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/APSInt.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/IR/Constants.h" 24 #include "llvm/IR/DerivedTypes.h" 25 #include "llvm/IR/Function.h" 26 #include "llvm/IR/GetElementPtrTypeIterator.h" 27 #include "llvm/IR/GlobalAlias.h" 28 #include "llvm/IR/GlobalVariable.h" 29 #include "llvm/IR/Instructions.h" 30 #include "llvm/IR/Operator.h" 31 #include "llvm/IR/PatternMatch.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/ManagedStatic.h" 34 #include "llvm/Support/MathExtras.h" 35 using namespace llvm; 36 using namespace llvm::PatternMatch; 37 38 //===----------------------------------------------------------------------===// 39 // ConstantFold*Instruction Implementations 40 //===----------------------------------------------------------------------===// 41 42 /// Convert the specified vector Constant node to the specified vector type. 43 /// At this point, we know that the elements of the input vector constant are 44 /// all simple integer or FP values. 45 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 46 47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 48 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 49 50 // If this cast changes element count then we can't handle it here: 51 // doing so requires endianness information. This should be handled by 52 // Analysis/ConstantFolding.cpp 53 unsigned NumElts = DstTy->getNumElements(); 54 if (NumElts != CV->getType()->getVectorNumElements()) 55 return nullptr; 56 57 Type *DstEltTy = DstTy->getElementType(); 58 59 SmallVector<Constant*, 16> Result; 60 Type *Ty = IntegerType::get(CV->getContext(), 32); 61 for (unsigned i = 0; i != NumElts; ++i) { 62 Constant *C = 63 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 64 C = ConstantExpr::getBitCast(C, DstEltTy); 65 Result.push_back(C); 66 } 67 68 return ConstantVector::get(Result); 69 } 70 71 /// This function determines which opcode to use to fold two constant cast 72 /// expressions together. It uses CastInst::isEliminableCastPair to determine 73 /// the opcode. Consequently its just a wrapper around that function. 74 /// @brief Determine if it is valid to fold a cast of a cast 75 static unsigned 76 foldConstantCastPair( 77 unsigned opc, ///< opcode of the second cast constant expression 78 ConstantExpr *Op, ///< the first cast constant expression 79 Type *DstTy ///< destination type of the first cast 80 ) { 81 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 82 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 83 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 84 85 // The types and opcodes for the two Cast constant expressions 86 Type *SrcTy = Op->getOperand(0)->getType(); 87 Type *MidTy = Op->getType(); 88 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 89 Instruction::CastOps secondOp = Instruction::CastOps(opc); 90 91 // Assume that pointers are never more than 64 bits wide, and only use this 92 // for the middle type. Otherwise we could end up folding away illegal 93 // bitcasts between address spaces with different sizes. 94 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 95 96 // Let CastInst::isEliminableCastPair do the heavy lifting. 97 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 98 nullptr, FakeIntPtrTy, nullptr); 99 } 100 101 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 102 Type *SrcTy = V->getType(); 103 if (SrcTy == DestTy) 104 return V; // no-op cast 105 106 // Check to see if we are casting a pointer to an aggregate to a pointer to 107 // the first element. If so, return the appropriate GEP instruction. 108 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 109 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 110 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 111 && PTy->getElementType()->isSized()) { 112 SmallVector<Value*, 8> IdxList; 113 Value *Zero = 114 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 115 IdxList.push_back(Zero); 116 Type *ElTy = PTy->getElementType(); 117 while (ElTy != DPTy->getElementType()) { 118 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 119 if (STy->getNumElements() == 0) break; 120 ElTy = STy->getElementType(0); 121 IdxList.push_back(Zero); 122 } else if (SequentialType *STy = 123 dyn_cast<SequentialType>(ElTy)) { 124 ElTy = STy->getElementType(); 125 IdxList.push_back(Zero); 126 } else { 127 break; 128 } 129 } 130 131 if (ElTy == DPTy->getElementType()) 132 // This GEP is inbounds because all indices are zero. 133 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), 134 V, IdxList); 135 } 136 137 // Handle casts from one vector constant to another. We know that the src 138 // and dest type have the same size (otherwise its an illegal cast). 139 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 140 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 141 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 142 "Not cast between same sized vectors!"); 143 SrcTy = nullptr; 144 // First, check for null. Undef is already handled. 145 if (isa<ConstantAggregateZero>(V)) 146 return Constant::getNullValue(DestTy); 147 148 // Handle ConstantVector and ConstantAggregateVector. 149 return BitCastConstantVector(V, DestPTy); 150 } 151 152 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 153 // This allows for other simplifications (although some of them 154 // can only be handled by Analysis/ConstantFolding.cpp). 155 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 156 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 157 } 158 159 // Finally, implement bitcast folding now. The code below doesn't handle 160 // bitcast right. 161 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 162 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 163 164 // Handle integral constant input. 165 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 166 if (DestTy->isIntegerTy()) 167 // Integral -> Integral. This is a no-op because the bit widths must 168 // be the same. Consequently, we just fold to V. 169 return V; 170 171 // See note below regarding the PPC_FP128 restriction. 172 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 173 return ConstantFP::get(DestTy->getContext(), 174 APFloat(DestTy->getFltSemantics(), 175 CI->getValue())); 176 177 // Otherwise, can't fold this (vector?) 178 return nullptr; 179 } 180 181 // Handle ConstantFP input: FP -> Integral. 182 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 183 // PPC_FP128 is really the sum of two consecutive doubles, where the first 184 // double is always stored first in memory, regardless of the target 185 // endianness. The memory layout of i128, however, depends on the target 186 // endianness, and so we can't fold this without target endianness 187 // information. This should instead be handled by 188 // Analysis/ConstantFolding.cpp 189 if (FP->getType()->isPPC_FP128Ty()) 190 return nullptr; 191 192 // Make sure dest type is compatible with the folded integer constant. 193 if (!DestTy->isIntegerTy()) 194 return nullptr; 195 196 return ConstantInt::get(FP->getContext(), 197 FP->getValueAPF().bitcastToAPInt()); 198 } 199 200 return nullptr; 201 } 202 203 204 /// V is an integer constant which only has a subset of its bytes used. 205 /// The bytes used are indicated by ByteStart (which is the first byte used, 206 /// counting from the least significant byte) and ByteSize, which is the number 207 /// of bytes used. 208 /// 209 /// This function analyzes the specified constant to see if the specified byte 210 /// range can be returned as a simplified constant. If so, the constant is 211 /// returned, otherwise null is returned. 212 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 213 unsigned ByteSize) { 214 assert(C->getType()->isIntegerTy() && 215 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 216 "Non-byte sized integer input"); 217 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 218 assert(ByteSize && "Must be accessing some piece"); 219 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 220 assert(ByteSize != CSize && "Should not extract everything"); 221 222 // Constant Integers are simple. 223 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 224 APInt V = CI->getValue(); 225 if (ByteStart) 226 V.lshrInPlace(ByteStart*8); 227 V = V.trunc(ByteSize*8); 228 return ConstantInt::get(CI->getContext(), V); 229 } 230 231 // In the input is a constant expr, we might be able to recursively simplify. 232 // If not, we definitely can't do anything. 233 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 234 if (!CE) return nullptr; 235 236 switch (CE->getOpcode()) { 237 default: return nullptr; 238 case Instruction::Or: { 239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 240 if (!RHS) 241 return nullptr; 242 243 // X | -1 -> -1. 