InstCombineCompares.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===- InstCombineCompares.cpp --------------------------------------------===// 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 the visitICmp and visitFCmp functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Analysis/ConstantFolding.h" 16#include "llvm/Analysis/InstructionSimplify.h" 17#include "llvm/Analysis/MemoryBuiltins.h" 18#include "llvm/IR/ConstantRange.h" 19#include "llvm/IR/DataLayout.h" 20#include "llvm/IR/GetElementPtrTypeIterator.h" 21#include "llvm/IR/IntrinsicInst.h" 22#include "llvm/IR/PatternMatch.h" 23#include "llvm/Target/TargetLibraryInfo.h" 24using namespace llvm; 25using namespace PatternMatch; 26 27static ConstantInt *getOne(Constant *C) { 28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1); 29} 30 31static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 32 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 33} 34 35static bool HasAddOverflow(ConstantInt *Result, 36 ConstantInt *In1, ConstantInt *In2, 37 bool IsSigned) { 38 if (!IsSigned) 39 return Result->getValue().ult(In1->getValue()); 40 41 if (In2->isNegative()) 42 return Result->getValue().sgt(In1->getValue()); 43 return Result->getValue().slt(In1->getValue()); 44} 45 46/// AddWithOverflow - Compute Result = In1+In2, returning true if the result 47/// overflowed for this type. 48static bool AddWithOverflow(Constant *&Result, Constant *In1, 49 Constant *In2, bool IsSigned = false) { 50 Result = ConstantExpr::getAdd(In1, In2); 51 52 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 53 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 54 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 55 if (HasAddOverflow(ExtractElement(Result, Idx), 56 ExtractElement(In1, Idx), 57 ExtractElement(In2, Idx), 58 IsSigned)) 59 return true; 60 } 61 return false; 62 } 63 64 return HasAddOverflow(cast<ConstantInt>(Result), 65 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 66 IsSigned); 67} 68 69static bool HasSubOverflow(ConstantInt *Result, 70 ConstantInt *In1, ConstantInt *In2, 71 bool IsSigned) { 72 if (!IsSigned) 73 return Result->getValue().ugt(In1->getValue()); 74 75 if (In2->isNegative()) 76 return Result->getValue().slt(In1->getValue()); 77 78 return Result->getValue().sgt(In1->getValue()); 79} 80 81/// SubWithOverflow - Compute Result = In1-In2, returning true if the result 82/// overflowed for this type. 83static bool SubWithOverflow(Constant *&Result, Constant *In1, 84 Constant *In2, bool IsSigned = false) { 85 Result = ConstantExpr::getSub(In1, In2); 86 87 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 88 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 89 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 90 if (HasSubOverflow(ExtractElement(Result, Idx), 91 ExtractElement(In1, Idx), 92 ExtractElement(In2, Idx), 93 IsSigned)) 94 return true; 95 } 96 return false; 97 } 98 99 return HasSubOverflow(cast<ConstantInt>(Result), 100 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 101 IsSigned); 102} 103 104/// isSignBitCheck - Given an exploded icmp instruction, return true if the 105/// comparison only checks the sign bit. If it only checks the sign bit, set 106/// TrueIfSigned if the result of the comparison is true when the input value is 107/// signed. 108static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 109 bool &TrueIfSigned) { 110 switch (pred) { 111 case ICmpInst::ICMP_SLT: // True if LHS s< 0 112 TrueIfSigned = true; 113 return RHS->isZero(); 114 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 115 TrueIfSigned = true; 116 return RHS->isAllOnesValue(); 117 case ICmpInst::ICMP_SGT: // True if LHS s> -1 118 TrueIfSigned = false; 119 return RHS->isAllOnesValue(); 120 case ICmpInst::ICMP_UGT: 121 // True if LHS u> RHS and RHS == high-bit-mask - 1 122 TrueIfSigned = true; 123 return RHS->isMaxValue(true); 124 case ICmpInst::ICMP_UGE: 125 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 126 TrueIfSigned = true; 127 return RHS->getValue().isSignBit(); 128 default: 129 return false; 130 } 131} 132 133/// Returns true if the exploded icmp can be expressed as a signed comparison 134/// to zero and updates the predicate accordingly. 135/// The signedness of the comparison is preserved. 136static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) { 137 if (!ICmpInst::isSigned(pred)) 138 return false; 139 140 if (RHS->isZero()) 141 return ICmpInst::isRelational(pred); 142 143 if (RHS->isOne()) { 144 if (pred == ICmpInst::ICMP_SLT) { 145 pred = ICmpInst::ICMP_SLE; 146 return true; 147 } 148 } else if (RHS->isAllOnesValue()) { 149 if (pred == ICmpInst::ICMP_SGT) { 150 pred = ICmpInst::ICMP_SGE; 151 return true; 152 } 153 } 154 155 return false; 156} 157 158// isHighOnes - Return true if the constant is of the form 1+0+. 159// This is the same as lowones(~X). 160static bool isHighOnes(const ConstantInt *CI) { 161 return (~CI->getValue() + 1).isPowerOf2(); 162} 163 164/// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a 165/// set of known zero and one bits, compute the maximum and minimum values that 166/// could have the specified known zero and known one bits, returning them in 167/// min/max. 168static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero, 169 const APInt& KnownOne, 170 APInt& Min, APInt& Max) { 171 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 172 KnownZero.getBitWidth() == Min.getBitWidth() && 173 KnownZero.getBitWidth() == Max.getBitWidth() && 174 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 175 APInt UnknownBits = ~(KnownZero|KnownOne); 176 177 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 178 // bit if it is unknown. 179 Min = KnownOne; 180 Max = KnownOne|UnknownBits; 181 182 if (UnknownBits.isNegative()) { // Sign bit is unknown 183 Min.setBit(Min.getBitWidth()-1); 184 Max.clearBit(Max.getBitWidth()-1); 185 } 186} 187 188// ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and 189// a set of known zero and one bits, compute the maximum and minimum values that 190// could have the specified known zero and known one bits, returning them in 191// min/max. 192static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, 193 const APInt &KnownOne, 194 APInt &Min, APInt &Max) { 195 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 196 KnownZero.getBitWidth() == Min.getBitWidth() && 197 KnownZero.getBitWidth() == Max.getBitWidth() && 198 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 199 APInt UnknownBits = ~(KnownZero|KnownOne); 200 201 // The minimum value is when the unknown bits are all zeros. 202 Min = KnownOne; 203 // The maximum value is when the unknown bits are all ones. 204 Max = KnownOne|UnknownBits; 205} 206 207 208 209/// FoldCmpLoadFromIndexedGlobal - Called we see this pattern: 210/// cmp pred (load (gep GV, ...)), cmpcst 211/// where GV is a global variable with a constant initializer. Try to simplify 212/// this into some simple computation that does not need the load. For example 213/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 214/// 215/// If AndCst is non-null, then the loaded value is masked with that constant 216/// before doing the comparison. This handles cases like "A[i]&4 == 0". 217Instruction *InstCombiner:: 218FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, 219 CmpInst &ICI, ConstantInt *AndCst) { 220 // We need TD information to know the pointer size unless this is inbounds. 221 if (!GEP->isInBounds() && DL == 0) 222 return 0; 223 224 Constant *Init = GV->getInitializer(); 225 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init)) 226 return 0; 227 228 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); 229 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays. 230 231 // There are many forms of this optimization we can handle, for now, just do 232 // the simple index into a single-dimensional array. 233 // 234 // Require: GEP GV, 0, i {{, constant indices}} 235 if (GEP->getNumOperands() < 3 || 236 !isa<ConstantInt>(GEP->getOperand(1)) || 237 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 238 isa<Constant>(GEP->getOperand(2))) 239 return 0; 240 241 // Check that indices after the variable are constants and in-range for the 242 // type they index. Collect the indices. This is typically for arrays of 243 // structs. 244 SmallVector<unsigned, 4> LaterIndices; 245 246 Type *EltTy = Init->getType()->getArrayElementType(); 247 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 248 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 249 if (Idx == 0) return 0; // Variable index. 250 251 uint64_t IdxVal = Idx->getZExtValue(); 252 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 253 254 if (StructType *STy = dyn_cast<StructType>(EltTy)) 255 EltTy = STy->getElementType(IdxVal); 256 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 257 if (IdxVal >= ATy->getNumElements()) return 0; 258 EltTy = ATy->getElementType(); 259 } else { 260 return 0; // Unknown type. 261 } 262 263 LaterIndices.push_back(IdxVal); 264 } 265 266 enum { Overdefined = -3, Undefined = -2 }; 267 268 // Variables for our state machines. 269 270 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 271 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 272 // and 87 is the second (and last) index. FirstTrueElement is -2 when 273 // undefined, otherwise set to the first true element. SecondTrueElement is 274 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 275 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 276 277 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 278 // form "i != 47 & i != 87". Same state transitions as for true elements. 279 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 280 281 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 282 /// define a state machine that triggers for ranges of values that the index 283 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 284 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 285 /// index in the range (inclusive). We use -2 for undefined here because we 286 /// use relative comparisons and don't want 0-1 to match -1. 287 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 288 289 // MagicBitvector - This is a magic bitvector where we set a bit if the 290 // comparison is true for element 'i'. If there are 64 elements or less in 291 // the array, this will fully represent all the comparison results. 292 uint64_t MagicBitvector = 0; 293 294 295 // Scan the array and see if one of our patterns matches. 296 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 297 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { 298 Constant *Elt = Init->getAggregateElement(i); 299 if (Elt == 0) return 0; 300 301 // If this is indexing an array of structures, get the structure element. 302 if (!LaterIndices.empty()) 303 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); 304 305 // If the element is masked, handle it. 306 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 307 308 // Find out if the comparison would be true or false for the i'th element. 309 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 310 CompareRHS, DL, TLI); 311 // If the result is undef for this element, ignore it. 312 if (isa<UndefValue>(C)) { 313 // Extend range state machines to cover this element in case there is an 314 // undef in the middle of the range. 315 if (TrueRangeEnd == (int)i-1) 316 TrueRangeEnd = i; 317 if (FalseRangeEnd == (int)i-1) 318 FalseRangeEnd = i; 319 continue; 320 } 321 322 // If we can't compute the result for any of the elements, we have to give 323 // up evaluating the entire conditional. 324 if (!isa<ConstantInt>(C)) return 0; 325 326 // Otherwise, we know if the comparison is true or false for this element, 327 // update our state machines. 328 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 329 330 // State machine for single/double/range index comparison. 331 if (IsTrueForElt) { 332 // Update the TrueElement state machine. 333 if (FirstTrueElement == Undefined) 334 FirstTrueElement = TrueRangeEnd = i; // First true element. 335 else { 336 // Update double-compare state machine. 337 if (SecondTrueElement == Undefined) 338 SecondTrueElement = i; 339 else 340 SecondTrueElement = Overdefined; 341 342 // Update range state machine. 343 if (TrueRangeEnd == (int)i-1) 344 TrueRangeEnd = i; 345 else 346 TrueRangeEnd = Overdefined; 347 } 348 } else { 349 // Update the FalseElement state machine. 350 if (FirstFalseElement == Undefined) 351 FirstFalseElement = FalseRangeEnd = i; // First false element. 352 else { 353 // Update double-compare state machine. 354 if (SecondFalseElement == Undefined) 355 SecondFalseElement = i; 356 else 357 SecondFalseElement = Overdefined; 358 359 // Update range state machine. 360 if (FalseRangeEnd == (int)i-1) 361 FalseRangeEnd = i; 362 else 363 FalseRangeEnd = Overdefined; 364 } 365 } 366 367 368 // If this element is in range, update our magic bitvector. 369 if (i < 64 && IsTrueForElt) 370 MagicBitvector |= 1ULL << i; 371 372 // If all of our states become overdefined, bail out early. Since the 373 // predicate is expensive, only check it every 8 elements. This is only 374 // really useful for really huge arrays. 375 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 376 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 377 FalseRangeEnd == Overdefined) 378 return 0; 379 } 380 381 // Now that we've scanned the entire array, emit our new comparison(s). We 382 // order the state machines in complexity of the generated code. 383 Value *Idx = GEP->getOperand(2); 384 385 // If the index is larger than the pointer size of the target, truncate the 386 // index down like the GEP would do implicitly. We don't have to do this for 387 // an inbounds GEP because the index can't be out of range. 388 if (!GEP->isInBounds()) { 389 Type *IntPtrTy = DL->getIntPtrType(GEP->getType()); 390 unsigned PtrSize = IntPtrTy->getIntegerBitWidth(); 391 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize) 392 Idx = Builder->CreateTrunc(Idx, IntPtrTy); 393 } 394 395 // If the comparison is only true for one or two elements, emit direct 396 // comparisons. 397 if (SecondTrueElement != Overdefined) { 398 // None true -> false. 399 if (FirstTrueElement == Undefined) 400 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 401 402 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 403 404 // True for one element -> 'i == 47'. 405 if (SecondTrueElement == Undefined) 406 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 407 408 // True for two elements -> 'i == 47 | i == 72'. 409 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 410 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 411 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 412 return BinaryOperator::CreateOr(C1, C2); 413 } 414 415 // If the comparison is only false for one or two elements, emit direct 416 // comparisons. 417 if (SecondFalseElement != Overdefined) { 418 // None false -> true. 419 if (FirstFalseElement == Undefined) 420 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 421 422 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 423 424 // False for one element -> 'i != 47'. 425 if (SecondFalseElement == Undefined) 426 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 427 428 // False for two elements -> 'i != 47 & i != 72'. 429 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 430 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 431 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 432 return BinaryOperator::CreateAnd(C1, C2); 433 } 434 435 // If the comparison can be replaced with a range comparison for the elements 436 // where it is true, emit the range check. 437 if (TrueRangeEnd != Overdefined) { 438 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 439 440 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 441 if (FirstTrueElement) { 442 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 443 Idx = Builder->CreateAdd(Idx, Offs); 444 } 445 446 Value *End = ConstantInt::get(Idx->getType(), 447 TrueRangeEnd-FirstTrueElement+1); 448 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 449 } 450 451 // False range check. 