ConstantFold.cpp revision 30cb6dda5ae441f27edef302cbb33936fbafba6d
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/Operator.h" 28#include "llvm/ADT/SmallVector.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 ConstantVector 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(ConstantVector *CV, 45 const VectorType *DstTy) { 46 47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 48 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 49 50 // If this cast changes element count then we can't handle it here: 51 // doing so requires endianness information. This should be handled by 52 // Analysis/ConstantFolding.cpp 53 unsigned NumElts = DstTy->getNumElements(); 54 if (NumElts != CV->getNumOperands()) 55 return 0; 56 57 // Check to verify that all elements of the input are simple. 58 for (unsigned i = 0; i != NumElts; ++i) { 59 if (!isa<ConstantInt>(CV->getOperand(i)) && 60 !isa<ConstantFP>(CV->getOperand(i))) 61 return 0; 62 } 63 64 // Bitcast each element now. 65 std::vector<Constant*> Result; 66 const Type *DstEltTy = DstTy->getElementType(); 67 for (unsigned i = 0; i != NumElts; ++i) 68 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i), 69 DstEltTy)); 70 return ConstantVector::get(Result); 71} 72 73/// This function determines which opcode to use to fold two constant cast 74/// expressions together. It uses CastInst::isEliminableCastPair to determine 75/// the opcode. Consequently its just a wrapper around that function. 76/// @brief Determine if it is valid to fold a cast of a cast 77static unsigned 78foldConstantCastPair( 79 unsigned opc, ///< opcode of the second cast constant expression 80 ConstantExpr *Op, ///< the first cast constant expression 81 const Type *DstTy ///< desintation type of the first cast 82) { 83 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 84 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 85 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 86 87 // The the types and opcodes for the two Cast constant expressions 88 const Type *SrcTy = Op->getOperand(0)->getType(); 89 const Type *MidTy = Op->getType(); 90 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 91 Instruction::CastOps secondOp = Instruction::CastOps(opc); 92 93 // Let CastInst::isEliminableCastPair do the heavy lifting. 94 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 95 Type::getInt64Ty(DstTy->getContext())); 96} 97 98static Constant *FoldBitCast(Constant *V, const Type *DestTy) { 99 const Type *SrcTy = V->getType(); 100 if (SrcTy == DestTy) 101 return V; // no-op cast 102 103 // Check to see if we are casting a pointer to an aggregate to a pointer to 104 // the first element. If so, return the appropriate GEP instruction. 105 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) 106 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) 107 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) { 108 SmallVector<Value*, 8> IdxList; 109 Value *Zero = 110 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext())); 111 IdxList.push_back(Zero); 112 const Type *ElTy = PTy->getElementType(); 113 while (ElTy != DPTy->getElementType()) { 114 if (const StructType *STy = dyn_cast<StructType>(ElTy)) { 115 if (STy->getNumElements() == 0) break; 116 ElTy = STy->getElementType(0); 117 IdxList.push_back(Zero); 118 } else if (const SequentialType *STy = 119 dyn_cast<SequentialType>(ElTy)) { 120 if (ElTy->isPointerTy()) break; // Can't index into pointers! 121 ElTy = STy->getElementType(); 122 IdxList.push_back(Zero); 123 } else { 124 break; 125 } 126 } 127 128 if (ElTy == DPTy->getElementType()) 129 // This GEP is inbounds because all indices are zero. 130 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0], 131 IdxList.size()); 132 } 133 134 // Handle casts from one vector constant to another. We know that the src 135 // and dest type have the same size (otherwise its an illegal cast). 136 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 137 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 138 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() && 139 "Not cast between same sized vectors!"); 140 SrcTy = NULL; 141 // First, check for null. Undef is already handled. 142 if (isa<ConstantAggregateZero>(V)) 143 return Constant::getNullValue(DestTy); 144 145 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 146 return BitCastConstantVector(CV, DestPTy); 147 } 148 149 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 150 // This allows for other simplifications (although some of them 151 // can only be handled by Analysis/ConstantFolding.cpp). 152 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 153 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 154 } 155 156 // Finally, implement bitcast folding now. The code below doesn't handle 157 // bitcast right. 158 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 159 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 160 161 // Handle integral constant input. 162 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 163 if (DestTy->isIntegerTy()) 164 // Integral -> Integral. This is a no-op because the bit widths must 165 // be the same. Consequently, we just fold to V. 166 return V; 167 168 if (DestTy->isFloatingPointTy()) 169 return ConstantFP::get(DestTy->getContext(), 170 APFloat(CI->getValue(), 171 !DestTy->isPPC_FP128Ty())); 172 173 // Otherwise, can't fold this (vector?) 174 return 0; 175 } 176 177 // Handle ConstantFP input: FP -> Integral. 178 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) 179 return ConstantInt::get(FP->getContext(), 180 FP->getValueAPF().bitcastToAPInt()); 181 182 return 0; 183} 184 185 186/// ExtractConstantBytes - V is an integer constant which only has a subset of 187/// its bytes used. The bytes used are indicated by ByteStart (which is the 188/// first byte used, counting from the least significant byte) and ByteSize, 189/// which is the number of bytes used. 190/// 191/// This function analyzes the specified constant to see if the specified byte 192/// range can be returned as a simplified constant. If so, the constant is 193/// returned, otherwise null is returned. 194/// 195static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 196 unsigned ByteSize) { 197 assert(C->getType()->isIntegerTy() && 198 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 199 "Non-byte sized integer input"); 200 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 201 assert(ByteSize && "Must be accessing some piece"); 202 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 203 assert(ByteSize != CSize && "Should not extract everything"); 204 205 // Constant Integers are simple. 206 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 207 APInt V = CI->getValue(); 208 if (ByteStart) 209 V = V.lshr(ByteStart*8); 210 V = V.trunc(ByteSize*8); 211 return ConstantInt::get(CI->getContext(), V); 212 } 213 214 // In the input is a constant expr, we might be able to recursively simplify. 215 // If not, we definitely can't do anything. 216 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 217 if (CE == 0) return 0; 218 219 switch (CE->getOpcode()) { 220 default: return 0; 221 case Instruction::Or: { 222 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 223 if (RHS == 0) 224 return 0; 225 226 // X | -1 -> -1. 227 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 228 if (RHSC->isAllOnesValue()) 229 return RHSC; 230 231 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 232 if (LHS == 0) 233 return 0; 234 return ConstantExpr::getOr(LHS, RHS); 235 } 236 case Instruction::And: { 237 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 238 if (RHS == 0) 239 return 0; 240 241 // X & 0 -> 0. 242 if (RHS->isNullValue()) 243 return RHS; 244 245 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 246 if (LHS == 0) 247 return 0; 248 return ConstantExpr::getAnd(LHS, RHS); 249 } 250 case Instruction::LShr: { 251 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 252 if (Amt == 0) 253 return 0; 254 unsigned ShAmt = Amt->getZExtValue(); 255 // Cannot analyze non-byte shifts. 256 if ((ShAmt & 7) != 0) 257 return 0; 258 ShAmt >>= 3; 259 260 // If the extract is known to be all zeros, return zero. 261 if (ByteStart >= CSize-ShAmt) 262 return Constant::getNullValue(IntegerType::get(CE->getContext(), 263 ByteSize*8)); 264 // If the extract is known to be fully in the input, extract it. 265 if (ByteStart+ByteSize+ShAmt <= CSize) 266 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize); 267 268 // TODO: Handle the 'partially zero' case. 269 return 0; 270 } 271 272 case Instruction::Shl: { 273 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 274 if (Amt == 0) 275 return 0; 276 unsigned ShAmt = Amt->getZExtValue(); 277 // Cannot analyze non-byte shifts. 278 if ((ShAmt & 7) != 0) 279 return 0; 280 ShAmt >>= 3; 281 282 // If the extract is known to be all zeros, return zero. 283 if (ByteStart+ByteSize <= ShAmt) 284 return Constant::getNullValue(IntegerType::get(CE->getContext(), 285 ByteSize*8)); 286 // If the extract is known to be fully in the input, extract it. 287 if (ByteStart >= ShAmt) 288 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize); 289 290 // TODO: Handle the 'partially zero' case. 291 return 0; 292 } 293 294 case Instruction::ZExt: { 295 unsigned SrcBitSize = 296 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 297 298 // If extracting something that is completely zero, return 0. 299 if (ByteStart*8 >= SrcBitSize) 300 return Constant::getNullValue(IntegerType::get(CE->getContext(), 301 ByteSize*8)); 302 303 // If exactly extracting the input, return it. 304 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 305 return CE->getOperand(0); 306 307 // If extracting something completely in the input, if if the input is a 308 // multiple of 8 bits, recurse. 309 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 310 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 311 312 // Otherwise, if extracting a subset of the input, which is not multiple of 313 // 8 bits, do a shift and trunc to get the bits. 314 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 315 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 316 Constant *Res = CE->getOperand(0); 317 if (ByteStart) 318 Res = ConstantExpr::getLShr(Res, 319 ConstantInt::get(Res->getType(), ByteStart*8)); 320 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 321 ByteSize*8)); 322 } 323 324 // TODO: Handle the 'partially zero' case. 325 return 0; 326 } 327 } 328} 329 330/// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof 331/// on Ty, with any known factors factored out. If Folded is false, 332/// return null if no factoring was possible, to avoid endlessly 333/// bouncing an unfoldable expression back into the top-level folder. 