ConstantFold.cpp revision 261991
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/GlobalAlias.h" 26#include "llvm/IR/GlobalVariable.h" 27#include "llvm/IR/Instructions.h" 28#include "llvm/IR/Operator.h" 29#include "llvm/Support/Compiler.h" 30#include "llvm/Support/ErrorHandling.h" 31#include "llvm/Support/GetElementPtrTypeIterator.h" 32#include "llvm/Support/ManagedStatic.h" 33#include "llvm/Support/MathExtras.h" 34#include <limits> 35using namespace llvm; 36 37//===----------------------------------------------------------------------===// 38// ConstantFold*Instruction Implementations 39//===----------------------------------------------------------------------===// 40 41/// BitCastConstantVector - Convert the specified vector Constant node to the 42/// specified vector type. At this point, we know that the elements of the 43/// input vector constant are all simple integer or FP values. 44static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 45 46 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 47 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 48 49 // If this cast changes element count then we can't handle it here: 50 // doing so requires endianness information. This should be handled by 51 // Analysis/ConstantFolding.cpp 52 unsigned NumElts = DstTy->getNumElements(); 53 if (NumElts != CV->getType()->getVectorNumElements()) 54 return 0; 55 56 Type *DstEltTy = DstTy->getElementType(); 57 58 SmallVector<Constant*, 16> Result; 59 Type *Ty = IntegerType::get(CV->getContext(), 32); 60 for (unsigned i = 0; i != NumElts; ++i) { 61 Constant *C = 62 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 63 C = ConstantExpr::getBitCast(C, DstEltTy); 64 Result.push_back(C); 65 } 66 67 return ConstantVector::get(Result); 68} 69 70/// This function determines which opcode to use to fold two constant cast 71/// expressions together. It uses CastInst::isEliminableCastPair to determine 72/// the opcode. Consequently its just a wrapper around that function. 73/// @brief Determine if it is valid to fold a cast of a cast 74static unsigned 75foldConstantCastPair( 76 unsigned opc, ///< opcode of the second cast constant expression 77 ConstantExpr *Op, ///< the first cast constant expression 78 Type *DstTy ///< destination type of the first cast 79) { 80 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 81 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 82 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 83 84 // The the types and opcodes for the two Cast constant expressions 85 Type *SrcTy = Op->getOperand(0)->getType(); 86 Type *MidTy = Op->getType(); 87 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 88 Instruction::CastOps secondOp = Instruction::CastOps(opc); 89 90 // Assume that pointers are never more than 64 bits wide, and only use this 91 // for the middle type. Otherwise we could end up folding away illegal 92 // bitcasts between address spaces with different sizes. 93 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 94 95 // Let CastInst::isEliminableCastPair do the heavy lifting. 96 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 97 0, FakeIntPtrTy, 0); 98} 99 100static Constant *FoldBitCast(Constant *V, Type *DestTy) { 101 Type *SrcTy = V->getType(); 102 if (SrcTy == DestTy) 103 return V; // no-op cast 104 105 // Check to see if we are casting a pointer to an aggregate to a pointer to 106 // the first element. If so, return the appropriate GEP instruction. 107 if (PointerType *PTy = dyn_cast<PointerType>(V->getType())) 108 if (PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 109 if (PTy->getAddressSpace() == DPTy->getAddressSpace() 110 && DPTy->getElementType()->isSized()) { 111 SmallVector<Value*, 8> IdxList; 112 Value *Zero = 113 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 114 IdxList.push_back(Zero); 115 Type *ElTy = PTy->getElementType(); 116 while (ElTy != DPTy->getElementType()) { 117 if (StructType *STy = dyn_cast<StructType>(ElTy)) { 118 if (STy->getNumElements() == 0) break; 119 ElTy = STy->getElementType(0); 120 IdxList.push_back(Zero); 121 } else if (SequentialType *STy = 122 dyn_cast<SequentialType>(ElTy)) { 123 if (ElTy->isPointerTy()) break; // Can't index into pointers! 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(V, IdxList); 134 } 135 136 // Handle casts from one vector constant to another. We know that the src 137 // and dest type have the same size (otherwise its an illegal cast). 138 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 139 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 140 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 141 "Not cast between same sized vectors!"); 142 SrcTy = NULL; 143 // First, check for null. Undef is already handled. 144 if (isa<ConstantAggregateZero>(V)) 145 return Constant::getNullValue(DestTy); 146 147 // Handle ConstantVector and ConstantAggregateVector. 148 return BitCastConstantVector(V, DestPTy); 149 } 150 151 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 152 // This allows for other simplifications (although some of them 153 // can only be handled by Analysis/ConstantFolding.cpp). 154 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 155 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 156 } 157 158 // Finally, implement bitcast folding now. The code below doesn't handle 159 // bitcast right. 160 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 161 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 162 163 // Handle integral constant input. 164 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 165 if (DestTy->isIntegerTy()) 166 // Integral -> Integral. This is a no-op because the bit widths must 167 // be the same. Consequently, we just fold to V. 168 return V; 169 170 if (DestTy->isFloatingPointTy()) 171 return ConstantFP::get(DestTy->getContext(), 172 APFloat(DestTy->getFltSemantics(), 173 CI->getValue())); 174 175 // Otherwise, can't fold this (vector?) 176 return 0; 177 } 178 179 // Handle ConstantFP input: FP -> Integral. 180 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 181 return ConstantInt::get(FP->getContext(), 182 FP->getValueAPF().bitcastToAPInt()); 183 184 return 0; 185} 186 187 188/// ExtractConstantBytes - V is an integer constant which only has a subset of 189/// its bytes used. The bytes used are indicated by ByteStart (which is the 190/// first byte used, counting from the least significant byte) and ByteSize, 191/// which is the number of bytes used. 192/// 193/// This function analyzes the specified constant to see if the specified byte 194/// range can be returned as a simplified constant. If so, the constant is 195/// returned, otherwise null is returned. 196/// 197static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 198 unsigned ByteSize) { 199 assert(C->getType()->isIntegerTy() && 200 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 201 "Non-byte sized integer input"); 202 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 203 assert(ByteSize && "Must be accessing some piece"); 204 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 205 assert(ByteSize != CSize && "Should not extract everything"); 206 207 // Constant Integers are simple. 208 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 209 APInt V = CI->getValue(); 210 if (ByteStart) 211 V = V.lshr(ByteStart*8); 212 V = V.trunc(ByteSize*8); 213 return ConstantInt::get(CI->getContext(), V); 214 } 215 216 // In the input is a constant expr, we might be able to recursively simplify. 217 // If not, we definitely can't do anything. 218 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 219 if (CE == 0) return 0; 220 221 switch (CE->getOpcode()) { 222 default: return 0; 223 case Instruction::Or: { 224 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 225 if (RHS == 0) 226 return 0; 227 228 // X | -1 -> -1. 229 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 230 if (RHSC->isAllOnesValue()) 231 return RHSC; 232 233 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 234 if (LHS == 0) 235 return 0; 236 return ConstantExpr::getOr(LHS, RHS); 237 } 238 case Instruction::And: { 239 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 240 if (RHS == 0) 241 return 0; 242 243 // X & 0 -> 0. 244 if (RHS->isNullValue()) 245 return RHS; 246 247 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 248 if (LHS == 0) 249 return 0; 250 return ConstantExpr::getAnd(LHS, RHS); 251 } 252 case Instruction::LShr: { 253 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 254 if (Amt == 0) 255 return 0; 256 unsigned ShAmt = Amt->getZExtValue(); 257 // Cannot analyze non-byte shifts. 258 if ((ShAmt & 7) != 0) 259 return 0; 260 ShAmt >>= 3; 261 262 // If the extract is known to be all zeros, return zero. 263 if (ByteStart >= CSize-ShAmt) 264 return Constant::getNullValue(IntegerType::get(CE->getContext(), 265 ByteSize*8)); 266 // If the extract is known to be fully in the input, extract it. 267 if (ByteStart+ByteSize+ShAmt <= CSize) 268 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 269 270 // TODO: Handle the 'partially zero' case. 271 return 0; 272 } 273 274 case Instruction::Shl: { 275 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 276 if (Amt == 0) 277 return 0; 278 unsigned ShAmt = Amt->getZExtValue(); 279 // Cannot analyze non-byte shifts. 280 if ((ShAmt & 7) != 0) 281 return 0; 282 ShAmt >>= 3; 283 284 // If the extract is known to be all zeros, return zero. 285 if (ByteStart+ByteSize <= ShAmt) 286 return Constant::getNullValue(IntegerType::get(CE->getContext(), 287 ByteSize*8)); 288 // If the extract is known to be fully in the input, extract it. 289 if (ByteStart >= ShAmt) 290 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 291 292 // TODO: Handle the 'partially zero' case. 293 return 0; 294 } 295 296 case Instruction::ZExt: { 297 unsigned SrcBitSize = 298 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 299 300 // If extracting something that is completely zero, return 0. 