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/DataLayout.h" 19 #include "llvm/IR/IntrinsicInst.h" 20 #include "llvm/Support/ConstantRange.h" 21 #include "llvm/Support/GetElementPtrTypeIterator.h" 22 #include "llvm/Support/PatternMatch.h" 23 #include "llvm/Target/TargetLibraryInfo.h" 24 using namespace llvm; 25 using namespace PatternMatch; 26 27 static ConstantInt *getOne(Constant *C) { 28 return ConstantInt::get(cast<IntegerType>(C->getType()), 1); 29 } 30 31 /// AddOne - Add one to a ConstantInt 32 static Constant *AddOne(Constant *C) { 33 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 34 } 35 /// SubOne - Subtract one from a ConstantInt 36 static Constant *SubOne(Constant *C) { 37 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 38 } 39 40 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) { 41 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx)); 42 } 43 44 static bool HasAddOverflow(ConstantInt *Result, 45 ConstantInt *In1, ConstantInt *In2, 46 bool IsSigned) { 47 if (!IsSigned) 48 return Result->getValue().ult(In1->getValue()); 49 50 if (In2->isNegative()) 51 return Result->getValue().sgt(In1->getValue()); 52 return Result->getValue().slt(In1->getValue()); 53 } 54 55 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result 56 /// overflowed for this type. 57 static bool AddWithOverflow(Constant *&Result, Constant *In1, 58 Constant *In2, bool IsSigned = false) { 59 Result = ConstantExpr::getAdd(In1, In2); 60 61 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 62 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 63 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 64 if (HasAddOverflow(ExtractElement(Result, Idx), 65 ExtractElement(In1, Idx), 66 ExtractElement(In2, Idx), 67 IsSigned)) 68 return true; 69 } 70 return false; 71 } 72 73 return HasAddOverflow(cast<ConstantInt>(Result), 74 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 75 IsSigned); 76 } 77 78 static bool HasSubOverflow(ConstantInt *Result, 79 ConstantInt *In1, ConstantInt *In2, 80 bool IsSigned) { 81 if (!IsSigned) 82 return Result->getValue().ugt(In1->getValue()); 83 84 if (In2->isNegative()) 85 return Result->getValue().slt(In1->getValue()); 86 87 return Result->getValue().sgt(In1->getValue()); 88 } 89 90 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result 91 /// overflowed for this type. 92 static bool SubWithOverflow(Constant *&Result, Constant *In1, 93 Constant *In2, bool IsSigned = false) { 94 Result = ConstantExpr::getSub(In1, In2); 95 96 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) { 97 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 98 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); 99 if (HasSubOverflow(ExtractElement(Result, Idx), 100 ExtractElement(In1, Idx), 101 ExtractElement(In2, Idx), 102 IsSigned)) 103 return true; 104 } 105 return false; 106 } 107 108 return HasSubOverflow(cast<ConstantInt>(Result), 109 cast<ConstantInt>(In1), cast<ConstantInt>(In2), 110 IsSigned); 111 } 112 113 /// isSignBitCheck - Given an exploded icmp instruction, return true if the 114 /// comparison only checks the sign bit. If it only checks the sign bit, set 115 /// TrueIfSigned if the result of the comparison is true when the input value is 116 /// signed. 117 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS, 118 bool &TrueIfSigned) { 119 switch (pred) { 120 case ICmpInst::ICMP_SLT: // True if LHS s< 0 121 TrueIfSigned = true; 122 return RHS->isZero(); 123 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 124 TrueIfSigned = true; 125 return RHS->isAllOnesValue(); 126 case ICmpInst::ICMP_SGT: // True if LHS s> -1 127 TrueIfSigned = false; 128 return RHS->isAllOnesValue(); 129 case ICmpInst::ICMP_UGT: 130 // True if LHS u> RHS and RHS == high-bit-mask - 1 131 TrueIfSigned = true; 132 return RHS->isMaxValue(true); 133 case ICmpInst::ICMP_UGE: 134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 135 TrueIfSigned = true; 136 return RHS->getValue().isSignBit(); 137 default: 138 return false; 139 } 140 } 141 142 /// Returns true if the exploded icmp can be expressed as a signed comparison 143 /// to zero and updates the predicate accordingly. 144 /// The signedness of the comparison is preserved. 145 static bool isSignTest(ICmpInst::Predicate &pred, const ConstantInt *RHS) { 146 if (!ICmpInst::isSigned(pred)) 147 return false; 148 149 if (RHS->isZero()) 150 return ICmpInst::isRelational(pred); 151 152 if (RHS->isOne()) { 153 if (pred == ICmpInst::ICMP_SLT) { 154 pred = ICmpInst::ICMP_SLE; 155 return true; 156 } 157 } else if (RHS->isAllOnesValue()) { 158 if (pred == ICmpInst::ICMP_SGT) { 159 pred = ICmpInst::ICMP_SGE; 160 return true; 161 } 162 } 163 164 return false; 165 } 166 167 // isHighOnes - Return true if the constant is of the form 1+0+. 168 // This is the same as lowones(~X). 169 static bool isHighOnes(const ConstantInt *CI) { 170 return (~CI->getValue() + 1).isPowerOf2(); 171 } 172 173 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a 174 /// set of known zero and one bits, compute the maximum and minimum values that 175 /// could have the specified known zero and known one bits, returning them in 176 /// min/max. 177 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero, 178 const APInt& KnownOne, 179 APInt& Min, APInt& Max) { 180 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 181 KnownZero.getBitWidth() == Min.getBitWidth() && 182 KnownZero.getBitWidth() == Max.getBitWidth() && 183 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 184 APInt UnknownBits = ~(KnownZero|KnownOne); 185 186 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 187 // bit if it is unknown. 188 Min = KnownOne; 189 Max = KnownOne|UnknownBits; 190 191 if (UnknownBits.isNegative()) { // Sign bit is unknown 192 Min.setBit(Min.getBitWidth()-1); 193 Max.clearBit(Max.getBitWidth()-1); 194 } 195 } 196 197 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and 198 // a set of known zero and one bits, compute the maximum and minimum values that 199 // could have the specified known zero and known one bits, returning them in 200 // min/max. 201 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, 202 const APInt &KnownOne, 203 APInt &Min, APInt &Max) { 204 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && 205 KnownZero.getBitWidth() == Min.getBitWidth() && 206 KnownZero.getBitWidth() == Max.getBitWidth() && 207 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 208 APInt UnknownBits = ~(KnownZero|KnownOne); 209 210 // The minimum value is when the unknown bits are all zeros. 211 Min = KnownOne; 212 // The maximum value is when the unknown bits are all ones. 213 Max = KnownOne|UnknownBits; 214 } 215 216 217 218 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern: 219 /// cmp pred (load (gep GV, ...)), cmpcst 220 /// where GV is a global variable with a constant initializer. Try to simplify 221 /// this into some simple computation that does not need the load. For example 222 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 223 /// 224 /// If AndCst is non-null, then the loaded value is masked with that constant 225 /// before doing the comparison. This handles cases like "A[i]&4 == 0". 226 Instruction *InstCombiner:: 227 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV, 228 CmpInst &ICI, ConstantInt *AndCst) { 229 // We need TD information to know the pointer size unless this is inbounds. 230 if (!GEP->isInBounds() && TD == 0) 231 return 0; 232 233 Constant *Init = GV->getInitializer(); 234 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init)) 235 return 0; 236 237 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); 238 if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays. 239 240 // There are many forms of this optimization we can handle, for now, just do 241 // the simple index into a single-dimensional array. 242 // 243 // Require: GEP GV, 0, i {{, constant indices}} 244 if (GEP->getNumOperands() < 3 || 245 !isa<ConstantInt>(GEP->getOperand(1)) || 246 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 247 isa<Constant>(GEP->getOperand(2))) 248 return 0; 249 250 // Check that indices after the variable are constants and in-range for the 251 // type they index. Collect the indices. This is typically for arrays of 252 // structs. 253 SmallVector<unsigned, 4> LaterIndices; 254 255 Type *EltTy = Init->getType()->getArrayElementType(); 256 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 257 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 258 if (Idx == 0) return 0; // Variable index. 259 260 uint64_t IdxVal = Idx->getZExtValue(); 261 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index. 262 263 if (StructType *STy = dyn_cast<StructType>(EltTy)) 264 EltTy = STy->getElementType(IdxVal); 265 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 266 if (IdxVal >= ATy->getNumElements()) return 0; 267 EltTy = ATy->getElementType(); 268 } else { 269 return 0; // Unknown type. 270 } 271 272 LaterIndices.push_back(IdxVal); 273 } 274 275 enum { Overdefined = -3, Undefined = -2 }; 276 277 // Variables for our state machines. 278 279 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 280 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 281 // and 87 is the second (and last) index. FirstTrueElement is -2 when 282 // undefined, otherwise set to the first true element. SecondTrueElement is 283 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 284 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 285 286 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 287 // form "i != 47 & i != 87". Same state transitions as for true elements. 288 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 289 290 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 291 /// define a state machine that triggers for ranges of values that the index 292 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 293 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 294 /// index in the range (inclusive). We use -2 for undefined here because we 295 /// use relative comparisons and don't want 0-1 to match -1. 296 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 297 298 // MagicBitvector - This is a magic bitvector where we set a bit if the 299 // comparison is true for element 'i'. If there are 64 elements or less in 300 // the array, this will fully represent all the comparison results. 301 uint64_t MagicBitvector = 0; 302 303 304 // Scan the array and see if one of our patterns matches. 305 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 306 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { 307 Constant *Elt = Init->getAggregateElement(i); 308 if (Elt == 0) return 0; 309 310 // If this is indexing an array of structures, get the structure element. 311 if (!LaterIndices.empty()) 312 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); 313 314 // If the element is masked, handle it. 315 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 316 317 // Find out if the comparison would be true or false for the i'th element. 318 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 319 CompareRHS, TD, TLI); 320 // If the result is undef for this element, ignore it. 321 if (isa<UndefValue>(C)) { 322 // Extend range state machines to cover this element in case there is an 323 // undef in the middle of the range. 324 if (TrueRangeEnd == (int)i-1) 325 TrueRangeEnd = i; 326 if (FalseRangeEnd == (int)i-1) 327 FalseRangeEnd = i; 328 continue; 329 } 330 331 // If we can't compute the result for any of the elements, we have to give 332 // up evaluating the entire conditional. 333 if (!isa<ConstantInt>(C)) return 0; 334 335 // Otherwise, we know if the comparison is true or false for this element, 336 // update our state machines. 337 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 338 339 // State machine for single/double/range index comparison. 340 if (IsTrueForElt) { 341 // Update the TrueElement state machine. 342 if (FirstTrueElement == Undefined) 343 FirstTrueElement = TrueRangeEnd = i; // First true element. 344 else { 345 // Update double-compare state machine. 346 if (SecondTrueElement == Undefined) 347 SecondTrueElement = i; 348 else 349 SecondTrueElement = Overdefined; 350 351 // Update range state machine. 352 if (TrueRangeEnd == (int)i-1) 353 TrueRangeEnd = i; 354 else 355 TrueRangeEnd = Overdefined; 356 } 357 } else { 358 // Update the FalseElement state machine. 359 if (FirstFalseElement == Undefined) 360 FirstFalseElement = FalseRangeEnd = i; // First false element. 361 else { 362 // Update double-compare state machine. 363 if (SecondFalseElement == Undefined) 364 SecondFalseElement = i; 365 else 366 SecondFalseElement = Overdefined; 367 368 // Update range state machine. 369 if (FalseRangeEnd == (int)i-1) 370 FalseRangeEnd = i; 371 else 372 FalseRangeEnd = Overdefined; 373 } 374 } 375 376 377 // If this element is in range, update our magic bitvector. 378 if (i < 64 && IsTrueForElt) 379 MagicBitvector |= 1ULL << i; 380 381 // If all of our states become overdefined, bail out early. Since the 382 // predicate is expensive, only check it every 8 elements. This is only 383 // really useful for really huge arrays. 384 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 385 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 386 FalseRangeEnd == Overdefined) 387 return 0; 388 } 389 390 // Now that we've scanned the entire array, emit our new comparison(s). We 391 // order the state machines in complexity of the generated code. 392 Value *Idx = GEP->getOperand(2); 393 394 // If the index is larger than the pointer size of the target, truncate the 395 // index down like the GEP would do implicitly. We don't have to do this for 396 // an inbounds GEP because the index can't be out of range. 397 if (!GEP->isInBounds()) { 398 Type *IntPtrTy = TD->getIntPtrType(GEP->getType()); 399 unsigned PtrSize = IntPtrTy->getIntegerBitWidth(); 400 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize) 401 Idx = Builder->CreateTrunc(Idx, IntPtrTy); 402 } 403 404 // If the comparison is only true for one or two elements, emit direct 405 // comparisons. 406 if (SecondTrueElement != Overdefined) { 407 // None true -> false. 408 if (FirstTrueElement == Undefined) 409 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 410 411 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 412 413 // True for one element -> 'i == 47'. 414 if (SecondTrueElement == Undefined) 415 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 416 417 // True for two elements -> 'i == 47 | i == 72'. 418 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); 419 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 420 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); 421 return BinaryOperator::CreateOr(C1, C2); 422 } 423 424 // If the comparison is only false for one or two elements, emit direct 425 // comparisons. 426 if (SecondFalseElement != Overdefined) { 427 // None false -> true. 428 if (FirstFalseElement == Undefined) 429 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 430 431 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 432 433 // False for one element -> 'i != 47'. 434 if (SecondFalseElement == Undefined) 435 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 436 437 // False for two elements -> 'i != 47 & i != 72'. 438 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); 439 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 440 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); 441 return BinaryOperator::CreateAnd(C1, C2); 442 } 443 444 // If the comparison can be replaced with a range comparison for the elements 445 // where it is true, emit the range check. 446 if (TrueRangeEnd != Overdefined) { 447 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 448 449 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 450 if (FirstTrueElement) { 451 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 452 Idx = Builder->CreateAdd(Idx, Offs); 453 } 454 455 Value *End = ConstantInt::get(Idx->getType(), 456 TrueRangeEnd-FirstTrueElement+1); 457 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 458 } 459 460 // False range check. 