ConstantFold.cpp revision 2c3e0051c31c3f5b2328b447eadf1cf9c4427442
1//===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements folding of constants for LLVM. This implements the 11// (internal) ConstantFold.h interface, which is used by the 12// ConstantExpr::get* methods to automatically fold constants when possible. 13// 14// The current constant folding implementation is implemented in two pieces: the 15// pieces that don't need DataLayout, and the pieces that do. This is to avoid 16// a dependence in IR on Target. 17// 18//===----------------------------------------------------------------------===// 19 20#include "ConstantFold.h" 21#include "llvm/ADT/SmallVector.h" 22#include "llvm/IR/Constants.h" 23#include "llvm/IR/DerivedTypes.h" 24#include "llvm/IR/Function.h" 25#include "llvm/IR/GetElementPtrTypeIterator.h" 26#include "llvm/IR/GlobalAlias.h" 27#include "llvm/IR/GlobalVariable.h" 28#include "llvm/IR/Instructions.h" 29#include "llvm/IR/Operator.h" 30#include "llvm/IR/PatternMatch.h" 31#include "llvm/Support/Compiler.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/ManagedStatic.h" 34#include "llvm/Support/MathExtras.h" 35#include <limits> 36using namespace llvm; 37using namespace llvm::PatternMatch; 38 39//===----------------------------------------------------------------------===// 40// ConstantFold*Instruction Implementations 41//===----------------------------------------------------------------------===// 42 43/// BitCastConstantVector - Convert the specified vector Constant node to the 44/// specified vector type. At this point, we know that the elements of the 45/// input vector constant are all simple integer or FP values. 46static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 47 48 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 49 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 50 51 // If this cast changes element count then we can't handle it here: 52 // doing so requires endianness information. This should be handled by 53 // Analysis/ConstantFolding.cpp 54 unsigned NumElts = DstTy->getNumElements(); 55 if (NumElts != CV->getType()->getVectorNumElements()) 56 return nullptr; 57 58 Type *DstEltTy = DstTy->getElementType(); 59 60 SmallVector<Constant*, 16> Result; 61 Type *Ty = IntegerType::get(CV->getContext(), 32); 62 for (unsigned i = 0; i != NumElts; ++i) { 63 Constant *C = 64 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 65 C = ConstantExpr::getBitCast(C, DstEltTy); 66 Result.push_back(C); 67 } 68 69 return ConstantVector::get(Result); 70} 71 72/// This function determines which opcode to use to fold two constant cast 73/// expressions together. It uses CastInst::isEliminableCastPair to determine 74/// the opcode. Consequently its just a wrapper around that function. 75/// @brief Determine if it is valid to fold a cast of a cast 76static unsigned 77foldConstantCastPair( 78 unsigned opc, ///< opcode of the second cast constant expression 79 ConstantExpr *Op, ///< the first cast constant expression 80 Type *DstTy ///< destination type of the first cast 81) { 82 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 83 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 84 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 85 86 // The the types and opcodes for the two Cast constant expressions 87 Type *SrcTy = Op->getOperand(0)->getType(); 88 Type *MidTy = Op->getType(); 89 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 90 Instruction::CastOps secondOp = Instruction::CastOps(opc); 91 92 // Assume that pointers are never more than 64 bits wide, and only use this 93 // for the middle type. Otherwise we could end up folding away illegal 94 // bitcasts between address spaces with different sizes. 95 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 96 97 // Let CastInst::isEliminableCastPair do the heavy lifting. 98 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 99 nullptr, FakeIntPtrTy, nullptr); 100} 101 102static Constant *FoldBitCast(Constant *V, Type *DestTy) { 103 Type *SrcTy = V->getType(); 104 if (SrcTy == DestTy) 105 return V; // no-op cast 106 107 // Check to see if we are casting a pointer to an aggregate to a pointer to 108 // the first element. If so, return the appropriate GEP instruction. 109 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 110 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 111 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 112 && DPTy->getElementType()->isSized()) { 113 SmallVector<Value*, 8> IdxList; 114 Value *Zero = 115 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 116 IdxList.push_back(Zero); 117 Type *ElTy = PTy->getElementType(); 118 while (ElTy != DPTy->getElementType()) { 119 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 120 if (STy->getNumElements() == 0) break; 121 ElTy = STy->getElementType(0); 122 IdxList.push_back(Zero); 123 } else if (SequentialType *STy = 124 dyn_cast<SequentialType>(ElTy)) { 125 if (ElTy->isPointerTy()) break; // Can't index into pointers! 126 ElTy = STy->getElementType(); 127 IdxList.push_back(Zero); 128 } else { 129 break; 130 } 131 } 132 133 if (ElTy == DPTy->getElementType()) 134 // This GEP is inbounds because all indices are zero. 135 return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(), 136 V, IdxList); 137 } 138 139 // Handle casts from one vector constant to another. We know that the src 140 // and dest type have the same size (otherwise its an illegal cast). 141 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 142 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 143 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 144 "Not cast between same sized vectors!"); 145 SrcTy = nullptr; 146 // First, check for null. Undef is already handled. 147 if (isa<ConstantAggregateZero>(V)) 148 return Constant::getNullValue(DestTy); 149 150 // Handle ConstantVector and ConstantAggregateVector. 151 return BitCastConstantVector(V, DestPTy); 152 } 153 154 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 155 // This allows for other simplifications (although some of them 156 // can only be handled by Analysis/ConstantFolding.cpp). 157 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 158 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 159 } 160 161 // Finally, implement bitcast folding now. The code below doesn't handle 162 // bitcast right. 163 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 164 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 165 166 // Handle integral constant input. 167 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 168 if (DestTy->isIntegerTy()) 169 // Integral -> Integral. This is a no-op because the bit widths must 170 // be the same. Consequently, we just fold to V. 171 return V; 172 173 // See note below regarding the PPC_FP128 restriction. 174 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 175 return ConstantFP::get(DestTy->getContext(), 176 APFloat(DestTy->getFltSemantics(), 177 CI->getValue())); 178 179 // Otherwise, can't fold this (vector?) 180 return nullptr; 181 } 182 183 // Handle ConstantFP input: FP -> Integral. 184 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 185 // PPC_FP128 is really the sum of two consecutive doubles, where the first 186 // double is always stored first in memory, regardless of the target 187 // endianness. The memory layout of i128, however, depends on the target 188 // endianness, and so we can't fold this without target endianness 189 // information. This should instead be handled by 190 // Analysis/ConstantFolding.cpp 191 if (FP->getType()->isPPC_FP128Ty()) 192 return nullptr; 193 194 return ConstantInt::get(FP->getContext(), 195 FP->getValueAPF().bitcastToAPInt()); 196 } 197 198 return nullptr; 199} 200 201 202/// ExtractConstantBytes - V is an integer constant which only has a subset of 203/// its bytes used. The bytes used are indicated by ByteStart (which is the 204/// first byte used, counting from the least significant byte) and ByteSize, 205/// which is the number of bytes used. 206/// 207/// This function analyzes the specified constant to see if the specified byte 208/// range can be returned as a simplified constant. If so, the constant is 209/// returned, otherwise null is returned. 210/// 211static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 212 unsigned ByteSize) { 213 assert(C->getType()->isIntegerTy() && 214 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 215 "Non-byte sized integer input"); 216 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 217 assert(ByteSize && "Must be accessing some piece"); 218 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 219 assert(ByteSize != CSize && "Should not extract everything"); 220 221 // Constant Integers are simple. 222 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 223 APInt V = CI->getValue(); 224 if (ByteStart) 225 V = V.lshr(ByteStart*8); 226 V = V.trunc(ByteSize*8); 227 return ConstantInt::get(CI->getContext(), V); 228 } 229 230 // In the input is a constant expr, we might be able to recursively simplify. 231 // If not, we definitely can't do anything. 232 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 233 if (!CE) return nullptr; 234 235 switch (CE->getOpcode()) { 236 default: return nullptr; 237 case Instruction::Or: { 238 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 239 if (!RHS) 240 return nullptr; 241 242 // X | -1 -> -1. 243 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 244 if (RHSC->isAllOnesValue()) 245 return RHSC; 246 247 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 248 if (!LHS) 249 return nullptr; 250 return ConstantExpr::getOr(LHS, RHS); 251 } 252 case Instruction::And: { 253 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 254 if (!RHS) 255 return nullptr; 256 257 // X & 0 -> 0. 258 if (RHS->isNullValue()) 259 return RHS; 260 261 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 262 if (!LHS) 263 return nullptr; 264 return ConstantExpr::getAnd(LHS, RHS); 265 } 266 case Instruction::LShr: { 267 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 268 if (!Amt) 269 return nullptr; 270 unsigned ShAmt = Amt->getZExtValue(); 271 // Cannot analyze non-byte shifts. 272 if ((ShAmt & 7) != 0) 273 return nullptr; 274 ShAmt >>= 3; 275 276 // If the extract is known to be all zeros, return zero. 277 if (ByteStart >= CSize-ShAmt) 278 return Constant::getNullValue(IntegerType::get(CE->getContext(), 279 ByteSize*8)); 280 // If the extract is known to be fully in the input, extract it. 281 if (ByteStart+ByteSize+ShAmt <= CSize) 282 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 283 284 // TODO: Handle the 'partially zero' case. 285 return nullptr; 286 } 287 288 case Instruction::Shl: { 289 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 290 if (!Amt) 291 return nullptr; 292 unsigned ShAmt = Amt->getZExtValue(); 293 // Cannot analyze non-byte shifts. 