301 if (ByteStart*8 >= SrcBitSize) 302 return Constant::getNullValue(IntegerType::get(CE->getContext(), 303 ByteSize*8)); 304 305 // If exactly extracting the input, return it. 306 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 307 return CE->getOperand(0); 308 309 // If extracting something completely in the input, if if the input is a 310 // multiple of 8 bits, recurse. 311 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 312 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 313 314 // Otherwise, if extracting a subset of the input, which is not multiple of 315 // 8 bits, do a shift and trunc to get the bits. 316 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 317 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 318 Constant *Res = CE->getOperand(0); 319 if (ByteStart) 320 Res = ConstantExpr::getLShr(Res, 321 ConstantInt::get(Res->getType(), ByteStart*8)); 322 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 323 ByteSize*8)); 324 } 325 326 // TODO: Handle the 'partially zero' case. 327 return 0; 328 } 329 } 330} 331 332/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 333/// on Ty, with any known factors factored out. If Folded is false, 334/// return null if no factoring was possible, to avoid endlessly 335/// bouncing an unfoldable expression back into the top-level folder. 336/// 337static Constant *getFoldedSizeOf(Type *Ty, Type *DestTy, 338 bool Folded) { 339 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 340 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 341 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 342 return ConstantExpr::getNUWMul(E, N); 343 } 344 345 if (StructType *STy = dyn_cast<StructType>(Ty)) 346 if (!STy->isPacked()) { 347 unsigned NumElems = STy->getNumElements(); 348 // An empty struct has size zero. 349 if (NumElems == 0) 350 return ConstantExpr::getNullValue(DestTy); 351 // Check for a struct with all members having the same size. 352 Constant *MemberSize = 353 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 354 bool AllSame = true; 355 for (unsigned i = 1; i != NumElems; ++i) 356 if (MemberSize != 357 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 358 AllSame = false; 359 break; 360 } 361 if (AllSame) { 362 Constant *N = ConstantInt::get(DestTy, NumElems); 363 return ConstantExpr::getNUWMul(MemberSize, N); 364 } 365 } 366 367 // Pointer size doesn't depend on the pointee type, so canonicalize them 368 // to an arbitrary pointee. 369 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 370 if (!PTy->getElementType()->isIntegerTy(1)) 371 return 372 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 373 PTy->getAddressSpace()), 374 DestTy, true); 375 376 // If there's no interesting folding happening, bail so that we don't create 377 // a constant that looks like it needs folding but really doesn't. 378 if (!Folded) 379 return 0; 380 381 // Base case: Get a regular sizeof expression. 382 Constant *C = ConstantExpr::getSizeOf(Ty); 383 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 384 DestTy, false), 385 C, DestTy); 386 return C; 387} 388 389/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 390/// on Ty, with any known factors factored out. If Folded is false, 391/// return null if no factoring was possible, to avoid endlessly 392/// bouncing an unfoldable expression back into the top-level folder. 393/// 394static Constant *getFoldedAlignOf(Type *Ty, Type *DestTy, 395 bool Folded) { 396 // The alignment of an array is equal to the alignment of the 397 // array element. Note that this is not always true for vectors. 398 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 399 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 400 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 401 DestTy, 402 false), 403 C, DestTy); 404 return C; 405 } 406 407 if (StructType *STy = dyn_cast<StructType>(Ty)) { 408 // Packed structs always have an alignment of 1. 409 if (STy->isPacked()) 410 return ConstantInt::get(DestTy, 1); 411 412 // Otherwise, struct alignment is the maximum alignment of any member. 413 // Without target data, we can't compare much, but we can check to see 414 // if all the members have the same alignment. 415 unsigned NumElems = STy->getNumElements(); 416 // An empty struct has minimal alignment. 417 if (NumElems == 0) 418 return ConstantInt::get(DestTy, 1); 419 // Check for a struct with all members having the same alignment. 420 Constant *MemberAlign = 421 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 422 bool AllSame = true; 423 for (unsigned i = 1; i != NumElems; ++i) 424 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 425 AllSame = false; 426 break; 427 } 428 if (AllSame) 429 return MemberAlign; 430 } 431 432 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 433 // to an arbitrary pointee. 434 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 435 if (!PTy->getElementType()->isIntegerTy(1)) 436 return 437 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 438 1), 439 PTy->getAddressSpace()), 440 DestTy, true); 441 442 // If there's no interesting folding happening, bail so that we don't create 443 // a constant that looks like it needs folding but really doesn't. 444 if (!Folded) 445 return 0; 446 447 // Base case: Get a regular alignof expression. 448 Constant *C = ConstantExpr::getAlignOf(Ty); 449 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 450 DestTy, false), 451 C, DestTy); 452 return C; 453} 454 455/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 456/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 457/// return null if no factoring was possible, to avoid endlessly 458/// bouncing an unfoldable expression back into the top-level folder. 459/// 460static Constant *getFoldedOffsetOf(Type *Ty, Constant *FieldNo, 461 Type *DestTy, 462 bool Folded) { 463 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 464 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 465 DestTy, false), 466 FieldNo, DestTy); 467 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 468 return ConstantExpr::getNUWMul(E, N); 469 } 470 471 if (StructType *STy = dyn_cast<StructType>(Ty)) 472 if (!STy->isPacked()) { 473 unsigned NumElems = STy->getNumElements(); 474 // An empty struct has no members. 475 if (NumElems == 0) 476 return 0; 477 // Check for a struct with all members having the same size. 478 Constant *MemberSize = 479 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 480 bool AllSame = true; 481 for (unsigned i = 1; i != NumElems; ++i) 482 if (MemberSize != 483 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 484 AllSame = false; 485 break; 486 } 487 if (AllSame) { 488 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 489 false, 490 DestTy, 491 false), 492 FieldNo, DestTy); 493 return ConstantExpr::getNUWMul(MemberSize, N); 494 } 495 } 496 497 // If there's no interesting folding happening, bail so that we don't create 498 // a constant that looks like it needs folding but really doesn't. 499 if (!Folded) 500 return 0; 501 502 // Base case: Get a regular offsetof expression. 503 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 504 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 505 DestTy, false), 506 C, DestTy); 507 return C; 508} 509 510Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 511 Type *DestTy) { 512 if (isa<UndefValue>(V)) { 513 // zext(undef) = 0, because the top bits will be zero. 514 // sext(undef) = 0, because the top bits will all be the same. 515 // [us]itofp(undef) = 0, because the result value is bounded. 516 if (opc == Instruction::ZExt || opc == Instruction::SExt || 517 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 518 return Constant::getNullValue(DestTy); 519 return UndefValue::get(DestTy); 520 } 521 522 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 523 return Constant::getNullValue(DestTy); 524 525 // If the cast operand is a constant expression, there's a few things we can 526 // do to try to simplify it. 527 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 528 if (CE->isCast()) { 529 // Try hard to fold cast of cast because they are often eliminable. 530 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 531 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 532 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 533 // If all of the indexes in the GEP are null values, there is no pointer 534 // adjustment going on. We might as well cast the source pointer. 535 bool isAllNull = true; 536 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 537 if (!CE->getOperand(i)->isNullValue()) { 538 isAllNull = false; 539 break; 540 } 541 if (isAllNull) 542 // This is casting one pointer type to another, always BitCast 543 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 544 } 545 } 546 547 // If the cast operand is a constant vector, perform the cast by 548 // operating on each element. In the cast of bitcasts, the element 549 // count may be mismatched; don't attempt to handle that here. 550 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 551 DestTy->isVectorTy() && 552 DestTy->getVectorNumElements() == V->getType()->getVectorNumElements()) { 553 SmallVector<Constant*, 16> res; 554 VectorType *DestVecTy = cast<VectorType>(DestTy); 555 Type *DstEltTy = DestVecTy->getElementType(); 556 Type *Ty = IntegerType::get(V->getContext(), 32); 557 for (unsigned i = 0, e = V->getType()->getVectorNumElements(); i != e; ++i) { 558 Constant *C = 559 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 560 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 561 } 562 return ConstantVector::get(res); 563 } 564 565 // We actually have to do a cast now. Perform the cast according to the 566 // opcode specified. 