1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements folding of constants for LLVM. This implements the 11 // (internal) ConstantFold.h interface, which is used by the 12 // ConstantExpr::get* methods to automatically fold constants when possible. 13 // 14 // The current constant folding implementation is implemented in two pieces: the 15 // pieces that don't need DataLayout, and the pieces that do. This is to avoid 16 // a dependence in IR on Target. 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "ConstantFold.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/IR/Constants.h" 23 #include "llvm/IR/DerivedTypes.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/GetElementPtrTypeIterator.h" 26 #include "llvm/IR/GlobalAlias.h" 27 #include "llvm/IR/GlobalVariable.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/Operator.h" 30 #include "llvm/IR/PatternMatch.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/ManagedStatic.h" 33 #include "llvm/Support/MathExtras.h" 34 using namespace llvm; 35 using namespace llvm::PatternMatch; 36 37 //===----------------------------------------------------------------------===// 38 // ConstantFold*Instruction Implementations 39 //===----------------------------------------------------------------------===// 40 41 /// Convert the specified vector Constant node to the specified vector type. 42 /// At this point, we know that the elements of the input vector constant are 43 /// all simple integer or FP values. 44 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 45 46 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 47 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 48 49 // If this cast changes element count then we can't handle it here: 50 // doing so requires endianness information. This should be handled by 51 // Analysis/ConstantFolding.cpp 52 unsigned NumElts = DstTy->getNumElements(); 53 if (NumElts != CV->getType()->getVectorNumElements()) 54 return nullptr; 55 56 Type *DstEltTy = DstTy->getElementType(); 57 58 SmallVector<Constant*, 16> Result; 59 Type *Ty = IntegerType::get(CV->getContext(), 32); 60 for (unsigned i = 0; i != NumElts; ++i) { 61 Constant *C = 62 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 63 C = ConstantExpr::getBitCast(C, DstEltTy); 64 Result.push_back(C); 65 } 66 67 return ConstantVector::get(Result); 68 } 69 70 /// This function determines which opcode to use to fold two constant cast 71 /// expressions together. It uses CastInst::isEliminableCastPair to determine 72 /// the opcode. Consequently its just a wrapper around that function. 73 /// @brief Determine if it is valid to fold a cast of a cast 74 static unsigned 75 foldConstantCastPair( 76 unsigned opc, ///< opcode of the second cast constant expression 77 ConstantExpr *Op, ///< the first cast constant expression 78 Type *DstTy ///< destination type of the first cast 79 ) { 80 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 81 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 82 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 83 84 // The types and opcodes for the two Cast constant expressions 85 Type *SrcTy = Op->getOperand(0)->getType(); 86 Type *MidTy = Op->getType(); 87 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 88 Instruction::CastOps secondOp = Instruction::CastOps(opc); 89 90 // Assume that pointers are never more than 64 bits wide, and only use this 91 // for the middle type. Otherwise we could end up folding away illegal 92 // bitcasts between address spaces with different sizes. 93 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 94 95 // Let CastInst::isEliminableCastPair do the heavy lifting. 96 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 97 nullptr, FakeIntPtrTy, nullptr); 98 } 99 100 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 101 Type *SrcTy = V->getType(); 102 if (SrcTy == DestTy) 103 return V; // no-op cast 104 105 // Check to see if we are casting a pointer to an aggregate to a pointer to 106 // the first element. If so, return the appropriate GEP instruction. 107 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 108 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 109 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 110 && PTy->getElementType()->isSized()) { 111 SmallVector<Value*, 8> IdxList; 112 Value *Zero = 113 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 114 IdxList.push_back(Zero); 115 Type *ElTy = PTy->getElementType(); 116 while (ElTy != DPTy->getElementType()) { 117 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 118 if (STy->getNumElements() == 0) break; 119 ElTy = STy->getElementType(0); 120 IdxList.push_back(Zero); 121 } else if (SequentialType *STy = 122 dyn_cast<SequentialType>(ElTy)) { 123 ElTy = STy->getElementType(); 124 IdxList.push_back(Zero); 125 } else { 126 break; 127 } 128 } 129 130 if (ElTy == DPTy->getElementType()) 131 // This GEP is inbounds because all indices are zero. 132 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), 133 V, IdxList); 134 } 135 136 // Handle casts from one vector constant to another. We know that the src 137 // and dest type have the same size (otherwise its an illegal cast). 138 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 139 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 140 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 141 "Not cast between same sized vectors!"); 142 SrcTy = nullptr; 143 // First, check for null. Undef is already handled. 144 if (isa<ConstantAggregateZero>(V)) 145 return Constant::getNullValue(DestTy); 146 147 // Handle ConstantVector and ConstantAggregateVector. 148 return BitCastConstantVector(V, DestPTy); 149 } 150 151 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 152 // This allows for other simplifications (although some of them 153 // can only be handled by Analysis/ConstantFolding.cpp). 154 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 155 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 156 } 157 158 // Finally, implement bitcast folding now. The code below doesn't handle 159 // bitcast right. 160 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 161 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 162 163 // Handle integral constant input. 164 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 165 if (DestTy->isIntegerTy()) 166 // Integral -> Integral. This is a no-op because the bit widths must 167 // be the same. Consequently, we just fold to V. 168 return V; 169 170 // See note below regarding the PPC_FP128 restriction. 171 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 172 return ConstantFP::get(DestTy->getContext(), 173 APFloat(DestTy->getFltSemantics(), 174 CI->getValue())); 175 176 // Otherwise, can't fold this (vector?) 177 return nullptr; 178 } 179 180 // Handle ConstantFP input: FP -> Integral. 181 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 182 // PPC_FP128 is really the sum of two consecutive doubles, where the first 183 // double is always stored first in memory, regardless of the target 184 // endianness. The memory layout of i128, however, depends on the target 185 // endianness, and so we can't fold this without target endianness 186 // information. This should instead be handled by 187 // Analysis/ConstantFolding.cpp 188 if (FP->getType()->isPPC_FP128Ty()) 189 return nullptr; 190 191 // Make sure dest type is compatible with the folded integer constant. 192 if (!DestTy->isIntegerTy()) 193 return nullptr; 194 195 return ConstantInt::get(FP->getContext(), 196 FP->getValueAPF().bitcastToAPInt()); 197 } 198 199 return nullptr; 200 } 201 202 203 /// V is an integer constant which only has a subset of its bytes used. 204 /// The bytes used are indicated by ByteStart (which is the first byte used, 205 /// counting from the least significant byte) and ByteSize, which is the number 206 /// of bytes used. 207 /// 208 /// This function analyzes the specified constant to see if the specified byte 209 /// range can be returned as a simplified constant. If so, the constant is 210 /// returned, otherwise null is returned. 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 /// Return a ConstantExpr with type DestTy for sizeof on Ty, with any known 347 /// factors factored out. If Folded is false, return null if no factoring was 348 /// possible, to avoid endlessly bouncing an unfoldable expression back into the 349 /// top-level folder. 350 static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 351 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, 407 bool Folded) { 408 // The alignment of an array is equal to the alignment of the 409 // array element. Note that this is not always true for vectors. 410 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 411 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 412 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 413 DestTy, 414 false), 415 C, DestTy); 416 return C; 417 } 418 419 if (StructType *STy = dyn_cast<StructType>(Ty)) { 420 // Packed structs always have an alignment of 1. 421 if (STy->isPacked()) 422 return ConstantInt::get(DestTy, 1); 423 424 // Otherwise, struct alignment is the maximum alignment of any member. 425 // Without target data, we can't compare much, but we can check to see 426 // if all the members have the same alignment. 427 unsigned NumElems = STy->getNumElements(); 428 // An empty struct has minimal alignment. 429 if (NumElems == 0) 430 return ConstantInt::get(DestTy, 1); 431 // Check for a struct with all members having the same alignment. 432 Constant *MemberAlign = 433 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 434 bool AllSame = true; 435 for (unsigned i = 1; i != NumElems; ++i) 436 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 437 AllSame = false; 438 break; 439 } 440 if (AllSame) 441 return MemberAlign; 442 } 443 444 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 445 // to an arbitrary pointee. 