244 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 245 if (RHSC->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, bool Folded) { 352 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 353 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 354 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 355 return ConstantExpr::getNUWMul(E, N); 356 } 357 358 if (StructType *STy = dyn_cast<StructType>(Ty)) 359 if (!STy->isPacked()) { 360 unsigned NumElems = STy->getNumElements(); 361 // An empty struct has size zero. 362 if (NumElems == 0) 363 return ConstantExpr::getNullValue(DestTy); 364 // Check for a struct with all members having the same size. 365 Constant *MemberSize = 366 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 367 bool AllSame = true; 368 for (unsigned i = 1; i != NumElems; ++i) 369 if (MemberSize != 370 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 371 AllSame = false; 372 break; 373 } 374 if (AllSame) { 375 Constant *N = ConstantInt::get(DestTy, NumElems); 376 return ConstantExpr::getNUWMul(MemberSize, N); 377 } 378 } 379 380 // Pointer size doesn't depend on the pointee type, so canonicalize them 381 // to an arbitrary pointee. 382 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 383 if (!PTy->getElementType()->isIntegerTy(1)) 384 return 385 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 386 PTy->getAddressSpace()), 387 DestTy, true); 388 389 // If there's no interesting folding happening, bail so that we don't create 390 // a constant that looks like it needs folding but really doesn't. 391 if (!Folded) 392 return nullptr; 393 394 // Base case: Get a regular sizeof expression. 395 Constant *C = ConstantExpr::getSizeOf(Ty); 396 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 397 DestTy, false), 398 C, DestTy); 399 return C; 400 } 401 402 /// Return a ConstantExpr with type DestTy for alignof on Ty, with any known 403 /// factors factored out. If Folded is false, return null if no factoring was 404 /// possible, to avoid endlessly bouncing an unfoldable expression back into the 405 /// top-level folder. 406 static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, bool Folded) { 407 // The alignment of an array is equal to the alignment of the 408 // array element. Note that this is not always true for vectors. 409 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 410 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 411 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 412 DestTy, 413 false), 414 C, DestTy); 415 return C; 416 } 417 418 if (StructType *STy = dyn_cast<StructType>(Ty)) { 419 // Packed structs always have an alignment of 1. 420 if (STy->isPacked()) 421 return ConstantInt::get(DestTy, 1); 422 423 // Otherwise, struct alignment is the maximum alignment of any member. 424 // Without target data, we can't compare much, but we can check to see 425 // if all the members have the same alignment. 426 unsigned NumElems = STy->getNumElements(); 427 // An empty struct has minimal alignment. 428 if (NumElems == 0) 429 return ConstantInt::get(DestTy, 1); 430 // Check for a struct with all members having the same alignment. 431 Constant *MemberAlign = 432 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 433 bool AllSame = true; 434 for (unsigned i = 1; i != NumElems; ++i) 435 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 436 AllSame = false; 437 break; 438 } 439 if (AllSame) 440 return MemberAlign; 441 } 442 443 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 444 // to an arbitrary pointee. 445 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 446 if (!PTy->getElementType()->isIntegerTy(1)) 447 return 448 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 449 1), 450 PTy->getAddressSpace()), 451 DestTy, true); 452 453 // If there's no interesting folding happening, bail so that we don't create 454 // a constant that looks like it needs folding but really doesn't. 455 if (!Folded) 456 return nullptr; 457 458 // Base case: Get a regular alignof expression. 459 Constant *C = ConstantExpr::getAlignOf(Ty); 460 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 461 DestTy, false), 462 C, DestTy); 463 return C; 464 } 465 466 /// Return a ConstantExpr with type DestTy for offsetof on Ty and FieldNo, with 467 /// any known factors factored out. If Folded is false, return null if no 468 /// factoring was possible, to avoid endlessly bouncing an unfoldable expression 469 /// back into the top-level folder. 470 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, Type *DestTy, 471 bool Folded) { 472 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 473 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 474 DestTy, false), 475 FieldNo, DestTy); 476 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 477 return ConstantExpr::getNUWMul(E, N); 478 } 479 480 if (StructType *STy = dyn_cast<StructType>(Ty)) 481 if (!STy->isPacked()) { 482 unsigned NumElems = STy->getNumElements(); 483 // An empty struct has no members. 484 if (NumElems == 0) 485 return nullptr; 486 // Check for a struct with all members having the same size. 487 Constant *MemberSize = 488 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 489 bool AllSame = true; 490 for (unsigned i = 1; i != NumElems; ++i) 491 if (MemberSize != 492 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 493 AllSame = false; 494 break; 495 } 496 if (AllSame) { 497 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 498 false, 499 DestTy, 500 false), 501 FieldNo, DestTy); 502 return ConstantExpr::getNUWMul(MemberSize, N); 503 } 504 } 505 506 // If there's no interesting folding happening, bail so that we don't create 507 // a constant that looks like it needs folding but really doesn't. 508 if (!Folded) 509 return nullptr; 510 511 // Base case: Get a regular offsetof expression. 512 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 513 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 514 DestTy, false), 515 C, DestTy); 516 return C; 517 } 518 519 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 520 Type *DestTy) { 521 if (isa<UndefValue>(V)) { 522 // zext(undef) = 0, because the top bits will be zero. 523 // sext(undef) = 0, because the top bits will all be the same. 524 // [us]itofp(undef) = 0, because the result value is bounded. 525 if (opc == Instruction::ZExt || opc == Instruction::SExt || 526 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 527 return Constant::getNullValue(DestTy); 528 return UndefValue::get(DestTy); 529 } 530 531 if (V->isNullValue() && !DestTy->isX86_MMXTy() && 532 opc != Instruction::AddrSpaceCast) 533 return Constant::getNullValue(DestTy); 534 535 // If the cast operand is a constant expression, there's a few things we can 536 // do to try to simplify it. 537 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 538 if (CE->isCast()) { 539 // Try hard to fold cast of cast because they are often eliminable. 540 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 541 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 542 } else if (CE->getOpcode() == Instruction::GetElementPtr && 543 // Do not fold addrspacecast (gep 0, .., 0). It might make the 544 // addrspacecast uncanonicalized. 545 opc != Instruction::AddrSpaceCast && 546 // Do not fold bitcast (gep) with inrange index, as this loses 547 // information. 548 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue()) { 549 // If all of the indexes in the GEP are null values, there is no pointer 550 // adjustment going on. We might as well cast the source pointer. 551 bool isAllNull = true; 552 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 553 if (!CE->getOperand(i)->isNullValue()) { 554 isAllNull = false; 555 break; 556 } 557 if (isAllNull) 558 // This is casting one pointer type to another, always BitCast 559 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 560 } 561 } 562 563 // If the cast operand is a constant vector, perform the cast by 564 // operating on each element. In the cast of bitcasts, the element 565 // count may be mismatched; don't attempt to handle that here. 566 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 567 DestTy->isVectorTy() && 568 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 569 SmallVector<Constant*, 16> res; 570 VectorType *DestVecTy = cast<VectorType>(DestTy); 571 Type *DstEltTy = DestVecTy->getElementType(); 572 Type *Ty = IntegerType::get(V->getContext(), 32); 573 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 574 Constant *C = 575 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 576 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 577 } 578 return ConstantVector::get(res); 579 } 580 581 // We actually have to do a cast now. Perform the cast according to the 582 // opcode specified. 