567 switch (opc) { 568 default: 569 llvm_unreachable("Failed to cast constant expression"); 570 case Instruction::FPTrunc: 571 case Instruction::FPExt: 572 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 573 bool ignored; 574 APFloat Val = FPC->getValueAPF(); 575 Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf : 576 DestTy->isFloatTy() ? APFloat::IEEEsingle : 577 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 578 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 579 DestTy->isFP128Ty() ? APFloat::IEEEquad : 580 DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble : 581 APFloat::Bogus, 582 APFloat::rmNearestTiesToEven, &ignored); 583 return ConstantFP::get(V->getContext(), Val); 584 } 585 return 0; // Can't fold. 586 case Instruction::FPToUI: 587 case Instruction::FPToSI: 588 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 589 const APFloat &V = FPC->getValueAPF(); 590 bool ignored; 591 uint64_t x[2]; 592 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 593 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 594 APFloat::rmTowardZero, &ignored); 595 APInt Val(DestBitWidth, x); 596 return ConstantInt::get(FPC->getContext(), Val); 597 } 598 return 0; // Can't fold. 599 case Instruction::IntToPtr: //always treated as unsigned 600 if (V->isNullValue()) // Is it an integral null value? 601 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 602 return 0; // Other pointer types cannot be casted 603 case Instruction::PtrToInt: // always treated as unsigned 604 // Is it a null pointer value? 605 if (V->isNullValue()) 606 return ConstantInt::get(DestTy, 0); 607 // If this is a sizeof-like expression, pull out multiplications by 608 // known factors to expose them to subsequent folding. If it's an 609 // alignof-like expression, factor out known factors. 610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 611 if (CE->getOpcode() == Instruction::GetElementPtr && 612 CE->getOperand(0)->isNullValue()) { 613 Type *Ty = 614 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 615 if (CE->getNumOperands() == 2) { 616 // Handle a sizeof-like expression. 617 Constant *Idx = CE->getOperand(1); 618 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 619 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 620 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 621 DestTy, false), 622 Idx, DestTy); 623 return ConstantExpr::getMul(C, Idx); 624 } 625 } else if (CE->getNumOperands() == 3 && 626 CE->getOperand(1)->isNullValue()) { 627 // Handle an alignof-like expression. 628 if (StructType *STy = dyn_cast<StructType>(Ty)) 629 if (!STy->isPacked()) { 630 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 631 if (CI->isOne() && 632 STy->getNumElements() == 2 && 633 STy->getElementType(0)->isIntegerTy(1)) { 634 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 635 } 636 } 637 // Handle an offsetof-like expression. 638 if (Ty->isStructTy() || Ty->isArrayTy()) { 639 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 640 DestTy, false)) 641 return C; 642 } 643 } 644 } 645 // Other pointer types cannot be casted 646 return 0; 647 case Instruction::UIToFP: 648 case Instruction::SIToFP: 649 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 650 APInt api = CI->getValue(); 651 APFloat apf(DestTy->getFltSemantics(), 652 APInt::getNullValue(DestTy->getPrimitiveSizeInBits())); 653 (void)apf.convertFromAPInt(api, 654 opc==Instruction::SIToFP, 655 APFloat::rmNearestTiesToEven); 656 return ConstantFP::get(V->getContext(), apf); 657 } 658 return 0; 659 case Instruction::ZExt: 660 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 661 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 662 return ConstantInt::get(V->getContext(), 663 CI->getValue().zext(BitWidth)); 664 } 665 return 0; 666 case Instruction::SExt: 667 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 668 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 669 return ConstantInt::get(V->getContext(), 670 CI->getValue().sext(BitWidth)); 671 } 672 return 0; 673 case Instruction::Trunc: { 674 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 675 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 676 return ConstantInt::get(V->getContext(), 677 CI->getValue().trunc(DestBitWidth)); 678 } 679 680 // The input must be a constantexpr. See if we can simplify this based on 681 // the bytes we are demanding. Only do this if the source and dest are an 682 // even multiple of a byte. 683 if ((DestBitWidth & 7) == 0 && 684 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 685 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 686 return Res; 687 688 return 0; 689 } 690 case Instruction::BitCast: 691 return FoldBitCast(V, DestTy); 692 case Instruction::AddrSpaceCast: 693 return 0; 694 } 695} 696 697Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 698 Constant *V1, Constant *V2) { 699 // Check for i1 and vector true/false conditions. 700 if (Cond->isNullValue()) return V2; 701 if (Cond->isAllOnesValue()) return V1; 702 703 // If the condition is a vector constant, fold the result elementwise. 704 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 705 SmallVector<Constant*, 16> Result; 706 Type *Ty = IntegerType::get(CondV->getContext(), 32); 707 for (unsigned i = 0, e = V1->getType()->getVectorNumElements(); i != e;++i){ 708 ConstantInt *Cond = dyn_cast<ConstantInt>(CondV->getOperand(i)); 709 if (Cond == 0) break; 710 711 Constant *V = Cond->isNullValue() ? V2 : V1; 712 Constant *Res = ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 713 Result.push_back(Res); 714 } 715 716 // If we were able to build the vector, return it. 717 if (Result.size() == V1->getType()->getVectorNumElements()) 718 return ConstantVector::get(Result); 719 } 720 721 if (isa<UndefValue>(Cond)) { 722 if (isa<UndefValue>(V1)) return V1; 723 return V2; 724 } 725 if (isa<UndefValue>(V1)) return V2; 726 if (isa<UndefValue>(V2)) return V1; 727 if (V1 == V2) return V1; 728 729 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 730 if (TrueVal->getOpcode() == Instruction::Select) 731 if (TrueVal->getOperand(0) == Cond) 732 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 733 } 734 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 735 if (FalseVal->getOpcode() == Instruction::Select) 736 if (FalseVal->getOperand(0) == Cond) 737 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 738 } 739 740 return 0; 741} 742 743Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 744 Constant *Idx) { 745 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 746 return UndefValue::get(Val->getType()->getVectorElementType()); 747 if (Val->isNullValue()) // ee(zero, x) -> zero 748 return Constant::getNullValue(Val->getType()->getVectorElementType()); 749 // ee({w,x,y,z}, undef) -> undef 750 if (isa<UndefValue>(Idx)) 751 return UndefValue::get(Val->getType()->getVectorElementType()); 752 753 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 754 uint64_t Index = CIdx->getZExtValue(); 755 // ee({w,x,y,z}, wrong_value) -> undef 756 if (Index >= Val->getType()->getVectorNumElements()) 757 return UndefValue::get(Val->getType()->getVectorElementType()); 758 return Val->getAggregateElement(Index); 759 } 760 return 0; 761} 762 763Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 764 Constant *Elt, 765 Constant *Idx) { 766 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 767 if (!CIdx) return 0; 768 const APInt &IdxVal = CIdx->getValue(); 769 770 SmallVector<Constant*, 16> Result; 771 Type *Ty = IntegerType::get(Val->getContext(), 32); 772 for (unsigned i = 0, e = Val->getType()->getVectorNumElements(); i != e; ++i){ 773 if (i == IdxVal) { 774 Result.push_back(Elt); 775 continue; 776 } 777 778 Constant *C = 779 ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 780 Result.push_back(C); 781 } 782 783 return ConstantVector::get(Result); 784} 785 786Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 787 Constant *V2, 788 Constant *Mask) { 789 unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); 790 Type *EltTy = V1->getType()->getVectorElementType(); 791 792 // Undefined shuffle mask -> undefined value. 793 if (isa<UndefValue>(Mask)) 794 return UndefValue::get(VectorType::get(EltTy, MaskNumElts)); 795 796 // Don't break the bitcode reader hack. 797 if (isa<ConstantExpr>(Mask)) return 0; 798 799 unsigned SrcNumElts = V1->getType()->getVectorNumElements(); 800 801 // Loop over the shuffle mask, evaluating each element. 802 SmallVector<Constant*, 32> Result; 803 for (unsigned i = 0; i != MaskNumElts; ++i) { 804 int Elt = ShuffleVectorInst::getMaskValue(Mask, i); 805 if (Elt == -1) { 806 Result.push_back(UndefValue::get(EltTy)); 807 continue; 808 } 809 Constant *InElt; 810 if (unsigned(Elt) >= SrcNumElts*2) 811 InElt = UndefValue::get(EltTy); 812 else if (unsigned(Elt) >= SrcNumElts) { 813 Type *Ty = IntegerType::get(V2->getContext(), 32); 814 InElt = 815 ConstantExpr::getExtractElement(V2, 816 ConstantInt::get(Ty, Elt - SrcNumElts)); 817 } else { 818 Type *Ty = IntegerType::get(V1->getContext(), 32); 819 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 820 } 821 Result.push_back(InElt); 822 } 823 824 return ConstantVector::get(Result); 825} 826 827Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 828 ArrayRef<unsigned> Idxs) { 829 // Base case: no indices, so return the entire value. 830 if (Idxs.empty()) 831 return Agg; 832 833 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 834 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 835 836 return 0; 837} 838 839Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 840 Constant *Val, 841 ArrayRef<unsigned> Idxs) { 842 // Base case: no indices, so replace the entire value. 843 if (Idxs.empty()) 844 return Val; 845 846 unsigned NumElts; 847 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 848 NumElts = ST->getNumElements(); 849 else if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 850 NumElts = AT->getNumElements(); 851 else 852 NumElts = Agg->getType()->getVectorNumElements(); 853 854 SmallVector<Constant*, 32> Result; 855 for (unsigned i = 0; i != NumElts; ++i) { 856 Constant *C = Agg->getAggregateElement(i); 857 if (C == 0) return 0; 858 859 if (Idxs[0] == i) 860 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 861 862 Result.push_back(C); 863 } 864 865 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 866 return ConstantStruct::get(ST, Result); 867 if (ArrayType *AT = dyn_cast<ArrayType>(Agg->getType())) 868 return ConstantArray::get(AT, Result); 869 return ConstantVector::get(Result); 870} 871 872 873Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 874 Constant *C1, Constant *C2) { 875 // Handle UndefValue up front. 