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 /// Determine if it is valid to fold a cast of a cast 75 static unsigned 76 foldConstantCastPair( 77 unsigned opc, ///< opcode of the second cast constant expression 78 ConstantExpr *Op, ///< the first cast constant expression 79 Type *DstTy ///< destination type of the first cast 80 ) { 81 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 82 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 83 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 84 85 // The types and opcodes for the two Cast constant expressions 86 Type *SrcTy = Op->getOperand(0)->getType(); 87 Type *MidTy = Op->getType(); 88 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 89 Instruction::CastOps secondOp = Instruction::CastOps(opc); 90 91 // Assume that pointers are never more than 64 bits wide, and only use this 92 // for the middle type. Otherwise we could end up folding away illegal 93 // bitcasts between address spaces with different sizes. 94 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 95 96 // Let CastInst::isEliminableCastPair do the heavy lifting. 97 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 98 nullptr, FakeIntPtrTy, nullptr); 99 } 100 101 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 102 Type *SrcTy = V->getType(); 103 if (SrcTy == DestTy) 104 return V; // no-op cast 105 106 // Check to see if we are casting a pointer to an aggregate to a pointer to 107 // the first element. If so, return the appropriate GEP instruction. 108 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 109 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 110 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 111 && PTy->getElementType()->isSized()) { 112 SmallVector<Value*, 8> IdxList; 113 Value *Zero = 114 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 115 IdxList.push_back(Zero); 116 Type *ElTy = PTy->getElementType(); 117 while (ElTy != DPTy->getElementType()) { 118 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 119 if (STy->getNumElements() == 0) break; 120 ElTy = STy->getElementType(0); 121 IdxList.push_back(Zero); 122 } else if (SequentialType *STy = 123 dyn_cast<SequentialType>(ElTy)) { 124 ElTy = STy->getElementType(); 125 IdxList.push_back(Zero); 126 } else { 127 break; 128 } 129 } 130 131 if (ElTy == DPTy->getElementType()) 132 // This GEP is inbounds because all indices are zero. 133 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), 134 V, IdxList); 135 } 136 137 // Handle casts from one vector constant to another. We know that the src 138 // and dest type have the same size (otherwise its an illegal cast). 139 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 140 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 141 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 142 "Not cast between same sized vectors!"); 143 SrcTy = nullptr; 144 // First, check for null. Undef is already handled. 145 if (isa<ConstantAggregateZero>(V)) 146 return Constant::getNullValue(DestTy); 147 148 // Handle ConstantVector and ConstantAggregateVector. 149 return BitCastConstantVector(V, DestPTy); 150 } 151 152 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 153 // This allows for other simplifications (although some of them 154 // can only be handled by Analysis/ConstantFolding.cpp). 155 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 156 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 157 } 158 159 // Finally, implement bitcast folding now. The code below doesn't handle 160 // bitcast right. 161 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 162 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 163 164 // Handle integral constant input. 165 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 166 if (DestTy->isIntegerTy()) 167 // Integral -> Integral. This is a no-op because the bit widths must 168 // be the same. Consequently, we just fold to V. 169 return V; 170 171 // See note below regarding the PPC_FP128 restriction. 172 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 173 return ConstantFP::get(DestTy->getContext(), 174 APFloat(DestTy->getFltSemantics(), 175 CI->getValue())); 176 177 // Otherwise, can't fold this (vector?) 178 return nullptr; 179 } 180 181 // Handle ConstantFP input: FP -> Integral. 182 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 183 // PPC_FP128 is really the sum of two consecutive doubles, where the first 184 // double is always stored first in memory, regardless of the target 185 // endianness. The memory layout of i128, however, depends on the target 186 // endianness, and so we can't fold this without target endianness 187 // information. This should instead be handled by 188 // Analysis/ConstantFolding.cpp 189 if (FP->getType()->isPPC_FP128Ty()) 190 return nullptr; 191 192 // Make sure dest type is compatible with the folded integer constant. 193 if (!DestTy->isIntegerTy()) 194 return nullptr; 195 196 return ConstantInt::get(FP->getContext(), 197 FP->getValueAPF().bitcastToAPInt()); 198 } 199 200 return nullptr; 201 } 202 203 204 /// V is an integer constant which only has a subset of its bytes used. 205 /// The bytes used are indicated by ByteStart (which is the first byte used, 206 /// counting from the least significant byte) and ByteSize, which is the number 207 /// of bytes used. 208 /// 209 /// This function analyzes the specified constant to see if the specified byte 210 /// range can be returned as a simplified constant. If so, the constant is 211 /// returned, otherwise null is returned. 212 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 213 unsigned ByteSize) { 214 assert(C->getType()->isIntegerTy() && 215 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 216 "Non-byte sized integer input"); 217 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 218 assert(ByteSize && "Must be accessing some piece"); 219 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 220 assert(ByteSize != CSize && "Should not extract everything"); 221 222 // Constant Integers are simple. 223 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 224 APInt V = CI->getValue(); 225 if (ByteStart) 226 V.lshrInPlace(ByteStart*8); 227 V = V.trunc(ByteSize*8); 228 return ConstantInt::get(CI->getContext(), V); 229 } 230 231 // In the input is a constant expr, we might be able to recursively simplify. 232 // If not, we definitely can't do anything. 233 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 234 if (!CE) return nullptr; 235 236 switch (CE->getOpcode()) { 237 default: return nullptr; 238 case Instruction::Or: { 239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 240 if (!RHS) 241 return nullptr; 242 243 // X | -1 -> -1. 244 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 245 if (RHSC->isMinusOne()) 246 return RHSC; 247 248 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 249 if (!LHS) 250 return nullptr; 251 return ConstantExpr::getOr(LHS, RHS); 252 } 253 case Instruction::And: { 254 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 255 if (!RHS) 256 return nullptr; 257 258 // X & 0 -> 0. 259 if (RHS->isNullValue()) 260 return RHS; 261 262 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 263 if (!LHS) 264 return nullptr; 265 return ConstantExpr::getAnd(LHS, RHS); 266 } 267 case Instruction::LShr: { 268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 269 if (!Amt) 270 return nullptr; 271 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 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 // Do not fold if the gep type is a vector, as bitcasting 550 // operand 0 of a vector gep will result in a bitcast between 551 // different sizes. 552 !CE->getType()->isVectorTy()) { 553 // If all of the indexes in the GEP are null values, there is no pointer 554 // adjustment going on. We might as well cast the source pointer. 555 bool isAllNull = true; 556 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 557 if (!CE->getOperand(i)->isNullValue()) { 558 isAllNull = false; 559 break; 560 } 561 if (isAllNull) 562 // This is casting one pointer type to another, always BitCast 563 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 564 } 565 } 566 567 // If the cast operand is a constant vector, perform the cast by 568 // operating on each element. In the cast of bitcasts, the element 569 // count may be mismatched; don't attempt to handle that here. 570 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 571 DestTy->isVectorTy() && 572 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 573 SmallVector<Constant*, 16> res; 574 VectorType *DestVecTy = cast<VectorType>(DestTy); 575 Type *DstEltTy = DestVecTy->getElementType(); 576 Type *Ty = IntegerType::get(V->getContext(), 32); 577 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 578 Constant *C = 579 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 580 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 581 } 582 return ConstantVector::get(res); 583 } 584 585 // We actually have to do a cast now. Perform the cast according to the 586 // opcode specified. 