583 switch (opc) { 584 default: 585 llvm_unreachable("Failed to cast constant expression"); 586 case Instruction::FPTrunc: 587 case Instruction::FPExt: 588 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 589 bool ignored; 590 APFloat Val = FPC->getValueAPF(); 591 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() : 592 DestTy->isFloatTy() ? APFloat::IEEEsingle() : 593 DestTy->isDoubleTy() ? APFloat::IEEEdouble() : 594 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() : 595 DestTy->isFP128Ty() ? APFloat::IEEEquad() : 596 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() : 597 APFloat::Bogus(), 598 APFloat::rmNearestTiesToEven, &ignored); 599 return ConstantFP::get(V->getContext(), Val); 600 } 601 return nullptr; // Can't fold. 602 case Instruction::FPToUI: 603 case Instruction::FPToSI: 604 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 605 const APFloat &V = FPC->getValueAPF(); 606 bool ignored; 607 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 608 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI); 609 if (APFloat::opInvalidOp == 610 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) { 611 // Undefined behavior invoked - the destination type can't represent 612 // the input constant. 613 return UndefValue::get(DestTy); 614 } 615 return ConstantInt::get(FPC->getContext(), IntVal); 616 } 617 return nullptr; // Can't fold. 618 case Instruction::IntToPtr: //always treated as unsigned 619 if (V->isNullValue()) // Is it an integral null value? 620 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 621 return nullptr; // Other pointer types cannot be casted 622 case Instruction::PtrToInt: // always treated as unsigned 623 // Is it a null pointer value? 624 if (V->isNullValue()) 625 return ConstantInt::get(DestTy, 0); 626 // If this is a sizeof-like expression, pull out multiplications by 627 // known factors to expose them to subsequent folding. If it's an 628 // alignof-like expression, factor out known factors. 629 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 630 if (CE->getOpcode() == Instruction::GetElementPtr && 631 CE->getOperand(0)->isNullValue()) { 632 GEPOperator *GEPO = cast<GEPOperator>(CE); 633 Type *Ty = GEPO->getSourceElementType(); 634 if (CE->getNumOperands() == 2) { 635 // Handle a sizeof-like expression. 636 Constant *Idx = CE->getOperand(1); 637 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 638 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 639 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 640 DestTy, false), 641 Idx, DestTy); 642 return ConstantExpr::getMul(C, Idx); 643 } 644 } else if (CE->getNumOperands() == 3 && 645 CE->getOperand(1)->isNullValue()) { 646 // Handle an alignof-like expression. 647 if (StructType *STy = dyn_cast<StructType>(Ty)) 648 if (!STy->isPacked()) { 649 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 650 if (CI->isOne() && 651 STy->getNumElements() == 2 && 652 STy->getElementType(0)->isIntegerTy(1)) { 653 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 654 } 655 } 656 // Handle an offsetof-like expression. 657 if (Ty->isStructTy() || Ty->isArrayTy()) { 658 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 659 DestTy, false)) 660 return C; 661 } 662 } 663 } 664 // Other pointer types cannot be casted 665 return nullptr; 666 case Instruction::UIToFP: 667 case Instruction::SIToFP: 668 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 669 const APInt &api = CI->getValue(); 670 APFloat apf(DestTy->getFltSemantics(), 671 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 672 if (APFloat::opOverflow & 673 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 674 APFloat::rmNearestTiesToEven)) { 675 // Undefined behavior invoked - the destination type can't represent 676 // the input constant. 677 return UndefValue::get(DestTy); 678 } 679 return ConstantFP::get(V->getContext(), apf); 680 } 681 return nullptr; 682 case Instruction::ZExt: 683 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 684 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 685 return ConstantInt::get(V->getContext(), 686 CI->getValue().zext(BitWidth)); 687 } 688 return nullptr; 689 case Instruction::SExt: 690 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 691 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 692 return ConstantInt::get(V->getContext(), 693 CI->getValue().sext(BitWidth)); 694 } 695 return nullptr; 696 case Instruction::Trunc: { 697 if (V->getType()->isVectorTy()) 698 return nullptr; 699 700 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 701 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 702 return ConstantInt::get(V->getContext(), 703 CI->getValue().trunc(DestBitWidth)); 704 } 705 706 // The input must be a constantexpr. See if we can simplify this based on 707 // the bytes we are demanding. Only do this if the source and dest are an 708 // even multiple of a byte. 709 if ((DestBitWidth & 7) == 0 && 710 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 711 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 712 return Res; 713 714 return nullptr; 715 } 716 case Instruction::BitCast: 717 return FoldBitCast(V, DestTy); 718 case Instruction::AddrSpaceCast: 719 return nullptr; 720 } 721 } 722 723 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 724 Constant *V1, Constant *V2) { 725 // Check for i1 and vector true/false conditions. 726 if (Cond->isNullValue()) return V2; 727 if (Cond->isAllOnesValue()) return V1; 728 729 // If the condition is a vector constant, fold the result elementwise. 730 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 731 SmallVector<Constant*, 16> Result; 732 Type *Ty = IntegerType::get(CondV->getContext(), 32); 733 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 734 Constant *V; 735 Constant *V1Element = ConstantExpr::getExtractElement(V1, 736 ConstantInt::get(Ty, i)); 737 Constant *V2Element = ConstantExpr::getExtractElement(V2, 738 ConstantInt::get(Ty, i)); 739 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 740 if (V1Element == V2Element) { 741 V = V1Element; 742 } else if (isa<UndefValue>(Cond)) { 743 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 744 } else { 745 if (!isa<ConstantInt>(Cond)) break; 746 V = Cond->isNullValue() ? V2Element : V1Element; 747 } 748 Result.push_back(V); 749 } 750 751 // If we were able to build the vector, return it. 752 if (Result.size() == V1->getType()->getVectorNumElements()) 753 return ConstantVector::get(Result); 754 } 755 756 if (isa<UndefValue>(Cond)) { 757 if (isa<UndefValue>(V1)) return V1; 758 return V2; 759 } 760 if (isa<UndefValue>(V1)) return V2; 761 if (isa<UndefValue>(V2)) return V1; 762 if (V1 == V2) return V1; 763 764 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 765 if (TrueVal->getOpcode() == Instruction::Select) 766 if (TrueVal->getOperand(0) == Cond) 767 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 768 } 769 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 770 if (FalseVal->getOpcode() == Instruction::Select) 771 if (FalseVal->getOperand(0) == Cond) 772 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 773 } 774 775 return nullptr; 776 } 777 778 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 779 Constant *Idx) { 780 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 781 return UndefValue::get(Val->getType()->getVectorElementType()); 782 if (Val->isNullValue()) // ee(zero, x) -> zero 783 return Constant::getNullValue(Val->getType()->getVectorElementType()); 784 // ee({w,x,y,z}, undef) -> undef 785 if (isa<UndefValue>(Idx)) 786 return UndefValue::get(Val->getType()->getVectorElementType()); 787 788 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 789 // ee({w,x,y,z}, wrong_value) -> undef 790 if (CIdx->uge(Val->getType()->getVectorNumElements())) 791 return UndefValue::get(Val->getType()->getVectorElementType()); 792 return Val->getAggregateElement(CIdx->getZExtValue()); 793 } 794 return nullptr; 795 } 796 797 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 798 Constant *Elt, 799 Constant *Idx) { 800 if (isa<UndefValue>(Idx)) 801 return UndefValue::get(Val->getType()); 802 803 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 804 if (!CIdx) return nullptr; 805 806 unsigned NumElts = Val->getType()->getVectorNumElements(); 807 if (CIdx->uge(NumElts)) 808 return UndefValue::get(Val->getType()); 809 810 SmallVector<Constant*, 16> Result; 811 Result.reserve(NumElts); 812 auto *Ty = Type::getInt32Ty(Val->getContext()); 813 uint64_t IdxVal = CIdx->getZExtValue(); 814 for (unsigned i = 0; i != NumElts; ++i) { 815 if (i == IdxVal) { 816 Result.push_back(Elt); 817 continue; 818 } 819 820 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 821 Result.push_back(C); 822 } 823 824 return ConstantVector::get(Result); 825 } 826 827 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 828 Constant *V2, 829 Constant *Mask) { 830 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 831 Type *EltTy = V1->getType()->getVectorElementType(); 832 833 // Undefined shuffle mask -> undefined value. 