294 if ((ShAmt & 7) != 0) 295 return nullptr; 296 ShAmt >>= 3; 297 298 // If the extract is known to be all zeros, return zero. 299 if (ByteStart+ByteSize <= ShAmt) 300 return Constant::getNullValue(IntegerType::get(CE->getContext(), 301 ByteSize*8)); 302 // If the extract is known to be fully in the input, extract it. 303 if (ByteStart >= ShAmt) 304 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 305 306 // TODO: Handle the 'partially zero' case. 307 return nullptr; 308 } 309 310 case Instruction::ZExt: { 311 unsigned SrcBitSize = 312 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 313 314 // If extracting something that is completely zero, return 0. 315 if (ByteStart*8 >= SrcBitSize) 316 return Constant::getNullValue(IntegerType::get(CE->getContext(), 317 ByteSize*8)); 318 319 // If exactly extracting the input, return it. 320 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 321 return CE->getOperand(0); 322 323 // If extracting something completely in the input, if if the input is a 324 // multiple of 8 bits, recurse. 325 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 326 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 327 328 // Otherwise, if extracting a subset of the input, which is not multiple of 329 // 8 bits, do a shift and trunc to get the bits. 330 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 331 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 332 Constant *Res = CE->getOperand(0); 333 if (ByteStart) 334 Res = ConstantExpr::getLShr(Res, 335 ConstantInt::get(Res->getType(), ByteStart*8)); 336 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 337 ByteSize*8)); 338 } 339 340 // TODO: Handle the 'partially zero' case. 341 return nullptr; 342 } 343 } 344} 345 346/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 347/// on Ty, with any known factors factored out. If Folded is false, 348/// return null if no factoring was possible, to avoid endlessly 349/// bouncing an unfoldable expression back into the top-level folder. 350/// 351static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 352 bool Folded) { 353 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 354 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 355 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 356 return ConstantExpr::getNUWMul(E, N); 357 } 358 359 if (StructType *STy = dyn_cast<StructType>(Ty)) 360 if (!STy->isPacked()) { 361 unsigned NumElems = STy->getNumElements(); 362 // An empty struct has size zero. 363 if (NumElems == 0) 364 return ConstantExpr::getNullValue(DestTy); 365 // Check for a struct with all members having the same size. 366 Constant *MemberSize = 367 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 368 bool AllSame = true; 369 for (unsigned i = 1; i != NumElems; ++i) 370 if (MemberSize != 371 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 372 AllSame = false; 373 break; 374 } 375 if (AllSame) { 376 Constant *N = ConstantInt::get(DestTy, NumElems); 377 return ConstantExpr::getNUWMul(MemberSize, N); 378 } 379 } 380 381 // Pointer size doesn't depend on the pointee type, so canonicalize them 382 // to an arbitrary pointee. 383 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 384 if (!PTy->getElementType()->isIntegerTy(1)) 385 return 386 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 387 PTy->getAddressSpace()), 388 DestTy, true); 389 390 // If there's no interesting folding happening, bail so that we don't create 391 // a constant that looks like it needs folding but really doesn't. 392 if (!Folded) 393 return nullptr; 394 395 // Base case: Get a regular sizeof expression. 396 Constant *C = ConstantExpr::getSizeOf(Ty); 397 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 398 DestTy, false), 399 C, DestTy); 400 return C; 401} 402 403/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 404/// on Ty, with any known factors factored out. If Folded is false, 405/// return null if no factoring was possible, to avoid endlessly 406/// bouncing an unfoldable expression back into the top-level folder. 407/// 408static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 409 bool Folded) { 410 // The alignment of an array is equal to the alignment of the 411 // array element. Note that this is not always true for vectors. 412 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 413 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 414 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 415 DestTy, 416 false), 417 C, DestTy); 418 return C; 419 } 420 421 if (StructType *STy = dyn_cast<StructType>(Ty)) { 422 // Packed structs always have an alignment of 1. 423 if (STy->isPacked()) 424 return ConstantInt::get(DestTy, 1); 425 426 // Otherwise, struct alignment is the maximum alignment of any member. 427 // Without target data, we can't compare much, but we can check to see 428 // if all the members have the same alignment. 429 unsigned NumElems = STy->getNumElements(); 430 // An empty struct has minimal alignment. 431 if (NumElems == 0) 432 return ConstantInt::get(DestTy, 1); 433 // Check for a struct with all members having the same alignment. 434 Constant *MemberAlign = 435 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 436 bool AllSame = true; 437 for (unsigned i = 1; i != NumElems; ++i) 438 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 439 AllSame = false; 440 break; 441 } 442 if (AllSame) 443 return MemberAlign; 444 } 445 446 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 447 // to an arbitrary pointee. 448 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 449 if (!PTy->getElementType()->isIntegerTy(1)) 450 return 451 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 452 1), 453 PTy->getAddressSpace()), 454 DestTy, true); 455 456 // If there's no interesting folding happening, bail so that we don't create 457 // a constant that looks like it needs folding but really doesn't. 458 if (!Folded) 459 return nullptr; 460 461 // Base case: Get a regular alignof expression. 462 Constant *C = ConstantExpr::getAlignOf(Ty); 463 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 464 DestTy, false), 465 C, DestTy); 466 return C; 467} 468 469/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 470/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 471/// return null if no factoring was possible, to avoid endlessly 472/// bouncing an unfoldable expression back into the top-level folder. 473/// 474static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 475 Type *DestTy, 476 bool Folded) { 477 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 478 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 479 DestTy, false), 480 FieldNo, DestTy); 481 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 482 return ConstantExpr::getNUWMul(E, N); 483 } 484 485 if (StructType *STy = dyn_cast<StructType>(Ty)) 486 if (!STy->isPacked()) { 487 unsigned NumElems = STy->getNumElements(); 488 // An empty struct has no members. 489 if (NumElems == 0) 490 return nullptr; 491 // Check for a struct with all members having the same size. 492 Constant *MemberSize = 493 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 494 bool AllSame = true; 495 for (unsigned i = 1; i != NumElems; ++i) 496 if (MemberSize != 497 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 498 AllSame = false; 499 break; 500 } 501 if (AllSame) { 502 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 503 false, 504 DestTy, 505 false), 506 FieldNo, DestTy); 507 return ConstantExpr::getNUWMul(MemberSize, N); 508 } 509 } 510 511 // If there's no interesting folding happening, bail so that we don't create 512 // a constant that looks like it needs folding but really doesn't. 513 if (!Folded) 514 return nullptr; 515 516 // Base case: Get a regular offsetof expression. 517 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 518 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 519 DestTy, false), 520 C, DestTy); 521 return C; 522} 523 524Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 525 Type *DestTy) { 526 if (isa<UndefValue>(V)) { 527 // zext(undef) = 0, because the top bits will be zero. 528 // sext(undef) = 0, because the top bits will all be the same. 529 // [us]itofp(undef) = 0, because the result value is bounded. 530 if (opc == Instruction::ZExt || opc == Instruction::SExt || 531 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 532 return Constant::getNullValue(DestTy); 533 return UndefValue::get(DestTy); 534 } 535 536 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 537 return Constant::getNullValue(DestTy); 538 539 // If the cast operand is a constant expression, there's a few things we can 540 // do to try to simplify it. 541 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 542 if (CE->isCast()) { 543 // Try hard to fold cast of cast because they are often eliminable. 544 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 545 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 546 } else if (CE->getOpcode() == Instruction::GetElementPtr && 547 // Do not fold addrspacecast (gep 0, .., 0). It might make the 548 // addrspacecast uncanonicalized. 549 opc != Instruction::AddrSpaceCast) { 550 // If all of the indexes in the GEP are null values, there is no pointer 551 // adjustment going on. We might as well cast the source pointer. 552 bool isAllNull = true; 553 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 554 if (!CE->getOperand(i)->isNullValue()) { 555 isAllNull = false; 556 break; 557 } 558 if (isAllNull) 559 // This is casting one pointer type to another, always BitCast 560 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 561 } 562 } 563 564 // If the cast operand is a constant vector, perform the cast by 565 // operating on each element. In the cast of bitcasts, the element 566 // count may be mismatched; don't attempt to handle that here. 567 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 568 DestTy->isVectorTy() && 569 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 570 SmallVector<Constant*, 16> res; 571 VectorType *DestVecTy = cast<VectorType>(DestTy); 572 Type *DstEltTy = DestVecTy->getElementType(); 573 Type *Ty = IntegerType::get(V->getContext(), 32); 574 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 575 Constant *C = 576 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 577 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 578 } 579 return ConstantVector::get(res); 580 } 581 582 // We actually have to do a cast now. Perform the cast according to the 583 // opcode specified. 