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 "InstCombineInternal.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/SetVector.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/Analysis/ConstantFolding.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/TargetLibraryInfo.h" 21 #include "llvm/IR/ConstantRange.h" 22 #include "llvm/IR/DataLayout.h" 23 #include "llvm/IR/GetElementPtrTypeIterator.h" 24 #include "llvm/IR/IntrinsicInst.h" 25 #include "llvm/IR/PatternMatch.h" 26 #include "llvm/Support/Debug.h" 27 #include "llvm/Support/KnownBits.h" 28 29 using namespace llvm; 30 using namespace PatternMatch; 31 32 #define DEBUG_TYPE "instcombine" 33 34 // How many times is a select replaced by one of its operands? 35 STATISTIC(NumSel, "Number of select opts"); 36 37 38 /// Compute Result = In1+In2, returning true if the result overflowed for this 39 /// type. 40 static bool addWithOverflow(APInt &Result, const APInt &In1, 41 const APInt &In2, bool IsSigned = false) { 42 bool Overflow; 43 if (IsSigned) 44 Result = In1.sadd_ov(In2, Overflow); 45 else 46 Result = In1.uadd_ov(In2, Overflow); 47 48 return Overflow; 49 } 50 51 /// Compute Result = In1-In2, returning true if the result overflowed for this 52 /// type. 53 static bool subWithOverflow(APInt &Result, const APInt &In1, 54 const APInt &In2, bool IsSigned = false) { 55 bool Overflow; 56 if (IsSigned) 57 Result = In1.ssub_ov(In2, Overflow); 58 else 59 Result = In1.usub_ov(In2, Overflow); 60 61 return Overflow; 62 } 63 64 /// Given an icmp instruction, return true if any use of this comparison is a 65 /// branch on sign bit comparison. 66 static bool hasBranchUse(ICmpInst &I) { 67 for (auto *U : I.users()) 68 if (isa<BranchInst>(U)) 69 return true; 70 return false; 71 } 72 73 /// Given an exploded icmp instruction, return true if the comparison only 74 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the 75 /// result of the comparison is true when the input value is signed. 76 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS, 77 bool &TrueIfSigned) { 78 switch (Pred) { 79 case ICmpInst::ICMP_SLT: // True if LHS s< 0 80 TrueIfSigned = true; 81 return RHS.isNullValue(); 82 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 83 TrueIfSigned = true; 84 return RHS.isAllOnesValue(); 85 case ICmpInst::ICMP_SGT: // True if LHS s> -1 86 TrueIfSigned = false; 87 return RHS.isAllOnesValue(); 88 case ICmpInst::ICMP_UGT: 89 // True if LHS u> RHS and RHS == high-bit-mask - 1 90 TrueIfSigned = true; 91 return RHS.isMaxSignedValue(); 92 case ICmpInst::ICMP_UGE: 93 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) 94 TrueIfSigned = true; 95 return RHS.isSignMask(); 96 default: 97 return false; 98 } 99 } 100 101 /// Returns true if the exploded icmp can be expressed as a signed comparison 102 /// to zero and updates the predicate accordingly. 103 /// The signedness of the comparison is preserved. 104 /// TODO: Refactor with decomposeBitTestICmp()? 105 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) { 106 if (!ICmpInst::isSigned(Pred)) 107 return false; 108 109 if (C.isNullValue()) 110 return ICmpInst::isRelational(Pred); 111 112 if (C.isOneValue()) { 113 if (Pred == ICmpInst::ICMP_SLT) { 114 Pred = ICmpInst::ICMP_SLE; 115 return true; 116 } 117 } else if (C.isAllOnesValue()) { 118 if (Pred == ICmpInst::ICMP_SGT) { 119 Pred = ICmpInst::ICMP_SGE; 120 return true; 121 } 122 } 123 124 return false; 125 } 126 127 /// Given a signed integer type and a set of known zero and one bits, compute 128 /// the maximum and minimum values that could have the specified known zero and 129 /// known one bits, returning them in Min/Max. 130 /// TODO: Move to method on KnownBits struct? 131 static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known, 132 APInt &Min, APInt &Max) { 133 assert(Known.getBitWidth() == Min.getBitWidth() && 134 Known.getBitWidth() == Max.getBitWidth() && 135 "KnownZero, KnownOne and Min, Max must have equal bitwidth."); 136 APInt UnknownBits = ~(Known.Zero|Known.One); 137 138 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign 139 // bit if it is unknown. 140 Min = Known.One; 141 Max = Known.One|UnknownBits; 142 143 if (UnknownBits.isNegative()) { // Sign bit is unknown 144 Min.setSignBit(); 145 Max.clearSignBit(); 146 } 147 } 148 149 /// Given an unsigned integer type and a set of known zero and one bits, compute 150 /// the maximum and minimum values that could have the specified known zero and 151 /// known one bits, returning them in Min/Max. 152 /// TODO: Move to method on KnownBits struct? 153 static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known, 154 APInt &Min, APInt &Max) { 155 assert(Known.getBitWidth() == Min.getBitWidth() && 156 Known.getBitWidth() == Max.getBitWidth() && 157 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); 158 APInt UnknownBits = ~(Known.Zero|Known.One); 159 160 // The minimum value is when the unknown bits are all zeros. 161 Min = Known.One; 162 // The maximum value is when the unknown bits are all ones. 163 Max = Known.One|UnknownBits; 164 } 165 166 /// This is called when we see this pattern: 167 /// cmp pred (load (gep GV, ...)), cmpcst 168 /// where GV is a global variable with a constant initializer. Try to simplify 169 /// this into some simple computation that does not need the load. For example 170 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". 171 /// 172 /// If AndCst is non-null, then the loaded value is masked with that constant 173 /// before doing the comparison. This handles cases like "A[i]&4 == 0". 174 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, 175 GlobalVariable *GV, 176 CmpInst &ICI, 177 ConstantInt *AndCst) { 178 Constant *Init = GV->getInitializer(); 179 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init)) 180 return nullptr; 181 182 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); 183 // Don't blow up on huge arrays. 184 if (ArrayElementCount > MaxArraySizeForCombine) 185 return nullptr; 186 187 // There are many forms of this optimization we can handle, for now, just do 188 // the simple index into a single-dimensional array. 189 // 190 // Require: GEP GV, 0, i {{, constant indices}} 191 if (GEP->getNumOperands() < 3 || 192 !isa<ConstantInt>(GEP->getOperand(1)) || 193 !cast<ConstantInt>(GEP->getOperand(1))->isZero() || 194 isa<Constant>(GEP->getOperand(2))) 195 return nullptr; 196 197 // Check that indices after the variable are constants and in-range for the 198 // type they index. Collect the indices. This is typically for arrays of 199 // structs. 200 SmallVector<unsigned, 4> LaterIndices; 201 202 Type *EltTy = Init->getType()->getArrayElementType(); 203 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { 204 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i)); 205 if (!Idx) return nullptr; // Variable index. 206 207 uint64_t IdxVal = Idx->getZExtValue(); 208 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index. 209 210 if (StructType *STy = dyn_cast<StructType>(EltTy)) 211 EltTy = STy->getElementType(IdxVal); 212 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) { 213 if (IdxVal >= ATy->getNumElements()) return nullptr; 214 EltTy = ATy->getElementType(); 215 } else { 216 return nullptr; // Unknown type. 217 } 218 219 LaterIndices.push_back(IdxVal); 220 } 221 222 enum { Overdefined = -3, Undefined = -2 }; 223 224 // Variables for our state machines. 225 226 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form 227 // "i == 47 | i == 87", where 47 is the first index the condition is true for, 228 // and 87 is the second (and last) index. FirstTrueElement is -2 when 229 // undefined, otherwise set to the first true element. SecondTrueElement is 230 // -2 when undefined, -3 when overdefined and >= 0 when that index is true. 231 int FirstTrueElement = Undefined, SecondTrueElement = Undefined; 232 233 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the 234 // form "i != 47 & i != 87". Same state transitions as for true elements. 235 int FirstFalseElement = Undefined, SecondFalseElement = Undefined; 236 237 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these 238 /// define a state machine that triggers for ranges of values that the index 239 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. 240 /// This is -2 when undefined, -3 when overdefined, and otherwise the last 241 /// index in the range (inclusive). We use -2 for undefined here because we 242 /// use relative comparisons and don't want 0-1 to match -1. 243 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; 244 245 // MagicBitvector - This is a magic bitvector where we set a bit if the 246 // comparison is true for element 'i'. If there are 64 elements or less in 247 // the array, this will fully represent all the comparison results. 248 uint64_t MagicBitvector = 0; 249 250 // Scan the array and see if one of our patterns matches. 251 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1)); 252 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { 253 Constant *Elt = Init->getAggregateElement(i); 254 if (!Elt) return nullptr; 255 256 // If this is indexing an array of structures, get the structure element. 257 if (!LaterIndices.empty()) 258 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); 259 260 // If the element is masked, handle it. 261 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); 262 263 // Find out if the comparison would be true or false for the i'th element. 264 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, 265 CompareRHS, DL, &TLI); 266 // If the result is undef for this element, ignore it. 267 if (isa<UndefValue>(C)) { 268 // Extend range state machines to cover this element in case there is an 269 // undef in the middle of the range. 270 if (TrueRangeEnd == (int)i-1) 271 TrueRangeEnd = i; 272 if (FalseRangeEnd == (int)i-1) 273 FalseRangeEnd = i; 274 continue; 275 } 276 277 // If we can't compute the result for any of the elements, we have to give 278 // up evaluating the entire conditional. 279 if (!isa<ConstantInt>(C)) return nullptr; 280 281 // Otherwise, we know if the comparison is true or false for this element, 282 // update our state machines. 283 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero(); 284 285 // State machine for single/double/range index comparison. 286 if (IsTrueForElt) { 287 // Update the TrueElement state machine. 288 if (FirstTrueElement == Undefined) 289 FirstTrueElement = TrueRangeEnd = i; // First true element. 290 else { 291 // Update double-compare state machine. 292 if (SecondTrueElement == Undefined) 293 SecondTrueElement = i; 294 else 295 SecondTrueElement = Overdefined; 296 297 // Update range state machine. 298 if (TrueRangeEnd == (int)i-1) 299 TrueRangeEnd = i; 300 else 301 TrueRangeEnd = Overdefined; 302 } 303 } else { 304 // Update the FalseElement state machine. 305 if (FirstFalseElement == Undefined) 306 FirstFalseElement = FalseRangeEnd = i; // First false element. 307 else { 308 // Update double-compare state machine. 309 if (SecondFalseElement == Undefined) 310 SecondFalseElement = i; 311 else 312 SecondFalseElement = Overdefined; 313 314 // Update range state machine. 315 if (FalseRangeEnd == (int)i-1) 316 FalseRangeEnd = i; 317 else 318 FalseRangeEnd = Overdefined; 319 } 320 } 321 322 // If this element is in range, update our magic bitvector. 323 if (i < 64 && IsTrueForElt) 324 MagicBitvector |= 1ULL << i; 325 326 // If all of our states become overdefined, bail out early. Since the 327 // predicate is expensive, only check it every 8 elements. This is only 328 // really useful for really huge arrays. 329 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && 330 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && 331 FalseRangeEnd == Overdefined) 332 return nullptr; 333 } 334 335 // Now that we've scanned the entire array, emit our new comparison(s). We 336 // order the state machines in complexity of the generated code. 337 Value *Idx = GEP->getOperand(2); 338 339 // If the index is larger than the pointer size of the target, truncate the 340 // index down like the GEP would do implicitly. We don't have to do this for 341 // an inbounds GEP because the index can't be out of range. 342 if (!GEP->isInBounds()) { 343 Type *IntPtrTy = DL.getIntPtrType(GEP->getType()); 344 unsigned PtrSize = IntPtrTy->getIntegerBitWidth(); 345 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize) 346 Idx = Builder.CreateTrunc(Idx, IntPtrTy); 347 } 348 349 // If the comparison is only true for one or two elements, emit direct 350 // comparisons. 351 if (SecondTrueElement != Overdefined) { 352 // None true -> false. 353 if (FirstTrueElement == Undefined) 354 return replaceInstUsesWith(ICI, Builder.getFalse()); 355 356 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); 357 358 // True for one element -> 'i == 47'. 359 if (SecondTrueElement == Undefined) 360 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); 361 362 // True for two elements -> 'i == 47 | i == 72'. 363 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx); 364 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); 365 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx); 366 return BinaryOperator::CreateOr(C1, C2); 367 } 368 369 // If the comparison is only false for one or two elements, emit direct 370 // comparisons. 371 if (SecondFalseElement != Overdefined) { 372 // None false -> true. 373 if (FirstFalseElement == Undefined) 374 return replaceInstUsesWith(ICI, Builder.getTrue()); 375 376 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); 377 378 // False for one element -> 'i != 47'. 379 if (SecondFalseElement == Undefined) 380 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); 381 382 // False for two elements -> 'i != 47 & i != 72'. 383 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx); 384 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); 385 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx); 386 return BinaryOperator::CreateAnd(C1, C2); 387 } 388 389 // If the comparison can be replaced with a range comparison for the elements 390 // where it is true, emit the range check. 391 if (TrueRangeEnd != Overdefined) { 392 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); 393 394 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). 395 if (FirstTrueElement) { 396 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement); 397 Idx = Builder.CreateAdd(Idx, Offs); 398 } 399 400 Value *End = ConstantInt::get(Idx->getType(), 401 TrueRangeEnd-FirstTrueElement+1); 402 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); 403 } 404 405 // False range check. 406 if (FalseRangeEnd != Overdefined) { 407 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); 408 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). 409 if (FirstFalseElement) { 410 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); 411 Idx = Builder.CreateAdd(Idx, Offs); 412 } 413 414 Value *End = ConstantInt::get(Idx->getType(), 415 FalseRangeEnd-FirstFalseElement); 416 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); 417 } 418 419 // If a magic bitvector captures the entire comparison state 420 // of this load, replace it with computation that does: 421 // ((magic_cst >> i) & 1) != 0 422 { 423 Type *Ty = nullptr; 424 425 // Look for an appropriate type: 426 // - The type of Idx if the magic fits 427 // - The smallest fitting legal type 428 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth()) 429 Ty = Idx->getType(); 430 else 431 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount); 432 433 if (Ty) { 434 Value *V = Builder.CreateIntCast(Idx, Ty, false); 435 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); 436 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V); 437 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); 438 } 439 } 440 441 return nullptr; 442 } 443 444 /// Return a value that can be used to compare the *offset* implied by a GEP to 445 /// zero. For example, if we have &A[i], we want to return 'i' for 446 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales 447 /// are involved. The above expression would also be legal to codegen as 448 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32). 449 /// This latter form is less amenable to optimization though, and we are allowed 450 /// to generate the first by knowing that pointer arithmetic doesn't overflow. 451 /// 452 /// If we can't emit an optimized form for this expression, this returns null. 453 /// 454 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC, 455 const DataLayout &DL) { 456 gep_type_iterator GTI = gep_type_begin(GEP); 457 458 // Check to see if this gep only has a single variable index. If so, and if 459 // any constant indices are a multiple of its scale, then we can compute this 460 // in terms of the scale of the variable index. For example, if the GEP 461 // implies an offset of "12 + i*4", then we can codegen this as "3 + i", 462 // because the expression will cross zero at the same point. 463 unsigned i, e = GEP->getNumOperands(); 464 int64_t Offset = 0; 465 for (i = 1; i != e; ++i, ++GTI) { 466 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 467 // Compute the aggregate offset of constant indices. 468 if (CI->isZero()) continue; 469 470 // Handle a struct index, which adds its field offset to the pointer. 471 if (StructType *STy = GTI.getStructTypeOrNull()) { 472 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 473 } else { 474 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); 475 Offset += Size*CI->getSExtValue(); 476 } 477 } else { 478 // Found our variable index. 479 break; 480 } 481 } 482 483 // If there are no variable indices, we must have a constant offset, just 484 // evaluate it the general way. 485 if (i == e) return nullptr; 486 487 Value *VariableIdx = GEP->getOperand(i); 488 // Determine the scale factor of the variable element. For example, this is 489 // 4 if the variable index is into an array of i32. 490 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType()); 491 492 // Verify that there are no other variable indices. If so, emit the hard way. 493 for (++i, ++GTI; i != e; ++i, ++GTI) { 494 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i)); 495 if (!CI) return nullptr; 496 497 // Compute the aggregate offset of constant indices. 498 if (CI->isZero()) continue; 499 500 // Handle a struct index, which adds its field offset to the pointer. 501 if (StructType *STy = GTI.getStructTypeOrNull()) { 502 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); 503 } else { 504 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); 505 Offset += Size*CI->getSExtValue(); 506 } 507 } 508 509 // Okay, we know we have a single variable index, which must be a 510 // pointer/array/vector index. If there is no offset, life is simple, return 511 // the index. 512 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType()); 513 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth(); 514 if (Offset == 0) { 515 // Cast to intptrty in case a truncation occurs. If an extension is needed, 516 // we don't need to bother extending: the extension won't affect where the 517 // computation crosses zero. 518 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { 519 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy); 520 } 521 return VariableIdx; 522 } 523 524 // Otherwise, there is an index. The computation we will do will be modulo 525 // the pointer size, so get it. 526 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); 527 528 Offset &= PtrSizeMask; 529 VariableScale &= PtrSizeMask; 530 531 // To do this transformation, any constant index must be a multiple of the 532 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", 533 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a 534 // multiple of the variable scale. 535 int64_t NewOffs = Offset / (int64_t)VariableScale; 536 if (Offset != NewOffs*(int64_t)VariableScale) 537 return nullptr; 538 539 // Okay, we can do this evaluation. Start by converting the index to intptr. 540 if (VariableIdx->getType() != IntPtrTy) 541 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy, 542 true /*Signed*/); 543 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); 544 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset"); 545 } 546 547 /// Returns true if we can rewrite Start as a GEP with pointer Base 548 /// and some integer offset. The nodes that need to be re-written 549 /// for this transformation will be added to Explored. 550 static bool canRewriteGEPAsOffset(Value *Start, Value *Base, 551 const DataLayout &DL, 552 SetVector<Value *> &Explored) { 553 SmallVector<Value *, 16> WorkList(1, Start); 554 Explored.insert(Base); 555 556 // The following traversal gives us an order which can be used 557 // when doing the final transformation. Since in the final 558 // transformation we create the PHI replacement instructions first, 559 // we don't have to get them in any particular order. 560 // 561 // However, for other instructions we will have to traverse the 562 // operands of an instruction first, which means that we have to 563 // do a post-order traversal. 