834 if (isa<UndefValue>(Mask)) 835 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 836 837 // Don't break the bitcode reader hack. 838 if (isa<ConstantExpr>(Mask)) return nullptr; 839 840 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 841 842 // Loop over the shuffle mask, evaluating each element. 843 SmallVector<Constant*, 32> Result; 844 for (unsigned i = 0; i != MaskNumElts; ++i) { 845 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 846 if (Elt == -1) { 847 Result.push_back(UndefValue::get(EltTy)); 848 continue; 849 } 850 Constant *InElt; 851 if (unsigned(Elt) >= SrcNumElts*2) 852 InElt = UndefValue::get(EltTy); 853 else if (unsigned(Elt) >= SrcNumElts) { 854 Type *Ty = IntegerType::get(V2->getContext(), 32); 855 InElt = 856 ConstantExpr::getExtractElement(V2, 857 ConstantInt::get(Ty, Elt - SrcNumElts)); 858 } else { 859 Type *Ty = IntegerType::get(V1->getContext(), 32); 860 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 861 } 862 Result.push_back(InElt); 863 } 864 865 return ConstantVector::get(Result); 866 } 867 868 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 869 ArrayRef<unsigned> Idxs) { 870 // Base case: no indices, so return the entire value. 871 if (Idxs.empty()) 872 return Agg; 873 874 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 875 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 876 877 return nullptr; 878 } 879 880 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 881 Constant *Val, 882 ArrayRef<unsigned> Idxs) { 883 // Base case: no indices, so replace the entire value. 884 if (Idxs.empty()) 885 return Val; 886 887 unsigned NumElts; 888 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 889 NumElts = ST->getNumElements(); 890 else 891 NumElts = cast<SequentialType>(Agg->getType())->getNumElements(); 892 893 SmallVector<Constant*, 32> Result; 894 for (unsigned i = 0; i != NumElts; ++i) { 895 Constant *C = Agg->getAggregateElement(i); 896 if (!C) return nullptr; 897 898 if (Idxs[0] == i) 899 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 900 901 Result.push_back(C); 902 } 903 904 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 905 return ConstantStruct::get(ST, Result); 906 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 907 return ConstantArray::get(AT, Result); 908 return ConstantVector::get(Result); 909 } 910 911 912 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 913 Constant *C1, Constant *C2) { 914 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 915 916 // Handle UndefValue up front. 917 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 918 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 919 case Instruction::Xor: 920 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 921 // Handle undef ^ undef -> 0 special case. This is a common 922 // idiom (misuse). 923 return Constant::getNullValue(C1->getType()); 924 LLVM_FALLTHROUGH; 925 case Instruction::Add: 926 case Instruction::Sub: 927 return UndefValue::get(C1->getType()); 928 case Instruction::And: 929 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 930 return C1; 931 return Constant::getNullValue(C1->getType()); // undef & X -> 0 932 case Instruction::Mul: { 933 // undef * undef -> undef 934 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 935 return C1; 936 const APInt *CV; 937 // X * undef -> undef if X is odd 938 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 939 if ((*CV)[0]) 940 return UndefValue::get(C1->getType()); 941 942 // X * undef -> 0 otherwise 943 return Constant::getNullValue(C1->getType()); 944 } 945 case Instruction::SDiv: 946 case Instruction::UDiv: 947 // X / undef -> undef 948 if (isa<UndefValue>(C2)) 949 return C2; 950 // undef / 0 -> undef 951 // undef / 1 -> undef 952 if (match(C2, m_Zero()) || match(C2, m_One())) 953 return C1; 954 // undef / X -> 0 otherwise 955 return Constant::getNullValue(C1->getType()); 956 case Instruction::URem: 957 case Instruction::SRem: 958 // X % undef -> undef 959 if (match(C2, m_Undef())) 960 return C2; 961 // undef % 0 -> undef 962 if (match(C2, m_Zero())) 963 return C1; 964 // undef % X -> 0 otherwise 965 return Constant::getNullValue(C1->getType()); 966 case Instruction::Or: // X | undef -> -1 967 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 968 return C1; 969 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 970 case Instruction::LShr: 971 // X >>l undef -> undef 972 if (isa<UndefValue>(C2)) 973 return C2; 974 // undef >>l 0 -> undef 975 if (match(C2, m_Zero())) 976 return C1; 977 // undef >>l X -> 0 978 return Constant::getNullValue(C1->getType()); 979 case Instruction::AShr: 980 // X >>a undef -> undef 981 if (isa<UndefValue>(C2)) 982 return C2; 983 // undef >>a 0 -> undef 984 if (match(C2, m_Zero())) 985 return C1; 986 // TODO: undef >>a X -> undef if the shift is exact 987 // undef >>a X -> 0 988 return Constant::getNullValue(C1->getType()); 989 case Instruction::Shl: 990 // X << undef -> undef 991 if (isa<UndefValue>(C2)) 992 return C2; 993 // undef << 0 -> undef 994 if (match(C2, m_Zero())) 995 return C1; 996 // undef << X -> 0 997 return Constant::getNullValue(C1->getType()); 998 case Instruction::FAdd: 999 case Instruction::FSub: 1000 case Instruction::FMul: 1001 case Instruction::FDiv: 1002 case Instruction::FRem: 1003 // TODO: UNDEF handling for binary float instructions. 1004 return nullptr; 1005 case Instruction::BinaryOpsEnd: 1006 llvm_unreachable("Invalid BinaryOp"); 1007 } 1008 } 1009 1010 // At this point neither constant should be an UndefValue. 1011 assert(!isa<UndefValue>(C1) && !isa<UndefValue>(C2) && 1012 "Unexpected UndefValue"); 1013 1014 // Handle simplifications when the RHS is a constant int. 1015 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1016 switch (Opcode) { 1017 case Instruction::Add: 1018 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1019 break; 1020 case Instruction::Sub: 1021 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1022 break; 1023 case Instruction::Mul: 1024 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1025 if (CI2->equalsInt(1)) 1026 return C1; // X * 1 == X 1027 break; 1028 case Instruction::UDiv: 1029 case Instruction::SDiv: 1030 if (CI2->equalsInt(1)) 1031 return C1; // X / 1 == X 1032 if (CI2->equalsInt(0)) 1033 return UndefValue::get(CI2->getType()); // X / 0 == undef 1034 break; 1035 case Instruction::URem: 1036 case Instruction::SRem: 1037 if (CI2->equalsInt(1)) 1038 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1039 if (CI2->equalsInt(0)) 1040 return UndefValue::get(CI2->getType()); // X % 0 == undef 1041 break; 1042 case Instruction::And: 1043 if (CI2->isZero()) return C2; // X & 0 == 0 1044 if (CI2->isAllOnesValue()) 1045 return C1; // X & -1 == X 1046 1047 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1048 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1049 if (CE1->getOpcode() == Instruction::ZExt) { 1050 unsigned DstWidth = CI2->getType()->getBitWidth(); 1051 unsigned SrcWidth = 1052 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1053 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1054 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1055 return C1; 1056 } 1057 1058 // If and'ing the address of a global with a constant, fold it. 1059 if (CE1->getOpcode() == Instruction::PtrToInt && 1060 isa<GlobalValue>(CE1->getOperand(0))) { 1061 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1062 1063 // Functions are at least 4-byte aligned. 1064 unsigned GVAlign = GV->getAlignment(); 1065 if (isa<Function>(GV)) 1066 GVAlign = std::max(GVAlign, 4U); 1067 1068 if (GVAlign > 1) { 1069 unsigned DstWidth = CI2->getType()->getBitWidth(); 1070 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1071 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1072 1073 // If checking bits we know are clear, return zero. 1074 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1075 return Constant::getNullValue(CI2->getType()); 1076 } 1077 } 1078 } 1079 break; 1080 case Instruction::Or: 1081 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1082 if (CI2->isAllOnesValue()) 1083 return C2; // X | -1 == -1 1084 break; 1085 case Instruction::Xor: 1086 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1087 1088 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1089 switch (CE1->getOpcode()) { 1090 default: break; 1091 case Instruction::ICmp: 1092 case Instruction::FCmp: 1093 // cmp pred ^ true -> cmp !pred 1094 assert(CI2->equalsInt(1)); 1095 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1096 pred = CmpInst::getInversePredicate(pred); 1097 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1098 CE1->getOperand(1)); 1099 } 1100 } 1101 break; 1102 case Instruction::AShr: 1103 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1104 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1105 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1106 return ConstantExpr::getLShr(C1, C2); 1107 break; 1108 } 1109 } else if (isa<ConstantInt>(C1)) { 1110 // If C1 is a ConstantInt and C2 is not, swap the operands. 