584 switch (opc) { 585 default: 586 llvm_unreachable("Failed to cast constant expression"); 587 case Instruction::FPTrunc: 588 case Instruction::FPExt: 589 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 590 bool ignored; 591 APFloat Val = FPC->getValueAPF(); 592 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 593 DestTy->isFloatTy() ? APFloat::IEEEsingle : 594 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 595 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 596 DestTy->isFP128Ty() ? APFloat::IEEEquad : 597 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble : 598 APFloat::Bogus, 599 APFloat::rmNearestTiesToEven, &ignored); 600 return ConstantFP::get(V->getContext(), Val); 601 } 602 return nullptr; // Can't fold. 603 case Instruction::FPToUI: 604 case Instruction::FPToSI: 605 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 606 const APFloat &V = FPC->getValueAPF(); 607 bool ignored; 608 uint64_t x[2]; 609 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 610 if (APFloat::opInvalidOp == 611 V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 612 APFloat::rmTowardZero, &ignored)) { 613 // Undefined behavior invoked - the destination type can't represent 614 // the input constant. 615 return UndefValue::get(DestTy); 616 } 617 APInt Val(DestBitWidth, x); 618 return ConstantInt::get(FPC->getContext(), Val); 619 } 620 return nullptr; // Can't fold. 621 case Instruction::IntToPtr: //always treated as unsigned 622 if (V->isNullValue()) // Is it an integral null value? 623 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 624 return nullptr; // Other pointer types cannot be casted 625 case Instruction::PtrToInt: // always treated as unsigned 626 // Is it a null pointer value? 627 if (V->isNullValue()) 628 return ConstantInt::get(DestTy, 0); 629 // If this is a sizeof-like expression, pull out multiplications by 630 // known factors to expose them to subsequent folding. If it's an 631 // alignof-like expression, factor out known factors. 632 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 633 if (CE->getOpcode() == Instruction::GetElementPtr && 634 CE->getOperand(0)->isNullValue()) { 635 Type *Ty = 636 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 637 if (CE->getNumOperands() == 2) { 638 // Handle a sizeof-like expression. 639 Constant *Idx = CE->getOperand(1); 640 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 641 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 642 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 643 DestTy, false), 644 Idx, DestTy); 645 return ConstantExpr::getMul(C, Idx); 646 } 647 } else if (CE->getNumOperands() == 3 && 648 CE->getOperand(1)->isNullValue()) { 649 // Handle an alignof-like expression. 650 if (StructType *STy = dyn_cast<StructType>(Ty)) 651 if (!STy->isPacked()) { 652 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 653 if (CI->isOne() && 654 STy->getNumElements() == 2 && 655 STy->getElementType(0)->isIntegerTy(1)) { 656 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 657 } 658 } 659 // Handle an offsetof-like expression. 660 if (Ty->isStructTy() || Ty->isArrayTy()) { 661 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 662 DestTy, false)) 663 return C; 664 } 665 } 666 } 667 // Other pointer types cannot be casted 668 return nullptr; 669 case Instruction::UIToFP: 670 case Instruction::SIToFP: 671 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 672 APInt api = CI->getValue(); 673 APFloat apf(DestTy->getFltSemantics(), 674 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 675 if (APFloat::opOverflow & 676 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 677 APFloat::rmNearestTiesToEven)) { 678 // Undefined behavior invoked - the destination type can't represent 679 // the input constant. 680 return UndefValue::get(DestTy); 681 } 682 return ConstantFP::get(V->getContext(), apf); 683 } 684 return nullptr; 685 case Instruction::ZExt: 686 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 687 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 688 return ConstantInt::get(V->getContext(), 689 CI->getValue().zext(BitWidth)); 690 } 691 return nullptr; 692 case Instruction::SExt: 693 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 694 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 695 return ConstantInt::get(V->getContext(), 696 CI->getValue().sext(BitWidth)); 697 } 698 return nullptr; 699 case Instruction::Trunc: { 700 if (V->getType()->isVectorTy()) 701 return nullptr; 702 703 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 704 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 705 return ConstantInt::get(V->getContext(), 706 CI->getValue().trunc(DestBitWidth)); 707 } 708 709 // The input must be a constantexpr. See if we can simplify this based on 710 // the bytes we are demanding. Only do this if the source and dest are an 711 // even multiple of a byte. 712 if ((DestBitWidth & 7) == 0 && 713 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 714 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 715 return Res; 716 717 return nullptr; 718 } 719 case Instruction::BitCast: 720 return FoldBitCast(V, DestTy); 721 case Instruction::AddrSpaceCast: 722 return nullptr; 723 } 724} 725 726Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 727 Constant *V1, Constant *V2) { 728 // Check for i1 and vector true/false conditions. 729 if (Cond->isNullValue()) return V2; 730 if (Cond->isAllOnesValue()) return V1; 731 732 // If the condition is a vector constant, fold the result elementwise. 733 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 734 SmallVector<Constant*, 16> Result; 735 Type *Ty = IntegerType::get(CondV->getContext(), 32); 736 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 737 Constant *V; 738 Constant *V1Element = ConstantExpr::getExtractElement(V1, 739 ConstantInt::get(Ty, i)); 740 Constant *V2Element = ConstantExpr::getExtractElement(V2, 741 ConstantInt::get(Ty, i)); 742 Constant *Cond = dyn_cast<Constant>(CondV->getOperand(i)); 743 if (V1Element == V2Element) { 744 V = V1Element; 745 } else if (isa<UndefValue>(Cond)) { 746 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 747 } else { 748 if (!isa<ConstantInt>(Cond)) break; 749 V = Cond->isNullValue() ? V2Element : V1Element; 750 } 751 Result.push_back(V); 752 } 753 754 // If we were able to build the vector, return it. 755 if (Result.size() == V1->getType()->getVectorNumElements()) 756 return ConstantVector::get(Result); 757 } 758 759 if (isa<UndefValue>(Cond)) { 760 if (isa<UndefValue>(V1)) return V1; 761 return V2; 762 } 763 if (isa<UndefValue>(V1)) return V2; 764 if (isa<UndefValue>(V2)) return V1; 765 if (V1 == V2) return V1; 766 767 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 768 if (TrueVal->getOpcode() == Instruction::Select) 769 if (TrueVal->getOperand(0) == Cond) 770 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 771 } 772 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 773 if (FalseVal->getOpcode() == Instruction::Select) 774 if (FalseVal->getOperand(0) == Cond) 775 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 776 } 777 778 return nullptr; 779} 780 781Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 782 Constant *Idx) { 783 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 784 return UndefValue::get(Val->getType()->getVectorElementType()); 785 if (Val->isNullValue()) // ee(zero, x) -> zero 786 return Constant::getNullValue(Val->getType()->getVectorElementType()); 787 // ee({w,x,y,z}, undef) -> undef 788 if (isa<UndefValue>(Idx)) 789 return UndefValue::get(Val->getType()->getVectorElementType()); 790 791 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 792 uint64_t Index = CIdx->getZExtValue(); 793 // ee({w,x,y,z}, wrong_value) -> undef 794 if (Index >= Val->getType()->getVectorNumElements()) 795 return UndefValue::get(Val->getType()->getVectorElementType()); 796 return Val->getAggregateElement(Index); 797 } 798 return nullptr; 799} 800 801Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 802 Constant *Elt, 803 Constant *Idx) { 804 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 805 if (!CIdx) return nullptr; 806 const APInt &IdxVal = CIdx->getValue(); 807 808 SmallVector<Constant*, 16> Result; 809 Type *Ty = IntegerType::get(Val->getContext(), 32); 810 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){ 811 if (i == IdxVal) { 812 Result.push_back(Elt); 813 continue; 814 } 815 816 Constant *C = 817 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 818 Result.push_back(C); 819 } 820 821 return ConstantVector::get(Result); 822} 823 824Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 825 Constant *V2, 826 Constant *Mask) { 827 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 828 Type *EltTy = V1->getType()->getVectorElementType(); 829 830 // Undefined shuffle mask -> undefined value. 831 if (isa<UndefValue>(Mask)) 832 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 833 834 // Don't break the bitcode reader hack. 835 if (isa<ConstantExpr>(Mask)) return nullptr; 836 837 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 838 839 // Loop over the shuffle mask, evaluating each element. 840 SmallVector<Constant*, 32> Result; 841 for (unsigned i = 0; i != MaskNumElts; ++i) { 842 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 843 if (Elt == -1) { 844 Result.push_back(UndefValue::get(EltTy)); 845 continue; 846 } 847 Constant *InElt; 848 if (unsigned(Elt) >= SrcNumElts*2) 849 InElt = UndefValue::get(EltTy); 850 else if (unsigned(Elt) >= SrcNumElts) { 851 Type *Ty = IntegerType::get(V2->getContext(), 32); 852 InElt = 853 ConstantExpr::getExtractElement(V2, 854 ConstantInt::get(Ty, Elt - SrcNumElts)); 855 } else { 856 Type *Ty = IntegerType::get(V1->getContext(), 32); 857 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 858 } 859 Result.push_back(InElt); 860 } 861 862 return ConstantVector::get(Result); 863} 864 865Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 866 ArrayRef<unsigned> Idxs) { 867 // Base case: no indices, so return the entire value. 868 if (Idxs.empty()) 869 return Agg; 870 871 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 872 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 873 874 return nullptr; 875} 876 877Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 878 Constant *Val, 879 ArrayRef<unsigned> Idxs) { 880 // Base case: no indices, so replace the entire value. 