ConstantFold.cpp revision 093399cbf3bcdb31d04b3bf5c5691fc88c25da48
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 TargetData, and the pieces that do. This is to avoid 16// a dependence in VMCore on Target. 17// 18//===----------------------------------------------------------------------===// 19 20#include "ConstantFold.h" 21#include "llvm/Constants.h" 22#include "llvm/Instructions.h" 23#include "llvm/DerivedTypes.h" 24#include "llvm/Function.h" 25#include "llvm/GlobalAlias.h" 26#include "llvm/GlobalVariable.h" 27#include "llvm/ADT/SmallVector.h" 28#include "llvm/Support/Compiler.h" 29#include "llvm/Support/ErrorHandling.h" 30#include "llvm/Support/GetElementPtrTypeIterator.h" 31#include "llvm/Support/ManagedStatic.h" 32#include "llvm/Support/MathExtras.h" 33#include <limits> 34using namespace llvm; 35 36//===----------------------------------------------------------------------===// 37// ConstantFold*Instruction Implementations 38//===----------------------------------------------------------------------===// 39 40/// BitCastConstantVector - Convert the specified ConstantVector node to the 41/// specified vector type. At this point, we know that the elements of the 42/// input vector constant are all simple integer or FP values. 43static Constant *BitCastConstantVector(ConstantVector *CV, 44 const 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->getNumOperands()) 54 return 0; 55 56 // Check to verify that all elements of the input are simple. 57 for (unsigned i = 0; i != NumElts; ++i) { 58 if (!isa<ConstantInt>(CV->getOperand(i)) && 59 !isa<ConstantFP>(CV->getOperand(i))) 60 return 0; 61 } 62 63 // Bitcast each element now. 64 std::vector<Constant*> Result; 65 const Type *DstEltTy = DstTy->getElementType(); 66 for (unsigned i = 0; i != NumElts; ++i) 67 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), 68 DstEltTy)); 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 const Type *DstTy ///< desintation 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 const Type *SrcTy = Op->getOperand(0)->getType(); 88 const Type *MidTy = Op->getType(); 89 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 90 Instruction::CastOps secondOp = Instruction::CastOps(opc); 91 92 // Let CastInst::isEliminableCastPair do the heavy lifting. 93 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 94 Type::getInt64Ty(DstTy->getContext())); 95} 96 97static Constant *FoldBitCast(Constant *V, const Type *DestTy) { 98 const Type *SrcTy = V->getType(); 99 if (SrcTy == DestTy) 100 return V; // no-op cast 101 102 // Check to see if we are casting a pointer to an aggregate to a pointer to 103 // the first element. If so, return the appropriate GEP instruction. 104 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) 105 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 106 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { 107 SmallVector<Value*, 8> IdxList; 108 Value *Zero = 109 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 110 IdxList.push_back(Zero); 111 const Type *ElTy = PTy->getElementType(); 112 while (ElTy != DPTy->getElementType()) { 113 if (const StructType *STy = dyn_cast<StructType>(ElTy)) { 114 if (STy->getNumElements() == 0) break; 115 ElTy = STy->getElementType(0); 116 IdxList.push_back(Zero); 117 } else if (const SequentialType *STy = 118 dyn_cast<SequentialType>(ElTy)) { 119 if (ElTy->isPointerTy()) break; // Can't index into pointers! 120 ElTy = STy->getElementType(); 121 IdxList.push_back(Zero); 122 } else { 123 break; 124 } 125 } 126 127 if (ElTy == DPTy->getElementType()) 128 // This GEP is inbounds because all indices are zero. 129 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], 130 IdxList.size()); 131 } 132 133 // Handle casts from one vector constant to another. We know that the src 134 // and dest type have the same size (otherwise its an illegal cast). 135 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 136 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 137 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 138 "Not cast between same sized vectors!"); 139 SrcTy = NULL; 140 // First, check for null. Undef is already handled. 141 if (isa<ConstantAggregateZero>(V)) 142 return Constant::getNullValue(DestTy); 143 144 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 145 return BitCastConstantVector(CV, DestPTy); 146 } 147 148 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 149 // This allows for other simplifications (although some of them 150 // can only be handled by Analysis/ConstantFolding.cpp). 151 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 152 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 153 } 154 155 // Finally, implement bitcast folding now. The code below doesn't handle 156 // bitcast right. 157 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 158 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 159 160 // Handle integral constant input. 161 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 162 if (DestTy->isIntegerTy()) 163 // Integral -> Integral. This is a no-op because the bit widths must 164 // be the same. Consequently, we just fold to V. 165 return V; 166 167 if (DestTy->isFloatingPointTy()) 168 return ConstantFP::get(DestTy->getContext(), 169 APFloat(CI->getValue(), 170 !DestTy->isPPC_FP128Ty())); 171 172 // Otherwise, can't fold this (vector?) 173 return 0; 174 } 175 176 // Handle ConstantFP input: FP -> Integral. 177 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 178 return ConstantInt::get(FP->getContext(), 179 FP->getValueAPF().bitcastToAPInt()); 180 181 return 0; 182} 183 184 185/// ExtractConstantBytes - V is an integer constant which only has a subset of 186/// its bytes used. The bytes used are indicated by ByteStart (which is the 187/// first byte used, counting from the least significant byte) and ByteSize, 188/// which is the number of bytes used. 189/// 190/// This function analyzes the specified constant to see if the specified byte 191/// range can be returned as a simplified constant. If so, the constant is 192/// returned, otherwise null is returned. 193/// 194static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 195 unsigned ByteSize) { 196 assert(C->getType()->isIntegerTy() && 197 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 198 "Non-byte sized integer input"); 199 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 200 assert(ByteSize && "Must be accessing some piece"); 201 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 202 assert(ByteSize != CSize && "Should not extract everything"); 203 204 // Constant Integers are simple. 205 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 206 APInt V = CI->getValue(); 207 if (ByteStart) 208 V = V.lshr(ByteStart*8); 209 V = V.trunc(ByteSize*8); 210 return ConstantInt::get(CI->getContext(), V); 211 } 212 213 // In the input is a constant expr, we might be able to recursively simplify. 214 // If not, we definitely can't do anything. 215 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 216 if (CE == 0) return 0; 217 218 switch (CE->getOpcode()) { 219 default: return 0; 220 case Instruction::Or: { 221 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 222 if (RHS == 0) 223 return 0; 224 225 // X | -1 -> -1. 226 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 227 if (RHSC->isAllOnesValue()) 228 return RHSC; 229 230 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 231 if (LHS == 0) 232 return 0; 233 return ConstantExpr::getOr(LHS, RHS); 234 } 235 case Instruction::And: { 236 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 237 if (RHS == 0) 238 return 0; 239 240 // X & 0 -> 0. 241 if (RHS->isNullValue()) 242 return RHS; 243 244 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 245 if (LHS == 0) 246 return 0; 247 return ConstantExpr::getAnd(LHS, RHS); 248 } 249 case Instruction::LShr: { 250 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 251 if (Amt == 0) 252 return 0; 253 unsigned ShAmt = Amt->getZExtValue(); 254 // Cannot analyze non-byte shifts. 255 if ((ShAmt & 7) != 0) 256 return 0; 257 ShAmt >>= 3; 258 259 // If the extract is known to be all zeros, return zero. 260 if (ByteStart >= CSize-ShAmt) 261 return Constant::getNullValue(IntegerType::get(CE->getContext(), 262 ByteSize*8)); 263 // If the extract is known to be fully in the input, extract it. 264 if (ByteStart+ByteSize+ShAmt <= CSize) 265 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 266 267 // TODO: Handle the 'partially zero' case. 268 return 0; 269 } 270 271 case Instruction::Shl: { 272 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 273 if (Amt == 0) 274 return 0; 275 unsigned ShAmt = Amt->getZExtValue(); 276 // Cannot analyze non-byte shifts. 277 if ((ShAmt & 7) != 0) 278 return 0; 279 ShAmt >>= 3; 280 281 // If the extract is known to be all zeros, return zero. 282 if (ByteStart+ByteSize <= ShAmt) 283 return Constant::getNullValue(IntegerType::get(CE->getContext(), 284 ByteSize*8)); 285 // If the extract is known to be fully in the input, extract it. 286 if (ByteStart >= ShAmt) 287 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 288 289 // TODO: Handle the 'partially zero' case. 290 return 0; 291 } 292 293 case Instruction::ZExt: { 294 unsigned SrcBitSize = 295 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 296 297 // If extracting something that is completely zero, return 0. 298 if (ByteStart*8 >= SrcBitSize) 299 return Constant::getNullValue(IntegerType::get(CE->getContext(), 300 ByteSize*8)); 301 302 // If exactly extracting the input, return it. 303 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 304 return CE->getOperand(0); 305 306 // If extracting something completely in the input, if if the input is a 307 // multiple of 8 bits, recurse. 308 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 309 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 310 311 // Otherwise, if extracting a subset of the input, which is not multiple of 312 // 8 bits, do a shift and trunc to get the bits. 313 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 314 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 315 Constant *Res = CE->getOperand(0); 316 if (ByteStart) 317 Res = ConstantExpr::getLShr(Res, 318 ConstantInt::get(Res->getType(), ByteStart*8)); 319 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 320 ByteSize*8)); 321 } 322 323 // TODO: Handle the 'partially zero' case. 324 return 0; 325 } 326 } 327} 328 329/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 330/// on Ty, with any known factors factored out. If Folded is false, 331/// return null if no factoring was possible, to avoid endlessly 332/// bouncing an unfoldable expression back into the top-level folder. 