1111 if (Instruction::isCommutative(Opcode)) 1112 return ConstantExpr::get(Opcode, C2, C1); 1113 } 1114 1115 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1116 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1117 const APInt &C1V = CI1->getValue(); 1118 const APInt &C2V = CI2->getValue(); 1119 switch (Opcode) { 1120 default: 1121 break; 1122 case Instruction::Add: 1123 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1124 case Instruction::Sub: 1125 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1126 case Instruction::Mul: 1127 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1128 case Instruction::UDiv: 1129 assert(!CI2->isNullValue() && "Div by zero handled above"); 1130 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1131 case Instruction::SDiv: 1132 assert(!CI2->isNullValue() && "Div by zero handled above"); 1133 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1134 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1135 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1136 case Instruction::URem: 1137 assert(!CI2->isNullValue() && "Div by zero handled above"); 1138 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1139 case Instruction::SRem: 1140 assert(!CI2->isNullValue() && "Div by zero handled above"); 1141 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1142 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1143 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1144 case Instruction::And: 1145 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1146 case Instruction::Or: 1147 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1148 case Instruction::Xor: 1149 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1150 case Instruction::Shl: 1151 if (C2V.ult(C1V.getBitWidth())) 1152 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1153 return UndefValue::get(C1->getType()); // too big shift is undef 1154 case Instruction::LShr: 1155 if (C2V.ult(C1V.getBitWidth())) 1156 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1157 return UndefValue::get(C1->getType()); // too big shift is undef 1158 case Instruction::AShr: 1159 if (C2V.ult(C1V.getBitWidth())) 1160 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1161 return UndefValue::get(C1->getType()); // too big shift is undef 1162 } 1163 } 1164 1165 switch (Opcode) { 1166 case Instruction::SDiv: 1167 case Instruction::UDiv: 1168 case Instruction::URem: 1169 case Instruction::SRem: 1170 case Instruction::LShr: 1171 case Instruction::AShr: 1172 case Instruction::Shl: 1173 if (CI1->equalsInt(0)) return C1; 1174 break; 1175 default: 1176 break; 1177 } 1178 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1179 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1180 const APFloat &C1V = CFP1->getValueAPF(); 1181 const APFloat &C2V = CFP2->getValueAPF(); 1182 APFloat C3V = C1V; // copy for modification 1183 switch (Opcode) { 1184 default: 1185 break; 1186 case Instruction::FAdd: 1187 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1188 return ConstantFP::get(C1->getContext(), C3V); 1189 case Instruction::FSub: 1190 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1191 return ConstantFP::get(C1->getContext(), C3V); 1192 case Instruction::FMul: 1193 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1194 return ConstantFP::get(C1->getContext(), C3V); 1195 case Instruction::FDiv: 1196 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1197 return ConstantFP::get(C1->getContext(), C3V); 1198 case Instruction::FRem: 1199 (void)C3V.mod(C2V); 1200 return ConstantFP::get(C1->getContext(), C3V); 1201 } 1202 } 1203 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1204 // Perform elementwise folding. 1205 SmallVector<Constant*, 16> Result; 1206 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1207 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1208 Constant *ExtractIdx = ConstantInt::get(Ty, i); 1209 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx); 1210 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx); 1211 1212 // If any element of a divisor vector is zero, the whole op is undef. 1213 if ((Opcode == Instruction::SDiv || Opcode == Instruction::UDiv || 1214 Opcode == Instruction::SRem || Opcode == Instruction::URem) && 1215 RHS->isNullValue()) 1216 return UndefValue::get(VTy); 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 /// This type is zero-sized if it's an array or structure of zero-sized types. 1277 /// 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 /// Compare the two constants as though they were getelementptr indices. 1294 /// 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 /// This function determines if there is anything we can decide about the two 1334 /// constants provided. This doesn't need to handle simple things like 1335 /// 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 /// This function determines if there is anything we can decide about the two 1421 /// constants provided. This doesn't need to handle simple things like integer 1422 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1423 /// If we can determine that the two constants have a particular relation to 1424 /// each other, we should return the corresponding ICmp predicate, otherwise 1425 /// 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 // We can't evaluate floating point casts or truncations. 1535 if (CE1Op0->getType()->isFloatingPointTy()) 1536 break; 1537 1538 // If the cast is not actually changing bits, and the second operand is a 1539 // null pointer, do the comparison with the pre-casted value. 1540 if (V2->isNullValue() && 1541 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1542 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1543 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1544 return evaluateICmpRelation(CE1Op0, 1545 Constant::getNullValue(CE1Op0->getType()), 1546 isSigned); 1547 } 1548 break; 1549 1550 case Instruction::GetElementPtr: { 1551 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1552 // Ok, since this is a getelementptr, we know that the constant has a 1553 // pointer type. Check the various cases. 1554 if (isa<ConstantPointerNull>(V2)) { 1555 // If we are comparing a GEP to a null pointer, check to see if the base 1556 // of the GEP equals the null pointer. 1557 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1558 if (GV->hasExternalWeakLinkage()) 1559 // Weak linkage GVals could be zero or not. We're comparing that 1560 // to null pointer so its greater-or-equal 1561 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1562 else 1563 // If its not weak linkage, the GVal must have a non-zero address 1564 // so the result is greater-than 1565 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1566 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1567 // If we are indexing from a null pointer, check to see if we have any 1568 // non-zero indices. 1569 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1570 if (!CE1->getOperand(i)->isNullValue()) 1571 // Offsetting from null, must not be equal. 1572 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1573 // Only zero indexes from null, must still be zero. 1574 return ICmpInst::ICMP_EQ; 1575 } 1576 // Otherwise, we can't really say if the first operand is null or not. 1577 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1578 if (isa<ConstantPointerNull>(CE1Op0)) { 1579 if (GV2->hasExternalWeakLinkage()) 1580 // Weak linkage GVals could be zero or not. We're comparing it to 1581 // a null pointer, so its less-or-equal 1582 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1583 else 1584 // If its not weak linkage, the GVal must have a non-zero address 1585 // so the result is less-than 1586 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1587 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1588 if (GV == GV2) { 1589 // If this is a getelementptr of the same global, then it must be 1590 // different. Because the types must match, the getelementptr could 1591 // only have at most one index, and because we fold getelementptr's 1592 // with a single zero index, it must be nonzero. 