446 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 447 if (!PTy->getElementType()->isIntegerTy(1)) 448 return 449 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 450 1), 451 PTy->getAddressSpace()), 452 DestTy, true); 453 454 // If there's no interesting folding happening, bail so that we don't create 455 // a constant that looks like it needs folding but really doesn't. 456 if (!Folded) 457 return nullptr; 458 459 // Base case: Get a regular alignof expression. 460 Constant *C = ConstantExpr::getAlignOf(Ty); 461 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 462 DestTy, false), 463 C, DestTy); 464 return C; 465 } 466 467 /// Return a ConstantExpr with type DestTy for offsetof on Ty and FieldNo, with 468 /// any known factors factored out. If Folded is false, return null if no 469 /// factoring was possible, to avoid endlessly bouncing an unfoldable expression 470 /// back into the top-level folder. 471 static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 472 Type *DestTy, 473 bool Folded) { 474 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 475 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 476 DestTy, false), 477 FieldNo, DestTy); 478 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 479 return ConstantExpr::getNUWMul(E, N); 480 } 481 482 if (StructType *STy = dyn_cast<StructType>(Ty)) 483 if (!STy->isPacked()) { 484 unsigned NumElems = STy->getNumElements(); 485 // An empty struct has no members. 486 if (NumElems == 0) 487 return nullptr; 488 // Check for a struct with all members having the same size. 489 Constant *MemberSize = 490 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 491 bool AllSame = true; 492 for (unsigned i = 1; i != NumElems; ++i) 493 if (MemberSize != 494 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 495 AllSame = false; 496 break; 497 } 498 if (AllSame) { 499 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 500 false, 501 DestTy, 502 false), 503 FieldNo, DestTy); 504 return ConstantExpr::getNUWMul(MemberSize, N); 505 } 506 } 507 508 // If there's no interesting folding happening, bail so that we don't create 509 // a constant that looks like it needs folding but really doesn't. 510 if (!Folded) 511 return nullptr; 512 513 // Base case: Get a regular offsetof expression. 514 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 515 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 516 DestTy, false), 517 C, DestTy); 518 return C; 519 } 520 521 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 522 Type *DestTy) { 523 if (isa<UndefValue>(V)) { 524 // zext(undef) = 0, because the top bits will be zero. 525 // sext(undef) = 0, because the top bits will all be the same. 526 // [us]itofp(undef) = 0, because the result value is bounded. 527 if (opc == Instruction::ZExt || opc == Instruction::SExt || 528 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 529 return Constant::getNullValue(DestTy); 530 return UndefValue::get(DestTy); 531 } 532 533 if (V->isNullValue() && !DestTy->isX86_MMXTy() && 534 opc != Instruction::AddrSpaceCast) 535 return Constant::getNullValue(DestTy); 536 537 // If the cast operand is a constant expression, there's a few things we can 538 // do to try to simplify it. 539 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 540 if (CE->isCast()) { 541 // Try hard to fold cast of cast because they are often eliminable. 542 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 543 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 544 } else if (CE->getOpcode() == Instruction::GetElementPtr && 545 // Do not fold addrspacecast (gep 0, .., 0). It might make the 546 // addrspacecast uncanonicalized. 547 opc != Instruction::AddrSpaceCast && 548 // Do not fold bitcast (gep) with inrange index, as this loses 549 // information. 550 !cast<GEPOperator>(CE)->getInRangeIndex().hasValue()) { 551 // If all of the indexes in the GEP are null values, there is no pointer 552 // adjustment going on. We might as well cast the source pointer. 553 bool isAllNull = true; 554 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 555 if (!CE->getOperand(i)->isNullValue()) { 556 isAllNull = false; 557 break; 558 } 559 if (isAllNull) 560 // This is casting one pointer type to another, always BitCast 561 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 562 } 563 } 564 565 // If the cast operand is a constant vector, perform the cast by 566 // operating on each element. In the cast of bitcasts, the element 567 // count may be mismatched; don't attempt to handle that here. 568 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 569 DestTy->isVectorTy() && 570 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 571 SmallVector<Constant*, 16> res; 572 VectorType *DestVecTy = cast<VectorType>(DestTy); 573 Type *DstEltTy = DestVecTy->getElementType(); 574 Type *Ty = IntegerType::get(V->getContext(), 32); 575 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 576 Constant *C = 577 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 578 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 579 } 580 return ConstantVector::get(res); 581 } 582 583 // We actually have to do a cast now. Perform the cast according to the 584 // opcode specified. 585 switch (opc) { 586 default: 587 llvm_unreachable("Failed to cast constant expression"); 588 case Instruction::FPTrunc: 589 case Instruction::FPExt: 590 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 591 bool ignored; 592 APFloat Val = FPC->getValueAPF(); 593 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() : 594 DestTy->isFloatTy() ? APFloat::IEEEsingle() : 595 DestTy->isDoubleTy() ? APFloat::IEEEdouble() : 596 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() : 597 DestTy->isFP128Ty() ? APFloat::IEEEquad() : 598 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() : 599 APFloat::Bogus(), 600 APFloat::rmNearestTiesToEven, &ignored); 601 return ConstantFP::get(V->getContext(), Val); 602 } 603 return nullptr; // Can't fold. 604 case Instruction::FPToUI: 605 case Instruction::FPToSI: 606 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 607 const APFloat &V = FPC->getValueAPF(); 608 bool ignored; 609 uint64_t x[2]; 610 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 611 if (APFloat::opInvalidOp == 612 V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 613 APFloat::rmTowardZero, &ignored)) { 614 // Undefined behavior invoked - the destination type can't represent 615 // the input constant. 616 return UndefValue::get(DestTy); 617 } 618 APInt Val(DestBitWidth, x); 619 return ConstantInt::get(FPC->getContext(), Val); 620 } 621 return nullptr; // Can't fold. 622 case Instruction::IntToPtr: //always treated as unsigned 623 if (V->isNullValue()) // Is it an integral null value? 624 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 625 return nullptr; // Other pointer types cannot be casted 626 case Instruction::PtrToInt: // always treated as unsigned 627 // Is it a null pointer value? 628 if (V->isNullValue()) 629 return ConstantInt::get(DestTy, 0); 630 // If this is a sizeof-like expression, pull out multiplications by 631 // known factors to expose them to subsequent folding. If it's an 632 // alignof-like expression, factor out known factors. 633 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 634 if (CE->getOpcode() == Instruction::GetElementPtr && 635 CE->getOperand(0)->isNullValue()) { 636 GEPOperator *GEPO = cast<GEPOperator>(CE); 637 Type *Ty = GEPO->getSourceElementType(); 638 if (CE->getNumOperands() == 2) { 639 // Handle a sizeof-like expression. 640 Constant *Idx = CE->getOperand(1); 641 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 642 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 643 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 644 DestTy, false), 645 Idx, DestTy); 646 return ConstantExpr::getMul(C, Idx); 647 } 648 } else if (CE->getNumOperands() == 3 && 649 CE->getOperand(1)->isNullValue()) { 650 // Handle an alignof-like expression. 651 if (StructType *STy = dyn_cast<StructType>(Ty)) 652 if (!STy->isPacked()) { 653 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 654 if (CI->isOne() && 655 STy->getNumElements() == 2 && 656 STy->getElementType(0)->isIntegerTy(1)) { 657 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 658 } 659 } 660 // Handle an offsetof-like expression. 661 if (Ty->isStructTy() || Ty->isArrayTy()) { 662 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 663 DestTy, false)) 664 return C; 665 } 666 } 667 } 668 // Other pointer types cannot be casted 669 return nullptr; 670 case Instruction::UIToFP: 671 case Instruction::SIToFP: 672 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 673 const APInt &api = CI->getValue(); 674 APFloat apf(DestTy->getFltSemantics(), 675 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 676 if (APFloat::opOverflow & 677 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 678 APFloat::rmNearestTiesToEven)) { 679 // Undefined behavior invoked - the destination type can't represent 680 // the input constant. 681 return UndefValue::get(DestTy); 682 } 683 return ConstantFP::get(V->getContext(), apf); 684 } 685 return nullptr; 686 case Instruction::ZExt: 687 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 688 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 689 return ConstantInt::get(V->getContext(), 690 CI->getValue().zext(BitWidth)); 691 } 692 return nullptr; 693 case Instruction::SExt: 694 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 695 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 696 return ConstantInt::get(V->getContext(), 697 CI->getValue().sext(BitWidth)); 698 } 699 return nullptr; 700 case Instruction::Trunc: { 701 if (V->getType()->isVectorTy()) 702 return nullptr; 703 704 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 705 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 706 return ConstantInt::get(V->getContext(), 707 CI->getValue().trunc(DestBitWidth)); 708 } 709 710 // The input must be a constantexpr. See if we can simplify this based on 711 // the bytes we are demanding. Only do this if the source and dest are an 712 // even multiple of a byte. 713 if ((DestBitWidth & 7) == 0 && 714 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 715 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 716 return Res; 717 718 return nullptr; 719 } 720 case Instruction::BitCast: 721 return FoldBitCast(V, DestTy); 722 case Instruction::AddrSpaceCast: 723 return nullptr; 724 } 725 } 726 727 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 728 Constant *V1, Constant *V2) { 729 // Check for i1 and vector true/false conditions. 