333/// 334static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy, 335 bool Folded) { 336 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 337 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 338 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 339 return ConstantExpr::getNUWMul(E, N); 340 } 341 342 if (const StructType *STy = dyn_cast<StructType>(Ty)) 343 if (!STy->isPacked()) { 344 unsigned NumElems = STy->getNumElements(); 345 // An empty struct has size zero. 346 if (NumElems == 0) 347 return ConstantExpr::getNullValue(DestTy); 348 // Check for a struct with all members having the same size. 349 Constant *MemberSize = 350 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 351 bool AllSame = true; 352 for (unsigned i = 1; i != NumElems; ++i) 353 if (MemberSize != 354 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 355 AllSame = false; 356 break; 357 } 358 if (AllSame) { 359 Constant *N = ConstantInt::get(DestTy, NumElems); 360 return ConstantExpr::getNUWMul(MemberSize, N); 361 } 362 } 363 364 // Pointer size doesn't depend on the pointee type, so canonicalize them 365 // to an arbitrary pointee. 366 if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) 367 if (!PTy->getElementType()->isIntegerTy(1)) 368 return 369 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 370 PTy->getAddressSpace()), 371 DestTy, true); 372 373 // If there's no interesting folding happening, bail so that we don't create 374 // a constant that looks like it needs folding but really doesn't. 375 if (!Folded) 376 return 0; 377 378 // Base case: Get a regular sizeof expression. 379 Constant *C = ConstantExpr::getSizeOf(Ty); 380 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 381 DestTy, false), 382 C, DestTy); 383 return C; 384} 385 386/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 387/// on Ty, with any known factors factored out. If Folded is false, 388/// return null if no factoring was possible, to avoid endlessly 389/// bouncing an unfoldable expression back into the top-level folder. 390/// 391static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, 392 bool Folded) { 393 // The alignment of an array is equal to the alignment of the 394 // array element. Note that this is not always true for vectors. 395 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 396 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 397 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 398 DestTy, 399 false), 400 C, DestTy); 401 return C; 402 } 403 404 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 405 // Packed structs always have an alignment of 1. 406 if (STy->isPacked()) 407 return ConstantInt::get(DestTy, 1); 408 409 // Otherwise, struct alignment is the maximum alignment of any member. 410 // Without target data, we can't compare much, but we can check to see 411 // if all the members have the same alignment. 412 unsigned NumElems = STy->getNumElements(); 413 // An empty struct has minimal alignment. 414 if (NumElems == 0) 415 return ConstantInt::get(DestTy, 1); 416 // Check for a struct with all members having the same alignment. 417 Constant *MemberAlign = 418 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 419 bool AllSame = true; 420 for (unsigned i = 1; i != NumElems; ++i) 421 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 422 AllSame = false; 423 break; 424 } 425 if (AllSame) 426 return MemberAlign; 427 } 428 429 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 430 // to an arbitrary pointee. 431 if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) 432 if (!PTy->getElementType()->isIntegerTy(1)) 433 return 434 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 435 1), 436 PTy->getAddressSpace()), 437 DestTy, true); 438 439 // If there's no interesting folding happening, bail so that we don't create 440 // a constant that looks like it needs folding but really doesn't. 441 if (!Folded) 442 return 0; 443 444 // Base case: Get a regular alignof expression. 445 Constant *C = ConstantExpr::getAlignOf(Ty); 446 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 447 DestTy, false), 448 C, DestTy); 449 return C; 450} 451 452/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 453/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 454/// return null if no factoring was possible, to avoid endlessly 455/// bouncing an unfoldable expression back into the top-level folder. 456/// 457static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo, 458 const Type *DestTy, 459 bool Folded) { 460 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 461 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 462 DestTy, false), 463 FieldNo, DestTy); 464 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 465 return ConstantExpr::getNUWMul(E, N); 466 } 467 468 if (const StructType *STy = dyn_cast<StructType>(Ty)) 469 if (!STy->isPacked()) { 470 unsigned NumElems = STy->getNumElements(); 471 // An empty struct has no members. 472 if (NumElems == 0) 473 return 0; 474 // Check for a struct with all members having the same size. 475 Constant *MemberSize = 476 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 477 bool AllSame = true; 478 for (unsigned i = 1; i != NumElems; ++i) 479 if (MemberSize != 480 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 481 AllSame = false; 482 break; 483 } 484 if (AllSame) { 485 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 486 false, 487 DestTy, 488 false), 489 FieldNo, DestTy); 490 return ConstantExpr::getNUWMul(MemberSize, N); 491 } 492 } 493 494 // If there's no interesting folding happening, bail so that we don't create 495 // a constant that looks like it needs folding but really doesn't. 496 if (!Folded) 497 return 0; 498 499 // Base case: Get a regular offsetof expression. 500 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 501 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 502 DestTy, false), 503 C, DestTy); 504 return C; 505} 506 507Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 508 const Type *DestTy) { 509 if (isa<UndefValue>(V)) { 510 // zext(undef) = 0, because the top bits will be zero. 511 // sext(undef) = 0, because the top bits will all be the same. 512 // [us]itofp(undef) = 0, because the result value is bounded. 513 if (opc == Instruction::ZExt || opc == Instruction::SExt || 514 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 515 return Constant::getNullValue(DestTy); 516 return UndefValue::get(DestTy); 517 } 518 519 // No compile-time operations on this type yet. 520 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) 521 return 0; 522 523 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 524 return Constant::getNullValue(DestTy); 525 526 // If the cast operand is a constant expression, there's a few things we can 527 // do to try to simplify it. 528 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 529 if (CE->isCast()) { 530 // Try hard to fold cast of cast because they are often eliminable. 531 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 532 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 533 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 534 // If all of the indexes in the GEP are null values, there is no pointer 535 // adjustment going on. We might as well cast the source pointer. 536 bool isAllNull = true; 537 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 538 if (!CE->getOperand(i)->isNullValue()) { 539 isAllNull = false; 540 break; 541 } 542 if (isAllNull) 543 // This is casting one pointer type to another, always BitCast 544 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 545 } 546 } 547 548 // If the cast operand is a constant vector, perform the cast by 549 // operating on each element. In the cast of bitcasts, the element 550 // count may be mismatched; don't attempt to handle that here. 551 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 552 if (DestTy->isVectorTy() && 553 cast<VectorType>(DestTy)->getNumElements() == 554 CV->getType()->getNumElements()) { 555 std::vector<Constant*> res; 556 const VectorType *DestVecTy = cast<VectorType>(DestTy); 557 const Type *DstEltTy = DestVecTy->getElementType(); 558 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 559 res.push_back(ConstantExpr::getCast(opc, 560 CV->getOperand(i), DstEltTy)); 561 return ConstantVector::get(DestVecTy, res); 562 } 563 564 // We actually have to do a cast now. Perform the cast according to the 565 // opcode specified. 566 switch (opc) { 567 default: 568 llvm_unreachable("Failed to cast constant expression"); 569 case Instruction::FPTrunc: 570 case Instruction::FPExt: 571 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 572 bool ignored; 573 APFloat Val = FPC->getValueAPF(); 574 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle : 575 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 576 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 577 DestTy->isFP128Ty() ? APFloat::IEEEquad : 578 APFloat::Bogus, 579 APFloat::rmNearestTiesToEven, &ignored); 580 return ConstantFP::get(V->getContext(), Val); 581 } 582 return 0; // Can't fold. 583 case Instruction::FPToUI: 584 case Instruction::FPToSI: 585 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 586 const APFloat &V = FPC->getValueAPF(); 587 bool ignored; 588 uint64_t x[2]; 589 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 590 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 591 APFloat::rmTowardZero, &ignored); 592 APInt Val(DestBitWidth, 2, x); 593 return ConstantInt::get(FPC->getContext(), Val); 594 } 595 return 0; // Can't fold. 596 case Instruction::IntToPtr: //always treated as unsigned 597 if (V->isNullValue()) // Is it an integral null value? 598 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 599 return 0; // Other pointer types cannot be casted 600 case Instruction::PtrToInt: // always treated as unsigned 601 // Is it a null pointer value? 602 if (V->isNullValue()) 603 return ConstantInt::get(DestTy, 0); 604 // If this is a sizeof-like expression, pull out multiplications by 605 // known factors to expose them to subsequent folding. If it's an 606 // alignof-like expression, factor out known factors. 607 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 608 if (CE->getOpcode() == Instruction::GetElementPtr && 609 CE->getOperand(0)->isNullValue()) { 610 const Type *Ty = 611 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 612 if (CE->getNumOperands() == 2) { 613 // Handle a sizeof-like expression. 614 Constant *Idx = CE->getOperand(1); 615 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 616 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 617 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 618 DestTy, false), 619 Idx, DestTy); 620 return ConstantExpr::getMul(C, Idx); 621 } 622 } else if (CE->getNumOperands() == 3 && 623 CE->getOperand(1)->isNullValue()) { 624 // Handle an alignof-like expression. 