587 switch (opc) { 588 default: 589 llvm_unreachable("Failed to cast constant expression"); 590 case Instruction::FPTrunc: 591 case Instruction::FPExt: 592 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 593 bool ignored; 594 APFloat Val = FPC->getValueAPF(); 595 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() : 596 DestTy->isFloatTy() ? APFloat::IEEEsingle() : 597 DestTy->isDoubleTy() ? APFloat::IEEEdouble() : 598 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() : 599 DestTy->isFP128Ty() ? APFloat::IEEEquad() : 600 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() : 601 APFloat::Bogus(), 602 APFloat::rmNearestTiesToEven, &ignored); 603 return ConstantFP::get(V->getContext(), Val); 604 } 605 return nullptr; // Can't fold. 606 case Instruction::FPToUI: 607 case Instruction::FPToSI: 608 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 609 const APFloat &V = FPC->getValueAPF(); 610 bool ignored; 611 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 612 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI); 613 if (APFloat::opInvalidOp == 614 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) { 615 // Undefined behavior invoked - the destination type can't represent 616 // the input constant. 617 return UndefValue::get(DestTy); 618 } 619 return ConstantInt::get(FPC->getContext(), IntVal); 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 // FIXME: Looks like getFoldedSizeOf(), getFoldedOffsetOf() and 637 // getFoldedAlignOf() don't handle the case when DestTy is a vector of 638 // pointers yet. We end up in asserts in CastInst::getCastOpcode (see 639 // test/Analysis/ConstantFolding/cast-vector.ll). I've only seen this 640 // happen in one "real" C-code test case, so it does not seem to be an 641 // important optimization to handle vectors here. For now, simply bail 642 // out. 643 if (DestTy->isVectorTy()) 644 return nullptr; 645 GEPOperator *GEPO = cast<GEPOperator>(CE); 646 Type *Ty = GEPO->getSourceElementType(); 647 if (CE->getNumOperands() == 2) { 648 // Handle a sizeof-like expression. 649 Constant *Idx = CE->getOperand(1); 650 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 651 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 652 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 653 DestTy, false), 654 Idx, DestTy); 655 return ConstantExpr::getMul(C, Idx); 656 } 657 } else if (CE->getNumOperands() == 3 && 658 CE->getOperand(1)->isNullValue()) { 659 // Handle an alignof-like expression. 660 if (StructType *STy = dyn_cast<StructType>(Ty)) 661 if (!STy->isPacked()) { 662 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 663 if (CI->isOne() && 664 STy->getNumElements() == 2 && 665 STy->getElementType(0)->isIntegerTy(1)) { 666 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 667 } 668 } 669 // Handle an offsetof-like expression. 670 if (Ty->isStructTy() || Ty->isArrayTy()) { 671 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 672 DestTy, false)) 673 return C; 674 } 675 } 676 } 677 // Other pointer types cannot be casted 678 return nullptr; 679 case Instruction::UIToFP: 680 case Instruction::SIToFP: 681 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 682 const APInt &api = CI->getValue(); 683 APFloat apf(DestTy->getFltSemantics(), 684 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 685 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 686 APFloat::rmNearestTiesToEven); 687 return ConstantFP::get(V->getContext(), apf); 688 } 689 return nullptr; 690 case Instruction::ZExt: 691 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 692 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 693 return ConstantInt::get(V->getContext(), 694 CI->getValue().zext(BitWidth)); 695 } 696 return nullptr; 697 case Instruction::SExt: 698 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 699 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 700 return ConstantInt::get(V->getContext(), 701 CI->getValue().sext(BitWidth)); 702 } 703 return nullptr; 704 case Instruction::Trunc: { 705 if (V->getType()->isVectorTy()) 706 return nullptr; 707 708 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 709 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 710 return ConstantInt::get(V->getContext(), 711 CI->getValue().trunc(DestBitWidth)); 712 } 713 714 // The input must be a constantexpr. See if we can simplify this based on 715 // the bytes we are demanding. Only do this if the source and dest are an 716 // even multiple of a byte. 717 if ((DestBitWidth & 7) == 0 && 718 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 719 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 720 return Res; 721 722 return nullptr; 723 } 724 case Instruction::BitCast: 725 return FoldBitCast(V, DestTy); 726 case Instruction::AddrSpaceCast: 727 return nullptr; 728 } 729 } 730 731 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 732 Constant *V1, Constant *V2) { 733 // Check for i1 and vector true/false conditions. 734 if (Cond->isNullValue()) return V2; 735 if (Cond->isAllOnesValue()) return V1; 736 737 // If the condition is a vector constant, fold the result elementwise. 738 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 739 SmallVector<Constant*, 16> Result; 740 Type *Ty = IntegerType::get(CondV->getContext(), 32); 741 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 742 Constant *V; 743 Constant *V1Element = ConstantExpr::getExtractElement(V1, 744 ConstantInt::get(Ty, i)); 745 Constant *V2Element = ConstantExpr::getExtractElement(V2, 746 ConstantInt::get(Ty, i)); 747 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 748 if (V1Element == V2Element) { 749 V = V1Element; 750 } else if (isa<UndefValue>(Cond)) { 751 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 752 } else { 753 if (!isa<ConstantInt>(Cond)) break; 754 V = Cond->isNullValue() ? V2Element : V1Element; 755 } 756 Result.push_back(V); 757 } 758 759 // If we were able to build the vector, return it. 760 if (Result.size() == V1->getType()->getVectorNumElements()) 761 return ConstantVector::get(Result); 762 } 763 764 if (isa<UndefValue>(Cond)) { 765 if (isa<UndefValue>(V1)) return V1; 766 return V2; 767 } 768 if (isa<UndefValue>(V1)) return V2; 769 if (isa<UndefValue>(V2)) return V1; 770 if (V1 == V2) return V1; 771 772 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 773 if (TrueVal->getOpcode() == Instruction::Select) 774 if (TrueVal->getOperand(0) == Cond) 775 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 776 } 777 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 778 if (FalseVal->getOpcode() == Instruction::Select) 779 if (FalseVal->getOperand(0) == Cond) 780 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 781 } 782 783 return nullptr; 784 } 785 786 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 787 Constant *Idx) { 788 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 789 return UndefValue::get(Val->getType()->getVectorElementType()); 790 if (Val->isNullValue()) // ee(zero, x) -> zero 791 return Constant::getNullValue(Val->getType()->getVectorElementType()); 792 // ee({w,x,y,z}, undef) -> undef 793 if (isa<UndefValue>(Idx)) 794 return UndefValue::get(Val->getType()->getVectorElementType()); 795 796 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 797 // ee({w,x,y,z}, wrong_value) -> undef 798 if (CIdx->uge(Val->getType()->getVectorNumElements())) 799 return UndefValue::get(Val->getType()->getVectorElementType()); 800 return Val->getAggregateElement(CIdx->getZExtValue()); 801 } 802 return nullptr; 803 } 804 805 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 806 Constant *Elt, 807 Constant *Idx) { 808 if (isa<UndefValue>(Idx)) 809 return UndefValue::get(Val->getType()); 810 811 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 812 if (!CIdx) return nullptr; 813 814 unsigned NumElts = Val->getType()->getVectorNumElements(); 815 if (CIdx->uge(NumElts)) 816 return UndefValue::get(Val->getType()); 817 818 SmallVector<Constant*, 16> Result; 819 Result.reserve(NumElts); 820 auto *Ty = Type::getInt32Ty(Val->getContext()); 821 uint64_t IdxVal = CIdx->getZExtValue(); 822 for (unsigned i = 0; i != NumElts; ++i) { 823 if (i == IdxVal) { 824 Result.push_back(Elt); 825 continue; 826 } 827 828 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 829 Result.push_back(C); 830 } 831 832 return ConstantVector::get(Result); 833 } 834 835 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 836 Constant *V2, 837 Constant *Mask) { 838 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 839 Type *EltTy = V1->getType()->getVectorElementType(); 840 841 // Undefined shuffle mask -> undefined value. 842 if (isa<UndefValue>(Mask)) 843 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 844 845 // Don't break the bitcode reader hack. 846 if (isa<ConstantExpr>(Mask)) return nullptr; 847 848 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 849 850 // Loop over the shuffle mask, evaluating each element. 851 SmallVector<Constant*, 32> Result; 852 for (unsigned i = 0; i != MaskNumElts; ++i) { 853 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 854 if (Elt == -1) { 855 Result.push_back(UndefValue::get(EltTy)); 856 continue; 857 } 858 Constant *InElt; 859 if (unsigned(Elt) >= SrcNumElts*2) 860 InElt = UndefValue::get(EltTy); 861 else if (unsigned(Elt) >= SrcNumElts) { 862 Type *Ty = IntegerType::get(V2->getContext(), 32); 863 InElt = 864 ConstantExpr::getExtractElement(V2, 865 ConstantInt::get(Ty, Elt - SrcNumElts)); 866 } else { 867 Type *Ty = IntegerType::get(V1->getContext(), 32); 868 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 869 } 870 Result.push_back(InElt); 871 } 872 873 return ConstantVector::get(Result); 874 } 875 876 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 877 ArrayRef<unsigned> Idxs) { 878 // Base case: no indices, so return the entire value. 879 if (Idxs.empty()) 880 return Agg; 881 882 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 883 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 884 885 return nullptr; 886 } 887 888 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 889 Constant *Val, 890 ArrayRef<unsigned> Idxs) { 891 // Base case: no indices, so replace the entire value. 