876 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 877 switch (Opcode) { 878 case Instruction::Xor: 879 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 880 // Handle undef ^ undef -> 0 special case. This is a common 881 // idiom (misuse). 882 return Constant::getNullValue(C1->getType()); 883 // Fallthrough 884 case Instruction::Add: 885 case Instruction::Sub: 886 return UndefValue::get(C1->getType()); 887 case Instruction::And: 888 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 889 return C1; 890 return Constant::getNullValue(C1->getType()); // undef & X -> 0 891 case Instruction::Mul: { 892 ConstantInt *CI; 893 // X * undef -> undef if X is odd or undef 894 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) || 895 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) || 896 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 897 return UndefValue::get(C1->getType()); 898 899 // X * undef -> 0 otherwise 900 return Constant::getNullValue(C1->getType()); 901 } 902 case Instruction::UDiv: 903 case Instruction::SDiv: 904 // undef / 1 -> undef 905 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) 906 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) 907 if (CI2->isOne()) 908 return C1; 909 // FALL THROUGH 910 case Instruction::URem: 911 case Instruction::SRem: 912 if (!isa<UndefValue>(C2)) // undef / X -> 0 913 return Constant::getNullValue(C1->getType()); 914 return C2; // X / undef -> undef 915 case Instruction::Or: // X | undef -> -1 916 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 917 return C1; 918 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 919 case Instruction::LShr: 920 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 921 return C1; // undef lshr undef -> undef 922 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 923 // undef lshr X -> 0 924 case Instruction::AShr: 925 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones 926 return Constant::getAllOnesValue(C1->getType()); 927 else if (isa<UndefValue>(C1)) 928 return C1; // undef ashr undef -> undef 929 else 930 return C1; // X ashr undef --> X 931 case Instruction::Shl: 932 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 933 return C1; // undef shl undef -> undef 934 // undef << X -> 0 or X << undef -> 0 935 return Constant::getNullValue(C1->getType()); 936 } 937 } 938 939 // Handle simplifications when the RHS is a constant int. 940 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 941 switch (Opcode) { 942 case Instruction::Add: 943 if (CI2->equalsInt(0)) return C1; // X + 0 == X 944 break; 945 case Instruction::Sub: 946 if (CI2->equalsInt(0)) return C1; // X - 0 == X 947 break; 948 case Instruction::Mul: 949 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 950 if (CI2->equalsInt(1)) 951 return C1; // X * 1 == X 952 break; 953 case Instruction::UDiv: 954 case Instruction::SDiv: 955 if (CI2->equalsInt(1)) 956 return C1; // X / 1 == X 957 if (CI2->equalsInt(0)) 958 return UndefValue::get(CI2->getType()); // X / 0 == undef 959 break; 960 case Instruction::URem: 961 case Instruction::SRem: 962 if (CI2->equalsInt(1)) 963 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 964 if (CI2->equalsInt(0)) 965 return UndefValue::get(CI2->getType()); // X % 0 == undef 966 break; 967 case Instruction::And: 968 if (CI2->isZero()) return C2; // X & 0 == 0 969 if (CI2->isAllOnesValue()) 970 return C1; // X & -1 == X 971 972 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 973 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 974 if (CE1->getOpcode() == Instruction::ZExt) { 975 unsigned DstWidth = CI2->getType()->getBitWidth(); 976 unsigned SrcWidth = 977 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 978 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 979 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 980 return C1; 981 } 982 983 // If and'ing the address of a global with a constant, fold it. 984 if (CE1->getOpcode() == Instruction::PtrToInt && 985 isa<GlobalValue>(CE1->getOperand(0))) { 986 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 987 988 // Functions are at least 4-byte aligned. 989 unsigned GVAlign = GV->getAlignment(); 990 if (isa<Function>(GV)) 991 GVAlign = std::max(GVAlign, 4U); 992 993 if (GVAlign > 1) { 994 unsigned DstWidth = CI2->getType()->getBitWidth(); 995 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 996 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 997 998 // If checking bits we know are clear, return zero. 999 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1000 return Constant::getNullValue(CI2->getType()); 1001 } 1002 } 1003 } 1004 break; 1005 case Instruction::Or: 1006 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1007 if (CI2->isAllOnesValue()) 1008 return C2; // X | -1 == -1 1009 break; 1010 case Instruction::Xor: 1011 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1012 1013 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1014 switch (CE1->getOpcode()) { 1015 default: break; 1016 case Instruction::ICmp: 1017 case Instruction::FCmp: 1018 // cmp pred ^ true -> cmp !pred 1019 assert(CI2->equalsInt(1)); 1020 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1021 pred = CmpInst::getInversePredicate(pred); 1022 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1023 CE1->getOperand(1)); 1024 } 1025 } 1026 break; 1027 case Instruction::AShr: 1028 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1029 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1030 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1031 return ConstantExpr::getLShr(C1, C2); 1032 break; 1033 } 1034 } else if (isa<ConstantInt>(C1)) { 1035 // If C1 is a ConstantInt and C2 is not, swap the operands. 1036 if (Instruction::isCommutative(Opcode)) 1037 return ConstantExpr::get(Opcode, C2, C1); 1038 } 1039 1040 // At this point we know neither constant is an UndefValue. 1041 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1042 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1043 const APInt &C1V = CI1->getValue(); 1044 const APInt &C2V = CI2->getValue(); 1045 switch (Opcode) { 1046 default: 1047 break; 1048 case Instruction::Add: 1049 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1050 case Instruction::Sub: 1051 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1052 case Instruction::Mul: 1053 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1054 case Instruction::UDiv: 1055 assert(!CI2->isNullValue() && "Div by zero handled above"); 1056 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1057 case Instruction::SDiv: 1058 assert(!CI2->isNullValue() && "Div by zero handled above"); 1059 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1060 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1061 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1062 case Instruction::URem: 1063 assert(!CI2->isNullValue() && "Div by zero handled above"); 1064 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1065 case Instruction::SRem: 1066 assert(!CI2->isNullValue() && "Div by zero handled above"); 1067 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1068 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1069 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1070 case Instruction::And: 1071 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1072 case Instruction::Or: 1073 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1074 case Instruction::Xor: 1075 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1076 case Instruction::Shl: { 1077 uint32_t shiftAmt = C2V.getZExtValue(); 1078 if (shiftAmt < C1V.getBitWidth()) 1079 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1080 else 1081 return UndefValue::get(C1->getType()); // too big shift is undef 1082 } 1083 case Instruction::LShr: { 1084 uint32_t shiftAmt = C2V.getZExtValue(); 1085 if (shiftAmt < C1V.getBitWidth()) 1086 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1087 else 1088 return UndefValue::get(C1->getType()); // too big shift is undef 1089 } 1090 case Instruction::AShr: { 1091 uint32_t shiftAmt = C2V.getZExtValue(); 1092 if (shiftAmt < C1V.getBitWidth()) 1093 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1094 else 1095 return UndefValue::get(C1->getType()); // too big shift is undef 1096 } 1097 } 1098 } 1099 1100 switch (Opcode) { 1101 case Instruction::SDiv: 1102 case Instruction::UDiv: 1103 case Instruction::URem: 1104 case Instruction::SRem: 1105 case Instruction::LShr: 1106 case Instruction::AShr: 1107 case Instruction::Shl: 1108 if (CI1->equalsInt(0)) return C1; 1109 break; 1110 default: 1111 break; 1112 } 1113 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1114 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1115 APFloat C1V = CFP1->getValueAPF(); 1116 APFloat C2V = CFP2->getValueAPF(); 1117 APFloat C3V = C1V; // copy for modification 1118 switch (Opcode) { 1119 default: 1120 break; 1121 case Instruction::FAdd: 1122 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1123 return ConstantFP::get(C1->getContext(), C3V); 1124 case Instruction::FSub: 1125 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1126 return ConstantFP::get(C1->getContext(), C3V); 1127 case Instruction::FMul: 1128 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1129 return ConstantFP::get(C1->getContext(), C3V); 1130 case Instruction::FDiv: 1131 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1132 return ConstantFP::get(C1->getContext(), C3V); 1133 case Instruction::FRem: 1134 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1135 return ConstantFP::get(C1->getContext(), C3V); 1136 } 1137 } 1138 } else if (VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1139 // Perform elementwise folding. 