625 if (const StructType *STy = dyn_cast<StructType>(Ty)) 626 if (!STy->isPacked()) { 627 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 628 if (CI->isOne() && 629 STy->getNumElements() == 2 && 630 STy->getElementType(0)->isIntegerTy(1)) { 631 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 632 } 633 } 634 // Handle an offsetof-like expression. 635 if (Ty->isStructTy() || Ty->isArrayTy()) { 636 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 637 DestTy, false)) 638 return C; 639 } 640 } 641 } 642 // Other pointer types cannot be casted 643 return 0; 644 case Instruction::UIToFP: 645 case Instruction::SIToFP: 646 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 647 APInt api = CI->getValue(); 648 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true); 649 (void)apf.convertFromAPInt(api, 650 opc==Instruction::SIToFP, 651 APFloat::rmNearestTiesToEven); 652 return ConstantFP::get(V->getContext(), apf); 653 } 654 return 0; 655 case Instruction::ZExt: 656 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 657 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 658 return ConstantInt::get(V->getContext(), 659 CI->getValue().zext(BitWidth)); 660 } 661 return 0; 662 case Instruction::SExt: 663 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 664 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 665 return ConstantInt::get(V->getContext(), 666 CI->getValue().sext(BitWidth)); 667 } 668 return 0; 669 case Instruction::Trunc: { 670 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 671 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 672 return ConstantInt::get(V->getContext(), 673 CI->getValue().trunc(DestBitWidth)); 674 } 675 676 // The input must be a constantexpr. See if we can simplify this based on 677 // the bytes we are demanding. Only do this if the source and dest are an 678 // even multiple of a byte. 679 if ((DestBitWidth & 7) == 0 && 680 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 681 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 682 return Res; 683 684 return 0; 685 } 686 case Instruction::BitCast: 687 return FoldBitCast(V, DestTy); 688 } 689} 690 691Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 692 Constant *V1, Constant *V2) { 693 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 694 return CB->getZExtValue() ? V1 : V2; 695 696 // Check for zero aggregate and ConstantVector of zeros 697 if (Cond->isNullValue()) return V2; 698 699 if (ConstantVector* CondV = dyn_cast<ConstantVector>(Cond)) { 700 701 if (CondV->isAllOnesValue()) return V1; 702 703 const VectorType *VTy = cast<VectorType>(V1->getType()); 704 ConstantVector *CP1 = dyn_cast<ConstantVector>(V1); 705 ConstantVector *CP2 = dyn_cast<ConstantVector>(V2); 706 707 if ((CP1 || isa<ConstantAggregateZero>(V1)) && 708 (CP2 || isa<ConstantAggregateZero>(V2))) { 709 710 // Find the element type of the returned vector 711 const Type *EltTy = VTy->getElementType(); 712 unsigned NumElem = VTy->getNumElements(); 713 std::vector<Constant*> Res(NumElem); 714 715 bool Valid = true; 716 for (unsigned i = 0; i < NumElem; ++i) { 717 ConstantInt* c = dyn_cast<ConstantInt>(CondV->getOperand(i)); 718 if (!c) { 719 Valid = false; 720 break; 721 } 722 Constant *C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 723 Constant *C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 724 Res[i] = c->getZExtValue() ? C1 : C2; 725 } 726 // If we were able to build the vector, return it 727 if (Valid) return ConstantVector::get(Res); 728 } 729 } 730 731 732 if (isa<UndefValue>(V1)) return V2; 733 if (isa<UndefValue>(V2)) return V1; 734 if (isa<UndefValue>(Cond)) return V1; 735 if (V1 == V2) return V1; 736 737 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 738 if (TrueVal->getOpcode() == Instruction::Select) 739 if (TrueVal->getOperand(0) == Cond) 740 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 741 } 742 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 743 if (FalseVal->getOpcode() == Instruction::Select) 744 if (FalseVal->getOperand(0) == Cond) 745 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 746 } 747 748 return 0; 749} 750 751Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 752 Constant *Idx) { 753 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 754 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 755 if (Val->isNullValue()) // ee(zero, x) -> zero 756 return Constant::getNullValue( 757 cast<VectorType>(Val->getType())->getElementType()); 758 759 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 760 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 761 return CVal->getOperand(CIdx->getZExtValue()); 762 } else if (isa<UndefValue>(Idx)) { 763 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 764 return CVal->getOperand(0); 765 } 766 } 767 return 0; 768} 769 770Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 771 Constant *Elt, 772 Constant *Idx) { 773 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 774 if (!CIdx) return 0; 775 APInt idxVal = CIdx->getValue(); 776 if (isa<UndefValue>(Val)) { 777 // Insertion of scalar constant into vector undef 778 // Optimize away insertion of undef 779 if (isa<UndefValue>(Elt)) 780 return Val; 781 // Otherwise break the aggregate undef into multiple undefs and do 782 // the insertion 783 unsigned numOps = 784 cast<VectorType>(Val->getType())->getNumElements(); 785 std::vector<Constant*> Ops; 786 Ops.reserve(numOps); 787 for (unsigned i = 0; i < numOps; ++i) { 788 Constant *Op = 789 (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); 790 Ops.push_back(Op); 791 } 792 return ConstantVector::get(Ops); 793 } 794 if (isa<ConstantAggregateZero>(Val)) { 795 // Insertion of scalar constant into vector aggregate zero 796 // Optimize away insertion of zero 797 if (Elt->isNullValue()) 798 return Val; 799 // Otherwise break the aggregate zero into multiple zeros and do 800 // the insertion 801 unsigned numOps = 802 cast<VectorType>(Val->getType())->getNumElements(); 803 std::vector<Constant*> Ops; 804 Ops.reserve(numOps); 805 for (unsigned i = 0; i < numOps; ++i) { 806 Constant *Op = 807 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); 808 Ops.push_back(Op); 809 } 810 return ConstantVector::get(Ops); 811 } 812 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 813 // Insertion of scalar constant into vector constant 814 std::vector<Constant*> Ops; 815 Ops.reserve(CVal->getNumOperands()); 816 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 817 Constant *Op = 818 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 819 Ops.push_back(Op); 820 } 821 return ConstantVector::get(Ops); 822 } 823 824 return 0; 825} 826 827/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 828/// return the specified element value. Otherwise return null. 829static Constant *GetVectorElement(Constant *C, unsigned EltNo) { 830 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) 831 return CV->getOperand(EltNo); 832 833 const Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 834 if (isa<ConstantAggregateZero>(C)) 835 return Constant::getNullValue(EltTy); 836 if (isa<UndefValue>(C)) 837 return UndefValue::get(EltTy); 838 return 0; 839} 840 841Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 842 Constant *V2, 843 Constant *Mask) { 844 // Undefined shuffle mask -> undefined value. 845 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 846 847 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 848 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 849 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 850 851 // Loop over the shuffle mask, evaluating each element. 852 SmallVector<Constant*, 32> Result; 853 for (unsigned i = 0; i != MaskNumElts; ++i) { 854 Constant *InElt = GetVectorElement(Mask, i); 855 if (InElt == 0) return 0; 856 857 if (isa<UndefValue>(InElt)) 858 InElt = UndefValue::get(EltTy); 859 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 860 unsigned Elt = CI->getZExtValue(); 861 if (Elt >= SrcNumElts*2) 862 InElt = UndefValue::get(EltTy); 863 else if (Elt >= SrcNumElts) 864 InElt = GetVectorElement(V2, Elt - SrcNumElts); 865 else 866 InElt = GetVectorElement(V1, Elt); 867 if (InElt == 0) return 0; 868 } else { 869 // Unknown value. 870 return 0; 871 } 872 Result.push_back(InElt); 873 } 874 875 return ConstantVector::get(Result); 876} 877 878Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 879 const unsigned *Idxs, 880 unsigned NumIdx) { 881 // Base case: no indices, so return the entire value. 882 if (NumIdx == 0) 883 return Agg; 884 885 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 886 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), 887 Idxs, 888 Idxs + NumIdx)); 889 890 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 891 return 892 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 893 Idxs, 894 Idxs + NumIdx)); 895 896 // Otherwise recurse. 897 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) 898 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs), 899 Idxs+1, NumIdx-1); 900 901 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) 902 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs), 903 Idxs+1, NumIdx-1); 904 ConstantVector *CV = cast<ConstantVector>(Agg); 905 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs), 906 Idxs+1, NumIdx-1); 907} 908 909Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 910 Constant *Val, 911 const unsigned *Idxs, 912 unsigned NumIdx) { 913 // Base case: no indices, so replace the entire value. 914 if (NumIdx == 0) 915 return Val; 916 917 if (isa<UndefValue>(Agg)) { 918 // Insertion of constant into aggregate undef 919 // Optimize away insertion of undef. 920 if (isa<UndefValue>(Val)) 921 return Agg; 922 923 // Otherwise break the aggregate undef into multiple undefs and do 924 // the insertion. 925 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 926 unsigned numOps; 927 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 928 numOps = AR->getNumElements(); 929 else 930 numOps = cast<StructType>(AggTy)->getNumElements(); 931 932 std::vector<Constant*> Ops(numOps); 933 for (unsigned i = 0; i < numOps; ++i) { 934 const Type *MemberTy = AggTy->getTypeAtIndex(i); 935 Constant *Op = 936 (*Idxs == i) ? 937 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy), 938 Val, Idxs+1, NumIdx-1) : 939 UndefValue::get(MemberTy); 940 Ops[i] = Op; 941 } 942 943 if (const StructType* ST = dyn_cast<StructType>(AggTy)) 944 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); 945 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 946 } 947 948 if (isa<ConstantAggregateZero>(Agg)) { 949 // Insertion of constant into aggregate zero 950 // Optimize away insertion of zero. 951 if (Val->isNullValue()) 952 return Agg; 953 954 // Otherwise break the aggregate zero into multiple zeros and do 955 // the insertion. 