461 if (FalseRangeEnd != Overdefined) { 462 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 463 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 464 if (FirstFalseElement) { 465 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 466 Idx = Builder->CreateAdd(Idx, Offs); 467 } 468 469 Value *End = ConstantInt::get(Idx->getType(), 470 FalseRangeEnd-FirstFalseElement); 471 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 472 } 473 474 475 // If a magic bitvector captures the entire comparison state 476 // of this load, replace it with computation that does: 477 // ((magic_cst >> i) & 1) != 0 478 { 479 Type *Ty = 0; 480 481 // Look for an appropriate type: 482 // - The type of Idx if the magic fits 483 // - The smallest fitting legal type if we have a DataLayout 484 // - Default to i32 485 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth()) 486 Ty = Idx->getType(); 487 else if (TD) 488 Ty = TD->getSmallestLegalIntType(Init->getContext(), ArrayElementCount); 489 else if (ArrayElementCount <= 32) 490 Ty = Type::getInt32Ty(Init->getContext()); 491 492 if (Ty != 0) { 493 Value *V = Builder->CreateIntCast(Idx, Ty, false); 494 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 495 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); 496 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 497 } 498 } 499 500 return 0; 501 } 502 503 504 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare 505 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we 506 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can 507 /// be complex, and scales are involved. The above expression would also be 508 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). 509 /// This later form is less amenable to optimization though, and we are allowed 510 /// to generate the first by knowing that pointer arithmetic doesn't overflow. 511 /// 512 /// If we can't emit an optimized form for this expression, this returns null. 513 /// 514 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) { 515 DataLayout &TD = *IC.getDataLayout(); 516 gep_type_iterator GTI = gep_type_begin(GEP); 517 518 // Check to see if this gep only has a single variable index. If so, and if 519 // any constant indices are a multiple of its scale, then we can compute this 520 // in terms of the scale of the variable index. For example, if the GEP 521 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 522 // because the expression will cross zero at the same point. 523 unsigned i, e = GEP->getNumOperands(); 524 int64_t Offset = 0; 525 for (i = 1; i != e; ++i, ++GTI) { 526 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 527 // Compute the aggregate offset of constant indices. 528 if (CI->isZero()) continue; 529 530 // Handle a struct index, which adds its field offset to the pointer. 531 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 532 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 533 } else { 534 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 535 Offset += Size*CI->getSExtValue(); 536 } 537 } else { 538 // Found our variable index. 539 break; 540 } 541 } 542 543 // If there are no variable indices, we must have a constant offset, just 544 // evaluate it the general way. 545 if (i == e) return 0; 546 547 Value *VariableIdx = GEP->getOperand(i); 548 // Determine the scale factor of the variable element. For example, this is 549 // 4 if the variable index is into an array of i32. 550 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType()); 551 552 // Verify that there are no other variable indices. If so, emit the hard way. 553 for (++i, ++GTI; i != e; ++i, ++GTI) { 554 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 555 if (!CI) return 0; 556 557 // Compute the aggregate offset of constant indices. 558 if (CI->isZero()) continue; 559 560 // Handle a struct index, which adds its field offset to the pointer. 561 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 562 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 563 } else { 564 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 565 Offset += Size*CI->getSExtValue(); 566 } 567 } 568 569 570 571 // Okay, we know we have a single variable index, which must be a 572 // pointer/array/vector index. If there is no offset, life is simple, return 573 // the index. 574 Type *IntPtrTy = TD.getIntPtrType(GEP->getOperand(0)->getType()); 575 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth(); 576 if (Offset == 0) { 577 // Cast to intptrty in case a truncation occurs. If an extension is needed, 578 // we don't need to bother extending: the extension won't affect where the 579 // computation crosses zero. 580 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { 581 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy); 582 } 583 return VariableIdx; 584 } 585 586 // Otherwise, there is an index. The computation we will do will be modulo 587 // the pointer size, so get it. 588 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 589 590 Offset &= PtrSizeMask; 591 VariableScale &= PtrSizeMask; 592 593 // To do this transformation, any constant index must be a multiple of the 594 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 595 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 596 // multiple of the variable scale. 597 int64_t NewOffs = Offset / (int64_t)VariableScale; 598 if (Offset != NewOffs*(int64_t)VariableScale) 599 return 0; 600 601 // Okay, we can do this evaluation. Start by converting the index to intptr. 602 if (VariableIdx->getType() != IntPtrTy) 603 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy, 604 true /*Signed*/); 605 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 606 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset"); 607 } 608 609 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something 610 /// else. At this point we know that the GEP is on the LHS of the comparison. 611 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 612 ICmpInst::Predicate Cond, 613 Instruction &I) { 614 // Don't transform signed compares of GEPs into index compares. Even if the 615 // GEP is inbounds, the final add of the base pointer can have signed overflow 616 // and would change the result of the icmp. 617 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be 618 // the maximum signed value for the pointer type. 619 if (ICmpInst::isSigned(Cond)) 620 return 0; 621 622 // Look through bitcasts. 623 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS)) 624 RHS = BCI->getOperand(0); 625 626 Value *PtrBase = GEPLHS->getOperand(0); 627 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) { 628 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 629 // This transformation (ignoring the base and scales) is valid because we 630 // know pointers can't overflow since the gep is inbounds. See if we can 631 // output an optimized form. 632 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this); 633 634 // If not, synthesize the offset the hard way. 635 if (Offset == 0) 636 Offset = EmitGEPOffset(GEPLHS); 637 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 638 Constant::getNullValue(Offset->getType())); 639 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 640 // If the base pointers are different, but the indices are the same, just 641 // compare the base pointer. 642 if (PtrBase != GEPRHS->getOperand(0)) { 643 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 644 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 645 GEPRHS->getOperand(0)->getType(); 646 if (IndicesTheSame) 647 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 648 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 649 IndicesTheSame = false; 650 break; 651 } 652 653 // If all indices are the same, just compare the base pointers. 654 if (IndicesTheSame) 655 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 656 657 // If we're comparing GEPs with two base pointers that only differ in type 658 // and both GEPs have only constant indices or just one use, then fold 659 // the compare with the adjusted indices. 660 if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() && 661 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && 662 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && 663 PtrBase->stripPointerCasts() == 664 GEPRHS->getOperand(0)->stripPointerCasts()) { 665 Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond), 666 EmitGEPOffset(GEPLHS), 667 EmitGEPOffset(GEPRHS)); 668 return ReplaceInstUsesWith(I, Cmp); 669 } 670 671 // Otherwise, the base pointers are different and the indices are 672 // different, bail out. 673 return 0; 674 } 675 676 // If one of the GEPs has all zero indices, recurse. 677 bool AllZeros = true; 678 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 679 if (!isa<Constant>(GEPLHS->getOperand(i)) || 680 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) { 681 AllZeros = false; 682 break; 683 } 684 if (AllZeros) 685 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 686 ICmpInst::getSwappedPredicate(Cond), I); 687 688 // If the other GEP has all zero indices, recurse. 689 AllZeros = true; 690 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 691 if (!isa<Constant>(GEPRHS->getOperand(i)) || 692 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) { 693 AllZeros = false; 694 break; 695 } 696 if (AllZeros) 697 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 698 699 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); 700 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 701 // If the GEPs only differ by one index, compare it. 702 unsigned NumDifferences = 0; // Keep track of # differences. 703 unsigned DiffOperand = 0; // The operand that differs. 704 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 705 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 706 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 707 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 708 // Irreconcilable differences. 709 NumDifferences = 2; 710 break; 711 } else { 712 if (NumDifferences++) break; 713 DiffOperand = i; 714 } 715 } 716 717 if (NumDifferences == 0) // SAME GEP? 718 return ReplaceInstUsesWith(I, // No comparison is needed here. 719 Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond))); 720 721 else if (NumDifferences == 1 && GEPsInBounds) { 722 Value *LHSV = GEPLHS->getOperand(DiffOperand); 723 Value *RHSV = GEPRHS->getOperand(DiffOperand); 724 // Make sure we do a signed comparison here. 725 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 726 } 727 } 728 729 // Only lower this if the icmp is the only user of the GEP or if we expect 730 // the result to fold to a constant! 731 if (TD && 732 GEPsInBounds && 733 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 734 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 735 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 736 Value *L = EmitGEPOffset(GEPLHS); 737 Value *R = EmitGEPOffset(GEPRHS); 738 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 739 } 740 } 741 return 0; 742 } 743 744 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X". 745 Instruction *InstCombiner::FoldICmpAddOpCst(Instruction &ICI, 746 Value *X, ConstantInt *CI, 747 ICmpInst::Predicate Pred) { 748 // If we have X+0, exit early (simplifying logic below) and let it get folded 749 // elsewhere. icmp X+0, X -> icmp X, X 750 if (CI->isZero()) { 751 bool isTrue = ICmpInst::isTrueWhenEqual(Pred); 752 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue)); 753 } 754 755 // (X+4) == X -> false. 756 if (Pred == ICmpInst::ICMP_EQ) 757 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 758 759 // (X+4) != X -> true. 760 if (Pred == ICmpInst::ICMP_NE) 761 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 762 763 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 764 // so the values can never be equal. Similarly for all other "or equals" 765 // operators. 766 767 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 768 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 769 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 770 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 771 Value *R = 772 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 773 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 774 } 775 776 // (X+1) >u X --> X <u (0-1) --> X != 255 777 // (X+2) >u X --> X <u (0-2) --> X <u 254 778 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 779 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 780 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 781 782 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 783 ConstantInt *SMax = ConstantInt::get(X->getContext(), 784 APInt::getSignedMaxValue(BitWidth)); 785 786 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 787 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 788 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 789 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 790 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 791 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 792 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 793 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 794 795 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 796 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 797 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 798 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 799 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 800 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 801 802 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 803 Constant *C = Builder->getInt(CI->getValue()-1); 804 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 805 } 806 807 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS 808 /// and CmpRHS are both known to be integer constants. 809 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 810 ConstantInt *DivRHS) { 811 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1)); 812 const APInt &CmpRHSV = CmpRHS->getValue(); 813 814 // FIXME: If the operand types don't match the type of the divide 815 // then don't attempt this transform. The code below doesn't have the 816 // logic to deal with a signed divide and an unsigned compare (and 817 // vice versa). This is because (x /s C1) <s C2 produces different 818 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even 819 // (x /u C1) <u C2. Simply casting the operands and result won't 820 // work. :( The if statement below tests that condition and bails 821 // if it finds it. 822 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv; 823 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned()) 824 return 0; 825 if (DivRHS->isZero()) 826 return 0; // The ProdOV computation fails on divide by zero. 827 if (DivIsSigned && DivRHS->isAllOnesValue()) 828 return 0; // The overflow computation also screws up here 829 if (DivRHS->isOne()) { 830 // This eliminates some funny cases with INT_MIN. 831 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X. 832 return &ICI; 833 } 834 835 // Compute Prod = CI * DivRHS. We are essentially solving an equation 836 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and 837 // C2 (CI). By solving for X we can turn this into a range check 838 // instead of computing a divide. 839 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); 840 841 // Determine if the product overflows by seeing if the product is 842 // not equal to the divide. Make sure we do the same kind of divide 843 // as in the LHS instruction that we're folding. 844 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) : 845 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; 846 847 // Get the ICmp opcode 848 ICmpInst::Predicate Pred = ICI.getPredicate(); 849 850 /// If the division is known to be exact, then there is no remainder from the 851 /// divide, so the covered range size is unit, otherwise it is the divisor. 852 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS; 853 854 // Figure out the interval that is being checked. For example, a comparison 855 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 856 // Compute this interval based on the constants involved and the signedness of 857 // the compare/divide. This computes a half-open interval, keeping track of 858 // whether either value in the interval overflows. After analysis each 859 // overflow variable is set to 0 if it's corresponding bound variable is valid 860 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 861 int LoOverflow = 0, HiOverflow = 0; 862 Constant *LoBound = 0, *HiBound = 0; 863 864 if (!DivIsSigned) { // udiv 865 // e.g. X/5 op 3 --> [15, 20) 866 LoBound = Prod; 867 HiOverflow = LoOverflow = ProdOV; 868 if (!HiOverflow) { 869 // If this is not an exact divide, then many values in the range collapse 870 // to the same result value. 