564 while (!WorkList.empty()) { 565 SetVector<PHINode *> PHIs; 566 567 while (!WorkList.empty()) { 568 if (Explored.size() >= 100) 569 return false; 570 571 Value *V = WorkList.back(); 572 573 if (Explored.count(V) != 0) { 574 WorkList.pop_back(); 575 continue; 576 } 577 578 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) && 579 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V)) 580 // We've found some value that we can't explore which is different from 581 // the base. Therefore we can't do this transformation. 582 return false; 583 584 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) { 585 auto *CI = dyn_cast<CastInst>(V); 586 if (!CI->isNoopCast(DL)) 587 return false; 588 589 if (Explored.count(CI->getOperand(0)) == 0) 590 WorkList.push_back(CI->getOperand(0)); 591 } 592 593 if (auto *GEP = dyn_cast<GEPOperator>(V)) { 594 // We're limiting the GEP to having one index. This will preserve 595 // the original pointer type. We could handle more cases in the 596 // future. 597 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() || 598 GEP->getType() != Start->getType()) 599 return false; 600 601 if (Explored.count(GEP->getOperand(0)) == 0) 602 WorkList.push_back(GEP->getOperand(0)); 603 } 604 605 if (WorkList.back() == V) { 606 WorkList.pop_back(); 607 // We've finished visiting this node, mark it as such. 608 Explored.insert(V); 609 } 610 611 if (auto *PN = dyn_cast<PHINode>(V)) { 612 // We cannot transform PHIs on unsplittable basic blocks. 613 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator())) 614 return false; 615 Explored.insert(PN); 616 PHIs.insert(PN); 617 } 618 } 619 620 // Explore the PHI nodes further. 621 for (auto *PN : PHIs) 622 for (Value *Op : PN->incoming_values()) 623 if (Explored.count(Op) == 0) 624 WorkList.push_back(Op); 625 } 626 627 // Make sure that we can do this. Since we can't insert GEPs in a basic 628 // block before a PHI node, we can't easily do this transformation if 629 // we have PHI node users of transformed instructions. 630 for (Value *Val : Explored) { 631 for (Value *Use : Val->uses()) { 632 633 auto *PHI = dyn_cast<PHINode>(Use); 634 auto *Inst = dyn_cast<Instruction>(Val); 635 636 if (Inst == Base || Inst == PHI || !Inst || !PHI || 637 Explored.count(PHI) == 0) 638 continue; 639 640 if (PHI->getParent() == Inst->getParent()) 641 return false; 642 } 643 } 644 return true; 645 } 646 647 // Sets the appropriate insert point on Builder where we can add 648 // a replacement Instruction for V (if that is possible). 649 static void setInsertionPoint(IRBuilder<> &Builder, Value *V, 650 bool Before = true) { 651 if (auto *PHI = dyn_cast<PHINode>(V)) { 652 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt()); 653 return; 654 } 655 if (auto *I = dyn_cast<Instruction>(V)) { 656 if (!Before) 657 I = &*std::next(I->getIterator()); 658 Builder.SetInsertPoint(I); 659 return; 660 } 661 if (auto *A = dyn_cast<Argument>(V)) { 662 // Set the insertion point in the entry block. 663 BasicBlock &Entry = A->getParent()->getEntryBlock(); 664 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt()); 665 return; 666 } 667 // Otherwise, this is a constant and we don't need to set a new 668 // insertion point. 669 assert(isa<Constant>(V) && "Setting insertion point for unknown value!"); 670 } 671 672 /// Returns a re-written value of Start as an indexed GEP using Base as a 673 /// pointer. 674 static Value *rewriteGEPAsOffset(Value *Start, Value *Base, 675 const DataLayout &DL, 676 SetVector<Value *> &Explored) { 677 // Perform all the substitutions. This is a bit tricky because we can 678 // have cycles in our use-def chains. 679 // 1. Create the PHI nodes without any incoming values. 680 // 2. Create all the other values. 681 // 3. Add the edges for the PHI nodes. 682 // 4. Emit GEPs to get the original pointers. 683 // 5. Remove the original instructions. 684 Type *IndexType = IntegerType::get( 685 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType())); 686 687 DenseMap<Value *, Value *> NewInsts; 688 NewInsts[Base] = ConstantInt::getNullValue(IndexType); 689 690 // Create the new PHI nodes, without adding any incoming values. 691 for (Value *Val : Explored) { 692 if (Val == Base) 693 continue; 694 // Create empty phi nodes. This avoids cyclic dependencies when creating 695 // the remaining instructions. 696 if (auto *PHI = dyn_cast<PHINode>(Val)) 697 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(), 698 PHI->getName() + ".idx", PHI); 699 } 700 IRBuilder<> Builder(Base->getContext()); 701 702 // Create all the other instructions. 703 for (Value *Val : Explored) { 704 705 if (NewInsts.find(Val) != NewInsts.end()) 706 continue; 707 708 if (auto *CI = dyn_cast<CastInst>(Val)) { 709 NewInsts[CI] = NewInsts[CI->getOperand(0)]; 710 continue; 711 } 712 if (auto *GEP = dyn_cast<GEPOperator>(Val)) { 713 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)] 714 : GEP->getOperand(1); 715 setInsertionPoint(Builder, GEP); 716 // Indices might need to be sign extended. GEPs will magically do 717 // this, but we need to do it ourselves here. 718 if (Index->getType()->getScalarSizeInBits() != 719 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) { 720 Index = Builder.CreateSExtOrTrunc( 721 Index, NewInsts[GEP->getOperand(0)]->getType(), 722 GEP->getOperand(0)->getName() + ".sext"); 723 } 724 725 auto *Op = NewInsts[GEP->getOperand(0)]; 726 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero()) 727 NewInsts[GEP] = Index; 728 else 729 NewInsts[GEP] = Builder.CreateNSWAdd( 730 Op, Index, GEP->getOperand(0)->getName() + ".add"); 731 continue; 732 } 733 if (isa<PHINode>(Val)) 734 continue; 735 736 llvm_unreachable("Unexpected instruction type"); 737 } 738 739 // Add the incoming values to the PHI nodes. 740 for (Value *Val : Explored) { 741 if (Val == Base) 742 continue; 743 // All the instructions have been created, we can now add edges to the 744 // phi nodes. 745 if (auto *PHI = dyn_cast<PHINode>(Val)) { 746 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]); 747 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) { 748 Value *NewIncoming = PHI->getIncomingValue(I); 749 750 if (NewInsts.find(NewIncoming) != NewInsts.end()) 751 NewIncoming = NewInsts[NewIncoming]; 752 753 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I)); 754 } 755 } 756 } 757 758 for (Value *Val : Explored) { 759 if (Val == Base) 760 continue; 761 762 // Depending on the type, for external users we have to emit 763 // a GEP or a GEP + ptrtoint. 764 setInsertionPoint(Builder, Val, false); 765 766 // If required, create an inttoptr instruction for Base. 767 Value *NewBase = Base; 768 if (!Base->getType()->isPointerTy()) 769 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(), 770 Start->getName() + "to.ptr"); 771 772 Value *GEP = Builder.CreateInBoundsGEP( 773 Start->getType()->getPointerElementType(), NewBase, 774 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr"); 775 776 if (!Val->getType()->isPointerTy()) { 777 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(), 778 Val->getName() + ".conv"); 779 GEP = Cast; 780 } 781 Val->replaceAllUsesWith(GEP); 782 } 783 784 return NewInsts[Start]; 785 } 786 787 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express 788 /// the input Value as a constant indexed GEP. Returns a pair containing 789 /// the GEPs Pointer and Index. 790 static std::pair<Value *, Value *> 791 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) { 792 Type *IndexType = IntegerType::get(V->getContext(), 793 DL.getIndexTypeSizeInBits(V->getType())); 794 795 Constant *Index = ConstantInt::getNullValue(IndexType); 796 while (true) { 797 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 798 // We accept only inbouds GEPs here to exclude the possibility of 799 // overflow. 800 if (!GEP->isInBounds()) 801 break; 802 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 && 803 GEP->getType() == V->getType()) { 804 V = GEP->getOperand(0); 805 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1)); 806 Index = ConstantExpr::getAdd( 807 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType)); 808 continue; 809 } 810 break; 811 } 812 if (auto *CI = dyn_cast<IntToPtrInst>(V)) { 813 if (!CI->isNoopCast(DL)) 814 break; 815 V = CI->getOperand(0); 816 continue; 817 } 818 if (auto *CI = dyn_cast<PtrToIntInst>(V)) { 819 if (!CI->isNoopCast(DL)) 820 break; 821 V = CI->getOperand(0); 822 continue; 823 } 824 break; 825 } 826 return {V, Index}; 827 } 828 829 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant. 830 /// We can look through PHIs, GEPs and casts in order to determine a common base 831 /// between GEPLHS and RHS. 832 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS, 833 ICmpInst::Predicate Cond, 834 const DataLayout &DL) { 835 if (!GEPLHS->hasAllConstantIndices()) 836 return nullptr; 837 838 // Make sure the pointers have the same type. 839 if (GEPLHS->getType() != RHS->getType()) 840 return nullptr; 841 842 Value *PtrBase, *Index; 843 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL); 844 845 // The set of nodes that will take part in this transformation. 846 SetVector<Value *> Nodes; 847 848 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes)) 849 return nullptr; 850 851 // We know we can re-write this as 852 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) 853 // Since we've only looked through inbouds GEPs we know that we 854 // can't have overflow on either side. We can therefore re-write 855 // this as: 856 // OFFSET1 cmp OFFSET2 857 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes); 858 859 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written 860 // GEP having PtrBase as the pointer base, and has returned in NewRHS the 861 // offset. Since Index is the offset of LHS to the base pointer, we will now 862 // compare the offsets instead of comparing the pointers. 863 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS); 864 } 865 866 /// Fold comparisons between a GEP instruction and something else. At this point 867 /// we know that the GEP is on the LHS of the comparison. 868 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 869 ICmpInst::Predicate Cond, 870 Instruction &I) { 871 // Don't transform signed compares of GEPs into index compares. Even if the 872 // GEP is inbounds, the final add of the base pointer can have signed overflow 873 // and would change the result of the icmp. 874 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be 875 // the maximum signed value for the pointer type. 876 if (ICmpInst::isSigned(Cond)) 877 return nullptr; 878 879 // Look through bitcasts and addrspacecasts. We do not however want to remove 880 // 0 GEPs. 881 if (!isa<GetElementPtrInst>(RHS)) 882 RHS = RHS->stripPointerCasts(); 883 884 Value *PtrBase = GEPLHS->getOperand(0); 885 if (PtrBase == RHS && GEPLHS->isInBounds()) { 886 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). 887 // This transformation (ignoring the base and scales) is valid because we 888 // know pointers can't overflow since the gep is inbounds. See if we can 889 // output an optimized form. 890 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL); 891 892 // If not, synthesize the offset the hard way. 893 if (!Offset) 894 Offset = EmitGEPOffset(GEPLHS); 895 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, 896 Constant::getNullValue(Offset->getType())); 897 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) { 898 // If the base pointers are different, but the indices are the same, just 899 // compare the base pointer. 900 if (PtrBase != GEPRHS->getOperand(0)) { 901 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); 902 IndicesTheSame &= GEPLHS->getOperand(0)->getType() == 903 GEPRHS->getOperand(0)->getType(); 904 if (IndicesTheSame) 905 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) 906 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 907 IndicesTheSame = false; 908 break; 909 } 910 911 // If all indices are the same, just compare the base pointers. 912 if (IndicesTheSame) 913 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0)); 914 915 // If we're comparing GEPs with two base pointers that only differ in type 916 // and both GEPs have only constant indices or just one use, then fold 917 // the compare with the adjusted indices. 918 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() && 919 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && 920 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && 921 PtrBase->stripPointerCasts() == 922 GEPRHS->getOperand(0)->stripPointerCasts()) { 923 Value *LOffset = EmitGEPOffset(GEPLHS); 924 Value *ROffset = EmitGEPOffset(GEPRHS); 925 926 // If we looked through an addrspacecast between different sized address 927 // spaces, the LHS and RHS pointers are different sized 928 // integers. Truncate to the smaller one. 929 Type *LHSIndexTy = LOffset->getType(); 930 Type *RHSIndexTy = ROffset->getType(); 931 if (LHSIndexTy != RHSIndexTy) { 932 if (LHSIndexTy->getPrimitiveSizeInBits() < 933 RHSIndexTy->getPrimitiveSizeInBits()) { 934 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy); 935 } else 936 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy); 937 } 938 939 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond), 940 LOffset, ROffset); 941 return replaceInstUsesWith(I, Cmp); 942 } 943 944 // Otherwise, the base pointers are different and the indices are 945 // different. Try convert this to an indexed compare by looking through 946 // PHIs/casts. 947 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL); 948 } 949 950 // If one of the GEPs has all zero indices, recurse. 951 if (GEPLHS->hasAllZeroIndices()) 952 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0), 953 ICmpInst::getSwappedPredicate(Cond), I); 954 955 // If the other GEP has all zero indices, recurse. 956 if (GEPRHS->hasAllZeroIndices()) 957 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); 958 959 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); 960 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { 961 // If the GEPs only differ by one index, compare it. 962 unsigned NumDifferences = 0; // Keep track of # differences. 963 unsigned DiffOperand = 0; // The operand that differs. 964 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) 965 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { 966 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != 967 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { 968 // Irreconcilable differences. 969 NumDifferences = 2; 970 break; 971 } else { 972 if (NumDifferences++) break; 973 DiffOperand = i; 974 } 975 } 976 977 if (NumDifferences == 0) // SAME GEP? 978 return replaceInstUsesWith(I, // No comparison is needed here. 979 Builder.getInt1(ICmpInst::isTrueWhenEqual(Cond))); 980 981 else if (NumDifferences == 1 && GEPsInBounds) { 982 Value *LHSV = GEPLHS->getOperand(DiffOperand); 983 Value *RHSV = GEPRHS->getOperand(DiffOperand); 984 // Make sure we do a signed comparison here. 985 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); 986 } 987 } 988 989 // Only lower this if the icmp is the only user of the GEP or if we expect 990 // the result to fold to a constant! 991 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) && 992 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) { 993 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) 994 Value *L = EmitGEPOffset(GEPLHS); 995 Value *R = EmitGEPOffset(GEPRHS); 996 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); 997 } 998 } 999 1000 // Try convert this to an indexed compare by looking through PHIs/casts as a 1001 // last resort. 1002 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL); 1003 } 1004 1005 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI, 1006 const AllocaInst *Alloca, 1007 const Value *Other) { 1008 assert(ICI.isEquality() && "Cannot fold non-equality comparison."); 1009 1010 // It would be tempting to fold away comparisons between allocas and any 1011 // pointer not based on that alloca (e.g. an argument). However, even 1012 // though such pointers cannot alias, they can still compare equal. 1013 // 1014 // But LLVM doesn't specify where allocas get their memory, so if the alloca 1015 // doesn't escape we can argue that it's impossible to guess its value, and we 1016 // can therefore act as if any such guesses are wrong. 1017 // 1018 // The code below checks that the alloca doesn't escape, and that it's only 1019 // used in a comparison once (the current instruction). The 1020 // single-comparison-use condition ensures that we're trivially folding all 1021 // comparisons against the alloca consistently, and avoids the risk of 1022 // erroneously folding a comparison of the pointer with itself. 1023 1024 unsigned MaxIter = 32; // Break cycles and bound to constant-time. 1025 1026 SmallVector<const Use *, 32> Worklist; 1027 for (const Use &U : Alloca->uses()) { 1028 if (Worklist.size() >= MaxIter) 1029 return nullptr; 1030 Worklist.push_back(&U); 1031 } 1032 1033 unsigned NumCmps = 0; 1034 while (!Worklist.empty()) { 1035 assert(Worklist.size() <= MaxIter); 1036 const Use *U = Worklist.pop_back_val(); 1037 const Value *V = U->getUser(); 1038 --MaxIter; 1039 1040 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) || 1041 isa<SelectInst>(V)) { 1042 // Track the uses. 1043 } else if (isa<LoadInst>(V)) { 1044 // Loading from the pointer doesn't escape it. 1045 continue; 1046 } else if (const auto *SI = dyn_cast<StoreInst>(V)) { 1047 // Storing *to* the pointer is fine, but storing the pointer escapes it. 1048 if (SI->getValueOperand() == U->get()) 1049 return nullptr; 1050 continue; 1051 } else if (isa<ICmpInst>(V)) { 1052 if (NumCmps++) 1053 return nullptr; // Found more than one cmp. 1054 continue; 1055 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) { 1056 switch (Intrin->getIntrinsicID()) { 1057 // These intrinsics don't escape or compare the pointer. Memset is safe 1058 // because we don't allow ptrtoint. Memcpy and memmove are safe because 1059 // we don't allow stores, so src cannot point to V. 1060 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: 1061 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset: 1062 continue; 1063 default: 1064 return nullptr; 1065 } 1066 } else { 1067 return nullptr; 1068 } 1069 for (const Use &U : V->uses()) { 1070 if (Worklist.size() >= MaxIter) 1071 return nullptr; 1072 Worklist.push_back(&U); 1073 } 1074 } 1075 1076 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType()); 1077 return replaceInstUsesWith( 1078 ICI, 1079 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate()))); 1080 } 1081 1082 /// Fold "icmp pred (X+CI), X". 1083 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, ConstantInt *CI, 1084 ICmpInst::Predicate Pred) { 1085 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, 1086 // so the values can never be equal. Similarly for all other "or equals" 1087 // operators. 1088 1089 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 1090 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 1091 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 1092 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 1093 Value *R = 1094 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); 1095 return new ICmpInst(ICmpInst::ICMP_UGT, X, R); 1096 } 1097 1098 // (X+1) >u X --> X <u (0-1) --> X != 255 1099 // (X+2) >u X --> X <u (0-2) --> X <u 254 1100 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 1101 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 1102 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); 1103 1104 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); 1105 ConstantInt *SMax = ConstantInt::get(X->getContext(), 1106 APInt::getSignedMaxValue(BitWidth)); 1107 1108 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 1109 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 1110 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 1111 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 1112 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 1113 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 1114 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 1115 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); 1116 1117 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 1118 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 1119 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 1120 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 1121 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 1122 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 1123 1124 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); 1125 Constant *C = Builder.getInt(CI->getValue() - 1); 1126 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); 1127 } 1128 1129 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" -> 1130 /// (icmp eq/ne A, Log2(AP2/AP1)) -> 1131 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)). 1132 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A, 1133 const APInt &AP1, 1134 const APInt &AP2) { 1135 assert(I.isEquality() && "Cannot fold icmp gt/lt"); 1136 1137 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) { 1138 if (I.getPredicate() == I.ICMP_NE) 1139 Pred = CmpInst::getInversePredicate(Pred); 1140 return new ICmpInst(Pred, LHS, RHS); 1141 }; 1142 1143 // Don't bother doing any work for cases which InstSimplify handles. 