732 if (C1->getType()->isIntegerTy(1)) { 1733 switch(pred) { 1734 case ICmpInst::ICMP_EQ: 1735 if (isa<ConstantInt>(C2)) 1736 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1737 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1738 case ICmpInst::ICMP_NE: 1739 return ConstantExpr::getXor(C1, C2); 1740 default: 1741 break; 1742 } 1743 } 1744 1745 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1746 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1747 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1748 switch (pred) { 1749 default: llvm_unreachable("Invalid ICmp Predicate"); 1750 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1751 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1752 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1753 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1754 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1755 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1756 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1757 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1758 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1759 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1760 } 1761 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1762 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1763 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1764 APFloat::cmpResult R = C1V.compare(C2V); 1765 switch (pred) { 1766 default: llvm_unreachable("Invalid FCmp Predicate"); 1767 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1768 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1769 case FCmpInst::FCMP_UNO: 1770 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1771 case FCmpInst::FCMP_ORD: 1772 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1773 case FCmpInst::FCMP_UEQ: 1774 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1775 R==APFloat::cmpEqual); 1776 case FCmpInst::FCMP_OEQ: 1777 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1778 case FCmpInst::FCMP_UNE: 1779 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1780 case FCmpInst::FCMP_ONE: 1781 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1782 R==APFloat::cmpGreaterThan); 1783 case FCmpInst::FCMP_ULT: 1784 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1785 R==APFloat::cmpLessThan); 1786 case FCmpInst::FCMP_OLT: 1787 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1788 case FCmpInst::FCMP_UGT: 1789 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1790 R==APFloat::cmpGreaterThan); 1791 case FCmpInst::FCMP_OGT: 1792 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1793 case FCmpInst::FCMP_ULE: 1794 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1795 case FCmpInst::FCMP_OLE: 1796 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1797 R==APFloat::cmpEqual); 1798 case FCmpInst::FCMP_UGE: 1799 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1800 case FCmpInst::FCMP_OGE: 1801 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1802 R==APFloat::cmpEqual); 1803 } 1804 } else if (C1->getType()->isVectorTy()) { 1805 // If we can constant fold the comparison of each element, constant fold 1806 // the whole vector comparison. 1807 SmallVector<Constant*, 4> ResElts; 1808 Type *Ty = IntegerType::get(C1->getContext(), 32); 1809 // Compare the elements, producing an i1 result or constant expr. 1810 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1811 Constant *C1E = 1812 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1813 Constant *C2E = 1814 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1815 1816 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1817 } 1818 1819 return ConstantVector::get(ResElts); 1820 } 1821 1822 if (C1->getType()->isFloatingPointTy() && 1823 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1824 // infinite recursive loop 1825 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1826 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1827 switch (evaluateFCmpRelation(C1, C2)) { 1828 default: llvm_unreachable("Unknown relation!"); 1829 case FCmpInst::FCMP_UNO: 1830 case FCmpInst::FCMP_ORD: 1831 case FCmpInst::FCMP_UEQ: 1832 case FCmpInst::FCMP_UNE: 1833 case FCmpInst::FCMP_ULT: 1834 case FCmpInst::FCMP_UGT: 1835 case FCmpInst::FCMP_ULE: 1836 case FCmpInst::FCMP_UGE: 1837 case FCmpInst::FCMP_TRUE: 1838 case FCmpInst::FCMP_FALSE: 1839 case FCmpInst::BAD_FCMP_PREDICATE: 1840 break; // Couldn't determine anything about these constants. 1841 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1842 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1843 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1844 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1845 break; 1846 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1847 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1848 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1849 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1850 break; 1851 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1852 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1853 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1854 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1855 break; 1856 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1857 // We can only partially decide this relation. 