334/// 335static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy, 336 bool Folded) { 337 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 338 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements()); 339 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 340 return ConstantExpr::getNUWMul(E, N); 341 } 342 343 if (const StructType *STy = dyn_cast<StructType>(Ty)) 344 if (!STy->isPacked()) { 345 unsigned NumElems = STy->getNumElements(); 346 // An empty struct has size zero. 347 if (NumElems == 0) 348 return ConstantExpr::getNullValue(DestTy); 349 // Check for a struct with all members having the same size. 350 Constant *MemberSize = 351 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 352 bool AllSame = true; 353 for (unsigned i = 1; i != NumElems; ++i) 354 if (MemberSize != 355 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 356 AllSame = false; 357 break; 358 } 359 if (AllSame) { 360 Constant *N = ConstantInt::get(DestTy, NumElems); 361 return ConstantExpr::getNUWMul(MemberSize, N); 362 } 363 } 364 365 // Pointer size doesn't depend on the pointee type, so canonicalize them 366 // to an arbitrary pointee. 367 if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) 368 if (!PTy->getElementType()->isIntegerTy(1)) 369 return 370 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1), 371 PTy->getAddressSpace()), 372 DestTy, true); 373 374 // If there's no interesting folding happening, bail so that we don't create 375 // a constant that looks like it needs folding but really doesn't. 376 if (!Folded) 377 return 0; 378 379 // Base case: Get a regular sizeof expression. 380 Constant *C = ConstantExpr::getSizeOf(Ty); 381 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 382 DestTy, false), 383 C, DestTy); 384 return C; 385} 386 387/// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof 388/// on Ty, with any known factors factored out. If Folded is false, 389/// return null if no factoring was possible, to avoid endlessly 390/// bouncing an unfoldable expression back into the top-level folder. 391/// 392static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy, 393 bool Folded) { 394 // The alignment of an array is equal to the alignment of the 395 // array element. Note that this is not always true for vectors. 396 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 397 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType()); 398 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 399 DestTy, 400 false), 401 C, DestTy); 402 return C; 403 } 404 405 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 406 // Packed structs always have an alignment of 1. 407 if (STy->isPacked()) 408 return ConstantInt::get(DestTy, 1); 409 410 // Otherwise, struct alignment is the maximum alignment of any member. 411 // Without target data, we can't compare much, but we can check to see 412 // if all the members have the same alignment. 413 unsigned NumElems = STy->getNumElements(); 414 // An empty struct has minimal alignment. 415 if (NumElems == 0) 416 return ConstantInt::get(DestTy, 1); 417 // Check for a struct with all members having the same alignment. 418 Constant *MemberAlign = 419 getFoldedAlignOf(STy->getElementType(0), DestTy, true); 420 bool AllSame = true; 421 for (unsigned i = 1; i != NumElems; ++i) 422 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) { 423 AllSame = false; 424 break; 425 } 426 if (AllSame) 427 return MemberAlign; 428 } 429 430 // Pointer alignment doesn't depend on the pointee type, so canonicalize them 431 // to an arbitrary pointee. 432 if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) 433 if (!PTy->getElementType()->isIntegerTy(1)) 434 return 435 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(), 436 1), 437 PTy->getAddressSpace()), 438 DestTy, true); 439 440 // If there's no interesting folding happening, bail so that we don't create 441 // a constant that looks like it needs folding but really doesn't. 442 if (!Folded) 443 return 0; 444 445 // Base case: Get a regular alignof expression. 446 Constant *C = ConstantExpr::getAlignOf(Ty); 447 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 448 DestTy, false), 449 C, DestTy); 450 return C; 451} 452 453/// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof 454/// on Ty and FieldNo, with any known factors factored out. If Folded is false, 455/// return null if no factoring was possible, to avoid endlessly 456/// bouncing an unfoldable expression back into the top-level folder. 457/// 458static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo, 459 const Type *DestTy, 460 bool Folded) { 461 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 462 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false, 463 DestTy, false), 464 FieldNo, DestTy); 465 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true); 466 return ConstantExpr::getNUWMul(E, N); 467 } 468 469 if (const StructType *STy = dyn_cast<StructType>(Ty)) 470 if (!STy->isPacked()) { 471 unsigned NumElems = STy->getNumElements(); 472 // An empty struct has no members. 473 if (NumElems == 0) 474 return 0; 475 // Check for a struct with all members having the same size. 476 Constant *MemberSize = 477 getFoldedSizeOf(STy->getElementType(0), DestTy, true); 478 bool AllSame = true; 479 for (unsigned i = 1; i != NumElems; ++i) 480 if (MemberSize != 481 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) { 482 AllSame = false; 483 break; 484 } 485 if (AllSame) { 486 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, 487 false, 488 DestTy, 489 false), 490 FieldNo, DestTy); 491 return ConstantExpr::getNUWMul(MemberSize, N); 492 } 493 } 494 495 // If there's no interesting folding happening, bail so that we don't create 496 // a constant that looks like it needs folding but really doesn't. 497 if (!Folded) 498 return 0; 499 500 // Base case: Get a regular offsetof expression. 501 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo); 502 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 503 DestTy, false), 504 C, DestTy); 505 return C; 506} 507 508Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 509 const Type *DestTy) { 510 if (isa<UndefValue>(V)) { 511 // zext(undef) = 0, because the top bits will be zero. 512 // sext(undef) = 0, because the top bits will all be the same. 513 // [us]itofp(undef) = 0, because the result value is bounded. 514 if (opc == Instruction::ZExt || opc == Instruction::SExt || 515 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 516 return Constant::getNullValue(DestTy); 517 return UndefValue::get(DestTy); 518 } 519 520 // No compile-time operations on this type yet. 521 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty()) 522 return 0; 523 524 if (V->isNullValue() && !DestTy->isX86_MMXTy()) 525 return Constant::getNullValue(DestTy); 526 527 // If the cast operand is a constant expression, there's a few things we can 528 // do to try to simplify it. 529 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 530 if (CE->isCast()) { 531 // Try hard to fold cast of cast because they are often eliminable. 532 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 533 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 534 } else if (CE->getOpcode() == Instruction::GetElementPtr) { 535 // If all of the indexes in the GEP are null values, there is no pointer 536 // adjustment going on. We might as well cast the source pointer. 537 bool isAllNull = true; 538 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 539 if (!CE->getOperand(i)->isNullValue()) { 540 isAllNull = false; 541 break; 542 } 543 if (isAllNull) 544 // This is casting one pointer type to another, always BitCast 545 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 546 } 547 } 548 549 // If the cast operand is a constant vector, perform the cast by 550 // operating on each element. In the cast of bitcasts, the element 551 // count may be mismatched; don't attempt to handle that here. 552 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) 553 if (DestTy->isVectorTy() && 554 cast<VectorType>(DestTy)->getNumElements() == 555 CV->getType()->getNumElements()) { 556 std::vector<Constant*> res; 557 const VectorType *DestVecTy = cast<VectorType>(DestTy); 558 const Type *DstEltTy = DestVecTy->getElementType(); 559 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i) 560 res.push_back(ConstantExpr::getCast(opc, 561 CV->getOperand(i), DstEltTy)); 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->isFloatTy() ? APFloat::IEEEsingle : 576 DestTy->isDoubleTy() ? APFloat::IEEEdouble : 577 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended : 578 DestTy->isFP128Ty() ? APFloat::IEEEquad : 579 APFloat::Bogus, 580 APFloat::rmNearestTiesToEven, &ignored); 581 return ConstantFP::get(V->getContext(), Val); 582 } 583 return 0; // Can't fold. 584 case Instruction::FPToUI: 585 case Instruction::FPToSI: 586 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 587 const APFloat &V = FPC->getValueAPF(); 588 bool ignored; 589 uint64_t x[2]; 590 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 591 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI, 592 APFloat::rmTowardZero, &ignored); 593 APInt Val(DestBitWidth, 2, x); 594 return ConstantInt::get(FPC->getContext(), Val); 595 } 596 return 0; // Can't fold. 597 case Instruction::IntToPtr: //always treated as unsigned 598 if (V->isNullValue()) // Is it an integral null value? 599 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 600 return 0; // Other pointer types cannot be casted 601 case Instruction::PtrToInt: // always treated as unsigned 602 // Is it a null pointer value? 603 if (V->isNullValue()) 604 return ConstantInt::get(DestTy, 0); 605 // If this is a sizeof-like expression, pull out multiplications by 606 // known factors to expose them to subsequent folding. If it's an 607 // alignof-like expression, factor out known factors. 608 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 609 if (CE->getOpcode() == Instruction::GetElementPtr && 610 CE->getOperand(0)->isNullValue()) { 611 const Type *Ty = 612 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 613 if (CE->getNumOperands() == 2) { 614 // Handle a sizeof-like expression. 