1144 if (AP2.isNullValue()) 1145 return nullptr; 1146 1147 bool IsAShr = isa<AShrOperator>(I.getOperand(0)); 1148 if (IsAShr) { 1149 if (AP2.isAllOnesValue()) 1150 return nullptr; 1151 if (AP2.isNegative() != AP1.isNegative()) 1152 return nullptr; 1153 if (AP2.sgt(AP1)) 1154 return nullptr; 1155 } 1156 1157 if (!AP1) 1158 // 'A' must be large enough to shift out the highest set bit. 1159 return getICmp(I.ICMP_UGT, A, 1160 ConstantInt::get(A->getType(), AP2.logBase2())); 1161 1162 if (AP1 == AP2) 1163 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType())); 1164 1165 int Shift; 1166 if (IsAShr && AP1.isNegative()) 1167 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes(); 1168 else 1169 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros(); 1170 1171 if (Shift > 0) { 1172 if (IsAShr && AP1 == AP2.ashr(Shift)) { 1173 // There are multiple solutions if we are comparing against -1 and the LHS 1174 // of the ashr is not a power of two. 1175 if (AP1.isAllOnesValue() && !AP2.isPowerOf2()) 1176 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift)); 1177 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift)); 1178 } else if (AP1 == AP2.lshr(Shift)) { 1179 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift)); 1180 } 1181 } 1182 1183 // Shifting const2 will never be equal to const1. 1184 // FIXME: This should always be handled by InstSimplify? 1185 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE); 1186 return replaceInstUsesWith(I, TorF); 1187 } 1188 1189 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" -> 1190 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)). 1191 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A, 1192 const APInt &AP1, 1193 const APInt &AP2) { 1194 assert(I.isEquality() && "Cannot fold icmp gt/lt"); 1195 1196 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) { 1197 if (I.getPredicate() == I.ICMP_NE) 1198 Pred = CmpInst::getInversePredicate(Pred); 1199 return new ICmpInst(Pred, LHS, RHS); 1200 }; 1201 1202 // Don't bother doing any work for cases which InstSimplify handles. 1203 if (AP2.isNullValue()) 1204 return nullptr; 1205 1206 unsigned AP2TrailingZeros = AP2.countTrailingZeros(); 1207 1208 if (!AP1 && AP2TrailingZeros != 0) 1209 return getICmp( 1210 I.ICMP_UGE, A, 1211 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros)); 1212 1213 if (AP1 == AP2) 1214 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType())); 1215 1216 // Get the distance between the lowest bits that are set. 1217 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros; 1218 1219 if (Shift > 0 && AP2.shl(Shift) == AP1) 1220 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift)); 1221 1222 // Shifting const2 will never be equal to const1. 1223 // FIXME: This should always be handled by InstSimplify? 1224 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE); 1225 return replaceInstUsesWith(I, TorF); 1226 } 1227 1228 /// The caller has matched a pattern of the form: 1229 /// I = icmp ugt (add (add A, B), CI2), CI1 1230 /// If this is of the form: 1231 /// sum = a + b 1232 /// if (sum+128 >u 255) 1233 /// Then replace it with llvm.sadd.with.overflow.i8. 1234 /// 1235 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, 1236 ConstantInt *CI2, ConstantInt *CI1, 1237 InstCombiner &IC) { 1238 // The transformation we're trying to do here is to transform this into an 1239 // llvm.sadd.with.overflow. To do this, we have to replace the original add 1240 // with a narrower add, and discard the add-with-constant that is part of the 1241 // range check (if we can't eliminate it, this isn't profitable). 1242 1243 // In order to eliminate the add-with-constant, the compare can be its only 1244 // use. 1245 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0)); 1246 if (!AddWithCst->hasOneUse()) 1247 return nullptr; 1248 1249 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. 1250 if (!CI2->getValue().isPowerOf2()) 1251 return nullptr; 1252 unsigned NewWidth = CI2->getValue().countTrailingZeros(); 1253 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) 1254 return nullptr; 1255 1256 // The width of the new add formed is 1 more than the bias. 1257 ++NewWidth; 1258 1259 // Check to see that CI1 is an all-ones value with NewWidth bits. 1260 if (CI1->getBitWidth() == NewWidth || 1261 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) 1262 return nullptr; 1263 1264 // This is only really a signed overflow check if the inputs have been 1265 // sign-extended; check for that condition. For example, if CI2 is 2^31 and 1266 // the operands of the add are 64 bits wide, we need at least 33 sign bits. 1267 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; 1268 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits || 1269 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits) 1270 return nullptr; 1271 1272 // In order to replace the original add with a narrower 1273 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant 1274 // and truncates that discard the high bits of the add. Verify that this is 1275 // the case. 1276 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0)); 1277 for (User *U : OrigAdd->users()) { 1278 if (U == AddWithCst) 1279 continue; 1280 1281 // Only accept truncates for now. We would really like a nice recursive 1282 // predicate like SimplifyDemandedBits, but which goes downwards the use-def 1283 // chain to see which bits of a value are actually demanded. If the 1284 // original add had another add which was then immediately truncated, we 1285 // could still do the transformation. 1286 TruncInst *TI = dyn_cast<TruncInst>(U); 1287 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth) 1288 return nullptr; 1289 } 1290 1291 // If the pattern matches, truncate the inputs to the narrower type and 1292 // use the sadd_with_overflow intrinsic to efficiently compute both the 1293 // result and the overflow bit. 1294 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); 1295 Value *F = Intrinsic::getDeclaration(I.getModule(), 1296 Intrinsic::sadd_with_overflow, NewType); 1297 1298 InstCombiner::BuilderTy &Builder = IC.Builder; 1299 1300 // Put the new code above the original add, in case there are any uses of the 1301 // add between the add and the compare. 1302 Builder.SetInsertPoint(OrigAdd); 1303 1304 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc"); 1305 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc"); 1306 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd"); 1307 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result"); 1308 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType()); 1309 1310 // The inner add was the result of the narrow add, zero extended to the 1311 // wider type. Replace it with the result computed by the intrinsic. 1312 IC.replaceInstUsesWith(*OrigAdd, ZExt); 1313 1314 // The original icmp gets replaced with the overflow value. 1315 return ExtractValueInst::Create(Call, 1, "sadd.overflow"); 1316 } 1317 1318 // Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0) 1319 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) { 1320 CmpInst::Predicate Pred = Cmp.getPredicate(); 1321 Value *X = Cmp.getOperand(0); 1322 1323 if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) { 1324 Value *A, *B; 1325 SelectPatternResult SPR = matchSelectPattern(X, A, B); 1326 if (SPR.Flavor == SPF_SMIN) { 1327 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT)) 1328 return new ICmpInst(Pred, B, Cmp.getOperand(1)); 1329 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT)) 1330 return new ICmpInst(Pred, A, Cmp.getOperand(1)); 1331 } 1332 } 1333 return nullptr; 1334 } 1335 1336 // Fold icmp Pred X, C. 1337 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) { 1338 CmpInst::Predicate Pred = Cmp.getPredicate(); 1339 Value *X = Cmp.getOperand(0); 1340 1341 const APInt *C; 1342 if (!match(Cmp.getOperand(1), m_APInt(C))) 1343 return nullptr; 1344 1345 Value *A = nullptr, *B = nullptr; 1346 1347 // Match the following pattern, which is a common idiom when writing 1348 // overflow-safe integer arithmetic functions. The source performs an addition 1349 // in wider type and explicitly checks for overflow using comparisons against 1350 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic. 1351 // 1352 // TODO: This could probably be generalized to handle other overflow-safe 1353 // operations if we worked out the formulas to compute the appropriate magic 1354 // constants. 1355 // 1356 // sum = a + b 1357 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 1358 { 1359 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI 1360 if (Pred == ICmpInst::ICMP_UGT && 1361 match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) 1362 if (Instruction *Res = processUGT_ADDCST_ADD( 1363 Cmp, A, B, CI2, cast<ConstantInt>(Cmp.getOperand(1)), *this)) 1364 return Res; 1365 } 1366 1367 // FIXME: Use m_APInt to allow folds for splat constants. 1368 ConstantInt *CI = dyn_cast<ConstantInt>(Cmp.getOperand(1)); 1369 if (!CI) 1370 return nullptr; 1371 1372 // Canonicalize icmp instructions based on dominating conditions. 1373 BasicBlock *Parent = Cmp.getParent(); 1374 BasicBlock *Dom = Parent->getSinglePredecessor(); 1375 auto *BI = Dom ? dyn_cast<BranchInst>(Dom->getTerminator()) : nullptr; 1376 ICmpInst::Predicate Pred2; 1377 BasicBlock *TrueBB, *FalseBB; 1378 ConstantInt *CI2; 1379 if (BI && match(BI, m_Br(m_ICmp(Pred2, m_Specific(X), m_ConstantInt(CI2)), 1380 TrueBB, FalseBB)) && 1381 TrueBB != FalseBB) { 1382 ConstantRange CR = 1383 ConstantRange::makeAllowedICmpRegion(Pred, CI->getValue()); 1384 ConstantRange DominatingCR = 1385 (Parent == TrueBB) 1386 ? ConstantRange::makeExactICmpRegion(Pred2, CI2->getValue()) 1387 : ConstantRange::makeExactICmpRegion( 1388 CmpInst::getInversePredicate(Pred2), CI2->getValue()); 1389 ConstantRange Intersection = DominatingCR.intersectWith(CR); 1390 ConstantRange Difference = DominatingCR.difference(CR); 1391 if (Intersection.isEmptySet()) 1392 return replaceInstUsesWith(Cmp, Builder.getFalse()); 1393 if (Difference.isEmptySet()) 1394 return replaceInstUsesWith(Cmp, Builder.getTrue()); 1395 1396 // If this is a normal comparison, it demands all bits. If it is a sign 1397 // bit comparison, it only demands the sign bit. 1398 bool UnusedBit; 1399 bool IsSignBit = isSignBitCheck(Pred, CI->getValue(), UnusedBit); 1400 1401 // Canonicalizing a sign bit comparison that gets used in a branch, 1402 // pessimizes codegen by generating branch on zero instruction instead 1403 // of a test and branch. So we avoid canonicalizing in such situations 1404 // because test and branch instruction has better branch displacement 1405 // than compare and branch instruction. 1406 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp))) 1407 return nullptr; 1408 1409 if (auto *AI = Intersection.getSingleElement()) 1410 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*AI)); 1411 if (auto *AD = Difference.getSingleElement()) 1412 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*AD)); 1413 } 1414 1415 return nullptr; 1416 } 1417 1418 /// Fold icmp (trunc X, Y), C. 1419 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp, 1420 TruncInst *Trunc, 1421 const APInt &C) { 1422 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1423 Value *X = Trunc->getOperand(0); 1424 if (C.isOneValue() && C.getBitWidth() > 1) { 1425 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1 1426 Value *V = nullptr; 1427 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V)))) 1428 return new ICmpInst(ICmpInst::ICMP_SLT, V, 1429 ConstantInt::get(V->getType(), 1)); 1430 } 1431 1432 if (Cmp.isEquality() && Trunc->hasOneUse()) { 1433 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all 1434 // of the high bits truncated out of x are known. 1435 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(), 1436 SrcBits = X->getType()->getScalarSizeInBits(); 1437 KnownBits Known = computeKnownBits(X, 0, &Cmp); 1438 1439 // If all the high bits are known, we can do this xform. 1440 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) { 1441 // Pull in the high bits from known-ones set. 1442 APInt NewRHS = C.zext(SrcBits); 1443 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits); 1444 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS)); 1445 } 1446 } 1447 1448 return nullptr; 1449 } 1450 1451 /// Fold icmp (xor X, Y), C. 1452 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp, 1453 BinaryOperator *Xor, 1454 const APInt &C) { 1455 Value *X = Xor->getOperand(0); 1456 Value *Y = Xor->getOperand(1); 1457 const APInt *XorC; 1458 if (!match(Y, m_APInt(XorC))) 1459 return nullptr; 1460 1461 // If this is a comparison that tests the signbit (X < 0) or (x > -1), 1462 // fold the xor. 1463 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1464 bool TrueIfSigned = false; 1465 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) { 1466 1467 // If the sign bit of the XorCst is not set, there is no change to 1468 // the operation, just stop using the Xor. 1469 if (!XorC->isNegative()) { 1470 Cmp.setOperand(0, X); 1471 Worklist.Add(Xor); 1472 return &Cmp; 1473 } 1474 1475 // Emit the opposite comparison. 1476 if (TrueIfSigned) 1477 return new ICmpInst(ICmpInst::ICMP_SGT, X, 1478 ConstantInt::getAllOnesValue(X->getType())); 1479 else 1480 return new ICmpInst(ICmpInst::ICMP_SLT, X, 1481 ConstantInt::getNullValue(X->getType())); 1482 } 1483 1484 if (Xor->hasOneUse()) { 1485 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask)) 1486 if (!Cmp.isEquality() && XorC->isSignMask()) { 1487 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate() 1488 : Cmp.getSignedPredicate(); 1489 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC)); 1490 } 1491 1492 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask)) 1493 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) { 1494 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate() 1495 : Cmp.getSignedPredicate(); 1496 Pred = Cmp.getSwappedPredicate(Pred); 1497 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC)); 1498 } 1499 } 1500 1501 // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C) 1502 // iff -C is a power of 2 1503 if (Pred == ICmpInst::ICMP_UGT && *XorC == ~C && (C + 1).isPowerOf2()) 1504 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y); 1505 1506 // (icmp ult (xor X, C), -C) -> (icmp uge X, C) 1507 // iff -C is a power of 2 1508 if (Pred == ICmpInst::ICMP_ULT && *XorC == -C && C.isPowerOf2()) 1509 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y); 1510 1511 return nullptr; 1512 } 1513 1514 /// Fold icmp (and (sh X, Y), C2), C1. 1515 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And, 1516 const APInt &C1, const APInt &C2) { 1517 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0)); 1518 if (!Shift || !Shift->isShift()) 1519 return nullptr; 1520 1521 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could 1522 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in 1523 // code produced by the clang front-end, for bitfield access. 1524 // This seemingly simple opportunity to fold away a shift turns out to be 1525 // rather complicated. See PR17827 for details. 1526 unsigned ShiftOpcode = Shift->getOpcode(); 1527 bool IsShl = ShiftOpcode == Instruction::Shl; 1528 const APInt *C3; 1529 if (match(Shift->getOperand(1), m_APInt(C3))) { 1530 bool CanFold = false; 1531 if (ShiftOpcode == Instruction::Shl) { 1532 // For a left shift, we can fold if the comparison is not signed. We can 1533 // also fold a signed comparison if the mask value and comparison value 1534 // are not negative. These constraints may not be obvious, but we can 1535 // prove that they are correct using an SMT solver. 1536 if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative())) 1537 CanFold = true; 1538 } else { 1539 bool IsAshr = ShiftOpcode == Instruction::AShr; 1540 // For a logical right shift, we can fold if the comparison is not signed. 1541 // We can also fold a signed comparison if the shifted mask value and the 1542 // shifted comparison value are not negative. These constraints may not be 1543 // obvious, but we can prove that they are correct using an SMT solver. 1544 // For an arithmetic shift right we can do the same, if we ensure 1545 // the And doesn't use any bits being shifted in. Normally these would 1546 // be turned into lshr by SimplifyDemandedBits, but not if there is an 1547 // additional user. 1548 if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) { 1549 if (!Cmp.isSigned() || 1550 (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative())) 1551 CanFold = true; 1552 } 1553 } 1554 1555 if (CanFold) { 1556 APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3); 1557 APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3); 1558 // Check to see if we are shifting out any of the bits being compared. 1559 if (SameAsC1 != C1) { 1560 // If we shifted bits out, the fold is not going to work out. As a 1561 // special case, check to see if this means that the result is always 1562 // true or false now. 1563 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ) 1564 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType())); 1565 if (Cmp.getPredicate() == ICmpInst::ICMP_NE) 1566 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType())); 1567 } else { 1568 Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst)); 1569 APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3); 1570 And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst)); 1571 And->setOperand(0, Shift->getOperand(0)); 1572 Worklist.Add(Shift); // Shift is dead. 1573 return &Cmp; 1574 } 1575 } 1576 } 1577 1578 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is 1579 // preferable because it allows the C2 << Y expression to be hoisted out of a 1580 // loop if Y is invariant and X is not. 1581 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() && 1582 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) { 1583 // Compute C2 << Y. 1584 Value *NewShift = 1585 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1)) 1586 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1)); 1587 1588 // Compute X & (C2 << Y). 1589 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift); 1590 Cmp.setOperand(0, NewAnd); 1591 return &Cmp; 1592 } 1593 1594 return nullptr; 1595 } 1596 1597 /// Fold icmp (and X, C2), C1. 1598 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp, 1599 BinaryOperator *And, 1600 const APInt &C1) { 1601 const APInt *C2; 1602 if (!match(And->getOperand(1), m_APInt(C2))) 1603 return nullptr; 1604 1605 if (!And->hasOneUse()) 1606 return nullptr; 1607 1608 // If the LHS is an 'and' of a truncate and we can widen the and/compare to 1609 // the input width without changing the value produced, eliminate the cast: 1610 // 1611 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1' 1612 // 1613 // We can do this transformation if the constants do not have their sign bits 1614 // set or if it is an equality comparison. Extending a relational comparison 1615 // when we're checking the sign bit would not work. 1616 Value *W; 1617 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) && 1618 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) { 1619 // TODO: Is this a good transform for vectors? Wider types may reduce 1620 // throughput. Should this transform be limited (even for scalars) by using 1621 // shouldChangeType()? 1622 if (!Cmp.getType()->isVectorTy()) { 1623 Type *WideType = W->getType(); 1624 unsigned WideScalarBits = WideType->getScalarSizeInBits(); 1625 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits)); 1626 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits)); 1627 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName()); 1628 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1); 1629 } 1630 } 1631 1632 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2)) 1633 return I; 1634 1635 // (icmp pred (and (or (lshr A, B), A), 1), 0) --> 1636 // (icmp pred (and A, (or (shl 1, B), 1), 0)) 1637 // 1638 // iff pred isn't signed 1639 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() && 1640 match(And->getOperand(1), m_One())) { 1641 Constant *One = cast<Constant>(And->getOperand(1)); 1642 Value *Or = And->getOperand(0); 1643 Value *A, *B, *LShr; 1644 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) && 1645 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) { 1646 unsigned UsesRemoved = 0; 1647 if (And->hasOneUse()) 1648 ++UsesRemoved; 1649 if (Or->hasOneUse()) 1650 ++UsesRemoved; 1651 if (LShr->hasOneUse()) 1652 ++UsesRemoved; 1653 1654 // Compute A & ((1 << B) | 1) 1655 Value *NewOr = nullptr; 1656 if (auto *C = dyn_cast<Constant>(B)) { 1657 if (UsesRemoved >= 1) 1658 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One); 1659 } else { 1660 if (UsesRemoved >= 3) 1661 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(), 1662 /*HasNUW=*/true), 1663 One, Or->getName()); 1664 } 1665 if (NewOr) { 1666 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName()); 1667 Cmp.setOperand(0, NewAnd); 1668 return &Cmp; 1669 } 1670 } 1671 } 1672 1673 return nullptr; 1674 } 1675 1676 /// Fold icmp (and X, Y), C. 1677 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp, 1678 BinaryOperator *And, 1679 const APInt &C) { 1680 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C)) 1681 return I; 1682 1683 // TODO: These all require that Y is constant too, so refactor with the above. 