871 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false); 872 } 873 874 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0. 875 if (CmpRHSV == 0) { // (X / pos) op 0 876 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 877 LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); 878 HiBound = RangeSize; 879 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos 880 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 881 HiOverflow = LoOverflow = ProdOV; 882 if (!HiOverflow) 883 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true); 884 } else { // (X / pos) op neg 885 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 886 HiBound = AddOne(Prod); 887 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 888 if (!LoOverflow) { 889 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 890 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 891 } 892 } 893 } else if (DivRHS->isNegative()) { // Divisor is < 0. 894 if (DivI->isExact()) 895 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 896 if (CmpRHSV == 0) { // (X / neg) op 0 897 // e.g. X/-5 op 0 --> [-4, 5) 898 LoBound = AddOne(RangeSize); 899 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize)); 900 if (HiBound == DivRHS) { // -INTMIN = INTMIN 901 HiOverflow = 1; // [INTMIN+1, overflow) 902 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN 903 } 904 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos 905 // e.g. X/-5 op 3 --> [-19, -14) 906 HiBound = AddOne(Prod); 907 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 908 if (!LoOverflow) 909 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 910 } else { // (X / neg) op neg 911 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 912 LoOverflow = HiOverflow = ProdOV; 913 if (!HiOverflow) 914 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true); 915 } 916 917 // Dividing by a negative swaps the condition. LT <-> GT 918 Pred = ICmpInst::getSwappedPredicate(Pred); 919 } 920 921 Value *X = DivI->getOperand(0); 922 switch (Pred) { 923 default: llvm_unreachable("Unhandled icmp opcode!"); 924 case ICmpInst::ICMP_EQ: 925 if (LoOverflow && HiOverflow) 926 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 927 if (HiOverflow) 928 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 929 ICmpInst::ICMP_UGE, X, LoBound); 930 if (LoOverflow) 931 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 932 ICmpInst::ICMP_ULT, X, HiBound); 933 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 934 DivIsSigned, true)); 935 case ICmpInst::ICMP_NE: 936 if (LoOverflow && HiOverflow) 937 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 938 if (HiOverflow) 939 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 940 ICmpInst::ICMP_ULT, X, LoBound); 941 if (LoOverflow) 942 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 943 ICmpInst::ICMP_UGE, X, HiBound); 944 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound, 945 DivIsSigned, false)); 946 case ICmpInst::ICMP_ULT: 947 case ICmpInst::ICMP_SLT: 948 if (LoOverflow == +1) // Low bound is greater than input range. 949 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 950 if (LoOverflow == -1) // Low bound is less than input range. 951 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 952 return new ICmpInst(Pred, X, LoBound); 953 case ICmpInst::ICMP_UGT: 954 case ICmpInst::ICMP_SGT: 955 if (HiOverflow == +1) // High bound greater than input range. 956 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 957 if (HiOverflow == -1) // High bound less than input range. 958 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 959 if (Pred == ICmpInst::ICMP_UGT) 960 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); 961 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); 962 } 963 } 964 965 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)". 966 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr, 967 ConstantInt *ShAmt) { 968 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue(); 969 970 // Check that the shift amount is in range. If not, don't perform 971 // undefined shifts. When the shift is visited it will be 972 // simplified. 973 uint32_t TypeBits = CmpRHSV.getBitWidth(); 974 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 975 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 976 return 0; 977 978 if (!ICI.isEquality()) { 979 // If we have an unsigned comparison and an ashr, we can't simplify this. 980 // Similarly for signed comparisons with lshr. 981 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr)) 982 return 0; 983 984 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv 985 // by a power of 2. Since we already have logic to simplify these, 986 // transform to div and then simplify the resultant comparison. 987 if (Shr->getOpcode() == Instruction::AShr && 988 (!Shr->isExact() || ShAmtVal == TypeBits - 1)) 989 return 0; 990 991 // Revisit the shift (to delete it). 992 Worklist.Add(Shr); 993 994 Constant *DivCst = 995 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); 996 997 Value *Tmp = 998 Shr->getOpcode() == Instruction::AShr ? 999 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) : 1000 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()); 1001 1002 ICI.setOperand(0, Tmp); 1003 1004 // If the builder folded the binop, just return it. 1005 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp); 1006 if (TheDiv == 0) 1007 return &ICI; 1008 1009 // Otherwise, fold this div/compare. 1010 assert(TheDiv->getOpcode() == Instruction::SDiv || 1011 TheDiv->getOpcode() == Instruction::UDiv); 1012 1013 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst)); 1014 assert(Res && "This div/cst should have folded!"); 1015 return Res; 1016 } 1017 1018 1019 // If we are comparing against bits always shifted out, the 1020 // comparison cannot succeed. 1021 APInt Comp = CmpRHSV << ShAmtVal; 1022 ConstantInt *ShiftedCmpRHS = Builder->getInt(Comp); 1023 if (Shr->getOpcode() == Instruction::LShr) 1024 Comp = Comp.lshr(ShAmtVal); 1025 else 1026 Comp = Comp.ashr(ShAmtVal); 1027 1028 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero. 1029 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1030 Constant *Cst = Builder->getInt1(IsICMP_NE); 1031 return ReplaceInstUsesWith(ICI, Cst); 1032 } 1033 1034 // Otherwise, check to see if the bits shifted out are known to be zero. 1035 // If so, we can compare against the unshifted value: 1036 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 1037 if (Shr->hasOneUse() && Shr->isExact()) 1038 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS); 1039 1040 if (Shr->hasOneUse()) { 1041 // Otherwise strength reduce the shift into an and. 1042 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 1043 Constant *Mask = Builder->getInt(Val); 1044 1045 Value *And = Builder->CreateAnd(Shr->getOperand(0), 1046 Mask, Shr->getName()+".mask"); 1047 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS); 1048 } 1049 return 0; 1050 } 1051 1052 1053 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)". 1054 /// 1055 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 1056 Instruction *LHSI, 1057 ConstantInt *RHS) { 1058 const APInt &RHSV = RHS->getValue(); 1059 1060 switch (LHSI->getOpcode()) { 1061 case Instruction::Trunc: 1062 if (ICI.isEquality() && LHSI->hasOneUse()) { 1063 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1064 // of the high bits truncated out of x are known. 1065 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(), 1066 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1067 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); 1068 ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne); 1069 1070 // If all the high bits are known, we can do this xform. 1071 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) { 1072 // Pull in the high bits from known-ones set. 1073 APInt NewRHS = RHS->getValue().zext(SrcBits); 1074 NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits); 1075 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1076 Builder->getInt(NewRHS)); 1077 } 1078 } 1079 break; 1080 1081 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI) 1082 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1083 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1084 // fold the xor. 1085 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) || 1086 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) { 1087 Value *CompareVal = LHSI->getOperand(0); 1088 1089 // If the sign bit of the XorCST is not set, there is no change to 1090 // the operation, just stop using the Xor. 1091 if (!XorCST->isNegative()) { 1092 ICI.setOperand(0, CompareVal); 1093 Worklist.Add(LHSI); 1094 return &ICI; 1095 } 1096 1097 // Was the old condition true if the operand is positive? 1098 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT; 1099 1100 // If so, the new one isn't. 1101 isTrueIfPositive ^= true; 1102 1103 if (isTrueIfPositive) 1104 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, 1105 SubOne(RHS)); 1106 else 1107 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, 1108 AddOne(RHS)); 1109 } 1110 1111 if (LHSI->hasOneUse()) { 1112 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit)) 1113 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) { 1114 const APInt &SignBit = XorCST->getValue(); 1115 ICmpInst::Predicate Pred = ICI.isSigned() 1116 ? ICI.getUnsignedPredicate() 1117 : ICI.getSignedPredicate(); 1118 return new ICmpInst(Pred, LHSI->getOperand(0), 1119 Builder->getInt(RHSV ^ SignBit)); 1120 } 1121 1122 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A) 1123 if (!ICI.isEquality() && XorCST->isMaxValue(true)) { 1124 const APInt &NotSignBit = XorCST->getValue(); 1125 ICmpInst::Predicate Pred = ICI.isSigned() 1126 ? ICI.getUnsignedPredicate() 1127 : ICI.getSignedPredicate(); 1128 Pred = ICI.getSwappedPredicate(Pred); 1129 return new ICmpInst(Pred, LHSI->getOperand(0), 1130 Builder->getInt(RHSV ^ NotSignBit)); 1131 } 1132 } 1133 1134 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C) 1135 // iff -C is a power of 2 1136 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && 1137 XorCST->getValue() == ~RHSV && (RHSV + 1).isPowerOf2()) 1138 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), XorCST); 1139 1140 // (icmp ult (xor X, C), -C) -> (icmp uge X, C) 1141 // iff -C is a power of 2 1142 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && 1143 XorCST->getValue() == -RHSV && RHSV.isPowerOf2()) 1144 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), XorCST); 1145 } 1146 break; 1147 case Instruction::And: // (icmp pred (and X, AndCST), RHS) 1148 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) && 1149 LHSI->getOperand(0)->hasOneUse()) { 1150 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1)); 1151 1152 // If the LHS is an AND of a truncating cast, we can widen the 1153 // and/compare to be the input width without changing the value 1154 // produced, eliminating a cast. 1155 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) { 1156 // We can do this transformation if either the AND constant does not 1157 // have its sign bit set or if it is an equality comparison. 1158 // Extending a relational comparison when we're checking the sign 1159 // bit would not work. 1160 if (ICI.isEquality() || 1161 (!AndCST->isNegative() && RHSV.isNonNegative())) { 1162 Value *NewAnd = 1163 Builder->CreateAnd(Cast->getOperand(0), 1164 ConstantExpr::getZExt(AndCST, Cast->getSrcTy())); 1165 NewAnd->takeName(LHSI); 1166 return new ICmpInst(ICI.getPredicate(), NewAnd, 1167 ConstantExpr::getZExt(RHS, Cast->getSrcTy())); 1168 } 1169 } 1170 1171 // If the LHS is an AND of a zext, and we have an equality compare, we can 1172 // shrink the and/compare to the smaller type, eliminating the cast. 1173 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) { 1174 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy()); 1175 // Make sure we don't compare the upper bits, SimplifyDemandedBits 1176 // should fold the icmp to true/false in that case. 1177 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) { 1178 Value *NewAnd = 1179 Builder->CreateAnd(Cast->getOperand(0), 1180 ConstantExpr::getTrunc(AndCST, Ty)); 1181 NewAnd->takeName(LHSI); 1182 return new ICmpInst(ICI.getPredicate(), NewAnd, 1183 ConstantExpr::getTrunc(RHS, Ty)); 1184 } 1185 } 1186 1187 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare 1188 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This 1189 // happens a LOT in code produced by the C front-end, for bitfield 1190 // access. 1191 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0)); 1192 if (Shift && !Shift->isShift()) 1193 Shift = 0; 1194 1195 ConstantInt *ShAmt; 1196 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0; 1197 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift. 1198 Type *AndTy = AndCST->getType(); // Type of the and. 1199 1200 // We can fold this as long as we can't shift unknown bits 1201 // into the mask. This can only happen with signed shift 1202 // rights, as they sign-extend. 1203 if (ShAmt) { 1204 bool CanFold = Shift->isLogicalShift(); 1205 if (!CanFold) { 1206 // To test for the bad case of the signed shr, see if any 1207 // of the bits shifted in could be tested after the mask. 1208 uint32_t TyBits = Ty->getPrimitiveSizeInBits(); 1209 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits); 1210 1211 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits(); 1212 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) & 1213 AndCST->getValue()) == 0) 1214 CanFold = true; 1215 } 1216 1217 if (CanFold) { 1218 Constant *NewCst; 1219 if (Shift->getOpcode() == Instruction::Shl) 1220 NewCst = ConstantExpr::getLShr(RHS, ShAmt); 1221 else 1222 NewCst = ConstantExpr::getShl(RHS, ShAmt); 1223 1224 // Check to see if we are shifting out any of the bits being 1225 // compared. 1226 if (ConstantExpr::get(Shift->getOpcode(), 1227 NewCst, ShAmt) != RHS) { 1228 // If we shifted bits out, the fold is not going to work out. 1229 // As a special case, check to see if this means that the 1230 // result is always true or false now. 1231 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1232 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 1233 if (ICI.getPredicate() == ICmpInst::ICMP_NE) 1234 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 1235 } else { 1236 ICI.setOperand(1, NewCst); 1237 Constant *NewAndCST; 1238 if (Shift->getOpcode() == Instruction::Shl) 1239 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt); 1240 else 1241 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt); 1242 LHSI->setOperand(1, NewAndCST); 1243 LHSI->setOperand(0, Shift->getOperand(0)); 1244 Worklist.Add(Shift); // Shift is dead. 1245 return &ICI; 1246 } 1247 } 1248 } 1249 1250 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is 1251 // preferable because it allows the C<<Y expression to be hoisted out 1252 // of a loop if Y is invariant and X is not. 1253 if (Shift && Shift->hasOneUse() && RHSV == 0 && 1254 ICI.isEquality() && !Shift->isArithmeticShift() && 1255 !isa<Constant>(Shift->getOperand(0))) { 1256 // Compute C << Y. 1257 Value *NS; 1258 if (Shift->getOpcode() == Instruction::LShr) { 1259 NS = Builder->CreateShl(AndCST, Shift->getOperand(1)); 1260 } else { 1261 // Insert a logical shift. 1262 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1)); 1263 } 1264 1265 // Compute X & (C << Y). 1266 Value *NewAnd = 1267 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName()); 1268 1269 ICI.setOperand(0, NewAnd); 1270 return &ICI; 1271 } 1272 1273 // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any 1274 // bit set in (X & AndCST) will produce a result greater than RHSV. 1275 if (ICI.getPredicate() == ICmpInst::ICMP_UGT) { 1276 unsigned NTZ = AndCST->getValue().countTrailingZeros(); 1277 if ((NTZ < AndCST->getBitWidth()) && 1278 APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV)) 1279 return new ICmpInst(ICmpInst::ICMP_NE, LHSI, 1280 Constant::getNullValue(RHS->getType())); 1281 } 1282 } 1283 1284 // Try to optimize things like "A[i]&42 == 0" to index computations. 