892 if (Idxs.empty()) 893 return Val; 894 895 unsigned NumElts; 896 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 897 NumElts = ST->getNumElements(); 898 else 899 NumElts = cast<SequentialType>(Agg->getType())->getNumElements(); 900 901 SmallVector<Constant*, 32> Result; 902 for (unsigned i = 0; i != NumElts; ++i) { 903 Constant *C = Agg->getAggregateElement(i); 904 if (!C) return nullptr; 905 906 if (Idxs[0] == i) 907 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 908 909 Result.push_back(C); 910 } 911 912 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 913 return ConstantStruct::get(ST, Result); 914 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 915 return ConstantArray::get(AT, Result); 916 return ConstantVector::get(Result); 917 } 918 919 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1, 920 Constant *C2) { 921 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 922 923 // Handle scalar UndefValue. Vectors are always evaluated per element. 924 bool HasScalarUndef = !C1->getType()->isVectorTy() && 925 (isa<UndefValue>(C1) || isa<UndefValue>(C2)); 926 if (HasScalarUndef) { 927 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 928 case Instruction::Xor: 929 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 930 // Handle undef ^ undef -> 0 special case. This is a common 931 // idiom (misuse). 932 return Constant::getNullValue(C1->getType()); 933 LLVM_FALLTHROUGH; 934 case Instruction::Add: 935 case Instruction::Sub: 936 return UndefValue::get(C1->getType()); 937 case Instruction::And: 938 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 939 return C1; 940 return Constant::getNullValue(C1->getType()); // undef & X -> 0 941 case Instruction::Mul: { 942 // undef * undef -> undef 943 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 944 return C1; 945 const APInt *CV; 946 // X * undef -> undef if X is odd 947 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 948 if ((*CV)[0]) 949 return UndefValue::get(C1->getType()); 950 951 // X * undef -> 0 otherwise 952 return Constant::getNullValue(C1->getType()); 953 } 954 case Instruction::SDiv: 955 case Instruction::UDiv: 956 // X / undef -> undef 957 if (isa<UndefValue>(C2)) 958 return C2; 959 // undef / 0 -> undef 960 // undef / 1 -> undef 961 if (match(C2, m_Zero()) || match(C2, m_One())) 962 return C1; 963 // undef / X -> 0 otherwise 964 return Constant::getNullValue(C1->getType()); 965 case Instruction::URem: 966 case Instruction::SRem: 967 // X % undef -> undef 968 if (match(C2, m_Undef())) 969 return C2; 970 // undef % 0 -> undef 971 if (match(C2, m_Zero())) 972 return C1; 973 // undef % X -> 0 otherwise 974 return Constant::getNullValue(C1->getType()); 975 case Instruction::Or: // X | undef -> -1 976 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 977 return C1; 978 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 979 case Instruction::LShr: 980 // X >>l undef -> undef 981 if (isa<UndefValue>(C2)) 982 return C2; 983 // undef >>l 0 -> undef 984 if (match(C2, m_Zero())) 985 return C1; 986 // undef >>l X -> 0 987 return Constant::getNullValue(C1->getType()); 988 case Instruction::AShr: 989 // X >>a undef -> undef 990 if (isa<UndefValue>(C2)) 991 return C2; 992 // undef >>a 0 -> undef 993 if (match(C2, m_Zero())) 994 return C1; 995 // TODO: undef >>a X -> undef if the shift is exact 996 // undef >>a X -> 0 997 return Constant::getNullValue(C1->getType()); 998 case Instruction::Shl: 999 // X << undef -> undef 1000 if (isa<UndefValue>(C2)) 1001 return C2; 1002 // undef << 0 -> undef 1003 if (match(C2, m_Zero())) 1004 return C1; 1005 // undef << X -> 0 1006 return Constant::getNullValue(C1->getType()); 1007 case Instruction::FAdd: 1008 case Instruction::FSub: 1009 case Instruction::FMul: 1010 case Instruction::FDiv: 1011 case Instruction::FRem: 1012 // [any flop] undef, undef -> undef 1013 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1014 return C1; 1015 // [any flop] C, undef -> NaN 1016 // [any flop] undef, C -> NaN 1017 // We could potentially specialize NaN/Inf constants vs. 'normal' 1018 // constants (possibly differently depending on opcode and operand). This 1019 // would allow returning undef sometimes. But it is always safe to fold to 1020 // NaN because we can choose the undef operand as NaN, and any FP opcode 1021 // with a NaN operand will propagate NaN. 1022 return ConstantFP::getNaN(C1->getType()); 1023 case Instruction::BinaryOpsEnd: 1024 llvm_unreachable("Invalid BinaryOp"); 1025 } 1026 } 1027 1028 // Neither constant should be UndefValue, unless these are vector constants. 1029 assert(!HasScalarUndef && "Unexpected UndefValue"); 1030 1031 // Handle simplifications when the RHS is a constant int. 1032 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1033 switch (Opcode) { 1034 case Instruction::Add: 1035 if (CI2->isZero()) return C1; // X + 0 == X 1036 break; 1037 case Instruction::Sub: 1038 if (CI2->isZero()) return C1; // X - 0 == X 1039 break; 1040 case Instruction::Mul: 1041 if (CI2->isZero()) return C2; // X * 0 == 0 1042 if (CI2->isOne()) 1043 return C1; // X * 1 == X 1044 break; 1045 case Instruction::UDiv: 1046 case Instruction::SDiv: 1047 if (CI2->isOne()) 1048 return C1; // X / 1 == X 1049 if (CI2->isZero()) 1050 return UndefValue::get(CI2->getType()); // X / 0 == undef 1051 break; 1052 case Instruction::URem: 1053 case Instruction::SRem: 1054 if (CI2->isOne()) 1055 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1056 if (CI2->isZero()) 1057 return UndefValue::get(CI2->getType()); // X % 0 == undef 1058 break; 1059 case Instruction::And: 1060 if (CI2->isZero()) return C2; // X & 0 == 0 1061 if (CI2->isMinusOne()) 1062 return C1; // X & -1 == X 1063 1064 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1065 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1066 if (CE1->getOpcode() == Instruction::ZExt) { 1067 unsigned DstWidth = CI2->getType()->getBitWidth(); 1068 unsigned SrcWidth = 1069 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1070 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1071 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1072 return C1; 1073 } 1074 1075 // If and'ing the address of a global with a constant, fold it. 1076 if (CE1->getOpcode() == Instruction::PtrToInt && 1077 isa<GlobalValue>(CE1->getOperand(0))) { 1078 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1079 1080 // Functions are at least 4-byte aligned. 1081 unsigned GVAlign = GV->getAlignment(); 1082 if (isa<Function>(GV)) 1083 GVAlign = std::max(GVAlign, 4U); 1084 1085 if (GVAlign > 1) { 1086 unsigned DstWidth = CI2->getType()->getBitWidth(); 1087 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1088 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1089 1090 // If checking bits we know are clear, return zero. 1091 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1092 return Constant::getNullValue(CI2->getType()); 1093 } 1094 } 1095 } 1096 break; 1097 case Instruction::Or: 1098 if (CI2->isZero()) return C1; // X | 0 == X 1099 if (CI2->isMinusOne()) 1100 return C2; // X | -1 == -1 1101 break; 1102 case Instruction::Xor: 1103 if (CI2->isZero()) return C1; // X ^ 0 == X 1104 1105 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1106 switch (CE1->getOpcode()) { 1107 default: break; 1108 case Instruction::ICmp: 1109 case Instruction::FCmp: 1110 // cmp pred ^ true -> cmp !pred 1111 assert(CI2->isOne()); 1112 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1113 pred = CmpInst::getInversePredicate(pred); 1114 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1115 CE1->getOperand(1)); 1116 } 1117 } 1118 break; 1119 case Instruction::AShr: 1120 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1121 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1122 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1123 return ConstantExpr::getLShr(C1, C2); 1124 break; 1125 } 1126 } else if (isa<ConstantInt>(C1)) { 1127 // If C1 is a ConstantInt and C2 is not, swap the operands. 1128 if (Instruction::isCommutative(Opcode)) 1129 return ConstantExpr::get(Opcode, C2, C1); 1130 } 1131 1132 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1133 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1134 const APInt &C1V = CI1->getValue(); 1135 const APInt &C2V = CI2->getValue(); 1136 switch (Opcode) { 1137 default: 1138 break; 1139 case Instruction::Add: 1140 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1141 case Instruction::Sub: 1142 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1143 case Instruction::Mul: 1144 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1145 case Instruction::UDiv: 1146 assert(!CI2->isZero() && "Div by zero handled above"); 1147 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1148 case Instruction::SDiv: 1149 assert(!CI2->isZero() && "Div by zero handled above"); 1150 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1151 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1152 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1153 case Instruction::URem: 1154 assert(!CI2->isZero() && "Div by zero handled above"); 1155 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1156 case Instruction::SRem: 1157 assert(!CI2->isZero() && "Div by zero handled above"); 1158 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1159 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1160 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1161 case Instruction::And: 1162 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1163 