1684 1685 // Try to optimize things like "A[i] & 42 == 0" to index computations. 1686 Value *X = And->getOperand(0); 1687 Value *Y = And->getOperand(1); 1688 if (auto *LI = dyn_cast<LoadInst>(X)) 1689 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) 1690 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 1691 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 1692 !LI->isVolatile() && isa<ConstantInt>(Y)) { 1693 ConstantInt *C2 = cast<ConstantInt>(Y); 1694 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2)) 1695 return Res; 1696 } 1697 1698 if (!Cmp.isEquality()) 1699 return nullptr; 1700 1701 // X & -C == -C -> X > u ~C 1702 // X & -C != -C -> X <= u ~C 1703 // iff C is a power of 2 1704 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) { 1705 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT 1706 : CmpInst::ICMP_ULE; 1707 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1)))); 1708 } 1709 1710 // (X & C2) == 0 -> (trunc X) >= 0 1711 // (X & C2) != 0 -> (trunc X) < 0 1712 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type. 1713 const APInt *C2; 1714 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) { 1715 int32_t ExactLogBase2 = C2->exactLogBase2(); 1716 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) { 1717 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1); 1718 if (And->getType()->isVectorTy()) 1719 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements()); 1720 Value *Trunc = Builder.CreateTrunc(X, NTy); 1721 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE 1722 : CmpInst::ICMP_SLT; 1723 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy)); 1724 } 1725 } 1726 1727 return nullptr; 1728 } 1729 1730 /// Fold icmp (or X, Y), C. 1731 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or, 1732 const APInt &C) { 1733 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1734 if (C.isOneValue()) { 1735 // icmp slt signum(V) 1 --> icmp slt V, 1 1736 Value *V = nullptr; 1737 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V)))) 1738 return new ICmpInst(ICmpInst::ICMP_SLT, V, 1739 ConstantInt::get(V->getType(), 1)); 1740 } 1741 1742 // X | C == C --> X <=u C 1743 // X | C != C --> X >u C 1744 // iff C+1 is a power of 2 (C is a bitmask of the low bits) 1745 if (Cmp.isEquality() && Cmp.getOperand(1) == Or->getOperand(1) && 1746 (C + 1).isPowerOf2()) { 1747 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT; 1748 return new ICmpInst(Pred, Or->getOperand(0), Or->getOperand(1)); 1749 } 1750 1751 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse()) 1752 return nullptr; 1753 1754 Value *P, *Q; 1755 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { 1756 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 1757 // -> and (icmp eq P, null), (icmp eq Q, null). 1758 Value *CmpP = 1759 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType())); 1760 Value *CmpQ = 1761 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType())); 1762 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 1763 return BinaryOperator::Create(BOpc, CmpP, CmpQ); 1764 } 1765 1766 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to 1767 // a shorter form that has more potential to be folded even further. 1768 Value *X1, *X2, *X3, *X4; 1769 if (match(Or->getOperand(0), m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) && 1770 match(Or->getOperand(1), m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) { 1771 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4) 1772 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4) 1773 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2); 1774 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4); 1775 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; 1776 return BinaryOperator::Create(BOpc, Cmp12, Cmp34); 1777 } 1778 1779 return nullptr; 1780 } 1781 1782 /// Fold icmp (mul X, Y), C. 1783 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp, 1784 BinaryOperator *Mul, 1785 const APInt &C) { 1786 const APInt *MulC; 1787 if (!match(Mul->getOperand(1), m_APInt(MulC))) 1788 return nullptr; 1789 1790 // If this is a test of the sign bit and the multiply is sign-preserving with 1791 // a constant operand, use the multiply LHS operand instead. 1792 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1793 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) { 1794 if (MulC->isNegative()) 1795 Pred = ICmpInst::getSwappedPredicate(Pred); 1796 return new ICmpInst(Pred, Mul->getOperand(0), 1797 Constant::getNullValue(Mul->getType())); 1798 } 1799 1800 return nullptr; 1801 } 1802 1803 /// Fold icmp (shl 1, Y), C. 1804 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl, 1805 const APInt &C) { 1806 Value *Y; 1807 if (!match(Shl, m_Shl(m_One(), m_Value(Y)))) 1808 return nullptr; 1809 1810 Type *ShiftType = Shl->getType(); 1811 unsigned TypeBits = C.getBitWidth(); 1812 bool CIsPowerOf2 = C.isPowerOf2(); 1813 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1814 if (Cmp.isUnsigned()) { 1815 // (1 << Y) pred C -> Y pred Log2(C) 1816 if (!CIsPowerOf2) { 1817 // (1 << Y) < 30 -> Y <= 4 1818 // (1 << Y) <= 30 -> Y <= 4 1819 // (1 << Y) >= 30 -> Y > 4 1820 // (1 << Y) > 30 -> Y > 4 1821 if (Pred == ICmpInst::ICMP_ULT) 1822 Pred = ICmpInst::ICMP_ULE; 1823 else if (Pred == ICmpInst::ICMP_UGE) 1824 Pred = ICmpInst::ICMP_UGT; 1825 } 1826 1827 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31 1828 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31 1829 unsigned CLog2 = C.logBase2(); 1830 if (CLog2 == TypeBits - 1) { 1831 if (Pred == ICmpInst::ICMP_UGE) 1832 Pred = ICmpInst::ICMP_EQ; 1833 else if (Pred == ICmpInst::ICMP_ULT) 1834 Pred = ICmpInst::ICMP_NE; 1835 } 1836 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2)); 1837 } else if (Cmp.isSigned()) { 1838 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1); 1839 if (C.isAllOnesValue()) { 1840 // (1 << Y) <= -1 -> Y == 31 1841 if (Pred == ICmpInst::ICMP_SLE) 1842 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne); 1843 1844 // (1 << Y) > -1 -> Y != 31 1845 if (Pred == ICmpInst::ICMP_SGT) 1846 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne); 1847 } else if (!C) { 1848 // (1 << Y) < 0 -> Y == 31 1849 // (1 << Y) <= 0 -> Y == 31 1850 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 1851 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne); 1852 1853 // (1 << Y) >= 0 -> Y != 31 1854 // (1 << Y) > 0 -> Y != 31 1855 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) 1856 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne); 1857 } 1858 } else if (Cmp.isEquality() && CIsPowerOf2) { 1859 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2())); 1860 } 1861 1862 return nullptr; 1863 } 1864 1865 /// Fold icmp (shl X, Y), C. 1866 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp, 1867 BinaryOperator *Shl, 1868 const APInt &C) { 1869 const APInt *ShiftVal; 1870 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal))) 1871 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal); 1872 1873 const APInt *ShiftAmt; 1874 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt))) 1875 return foldICmpShlOne(Cmp, Shl, C); 1876 1877 // Check that the shift amount is in range. If not, don't perform undefined 1878 // shifts. When the shift is visited, it will be simplified. 1879 unsigned TypeBits = C.getBitWidth(); 1880 if (ShiftAmt->uge(TypeBits)) 1881 return nullptr; 1882 1883 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1884 Value *X = Shl->getOperand(0); 1885 Type *ShType = Shl->getType(); 1886 1887 // NSW guarantees that we are only shifting out sign bits from the high bits, 1888 // so we can ASHR the compare constant without needing a mask and eliminate 1889 // the shift. 1890 if (Shl->hasNoSignedWrap()) { 1891 if (Pred == ICmpInst::ICMP_SGT) { 1892 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt) 1893 APInt ShiftedC = C.ashr(*ShiftAmt); 1894 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC)); 1895 } 1896 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && 1897 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) { 1898 APInt ShiftedC = C.ashr(*ShiftAmt); 1899 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC)); 1900 } 1901 if (Pred == ICmpInst::ICMP_SLT) { 1902 // SLE is the same as above, but SLE is canonicalized to SLT, so convert: 1903 // (X << S) <=s C is equiv to X <=s (C >> S) for all C 1904 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX 1905 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN 1906 assert(!C.isMinSignedValue() && "Unexpected icmp slt"); 1907 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1; 1908 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC)); 1909 } 1910 // If this is a signed comparison to 0 and the shift is sign preserving, 1911 // use the shift LHS operand instead; isSignTest may change 'Pred', so only 1912 // do that if we're sure to not continue on in this function. 1913 if (isSignTest(Pred, C)) 1914 return new ICmpInst(Pred, X, Constant::getNullValue(ShType)); 1915 } 1916 1917 // NUW guarantees that we are only shifting out zero bits from the high bits, 1918 // so we can LSHR the compare constant without needing a mask and eliminate 1919 // the shift. 1920 if (Shl->hasNoUnsignedWrap()) { 1921 if (Pred == ICmpInst::ICMP_UGT) { 1922 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt) 1923 APInt ShiftedC = C.lshr(*ShiftAmt); 1924 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC)); 1925 } 1926 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && 1927 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) { 1928 APInt ShiftedC = C.lshr(*ShiftAmt); 1929 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC)); 1930 } 1931 if (Pred == ICmpInst::ICMP_ULT) { 1932 // ULE is the same as above, but ULE is canonicalized to ULT, so convert: 1933 // (X << S) <=u C is equiv to X <=u (C >> S) for all C 1934 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u 1935 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0 1936 assert(C.ugt(0) && "ult 0 should have been eliminated"); 1937 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1; 1938 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC)); 1939 } 1940 } 1941 1942 if (Cmp.isEquality() && Shl->hasOneUse()) { 1943 // Strength-reduce the shift into an 'and'. 1944 Constant *Mask = ConstantInt::get( 1945 ShType, 1946 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue())); 1947 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask"); 1948 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt)); 1949 return new ICmpInst(Pred, And, LShrC); 1950 } 1951 1952 // Otherwise, if this is a comparison of the sign bit, simplify to and/test. 1953 bool TrueIfSigned = false; 1954 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) { 1955 // (X << 31) <s 0 --> (X & 1) != 0 1956 Constant *Mask = ConstantInt::get( 1957 ShType, 1958 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1)); 1959 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask"); 1960 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, 1961 And, Constant::getNullValue(ShType)); 1962 } 1963 1964 // Transform (icmp pred iM (shl iM %v, N), C) 1965 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N)) 1966 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N. 1967 // This enables us to get rid of the shift in favor of a trunc that may be 1968 // free on the target. It has the additional benefit of comparing to a 1969 // smaller constant that may be more target-friendly. 1970 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1); 1971 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt && 1972 DL.isLegalInteger(TypeBits - Amt)) { 1973 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt); 1974 if (ShType->isVectorTy()) 1975 TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements()); 1976 Constant *NewC = 1977 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt)); 1978 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC); 1979 } 1980 1981 return nullptr; 1982 } 1983 1984 /// Fold icmp ({al}shr X, Y), C. 1985 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp, 1986 BinaryOperator *Shr, 1987 const APInt &C) { 1988 // An exact shr only shifts out zero bits, so: 1989 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0 1990 Value *X = Shr->getOperand(0); 1991 CmpInst::Predicate Pred = Cmp.getPredicate(); 1992 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() && 1993 C.isNullValue()) 1994 return new ICmpInst(Pred, X, Cmp.getOperand(1)); 1995 1996 const APInt *ShiftVal; 1997 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal))) 1998 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal); 1999 2000 const APInt *ShiftAmt; 2001 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt))) 2002 return nullptr; 2003 2004 // Check that the shift amount is in range. If not, don't perform undefined 2005 // shifts. When the shift is visited it will be simplified. 2006 unsigned TypeBits = C.getBitWidth(); 2007 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits); 2008 if (ShAmtVal >= TypeBits || ShAmtVal == 0) 2009 return nullptr; 2010 2011 bool IsAShr = Shr->getOpcode() == Instruction::AShr; 2012 bool IsExact = Shr->isExact(); 2013 Type *ShrTy = Shr->getType(); 2014 // TODO: If we could guarantee that InstSimplify would handle all of the 2015 // constant-value-based preconditions in the folds below, then we could assert 2016 // those conditions rather than checking them. This is difficult because of 2017 // undef/poison (PR34838). 2018 if (IsAShr) { 2019 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) { 2020 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC) 2021 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC) 2022 APInt ShiftedC = C.shl(ShAmtVal); 2023 if (ShiftedC.ashr(ShAmtVal) == C) 2024 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC)); 2025 } 2026 if (Pred == CmpInst::ICMP_SGT) { 2027 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1 2028 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1; 2029 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() && 2030 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1)) 2031 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC)); 2032 } 2033 } else { 2034 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) { 2035 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC) 2036 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC) 2037 APInt ShiftedC = C.shl(ShAmtVal); 2038 if (ShiftedC.lshr(ShAmtVal) == C) 2039 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC)); 2040 } 2041 if (Pred == CmpInst::ICMP_UGT) { 2042 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1 2043 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1; 2044 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1)) 2045 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC)); 2046 } 2047 } 2048 2049 if (!Cmp.isEquality()) 2050 return nullptr; 2051 2052 // Handle equality comparisons of shift-by-constant. 2053 2054 // If the comparison constant changes with the shift, the comparison cannot 2055 // succeed (bits of the comparison constant cannot match the shifted value). 2056 // This should be known by InstSimplify and already be folded to true/false. 2057 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || 2058 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && 2059 "Expected icmp+shr simplify did not occur."); 2060 2061 // If the bits shifted out are known zero, compare the unshifted value: 2062 // (X & 4) >> 1 == 2 --> (X & 4) == 4. 2063 if (Shr->isExact()) 2064 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal)); 2065 2066 if (Shr->hasOneUse()) { 2067 // Canonicalize the shift into an 'and': 2068 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt) 2069 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); 2070 Constant *Mask = ConstantInt::get(ShrTy, Val); 2071 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask"); 2072 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal)); 2073 } 2074 2075 return nullptr; 2076 } 2077 2078 /// Fold icmp (udiv X, Y), C. 2079 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp, 2080 BinaryOperator *UDiv, 2081 const APInt &C) { 2082 const APInt *C2; 2083 if (!match(UDiv->getOperand(0), m_APInt(C2))) 2084 return nullptr; 2085 2086 assert(*C2 != 0 && "udiv 0, X should have been simplified already."); 2087 2088 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1)) 2089 Value *Y = UDiv->getOperand(1); 2090 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) { 2091 assert(!C.isMaxValue() && 2092 "icmp ugt X, UINT_MAX should have been simplified already."); 2093 return new ICmpInst(ICmpInst::ICMP_ULE, Y, 2094 ConstantInt::get(Y->getType(), C2->udiv(C + 1))); 2095 } 2096 2097 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C) 2098 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) { 2099 assert(C != 0 && "icmp ult X, 0 should have been simplified already."); 2100 return new ICmpInst(ICmpInst::ICMP_UGT, Y, 2101 ConstantInt::get(Y->getType(), C2->udiv(C))); 2102 } 2103 2104 return nullptr; 2105 } 2106 2107 /// Fold icmp ({su}div X, Y), C. 2108 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp, 2109 BinaryOperator *Div, 2110 const APInt &C) { 2111 // Fold: icmp pred ([us]div X, C2), C -> range test 2112 // Fold this div into the comparison, producing a range check. 2113 // Determine, based on the divide type, what the range is being 2114 // checked. If there is an overflow on the low or high side, remember 2115 // it, otherwise compute the range [low, hi) bounding the new value. 2116 // See: InsertRangeTest above for the kinds of replacements possible. 2117 const APInt *C2; 2118 if (!match(Div->getOperand(1), m_APInt(C2))) 2119 return nullptr; 2120 2121 // FIXME: If the operand types don't match the type of the divide 2122 // then don't attempt this transform. The code below doesn't have the 2123 // logic to deal with a signed divide and an unsigned compare (and 2124 // vice versa). This is because (x /s C2) <s C produces different 2125 // results than (x /s C2) <u C or (x /u C2) <s C or even 2126 // (x /u C2) <u C. Simply casting the operands and result won't 2127 // work. :( The if statement below tests that condition and bails 2128 // if it finds it. 2129 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv; 2130 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned()) 2131 return nullptr; 2132 2133 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with 2134 // INT_MIN will also fail if the divisor is 1. Although folds of all these 2135 // division-by-constant cases should be present, we can not assert that they 2136 // have happened before we reach this icmp instruction. 2137 if (C2->isNullValue() || C2->isOneValue() || 2138 (DivIsSigned && C2->isAllOnesValue())) 2139 return nullptr; 2140 2141 // Compute Prod = C * C2. We are essentially solving an equation of 2142 // form X / C2 = C. We solve for X by multiplying C2 and C. 2143 // By solving for X, we can turn this into a range check instead of computing 2144 // a divide. 2145 APInt Prod = C * *C2; 2146 2147 // Determine if the product overflows by seeing if the product is not equal to 2148 // the divide. Make sure we do the same kind of divide as in the LHS 2149 // instruction that we're folding. 2150 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C; 2151 2152 ICmpInst::Predicate Pred = Cmp.getPredicate(); 2153 2154 // If the division is known to be exact, then there is no remainder from the 2155 // divide, so the covered range size is unit, otherwise it is the divisor. 2156 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2; 2157 2158 // Figure out the interval that is being checked. For example, a comparison 2159 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). 2160 // Compute this interval based on the constants involved and the signedness of 2161 // the compare/divide. This computes a half-open interval, keeping track of 2162 // whether either value in the interval overflows. After analysis each 2163 // overflow variable is set to 0 if it's corresponding bound variable is valid 2164 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. 2165 int LoOverflow = 0, HiOverflow = 0; 2166 APInt LoBound, HiBound; 2167 2168 if (!DivIsSigned) { // udiv 2169 // e.g. X/5 op 3 --> [15, 20) 2170 LoBound = Prod; 2171 HiOverflow = LoOverflow = ProdOV; 2172 if (!HiOverflow) { 2173 // If this is not an exact divide, then many values in the range collapse 2174 // to the same result value. 2175 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false); 2176 } 2177 } else if (C2->isStrictlyPositive()) { // Divisor is > 0. 