730 if (Cond->isNullValue()) return V2; 731 if (Cond->isAllOnesValue()) return V1; 732 733 // If the condition is a vector constant, fold the result elementwise. 734 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 735 SmallVector<Constant*, 16> Result; 736 Type *Ty = IntegerType::get(CondV->getContext(), 32); 737 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 738 Constant *V; 739 Constant *V1Element = ConstantExpr::getExtractElement(V1, 740 ConstantInt::get(Ty, i)); 741 Constant *V2Element = ConstantExpr::getExtractElement(V2, 742 ConstantInt::get(Ty, i)); 743 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 744 if (V1Element == V2Element) { 745 V = V1Element; 746 } else if (isa<UndefValue>(Cond)) { 747 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 748 } else { 749 if (!isa<ConstantInt>(Cond)) break; 750 V = Cond->isNullValue() ? V2Element : V1Element; 751 } 752 Result.push_back(V); 753 } 754 755 // If we were able to build the vector, return it. 756 if (Result.size() == V1->getType()->getVectorNumElements()) 757 return ConstantVector::get(Result); 758 } 759 760 if (isa<UndefValue>(Cond)) { 761 if (isa<UndefValue>(V1)) return V1; 762 return V2; 763 } 764 if (isa<UndefValue>(V1)) return V2; 765 if (isa<UndefValue>(V2)) return V1; 766 if (V1 == V2) return V1; 767 768 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 769 if (TrueVal->getOpcode() == Instruction::Select) 770 if (TrueVal->getOperand(0) == Cond) 771 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 772 } 773 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 774 if (FalseVal->getOpcode() == Instruction::Select) 775 if (FalseVal->getOperand(0) == Cond) 776 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 777 } 778 779 return nullptr; 780 } 781 782 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 783 Constant *Idx) { 784 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 785 return UndefValue::get(Val->getType()->getVectorElementType()); 786 if (Val->isNullValue()) // ee(zero, x) -> zero 787 return Constant::getNullValue(Val->getType()->getVectorElementType()); 788 // ee({w,x,y,z}, undef) -> undef 789 if (isa<UndefValue>(Idx)) 790 return UndefValue::get(Val->getType()->getVectorElementType()); 791 792 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 793 // ee({w,x,y,z}, wrong_value) -> undef 794 if (CIdx->uge(Val->getType()->getVectorNumElements())) 795 return UndefValue::get(Val->getType()->getVectorElementType()); 796 return Val->getAggregateElement(CIdx->getZExtValue()); 797 } 798 return nullptr; 799 } 800 801 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 802 Constant *Elt, 803 Constant *Idx) { 804 if (isa<UndefValue>(Idx)) 805 return UndefValue::get(Val->getType()); 806 807 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 808 if (!CIdx) return nullptr; 809 810 unsigned NumElts = Val->getType()->getVectorNumElements(); 811 if (CIdx->uge(NumElts)) 812 return UndefValue::get(Val->getType()); 813 814 SmallVector<Constant*, 16> Result; 815 Result.reserve(NumElts); 816 auto *Ty = Type::getInt32Ty(Val->getContext()); 817 uint64_t IdxVal = CIdx->getZExtValue(); 818 for (unsigned i = 0; i != NumElts; ++i) { 819 if (i == IdxVal) { 820 Result.push_back(Elt); 821 continue; 822 } 823 824 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 825 Result.push_back(C); 826 } 827 828 return ConstantVector::get(Result); 829 } 830 831 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 832 Constant *V2, 833 Constant *Mask) { 834 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 835 Type *EltTy = V1->getType()->getVectorElementType(); 836 837 // Undefined shuffle mask -> undefined value. 838 if (isa<UndefValue>(Mask)) 839 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 840 841 // Don't break the bitcode reader hack. 842 if (isa<ConstantExpr>(Mask)) return nullptr; 843 844 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 845 846 // Loop over the shuffle mask, evaluating each element. 847 SmallVector<Constant*, 32> Result; 848 for (unsigned i = 0; i != MaskNumElts; ++i) { 849 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 850 if (Elt == -1) { 851 Result.push_back(UndefValue::get(EltTy)); 852 continue; 853 } 854 Constant *InElt; 855 if (unsigned(Elt) >= SrcNumElts*2) 856 InElt = UndefValue::get(EltTy); 857 else if (unsigned(Elt) >= SrcNumElts) { 858 Type *Ty = IntegerType::get(V2->getContext(), 32); 859 InElt = 860 ConstantExpr::getExtractElement(V2, 861 ConstantInt::get(Ty, Elt - SrcNumElts)); 862 } else { 863 Type *Ty = IntegerType::get(V1->getContext(), 32); 864 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 865 } 866 Result.push_back(InElt); 867 } 868 869 return ConstantVector::get(Result); 870 } 871 872 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 873 ArrayRef<unsigned> Idxs) { 874 // Base case: no indices, so return the entire value. 875 if (Idxs.empty()) 876 return Agg; 877 878 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 879 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 880 881 return nullptr; 882 } 883 884 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 885 Constant *Val, 886 ArrayRef<unsigned> Idxs) { 887 // Base case: no indices, so replace the entire value. 888 if (Idxs.empty()) 889 return Val; 890 891 unsigned NumElts; 892 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 893 NumElts = ST->getNumElements(); 894 else 895 NumElts = cast<SequentialType>(Agg->getType())->getNumElements(); 896 897 SmallVector<Constant*, 32> Result; 898 for (unsigned i = 0; i != NumElts; ++i) { 899 Constant *C = Agg->getAggregateElement(i); 900 if (!C) return nullptr; 901 902 if (Idxs[0] == i) 903 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 904 905 Result.push_back(C); 906 } 907 908 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 909 return ConstantStruct::get(ST, Result); 910 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 911 return ConstantArray::get(AT, Result); 912 return ConstantVector::get(Result); 913 } 914 915 916 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 917 Constant *C1, Constant *C2) { 918 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 919 920 // Handle UndefValue up front. 921 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 922 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 923 case Instruction::Xor: 924 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 925 // Handle undef ^ undef -> 0 special case. This is a common 926 // idiom (misuse). 927 return Constant::getNullValue(C1->getType()); 928 LLVM_FALLTHROUGH; 929 case Instruction::Add: 930 case Instruction::Sub: 931 return UndefValue::get(C1->getType()); 932 case Instruction::And: 933 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 934 return C1; 935 return Constant::getNullValue(C1->getType()); // undef & X -> 0 936 case Instruction::Mul: { 937 // undef * undef -> undef 938 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 939 return C1; 940 const APInt *CV; 941 // X * undef -> undef if X is odd 942 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 943 if ((*CV)[0]) 944 return UndefValue::get(C1->getType()); 945 946 // X * undef -> 0 otherwise 947 return Constant::getNullValue(C1->getType()); 948 } 949 case Instruction::SDiv: 950 case Instruction::UDiv: 951 // X / undef -> undef 952 if (isa<UndefValue>(C2)) 953 return C2; 954 // undef / 0 -> undef 955 // undef / 1 -> undef 956 if (match(C2, m_Zero()) || match(C2, m_One())) 957 return C1; 958 // undef / X -> 0 otherwise 959 return Constant::getNullValue(C1->getType()); 960 case Instruction::URem: 961 case Instruction::SRem: 962 // X % undef -> undef 963 if (match(C2, m_Undef())) 964 return C2; 965 // undef % 0 -> undef 966 if (match(C2, m_Zero())) 967 return C1; 968 // undef % X -> 0 otherwise 969 return Constant::getNullValue(C1->getType()); 970 case Instruction::Or: // X | undef -> -1 971 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 972 return C1; 973 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 974 case Instruction::LShr: 975 // X >>l undef -> undef 976 if (isa<UndefValue>(C2)) 977 return C2; 978 // undef >>l 0 -> undef 979 if (match(C2, m_Zero())) 980 return C1; 981 // undef >>l X -> 0 982 return Constant::getNullValue(C1->getType()); 983 case Instruction::AShr: 984 // X >>a undef -> undef 985 if (isa<UndefValue>(C2)) 986 return C2; 987 // undef >>a 0 -> undef 988 if (match(C2, m_Zero())) 989 return C1; 990 // TODO: undef >>a X -> undef if the shift is exact 991 // undef >>a X -> 0 992 return Constant::getNullValue(C1->getType()); 993 case Instruction::Shl: 994 // X << undef -> undef 995 if (isa<UndefValue>(C2)) 996 return C2; 997 // undef << 0 -> undef 998 if (match(C2, m_Zero())) 999 return C1; 1000 // undef << X -> 0 1001 return Constant::getNullValue(C1->getType()); 1002 case Instruction::FAdd: 1003 case Instruction::FSub: 1004 case Instruction::FMul: 1005 case Instruction::FDiv: 1006 case Instruction::FRem: 1007 // TODO: UNDEF handling for binary float instructions. 1008 return nullptr; 1009 case Instruction::BinaryOpsEnd: 1010 llvm_unreachable("Invalid BinaryOp"); 1011 } 1012 } 1013 1014 // At this point neither constant should be an UndefValue. 