1140 SmallVector<Constant*, 16> Result; 1141 Type *Ty = IntegerType::get(VTy->getContext(), 32); 1142 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1143 Constant *LHS = 1144 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1145 Constant *RHS = 1146 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1147 1148 Result.push_back(ConstantExpr::get(Opcode, LHS, RHS)); 1149 } 1150 1151 return ConstantVector::get(Result); 1152 } 1153 1154 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1155 // There are many possible foldings we could do here. We should probably 1156 // at least fold add of a pointer with an integer into the appropriate 1157 // getelementptr. This will improve alias analysis a bit. 1158 1159 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1160 // (a + (b + c)). 1161 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1162 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1163 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1164 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1165 } 1166 } else if (isa<ConstantExpr>(C2)) { 1167 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1168 // other way if possible. 1169 if (Instruction::isCommutative(Opcode)) 1170 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1171 } 1172 1173 // i1 can be simplified in many cases. 1174 if (C1->getType()->isIntegerTy(1)) { 1175 switch (Opcode) { 1176 case Instruction::Add: 1177 case Instruction::Sub: 1178 return ConstantExpr::getXor(C1, C2); 1179 case Instruction::Mul: 1180 return ConstantExpr::getAnd(C1, C2); 1181 case Instruction::Shl: 1182 case Instruction::LShr: 1183 case Instruction::AShr: 1184 // We can assume that C2 == 0. If it were one the result would be 1185 // undefined because the shift value is as large as the bitwidth. 1186 return C1; 1187 case Instruction::SDiv: 1188 case Instruction::UDiv: 1189 // We can assume that C2 == 1. If it were zero the result would be 1190 // undefined through division by zero. 1191 return C1; 1192 case Instruction::URem: 1193 case Instruction::SRem: 1194 // We can assume that C2 == 1. If it were zero the result would be 1195 // undefined through division by zero. 1196 return ConstantInt::getFalse(C1->getContext()); 1197 default: 1198 break; 1199 } 1200 } 1201 1202 // We don't know how to fold this. 1203 return 0; 1204} 1205 1206/// isZeroSizedType - This type is zero sized if its an array or structure of 1207/// zero sized types. The only leaf zero sized type is an empty structure. 1208static bool isMaybeZeroSizedType(Type *Ty) { 1209 if (StructType *STy = dyn_cast<StructType>(Ty)) { 1210 if (STy->isOpaque()) return true; // Can't say. 1211 1212 // If all of elements have zero size, this does too. 1213 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1214 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1215 return true; 1216 1217 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1218 return isMaybeZeroSizedType(ATy->getElementType()); 1219 } 1220 return false; 1221} 1222 1223/// IdxCompare - Compare the two constants as though they were getelementptr 1224/// indices. This allows coersion of the types to be the same thing. 1225/// 1226/// If the two constants are the "same" (after coersion), return 0. If the 1227/// first is less than the second, return -1, if the second is less than the 1228/// first, return 1. If the constants are not integral, return -2. 1229/// 1230static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) { 1231 if (C1 == C2) return 0; 1232 1233 // Ok, we found a different index. If they are not ConstantInt, we can't do 1234 // anything with them. 1235 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1236 return -2; // don't know! 1237 1238 // Ok, we have two differing integer indices. Sign extend them to be the same 1239 // type. Long is always big enough, so we use it. 1240 if (!C1->getType()->isIntegerTy(64)) 1241 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1242 1243 if (!C2->getType()->isIntegerTy(64)) 1244 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1245 1246 if (C1 == C2) return 0; // They are equal 1247 1248 // If the type being indexed over is really just a zero sized type, there is 1249 // no pointer difference being made here. 1250 if (isMaybeZeroSizedType(ElTy)) 1251 return -2; // dunno. 1252 1253 // If they are really different, now that they are the same type, then we 1254 // found a difference! 1255 if (cast<ConstantInt>(C1)->getSExtValue() < 1256 cast<ConstantInt>(C2)->getSExtValue()) 1257 return -1; 1258 else 1259 return 1; 1260} 1261 1262/// evaluateFCmpRelation - This function determines if there is anything we can 1263/// decide about the two constants provided. This doesn't need to handle simple 1264/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1265/// If we can determine that the two constants have a particular relation to 1266/// each other, we should return the corresponding FCmpInst predicate, 1267/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1268/// ConstantFoldCompareInstruction. 1269/// 1270/// To simplify this code we canonicalize the relation so that the first 1271/// operand is always the most "complex" of the two. We consider ConstantFP 1272/// to be the simplest, and ConstantExprs to be the most complex. 1273static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1274 assert(V1->getType() == V2->getType() && 1275 "Cannot compare values of different types!"); 1276 1277 // Handle degenerate case quickly 1278 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1279 1280 if (!isa<ConstantExpr>(V1)) { 1281 if (!isa<ConstantExpr>(V2)) { 1282 // We distilled thisUse the standard constant folder for a few cases 1283 ConstantInt *R = 0; 1284 R = dyn_cast<ConstantInt>( 1285 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1286 if (R && !R->isZero()) 1287 return FCmpInst::FCMP_OEQ; 1288 R = dyn_cast<ConstantInt>( 1289 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1290 if (R && !R->isZero()) 1291 return FCmpInst::FCMP_OLT; 1292 R = dyn_cast<ConstantInt>( 1293 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1294 if (R && !R->isZero()) 1295 return FCmpInst::FCMP_OGT; 1296 1297 // Nothing more we can do 1298 return FCmpInst::BAD_FCMP_PREDICATE; 1299 } 1300 1301 // If the first operand is simple and second is ConstantExpr, swap operands. 1302 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1303 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1304 return FCmpInst::getSwappedPredicate(SwappedRelation); 1305 } else { 1306 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1307 // constantexpr or a simple constant. 1308 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1309 switch (CE1->getOpcode()) { 1310 case Instruction::FPTrunc: 1311 case Instruction::FPExt: 1312 case Instruction::UIToFP: 1313 case Instruction::SIToFP: 1314 // We might be able to do something with these but we don't right now. 1315 break; 1316 default: 1317 break; 1318 } 1319 } 1320 // There are MANY other foldings that we could perform here. They will 1321 // probably be added on demand, as they seem needed. 1322 return FCmpInst::BAD_FCMP_PREDICATE; 1323} 1324 1325/// evaluateICmpRelation - This function determines if there is anything we can 1326/// decide about the two constants provided. This doesn't need to handle simple 1327/// things like integer comparisons, but should instead handle ConstantExprs 1328/// and GlobalValues. If we can determine that the two constants have a 1329/// particular relation to each other, we should return the corresponding ICmp 1330/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1331/// 1332/// To simplify this code we canonicalize the relation so that the first 1333/// operand is always the most "complex" of the two. We consider simple 1334/// constants (like ConstantInt) to be the simplest, followed by 1335/// GlobalValues, followed by ConstantExpr's (the most complex). 1336/// 1337static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1338 bool isSigned) { 1339 assert(V1->getType() == V2->getType() && 1340 "Cannot compare different types of values!"); 1341 if (V1 == V2) return ICmpInst::ICMP_EQ; 1342 1343 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1344 !isa<BlockAddress>(V1)) { 1345 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1346 !isa<BlockAddress>(V2)) { 1347 // We distilled this down to a simple case, use the standard constant 1348 // folder. 1349 ConstantInt *R = 0; 1350 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1351 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1352 if (R && !R->isZero()) 1353 return pred; 1354 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1355 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1356 if (R && !R->isZero()) 1357 return pred; 1358 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1359 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1360 if (R && !R->isZero()) 1361 return pred; 1362 1363 // If we couldn't figure it out, bail. 1364 return ICmpInst::BAD_ICMP_PREDICATE; 1365 } 1366 1367 // If the first operand is simple, swap operands. 1368 ICmpInst::Predicate SwappedRelation = 1369 evaluateICmpRelation(V2, V1, isSigned); 1370 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1371 return ICmpInst::getSwappedPredicate(SwappedRelation); 1372 1373 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1374 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1375 ICmpInst::Predicate SwappedRelation = 1376 evaluateICmpRelation(V2, V1, isSigned); 1377 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1378 return ICmpInst::getSwappedPredicate(SwappedRelation); 1379 return ICmpInst::BAD_ICMP_PREDICATE; 1380 } 1381 1382 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1383 // constant (which, since the types must match, means that it's a 1384 // ConstantPointerNull). 