case Instruction::Or: 1164 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1165 case Instruction::Xor: 1166 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1167 case Instruction::Shl: 1168 if (C2V.ult(C1V.getBitWidth())) 1169 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1170 return UndefValue::get(C1->getType()); // too big shift is undef 1171 case Instruction::LShr: 1172 if (C2V.ult(C1V.getBitWidth())) 1173 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1174 return UndefValue::get(C1->getType()); // too big shift is undef 1175 case Instruction::AShr: 1176 if (C2V.ult(C1V.getBitWidth())) 1177 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1178 return UndefValue::get(C1->getType()); // too big shift is undef 1179 } 1180 } 1181 1182 switch (Opcode) { 1183 case Instruction::SDiv: 1184 case Instruction::UDiv: 1185 case Instruction::URem: 1186 case Instruction::SRem: 1187 case Instruction::LShr: 1188 case Instruction::AShr: 1189 case Instruction::Shl: 1190 if (CI1->isZero()) return C1; 1191 break; 1192 default: 1193 break; 1194 } 1195 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1196 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1197 const APFloat &C1V = CFP1->getValueAPF(); 1198 const APFloat &C2V = CFP2->getValueAPF(); 1199 APFloat C3V = C1V; // copy for modification 1200 switch (Opcode) { 1201 default: 1202 break; 1203 case Instruction::FAdd: 1204 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1205 return ConstantFP::get(C1->getContext(), C3V); 1206 case Instruction::FSub: 1207 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1208 return ConstantFP::get(C1->getContext(), C3V); 1209 case Instruction::FMul: 1210 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1211 return ConstantFP::get(C1->getContext(), C3V); 1212 case Instruction::FDiv: 1213 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1214 return ConstantFP::get(C1->getContext(), C3V); 1215 case Instruction::FRem: 1216 (void)C3V.mod(C2V); 1217 return ConstantFP::get(C1->getContext(), C3V); 1218 } 1219 } 1220 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1221 // Fold each element and create a vector constant from those constants. 1222 SmallVector<Constant*, 16> Result; 1223 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1224 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1225 Constant *ExtractIdx = ConstantInt::get(Ty, i); 1226 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx); 1227 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx); 1228 1229 // If any element of a divisor vector is zero, the whole op is undef. 1230 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue()) 1231 return UndefValue::get(VTy); 1232 1233 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1234 } 1235 1236 return ConstantVector::get(Result); 1237 } 1238 1239 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1240 // There are many possible foldings we could do here. We should probably 1241 // at least fold add of a pointer with an integer into the appropriate 1242 // getelementptr. This will improve alias analysis a bit. 1243 1244 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1245 // (a + (b + c)). 1246 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1247 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1248 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1249 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1250 } 1251 } else if (isa<ConstantExpr>(C2)) { 1252 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1253 // other way if possible. 1254 if (Instruction::isCommutative(Opcode)) 1255 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1256 } 1257 1258 // i1 can be simplified in many cases. 1259 if (C1->getType()->isIntegerTy(1)) { 1260 switch (Opcode) { 1261 case Instruction::Add: 1262 case Instruction::Sub: 1263 return ConstantExpr::getXor(C1, C2); 1264 case Instruction::Mul: 1265 return ConstantExpr::getAnd(C1, C2); 1266 case Instruction::Shl: 1267 case Instruction::LShr: 1268 case Instruction::AShr: 1269 // We can assume that C2 == 0. If it were one the result would be 1270 // undefined because the shift value is as large as the bitwidth. 1271 return C1; 1272 case Instruction::SDiv: 1273 case Instruction::UDiv: 1274 // We can assume that C2 == 1. If it were zero the result would be 1275 // undefined through division by zero. 1276 return C1; 1277 case Instruction::URem: 1278 case Instruction::SRem: 1279 // We can assume that C2 == 1. If it were zero the result would be 1280 // undefined through division by zero. 1281 return ConstantInt::getFalse(C1->getContext()); 1282 default: 1283 break; 1284 } 1285 } 1286 1287 // We don't know how to fold this. 1288 return nullptr; 1289 } 1290 1291 /// This type is zero-sized if it's an array or structure of zero-sized types. 1292 /// The only leaf zero-sized type is an empty structure. 1293 static bool isMaybeZeroSizedType(Type *Ty) { 1294 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1295 if (STy->isOpaque()) return true; // Can't say. 1296 1297 // If all of elements have zero size, this does too. 1298 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1299 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1300 return true; 1301 1302 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1303 return isMaybeZeroSizedType(ATy->getElementType()); 1304 } 1305 return false; 1306 } 1307 1308 /// Compare the two constants as though they were getelementptr indices. 1309 /// This allows coercion of the types to be the same thing. 1310 /// 1311 /// If the two constants are the "same" (after coercion), return 0. If the 1312 /// first is less than the second, return -1, if the second is less than the 1313 /// first, return 1. If the constants are not integral, return -2. 1314 /// 1315 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1316 if (C1 == C2) return 0; 1317 1318 // Ok, we found a different index. If they are not ConstantInt, we can't do 1319 // anything with them. 1320 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1321 return -2; // don't know! 1322 1323 // We cannot compare the indices if they don't fit in an int64_t. 1324 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 || 1325 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64) 1326 return -2; // don't know! 1327 1328 // Ok, we have two differing integer indices. Sign extend them to be the same 1329 // type. 1330 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue(); 1331 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue(); 1332 1333 if (C1Val == C2Val) return 0; // They are equal 1334 1335 // If the type being indexed over is really just a zero sized type, there is 1336 // no pointer difference being made here. 1337 if (isMaybeZeroSizedType(ElTy)) 1338 return -2; // dunno. 1339 1340 // If they are really different, now that they are the same type, then we 1341 // found a difference! 1342 if (C1Val < C2Val) 1343 return -1; 1344 else 1345 return 1; 1346 } 1347 1348 /// This function determines if there is anything we can decide about the two 1349 /// constants provided. This doesn't need to handle simple things like 1350 /// ConstantFP comparisons, but should instead handle ConstantExprs. 1351 /// If we can determine that the two constants have a particular relation to 1352 /// each other, we should return the corresponding FCmpInst predicate, 1353 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1354 /// ConstantFoldCompareInstruction. 1355 /// 1356 /// To simplify this code we canonicalize the relation so that the first 1357 /// operand is always the most "complex" of the two. We consider ConstantFP 1358 /// to be the simplest, and ConstantExprs to be the most complex. 1359 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1360 assert(V1->getType() == V2->getType() && 1361 "Cannot compare values of different types!"); 1362 1363 // Handle degenerate case quickly 1364 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1365 1366 if (!isa<ConstantExpr>(V1)) { 1367 if (!isa<ConstantExpr>(V2)) { 1368 // Simple case, use the standard constant folder. 1369 ConstantInt *R = nullptr; 1370 R = dyn_cast<ConstantInt>( 1371 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1372 if (R && !R->isZero()) 1373 return FCmpInst::FCMP_OEQ; 1374 R = dyn_cast<ConstantInt>( 1375 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1376 if (R && !R->isZero()) 1377 return FCmpInst::FCMP_OLT; 1378 R = dyn_cast<ConstantInt>( 1379 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1380 if (R && !R->isZero()) 1381 return FCmpInst::FCMP_OGT; 1382 1383 // Nothing more we can do 1384 return FCmpInst::BAD_FCMP_PREDICATE; 1385 } 1386 1387 // If the first operand is simple and second is ConstantExpr, swap operands. 1388 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1389 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1390 return FCmpInst::getSwappedPredicate(SwappedRelation); 1391 } else { 1392 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1393 // constantexpr or a simple constant. 1394 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1395 switch (CE1->getOpcode()) { 1396 case Instruction::FPTrunc: 1397 case Instruction::FPExt: 1398 case Instruction::UIToFP: 1399 case Instruction::SIToFP: 1400 // We might be able to do something with these but we don't right now. 1401 break; 1402 default: 1403 break; 1404 } 1405 } 1406 // There are MANY other foldings that we could perform here. They will 1407 // probably be added on demand, as they seem needed. 