1015 assert(!isa<UndefValue>(C1) && !isa<UndefValue>(C2) && 1016 "Unexpected UndefValue"); 1017 1018 // Handle simplifications when the RHS is a constant int. 1019 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1020 switch (Opcode) { 1021 case Instruction::Add: 1022 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1023 break; 1024 case Instruction::Sub: 1025 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1026 break; 1027 case Instruction::Mul: 1028 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1029 if (CI2->equalsInt(1)) 1030 return C1; // X * 1 == X 1031 break; 1032 case Instruction::UDiv: 1033 case Instruction::SDiv: 1034 if (CI2->equalsInt(1)) 1035 return C1; // X / 1 == X 1036 if (CI2->equalsInt(0)) 1037 return UndefValue::get(CI2->getType()); // X / 0 == undef 1038 break; 1039 case Instruction::URem: 1040 case Instruction::SRem: 1041 if (CI2->equalsInt(1)) 1042 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1043 if (CI2->equalsInt(0)) 1044 return UndefValue::get(CI2->getType()); // X % 0 == undef 1045 break; 1046 case Instruction::And: 1047 if (CI2->isZero()) return C2; // X & 0 == 0 1048 if (CI2->isAllOnesValue()) 1049 return C1; // X & -1 == X 1050 1051 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1052 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1053 if (CE1->getOpcode() == Instruction::ZExt) { 1054 unsigned DstWidth = CI2->getType()->getBitWidth(); 1055 unsigned SrcWidth = 1056 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1057 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1058 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1059 return C1; 1060 } 1061 1062 // If and'ing the address of a global with a constant, fold it. 1063 if (CE1->getOpcode() == Instruction::PtrToInt && 1064 isa<GlobalValue>(CE1->getOperand(0))) { 1065 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1066 1067 // Functions are at least 4-byte aligned. 1068 unsigned GVAlign = GV->getAlignment(); 1069 if (isa<Function>(GV)) 1070 GVAlign = std::max(GVAlign, 4U); 1071 1072 if (GVAlign > 1) { 1073 unsigned DstWidth = CI2->getType()->getBitWidth(); 1074 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1075 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1076 1077 // If checking bits we know are clear, return zero. 1078 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1079 return Constant::getNullValue(CI2->getType()); 1080 } 1081 } 1082 } 1083 break; 1084 case Instruction::Or: 1085 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1086 if (CI2->isAllOnesValue()) 1087 return C2; // X | -1 == -1 1088 break; 1089 case Instruction::Xor: 1090 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1091 1092 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1093 switch (CE1->getOpcode()) { 1094 default: break; 1095 case Instruction::ICmp: 1096 case Instruction::FCmp: 1097 // cmp pred ^ true -> cmp !pred 1098 assert(CI2->equalsInt(1)); 1099 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1100 pred = CmpInst::getInversePredicate(pred); 1101 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1102 CE1->getOperand(1)); 1103 } 1104 } 1105 break; 1106 case Instruction::AShr: 1107 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1108 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1109 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1110 return ConstantExpr::getLShr(C1, C2); 1111 break; 1112 } 1113 } else if (isa<ConstantInt>(C1)) { 1114 // If C1 is a ConstantInt and C2 is not, swap the operands. 1115 if (Instruction::isCommutative(Opcode)) 1116 return ConstantExpr::get(Opcode, C2, C1); 1117 } 1118 1119 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1120 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1121 const APInt &C1V = CI1->getValue(); 1122 const APInt &C2V = CI2->getValue(); 1123 switch (Opcode) { 1124 default: 1125 break; 1126 case Instruction::Add: 1127 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1128 case Instruction::Sub: 1129 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1130 case Instruction::Mul: 1131 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1132 case Instruction::UDiv: 1133 assert(!CI2->isNullValue() && "Div by zero handled above"); 1134 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1135 case Instruction::SDiv: 1136 assert(!CI2->isNullValue() && "Div by zero handled above"); 1137 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1138 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1139 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1140 case Instruction::URem: 1141 assert(!CI2->isNullValue() && "Div by zero handled above"); 1142 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1143 case Instruction::SRem: 1144 assert(!CI2->isNullValue() && "Div by zero handled above"); 1145 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1146 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1147 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1148 case Instruction::And: 1149 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1150 case Instruction::Or: 1151 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1152 case Instruction::Xor: 1153 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1154 case Instruction::Shl: 1155 if (C2V.ult(C1V.getBitWidth())) 1156 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1157 return UndefValue::get(C1->getType()); // too big shift is undef 1158 case Instruction::LShr: 1159 if (C2V.ult(C1V.getBitWidth())) 1160 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1161 return UndefValue::get(C1->getType()); // too big shift is undef 1162 case Instruction::AShr: 1163 if (C2V.ult(C1V.getBitWidth())) 1164 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1165 return UndefValue::get(C1->getType()); // too big shift is undef 1166 } 1167 } 1168 1169 switch (Opcode) { 1170 case Instruction::SDiv: 1171 case Instruction::UDiv: 1172 case Instruction::URem: 1173 case Instruction::SRem: 1174 case Instruction::LShr: 1175 case Instruction::AShr: 1176 case Instruction::Shl: 1177 if (CI1->equalsInt(0)) return C1; 1178 break; 1179 default: 1180 break; 1181 } 1182 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1183 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1184 const APFloat &C1V = CFP1->getValueAPF(); 1185 const APFloat &C2V = CFP2->getValueAPF(); 1186 APFloat C3V = C1V; // copy for modification 1187 switch (Opcode) { 1188 default: 1189 break; 1190 case Instruction::FAdd: 1191 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1192 return ConstantFP::get(C1->getContext(), C3V); 1193 case Instruction::FSub: 1194 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1195 return ConstantFP::get(C1->getContext(), C3V); 1196 case Instruction::FMul: 1197 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1198 return ConstantFP::get(C1->getContext(), C3V); 1199 case Instruction::FDiv: 1200 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1201 return ConstantFP::get(C1->getContext(), C3V); 1202 case Instruction::FRem: 1203 (void)C3V.mod(C2V); 1204 return ConstantFP::get(C1->getContext(), C3V); 1205 } 1206 } 1207 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1208 // Perform elementwise folding. 1209 SmallVector<Constant*, 16> Result; 1210 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1211 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1212 Constant *LHS = 1213 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1214 Constant *RHS = 1215 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1216 1217 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1218 } 1219 1220 return ConstantVector::get(Result); 1221 } 1222 1223 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1224 // There are many possible foldings we could do here. We should probably 1225 // at least fold add of a pointer with an integer into the appropriate 1226 // getelementptr. This will improve alias analysis a bit. 1227 1228 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1229 // (a + (b + c)). 1230 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1231 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1232 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1233 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1234 } 1235 } else if (isa<ConstantExpr>(C2)) { 1236 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1237 // other way if possible. 1238 if (Instruction::isCommutative(Opcode)) 1239 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1240 } 1241 1242 // i1 can be simplified in many cases. 1243 if (C1->getType()->isIntegerTy(1)) { 1244 switch (Opcode) { 1245 case Instruction::Add: 1246 case Instruction::Sub: 1247 return ConstantExpr::getXor(C1, C2); 1248 case Instruction::Mul: 1249 return ConstantExpr::getAnd(C1, C2); 1250 case Instruction::Shl: 1251 case Instruction::LShr: 1252 case Instruction::AShr: 1253 // We can assume that C2 == 0. If it were one the result would be 1254 // undefined because the shift value is as large as the bitwidth. 1255 return C1; 1256 case Instruction::SDiv: 1257 case Instruction::UDiv: 1258 // We can assume that C2 == 1. If it were zero the result would be 1259 // undefined through division by zero. 1260 return C1; 1261 case Instruction::URem: 1262 case Instruction::SRem: 1263 // We can assume that C2 == 1. If it were zero the result would be 1264 // undefined through division by zero. 1265 return ConstantInt::getFalse(C1->getContext()); 1266 default: 1267 break; 1268 } 1269 } 1270 1271 // We don't know how to fold this. 1272 return nullptr; 1273 } 1274 1275 /// This type is zero-sized if it's an array or structure of zero-sized types. 1276 /// The only leaf zero-sized type is an empty structure. 