1385 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1386 // Don't try to decide equality of aliases. 1387 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2)) 1388 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 1389 return ICmpInst::ICMP_NE; 1390 } else if (isa<BlockAddress>(V2)) { 1391 return ICmpInst::ICMP_NE; // Globals never equal labels. 1392 } else { 1393 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1394 // GlobalVals can never be null unless they have external weak linkage. 1395 // We don't try to evaluate aliases here. 1396 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1397 return ICmpInst::ICMP_NE; 1398 } 1399 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1400 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1401 ICmpInst::Predicate SwappedRelation = 1402 evaluateICmpRelation(V2, V1, isSigned); 1403 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1404 return ICmpInst::getSwappedPredicate(SwappedRelation); 1405 return ICmpInst::BAD_ICMP_PREDICATE; 1406 } 1407 1408 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1409 // constant (which, since the types must match, means that it is a 1410 // ConstantPointerNull). 1411 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1412 // Block address in another function can't equal this one, but block 1413 // addresses in the current function might be the same if blocks are 1414 // empty. 1415 if (BA2->getFunction() != BA->getFunction()) 1416 return ICmpInst::ICMP_NE; 1417 } else { 1418 // Block addresses aren't null, don't equal the address of globals. 1419 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1420 "Canonicalization guarantee!"); 1421 return ICmpInst::ICMP_NE; 1422 } 1423 } else { 1424 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1425 // constantexpr, a global, block address, or a simple constant. 1426 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1427 Constant *CE1Op0 = CE1->getOperand(0); 1428 1429 switch (CE1->getOpcode()) { 1430 case Instruction::Trunc: 1431 case Instruction::FPTrunc: 1432 case Instruction::FPExt: 1433 case Instruction::FPToUI: 1434 case Instruction::FPToSI: 1435 break; // We can't evaluate floating point casts or truncations. 1436 1437 case Instruction::UIToFP: 1438 case Instruction::SIToFP: 1439 case Instruction::BitCast: 1440 case Instruction::ZExt: 1441 case Instruction::SExt: 1442 // If the cast is not actually changing bits, and the second operand is a 1443 // null pointer, do the comparison with the pre-casted value. 1444 if (V2->isNullValue() && 1445 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1446 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1447 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1448 return evaluateICmpRelation(CE1Op0, 1449 Constant::getNullValue(CE1Op0->getType()), 1450 isSigned); 1451 } 1452 break; 1453 1454 case Instruction::GetElementPtr: 1455 // Ok, since this is a getelementptr, we know that the constant has a 1456 // pointer type. Check the various cases. 1457 if (isa<ConstantPointerNull>(V2)) { 1458 // If we are comparing a GEP to a null pointer, check to see if the base 1459 // of the GEP equals the null pointer. 1460 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1461 if (GV->hasExternalWeakLinkage()) 1462 // Weak linkage GVals could be zero or not. We're comparing that 1463 // to null pointer so its greater-or-equal 1464 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1465 else 1466 // If its not weak linkage, the GVal must have a non-zero address 1467 // so the result is greater-than 1468 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1469 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1470 // If we are indexing from a null pointer, check to see if we have any 1471 // non-zero indices. 1472 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1473 if (!CE1->getOperand(i)->isNullValue()) 1474 // Offsetting from null, must not be equal. 1475 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1476 // Only zero indexes from null, must still be zero. 1477 return ICmpInst::ICMP_EQ; 1478 } 1479 // Otherwise, we can't really say if the first operand is null or not. 1480 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1481 if (isa<ConstantPointerNull>(CE1Op0)) { 1482 if (GV2->hasExternalWeakLinkage()) 1483 // Weak linkage GVals could be zero or not. We're comparing it to 1484 // a null pointer, so its less-or-equal 1485 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1486 else 1487 // If its not weak linkage, the GVal must have a non-zero address 1488 // so the result is less-than 1489 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1490 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1491 if (GV == GV2) { 1492 // If this is a getelementptr of the same global, then it must be 1493 // different. Because the types must match, the getelementptr could 1494 // only have at most one index, and because we fold getelementptr's 1495 // with a single zero index, it must be nonzero. 1496 assert(CE1->getNumOperands() == 2 && 1497 !CE1->getOperand(1)->isNullValue() && 1498 "Surprising getelementptr!"); 1499 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1500 } else { 1501 // If they are different globals, we don't know what the value is. 1502 return ICmpInst::BAD_ICMP_PREDICATE; 1503 } 1504 } 1505 } else { 1506 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1507 Constant *CE2Op0 = CE2->getOperand(0); 1508 1509 // There are MANY other foldings that we could perform here. They will 1510 // probably be added on demand, as they seem needed. 1511 switch (CE2->getOpcode()) { 1512 default: break; 1513 case Instruction::GetElementPtr: 1514 // By far the most common case to handle is when the base pointers are 1515 // obviously to the same global. 1516 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1517 if (CE1Op0 != CE2Op0) // Don't know relative ordering. 1518 return ICmpInst::BAD_ICMP_PREDICATE; 1519 // Ok, we know that both getelementptr instructions are based on the 1520 // same global. From this, we can precisely determine the relative 1521 // ordering of the resultant pointers. 1522 unsigned i = 1; 1523 1524 // The logic below assumes that the result of the comparison 1525 // can be determined by finding the first index that differs. 1526 // This doesn't work if there is over-indexing in any 1527 // subsequent indices, so check for that case first. 1528 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1529 !CE2->isGEPWithNoNotionalOverIndexing()) 1530 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1531 1532 // Compare all of the operands the GEP's have in common. 1533 gep_type_iterator GTI = gep_type_begin(CE1); 1534 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1535 ++i, ++GTI) 1536 switch (IdxCompare(CE1->getOperand(i), 1537 CE2->getOperand(i), GTI.getIndexedType())) { 1538 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1539 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1540 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1541 } 1542 1543 // Ok, we ran out of things they have in common. If any leftovers 1544 // are non-zero then we have a difference, otherwise we are equal. 1545 for (; i < CE1->getNumOperands(); ++i) 1546 if (!CE1->getOperand(i)->isNullValue()) { 1547 if (isa<ConstantInt>(CE1->getOperand(i))) 1548 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1549 else 1550 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1551 } 1552 1553 for (; i < CE2->getNumOperands(); ++i) 1554 if (!CE2->getOperand(i)->isNullValue()) { 1555 if (isa<ConstantInt>(CE2->getOperand(i))) 1556 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1557 else 1558 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1559 } 1560 return ICmpInst::ICMP_EQ; 1561 } 1562 } 1563 } 1564 default: 1565 break; 1566 } 1567 } 1568 1569 return ICmpInst::BAD_ICMP_PREDICATE; 1570} 1571 1572Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1573 Constant *C1, Constant *C2) { 1574 Type *ResultTy; 1575 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1576 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1577 VT->getNumElements()); 1578 else 1579 ResultTy = Type::getInt1Ty(C1->getContext()); 1580 1581 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1582 if (pred == FCmpInst::FCMP_FALSE) 1583 return Constant::getNullValue(ResultTy); 1584 1585 if (pred == FCmpInst::FCMP_TRUE) 1586 return Constant::getAllOnesValue(ResultTy); 1587 1588 // Handle some degenerate cases first 1589 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1590 // For EQ and NE, we can always pick a value for the undef to make the 1591 // predicate pass or fail, so we can return undef. 1592 // Also, if both operands are undef, we can return undef. 1593 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) || 1594 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 1595 return UndefValue::get(ResultTy); 1596 // Otherwise, pick the same value as the non-undef operand, and fold 1597 // it to true or false. 