615 Constant *Idx = CE->getOperand(1); 616 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne(); 617 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) { 618 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true, 619 DestTy, false), 620 Idx, DestTy); 621 return ConstantExpr::getMul(C, Idx); 622 } 623 } else if (CE->getNumOperands() == 3 && 624 CE->getOperand(1)->isNullValue()) { 625 // Handle an alignof-like expression. 626 if (const StructType *STy = dyn_cast<StructType>(Ty)) 627 if (!STy->isPacked()) { 628 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2)); 629 if (CI->isOne() && 630 STy->getNumElements() == 2 && 631 STy->getElementType(0)->isIntegerTy(1)) { 632 return getFoldedAlignOf(STy->getElementType(1), DestTy, false); 633 } 634 } 635 // Handle an offsetof-like expression. 636 if (Ty->isStructTy() || Ty->isArrayTy()) { 637 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2), 638 DestTy, false)) 639 return C; 640 } 641 } 642 } 643 // Other pointer types cannot be casted 644 return 0; 645 case Instruction::UIToFP: 646 case Instruction::SIToFP: 647 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 648 APInt api = CI->getValue(); 649 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true); 650 (void)apf.convertFromAPInt(api, 651 opc==Instruction::SIToFP, 652 APFloat::rmNearestTiesToEven); 653 return ConstantFP::get(V->getContext(), apf); 654 } 655 return 0; 656 case Instruction::ZExt: 657 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 658 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 659 return ConstantInt::get(V->getContext(), 660 CI->getValue().zext(BitWidth)); 661 } 662 return 0; 663 case Instruction::SExt: 664 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 665 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 666 return ConstantInt::get(V->getContext(), 667 CI->getValue().sext(BitWidth)); 668 } 669 return 0; 670 case Instruction::Trunc: { 671 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 672 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 673 return ConstantInt::get(V->getContext(), 674 CI->getValue().trunc(DestBitWidth)); 675 } 676 677 // The input must be a constantexpr. See if we can simplify this based on 678 // the bytes we are demanding. Only do this if the source and dest are an 679 // even multiple of a byte. 680 if ((DestBitWidth & 7) == 0 && 681 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 682 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 683 return Res; 684 685 return 0; 686 } 687 case Instruction::BitCast: 688 return FoldBitCast(V, DestTy); 689 } 690} 691 692Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 693 Constant *V1, Constant *V2) { 694 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond)) 695 return CB->getZExtValue() ? V1 : V2; 696 697 // Check for zero aggregate and ConstantVector of zeros 698 if (Cond->isNullValue()) return V2; 699 700 if (ConstantVector* CondV = dyn_cast<ConstantVector>(Cond)) { 701 702 if (CondV->isAllOnesValue()) return V1; 703 704 const VectorType *VTy = cast<VectorType>(V1->getType()); 705 ConstantVector *CP1 = dyn_cast<ConstantVector>(V1); 706 ConstantVector *CP2 = dyn_cast<ConstantVector>(V2); 707 708 if ((CP1 || isa<ConstantAggregateZero>(V1)) && 709 (CP2 || isa<ConstantAggregateZero>(V2))) { 710 711 // Find the element type of the returned vector 712 const Type *EltTy = VTy->getElementType(); 713 unsigned NumElem = VTy->getNumElements(); 714 std::vector<Constant*> Res(NumElem); 715 716 bool Valid = true; 717 for (unsigned i = 0; i < NumElem; ++i) { 718 ConstantInt* c = dyn_cast<ConstantInt>(CondV->getOperand(i)); 719 if (!c) { 720 Valid = false; 721 break; 722 } 723 Constant *C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 724 Constant *C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 725 Res[i] = c->getZExtValue() ? C1 : C2; 726 } 727 // If we were able to build the vector, return it 728 if (Valid) return ConstantVector::get(Res); 729 } 730 } 731 732 733 if (isa<UndefValue>(V1)) return V2; 734 if (isa<UndefValue>(V2)) return V1; 735 if (isa<UndefValue>(Cond)) return V1; 736 if (V1 == V2) return V1; 737 738 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) { 739 if (TrueVal->getOpcode() == Instruction::Select) 740 if (TrueVal->getOperand(0) == Cond) 741 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2); 742 } 743 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) { 744 if (FalseVal->getOpcode() == Instruction::Select) 745 if (FalseVal->getOperand(0) == Cond) 746 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2)); 747 } 748 749 return 0; 750} 751 752Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 753 Constant *Idx) { 754 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef 755 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType()); 756 if (Val->isNullValue()) // ee(zero, x) -> zero 757 return Constant::getNullValue( 758 cast<VectorType>(Val->getType())->getElementType()); 759 760 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 761 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) { 762 return CVal->getOperand(CIdx->getZExtValue()); 763 } else if (isa<UndefValue>(Idx)) { 764 // ee({w,x,y,z}, undef) -> w (an arbitrary value). 765 return CVal->getOperand(0); 766 } 767 } 768 return 0; 769} 770 771Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 772 Constant *Elt, 773 Constant *Idx) { 774 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 775 if (!CIdx) return 0; 776 APInt idxVal = CIdx->getValue(); 777 if (isa<UndefValue>(Val)) { 778 // Insertion of scalar constant into vector undef 779 // Optimize away insertion of undef 780 if (isa<UndefValue>(Elt)) 781 return Val; 782 // Otherwise break the aggregate undef into multiple undefs and do 783 // the insertion 784 unsigned numOps = 785 cast<VectorType>(Val->getType())->getNumElements(); 786 std::vector<Constant*> Ops; 787 Ops.reserve(numOps); 788 for (unsigned i = 0; i < numOps; ++i) { 789 Constant *Op = 790 (idxVal == i) ? Elt : UndefValue::get(Elt->getType()); 791 Ops.push_back(Op); 792 } 793 return ConstantVector::get(Ops); 794 } 795 if (isa<ConstantAggregateZero>(Val)) { 796 // Insertion of scalar constant into vector aggregate zero 797 // Optimize away insertion of zero 798 if (Elt->isNullValue()) 799 return Val; 800 // Otherwise break the aggregate zero into multiple zeros and do 801 // the insertion 802 unsigned numOps = 803 cast<VectorType>(Val->getType())->getNumElements(); 804 std::vector<Constant*> Ops; 805 Ops.reserve(numOps); 806 for (unsigned i = 0; i < numOps; ++i) { 807 Constant *Op = 808 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType()); 809 Ops.push_back(Op); 810 } 811 return ConstantVector::get(Ops); 812 } 813 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) { 814 // Insertion of scalar constant into vector constant 815 std::vector<Constant*> Ops; 816 Ops.reserve(CVal->getNumOperands()); 817 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) { 818 Constant *Op = 819 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i)); 820 Ops.push_back(Op); 821 } 822 return ConstantVector::get(Ops); 823 } 824 825 return 0; 826} 827 828/// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef 829/// return the specified element value. Otherwise return null. 830static Constant *GetVectorElement(Constant *C, unsigned EltNo) { 831 if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) 832 return CV->getOperand(EltNo); 833 834 const Type *EltTy = cast<VectorType>(C->getType())->getElementType(); 835 if (isa<ConstantAggregateZero>(C)) 836 return Constant::getNullValue(EltTy); 837 if (isa<UndefValue>(C)) 838 return UndefValue::get(EltTy); 839 return 0; 840} 841 842Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, 843 Constant *V2, 844 Constant *Mask) { 845 // Undefined shuffle mask -> undefined value. 846 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType()); 847 848 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements(); 849 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements(); 850 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType(); 851 852 // Loop over the shuffle mask, evaluating each element. 853 SmallVector<Constant*, 32> Result; 854 for (unsigned i = 0; i != MaskNumElts; ++i) { 855 Constant *InElt = GetVectorElement(Mask, i); 856 if (InElt == 0) return 0; 857 858 if (isa<UndefValue>(InElt)) 859 InElt = UndefValue::get(EltTy); 860 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) { 861 unsigned Elt = CI->getZExtValue(); 862 if (Elt >= SrcNumElts*2) 863 InElt = UndefValue::get(EltTy); 864 else if (Elt >= SrcNumElts) 865 InElt = GetVectorElement(V2, Elt - SrcNumElts); 866 else 867 InElt = GetVectorElement(V1, Elt); 868 if (InElt == 0) return 0; 869 } else { 870 // Unknown value. 871 return 0; 872 } 873 Result.push_back(InElt); 874 } 875 876 return ConstantVector::get(Result); 877} 878 879Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 880 const unsigned *Idxs, 881 unsigned NumIdx) { 882 // Base case: no indices, so return the entire value. 883 if (NumIdx == 0) 884 return Agg; 885 886 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef 887 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(), 888 Idxs, 889 Idxs + NumIdx)); 890 891 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0 892 return 893 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(), 894 Idxs, 895 Idxs + NumIdx)); 896 897 // Otherwise recurse. 898 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) 899 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs), 900 Idxs+1, NumIdx-1); 901 902 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) 903 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs), 904 Idxs+1, NumIdx-1); 905 ConstantVector *CV = cast<ConstantVector>(Agg); 906 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs), 907 Idxs+1, NumIdx-1); 908} 909 910Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 911 Constant *Val, 912 const unsigned *Idxs, 913 unsigned NumIdx) { 914 // Base case: no indices, so replace the entire value. 915 if (NumIdx == 0) 916 return Val; 917 918 if (isa<UndefValue>(Agg)) { 919 // Insertion of constant into aggregate undef 920 // Optimize away insertion of undef. 921 if (isa<UndefValue>(Val)) 922 return Agg; 923 924 // Otherwise break the aggregate undef into multiple undefs and do 925 // the insertion. 926 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 927 unsigned numOps; 928 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 929 numOps = AR->getNumElements(); 930 else 931 numOps = cast<StructType>(AggTy)->getNumElements(); 932 933 std::vector<Constant*> Ops(numOps); 934 for (unsigned i = 0; i < numOps; ++i) { 935 const Type *MemberTy = AggTy->getTypeAtIndex(i); 936 Constant *Op = 937 (*Idxs == i) ? 