1858 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1859 Result = 0; 1860 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1861 Result = 1; 1862 break; 1863 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1864 // We can only partially decide this relation. 1865 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1866 Result = 0; 1867 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1868 Result = 1; 1869 break; 1870 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1871 // We can only partially decide this relation. 1872 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1873 Result = 0; 1874 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1875 Result = 1; 1876 break; 1877 } 1878 1879 // If we evaluated the result, return it now. 1880 if (Result != -1) 1881 return ConstantInt::get(ResultTy, Result); 1882 1883 } else { 1884 // Evaluate the relation between the two constants, per the predicate. 1885 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1886 switch (evaluateICmpRelation(C1, C2, 1887 CmpInst::isSigned((CmpInst::Predicate)pred))) { 1888 default: llvm_unreachable("Unknown relational!"); 1889 case ICmpInst::BAD_ICMP_PREDICATE: 1890 break; // Couldn't determine anything about these constants. 1891 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1892 // If we know the constants are equal, we can decide the result of this 1893 // computation precisely. 1894 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1895 break; 1896 case ICmpInst::ICMP_ULT: 1897 switch (pred) { 1898 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1899 Result = 1; break; 1900 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1901 Result = 0; break; 1902 } 1903 break; 1904 case ICmpInst::ICMP_SLT: 1905 switch (pred) { 1906 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1907 Result = 1; break; 1908 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1909 Result = 0; break; 1910 } 1911 break; 1912 case ICmpInst::ICMP_UGT: 1913 switch (pred) { 1914 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1915 Result = 1; break; 1916 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1917 Result = 0; break; 1918 } 1919 break; 1920 case ICmpInst::ICMP_SGT: 1921 switch (pred) { 1922 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1923 Result = 1; break; 1924 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1925 Result = 0; break; 1926 } 1927 break; 1928 case ICmpInst::ICMP_ULE: 1929 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1930 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1931 break; 1932 case ICmpInst::ICMP_SLE: 1933 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1934 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1935 break; 1936 case ICmpInst::ICMP_UGE: 1937 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1938 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1939 break; 1940 case ICmpInst::ICMP_SGE: 1941 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1942 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1943 break; 1944 case ICmpInst::ICMP_NE: 1945 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1946 if (pred == ICmpInst::ICMP_NE) Result = 1; 1947 break; 1948 } 1949 1950 // If we evaluated the result, return it now. 1951 if (Result != -1) 1952 return ConstantInt::get(ResultTy, Result); 1953 1954 // If the right hand side is a bitcast, try using its inverse to simplify 1955 // it by moving it to the left hand side. We can't do this if it would turn 1956 // a vector compare into a scalar compare or visa versa. 1957 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1958 Constant *CE2Op0 = CE2->getOperand(0); 1959 if (CE2->getOpcode() == Instruction::BitCast && 1960 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1961 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1962 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1963 } 1964 } 1965 1966 // If the left hand side is an extension, try eliminating it. 1967 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1968 if ((CE1->getOpcode() == Instruction::SExt && 1969 ICmpInst::isSigned((ICmpInst::Predicate)pred)) || 1970 (CE1->getOpcode() == Instruction::ZExt && 1971 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){ 1972 Constant *CE1Op0 = CE1->getOperand(0); 1973 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1974 if (CE1Inverse == CE1Op0) { 1975 // Check whether we can safely truncate the right hand side. 1976 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1977 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1978 C2->getType()) == C2) 1979 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1980 } 1981 } 1982 } 1983 1984 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1985 (C1->isNullValue() && !C2->isNullValue())) { 1986 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1987 // other way if possible. 1988 // Also, if C1 is null and C2 isn't, flip them around. 1989 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1990 return ConstantExpr::getICmp(pred, C2, C1); 1991 } 1992 } 1993 return nullptr; 1994 } 1995 1996 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1997 /// is "inbounds". 