956 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 957 unsigned numOps; 958 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 959 numOps = AR->getNumElements(); 960 else 961 numOps = cast<StructType>(AggTy)->getNumElements(); 962 963 std::vector<Constant*> Ops(numOps); 964 for (unsigned i = 0; i < numOps; ++i) { 965 const Type *MemberTy = AggTy->getTypeAtIndex(i); 966 Constant *Op = 967 (*Idxs == i) ? 968 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy), 969 Val, Idxs+1, NumIdx-1) : 970 Constant::getNullValue(MemberTy); 971 Ops[i] = Op; 972 } 973 974 if (const StructType *ST = dyn_cast<StructType>(AggTy)) 975 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); 976 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 977 } 978 979 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 980 // Insertion of constant into aggregate constant. 981 std::vector<Constant*> Ops(Agg->getNumOperands()); 982 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 983 Constant *Op = cast<Constant>(Agg->getOperand(i)); 984 if (*Idxs == i) 985 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1); 986 Ops[i] = Op; 987 } 988 989 if (const StructType* ST = dyn_cast<StructType>(Agg->getType())) 990 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked()); 991 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 992 } 993 994 return 0; 995} 996 997 998Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 999 Constant *C1, Constant *C2) { 1000 // No compile-time operations on this type yet. 1001 if (C1->getType()->isPPC_FP128Ty()) 1002 return 0; 1003 1004 // Handle UndefValue up front. 1005 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1006 switch (Opcode) { 1007 case Instruction::Xor: 1008 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1009 // Handle undef ^ undef -> 0 special case. This is a common 1010 // idiom (misuse). 1011 return Constant::getNullValue(C1->getType()); 1012 // Fallthrough 1013 case Instruction::Add: 1014 case Instruction::Sub: 1015 return UndefValue::get(C1->getType()); 1016 case Instruction::Mul: 1017 case Instruction::And: 1018 return Constant::getNullValue(C1->getType()); 1019 case Instruction::UDiv: 1020 case Instruction::SDiv: 1021 case Instruction::URem: 1022 case Instruction::SRem: 1023 if (!isa<UndefValue>(C2)) // undef / X -> 0 1024 return Constant::getNullValue(C1->getType()); 1025 return C2; // X / undef -> undef 1026 case Instruction::Or: // X | undef -> -1 1027 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType())) 1028 return Constant::getAllOnesValue(PTy); 1029 return Constant::getAllOnesValue(C1->getType()); 1030 case Instruction::LShr: 1031 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 1032 return C1; // undef lshr undef -> undef 1033 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 1034 // undef lshr X -> 0 1035 case Instruction::AShr: 1036 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones 1037 return Constant::getAllOnesValue(C1->getType()); 1038 else if (isa<UndefValue>(C1)) 1039 return C1; // undef ashr undef -> undef 1040 else 1041 return C1; // X ashr undef --> X 1042 case Instruction::Shl: 1043 // undef << X -> 0 or X << undef -> 0 1044 return Constant::getNullValue(C1->getType()); 1045 } 1046 } 1047 1048 // Handle simplifications when the RHS is a constant int. 1049 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1050 switch (Opcode) { 1051 case Instruction::Add: 1052 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1053 break; 1054 case Instruction::Sub: 1055 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1056 break; 1057 case Instruction::Mul: 1058 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1059 if (CI2->equalsInt(1)) 1060 return C1; // X * 1 == X 1061 break; 1062 case Instruction::UDiv: 1063 case Instruction::SDiv: 1064 if (CI2->equalsInt(1)) 1065 return C1; // X / 1 == X 1066 if (CI2->equalsInt(0)) 1067 return UndefValue::get(CI2->getType()); // X / 0 == undef 1068 break; 1069 case Instruction::URem: 1070 case Instruction::SRem: 1071 if (CI2->equalsInt(1)) 1072 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1073 if (CI2->equalsInt(0)) 1074 return UndefValue::get(CI2->getType()); // X % 0 == undef 1075 break; 1076 case Instruction::And: 1077 if (CI2->isZero()) return C2; // X & 0 == 0 1078 if (CI2->isAllOnesValue()) 1079 return C1; // X & -1 == X 1080 1081 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1082 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1083 if (CE1->getOpcode() == Instruction::ZExt) { 1084 unsigned DstWidth = CI2->getType()->getBitWidth(); 1085 unsigned SrcWidth = 1086 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1087 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1088 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1089 return C1; 1090 } 1091 1092 // If and'ing the address of a global with a constant, fold it. 1093 if (CE1->getOpcode() == Instruction::PtrToInt && 1094 isa<GlobalValue>(CE1->getOperand(0))) { 1095 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1096 1097 // Functions are at least 4-byte aligned. 1098 unsigned GVAlign = GV->getAlignment(); 1099 if (isa<Function>(GV)) 1100 GVAlign = std::max(GVAlign, 4U); 1101 1102 if (GVAlign > 1) { 1103 unsigned DstWidth = CI2->getType()->getBitWidth(); 1104 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1105 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1106 1107 // If checking bits we know are clear, return zero. 1108 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1109 return Constant::getNullValue(CI2->getType()); 1110 } 1111 } 1112 } 1113 break; 1114 case Instruction::Or: 1115 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1116 if (CI2->isAllOnesValue()) 1117 return C2; // X | -1 == -1 1118 break; 1119 case Instruction::Xor: 1120 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1121 1122 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1123 switch (CE1->getOpcode()) { 1124 default: break; 1125 case Instruction::ICmp: 1126 case Instruction::FCmp: 1127 // cmp pred ^ true -> cmp !pred 1128 assert(CI2->equalsInt(1)); 1129 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1130 pred = CmpInst::getInversePredicate(pred); 1131 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1132 CE1->getOperand(1)); 1133 } 1134 } 1135 break; 1136 case Instruction::AShr: 1137 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1138 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1139 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1140 return ConstantExpr::getLShr(C1, C2); 1141 break; 1142 } 1143 } else if (isa<ConstantInt>(C1)) { 1144 // If C1 is a ConstantInt and C2 is not, swap the operands. 1145 if (Instruction::isCommutative(Opcode)) 1146 return ConstantExpr::get(Opcode, C2, C1); 1147 } 1148 1149 // At this point we know neither constant is an UndefValue. 1150 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1151 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1152 using namespace APIntOps; 1153 const APInt &C1V = CI1->getValue(); 1154 const APInt &C2V = CI2->getValue(); 1155 switch (Opcode) { 1156 default: 1157 break; 1158 case Instruction::Add: 1159 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1160 case Instruction::Sub: 1161 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1162 case Instruction::Mul: 1163 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1164 case Instruction::UDiv: 1165 assert(!CI2->isNullValue() && "Div by zero handled above"); 1166 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1167 case Instruction::SDiv: 1168 assert(!CI2->isNullValue() && "Div by zero handled above"); 1169 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1170 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1171 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1172 case Instruction::URem: 1173 assert(!CI2->isNullValue() && "Div by zero handled above"); 1174 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1175 case Instruction::SRem: 1176 assert(!CI2->isNullValue() && "Div by zero handled above"); 1177 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1178 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1179 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1180 case Instruction::And: 1181 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1182 case Instruction::Or: 1183 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1184 case Instruction::Xor: 1185 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1186 case Instruction::Shl: { 1187 uint32_t shiftAmt = C2V.getZExtValue(); 1188 if (shiftAmt < C1V.getBitWidth()) 1189 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1190 else 1191 return UndefValue::get(C1->getType()); // too big shift is undef 1192 } 1193 case Instruction::LShr: { 1194 uint32_t shiftAmt = C2V.getZExtValue(); 1195 if (shiftAmt < C1V.getBitWidth()) 1196 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1197 else 1198 return UndefValue::get(C1->getType()); // too big shift is undef 1199 } 1200 case Instruction::AShr: { 1201 uint32_t shiftAmt = C2V.getZExtValue(); 1202 if (shiftAmt < C1V.getBitWidth()) 1203 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1204 else 1205 return UndefValue::get(C1->getType()); // too big shift is undef 1206 } 1207 } 1208 } 1209 1210 switch (Opcode) { 1211 case Instruction::SDiv: 1212 case Instruction::UDiv: 1213 case Instruction::URem: 1214 case Instruction::SRem: 1215 case Instruction::LShr: 1216 case Instruction::AShr: 1217 case Instruction::Shl: 1218 if (CI1->equalsInt(0)) return C1; 1219 break; 1220 default: 1221 break; 1222 } 1223 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1224 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1225 APFloat C1V = CFP1->getValueAPF(); 1226 APFloat C2V = CFP2->getValueAPF(); 1227 APFloat C3V = C1V; // copy for modification 1228 switch (Opcode) { 1229 default: 1230 break; 1231 case Instruction::FAdd: 1232 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1233 return ConstantFP::get(C1->getContext(), C3V); 1234 case Instruction::FSub: 1235 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1236 return ConstantFP::get(C1->getContext(), C3V); 1237 case Instruction::FMul: 1238 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1239 return ConstantFP::get(C1->getContext(), C3V); 1240 case Instruction::FDiv: 1241 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1242 return ConstantFP::get(C1->getContext(), C3V); 1243 case Instruction::FRem: 1244 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1245 