881 if (Idxs.empty()) 882 return Val; 883 884 unsigned NumElts; 885 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 886 NumElts = ST->getNumElements(); 887 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 888 NumElts = AT->getNumElements(); 889 else 890 NumElts = Agg->getType()->getVectorNumElements(); 891 892 SmallVector<Constant*, 32> Result; 893 for (unsigned i = 0; i != NumElts; ++i) { 894 Constant *C = Agg->getAggregateElement(i); 895 if (!C) return nullptr; 896 897 if (Idxs[0] == i) 898 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 899 900 Result.push_back(C); 901 } 902 903 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 904 return ConstantStruct::get(ST, Result); 905 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 906 return ConstantArray::get(AT, Result); 907 return ConstantVector::get(Result); 908} 909 910 911Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 912 Constant *C1, Constant *C2) { 913 // Handle UndefValue up front. 914 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 915 switch (Opcode) { 916 case Instruction::Xor: 917 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 918 // Handle undef ^ undef -> 0 special case. This is a common 919 // idiom (misuse). 920 return Constant::getNullValue(C1->getType()); 921 // Fallthrough 922 case Instruction::Add: 923 case Instruction::Sub: 924 return UndefValue::get(C1->getType()); 925 case Instruction::And: 926 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 927 return C1; 928 return Constant::getNullValue(C1->getType()); // undef & X -> 0 929 case Instruction::Mul: { 930 // undef * undef -> undef 931 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 932 return C1; 933 const APInt *CV; 934 // X * undef -> undef if X is odd 935 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 936 if ((*CV)[0]) 937 return UndefValue::get(C1->getType()); 938 939 // X * undef -> 0 otherwise 940 return Constant::getNullValue(C1->getType()); 941 } 942 case Instruction::SDiv: 943 case Instruction::UDiv: 944 // X / undef -> undef 945 if (match(C1, m_Zero())) 946 return C2; 947 // undef / 0 -> undef 948 // undef / 1 -> undef 949 if (match(C2, m_Zero()) || match(C2, m_One())) 950 return C1; 951 // undef / X -> 0 otherwise 952 return Constant::getNullValue(C1->getType()); 953 case Instruction::URem: 954 case Instruction::SRem: 955 // X % undef -> undef 956 if (match(C2, m_Undef())) 957 return C2; 958 // undef % 0 -> undef 959 if (match(C2, m_Zero())) 960 return C1; 961 // undef % X -> 0 otherwise 962 return Constant::getNullValue(C1->getType()); 963 case Instruction::Or: // X | undef -> -1 964 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 965 return C1; 966 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 967 case Instruction::LShr: 968 // X >>l undef -> undef 969 if (isa<UndefValue>(C2)) 970 return C2; 971 // undef >>l 0 -> undef 972 if (match(C2, m_Zero())) 973 return C1; 974 // undef >>l X -> 0 975 return Constant::getNullValue(C1->getType()); 976 case Instruction::AShr: 977 // X >>a undef -> undef 978 if (isa<UndefValue>(C2)) 979 return C2; 980 // undef >>a 0 -> undef 981 if (match(C2, m_Zero())) 982 return C1; 983 // TODO: undef >>a X -> undef if the shift is exact 984 // undef >>a X -> 0 985 return Constant::getNullValue(C1->getType()); 986 case Instruction::Shl: 987 // X << undef -> undef 988 if (isa<UndefValue>(C2)) 989 return C2; 990 // undef << 0 -> undef 991 if (match(C2, m_Zero())) 992 return C1; 993 // undef << X -> 0 994 return Constant::getNullValue(C1->getType()); 995 } 996 } 997 998 // Handle simplifications when the RHS is a constant int. 999 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1000 switch (Opcode) { 1001 case Instruction::Add: 1002 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1003 break; 1004 case Instruction::Sub: 1005 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1006 break; 1007 case Instruction::Mul: 1008 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1009 if (CI2->equalsInt(1)) 1010 return C1; // X * 1 == X 1011 break; 1012 case Instruction::UDiv: 1013 case Instruction::SDiv: 1014 if (CI2->equalsInt(1)) 1015 return C1; // X / 1 == X 1016 if (CI2->equalsInt(0)) 1017 return UndefValue::get(CI2->getType()); // X / 0 == undef 1018 break; 1019 case Instruction::URem: 1020 case Instruction::SRem: 1021 if (CI2->equalsInt(1)) 1022 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1023 if (CI2->equalsInt(0)) 1024 return UndefValue::get(CI2->getType()); // X % 0 == undef 1025 break; 1026 case Instruction::And: 1027 if (CI2->isZero()) return C2; // X & 0 == 0 1028 if (CI2->isAllOnesValue()) 1029 return C1; // X & -1 == X 1030 1031 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1032 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1033 if (CE1->getOpcode() == Instruction::ZExt) { 1034 unsigned DstWidth = CI2->getType()->getBitWidth(); 1035 unsigned SrcWidth = 1036 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1037 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1038 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1039 return C1; 1040 } 1041 1042 // If and'ing the address of a global with a constant, fold it. 1043 if (CE1->getOpcode() == Instruction::PtrToInt && 1044 isa<GlobalValue>(CE1->getOperand(0))) { 1045 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1046 1047 // Functions are at least 4-byte aligned. 1048 unsigned GVAlign = GV->getAlignment(); 1049 if (isa<Function>(GV)) 1050 GVAlign = std::max(GVAlign, 4U); 1051 1052 if (GVAlign > 1) { 1053 unsigned DstWidth = CI2->getType()->getBitWidth(); 1054 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1055 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1056 1057 // If checking bits we know are clear, return zero. 1058 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1059 return Constant::getNullValue(CI2->getType()); 1060 } 1061 } 1062 } 1063 break; 1064 case Instruction::Or: 1065 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1066 if (CI2->isAllOnesValue()) 1067 return C2; // X | -1 == -1 1068 break; 1069 case Instruction::Xor: 1070 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1071 1072 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1073 switch (CE1->getOpcode()) { 1074 default: break; 1075 case Instruction::ICmp: 1076 case Instruction::FCmp: 1077 // cmp pred ^ true -> cmp !pred 1078 assert(CI2->equalsInt(1)); 1079 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1080 pred = CmpInst::getInversePredicate(pred); 1081 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1082 CE1->getOperand(1)); 1083 } 1084 } 1085 break; 1086 case Instruction::AShr: 1087 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1088 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1089 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1090 return ConstantExpr::getLShr(C1, C2); 1091 break; 1092 } 1093 } else if (isa<ConstantInt>(C1)) { 1094 // If C1 is a ConstantInt and C2 is not, swap the operands. 1095 if (Instruction::isCommutative(Opcode)) 1096 return ConstantExpr::get(Opcode, C2, C1); 1097 } 1098 1099 // At this point we know neither constant is an UndefValue. 1100 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1101 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1102 const APInt &C1V = CI1->getValue(); 1103 const APInt &C2V = CI2->getValue(); 1104 switch (Opcode) { 1105 default: 1106 break; 1107 case Instruction::Add: 1108 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1109 case Instruction::Sub: 1110 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1111 case Instruction::Mul: 1112 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1113 case Instruction::UDiv: 1114 assert(!CI2->isNullValue() && "Div by zero handled above"); 1115 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1116 case Instruction::SDiv: 1117 assert(!CI2->isNullValue() && "Div by zero handled above"); 1118 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1119 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1120 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1121 case Instruction::URem: 1122 assert(!CI2->isNullValue() && "Div by zero handled above"); 1123 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1124 case Instruction::SRem: 1125 assert(!CI2->isNullValue() && "Div by zero handled above"); 1126 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1127 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1128 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1129 case Instruction::And: 1130 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1131 case Instruction::Or: 1132 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1133 case Instruction::Xor: 1134 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1135 case Instruction::Shl: 1136 if (C2V.ult(C1V.getBitWidth())) 1137 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1138 return UndefValue::get(C1->getType()); // too big shift is undef 1139 case Instruction::LShr: 1140 if (C2V.ult(C1V.getBitWidth())) 1141 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1142 return UndefValue::get(C1->getType()); // too big shift is undef 1143 case Instruction::AShr: 1144 if (C2V.ult(C1V.getBitWidth())) 1145 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1146 return UndefValue::get(C1->getType()); // too big shift is undef 1147 } 1148 } 1149 1150 switch (Opcode) { 1151 case Instruction::SDiv: 1152 case Instruction::UDiv: 1153 case Instruction::URem: 1154 case Instruction::SRem: 1155 case Instruction::LShr: 1156 case Instruction::AShr: 1157 case Instruction::Shl: 1158 if (CI1->equalsInt(0)) return C1; 1159 break; 1160 default: 1161 break; 1162 } 1163 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1164 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1165 APFloat C1V = CFP1->getValueAPF(); 1166 APFloat C2V = CFP2->getValueAPF(); 1167 APFloat C3V = C1V; // copy for modification 1168 switch (Opcode) { 1169 default: 1170 break; 1171 case Instruction::FAdd: 1172 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1173 return ConstantFP::get(C1->getContext(), C3V); 1174 case Instruction::FSub: 1175 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1176 return ConstantFP::get(C1->getContext(), C3V); 1177 case Instruction::FMul: 1178 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1179 return ConstantFP::get(C1->getContext(), C3V); 1180 case Instruction::FDiv: 1181 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1182 return ConstantFP::get(C1->getContext(), C3V); 1183 case Instruction::FRem: 1184 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1185 return ConstantFP::get(C1->getContext(), C3V); 1186 } 1187 } 1188 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1189 // Perform elementwise folding. 