938 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy), 939 Val, Idxs+1, NumIdx-1) : 940 UndefValue::get(MemberTy); 941 Ops[i] = Op; 942 } 943 944 if (const StructType* ST = dyn_cast<StructType>(AggTy)) 945 return ConstantStruct::get(ST, Ops); 946 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 947 } 948 949 if (isa<ConstantAggregateZero>(Agg)) { 950 // Insertion of constant into aggregate zero 951 // Optimize away insertion of zero. 952 if (Val->isNullValue()) 953 return Agg; 954 955 // Otherwise break the aggregate zero into multiple zeros and do 956 // the insertion. 957 const CompositeType *AggTy = cast<CompositeType>(Agg->getType()); 958 unsigned numOps; 959 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy)) 960 numOps = AR->getNumElements(); 961 else 962 numOps = cast<StructType>(AggTy)->getNumElements(); 963 964 std::vector<Constant*> Ops(numOps); 965 for (unsigned i = 0; i < numOps; ++i) { 966 const Type *MemberTy = AggTy->getTypeAtIndex(i); 967 Constant *Op = 968 (*Idxs == i) ? 969 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy), 970 Val, Idxs+1, NumIdx-1) : 971 Constant::getNullValue(MemberTy); 972 Ops[i] = Op; 973 } 974 975 if (const StructType *ST = dyn_cast<StructType>(AggTy)) 976 return ConstantStruct::get(ST, Ops); 977 return ConstantArray::get(cast<ArrayType>(AggTy), Ops); 978 } 979 980 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) { 981 // Insertion of constant into aggregate constant. 982 std::vector<Constant*> Ops(Agg->getNumOperands()); 983 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) { 984 Constant *Op = cast<Constant>(Agg->getOperand(i)); 985 if (*Idxs == i) 986 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1); 987 Ops[i] = Op; 988 } 989 990 if (const StructType* ST = dyn_cast<StructType>(Agg->getType())) 991 return ConstantStruct::get(ST, Ops); 992 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops); 993 } 994 995 return 0; 996} 997 998 999Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, 1000 Constant *C1, Constant *C2) { 1001 // No compile-time operations on this type yet. 1002 if (C1->getType()->isPPC_FP128Ty()) 1003 return 0; 1004 1005 // Handle UndefValue up front. 1006 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1007 switch (Opcode) { 1008 case Instruction::Xor: 1009 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 1010 // Handle undef ^ undef -> 0 special case. This is a common 1011 // idiom (misuse). 1012 return Constant::getNullValue(C1->getType()); 1013 // Fallthrough 1014 case Instruction::Add: 1015 case Instruction::Sub: 1016 return UndefValue::get(C1->getType()); 1017 case Instruction::And: 1018 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 1019 return C1; 1020 return Constant::getNullValue(C1->getType()); // undef & X -> 0 1021 case Instruction::Mul: { 1022 ConstantInt *CI; 1023 // X * undef -> undef if X is odd or undef 1024 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) || 1025 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) || 1026 (isa<UndefValue>(C1) && isa<UndefValue>(C2))) 1027 return UndefValue::get(C1->getType()); 1028 1029 // X * undef -> 0 otherwise 1030 return Constant::getNullValue(C1->getType()); 1031 } 1032 case Instruction::UDiv: 1033 case Instruction::SDiv: 1034 // undef / 1 -> undef 1035 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv) 1036 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) 1037 if (CI2->isOne()) 1038 return C1; 1039 // FALL THROUGH 1040 case Instruction::URem: 1041 case Instruction::SRem: 1042 if (!isa<UndefValue>(C2)) // undef / X -> 0 1043 return Constant::getNullValue(C1->getType()); 1044 return C2; // X / undef -> undef 1045 case Instruction::Or: // X | undef -> -1 1046 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 1047 return C1; 1048 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 1049 case Instruction::LShr: 1050 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 1051 return C1; // undef lshr undef -> undef 1052 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0 1053 // undef lshr X -> 0 1054 case Instruction::AShr: 1055 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones 1056 return Constant::getAllOnesValue(C1->getType()); 1057 else if (isa<UndefValue>(C1)) 1058 return C1; // undef ashr undef -> undef 1059 else 1060 return C1; // X ashr undef --> X 1061 case Instruction::Shl: 1062 if (isa<UndefValue>(C2) && isa<UndefValue>(C1)) 1063 return C1; // undef shl undef -> undef 1064 // undef << X -> 0 or X << undef -> 0 1065 return Constant::getNullValue(C1->getType()); 1066 } 1067 } 1068 1069 // Handle simplifications when the RHS is a constant int. 1070 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1071 switch (Opcode) { 1072 case Instruction::Add: 1073 if (CI2->equalsInt(0)) return C1; // X + 0 == X 1074 break; 1075 case Instruction::Sub: 1076 if (CI2->equalsInt(0)) return C1; // X - 0 == X 1077 break; 1078 case Instruction::Mul: 1079 if (CI2->equalsInt(0)) return C2; // X * 0 == 0 1080 if (CI2->equalsInt(1)) 1081 return C1; // X * 1 == X 1082 break; 1083 case Instruction::UDiv: 1084 case Instruction::SDiv: 1085 if (CI2->equalsInt(1)) 1086 return C1; // X / 1 == X 1087 if (CI2->equalsInt(0)) 1088 return UndefValue::get(CI2->getType()); // X / 0 == undef 1089 break; 1090 case Instruction::URem: 1091 case Instruction::SRem: 1092 if (CI2->equalsInt(1)) 1093 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 1094 if (CI2->equalsInt(0)) 1095 return UndefValue::get(CI2->getType()); // X % 0 == undef 1096 break; 1097 case Instruction::And: 1098 if (CI2->isZero()) return C2; // X & 0 == 0 1099 if (CI2->isAllOnesValue()) 1100 return C1; // X & -1 == X 1101 1102 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1103 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1104 if (CE1->getOpcode() == Instruction::ZExt) { 1105 unsigned DstWidth = CI2->getType()->getBitWidth(); 1106 unsigned SrcWidth = 1107 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1108 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1109 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1110 return C1; 1111 } 1112 1113 // If and'ing the address of a global with a constant, fold it. 1114 if (CE1->getOpcode() == Instruction::PtrToInt && 1115 isa<GlobalValue>(CE1->getOperand(0))) { 1116 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1117 1118 // Functions are at least 4-byte aligned. 1119 unsigned GVAlign = GV->getAlignment(); 1120 if (isa<Function>(GV)) 1121 GVAlign = std::max(GVAlign, 4U); 1122 1123 if (GVAlign > 1) { 1124 unsigned DstWidth = CI2->getType()->getBitWidth(); 1125 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign)); 1126 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1127 1128 // If checking bits we know are clear, return zero. 1129 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1130 return Constant::getNullValue(CI2->getType()); 1131 } 1132 } 1133 } 1134 break; 1135 case Instruction::Or: 1136 if (CI2->equalsInt(0)) return C1; // X | 0 == X 1137 if (CI2->isAllOnesValue()) 1138 return C2; // X | -1 == -1 1139 break; 1140 case Instruction::Xor: 1141 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X 1142 1143 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1144 switch (CE1->getOpcode()) { 1145 default: break; 1146 case Instruction::ICmp: 1147 case Instruction::FCmp: 1148 // cmp pred ^ true -> cmp !pred 1149 assert(CI2->equalsInt(1)); 1150 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1151 pred = CmpInst::getInversePredicate(pred); 1152 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1153 CE1->getOperand(1)); 1154 } 1155 } 1156 break; 1157 case Instruction::AShr: 1158 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1159 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1160 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1161 return ConstantExpr::getLShr(C1, C2); 1162 break; 1163 } 1164 } else if (isa<ConstantInt>(C1)) { 1165 // If C1 is a ConstantInt and C2 is not, swap the operands. 1166 if (Instruction::isCommutative(Opcode)) 1167 return ConstantExpr::get(Opcode, C2, C1); 1168 } 1169 1170 // At this point we know neither constant is an UndefValue. 1171 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1172 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1173 using namespace APIntOps; 1174 const APInt &C1V = CI1->getValue(); 1175 const APInt &C2V = CI2->getValue(); 1176 switch (Opcode) { 1177 default: 1178 break; 1179 case Instruction::Add: 1180 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1181 case Instruction::Sub: 1182 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1183 case Instruction::Mul: 1184 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1185 case Instruction::UDiv: 1186 assert(!CI2->isNullValue() && "Div by zero handled above"); 1187 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1188 case Instruction::SDiv: 1189 assert(!CI2->isNullValue() && "Div by zero handled above"); 1190 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1191 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef 1192 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1193 case Instruction::URem: 1194 assert(!CI2->isNullValue() && "Div by zero handled above"); 1195 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1196 case Instruction::SRem: 1197 assert(!CI2->isNullValue() && "Div by zero handled above"); 1198 if (C2V.isAllOnesValue() && C1V.isMinSignedValue()) 1199 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef 1200 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1201 case Instruction::And: 1202 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1203 case Instruction::Or: 1204 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1205 case Instruction::Xor: 1206 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1207 case Instruction::Shl: { 1208 uint32_t shiftAmt = C2V.getZExtValue(); 1209 if (shiftAmt < C1V.getBitWidth()) 1210 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt)); 1211 else 1212 return UndefValue::get(C1->getType()); // too big shift is undef 1213 } 1214 case Instruction::LShr: { 1215 uint32_t shiftAmt = C2V.getZExtValue(); 1216 if (shiftAmt < C1V.getBitWidth()) 1217 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt)); 1218 else 1219 return UndefValue::get(C1->getType()); // too big shift is undef 1220 } 1221 case Instruction::AShr: { 1222 uint32_t shiftAmt = C2V.getZExtValue(); 1223 if (shiftAmt < C1V.getBitWidth()) 1224 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt)); 1225 else 1226 return UndefValue::get(C1->getType()); // too big shift is undef 1227 } 1228 } 1229 } 1230 1231 switch (Opcode) { 1232 case Instruction::SDiv: 1233 case Instruction::UDiv: 1234 case Instruction::URem: 1235 case Instruction::SRem: 1236 case Instruction::LShr: 1237 case Instruction::AShr: 1238 case Instruction::Shl: 1239 if (CI1->equalsInt(0)) return C1; 1240 break; 1241 default: 1242 break; 1243 } 1244 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1245 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1246 APFloat C1V = CFP1->getValueAPF(); 1247 APFloat C2V = CFP2->getValueAPF(); 1248 APFloat C3V = C1V; // copy for modification 1249 switch (Opcode) { 1250 default: 1251 break; 1252 case Instruction::FAdd: 1253 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1254 return ConstantFP::get(C1->getContext(), C3V); 1255 case Instruction::FSub: 1256 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1257 return ConstantFP::get(C1->getContext(), C3V); 1258 case Instruction::FMul: 1259 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1260 return ConstantFP::get(C1->getContext(), C3V); 1261 case Instruction::FDiv: 1262 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1263 return ConstantFP::get(C1->getContext(), C3V); 1264 case Instruction::FRem: 1265 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven); 1266 return ConstantFP::get(C1->getContext(), C3V); 1267 } 1268 } 1269 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) { 1270 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1); 1271 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2); 1272 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) && 1273 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) { 1274 std::vector<Constant*> Res; 1275 const Type* EltTy = VTy->getElementType(); 1276 Constant *C1 = 0; 1277 Constant *C2 = 0; 1278 switch (Opcode) { 1279 default: 1280 break; 1281 case Instruction::Add: 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::getAdd(C1, C2)); 1286 } 1287 return ConstantVector::get(Res); 1288 case Instruction::FAdd: 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::getFAdd(C1, C2)); 1293 } 1294 return ConstantVector::get(Res); 1295 case Instruction::Sub: 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::getSub(C1, C2)); 1300 } 1301 return ConstantVector::get(Res); 1302 case Instruction::FSub: 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::getFSub(C1, C2)); 1307 } 1308 return ConstantVector::get(Res); 1309 case Instruction::Mul: 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::getMul(C1, C2)); 1314 } 1315 return ConstantVector::get(Res); 1316 case Instruction::FMul: 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::getFMul(C1, C2)); 1321 } 1322 return ConstantVector::get(Res); 1323 case Instruction::UDiv: 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::getUDiv(C1, C2)); 1328 } 1329 return ConstantVector::get(Res); 1330 case Instruction::SDiv: 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::getSDiv(C1, C2)); 1335 } 1336 return ConstantVector::get(Res); 1337 case Instruction::FDiv: 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::getFDiv(C1, C2)); 1342 } 1343 return ConstantVector::get(Res); 1344 case Instruction::URem: 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::getURem(C1, C2)); 1349 } 1350 return ConstantVector::get(Res); 1351 case Instruction::SRem: 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::getSRem(C1, C2)); 1356 } 1357 return ConstantVector::get(Res); 1358 case Instruction::FRem: 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::getFRem(C1, C2)); 1363 } 1364 return ConstantVector::get(Res); 1365 case Instruction::And: 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::getAnd(C1, C2)); 1370 } 1371 return ConstantVector::get(Res); 1372 case Instruction::Or: 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::getOr(C1, C2)); 1377 } 1378 return ConstantVector::get(Res); 1379 case Instruction::Xor: 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::getXor(C1, C2)); 1384 } 1385 return ConstantVector::get(Res); 1386 case Instruction::LShr: 1387 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1388 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1389 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1390 Res.push_back(ConstantExpr::getLShr(C1, C2)); 1391 } 1392 return ConstantVector::get(Res); 1393 case Instruction::AShr: 1394 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1395 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1396 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1397 Res.push_back(ConstantExpr::getAShr(C1, C2)); 1398 } 1399 return ConstantVector::get(Res); 1400 case Instruction::Shl: 1401 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 1402 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy); 1403 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy); 1404 Res.push_back(ConstantExpr::getShl(C1, C2)); 1405 } 1406 return ConstantVector::get(Res); 1407 } 1408 } 1409 } 1410 1411 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1412 // There are many possible foldings we could do here. We should probably 1413 // at least fold add of a pointer with an integer into the appropriate 1414 // getelementptr. This will improve alias analysis a bit. 1415 1416 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1417 // (a + (b + c)). 1418 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1419 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1420 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1421 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1422 } 1423 } else if (isa<ConstantExpr>(C2)) { 1424 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1425 // other way if possible. 1426 if (Instruction::isCommutative(Opcode)) 1427 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1428 } 1429 1430 // i1 can be simplified in many cases. 1431 if (C1->getType()->isIntegerTy(1)) { 1432 switch (Opcode) { 1433 case Instruction::Add: 1434 case Instruction::Sub: 1435 return ConstantExpr::getXor(C1, C2); 1436 case Instruction::Mul: 1437 return ConstantExpr::getAnd(C1, C2); 1438 case Instruction::Shl: 1439 case Instruction::LShr: 1440 case Instruction::AShr: 1441 // We can assume that C2 == 0. If it were one the result would be 1442 // undefined because the shift value is as large as the bitwidth. 1443 return C1; 1444 case Instruction::SDiv: 1445 case Instruction::UDiv: 1446 // We can assume that C2 == 1. If it were zero the result would be 1447 // undefined through division by zero. 1448 return C1; 1449 case Instruction::URem: 1450 case Instruction::SRem: 1451 // We can assume that C2 == 1. If it were zero the result would be 1452 // undefined through division by zero. 1453 return ConstantInt::getFalse(C1->getContext()); 1454 default: 1455 break; 1456 } 1457 } 1458 1459 // We don't know how to fold this. 1460 return 0; 1461} 1462 1463/// isZeroSizedType - This type is zero sized if its an array or structure of 1464/// zero sized types. The only leaf zero sized type is an empty structure. 1465static bool isMaybeZeroSizedType(const Type *Ty) { 1466 if (Ty->isOpaqueTy()) return true; // Can't say. 1467 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1468 1469 // If all of elements have zero size, this does too. 1470 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 1471 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false; 1472 return true; 1473 1474 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1475 return isMaybeZeroSizedType(ATy->getElementType()); 1476 } 1477 return false; 1478} 1479 1480/// IdxCompare - Compare the two constants as though they were getelementptr 1481/// indices. This allows coersion of the types to be the same thing. 1482/// 1483/// If the two constants are the "same" (after coersion), return 0. If the 1484/// first is less than the second, return -1, if the second is less than the 1485/// first, return 1. If the constants are not integral, return -2. 1486/// 1487static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) { 1488 if (C1 == C2) return 0; 1489 1490 // Ok, we found a different index. If they are not ConstantInt, we can't do 1491 // anything with them. 1492 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2)) 1493 return -2; // don't know! 1494 1495 // Ok, we have two differing integer indices. Sign extend them to be the same 1496 // type. Long is always big enough, so we use it. 1497 if (!C1->getType()->isIntegerTy(64)) 1498 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext())); 1499 1500 if (!C2->getType()->isIntegerTy(64)) 1501 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext())); 1502 1503 if (C1 == C2) return 0; // They are equal 1504 1505 // If the type being indexed over is really just a zero sized type, there is 1506 // no pointer difference being made here. 1507 if (isMaybeZeroSizedType(ElTy)) 1508 return -2; // dunno. 1509 1510 // If they are really different, now that they are the same type, then we 1511 // found a difference! 1512 if (cast<ConstantInt>(C1)->getSExtValue() < 1513 cast<ConstantInt>(C2)->getSExtValue()) 1514 return -1; 1515 else 1516 return 1; 1517} 1518 1519/// evaluateFCmpRelation - This function determines if there is anything we can 1520/// decide about the two constants provided. This doesn't need to handle simple 1521/// things like ConstantFP comparisons, but should instead handle ConstantExprs. 1522/// If we can determine that the two constants have a particular relation to 1523/// each other, we should return the corresponding FCmpInst predicate, 1524/// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1525/// ConstantFoldCompareInstruction. 1526/// 1527/// To simplify this code we canonicalize the relation so that the first 1528/// operand is always the most "complex" of the two. We consider ConstantFP 1529/// to be the simplest, and ConstantExprs to be the most complex. 1530static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1531 assert(V1->getType() == V2->getType() && 1532 "Cannot compare values of different types!"); 1533 1534 // No compile-time operations on this type yet. 1535 if (V1->getType()->isPPC_FP128Ty()) 1536 return FCmpInst::BAD_FCMP_PREDICATE; 1537 1538 // Handle degenerate case quickly 1539 if (V1 == V2) return FCmpInst::FCMP_OEQ; 1540 1541 if (!isa<ConstantExpr>(V1)) { 1542 if (!isa<ConstantExpr>(V2)) { 1543 // We distilled thisUse the standard constant folder for a few cases 1544 ConstantInt *R = 0; 1545 R = dyn_cast<ConstantInt>( 1546 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1547 if (R && !R->isZero()) 1548 return FCmpInst::FCMP_OEQ; 1549 R = dyn_cast<ConstantInt>( 1550 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1551 if (R && !R->isZero()) 1552 return FCmpInst::FCMP_OLT; 1553 R = dyn_cast<ConstantInt>( 1554 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1555 if (R && !R->isZero()) 1556 return FCmpInst::FCMP_OGT; 1557 1558 // Nothing more we can do 1559 return FCmpInst::BAD_FCMP_PREDICATE; 1560 } 1561 1562 // If the first operand is simple and second is ConstantExpr, swap operands. 