1285 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) { 1286 if (GetElementPtrInst *GEP = 1287 dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1288 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1289 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1290 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) { 1291 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1)); 1292 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C)) 1293 return Res; 1294 } 1295 } 1296 1297 // X & -C == -C -> X > u ~C 1298 // X & -C != -C -> X <= u ~C 1299 // iff C is a power of 2 1300 if (ICI.isEquality() && RHS == LHSI->getOperand(1) && (-RHSV).isPowerOf2()) 1301 return new ICmpInst( 1302 ICI.getPredicate() == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT 1303 : ICmpInst::ICMP_ULE, 1304 LHSI->getOperand(0), SubOne(RHS)); 1305 break; 1306 1307 case Instruction::Or: { 1308 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse()) 1309 break; 1310 Value *P, *Q; 1311 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1312 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1313 // -> and (icmp eq P, null), (icmp eq Q, null). 1314 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P, 1315 Constant::getNullValue(P->getType())); 1316 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q, 1317 Constant::getNullValue(Q->getType())); 1318 Instruction *Op; 1319 if (ICI.getPredicate() == ICmpInst::ICMP_EQ) 1320 Op = BinaryOperator::CreateAnd(ICIP, ICIQ); 1321 else 1322 Op = BinaryOperator::CreateOr(ICIP, ICIQ); 1323 return Op; 1324 } 1325 break; 1326 } 1327 1328 case Instruction::Mul: { // (icmp pred (mul X, Val), CI) 1329 ConstantInt *Val = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1330 if (!Val) break; 1331 1332 // If this is a signed comparison to 0 and the mul is sign preserving, 1333 // use the mul LHS operand instead. 1334 ICmpInst::Predicate pred = ICI.getPredicate(); 1335 if (isSignTest(pred, RHS) && !Val->isZero() && 1336 cast<BinaryOperator>(LHSI)->hasNoSignedWrap()) 1337 return new ICmpInst(Val->isNegative() ? 1338 ICmpInst::getSwappedPredicate(pred) : pred, 1339 LHSI->getOperand(0), 1340 Constant::getNullValue(RHS->getType())); 1341 1342 break; 1343 } 1344 1345 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI) 1346 uint32_t TypeBits = RHSV.getBitWidth(); 1347 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1348 if (!ShAmt) { 1349 Value *X; 1350 // (1 << X) pred P2 -> X pred Log2(P2) 1351 if (match(LHSI, m_Shl(m_One(), m_Value(X)))) { 1352 bool RHSVIsPowerOf2 = RHSV.isPowerOf2(); 1353 ICmpInst::Predicate Pred = ICI.getPredicate(); 1354 if (ICI.isUnsigned()) { 1355 if (!RHSVIsPowerOf2) { 1356 // (1 << X) < 30 -> X <= 4 1357 // (1 << X) <= 30 -> X <= 4 1358 // (1 << X) >= 30 -> X > 4 1359 // (1 << X) > 30 -> X > 4 1360 if (Pred == ICmpInst::ICMP_ULT) 1361 Pred = ICmpInst::ICMP_ULE; 1362 else if (Pred == ICmpInst::ICMP_UGE) 1363 Pred = ICmpInst::ICMP_UGT; 1364 } 1365 unsigned RHSLog2 = RHSV.logBase2(); 1366 1367 // (1 << X) >= 2147483648 -> X >= 31 -> X == 31 1368 // (1 << X) > 2147483648 -> X > 31 -> false 1369 // (1 << X) <= 2147483648 -> X <= 31 -> true 1370 // (1 << X) < 2147483648 -> X < 31 -> X != 31 1371 if (RHSLog2 == TypeBits-1) { 1372 if (Pred == ICmpInst::ICMP_UGE) 1373 Pred = ICmpInst::ICMP_EQ; 1374 else if (Pred == ICmpInst::ICMP_UGT) 1375 return ReplaceInstUsesWith(ICI, Builder->getFalse()); 1376 else if (Pred == ICmpInst::ICMP_ULE) 1377 return ReplaceInstUsesWith(ICI, Builder->getTrue()); 1378 else if (Pred == ICmpInst::ICMP_ULT) 1379 Pred = ICmpInst::ICMP_NE; 1380 } 1381 1382 return new ICmpInst(Pred, X, 1383 ConstantInt::get(RHS->getType(), RHSLog2)); 1384 } else if (ICI.isSigned()) { 1385 if (RHSV.isAllOnesValue()) { 1386 // (1 << X) <= -1 -> X == 31 1387 if (Pred == ICmpInst::ICMP_SLE) 1388 return new ICmpInst(ICmpInst::ICMP_EQ, X, 1389 ConstantInt::get(RHS->getType(), TypeBits-1)); 1390 1391 // (1 << X) > -1 -> X != 31 1392 if (Pred == ICmpInst::ICMP_SGT) 1393 return new ICmpInst(ICmpInst::ICMP_NE, X, 1394 ConstantInt::get(RHS->getType(), TypeBits-1)); 1395 } else if (!RHSV) { 1396 // (1 << X) < 0 -> X == 31 1397 // (1 << X) <= 0 -> X == 31 1398 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 1399 return new ICmpInst(ICmpInst::ICMP_EQ, X, 1400 ConstantInt::get(RHS->getType(), TypeBits-1)); 1401 1402 // (1 << X) >= 0 -> X != 31 1403 // (1 << X) > 0 -> X != 31 1404 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) 1405 return new ICmpInst(ICmpInst::ICMP_NE, X, 1406 ConstantInt::get(RHS->getType(), TypeBits-1)); 1407 } 1408 } else if (ICI.isEquality()) { 1409 if (RHSVIsPowerOf2) 1410 return new ICmpInst( 1411 Pred, X, ConstantInt::get(RHS->getType(), RHSV.logBase2())); 1412 1413 return ReplaceInstUsesWith( 1414 ICI, Pred == ICmpInst::ICMP_EQ ? Builder->getFalse() 1415 : Builder->getTrue()); 1416 } 1417 } 1418 break; 1419 } 1420 1421 // Check that the shift amount is in range. If not, don't perform 1422 // undefined shifts. When the shift is visited it will be 1423 // simplified. 1424 if (ShAmt->uge(TypeBits)) 1425 break; 1426 1427 if (ICI.isEquality()) { 1428 // If we are comparing against bits always shifted out, the 1429 // comparison cannot succeed. 1430 Constant *Comp = 1431 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), 1432 ShAmt); 1433 if (Comp != RHS) {// Comparing against a bit that we know is zero. 1434 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1435 Constant *Cst = Builder->getInt1(IsICMP_NE); 1436 return ReplaceInstUsesWith(ICI, Cst); 1437 } 1438 1439 // If the shift is NUW, then it is just shifting out zeros, no need for an 1440 // AND. 1441 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap()) 1442 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1443 ConstantExpr::getLShr(RHS, ShAmt)); 1444 1445 // If the shift is NSW and we compare to 0, then it is just shifting out 1446 // sign bits, no need for an AND either. 1447 if (cast<BinaryOperator>(LHSI)->hasNoSignedWrap() && RHSV == 0) 1448 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0), 1449 ConstantExpr::getLShr(RHS, ShAmt)); 1450 1451 if (LHSI->hasOneUse()) { 1452 // Otherwise strength reduce the shift into an and. 1453 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits); 1454 Constant *Mask = Builder->getInt(APInt::getLowBitsSet(TypeBits, 1455 TypeBits - ShAmtVal)); 1456 1457 Value *And = 1458 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask"); 1459 return new ICmpInst(ICI.getPredicate(), And, 1460 ConstantExpr::getLShr(RHS, ShAmt)); 1461 } 1462 } 1463 1464 // If this is a signed comparison to 0 and the shift is sign preserving, 1465 // use the shift LHS operand instead. 1466 ICmpInst::Predicate pred = ICI.getPredicate(); 1467 if (isSignTest(pred, RHS) && 1468 cast<BinaryOperator>(LHSI)->hasNoSignedWrap()) 1469 return new ICmpInst(pred, 1470 LHSI->getOperand(0), 1471 Constant::getNullValue(RHS->getType())); 1472 1473 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1474 bool TrueIfSigned = false; 1475 if (LHSI->hasOneUse() && 1476 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) { 1477 // (X << 31) <s 0 --> (X&1) != 0 1478 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(), 1479 APInt::getOneBitSet(TypeBits, 1480 TypeBits-ShAmt->getZExtValue()-1)); 1481 Value *And = 1482 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask"); 1483 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1484 And, Constant::getNullValue(And->getType())); 1485 } 1486 1487 // Transform (icmp pred iM (shl iM %v, N), CI) 1488 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (CI>>N)) 1489 // Transform the shl to a trunc if (trunc (CI>>N)) has no loss and M-N. 1490 // This enables to get rid of the shift in favor of a trunc which can be 1491 // free on the target. It has the additional benefit of comparing to a 1492 // smaller constant, which will be target friendly. 1493 unsigned Amt = ShAmt->getLimitedValue(TypeBits-1); 1494 if (LHSI->hasOneUse() && 1495 Amt != 0 && RHSV.countTrailingZeros() >= Amt) { 1496 Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt); 1497 Constant *NCI = ConstantExpr::getTrunc( 1498 ConstantExpr::getAShr(RHS, 1499 ConstantInt::get(RHS->getType(), Amt)), 1500 NTy); 1501 return new ICmpInst(ICI.getPredicate(), 1502 Builder->CreateTrunc(LHSI->getOperand(0), NTy), 1503 NCI); 1504 } 1505 1506 break; 1507 } 1508 1509 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI) 1510 case Instruction::AShr: { 1511 // Handle equality comparisons of shift-by-constant. 1512 BinaryOperator *BO = cast<BinaryOperator>(LHSI); 1513 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) { 1514 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt)) 1515 return Res; 1516 } 1517 1518 // Handle exact shr's. 1519 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) { 1520 if (RHSV.isMinValue()) 1521 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS); 1522 } 1523 break; 1524 } 1525 1526 case Instruction::SDiv: 1527 case Instruction::UDiv: 1528 // Fold: icmp pred ([us]div X, C1), C2 -> range test 1529 // Fold this div into the comparison, producing a range check. 1530 // Determine, based on the divide type, what the range is being 1531 // checked. If there is an overflow on the low or high side, remember 1532 // it, otherwise compute the range [low, hi) bounding the new value. 1533 // See: InsertRangeTest above for the kinds of replacements possible. 1534 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) 1535 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI), 1536 DivRHS)) 1537 return R; 1538 break; 1539 1540 case Instruction::Sub: { 1541 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(0)); 1542 if (!LHSC) break; 1543 const APInt &LHSV = LHSC->getValue(); 1544 1545 // C1-X <u C2 -> (X|(C2-1)) == C1 1546 // iff C1 & (C2-1) == C2-1 1547 // C2 is a power of 2 1548 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() && 1549 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == (RHSV - 1)) 1550 return new ICmpInst(ICmpInst::ICMP_EQ, 1551 Builder->CreateOr(LHSI->getOperand(1), RHSV - 1), 1552 LHSC); 1553 1554 // C1-X >u C2 -> (X|C2) != C1 1555 // iff C1 & C2 == C2 1556 // C2+1 is a power of 2 1557 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() && 1558 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == RHSV) 1559 return new ICmpInst(ICmpInst::ICMP_NE, 1560 Builder->CreateOr(LHSI->getOperand(1), RHSV), LHSC); 1561 break; 1562 } 1563 1564 case Instruction::Add: 1565 // Fold: icmp pred (add X, C1), C2 1566 if (!ICI.isEquality()) { 1567 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1)); 1568 if (!LHSC) break; 1569 const APInt &LHSV = LHSC->getValue(); 1570 1571 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV) 1572 .subtract(LHSV); 1573 1574 if (ICI.isSigned()) { 1575 if (CR.getLower().isSignBit()) { 1576 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0), 1577 Builder->getInt(CR.getUpper())); 1578 } else if (CR.getUpper().isSignBit()) { 1579 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0), 1580 Builder->getInt(CR.getLower())); 1581 } 1582 } else { 1583 if (CR.getLower().isMinValue()) { 1584 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0), 1585 Builder->getInt(CR.getUpper())); 1586 } else if (CR.getUpper().isMinValue()) { 1587 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0), 1588 Builder->getInt(CR.getLower())); 1589 } 1590 } 1591 1592 // X-C1 <u C2 -> (X & -C2) == C1 1593 // iff C1 & (C2-1) == 0 1594 // C2 is a power of 2 1595 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && LHSI->hasOneUse() && 1596 RHSV.isPowerOf2() && (LHSV & (RHSV - 1)) == 0) 1597 return new ICmpInst(ICmpInst::ICMP_EQ, 1598 Builder->CreateAnd(LHSI->getOperand(0), -RHSV), 1599 ConstantExpr::getNeg(LHSC)); 1600 1601 // X-C1 >u C2 -> (X & ~C2) != C1 1602 // iff C1 & C2 == 0 1603 // C2+1 is a power of 2 1604 if (ICI.getPredicate() == ICmpInst::ICMP_UGT && LHSI->hasOneUse() && 1605 (RHSV + 1).isPowerOf2() && (LHSV & RHSV) == 0) 1606 return new ICmpInst(ICmpInst::ICMP_NE, 1607 Builder->CreateAnd(LHSI->getOperand(0), ~RHSV), 1608 ConstantExpr::getNeg(LHSC)); 1609 } 1610 break; 1611 } 1612 1613 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS. 1614 if (ICI.isEquality()) { 1615 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE; 1616 1617 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and 1618 // the second operand is a constant, simplify a bit. 1619 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) { 1620 switch (BO->getOpcode()) { 1621 case Instruction::SRem: 1622 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 1623 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){ 1624 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue(); 1625 if (V.sgt(1) && V.isPowerOf2()) { 1626 Value *NewRem = 1627 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1), 1628 BO->getName()); 1629 return new ICmpInst(ICI.getPredicate(), NewRem, 1630 Constant::getNullValue(BO->getType())); 1631 } 1632 } 1633 break; 1634 case Instruction::Add: 1635 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 1636 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1637 if (BO->hasOneUse()) 1638 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1639 ConstantExpr::getSub(RHS, BOp1C)); 1640 } else if (RHSV == 0) { 1641 // Replace ((add A, B) != 0) with (A != -B) if A or B is 1642 // efficiently invertible, or if the add has just this one use. 1643 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 1644 1645 if (Value *NegVal = dyn_castNegVal(BOp1)) 1646 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal); 1647 if (Value *NegVal = dyn_castNegVal(BOp0)) 1648 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1); 1649 if (BO->hasOneUse()) { 1650 Value *Neg = Builder->CreateNeg(BOp1); 1651 Neg->takeName(BO); 1652 return new ICmpInst(ICI.getPredicate(), BOp0, Neg); 1653 } 1654 } 1655 break; 1656 case Instruction::Xor: 1657 // For the xor case, we can xor two constants together, eliminating 1658 // the explicit xor. 1659 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) { 1660 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1661 ConstantExpr::getXor(RHS, BOC)); 1662 } else if (RHSV == 0) { 1663 // Replace ((xor A, B) != 0) with (A != B) 1664 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1665 BO->getOperand(1)); 1666 } 1667 break; 1668 case Instruction::Sub: 1669 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants. 1670 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) { 1671 if (BO->hasOneUse()) 1672 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1), 1673 ConstantExpr::getSub(BOp0C, RHS)); 1674 } else if (RHSV == 0) { 1675 // Replace ((sub A, B) != 0) with (A != B) 1676 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1677 BO->getOperand(1)); 1678 } 1679 break; 1680 case Instruction::Or: 1681 // If bits are being or'd in that are not present in the constant we 1682 // are comparing against, then the comparison could never succeed! 1683 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1684 Constant *NotCI = ConstantExpr::getNot(RHS); 1685 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue()) 1686 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE)); 1687 } 1688 break; 1689 1690 case Instruction::And: 1691 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1692 // If bits are being compared against that are and'd out, then the 1693 // comparison can never succeed! 1694 if ((RHSV & ~BOC->getValue()) != 0) 1695 return ReplaceInstUsesWith(ICI, Builder->getInt1(isICMP_NE)); 1696 1697 // If we have ((X & C) == C), turn it into ((X & C) != 0). 1698 if (RHS == BOC && RHSV.isPowerOf2()) 1699 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : 1700 ICmpInst::ICMP_NE, LHSI, 1701 Constant::getNullValue(RHS->getType())); 1702 1703 // Don't perform the following transforms if the AND has multiple uses 1704 if (!BO->hasOneUse()) 1705 break; 1706 1707 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 1708 if (BOC->getValue().isSignBit()) { 1709 Value *X = BO->getOperand(0); 1710 Constant *Zero = Constant::getNullValue(X->getType()); 1711 ICmpInst::Predicate pred = isICMP_NE ? 1712 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 1713 return new ICmpInst(pred, X, Zero); 1714 } 1715 1716 // ((X & ~7) == 0) --> X < 8 1717 if (RHSV == 0 && isHighOnes(BOC)) { 1718 Value *X = BO->getOperand(0); 1719 Constant *NegX = ConstantExpr::getNeg(BOC); 1720 ICmpInst::Predicate pred = isICMP_NE ? 