2178 if (C.isNullValue()) { // (X / pos) op 0 2179 // Can't overflow. e.g. X/2 op 0 --> [-1, 2) 2180 LoBound = -(RangeSize - 1); 2181 HiBound = RangeSize; 2182 } else if (C.isStrictlyPositive()) { // (X / pos) op pos 2183 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) 2184 HiOverflow = LoOverflow = ProdOV; 2185 if (!HiOverflow) 2186 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true); 2187 } else { // (X / pos) op neg 2188 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) 2189 HiBound = Prod + 1; 2190 LoOverflow = HiOverflow = ProdOV ? -1 : 0; 2191 if (!LoOverflow) { 2192 APInt DivNeg = -RangeSize; 2193 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; 2194 } 2195 } 2196 } else if (C2->isNegative()) { // Divisor is < 0. 2197 if (Div->isExact()) 2198 RangeSize.negate(); 2199 if (C.isNullValue()) { // (X / neg) op 0 2200 // e.g. X/-5 op 0 --> [-4, 5) 2201 LoBound = RangeSize + 1; 2202 HiBound = -RangeSize; 2203 if (HiBound == *C2) { // -INTMIN = INTMIN 2204 HiOverflow = 1; // [INTMIN+1, overflow) 2205 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN 2206 } 2207 } else if (C.isStrictlyPositive()) { // (X / neg) op pos 2208 // e.g. X/-5 op 3 --> [-19, -14) 2209 HiBound = Prod + 1; 2210 HiOverflow = LoOverflow = ProdOV ? -1 : 0; 2211 if (!LoOverflow) 2212 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; 2213 } else { // (X / neg) op neg 2214 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) 2215 LoOverflow = HiOverflow = ProdOV; 2216 if (!HiOverflow) 2217 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true); 2218 } 2219 2220 // Dividing by a negative swaps the condition. LT <-> GT 2221 Pred = ICmpInst::getSwappedPredicate(Pred); 2222 } 2223 2224 Value *X = Div->getOperand(0); 2225 switch (Pred) { 2226 default: llvm_unreachable("Unhandled icmp opcode!"); 2227 case ICmpInst::ICMP_EQ: 2228 if (LoOverflow && HiOverflow) 2229 return replaceInstUsesWith(Cmp, Builder.getFalse()); 2230 if (HiOverflow) 2231 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 2232 ICmpInst::ICMP_UGE, X, 2233 ConstantInt::get(Div->getType(), LoBound)); 2234 if (LoOverflow) 2235 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 2236 ICmpInst::ICMP_ULT, X, 2237 ConstantInt::get(Div->getType(), HiBound)); 2238 return replaceInstUsesWith( 2239 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true)); 2240 case ICmpInst::ICMP_NE: 2241 if (LoOverflow && HiOverflow) 2242 return replaceInstUsesWith(Cmp, Builder.getTrue()); 2243 if (HiOverflow) 2244 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : 2245 ICmpInst::ICMP_ULT, X, 2246 ConstantInt::get(Div->getType(), LoBound)); 2247 if (LoOverflow) 2248 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : 2249 ICmpInst::ICMP_UGE, X, 2250 ConstantInt::get(Div->getType(), HiBound)); 2251 return replaceInstUsesWith(Cmp, 2252 insertRangeTest(X, LoBound, HiBound, 2253 DivIsSigned, false)); 2254 case ICmpInst::ICMP_ULT: 2255 case ICmpInst::ICMP_SLT: 2256 if (LoOverflow == +1) // Low bound is greater than input range. 2257 return replaceInstUsesWith(Cmp, Builder.getTrue()); 2258 if (LoOverflow == -1) // Low bound is less than input range. 2259 return replaceInstUsesWith(Cmp, Builder.getFalse()); 2260 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound)); 2261 case ICmpInst::ICMP_UGT: 2262 case ICmpInst::ICMP_SGT: 2263 if (HiOverflow == +1) // High bound greater than input range. 2264 return replaceInstUsesWith(Cmp, Builder.getFalse()); 2265 if (HiOverflow == -1) // High bound less than input range. 2266 return replaceInstUsesWith(Cmp, Builder.getTrue()); 2267 if (Pred == ICmpInst::ICMP_UGT) 2268 return new ICmpInst(ICmpInst::ICMP_UGE, X, 2269 ConstantInt::get(Div->getType(), HiBound)); 2270 return new ICmpInst(ICmpInst::ICMP_SGE, X, 2271 ConstantInt::get(Div->getType(), HiBound)); 2272 } 2273 2274 return nullptr; 2275 } 2276 2277 /// Fold icmp (sub X, Y), C. 2278 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp, 2279 BinaryOperator *Sub, 2280 const APInt &C) { 2281 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1); 2282 ICmpInst::Predicate Pred = Cmp.getPredicate(); 2283 2284 // The following transforms are only worth it if the only user of the subtract 2285 // is the icmp. 2286 if (!Sub->hasOneUse()) 2287 return nullptr; 2288 2289 if (Sub->hasNoSignedWrap()) { 2290 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y) 2291 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue()) 2292 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y); 2293 2294 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y) 2295 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue()) 2296 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y); 2297 2298 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y) 2299 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue()) 2300 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y); 2301 2302 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y) 2303 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue()) 2304 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y); 2305 } 2306 2307 const APInt *C2; 2308 if (!match(X, m_APInt(C2))) 2309 return nullptr; 2310 2311 // C2 - Y <u C -> (Y | (C - 1)) == C2 2312 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2 2313 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && 2314 (*C2 & (C - 1)) == (C - 1)) 2315 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X); 2316 2317 // C2 - Y >u C -> (Y | C) != C2 2318 // iff C2 & C == C and C + 1 is a power of 2 2319 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C) 2320 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X); 2321 2322 return nullptr; 2323 } 2324 2325 /// Fold icmp (add X, Y), C. 2326 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp, 2327 BinaryOperator *Add, 2328 const APInt &C) { 2329 Value *Y = Add->getOperand(1); 2330 const APInt *C2; 2331 if (Cmp.isEquality() || !match(Y, m_APInt(C2))) 2332 return nullptr; 2333 2334 // Fold icmp pred (add X, C2), C. 2335 Value *X = Add->getOperand(0); 2336 Type *Ty = Add->getType(); 2337 CmpInst::Predicate Pred = Cmp.getPredicate(); 2338 2339 // If the add does not wrap, we can always adjust the compare by subtracting 2340 // the constants. Equality comparisons are handled elsewhere. SGE/SLE are 2341 // canonicalized to SGT/SLT. 2342 if (Add->hasNoSignedWrap() && 2343 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) { 2344 bool Overflow; 2345 APInt NewC = C.ssub_ov(*C2, Overflow); 2346 // If there is overflow, the result must be true or false. 2347 // TODO: Can we assert there is no overflow because InstSimplify always 2348 // handles those cases? 2349 if (!Overflow) 2350 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2) 2351 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC)); 2352 } 2353 2354 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2); 2355 const APInt &Upper = CR.getUpper(); 2356 const APInt &Lower = CR.getLower(); 2357 if (Cmp.isSigned()) { 2358 if (Lower.isSignMask()) 2359 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper)); 2360 if (Upper.isSignMask()) 2361 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower)); 2362 } else { 2363 if (Lower.isMinValue()) 2364 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper)); 2365 if (Upper.isMinValue()) 2366 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower)); 2367 } 2368 2369 if (!Add->hasOneUse()) 2370 return nullptr; 2371 2372 // X+C <u C2 -> (X & -C2) == C 2373 // iff C & (C2-1) == 0 2374 // C2 is a power of 2 2375 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0) 2376 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C), 2377 ConstantExpr::getNeg(cast<Constant>(Y))); 2378 2379 // X+C >u C2 -> (X & ~C2) != C 2380 // iff C & C2 == 0 2381 // C2+1 is a power of 2 2382 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0) 2383 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C), 2384 ConstantExpr::getNeg(cast<Constant>(Y))); 2385 2386 return nullptr; 2387 } 2388 2389 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, 2390 Value *&RHS, ConstantInt *&Less, 2391 ConstantInt *&Equal, 2392 ConstantInt *&Greater) { 2393 // TODO: Generalize this to work with other comparison idioms or ensure 2394 // they get canonicalized into this form. 2395 2396 // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32 2397 // Greater), where Equal, Less and Greater are placeholders for any three 2398 // constants. 2399 ICmpInst::Predicate PredA, PredB; 2400 if (match(SI->getTrueValue(), m_ConstantInt(Equal)) && 2401 match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) && 2402 PredA == ICmpInst::ICMP_EQ && 2403 match(SI->getFalseValue(), 2404 m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)), 2405 m_ConstantInt(Less), m_ConstantInt(Greater))) && 2406 PredB == ICmpInst::ICMP_SLT) { 2407 return true; 2408 } 2409 return false; 2410 } 2411 2412 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp, 2413 SelectInst *Select, 2414 ConstantInt *C) { 2415 2416 assert(C && "Cmp RHS should be a constant int!"); 2417 // If we're testing a constant value against the result of a three way 2418 // comparison, the result can be expressed directly in terms of the 2419 // original values being compared. Note: We could possibly be more 2420 // aggressive here and remove the hasOneUse test. The original select is 2421 // really likely to simplify or sink when we remove a test of the result. 2422 Value *OrigLHS, *OrigRHS; 2423 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan; 2424 if (Cmp.hasOneUse() && 2425 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal, 2426 C3GreaterThan)) { 2427 assert(C1LessThan && C2Equal && C3GreaterThan); 2428 2429 bool TrueWhenLessThan = 2430 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C) 2431 ->isAllOnesValue(); 2432 bool TrueWhenEqual = 2433 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C) 2434 ->isAllOnesValue(); 2435 bool TrueWhenGreaterThan = 2436 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C) 2437 ->isAllOnesValue(); 2438 2439 // This generates the new instruction that will replace the original Cmp 2440 // Instruction. Instead of enumerating the various combinations when 2441 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus 2442 // false, we rely on chaining of ORs and future passes of InstCombine to 2443 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b). 2444 2445 // When none of the three constants satisfy the predicate for the RHS (C), 2446 // the entire original Cmp can be simplified to a false. 2447 Value *Cond = Builder.getFalse(); 2448 if (TrueWhenLessThan) 2449 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT, OrigLHS, OrigRHS)); 2450 if (TrueWhenEqual) 2451 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ, OrigLHS, OrigRHS)); 2452 if (TrueWhenGreaterThan) 2453 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT, OrigLHS, OrigRHS)); 2454 2455 return replaceInstUsesWith(Cmp, Cond); 2456 } 2457 return nullptr; 2458 } 2459 2460 Instruction *InstCombiner::foldICmpBitCastConstant(ICmpInst &Cmp, 2461 BitCastInst *Bitcast, 2462 const APInt &C) { 2463 // Folding: icmp <pred> iN X, C 2464 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN 2465 // and C is a splat of a K-bit pattern 2466 // and SC is a constant vector = <C', C', C', ..., C'> 2467 // Into: 2468 // %E = extractelement <M x iK> %vec, i32 C' 2469 // icmp <pred> iK %E, trunc(C) 2470 if (!Bitcast->getType()->isIntegerTy() || 2471 !Bitcast->getSrcTy()->isIntOrIntVectorTy()) 2472 return nullptr; 2473 2474 Value *BCIOp = Bitcast->getOperand(0); 2475 Value *Vec = nullptr; // 1st vector arg of the shufflevector 2476 Constant *Mask = nullptr; // Mask arg of the shufflevector 2477 if (match(BCIOp, 2478 m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) { 2479 // Check whether every element of Mask is the same constant 2480 if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) { 2481 auto *VecTy = cast<VectorType>(BCIOp->getType()); 2482 auto *EltTy = cast<IntegerType>(VecTy->getElementType()); 2483 auto Pred = Cmp.getPredicate(); 2484 if (C.isSplat(EltTy->getBitWidth())) { 2485 // Fold the icmp based on the value of C 2486 // If C is M copies of an iK sized bit pattern, 2487 // then: 2488 // => %E = extractelement <N x iK> %vec, i32 Elem 2489 // icmp <pred> iK %SplatVal, <pattern> 2490 Value *Extract = Builder.CreateExtractElement(Vec, Elem); 2491 Value *NewC = ConstantInt::get(EltTy, C.trunc(EltTy->getBitWidth())); 2492 return new ICmpInst(Pred, Extract, NewC); 2493 } 2494 } 2495 } 2496 return nullptr; 2497 } 2498 2499 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C 2500 /// where X is some kind of instruction. 2501 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) { 2502 const APInt *C; 2503 if (!match(Cmp.getOperand(1), m_APInt(C))) 2504 return nullptr; 2505 2506 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) { 2507 switch (BO->getOpcode()) { 2508 case Instruction::Xor: 2509 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C)) 2510 return I; 2511 break; 2512 case Instruction::And: 2513 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C)) 2514 return I; 2515 break; 2516 case Instruction::Or: 2517 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C)) 2518 return I; 2519 break; 2520 case Instruction::Mul: 2521 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C)) 2522 return I; 2523 break; 2524 case Instruction::Shl: 2525 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C)) 2526 return I; 2527 break; 2528 case Instruction::LShr: 2529 case Instruction::AShr: 2530 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C)) 2531 return I; 2532 break; 2533 case Instruction::UDiv: 2534 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C)) 2535 return I; 2536 LLVM_FALLTHROUGH; 2537 case Instruction::SDiv: 2538 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C)) 2539 return I; 2540 break; 2541 case Instruction::Sub: 2542 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C)) 2543 return I; 2544 break; 2545 case Instruction::Add: 2546 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C)) 2547 return I; 2548 break; 2549 default: 2550 break; 2551 } 2552 // TODO: These folds could be refactored to be part of the above calls. 2553 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C)) 2554 return I; 2555 } 2556 2557 // Match against CmpInst LHS being instructions other than binary operators. 2558 2559 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) { 2560 // For now, we only support constant integers while folding the 2561 // ICMP(SELECT)) pattern. We can extend this to support vector of integers 2562 // similar to the cases handled by binary ops above. 2563 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1))) 2564 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS)) 2565 return I; 2566 } 2567 2568 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) { 2569 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C)) 2570 return I; 2571 } 2572 2573 if (auto *BCI = dyn_cast<BitCastInst>(Cmp.getOperand(0))) { 2574 if (Instruction *I = foldICmpBitCastConstant(Cmp, BCI, *C)) 2575 return I; 2576 } 2577 2578 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, *C)) 2579 return I; 2580 2581 return nullptr; 2582 } 2583 2584 /// Fold an icmp equality instruction with binary operator LHS and constant RHS: 2585 /// icmp eq/ne BO, C. 2586 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp, 2587 BinaryOperator *BO, 2588 const APInt &C) { 2589 // TODO: Some of these folds could work with arbitrary constants, but this 2590 // function is limited to scalar and vector splat constants. 2591 if (!Cmp.isEquality()) 2592 return nullptr; 2593 2594 ICmpInst::Predicate Pred = Cmp.getPredicate(); 2595 bool isICMP_NE = Pred == ICmpInst::ICMP_NE; 2596 Constant *RHS = cast<Constant>(Cmp.getOperand(1)); 2597 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); 2598 2599 switch (BO->getOpcode()) { 2600 case Instruction::SRem: 2601 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. 2602 if (C.isNullValue() && BO->hasOneUse()) { 2603 const APInt *BOC; 2604 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) { 2605 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName()); 2606 return new ICmpInst(Pred, NewRem, 2607 Constant::getNullValue(BO->getType())); 2608 } 2609 } 2610 break; 2611 case Instruction::Add: { 2612 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. 2613 const APInt *BOC; 2614 if (match(BOp1, m_APInt(BOC))) { 2615 if (BO->hasOneUse()) { 2616 Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1)); 2617 return new ICmpInst(Pred, BOp0, SubC); 2618 } 2619 } else if (C.isNullValue()) { 2620 // Replace ((add A, B) != 0) with (A != -B) if A or B is 2621 // efficiently invertible, or if the add has just this one use. 2622 if (Value *NegVal = dyn_castNegVal(BOp1)) 2623 return new ICmpInst(Pred, BOp0, NegVal); 2624 if (Value *NegVal = dyn_castNegVal(BOp0)) 2625 return new ICmpInst(Pred, NegVal, BOp1); 2626 if (BO->hasOneUse()) { 2627 Value *Neg = Builder.CreateNeg(BOp1); 2628 Neg->takeName(BO); 2629 return new ICmpInst(Pred, BOp0, Neg); 2630 } 2631 } 2632 break; 2633 } 2634 case Instruction::Xor: 2635 if (BO->hasOneUse()) { 2636 if (Constant *BOC = dyn_cast<Constant>(BOp1)) { 2637 // For the xor case, we can xor two constants together, eliminating 2638 // the explicit xor. 2639 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC)); 2640 } else if (C.isNullValue()) { 2641 // Replace ((xor A, B) != 0) with (A != B) 2642 return new ICmpInst(Pred, BOp0, BOp1); 2643 } 2644 } 2645 break; 2646 case Instruction::Sub: 2647 if (BO->hasOneUse()) { 2648 const APInt *BOC; 2649 if (match(BOp0, m_APInt(BOC))) { 2650 // Replace ((sub BOC, B) != C) with (B != BOC-C). 2651 Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS); 2652 return new ICmpInst(Pred, BOp1, SubC); 2653 } else if (C.isNullValue()) { 2654 // Replace ((sub A, B) != 0) with (A != B). 2655 return new ICmpInst(Pred, BOp0, BOp1); 2656 } 2657 } 2658 break; 2659 case Instruction::Or: { 2660 const APInt *BOC; 2661 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) { 2662 // Comparing if all bits outside of a constant mask are set? 2663 // Replace (X | C) == -1 with (X & ~C) == ~C. 2664 // This removes the -1 constant. 2665 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1)); 2666 Value *And = Builder.CreateAnd(BOp0, NotBOC); 2667 return new ICmpInst(Pred, And, NotBOC); 2668 } 2669 break; 2670 } 2671 case Instruction::And: { 2672 const APInt *BOC; 2673 if (match(BOp1, m_APInt(BOC))) { 2674 // If we have ((X & C) == C), turn it into ((X & C) != 0). 2675 if (C == *BOC && C.isPowerOf2()) 2676 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, 2677 BO, Constant::getNullValue(RHS->getType())); 2678 2679 // Don't perform the following transforms if the AND has multiple uses 2680 if (!BO->hasOneUse()) 2681 break; 2682 2683 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 2684 if (BOC->isSignMask()) { 2685 Constant *Zero = Constant::getNullValue(BOp0->getType()); 2686 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; 2687 return new ICmpInst(NewPred, BOp0, Zero); 2688 } 2689 2690 // ((X & ~7) == 0) --> X < 8 2691 if (C.isNullValue() && (~(*BOC) + 1).isPowerOf2()) { 2692 Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1)); 2693 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 2694 return new ICmpInst(NewPred, BOp0, NegBOC); 2695 } 2696 } 2697 break; 2698 } 2699 case Instruction::Mul: 2700 if (C.isNullValue() && BO->hasNoSignedWrap()) { 2701 const APInt *BOC; 2702 if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) { 2703 // The trivial case (mul X, 0) is handled by InstSimplify. 2704 // General case : (mul X, C) != 0 iff X != 0 2705 // (mul X, C) == 0 iff X == 0 2706 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType())); 2707 } 2708 } 2709 break; 2710 case Instruction::UDiv: 2711 if (C.isNullValue()) { 2712 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A) 2713 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT; 2714 return new ICmpInst(NewPred, BOp1, BOp0); 2715 } 2716 break; 2717 default: 2718 break; 2719 } 2720 return nullptr; 2721 } 2722 2723 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C. 2724 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp, 2725 const APInt &C) { 2726 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)); 2727 if (!II || !Cmp.isEquality()) 2728 return nullptr; 2729 2730 // Handle icmp {eq|ne} <intrinsic>, Constant. 