1998 template<typename IndexTy> 1999 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 2000 // No indices means nothing that could be out of bounds. 2001 if (Idxs.empty()) return true; 2002 2003 // If the first index is zero, it's in bounds. 2004 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2005 2006 // If the first index is one and all the rest are zero, it's in bounds, 2007 // by the one-past-the-end rule. 2008 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2009 return false; 2010 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 2011 if (!cast<Constant>(Idxs[i])->isNullValue()) 2012 return false; 2013 return true; 2014 } 2015 2016 /// \brief Test whether a given ConstantInt is in-range for a SequentialType. 2017 static bool isIndexInRangeOfSequentialType(SequentialType *STy, 2018 const ConstantInt *CI) { 2019 // And indices are valid when indexing along a pointer 2020 if (isa<PointerType>(STy)) 2021 return true; 2022 2023 uint64_t NumElements = 0; 2024 // Determine the number of elements in our sequential type. 2025 if (auto *ATy = dyn_cast<ArrayType>(STy)) 2026 NumElements = ATy->getNumElements(); 2027 else if (auto *VTy = dyn_cast<VectorType>(STy)) 2028 NumElements = VTy->getNumElements(); 2029 2030 assert((isa<ArrayType>(STy) || NumElements > 0) && 2031 "didn't expect non-array type to have zero elements!"); 2032 2033 // We cannot bounds check the index if it doesn't fit in an int64_t. 2034 if (CI->getValue().getActiveBits() > 64) 2035 return false; 2036 2037 // A negative index or an index past the end of our sequential type is 2038 // considered out-of-range. 2039 int64_t IndexVal = CI->getSExtValue(); 2040 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2041 return false; 2042 2043 // Otherwise, it is in-range. 2044 return true; 2045 } 2046 2047 template<typename IndexTy> 2048 static Constant *ConstantFoldGetElementPtrImpl(Type *PointeeTy, Constant *C, 2049 bool inBounds, 2050 ArrayRef<IndexTy> Idxs) { 2051 if (Idxs.empty()) return C; 2052 Constant *Idx0 = cast<Constant>(Idxs[0]); 2053 if ((Idxs.size() == 1 && Idx0->isNullValue())) 2054 return C; 2055 2056 if (isa<UndefValue>(C)) { 2057 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2058 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2059 assert(Ty && "Invalid indices for GEP!"); 2060 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2061 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2062 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 2063 return UndefValue::get(GEPTy); 2064 } 2065 2066 if (C->isNullValue()) { 2067 bool isNull = true; 2068 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2069 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2070 isNull = false; 2071 break; 2072 } 2073 if (isNull) { 2074 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2075 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2076 2077 assert(Ty && "Invalid indices for GEP!"); 2078 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2079 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2080 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 2081 return Constant::getNullValue(GEPTy); 2082 } 2083 } 2084 2085 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2086 // Combine Indices - If the source pointer to this getelementptr instruction 2087 // is a getelementptr instruction, combine the indices of the two 2088 // getelementptr instructions into a single instruction. 2089 // 2090 if (CE->getOpcode() == Instruction::GetElementPtr) { 2091 Type *LastTy = nullptr; 2092 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2093 I != E; ++I) 2094 LastTy = *I; 2095 2096 // We cannot combine indices if doing so would take us outside of an 2097 // array or vector. Doing otherwise could trick us if we evaluated such a 2098 // GEP as part of a load. 2099 // 2100 // e.g. Consider if the original GEP was: 2101 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2102 // i32 0, i32 0, i64 0) 2103 // 2104 // If we then tried to offset it by '8' to get to the third element, 2105 // an i8, we should *not* get: 2106 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2107 // i32 0, i32 0, i64 8) 2108 // 2109 // This GEP tries to index array element '8 which runs out-of-bounds. 2110 // Subsequent evaluation would get confused and produce erroneous results. 2111 // 2112 // The following prohibits such a GEP from being formed by checking to see 2113 // if the index is in-range with respect to an array or vector. 2114 bool PerformFold = false; 2115 if (Idx0->isNullValue()) 2116 PerformFold = true; 2117 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy)) 2118 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2119 PerformFold = isIndexInRangeOfSequentialType(STy, CI); 2120 2121 if (PerformFold) { 2122 SmallVector<Value*, 16> NewIndices; 2123 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2124 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2125 2126 // Add the last index of the source with the first index of the new GEP. 2127 // Make sure to handle the case when they are actually different types. 2128 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2129 // Otherwise it must be an array. 