1598 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1599 } 1600 1601 // icmp eq/ne(null,GV) -> false/true 1602 if (C1->isNullValue()) { 1603 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1604 // Don't try to evaluate aliases. External weak GV can be null. 1605 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1606 if (pred == ICmpInst::ICMP_EQ) 1607 return ConstantInt::getFalse(C1->getContext()); 1608 else if (pred == ICmpInst::ICMP_NE) 1609 return ConstantInt::getTrue(C1->getContext()); 1610 } 1611 // icmp eq/ne(GV,null) -> false/true 1612 } else if (C2->isNullValue()) { 1613 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1614 // Don't try to evaluate aliases. External weak GV can be null. 1615 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1616 if (pred == ICmpInst::ICMP_EQ) 1617 return ConstantInt::getFalse(C1->getContext()); 1618 else if (pred == ICmpInst::ICMP_NE) 1619 return ConstantInt::getTrue(C1->getContext()); 1620 } 1621 } 1622 1623 // If the comparison is a comparison between two i1's, simplify it. 1624 if (C1->getType()->isIntegerTy(1)) { 1625 switch(pred) { 1626 case ICmpInst::ICMP_EQ: 1627 if (isa<ConstantInt>(C2)) 1628 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1629 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1630 case ICmpInst::ICMP_NE: 1631 return ConstantExpr::getXor(C1, C2); 1632 default: 1633 break; 1634 } 1635 } 1636 1637 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1638 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1639 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1640 switch (pred) { 1641 default: llvm_unreachable("Invalid ICmp Predicate"); 1642 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1643 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1644 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1645 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1646 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1647 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1648 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1649 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1650 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1651 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1652 } 1653 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1654 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1655 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1656 APFloat::cmpResult R = C1V.compare(C2V); 1657 switch (pred) { 1658 default: llvm_unreachable("Invalid FCmp Predicate"); 1659 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1660 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1661 case FCmpInst::FCMP_UNO: 1662 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1663 case FCmpInst::FCMP_ORD: 1664 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1665 case FCmpInst::FCMP_UEQ: 1666 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1667 R==APFloat::cmpEqual); 1668 case FCmpInst::FCMP_OEQ: 1669 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1670 case FCmpInst::FCMP_UNE: 1671 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1672 case FCmpInst::FCMP_ONE: 1673 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1674 R==APFloat::cmpGreaterThan); 1675 case FCmpInst::FCMP_ULT: 1676 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1677 R==APFloat::cmpLessThan); 1678 case FCmpInst::FCMP_OLT: 1679 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1680 case FCmpInst::FCMP_UGT: 1681 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1682 R==APFloat::cmpGreaterThan); 1683 case FCmpInst::FCMP_OGT: 1684 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1685 case FCmpInst::FCMP_ULE: 1686 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1687 case FCmpInst::FCMP_OLE: 1688 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1689 R==APFloat::cmpEqual); 1690 case FCmpInst::FCMP_UGE: 1691 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1692 case FCmpInst::FCMP_OGE: 1693 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1694 R==APFloat::cmpEqual); 1695 } 1696 } else if (C1->getType()->isVectorTy()) { 1697 // If we can constant fold the comparison of each element, constant fold 1698 // the whole vector comparison. 1699 SmallVector<Constant*, 4> ResElts; 1700 Type *Ty = IntegerType::get(C1->getContext(), 32); 1701 // Compare the elements, producing an i1 result or constant expr. 1702 for (unsigned i = 0, e = C1->getType()->getVectorNumElements(); i != e;++i){ 1703 Constant *C1E = 1704 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, i)); 1705 Constant *C2E = 1706 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, i)); 1707 1708 ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E)); 1709 } 1710 1711 return ConstantVector::get(ResElts); 1712 } 1713 1714 if (C1->getType()->isFloatingPointTy()) { 1715 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1716 switch (evaluateFCmpRelation(C1, C2)) { 1717 default: llvm_unreachable("Unknown relation!"); 1718 case FCmpInst::FCMP_UNO: 1719 case FCmpInst::FCMP_ORD: 1720 case FCmpInst::FCMP_UEQ: 1721 case FCmpInst::FCMP_UNE: 1722 case FCmpInst::FCMP_ULT: 1723 case FCmpInst::FCMP_UGT: 1724 case FCmpInst::FCMP_ULE: 1725 case FCmpInst::FCMP_UGE: 1726 case FCmpInst::FCMP_TRUE: 1727 case FCmpInst::FCMP_FALSE: 1728 case FCmpInst::BAD_FCMP_PREDICATE: 1729 break; // Couldn't determine anything about these constants. 1730 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1731 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1732 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1733 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1734 break; 1735 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1736 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1737 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1738 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1739 break; 1740 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1741 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1742 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1743 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1744 break; 1745 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1746 // We can only partially decide this relation. 1747 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1748 Result = 0; 1749 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1750 Result = 1; 1751 break; 1752 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1753 // We can only partially decide this relation. 1754 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1755 Result = 0; 1756 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1757 Result = 1; 1758 break; 1759 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1760 // We can only partially decide this relation. 1761 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 1762 Result = 0; 1763 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 1764 Result = 1; 1765 break; 1766 } 1767 1768 // If we evaluated the result, return it now. 1769 if (Result != -1) 1770 return ConstantInt::get(ResultTy, Result); 1771 1772 } else { 1773 // Evaluate the relation between the two constants, per the predicate. 1774 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1775 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 1776 default: llvm_unreachable("Unknown relational!"); 1777 case ICmpInst::BAD_ICMP_PREDICATE: 1778 break; // Couldn't determine anything about these constants. 1779 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1780 // If we know the constants are equal, we can decide the result of this 1781 // computation precisely. 1782 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 1783 break; 1784 case ICmpInst::ICMP_ULT: 1785 switch (pred) { 1786 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1787 Result = 1; break; 1788 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1789 Result = 0; break; 1790 } 1791 break; 1792 case ICmpInst::ICMP_SLT: 1793 switch (pred) { 1794 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1795 Result = 1; break; 1796 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1797 Result = 0; break; 1798 } 1799 break; 1800 case ICmpInst::ICMP_UGT: 1801 switch (pred) { 1802 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1803 Result = 1; break; 1804 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1805 Result = 0; break; 1806 } 1807 break; 1808 case ICmpInst::ICMP_SGT: 1809 switch (pred) { 1810 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1811 Result = 1; break; 1812 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1813 Result = 0; break; 1814 } 1815 break; 1816 case ICmpInst::ICMP_ULE: 1817 if (pred == ICmpInst::ICMP_UGT) Result = 0; 1818 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 1819 break; 1820 case ICmpInst::ICMP_SLE: 1821 if (pred == ICmpInst::ICMP_SGT) Result = 0; 1822 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 1823 break; 1824 case ICmpInst::ICMP_UGE: 1825 if (pred == ICmpInst::ICMP_ULT) Result = 0; 1826 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 1827 break; 1828 case ICmpInst::ICMP_SGE: 1829 if (pred == ICmpInst::ICMP_SLT) Result = 0; 1830 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 1831 break; 1832 case ICmpInst::ICMP_NE: 1833 if (pred == ICmpInst::ICMP_EQ) Result = 0; 1834 if (pred == ICmpInst::ICMP_NE) Result = 1; 1835 break; 1836 } 1837 1838 // If we evaluated the result, return it now. 1839 if (Result != -1) 1840 return ConstantInt::get(ResultTy, Result); 1841 1842 // If the right hand side is a bitcast, try using its inverse to simplify 1843 // it by moving it to the left hand side. We can't do this if it would turn 1844 // a vector compare into a scalar compare or visa versa. 1845 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1846 Constant *CE2Op0 = CE2->getOperand(0); 1847 if (CE2->getOpcode() == Instruction::BitCast && 1848 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 1849 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1850 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 1851 } 1852 } 1853 1854 // If the left hand side is an extension, try eliminating it. 1855 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1856 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 1857 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 1858 Constant *CE1Op0 = CE1->getOperand(0); 1859 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1860 if (CE1Inverse == CE1Op0) { 1861 // Check whether we can safely truncate the right hand side. 1862 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1863 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1864 C2->getType()) == C2) 1865 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 1866 } 1867 } 1868 } 1869 1870 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1871 (C1->isNullValue() && !C2->isNullValue())) { 1872 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1873 // other way if possible. 