1277 static bool isMaybeZeroSizedType(Type *Ty) { 1278 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1279 if (STy->isOpaque()) return true; // Can't say. 1280 1281 // If all of elements have zero size, this does too. 1282 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1283 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1284 return true; 1285 1286 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1287 return isMaybeZeroSizedType(ATy->getElementType()); 1288 } 1289 return false; 1290 } 1291 1292 /// Compare the two constants as though they were getelementptr indices. 1293 /// This allows coercion of the types to be the same thing. 1294 /// 1295 /// If the two constants are the "same" (after coercion), return 0. If the 1296 /// first is less than the second, return -1, if the second is less than the 1297 /// first, return 1. If the constants are not integral, return -2. 1298 /// 1299 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1300 if (C1 == C2) return 0; 1301 1302 // Ok, we found a different index. If they are not ConstantInt, we can't do 1303 // anything with them. 1304 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1305 return -2; // don't know! 1306 1307 // We cannot compare the indices if they don't fit in an int64_t. 1308 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 || 1309 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64) 1310 return -2; // don't know! 1311 1312 // Ok, we have two differing integer indices. Sign extend them to be the same 1313 // type. 1314 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue(); 1315 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue(); 1316 1317 if (C1Val == C2Val) return 0; // They are equal 1318 1319 // If the type being indexed over is really just a zero sized type, there is 1320 // no pointer difference being made here. 1321 if (isMaybeZeroSizedType(ElTy)) 1322 return -2; // dunno. 1323 1324 // If they are really different, now that they are the same type, then we 1325 // found a difference! 1326 if (C1Val < C2Val) 1327 return -1; 1328 else 1329 return 1; 1330 } 1331 1332 /// This function determines if there is anything we can decide about the two 1333 /// constants provided. This doesn't need to handle simple things like 1334 /// ConstantFP comparisons, but should instead handle ConstantExprs. 1335 /// If we can determine that the two constants have a particular relation to 1336 /// each other, we should return the corresponding FCmpInst predicate, 1337 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1338 /// ConstantFoldCompareInstruction. 1339 /// 1340 /// To simplify this code we canonicalize the relation so that the first 1341 /// operand is always the most "complex" of the two. We consider ConstantFP 1342 /// to be the simplest, and ConstantExprs to be the most complex. 1343 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1344 assert(V1->getType() == V2->getType() && 1345 "Cannot compare values of different types!"); 1346 1347 // Handle degenerate case quickly 1348 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1349 1350 if (!isa<ConstantExpr>(V1)) { 1351 if (!isa<ConstantExpr>(V2)) { 1352 // Simple case, use the standard constant folder. 1353 ConstantInt *R = nullptr; 1354 R = dyn_cast<ConstantInt>( 1355 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1356 if (R && !R->isZero()) 1357 return FCmpInst::FCMP_OEQ; 1358 R = dyn_cast<ConstantInt>( 1359 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1360 if (R && !R->isZero()) 1361 return FCmpInst::FCMP_OLT; 1362 R = dyn_cast<ConstantInt>( 1363 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1364 if (R && !R->isZero()) 1365 return FCmpInst::FCMP_OGT; 1366 1367 // Nothing more we can do 1368 return FCmpInst::BAD_FCMP_PREDICATE; 1369 } 1370 1371 // If the first operand is simple and second is ConstantExpr, swap operands. 1372 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1373 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1374 return FCmpInst::getSwappedPredicate(SwappedRelation); 1375 } else { 1376 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1377 // constantexpr or a simple constant. 1378 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1379 switch (CE1->getOpcode()) { 1380 case Instruction::FPTrunc: 1381 case Instruction::FPExt: 1382 case Instruction::UIToFP: 1383 case Instruction::SIToFP: 1384 // We might be able to do something with these but we don't right now. 1385 break; 1386 default: 1387 break; 1388 } 1389 } 1390 // There are MANY other foldings that we could perform here. They will 1391 // probably be added on demand, as they seem needed. 1392 return FCmpInst::BAD_FCMP_PREDICATE; 1393 } 1394 1395 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1396 const GlobalValue *GV2) { 1397 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1398 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1399 return true; 1400 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1401 Type *Ty = GVar->getValueType(); 1402 // A global with opaque type might end up being zero sized. 1403 if (!Ty->isSized()) 1404 return true; 1405 // A global with an empty type might lie at the address of any other 1406 // global. 1407 if (Ty->isEmptyTy()) 1408 return true; 1409 } 1410 return false; 1411 }; 1412 // Don't try to decide equality of aliases. 1413 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1414 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1415 return ICmpInst::ICMP_NE; 1416 return ICmpInst::BAD_ICMP_PREDICATE; 1417 } 1418 1419 /// This function determines if there is anything we can decide about the two 1420 /// constants provided. This doesn't need to handle simple things like integer 1421 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1422 /// If we can determine that the two constants have a particular relation to 1423 /// each other, we should return the corresponding ICmp predicate, otherwise 1424 /// return ICmpInst::BAD_ICMP_PREDICATE. 1425 /// 1426 /// To simplify this code we canonicalize the relation so that the first 1427 /// operand is always the most "complex" of the two. We consider simple 1428 /// constants (like ConstantInt) to be the simplest, followed by 1429 /// GlobalValues, followed by ConstantExpr's (the most complex). 1430 /// 1431 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1432 bool isSigned) { 1433 assert(V1->getType() == V2->getType() && 1434 "Cannot compare different types of values!"); 1435 if (V1 == V2) return ICmpInst::ICMP_EQ; 1436 1437 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1438 !isa<BlockAddress>(V1)) { 1439 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1440 !isa<BlockAddress>(V2)) { 1441 // We distilled this down to a simple case, use the standard constant 1442 // folder. 1443 ConstantInt *R = nullptr; 1444 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1445 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1446 if (R && !R->isZero()) 1447 return pred; 1448 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1449 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1450 if (R && !R->isZero()) 1451 return pred; 1452 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1453 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1454 if (R && !R->isZero()) 1455 return pred; 1456 1457 // If we couldn't figure it out, bail. 1458 return ICmpInst::BAD_ICMP_PREDICATE; 1459 } 1460 1461 // If the first operand is simple, swap operands. 1462 ICmpInst::Predicate SwappedRelation = 1463 evaluateICmpRelation(V2, V1, isSigned); 1464 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1465 return ICmpInst::getSwappedPredicate(SwappedRelation); 1466 1467 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1468 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1469 ICmpInst::Predicate SwappedRelation = 1470 evaluateICmpRelation(V2, V1, isSigned); 1471 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1472 return ICmpInst::getSwappedPredicate(SwappedRelation); 1473 return ICmpInst::BAD_ICMP_PREDICATE; 1474 } 1475 1476 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1477 // constant (which, since the types must match, means that it's a 1478 // ConstantPointerNull). 1479 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1480 return areGlobalsPotentiallyEqual(GV, GV2); 1481 } else if (isa<BlockAddress>(V2)) { 1482 return ICmpInst::ICMP_NE; // Globals never equal labels. 1483 } else { 1484 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1485 // GlobalVals can never be null unless they have external weak linkage. 1486 // We don't try to evaluate aliases here. 1487 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1488 return ICmpInst::ICMP_NE; 1489 } 1490 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1491 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1492 ICmpInst::Predicate SwappedRelation = 1493 evaluateICmpRelation(V2, V1, isSigned); 1494 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1495 return ICmpInst::getSwappedPredicate(SwappedRelation); 1496 return ICmpInst::BAD_ICMP_PREDICATE; 1497 } 1498 1499 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1500 // constant (which, since the types must match, means that it is a 1501 // ConstantPointerNull). 1502 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1503 // Block address in another function can't equal this one, but block 1504 // addresses in the current function might be the same if blocks are 1505 // empty. 1506 if (BA2->getFunction() != BA->getFunction()) 1507 return ICmpInst::ICMP_NE; 1508 } else { 1509 // Block addresses aren't null, don't equal the address of globals. 1510 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1511 "Canonicalization guarantee!"); 1512 return ICmpInst::ICMP_NE; 1513 } 1514 } else { 1515 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1516 // constantexpr, a global, block address, or a simple constant. 