1408 return FCmpInst::BAD_FCMP_PREDICATE; 1409 } 1410 1411 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1412 const GlobalValue *GV2) { 1413 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1414 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1415 return true; 1416 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1417 Type *Ty = GVar->getValueType(); 1418 // A global with opaque type might end up being zero sized. 1419 if (!Ty->isSized()) 1420 return true; 1421 // A global with an empty type might lie at the address of any other 1422 // global. 1423 if (Ty->isEmptyTy()) 1424 return true; 1425 } 1426 return false; 1427 }; 1428 // Don't try to decide equality of aliases. 1429 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1430 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1431 return ICmpInst::ICMP_NE; 1432 return ICmpInst::BAD_ICMP_PREDICATE; 1433 } 1434 1435 /// This function determines if there is anything we can decide about the two 1436 /// constants provided. This doesn't need to handle simple things like integer 1437 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1438 /// If we can determine that the two constants have a particular relation to 1439 /// each other, we should return the corresponding ICmp predicate, otherwise 1440 /// return ICmpInst::BAD_ICMP_PREDICATE. 1441 /// 1442 /// To simplify this code we canonicalize the relation so that the first 1443 /// operand is always the most "complex" of the two. We consider simple 1444 /// constants (like ConstantInt) to be the simplest, followed by 1445 /// GlobalValues, followed by ConstantExpr's (the most complex). 1446 /// 1447 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1448 bool isSigned) { 1449 assert(V1->getType() == V2->getType() && 1450 "Cannot compare different types of values!"); 1451 if (V1 == V2) return ICmpInst::ICMP_EQ; 1452 1453 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1454 !isa<BlockAddress>(V1)) { 1455 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1456 !isa<BlockAddress>(V2)) { 1457 // We distilled this down to a simple case, use the standard constant 1458 // folder. 1459 ConstantInt *R = nullptr; 1460 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1461 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1462 if (R && !R->isZero()) 1463 return pred; 1464 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1465 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1466 if (R && !R->isZero()) 1467 return pred; 1468 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1469 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1470 if (R && !R->isZero()) 1471 return pred; 1472 1473 // If we couldn't figure it out, bail. 1474 return ICmpInst::BAD_ICMP_PREDICATE; 1475 } 1476 1477 // If the first operand is simple, swap operands. 1478 ICmpInst::Predicate SwappedRelation = 1479 evaluateICmpRelation(V2, V1, isSigned); 1480 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1481 return ICmpInst::getSwappedPredicate(SwappedRelation); 1482 1483 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1484 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1485 ICmpInst::Predicate SwappedRelation = 1486 evaluateICmpRelation(V2, V1, isSigned); 1487 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1488 return ICmpInst::getSwappedPredicate(SwappedRelation); 1489 return ICmpInst::BAD_ICMP_PREDICATE; 1490 } 1491 1492 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1493 // constant (which, since the types must match, means that it's a 1494 // ConstantPointerNull). 1495 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1496 return areGlobalsPotentiallyEqual(GV, GV2); 1497 } else if (isa<BlockAddress>(V2)) { 1498 return ICmpInst::ICMP_NE; // Globals never equal labels. 1499 } else { 1500 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1501 // GlobalVals can never be null unless they have external weak linkage. 1502 // We don't try to evaluate aliases here. 1503 // NOTE: We should not be doing this constant folding if null pointer 1504 // is considered valid for the function. But currently there is no way to 1505 // query it from the Constant type. 1506 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) && 1507 !NullPointerIsDefined(nullptr /* F */, 1508 GV->getType()->getAddressSpace())) 1509 return ICmpInst::ICMP_NE; 1510 } 1511 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1512 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1513 ICmpInst::Predicate SwappedRelation = 1514 evaluateICmpRelation(V2, V1, isSigned); 1515 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1516 return ICmpInst::getSwappedPredicate(SwappedRelation); 1517 return ICmpInst::BAD_ICMP_PREDICATE; 1518 } 1519 1520 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1521 // constant (which, since the types must match, means that it is a 1522 // ConstantPointerNull). 1523 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1524 // Block address in another function can't equal this one, but block 1525 // addresses in the current function might be the same if blocks are 1526 // empty. 1527 if (BA2->getFunction() != BA->getFunction()) 1528 return ICmpInst::ICMP_NE; 1529 } else { 1530 // Block addresses aren't null, don't equal the address of globals. 1531 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1532 "Canonicalization guarantee!"); 1533 return ICmpInst::ICMP_NE; 1534 } 1535 } else { 1536 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1537 // constantexpr, a global, block address, or a simple constant. 1538 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1539 Constant *CE1Op0 = CE1->getOperand(0); 1540 1541 switch (CE1->getOpcode()) { 1542 case Instruction::Trunc: 1543 case Instruction::FPTrunc: 1544 case Instruction::FPExt: 1545 case Instruction::FPToUI: 1546 case Instruction::FPToSI: 1547 break; // We can't evaluate floating point casts or truncations. 1548 1549 case Instruction::UIToFP: 1550 case Instruction::SIToFP: 1551 case Instruction::BitCast: 1552 case Instruction::ZExt: 1553 case Instruction::SExt: 1554 // We can't evaluate floating point casts or truncations. 1555 if (CE1Op0->getType()->isFloatingPointTy()) 1556 break; 1557 1558 // If the cast is not actually changing bits, and the second operand is a 1559 // null pointer, do the comparison with the pre-casted value. 1560 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) { 1561 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1562 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1563 return evaluateICmpRelation(CE1Op0, 1564 Constant::getNullValue(CE1Op0->getType()), 1565 isSigned); 1566 } 1567 break; 1568 1569 case Instruction::GetElementPtr: { 1570 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1571 // Ok, since this is a getelementptr, we know that the constant has a 1572 // pointer type. Check the various cases. 1573 if (isa<ConstantPointerNull>(V2)) { 1574 // If we are comparing a GEP to a null pointer, check to see if the base 1575 // of the GEP equals the null pointer. 1576 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1577 if (GV->hasExternalWeakLinkage()) 1578 // Weak linkage GVals could be zero or not. We're comparing that 1579 // to null pointer so its greater-or-equal 1580 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1581 else 1582 // If its not weak linkage, the GVal must have a non-zero address 1583 // so the result is greater-than 1584 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1585 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1586 // If we are indexing from a null pointer, check to see if we have any 1587 // non-zero indices. 1588 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1589 if (!CE1->getOperand(i)->isNullValue()) 1590 // Offsetting from null, must not be equal. 1591 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1592 // Only zero indexes from null, must still be zero. 1593 return ICmpInst::ICMP_EQ; 1594 } 1595 // Otherwise, we can't really say if the first operand is null or not. 1596 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1597 if (isa<ConstantPointerNull>(CE1Op0)) { 1598 if (GV2->hasExternalWeakLinkage()) 1599 // Weak linkage GVals could be zero or not. We're comparing it to 1600 // a null pointer, so its less-or-equal 1601 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1602 else 1603 // If its not weak linkage, the GVal must have a non-zero address 1604 // so the result is less-than 1605 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1606 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1607 if (GV == GV2) { 1608 // If this is a getelementptr of the same global, then it must be 1609 // different. Because the types must match, the getelementptr could 1610 // only have at most one index, and because we fold getelementptr's 1611 // with a single zero index, it must be nonzero. 1612 assert(CE1->getNumOperands() == 2 && 1613 !CE1->getOperand(1)->isNullValue() && 1614 "Surprising getelementptr!"); 1615 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1616 } else { 1617 if (CE1GEP->hasAllZeroIndices()) 1618 return areGlobalsPotentiallyEqual(GV, GV2); 1619 return ICmpInst::BAD_ICMP_PREDICATE; 1620 } 1621 } 1622 } else { 1623 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1624 Constant *CE2Op0 = CE2->getOperand(0); 1625 1626 // There are MANY other foldings that we could perform here. They will 1627 // probably be added on demand, as they seem needed. 1628 switch (CE2->getOpcode()) { 1629 default: break; 1630 case Instruction::GetElementPtr: 1631 // By far the most common case to handle is when the base pointers are 1632 // obviously to the same global. 