1593 assert(CE1->getNumOperands() == 2 && 1594 !CE1->getOperand(1)->isNullValue() && 1595 "Surprising getelementptr!"); 1596 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1597 } else { 1598 if (CE1GEP->hasAllZeroIndices()) 1599 return areGlobalsPotentiallyEqual(GV, GV2); 1600 return ICmpInst::BAD_ICMP_PREDICATE; 1601 } 1602 } 1603 } else { 1604 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1605 Constant *CE2Op0 = CE2->getOperand(0); 1606 1607 // There are MANY other foldings that we could perform here. They will 1608 // probably be added on demand, as they seem needed. 1609 switch (CE2->getOpcode()) { 1610 default: break; 1611 case Instruction::GetElementPtr: 1612 // By far the most common case to handle is when the base pointers are 1613 // obviously to the same global. 1614 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1615 // Don't know relative ordering, but check for inequality. 1616 if (CE1Op0 != CE2Op0) { 1617 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1618 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1619 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1620 cast<GlobalValue>(CE2Op0)); 1621 return ICmpInst::BAD_ICMP_PREDICATE; 1622 } 1623 // Ok, we know that both getelementptr instructions are based on the 1624 // same global. From this, we can precisely determine the relative 1625 // ordering of the resultant pointers. 1626 unsigned i = 1; 1627 1628 // The logic below assumes that the result of the comparison 1629 // can be determined by finding the first index that differs. 1630 // This doesn't work if there is over-indexing in any 1631 // subsequent indices, so check for that case first. 1632 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1633 !CE2->isGEPWithNoNotionalOverIndexing()) 1634 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1635 1636 // Compare all of the operands the GEP's have in common. 1637 gep_type_iterator GTI = gep_type_begin(CE1); 1638 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1639 ++i, ++GTI) 1640 switch (IdxCompare(CE1->getOperand(i), 1641 CE2->getOperand(i), GTI.getIndexedType())) { 1642 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1643 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1644 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1645 } 1646 1647 // Ok, we ran out of things they have in common. If any leftovers 1648 // are non-zero then we have a difference, otherwise we are equal. 1649 for (; i < CE1->getNumOperands(); ++i) 1650 if (!CE1->getOperand(i)->isNullValue()) { 1651 if (isa<ConstantInt>(CE1->getOperand(i))) 1652 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1653 else 1654 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1655 } 1656 1657 for (; i < CE2->getNumOperands(); ++i) 1658 if (!CE2->getOperand(i)->isNullValue()) { 1659 if (isa<ConstantInt>(CE2->getOperand(i))) 1660 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1661 else 1662 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1663 } 1664 return ICmpInst::ICMP_EQ; 1665 } 1666 } 1667 } 1668 } 1669 default: 1670 break; 1671 } 1672 } 1673 1674 return ICmpInst::BAD_ICMP_PREDICATE; 1675 } 1676 1677 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1678 Constant *C1, Constant *C2) { 1679 Type *ResultTy; 1680 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1681 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1682 VT->getNumElements()); 1683 else 1684 ResultTy = Type::getInt1Ty(C1->getContext()); 1685 1686 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1687 if (pred == FCmpInst::FCMP_FALSE) 1688 return Constant::getNullValue(ResultTy); 1689 1690 if (pred == FCmpInst::FCMP_TRUE) 1691 return Constant::getAllOnesValue(ResultTy); 1692 1693 // Handle some degenerate cases first 1694 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1695 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1696 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1697 // For EQ and NE, we can always pick a value for the undef to make the 1698 // predicate pass or fail, so we can return undef. 1699 // Also, if both operands are undef, we can return undef for int comparison. 1700 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1701 return UndefValue::get(ResultTy); 1702 1703 // Otherwise, for integer compare, pick the same value as the non-undef 1704 // operand, and fold it to true or false. 1705 if (isIntegerPredicate) 1706 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1707 1708 // Choosing NaN for the undef will always make unordered comparison succeed 1709 // and ordered comparison fails. 1710 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1711 } 1712 1713 // icmp eq/ne(null,GV) -> false/true 1714 if (C1->isNullValue()) { 1715 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1716 // Don't try to evaluate aliases. External weak GV can be null. 1717 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1718 if (pred == ICmpInst::ICMP_EQ) 1719 return ConstantInt::getFalse(C1->getContext()); 1720 else if (pred == ICmpInst::ICMP_NE) 1721 return ConstantInt::getTrue(C1->getContext()); 1722 } 1723 // icmp eq/ne(GV,null) -> false/true 1724 } else if (C2->isNullValue()) { 1725 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1726 // Don't try to evaluate aliases. External weak GV can be null. 1727 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1728 if (pred == ICmpInst::ICMP_EQ) 1729 return ConstantInt::getFalse(C1->getContext()); 1730 else if (pred == ICmpInst::ICMP_NE) 1731 return ConstantInt::getTrue(C1->getContext()); 1732 } 1733 } 1734 1735 // If the comparison is a comparison between two i1's, simplify it. 1736 if (C1->getType()->isIntegerTy(1)) { 1737 switch(pred) { 1738 case ICmpInst::ICMP_EQ: 1739 if (isa<ConstantInt>(C2)) 1740 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1741 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1742 case ICmpInst::ICMP_NE: 1743 return ConstantExpr::getXor(C1, C2); 1744 default: 1745 break; 1746 } 1747 } 1748 1749 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1750 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1751 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1752 switch (pred) { 1753 default: llvm_unreachable("Invalid ICmp Predicate"); 1754 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1755 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1756 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1757 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1758 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1759 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1760 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1761 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1762 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1763 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1764 } 1765 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1766 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1767 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1768 APFloat::cmpResult R = C1V.compare(C2V); 1769 switch (pred) { 1770 default: llvm_unreachable("Invalid FCmp Predicate"); 1771 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1772 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1773 case FCmpInst::FCMP_UNO: 1774 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1775 case FCmpInst::FCMP_ORD: 1776 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1777 case FCmpInst::FCMP_UEQ: 1778 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1779 R==APFloat::cmpEqual); 1780 case FCmpInst::FCMP_OEQ: 1781 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1782 case FCmpInst::FCMP_UNE: 1783 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1784 case FCmpInst::FCMP_ONE: 1785 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1786 R==APFloat::cmpGreaterThan); 1787 case FCmpInst::FCMP_ULT: 1788 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1789 R==APFloat::cmpLessThan); 1790 case FCmpInst::FCMP_OLT: 1791 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1792 case FCmpInst::FCMP_UGT: 1793 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1794 R==APFloat::cmpGreaterThan); 1795 case FCmpInst::FCMP_OGT: 1796 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1797 case FCmpInst::FCMP_ULE: 1798 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1799 case FCmpInst::FCMP_OLE: 1800 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1801 R==APFloat::cmpEqual); 1802 case FCmpInst::FCMP_UGE: 1803 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1804 case FCmpInst::FCMP_OGE: 1805 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1806 R==APFloat::cmpEqual); 1807 } 1808 } else if (C1->getType()->isVectorTy()) { 1809 // If we can constant fold the comparison of each element, constant fold 1810 // the whole vector comparison. 