1721 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 1722 return new ICmpInst(pred, X, NegX); 1723 } 1724 } 1725 break; 1726 case Instruction::Mul: 1727 if (RHSV == 0 && BO->hasNoSignedWrap()) { 1728 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) { 1729 // The trivial case (mul X, 0) is handled by InstSimplify 1730 // General case : (mul X, C) != 0 iff X != 0 1731 // (mul X, C) == 0 iff X == 0 1732 if (!BOC->isZero()) 1733 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), 1734 Constant::getNullValue(RHS->getType())); 1735 } 1736 } 1737 break; 1738 default: break; 1739 } 1740 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) { 1741 // Handle icmp {eq|ne} <intrinsic>, intcst. 1742 switch (II->getIntrinsicID()) { 1743 case Intrinsic::bswap: 1744 Worklist.Add(II); 1745 ICI.setOperand(0, II->getArgOperand(0)); 1746 ICI.setOperand(1, Builder->getInt(RHSV.byteSwap())); 1747 return &ICI; 1748 case Intrinsic::ctlz: 1749 case Intrinsic::cttz: 1750 // ctz(A) == bitwidth(a) -> A == 0 and likewise for != 1751 if (RHSV == RHS->getType()->getBitWidth()) { 1752 Worklist.Add(II); 1753 ICI.setOperand(0, II->getArgOperand(0)); 1754 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0)); 1755 return &ICI; 1756 } 1757 break; 1758 case Intrinsic::ctpop: 1759 // popcount(A) == 0 -> A == 0 and likewise for != 1760 if (RHS->isZero()) { 1761 Worklist.Add(II); 1762 ICI.setOperand(0, II->getArgOperand(0)); 1763 ICI.setOperand(1, RHS); 1764 return &ICI; 1765 } 1766 break; 1767 default: 1768 break; 1769 } 1770 } 1771 } 1772 return 0; 1773 } 1774 1775 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst). 1776 /// We only handle extending casts so far. 1777 /// 1778 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) { 1779 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0)); 1780 Value *LHSCIOp = LHSCI->getOperand(0); 1781 Type *SrcTy = LHSCIOp->getType(); 1782 Type *DestTy = LHSCI->getType(); 1783 Value *RHSCIOp; 1784 1785 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 1786 // integer type is the same size as the pointer type. 1787 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt && 1788 TD->getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) { 1789 Value *RHSOp = 0; 1790 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) { 1791 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 1792 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) { 1793 RHSOp = RHSC->getOperand(0); 1794 // If the pointer types don't match, insert a bitcast. 1795 if (LHSCIOp->getType() != RHSOp->getType()) 1796 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); 1797 } 1798 1799 if (RHSOp) 1800 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp); 1801 } 1802 1803 // The code below only handles extension cast instructions, so far. 1804 // Enforce this. 1805 if (LHSCI->getOpcode() != Instruction::ZExt && 1806 LHSCI->getOpcode() != Instruction::SExt) 1807 return 0; 1808 1809 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 1810 bool isSignedCmp = ICI.isSigned(); 1811 1812 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) { 1813 // Not an extension from the same type? 1814 RHSCIOp = CI->getOperand(0); 1815 if (RHSCIOp->getType() != LHSCIOp->getType()) 1816 return 0; 1817 1818 // If the signedness of the two casts doesn't agree (i.e. one is a sext 1819 // and the other is a zext), then we can't handle this. 1820 if (CI->getOpcode() != LHSCI->getOpcode()) 1821 return 0; 1822 1823 // Deal with equality cases early. 1824 if (ICI.isEquality()) 1825 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1826 1827 // A signed comparison of sign extended values simplifies into a 1828 // signed comparison. 1829 if (isSignedCmp && isSignedExt) 1830 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp); 1831 1832 // The other three cases all fold into an unsigned comparison. 1833 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 1834 } 1835 1836 // If we aren't dealing with a constant on the RHS, exit early 1837 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1)); 1838 if (!CI) 1839 return 0; 1840 1841 // Compute the constant that would happen if we truncated to SrcTy then 1842 // reextended to DestTy. 1843 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy); 1844 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), 1845 Res1, DestTy); 1846 1847 // If the re-extended constant didn't change... 1848 if (Res2 == CI) { 1849 // Deal with equality cases early. 1850 if (ICI.isEquality()) 1851 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1852 1853 // A signed comparison of sign extended values simplifies into a 1854 // signed comparison. 1855 if (isSignedExt && isSignedCmp) 1856 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1); 1857 1858 // The other three cases all fold into an unsigned comparison. 1859 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1); 1860 } 1861 1862 // The re-extended constant changed so the constant cannot be represented 1863 // in the shorter type. Consequently, we cannot emit a simple comparison. 1864 // All the cases that fold to true or false will have already been handled 1865 // by SimplifyICmpInst, so only deal with the tricky case. 1866 1867 if (isSignedCmp || !isSignedExt) 1868 return 0; 1869 1870 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 1871 // should have been folded away previously and not enter in here. 1872 1873 // We're performing an unsigned comp with a sign extended value. 1874 // This is true if the input is >= 0. [aka >s -1] 1875 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 1876 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName()); 1877 1878 // Finally, return the value computed. 1879 if (ICI.getPredicate() == ICmpInst::ICMP_ULT) 1880 return ReplaceInstUsesWith(ICI, Result); 1881 1882 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 1883 return BinaryOperator::CreateNot(Result); 1884 } 1885 1886 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form: 1887 /// I = icmp ugt (add (add A, B), CI2), CI1 1888 /// If this is of the form: 1889 /// sum = a + b 1890 /// if (sum+128 >u 255) 1891 /// Then replace it with llvm.sadd.with.overflow.i8. 1892 /// 1893 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1894 ConstantInt *CI2, ConstantInt *CI1, 1895 InstCombiner &IC) { 1896 // The transformation we're trying to do here is to transform this into an 1897 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1898 // with a narrower add, and discard the add-with-constant that is part of the 1899 // range check (if we can't eliminate it, this isn't profitable). 1900 1901 // In order to eliminate the add-with-constant, the compare can be its only 1902 // use. 1903 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1904 if (!AddWithCst->hasOneUse()) return 0; 1905 1906 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1907 if (!CI2->getValue().isPowerOf2()) return 0; 1908 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1909 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0; 1910 1911 // The width of the new add formed is 1 more than the bias. 1912 ++NewWidth; 1913 1914 // Check to see that CI1 is an all-ones value with NewWidth bits. 1915 if (CI1->getBitWidth() == NewWidth || 1916 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1917 return 0; 1918 1919 // This is only really a signed overflow check if the inputs have been 1920 // sign-extended; check for that condition. For example, if CI2 is 2^31 and 1921 // the operands of the add are 64 bits wide, we need at least 33 sign bits. 1922 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; 1923 if (IC.ComputeNumSignBits(A) < NeededSignBits || 1924 IC.ComputeNumSignBits(B) < NeededSignBits) 1925 return 0; 1926 1927 // In order to replace the original add with a narrower 1928 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1929 // and truncates that discard the high bits of the add. Verify that this is 1930 // the case. 1931 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1932 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end(); 1933 UI != E; ++UI) { 1934 if (*UI == AddWithCst) continue; 1935 1936 // Only accept truncates for now. We would really like a nice recursive 1937 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1938 // chain to see which bits of a value are actually demanded. If the 1939 // original add had another add which was then immediately truncated, we 1940 // could still do the transformation. 1941 TruncInst *TI = dyn_cast<TruncInst>(*UI); 1942 if (TI == 0 || 1943 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0; 1944 } 1945 1946 // If the pattern matches, truncate the inputs to the narrower type and 1947 // use the sadd_with_overflow intrinsic to efficiently compute both the 1948 // result and the overflow bit. 1949 Module *M = I.getParent()->getParent()->getParent(); 1950 1951 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1952 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow, 1953 NewType); 1954 1955 InstCombiner::BuilderTy *Builder = IC.Builder; 1956 1957 // Put the new code above the original add, in case there are any uses of the 1958 // add between the add and the compare. 1959 Builder->SetInsertPoint(OrigAdd); 1960 1961 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc"); 1962 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc"); 1963 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd"); 1964 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); 1965 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); 1966 1967 // The inner add was the result of the narrow add, zero extended to the 1968 // wider type. Replace it with the result computed by the intrinsic. 1969 IC.ReplaceInstUsesWith(*OrigAdd, ZExt); 1970 1971 // The original icmp gets replaced with the overflow value. 1972 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1973 } 1974 1975 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV, 1976 InstCombiner &IC) { 1977 // Don't bother doing this transformation for pointers, don't do it for 1978 // vectors. 1979 if (!isa<IntegerType>(OrigAddV->getType())) return 0; 1980 1981 // If the add is a constant expr, then we don't bother transforming it. 1982 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV); 1983 if (OrigAdd == 0) return 0; 1984 1985 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1); 1986 1987 // Put the new code above the original add, in case there are any uses of the 1988 // add between the add and the compare. 1989 InstCombiner::BuilderTy *Builder = IC.Builder; 1990 Builder->SetInsertPoint(OrigAdd); 1991 1992 Module *M = I.getParent()->getParent()->getParent(); 1993 Type *Ty = LHS->getType(); 1994 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); 1995 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd"); 1996 Value *Add = Builder->CreateExtractValue(Call, 0); 1997 1998 IC.ReplaceInstUsesWith(*OrigAdd, Add); 1999 2000 // The original icmp gets replaced with the overflow value. 2001 return ExtractValueInst::Create(Call, 1, "uadd.overflow"); 2002 } 2003 2004 // DemandedBitsLHSMask - When performing a comparison against a constant, 2005 // it is possible that not all the bits in the LHS are demanded. This helper 2006 // method computes the mask that IS demanded. 2007 static APInt DemandedBitsLHSMask(ICmpInst &I, 2008 unsigned BitWidth, bool isSignCheck) { 2009 if (isSignCheck) 2010 return APInt::getSignBit(BitWidth); 2011 2012 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1)); 2013 if (!CI) return APInt::getAllOnesValue(BitWidth); 2014 const APInt &RHS = CI->getValue(); 2015 2016 switch (I.getPredicate()) { 2017 // For a UGT comparison, we don't care about any bits that 2018 // correspond to the trailing ones of the comparand. The value of these 2019 // bits doesn't impact the outcome of the comparison, because any value 2020 // greater than the RHS must differ in a bit higher than these due to carry. 2021 case ICmpInst::ICMP_UGT: { 2022 unsigned trailingOnes = RHS.countTrailingOnes(); 2023 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); 2024 return ~lowBitsSet; 2025 } 2026 2027 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 2028 // Any value less than the RHS must differ in a higher bit because of carries. 2029 case ICmpInst::ICMP_ULT: { 2030 unsigned trailingZeros = RHS.countTrailingZeros(); 2031 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); 2032 return ~lowBitsSet; 2033 } 2034 2035 default: 2036 return APInt::getAllOnesValue(BitWidth); 2037 } 2038 2039 } 2040 2041 /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst 2042 /// should be swapped. 2043 /// The descision is based on how many times these two operands are reused 2044 /// as subtract operands and their positions in those instructions. 2045 /// The rational is that several architectures use the same instruction for 2046 /// both subtract and cmp, thus it is better if the order of those operands 2047 /// match. 2048 /// \return true if Op0 and Op1 should be swapped. 2049 static bool swapMayExposeCSEOpportunities(const Value * Op0, 2050 const Value * Op1) { 2051 // Filter out pointer value as those cannot appears directly in subtract. 2052 // FIXME: we may want to go through inttoptrs or bitcasts. 2053 if (Op0->getType()->isPointerTy()) 2054 return false; 2055 // Count every uses of both Op0 and Op1 in a subtract. 2056 // Each time Op0 is the first operand, count -1: swapping is bad, the 2057 // subtract has already the same layout as the compare. 2058 // Each time Op0 is the second operand, count +1: swapping is good, the 2059 // subtract has a diffrent layout as the compare. 2060 // At the end, if the benefit is greater than 0, Op0 should come second to 2061 // expose more CSE opportunities. 2062 int GlobalSwapBenefits = 0; 2063 for (Value::const_use_iterator UI = Op0->use_begin(), UIEnd = Op0->use_end(); UI != UIEnd; ++UI) { 2064 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(*UI); 2065 if (!BinOp || BinOp->getOpcode() != Instruction::Sub) 2066 continue; 2067 // If Op0 is the first argument, this is not beneficial to swap the 2068 // arguments. 2069 int LocalSwapBenefits = -1; 2070 unsigned Op1Idx = 1; 2071 if (BinOp->getOperand(Op1Idx) == Op0) { 2072 Op1Idx = 0; 2073 LocalSwapBenefits = 1; 2074 } 2075 if (BinOp->getOperand(Op1Idx) != Op1) 2076 continue; 2077 GlobalSwapBenefits += LocalSwapBenefits; 2078 } 2079 return GlobalSwapBenefits > 0; 2080 } 2081 2082 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 2083 bool Changed = false; 2084 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2085 unsigned Op0Cplxity = getComplexity(Op0); 2086 unsigned Op1Cplxity = getComplexity(Op1); 2087 2088 /// Orders the operands of the compare so that they are listed from most 2089 /// complex to least complex. This puts constants before unary operators, 2090 /// before binary operators. 2091 if (Op0Cplxity < Op1Cplxity || 2092 (Op0Cplxity == Op1Cplxity && 2093 swapMayExposeCSEOpportunities(Op0, Op1))) { 2094 I.swapOperands(); 2095 std::swap(Op0, Op1); 2096 Changed = true; 2097 } 2098 2099 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD)) 2100 return ReplaceInstUsesWith(I, V); 2101 2102 // comparing -val or val with non-zero is the same as just comparing val 2103 // ie, abs(val) != 0 -> val != 0 2104 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) 2105 { 2106 Value *Cond, *SelectTrue, *SelectFalse; 2107 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue), 2108 m_Value(SelectFalse)))) { 2109 if (Value *V = dyn_castNegVal(SelectTrue)) { 2110 if (V == SelectFalse) 2111 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 2112 } 2113 else if (Value *V = dyn_castNegVal(SelectFalse)) { 2114 if (V == SelectTrue) 2115 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 2116 } 2117 } 2118 } 2119 2120 Type *Ty = Op0->getType(); 2121 2122 // icmp's with boolean values can always be turned into bitwise operations 2123 if (Ty->isIntegerTy(1)) { 2124 switch (I.getPredicate()) { 2125 default: llvm_unreachable("Invalid icmp instruction!"); 2126 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B) 2127 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp"); 2128 return BinaryOperator::CreateNot(Xor); 2129 } 2130 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B 2131 return BinaryOperator::CreateXor(Op0, Op1); 2132 2133 case ICmpInst::ICMP_UGT: 2134 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult 2135 // FALL THROUGH 2136 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B 2137 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 2138 return BinaryOperator::CreateAnd(Not, Op1); 2139 } 2140 case ICmpInst::ICMP_SGT: 2141 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt 2142 // FALL THROUGH 2143 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B 2144 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 2145 return BinaryOperator::CreateAnd(Not, Op0); 2146 } 2147 case ICmpInst::ICMP_UGE: 2148 std::swap(Op0, Op1); // Change icmp uge -> icmp ule 2149 // FALL THROUGH 2150 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B 2151 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp"); 2152 return BinaryOperator::CreateOr(Not, Op1); 2153 } 2154 case ICmpInst::ICMP_SGE: 2155 std::swap(Op0, Op1); // Change icmp sge -> icmp sle 2156 // FALL THROUGH 2157 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B 2158 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp"); 2159 return BinaryOperator::CreateOr(Not, Op0); 2160 } 2161 } 2162 } 2163 2164 unsigned BitWidth = 0; 2165 if (Ty->isIntOrIntVectorTy()) 2166 BitWidth = Ty->getScalarSizeInBits(); 2167 else if (TD) // Pointers require TD info to get their size. 2168 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType()); 2169 2170 bool isSignBit = false; 2171 2172 // See if we are doing a comparison with a constant. 