1517 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1518 Constant *CE1Op0 = CE1->getOperand(0); 1519 1520 switch (CE1->getOpcode()) { 1521 case Instruction::Trunc: 1522 case Instruction::FPTrunc: 1523 case Instruction::FPExt: 1524 case Instruction::FPToUI: 1525 case Instruction::FPToSI: 1526 break; // We can't evaluate floating point casts or truncations. 1527 1528 case Instruction::UIToFP: 1529 case Instruction::SIToFP: 1530 case Instruction::BitCast: 1531 case Instruction::ZExt: 1532 case Instruction::SExt: 1533 // We can't evaluate floating point casts or truncations. 1534 if (CE1Op0->getType()->isFloatingPointTy()) 1535 break; 1536 1537 // If the cast is not actually changing bits, and the second operand is a 1538 // null pointer, do the comparison with the pre-casted value. 1539 if (V2->isNullValue() && 1540 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1541 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1542 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1543 return evaluateICmpRelation(CE1Op0, 1544 Constant::getNullValue(CE1Op0->getType()), 1545 isSigned); 1546 } 1547 break; 1548 1549 case Instruction::GetElementPtr: { 1550 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1551 // Ok, since this is a getelementptr, we know that the constant has a 1552 // pointer type. Check the various cases. 1553 if (isa<ConstantPointerNull>(V2)) { 1554 // If we are comparing a GEP to a null pointer, check to see if the base 1555 // of the GEP equals the null pointer. 1556 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1557 if (GV->hasExternalWeakLinkage()) 1558 // Weak linkage GVals could be zero or not. We're comparing that 1559 // to null pointer so its greater-or-equal 1560 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1561 else 1562 // If its not weak linkage, the GVal must have a non-zero address 1563 // so the result is greater-than 1564 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1565 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1566 // If we are indexing from a null pointer, check to see if we have any 1567 // non-zero indices. 1568 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1569 if (!CE1->getOperand(i)->isNullValue()) 1570 // Offsetting from null, must not be equal. 1571 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1572 // Only zero indexes from null, must still be zero. 1573 return ICmpInst::ICMP_EQ; 1574 } 1575 // Otherwise, we can't really say if the first operand is null or not. 1576 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1577 if (isa<ConstantPointerNull>(CE1Op0)) { 1578 if (GV2->hasExternalWeakLinkage()) 1579 // Weak linkage GVals could be zero or not. We're comparing it to 1580 // a null pointer, so its less-or-equal 1581 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1582 else 1583 // If its not weak linkage, the GVal must have a non-zero address 1584 // so the result is less-than 1585 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1586 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1587 if (GV == GV2) { 1588 // If this is a getelementptr of the same global, then it must be 1589 // different. Because the types must match, the getelementptr could 1590 // only have at most one index, and because we fold getelementptr's 1591 // with a single zero index, it must be nonzero. 1592 assert(CE1->getNumOperands() == 2 && 1593 !CE1->getOperand(1)->isNullValue() && 1594 "Surprising getelementptr!"); 1595 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1596 } else { 1597 if (CE1GEP->hasAllZeroIndices()) 1598 return areGlobalsPotentiallyEqual(GV, GV2); 1599 return ICmpInst::BAD_ICMP_PREDICATE; 1600 } 1601 } 1602 } else { 1603 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1604 Constant *CE2Op0 = CE2->getOperand(0); 1605 1606 // There are MANY other foldings that we could perform here. They will 1607 // probably be added on demand, as they seem needed. 1608 switch (CE2->getOpcode()) { 1609 default: break; 1610 case Instruction::GetElementPtr: 1611 // By far the most common case to handle is when the base pointers are 1612 // obviously to the same global. 1613 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1614 // Don't know relative ordering, but check for inequality. 1615 if (CE1Op0 != CE2Op0) { 1616 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1617 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1618 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1619 cast<GlobalValue>(CE2Op0)); 1620 return ICmpInst::BAD_ICMP_PREDICATE; 1621 } 1622 // Ok, we know that both getelementptr instructions are based on the 1623 // same global. From this, we can precisely determine the relative 1624 // ordering of the resultant pointers. 1625 unsigned i = 1; 1626 1627 // The logic below assumes that the result of the comparison 1628 // can be determined by finding the first index that differs. 1629 // This doesn't work if there is over-indexing in any 1630 // subsequent indices, so check for that case first. 1631 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1632 !CE2->isGEPWithNoNotionalOverIndexing()) 1633 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1634 1635 // Compare all of the operands the GEP's have in common. 1636 gep_type_iterator GTI = gep_type_begin(CE1); 1637 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1638 ++i, ++GTI) 1639 switch (IdxCompare(CE1->getOperand(i), 1640 CE2->getOperand(i), GTI.getIndexedType())) { 1641 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1642 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1643 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1644 } 1645 1646 // Ok, we ran out of things they have in common. If any leftovers 1647 // are non-zero then we have a difference, otherwise we are equal. 1648 for (; i < CE1->getNumOperands(); ++i) 1649 if (!CE1->getOperand(i)->isNullValue()) { 1650 if (isa<ConstantInt>(CE1->getOperand(i))) 1651 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1652 else 1653 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1654 } 1655 1656 for (; i < CE2->getNumOperands(); ++i) 1657 if (!CE2->getOperand(i)->isNullValue()) { 1658 if (isa<ConstantInt>(CE2->getOperand(i))) 1659 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1660 else 1661 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1662 } 1663 return ICmpInst::ICMP_EQ; 1664 } 1665 } 1666 } 1667 } 1668 default: 1669 break; 1670 } 1671 } 1672 1673 return ICmpInst::BAD_ICMP_PREDICATE; 1674 } 1675 1676 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1677 Constant *C1, Constant *C2) { 1678 Type *ResultTy; 1679 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1680 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1681 VT->getNumElements()); 1682 else 1683 ResultTy = Type::getInt1Ty(C1->getContext()); 1684 1685 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1686 if (pred == FCmpInst::FCMP_FALSE) 1687 return Constant::getNullValue(ResultTy); 1688 1689 if (pred == FCmpInst::FCMP_TRUE) 1690 return Constant::getAllOnesValue(ResultTy); 1691 1692 // Handle some degenerate cases first 1693 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1694 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1695 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1696 // For EQ and NE, we can always pick a value for the undef to make the 1697 // predicate pass or fail, so we can return undef. 1698 // Also, if both operands are undef, we can return undef for int comparison. 1699 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1700 return UndefValue::get(ResultTy); 1701 1702 // Otherwise, for integer compare, pick the same value as the non-undef 1703 // operand, and fold it to true or false. 1704 if (isIntegerPredicate) 1705 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1706 1707 // Choosing NaN for the undef will always make unordered comparison succeed 1708 // and ordered comparison fails. 1709 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1710 } 1711 1712 // icmp eq/ne(null,GV) -> false/true 1713 if (C1->isNullValue()) { 1714 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1715 // Don't try to evaluate aliases. External weak GV can be null. 1716 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1717 if (pred == ICmpInst::ICMP_EQ) 1718 return ConstantInt::getFalse(C1->getContext()); 1719 else if (pred == ICmpInst::ICMP_NE) 1720 return ConstantInt::getTrue(C1->getContext()); 1721 } 1722 // icmp eq/ne(GV,null) -> false/true 1723 } else if (C2->isNullValue()) { 1724 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1725 // Don't try to evaluate aliases. External weak GV can be null. 1726 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1727 if (pred == ICmpInst::ICMP_EQ) 1728 return ConstantInt::getFalse(C1->getContext()); 1729 else if (pred == ICmpInst::ICMP_NE) 1730 return ConstantInt::getTrue(C1->getContext()); 1731 } 1732 } 1733 1734 // If the comparison is a comparison between two i1's, simplify it. 1735 if (C1->getType()->isIntegerTy(1)) { 1736 switch(pred) { 1737 case ICmpInst::ICMP_EQ: 1738 if (isa<ConstantInt>(C2)) 1739 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1740 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1741 case ICmpInst::ICMP_NE: 1742 return ConstantExpr::getXor(C1, C2); 1743 default: 1744 break; 1745 } 1746 } 1747 1748 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1749 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1750 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1751 switch (pred) { 1752 default: llvm_unreachable("Invalid ICmp Predicate"); 1753 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1754 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1755 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1756 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1757 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1758 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1759 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1760 