452 if (FalseRangeEnd != Overdefined) { 453 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 454 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 455 if (FirstFalseElement) { 456 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 457 Idx = Builder->CreateAdd(Idx, Offs); 458 } 459 460 Value *End = ConstantInt::get(Idx->getType(), 461 FalseRangeEnd-FirstFalseElement); 462 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 463 } 464 465 466 // If a magic bitvector captures the entire comparison state 467 // of this load, replace it with computation that does: 468 // ((magic_cst >> i) & 1) != 0 469 { 470 Type *Ty = 0; 471 472 // Look for an appropriate type: 473 // - The type of Idx if the magic fits 474 // - The smallest fitting legal type if we have a DataLayout 475 // - Default to i32 476 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth()) 477 Ty = Idx->getType(); 478 else if (DL) 479 Ty = DL->getSmallestLegalIntType(Init->getContext(), ArrayElementCount); 480 else if (ArrayElementCount <= 32) 481 Ty = Type::getInt32Ty(Init->getContext()); 482 483 if (Ty != 0) { 484 Value *V = Builder->CreateIntCast(Idx, Ty, false); 485 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 486 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 487 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 488 } 489 } 490 491 return 0; 492} 493 494 495/// EvaluateGEPOffsetExpression - Return a value that can be used to compare 496/// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 497/// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 498/// be complex, and scales are involved. The above expression would also be 499/// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 500/// This later form is less amenable to optimization though, and we are allowed 501/// to generate the first by knowing that pointer arithmetic doesn't overflow. 502/// 503/// If we can't emit an optimized form for this expression, this returns null. 504/// 505static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) { 506 const DataLayout &DL = *IC.getDataLayout(); 507 gep_type_iterator GTI = gep_type_begin(GEP); 508 509 // Check to see if this gep only has a single variable index. If so, and if 510 // any constant indices are a multiple of its scale, then we can compute this 511 // in terms of the scale of the variable index. For example, if the GEP 512 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 513 // because the expression will cross zero at the same point. 514 unsigned i, e = GEP->getNumOperands(); 515 int64_t Offset = 0; 516 for (i = 1; i != e; ++i, ++GTI) { 517 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 518 // Compute the aggregate offset of constant indices. 519 if (CI->isZero()) continue; 520 521 // Handle a struct index, which adds its field offset to the pointer. 522 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 523 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 524 } else { 525 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); 526 Offset += Size*CI->getSExtValue(); 527 } 528 } else { 529 // Found our variable index. 530 break; 531 } 532 } 533 534 // If there are no variable indices, we must have a constant offset, just 535 // evaluate it the general way. 536 if (i == e) return 0; 537 538 Value *VariableIdx = GEP->getOperand(i); 539 // Determine the scale factor of the variable element. For example, this is 540 // 4 if the variable index is into an array of i32. 541 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType()); 542 543 // Verify that there are no other variable indices. If so, emit the hard way. 544 for (++i, ++GTI; i != e; ++i, ++GTI) { 545 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 546 if (!CI) return 0; 547 548 // Compute the aggregate offset of constant indices. 549 if (CI->isZero()) continue; 550 551 // Handle a struct index, which adds its field offset to the pointer. 552 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 553 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 554 } else { 555 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); 556 Offset += Size*CI->getSExtValue(); 557 } 558 } 559 560 561 562 // Okay, we know we have a single variable index, which must be a 563 // pointer/array/vector index. If there is no offset, life is simple, return 564 // the index. 565 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType()); 566 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth(); 567 if (Offset == 0) { 568 // Cast to intptrty in case a truncation occurs. If an extension is needed, 569 // we don't need to bother extending: the extension won't affect where the 570 // computation crosses zero. 571 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { 572 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy); 573 } 574 return VariableIdx; 575 } 576 577 // Otherwise, there is an index. The computation we will do will be modulo 578 // the pointer size, so get it. 579 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 580 581 Offset &= PtrSizeMask; 582 VariableScale &= PtrSizeMask; 583 584 // To do this transformation, any constant index must be a multiple of the 585 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 586 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 587 // multiple of the variable scale. 588 int64_t NewOffs = Offset / (int64_t)VariableScale; 589 if (Offset != NewOffs*(int64_t)VariableScale) 590 return 0; 591 592 // Okay, we can do this evaluation. Start by converting the index to intptr. 593 if (VariableIdx->getType() != IntPtrTy) 594 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy, 595 true /*Signed*/); 596 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 597 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset"); 598} 599 600/// FoldGEPICmp - Fold comparisons between a GEP instruction and something 601/// else. At this point we know that the GEP is on the LHS of the comparison. 602Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 603 ICmpInst::Predicate Cond, 604 Instruction &I) { 605 // Don't transform signed compares of GEPs into index compares. Even if the 606 // GEP is inbounds, the final add of the base pointer can have signed overflow 607 // and would change the result of the icmp. 608 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be 609 // the maximum signed value for the pointer type. 610 if (ICmpInst::isSigned(Cond)) 611 return 0; 612 613 // Look through bitcasts. 614 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 615 RHS = BCI->getOperand(0); 616 617 Value *PtrBase = GEPLHS->getOperand(0); 618 if (DL && PtrBase == RHS && GEPLHS->isInBounds()) { 619 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 620 // This transformation (ignoring the base and scales) is valid because we 621 // know pointers can't overflow since the gep is inbounds. See if we can 622 // output an optimized form. 623 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this); 624 625 // If not, synthesize the offset the hard way. 626 if (Offset == 0) 627 Offset = EmitGEPOffset(GEPLHS); 628 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 629 Constant::getNullValue(Offset->getType())); 630 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 631 // If the base pointers are different, but the indices are the same, just 632 // compare the base pointer. 633 if (PtrBase != GEPRHS->getOperand(0)) { 634 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 635 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 636 GEPRHS->getOperand(0)->getType(); 637 if (IndicesTheSame) 638 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 639 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 640 IndicesTheSame = false; 641 break; 642 } 643 644 // If all indices are the same, just compare the base pointers. 645 if (IndicesTheSame) 646 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 647 648 // If we're comparing GEPs with two base pointers that only differ in type 649 // and both GEPs have only constant indices or just one use, then fold 650 // the compare with the adjusted indices. 651 if (DL && GEPLHS->isInBounds() && GEPRHS->isInBounds() && 652 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && 653 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && 654 PtrBase->stripPointerCasts() == 655 GEPRHS->getOperand(0)->stripPointerCasts()) { 656 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond), 657 EmitGEPOffset(GEPLHS), 658 EmitGEPOffset(GEPRHS)); 659 return ReplaceInstUsesWith(I, Cmp); 660 } 661 662 // Otherwise, the base pointers are different and the indices are 663 // different, bail out. 664 return 0; 665 } 666 667 // If one of the GEPs has all zero indices, recurse. 668 bool AllZeros = true; 669 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 670 if (!isa<Constant>(GEPLHS->getOperand(i)) || 671 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 672 AllZeros = false; 673 break; 674 } 675 if (AllZeros) 676 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 677 ICmpInst::getSwappedPredicate(Cond), I); 678 679 // If the other GEP has all zero indices, recurse. 680 AllZeros = true; 681 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 682 if (!isa<Constant>(GEPRHS->getOperand(i)) || 683 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 684 AllZeros = false; 685 break; 686 } 687 if (AllZeros) 688 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 689 690 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); 691 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 692 // If the GEPs only differ by one index, compare it. 693 unsigned NumDifferences = 0; // Keep track of # differences. 694 unsigned DiffOperand = 0; // The operand that differs. 695 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 696 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 697 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 698 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 699 // Irreconcilable differences. 700 NumDifferences = 2; 701 break; 702 } else { 703 if (NumDifferences++) break; 704 DiffOperand = i; 705 } 706 } 707 708 if (NumDifferences == 0) // SAME GEP? 709 return ReplaceInstUsesWith(I, // No comparison is needed here. 710 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond))); 711 712 else if (NumDifferences == 1 && GEPsInBounds) { 713 Value *LHSV = GEPLHS->getOperand(DiffOperand); 714 Value *RHSV = GEPRHS->getOperand(DiffOperand); 715 // Make sure we do a signed comparison here. 716 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 717 } 718 } 719 720 // Only lower this if the icmp is the only user of the GEP or if we expect 721 // the result to fold to a constant! 722 if (DL && 723 GEPsInBounds && 724 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 725 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 726 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 727 Value *L = EmitGEPOffset(GEPLHS); 728 Value *R = EmitGEPOffset(GEPRHS); 729 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 730 } 731 } 732 return 0; 733} 734 735/// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 736Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI, 737 Value *X, ConstantInt *CI, 738 ICmpInst::Predicate Pred) { 739 // If we have X+0, exit early (simplifying logic below) and let it get folded 740 // elsewhere. icmp X+0, X -> icmp X, X 741 if (CI->isZero()) { 742 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 743 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 744 } 745 746 // (X+4) == X -> false. 747 if (Pred == ICmpInst::ICMP_EQ) 748 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 749 750 // (X+4) != X -> true. 751 if (Pred == ICmpInst::ICMP_NE) 752 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 753 754 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 755 // so the values can never be equal. Similarly for all other "or equals" 756 // operators. 757 758 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 759 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 760 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 761 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 762 Value *R = 763 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 764 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 765 } 766 767 // (X+1) >u X --> X <u (0-1) --> X != 255 768 // (X+2) >u X --> X <u (0-2) --> X <u 254 769 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 770 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 771 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 772 773 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 774 ConstantInt *SMax = ConstantInt::get(X->getContext(), 775 APInt::getSignedMaxValue(BitWidth)); 776 777 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 778 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 779 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 780 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 781 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 782 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 783 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 784 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 785 786 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 787 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 788 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 789 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 790 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 791 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 792 793 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 794 Constant *C = Builder->getInt(CI->getValue()-1); 795 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 796} 797 798/// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 799/// and CmpRHS are both known to be integer constants. 800Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 801 ConstantInt *DivRHS) { 802 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 803 const APInt &CmpRHSV = CmpRHS->getValue(); 804 805 // FIXME: If the operand types don't match the type of the divide 806 // then don't attempt this transform. The code below doesn't have the 807 // logic to deal with a signed divide and an unsigned compare (and 808 // vice versa). This is because (x /s C1) <s C2 produces different 809 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 810 // (x /u C1) <u C2. Simply casting the operands and result won't 811 // work. :( The if statement below tests that condition and bails 812 // if it finds it. 813 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 814 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 815 return 0; 816 if (DivRHS->isZero()) 817 return 0; // The ProdOV computation fails on divide by zero. 818 if (DivIsSigned && DivRHS->isAllOnesValue()) 819 return 0; // The overflow computation also screws up here 820 if (DivRHS->isOne()) { 821 // This eliminates some funny cases with INT_MIN. 822 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X. 823 return &ICI; 824 } 825 826 // Compute Prod = CI * DivRHS. We are essentially solving an equation 827 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 828 // C2 (CI). By solving for X we can turn this into a range check 829 // instead of computing a divide. 830 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 831 832 // Determine if the product overflows by seeing if the product is 833 // not equal to the divide. Make sure we do the same kind of divide 834 // as in the LHS instruction that we're folding. 835 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 836 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 837 838 // Get the ICmp opcode 839 ICmpInst::Predicate Pred = ICI.getPredicate(); 840 841 /// If the division is known to be exact, then there is no remainder from the 842 /// divide, so the covered range size is unit, otherwise it is the divisor. 843 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS; 844 845 // Figure out the interval that is being checked. For example, a comparison 846 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 847 // Compute this interval based on the constants involved and the signedness of 848 // the compare/divide. This computes a half-open interval, keeping track of 849 // whether either value in the interval overflows. After analysis each 850 // overflow variable is set to 0 if it's corresponding bound variable is valid 851 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 852 int LoOverflow = 0, HiOverflow = 0; 853 Constant *LoBound = 0, *HiBound = 0; 854 855 if (!DivIsSigned) { // udiv 856 // e.g. X/5 op 3 --> [15, 20) 857 LoBound = Prod; 858 HiOverflow = LoOverflow = ProdOV; 859 if (!HiOverflow) { 860 // If this is not an exact divide, then many values in the range collapse 861 // to the same result value. 