1190 SmallVector<Constant*, 16> Result; 1191 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1192 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1193 Constant *LHS = 1194 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1195 Constant *RHS = 1196 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1197 1198 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1199 } 1200 1201 return ConstantVector::get(Result); 1202 } 1203 1204 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1205 // There are many possible foldings we could do here. We should probably 1206 // at least fold add of a pointer with an integer into the appropriate 1207 // getelementptr. This will improve alias analysis a bit. 1208 1209 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1210 // (a + (b + c)). 1211 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1212 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1213 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1214 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1215 } 1216 } else if (isa<ConstantExpr>(C2)) { 1217 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1218 // other way if possible. 1219 if (Instruction::isCommutative(Opcode)) 1220 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1221 } 1222 1223 // i1 can be simplified in many cases. 1224 if (C1->getType()->isIntegerTy(1)) { 1225 switch (Opcode) { 1226 case Instruction::Add: 1227 case Instruction::Sub: 1228 return ConstantExpr::getXor(C1, C2); 1229 case Instruction::Mul: 1230 return ConstantExpr::getAnd(C1, C2); 1231 case Instruction::Shl: 1232 case Instruction::LShr: 1233 case Instruction::AShr: 1234 // We can assume that C2 == 0. If it were one the result would be 1235 // undefined because the shift value is as large as the bitwidth. 1236 return C1; 1237 case Instruction::SDiv: 1238 case Instruction::UDiv: 1239 // We can assume that C2 == 1. If it were zero the result would be 1240 // undefined through division by zero. 1241 return C1; 1242 case Instruction::URem: 1243 case Instruction::SRem: 1244 // We can assume that C2 == 1. If it were zero the result would be 1245 // undefined through division by zero. 1246 return ConstantInt::getFalse(C1->getContext()); 1247 default: 1248 break; 1249 } 1250 } 1251 1252 // We don't know how to fold this. 1253 return nullptr; 1254} 1255 1256/// isZeroSizedType - This type is zero sized if its an array or structure of 1257/// zero sized types. The only leaf zero sized type is an empty structure. 1258static bool isMaybeZeroSizedType(Type *Ty) { 1259 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1260 if (STy->isOpaque()) return true; // Can't say. 1261 1262 // If all of elements have zero size, this does too. 1263 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1264 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1265 return true; 1266 1267 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1268 return isMaybeZeroSizedType(ATy->getElementType()); 1269 } 1270 return false; 1271} 1272 1273/// IdxCompare - Compare the two constants as though they were getelementptr 1274/// indices. This allows coersion of the types to be the same thing. 1275/// 1276/// If the two constants are the "same" (after coersion), return 0. If the 1277/// first is less than the second, return -1, if the second is less than the 1278/// first, return 1. If the constants are not integral, return -2. 1279/// 1280static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1281 if (C1 == C2) return 0; 1282 1283 // Ok, we found a different index. If they are not ConstantInt, we can't do 1284 // anything with them. 1285 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1286 return -2; // don't know! 1287 1288 // We cannot compare the indices if they don't fit in an int64_t. 1289 if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 || 1290 cast<ConstantInt>(C2)->getValue().getActiveBits() > 64) 1291 return -2; // don't know! 1292 1293 // Ok, we have two differing integer indices. Sign extend them to be the same 1294 // type. 1295 int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue(); 1296 int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue(); 1297 1298 if (C1Val == C2Val) return 0; // They are equal 1299 1300 // If the type being indexed over is really just a zero sized type, there is 1301 // no pointer difference being made here. 1302 if (isMaybeZeroSizedType(ElTy)) 1303 return -2; // dunno. 1304 1305 // If they are really different, now that they are the same type, then we 1306 // found a difference! 1307 if (C1Val < C2Val) 1308 return -1; 1309 else 1310 return 1; 1311} 1312 1313/// evaluateFCmpRelation - This function determines if there is anything we can 1314/// decide about the two constants provided. This doesn't need to handle simple 1315/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1316/// If we can determine that the two constants have a particular relation to 1317/// each other, we should return the corresponding FCmpInst predicate, 1318/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1319/// ConstantFoldCompareInstruction. 1320/// 1321/// To simplify this code we canonicalize the relation so that the first 1322/// operand is always the most "complex" of the two. We consider ConstantFP 1323/// to be the simplest, and ConstantExprs to be the most complex. 1324static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1325 assert(V1->getType() == V2->getType() && 1326 "Cannot compare values of different types!"); 1327 1328 // Handle degenerate case quickly 1329 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1330 1331 if (!isa<ConstantExpr>(V1)) { 1332 if (!isa<ConstantExpr>(V2)) { 1333 // Simple case, use the standard constant folder. 1334 ConstantInt *R = nullptr; 1335 R = dyn_cast<ConstantInt>( 1336 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1337 if (R && !R->isZero()) 1338 return FCmpInst::FCMP_OEQ; 1339 R = dyn_cast<ConstantInt>( 1340 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1341 if (R && !R->isZero()) 1342 return FCmpInst::FCMP_OLT; 1343 R = dyn_cast<ConstantInt>( 1344 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1345 if (R && !R->isZero()) 1346 return FCmpInst::FCMP_OGT; 1347 1348 // Nothing more we can do 1349 return FCmpInst::BAD_FCMP_PREDICATE; 1350 } 1351 1352 // If the first operand is simple and second is ConstantExpr, swap operands. 1353 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1354 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1355 return FCmpInst::getSwappedPredicate(SwappedRelation); 1356 } else { 1357 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1358 // constantexpr or a simple constant. 1359 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1360 switch (CE1->getOpcode()) { 1361 case Instruction::FPTrunc: 1362 case Instruction::FPExt: 1363 case Instruction::UIToFP: 1364 case Instruction::SIToFP: 1365 // We might be able to do something with these but we don't right now. 1366 break; 1367 default: 1368 break; 1369 } 1370 } 1371 // There are MANY other foldings that we could perform here. They will 1372 // probably be added on demand, as they seem needed. 1373 return FCmpInst::BAD_FCMP_PREDICATE; 1374} 1375 1376static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1377 const GlobalValue *GV2) { 1378 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1379 if (GV->hasExternalWeakLinkage() || GV->hasWeakAnyLinkage()) 1380 return true; 1381 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1382 Type *Ty = GVar->getType()->getPointerElementType(); 1383 // A global with opaque type might end up being zero sized. 1384 if (!Ty->isSized()) 1385 return true; 1386 // A global with an empty type might lie at the address of any other 1387 // global. 1388 if (Ty->isEmptyTy()) 1389 return true; 1390 } 1391 return false; 1392 }; 1393 // Don't try to decide equality of aliases. 1394 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1395 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1396 return ICmpInst::ICMP_NE; 1397 return ICmpInst::BAD_ICMP_PREDICATE; 1398} 1399 1400/// evaluateICmpRelation - This function determines if there is anything we can 1401/// decide about the two constants provided. This doesn't need to handle simple 1402/// things like integer comparisons, but should instead handle ConstantExprs 1403/// and GlobalValues. If we can determine that the two constants have a 1404/// particular relation to each other, we should return the corresponding ICmp 1405/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1406/// 1407/// To simplify this code we canonicalize the relation so that the first 1408/// operand is always the most "complex" of the two. We consider simple 1409/// constants (like ConstantInt) to be the simplest, followed by 1410/// GlobalValues, followed by ConstantExpr's (the most complex). 1411/// 1412static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1413 bool isSigned) { 1414 assert(V1->getType() == V2->getType() && 1415 "Cannot compare different types of values!"); 1416 if (V1 == V2) return ICmpInst::ICMP_EQ; 1417 1418 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1419 !isa<BlockAddress>(V1)) { 1420 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1421 !isa<BlockAddress>(V2)) { 1422 // We distilled this down to a simple case, use the standard constant 1423 // folder. 1424 ConstantInt *R = nullptr; 1425 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1426 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1427 if (R && !R->isZero()) 1428 return pred; 1429 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1430 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1431 if (R && !R->isZero()) 1432 return pred; 1433 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1434 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1435 if (R && !R->isZero()) 1436 return pred; 1437 1438 // If we couldn't figure it out, bail. 1439 return ICmpInst::BAD_ICMP_PREDICATE; 1440 } 1441 1442 // If the first operand is simple, swap operands. 