2731 Type *Ty = II->getType(); 2732 switch (II->getIntrinsicID()) { 2733 case Intrinsic::bswap: 2734 Worklist.Add(II); 2735 Cmp.setOperand(0, II->getArgOperand(0)); 2736 Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap())); 2737 return &Cmp; 2738 2739 case Intrinsic::ctlz: 2740 case Intrinsic::cttz: 2741 // ctz(A) == bitwidth(A) -> A == 0 and likewise for != 2742 if (C == C.getBitWidth()) { 2743 Worklist.Add(II); 2744 Cmp.setOperand(0, II->getArgOperand(0)); 2745 Cmp.setOperand(1, ConstantInt::getNullValue(Ty)); 2746 return &Cmp; 2747 } 2748 break; 2749 2750 case Intrinsic::ctpop: { 2751 // popcount(A) == 0 -> A == 0 and likewise for != 2752 // popcount(A) == bitwidth(A) -> A == -1 and likewise for != 2753 bool IsZero = C.isNullValue(); 2754 if (IsZero || C == C.getBitWidth()) { 2755 Worklist.Add(II); 2756 Cmp.setOperand(0, II->getArgOperand(0)); 2757 auto *NewOp = 2758 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty); 2759 Cmp.setOperand(1, NewOp); 2760 return &Cmp; 2761 } 2762 break; 2763 } 2764 default: 2765 break; 2766 } 2767 2768 return nullptr; 2769 } 2770 2771 /// Handle icmp with constant (but not simple integer constant) RHS. 2772 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) { 2773 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2774 Constant *RHSC = dyn_cast<Constant>(Op1); 2775 Instruction *LHSI = dyn_cast<Instruction>(Op0); 2776 if (!RHSC || !LHSI) 2777 return nullptr; 2778 2779 switch (LHSI->getOpcode()) { 2780 case Instruction::GetElementPtr: 2781 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null 2782 if (RHSC->isNullValue() && 2783 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices()) 2784 return new ICmpInst( 2785 I.getPredicate(), LHSI->getOperand(0), 2786 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2787 break; 2788 case Instruction::PHI: 2789 // Only fold icmp into the PHI if the phi and icmp are in the same 2790 // block. If in the same block, we're encouraging jump threading. If 2791 // not, we are just pessimizing the code by making an i1 phi. 2792 if (LHSI->getParent() == I.getParent()) 2793 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI))) 2794 return NV; 2795 break; 2796 case Instruction::Select: { 2797 // If either operand of the select is a constant, we can fold the 2798 // comparison into the select arms, which will cause one to be 2799 // constant folded and the select turned into a bitwise or. 2800 Value *Op1 = nullptr, *Op2 = nullptr; 2801 ConstantInt *CI = nullptr; 2802 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) { 2803 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2804 CI = dyn_cast<ConstantInt>(Op1); 2805 } 2806 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) { 2807 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC); 2808 CI = dyn_cast<ConstantInt>(Op2); 2809 } 2810 2811 // We only want to perform this transformation if it will not lead to 2812 // additional code. This is true if either both sides of the select 2813 // fold to a constant (in which case the icmp is replaced with a select 2814 // which will usually simplify) or this is the only user of the 2815 // select (in which case we are trading a select+icmp for a simpler 2816 // select+icmp) or all uses of the select can be replaced based on 2817 // dominance information ("Global cases"). 2818 bool Transform = false; 2819 if (Op1 && Op2) 2820 Transform = true; 2821 else if (Op1 || Op2) { 2822 // Local case 2823 if (LHSI->hasOneUse()) 2824 Transform = true; 2825 // Global cases 2826 else if (CI && !CI->isZero()) 2827 // When Op1 is constant try replacing select with second operand. 2828 // Otherwise Op2 is constant and try replacing select with first 2829 // operand. 2830 Transform = 2831 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1); 2832 } 2833 if (Transform) { 2834 if (!Op1) 2835 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC, 2836 I.getName()); 2837 if (!Op2) 2838 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC, 2839 I.getName()); 2840 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2); 2841 } 2842 break; 2843 } 2844 case Instruction::IntToPtr: 2845 // icmp pred inttoptr(X), null -> icmp pred X, 0 2846 if (RHSC->isNullValue() && 2847 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType()) 2848 return new ICmpInst( 2849 I.getPredicate(), LHSI->getOperand(0), 2850 Constant::getNullValue(LHSI->getOperand(0)->getType())); 2851 break; 2852 2853 case Instruction::Load: 2854 // Try to optimize things like "A[i] > 4" to index computations. 2855 if (GetElementPtrInst *GEP = 2856 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 2857 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 2858 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 2859 !cast<LoadInst>(LHSI)->isVolatile()) 2860 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I)) 2861 return Res; 2862 } 2863 break; 2864 } 2865 2866 return nullptr; 2867 } 2868 2869 /// Some comparisons can be simplified. 2870 /// In this case, we are looking for comparisons that look like 2871 /// a check for a lossy truncation. 2872 /// Folds: 2873 /// x & (-1 >> y) SrcPred x to x DstPred (-1 >> y) 2874 /// The Mask can be a constant, too. 2875 /// For some predicates, the operands are commutative. 2876 /// For others, x can only be on a specific side. 2877 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I, 2878 InstCombiner::BuilderTy &Builder) { 2879 ICmpInst::Predicate SrcPred; 2880 Value *X, *M; 2881 auto m_Mask = m_CombineOr(m_LShr(m_AllOnes(), m_Value()), m_LowBitMask()); 2882 if (!match(&I, m_c_ICmp(SrcPred, 2883 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)), 2884 m_Deferred(X)))) 2885 return nullptr; 2886 2887 ICmpInst::Predicate DstPred; 2888 switch (SrcPred) { 2889 case ICmpInst::Predicate::ICMP_EQ: 2890 // x & (-1 >> y) == x -> x u<= (-1 >> y) 2891 DstPred = ICmpInst::Predicate::ICMP_ULE; 2892 break; 2893 case ICmpInst::Predicate::ICMP_NE: 2894 // x & (-1 >> y) != x -> x u> (-1 >> y) 2895 DstPred = ICmpInst::Predicate::ICMP_UGT; 2896 break; 2897 case ICmpInst::Predicate::ICMP_UGT: 2898 // x u> x & (-1 >> y) -> x u> (-1 >> y) 2899 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant"); 2900 DstPred = ICmpInst::Predicate::ICMP_UGT; 2901 break; 2902 case ICmpInst::Predicate::ICMP_UGE: 2903 // x & (-1 >> y) u>= x -> x u<= (-1 >> y) 2904 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant"); 2905 DstPred = ICmpInst::Predicate::ICMP_ULE; 2906 break; 2907 case ICmpInst::Predicate::ICMP_ULT: 2908 // x & (-1 >> y) u< x -> x u> (-1 >> y) 2909 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant"); 2910 DstPred = ICmpInst::Predicate::ICMP_UGT; 2911 break; 2912 case ICmpInst::Predicate::ICMP_ULE: 2913 // x u<= x & (-1 >> y) -> x u<= (-1 >> y) 2914 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant"); 2915 DstPred = ICmpInst::Predicate::ICMP_ULE; 2916 break; 2917 case ICmpInst::Predicate::ICMP_SGT: 2918 // x s> x & (-1 >> y) -> x s> (-1 >> y) 2919 if (X != I.getOperand(0)) // X must be on LHS of comparison! 2920 return nullptr; // Ignore the other case. 2921 DstPred = ICmpInst::Predicate::ICMP_SGT; 2922 break; 2923 case ICmpInst::Predicate::ICMP_SGE: 2924 // x & (-1 >> y) s>= x -> x s<= (-1 >> y) 2925 if (X != I.getOperand(1)) // X must be on RHS of comparison! 2926 return nullptr; // Ignore the other case. 2927 if (!match(M, m_Constant())) // Can not do this fold with non-constant. 2928 return nullptr; 2929 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements. 2930 return nullptr; 2931 DstPred = ICmpInst::Predicate::ICMP_SLE; 2932 break; 2933 case ICmpInst::Predicate::ICMP_SLT: 2934 // x & (-1 >> y) s< x -> x s> (-1 >> y) 2935 if (X != I.getOperand(1)) // X must be on RHS of comparison! 2936 return nullptr; // Ignore the other case. 2937 if (!match(M, m_Constant())) // Can not do this fold with non-constant. 2938 return nullptr; 2939 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements. 2940 return nullptr; 2941 DstPred = ICmpInst::Predicate::ICMP_SGT; 2942 break; 2943 case ICmpInst::Predicate::ICMP_SLE: 2944 // x s<= x & (-1 >> y) -> x s<= (-1 >> y) 2945 if (X != I.getOperand(0)) // X must be on LHS of comparison! 2946 return nullptr; // Ignore the other case. 2947 DstPred = ICmpInst::Predicate::ICMP_SLE; 2948 break; 2949 default: 2950 llvm_unreachable("All possible folds are handled."); 2951 } 2952 2953 return Builder.CreateICmp(DstPred, X, M); 2954 } 2955 2956 /// Some comparisons can be simplified. 2957 /// In this case, we are looking for comparisons that look like 2958 /// a check for a lossy signed truncation. 2959 /// Folds: (MaskedBits is a constant.) 2960 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x 2961 /// Into: 2962 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits) 2963 /// Where KeptBits = bitwidth(%x) - MaskedBits 2964 static Value * 2965 foldICmpWithTruncSignExtendedVal(ICmpInst &I, 2966 InstCombiner::BuilderTy &Builder) { 2967 ICmpInst::Predicate SrcPred; 2968 Value *X; 2969 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef. 2970 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use. 2971 if (!match(&I, m_c_ICmp(SrcPred, 2972 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)), 2973 m_APInt(C1))), 2974 m_Deferred(X)))) 2975 return nullptr; 2976 2977 // Potential handling of non-splats: for each element: 2978 // * if both are undef, replace with constant 0. 2979 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0. 2980 // * if both are not undef, and are different, bailout. 2981 // * else, only one is undef, then pick the non-undef one. 2982 2983 // The shift amount must be equal. 2984 if (*C0 != *C1) 2985 return nullptr; 2986 const APInt &MaskedBits = *C0; 2987 assert(MaskedBits != 0 && "shift by zero should be folded away already."); 2988 2989 ICmpInst::Predicate DstPred; 2990 switch (SrcPred) { 2991 case ICmpInst::Predicate::ICMP_EQ: 2992 // ((%x << MaskedBits) a>> MaskedBits) == %x 2993 // => 2994 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits) 2995 DstPred = ICmpInst::Predicate::ICMP_ULT; 2996 break; 2997 case ICmpInst::Predicate::ICMP_NE: 2998 // ((%x << MaskedBits) a>> MaskedBits) != %x 2999 // => 3000 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits) 3001 DstPred = ICmpInst::Predicate::ICMP_UGE; 3002 break; 3003 // FIXME: are more folds possible? 3004 default: 3005 return nullptr; 3006 } 3007 3008 auto *XType = X->getType(); 3009 const unsigned XBitWidth = XType->getScalarSizeInBits(); 3010 const APInt BitWidth = APInt(XBitWidth, XBitWidth); 3011 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched"); 3012 3013 // KeptBits = bitwidth(%x) - MaskedBits 3014 const APInt KeptBits = BitWidth - MaskedBits; 3015 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable"); 3016 // ICmpCst = (1 << KeptBits) 3017 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits); 3018 assert(ICmpCst.isPowerOf2()); 3019 // AddCst = (1 << (KeptBits-1)) 3020 const APInt AddCst = ICmpCst.lshr(1); 3021 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2()); 3022 3023 // T0 = add %x, AddCst 3024 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst)); 3025 // T1 = T0 DstPred ICmpCst 3026 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst)); 3027 3028 return T1; 3029 } 3030 3031 /// Try to fold icmp (binop), X or icmp X, (binop). 3032 /// TODO: A large part of this logic is duplicated in InstSimplify's 3033 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code 3034 /// duplication. 3035 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) { 3036 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3037 3038 // Special logic for binary operators. 3039 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0); 3040 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1); 3041 if (!BO0 && !BO1) 3042 return nullptr; 3043 3044 const CmpInst::Predicate Pred = I.getPredicate(); 3045 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; 3046 if (BO0 && isa<OverflowingBinaryOperator>(BO0)) 3047 NoOp0WrapProblem = 3048 ICmpInst::isEquality(Pred) || 3049 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) || 3050 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap()); 3051 if (BO1 && isa<OverflowingBinaryOperator>(BO1)) 3052 NoOp1WrapProblem = 3053 ICmpInst::isEquality(Pred) || 3054 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) || 3055 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap()); 3056 3057 // Analyze the case when either Op0 or Op1 is an add instruction. 3058 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). 3059 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 3060 if (BO0 && BO0->getOpcode() == Instruction::Add) { 3061 A = BO0->getOperand(0); 3062 B = BO0->getOperand(1); 3063 } 3064 if (BO1 && BO1->getOpcode() == Instruction::Add) { 3065 C = BO1->getOperand(0); 3066 D = BO1->getOperand(1); 3067 } 3068 3069 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. 3070 if ((A == Op1 || B == Op1) && NoOp0WrapProblem) 3071 return new ICmpInst(Pred, A == Op1 ? B : A, 3072 Constant::getNullValue(Op1->getType())); 3073 3074 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. 3075 if ((C == Op0 || D == Op0) && NoOp1WrapProblem) 3076 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()), 3077 C == Op0 ? D : C); 3078 3079 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow. 3080 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem && 3081 NoOp1WrapProblem && 3082 // Try not to increase register pressure. 3083 BO0->hasOneUse() && BO1->hasOneUse()) { 3084 // Determine Y and Z in the form icmp (X+Y), (X+Z). 3085 Value *Y, *Z; 3086 if (A == C) { 3087 // C + B == C + D -> B == D 3088 Y = B; 3089 Z = D; 3090 } else if (A == D) { 3091 // D + B == C + D -> B == C 3092 Y = B; 3093 Z = C; 3094 } else if (B == C) { 3095 // A + C == C + D -> A == D 3096 Y = A; 3097 Z = D; 3098 } else { 3099 assert(B == D); 3100 // A + D == C + D -> A == C 3101 Y = A; 3102 Z = C; 3103 } 3104 return new ICmpInst(Pred, Y, Z); 3105 } 3106 3107 // icmp slt (X + -1), Y -> icmp sle X, Y 3108 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT && 3109 match(B, m_AllOnes())) 3110 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1); 3111 3112 // icmp sge (X + -1), Y -> icmp sgt X, Y 3113 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE && 3114 match(B, m_AllOnes())) 3115 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1); 3116 3117 // icmp sle (X + 1), Y -> icmp slt X, Y 3118 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One())) 3119 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1); 3120 3121 // icmp sgt (X + 1), Y -> icmp sge X, Y 3122 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One())) 3123 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1); 3124 3125 // icmp sgt X, (Y + -1) -> icmp sge X, Y 3126 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT && 3127 match(D, m_AllOnes())) 3128 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C); 3129 3130 // icmp sle X, (Y + -1) -> icmp slt X, Y 3131 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE && 3132 match(D, m_AllOnes())) 3133 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C); 3134 3135 // icmp sge X, (Y + 1) -> icmp sgt X, Y 3136 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One())) 3137 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C); 3138 3139 // icmp slt X, (Y + 1) -> icmp sle X, Y 3140 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One())) 3141 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C); 3142 3143 // TODO: The subtraction-related identities shown below also hold, but 3144 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations 3145 // wouldn't happen even if they were implemented. 3146 // 3147 // icmp ult (X - 1), Y -> icmp ule X, Y 3148 // icmp uge (X - 1), Y -> icmp ugt X, Y 3149 // icmp ugt X, (Y - 1) -> icmp uge X, Y 3150 // icmp ule X, (Y - 1) -> icmp ult X, Y 3151 3152 // icmp ule (X + 1), Y -> icmp ult X, Y 3153 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One())) 3154 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1); 3155 3156 // icmp ugt (X + 1), Y -> icmp uge X, Y 3157 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One())) 3158 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1); 3159 3160 // icmp uge X, (Y + 1) -> icmp ugt X, Y 3161 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One())) 3162 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C); 3163 3164 // icmp ult X, (Y + 1) -> icmp ule X, Y 3165 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One())) 3166 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C); 3167 3168 // if C1 has greater magnitude than C2: 3169 // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y 3170 // s.t. C3 = C1 - C2 3171 // 3172 // if C2 has greater magnitude than C1: 3173 // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3) 3174 // s.t. C3 = C2 - C1 3175 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem && 3176 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) 3177 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B)) 3178 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) { 3179 const APInt &AP1 = C1->getValue(); 3180 const APInt &AP2 = C2->getValue(); 3181 if (AP1.isNegative() == AP2.isNegative()) { 3182 APInt AP1Abs = C1->getValue().abs(); 3183 APInt AP2Abs = C2->getValue().abs(); 3184 if (AP1Abs.uge(AP2Abs)) { 3185 ConstantInt *C3 = Builder.getInt(AP1 - AP2); 3186 Value *NewAdd = Builder.CreateNSWAdd(A, C3); 3187 return new ICmpInst(Pred, NewAdd, C); 3188 } else { 3189 ConstantInt *C3 = Builder.getInt(AP2 - AP1); 3190 Value *NewAdd = Builder.CreateNSWAdd(C, C3); 3191 return new ICmpInst(Pred, A, NewAdd); 3192 } 3193 } 3194 } 3195 3196 // Analyze the case when either Op0 or Op1 is a sub instruction. 3197 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). 3198 A = nullptr; 3199 B = nullptr; 3200 C = nullptr; 3201 D = nullptr; 3202 if (BO0 && BO0->getOpcode() == Instruction::Sub) { 3203 A = BO0->getOperand(0); 3204 B = BO0->getOperand(1); 3205 } 3206 if (BO1 && BO1->getOpcode() == Instruction::Sub) { 3207 C = BO1->getOperand(0); 3208 D = BO1->getOperand(1); 3209 } 3210 3211 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow. 3212 if (A == Op1 && NoOp0WrapProblem) 3213 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B); 3214 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow. 3215 if (C == Op0 && NoOp1WrapProblem) 3216 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType())); 3217 3218 // (A - B) >u A --> A <u B 3219 if (A == Op1 && Pred == ICmpInst::ICMP_UGT) 3220 return new ICmpInst(ICmpInst::ICMP_ULT, A, B); 3221 // C <u (C - D) --> C <u D 3222 if (C == Op0 && Pred == ICmpInst::ICMP_ULT) 3223 return new ICmpInst(ICmpInst::ICMP_ULT, C, D); 3224 3225 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow. 3226 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem && 3227 // Try not to increase register pressure. 3228 BO0->hasOneUse() && BO1->hasOneUse()) 3229 return new ICmpInst(Pred, A, C); 3230 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow. 3231 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem && 3232 // Try not to increase register pressure. 3233 BO0->hasOneUse() && BO1->hasOneUse()) 3234 return new ICmpInst(Pred, D, B); 3235 3236 // icmp (0-X) < cst --> x > -cst 3237 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) { 3238 Value *X; 3239 if (match(BO0, m_Neg(m_Value(X)))) 3240 if (Constant *RHSC = dyn_cast<Constant>(Op1)) 3241 if (RHSC->isNotMinSignedValue()) 3242 return new ICmpInst(I.getSwappedPredicate(), X, 3243 ConstantExpr::getNeg(RHSC)); 3244 } 3245 3246 BinaryOperator *SRem = nullptr; 3247 // icmp (srem X, Y), Y 3248 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1)) 3249 SRem = BO0; 3250 // icmp Y, (srem X, Y) 3251 else if (BO1 && BO1->getOpcode() == Instruction::SRem && 3252 Op0 == BO1->getOperand(1)) 3253 SRem = BO1; 3254 if (SRem) { 3255 // We don't check hasOneUse to avoid increasing register pressure because 3256 // the value we use is the same value this instruction was already using. 3257 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) { 3258 default: 3259 break; 3260 case ICmpInst::ICMP_EQ: 3261 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 3262 case ICmpInst::ICMP_NE: 3263 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 3264 case ICmpInst::ICMP_SGT: 3265 case ICmpInst::ICMP_SGE: 3266 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1), 3267 Constant::getAllOnesValue(SRem->getType())); 3268 case ICmpInst::ICMP_SLT: 3269 case ICmpInst::ICMP_SLE: 3270 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1), 3271 Constant::getNullValue(SRem->getType())); 3272 } 3273 } 3274 3275 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() && 3276 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) { 3277 switch (BO0->getOpcode()) { 3278 default: 3279 break; 3280 case Instruction::Add: 3281 case Instruction::Sub: 3282 case Instruction::Xor: { 3283 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b 3284 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0)); 3285 3286 const APInt *C; 3287 if (match(BO0->getOperand(1), m_APInt(C))) { 3288 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b 3289 if (C->isSignMask()) { 3290 ICmpInst::Predicate NewPred = 3291 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate(); 3292 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0)); 3293 } 3294 3295 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b 3296 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) { 3297 ICmpInst::Predicate NewPred = 3298 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate(); 3299 NewPred = I.getSwappedPredicate(NewPred); 3300 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0)); 3301 } 3302 } 3303 break; 3304 } 3305 case Instruction::Mul: { 3306 if (!I.isEquality()) 3307 break; 3308 3309 const APInt *C; 3310 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() && 3311 !C->isOneValue()) { 3312 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask) 3313 // Mask = -1 >> count-trailing-zeros(C). 