862 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false); 863 } 864 865 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 866 if (CmpRHSV == 0) { // (X / pos) op 0 867 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 868 LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); 869 HiBound = RangeSize; 870 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 871 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 872 HiOverflow = LoOverflow = ProdOV; 873 if (!HiOverflow) 874 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true); 875 } else { // (X / pos) op neg 876 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 877 HiBound = AddOne(Prod); 878 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 879 if (!LoOverflow) { 880 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 881 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 882 } 883 } 884 } else if (DivRHS->isNegative()) { // Divisor is < 0. 885 if (DivI->isExact()) 886 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 887 if (CmpRHSV == 0) { // (X / neg) op 0 888 // e.g. X/-5 op 0 --> [-4, 5) 889 LoBound = AddOne(RangeSize); 890 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 891 if (HiBound == DivRHS) { // -INTMIN = INTMIN 892 HiOverflow = 1; // [INTMIN+1, overflow) 893 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 894 } 895 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 896 // e.g. X/-5 op 3 --> [-19, -14) 897 HiBound = AddOne(Prod); 898 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 899 if (!LoOverflow) 900 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 901 } else { // (X / neg) op neg 902 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 903 LoOverflow = HiOverflow = ProdOV; 904 if (!HiOverflow) 905 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true); 906 } 907 908 // Dividing by a negative swaps the condition. LT <-> GT 909 Pred = ICmpInst::getSwappedPredicate(Pred); 910 } 911 912 Value *X = DivI->getOperand(0); 913 switch (Pred) { 914 default: llvm_unreachable("Unhandled icmp opcode!"); 915 case ICmpInst::ICMP_EQ: 916 if (LoOverflow && HiOverflow) 917 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 918 if (HiOverflow) 919 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 920 ICmpInst::ICMP_UGE, X, LoBound); 921 if (LoOverflow) 922 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 923 ICmpInst::ICMP_ULT, X, HiBound); 924 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 925 DivIsSigned, true)); 926 case ICmpInst::ICMP_NE: 927 if (LoOverflow && HiOverflow) 928 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 929 if (HiOverflow) 930 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 931 ICmpInst::ICMP_ULT, X, LoBound); 932 if (LoOverflow) 933 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 934 ICmpInst::ICMP_UGE, X, HiBound); 935 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 936 DivIsSigned, false)); 937 case ICmpInst::ICMP_ULT: 938 case ICmpInst::ICMP_SLT: 939 if (LoOverflow == +1) // Low bound is greater than input range. 940 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 941 if (LoOverflow == -1) // Low bound is less than input range. 942 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 943 return new ICmpInst(Pred, X, LoBound); 944 case ICmpInst::ICMP_UGT: 945 case ICmpInst::ICMP_SGT: 946 if (HiOverflow == +1) // High bound greater than input range. 947 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 948 if (HiOverflow == -1) // High bound less than input range. 949 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 950 if (Pred == ICmpInst::ICMP_UGT) 951 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 952 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 953 } 954} 955 956/// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)". 957Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr, 958 ConstantInt *ShAmt) { 959 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue(); 960 961 // Check that the shift amount is in range. If not, don't perform 962 // undefined shifts. When the shift is visited it will be 963 // simplified. 964 uint32_t TypeBits = CmpRHSV.getBitWidth(); 965 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 966 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 967 return 0; 968 969 if (!ICI.isEquality()) { 970 // If we have an unsigned comparison and an ashr, we can't simplify this. 971 // Similarly for signed comparisons with lshr. 972 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr)) 973 return 0; 974 975 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv 976 // by a power of 2. Since we already have logic to simplify these, 977 // transform to div and then simplify the resultant comparison. 978 if (Shr->getOpcode() == Instruction::AShr && 979 (!Shr->isExact() || ShAmtVal == TypeBits - 1)) 980 return 0; 981 982 // Revisit the shift (to delete it). 983 Worklist.Add(Shr); 984 985 Constant *DivCst = 986 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); 987 988 Value *Tmp = 989 Shr->getOpcode() == Instruction::AShr ? 990 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) : 991 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()); 992 993 ICI.setOperand(0, Tmp); 994 995 // If the builder folded the binop, just return it. 996 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp); 997 if (TheDiv == 0) 998 return &ICI; 999 1000 // Otherwise, fold this div/compare. 1001 assert(TheDiv->getOpcode() == Instruction::SDiv || 1002 TheDiv->getOpcode() == Instruction::UDiv); 1003 1004 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst)); 1005 assert(Res && "This div/cst should have folded!"); 1006 return Res; 1007 } 1008 1009 1010 // If we are comparing against bits always shifted out, the 1011 // comparison cannot succeed. 1012 APInt Comp = CmpRHSV << ShAmtVal; 1013 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp); 1014 if (Shr->getOpcode() == Instruction::LShr) 1015 Comp = Comp.lshr(ShAmtVal); 1016 else 1017 Comp = Comp.ashr(ShAmtVal); 1018 1019 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero. 1020 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1021 Constant *Cst = Builder->getInt1(IsICMP_NE); 1022 return ReplaceInstUsesWith(ICI, Cst); 1023 } 1024 1025 // Otherwise, check to see if the bits shifted out are known to be zero. 1026 // If so, we can compare against the unshifted value: 1027 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1028 if (Shr->hasOneUse() && Shr->isExact()) 1029 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS); 1030 1031 if (Shr->hasOneUse()) { 1032 // Otherwise strength reduce the shift into an and. 1033 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1034 Constant *Mask = Builder->getInt(Val); 1035 1036 Value *And = Builder->CreateAnd(Shr->getOperand(0), 1037 Mask, Shr->getName()+".mask"); 1038 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS); 1039 } 1040 return 0; 1041} 1042 1043 1044/// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 1045/// 1046Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 1047 Instruction *LHSI, 1048 ConstantInt *RHS) { 1049 const APInt &RHSV = RHS->getValue(); 1050 1051 switch (LHSI->getOpcode()) { 1052 case Instruction::Trunc: 1053 if (ICI.isEquality() && LHSI->hasOneUse()) { 1054 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1055 // of the high bits truncated out of x are known. 1056 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 1057 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1058 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 1059 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne); 1060 1061 // If all the high bits are known, we can do this xform. 1062 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 1063 // Pull in the high bits from known-ones set. 1064 APInt NewRHS = RHS->getValue().zext(SrcBits); 1065 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits); 1066 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1067 Builder->getInt(NewRHS)); 1068 } 1069 } 1070 break; 1071 1072 case Instruction::Xor: // (icmp pred (xor X, XorCst), CI) 1073 if (ConstantInt *XorCst = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1074 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1075 // fold the xor. 1076 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 1077 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 1078 Value *CompareVal = LHSI->getOperand(0); 1079 1080 // If the sign bit of the XorCst is not set, there is no change to 1081 // the operation, just stop using the Xor. 1082 if (!XorCst->isNegative()) { 1083 ICI.setOperand(0, CompareVal); 1084 Worklist.Add(LHSI); 1085 return &ICI; 1086 } 1087 1088 // Was the old condition true if the operand is positive? 1089 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 1090 1091 // If so, the new one isn't. 1092 isTrueIfPositive ^= true; 1093 1094 if (isTrueIfPositive) 1095 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 1096 SubOne(RHS)); 1097 else 1098 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 1099 AddOne(RHS)); 1100 } 1101 1102 if (LHSI->hasOneUse()) { 1103 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 1104 if (!ICI.isEquality() && XorCst->getValue().isSignBit()) { 1105 const APInt &SignBit = XorCst->getValue(); 1106 ICmpInst::Predicate Pred = ICI.isSigned() 1107 ? ICI.getUnsignedPredicate() 1108 : ICI.getSignedPredicate(); 1109 return new ICmpInst(Pred, LHSI->getOperand(0), 1110 Builder->getInt(RHSV ^ SignBit)); 1111 } 1112 1113 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 1114 if (!ICI.isEquality() && XorCst->isMaxValue(true)) { 1115 const APInt &NotSignBit = XorCst->getValue(); 1116 ICmpInst::Predicate Pred = ICI.isSigned() 1117 ? ICI.getUnsignedPredicate() 1118 : ICI.getSignedPredicate(); 1119 Pred = ICI.getSwappedPredicate(Pred); 1120 return new ICmpInst(Pred, LHSI->getOperand(0), 1121 Builder->getInt(RHSV ^ NotSignBit)); 1122 } 1123 } 1124 1125 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C) 1126 // iff -C is a power of 2 1127 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && 1128 XorCst->getValue() == ~RHSV && (RHSV + 1).isPowerOf2()) 1129 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCst); 1130 1131 // (icmp ult (xor X, C), -C) -> (icmp uge X, C) 1132 // iff -C is a power of 2 1133 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && 1134 XorCst->getValue() == -RHSV && RHSV.isPowerOf2()) 1135 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCst); 1136 } 1137 break; 1138 case Instruction::And: // (icmp pred (and X, AndCst), RHS) 1139 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1140 LHSI->getOperand(0)->hasOneUse()) { 1141 ConstantInt *AndCst = cast<ConstantInt>(LHSI->getOperand(1)); 1142 1143 // If the LHS is an AND of a truncating cast, we can widen the 1144 // and/compare to be the input width without changing the value 1145 // produced, eliminating a cast. 1146 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1147 // We can do this transformation if either the AND constant does not 1148 // have its sign bit set or if it is an equality comparison. 1149 // Extending a relational comparison when we're checking the sign 1150 // bit would not work. 1151 if (ICI.isEquality() || 1152 (!AndCst->isNegative() && RHSV.isNonNegative())) { 1153 Value *NewAnd = 1154 Builder->CreateAnd(Cast->getOperand(0), 1155 ConstantExpr::getZExt(AndCst, Cast->getSrcTy())); 1156 NewAnd->takeName(LHSI); 1157 return new ICmpInst(ICI.getPredicate(), NewAnd, 1158 ConstantExpr::getZExt(RHS, Cast->getSrcTy())); 1159 } 1160 } 1161 1162 // If the LHS is an AND of a zext, and we have an equality compare, we can 1163 // shrink the and/compare to the smaller type, eliminating the cast. 1164 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) { 1165 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy()); 1166 // Make sure we don't compare the upper bits, SimplifyDemandedBits 1167 // should fold the icmp to true/false in that case. 1168 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) { 1169 Value *NewAnd = 1170 Builder->CreateAnd(Cast->getOperand(0), 1171 ConstantExpr::getTrunc(AndCst, Ty)); 1172 NewAnd->takeName(LHSI); 1173 return new ICmpInst(ICI.getPredicate(), NewAnd, 1174 ConstantExpr::getTrunc(RHS, Ty)); 1175 } 1176 } 1177 1178 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1179 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1180 // happens a LOT in code produced by the C front-end, for bitfield 1181 // access. 1182 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1183 if (Shift && !Shift->isShift()) 1184 Shift = 0; 1185 1186 ConstantInt *ShAmt; 1187 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1188 1189 // This seemingly simple opportunity to fold away a shift turns out to 1190 // be rather complicated. See PR17827 1191 // ( http://llvm.org/bugs/show_bug.cgi?id=17827 ) for details. 1192 if (ShAmt) { 1193 bool CanFold = false; 1194 unsigned ShiftOpcode = Shift->getOpcode(); 1195 if (ShiftOpcode == Instruction::AShr) { 1196 // There may be some constraints that make this possible, 1197 // but nothing simple has been discovered yet. 1198 CanFold = false; 1199 } else if (ShiftOpcode == Instruction::Shl) { 1200 // For a left shift, we can fold if the comparison is not signed. 1201 // We can also fold a signed comparison if the mask value and 1202 // comparison value are not negative. These constraints may not be 1203 // obvious, but we can prove that they are correct using an SMT 1204 // solver. 1205 if (!ICI.isSigned() || (!AndCst->isNegative() && !RHS->isNegative())) 1206 CanFold = true; 1207 } else if (ShiftOpcode == Instruction::LShr) { 1208 // For a logical right shift, we can fold if the comparison is not 1209 // signed. We can also fold a signed comparison if the shifted mask 1210 // value and the shifted comparison value are not negative. 1211 // These constraints may not be obvious, but we can prove that they 1212 // are correct using an SMT solver. 1213 if (!ICI.isSigned()) 1214 CanFold = true; 1215 else { 1216 ConstantInt *ShiftedAndCst = 1217 cast<ConstantInt>(ConstantExpr::getShl(AndCst, ShAmt)); 1218 ConstantInt *ShiftedRHSCst = 1219 cast<ConstantInt>(ConstantExpr::getShl(RHS, ShAmt)); 1220 1221 if (!ShiftedAndCst->isNegative() && !ShiftedRHSCst->isNegative()) 1222 CanFold = true; 1223 } 1224 } 1225 1226 if (CanFold) { 1227 Constant *NewCst; 1228 if (ShiftOpcode == Instruction::Shl) 1229 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1230 else 1231 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1232 1233 // Check to see if we are shifting out any of the bits being 1234 // compared. 1235 if (ConstantExpr::get(ShiftOpcode, NewCst, ShAmt) != RHS) { 1236 // If we shifted bits out, the fold is not going to work out. 1237 // As a special case, check to see if this means that the 1238 // result is always true or false now. 1239 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1240 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 1241 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1242 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 1243 } else { 1244 ICI.setOperand(1, NewCst); 1245 Constant *NewAndCst; 1246 if (ShiftOpcode == Instruction::Shl) 1247 NewAndCst = ConstantExpr::getLShr(AndCst, ShAmt); 1248 else 1249 NewAndCst = ConstantExpr::getShl(AndCst, ShAmt); 1250 LHSI->setOperand(1, NewAndCst); 1251 LHSI->setOperand(0, Shift->getOperand(0)); 1252 Worklist.Add(Shift); // Shift is dead. 1253 return &ICI; 1254 } 1255 } 1256 } 1257 1258 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1259 // preferable because it allows the C<<Y expression to be hoisted out 1260 // of a loop if Y is invariant and X is not. 1261 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1262 ICI.isEquality() && !Shift->isArithmeticShift() && 1263 !isa<Constant>(Shift->getOperand(0))) { 1264 // Compute C << Y. 1265 Value *NS; 1266 if (Shift->getOpcode() == Instruction::LShr) { 1267 NS = Builder->CreateShl(AndCst, Shift->getOperand(1)); 1268 } else { 1269 // Insert a logical shift. 