return ConstantFP::get(C1->getContext(), C3V); 1246 } 1247 } 1248 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1249 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 1250 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 1251 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 1252 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 1253 std::vector<Constant*> Res; 1254 const Type* EltTy = VTy->getElementType(); 1255 Constant *C1 = 0; 1256 Constant *C2 = 0; 1257 switch (Opcode) { 1258 default: 1259 break; 1260 case Instruction::Add: 1261 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1262 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1263 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1264 Res.push_back(ConstantExpr::getAdd(C1, C2)); 1265 } 1266 return ConstantVector::get(Res); 1267 case Instruction::FAdd: 1268 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1269 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1270 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1271 Res.push_back(ConstantExpr::getFAdd(C1, C2)); 1272 } 1273 return ConstantVector::get(Res); 1274 case Instruction::Sub: 1275 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1276 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1277 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1278 Res.push_back(ConstantExpr::getSub(C1, C2)); 1279 } 1280 return ConstantVector::get(Res); 1281 case Instruction::FSub: 1282 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1283 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1284 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1285 Res.push_back(ConstantExpr::getFSub(C1, C2)); 1286 } 1287 return ConstantVector::get(Res); 1288 case Instruction::Mul: 1289 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1290 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1291 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1292 Res.push_back(ConstantExpr::getMul(C1, C2)); 1293 } 1294 return ConstantVector::get(Res); 1295 case Instruction::FMul: 1296 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1297 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1298 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1299 Res.push_back(ConstantExpr::getFMul(C1, C2)); 1300 } 1301 return ConstantVector::get(Res); 1302 case Instruction::UDiv: 1303 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1304 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1305 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1306 Res.push_back(ConstantExpr::getUDiv(C1, C2)); 1307 } 1308 return ConstantVector::get(Res); 1309 case Instruction::SDiv: 1310 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1311 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1312 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1313 Res.push_back(ConstantExpr::getSDiv(C1, C2)); 1314 } 1315 return ConstantVector::get(Res); 1316 case Instruction::FDiv: 1317 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1318 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1319 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1320 Res.push_back(ConstantExpr::getFDiv(C1, C2)); 1321 } 1322 return ConstantVector::get(Res); 1323 case Instruction::URem: 1324 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1325 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1326 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1327 Res.push_back(ConstantExpr::getURem(C1, C2)); 1328 } 1329 return ConstantVector::get(Res); 1330 case Instruction::SRem: 1331 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1332 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1333 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1334 Res.push_back(ConstantExpr::getSRem(C1, C2)); 1335 } 1336 return ConstantVector::get(Res); 1337 case Instruction::FRem: 1338 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1339 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1340 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1341 Res.push_back(ConstantExpr::getFRem(C1, C2)); 1342 } 1343 return ConstantVector::get(Res); 1344 case Instruction::And: 1345 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1346 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1347 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1348 Res.push_back(ConstantExpr::getAnd(C1, C2)); 1349 } 1350 return ConstantVector::get(Res); 1351 case Instruction::Or: 1352 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1353 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1354 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1355 Res.push_back(ConstantExpr::getOr(C1, C2)); 1356 } 1357 return ConstantVector::get(Res); 1358 case Instruction::Xor: 1359 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1360 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1361 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1362 Res.push_back(ConstantExpr::getXor(C1, C2)); 1363 } 1364 return ConstantVector::get(Res); 1365 case Instruction::LShr: 1366 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1367 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1368 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1369 Res.push_back(ConstantExpr::getLShr(C1, C2)); 1370 } 1371 return ConstantVector::get(Res); 1372 case Instruction::AShr: 1373 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1374 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1375 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1376 Res.push_back(ConstantExpr::getAShr(C1, C2)); 1377 } 1378 return ConstantVector::get(Res); 1379 case Instruction::Shl: 1380 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1381 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1382 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1383 Res.push_back(ConstantExpr::getShl(C1, C2)); 1384 } 1385 return ConstantVector::get(Res); 1386 } 1387 } 1388 } 1389 1390 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1391 // There are many possible foldings we could do here. We should probably 1392 // at least fold add of a pointer with an integer into the appropriate 1393 // getelementptr. This will improve alias analysis a bit. 1394 1395 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1396 // (a + (b + c)). 1397 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1398 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1399 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1400 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1401 } 1402 } else if (isa<ConstantExpr>(C2)) { 1403 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1404 // other way if possible. 1405 if (Instruction::isCommutative(Opcode)) 1406 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1407 } 1408 1409 // i1 can be simplified in many cases. 1410 if (C1->getType()->isIntegerTy(1)) { 1411 switch (Opcode) { 1412 case Instruction::Add: 1413 case Instruction::Sub: 1414 return ConstantExpr::getXor(C1, C2); 1415 case Instruction::Mul: 1416 return ConstantExpr::getAnd(C1, C2); 1417 case Instruction::Shl: 1418 case Instruction::LShr: 1419 case Instruction::AShr: 1420 // We can assume that C2 == 0. If it were one the result would be 1421 // undefined because the shift value is as large as the bitwidth. 1422 return C1; 1423 case Instruction::SDiv: 1424 case Instruction::UDiv: 1425 // We can assume that C2 == 1. If it were zero the result would be 1426 // undefined through division by zero. 1427 return C1; 1428 case Instruction::URem: 1429 case Instruction::SRem: 1430 // We can assume that C2 == 1. If it were zero the result would be 1431 // undefined through division by zero. 1432 return ConstantInt::getFalse(C1->getContext()); 1433 default: 1434 break; 1435 } 1436 } 1437 1438 // We don't know how to fold this. 1439 return 0; 1440} 1441 1442/// isZeroSizedType - This type is zero sized if its an array or structure of 1443/// zero sized types. The only leaf zero sized type is an empty structure. 1444static bool isMaybeZeroSizedType(const Type *Ty) { 1445 if (Ty->isOpaqueTy()) return true; // Can't say. 1446 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1447 1448 // If all of elements have zero size, this does too. 1449 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1450 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1451 return true; 1452 1453 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1454 return isMaybeZeroSizedType(ATy->getElementType()); 1455 } 1456 return false; 1457} 1458 1459/// IdxCompare - Compare the two constants as though they were getelementptr 1460/// indices. This allows coersion of the types to be the same thing. 1461/// 1462/// If the two constants are the "same" (after coersion), return 0. If the 1463/// first is less than the second, return -1, if the second is less than the 1464/// first, return 1. If the constants are not integral, return -2. 1465/// 1466static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) { 1467 if (C1 == C2) return 0; 1468 1469 // Ok, we found a different index. If they are not ConstantInt, we can't do 1470 // anything with them. 1471 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1472 return -2; // don't know! 1473 1474 // Ok, we have two differing integer indices. Sign extend them to be the same 1475 // type. Long is always big enough, so we use it. 1476 if (!C1->getType()->isIntegerTy(64)) 1477 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1478 1479 if (!C2->getType()->isIntegerTy(64)) 1480 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1481 1482 if (C1 == C2) return 0; // They are equal 1483 1484 // If the type being indexed over is really just a zero sized type, there is 1485 // no pointer difference being made here. 1486 if (isMaybeZeroSizedType(ElTy)) 1487 return -2; // dunno. 1488 1489 // If they are really different, now that they are the same type, then we 1490 // found a difference! 1491 if (cast<ConstantInt>(C1)->getSExtValue() < 1492 cast<ConstantInt>(C2)->getSExtValue()) 1493 return -1; 1494 else 1495 return 1; 1496} 1497 1498/// evaluateFCmpRelation - This function determines if there is anything we can 1499/// decide about the two constants provided. This doesn't need to handle simple 1500/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1501/// If we can determine that the two constants have a particular relation to 1502/// each other, we should return the corresponding FCmpInst predicate, 1503/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1504/// ConstantFoldCompareInstruction. 1505/// 1506/// To simplify this code we canonicalize the relation so that the first 1507/// operand is always the most "complex" of the two. We consider ConstantFP 1508/// to be the simplest, and ConstantExprs to be the most complex. 