2173 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2174 Value *A = 0, *B = 0; 2175 2176 // Match the following pattern, which is a common idiom when writing 2177 // overflow-safe integer arithmetic function. The source performs an 2178 // addition in wider type, and explicitly checks for overflow using 2179 // comparisons against INT_MIN and INT_MAX. Simplify this by using the 2180 // sadd_with_overflow intrinsic. 2181 // 2182 // TODO: This could probably be generalized to handle other overflow-safe 2183 // operations if we worked out the formulas to compute the appropriate 2184 // magic constants. 2185 // 2186 // sum = a + b 2187 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 2188 { 2189 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 2190 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2191 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 2192 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this)) 2193 return Res; 2194 } 2195 2196 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B) 2197 if (I.isEquality() && CI->isZero() && 2198 match(Op0, m_Sub(m_Value(A), m_Value(B)))) { 2199 // (icmp cond A B) if cond is equality 2200 return new ICmpInst(I.getPredicate(), A, B); 2201 } 2202 2203 // If we have an icmp le or icmp ge instruction, turn it into the 2204 // appropriate icmp lt or icmp gt instruction. This allows us to rely on 2205 // them being folded in the code below. The SimplifyICmpInst code has 2206 // already handled the edge cases for us, so we just assert on them. 2207 switch (I.getPredicate()) { 2208 default: break; 2209 case ICmpInst::ICMP_ULE: 2210 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE 2211 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, 2212 Builder->getInt(CI->getValue()+1)); 2213 case ICmpInst::ICMP_SLE: 2214 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE 2215 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2216 Builder->getInt(CI->getValue()+1)); 2217 case ICmpInst::ICMP_UGE: 2218 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE 2219 return new ICmpInst(ICmpInst::ICMP_UGT, Op0, 2220 Builder->getInt(CI->getValue()-1)); 2221 case ICmpInst::ICMP_SGE: 2222 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE 2223 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2224 Builder->getInt(CI->getValue()-1)); 2225 } 2226 2227 // If this comparison is a normal comparison, it demands all 2228 // bits, if it is a sign bit comparison, it only demands the sign bit. 2229 bool UnusedBit; 2230 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit); 2231 } 2232 2233 // See if we can fold the comparison based on range information we can get 2234 // by checking whether bits are known to be zero or one in the input. 2235 if (BitWidth != 0) { 2236 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0); 2237 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0); 2238 2239 if (SimplifyDemandedBits(I.getOperandUse(0), 2240 DemandedBitsLHSMask(I, BitWidth, isSignBit), 2241 Op0KnownZero, Op0KnownOne, 0)) 2242 return &I; 2243 if (SimplifyDemandedBits(I.getOperandUse(1), 2244 APInt::getAllOnesValue(BitWidth), 2245 Op1KnownZero, Op1KnownOne, 0)) 2246 return &I; 2247 2248 // Given the known and unknown bits, compute a range that the LHS could be 2249 // in. Compute the Min, Max and RHS values based on the known bits. For the 2250 // EQ and NE we use unsigned values. 2251 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 2252 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 2253 if (I.isSigned()) { 2254 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2255 Op0Min, Op0Max); 2256 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2257 Op1Min, Op1Max); 2258 } else { 2259 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne, 2260 Op0Min, Op0Max); 2261 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne, 2262 Op1Min, Op1Max); 2263 } 2264 2265 // If Min and Max are known to be the same, then SimplifyDemandedBits 2266 // figured out that the LHS is a constant. Just constant fold this now so 2267 // that code below can assume that Min != Max. 2268 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 2269 return new ICmpInst(I.getPredicate(), 2270 ConstantInt::get(Op0->getType(), Op0Min), Op1); 2271 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 2272 return new ICmpInst(I.getPredicate(), Op0, 2273 ConstantInt::get(Op1->getType(), Op1Min)); 2274 2275 // Based on the range information we know about the LHS, see if we can 2276 // simplify this comparison. For example, (x&4) < 8 is always true. 2277 switch (I.getPredicate()) { 2278 default: llvm_unreachable("Unknown icmp opcode!"); 2279 case ICmpInst::ICMP_EQ: { 2280 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2281 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2282 2283 // If all bits are known zero except for one, then we know at most one 2284 // bit is set. If the comparison is against zero, then this is a check 2285 // to see if *that* bit is set. 2286 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2287 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2288 // If the LHS is an AND with the same constant, look through it. 2289 Value *LHS = 0; 2290 ConstantInt *LHSC = 0; 2291 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2292 LHSC->getValue() != Op0KnownZeroInverted) 2293 LHS = Op0; 2294 2295 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2296 // then turn "((1 << x)&8) == 0" into "x != 3". 2297 Value *X = 0; 2298 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2299 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2300 return new ICmpInst(ICmpInst::ICMP_NE, X, 2301 ConstantInt::get(X->getType(), CmpVal)); 2302 } 2303 2304 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2305 // then turn "((8 >>u x)&1) == 0" into "x != 3". 2306 const APInt *CI; 2307 if (Op0KnownZeroInverted == 1 && 2308 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2309 return new ICmpInst(ICmpInst::ICMP_NE, X, 2310 ConstantInt::get(X->getType(), 2311 CI->countTrailingZeros())); 2312 } 2313 2314 break; 2315 } 2316 case ICmpInst::ICMP_NE: { 2317 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) 2318 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2319 2320 // If all bits are known zero except for one, then we know at most one 2321 // bit is set. If the comparison is against zero, then this is a check 2322 // to see if *that* bit is set. 2323 APInt Op0KnownZeroInverted = ~Op0KnownZero; 2324 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) { 2325 // If the LHS is an AND with the same constant, look through it. 2326 Value *LHS = 0; 2327 ConstantInt *LHSC = 0; 2328 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) || 2329 LHSC->getValue() != Op0KnownZeroInverted) 2330 LHS = Op0; 2331 2332 // If the LHS is 1 << x, and we know the result is a power of 2 like 8, 2333 // then turn "((1 << x)&8) != 0" into "x == 3". 2334 Value *X = 0; 2335 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 2336 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros(); 2337 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2338 ConstantInt::get(X->getType(), CmpVal)); 2339 } 2340 2341 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1, 2342 // then turn "((8 >>u x)&1) != 0" into "x == 3". 2343 const APInt *CI; 2344 if (Op0KnownZeroInverted == 1 && 2345 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) 2346 return new ICmpInst(ICmpInst::ICMP_EQ, X, 2347 ConstantInt::get(X->getType(), 2348 CI->countTrailingZeros())); 2349 } 2350 2351 break; 2352 } 2353 case ICmpInst::ICMP_ULT: 2354 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 2355 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2356 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 2357 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2358 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 2359 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2360 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2361 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C 2362 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2363 Builder->getInt(CI->getValue()-1)); 2364 2365 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear 2366 if (CI->isMinValue(true)) 2367 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, 2368 Constant::getAllOnesValue(Op0->getType())); 2369 } 2370 break; 2371 case ICmpInst::ICMP_UGT: 2372 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 2373 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2374 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 2375 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2376 2377 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 2378 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2379 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2380 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C 2381 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2382 Builder->getInt(CI->getValue()+1)); 2383 2384 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set 2385 if (CI->isMaxValue(true)) 2386 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, 2387 Constant::getNullValue(Op0->getType())); 2388 } 2389 break; 2390 case ICmpInst::ICMP_SLT: 2391 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 2392 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2393 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 2394 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2395 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 2396 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2397 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2398 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C 2399 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2400 Builder->getInt(CI->getValue()-1)); 2401 } 2402 break; 2403 case ICmpInst::ICMP_SGT: 2404 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 2405 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2406 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 2407 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2408 2409 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 2410 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 2411 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2412 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C 2413 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 2414 Builder->getInt(CI->getValue()+1)); 2415 } 2416 break; 2417 case ICmpInst::ICMP_SGE: 2418 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 2419 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 2420 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2421 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 2422 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2423 break; 2424 case ICmpInst::ICMP_SLE: 2425 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 2426 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 2427 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2428 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 2429 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2430 break; 2431 case ICmpInst::ICMP_UGE: 2432 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 2433 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 2434 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2435 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 2436 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2437 break; 2438 case ICmpInst::ICMP_ULE: 2439 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 2440 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 2441 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2442 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 2443 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2444 break; 2445 } 2446 2447 // Turn a signed comparison into an unsigned one if both operands 2448 // are known to have the same sign. 2449 if (I.isSigned() && 2450 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) || 2451 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative()))) 2452 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 2453 } 2454 2455 // Test if the ICmpInst instruction is used exclusively by a select as 2456 // part of a minimum or maximum operation. If so, refrain from doing 2457 // any other folding. This helps out other analyses which understand 2458 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 2459 // and CodeGen. And in this case, at least one of the comparison 2460 // operands has at least one user besides the compare (the select), 2461 // which would often largely negate the benefit of folding anyway. 2462 if (I.hasOneUse()) 2463 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin())) 2464 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) || 2465 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1)) 2466 return 0; 2467 2468 // See if we are doing a comparison between a constant and an instruction that 2469 // can be folded into the comparison. 2470 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) { 2471 // Since the RHS is a ConstantInt (CI), if the left hand side is an 2472 // instruction, see if that instruction also has constants so that the 2473 // instruction can be folded into the icmp 2474 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2475 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI)) 2476 return Res; 2477 } 2478 2479 // Handle icmp with constant (but not simple integer constant) RHS 2480 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 2481 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 2482 switch (LHSI->getOpcode()) { 2483 case Instruction::GetElementPtr: 2484 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2485 if (RHSC->isNullValue() && 2486 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2487 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2488 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2489 break; 2490 case Instruction::PHI: 2491 // Only fold icmp into the PHI if the phi and icmp are in the same 2492 // block. If in the same block, we're encouraging jump threading. If 2493 // not, we are just pessimizing the code by making an i1 phi. 2494 if (LHSI->getParent() == I.getParent()) 2495 if (Instruction *NV = FoldOpIntoPhi(I)) 2496 return NV; 2497 break; 2498 case Instruction::Select: { 2499 // If either operand of the select is a constant, we can fold the 2500 // comparison into the select arms, which will cause one to be 2501 // constant folded and the select turned into a bitwise or. 2502 Value *Op1 = 0, *Op2 = 0; 2503 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) 2504 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2505 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) 2506 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2507 2508 // We only want to perform this transformation if it will not lead to 2509 // additional code. This is true if either both sides of the select 2510 // fold to a constant (in which case the icmp is replaced with a select 2511 // which will usually simplify) or this is the only user of the 2512 // select (in which case we are trading a select+icmp for a simpler 2513 // select+icmp). 