1811 SmallVector<Constant*, 4> ResElts; 1812 Type *Ty = IntegerType::get(C1->getContext(), 32); 1813 // Compare the elements, producing an i1 result or constant expr. 1814 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1815 Constant *C1E = 1816 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1817 Constant *C2E = 1818 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1819 1820 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1821 } 1822 1823 return ConstantVector::get(ResElts); 1824 } 1825 1826 if (C1->getType()->isFloatingPointTy() && 1827 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1828 // infinite recursive loop 1829 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1830 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1831 switch (evaluateFCmpRelation(C1, C2)) { 1832 default: llvm_unreachable("Unknown relation!"); 1833 case FCmpInst::FCMP_UNO: 1834 case FCmpInst::FCMP_ORD: 1835 case FCmpInst::FCMP_UEQ: 1836 case FCmpInst::FCMP_UNE: 1837 case FCmpInst::FCMP_ULT: 1838 case FCmpInst::FCMP_UGT: 1839 case FCmpInst::FCMP_ULE: 1840 case FCmpInst::FCMP_UGE: 1841 case FCmpInst::FCMP_TRUE: 1842 case FCmpInst::FCMP_FALSE: 1843 case FCmpInst::BAD_FCMP_PREDICATE: 1844 break; // Couldn't determine anything about these constants. 1845 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1846 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1847 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1848 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1849 break; 1850 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1851 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1852 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1853 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1854 break; 1855 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1856 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1857 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1858 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1859 break; 1860 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1861 // We can only partially decide this relation. 1862 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1863 Result = 0; 1864 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1865 Result = 1; 1866 break; 1867 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1868 // We can only partially decide this relation. 1869 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1870 Result = 0; 1871 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1872 Result = 1; 1873 break; 1874 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1875 // We can only partially decide this relation. 1876 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1877 Result = 0; 1878 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1879 Result = 1; 1880 break; 1881 } 1882 1883 // If we evaluated the result, return it now. 1884 if (Result != -1) 1885 return ConstantInt::get(ResultTy, Result); 1886 1887 } else { 1888 // Evaluate the relation between the two constants, per the predicate. 1889 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1890 switch (evaluateICmpRelation(C1, C2, 1891 CmpInst::isSigned((CmpInst::Predicate)pred))) { 1892 default: llvm_unreachable("Unknown relational!"); 1893 case ICmpInst::BAD_ICMP_PREDICATE: 1894 break; // Couldn't determine anything about these constants. 1895 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1896 // If we know the constants are equal, we can decide the result of this 1897 // computation precisely. 1898 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1899 break; 1900 case ICmpInst::ICMP_ULT: 1901 switch (pred) { 1902 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1903 Result = 1; break; 1904 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1905 Result = 0; break; 1906 } 1907 break; 1908 case ICmpInst::ICMP_SLT: 1909 switch (pred) { 1910 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1911 Result = 1; break; 1912 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1913 Result = 0; break; 1914 } 1915 break; 1916 case ICmpInst::ICMP_UGT: 1917 switch (pred) { 1918 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1919 Result = 1; break; 1920 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1921 Result = 0; break; 1922 } 1923 break; 1924 case ICmpInst::ICMP_SGT: 1925 switch (pred) { 1926 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1927 Result = 1; break; 1928 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1929 Result = 0; break; 1930 } 1931 break; 1932 case ICmpInst::ICMP_ULE: 1933 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1934 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1935 break; 1936 case ICmpInst::ICMP_SLE: 1937 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1938 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1939 break; 1940 case ICmpInst::ICMP_UGE: 1941 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1942 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1943 break; 1944 case ICmpInst::ICMP_SGE: 1945 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1946 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1947 break; 1948 case ICmpInst::ICMP_NE: 1949 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1950 if (pred == ICmpInst::ICMP_NE) Result = 1; 1951 break; 1952 } 1953 1954 // If we evaluated the result, return it now. 1955 if (Result != -1) 1956 return ConstantInt::get(ResultTy, Result); 1957 1958 // If the right hand side is a bitcast, try using its inverse to simplify 1959 // it by moving it to the left hand side. We can't do this if it would turn 1960 // a vector compare into a scalar compare or visa versa. 1961 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1962 Constant *CE2Op0 = CE2->getOperand(0); 1963 if (CE2->getOpcode() == Instruction::BitCast && 1964 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1965 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1966 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1967 } 1968 } 1969 1970 // If the left hand side is an extension, try eliminating it. 1971 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1972 if ((CE1->getOpcode() == Instruction::SExt && 1973 ICmpInst::isSigned((ICmpInst::Predicate)pred)) || 1974 (CE1->getOpcode() == Instruction::ZExt && 1975 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){ 1976 Constant *CE1Op0 = CE1->getOperand(0); 1977 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1978 if (CE1Inverse == CE1Op0) { 1979 // Check whether we can safely truncate the right hand side. 1980 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1981 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1982 C2->getType()) == C2) 1983 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1984 } 1985 } 1986 } 1987 1988 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1989 (C1->isNullValue() && !C2->isNullValue())) { 1990 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1991 // other way if possible. 1992 // Also, if C1 is null and C2 isn't, flip them around. 1993 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1994 return ConstantExpr::getICmp(pred, C2, C1); 1995 } 1996 } 1997 return nullptr; 1998 } 1999 2000 /// Test whether the given sequence of *normalized* indices is "inbounds". 2001 template<typename IndexTy> 2002 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 2003 // No indices means nothing that could be out of bounds. 2004 if (Idxs.empty()) return true; 2005 2006 // If the first index is zero, it's in bounds. 2007 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2008 2009 // If the first index is one and all the rest are zero, it's in bounds, 2010 // by the one-past-the-end rule. 2011 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2012 return false; 2013 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 2014 if (!cast<Constant>(Idxs[i])->isNullValue()) 2015 return false; 2016 return true; 2017 } 2018 2019 /// Test whether a given ConstantInt is in-range for a SequentialType. 2020 static bool isIndexInRangeOfArrayType(uint64_t NumElements, 2021 const ConstantInt *CI) { 2022 // We cannot bounds check the index if it doesn't fit in an int64_t. 2023 if (CI->getValue().getActiveBits() > 64) 2024 return false; 2025 2026 // A negative index or an index past the end of our sequential type is 2027 // considered out-of-range. 2028 int64_t IndexVal = CI->getSExtValue(); 2029 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2030 return false; 2031 2032 // Otherwise, it is in-range. 