1509static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1510 assert(V1->getType() == V2->getType() && 1511 "Cannot compare values of different types!"); 1512 1513 // No compile-time operations on this type yet. 1514 if (V1->getType()->isPPC_FP128Ty()) 1515 return FCmpInst::BAD_FCMP_PREDICATE; 1516 1517 // Handle degenerate case quickly 1518 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1519 1520 if (!isa<ConstantExpr>(V1)) { 1521 if (!isa<ConstantExpr>(V2)) { 1522 // We distilled thisUse the standard constant folder for a few cases 1523 ConstantInt *R = 0; 1524 R = dyn_cast<ConstantInt>( 1525 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1526 if (R && !R->isZero()) 1527 return FCmpInst::FCMP_OEQ; 1528 R = dyn_cast<ConstantInt>( 1529 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1530 if (R && !R->isZero()) 1531 return FCmpInst::FCMP_OLT; 1532 R = dyn_cast<ConstantInt>( 1533 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1534 if (R && !R->isZero()) 1535 return FCmpInst::FCMP_OGT; 1536 1537 // Nothing more we can do 1538 return FCmpInst::BAD_FCMP_PREDICATE; 1539 } 1540 1541 // If the first operand is simple and second is ConstantExpr, swap operands. 1542 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1543 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1544 return FCmpInst::getSwappedPredicate(SwappedRelation); 1545 } else { 1546 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1547 // constantexpr or a simple constant. 1548 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1549 switch (CE1->getOpcode()) { 1550 case Instruction::FPTrunc: 1551 case Instruction::FPExt: 1552 case Instruction::UIToFP: 1553 case Instruction::SIToFP: 1554 // We might be able to do something with these but we don't right now. 1555 break; 1556 default: 1557 break; 1558 } 1559 } 1560 // There are MANY other foldings that we could perform here. They will 1561 // probably be added on demand, as they seem needed. 1562 return FCmpInst::BAD_FCMP_PREDICATE; 1563} 1564 1565/// evaluateICmpRelation - This function determines if there is anything we can 1566/// decide about the two constants provided. This doesn't need to handle simple 1567/// things like integer comparisons, but should instead handle ConstantExprs 1568/// and GlobalValues. If we can determine that the two constants have a 1569/// particular relation to each other, we should return the corresponding ICmp 1570/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1571/// 1572/// To simplify this code we canonicalize the relation so that the first 1573/// operand is always the most "complex" of the two. We consider simple 1574/// constants (like ConstantInt) to be the simplest, followed by 1575/// GlobalValues, followed by ConstantExpr's (the most complex). 1576/// 1577static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1578 bool isSigned) { 1579 assert(V1->getType() == V2->getType() && 1580 "Cannot compare different types of values!"); 1581 if (V1 == V2) return ICmpInst::ICMP_EQ; 1582 1583 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1584 !isa<BlockAddress>(V1)) { 1585 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1586 !isa<BlockAddress>(V2)) { 1587 // We distilled this down to a simple case, use the standard constant 1588 // folder. 1589 ConstantInt *R = 0; 1590 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1591 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1592 if (R && !R->isZero()) 1593 return pred; 1594 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1595 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1596 if (R && !R->isZero()) 1597 return pred; 1598 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1599 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1600 if (R && !R->isZero()) 1601 return pred; 1602 1603 // If we couldn't figure it out, bail. 1604 return ICmpInst::BAD_ICMP_PREDICATE; 1605 } 1606 1607 // If the first operand is simple, swap operands. 1608 ICmpInst::Predicate SwappedRelation = 1609 evaluateICmpRelation(V2, V1, isSigned); 1610 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1611 return ICmpInst::getSwappedPredicate(SwappedRelation); 1612 1613 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1614 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1615 ICmpInst::Predicate SwappedRelation = 1616 evaluateICmpRelation(V2, V1, isSigned); 1617 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1618 return ICmpInst::getSwappedPredicate(SwappedRelation); 1619 return ICmpInst::BAD_ICMP_PREDICATE; 1620 } 1621 1622 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1623 // constant (which, since the types must match, means that it's a 1624 // ConstantPointerNull). 1625 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1626 // Don't try to decide equality of aliases. 1627 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2)) 1628 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 1629 return ICmpInst::ICMP_NE; 1630 } else if (isa<BlockAddress>(V2)) { 1631 return ICmpInst::ICMP_NE; // Globals never equal labels. 1632 } else { 1633 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1634 // GlobalVals can never be null unless they have external weak linkage. 1635 // We don't try to evaluate aliases here. 1636 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1637 return ICmpInst::ICMP_NE; 1638 } 1639 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1640 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1641 ICmpInst::Predicate SwappedRelation = 1642 evaluateICmpRelation(V2, V1, isSigned); 1643 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1644 return ICmpInst::getSwappedPredicate(SwappedRelation); 1645 return ICmpInst::BAD_ICMP_PREDICATE; 1646 } 1647 1648 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1649 // constant (which, since the types must match, means that it is a 1650 // ConstantPointerNull). 1651 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1652 // Block address in another function can't equal this one, but block 1653 // addresses in the current function might be the same if blocks are 1654 // empty. 1655 if (BA2->getFunction() != BA->getFunction()) 1656 return ICmpInst::ICMP_NE; 1657 } else { 1658 // Block addresses aren't null, don't equal the address of globals. 1659 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1660 "Canonicalization guarantee!"); 1661 return ICmpInst::ICMP_NE; 1662 } 1663 } else { 1664 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1665 // constantexpr, a global, block address, or a simple constant. 1666 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1667 Constant *CE1Op0 = CE1->getOperand(0); 1668 1669 switch (CE1->getOpcode()) { 1670 case Instruction::Trunc: 1671 case Instruction::FPTrunc: 1672 case Instruction::FPExt: 1673 case Instruction::FPToUI: 1674 case Instruction::FPToSI: 1675 break; // We can't evaluate floating point casts or truncations. 1676 1677 case Instruction::UIToFP: 1678 case Instruction::SIToFP: 1679 case Instruction::BitCast: 1680 case Instruction::ZExt: 1681 case Instruction::SExt: 1682 // If the cast is not actually changing bits, and the second operand is a 1683 // null pointer, do the comparison with the pre-casted value. 1684 if (V2->isNullValue() && 1685 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1686 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1687 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1688 return evaluateICmpRelation(CE1Op0, 1689 Constant::getNullValue(CE1Op0->getType()), 1690 isSigned); 1691 } 1692 break; 1693 1694 case Instruction::GetElementPtr: 1695 // Ok, since this is a getelementptr, we know that the constant has a 1696 // pointer type. Check the various cases. 1697 if (isa<ConstantPointerNull>(V2)) { 1698 // If we are comparing a GEP to a null pointer, check to see if the base 1699 // of the GEP equals the null pointer. 1700 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1701 if (GV->hasExternalWeakLinkage()) 1702 // Weak linkage GVals could be zero or not. We're comparing that 1703 // to null pointer so its greater-or-equal 1704 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1705 else 1706 // If its not weak linkage, the GVal must have a non-zero address 1707 // so the result is greater-than 1708 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1709 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1710 // If we are indexing from a null pointer, check to see if we have any 1711 // non-zero indices. 1712 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1713 if (!CE1->getOperand(i)->isNullValue()) 1714 // Offsetting from null, must not be equal. 1715 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1716 // Only zero indexes from null, must still be zero. 1717 return ICmpInst::ICMP_EQ; 1718 } 1719 // Otherwise, we can't really say if the first operand is null or not. 1720 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1721 if (isa<ConstantPointerNull>(CE1Op0)) { 1722 if (GV2->hasExternalWeakLinkage()) 1723 // Weak linkage GVals could be zero or not. We're comparing it to 1724 // a null pointer, so its less-or-equal 1725 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1726 else 1727 // If its not weak linkage, the GVal must have a non-zero address 1728 // so the result is less-than 1729 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1730 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1731 if (GV == GV2) { 1732 // If this is a getelementptr of the same global, then it must be 1733 // different. Because the types must match, the getelementptr could 1734 // only have at most one index, and because we fold getelementptr's 1735 // with a single zero index, it must be nonzero. 1736 assert(CE1->getNumOperands() == 2 && 1737 !CE1->getOperand(1)->isNullValue() && 1738 "Suprising getelementptr!"); 1739 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1740 } else { 1741 // If they are different globals, we don't know what the value is, 1742 // but they can't be equal. 1743 return ICmpInst::ICMP_NE; 1744 } 1745 } 1746 } else { 1747 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1748 Constant *CE2Op0 = CE2->getOperand(0); 1749 1750 // There are MANY other foldings that we could perform here. They will 1751 // probably be added on demand, as they seem needed. 1752 switch (CE2->getOpcode()) { 1753 default: break; 1754 case Instruction::GetElementPtr: 1755 // By far the most common case to handle is when the base pointers are 1756 // obviously to the same or different globals. 