3314 if (unsigned TZs = C->countTrailingZeros()) { 3315 Constant *Mask = ConstantInt::get( 3316 BO0->getType(), 3317 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs)); 3318 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask); 3319 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask); 3320 return new ICmpInst(Pred, And1, And2); 3321 } 3322 // If there are no trailing zeros in the multiplier, just eliminate 3323 // the multiplies (no masking is needed): 3324 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y 3325 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0)); 3326 } 3327 break; 3328 } 3329 case Instruction::UDiv: 3330 case Instruction::LShr: 3331 if (I.isSigned() || !BO0->isExact() || !BO1->isExact()) 3332 break; 3333 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0)); 3334 3335 case Instruction::SDiv: 3336 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact()) 3337 break; 3338 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0)); 3339 3340 case Instruction::AShr: 3341 if (!BO0->isExact() || !BO1->isExact()) 3342 break; 3343 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0)); 3344 3345 case Instruction::Shl: { 3346 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap(); 3347 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap(); 3348 if (!NUW && !NSW) 3349 break; 3350 if (!NSW && I.isSigned()) 3351 break; 3352 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0)); 3353 } 3354 } 3355 } 3356 3357 if (BO0) { 3358 // Transform A & (L - 1) `ult` L --> L != 0 3359 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes()); 3360 auto BitwiseAnd = m_c_And(m_Value(), LSubOne); 3361 3362 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) { 3363 auto *Zero = Constant::getNullValue(BO0->getType()); 3364 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero); 3365 } 3366 } 3367 3368 if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder)) 3369 return replaceInstUsesWith(I, V); 3370 3371 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder)) 3372 return replaceInstUsesWith(I, V); 3373 3374 return nullptr; 3375 } 3376 3377 /// Fold icmp Pred min|max(X, Y), X. 3378 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) { 3379 ICmpInst::Predicate Pred = Cmp.getPredicate(); 3380 Value *Op0 = Cmp.getOperand(0); 3381 Value *X = Cmp.getOperand(1); 3382 3383 // Canonicalize minimum or maximum operand to LHS of the icmp. 3384 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) || 3385 match(X, m_c_SMax(m_Specific(Op0), m_Value())) || 3386 match(X, m_c_UMin(m_Specific(Op0), m_Value())) || 3387 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) { 3388 std::swap(Op0, X); 3389 Pred = Cmp.getSwappedPredicate(); 3390 } 3391 3392 Value *Y; 3393 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) { 3394 // smin(X, Y) == X --> X s<= Y 3395 // smin(X, Y) s>= X --> X s<= Y 3396 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE) 3397 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y); 3398 3399 // smin(X, Y) != X --> X s> Y 3400 // smin(X, Y) s< X --> X s> Y 3401 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT) 3402 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y); 3403 3404 // These cases should be handled in InstSimplify: 3405 // smin(X, Y) s<= X --> true 3406 // smin(X, Y) s> X --> false 3407 return nullptr; 3408 } 3409 3410 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) { 3411 // smax(X, Y) == X --> X s>= Y 3412 // smax(X, Y) s<= X --> X s>= Y 3413 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE) 3414 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y); 3415 3416 // smax(X, Y) != X --> X s< Y 3417 // smax(X, Y) s> X --> X s< Y 3418 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT) 3419 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y); 3420 3421 // These cases should be handled in InstSimplify: 3422 // smax(X, Y) s>= X --> true 3423 // smax(X, Y) s< X --> false 3424 return nullptr; 3425 } 3426 3427 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) { 3428 // umin(X, Y) == X --> X u<= Y 3429 // umin(X, Y) u>= X --> X u<= Y 3430 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE) 3431 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y); 3432 3433 // umin(X, Y) != X --> X u> Y 3434 // umin(X, Y) u< X --> X u> Y 3435 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT) 3436 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y); 3437 3438 // These cases should be handled in InstSimplify: 3439 // umin(X, Y) u<= X --> true 3440 // umin(X, Y) u> X --> false 3441 return nullptr; 3442 } 3443 3444 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) { 3445 // umax(X, Y) == X --> X u>= Y 3446 // umax(X, Y) u<= X --> X u>= Y 3447 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE) 3448 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y); 3449 3450 // umax(X, Y) != X --> X u< Y 3451 // umax(X, Y) u> X --> X u< Y 3452 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT) 3453 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y); 3454 3455 // These cases should be handled in InstSimplify: 3456 // umax(X, Y) u>= X --> true 3457 // umax(X, Y) u< X --> false 3458 return nullptr; 3459 } 3460 3461 return nullptr; 3462 } 3463 3464 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) { 3465 if (!I.isEquality()) 3466 return nullptr; 3467 3468 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3469 const CmpInst::Predicate Pred = I.getPredicate(); 3470 Value *A, *B, *C, *D; 3471 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 3472 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 3473 Value *OtherVal = A == Op1 ? B : A; 3474 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType())); 3475 } 3476 3477 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) { 3478 // A^c1 == C^c2 --> A == C^(c1^c2) 3479 ConstantInt *C1, *C2; 3480 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) && 3481 Op1->hasOneUse()) { 3482 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue()); 3483 Value *Xor = Builder.CreateXor(C, NC); 3484 return new ICmpInst(Pred, A, Xor); 3485 } 3486 3487 // A^B == A^D -> B == D 3488 if (A == C) 3489 return new ICmpInst(Pred, B, D); 3490 if (A == D) 3491 return new ICmpInst(Pred, B, C); 3492 if (B == C) 3493 return new ICmpInst(Pred, A, D); 3494 if (B == D) 3495 return new ICmpInst(Pred, A, C); 3496 } 3497 } 3498 3499 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) { 3500 // A == (A^B) -> B == 0 3501 Value *OtherVal = A == Op0 ? B : A; 3502 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType())); 3503 } 3504 3505 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 3506 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) && 3507 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) { 3508 Value *X = nullptr, *Y = nullptr, *Z = nullptr; 3509 3510 if (A == C) { 3511 X = B; 3512 Y = D; 3513 Z = A; 3514 } else if (A == D) { 3515 X = B; 3516 Y = C; 3517 Z = A; 3518 } else if (B == C) { 3519 X = A; 3520 Y = D; 3521 Z = B; 3522 } else if (B == D) { 3523 X = A; 3524 Y = C; 3525 Z = B; 3526 } 3527 3528 if (X) { // Build (X^Y) & Z 3529 Op1 = Builder.CreateXor(X, Y); 3530 Op1 = Builder.CreateAnd(Op1, Z); 3531 I.setOperand(0, Op1); 3532 I.setOperand(1, Constant::getNullValue(Op1->getType())); 3533 return &I; 3534 } 3535 } 3536 3537 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B) 3538 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B) 3539 ConstantInt *Cst1; 3540 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) && 3541 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) || 3542 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) && 3543 match(Op1, m_ZExt(m_Value(A))))) { 3544 APInt Pow2 = Cst1->getValue() + 1; 3545 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) && 3546 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth()) 3547 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType())); 3548 } 3549 3550 // (A >> C) == (B >> C) --> (A^B) u< (1 << C) 3551 // For lshr and ashr pairs. 3552 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) && 3553 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) || 3554 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) && 3555 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) { 3556 unsigned TypeBits = Cst1->getBitWidth(); 3557 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits); 3558 if (ShAmt < TypeBits && ShAmt != 0) { 3559 ICmpInst::Predicate NewPred = 3560 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; 3561 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted"); 3562 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt); 3563 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal)); 3564 } 3565 } 3566 3567 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0 3568 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) && 3569 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) { 3570 unsigned TypeBits = Cst1->getBitWidth(); 3571 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits); 3572 if (ShAmt < TypeBits && ShAmt != 0) { 3573 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted"); 3574 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt); 3575 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal), 3576 I.getName() + ".mask"); 3577 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType())); 3578 } 3579 } 3580 3581 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to 3582 // "icmp (and X, mask), cst" 3583 uint64_t ShAmt = 0; 3584 if (Op0->hasOneUse() && 3585 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) && 3586 match(Op1, m_ConstantInt(Cst1)) && 3587 // Only do this when A has multiple uses. This is most important to do 3588 // when it exposes other optimizations. 3589 !A->hasOneUse()) { 3590 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits(); 3591 3592 if (ShAmt < ASize) { 3593 APInt MaskV = 3594 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits()); 3595 MaskV <<= ShAmt; 3596 3597 APInt CmpV = Cst1->getValue().zext(ASize); 3598 CmpV <<= ShAmt; 3599 3600 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV)); 3601 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV)); 3602 } 3603 } 3604 3605 // If both operands are byte-swapped or bit-reversed, just compare the 3606 // original values. 3607 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant() 3608 // and handle more intrinsics. 3609 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) || 3610 (match(Op0, m_BitReverse(m_Value(A))) && 3611 match(Op1, m_BitReverse(m_Value(B))))) 3612 return new ICmpInst(Pred, A, B); 3613 3614 return nullptr; 3615 } 3616 3617 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so 3618 /// far. 3619 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) { 3620 const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0)); 3621 Value *LHSCIOp = LHSCI->getOperand(0); 3622 Type *SrcTy = LHSCIOp->getType(); 3623 Type *DestTy = LHSCI->getType(); 3624 Value *RHSCIOp; 3625 3626 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the 3627 // integer type is the same size as the pointer type. 3628 const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool { 3629 if (isa<VectorType>(SrcTy)) { 3630 SrcTy = cast<VectorType>(SrcTy)->getElementType(); 3631 DestTy = cast<VectorType>(DestTy)->getElementType(); 3632 } 3633 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth(); 3634 }; 3635 if (LHSCI->getOpcode() == Instruction::PtrToInt && 3636 CompatibleSizes(SrcTy, DestTy)) { 3637 Value *RHSOp = nullptr; 3638 if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) { 3639 Value *RHSCIOp = RHSC->getOperand(0); 3640 if (RHSCIOp->getType()->getPointerAddressSpace() == 3641 LHSCIOp->getType()->getPointerAddressSpace()) { 3642 RHSOp = RHSC->getOperand(0); 3643 // If the pointer types don't match, insert a bitcast. 3644 if (LHSCIOp->getType() != RHSOp->getType()) 3645 RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType()); 3646 } 3647 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) { 3648 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); 3649 } 3650 3651 if (RHSOp) 3652 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp); 3653 } 3654 3655 // The code below only handles extension cast instructions, so far. 3656 // Enforce this. 3657 if (LHSCI->getOpcode() != Instruction::ZExt && 3658 LHSCI->getOpcode() != Instruction::SExt) 3659 return nullptr; 3660 3661 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; 3662 bool isSignedCmp = ICmp.isSigned(); 3663 3664 if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) { 3665 // Not an extension from the same type? 3666 RHSCIOp = CI->getOperand(0); 3667 if (RHSCIOp->getType() != LHSCIOp->getType()) 3668 return nullptr; 3669 3670 // If the signedness of the two casts doesn't agree (i.e. one is a sext 3671 // and the other is a zext), then we can't handle this. 3672 if (CI->getOpcode() != LHSCI->getOpcode()) 3673 return nullptr; 3674 3675 // Deal with equality cases early. 3676 if (ICmp.isEquality()) 3677 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp); 3678 3679 // A signed comparison of sign extended values simplifies into a 3680 // signed comparison. 3681 if (isSignedCmp && isSignedExt) 3682 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp); 3683 3684 // The other three cases all fold into an unsigned comparison. 3685 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp); 3686 } 3687 3688 // If we aren't dealing with a constant on the RHS, exit early. 3689 auto *C = dyn_cast<Constant>(ICmp.getOperand(1)); 3690 if (!C) 3691 return nullptr; 3692 3693 // Compute the constant that would happen if we truncated to SrcTy then 3694 // re-extended to DestTy. 3695 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy); 3696 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy); 3697 3698 // If the re-extended constant didn't change... 3699 if (Res2 == C) { 3700 // Deal with equality cases early. 3701 if (ICmp.isEquality()) 3702 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1); 3703 3704 // A signed comparison of sign extended values simplifies into a 3705 // signed comparison. 3706 if (isSignedExt && isSignedCmp) 3707 return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1); 3708 3709 // The other three cases all fold into an unsigned comparison. 3710 return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1); 3711 } 3712 3713 // The re-extended constant changed, partly changed (in the case of a vector), 3714 // or could not be determined to be equal (in the case of a constant 3715 // expression), so the constant cannot be represented in the shorter type. 3716 // Consequently, we cannot emit a simple comparison. 3717 // All the cases that fold to true or false will have already been handled 3718 // by SimplifyICmpInst, so only deal with the tricky case. 3719 3720 if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C)) 3721 return nullptr; 3722 3723 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases 3724 // should have been folded away previously and not enter in here. 3725 3726 // We're performing an unsigned comp with a sign extended value. 3727 // This is true if the input is >= 0. [aka >s -1] 3728 Constant *NegOne = Constant::getAllOnesValue(SrcTy); 3729 Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName()); 3730 3731 // Finally, return the value computed. 3732 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT) 3733 return replaceInstUsesWith(ICmp, Result); 3734 3735 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); 3736 return BinaryOperator::CreateNot(Result); 3737 } 3738 3739 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS, 3740 Value *RHS, Instruction &OrigI, 3741 Value *&Result, Constant *&Overflow) { 3742 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS)) 3743 std::swap(LHS, RHS); 3744 3745 auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) { 3746 Result = OpResult; 3747 Overflow = OverflowVal; 3748 if (ReuseName) 3749 Result->takeName(&OrigI); 3750 return true; 3751 }; 3752 3753 // If the overflow check was an add followed by a compare, the insertion point 3754 // may be pointing to the compare. We want to insert the new instructions 3755 // before the add in case there are uses of the add between the add and the 3756 // compare. 3757 Builder.SetInsertPoint(&OrigI); 3758 3759 switch (OCF) { 3760 case OCF_INVALID: 3761 llvm_unreachable("bad overflow check kind!"); 3762 3763 case OCF_UNSIGNED_ADD: { 3764 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI); 3765 if (OR == OverflowResult::NeverOverflows) 3766 return SetResult(Builder.CreateNUWAdd(LHS, RHS), Builder.getFalse(), 3767 true); 3768 3769 if (OR == OverflowResult::AlwaysOverflows) 3770 return SetResult(Builder.CreateAdd(LHS, RHS), Builder.getTrue(), true); 3771 3772 // Fall through uadd into sadd 3773 LLVM_FALLTHROUGH; 3774 } 3775 case OCF_SIGNED_ADD: { 3776 // X + 0 -> {X, false} 3777 if (match(RHS, m_Zero())) 3778 return SetResult(LHS, Builder.getFalse(), false); 3779 3780 // We can strength reduce this signed add into a regular add if we can prove 3781 // that it will never overflow. 3782 if (OCF == OCF_SIGNED_ADD) 3783 if (willNotOverflowSignedAdd(LHS, RHS, OrigI)) 3784 return SetResult(Builder.CreateNSWAdd(LHS, RHS), Builder.getFalse(), 3785 true); 3786 break; 3787 } 3788 3789 case OCF_UNSIGNED_SUB: 3790 case OCF_SIGNED_SUB: { 3791 // X - 0 -> {X, false} 3792 if (match(RHS, m_Zero())) 3793 return SetResult(LHS, Builder.getFalse(), false); 3794 3795 if (OCF == OCF_SIGNED_SUB) { 3796 if (willNotOverflowSignedSub(LHS, RHS, OrigI)) 3797 return SetResult(Builder.CreateNSWSub(LHS, RHS), Builder.getFalse(), 3798 true); 3799 } else { 3800 if (willNotOverflowUnsignedSub(LHS, RHS, OrigI)) 3801 return SetResult(Builder.CreateNUWSub(LHS, RHS), Builder.getFalse(), 3802 true); 3803 } 3804 break; 3805 } 3806 3807 case OCF_UNSIGNED_MUL: { 3808 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI); 3809 if (OR == OverflowResult::NeverOverflows) 3810 return SetResult(Builder.CreateNUWMul(LHS, RHS), Builder.getFalse(), 3811 true); 3812 if (OR == OverflowResult::AlwaysOverflows) 3813 return SetResult(Builder.CreateMul(LHS, RHS), Builder.getTrue(), true); 3814 LLVM_FALLTHROUGH; 3815 } 3816 case OCF_SIGNED_MUL: 3817 // X * undef -> undef 3818 if (isa<UndefValue>(RHS)) 3819 return SetResult(RHS, UndefValue::get(Builder.getInt1Ty()), false); 3820 3821 // X * 0 -> {0, false} 3822 if (match(RHS, m_Zero())) 3823 return SetResult(RHS, Builder.getFalse(), false); 3824 3825 // X * 1 -> {X, false} 3826 if (match(RHS, m_One())) 3827 return SetResult(LHS, Builder.getFalse(), false); 3828 3829 if (OCF == OCF_SIGNED_MUL) 3830 if (willNotOverflowSignedMul(LHS, RHS, OrigI)) 3831 return SetResult(Builder.CreateNSWMul(LHS, RHS), Builder.getFalse(), 3832 true); 3833 break; 3834 } 3835 3836 return false; 3837 } 3838 3839 /// Recognize and process idiom involving test for multiplication 3840 /// overflow. 3841 /// 3842 /// The caller has matched a pattern of the form: 3843 /// I = cmp u (mul(zext A, zext B), V 3844 /// The function checks if this is a test for overflow and if so replaces 3845 /// multiplication with call to 'mul.with.overflow' intrinsic. 3846 /// 3847 /// \param I Compare instruction. 3848 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of 3849 /// the compare instruction. Must be of integer type. 3850 /// \param OtherVal The other argument of compare instruction. 3851 /// \returns Instruction which must replace the compare instruction, NULL if no 3852 /// replacement required. 3853 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal, 3854 Value *OtherVal, InstCombiner &IC) { 3855 // Don't bother doing this transformation for pointers, don't do it for 3856 // vectors. 3857 if (!isa<IntegerType>(MulVal->getType())) 3858 return nullptr; 3859 3860 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal); 3861 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal); 3862 auto *MulInstr = dyn_cast<Instruction>(MulVal); 3863 if (!MulInstr) 3864 return nullptr; 3865 assert(MulInstr->getOpcode() == Instruction::Mul); 3866 3867 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)), 3868 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1)); 3869 assert(LHS->getOpcode() == Instruction::ZExt); 3870 assert(RHS->getOpcode() == Instruction::ZExt); 3871 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0); 3872 3873 // Calculate type and width of the result produced by mul.with.overflow. 3874 Type *TyA = A->getType(), *TyB = B->getType(); 3875 unsigned WidthA = TyA->getPrimitiveSizeInBits(), 3876 WidthB = TyB->getPrimitiveSizeInBits(); 3877 unsigned MulWidth; 3878 Type *MulType; 3879 if (WidthB > WidthA) { 3880 MulWidth = WidthB; 3881 MulType = TyB; 3882 } else { 3883 MulWidth = WidthA; 3884 MulType = TyA; 3885 } 3886 3887 // In order to replace the original mul with a narrower mul.with.overflow, 3888 // all uses must ignore upper bits of the product. The number of used low 3889 // bits must be not greater than the width of mul.with.overflow. 