1270 NS = Builder->CreateLShr(AndCst, Shift->getOperand(1)); 1271 } 1272 1273 // Compute X & (C << Y). 1274 Value *NewAnd = 1275 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1276 1277 ICI.setOperand(0, NewAnd); 1278 return &ICI; 1279 } 1280 1281 // Replace ((X & AndCst) > RHSV) with ((X & AndCst) != 0), if any 1282 // bit set in (X & AndCst) will produce a result greater than RHSV. 1283 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) { 1284 unsigned NTZ = AndCst->getValue().countTrailingZeros(); 1285 if ((NTZ < AndCst->getBitWidth()) && 1286 APInt::getOneBitSet(AndCst->getBitWidth(), NTZ).ugt(RHSV)) 1287 return new ICmpInst(ICmpInst::ICMP_NE, LHSI, 1288 Constant::getNullValue(RHS->getType())); 1289 } 1290 } 1291 1292 // Try to optimize things like "A[i]&42 == 0" to index computations. 1293 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1294 if (GetElementPtrInst *GEP = 1295 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1296 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1297 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1298 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1299 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1300 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1301 return Res; 1302 } 1303 } 1304 1305 // X & -C == -C -> X > u ~C 1306 // X & -C != -C -> X <= u ~C 1307 // iff C is a power of 2 1308 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2()) 1309 return new ICmpInst( 1310 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT 1311 : ICmpInst::ICMP_ULE, 1312 LHSI->getOperand(0), SubOne(RHS)); 1313 break; 1314 1315 case Instruction::Or: { 1316 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1317 break; 1318 Value *P, *Q; 1319 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1320 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1321 // -> and (icmp eq P, null), (icmp eq Q, null). 1322 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1323 Constant::getNullValue(P->getType())); 1324 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1325 Constant::getNullValue(Q->getType())); 1326 Instruction *Op; 1327 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1328 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1329 else 1330 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1331 return Op; 1332 } 1333 break; 1334 } 1335 1336 case Instruction::Mul: { // (icmp pred (mul X, Val), CI) 1337 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1338 if (!Val) break; 1339 1340 // If this is a signed comparison to 0 and the mul is sign preserving, 1341 // use the mul LHS operand instead. 1342 ICmpInst::Predicate pred = ICI.getPredicate(); 1343 if (isSignTest(pred, RHS) && !Val->isZero() && 1344 cast<BinaryOperator>(LHSI)->hasNoSignedWrap()) 1345 return new ICmpInst(Val->isNegative() ? 1346 ICmpInst::getSwappedPredicate(pred) : pred, 1347 LHSI->getOperand(0), 1348 Constant::getNullValue(RHS->getType())); 1349 1350 break; 1351 } 1352 1353 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1354 uint32_t TypeBits = RHSV.getBitWidth(); 1355 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1356 if (!ShAmt) { 1357 Value *X; 1358 // (1 << X) pred P2 -> X pred Log2(P2) 1359 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) { 1360 bool RHSVIsPowerOf2 = RHSV.isPowerOf2(); 1361 ICmpInst::Predicate Pred = ICI.getPredicate(); 1362 if (ICI.isUnsigned()) { 1363 if (!RHSVIsPowerOf2) { 1364 // (1 << X) < 30 -> X <= 4 1365 // (1 << X) <= 30 -> X <= 4 1366 // (1 << X) >= 30 -> X > 4 1367 // (1 << X) > 30 -> X > 4 1368 if (Pred == ICmpInst::ICMP_ULT) 1369 Pred = ICmpInst::ICMP_ULE; 1370 else if (Pred == ICmpInst::ICMP_UGE) 1371 Pred = ICmpInst::ICMP_UGT; 1372 } 1373 unsigned RHSLog2 = RHSV.logBase2(); 1374 1375 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31 1376 // (1 << X) > 2147483648 -> X > 31 -> false 1377 // (1 << X) <= 2147483648 -> X <= 31 -> true 1378 // (1 << X) < 2147483648 -> X < 31 -> X != 31 1379 if (RHSLog2 == TypeBits-1) { 1380 if (Pred == ICmpInst::ICMP_UGE) 1381 Pred = ICmpInst::ICMP_EQ; 1382 else if (Pred == ICmpInst::ICMP_UGT) 1383 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 1384 else if (Pred == ICmpInst::ICMP_ULE) 1385 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 1386 else if (Pred == ICmpInst::ICMP_ULT) 1387 Pred = ICmpInst::ICMP_NE; 1388 } 1389 1390 return new ICmpInst(Pred, X, 1391 ConstantInt::get(RHS->getType(), RHSLog2)); 1392 } else if (ICI.isSigned()) { 1393 if (RHSV.isAllOnesValue()) { 1394 // (1 << X) <= -1 -> X == 31 1395 if (Pred == ICmpInst::ICMP_SLE) 1396 return new ICmpInst(ICmpInst::ICMP_EQ, X, 1397 ConstantInt::get(RHS->getType(), TypeBits-1)); 1398 1399 // (1 << X) > -1 -> X != 31 1400 if (Pred == ICmpInst::ICMP_SGT) 1401 return new ICmpInst(ICmpInst::ICMP_NE, X, 1402 ConstantInt::get(RHS->getType(), TypeBits-1)); 1403 } else if (!RHSV) { 1404 // (1 << X) < 0 -> X == 31 1405 // (1 << X) <= 0 -> X == 31 1406 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 1407 return new ICmpInst(ICmpInst::ICMP_EQ, X, 1408 ConstantInt::get(RHS->getType(), TypeBits-1)); 1409 1410 // (1 << X) >= 0 -> X != 31 1411 // (1 << X) > 0 -> X != 31 1412 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) 1413 return new ICmpInst(ICmpInst::ICMP_NE, X, 1414 ConstantInt::get(RHS->getType(), TypeBits-1)); 1415 } 1416 } else if (ICI.isEquality()) { 1417 if (RHSVIsPowerOf2) 1418 return new ICmpInst( 1419 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2())); 1420 1421 return ReplaceInstUsesWith( 1422 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse() 1423 : Builder->getTrue()); 1424 } 1425 } 1426 break; 1427 } 1428 1429 // Check that the shift amount is in range. If not, don't perform 1430 // undefined shifts. When the shift is visited it will be 1431 // simplified. 1432 if (ShAmt->uge(TypeBits)) 1433 break; 1434 1435 if (ICI.isEquality()) { 1436 // If we are comparing against bits always shifted out, the 1437 // comparison cannot succeed. 1438 Constant *Comp = 1439 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1440 ShAmt); 1441 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1442 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1443 Constant *Cst = Builder->getInt1(IsICMP_NE); 1444 return ReplaceInstUsesWith(ICI, Cst); 1445 } 1446 1447 // If the shift is NUW, then it is just shifting out zeros, no need for an 1448 // AND. 1449 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap()) 1450 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1451 ConstantExpr::getLShr(RHS, ShAmt)); 1452 1453 // If the shift is NSW and we compare to 0, then it is just shifting out 1454 // sign bits, no need for an AND either. 1455 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0) 1456 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1457 ConstantExpr::getLShr(RHS, ShAmt)); 1458 1459 if (LHSI->hasOneUse()) { 1460 // Otherwise strength reduce the shift into an and. 1461 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1462 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits, 1463 TypeBits - ShAmtVal)); 1464 1465 Value *And = 1466 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1467 return new ICmpInst(ICI.getPredicate(), And, 1468 ConstantExpr::getLShr(RHS, ShAmt)); 1469 } 1470 } 1471 1472 // If this is a signed comparison to 0 and the shift is sign preserving, 1473 // use the shift LHS operand instead. 1474 ICmpInst::Predicate pred = ICI.getPredicate(); 1475 if (isSignTest(pred, RHS) && 1476 cast<BinaryOperator>(LHSI)->hasNoSignedWrap()) 1477 return new ICmpInst(pred, 1478 LHSI->getOperand(0), 1479 Constant::getNullValue(RHS->getType())); 1480 1481 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1482 bool TrueIfSigned = false; 1483 if (LHSI->hasOneUse() && 1484 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1485 // (X << 31) <s 0 --> (X&1) != 0 1486 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(), 1487 APInt::getOneBitSet(TypeBits, 1488 TypeBits-ShAmt->getZExtValue()-1)); 1489 Value *And = 1490 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1491 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1492 And, Constant::getNullValue(And->getType())); 1493 } 1494 1495 // Transform (icmp pred iM (shl iM %v, N), CI) 1496 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N)) 1497 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N. 1498 // This enables to get rid of the shift in favor of a trunc which can be 1499 // free on the target. It has the additional benefit of comparing to a 1500 // smaller constant, which will be target friendly. 1501 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1); 1502 if (LHSI->hasOneUse() && 1503 Amt != 0 && RHSV.countTrailingZeros() >= Amt) { 1504 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt); 1505 Constant *NCI = ConstantExpr::getTrunc( 1506 ConstantExpr::getAShr(RHS, 1507 ConstantInt::get(RHS->getType(), Amt)), 1508 NTy); 1509 return new ICmpInst(ICI.getPredicate(), 1510 Builder->CreateTrunc(LHSI->getOperand(0), NTy), 1511 NCI); 1512 } 1513 1514 break; 1515 } 1516 1517 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1518 case Instruction::AShr: { 1519 // Handle equality comparisons of shift-by-constant. 1520 BinaryOperator *BO = cast<BinaryOperator>(LHSI); 1521 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1522 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt)) 1523 return Res; 1524 } 1525 1526 // Handle exact shr's. 1527 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) { 1528 if (RHSV.isMinValue()) 1529 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS); 1530 } 1531 break; 1532 } 1533 1534 case Instruction::SDiv: 1535 case Instruction::UDiv: 1536 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1537 // Fold this div into the comparison, producing a range check. 1538 // Determine, based on the divide type, what the range is being 1539 // checked. If there is an overflow on the low or high side, remember 1540 // it, otherwise compute the range [low, hi) bounding the new value. 1541 // See: InsertRangeTest above for the kinds of replacements possible. 1542 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1543 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1544 DivRHS)) 1545 return R; 1546 break; 1547 1548 case Instruction::Sub: { 1549 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0)); 1550 if (!LHSC) break; 1551 const APInt &LHSV = LHSC->getValue(); 1552 1553 // C1-X <u C2 -> (X|(C2-1)) == C1 1554 // iff C1 & (C2-1) == C2-1 1555 // C2 is a power of 2 1556 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() && 1557 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1)) 1558 return new ICmpInst(ICmpInst::ICMP_EQ, 1559 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1), 1560 LHSC); 1561 1562 // C1-X >u C2 -> (X|C2) != C1 1563 // iff C1 & C2 == C2 1564 // C2+1 is a power of 2 1565 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() && 1566 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV) 1567 return new ICmpInst(ICmpInst::ICMP_NE, 1568 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC); 1569 break; 1570 } 1571 1572 case Instruction::Add: 1573 // Fold: icmp pred (add X, C1), C2 1574 if (!ICI.isEquality()) { 1575 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1576 if (!LHSC) break; 1577 const APInt &LHSV = LHSC->getValue(); 1578 1579 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1580 .subtract(LHSV); 1581 1582 if (ICI.isSigned()) { 1583 if (CR.getLower().isSignBit()) { 1584 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1585 Builder->getInt(CR.getUpper())); 1586 } else if (CR.getUpper().isSignBit()) { 1587 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1588 Builder->getInt(CR.getLower())); 1589 } 1590 } else { 1591 if (CR.getLower().isMinValue()) { 1592 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1593 Builder->getInt(CR.getUpper())); 1594 } else if (CR.getUpper().isMinValue()) { 1595 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1596 Builder->getInt(CR.getLower())); 1597 } 1598 } 1599 1600 // X-C1 <u C2 -> (X & -C2) == C1 1601 // iff C1 & (C2-1) == 0 1602 // C2 is a power of 2 1603 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() && 1604 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0) 1605 return new ICmpInst(ICmpInst::ICMP_EQ, 1606 Builder->CreateAnd(LHSI->getOperand(0), -RHSV), 1607 ConstantExpr::getNeg(LHSC)); 1608 1609 // X-C1 >u C2 -> (X & ~C2) != C1 1610 // iff C1 & C2 == 0 1611 // C2+1 is a power of 2 1612 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() && 1613 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0) 1614 return new ICmpInst(ICmpInst::ICMP_NE, 1615 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV), 1616 ConstantExpr::getNeg(LHSC)); 1617 } 1618 break; 1619 } 1620 1621 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1622 if (ICI.isEquality()) { 1623 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1624 1625 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1626 // the second operand is a constant, simplify a bit. 1627 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1628 switch (BO->getOpcode()) { 1629 case Instruction::SRem: 1630 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1631 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1632 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1633 if (V.sgt(1) && V.isPowerOf2()) { 1634 Value *NewRem = 1635 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1636 BO->getName()); 1637 return new ICmpInst(ICI.getPredicate(), NewRem, 1638 Constant::getNullValue(BO->getType())); 1639 } 1640 } 1641 break; 1642 case Instruction::Add: 1643 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1644 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1645 if (BO->hasOneUse()) 1646 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1647 ConstantExpr::getSub(RHS, BOp1C)); 1648 } else if (RHSV == 0) { 1649 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1650 // efficiently invertible, or if the add has just this one use. 1651 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1652 1653 if (Value *NegVal = dyn_castNegVal(BOp1)) 1654 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1655 if (Value *NegVal = dyn_castNegVal(BOp0)) 1656 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1657 if (BO->hasOneUse()) { 1658 Value *Neg = Builder->CreateNeg(BOp1); 1659 Neg->takeName(BO); 1660 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1661 } 1662 } 1663 break; 1664 case Instruction::Xor: 1665 // For the xor case, we can xor two constants together, eliminating 1666 // the explicit xor. 1667 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) { 1668 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1669 ConstantExpr::getXor(RHS, BOC)); 1670 } else if (RHSV == 0) { 1671 // Replace ((xor A, B) != 0) with (A != B) 1672 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1673 BO->getOperand(1)); 1674 } 1675 break; 1676 case Instruction::Sub: 1677 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants. 1678 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) { 1679 if (BO->hasOneUse()) 1680 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1), 1681 ConstantExpr::getSub(BOp0C, RHS)); 1682 } else if (RHSV == 0) { 1683 // Replace ((sub A, B) != 0) with (A != B) 1684 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1685 BO->getOperand(1)); 1686 } 1687 break; 1688 case Instruction::Or: 1689 // If bits are being or'd in that are not present in the constant we 1690 // are comparing against, then the comparison could never succeed! 1691 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1692 Constant *NotCI = ConstantExpr::getNot(RHS); 1693 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1694 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE)); 1695 } 1696 break; 1697 1698 case Instruction::And: 1699 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1700 // If bits are being compared against that are and'd out, then the 1701 // comparison can never succeed! 