1874 // Also, if C1 is null and C2 isn't, flip them around. 1875 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 1876 return ConstantExpr::getICmp(pred, C2, C1); 1877 } 1878 } 1879 return 0; 1880} 1881 1882/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 1883/// is "inbounds". 1884template<typename IndexTy> 1885static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1886 // No indices means nothing that could be out of bounds. 1887 if (Idxs.empty()) return true; 1888 1889 // If the first index is zero, it's in bounds. 1890 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 1891 1892 // If the first index is one and all the rest are zero, it's in bounds, 1893 // by the one-past-the-end rule. 1894 if (!cast<ConstantInt>(Idxs[0])->isOne()) 1895 return false; 1896 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 1897 if (!cast<Constant>(Idxs[i])->isNullValue()) 1898 return false; 1899 return true; 1900} 1901 1902/// \brief Test whether a given ConstantInt is in-range for a SequentialType. 1903static bool isIndexInRangeOfSequentialType(const SequentialType *STy, 1904 const ConstantInt *CI) { 1905 if (const PointerType *PTy = dyn_cast<PointerType>(STy)) 1906 // Only handle pointers to sized types, not pointers to functions. 1907 return PTy->getElementType()->isSized(); 1908 1909 uint64_t NumElements = 0; 1910 // Determine the number of elements in our sequential type. 1911 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 1912 NumElements = ATy->getNumElements(); 1913 else if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 1914 NumElements = VTy->getNumElements(); 1915 1916 assert((isa<ArrayType>(STy) || NumElements > 0) && 1917 "didn't expect non-array type to have zero elements!"); 1918 1919 // We cannot bounds check the index if it doesn't fit in an int64_t. 1920 if (CI->getValue().getActiveBits() > 64) 1921 return false; 1922 1923 // A negative index or an index past the end of our sequential type is 1924 // considered out-of-range. 1925 int64_t IndexVal = CI->getSExtValue(); 1926 if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements)) 1927 return false; 1928 1929 // Otherwise, it is in-range. 1930 return true; 1931} 1932 1933template<typename IndexTy> 1934static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 1935 bool inBounds, 1936 ArrayRef<IndexTy> Idxs) { 1937 if (Idxs.empty()) return C; 1938 Constant *Idx0 = cast<Constant>(Idxs[0]); 1939 if ((Idxs.size() == 1 && Idx0->isNullValue())) 1940 return C; 1941 1942 if (isa<UndefValue>(C)) { 1943 PointerType *Ptr = cast<PointerType>(C->getType()); 1944 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 1945 assert(Ty != 0 && "Invalid indices for GEP!"); 1946 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 1947 } 1948 1949 if (C->isNullValue()) { 1950 bool isNull = true; 1951 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 1952 if (!cast<Constant>(Idxs[i])->isNullValue()) { 1953 isNull = false; 1954 break; 1955 } 1956 if (isNull) { 1957 PointerType *Ptr = cast<PointerType>(C->getType()); 1958 Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs); 1959 assert(Ty != 0 && "Invalid indices for GEP!"); 1960 return ConstantPointerNull::get(PointerType::get(Ty, 1961 Ptr->getAddressSpace())); 1962 } 1963 } 1964 1965 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 1966 // Combine Indices - If the source pointer to this getelementptr instruction 1967 // is a getelementptr instruction, combine the indices of the two 1968 // getelementptr instructions into a single instruction. 1969 // 1970 if (CE->getOpcode() == Instruction::GetElementPtr) { 1971 Type *LastTy = 0; 1972 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 1973 I != E; ++I) 1974 LastTy = *I; 1975 1976 // We cannot combine indices if doing so would take us outside of an 1977 // array or vector. Doing otherwise could trick us if we evaluated such a 1978 // GEP as part of a load. 1979 // 1980 // e.g. Consider if the original GEP was: 1981 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 1982 // i32 0, i32 0, i64 0) 1983 // 1984 // If we then tried to offset it by '8' to get to the third element, 1985 // an i8, we should *not* get: 1986 // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c, 1987 // i32 0, i32 0, i64 8) 1988 // 1989 // This GEP tries to index array element '8 which runs out-of-bounds. 1990 // Subsequent evaluation would get confused and produce erroneous results. 1991 // 1992 // The following prohibits such a GEP from being formed by checking to see 1993 // if the index is in-range with respect to an array or vector. 1994 bool PerformFold = false; 1995 if (Idx0->isNullValue()) 1996 PerformFold = true; 1997 else if (SequentialType *STy = dyn_cast_or_null<SequentialType>(LastTy)) 1998 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx0)) 1999 PerformFold = isIndexInRangeOfSequentialType(STy, CI); 2000 2001 if (PerformFold) { 2002 SmallVector<Value*, 16> NewIndices; 2003 NewIndices.reserve(Idxs.size() + CE->getNumOperands()); 2004 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 2005 NewIndices.push_back(CE->getOperand(i)); 2006 2007 // Add the last index of the source with the first index of the new GEP. 2008 // Make sure to handle the case when they are actually different types. 2009 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2010 // Otherwise it must be an array. 2011 if (!Idx0->isNullValue()) { 2012 Type *IdxTy = Combined->getType(); 2013 if (IdxTy != Idx0->getType()) { 2014 Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 2015 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 2016 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 2017 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2018 } else { 2019 Combined = 2020 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2021 } 2022 } 2023 2024 NewIndices.push_back(Combined); 2025 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 2026 return 2027 ConstantExpr::getGetElementPtr(CE->getOperand(0), NewIndices, 2028 inBounds && 2029 cast<GEPOperator>(CE)->isInBounds()); 2030 } 2031 } 2032 2033 // Attempt to fold casts to the same type away. For example, folding: 2034 // 2035 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2036 // i64 0, i64 0) 2037 // into: 2038 // 2039 // i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2040 // 2041 // Don't fold if the cast is changing address spaces. 2042 if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) { 2043 PointerType *SrcPtrTy = 2044 dyn_cast<PointerType>(CE->getOperand(0)->getType()); 2045 PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType()); 2046 if (SrcPtrTy && DstPtrTy) { 2047 ArrayType *SrcArrayTy = 2048 dyn_cast<ArrayType>(SrcPtrTy->getElementType()); 2049 ArrayType *DstArrayTy = 2050 dyn_cast<ArrayType>(DstPtrTy->getElementType()); 2051 if (SrcArrayTy && DstArrayTy 2052 && SrcArrayTy->getElementType() == DstArrayTy->getElementType() 2053 && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace()) 2054 return ConstantExpr::getGetElementPtr((Constant*)CE->getOperand(0), 2055 Idxs, inBounds); 2056 } 2057 } 2058 } 2059 2060 // Check to see if any array indices are not within the corresponding 2061 // notional array or vector bounds. If so, try to determine if they can be 2062 // factored out into preceding dimensions. 2063 bool Unknown = false; 2064 SmallVector<Constant *, 8> NewIdxs; 2065 Type *Ty = C->getType(); 2066 Type *Prev = 0; 2067 for (unsigned i = 0, e = Idxs.size(); i != e; 2068 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2069 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2070 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) 2071 if (CI->getSExtValue() > 0 && 2072 !isIndexInRangeOfSequentialType(cast<SequentialType>(Ty), CI)) { 2073 if (isa<SequentialType>(Prev)) { 2074 // It's out of range, but we can factor it into the prior 2075 // dimension. 2076 NewIdxs.resize(Idxs.size()); 2077 uint64_t NumElements = 0; 2078 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2079 NumElements = ATy->getNumElements(); 2080 else 2081 NumElements = cast<VectorType>(Ty)->getNumElements(); 2082 2083 ConstantInt *Factor = ConstantInt::get(CI->getType(), NumElements); 2084 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2085 2086 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2087 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2088 2089 // Before adding, extend both operands to i64 to avoid 2090 // overflow trouble. 2091 if (!PrevIdx->getType()->isIntegerTy(64)) 2092 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2093 Type::getInt64Ty(Div->getContext())); 2094 if (!Div->getType()->isIntegerTy(64)) 2095 Div = ConstantExpr::getSExt(Div, 2096 Type::getInt64Ty(Div->getContext())); 2097 2098 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2099 } else { 2100 // It's out of range, but the prior dimension is a struct 2101 // so we can't do anything about it. 2102 Unknown = true; 2103 } 2104 } 2105 } else { 2106 // We don't know if it's in range or not. 2107 Unknown = true; 2108 } 2109 } 2110 2111 // If we did any factoring, start over with the adjusted indices. 2112 if (!NewIdxs.empty()) { 2113 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2114 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2115 return ConstantExpr::getGetElementPtr(C, NewIdxs, inBounds); 2116 } 2117 2118 // If all indices are known integers and normalized, we can do a simple 2119 // check for the "inbounds" property. 2120 if (!Unknown && !inBounds && 2121 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs)) 2122 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs); 2123 2124 return 0; 2125} 2126 2127Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2128 bool inBounds, 2129 ArrayRef<Constant *> Idxs) { 2130 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2131} 2132 2133Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2134 bool inBounds, 2135 ArrayRef<Value *> Idxs) { 2136 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs); 2137} 2138