1443 ICmpInst::Predicate SwappedRelation = 1444 evaluateICmpRelation(V2, V1, isSigned); 1445 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1446 return ICmpInst::getSwappedPredicate(SwappedRelation); 1447 1448 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1449 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1450 ICmpInst::Predicate SwappedRelation = 1451 evaluateICmpRelation(V2, V1, isSigned); 1452 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1453 return ICmpInst::getSwappedPredicate(SwappedRelation); 1454 return ICmpInst::BAD_ICMP_PREDICATE; 1455 } 1456 1457 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1458 // constant (which, since the types must match, means that it's a 1459 // ConstantPointerNull). 1460 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1461 return areGlobalsPotentiallyEqual(GV, GV2); 1462 } else if (isa<BlockAddress>(V2)) { 1463 return ICmpInst::ICMP_NE; // Globals never equal labels. 1464 } else { 1465 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1466 // GlobalVals can never be null unless they have external weak linkage. 1467 // We don't try to evaluate aliases here. 1468 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1469 return ICmpInst::ICMP_NE; 1470 } 1471 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1472 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1473 ICmpInst::Predicate SwappedRelation = 1474 evaluateICmpRelation(V2, V1, isSigned); 1475 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1476 return ICmpInst::getSwappedPredicate(SwappedRelation); 1477 return ICmpInst::BAD_ICMP_PREDICATE; 1478 } 1479 1480 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1481 // constant (which, since the types must match, means that it is a 1482 // ConstantPointerNull). 1483 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1484 // Block address in another function can't equal this one, but block 1485 // addresses in the current function might be the same if blocks are 1486 // empty. 1487 if (BA2->getFunction() != BA->getFunction()) 1488 return ICmpInst::ICMP_NE; 1489 } else { 1490 // Block addresses aren't null, don't equal the address of globals. 1491 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1492 "Canonicalization guarantee!"); 1493 return ICmpInst::ICMP_NE; 1494 } 1495 } else { 1496 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1497 // constantexpr, a global, block address, or a simple constant. 1498 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1499 Constant *CE1Op0 = CE1->getOperand(0); 1500 1501 switch (CE1->getOpcode()) { 1502 case Instruction::Trunc: 1503 case Instruction::FPTrunc: 1504 case Instruction::FPExt: 1505 case Instruction::FPToUI: 1506 case Instruction::FPToSI: 1507 break; // We can't evaluate floating point casts or truncations. 1508 1509 case Instruction::UIToFP: 1510 case Instruction::SIToFP: 1511 case Instruction::BitCast: 1512 case Instruction::ZExt: 1513 case Instruction::SExt: 1514 // If the cast is not actually changing bits, and the second operand is a 1515 // null pointer, do the comparison with the pre-casted value. 1516 if (V2->isNullValue() && 1517 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1518 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1519 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1520 return evaluateICmpRelation(CE1Op0, 1521 Constant::getNullValue(CE1Op0->getType()), 1522 isSigned); 1523 } 1524 break; 1525 1526 case Instruction::GetElementPtr: { 1527 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1528 // Ok, since this is a getelementptr, we know that the constant has a 1529 // pointer type. Check the various cases. 1530 if (isa<ConstantPointerNull>(V2)) { 1531 // If we are comparing a GEP to a null pointer, check to see if the base 1532 // of the GEP equals the null pointer. 1533 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1534 if (GV->hasExternalWeakLinkage()) 1535 // Weak linkage GVals could be zero or not. We're comparing that 1536 // to null pointer so its greater-or-equal 1537 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1538 else 1539 // If its not weak linkage, the GVal must have a non-zero address 1540 // so the result is greater-than 1541 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1542 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1543 // If we are indexing from a null pointer, check to see if we have any 1544 // non-zero indices. 1545 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1546 if (!CE1->getOperand(i)->isNullValue()) 1547 // Offsetting from null, must not be equal. 1548 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1549 // Only zero indexes from null, must still be zero. 1550 return ICmpInst::ICMP_EQ; 1551 } 1552 // Otherwise, we can't really say if the first operand is null or not. 1553 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1554 if (isa<ConstantPointerNull>(CE1Op0)) { 1555 if (GV2->hasExternalWeakLinkage()) 1556 // Weak linkage GVals could be zero or not. We're comparing it to 1557 // a null pointer, so its less-or-equal 1558 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1559 else 1560 // If its not weak linkage, the GVal must have a non-zero address 1561 // so the result is less-than 1562 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1563 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1564 if (GV == GV2) { 1565 // If this is a getelementptr of the same global, then it must be 1566 // different. Because the types must match, the getelementptr could 1567 // only have at most one index, and because we fold getelementptr's 1568 // with a single zero index, it must be nonzero. 1569 assert(CE1->getNumOperands() == 2 && 1570 !CE1->getOperand(1)->isNullValue() && 1571 "Surprising getelementptr!"); 1572 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1573 } else { 1574 if (CE1GEP->hasAllZeroIndices()) 1575 return areGlobalsPotentiallyEqual(GV, GV2); 1576 return ICmpInst::BAD_ICMP_PREDICATE; 1577 } 1578 } 1579 } else { 1580 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1581 Constant *CE2Op0 = CE2->getOperand(0); 1582 1583 // There are MANY other foldings that we could perform here. They will 1584 // probably be added on demand, as they seem needed. 1585 switch (CE2->getOpcode()) { 1586 default: break; 1587 case Instruction::GetElementPtr: 1588 // By far the most common case to handle is when the base pointers are 1589 // obviously to the same global. 1590 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1591 // Don't know relative ordering, but check for inequality. 1592 if (CE1Op0 != CE2Op0) { 1593 GEPOperator *CE2GEP = cast<GEPOperator>(CE2); 1594 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1595 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1596 cast<GlobalValue>(CE2Op0)); 1597 return ICmpInst::BAD_ICMP_PREDICATE; 1598 } 1599 // Ok, we know that both getelementptr instructions are based on the 1600 // same global. From this, we can precisely determine the relative 1601 // ordering of the resultant pointers. 1602 unsigned i = 1; 1603 1604 // The logic below assumes that the result of the comparison 1605 // can be determined by finding the first index that differs. 1606 // This doesn't work if there is over-indexing in any 1607 // subsequent indices, so check for that case first. 1608 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1609 !CE2->isGEPWithNoNotionalOverIndexing()) 1610 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1611 1612 // Compare all of the operands the GEP's have in common. 1613 gep_type_iterator GTI = gep_type_begin(CE1); 1614 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1615 ++i, ++GTI) 1616 switch (IdxCompare(CE1->getOperand(i), 1617 CE2->getOperand(i), GTI.getIndexedType())) { 1618 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1619 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1620 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1621 } 1622 1623 // Ok, we ran out of things they have in common. If any leftovers 1624 // are non-zero then we have a difference, otherwise we are equal. 1625 for (; i < CE1->getNumOperands(); ++i) 1626 if (!CE1->getOperand(i)->isNullValue()) { 1627 if (isa<ConstantInt>(CE1->getOperand(i))) 1628 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1629 else 1630 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1631 } 1632 1633 for (; i < CE2->getNumOperands(); ++i) 1634 if (!CE2->getOperand(i)->isNullValue()) { 1635 if (isa<ConstantInt>(CE2->getOperand(i))) 1636 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1637 else 1638 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1639 } 1640 return ICmpInst::ICMP_EQ; 1641 } 1642 } 1643 } 1644 } 1645 default: 1646 break; 1647 } 1648 } 1649 1650 return ICmpInst::BAD_ICMP_PREDICATE; 1651} 1652 1653Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1654 Constant *C1, Constant *C2) { 1655 Type *ResultTy; 1656 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1657 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1658 VT->getNumElements()); 1659 else 1660 ResultTy = Type::getInt1Ty(C1->getContext()); 1661 1662 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1663 if (pred == FCmpInst::FCMP_FALSE) 1664 return Constant::getNullValue(ResultTy); 1665 1666 if (pred == FCmpInst::FCMP_TRUE) 1667 return Constant::getAllOnesValue(ResultTy); 1668 1669 // Handle some degenerate cases first 1670 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1671 CmpInst::Predicate Predicate = CmpInst::Predicate(pred); 1672 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1673 // For EQ and NE, we can always pick a value for the undef to make the 1674 // predicate pass or fail, so we can return undef. 1675 // Also, if both operands are undef, we can return undef for int comparison. 1676 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1677 return UndefValue::get(ResultTy); 1678 1679 // Otherwise, for integer compare, pick the same value as the non-undef 1680 // operand, and fold it to true or false. 1681 if (isIntegerPredicate) 1682 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1683 1684 // Choosing NaN for the undef will always make unordered comparison succeed 1685 // and ordered comparison fails. 