1702 if ((RHSV & ~BOC->getValue()) != 0) 1703 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE)); 1704 1705 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1706 if (RHS == BOC && RHSV.isPowerOf2()) 1707 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1708 ICmpInst::ICMP_NE, LHSI, 1709 Constant::getNullValue(RHS->getType())); 1710 1711 // Don't perform the following transforms if the AND has multiple uses 1712 if (!BO->hasOneUse()) 1713 break; 1714 1715 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1716 if (BOC->getValue().isSignBit()) { 1717 Value *X = BO->getOperand(0); 1718 Constant *Zero = Constant::getNullValue(X->getType()); 1719 ICmpInst::Predicate pred = isICMP_NE ? 1720 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1721 return new ICmpInst(pred, X, Zero); 1722 } 1723 1724 // ((X & ~7) == 0) --> X < 8 1725 if (RHSV == 0 && isHighOnes(BOC)) { 1726 Value *X = BO->getOperand(0); 1727 Constant *NegX = ConstantExpr::getNeg(BOC); 1728 ICmpInst::Predicate pred = isICMP_NE ? 1729 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1730 return new ICmpInst(pred, X, NegX); 1731 } 1732 } 1733 break; 1734 case Instruction::Mul: 1735 if (RHSV == 0 && BO->hasNoSignedWrap()) { 1736 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1737 // The trivial case (mul X, 0) is handled by InstSimplify 1738 // General case : (mul X, C) != 0 iff X != 0 1739 // (mul X, C) == 0 iff X == 0 1740 if (!BOC->isZero()) 1741 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1742 Constant::getNullValue(RHS->getType())); 1743 } 1744 } 1745 break; 1746 default: break; 1747 } 1748 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1749 // Handle icmp {eq|ne} <intrinsic>, intcst. 1750 switch (II->getIntrinsicID()) { 1751 case Intrinsic::bswap: 1752 Worklist.Add(II); 1753 ICI.setOperand(0, II->getArgOperand(0)); 1754 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap())); 1755 return &ICI; 1756 case Intrinsic::ctlz: 1757 case Intrinsic::cttz: 1758 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1759 if (RHSV == RHS->getType()->getBitWidth()) { 1760 Worklist.Add(II); 1761 ICI.setOperand(0, II->getArgOperand(0)); 1762 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1763 return &ICI; 1764 } 1765 break; 1766 case Intrinsic::ctpop: 1767 // popcount(A) == 0 -> A == 0 and likewise for != 1768 if (RHS->isZero()) { 1769 Worklist.Add(II); 1770 ICI.setOperand(0, II->getArgOperand(0)); 1771 ICI.setOperand(1, RHS); 1772 return &ICI; 1773 } 1774 break; 1775 default: 1776 break; 1777 } 1778 } 1779 } 1780 return 0; 1781} 1782 1783/// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1784/// We only handle extending casts so far. 1785/// 1786Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1787 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1788 Value *LHSCIOp = LHSCI->getOperand(0); 1789 Type *SrcTy = LHSCIOp->getType(); 1790 Type *DestTy = LHSCI->getType(); 1791 Value *RHSCIOp; 1792 1793 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1794 // integer type is the same size as the pointer type. 1795 if (DL && LHSCI->getOpcode() == Instruction::PtrToInt && 1796 DL->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) { 1797 Value *RHSOp = 0; 1798 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1799 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1800 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1801 RHSOp = RHSC->getOperand(0); 1802 // If the pointer types don't match, insert a bitcast. 1803 if (LHSCIOp->getType() != RHSOp->getType()) 1804 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1805 } 1806 1807 if (RHSOp) 1808 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1809 } 1810 1811 // The code below only handles extension cast instructions, so far. 1812 // Enforce this. 1813 if (LHSCI->getOpcode() != Instruction::ZExt && 1814 LHSCI->getOpcode() != Instruction::SExt) 1815 return 0; 1816 1817 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1818 bool isSignedCmp = ICI.isSigned(); 1819 1820 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1821 // Not an extension from the same type? 1822 RHSCIOp = CI->getOperand(0); 1823 if (RHSCIOp->getType() != LHSCIOp->getType()) 1824 return 0; 1825 1826 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1827 // and the other is a zext), then we can't handle this. 1828 if (CI->getOpcode() != LHSCI->getOpcode()) 1829 return 0; 1830 1831 // Deal with equality cases early. 1832 if (ICI.isEquality()) 1833 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1834 1835 // A signed comparison of sign extended values simplifies into a 1836 // signed comparison. 1837 if (isSignedCmp && isSignedExt) 1838 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1839 1840 // The other three cases all fold into an unsigned comparison. 1841 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1842 } 1843 1844 // If we aren't dealing with a constant on the RHS, exit early 1845 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1846 if (!CI) 1847 return 0; 1848 1849 // Compute the constant that would happen if we truncated to SrcTy then 1850 // reextended to DestTy. 1851 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1852 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1853 Res1, DestTy); 1854 1855 // If the re-extended constant didn't change... 1856 if (Res2 == CI) { 1857 // Deal with equality cases early. 1858 if (ICI.isEquality()) 1859 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1860 1861 // A signed comparison of sign extended values simplifies into a 1862 // signed comparison. 1863 if (isSignedExt && isSignedCmp) 1864 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1865 1866 // The other three cases all fold into an unsigned comparison. 1867 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1868 } 1869 1870 // The re-extended constant changed so the constant cannot be represented 1871 // in the shorter type. Consequently, we cannot emit a simple comparison. 1872 // All the cases that fold to true or false will have already been handled 1873 // by SimplifyICmpInst, so only deal with the tricky case. 1874 1875 if (isSignedCmp || !isSignedExt) 1876 return 0; 1877 1878 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1879 // should have been folded away previously and not enter in here. 1880 1881 // We're performing an unsigned comp with a sign extended value. 1882 // This is true if the input is >= 0. [aka >s -1] 1883 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1884 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1885 1886 // Finally, return the value computed. 1887 if (ICI.getPredicate() == ICmpInst::ICMP_ULT) 1888 return ReplaceInstUsesWith(ICI, Result); 1889 1890 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 1891 return BinaryOperator::CreateNot(Result); 1892} 1893 1894/// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form: 1895/// I = icmp ugt (add (add A, B), CI2), CI1 1896/// If this is of the form: 1897/// sum = a + b 1898/// if (sum+128 >u 255) 1899/// Then replace it with llvm.sadd.with.overflow.i8. 1900/// 1901static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1902 ConstantInt *CI2, ConstantInt *CI1, 1903 InstCombiner &IC) { 1904 // The transformation we're trying to do here is to transform this into an 1905 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1906 // with a narrower add, and discard the add-with-constant that is part of the 1907 // range check (if we can't eliminate it, this isn't profitable). 1908 1909 // In order to eliminate the add-with-constant, the compare can be its only 1910 // use. 1911 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1912 if (!AddWithCst->hasOneUse()) return 0; 1913 1914 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1915 if (!CI2->getValue().isPowerOf2()) return 0; 1916 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1917 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0; 1918 1919 // The width of the new add formed is 1 more than the bias. 1920 ++NewWidth; 1921 1922 // Check to see that CI1 is an all-ones value with NewWidth bits. 1923 if (CI1->getBitWidth() == NewWidth || 1924 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1925 return 0; 1926 1927 // This is only really a signed overflow check if the inputs have been 1928 // sign-extended; check for that condition. For example, if CI2 is 2^31 and 1929 // the operands of the add are 64 bits wide, we need at least 33 sign bits. 1930 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; 1931 if (IC.ComputeNumSignBits(A) < NeededSignBits || 1932 IC.ComputeNumSignBits(B) < NeededSignBits) 1933 return 0; 1934 1935 // In order to replace the original add with a narrower 1936 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1937 // and truncates that discard the high bits of the add. Verify that this is 1938 // the case. 1939 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1940 for (User *U : OrigAdd->users()) { 1941 if (U == AddWithCst) continue; 1942 1943 // Only accept truncates for now. We would really like a nice recursive 1944 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1945 // chain to see which bits of a value are actually demanded. If the 1946 // original add had another add which was then immediately truncated, we 1947 // could still do the transformation. 1948 TruncInst *TI = dyn_cast<TruncInst>(U); 1949 if (TI == 0 || 1950 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0; 1951 } 1952 1953 // If the pattern matches, truncate the inputs to the narrower type and 1954 // use the sadd_with_overflow intrinsic to efficiently compute both the 1955 // result and the overflow bit. 1956 Module *M = I.getParent()->getParent()->getParent(); 1957 1958 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1959 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow, 1960 NewType); 1961 1962 InstCombiner::BuilderTy *Builder = IC.Builder; 1963 1964 // Put the new code above the original add, in case there are any uses of the 1965 // add between the add and the compare. 1966 Builder->SetInsertPoint(OrigAdd); 1967 1968 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc"); 1969 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc"); 1970 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd"); 1971 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); 1972 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); 1973 1974 // The inner add was the result of the narrow add, zero extended to the 1975 // wider type. Replace it with the result computed by the intrinsic. 1976 IC.ReplaceInstUsesWith(*OrigAdd, ZExt); 1977 1978 // The original icmp gets replaced with the overflow value. 1979 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1980} 1981 1982static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV, 1983 InstCombiner &IC) { 1984 // Don't bother doing this transformation for pointers, don't do it for 1985 // vectors. 1986 if (!isa<IntegerType>(OrigAddV->getType())) return 0; 1987 1988 // If the add is a constant expr, then we don't bother transforming it. 1989 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV); 1990 if (OrigAdd == 0) return 0; 1991 1992 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1); 1993 1994 // Put the new code above the original add, in case there are any uses of the 1995 // add between the add and the compare. 1996 InstCombiner::BuilderTy *Builder = IC.Builder; 1997 Builder->SetInsertPoint(OrigAdd); 1998 1999 Module *M = I.getParent()->getParent()->getParent(); 2000 Type *Ty = LHS->getType(); 2001 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); 2002 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd"); 2003 Value *Add = Builder->CreateExtractValue(Call, 0); 2004 2005 IC.ReplaceInstUsesWith(*OrigAdd, Add); 2006 2007 // The original icmp gets replaced with the overflow value. 2008 return ExtractValueInst::Create(Call, 1, "uadd.overflow"); 2009} 2010 2011// DemandedBitsLHSMask - When performing a comparison against a constant, 2012// it is possible that not all the bits in the LHS are demanded. This helper 2013// method computes the mask that IS demanded. 2014static APInt DemandedBitsLHSMask(ICmpInst &I, 2015 unsigned BitWidth, bool isSignCheck) { 2016 if (isSignCheck) 2017 return APInt::getSignBit(BitWidth); 2018 2019 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1)); 2020 if (!CI) return APInt::getAllOnesValue(BitWidth); 2021 const APInt &RHS = CI->getValue(); 2022 2023 switch (I.getPredicate()) { 2024 // For a UGT comparison, we don't care about any bits that 2025 // correspond to the trailing ones of the comparand. The value of these 2026 // bits doesn't impact the outcome of the comparison, because any value 2027 // greater than the RHS must differ in a bit higher than these due to carry. 2028 case ICmpInst::ICMP_UGT: { 2029 unsigned trailingOnes = RHS.countTrailingOnes(); 2030 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); 2031 return ~lowBitsSet; 2032 } 2033 2034 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 2035 // Any value less than the RHS must differ in a higher bit because of carries. 2036 case ICmpInst::ICMP_ULT: { 2037 unsigned trailingZeros = RHS.countTrailingZeros(); 2038 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); 2039 return ~lowBitsSet; 2040 } 2041 2042 default: 2043 return APInt::getAllOnesValue(BitWidth); 2044 } 2045 2046} 2047 2048/// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst 2049/// should be swapped. 2050/// The decision is based on how many times these two operands are reused 2051/// as subtract operands and their positions in those instructions. 2052/// The rational is that several architectures use the same instruction for 2053/// both subtract and cmp, thus it is better if the order of those operands 2054/// match. 2055/// \return true if Op0 and Op1 should be swapped. 2056static bool swapMayExposeCSEOpportunities(const Value * Op0, 2057 const Value * Op1) { 2058 // Filter out pointer value as those cannot appears directly in subtract. 2059 // FIXME: we may want to go through inttoptrs or bitcasts. 2060 if (Op0->getType()->isPointerTy()) 2061 return false; 2062 // Count every uses of both Op0 and Op1 in a subtract. 2063 // Each time Op0 is the first operand, count -1: swapping is bad, the 2064 // subtract has already the same layout as the compare. 2065 // Each time Op0 is the second operand, count +1: swapping is good, the 2066 // subtract has a different layout as the compare. 2067 // At the end, if the benefit is greater than 0, Op0 should come second to 2068 // expose more CSE opportunities. 2069 int GlobalSwapBenefits = 0; 2070 for (const User *U : Op0->users()) { 2071 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(U); 2072 if (!BinOp || BinOp->getOpcode() != Instruction::Sub) 2073 continue; 2074 // If Op0 is the first argument, this is not beneficial to swap the 2075 // arguments. 2076 int LocalSwapBenefits = -1; 2077 unsigned Op1Idx = 1; 2078 if (BinOp->getOperand(Op1Idx) == Op0) { 2079 Op1Idx = 0; 2080 LocalSwapBenefits = 1; 2081 } 2082 if (BinOp->getOperand(Op1Idx) != Op1) 2083 continue; 2084 GlobalSwapBenefits += LocalSwapBenefits; 2085 } 2086 return GlobalSwapBenefits > 0; 2087} 2088 2089Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 2090 bool Changed = false; 2091 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2092 unsigned Op0Cplxity = getComplexity(Op0); 2093 unsigned Op1Cplxity = getComplexity(Op1); 2094 2095 /// Orders the operands of the compare so that they are listed from most 2096 /// complex to least complex. This puts constants before unary operators, 2097 /// before binary operators. 