1757 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1758 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1759 return ICmpInst::ICMP_NE; 1760 // Ok, we know that both getelementptr instructions are based on the 1761 // same global. From this, we can precisely determine the relative 1762 // ordering of the resultant pointers. 1763 unsigned i = 1; 1764 1765 // The logic below assumes that the result of the comparison 1766 // can be determined by finding the first index that differs. 1767 // This doesn't work if there is over-indexing in any 1768 // subsequent indices, so check for that case first. 1769 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1770 !CE2->isGEPWithNoNotionalOverIndexing()) 1771 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1772 1773 // Compare all of the operands the GEP's have in common. 1774 gep_type_iterator GTI = gep_type_begin(CE1); 1775 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1776 ++i, ++GTI) 1777 switch (IdxCompare(CE1->getOperand(i), 1778 CE2->getOperand(i), GTI.getIndexedType())) { 1779 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1780 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1781 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1782 } 1783 1784 // Ok, we ran out of things they have in common. If any leftovers 1785 // are non-zero then we have a difference, otherwise we are equal. 1786 for (; i < CE1->getNumOperands(); ++i) 1787 if (!CE1->getOperand(i)->isNullValue()) { 1788 if (isa<ConstantInt>(CE1->getOperand(i))) 1789 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1790 else 1791 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1792 } 1793 1794 for (; i < CE2->getNumOperands(); ++i) 1795 if (!CE2->getOperand(i)->isNullValue()) { 1796 if (isa<ConstantInt>(CE2->getOperand(i))) 1797 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1798 else 1799 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1800 } 1801 return ICmpInst::ICMP_EQ; 1802 } 1803 } 1804 } 1805 default: 1806 break; 1807 } 1808 } 1809 1810 return ICmpInst::BAD_ICMP_PREDICATE; 1811} 1812 1813Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1814 Constant *C1, Constant *C2) { 1815 const Type *ResultTy; 1816 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1817 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1818 VT->getNumElements()); 1819 else 1820 ResultTy = Type::getInt1Ty(C1->getContext()); 1821 1822 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1823 if (pred == FCmpInst::FCMP_FALSE) 1824 return Constant::getNullValue(ResultTy); 1825 1826 if (pred == FCmpInst::FCMP_TRUE) 1827 return Constant::getAllOnesValue(ResultTy); 1828 1829 // Handle some degenerate cases first 1830 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1831 // For EQ and NE, we can always pick a value for the undef to make the 1832 // predicate pass or fail, so we can return undef. 1833 if (ICmpInst::isEquality(ICmpInst::Predicate(pred))) 1834 return UndefValue::get(ResultTy); 1835 // Otherwise, pick the same value as the non-undef operand, and fold 1836 // it to true or false. 1837 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1838 } 1839 1840 // No compile-time operations on this type yet. 1841 if (C1->getType()->isPPC_FP128Ty()) 1842 return 0; 1843 1844 // icmp eq/ne(null,GV) -> false/true 1845 if (C1->isNullValue()) { 1846 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1847 // Don't try to evaluate aliases. External weak GV can be null. 1848 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1849 if (pred == ICmpInst::ICMP_EQ) 1850 return ConstantInt::getFalse(C1->getContext()); 1851 else if (pred == ICmpInst::ICMP_NE) 1852 return ConstantInt::getTrue(C1->getContext()); 1853 } 1854 // icmp eq/ne(GV,null) -> false/true 1855 } else if (C2->isNullValue()) { 1856 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1857 // Don't try to evaluate aliases. External weak GV can be null. 1858 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1859 if (pred == ICmpInst::ICMP_EQ) 1860 return ConstantInt::getFalse(C1->getContext()); 1861 else if (pred == ICmpInst::ICMP_NE) 1862 return ConstantInt::getTrue(C1->getContext()); 1863 } 1864 } 1865 1866 // If the comparison is a comparison between two i1's, simplify it. 1867 if (C1->getType()->isIntegerTy(1)) { 1868 switch(pred) { 1869 case ICmpInst::ICMP_EQ: 1870 if (isa<ConstantInt>(C2)) 1871 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1872 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1873 case ICmpInst::ICMP_NE: 1874 return ConstantExpr::getXor(C1, C2); 1875 default: 1876 break; 1877 } 1878 } 1879 1880 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1881 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1882 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1883 switch (pred) { 1884 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1885 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1886 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1887 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1888 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1889 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1890 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1891 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1892 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1893 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1894 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1895 } 1896 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1897 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1898 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1899 APFloat::cmpResult R = C1V.compare(C2V); 1900 switch (pred) { 1901 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1902 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1903 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1904 case FCmpInst::FCMP_UNO: 1905 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1906 case FCmpInst::FCMP_ORD: 1907 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1908 case FCmpInst::FCMP_UEQ: 1909 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1910 R==APFloat::cmpEqual); 1911 case FCmpInst::FCMP_OEQ: 1912 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1913 case FCmpInst::FCMP_UNE: 1914 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1915 case FCmpInst::FCMP_ONE: 1916 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1917 R==APFloat::cmpGreaterThan); 1918 case FCmpInst::FCMP_ULT: 1919 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1920 R==APFloat::cmpLessThan); 1921 case FCmpInst::FCMP_OLT: 1922 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1923 case FCmpInst::FCMP_UGT: 1924 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1925 R==APFloat::cmpGreaterThan); 1926 case FCmpInst::FCMP_OGT: 1927 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1928 case FCmpInst::FCMP_ULE: 1929 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1930 case FCmpInst::FCMP_OLE: 1931 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1932 R==APFloat::cmpEqual); 1933 case FCmpInst::FCMP_UGE: 1934 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1935 case FCmpInst::FCMP_OGE: 1936 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1937 R==APFloat::cmpEqual); 1938 } 1939 } else if (C1->getType()->isVectorTy()) { 1940 SmallVector<Constant*, 16> C1Elts, C2Elts; 1941 C1->getVectorElements(C1Elts); 1942 C2->getVectorElements(C2Elts); 1943 if (C1Elts.empty() || C2Elts.empty()) 1944 return 0; 1945 1946 // If we can constant fold the comparison of each element, constant fold 1947 // the whole vector comparison. 1948 SmallVector<Constant*, 4> ResElts; 1949 // Compare the elements, producing an i1 result or constant expr. 1950 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) 1951 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1952 1953 return ConstantVector::get(ResElts); 1954 } 1955 1956 if (C1->getType()->isFloatingPointTy()) { 1957 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1958 switch (evaluateFCmpRelation(C1, C2)) { 1959 default: llvm_unreachable("Unknown relation!"); 1960 case FCmpInst::FCMP_UNO: 1961 case FCmpInst::FCMP_ORD: 1962 case FCmpInst::FCMP_UEQ: 1963 case FCmpInst::FCMP_UNE: 1964 case FCmpInst::FCMP_ULT: 1965 case FCmpInst::FCMP_UGT: 1966 case FCmpInst::FCMP_ULE: 1967 case FCmpInst::FCMP_UGE: 1968 case FCmpInst::FCMP_TRUE: 1969 case FCmpInst::FCMP_FALSE: 1970 case FCmpInst::BAD_FCMP_PREDICATE: 1971 break; // Couldn't determine anything about these constants. 1972 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1973 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1974 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1975 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1976 break; 1977 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1978 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1979 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 1980 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 1981 break; 1982 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1983 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 1984 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 1985 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1986 break; 1987 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1988 // We can only partially decide this relation. 1989 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1990 Result = 0; 1991 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1992 Result = 1; 1993 break; 1994 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1995 // We can only partially decide this relation. 1996 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 1997 Result = 0; 1998 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 1999 Result = 1; 2000 break; 2001 case FCmpInst::FCMP_ONE: // We know that C1 != C2 2002 // We can only partially decide this relation. 2003 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 2004 Result = 0; 2005 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 2006 Result = 1; 2007 break; 2008 } 2009 2010 // If we evaluated the result, return it now. 2011 if (Result != -1) 2012 return ConstantInt::get(ResultTy, Result); 2013 2014 } else { 2015 // Evaluate the relation between the two constants, per the predicate. 2016 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 2017 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 2018 default: llvm_unreachable("Unknown relational!"); 2019 case ICmpInst::BAD_ICMP_PREDICATE: 2020 break; // Couldn't determine anything about these constants. 2021 case ICmpInst::ICMP_EQ: // We know the constants are equal! 2022 // If we know the constants are equal, we can decide the result of this 2023 // computation precisely. 