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1761 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1762 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1763 } 1764 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1765 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1766 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1767 APFloat::cmpResult R = C1V.compare(C2V); 1768 switch (pred) { 1769 default: llvm_unreachable("Invalid FCmp Predicate"); 1770 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1771 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1772 case FCmpInst::FCMP_UNO: 1773 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1774 case FCmpInst::FCMP_ORD: 1775 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1776 case FCmpInst::FCMP_UEQ: 1777 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1778 R==APFloat::cmpEqual); 1779 case FCmpInst::FCMP_OEQ: 1780 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1781 case FCmpInst::FCMP_UNE: 1782 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1783 case FCmpInst::FCMP_ONE: 1784 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1785 R==APFloat::cmpGreaterThan); 1786 case FCmpInst::FCMP_ULT: 1787 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1788 R==APFloat::cmpLessThan); 1789 case FCmpInst::FCMP_OLT: 1790 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1791 case FCmpInst::FCMP_UGT: 1792 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1793 R==APFloat::cmpGreaterThan); 1794 case FCmpInst::FCMP_OGT: 1795 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1796 case FCmpInst::FCMP_ULE: 1797 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1798 case FCmpInst::FCMP_OLE: 1799 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1800 R==APFloat::cmpEqual); 1801 case FCmpInst::FCMP_UGE: 1802 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1803 case FCmpInst::FCMP_OGE: 1804 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1805 R==APFloat::cmpEqual); 1806 } 1807 } else if (C1->getType()->isVectorTy()) { 1808 // If we can constant fold the comparison of each element, constant fold 1809 // the whole vector comparison. 1810 SmallVector<Constant*, 4> ResElts; 1811 Type *Ty = IntegerType::get(C1->getContext(), 32); 1812 // Compare the elements, producing an i1 result or constant expr. 1813 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1814 Constant *C1E = 1815 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1816 Constant *C2E = 1817 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1818 1819 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1820 } 1821 1822 return ConstantVector::get(ResElts); 1823 } 1824 1825 if (C1->getType()->isFloatingPointTy() && 1826 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1827 // infinite recursive loop 1828 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1829 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1830 switch (evaluateFCmpRelation(C1, C2)) { 1831 default: llvm_unreachable("Unknown relation!"); 1832 case FCmpInst::FCMP_UNO: 1833 case FCmpInst::FCMP_ORD: 1834 case FCmpInst::FCMP_UEQ: 1835 case FCmpInst::FCMP_UNE: 1836 case FCmpInst::FCMP_ULT: 1837 case FCmpInst::FCMP_UGT: 1838 case FCmpInst::FCMP_ULE: 1839 case FCmpInst::FCMP_UGE: 1840 case FCmpInst::FCMP_TRUE: 1841 case FCmpInst::FCMP_FALSE: 1842 case FCmpInst::BAD_FCMP_PREDICATE: 1843 break; // Couldn't determine anything about these constants. 1844 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1845 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1846 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1847 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1848 break; 1849 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1850 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1851 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1852 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1853 break; 1854 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1855 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1856 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1857 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1858 break; 1859 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1860 // We can only partially decide this relation. 1861 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1862 Result = 0; 1863 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1864 Result = 1; 1865 break; 1866 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1867 // We can only partially decide this relation. 1868 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1869 Result = 0; 1870 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1871 Result = 1; 1872 break; 1873 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1874 // We can only partially decide this relation. 1875 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1876 Result = 0; 1877 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1878 Result = 1; 1879 break; 1880 } 1881 1882 // If we evaluated the result, return it now. 1883 if (Result != -1) 1884 return ConstantInt::get(ResultTy, Result); 1885 1886 } else { 1887 // Evaluate the relation between the two constants, per the predicate. 1888 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1889 switch (evaluateICmpRelation(C1, C2, 1890 CmpInst::isSigned((CmpInst::Predicate)pred))) { 1891 default: llvm_unreachable("Unknown relational!"); 1892 case ICmpInst::BAD_ICMP_PREDICATE: 1893 break; // Couldn't determine anything about these constants. 1894 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1895 // If we know the constants are equal, we can decide the result of this 1896 // computation precisely. 1897 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1898 break; 1899 case ICmpInst::ICMP_ULT: 1900 switch (pred) { 1901 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1902 Result = 1; break; 1903 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1904 Result = 0; break; 1905 } 1906 break; 1907 case ICmpInst::ICMP_SLT: 1908 switch (pred) { 1909 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1910 Result = 1; break; 1911 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1912 Result = 0; break; 1913 } 1914 break; 1915 case ICmpInst::ICMP_UGT: 1916 switch (pred) { 1917 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1918 Result = 1; break; 1919 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1920 Result = 0; break; 1921 } 1922 break; 1923 case ICmpInst::ICMP_SGT: 1924 switch (pred) { 1925 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1926 Result = 1; break; 1927 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1928 Result = 0; break; 1929 } 1930 break; 1931 case ICmpInst::ICMP_ULE: 1932 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1933 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1934 break; 1935 case ICmpInst::ICMP_SLE: 1936 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1937 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1938 break; 1939 case ICmpInst::ICMP_UGE: 1940 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1941 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1942 break; 1943 case ICmpInst::ICMP_SGE: 1944 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1945 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1946 break; 1947 case ICmpInst::ICMP_NE: 1948 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1949 if (pred == ICmpInst::ICMP_NE) Result = 1; 1950 break; 1951 } 1952 1953 // If we evaluated the result, return it now. 1954 if (Result != -1) 1955 return ConstantInt::get(ResultTy, Result); 1956 1957 // If the right hand side is a bitcast, try using its inverse to simplify 1958 // it by moving it to the left hand side. We can't do this if it would turn 1959 // a vector compare into a scalar compare or visa versa. 1960 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1961 Constant *CE2Op0 = CE2->getOperand(0); 1962 if (CE2->getOpcode() == Instruction::BitCast && 1963 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1964 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1965 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1966 } 1967 } 1968 1969 // If the left hand side is an extension, try eliminating it. 1970 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1971 if ((CE1->getOpcode() == Instruction::SExt && 1972 ICmpInst::isSigned((ICmpInst::Predicate)pred)) || 1973 (CE1->getOpcode() == Instruction::ZExt && 1974 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){ 1975 Constant *CE1Op0 = CE1->getOperand(0); 1976 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1977 if (CE1Inverse == CE1Op0) { 1978 // Check whether we can safely truncate the right hand side. 1979 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1980 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1981 C2->getType()) == C2) 1982 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1983 } 1984 } 1985 } 1986 1987 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1988 (C1->isNullValue() && !C2->isNullValue())) { 1989 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1990 // other way if possible. 1991 // Also, if C1 is null and C2 isn't, flip them around. 1992 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1993 return ConstantExpr::getICmp(pred, C2, C1); 1994 } 1995 } 1996 return nullptr; 1997 } 1998 1999 /// Test whether the given sequence of *normalized* indices is "inbounds". 2000 template<typename IndexTy> 2001 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 2002 // No indices means nothing that could be out of bounds. 2003 if (Idxs.empty()) return true; 2004 2005 // If the first index is zero, it's in bounds. 