1633 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1634 // Don't know relative ordering, but check for inequality. 1635 if (CE1Op0 != CE2Op0) { 1636 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1637 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1638 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1639 cast<GlobalValue>(CE2Op0)); 1640 return ICmpInst::BAD_ICMP_PREDICATE; 1641 } 1642 // Ok, we know that both getelementptr instructions are based on the 1643 // same global. From this, we can precisely determine the relative 1644 // ordering of the resultant pointers. 1645 unsigned i = 1; 1646 1647 // The logic below assumes that the result of the comparison 1648 // can be determined by finding the first index that differs. 1649 // This doesn't work if there is over-indexing in any 1650 // subsequent indices, so check for that case first. 1651 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1652 !CE2->isGEPWithNoNotionalOverIndexing()) 1653 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1654 1655 // Compare all of the operands the GEP's have in common. 1656 gep_type_iterator GTI = gep_type_begin(CE1); 1657 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1658 ++i, ++GTI) 1659 switch (IdxCompare(CE1->getOperand(i), 1660 CE2->getOperand(i), GTI.getIndexedType())) { 1661 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1662 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1663 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1664 } 1665 1666 // Ok, we ran out of things they have in common. If any leftovers 1667 // are non-zero then we have a difference, otherwise we are equal. 1668 for (; i < CE1->getNumOperands(); ++i) 1669 if (!CE1->getOperand(i)->isNullValue()) { 1670 if (isa<ConstantInt>(CE1->getOperand(i))) 1671 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1672 else 1673 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1674 } 1675 1676 for (; i < CE2->getNumOperands(); ++i) 1677 if (!CE2->getOperand(i)->isNullValue()) { 1678 if (isa<ConstantInt>(CE2->getOperand(i))) 1679 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1680 else 1681 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1682 } 1683 return ICmpInst::ICMP_EQ; 1684 } 1685 } 1686 } 1687 break; 1688 } 1689 default: 1690 break; 1691 } 1692 } 1693 1694 return ICmpInst::BAD_ICMP_PREDICATE; 1695 } 1696 1697 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1698 Constant *C1, Constant *C2) { 1699 Type *ResultTy; 1700 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1701 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1702 VT->getNumElements()); 1703 else 1704 ResultTy = Type::getInt1Ty(C1->getContext()); 1705 1706 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1707 if (pred == FCmpInst::FCMP_FALSE) 1708 return Constant::getNullValue(ResultTy); 1709 1710 if (pred == FCmpInst::FCMP_TRUE) 1711 return Constant::getAllOnesValue(ResultTy); 1712 1713 // Handle some degenerate cases first 1714 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1715 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1716 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1717 // For EQ and NE, we can always pick a value for the undef to make the 1718 // predicate pass or fail, so we can return undef. 1719 // Also, if both operands are undef, we can return undef for int comparison. 1720 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1721 return UndefValue::get(ResultTy); 1722 1723 // Otherwise, for integer compare, pick the same value as the non-undef 1724 // operand, and fold it to true or false. 1725 if (isIntegerPredicate) 1726 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1727 1728 // Choosing NaN for the undef will always make unordered comparison succeed 1729 // and ordered comparison fails. 1730 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1731 } 1732 1733 // icmp eq/ne(null,GV) -> false/true 1734 if (C1->isNullValue()) { 1735 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1736 // Don't try to evaluate aliases. External weak GV can be null. 1737 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1738 !NullPointerIsDefined(nullptr /* F */, 1739 GV->getType()->getAddressSpace())) { 1740 if (pred == ICmpInst::ICMP_EQ) 1741 return ConstantInt::getFalse(C1->getContext()); 1742 else if (pred == ICmpInst::ICMP_NE) 1743 return ConstantInt::getTrue(C1->getContext()); 1744 } 1745 // icmp eq/ne(GV,null) -> false/true 1746 } else if (C2->isNullValue()) { 1747 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1748 // Don't try to evaluate aliases. External weak GV can be null. 1749 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1750 !NullPointerIsDefined(nullptr /* F */, 1751 GV->getType()->getAddressSpace())) { 1752 if (pred == ICmpInst::ICMP_EQ) 1753 return ConstantInt::getFalse(C1->getContext()); 1754 else if (pred == ICmpInst::ICMP_NE) 1755 return ConstantInt::getTrue(C1->getContext()); 1756 } 1757 } 1758 1759 // If the comparison is a comparison between two i1's, simplify it. 1760 if (C1->getType()->isIntegerTy(1)) { 1761 switch(pred) { 1762 case ICmpInst::ICMP_EQ: 1763 if (isa<ConstantInt>(C2)) 1764 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1765 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1766 case ICmpInst::ICMP_NE: 1767 return ConstantExpr::getXor(C1, C2); 1768 default: 1769 break; 1770 } 1771 } 1772 1773 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1774 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1775 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1776 switch (pred) { 1777 default: llvm_unreachable("Invalid ICmp Predicate"); 1778 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1779 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1780 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1781 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1782 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1783 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1784 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1785 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1786 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1787 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1788 } 1789 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1790 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1791 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1792 APFloat::cmpResult R = C1V.compare(C2V); 1793 switch (pred) { 1794 default: llvm_unreachable("Invalid FCmp Predicate"); 1795 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1796 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1797 case FCmpInst::FCMP_UNO: 1798 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1799 case FCmpInst::FCMP_ORD: 1800 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1801 case FCmpInst::FCMP_UEQ: 1802 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1803 R==APFloat::cmpEqual); 1804 case FCmpInst::FCMP_OEQ: 1805 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1806 case FCmpInst::FCMP_UNE: 1807 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1808 case FCmpInst::FCMP_ONE: 1809 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1810 R==APFloat::cmpGreaterThan); 1811 case FCmpInst::FCMP_ULT: 1812 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1813 R==APFloat::cmpLessThan); 1814 case FCmpInst::FCMP_OLT: 1815 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1816 case FCmpInst::FCMP_UGT: 1817 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1818 R==APFloat::cmpGreaterThan); 1819 case FCmpInst::FCMP_OGT: 1820 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1821 case FCmpInst::FCMP_ULE: 1822 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1823 case FCmpInst::FCMP_OLE: 1824 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1825 R==APFloat::cmpEqual); 1826 case FCmpInst::FCMP_UGE: 1827 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1828 case FCmpInst::FCMP_OGE: 1829 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1830 R==APFloat::cmpEqual); 1831 } 1832 } else if (C1->getType()->isVectorTy()) { 1833 // If we can constant fold the comparison of each element, constant fold 1834 // the whole vector comparison. 1835 SmallVector<Constant*, 4> ResElts; 1836 Type *Ty = IntegerType::get(C1->getContext(), 32); 1837 // Compare the elements, producing an i1 result or constant expr. 1838 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1839 Constant *C1E = 1840 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1841 Constant *C2E = 1842 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1843 1844 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1845 } 1846 1847 return ConstantVector::get(ResElts); 1848 } 1849 1850 if (C1->getType()->isFloatingPointTy() && 1851 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1852 // infinite recursive loop 1853 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1854 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1855 switch (evaluateFCmpRelation(C1, C2)) { 1856 default: llvm_unreachable("Unknown relation!"); 1857 case FCmpInst::FCMP_UNO: 1858 case FCmpInst::FCMP_ORD: 1859 case FCmpInst::FCMP_UEQ: 1860 case FCmpInst::FCMP_UNE: 1861 case FCmpInst::FCMP_ULT: 1862 case FCmpInst::FCMP_UGT: 1863 case FCmpInst::FCMP_ULE: 1864 case FCmpInst::FCMP_UGE: 1865 case FCmpInst::FCMP_TRUE: 1866 case FCmpInst::FCMP_FALSE: 1867 case FCmpInst::BAD_FCMP_PREDICATE: 1868 break; // Couldn't determine anything about these constants. 