2130 if (!Idx0->isNullValue()) { 2131 Type *IdxTy = Combined->getType(); 2132 if (IdxTy != Idx0->getType()) { 2133 unsigned CommonExtendedWidth = 2134 std::max(IdxTy->getIntegerBitWidth(), 2135 Idx0->getType()->getIntegerBitWidth()); 2136 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2137 2138 Type *CommonTy = 2139 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2140 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2141 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2142 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2143 } else { 2144 Combined = 2145 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2146 } 2147 } 2148 2149 NewIndices.push_back(Combined); 2150 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2151 return ConstantExpr::getGetElementPtr( 2152 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2153 NewIndices, inBounds && cast<GEPOperator>(CE)->isInBounds()); 2154 } 2155 } 2156 2157 // Attempt to fold casts to the same type away. For example, folding: 2158 // 2159 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2160 // i64 0, i64 0) 2161 // into: 2162 // 2163 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2164 // 2165 // Don't fold if the cast is changing address spaces. 2166 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2167 PointerType *SrcPtrTy = 2168 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2169 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2170 if (SrcPtrTy && DstPtrTy) { 2171 ArrayType *SrcArrayTy = 2172 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2173 ArrayType *DstArrayTy = 2174 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2175 if (SrcArrayTy && DstArrayTy 2176 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2177 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2178 return ConstantExpr::getGetElementPtr( 2179 SrcArrayTy, (Constant *)CE->getOperand(0), Idxs, inBounds); 2180 } 2181 } 2182 } 2183 2184 // Check to see if any array indices are not within the corresponding 2185 // notional array or vector bounds. If so, try to determine if they can be 2186 // factored out into preceding dimensions. 2187 SmallVector<Constant *, 8> NewIdxs; 2188 Type *Ty = PointeeTy; 2189 Type *Prev = C->getType(); 2190 bool Unknown = !isa<ConstantInt>(Idxs[0]); 2191 for (unsigned i = 1, e = Idxs.size(); i != e; 2192 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2193 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2194 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) 2195 if (CI->getSExtValue() > 0 && 2196 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) { 2197 if (isa<SequentialType>(Prev)) { 2198 // It's out of range, but we can factor it into the prior 2199 // dimension. 2200 NewIdxs.resize(Idxs.size()); 2201 uint64_t NumElements = 0; 2202 if (auto *ATy = dyn_cast<ArrayType>(Ty)) 2203 NumElements = ATy->getNumElements(); 2204 else 2205 NumElements = cast<VectorType>(Ty)->getNumElements(); 2206 2207 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2208 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2209 2210 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2211 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2212 2213 unsigned CommonExtendedWidth = 2214 std::max(PrevIdx->getType()->getIntegerBitWidth(), 2215 Div->getType()->getIntegerBitWidth()); 2216 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2217 2218 // Before adding, extend both operands to i64 to avoid 2219 // overflow trouble. 2220 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth)) 2221 PrevIdx = ConstantExpr::getSExt( 2222 PrevIdx, 2223 Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2224 if (!Div->getType()->isIntegerTy(CommonExtendedWidth)) 2225 Div = ConstantExpr::getSExt( 2226 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2227 2228 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2229 } else { 2230 // It's out of range, but the prior dimension is a struct 2231 // so we can't do anything about it. 2232 Unknown = true; 2233 } 2234 } 2235 } else { 2236 // We don't know if it's in range or not. 2237 Unknown = true; 2238 } 2239 } 2240 2241 // If we did any factoring, start over with the adjusted indices. 2242 if (!NewIdxs.empty()) { 2243 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2244 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2245 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, inBounds); 2246 } 2247 2248 // If all indices are known integers and normalized, we can do a simple 2249 // check for the "inbounds" property. 2250 if (!Unknown && !inBounds) 2251 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2252 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2253 return ConstantExpr::getInBoundsGetElementPtr(PointeeTy, C, Idxs); 2254 2255 return nullptr; 2256 } 2257 2258 Constant *llvm::ConstantFoldGetElementPtr(Type *Ty, Constant *C, 2259 bool inBounds, 2260 ArrayRef<Constant *> Idxs) { 2261 return ConstantFoldGetElementPtrImpl(Ty, C, inBounds, Idxs); 2262 } 2263 2264 Constant *llvm::ConstantFoldGetElementPtr(Type *Ty, Constant *C, 2265 bool inBounds, 2266 ArrayRef<Value *> Idxs) { 2267 return ConstantFoldGetElementPtrImpl(Ty, C, inBounds, Idxs); 2268 } 2269