2033 return true; 2034 } 2035 2036 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, 2037 bool InBounds, 2038 Optional<unsigned> InRangeIndex, 2039 ArrayRef<Value *> Idxs) { 2040 if (Idxs.empty()) return C; 2041 2042 if (isa<UndefValue>(C)) { 2043 Type *GEPTy = GetElementPtrInst::getGEPReturnType( 2044 C, makeArrayRef((Value * const *)Idxs.data(), Idxs.size())); 2045 return UndefValue::get(GEPTy); 2046 } 2047 2048 Constant *Idx0 = cast<Constant>(Idxs[0]); 2049 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0))) 2050 return C; 2051 2052 if (C->isNullValue()) { 2053 bool isNull = true; 2054 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2055 if (!isa<UndefValue>(Idxs[i]) && 2056 !cast<Constant>(Idxs[i])->isNullValue()) { 2057 isNull = false; 2058 break; 2059 } 2060 if (isNull) { 2061 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2062 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2063 2064 assert(Ty && "Invalid indices for GEP!"); 2065 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2066 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2067 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 2068 return Constant::getNullValue(GEPTy); 2069 } 2070 } 2071 2072 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2073 // Combine Indices - If the source pointer to this getelementptr instruction 2074 // is a getelementptr instruction, combine the indices of the two 2075 // getelementptr instructions into a single instruction. 2076 // 2077 if (CE->getOpcode() == Instruction::GetElementPtr) { 2078 gep_type_iterator LastI = gep_type_end(CE); 2079 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2080 I != E; ++I) 2081 LastI = I; 2082 2083 // We cannot combine indices if doing so would take us outside of an 2084 // array or vector. Doing otherwise could trick us if we evaluated such a 2085 // GEP as part of a load. 2086 // 2087 // e.g. Consider if the original GEP was: 2088 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2089 // i32 0, i32 0, i64 0) 2090 // 2091 // If we then tried to offset it by '8' to get to the third element, 2092 // an i8, we should *not* get: 2093 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2094 // i32 0, i32 0, i64 8) 2095 // 2096 // This GEP tries to index array element '8 which runs out-of-bounds. 2097 // Subsequent evaluation would get confused and produce erroneous results. 2098 // 2099 // The following prohibits such a GEP from being formed by checking to see 2100 // if the index is in-range with respect to an array or vector. 2101 bool PerformFold = false; 2102 if (Idx0->isNullValue()) 2103 PerformFold = true; 2104 else if (LastI.isSequential()) 2105 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2106 PerformFold = 2107 !LastI.isBoundedSequential() || 2108 isIndexInRangeOfArrayType(LastI.getSequentialNumElements(), CI); 2109 2110 if (PerformFold) { 2111 SmallVector<Value*, 16> NewIndices; 2112 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2113 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2114 2115 // Add the last index of the source with the first index of the new GEP. 2116 // Make sure to handle the case when they are actually different types. 2117 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2118 // Otherwise it must be an array. 2119 if (!Idx0->isNullValue()) { 2120 Type *IdxTy = Combined->getType(); 2121 if (IdxTy != Idx0->getType()) { 2122 unsigned CommonExtendedWidth = 2123 std::max(IdxTy->getIntegerBitWidth(), 2124 Idx0->getType()->getIntegerBitWidth()); 2125 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2126 2127 Type *CommonTy = 2128 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2129 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2130 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2131 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2132 } else { 2133 Combined = 2134 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2135 } 2136 } 2137 2138 NewIndices.push_back(Combined); 2139 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2140 2141 // The combined GEP normally inherits its index inrange attribute from 2142 // the inner GEP, but if the inner GEP's last index was adjusted by the 2143 // outer GEP, any inbounds attribute on that index is invalidated. 2144 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex(); 2145 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue()) 2146 IRIndex = None; 2147 2148 return ConstantExpr::getGetElementPtr( 2149 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2150 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(), 2151 IRIndex); 2152 } 2153 } 2154 2155 // Attempt to fold casts to the same type away. For example, folding: 2156 // 2157 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2158 // i64 0, i64 0) 2159 // into: 2160 // 2161 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2162 // 2163 // Don't fold if the cast is changing address spaces. 2164 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2165 PointerType *SrcPtrTy = 2166 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2167 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2168 if (SrcPtrTy && DstPtrTy) { 2169 ArrayType *SrcArrayTy = 2170 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2171 ArrayType *DstArrayTy = 2172 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2173 if (SrcArrayTy && DstArrayTy 2174 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2175 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2176 return ConstantExpr::getGetElementPtr(SrcArrayTy, 2177 (Constant *)CE->getOperand(0), 2178 Idxs, InBounds, InRangeIndex); 2179 } 2180 } 2181 } 2182 2183 // Check to see if any array indices are not within the corresponding 2184 // notional array or vector bounds. If so, try to determine if they can be 2185 // factored out into preceding dimensions. 2186 SmallVector<Constant *, 8> NewIdxs; 2187 Type *Ty = PointeeTy; 2188 Type *Prev = C->getType(); 2189 bool Unknown = !isa<ConstantInt>(Idxs[0]); 2190 for (unsigned i = 1, e = Idxs.size(); i != e; 2191 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2192 auto *CI = dyn_cast<ConstantInt>(Idxs[i]); 2193 if (!CI) { 2194 // We don't know if it's in range or not. 2195 Unknown = true; 2196 continue; 2197 } 2198 if (InRangeIndex && i == *InRangeIndex + 1) { 2199 // If an index is marked inrange, we cannot apply this canonicalization to 2200 // the following index, as that will cause the inrange index to point to 2201 // the wrong element. 2202 continue; 2203 } 2204 if (isa<StructType>(Ty)) { 2205 // The verify makes sure that GEPs into a struct are in range. 2206 continue; 2207 } 2208 auto *STy = cast<SequentialType>(Ty); 2209 if (isa<VectorType>(STy)) { 2210 // There can be awkward padding in after a non-power of two vector. 2211 Unknown = true; 2212 continue; 2213 } 2214 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI)) 2215 // It's in range, skip to the next index. 2216 continue; 2217 if (isa<StructType>(Prev)) { 2218 // It's out of range, but the prior dimension is a struct 2219 // so we can't do anything about it. 2220 Unknown = true; 2221 continue; 2222 } 2223 if (CI->getSExtValue() < 0) { 2224 // It's out of range and negative, don't try to factor it. 2225 Unknown = true; 2226 continue; 2227 } 2228 // It's out of range, but we can factor it into the prior 2229 // dimension. 2230 NewIdxs.resize(Idxs.size()); 2231 // Determine the number of elements in our sequential type. 2232 uint64_t NumElements = STy->getArrayNumElements(); 2233 2234 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2235 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2236 2237 Constant *PrevIdx = NewIdxs[i-1] ? NewIdxs[i-1] : 2238 cast<Constant>(Idxs[i - 1]); 2239 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2240 2241 unsigned CommonExtendedWidth = 2242 std::max(PrevIdx->getType()->getIntegerBitWidth(), 2243 Div->getType()->getIntegerBitWidth()); 2244 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2245 2246 // Before adding, extend both operands to i64 to avoid 2247 // overflow trouble. 2248 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth)) 2249 PrevIdx = ConstantExpr::getSExt( 2250 PrevIdx, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2251 if (!Div->getType()->isIntegerTy(CommonExtendedWidth)) 2252 Div = ConstantExpr::getSExt( 2253 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2254 2255 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div); 2256 } 2257 2258 // If we did any factoring, start over with the adjusted indices. 2259 if (!NewIdxs.empty()) { 2260 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2261 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2262 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds, 2263 InRangeIndex); 2264 } 2265 2266 // If all indices are known integers and normalized, we can do a simple 2267 // check for the "inbounds" property. 2268 if (!Unknown && !InBounds) 2269 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2270 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2271 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs, 2272 /*InBounds=*/true, InRangeIndex); 2273 2274 return nullptr; 2275 } 2276