2006 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2007 2008 // If the first index is one and all the rest are zero, it's in bounds, 2009 // by the one-past-the-end rule. 2010 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2011 return false; 2012 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 2013 if (!cast<Constant>(Idxs[i])->isNullValue()) 2014 return false; 2015 return true; 2016 } 2017 2018 /// Test whether a given ConstantInt is in-range for a SequentialType. 2019 static bool isIndexInRangeOfArrayType(uint64_t NumElements, 2020 const ConstantInt *CI) { 2021 // We cannot bounds check the index if it doesn't fit in an int64_t. 2022 if (CI->getValue().getActiveBits() > 64) 2023 return false; 2024 2025 // A negative index or an index past the end of our sequential type is 2026 // considered out-of-range. 2027 int64_t IndexVal = CI->getSExtValue(); 2028 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2029 return false; 2030 2031 // Otherwise, it is in-range. 2032 return true; 2033 } 2034 2035 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, 2036 bool InBounds, 2037 Optional<unsigned> InRangeIndex, 2038 ArrayRef<Value *> Idxs) { 2039 if (Idxs.empty()) return C; 2040 Constant *Idx0 = cast<Constant>(Idxs[0]); 2041 if ((Idxs.size() == 1 && Idx0->isNullValue())) 2042 return C; 2043 2044 if (isa<UndefValue>(C)) { 2045 Type *GEPTy = GetElementPtrInst::getGEPReturnType( 2046 C, makeArrayRef((Value * const *)Idxs.data(), Idxs.size())); 2047 return UndefValue::get(GEPTy); 2048 } 2049 2050 if (C->isNullValue()) { 2051 bool isNull = true; 2052 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2053 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2054 isNull = false; 2055 break; 2056 } 2057 if (isNull) { 2058 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2059 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2060 2061 assert(Ty && "Invalid indices for GEP!"); 2062 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2063 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2064 GEPTy = VectorType::get(GEPTy, VT->getNumElements()); 2065 return Constant::getNullValue(GEPTy); 2066 } 2067 } 2068 2069 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2070 // Combine Indices - If the source pointer to this getelementptr instruction 2071 // is a getelementptr instruction, combine the indices of the two 2072 // getelementptr instructions into a single instruction. 2073 // 2074 if (CE->getOpcode() == Instruction::GetElementPtr) { 2075 gep_type_iterator LastI = gep_type_end(CE); 2076 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2077 I != E; ++I) 2078 LastI = I; 2079 2080 // We cannot combine indices if doing so would take us outside of an 2081 // array or vector. Doing otherwise could trick us if we evaluated such a 2082 // GEP as part of a load. 2083 // 2084 // e.g. Consider if the original GEP was: 2085 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2086 // i32 0, i32 0, i64 0) 2087 // 2088 // If we then tried to offset it by '8' to get to the third element, 2089 // an i8, we should *not* get: 2090 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2091 // i32 0, i32 0, i64 8) 2092 // 2093 // This GEP tries to index array element '8 which runs out-of-bounds. 2094 // Subsequent evaluation would get confused and produce erroneous results. 2095 // 2096 // The following prohibits such a GEP from being formed by checking to see 2097 // if the index is in-range with respect to an array or vector. 2098 bool PerformFold = false; 2099 if (Idx0->isNullValue()) 2100 PerformFold = true; 2101 else if (LastI.isSequential()) 2102 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2103 PerformFold = 2104 !LastI.isBoundedSequential() || 2105 isIndexInRangeOfArrayType(LastI.getSequentialNumElements(), CI); 2106 2107 if (PerformFold) { 2108 SmallVector<Value*, 16> NewIndices; 2109 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2110 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2111 2112 // Add the last index of the source with the first index of the new GEP. 2113 // Make sure to handle the case when they are actually different types. 2114 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2115 // Otherwise it must be an array. 2116 if (!Idx0->isNullValue()) { 2117 Type *IdxTy = Combined->getType(); 2118 if (IdxTy != Idx0->getType()) { 2119 unsigned CommonExtendedWidth = 2120 std::max(IdxTy->getIntegerBitWidth(), 2121 Idx0->getType()->getIntegerBitWidth()); 2122 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2123 2124 Type *CommonTy = 2125 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2126 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2127 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2128 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2129 } else { 2130 Combined = 2131 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2132 } 2133 } 2134 2135 NewIndices.push_back(Combined); 2136 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2137 2138 // The combined GEP normally inherits its index inrange attribute from 2139 // the inner GEP, but if the inner GEP's last index was adjusted by the 2140 // outer GEP, any inbounds attribute on that index is invalidated. 2141 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex(); 2142 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue()) 2143 IRIndex = None; 2144 2145 return ConstantExpr::getGetElementPtr( 2146 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2147 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(), 2148 IRIndex); 2149 } 2150 } 2151 2152 // Attempt to fold casts to the same type away. For example, folding: 2153 // 2154 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2155 // i64 0, i64 0) 2156 // into: 2157 // 2158 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2159 // 2160 // Don't fold if the cast is changing address spaces. 2161 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2162 PointerType *SrcPtrTy = 2163 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2164 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2165 if (SrcPtrTy && DstPtrTy) { 2166 ArrayType *SrcArrayTy = 2167 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2168 ArrayType *DstArrayTy = 2169 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2170 if (SrcArrayTy && DstArrayTy 2171 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2172 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2173 return ConstantExpr::getGetElementPtr(SrcArrayTy, 2174 (Constant *)CE->getOperand(0), 2175 Idxs, InBounds, InRangeIndex); 2176 } 2177 } 2178 } 2179 2180 // Check to see if any array indices are not within the corresponding 2181 // notional array or vector bounds. If so, try to determine if they can be 2182 // factored out into preceding dimensions. 2183 SmallVector<Constant *, 8> NewIdxs; 2184 Type *Ty = PointeeTy; 2185 Type *Prev = C->getType(); 2186 bool Unknown = !isa<ConstantInt>(Idxs[0]); 2187 for (unsigned i = 1, e = Idxs.size(); i != e; 2188 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2189 auto *CI = dyn_cast<ConstantInt>(Idxs[i]); 2190 if (!CI) { 2191 // We don't know if it's in range or not. 2192 Unknown = true; 2193 continue; 2194 } 2195 if (InRangeIndex && i == *InRangeIndex + 1) { 2196 // If an index is marked inrange, we cannot apply this canonicalization to 2197 // the following index, as that will cause the inrange index to point to 2198 // the wrong element. 2199 continue; 2200 } 2201 if (isa<StructType>(Ty)) { 2202 // The verify makes sure that GEPs into a struct are in range. 2203 continue; 2204 } 2205 auto *STy = cast<SequentialType>(Ty); 2206 if (isa<VectorType>(STy)) { 2207 // There can be awkward padding in after a non-power of two vector. 2208 Unknown = true; 2209 continue; 2210 } 2211 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI)) 2212 // It's in range, skip to the next index. 2213 continue; 2214 if (isa<StructType>(Prev)) { 2215 // It's out of range, but the prior dimension is a struct 2216 // so we can't do anything about it. 2217 Unknown = true; 2218 continue; 2219 } 2220 if (CI->getSExtValue() < 0) { 2221 // It's out of range and negative, don't try to factor it. 2222 Unknown = true; 2223 continue; 2224 } 2225 // It's out of range, but we can factor it into the prior 2226 // dimension. 2227 NewIdxs.resize(Idxs.size()); 2228 // Determine the number of elements in our sequential type. 2229 uint64_t NumElements = STy->getArrayNumElements(); 2230 2231 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2232 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2233 2234 Constant *PrevIdx = cast<Constant>(Idxs[i - 1]); 2235 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2236 2237 unsigned CommonExtendedWidth = 2238 std::max(PrevIdx->getType()->getIntegerBitWidth(), 2239 Div->getType()->getIntegerBitWidth()); 2240 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2241 2242 // Before adding, extend both operands to i64 to avoid 2243 // overflow trouble. 2244 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth)) 2245 PrevIdx = ConstantExpr::getSExt( 2246 PrevIdx, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2247 if (!Div->getType()->isIntegerTy(CommonExtendedWidth)) 2248 Div = ConstantExpr::getSExt( 2249 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2250 2251 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div); 2252 } 2253 2254 // If we did any factoring, start over with the adjusted indices. 2255 if (!NewIdxs.empty()) { 2256 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2257 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2258 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds, 2259 InRangeIndex); 2260 } 2261 2262 // If all indices are known integers and normalized, we can do a simple 2263 // check for the "inbounds" property. 2264 if (!Unknown && !InBounds) 2265 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2266 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2267 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs, 2268 /*InBounds=*/true, InRangeIndex); 2269 2270 return nullptr; 2271 } 2272