2098 if (Op0Cplxity < Op1Cplxity || 2099 (Op0Cplxity == Op1Cplxity && 2100 swapMayExposeCSEOpportunities(Op0, Op1))) { 2101 I.swapOperands(); 2102 std::swap(Op0, Op1); 2103 Changed = true; 2104 } 2105 2106 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, DL)) 2107 return ReplaceInstUsesWith(I, V); 2108 2109 // comparing -val or val with non-zero is the same as just comparing val 2110 // ie, abs(val) != 0 -> val != 0 2111 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) 2112 { 2113 Value *Cond, *SelectTrue, *SelectFalse; 2114 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue), 2115 m_Value(SelectFalse)))) { 2116 if (Value *V = dyn_castNegVal(SelectTrue)) { 2117 if (V == SelectFalse) 2118 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 2119 } 2120 else if (Value *V = dyn_castNegVal(SelectFalse)) { 2121 if (V == SelectTrue) 2122 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 2123 } 2124 } 2125 } 2126 2127 Type *Ty = Op0->getType(); 2128 2129 // icmp's with boolean values can always be turned into bitwise operations 2130 if (Ty->isIntegerTy(1)) { 2131 switch (I.getPredicate()) { 2132 default: llvm_unreachable("Invalid icmp instruction!"); 2133 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 2134 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 2135 return BinaryOperator::CreateNot(Xor); 2136 } 2137 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 2138 return BinaryOperator::CreateXor(Op0, Op1); 2139 2140 case ICmpInst::ICMP_UGT: 2141 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 2142 // FALL THROUGH 2143 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 2144 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 2145 return BinaryOperator::CreateAnd(Not, Op1); 2146 } 2147 case ICmpInst::ICMP_SGT: 2148 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 2149 // FALL THROUGH 2150 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 2151 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 2152 return BinaryOperator::CreateAnd(Not, Op0); 2153 } 2154 case ICmpInst::ICMP_UGE: 2155 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 2156 // FALL THROUGH 2157 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 2158 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 2159 return BinaryOperator::CreateOr(Not, Op1); 2160 } 2161 case ICmpInst::ICMP_SGE: 2162 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 2163 // FALL THROUGH 2164 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 2165 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 2166 return BinaryOperator::CreateOr(Not, Op0); 2167 } 2168 } 2169 } 2170 2171 unsigned BitWidth = 0; 2172 if (Ty->isIntOrIntVectorTy()) 2173 BitWidth = Ty->getScalarSizeInBits(); 2174 else if (DL) // Pointers require DL info to get their size. 2175 BitWidth = DL->getTypeSizeInBits(Ty->getScalarType()); 2176 2177 bool isSignBit = false; 2178 2179 // See if we are doing a comparison with a constant. 2180 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2181 Value *A = 0, *B = 0; 2182 2183 // Match the following pattern, which is a common idiom when writing 2184 // overflow-safe integer arithmetic function. The source performs an 2185 // addition in wider type, and explicitly checks for overflow using 2186 // comparisons against INT_MIN and INT_MAX. Simplify this by using the 2187 // sadd_with_overflow intrinsic. 2188 // 2189 // TODO: This could probably be generalized to handle other overflow-safe 2190 // operations if we worked out the formulas to compute the appropriate 2191 // magic constants. 2192 // 2193 // sum = a + b 2194 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 2195 { 2196 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 2197 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2198 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 2199 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this)) 2200 return Res; 2201 } 2202 2203 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 2204 if (I.isEquality() && CI->isZero() && 2205 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 2206 // (icmp cond A B) if cond is equality 2207 return new ICmpInst(I.getPredicate(), A, B); 2208 } 2209 2210 // If we have an icmp le or icmp ge instruction, turn it into the 2211 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 2212 // them being folded in the code below. The SimplifyICmpInst code has 2213 // already handled the edge cases for us, so we just assert on them. 2214 switch (I.getPredicate()) { 2215 default: break; 2216 case ICmpInst::ICMP_ULE: 2217 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 2218 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 2219 Builder->getInt(CI->getValue()+1)); 2220 case ICmpInst::ICMP_SLE: 2221 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 2222 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2223 Builder->getInt(CI->getValue()+1)); 2224 case ICmpInst::ICMP_UGE: 2225 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 2226 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 2227 Builder->getInt(CI->getValue()-1)); 2228 case ICmpInst::ICMP_SGE: 2229 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 2230 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2231 Builder->getInt(CI->getValue()-1)); 2232 } 2233 2234 // If this comparison is a normal comparison, it demands all 2235 // bits, if it is a sign bit comparison, it only demands the sign bit. 2236 bool UnusedBit; 2237 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 2238 } 2239 2240 // See if we can fold the comparison based on range information we can get 2241 // by checking whether bits are known to be zero or one in the input. 2242 if (BitWidth != 0) { 2243 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 2244 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 2245 2246 if (SimplifyDemandedBits(I.getOperandUse(0), 2247 DemandedBitsLHSMask(I, BitWidth, isSignBit), 2248 Op0KnownZero, Op0KnownOne, 0)) 2249 return &I; 2250 if (SimplifyDemandedBits(I.getOperandUse(1), 2251 APInt::getAllOnesValue(BitWidth), 2252 Op1KnownZero, Op1KnownOne, 0)) 2253 return &I; 2254 2255 // Given the known and unknown bits, compute a range that the LHS could be 2256 // in. Compute the Min, Max and RHS values based on the known bits. For the 2257 // EQ and NE we use unsigned values. 2258 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 2259 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 2260 if (I.isSigned()) { 2261 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2262 Op0Min, Op0Max); 2263 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2264 Op1Min, Op1Max); 2265 } else { 2266 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2267 Op0Min, Op0Max); 2268 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2269 Op1Min, Op1Max); 2270 } 2271 2272 // If Min and Max are known to be the same, then SimplifyDemandedBits 2273 // figured out that the LHS is a constant. Just constant fold this now so 2274 // that code below can assume that Min != Max. 2275 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 2276 return new ICmpInst(I.getPredicate(), 2277 ConstantInt::get(Op0->getType(), Op0Min), Op1); 2278 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 2279 return new ICmpInst(I.getPredicate(), Op0, 2280 ConstantInt::get(Op1->getType(), Op1Min)); 2281 2282 // Based on the range information we know about the LHS, see if we can 2283 // simplify this comparison. For example, (x&4) < 8 is always true. 2284 switch (I.getPredicate()) { 2285 default: llvm_unreachable("Unknown icmp opcode!"); 2286 case ICmpInst::ICMP_EQ: { 2287 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2288 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2289 2290 // If all bits are known zero except for one, then we know at most one 2291 // bit is set. If the comparison is against zero, then this is a check 2292 // to see if *that* bit is set. 2293 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2294 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2295 // If the LHS is an AND with the same constant, look through it. 2296 Value *LHS = 0; 2297 ConstantInt *LHSC = 0; 2298 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2299 LHSC->getValue() != Op0KnownZeroInverted) 2300 LHS = Op0; 2301 2302 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2303 // then turn "((1 << x)&8) == 0" into "x != 3". 2304 Value *X = 0; 2305 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2306 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2307 return new ICmpInst(ICmpInst::ICMP_NE, X, 2308 ConstantInt::get(X->getType(), CmpVal)); 2309 } 2310 2311 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2312 // then turn "((8 >>u x)&1) == 0" into "x != 3". 2313 const APInt *CI; 2314 if (Op0KnownZeroInverted == 1 && 2315 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2316 return new ICmpInst(ICmpInst::ICMP_NE, X, 2317 ConstantInt::get(X->getType(), 2318 CI->countTrailingZeros())); 2319 } 2320 2321 break; 2322 } 2323 case ICmpInst::ICMP_NE: { 2324 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2325 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2326 2327 // If all bits are known zero except for one, then we know at most one 2328 // bit is set. If the comparison is against zero, then this is a check 2329 // to see if *that* bit is set. 2330 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2331 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2332 // If the LHS is an AND with the same constant, look through it. 2333 Value *LHS = 0; 2334 ConstantInt *LHSC = 0; 2335 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2336 LHSC->getValue() != Op0KnownZeroInverted) 2337 LHS = Op0; 2338 2339 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2340 // then turn "((1 << x)&8) != 0" into "x == 3". 2341 Value *X = 0; 2342 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2343 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2344 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2345 ConstantInt::get(X->getType(), CmpVal)); 2346 } 2347 2348 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2349 // then turn "((8 >>u x)&1) != 0" into "x == 3". 2350 const APInt *CI; 2351 if (Op0KnownZeroInverted == 1 && 2352 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2353 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2354 ConstantInt::get(X->getType(), 2355 CI->countTrailingZeros())); 2356 } 2357 2358 break; 2359 } 2360 case ICmpInst::ICMP_ULT: 2361 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 2362 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2363 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 2364 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2365 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 2366 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2367 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2368 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 2369 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2370 Builder->getInt(CI->getValue()-1)); 2371 2372 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 2373 if (CI->isMinValue(true)) 2374 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2375 Constant::getAllOnesValue(Op0->getType())); 2376 } 2377 break; 2378 case ICmpInst::ICMP_UGT: 2379 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 2380 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2381 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 2382 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2383 2384 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 2385 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2386 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2387 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 2388 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2389 Builder->getInt(CI->getValue()+1)); 2390 2391 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 2392 if (CI->isMaxValue(true)) 2393 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2394 Constant::getNullValue(Op0->getType())); 2395 } 2396 break; 2397 case ICmpInst::ICMP_SLT: 2398 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 2399 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2400 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 2401 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2402 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 2403 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2404 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2405 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 2406 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2407 Builder->getInt(CI->getValue()-1)); 2408 } 2409 break; 2410 case ICmpInst::ICMP_SGT: 2411 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 2412 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2413 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 2414 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2415 2416 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 2417 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2418 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2419 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 2420 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2421 Builder->getInt(CI->getValue()+1)); 2422 } 2423 break; 2424 case ICmpInst::ICMP_SGE: 2425 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 2426 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 2427 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2428 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 2429 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2430 break; 2431 case ICmpInst::ICMP_SLE: 2432 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 2433 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 2434 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2435 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 2436 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2437 break; 2438 case ICmpInst::ICMP_UGE: 2439 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 2440 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 2441 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2442 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 2443 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2444 break; 2445 case ICmpInst::ICMP_ULE: 2446 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 2447 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 2448 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2449 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 2450 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2451 break; 2452 } 2453 2454 // Turn a signed comparison into an unsigned one if both operands 2455 // are known to have the same sign. 2456 if (I.isSigned() && 2457 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 2458 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 2459 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 2460 } 2461 2462 // Test if the ICmpInst instruction is used exclusively by a select as 2463 // part of a minimum or maximum operation. If so, refrain from doing 2464 // any other folding. This helps out other analyses which understand 2465 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 2466 // and CodeGen. And in this case, at least one of the comparison 2467 // operands has at least one user besides the compare (the select), 2468 // which would often largely negate the benefit of folding anyway. 2469 if (I.hasOneUse()) 2470 if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin())) 2471 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 2472 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 2473 return 0; 2474 2475 // See if we are doing a comparison between a constant and an instruction that 2476 // can be folded into the comparison. 2477 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2478 // Since the RHS is a ConstantInt (CI), if the left hand side is an 2479 // instruction, see if that instruction also has constants so that the 2480 // instruction can be folded into the icmp 2481 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2482 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 2483 return Res; 2484 } 2485 2486 // Handle icmp with constant (but not simple integer constant) RHS 2487 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2488 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2489 switch (LHSI->getOpcode()) { 2490 case Instruction::GetElementPtr: 2491 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2492 if (RHSC->isNullValue() && 2493 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2494 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2495 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2496 break; 2497 case Instruction::PHI: 2498 // Only fold icmp into the PHI if the phi and icmp are in the same 2499 // block. If in the same block, we're encouraging jump threading. If 2500 // not, we are just pessimizing the code by making an i1 phi. 2501 if (LHSI->getParent() == I.getParent()) 2502 if (Instruction *NV = FoldOpIntoPhi(I)) 2503 return NV; 2504 break; 2505 case Instruction::Select: { 2506 // If either operand of the select is a constant, we can fold the 2507 // comparison into the select arms, which will cause one to be 2508 // constant folded and the select turned into a bitwise or. 2509 Value *Op1 = 0, *Op2 = 0; 2510 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 2511 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2512 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 2513 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2514 2515 // We only want to perform this transformation if it will not lead to 2516 // additional code. This is true if either both sides of the select 2517 // fold to a constant (in which case the icmp is replaced with a select 2518 // which will usually simplify) or this is the only user of the 2519 // select (in which case we are trading a select+icmp for a simpler 2520 // select+icmp). 