2514 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) { 2515 if (!Op1) 2516 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1), 2517 RHSC, I.getName()); 2518 if (!Op2) 2519 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2), 2520 RHSC, I.getName()); 2521 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2522 } 2523 break; 2524 } 2525 case Instruction::IntToPtr: 2526 // icmp pred inttoptr(X), null -> icmp pred X, 0 2527 if (RHSC->isNullValue() && TD && 2528 TD->getIntPtrType(RHSC->getType()) == 2529 LHSI->getOperand(0)->getType()) 2530 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0), 2531 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2532 break; 2533 2534 case Instruction::Load: 2535 // Try to optimize things like "A[i] > 4" to index computations. 2536 if (GetElementPtrInst *GEP = 2537 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2538 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2539 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2540 !cast<LoadInst>(LHSI)->isVolatile()) 2541 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2542 return Res; 2543 } 2544 break; 2545 } 2546 } 2547 2548 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 2549 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 2550 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I)) 2551 return NI; 2552 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 2553 if (Instruction *NI = FoldGEPICmp(GEP, Op0, 2554 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 2555 return NI; 2556 2557 // Test to see if the operands of the icmp are casted versions of other 2558 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 2559 // now. 2560 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 2561 if (Op0->getType()->isPointerTy() && 2562 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 2563 // We keep moving the cast from the left operand over to the right 2564 // operand, where it can often be eliminated completely. 2565 Op0 = CI->getOperand(0); 2566 2567 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 2568 // so eliminate it as well. 2569 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 2570 Op1 = CI2->getOperand(0); 2571 2572 // If Op1 is a constant, we can fold the cast into the constant. 2573 if (Op0->getType() != Op1->getType()) { 2574 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 2575 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 2576 } else { 2577 // Otherwise, cast the RHS right before the icmp 2578 Op1 = Builder->CreateBitCast(Op1, Op0->getType()); 2579 } 2580 } 2581 return new ICmpInst(I.getPredicate(), Op0, Op1); 2582 } 2583 } 2584 2585 if (isa<CastInst>(Op0)) { 2586 // Handle the special case of: icmp (cast bool to X), <cst> 2587 // This comes up when you have code like 2588 // int X = A < B; 2589 // if (X) ... 2590 // For generality, we handle any zero-extension of any operand comparison 2591 // with a constant or another cast from the same type. 2592 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 2593 if (Instruction *R = visitICmpInstWithCastAndCast(I)) 2594 return R; 2595 } 2596 2597 // Special logic for binary operators. 2598 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0); 2599 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1); 2600 if (BO0 || BO1) { 2601 CmpInst::Predicate Pred = I.getPredicate(); 2602 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; 2603 if (BO0 && isa<OverflowingBinaryOperator>(BO0)) 2604 NoOp0WrapProblem = ICmpInst::isEquality(Pred) || 2605 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) || 2606 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap()); 2607 if (BO1 && isa<OverflowingBinaryOperator>(BO1)) 2608 NoOp1WrapProblem = ICmpInst::isEquality(Pred) || 2609 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) || 2610 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap()); 2611 2612 // Analyze the case when either Op0 or Op1 is an add instruction. 2613 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). 2614 Value *A = 0, *B = 0, *C = 0, *D = 0; 2615 if (BO0 && BO0->getOpcode() == Instruction::Add) 2616 A = BO0->getOperand(0), B = BO0->getOperand(1); 2617 if (BO1 && BO1->getOpcode() == Instruction::Add) 2618 C = BO1->getOperand(0), D = BO1->getOperand(1); 2619 2620 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 2621 if ((A == Op1 || B == Op1) && NoOp0WrapProblem) 2622 return new ICmpInst(Pred, A == Op1 ? B : A, 2623 Constant::getNullValue(Op1->getType())); 2624 2625 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 2626 if ((C == Op0 || D == Op0) && NoOp1WrapProblem) 2627 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()), 2628 C == Op0 ? D : C); 2629 2630 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow. 2631 if (A && C && (A == C || A == D || B == C || B == D) && 2632 NoOp0WrapProblem && NoOp1WrapProblem && 2633 // Try not to increase register pressure. 2634 BO0->hasOneUse() && BO1->hasOneUse()) { 2635 // Determine Y and Z in the form icmp (X+Y), (X+Z). 2636 Value *Y, *Z; 2637 if (A == C) { 2638 // C + B == C + D -> B == D 2639 Y = B; 2640 Z = D; 2641 } else if (A == D) { 2642 // D + B == C + D -> B == C 2643 Y = B; 2644 Z = C; 2645 } else if (B == C) { 2646 // A + C == C + D -> A == D 2647 Y = A; 2648 Z = D; 2649 } else { 2650 assert(B == D); 2651 // A + D == C + D -> A == C 2652 Y = A; 2653 Z = C; 2654 } 2655 return new ICmpInst(Pred, Y, Z); 2656 } 2657 2658 // icmp slt (X + -1), Y -> icmp sle X, Y 2659 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT && 2660 match(B, m_AllOnes())) 2661 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1); 2662 2663 // icmp sge (X + -1), Y -> icmp sgt X, Y 2664 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE && 2665 match(B, m_AllOnes())) 2666 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1); 2667 2668 // icmp sle (X + 1), Y -> icmp slt X, Y 2669 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && 2670 match(B, m_One())) 2671 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1); 2672 2673 // icmp sgt (X + 1), Y -> icmp sge X, Y 2674 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && 2675 match(B, m_One())) 2676 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1); 2677 2678 // if C1 has greater magnitude than C2: 2679 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y 2680 // s.t. C3 = C1 - C2 2681 // 2682 // if C2 has greater magnitude than C1: 2683 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3) 2684 // s.t. C3 = C2 - C1 2685 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem && 2686 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) 2687 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B)) 2688 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) { 2689 const APInt &AP1 = C1->getValue(); 2690 const APInt &AP2 = C2->getValue(); 2691 if (AP1.isNegative() == AP2.isNegative()) { 2692 APInt AP1Abs = C1->getValue().abs(); 2693 APInt AP2Abs = C2->getValue().abs(); 2694 if (AP1Abs.uge(AP2Abs)) { 2695 ConstantInt *C3 = Builder->getInt(AP1 - AP2); 2696 Value *NewAdd = Builder->CreateNSWAdd(A, C3); 2697 return new ICmpInst(Pred, NewAdd, C); 2698 } else { 2699 ConstantInt *C3 = Builder->getInt(AP2 - AP1); 2700 Value *NewAdd = Builder->CreateNSWAdd(C, C3); 2701 return new ICmpInst(Pred, A, NewAdd); 2702 } 2703 } 2704 } 2705 2706 2707 // Analyze the case when either Op0 or Op1 is a sub instruction. 2708 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). 2709 A = 0; B = 0; C = 0; D = 0; 2710 if (BO0 && BO0->getOpcode() == Instruction::Sub) 2711 A = BO0->getOperand(0), B = BO0->getOperand(1); 2712 if (BO1 && BO1->getOpcode() == Instruction::Sub) 2713 C = BO1->getOperand(0), D = BO1->getOperand(1); 2714 2715 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow. 2716 if (A == Op1 && NoOp0WrapProblem) 2717 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B); 2718 2719 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow. 2720 if (C == Op0 && NoOp1WrapProblem) 2721 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType())); 2722 2723 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow. 2724 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem && 2725 // Try not to increase register pressure. 2726 BO0->hasOneUse() && BO1->hasOneUse()) 2727 return new ICmpInst(Pred, A, C); 2728 2729 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow. 2730 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem && 2731 // Try not to increase register pressure. 2732 BO0->hasOneUse() && BO1->hasOneUse()) 2733 return new ICmpInst(Pred, D, B); 2734 2735 BinaryOperator *SRem = NULL; 2736 // icmp (srem X, Y), Y 2737 if (BO0 && BO0->getOpcode() == Instruction::SRem && 2738 Op1 == BO0->getOperand(1)) 2739 SRem = BO0; 2740 // icmp Y, (srem X, Y) 2741 else if (BO1 && BO1->getOpcode() == Instruction::SRem && 2742 Op0 == BO1->getOperand(1)) 2743 SRem = BO1; 2744 if (SRem) { 2745 // We don't check hasOneUse to avoid increasing register pressure because 2746 // the value we use is the same value this instruction was already using. 2747 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) { 2748 default: break; 2749 case ICmpInst::ICMP_EQ: 2750 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 2751 case ICmpInst::ICMP_NE: 2752 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 2753 case ICmpInst::ICMP_SGT: 2754 case ICmpInst::ICMP_SGE: 2755 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1), 2756 Constant::getAllOnesValue(SRem->getType())); 2757 case ICmpInst::ICMP_SLT: 2758 case ICmpInst::ICMP_SLE: 2759 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1), 2760 Constant::getNullValue(SRem->getType())); 2761 } 2762 } 2763 2764 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && 2765 BO0->hasOneUse() && BO1->hasOneUse() && 2766 BO0->getOperand(1) == BO1->getOperand(1)) { 2767 switch (BO0->getOpcode()) { 2768 default: break; 2769 case Instruction::Add: 2770 case Instruction::Sub: 2771 case Instruction::Xor: 2772 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 2773 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2774 BO1->getOperand(0)); 2775 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b 2776 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2777 if (CI->getValue().isSignBit()) { 2778 ICmpInst::Predicate Pred = I.isSigned() 2779 ? I.getUnsignedPredicate() 2780 : I.getSignedPredicate(); 2781 return new ICmpInst(Pred, BO0->getOperand(0), 2782 BO1->getOperand(0)); 2783 } 2784 2785 if (CI->isMaxValue(true)) { 2786 ICmpInst::Predicate Pred = I.isSigned() 2787 ? I.getUnsignedPredicate() 2788 : I.getSignedPredicate(); 2789 Pred = I.getSwappedPredicate(Pred); 2790 return new ICmpInst(Pred, BO0->getOperand(0), 2791 BO1->getOperand(0)); 2792 } 2793 } 2794 break; 2795 case Instruction::Mul: 2796 if (!I.isEquality()) 2797 break; 2798 2799 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) { 2800 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask 2801 // Mask = -1 >> count-trailing-zeros(Cst). 2802 if (!CI->isZero() && !CI->isOne()) { 2803 const APInt &AP = CI->getValue(); 2804 ConstantInt *Mask = ConstantInt::get(I.getContext(), 2805 APInt::getLowBitsSet(AP.getBitWidth(), 2806 AP.getBitWidth() - 2807 AP.countTrailingZeros())); 2808 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask); 2809 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask); 2810 return new ICmpInst(I.getPredicate(), And1, And2); 2811 } 2812 } 2813 break; 2814 case Instruction::UDiv: 2815 case Instruction::LShr: 2816 if (I.isSigned()) 2817 break; 2818 // fall-through 2819 case Instruction::SDiv: 2820 case Instruction::AShr: 2821 if (!BO0->isExact() || !BO1->isExact()) 2822 break; 2823 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2824 BO1->getOperand(0)); 2825 case Instruction::Shl: { 2826 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap(); 2827 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap(); 2828 if (!NUW && !NSW) 2829 break; 2830 if (!NSW && I.isSigned()) 2831 break; 2832 return new ICmpInst(I.getPredicate(), BO0->getOperand(0), 2833 BO1->getOperand(0)); 2834 } 2835 } 2836 } 2837 } 2838 2839 { Value *A, *B; 2840 // Transform (A & ~B) == 0 --> (A & B) != 0 2841 // and (A & ~B) != 0 --> (A & B) == 0 2842 // if A is a power of 2. 2843 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) && 2844 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(A) && I.isEquality()) 2845 return new ICmpInst(I.getInversePredicate(), 2846 Builder->CreateAnd(A, B), 2847 Op1); 2848 2849 // ~x < ~y --> y < x 2850 // ~x < cst --> ~cst < x 2851 if (match(Op0, m_Not(m_Value(A)))) { 2852 if (match(Op1, m_Not(m_Value(B)))) 2853 return new ICmpInst(I.getPredicate(), B, A); 2854 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) 2855 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A); 2856 } 2857 2858 // (a+b) <u a --> llvm.uadd.with.overflow. 2859 // (a+b) <u b --> llvm.uadd.with.overflow. 2860 if (I.getPredicate() == ICmpInst::ICMP_ULT && 2861 match(Op0, m_Add(m_Value(A), m_Value(B))) && 2862 (Op1 == A || Op1 == B)) 2863 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this)) 2864 return R; 2865 2866 // a >u (a+b) --> llvm.uadd.with.overflow. 2867 // b >u (a+b) --> llvm.uadd.with.overflow. 2868 if (I.getPredicate() == ICmpInst::ICMP_UGT && 2869 match(Op1, m_Add(m_Value(A), m_Value(B))) && 2870 (Op0 == A || Op0 == B)) 2871 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this)) 2872 return R; 2873 } 2874 2875 if (I.isEquality()) { 2876 Value *A, *B, *C, *D; 2877 2878 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 2879 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 2880 Value *OtherVal = A == Op1 ? B : A; 2881 return new ICmpInst(I.getPredicate(), OtherVal, 2882 Constant::getNullValue(A->getType())); 2883 } 2884 2885 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 2886 // A^c1 == C^c2 --> A == C^(c1^c2) 2887 ConstantInt *C1, *C2; 2888 if (match(B, m_ConstantInt(C1)) && 2889 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) { 2890 Constant *NC = Builder->getInt(C1->getValue() ^ C2->getValue()); 2891 Value *Xor = Builder->CreateXor(C, NC); 2892 return new ICmpInst(I.getPredicate(), A, Xor); 2893 } 2894 2895 // A^B == A^D -> B == D 2896 if (A == C) return new ICmpInst(I.getPredicate(), B, D); 2897 if (A == D) return new ICmpInst(I.getPredicate(), B, C); 2898 if (B == C) return new ICmpInst(I.getPredicate(), A, D); 2899 if (B == D) return new ICmpInst(I.getPredicate(), A, C); 2900 } 2901 } 2902 2903 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && 2904 (A == Op0 || B == Op0)) { 2905 // A == (A^B) -> B == 0 2906 Value *OtherVal = A == Op0 ? B : A; 2907 return new ICmpInst(I.getPredicate(), OtherVal, 2908 Constant::getNullValue(A->getType())); 2909 } 2910 2911 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 2912 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) && 2913 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) { 2914 Value *X = 0, *Y = 0, *Z = 0; 2915 2916 if (A == C) { 2917 X = B; Y = D; Z = A; 2918 } else if (A == D) { 2919 X = B; Y = C; Z = A; 2920 } else if (B == C) { 2921 X = A; Y = D; Z = B; 2922 } else if (B == D) { 2923 X = A; Y = C; Z = B; 2924 } 2925 2926 if (X) { // Build (X^Y) & Z 2927 Op1 = Builder->CreateXor(X, Y); 2928 Op1 = Builder->CreateAnd(Op1, Z); 2929 I.setOperand(0, Op1); 2930 I.setOperand(1, Constant::getNullValue(Op1->getType())); 2931 return &I; 2932 } 2933 } 2934 2935 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B) 2936 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B) 2937 ConstantInt *Cst1; 2938 if ((Op0->hasOneUse() && 2939 match(Op0, m_ZExt(m_Value(A))) && 2940 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) || 2941 (Op1->hasOneUse() && 2942 match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) && 2943 match(Op1, m_ZExt(m_Value(A))))) { 2944 APInt Pow2 = Cst1->getValue() + 1; 2945 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) && 2946 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth()) 2947 return new ICmpInst(I.getPredicate(), A, 2948 Builder->CreateTrunc(B, A->getType())); 2949 } 2950 2951 // (A >> C) == (B >> C) --> (A^B) u< (1 << C) 2952 // For lshr and ashr pairs. 