1563 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1564 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1565 return FCmpInst::getSwappedPredicate(SwappedRelation); 1566 } else { 1567 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1568 // constantexpr or a simple constant. 1569 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1570 switch (CE1->getOpcode()) { 1571 case Instruction::FPTrunc: 1572 case Instruction::FPExt: 1573 case Instruction::UIToFP: 1574 case Instruction::SIToFP: 1575 // We might be able to do something with these but we don't right now. 1576 break; 1577 default: 1578 break; 1579 } 1580 } 1581 // There are MANY other foldings that we could perform here. They will 1582 // probably be added on demand, as they seem needed. 1583 return FCmpInst::BAD_FCMP_PREDICATE; 1584} 1585 1586/// evaluateICmpRelation - This function determines if there is anything we can 1587/// decide about the two constants provided. This doesn't need to handle simple 1588/// things like integer comparisons, but should instead handle ConstantExprs 1589/// and GlobalValues. If we can determine that the two constants have a 1590/// particular relation to each other, we should return the corresponding ICmp 1591/// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE. 1592/// 1593/// To simplify this code we canonicalize the relation so that the first 1594/// operand is always the most "complex" of the two. We consider simple 1595/// constants (like ConstantInt) to be the simplest, followed by 1596/// GlobalValues, followed by ConstantExpr's (the most complex). 1597/// 1598static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1599 bool isSigned) { 1600 assert(V1->getType() == V2->getType() && 1601 "Cannot compare different types of values!"); 1602 if (V1 == V2) return ICmpInst::ICMP_EQ; 1603 1604 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1605 !isa<BlockAddress>(V1)) { 1606 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1607 !isa<BlockAddress>(V2)) { 1608 // We distilled this down to a simple case, use the standard constant 1609 // folder. 1610 ConstantInt *R = 0; 1611 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1612 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1613 if (R && !R->isZero()) 1614 return pred; 1615 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1616 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1617 if (R && !R->isZero()) 1618 return pred; 1619 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1620 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1621 if (R && !R->isZero()) 1622 return pred; 1623 1624 // If we couldn't figure it out, bail. 1625 return ICmpInst::BAD_ICMP_PREDICATE; 1626 } 1627 1628 // If the first operand is simple, swap operands. 1629 ICmpInst::Predicate SwappedRelation = 1630 evaluateICmpRelation(V2, V1, isSigned); 1631 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1632 return ICmpInst::getSwappedPredicate(SwappedRelation); 1633 1634 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1635 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1636 ICmpInst::Predicate SwappedRelation = 1637 evaluateICmpRelation(V2, V1, isSigned); 1638 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1639 return ICmpInst::getSwappedPredicate(SwappedRelation); 1640 return ICmpInst::BAD_ICMP_PREDICATE; 1641 } 1642 1643 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1644 // constant (which, since the types must match, means that it's a 1645 // ConstantPointerNull). 1646 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1647 // Don't try to decide equality of aliases. 1648 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2)) 1649 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage()) 1650 return ICmpInst::ICMP_NE; 1651 } else if (isa<BlockAddress>(V2)) { 1652 return ICmpInst::ICMP_NE; // Globals never equal labels. 1653 } else { 1654 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1655 // GlobalVals can never be null unless they have external weak linkage. 1656 // We don't try to evaluate aliases here. 1657 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV)) 1658 return ICmpInst::ICMP_NE; 1659 } 1660 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1661 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1662 ICmpInst::Predicate SwappedRelation = 1663 evaluateICmpRelation(V2, V1, isSigned); 1664 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1665 return ICmpInst::getSwappedPredicate(SwappedRelation); 1666 return ICmpInst::BAD_ICMP_PREDICATE; 1667 } 1668 1669 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1670 // constant (which, since the types must match, means that it is a 1671 // ConstantPointerNull). 1672 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1673 // Block address in another function can't equal this one, but block 1674 // addresses in the current function might be the same if blocks are 1675 // empty. 1676 if (BA2->getFunction() != BA->getFunction()) 1677 return ICmpInst::ICMP_NE; 1678 } else { 1679 // Block addresses aren't null, don't equal the address of globals. 1680 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1681 "Canonicalization guarantee!"); 1682 return ICmpInst::ICMP_NE; 1683 } 1684 } else { 1685 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1686 // constantexpr, a global, block address, or a simple constant. 1687 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1688 Constant *CE1Op0 = CE1->getOperand(0); 1689 1690 switch (CE1->getOpcode()) { 1691 case Instruction::Trunc: 1692 case Instruction::FPTrunc: 1693 case Instruction::FPExt: 1694 case Instruction::FPToUI: 1695 case Instruction::FPToSI: 1696 break; // We can't evaluate floating point casts or truncations. 1697 1698 case Instruction::UIToFP: 1699 case Instruction::SIToFP: 1700 case Instruction::BitCast: 1701 case Instruction::ZExt: 1702 case Instruction::SExt: 1703 // If the cast is not actually changing bits, and the second operand is a 1704 // null pointer, do the comparison with the pre-casted value. 1705 if (V2->isNullValue() && 1706 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) { 1707 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1708 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1709 return evaluateICmpRelation(CE1Op0, 1710 Constant::getNullValue(CE1Op0->getType()), 1711 isSigned); 1712 } 1713 break; 1714 1715 case Instruction::GetElementPtr: 1716 // Ok, since this is a getelementptr, we know that the constant has a 1717 // pointer type. Check the various cases. 1718 if (isa<ConstantPointerNull>(V2)) { 1719 // If we are comparing a GEP to a null pointer, check to see if the base 1720 // of the GEP equals the null pointer. 1721 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1722 if (GV->hasExternalWeakLinkage()) 1723 // Weak linkage GVals could be zero or not. We're comparing that 1724 // to null pointer so its greater-or-equal 1725 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1726 else 1727 // If its not weak linkage, the GVal must have a non-zero address 1728 // so the result is greater-than 1729 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1730 } else if (isa<ConstantPointerNull>(CE1Op0)) { 1731 // If we are indexing from a null pointer, check to see if we have any 1732 // non-zero indices. 1733 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i) 1734 if (!CE1->getOperand(i)->isNullValue()) 1735 // Offsetting from null, must not be equal. 1736 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1737 // Only zero indexes from null, must still be zero. 1738 return ICmpInst::ICMP_EQ; 1739 } 1740 // Otherwise, we can't really say if the first operand is null or not. 1741 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1742 if (isa<ConstantPointerNull>(CE1Op0)) { 1743 if (GV2->hasExternalWeakLinkage()) 1744 // Weak linkage GVals could be zero or not. We're comparing it to 1745 // a null pointer, so its less-or-equal 1746 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1747 else 1748 // If its not weak linkage, the GVal must have a non-zero address 1749 // so the result is less-than 1750 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1751 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1752 if (GV == GV2) { 1753 // If this is a getelementptr of the same global, then it must be 1754 // different. Because the types must match, the getelementptr could 1755 // only have at most one index, and because we fold getelementptr's 1756 // with a single zero index, it must be nonzero. 1757 assert(CE1->getNumOperands() == 2 && 1758 !CE1->getOperand(1)->isNullValue() && 1759 "Surprising getelementptr!"); 1760 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1761 } else { 1762 // If they are different globals, we don't know what the value is, 1763 // but they can't be equal. 1764 return ICmpInst::ICMP_NE; 1765 } 1766 } 1767 } else { 1768 ConstantExpr *CE2 = cast<ConstantExpr>(V2); 1769 Constant *CE2Op0 = CE2->getOperand(0); 1770 1771 // There are MANY other foldings that we could perform here. They will 1772 // probably be added on demand, as they seem needed. 1773 switch (CE2->getOpcode()) { 1774 default: break; 1775 case Instruction::GetElementPtr: 1776 // By far the most common case to handle is when the base pointers are 1777 // obviously to the same or different globals. 1778 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1779 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal 1780 return ICmpInst::ICMP_NE; 1781 // Ok, we know that both getelementptr instructions are based on the 1782 // same global. From this, we can precisely determine the relative 1783 // ordering of the resultant pointers. 1784 unsigned i = 1; 1785 1786 // The logic below assumes that the result of the comparison 1787 // can be determined by finding the first index that differs. 1788 // This doesn't work if there is over-indexing in any 1789 // subsequent indices, so check for that case first. 1790 if (!CE1->isGEPWithNoNotionalOverIndexing() || 1791 !CE2->isGEPWithNoNotionalOverIndexing()) 1792 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1793 1794 // Compare all of the operands the GEP's have in common. 