1686 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1687 } 1688 1689 // icmp eq/ne(null,GV) -> false/true 1690 if (C1->isNullValue()) { 1691 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1692 // Don't try to evaluate aliases. External weak GV can be null. 1693 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1694 if (pred == ICmpInst::ICMP_EQ) 1695 return ConstantInt::getFalse(C1->getContext()); 1696 else if (pred == ICmpInst::ICMP_NE) 1697 return ConstantInt::getTrue(C1->getContext()); 1698 } 1699 // icmp eq/ne(GV,null) -> false/true 1700 } else if (C2->isNullValue()) { 1701 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1702 // Don't try to evaluate aliases. External weak GV can be null. 1703 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1704 if (pred == ICmpInst::ICMP_EQ) 1705 return ConstantInt::getFalse(C1->getContext()); 1706 else if (pred == ICmpInst::ICMP_NE) 1707 return ConstantInt::getTrue(C1->getContext()); 1708 } 1709 } 1710 1711 // If the comparison is a comparison between two i1's, simplify it. 1712 if (C1->getType()->isIntegerTy(1)) { 1713 switch(pred) { 1714 case ICmpInst::ICMP_EQ: 1715 if (isa<ConstantInt>(C2)) 1716 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1717 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1718 case ICmpInst::ICMP_NE: 1719 return ConstantExpr::getXor(C1, C2); 1720 default: 1721 break; 1722 } 1723 } 1724 1725 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1726 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1727 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1728 switch (pred) { 1729 default: llvm_unreachable("Invalid ICmp Predicate"); 1730 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1731 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1732 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1733 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1734 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1735 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1736 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1737 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1738 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1739 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1740 } 1741 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1742 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1743 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1744 APFloat::cmpResult R = C1V.compare(C2V); 1745 switch (pred) { 1746 default: llvm_unreachable("Invalid FCmp Predicate"); 1747 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1748 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1749 case FCmpInst::FCMP_UNO: 1750 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1751 case FCmpInst::FCMP_ORD: 1752 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1753 case FCmpInst::FCMP_UEQ: 1754 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1755 R==APFloat::cmpEqual); 1756 case FCmpInst::FCMP_OEQ: 1757 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1758 case FCmpInst::FCMP_UNE: 1759 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1760 case FCmpInst::FCMP_ONE: 1761 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1762 R==APFloat::cmpGreaterThan); 1763 case FCmpInst::FCMP_ULT: 1764 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1765 R==APFloat::cmpLessThan); 1766 case FCmpInst::FCMP_OLT: 1767 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1768 case FCmpInst::FCMP_UGT: 1769 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1770 R==APFloat::cmpGreaterThan); 1771 case FCmpInst::FCMP_OGT: 1772 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1773 case FCmpInst::FCMP_ULE: 1774 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1775 case FCmpInst::FCMP_OLE: 1776 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1777 R==APFloat::cmpEqual); 1778 case FCmpInst::FCMP_UGE: 1779 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1780 case FCmpInst::FCMP_OGE: 1781 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1782 R==APFloat::cmpEqual); 1783 } 1784 } else if (C1->getType()->isVectorTy()) { 1785 // If we can constant fold the comparison of each element, constant fold 1786 // the whole vector comparison. 1787 SmallVector<Constant*, 4> ResElts; 1788 Type *Ty = IntegerType::get(C1->getContext(), 32); 1789 // Compare the elements, producing an i1 result or constant expr. 1790 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1791 Constant *C1E = 1792 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1793 Constant *C2E = 1794 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1795 1796 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1797 } 1798 1799 return ConstantVector::get(ResElts); 1800 } 1801 1802 if (C1->getType()->isFloatingPointTy() && 1803 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1804 // infinite recursive loop 1805 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1806 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1807 switch (evaluateFCmpRelation(C1, C2)) { 1808 default: llvm_unreachable("Unknown relation!"); 1809 case FCmpInst::FCMP_UNO: 1810 case FCmpInst::FCMP_ORD: 1811 case FCmpInst::FCMP_UEQ: 1812 case FCmpInst::FCMP_UNE: 1813 case FCmpInst::FCMP_ULT: 1814 case FCmpInst::FCMP_UGT: 1815 case FCmpInst::FCMP_ULE: 1816 case FCmpInst::FCMP_UGE: 1817 case FCmpInst::FCMP_TRUE: 1818 case FCmpInst::FCMP_FALSE: 1819 case FCmpInst::BAD_FCMP_PREDICATE: 1820 break; // Couldn't determine anything about these constants. 1821 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1822 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1823 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1824 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1825 break; 1826 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1827 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1828 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1829 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1830 break; 1831 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1832 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1833 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1834 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1835 break; 1836 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1837 // We can only partially decide this relation. 1838 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1839 Result = 0; 1840 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1841 Result = 1; 1842 break; 1843 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1844 // We can only partially decide this relation. 1845 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1846 Result = 0; 1847 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1848 Result = 1; 1849 break; 1850 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1851 // We can only partially decide this relation. 1852 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1853 Result = 0; 1854 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1855 Result = 1; 1856 break; 1857 } 1858 1859 // If we evaluated the result, return it now. 1860 if (Result != -1) 1861 return ConstantInt::get(ResultTy, Result); 1862 1863 } else { 1864 // Evaluate the relation between the two constants, per the predicate. 1865 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1866 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 1867 default: llvm_unreachable("Unknown relational!"); 1868 case ICmpInst::BAD_ICMP_PREDICATE: 1869 break; // Couldn't determine anything about these constants. 1870 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1871 // If we know the constants are equal, we can decide the result of this 1872 // computation precisely. 1873 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1874 break; 1875 case ICmpInst::ICMP_ULT: 1876 switch (pred) { 1877 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1878 Result = 1; break; 1879 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1880 Result = 0; break; 1881 } 1882 break; 1883 case ICmpInst::ICMP_SLT: 1884 switch (pred) { 1885 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1886 Result = 1; break; 1887 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1888 Result = 0; break; 1889 } 1890 break; 1891 case ICmpInst::ICMP_UGT: 1892 switch (pred) { 1893 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1894 Result = 1; break; 1895 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1896 Result = 0; break; 1897 } 1898 break; 1899 case ICmpInst::ICMP_SGT: 1900 switch (pred) { 1901 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1902 Result = 1; break; 1903 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1904 Result = 0; break; 1905 } 1906 break; 1907 case ICmpInst::ICMP_ULE: 1908 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1909 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1910 break; 1911 case ICmpInst::ICMP_SLE: 1912 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1913 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1914 break; 1915 case ICmpInst::ICMP_UGE: 1916 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1917 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1918 break; 1919 case ICmpInst::ICMP_SGE: 1920 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1921 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1922 break; 1923 case ICmpInst::ICMP_NE: 1924 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1925 if (pred == ICmpInst::ICMP_NE) Result = 1; 1926 break; 1927 } 1928 1929 // If we evaluated the result, return it now. 1930 if (Result != -1) 1931 return ConstantInt::get(ResultTy, Result); 1932 1933 // If the right hand side is a bitcast, try using its inverse to simplify 1934 // it by moving it to the left hand side. We can't do this if it would turn 1935 // a vector compare into a scalar compare or visa versa. 1936 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1937 Constant *CE2Op0 = CE2->getOperand(0); 1938 if (CE2->getOpcode() == Instruction::BitCast && 1939 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1940 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1941 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1942 } 1943 } 1944 1945 // If the left hand side is an extension, try eliminating it. 1946 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1947 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 1948 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 1949 Constant *CE1Op0 = CE1->getOperand(0); 1950 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1951 if (CE1Inverse == CE1Op0) { 1952 // Check whether we can safely truncate the right hand side. 1953 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1954 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1955 C2->getType()) == C2) 1956 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1957 } 1958 } 1959 } 1960 1961 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1962 (C1->isNullValue() && !C2->isNullValue())) { 1963 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1964 // other way if possible. 1965 // Also, if C1 is null and C2 isn't, flip them around. 