2024 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 2025 break; 2026 case ICmpInst::ICMP_ULT: 2027 switch (pred) { 2028 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 2029 Result = 1; break; 2030 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 2031 Result = 0; break; 2032 } 2033 break; 2034 case ICmpInst::ICMP_SLT: 2035 switch (pred) { 2036 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 2037 Result = 1; break; 2038 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 2039 Result = 0; break; 2040 } 2041 break; 2042 case ICmpInst::ICMP_UGT: 2043 switch (pred) { 2044 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 2045 Result = 1; break; 2046 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 2047 Result = 0; break; 2048 } 2049 break; 2050 case ICmpInst::ICMP_SGT: 2051 switch (pred) { 2052 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 2053 Result = 1; break; 2054 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 2055 Result = 0; break; 2056 } 2057 break; 2058 case ICmpInst::ICMP_ULE: 2059 if (pred == ICmpInst::ICMP_UGT) Result = 0; 2060 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 2061 break; 2062 case ICmpInst::ICMP_SLE: 2063 if (pred == ICmpInst::ICMP_SGT) Result = 0; 2064 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 2065 break; 2066 case ICmpInst::ICMP_UGE: 2067 if (pred == ICmpInst::ICMP_ULT) Result = 0; 2068 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 2069 break; 2070 case ICmpInst::ICMP_SGE: 2071 if (pred == ICmpInst::ICMP_SLT) Result = 0; 2072 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 2073 break; 2074 case ICmpInst::ICMP_NE: 2075 if (pred == ICmpInst::ICMP_EQ) Result = 0; 2076 if (pred == ICmpInst::ICMP_NE) Result = 1; 2077 break; 2078 } 2079 2080 // If we evaluated the result, return it now. 2081 if (Result != -1) 2082 return ConstantInt::get(ResultTy, Result); 2083 2084 // If the right hand side is a bitcast, try using its inverse to simplify 2085 // it by moving it to the left hand side. We can't do this if it would turn 2086 // a vector compare into a scalar compare or visa versa. 2087 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 2088 Constant *CE2Op0 = CE2->getOperand(0); 2089 if (CE2->getOpcode() == Instruction::BitCast && 2090 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 2091 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 2092 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 2093 } 2094 } 2095 2096 // If the left hand side is an extension, try eliminating it. 2097 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 2098 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 2099 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 2100 Constant *CE1Op0 = CE1->getOperand(0); 2101 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 2102 if (CE1Inverse == CE1Op0) { 2103 // Check whether we can safely truncate the right hand side. 2104 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 2105 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) { 2106 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 2107 } 2108 } 2109 } 2110 } 2111 2112 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 2113 (C1->isNullValue() && !C2->isNullValue())) { 2114 // If C2 is a constant expr and C1 isn't, flip them around and fold the 2115 // other way if possible. 2116 // Also, if C1 is null and C2 isn't, flip them around. 2117 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 2118 return ConstantExpr::getICmp(pred, C2, C1); 2119 } 2120 } 2121 return 0; 2122} 2123 2124/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 2125/// is "inbounds". 2126template<typename IndexTy> 2127static bool isInBoundsIndices(IndexTy const *Idxs, size_t NumIdx) { 2128 // No indices means nothing that could be out of bounds. 2129 if (NumIdx == 0) return true; 2130 2131 // If the first index is zero, it's in bounds. 2132 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2133 2134 // If the first index is one and all the rest are zero, it's in bounds, 2135 // by the one-past-the-end rule. 2136 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2137 return false; 2138 for (unsigned i = 1, e = NumIdx; i != e; ++i) 2139 if (!cast<Constant>(Idxs[i])->isNullValue()) 2140 return false; 2141 return true; 2142} 2143 2144template<typename IndexTy> 2145static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 2146 bool inBounds, 2147 IndexTy const *Idxs, 2148 unsigned NumIdx) { 2149 Constant *Idx0 = cast<Constant>(Idxs[0]); 2150 if (NumIdx == 0 || 2151 (NumIdx == 1 && Idx0->isNullValue())) 2152 return C; 2153 2154 if (isa<UndefValue>(C)) { 2155 const PointerType *Ptr = cast<PointerType>(C->getType()); 2156 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, Idxs+NumIdx); 2157 assert(Ty != 0 && "Invalid indices for GEP!"); 2158 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 2159 } 2160 2161 if (C->isNullValue()) { 2162 bool isNull = true; 2163 for (unsigned i = 0, e = NumIdx; i != e; ++i) 2164 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2165 isNull = false; 2166 break; 2167 } 2168 if (isNull) { 2169 const PointerType *Ptr = cast<PointerType>(C->getType()); 2170 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, 2171 Idxs+NumIdx); 2172 assert(Ty != 0 && "Invalid indices for GEP!"); 2173 return ConstantPointerNull::get(PointerType::get(Ty, 2174 Ptr->getAddressSpace())); 2175 } 2176 } 2177 2178 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2179 // Combine Indices - If the source pointer to this getelementptr instruction 2180 // is a getelementptr instruction, combine the indices of the two 2181 // getelementptr instructions into a single instruction. 2182 // 2183 if (CE->getOpcode() == Instruction::GetElementPtr) { 2184 const Type *LastTy = 0; 2185 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2186 I != E; ++I) 2187 LastTy = *I; 2188 2189 if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) { 2190 SmallVector<Value*, 16> NewIndices; 2191 NewIndices.reserve(NumIdx + CE->getNumOperands()); 2192 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 2193 NewIndices.push_back(CE->getOperand(i)); 2194 2195 // Add the last index of the source with the first index of the new GEP. 2196 // Make sure to handle the case when they are actually different types. 2197 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2198 // Otherwise it must be an array. 2199 if (!Idx0->isNullValue()) { 2200 const Type *IdxTy = Combined->getType(); 2201 if (IdxTy != Idx0->getType()) { 2202 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 2203 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 2204 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 2205 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2206 } else { 2207 Combined = 2208 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2209 } 2210 } 2211 2212 NewIndices.push_back(Combined); 2213 NewIndices.append(Idxs+1, Idxs+NumIdx); 2214 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ? 2215 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), 2216 &NewIndices[0], 2217 NewIndices.size()) : 2218 ConstantExpr::getGetElementPtr(CE->getOperand(0), 2219 &NewIndices[0], 2220 NewIndices.size()); 2221 } 2222 } 2223 2224 // Implement folding of: 2225 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2226 // i64 0, i64 0) 2227 // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2228 // 2229 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { 2230 if (const PointerType *SPT = 2231 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 2232 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 2233 if (const ArrayType *CAT = 2234 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 2235 if (CAT->getElementType() == SAT->getElementType()) 2236 return inBounds ? 2237 ConstantExpr::getInBoundsGetElementPtr( 2238 (Constant*)CE->getOperand(0), Idxs, NumIdx) : 2239 ConstantExpr::getGetElementPtr( 2240 (Constant*)CE->getOperand(0), Idxs, NumIdx); 2241 } 2242 } 2243 2244 // Check to see if any array indices are not within the corresponding 2245 // notional array bounds. If so, try to determine if they can be factored 2246 // out into preceding dimensions. 2247 bool Unknown = false; 2248 SmallVector<Constant *, 8> NewIdxs; 2249 const Type *Ty = C->getType(); 2250 const Type *Prev = 0; 2251 for (unsigned i = 0; i != NumIdx; 2252 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2253 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2254 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2255 if (ATy->getNumElements() <= INT64_MAX && 2256 ATy->getNumElements() != 0 && 2257 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 2258 if (isa<SequentialType>(Prev)) { 2259 // It's out of range, but we can factor it into the prior 2260 // dimension. 2261 NewIdxs.resize(NumIdx); 2262 ConstantInt *Factor = ConstantInt::get(CI->getType(), 2263 ATy->getNumElements()); 2264 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2265 2266 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2267 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2268 2269 // Before adding, extend both operands to i64 to avoid 2270 // overflow trouble. 2271 if (!PrevIdx->getType()->isIntegerTy(64)) 2272 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2273 Type::getInt64Ty(Div->getContext())); 2274 if (!Div->getType()->isIntegerTy(64)) 2275 Div = ConstantExpr::getSExt(Div, 2276 Type::getInt64Ty(Div->getContext())); 2277 2278 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2279 } else { 2280 // It's out of range, but the prior dimension is a struct 2281 // so we can't do anything about it. 2282 Unknown = true; 2283 } 2284 } 2285 } else { 2286 // We don't know if it's in range or not. 2287 Unknown = true; 2288 } 2289 } 2290 2291 // If we did any factoring, start over with the adjusted indices. 2292 if (!NewIdxs.empty()) { 2293 for (unsigned i = 0; i != NumIdx; ++i) 2294 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2295 return inBounds ? 2296 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), 2297 NewIdxs.size()) : 2298 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); 2299 } 2300 2301 // If all indices are known integers and normalized, we can do a simple 2302 // check for the "inbounds" property. 2303 if (!Unknown && !inBounds && 2304 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx)) 2305 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx); 2306 2307 return 0; 2308} 2309 2310Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2311 bool inBounds, 2312 Constant* const *Idxs, 2313 unsigned NumIdx) { 2314 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx); 2315} 2316 2317Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2318 bool inBounds, 2319 Value* const *Idxs, 2320 unsigned NumIdx) { 2321 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx); 2322} 2323