1795 gep_type_iterator GTI = gep_type_begin(CE1); 1796 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands(); 1797 ++i, ++GTI) 1798 switch (IdxCompare(CE1->getOperand(i), 1799 CE2->getOperand(i), GTI.getIndexedType())) { 1800 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT; 1801 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT; 1802 case -2: return ICmpInst::BAD_ICMP_PREDICATE; 1803 } 1804 1805 // Ok, we ran out of things they have in common. If any leftovers 1806 // are non-zero then we have a difference, otherwise we are equal. 1807 for (; i < CE1->getNumOperands(); ++i) 1808 if (!CE1->getOperand(i)->isNullValue()) { 1809 if (isa<ConstantInt>(CE1->getOperand(i))) 1810 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1811 else 1812 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1813 } 1814 1815 for (; i < CE2->getNumOperands(); ++i) 1816 if (!CE2->getOperand(i)->isNullValue()) { 1817 if (isa<ConstantInt>(CE2->getOperand(i))) 1818 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1819 else 1820 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal. 1821 } 1822 return ICmpInst::ICMP_EQ; 1823 } 1824 } 1825 } 1826 default: 1827 break; 1828 } 1829 } 1830 1831 return ICmpInst::BAD_ICMP_PREDICATE; 1832} 1833 1834Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred, 1835 Constant *C1, Constant *C2) { 1836 const Type *ResultTy; 1837 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1838 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1839 VT->getNumElements()); 1840 else 1841 ResultTy = Type::getInt1Ty(C1->getContext()); 1842 1843 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1844 if (pred == FCmpInst::FCMP_FALSE) 1845 return Constant::getNullValue(ResultTy); 1846 1847 if (pred == FCmpInst::FCMP_TRUE) 1848 return Constant::getAllOnesValue(ResultTy); 1849 1850 // Handle some degenerate cases first 1851 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1852 // For EQ and NE, we can always pick a value for the undef to make the 1853 // predicate pass or fail, so we can return undef. 1854 if (ICmpInst::isEquality(ICmpInst::Predicate(pred))) 1855 return UndefValue::get(ResultTy); 1856 // Otherwise, pick the same value as the non-undef operand, and fold 1857 // it to true or false. 1858 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred)); 1859 } 1860 1861 // No compile-time operations on this type yet. 1862 if (C1->getType()->isPPC_FP128Ty()) 1863 return 0; 1864 1865 // icmp eq/ne(null,GV) -> false/true 1866 if (C1->isNullValue()) { 1867 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2)) 1868 // Don't try to evaluate aliases. External weak GV can be null. 1869 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1870 if (pred == ICmpInst::ICMP_EQ) 1871 return ConstantInt::getFalse(C1->getContext()); 1872 else if (pred == ICmpInst::ICMP_NE) 1873 return ConstantInt::getTrue(C1->getContext()); 1874 } 1875 // icmp eq/ne(GV,null) -> false/true 1876 } else if (C2->isNullValue()) { 1877 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) 1878 // Don't try to evaluate aliases. External weak GV can be null. 1879 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) { 1880 if (pred == ICmpInst::ICMP_EQ) 1881 return ConstantInt::getFalse(C1->getContext()); 1882 else if (pred == ICmpInst::ICMP_NE) 1883 return ConstantInt::getTrue(C1->getContext()); 1884 } 1885 } 1886 1887 // If the comparison is a comparison between two i1's, simplify it. 1888 if (C1->getType()->isIntegerTy(1)) { 1889 switch(pred) { 1890 case ICmpInst::ICMP_EQ: 1891 if (isa<ConstantInt>(C2)) 1892 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1893 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1894 case ICmpInst::ICMP_NE: 1895 return ConstantExpr::getXor(C1, C2); 1896 default: 1897 break; 1898 } 1899 } 1900 1901 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1902 APInt V1 = cast<ConstantInt>(C1)->getValue(); 1903 APInt V2 = cast<ConstantInt>(C2)->getValue(); 1904 switch (pred) { 1905 default: llvm_unreachable("Invalid ICmp Predicate"); return 0; 1906 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2); 1907 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2); 1908 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2)); 1909 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2)); 1910 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2)); 1911 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2)); 1912 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2)); 1913 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2)); 1914 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2)); 1915 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2)); 1916 } 1917 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1918 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF(); 1919 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF(); 1920 APFloat::cmpResult R = C1V.compare(C2V); 1921 switch (pred) { 1922 default: llvm_unreachable("Invalid FCmp Predicate"); return 0; 1923 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy); 1924 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy); 1925 case FCmpInst::FCMP_UNO: 1926 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered); 1927 case FCmpInst::FCMP_ORD: 1928 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered); 1929 case FCmpInst::FCMP_UEQ: 1930 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1931 R==APFloat::cmpEqual); 1932 case FCmpInst::FCMP_OEQ: 1933 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual); 1934 case FCmpInst::FCMP_UNE: 1935 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual); 1936 case FCmpInst::FCMP_ONE: 1937 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1938 R==APFloat::cmpGreaterThan); 1939 case FCmpInst::FCMP_ULT: 1940 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1941 R==APFloat::cmpLessThan); 1942 case FCmpInst::FCMP_OLT: 1943 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan); 1944 case FCmpInst::FCMP_UGT: 1945 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered || 1946 R==APFloat::cmpGreaterThan); 1947 case FCmpInst::FCMP_OGT: 1948 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan); 1949 case FCmpInst::FCMP_ULE: 1950 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan); 1951 case FCmpInst::FCMP_OLE: 1952 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan || 1953 R==APFloat::cmpEqual); 1954 case FCmpInst::FCMP_UGE: 1955 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan); 1956 case FCmpInst::FCMP_OGE: 1957 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan || 1958 R==APFloat::cmpEqual); 1959 } 1960 } else if (C1->getType()->isVectorTy()) { 1961 SmallVector<Constant*, 16> C1Elts, C2Elts; 1962 C1->getVectorElements(C1Elts); 1963 C2->getVectorElements(C2Elts); 1964 if (C1Elts.empty() || C2Elts.empty()) 1965 return 0; 1966 1967 // If we can constant fold the comparison of each element, constant fold 1968 // the whole vector comparison. 1969 SmallVector<Constant*, 4> ResElts; 1970 // Compare the elements, producing an i1 result or constant expr. 1971 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) 1972 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i])); 1973 1974 return ConstantVector::get(ResElts); 1975 } 1976 1977 if (C1->getType()->isFloatingPointTy()) { 1978 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1979 switch (evaluateFCmpRelation(C1, C2)) { 1980 default: llvm_unreachable("Unknown relation!"); 1981 case FCmpInst::FCMP_UNO: 1982 case FCmpInst::FCMP_ORD: 1983 case FCmpInst::FCMP_UEQ: 1984 case FCmpInst::FCMP_UNE: 1985 case FCmpInst::FCMP_ULT: 1986 case FCmpInst::FCMP_UGT: 1987 case FCmpInst::FCMP_ULE: 1988 case FCmpInst::FCMP_UGE: 1989 case FCmpInst::FCMP_TRUE: 1990 case FCmpInst::FCMP_FALSE: 1991 case FCmpInst::BAD_FCMP_PREDICATE: 1992 break; // Couldn't determine anything about these constants. 1993 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1994 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ || 1995 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE || 1996 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 1997 break; 1998 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1999 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 2000 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT || 2001 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE); 2002 break; 2003 case FCmpInst::FCMP_OGT: // We know that C1 > C2 2004 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE || 2005 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT || 2006 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE); 2007 break; 2008 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 2009 // We can only partially decide this relation. 2010 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 2011 Result = 0; 2012 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 2013 Result = 1; 2014 break; 2015 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 2016 // We can only partially decide this relation. 2017 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT) 2018 Result = 0; 2019 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT) 2020 Result = 1; 2021 break; 2022 case FCmpInst::FCMP_ONE: // We know that C1 != C2 2023 // We can only partially decide this relation. 2024 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) 2025 Result = 0; 2026 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE) 2027 Result = 1; 2028 break; 2029 } 2030 2031 // If we evaluated the result, return it now. 2032 if (Result != -1) 2033 return ConstantInt::get(ResultTy, Result); 2034 2035 } else { 2036 // Evaluate the relation between the two constants, per the predicate. 2037 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 2038 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) { 2039 default: llvm_unreachable("Unknown relational!"); 2040 case ICmpInst::BAD_ICMP_PREDICATE: 2041 break; // Couldn't determine anything about these constants. 2042 case ICmpInst::ICMP_EQ: // We know the constants are equal! 2043 // If we know the constants are equal, we can decide the result of this 2044 // computation precisely. 