2521 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 2522 if (!Op1) 2523 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 2524 RHSC, I.getName()); 2525 if (!Op2) 2526 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 2527 RHSC, I.getName()); 2528 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2529 } 2530 break; 2531 } 2532 case Instruction::IntToPtr: 2533 // icmp pred inttoptr(X), null -> icmp pred X, 0 2534 if (RHSC->isNullValue() && DL && 2535 DL->getIntPtrType(RHSC->getType()) == 2536 LHSI->getOperand(0)->getType()) 2537 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2538 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2539 break; 2540 2541 case Instruction::Load: 2542 // Try to optimize things like "A[i] > 4" to index computations. 2543 if (GetElementPtrInst *GEP = 2544 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2545 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2546 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2547 !cast<LoadInst>(LHSI)->isVolatile()) 2548 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2549 return Res; 2550 } 2551 break; 2552 } 2553 } 2554 2555 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 2556 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 2557 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 2558 return NI; 2559 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 2560 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 2561 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 2562 return NI; 2563 2564 // Test to see if the operands of the icmp are casted versions of other 2565 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 2566 // now. 2567 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 2568 if (Op0->getType()->isPointerTy() && 2569 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 2570 // We keep moving the cast from the left operand over to the right 2571 // operand, where it can often be eliminated completely. 2572 Op0 = CI->getOperand(0); 2573 2574 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 2575 // so eliminate it as well. 2576 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2577 Op1 = CI2->getOperand(0); 2578 2579 // If Op1 is a constant, we can fold the cast into the constant. 2580 if (Op0->getType() != Op1->getType()) { 2581 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2582 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2583 } else { 2584 // Otherwise, cast the RHS right before the icmp 2585 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2586 } 2587 } 2588 return new ICmpInst(I.getPredicate(), Op0, Op1); 2589 } 2590 } 2591 2592 if (isa<CastInst>(Op0)) { 2593 // Handle the special case of: icmp (cast bool to X), <cst> 2594 // This comes up when you have code like 2595 // int X = A < B; 2596 // if (X) ... 2597 // For generality, we handle any zero-extension of any operand comparison 2598 // with a constant or another cast from the same type. 2599 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2600 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2601 return R; 2602 } 2603 2604 // Special logic for binary operators. 2605 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0); 2606 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1); 2607 if (BO0 || BO1) { 2608 CmpInst::Predicate Pred = I.getPredicate(); 2609 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; 2610 if (BO0 && isa<OverflowingBinaryOperator>(BO0)) 2611 NoOp0WrapProblem = ICmpInst::isEquality(Pred) || 2612 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) || 2613 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap()); 2614 if (BO1 && isa<OverflowingBinaryOperator>(BO1)) 2615 NoOp1WrapProblem = ICmpInst::isEquality(Pred) || 2616 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) || 2617 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap()); 2618 2619 // Analyze the case when either Op0 or Op1 is an add instruction. 2620 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). 2621 Value *A = 0, *B = 0, *C = 0, *D = 0; 2622 if (BO0 && BO0->getOpcode() == Instruction::Add) 2623 A = BO0->getOperand(0), B = BO0->getOperand(1); 2624 if (BO1 && BO1->getOpcode() == Instruction::Add) 2625 C = BO1->getOperand(0), D = BO1->getOperand(1); 2626 2627 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2628 if ((A == Op1 || B == Op1) && NoOp0WrapProblem) 2629 return new ICmpInst(Pred, A == Op1 ? B : A, 2630 Constant::getNullValue(Op1->getType())); 2631 2632 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2633 if ((C == Op0 || D == Op0) && NoOp1WrapProblem) 2634 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()), 2635 C == Op0 ? D : C); 2636 2637 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow. 2638 if (A && C && (A == C || A == D || B == C || B == D) && 2639 NoOp0WrapProblem && NoOp1WrapProblem && 2640 // Try not to increase register pressure. 2641 BO0->hasOneUse() && BO1->hasOneUse()) { 2642 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2643 Value *Y, *Z; 2644 if (A == C) { 2645 // C + B == C + D -> B == D 2646 Y = B; 2647 Z = D; 2648 } else if (A == D) { 2649 // D + B == C + D -> B == C 2650 Y = B; 2651 Z = C; 2652 } else if (B == C) { 2653 // A + C == C + D -> A == D 2654 Y = A; 2655 Z = D; 2656 } else { 2657 assert(B == D); 2658 // A + D == C + D -> A == C 2659 Y = A; 2660 Z = C; 2661 } 2662 return new ICmpInst(Pred, Y, Z); 2663 } 2664 2665 // icmp slt (X + -1), Y -> icmp sle X, Y 2666 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT && 2667 match(B, m_AllOnes())) 2668 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1); 2669 2670 // icmp sge (X + -1), Y -> icmp sgt X, Y 2671 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE && 2672 match(B, m_AllOnes())) 2673 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1); 2674 2675 // icmp sle (X + 1), Y -> icmp slt X, Y 2676 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && 2677 match(B, m_One())) 2678 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1); 2679 2680 // icmp sgt (X + 1), Y -> icmp sge X, Y 2681 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && 2682 match(B, m_One())) 2683 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1); 2684 2685 // if C1 has greater magnitude than C2: 2686 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y 2687 // s.t. C3 = C1 - C2 2688 // 2689 // if C2 has greater magnitude than C1: 2690 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3) 2691 // s.t. C3 = C2 - C1 2692 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem && 2693 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) 2694 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B)) 2695 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) { 2696 const APInt &AP1 = C1->getValue(); 2697 const APInt &AP2 = C2->getValue(); 2698 if (AP1.isNegative() == AP2.isNegative()) { 2699 APInt AP1Abs = C1->getValue().abs(); 2700 APInt AP2Abs = C2->getValue().abs(); 2701 if (AP1Abs.uge(AP2Abs)) { 2702 ConstantInt *C3 = Builder->getInt(AP1 - AP2); 2703 Value *NewAdd = Builder->CreateNSWAdd(A, C3); 2704 return new ICmpInst(Pred, NewAdd, C); 2705 } else { 2706 ConstantInt *C3 = Builder->getInt(AP2 - AP1); 2707 Value *NewAdd = Builder->CreateNSWAdd(C, C3); 2708 return new ICmpInst(Pred, A, NewAdd); 2709 } 2710 } 2711 } 2712 2713 2714 // Analyze the case when either Op0 or Op1 is a sub instruction. 2715 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). 2716 A = 0; B = 0; C = 0; D = 0; 2717 if (BO0 && BO0->getOpcode() == Instruction::Sub) 2718 A = BO0->getOperand(0), B = BO0->getOperand(1); 2719 if (BO1 && BO1->getOpcode() == Instruction::Sub) 2720 C = BO1->getOperand(0), D = BO1->getOperand(1); 2721 2722 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow. 2723 if (A == Op1 && NoOp0WrapProblem) 2724 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B); 2725 2726 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow. 2727 if (C == Op0 && NoOp1WrapProblem) 2728 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType())); 2729 2730 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow. 2731 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem && 2732 // Try not to increase register pressure. 2733 BO0->hasOneUse() && BO1->hasOneUse()) 2734 return new ICmpInst(Pred, A, C); 2735 2736 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow. 2737 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem && 2738 // Try not to increase register pressure. 2739 BO0->hasOneUse() && BO1->hasOneUse()) 2740 return new ICmpInst(Pred, D, B); 2741 2742 BinaryOperator *SRem = NULL; 2743 // icmp (srem X, Y), Y 2744 if (BO0 && BO0->getOpcode() == Instruction::SRem && 2745 Op1 == BO0->getOperand(1)) 2746 SRem = BO0; 2747 // icmp Y, (srem X, Y) 2748 else if (BO1 && BO1->getOpcode() == Instruction::SRem && 2749 Op0 == BO1->getOperand(1)) 2750 SRem = BO1; 2751 if (SRem) { 2752 // We don't check hasOneUse to avoid increasing register pressure because 2753 // the value we use is the same value this instruction was already using. 2754 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) { 2755 default: break; 2756 case ICmpInst::ICMP_EQ: 2757 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2758 case ICmpInst::ICMP_NE: 2759 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2760 case ICmpInst::ICMP_SGT: 2761 case ICmpInst::ICMP_SGE: 2762 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1), 2763 Constant::getAllOnesValue(SRem->getType())); 2764 case ICmpInst::ICMP_SLT: 2765 case ICmpInst::ICMP_SLE: 2766 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1), 2767 Constant::getNullValue(SRem->getType())); 2768 } 2769 } 2770 2771 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && 2772 BO0->hasOneUse() && BO1->hasOneUse() && 2773 BO0->getOperand(1) == BO1->getOperand(1)) { 2774 switch (BO0->getOpcode()) { 2775 default: break; 2776 case Instruction::Add: 2777 case Instruction::Sub: 2778 case Instruction::Xor: 2779 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2780 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2781 BO1->getOperand(0)); 2782 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2783 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2784 if (CI->getValue().isSignBit()) { 2785 ICmpInst::Predicate Pred = I.isSigned() 2786 ? I.getUnsignedPredicate() 2787 : I.getSignedPredicate(); 2788 return new ICmpInst(Pred, BO0->getOperand(0), 2789 BO1->getOperand(0)); 2790 } 2791 2792 if (CI->isMaxValue(true)) { 2793 ICmpInst::Predicate Pred = I.isSigned() 2794 ? I.getUnsignedPredicate() 2795 : I.getSignedPredicate(); 2796 Pred = I.getSwappedPredicate(Pred); 2797 return new ICmpInst(Pred, BO0->getOperand(0), 2798 BO1->getOperand(0)); 2799 } 2800 } 2801 break; 2802 case Instruction::Mul: 2803 if (!I.isEquality()) 2804 break; 2805 2806 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2807 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2808 // Mask = -1 >> count-trailing-zeros(Cst). 2809 if (!CI->isZero() && !CI->isOne()) { 2810 const APInt &AP = CI->getValue(); 2811 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2812 APInt::getLowBitsSet(AP.getBitWidth(), 2813 AP.getBitWidth() - 2814 AP.countTrailingZeros())); 2815 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask); 2816 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask); 2817 return new ICmpInst(I.getPredicate(), And1, And2); 2818 } 2819 } 2820 break; 2821 case Instruction::UDiv: 2822 case Instruction::LShr: 2823 if (I.isSigned()) 2824 break; 2825 // fall-through 2826 case Instruction::SDiv: 2827 case Instruction::AShr: 2828 if (!BO0->isExact() || !BO1->isExact()) 2829 break; 2830 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2831 BO1->getOperand(0)); 2832 case Instruction::Shl: { 2833 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap(); 2834 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap(); 2835 if (!NUW && !NSW) 2836 break; 2837 if (!NSW && I.isSigned()) 2838 break; 2839 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2840 BO1->getOperand(0)); 2841 } 2842 } 2843 } 2844 } 2845 2846 { Value *A, *B; 2847 // Transform (A & ~B) == 0 --> (A & B) != 0 2848 // and (A & ~B) != 0 --> (A & B) == 0 2849 // if A is a power of 2. 2850 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) && 2851 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality()) 2852 return new ICmpInst(I.getInversePredicate(), 2853 Builder->CreateAnd(A, B), 2854 Op1); 2855 2856 // ~x < ~y --> y < x 2857 // ~x < cst --> ~cst < x 2858 if (match(Op0, m_Not(m_Value(A)))) { 2859 if (match(Op1, m_Not(m_Value(B)))) 2860 return new ICmpInst(I.getPredicate(), B, A); 2861 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) 2862 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A); 2863 } 2864 2865 // (a+b) <u a --> llvm.uadd.with.overflow. 2866 // (a+b) <u b --> llvm.uadd.with.overflow. 2867 if (I.getPredicate() == ICmpInst::ICMP_ULT && 2868 match(Op0, m_Add(m_Value(A), m_Value(B))) && 2869 (Op1 == A || Op1 == B)) 2870 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this)) 2871 return R; 2872 2873 // a >u (a+b) --> llvm.uadd.with.overflow. 2874 // b >u (a+b) --> llvm.uadd.with.overflow. 2875 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2876 match(Op1, m_Add(m_Value(A), m_Value(B))) && 2877 (Op0 == A || Op0 == B)) 2878 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this)) 2879 return R; 2880 } 2881 2882 if (I.isEquality()) { 2883 Value *A, *B, *C, *D; 2884 2885 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2886 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2887 Value *OtherVal = A == Op1 ? B : A; 2888 return new ICmpInst(I.getPredicate(), OtherVal, 2889 Constant::getNullValue(A->getType())); 2890 } 2891 2892 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2893 // A^c1 == C^c2 --> A == C^(c1^c2) 2894 ConstantInt *C1, *C2; 2895 if (match(B, m_ConstantInt(C1)) && 2896 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2897 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue()); 2898 Value *Xor = Builder->CreateXor(C, NC); 2899 return new ICmpInst(I.getPredicate(), A, Xor); 2900 } 2901 2902 // A^B == A^D -> B == D 2903 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2904 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2905 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2906 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2907 } 2908 } 2909 2910 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2911 (A == Op0 || B == Op0)) { 2912 // A == (A^B) -> B == 0 2913 Value *OtherVal = A == Op0 ? B : A; 2914 return new ICmpInst(I.getPredicate(), OtherVal, 2915 Constant::getNullValue(A->getType())); 2916 } 2917 2918 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2919 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) && 2920 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) { 2921 Value *X = 0, *Y = 0, *Z = 0; 2922 2923 if (A == C) { 2924 X = B; Y = D; Z = A; 2925 } else if (A == D) { 2926 X = B; Y = C; Z = A; 2927 } else if (B == C) { 2928 X = A; Y = D; Z = B; 2929 } else if (B == D) { 2930 X = A; Y = C; Z = B; 2931 } 2932 2933 if (X) { // Build (X^Y) & Z 2934 Op1 = Builder->CreateXor(X, Y); 2935 Op1 = Builder->CreateAnd(Op1, Z); 2936 I.setOperand(0, Op1); 2937 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2938 return &I; 2939 } 2940 } 2941 2942 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B) 2943 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B) 2944 ConstantInt *Cst1; 2945 if ((Op0->hasOneUse() && 2946 match(Op0, m_ZExt(m_Value(A))) && 2947 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) || 2948 (Op1->hasOneUse() && 2949 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) && 2950 match(Op1, m_ZExt(m_Value(A))))) { 2951 APInt Pow2 = Cst1->getValue() + 1; 2952 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) && 2953 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth()) 2954 return new ICmpInst(I.getPredicate(), A, 2955 Builder->CreateTrunc(B, A->getType())); 2956 } 2957 2958 // (A >> C) == (B >> C) --> (A^B) u< (1 << C) 2959 // For lshr and ashr pairs. 