1869 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1870 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1871 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1872 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1873 break; 1874 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1875 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1876 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1877 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1878 break; 1879 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1880 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1881 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1882 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1883 break; 1884 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1885 // We can only partially decide this relation. 1886 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1887 Result = 0; 1888 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1889 Result = 1; 1890 break; 1891 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1892 // We can only partially decide this relation. 1893 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1894 Result = 0; 1895 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1896 Result = 1; 1897 break; 1898 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1899 // We can only partially decide this relation. 1900 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1901 Result = 0; 1902 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1903 Result = 1; 1904 break; 1905 } 1906 1907 // If we evaluated the result, return it now. 1908 if (Result != -1) 1909 return ConstantInt::get(ResultTy, Result); 1910 1911 } else { 1912 // Evaluate the relation between the two constants, per the predicate. 1913 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1914 switch (evaluateICmpRelation(C1, C2, 1915 CmpInst::isSigned((CmpInst::Predicate)pred))) { 1916 default: llvm_unreachable("Unknown relational!"); 1917 case ICmpInst::BAD_ICMP_PREDICATE: 1918 break; // Couldn't determine anything about these constants. 1919 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1920 // If we know the constants are equal, we can decide the result of this 1921 // computation precisely. 1922 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1923 break; 1924 case ICmpInst::ICMP_ULT: 1925 switch (pred) { 1926 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1927 Result = 1; break; 1928 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1929 Result = 0; break; 1930 } 1931 break; 1932 case ICmpInst::ICMP_SLT: 1933 switch (pred) { 1934 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1935 Result = 1; break; 1936 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1937 Result = 0; break; 1938 } 1939 break; 1940 case ICmpInst::ICMP_UGT: 1941 switch (pred) { 1942 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1943 Result = 1; break; 1944 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1945 Result = 0; break; 1946 } 1947 break; 1948 case ICmpInst::ICMP_SGT: 1949 switch (pred) { 1950 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1951 Result = 1; break; 1952 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1953 Result = 0; break; 1954 } 1955 break; 1956 case ICmpInst::ICMP_ULE: 1957 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1958 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1959 break; 1960 case ICmpInst::ICMP_SLE: 1961 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1962 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1963 break; 1964 case ICmpInst::ICMP_UGE: 1965 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1966 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1967 break; 1968 case ICmpInst::ICMP_SGE: 1969 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1970 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1971 break; 1972 case ICmpInst::ICMP_NE: 1973 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1974 if (pred == ICmpInst::ICMP_NE) Result = 1; 1975 break; 1976 } 1977 1978 // If we evaluated the result, return it now. 1979 if (Result != -1) 1980 return ConstantInt::get(ResultTy, Result); 1981 1982 // If the right hand side is a bitcast, try using its inverse to simplify 1983 // it by moving it to the left hand side. We can't do this if it would turn 1984 // a vector compare into a scalar compare or visa versa. 1985 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1986 Constant *CE2Op0 = CE2->getOperand(0); 1987 if (CE2->getOpcode() == Instruction::BitCast && 1988 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1989 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1990 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1991 } 1992 } 1993 1994 // If the left hand side is an extension, try eliminating it. 1995 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1996 if ((CE1->getOpcode() == Instruction::SExt && 1997 ICmpInst::isSigned((ICmpInst::Predicate)pred)) || 1998 (CE1->getOpcode() == Instruction::ZExt && 1999 !ICmpInst::isSigned((ICmpInst::Predicate)pred))){ 2000 Constant *CE1Op0 = CE1->getOperand(0); 2001 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 2002 if (CE1Inverse == CE1Op0) { 2003 // Check whether we can safely truncate the right hand side. 2004 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 2005 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 2006 C2->getType()) == C2) 2007 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 2008 } 2009 } 2010 } 2011 2012 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 2013 (C1->isNullValue() && !C2->isNullValue())) { 2014 // If C2 is a constant expr and C1 isn't, flip them around and fold the 2015 // other way if possible. 2016 // Also, if C1 is null and C2 isn't, flip them around. 2017 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 2018 return ConstantExpr::getICmp(pred, C2, C1); 2019 } 2020 } 2021 return nullptr; 2022 } 2023 2024 /// Test whether the given sequence of *normalized* indices is "inbounds". 2025 template<typename IndexTy> 2026 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 2027 // No indices means nothing that could be out of bounds. 2028 if (Idxs.empty()) return true; 2029 2030 // If the first index is zero, it's in bounds. 2031 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2032 2033 // If the first index is one and all the rest are zero, it's in bounds, 2034 // by the one-past-the-end rule. 2035 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) { 2036 if (!CI->isOne()) 2037 return false; 2038 } else { 2039 auto *CV = cast<ConstantDataVector>(Idxs[0]); 2040 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue()); 2041 if (!CI || !CI->isOne()) 2042 return false; 2043 } 2044 2045 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 2046 if (!cast<Constant>(Idxs[i])->isNullValue()) 2047 return false; 2048 return true; 2049 } 2050 2051 /// Test whether a given ConstantInt is in-range for a SequentialType. 2052 static bool isIndexInRangeOfArrayType(uint64_t NumElements, 2053 const ConstantInt *CI) { 2054 // We cannot bounds check the index if it doesn't fit in an int64_t. 2055 if (CI->getValue().getMinSignedBits() > 64) 2056 return false; 2057 2058 // A negative index or an index past the end of our sequential type is 2059 // considered out-of-range. 2060 int64_t IndexVal = CI->getSExtValue(); 2061 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2062 return false; 2063 2064 // Otherwise, it is in-range. 2065 return true; 2066 } 2067 2068 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, 2069 bool InBounds, 2070 Optional<unsigned> InRangeIndex, 2071 ArrayRef<Value *> Idxs) { 2072 if (Idxs.empty()) return C; 2073 2074 Type *GEPTy = GetElementPtrInst::getGEPReturnType( 2075 C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size())); 2076 2077 if (isa<UndefValue>(C)) 2078 return UndefValue::get(GEPTy); 2079 2080 Constant *Idx0 = cast<Constant>(Idxs[0]); 2081 if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0))) 2082 return GEPTy->isVectorTy() && !C->getType()->isVectorTy() 2083 ? ConstantVector::getSplat( 2084 cast<VectorType>(GEPTy)->getNumElements(), C) 2085 : C; 2086 2087 if (C->isNullValue()) { 2088 bool isNull = true; 2089 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2090 if (!isa<UndefValue>(Idxs[i]) && 2091 !cast<Constant>(Idxs[i])->isNullValue()) { 2092 isNull = false; 2093 break; 2094 } 2095 if (isNull) { 2096 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2097 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2098 2099 assert(Ty && "Invalid indices for GEP!"); 2100 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2101 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2102 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2103 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements()); 2104 2105 // The GEP returns a vector of pointers when one of more of 2106 // its arguments is a vector. 2107 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) { 2108 if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) { 2109 GEPTy = VectorType::get(OrigGEPTy, VT->getNumElements()); 2110 break; 2111 } 2112 } 2113 2114 return Constant::getNullValue(GEPTy); 2115 } 2116 } 2117 2118 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2119 // Combine Indices - If the source pointer to this getelementptr instruction 2120 // is a getelementptr instruction, combine the indices of the two 2121 // getelementptr instructions into a single instruction. 