2953 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) && 2954 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) || 2955 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) && 2956 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) { 2957 unsigned TypeBits = Cst1->getBitWidth(); 2958 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits); 2959 if (ShAmt < TypeBits && ShAmt != 0) { 2960 ICmpInst::Predicate Pred = I.getPredicate() == ICmpInst::ICMP_NE 2961 ? ICmpInst::ICMP_UGE 2962 : ICmpInst::ICMP_ULT; 2963 Value *Xor = Builder->CreateXor(A, B, I.getName() + ".unshifted"); 2964 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt); 2965 return new ICmpInst(Pred, Xor, Builder->getInt(CmpVal)); 2966 } 2967 } 2968 2969 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to 2970 // "icmp (and X, mask), cst" 2971 uint64_t ShAmt = 0; 2972 if (Op0->hasOneUse() && 2973 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), 2974 m_ConstantInt(ShAmt))))) && 2975 match(Op1, m_ConstantInt(Cst1)) && 2976 // Only do this when A has multiple uses. This is most important to do 2977 // when it exposes other optimizations. 2978 !A->hasOneUse()) { 2979 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits(); 2980 2981 if (ShAmt < ASize) { 2982 APInt MaskV = 2983 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits()); 2984 MaskV <<= ShAmt; 2985 2986 APInt CmpV = Cst1->getValue().zext(ASize); 2987 CmpV <<= ShAmt; 2988 2989 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV)); 2990 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV)); 2991 } 2992 } 2993 } 2994 2995 { 2996 Value *X; ConstantInt *Cst; 2997 // icmp X+Cst, X 2998 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 2999 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate()); 3000 3001 // icmp X, X+Cst 3002 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 3003 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate()); 3004 } 3005 return Changed ? &I : 0; 3006 } 3007 3008 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible. 3009 /// 3010 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I, 3011 Instruction *LHSI, 3012 Constant *RHSC) { 3013 if (!isa<ConstantFP>(RHSC)) return 0; 3014 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 3015 3016 // Get the width of the mantissa. We don't want to hack on conversions that 3017 // might lose information from the integer, e.g. "i64 -> float" 3018 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 3019 if (MantissaWidth == -1) return 0; // Unknown. 3020 3021 // Check to see that the input is converted from an integer type that is small 3022 // enough that preserves all bits. TODO: check here for "known" sign bits. 3023 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 3024 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits(); 3025 3026 // If this is a uitofp instruction, we need an extra bit to hold the sign. 3027 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 3028 if (LHSUnsigned) 3029 ++InputSize; 3030 3031 // If the conversion would lose info, don't hack on this. 3032 if ((int)InputSize > MantissaWidth) 3033 return 0; 3034 3035 // Otherwise, we can potentially simplify the comparison. We know that it 3036 // will always come through as an integer value and we know the constant is 3037 // not a NAN (it would have been previously simplified). 3038 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 3039 3040 ICmpInst::Predicate Pred; 3041 switch (I.getPredicate()) { 3042 default: llvm_unreachable("Unexpected predicate!"); 3043 case FCmpInst::FCMP_UEQ: 3044 case FCmpInst::FCMP_OEQ: 3045 Pred = ICmpInst::ICMP_EQ; 3046 break; 3047 case FCmpInst::FCMP_UGT: 3048 case FCmpInst::FCMP_OGT: 3049 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 3050 break; 3051 case FCmpInst::FCMP_UGE: 3052 case FCmpInst::FCMP_OGE: 3053 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 3054 break; 3055 case FCmpInst::FCMP_ULT: 3056 case FCmpInst::FCMP_OLT: 3057 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 3058 break; 3059 case FCmpInst::FCMP_ULE: 3060 case FCmpInst::FCMP_OLE: 3061 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 3062 break; 3063 case FCmpInst::FCMP_UNE: 3064 case FCmpInst::FCMP_ONE: 3065 Pred = ICmpInst::ICMP_NE; 3066 break; 3067 case FCmpInst::FCMP_ORD: 3068 return ReplaceInstUsesWith(I, Builder->getTrue()); 3069 case FCmpInst::FCMP_UNO: 3070 return ReplaceInstUsesWith(I, Builder->getFalse()); 3071 } 3072 3073 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 3074 3075 // Now we know that the APFloat is a normal number, zero or inf. 3076 3077 // See if the FP constant is too large for the integer. For example, 3078 // comparing an i8 to 300.0. 3079 unsigned IntWidth = IntTy->getScalarSizeInBits(); 3080 3081 if (!LHSUnsigned) { 3082 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 3083 // and large values. 3084 APFloat SMax(RHS.getSemantics()); 3085 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 3086 APFloat::rmNearestTiesToEven); 3087 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 3088 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 3089 Pred == ICmpInst::ICMP_SLE) 3090 return ReplaceInstUsesWith(I, Builder->getTrue()); 3091 return ReplaceInstUsesWith(I, Builder->getFalse()); 3092 } 3093 } else { 3094 // If the RHS value is > UnsignedMax, fold the comparison. This handles 3095 // +INF and large values. 3096 APFloat UMax(RHS.getSemantics()); 3097 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 3098 APFloat::rmNearestTiesToEven); 3099 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 3100 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 3101 Pred == ICmpInst::ICMP_ULE) 3102 return ReplaceInstUsesWith(I, Builder->getTrue()); 3103 return ReplaceInstUsesWith(I, Builder->getFalse()); 3104 } 3105 } 3106 3107 if (!LHSUnsigned) { 3108 // See if the RHS value is < SignedMin. 3109 APFloat SMin(RHS.getSemantics()); 3110 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 3111 APFloat::rmNearestTiesToEven); 3112 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 3113 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 3114 Pred == ICmpInst::ICMP_SGE) 3115 return ReplaceInstUsesWith(I, Builder->getTrue()); 3116 return ReplaceInstUsesWith(I, Builder->getFalse()); 3117 } 3118 } else { 3119 // See if the RHS value is < UnsignedMin. 3120 APFloat SMin(RHS.getSemantics()); 3121 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true, 3122 APFloat::rmNearestTiesToEven); 3123 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0 3124 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || 3125 Pred == ICmpInst::ICMP_UGE) 3126 return ReplaceInstUsesWith(I, Builder->getTrue()); 3127 return ReplaceInstUsesWith(I, Builder->getFalse()); 3128 } 3129 } 3130 3131 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 3132 // [0, UMAX], but it may still be fractional. See if it is fractional by 3133 // casting the FP value to the integer value and back, checking for equality. 3134 // Don't do this for zero, because -0.0 is not fractional. 3135 Constant *RHSInt = LHSUnsigned 3136 ? ConstantExpr::getFPToUI(RHSC, IntTy) 3137 : ConstantExpr::getFPToSI(RHSC, IntTy); 3138 if (!RHS.isZero()) { 3139 bool Equal = LHSUnsigned 3140 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 3141 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 3142 if (!Equal) { 3143 // If we had a comparison against a fractional value, we have to adjust 3144 // the compare predicate and sometimes the value. RHSC is rounded towards 3145 // zero at this point. 3146 switch (Pred) { 3147 default: llvm_unreachable("Unexpected integer comparison!"); 3148 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 3149 return ReplaceInstUsesWith(I, Builder->getTrue()); 3150 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 3151 return ReplaceInstUsesWith(I, Builder->getFalse()); 3152 case ICmpInst::ICMP_ULE: 3153 // (float)int <= 4.4 --> int <= 4 3154 // (float)int <= -4.4 --> false 3155 if (RHS.isNegative()) 3156 return ReplaceInstUsesWith(I, Builder->getFalse()); 3157 break; 3158 case ICmpInst::ICMP_SLE: 3159 // (float)int <= 4.4 --> int <= 4 3160 // (float)int <= -4.4 --> int < -4 3161 if (RHS.isNegative()) 3162 Pred = ICmpInst::ICMP_SLT; 3163 break; 3164 case ICmpInst::ICMP_ULT: 3165 // (float)int < -4.4 --> false 3166 // (float)int < 4.4 --> int <= 4 3167 if (RHS.isNegative()) 3168 return ReplaceInstUsesWith(I, Builder->getFalse()); 3169 Pred = ICmpInst::ICMP_ULE; 3170 break; 3171 case ICmpInst::ICMP_SLT: 3172 // (float)int < -4.4 --> int < -4 3173 // (float)int < 4.4 --> int <= 4 3174 if (!RHS.isNegative()) 3175 Pred = ICmpInst::ICMP_SLE; 3176 break; 3177 case ICmpInst::ICMP_UGT: 3178 // (float)int > 4.4 --> int > 4 3179 // (float)int > -4.4 --> true 3180 if (RHS.isNegative()) 3181 return ReplaceInstUsesWith(I, Builder->getTrue()); 3182 break; 3183 case ICmpInst::ICMP_SGT: 3184 // (float)int > 4.4 --> int > 4 3185 // (float)int > -4.4 --> int >= -4 3186 if (RHS.isNegative()) 3187 Pred = ICmpInst::ICMP_SGE; 3188 break; 3189 case ICmpInst::ICMP_UGE: 3190 // (float)int >= -4.4 --> true 3191 // (float)int >= 4.4 --> int > 4 3192 if (RHS.isNegative()) 3193 return ReplaceInstUsesWith(I, Builder->getTrue()); 3194 Pred = ICmpInst::ICMP_UGT; 3195 break; 3196 case ICmpInst::ICMP_SGE: 3197 // (float)int >= -4.4 --> int >= -4 3198 // (float)int >= 4.4 --> int > 4 3199 if (!RHS.isNegative()) 3200 Pred = ICmpInst::ICMP_SGT; 3201 break; 3202 } 3203 } 3204 } 3205 3206 // Lower this FP comparison into an appropriate integer version of the 3207 // comparison. 3208 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 3209 } 3210 3211 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 3212 bool Changed = false; 3213 3214 /// Orders the operands of the compare so that they are listed from most 3215 /// complex to least complex. This puts constants before unary operators, 3216 /// before binary operators. 3217 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 3218 I.swapOperands(); 3219 Changed = true; 3220 } 3221 3222 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3223 3224 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD)) 3225 return ReplaceInstUsesWith(I, V); 3226 3227 // Simplify 'fcmp pred X, X' 3228 if (Op0 == Op1) { 3229 switch (I.getPredicate()) { 3230 default: llvm_unreachable("Unknown predicate!"); 3231 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 3232 case FCmpInst::FCMP_ULT: // True if unordered or less than 3233 case FCmpInst::FCMP_UGT: // True if unordered or greater than 3234 case FCmpInst::FCMP_UNE: // True if unordered or not equal 3235 // Canonicalize these to be 'fcmp uno %X, 0.0'. 3236 I.setPredicate(FCmpInst::FCMP_UNO); 3237 I.setOperand(1, Constant::getNullValue(Op0->getType())); 3238 return &I; 3239 3240 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 3241 case FCmpInst::FCMP_OEQ: // True if ordered and equal 3242 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 3243 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 3244 // Canonicalize these to be 'fcmp ord %X, 0.0'. 3245 I.setPredicate(FCmpInst::FCMP_ORD); 3246 I.setOperand(1, Constant::getNullValue(Op0->getType())); 3247 return &I; 3248 } 3249 } 3250 3251 // Handle fcmp with constant RHS 3252 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 3253 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 3254 switch (LHSI->getOpcode()) { 3255 case Instruction::FPExt: { 3256 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless 3257 FPExtInst *LHSExt = cast<FPExtInst>(LHSI); 3258 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC); 3259 if (!RHSF) 3260 break; 3261 3262 const fltSemantics *Sem; 3263 // FIXME: This shouldn't be here. 3264 if (LHSExt->getSrcTy()->isHalfTy()) 3265 Sem = &APFloat::IEEEhalf; 3266 else if (LHSExt->getSrcTy()->isFloatTy()) 3267 Sem = &APFloat::IEEEsingle; 3268 else if (LHSExt->getSrcTy()->isDoubleTy()) 3269 Sem = &APFloat::IEEEdouble; 3270 else if (LHSExt->getSrcTy()->isFP128Ty()) 3271 Sem = &APFloat::IEEEquad; 3272 else if (LHSExt->getSrcTy()->isX86_FP80Ty()) 3273 Sem = &APFloat::x87DoubleExtended; 3274 else if (LHSExt->getSrcTy()->isPPC_FP128Ty()) 3275 Sem = &APFloat::PPCDoubleDouble; 3276 else 3277 break; 3278 3279 bool Lossy; 3280 APFloat F = RHSF->getValueAPF(); 3281 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy); 3282 3283 // Avoid lossy conversions and denormals. Zero is a special case 3284 // that's OK to convert. 3285 APFloat Fabs = F; 3286 Fabs.clearSign(); 3287 if (!Lossy && 3288 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) != 3289 APFloat::cmpLessThan) || Fabs.isZero())) 3290 3291 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3292 ConstantFP::get(RHSC->getContext(), F)); 3293 break; 3294 } 3295 case Instruction::PHI: 3296 // Only fold fcmp into the PHI if the phi and fcmp are in the same 3297 // block. If in the same block, we're encouraging jump threading. If 3298 // not, we are just pessimizing the code by making an i1 phi. 3299 if (LHSI->getParent() == I.getParent()) 3300 if (Instruction *NV = FoldOpIntoPhi(I)) 3301 return NV; 3302 break; 3303 case Instruction::SIToFP: 3304 case Instruction::UIToFP: 3305 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC)) 3306 return NV; 3307 break; 3308 case Instruction::Select: { 3309 // If either operand of the select is a constant, we can fold the 3310 // comparison into the select arms, which will cause one to be 3311 // constant folded and the select turned into a bitwise or. 3312 Value *Op1 = 0, *Op2 = 0; 3313 if (LHSI->hasOneUse()) { 3314 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 3315 // Fold the known value into the constant operand. 3316 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 3317 // Insert a new FCmp of the other select operand. 3318 Op2 = Builder->CreateFCmp(I.getPredicate(), 3319 LHSI->getOperand(2), RHSC, I.getName()); 3320 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 3321 // Fold the known value into the constant operand. 3322 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC); 3323 // Insert a new FCmp of the other select operand. 3324 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1), 3325 RHSC, I.getName()); 3326 } 3327 } 3328 3329 if (Op1) 3330 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 3331 break; 3332 } 3333 case Instruction::FSub: { 3334 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C 3335 Value *Op; 3336 if (match(LHSI, m_FNeg(m_Value(Op)))) 3337 return new FCmpInst(I.getSwappedPredicate(), Op, 3338 ConstantExpr::getFNeg(RHSC)); 3339 break; 3340 } 3341 case Instruction::Load: 3342 if (GetElementPtrInst *GEP = 3343 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 3344 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 3345 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 3346 !cast<LoadInst>(LHSI)->isVolatile()) 3347 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I)) 3348 return Res; 3349 } 3350 break; 3351 case Instruction::Call: { 3352 CallInst *CI = cast<CallInst>(LHSI); 3353 LibFunc::Func Func; 3354 // Various optimization for fabs compared with zero. 3355 if (RHSC->isNullValue() && CI->getCalledFunction() && 3356 TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) && 3357 TLI->has(Func)) { 3358 if (Func == LibFunc::fabs || Func == LibFunc::fabsf || 3359 Func == LibFunc::fabsl) { 3360 switch (I.getPredicate()) { 3361 default: break; 3362 // fabs(x) < 0 --> false 3363 case FCmpInst::FCMP_OLT: 3364 return ReplaceInstUsesWith(I, Builder->getFalse()); 3365 // fabs(x) > 0 --> x != 0 3366 case FCmpInst::FCMP_OGT: 3367 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), 3368 RHSC); 3369 // fabs(x) <= 0 --> x == 0 3370 case FCmpInst::FCMP_OLE: 3371 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), 3372 RHSC); 3373 // fabs(x) >= 0 --> !isnan(x) 3374 case FCmpInst::FCMP_OGE: 3375 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), 3376 RHSC); 3377 // fabs(x) == 0 --> x == 0 3378 // fabs(x) != 0 --> x != 0 3379 case FCmpInst::FCMP_OEQ: 3380 case FCmpInst::FCMP_UEQ: 3381 case FCmpInst::FCMP_ONE: 3382 case FCmpInst::FCMP_UNE: 3383 return new FCmpInst(I.getPredicate(), CI->getArgOperand(0), 3384 RHSC); 3385 } 3386 } 3387 } 3388 } 3389 } 3390 } 3391 3392 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y 3393 Value *X, *Y; 3394 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 3395 return new FCmpInst(I.getSwappedPredicate(), X, Y); 3396 3397 // fcmp (fpext x), (fpext y) -> fcmp x, y 3398 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0)) 3399 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1)) 3400 if (LHSExt->getSrcTy() == RHSExt->getSrcTy()) 3401 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0), 3402 RHSExt->getOperand(0)); 3403 3404 return Changed ? &I : 0; 3405 } 3406