2045 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred); 2046 break; 2047 case ICmpInst::ICMP_ULT: 2048 switch (pred) { 2049 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 2050 Result = 1; break; 2051 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 2052 Result = 0; break; 2053 } 2054 break; 2055 case ICmpInst::ICMP_SLT: 2056 switch (pred) { 2057 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 2058 Result = 1; break; 2059 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 2060 Result = 0; break; 2061 } 2062 break; 2063 case ICmpInst::ICMP_UGT: 2064 switch (pred) { 2065 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 2066 Result = 1; break; 2067 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 2068 Result = 0; break; 2069 } 2070 break; 2071 case ICmpInst::ICMP_SGT: 2072 switch (pred) { 2073 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 2074 Result = 1; break; 2075 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 2076 Result = 0; break; 2077 } 2078 break; 2079 case ICmpInst::ICMP_ULE: 2080 if (pred == ICmpInst::ICMP_UGT) Result = 0; 2081 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1; 2082 break; 2083 case ICmpInst::ICMP_SLE: 2084 if (pred == ICmpInst::ICMP_SGT) Result = 0; 2085 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1; 2086 break; 2087 case ICmpInst::ICMP_UGE: 2088 if (pred == ICmpInst::ICMP_ULT) Result = 0; 2089 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1; 2090 break; 2091 case ICmpInst::ICMP_SGE: 2092 if (pred == ICmpInst::ICMP_SLT) Result = 0; 2093 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1; 2094 break; 2095 case ICmpInst::ICMP_NE: 2096 if (pred == ICmpInst::ICMP_EQ) Result = 0; 2097 if (pred == ICmpInst::ICMP_NE) Result = 1; 2098 break; 2099 } 2100 2101 // If we evaluated the result, return it now. 2102 if (Result != -1) 2103 return ConstantInt::get(ResultTy, Result); 2104 2105 // If the right hand side is a bitcast, try using its inverse to simplify 2106 // it by moving it to the left hand side. We can't do this if it would turn 2107 // a vector compare into a scalar compare or visa versa. 2108 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 2109 Constant *CE2Op0 = CE2->getOperand(0); 2110 if (CE2->getOpcode() == Instruction::BitCast && 2111 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) { 2112 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 2113 return ConstantExpr::getICmp(pred, Inverse, CE2Op0); 2114 } 2115 } 2116 2117 // If the left hand side is an extension, try eliminating it. 2118 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 2119 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) || 2120 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){ 2121 Constant *CE1Op0 = CE1->getOperand(0); 2122 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 2123 if (CE1Inverse == CE1Op0) { 2124 // Check whether we can safely truncate the right hand side. 2125 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 2126 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) { 2127 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse); 2128 } 2129 } 2130 } 2131 } 2132 2133 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 2134 (C1->isNullValue() && !C2->isNullValue())) { 2135 // If C2 is a constant expr and C1 isn't, flip them around and fold the 2136 // other way if possible. 2137 // Also, if C1 is null and C2 isn't, flip them around. 2138 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred); 2139 return ConstantExpr::getICmp(pred, C2, C1); 2140 } 2141 } 2142 return 0; 2143} 2144 2145/// isInBoundsIndices - Test whether the given sequence of *normalized* indices 2146/// is "inbounds". 2147template<typename IndexTy> 2148static bool isInBoundsIndices(IndexTy const *Idxs, size_t NumIdx) { 2149 // No indices means nothing that could be out of bounds. 2150 if (NumIdx == 0) return true; 2151 2152 // If the first index is zero, it's in bounds. 2153 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 2154 2155 // If the first index is one and all the rest are zero, it's in bounds, 2156 // by the one-past-the-end rule. 2157 if (!cast<ConstantInt>(Idxs[0])->isOne()) 2158 return false; 2159 for (unsigned i = 1, e = NumIdx; i != e; ++i) 2160 if (!cast<Constant>(Idxs[i])->isNullValue()) 2161 return false; 2162 return true; 2163} 2164 2165template<typename IndexTy> 2166static Constant *ConstantFoldGetElementPtrImpl(Constant *C, 2167 bool inBounds, 2168 IndexTy const *Idxs, 2169 unsigned NumIdx) { 2170 Constant *Idx0 = cast<Constant>(Idxs[0]); 2171 if (NumIdx == 0 || 2172 (NumIdx == 1 && Idx0->isNullValue())) 2173 return C; 2174 2175 if (isa<UndefValue>(C)) { 2176 const PointerType *Ptr = cast<PointerType>(C->getType()); 2177 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, Idxs+NumIdx); 2178 assert(Ty != 0 && "Invalid indices for GEP!"); 2179 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace())); 2180 } 2181 2182 if (C->isNullValue()) { 2183 bool isNull = true; 2184 for (unsigned i = 0, e = NumIdx; i != e; ++i) 2185 if (!cast<Constant>(Idxs[i])->isNullValue()) { 2186 isNull = false; 2187 break; 2188 } 2189 if (isNull) { 2190 const PointerType *Ptr = cast<PointerType>(C->getType()); 2191 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, 2192 Idxs+NumIdx); 2193 assert(Ty != 0 && "Invalid indices for GEP!"); 2194 return ConstantPointerNull::get(PointerType::get(Ty, 2195 Ptr->getAddressSpace())); 2196 } 2197 } 2198 2199 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2200 // Combine Indices - If the source pointer to this getelementptr instruction 2201 // is a getelementptr instruction, combine the indices of the two 2202 // getelementptr instructions into a single instruction. 2203 // 2204 if (CE->getOpcode() == Instruction::GetElementPtr) { 2205 const Type *LastTy = 0; 2206 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE); 2207 I != E; ++I) 2208 LastTy = *I; 2209 2210 if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) { 2211 SmallVector<Value*, 16> NewIndices; 2212 NewIndices.reserve(NumIdx + CE->getNumOperands()); 2213 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i) 2214 NewIndices.push_back(CE->getOperand(i)); 2215 2216 // Add the last index of the source with the first index of the new GEP. 2217 // Make sure to handle the case when they are actually different types. 2218 Constant *Combined = CE->getOperand(CE->getNumOperands()-1); 2219 // Otherwise it must be an array. 2220 if (!Idx0->isNullValue()) { 2221 const Type *IdxTy = Combined->getType(); 2222 if (IdxTy != Idx0->getType()) { 2223 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext()); 2224 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty); 2225 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty); 2226 Combined = ConstantExpr::get(Instruction::Add, C1, C2); 2227 } else { 2228 Combined = 2229 ConstantExpr::get(Instruction::Add, Idx0, Combined); 2230 } 2231 } 2232 2233 NewIndices.push_back(Combined); 2234 NewIndices.append(Idxs+1, Idxs+NumIdx); 2235 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ? 2236 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0), 2237 &NewIndices[0], 2238 NewIndices.size()) : 2239 ConstantExpr::getGetElementPtr(CE->getOperand(0), 2240 &NewIndices[0], 2241 NewIndices.size()); 2242 } 2243 } 2244 2245 // Implement folding of: 2246 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*), 2247 // i64 0, i64 0) 2248 // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0) 2249 // 2250 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) { 2251 if (const PointerType *SPT = 2252 dyn_cast<PointerType>(CE->getOperand(0)->getType())) 2253 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType())) 2254 if (const ArrayType *CAT = 2255 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType())) 2256 if (CAT->getElementType() == SAT->getElementType()) 2257 return inBounds ? 2258 ConstantExpr::getInBoundsGetElementPtr( 2259 (Constant*)CE->getOperand(0), Idxs, NumIdx) : 2260 ConstantExpr::getGetElementPtr( 2261 (Constant*)CE->getOperand(0), Idxs, NumIdx); 2262 } 2263 } 2264 2265 // Check to see if any array indices are not within the corresponding 2266 // notional array bounds. If so, try to determine if they can be factored 2267 // out into preceding dimensions. 2268 bool Unknown = false; 2269 SmallVector<Constant *, 8> NewIdxs; 2270 const Type *Ty = C->getType(); 2271 const Type *Prev = 0; 2272 for (unsigned i = 0; i != NumIdx; 2273 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) { 2274 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2275 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 2276 if (ATy->getNumElements() <= INT64_MAX && 2277 ATy->getNumElements() != 0 && 2278 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) { 2279 if (isa<SequentialType>(Prev)) { 2280 // It's out of range, but we can factor it into the prior 2281 // dimension. 2282 NewIdxs.resize(NumIdx); 2283 ConstantInt *Factor = ConstantInt::get(CI->getType(), 2284 ATy->getNumElements()); 2285 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor); 2286 2287 Constant *PrevIdx = cast<Constant>(Idxs[i-1]); 2288 Constant *Div = ConstantExpr::getSDiv(CI, Factor); 2289 2290 // Before adding, extend both operands to i64 to avoid 2291 // overflow trouble. 2292 if (!PrevIdx->getType()->isIntegerTy(64)) 2293 PrevIdx = ConstantExpr::getSExt(PrevIdx, 2294 Type::getInt64Ty(Div->getContext())); 2295 if (!Div->getType()->isIntegerTy(64)) 2296 Div = ConstantExpr::getSExt(Div, 2297 Type::getInt64Ty(Div->getContext())); 2298 2299 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div); 2300 } else { 2301 // It's out of range, but the prior dimension is a struct 2302 // so we can't do anything about it. 2303 Unknown = true; 2304 } 2305 } 2306 } else { 2307 // We don't know if it's in range or not. 2308 Unknown = true; 2309 } 2310 } 2311 2312 // If we did any factoring, start over with the adjusted indices. 2313 if (!NewIdxs.empty()) { 2314 for (unsigned i = 0; i != NumIdx; ++i) 2315 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2316 return inBounds ? 2317 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(), 2318 NewIdxs.size()) : 2319 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size()); 2320 } 2321 2322 // If all indices are known integers and normalized, we can do a simple 2323 // check for the "inbounds" property. 2324 if (!Unknown && !inBounds && 2325 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx)) 2326 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx); 2327 2328 return 0; 2329} 2330 2331Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2332 bool inBounds, 2333 Constant* const *Idxs, 2334 unsigned NumIdx) { 2335 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx); 2336} 2337 2338Constant *llvm::ConstantFoldGetElementPtr(Constant *C, 2339 bool inBounds, 2340 Value* const *Idxs, 2341 unsigned NumIdx) { 2342 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx); 2343} 2344