2960 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) && 2961 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) || 2962 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) && 2963 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) { 2964 unsigned TypeBits = Cst1->getBitWidth(); 2965 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits); 2966 if (ShAmt < TypeBits && ShAmt != 0) { 2967 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE 2968 ? ICmpInst::ICMP_UGE 2969 : ICmpInst::ICMP_ULT; 2970 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted"); 2971 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt); 2972 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal)); 2973 } 2974 } 2975 2976 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to 2977 // "icmp (and X, mask), cst" 2978 uint64_t ShAmt = 0; 2979 if (Op0->hasOneUse() && 2980 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), 2981 m_ConstantInt(ShAmt))))) && 2982 match(Op1, m_ConstantInt(Cst1)) && 2983 // Only do this when A has multiple uses. This is most important to do 2984 // when it exposes other optimizations. 2985 !A->hasOneUse()) { 2986 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits(); 2987 2988 if (ShAmt < ASize) { 2989 APInt MaskV = 2990 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits()); 2991 MaskV <<= ShAmt; 2992 2993 APInt CmpV = Cst1->getValue().zext(ASize); 2994 CmpV <<= ShAmt; 2995 2996 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV)); 2997 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV)); 2998 } 2999 } 3000 } 3001 3002 { 3003 Value *X; ConstantInt *Cst; 3004 // icmp X+Cst, X 3005 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 3006 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate()); 3007 3008 // icmp X, X+Cst 3009 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 3010 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate()); 3011 } 3012 return Changed ? &I : 0; 3013} 3014 3015/// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 3016/// 3017Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 3018 Instruction *LHSI, 3019 Constant *RHSC) { 3020 if (!isa<ConstantFP>(RHSC)) return 0; 3021 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 3022 3023 // Get the width of the mantissa. We don't want to hack on conversions that 3024 // might lose information from the integer, e.g. "i64 -> float" 3025 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 3026 if (MantissaWidth == -1) return 0; // Unknown. 3027 3028 // Check to see that the input is converted from an integer type that is small 3029 // enough that preserves all bits. TODO: check here for "known" sign bits. 3030 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 3031 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 3032 3033 // If this is a uitofp instruction, we need an extra bit to hold the sign. 3034 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 3035 if (LHSUnsigned) 3036 ++InputSize; 3037 3038 // If the conversion would lose info, don't hack on this. 3039 if ((int)InputSize > MantissaWidth) 3040 return 0; 3041 3042 // Otherwise, we can potentially simplify the comparison. We know that it 3043 // will always come through as an integer value and we know the constant is 3044 // not a NAN (it would have been previously simplified). 3045 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 3046 3047 ICmpInst::Predicate Pred; 3048 switch (I.getPredicate()) { 3049 default: llvm_unreachable("Unexpected predicate!"); 3050 case FCmpInst::FCMP_UEQ: 3051 case FCmpInst::FCMP_OEQ: 3052 Pred = ICmpInst::ICMP_EQ; 3053 break; 3054 case FCmpInst::FCMP_UGT: 3055 case FCmpInst::FCMP_OGT: 3056 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 3057 break; 3058 case FCmpInst::FCMP_UGE: 3059 case FCmpInst::FCMP_OGE: 3060 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 3061 break; 3062 case FCmpInst::FCMP_ULT: 3063 case FCmpInst::FCMP_OLT: 3064 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 3065 break; 3066 case FCmpInst::FCMP_ULE: 3067 case FCmpInst::FCMP_OLE: 3068 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 3069 break; 3070 case FCmpInst::FCMP_UNE: 3071 case FCmpInst::FCMP_ONE: 3072 Pred = ICmpInst::ICMP_NE; 3073 break; 3074 case FCmpInst::FCMP_ORD: 3075 return ReplaceInstUsesWith(I, Builder->getTrue()); 3076 case FCmpInst::FCMP_UNO: 3077 return ReplaceInstUsesWith(I, Builder->getFalse()); 3078 } 3079 3080 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 3081 3082 // Now we know that the APFloat is a normal number, zero or inf. 3083 3084 // See if the FP constant is too large for the integer. For example, 3085 // comparing an i8 to 300.0. 3086 unsigned IntWidth = IntTy->getScalarSizeInBits(); 3087 3088 if (!LHSUnsigned) { 3089 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 3090 // and large values. 3091 APFloat SMax(RHS.getSemantics()); 3092 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 3093 APFloat::rmNearestTiesToEven); 3094 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 3095 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 3096 Pred == ICmpInst::ICMP_SLE) 3097 return ReplaceInstUsesWith(I, Builder->getTrue()); 3098 return ReplaceInstUsesWith(I, Builder->getFalse()); 3099 } 3100 } else { 3101 // If the RHS value is > UnsignedMax, fold the comparison. This handles 3102 // +INF and large values. 3103 APFloat UMax(RHS.getSemantics()); 3104 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 3105 APFloat::rmNearestTiesToEven); 3106 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 3107 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 3108 Pred == ICmpInst::ICMP_ULE) 3109 return ReplaceInstUsesWith(I, Builder->getTrue()); 3110 return ReplaceInstUsesWith(I, Builder->getFalse()); 3111 } 3112 } 3113 3114 if (!LHSUnsigned) { 3115 // See if the RHS value is < SignedMin. 3116 APFloat SMin(RHS.getSemantics()); 3117 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 3118 APFloat::rmNearestTiesToEven); 3119 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 3120 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 3121 Pred == ICmpInst::ICMP_SGE) 3122 return ReplaceInstUsesWith(I, Builder->getTrue()); 3123 return ReplaceInstUsesWith(I, Builder->getFalse()); 3124 } 3125 } else { 3126 // See if the RHS value is < UnsignedMin. 3127 APFloat SMin(RHS.getSemantics()); 3128 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true, 3129 APFloat::rmNearestTiesToEven); 3130 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0 3131 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || 3132 Pred == ICmpInst::ICMP_UGE) 3133 return ReplaceInstUsesWith(I, Builder->getTrue()); 3134 return ReplaceInstUsesWith(I, Builder->getFalse()); 3135 } 3136 } 3137 3138 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 3139 // [0, UMAX], but it may still be fractional. See if it is fractional by 3140 // casting the FP value to the integer value and back, checking for equality. 3141 // Don't do this for zero, because -0.0 is not fractional. 3142 Constant *RHSInt = LHSUnsigned 3143 ? ConstantExpr::getFPToUI(RHSC, IntTy) 3144 : ConstantExpr::getFPToSI(RHSC, IntTy); 3145 if (!RHS.isZero()) { 3146 bool Equal = LHSUnsigned 3147 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 3148 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 3149 if (!Equal) { 3150 // If we had a comparison against a fractional value, we have to adjust 3151 // the compare predicate and sometimes the value. RHSC is rounded towards 3152 // zero at this point. 3153 switch (Pred) { 3154 default: llvm_unreachable("Unexpected integer comparison!"); 3155 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 3156 return ReplaceInstUsesWith(I, Builder->getTrue()); 3157 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 3158 return ReplaceInstUsesWith(I, Builder->getFalse()); 3159 case ICmpInst::ICMP_ULE: 3160 // (float)int <= 4.4 --> int <= 4 3161 // (float)int <= -4.4 --> false 3162 if (RHS.isNegative()) 3163 return ReplaceInstUsesWith(I, Builder->getFalse()); 3164 break; 3165 case ICmpInst::ICMP_SLE: 3166 // (float)int <= 4.4 --> int <= 4 3167 // (float)int <= -4.4 --> int < -4 3168 if (RHS.isNegative()) 3169 Pred = ICmpInst::ICMP_SLT; 3170 break; 3171 case ICmpInst::ICMP_ULT: 3172 // (float)int < -4.4 --> false 3173 // (float)int < 4.4 --> int <= 4 3174 if (RHS.isNegative()) 3175 return ReplaceInstUsesWith(I, Builder->getFalse()); 3176 Pred = ICmpInst::ICMP_ULE; 3177 break; 3178 case ICmpInst::ICMP_SLT: 3179 // (float)int < -4.4 --> int < -4 3180 // (float)int < 4.4 --> int <= 4 3181 if (!RHS.isNegative()) 3182 Pred = ICmpInst::ICMP_SLE; 3183 break; 3184 case ICmpInst::ICMP_UGT: 3185 // (float)int > 4.4 --> int > 4 3186 // (float)int > -4.4 --> true 3187 if (RHS.isNegative()) 3188 return ReplaceInstUsesWith(I, Builder->getTrue()); 3189 break; 3190 case ICmpInst::ICMP_SGT: 3191 // (float)int > 4.4 --> int > 4 3192 // (float)int > -4.4 --> int >= -4 3193 if (RHS.isNegative()) 3194 Pred = ICmpInst::ICMP_SGE; 3195 break; 3196 case ICmpInst::ICMP_UGE: 3197 // (float)int >= -4.4 --> true 3198 // (float)int >= 4.4 --> int > 4 3199 if (RHS.isNegative()) 3200 return ReplaceInstUsesWith(I, Builder->getTrue()); 3201 Pred = ICmpInst::ICMP_UGT; 3202 break; 3203 case ICmpInst::ICMP_SGE: 3204 // (float)int >= -4.4 --> int >= -4 3205 // (float)int >= 4.4 --> int > 4 3206 if (!RHS.isNegative()) 3207 Pred = ICmpInst::ICMP_SGT; 3208 break; 3209 } 3210 } 3211 } 3212 3213 // Lower this FP comparison into an appropriate integer version of the 3214 // comparison. 3215 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 3216} 3217 3218Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 3219 bool Changed = false; 3220 3221 /// Orders the operands of the compare so that they are listed from most 3222 /// complex to least complex. This puts constants before unary operators, 3223 /// before binary operators. 3224 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 3225 I.swapOperands(); 3226 Changed = true; 3227 } 3228 3229 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3230 3231 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, DL)) 3232 return ReplaceInstUsesWith(I, V); 3233 3234 // Simplify 'fcmp pred X, X' 3235 if (Op0 == Op1) { 3236 switch (I.getPredicate()) { 3237 default: llvm_unreachable("Unknown predicate!"); 3238 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 3239 case FCmpInst::FCMP_ULT: // True if unordered or less than 3240 case FCmpInst::FCMP_UGT: // True if unordered or greater than 3241 case FCmpInst::FCMP_UNE: // True if unordered or not equal 3242 // Canonicalize these to be 'fcmp uno %X, 0.0'. 3243 I.setPredicate(FCmpInst::FCMP_UNO); 3244 I.setOperand(1, Constant::getNullValue(Op0->getType())); 3245 return &I; 3246 3247 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 3248 case FCmpInst::FCMP_OEQ: // True if ordered and equal 3249 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 3250 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 3251 // Canonicalize these to be 'fcmp ord %X, 0.0'. 3252 I.setPredicate(FCmpInst::FCMP_ORD); 3253 I.setOperand(1, Constant::getNullValue(Op0->getType())); 3254 return &I; 3255 } 3256 } 3257 3258 // Handle fcmp with constant RHS 3259 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 3260 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 3261 switch (LHSI->getOpcode()) { 3262 case Instruction::FPExt: { 3263 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless 3264 FPExtInst *LHSExt = cast<FPExtInst>(LHSI); 3265 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC); 3266 if (!RHSF) 3267 break; 3268 3269 const fltSemantics *Sem; 3270 // FIXME: This shouldn't be here. 3271 if (LHSExt->getSrcTy()->isHalfTy()) 3272 Sem = &APFloat::IEEEhalf; 3273 else if (LHSExt->getSrcTy()->isFloatTy()) 3274 Sem = &APFloat::IEEEsingle; 3275 else if (LHSExt->getSrcTy()->isDoubleTy()) 3276 Sem = &APFloat::IEEEdouble; 3277 else if (LHSExt->getSrcTy()->isFP128Ty()) 3278 Sem = &APFloat::IEEEquad; 3279 else if (LHSExt->getSrcTy()->isX86_FP80Ty()) 3280 Sem = &APFloat::x87DoubleExtended; 3281 else if (LHSExt->getSrcTy()->isPPC_FP128Ty()) 3282 Sem = &APFloat::PPCDoubleDouble; 3283 else 3284 break; 3285 3286 bool Lossy; 3287 APFloat F = RHSF->getValueAPF(); 3288 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy); 3289 3290 // Avoid lossy conversions and denormals. Zero is a special case 3291 // that's OK to convert. 3292 APFloat Fabs = F; 3293 Fabs.clearSign(); 3294 if (!Lossy && 3295 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) != 3296 APFloat::cmpLessThan) || Fabs.isZero())) 3297 3298 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3299 ConstantFP::get(RHSC->getContext(), F)); 3300 break; 3301 } 3302 case Instruction::PHI: 3303 // Only fold fcmp into the PHI if the phi and fcmp are in the same 3304 // block. If in the same block, we're encouraging jump threading. If 3305 // not, we are just pessimizing the code by making an i1 phi. 3306 if (LHSI->getParent() == I.getParent()) 3307 if (Instruction *NV = FoldOpIntoPhi(I)) 3308 return NV; 3309 break; 3310 case Instruction::SIToFP: 3311 case Instruction::UIToFP: 3312 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 3313 return NV; 3314 break; 3315 case Instruction::FSub: { 3316 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C 3317 Value *Op; 3318 if (match(LHSI, m_FNeg(m_Value(Op)))) 3319 return new FCmpInst(I.getSwappedPredicate(), Op, 3320 ConstantExpr::getFNeg(RHSC)); 3321 break; 3322 } 3323 case Instruction::Load: 3324 if (GetElementPtrInst *GEP = 3325 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 3326 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 3327 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 3328 !cast<LoadInst>(LHSI)->isVolatile()) 3329 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 3330 return Res; 3331 } 3332 break; 3333 case Instruction::Call: { 3334 CallInst *CI = cast<CallInst>(LHSI); 3335 LibFunc::Func Func; 3336 // Various optimization for fabs compared with zero. 3337 if (RHSC->isNullValue() && CI->getCalledFunction() && 3338 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) && 3339 TLI->has(Func)) { 3340 if (Func == LibFunc::fabs || Func == LibFunc::fabsf || 3341 Func == LibFunc::fabsl) { 3342 switch (I.getPredicate()) { 3343 default: break; 3344 // fabs(x) < 0 --> false 3345 case FCmpInst::FCMP_OLT: 3346 return ReplaceInstUsesWith(I, Builder->getFalse()); 3347 // fabs(x) > 0 --> x != 0 3348 case FCmpInst::FCMP_OGT: 3349 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), 3350 RHSC); 3351 // fabs(x) <= 0 --> x == 0 3352 case FCmpInst::FCMP_OLE: 3353 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), 3354 RHSC); 3355 // fabs(x) >= 0 --> !isnan(x) 3356 case FCmpInst::FCMP_OGE: 3357 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), 3358 RHSC); 3359 // fabs(x) == 0 --> x == 0 3360 // fabs(x) != 0 --> x != 0 3361 case FCmpInst::FCMP_OEQ: 3362 case FCmpInst::FCMP_UEQ: 3363 case FCmpInst::FCMP_ONE: 3364 case FCmpInst::FCMP_UNE: 3365 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), 3366 RHSC); 3367 } 3368 } 3369 } 3370 } 3371 } 3372 } 3373 3374 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y 3375 Value *X, *Y; 3376 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 3377 return new FCmpInst(I.getSwappedPredicate(), X, Y); 3378 3379 // fcmp (fpext x), (fpext y) -> fcmp x, y 3380 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0)) 3381 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1)) 3382 if (LHSExt->getSrcTy() == RHSExt->getSrcTy()) 3383 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3384 RHSExt->getOperand(0)); 3385 3386 return Changed ? &I : 0; 3387} 3388