1966 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1967 return ConstantExpr::getICmp(pred, C2, C1); 1968 } 1969 } 1970 return nullptr; 1971} 1972 1973/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1974/// is "inbounds". 1975template<typename IndexTy> 1976static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1977 // No indices means nothing that could be out of bounds. 1978 if (Idxs.empty()) return true; 1979 1980 // If the first index is zero, it's in bounds. 1981 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 1982 1983 // If the first index is one and all the rest are zero, it's in bounds, 1984 // by the one-past-the-end rule. 1985 if (!cast<ConstantInt>(Idxs[0])->isOne()) 1986 return false; 1987 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 1988 if (!cast<Constant>(Idxs[i])->isNullValue()) 1989 return false; 1990 return true; 1991} 1992 1993/// \brief Test whether a given ConstantInt is in-range for a SequentialType. 1994static bool isIndexInRangeOfSequentialType(const SequentialType *STy, 1995 const ConstantInt *CI) { 1996 if (const PointerType *PTy = dyn_cast<PointerType>(STy)) 1997 // Only handle pointers to sized types, not pointers to functions. 1998 return PTy->getElementType()->isSized(); 1999 2000 uint64_t NumElements = 0; 2001 // Determine the number of elements in our sequential type. 2002 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 2003 NumElements = ATy->getNumElements(); 2004 else if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 2005 NumElements = VTy->getNumElements(); 2006 2007 assert((isa<ArrayType>(STy) || NumElements > 0) && 2008 "didn't expect non-array type to have zero elements!"); 2009 2010 // We cannot bounds check the index if it doesn't fit in an int64_t. 2011 if (CI->getValue().getActiveBits() > 64) 2012 return false; 2013 2014 // A negative index or an index past the end of our sequential type is 2015 // considered out-of-range. 2016 int64_t IndexVal = CI->getSExtValue(); 2017 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 2018 return false; 2019 2020 // Otherwise, it is in-range. 2021 return true; 2022} 2023 2024template<typename IndexTy> 2025static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 2026 bool inBounds, 2027 ArrayRef<IndexTy> Idxs) { 2028 if (Idxs.empty()) return C; 2029 Constant *Idx0 = cast<Constant>(Idxs[0]); 2030 if ((Idxs.size() == 1 && Idx0->isNullValue())) 2031 return C; 2032 2033 if (isa<UndefValue>(C)) { 2034 PointerType *Ptr = cast<PointerType>(C->getType()); 2035 Type *Ty = GetElementPtrInst::getIndexedType( 2036 cast<PointerType>(Ptr->getScalarType())->getElementType(), Idxs); 2037 assert(Ty && "Invalid indices for GEP!"); 2038 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 2039 } 2040 2041 if (C->isNullValue()) { 2042 bool isNull = true; 2043 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2044 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2045 isNull = false; 2046 break; 2047 } 2048 if (isNull) { 2049 PointerType *Ptr = cast<PointerType>(C->getType()); 2050 Type *Ty = GetElementPtrInst::getIndexedType( 2051 cast<PointerType>(Ptr->getScalarType())->getElementType(), Idxs); 2052 assert(Ty && "Invalid indices for GEP!"); 2053 return ConstantPointerNull::get(PointerType::get(Ty, 2054 Ptr->getAddressSpace())); 2055 } 2056 } 2057 2058 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2059 // Combine Indices - If the source pointer to this getelementptr instruction 2060 // is a getelementptr instruction, combine the indices of the two 2061 // getelementptr instructions into a single instruction. 2062 // 2063 if (CE->getOpcode() == Instruction::GetElementPtr) { 2064 Type *LastTy = nullptr; 2065 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2066 I != E; ++I) 2067 LastTy = *I; 2068 2069 // We cannot combine indices if doing so would take us outside of an 2070 // array or vector. Doing otherwise could trick us if we evaluated such a 2071 // GEP as part of a load. 2072 // 2073 // e.g. Consider if the original GEP was: 2074 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2075 // i32 0, i32 0, i64 0) 2076 // 2077 // If we then tried to offset it by '8' to get to the third element, 2078 // an i8, we should *not* get: 2079 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 2080 // i32 0, i32 0, i64 8) 2081 // 2082 // This GEP tries to index array element '8 which runs out-of-bounds. 2083 // Subsequent evaluation would get confused and produce erroneous results. 2084 // 2085 // The following prohibits such a GEP from being formed by checking to see 2086 // if the index is in-range with respect to an array or vector. 2087 bool PerformFold = false; 2088 if (Idx0->isNullValue()) 2089 PerformFold = true; 2090 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy)) 2091 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 2092 PerformFold = isIndexInRangeOfSequentialType(STy, CI); 2093 2094 if (PerformFold) { 2095 SmallVector<Value*, 16> NewIndices; 2096 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2097 NewIndices.append(CE->op_begin() + 1, CE->op_end() - 1); 2098 2099 // Add the last index of the source with the first index of the new GEP. 2100 // Make sure to handle the case when they are actually different types. 2101 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2102 // Otherwise it must be an array. 2103 if (!Idx0->isNullValue()) { 2104 Type *IdxTy = Combined->getType(); 2105 if (IdxTy != Idx0->getType()) { 2106 unsigned CommonExtendedWidth = 2107 std::max(IdxTy->getIntegerBitWidth(), 2108 Idx0->getType()->getIntegerBitWidth()); 2109 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2110 2111 Type *CommonTy = 2112 Type::getIntNTy(IdxTy->getContext(), CommonExtendedWidth); 2113 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 2114 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, CommonTy); 2115 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2116 } else { 2117 Combined = 2118 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2119 } 2120 } 2121 2122 NewIndices.push_back(Combined); 2123 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2124 return ConstantExpr::getGetElementPtr( 2125 cast<GEPOperator>(CE)->getSourceElementType(), CE->getOperand(0), 2126 NewIndices, inBounds && cast<GEPOperator>(CE)->isInBounds()); 2127 } 2128 } 2129 2130 // Attempt to fold casts to the same type away. For example, folding: 2131 // 2132 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2133 // i64 0, i64 0) 2134 // into: 2135 // 2136 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2137 // 2138 // Don't fold if the cast is changing address spaces. 2139 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2140 PointerType *SrcPtrTy = 2141 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2142 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2143 if (SrcPtrTy && DstPtrTy) { 2144 ArrayType *SrcArrayTy = 2145 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2146 ArrayType *DstArrayTy = 2147 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2148 if (SrcArrayTy && DstArrayTy 2149 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2150 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2151 return ConstantExpr::getGetElementPtr( 2152 SrcArrayTy, (Constant *)CE->getOperand(0), Idxs, inBounds); 2153 } 2154 } 2155 } 2156 2157 // Check to see if any array indices are not within the corresponding 2158 // notional array or vector bounds. If so, try to determine if they can be 2159 // factored out into preceding dimensions. 2160 bool Unknown = false; 2161 SmallVector<Constant *, 8> NewIdxs; 2162 Type *Ty = C->getType(); 2163 Type *Prev = nullptr; 2164 for (unsigned i = 0, e = Idxs.size(); i != e; 2165 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2166 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2167 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) 2168 if (CI->getSExtValue() > 0 && 2169 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) { 2170 if (isa<SequentialType>(Prev)) { 2171 // It's out of range, but we can factor it into the prior 2172 // dimension. 2173 NewIdxs.resize(Idxs.size()); 2174 uint64_t NumElements = 0; 2175 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2176 NumElements = ATy->getNumElements(); 2177 else 2178 NumElements = cast<VectorType>(Ty)->getNumElements(); 2179 2180 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2181 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2182 2183 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2184 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2185 2186 unsigned CommonExtendedWidth = 2187 std::max(PrevIdx->getType()->getIntegerBitWidth(), 2188 Div->getType()->getIntegerBitWidth()); 2189 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2190 2191 // Before adding, extend both operands to i64 to avoid 2192 // overflow trouble. 2193 if (!PrevIdx->getType()->isIntegerTy(CommonExtendedWidth)) 2194 PrevIdx = ConstantExpr::getSExt( 2195 PrevIdx, 2196 Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2197 if (!Div->getType()->isIntegerTy(CommonExtendedWidth)) 2198 Div = ConstantExpr::getSExt( 2199 Div, Type::getIntNTy(Div->getContext(), CommonExtendedWidth)); 2200 2201 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2202 } else { 2203 // It's out of range, but the prior dimension is a struct 2204 // so we can't do anything about it. 2205 Unknown = true; 2206 } 2207 } 2208 } else { 2209 // We don't know if it's in range or not. 2210 Unknown = true; 2211 } 2212 } 2213 2214 // If we did any factoring, start over with the adjusted indices. 2215 if (!NewIdxs.empty()) { 2216 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2217 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2218 return ConstantExpr::getGetElementPtr(nullptr, C, NewIdxs, inBounds); 2219 } 2220 2221 // If all indices are known integers and normalized, we can do a simple 2222 // check for the "inbounds" property. 2223 if (!Unknown && !inBounds) 2224 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2225 if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs)) 2226 return ConstantExpr::getInBoundsGetElementPtr(nullptr, C, Idxs); 2227 2228 return nullptr; 2229} 2230 2231Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2232 bool inBounds, 2233 ArrayRef<Constant *> Idxs) { 2234 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2235} 2236 2237Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2238 bool inBounds, 2239 ArrayRef<Value *> Idxs) { 2240 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2241} 2242