2122 // 2123 if (CE->getOpcode() == Instruction::GetElementPtr) { 2124 gep_type_iterator LastI = gep_type_end(CE); 2125 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2126 I != E; ++I) 2127 LastI = I; 2128 2129 // We cannot combine indices if doing so would take us outside of an 2130 // array or vector. Doing otherwise could trick us if we evaluated such a 2131 // GEP as part of a load. 2132 // 2133 // e.g. Consider if the original GEP was: 2134 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2135 // i32 0, i32 0, i64 0) 2136 // 2137 // If we then tried to offset it by '8' to get to the third element, 2138 // an i8, we should *not* get: 2139 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2140 // i32 0, i32 0, i64 8) 2141 // 2142 // This GEP tries to index array element '8 which runs out-of-bounds. 2143 // Subsequent evaluation would get confused and produce erroneous results. 2144 // 2145 // The following prohibits such a GEP from being formed by checking to see 2146 // if the index is in-range with respect to an array. 2147 // TODO: This code may be extended to handle vectors as well. 2148 bool PerformFold = false; 2149 if (Idx0->isNullValue()) 2150 PerformFold = true; 2151 else if (LastI.isSequential()) 2152 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2153 PerformFold = (!LastI.isBoundedSequential() || 2154 isIndexInRangeOfArrayType( 2155 LastI.getSequentialNumElements(), CI)) && 2156 !CE->getOperand(CE->getNumOperands() - 1) 2157 ->getType() 2158 ->isVectorTy(); 2159 2160 if (PerformFold) { 2161 SmallVector<Value*, 16> NewIndices; 2162 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2163 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2164 2165 // Add the last index of the source with the first index of the new GEP. 2166 // Make sure to handle the case when they are actually different types. 2167 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2168 // Otherwise it must be an array. 2169 if (!Idx0->isNullValue()) { 2170 Type *IdxTy = Combined->getType(); 2171 if (IdxTy != Idx0->getType()) { 2172 unsigned CommonExtendedWidth = 2173 std::max(IdxTy->getIntegerBitWidth(), 2174 Idx0->getType()->getIntegerBitWidth()); 2175 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2176 2177 Type *CommonTy = 2178 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2179 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2180 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2181 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2182 } else { 2183 Combined = 2184 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2185 } 2186 } 2187 2188 NewIndices.push_back(Combined); 2189 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2190 2191 // The combined GEP normally inherits its index inrange attribute from 2192 // the inner GEP, but if the inner GEP's last index was adjusted by the 2193 // outer GEP, any inbounds attribute on that index is invalidated. 2194 Optional<unsigned> IRIndex = cast<GEPOperator>(CE)->getInRangeIndex(); 2195 if (IRIndex && *IRIndex == CE->getNumOperands() - 2 && !Idx0->isNullValue()) 2196 IRIndex = None; 2197 2198 return ConstantExpr::getGetElementPtr( 2199 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2200 NewIndices, InBounds && cast<GEPOperator>(CE)->isInBounds(), 2201 IRIndex); 2202 } 2203 } 2204 2205 // Attempt to fold casts to the same type away. For example, folding: 2206 // 2207 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2208 // i64 0, i64 0) 2209 // into: 2210 // 2211 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2212 // 2213 // Don't fold if the cast is changing address spaces. 2214 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2215 PointerType *SrcPtrTy = 2216 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2217 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2218 if (SrcPtrTy && DstPtrTy) { 2219 ArrayType *SrcArrayTy = 2220 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2221 ArrayType *DstArrayTy = 2222 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2223 if (SrcArrayTy && DstArrayTy 2224 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2225 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2226 return ConstantExpr::getGetElementPtr(SrcArrayTy, 2227 (Constant *)CE->getOperand(0), 2228 Idxs, InBounds, InRangeIndex); 2229 } 2230 } 2231 } 2232 2233 // Check to see if any array indices are not within the corresponding 2234 // notional array or vector bounds. If so, try to determine if they can be 2235 // factored out into preceding dimensions. 2236 SmallVector<Constant *, 8> NewIdxs; 2237 Type *Ty = PointeeTy; 2238 Type *Prev = C->getType(); 2239 bool Unknown = 2240 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]); 2241 for (unsigned i = 1, e = Idxs.size(); i != e; 2242 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2243 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) { 2244 // We don't know if it's in range or not. 2245 Unknown = true; 2246 continue; 2247 } 2248 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1])) 2249 // Skip if the type of the previous index is not supported. 2250 continue; 2251 if (InRangeIndex && i == *InRangeIndex + 1) { 2252 // If an index is marked inrange, we cannot apply this canonicalization to 2253 // the following index, as that will cause the inrange index to point to 2254 // the wrong element. 2255 continue; 2256 } 2257 if (isa<StructType>(Ty)) { 2258 // The verify makes sure that GEPs into a struct are in range. 2259 continue; 2260 } 2261 auto *STy = cast<SequentialType>(Ty); 2262 if (isa<VectorType>(STy)) { 2263 // There can be awkward padding in after a non-power of two vector. 2264 Unknown = true; 2265 continue; 2266 } 2267 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2268 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI)) 2269 // It's in range, skip to the next index. 2270 continue; 2271 if (CI->getSExtValue() < 0) { 2272 // It's out of range and negative, don't try to factor it. 2273 Unknown = true; 2274 continue; 2275 } 2276 } else { 2277 auto *CV = cast<ConstantDataVector>(Idxs[i]); 2278 bool InRange = true; 2279 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) { 2280 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I)); 2281 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI); 2282 if (CI->getSExtValue() < 0) { 2283 Unknown = true; 2284 break; 2285 } 2286 } 2287 if (InRange || Unknown) 2288 // It's in range, skip to the next index. 2289 // It's out of range and negative, don't try to factor it. 2290 continue; 2291 } 2292 if (isa<StructType>(Prev)) { 2293 // It's out of range, but the prior dimension is a struct 2294 // so we can't do anything about it. 2295 Unknown = true; 2296 continue; 2297 } 2298 // It's out of range, but we can factor it into the prior 2299 // dimension. 2300 NewIdxs.resize(Idxs.size()); 2301 // Determine the number of elements in our sequential type. 2302 uint64_t NumElements = STy->getArrayNumElements(); 2303 2304 // Expand the current index or the previous index to a vector from a scalar 2305 // if necessary. 2306 Constant *CurrIdx = cast<Constant>(Idxs[i]); 2307 auto *PrevIdx = 2308 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]); 2309 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy(); 2310 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy(); 2311 bool UseVector = IsCurrIdxVector || IsPrevIdxVector; 2312 2313 if (!IsCurrIdxVector && IsPrevIdxVector) 2314 CurrIdx = ConstantDataVector::getSplat( 2315 PrevIdx->getType()->getVectorNumElements(), CurrIdx); 2316 2317 if (!IsPrevIdxVector && IsCurrIdxVector) 2318 PrevIdx = ConstantDataVector::getSplat( 2319 CurrIdx->getType()->getVectorNumElements(), PrevIdx); 2320 2321 Constant *Factor = 2322 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements); 2323 if (UseVector) 2324 Factor = ConstantDataVector::getSplat( 2325 IsPrevIdxVector ? PrevIdx->getType()->getVectorNumElements() 2326 : CurrIdx->getType()->getVectorNumElements(), 2327 Factor); 2328 2329 NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor); 2330 2331 Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor); 2332 2333 unsigned CommonExtendedWidth = 2334 std::max(PrevIdx->getType()->getScalarSizeInBits(), 2335 Div->getType()->getScalarSizeInBits()); 2336 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2337 2338 // Before adding, extend both operands to i64 to avoid 2339 // overflow trouble. 2340 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth); 2341 if (UseVector) 2342 ExtendedTy = VectorType::get( 2343 ExtendedTy, IsPrevIdxVector 2344 ? PrevIdx->getType()->getVectorNumElements() 2345 : CurrIdx->getType()->getVectorNumElements()); 2346 2347 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2348 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy); 2349 2350 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2351 Div = ConstantExpr::getSExt(Div, ExtendedTy); 2352 2353 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div); 2354 } 2355 2356 // If we did any factoring, start over with the adjusted indices. 2357 if (!NewIdxs.empty()) { 2358 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2359 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2360 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds, 2361 InRangeIndex); 2362 } 2363 2364 // If all indices are known integers and normalized, we can do a simple 2365 // check for the "inbounds" property. 2366 if (!Unknown && !InBounds) 2367 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2368 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2369 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs, 2370 /*InBounds=*/true, InRangeIndex); 2371 2372 return nullptr; 2373 } 2374