3890 if (MulVal->hasNUsesOrMore(2)) 3891 for (User *U : MulVal->users()) { 3892 if (U == &I) 3893 continue; 3894 if (TruncInst *TI = dyn_cast<TruncInst>(U)) { 3895 // Check if truncation ignores bits above MulWidth. 3896 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits(); 3897 if (TruncWidth > MulWidth) 3898 return nullptr; 3899 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) { 3900 // Check if AND ignores bits above MulWidth. 3901 if (BO->getOpcode() != Instruction::And) 3902 return nullptr; 3903 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 3904 const APInt &CVal = CI->getValue(); 3905 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth) 3906 return nullptr; 3907 } else { 3908 // In this case we could have the operand of the binary operation 3909 // being defined in another block, and performing the replacement 3910 // could break the dominance relation. 3911 return nullptr; 3912 } 3913 } else { 3914 // Other uses prohibit this transformation. 3915 return nullptr; 3916 } 3917 } 3918 3919 // Recognize patterns 3920 switch (I.getPredicate()) { 3921 case ICmpInst::ICMP_EQ: 3922 case ICmpInst::ICMP_NE: 3923 // Recognize pattern: 3924 // mulval = mul(zext A, zext B) 3925 // cmp eq/neq mulval, zext trunc mulval 3926 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal)) 3927 if (Zext->hasOneUse()) { 3928 Value *ZextArg = Zext->getOperand(0); 3929 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg)) 3930 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth) 3931 break; //Recognized 3932 } 3933 3934 // Recognize pattern: 3935 // mulval = mul(zext A, zext B) 3936 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits. 3937 ConstantInt *CI; 3938 Value *ValToMask; 3939 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) { 3940 if (ValToMask != MulVal) 3941 return nullptr; 3942 const APInt &CVal = CI->getValue() + 1; 3943 if (CVal.isPowerOf2()) { 3944 unsigned MaskWidth = CVal.logBase2(); 3945 if (MaskWidth == MulWidth) 3946 break; // Recognized 3947 } 3948 } 3949 return nullptr; 3950 3951 case ICmpInst::ICMP_UGT: 3952 // Recognize pattern: 3953 // mulval = mul(zext A, zext B) 3954 // cmp ugt mulval, max 3955 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) { 3956 APInt MaxVal = APInt::getMaxValue(MulWidth); 3957 MaxVal = MaxVal.zext(CI->getBitWidth()); 3958 if (MaxVal.eq(CI->getValue())) 3959 break; // Recognized 3960 } 3961 return nullptr; 3962 3963 case ICmpInst::ICMP_UGE: 3964 // Recognize pattern: 3965 // mulval = mul(zext A, zext B) 3966 // cmp uge mulval, max+1 3967 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) { 3968 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth); 3969 if (MaxVal.eq(CI->getValue())) 3970 break; // Recognized 3971 } 3972 return nullptr; 3973 3974 case ICmpInst::ICMP_ULE: 3975 // Recognize pattern: 3976 // mulval = mul(zext A, zext B) 3977 // cmp ule mulval, max 3978 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) { 3979 APInt MaxVal = APInt::getMaxValue(MulWidth); 3980 MaxVal = MaxVal.zext(CI->getBitWidth()); 3981 if (MaxVal.eq(CI->getValue())) 3982 break; // Recognized 3983 } 3984 return nullptr; 3985 3986 case ICmpInst::ICMP_ULT: 3987 // Recognize pattern: 3988 // mulval = mul(zext A, zext B) 3989 // cmp ule mulval, max + 1 3990 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) { 3991 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth); 3992 if (MaxVal.eq(CI->getValue())) 3993 break; // Recognized 3994 } 3995 return nullptr; 3996 3997 default: 3998 return nullptr; 3999 } 4000 4001 InstCombiner::BuilderTy &Builder = IC.Builder; 4002 Builder.SetInsertPoint(MulInstr); 4003 4004 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B) 4005 Value *MulA = A, *MulB = B; 4006 if (WidthA < MulWidth) 4007 MulA = Builder.CreateZExt(A, MulType); 4008 if (WidthB < MulWidth) 4009 MulB = Builder.CreateZExt(B, MulType); 4010 Value *F = Intrinsic::getDeclaration(I.getModule(), 4011 Intrinsic::umul_with_overflow, MulType); 4012 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul"); 4013 IC.Worklist.Add(MulInstr); 4014 4015 // If there are uses of mul result other than the comparison, we know that 4016 // they are truncation or binary AND. Change them to use result of 4017 // mul.with.overflow and adjust properly mask/size. 4018 if (MulVal->hasNUsesOrMore(2)) { 4019 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value"); 4020 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) { 4021 User *U = *UI++; 4022 if (U == &I || U == OtherVal) 4023 continue; 4024 if (TruncInst *TI = dyn_cast<TruncInst>(U)) { 4025 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth) 4026 IC.replaceInstUsesWith(*TI, Mul); 4027 else 4028 TI->setOperand(0, Mul); 4029 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) { 4030 assert(BO->getOpcode() == Instruction::And); 4031 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask) 4032 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1)); 4033 APInt ShortMask = CI->getValue().trunc(MulWidth); 4034 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask); 4035 Instruction *Zext = 4036 cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType())); 4037 IC.Worklist.Add(Zext); 4038 IC.replaceInstUsesWith(*BO, Zext); 4039 } else { 4040 llvm_unreachable("Unexpected Binary operation"); 4041 } 4042 IC.Worklist.Add(cast<Instruction>(U)); 4043 } 4044 } 4045 if (isa<Instruction>(OtherVal)) 4046 IC.Worklist.Add(cast<Instruction>(OtherVal)); 4047 4048 // The original icmp gets replaced with the overflow value, maybe inverted 4049 // depending on predicate. 4050 bool Inverse = false; 4051 switch (I.getPredicate()) { 4052 case ICmpInst::ICMP_NE: 4053 break; 4054 case ICmpInst::ICMP_EQ: 4055 Inverse = true; 4056 break; 4057 case ICmpInst::ICMP_UGT: 4058 case ICmpInst::ICMP_UGE: 4059 if (I.getOperand(0) == MulVal) 4060 break; 4061 Inverse = true; 4062 break; 4063 case ICmpInst::ICMP_ULT: 4064 case ICmpInst::ICMP_ULE: 4065 if (I.getOperand(1) == MulVal) 4066 break; 4067 Inverse = true; 4068 break; 4069 default: 4070 llvm_unreachable("Unexpected predicate"); 4071 } 4072 if (Inverse) { 4073 Value *Res = Builder.CreateExtractValue(Call, 1); 4074 return BinaryOperator::CreateNot(Res); 4075 } 4076 4077 return ExtractValueInst::Create(Call, 1); 4078 } 4079 4080 /// When performing a comparison against a constant, it is possible that not all 4081 /// the bits in the LHS are demanded. This helper method computes the mask that 4082 /// IS demanded. 4083 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) { 4084 const APInt *RHS; 4085 if (!match(I.getOperand(1), m_APInt(RHS))) 4086 return APInt::getAllOnesValue(BitWidth); 4087 4088 // If this is a normal comparison, it demands all bits. If it is a sign bit 4089 // comparison, it only demands the sign bit. 4090 bool UnusedBit; 4091 if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit)) 4092 return APInt::getSignMask(BitWidth); 4093 4094 switch (I.getPredicate()) { 4095 // For a UGT comparison, we don't care about any bits that 4096 // correspond to the trailing ones of the comparand. The value of these 4097 // bits doesn't impact the outcome of the comparison, because any value 4098 // greater than the RHS must differ in a bit higher than these due to carry. 4099 case ICmpInst::ICMP_UGT: 4100 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes()); 4101 4102 // Similarly, for a ULT comparison, we don't care about the trailing zeros. 4103 // Any value less than the RHS must differ in a higher bit because of carries. 4104 case ICmpInst::ICMP_ULT: 4105 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros()); 4106 4107 default: 4108 return APInt::getAllOnesValue(BitWidth); 4109 } 4110 } 4111 4112 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst 4113 /// should be swapped. 4114 /// The decision is based on how many times these two operands are reused 4115 /// as subtract operands and their positions in those instructions. 4116 /// The rationale is that several architectures use the same instruction for 4117 /// both subtract and cmp. Thus, it is better if the order of those operands 4118 /// match. 4119 /// \return true if Op0 and Op1 should be swapped. 4120 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) { 4121 // Filter out pointer values as those cannot appear directly in subtract. 4122 // FIXME: we may want to go through inttoptrs or bitcasts. 4123 if (Op0->getType()->isPointerTy()) 4124 return false; 4125 // If a subtract already has the same operands as a compare, swapping would be 4126 // bad. If a subtract has the same operands as a compare but in reverse order, 4127 // then swapping is good. 4128 int GoodToSwap = 0; 4129 for (const User *U : Op0->users()) { 4130 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0)))) 4131 GoodToSwap++; 4132 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1)))) 4133 GoodToSwap--; 4134 } 4135 return GoodToSwap > 0; 4136 } 4137 4138 /// Check that one use is in the same block as the definition and all 4139 /// other uses are in blocks dominated by a given block. 4140 /// 4141 /// \param DI Definition 4142 /// \param UI Use 4143 /// \param DB Block that must dominate all uses of \p DI outside 4144 /// the parent block 4145 /// \return true when \p UI is the only use of \p DI in the parent block 4146 /// and all other uses of \p DI are in blocks dominated by \p DB. 4147 /// 4148 bool InstCombiner::dominatesAllUses(const Instruction *DI, 4149 const Instruction *UI, 4150 const BasicBlock *DB) const { 4151 assert(DI && UI && "Instruction not defined\n"); 4152 // Ignore incomplete definitions. 4153 if (!DI->getParent()) 4154 return false; 4155 // DI and UI must be in the same block. 4156 if (DI->getParent() != UI->getParent()) 4157 return false; 4158 // Protect from self-referencing blocks. 4159 if (DI->getParent() == DB) 4160 return false; 4161 for (const User *U : DI->users()) { 4162 auto *Usr = cast<Instruction>(U); 4163 if (Usr != UI && !DT.dominates(DB, Usr->getParent())) 4164 return false; 4165 } 4166 return true; 4167 } 4168 4169 /// Return true when the instruction sequence within a block is select-cmp-br. 4170 static bool isChainSelectCmpBranch(const SelectInst *SI) { 4171 const BasicBlock *BB = SI->getParent(); 4172 if (!BB) 4173 return false; 4174 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator()); 4175 if (!BI || BI->getNumSuccessors() != 2) 4176 return false; 4177 auto *IC = dyn_cast<ICmpInst>(BI->getCondition()); 4178 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI)) 4179 return false; 4180 return true; 4181 } 4182 4183 /// True when a select result is replaced by one of its operands 4184 /// in select-icmp sequence. This will eventually result in the elimination 4185 /// of the select. 4186 /// 4187 /// \param SI Select instruction 4188 /// \param Icmp Compare instruction 4189 /// \param SIOpd Operand that replaces the select 4190 /// 4191 /// Notes: 4192 /// - The replacement is global and requires dominator information 4193 /// - The caller is responsible for the actual replacement 4194 /// 4195 /// Example: 4196 /// 4197 /// entry: 4198 /// %4 = select i1 %3, %C* %0, %C* null 4199 /// %5 = icmp eq %C* %4, null 4200 /// br i1 %5, label %9, label %7 4201 /// ... 4202 /// ; <label>:7 ; preds = %entry 4203 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0 4204 /// ... 4205 /// 4206 /// can be transformed to 4207 /// 4208 /// %5 = icmp eq %C* %0, null 4209 /// %6 = select i1 %3, i1 %5, i1 true 4210 /// br i1 %6, label %9, label %7 4211 /// ... 4212 /// ; <label>:7 ; preds = %entry 4213 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0! 4214 /// 4215 /// Similar when the first operand of the select is a constant or/and 4216 /// the compare is for not equal rather than equal. 4217 /// 4218 /// NOTE: The function is only called when the select and compare constants 4219 /// are equal, the optimization can work only for EQ predicates. This is not a 4220 /// major restriction since a NE compare should be 'normalized' to an equal 4221 /// compare, which usually happens in the combiner and test case 4222 /// select-cmp-br.ll checks for it. 4223 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI, 4224 const ICmpInst *Icmp, 4225 const unsigned SIOpd) { 4226 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!"); 4227 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) { 4228 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1); 4229 // The check for the single predecessor is not the best that can be 4230 // done. But it protects efficiently against cases like when SI's 4231 // home block has two successors, Succ and Succ1, and Succ1 predecessor 4232 // of Succ. Then SI can't be replaced by SIOpd because the use that gets 4233 // replaced can be reached on either path. So the uniqueness check 4234 // guarantees that the path all uses of SI (outside SI's parent) are on 4235 // is disjoint from all other paths out of SI. But that information 4236 // is more expensive to compute, and the trade-off here is in favor 4237 // of compile-time. It should also be noticed that we check for a single 4238 // predecessor and not only uniqueness. This to handle the situation when 4239 // Succ and Succ1 points to the same basic block. 4240 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) { 4241 NumSel++; 4242 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent()); 4243 return true; 4244 } 4245 } 4246 return false; 4247 } 4248 4249 /// Try to fold the comparison based on range information we can get by checking 4250 /// whether bits are known to be zero or one in the inputs. 4251 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) { 4252 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 4253 Type *Ty = Op0->getType(); 4254 ICmpInst::Predicate Pred = I.getPredicate(); 4255 4256 // Get scalar or pointer size. 4257 unsigned BitWidth = Ty->isIntOrIntVectorTy() 4258 ? Ty->getScalarSizeInBits() 4259 : DL.getIndexTypeSizeInBits(Ty->getScalarType()); 4260 4261 if (!BitWidth) 4262 return nullptr; 4263 4264 KnownBits Op0Known(BitWidth); 4265 KnownBits Op1Known(BitWidth); 4266 4267 if (SimplifyDemandedBits(&I, 0, 4268 getDemandedBitsLHSMask(I, BitWidth), 4269 Op0Known, 0)) 4270 return &I; 4271 4272 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth), 4273 Op1Known, 0)) 4274 return &I; 4275 4276 // Given the known and unknown bits, compute a range that the LHS could be 4277 // in. Compute the Min, Max and RHS values based on the known bits. For the 4278 // EQ and NE we use unsigned values. 4279 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); 4280 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); 4281 if (I.isSigned()) { 4282 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max); 4283 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max); 4284 } else { 4285 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max); 4286 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max); 4287 } 4288 4289 // If Min and Max are known to be the same, then SimplifyDemandedBits figured 4290 // out that the LHS or RHS is a constant. Constant fold this now, so that 4291 // code below can assume that Min != Max. 4292 if (!isa<Constant>(Op0) && Op0Min == Op0Max) 4293 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1); 4294 if (!isa<Constant>(Op1) && Op1Min == Op1Max) 4295 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min)); 4296 4297 // Based on the range information we know about the LHS, see if we can 4298 // simplify this comparison. For example, (x&4) < 8 is always true. 4299 switch (Pred) { 4300 default: 4301 llvm_unreachable("Unknown icmp opcode!"); 4302 case ICmpInst::ICMP_EQ: 4303 case ICmpInst::ICMP_NE: { 4304 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) { 4305 return Pred == CmpInst::ICMP_EQ 4306 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())) 4307 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4308 } 4309 4310 // If all bits are known zero except for one, then we know at most one bit 4311 // is set. If the comparison is against zero, then this is a check to see if 4312 // *that* bit is set. 4313 APInt Op0KnownZeroInverted = ~Op0Known.Zero; 4314 if (Op1Known.isZero()) { 4315 // If the LHS is an AND with the same constant, look through it. 4316 Value *LHS = nullptr; 4317 const APInt *LHSC; 4318 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) || 4319 *LHSC != Op0KnownZeroInverted) 4320 LHS = Op0; 4321 4322 Value *X; 4323 if (match(LHS, m_Shl(m_One(), m_Value(X)))) { 4324 APInt ValToCheck = Op0KnownZeroInverted; 4325 Type *XTy = X->getType(); 4326 if (ValToCheck.isPowerOf2()) { 4327 // ((1 << X) & 8) == 0 -> X != 3 4328 // ((1 << X) & 8) != 0 -> X == 3 4329 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros()); 4330 auto NewPred = ICmpInst::getInversePredicate(Pred); 4331 return new ICmpInst(NewPred, X, CmpC); 4332 } else if ((++ValToCheck).isPowerOf2()) { 4333 // ((1 << X) & 7) == 0 -> X >= 3 4334 // ((1 << X) & 7) != 0 -> X < 3 4335 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros()); 4336 auto NewPred = 4337 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT; 4338 return new ICmpInst(NewPred, X, CmpC); 4339 } 4340 } 4341 4342 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1. 4343 const APInt *CI; 4344 if (Op0KnownZeroInverted.isOneValue() && 4345 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) { 4346 // ((8 >>u X) & 1) == 0 -> X != 3 4347 // ((8 >>u X) & 1) != 0 -> X == 3 4348 unsigned CmpVal = CI->countTrailingZeros(); 4349 auto NewPred = ICmpInst::getInversePredicate(Pred); 4350 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal)); 4351 } 4352 } 4353 break; 4354 } 4355 case ICmpInst::ICMP_ULT: { 4356 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B) 4357 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4358 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B) 4359 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4360 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) 4361 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 4362 4363 const APInt *CmpC; 4364 if (match(Op1, m_APInt(CmpC))) { 4365 // A <u C -> A == C-1 if min(A)+1 == C 4366 if (*CmpC == Op0Min + 1) 4367 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 4368 ConstantInt::get(Op1->getType(), *CmpC - 1)); 4369 // X <u C --> X == 0, if the number of zero bits in the bottom of X 4370 // exceeds the log2 of C. 4371 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2()) 4372 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 4373 Constant::getNullValue(Op1->getType())); 4374 } 4375 break; 4376 } 4377 case ICmpInst::ICMP_UGT: { 4378 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B) 4379 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4380 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B) 4381 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4382 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) 4383 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 4384 4385 const APInt *CmpC; 4386 if (match(Op1, m_APInt(CmpC))) { 4387 // A >u C -> A == C+1 if max(a)-1 == C 4388 if (*CmpC == Op0Max - 1) 4389 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 4390 ConstantInt::get(Op1->getType(), *CmpC + 1)); 4391 // X >u C --> X != 0, if the number of zero bits in the bottom of X 4392 // exceeds the log2 of C. 4393 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits()) 4394 return new ICmpInst(ICmpInst::ICMP_NE, Op0, 4395 Constant::getNullValue(Op1->getType())); 4396 } 4397 break; 4398 } 4399 case ICmpInst::ICMP_SLT: { 4400 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C) 4401 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4402 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C) 4403 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4404 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) 4405 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 4406 const APInt *CmpC; 4407 if (match(Op1, m_APInt(CmpC))) { 4408 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C 4409 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 4410 ConstantInt::get(Op1->getType(), *CmpC - 1)); 4411 } 4412 break; 4413 } 4414 case ICmpInst::ICMP_SGT: { 4415 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B) 4416 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4417 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B) 4418 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4419 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) 4420 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); 4421 const APInt *CmpC; 4422 if (match(Op1, m_APInt(CmpC))) { 4423 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C 4424 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, 4425 ConstantInt::get(Op1->getType(), *CmpC + 1)); 4426 } 4427 break; 4428 } 4429 case ICmpInst::ICMP_SGE: 4430 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"); 4431 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B) 4432 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4433 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B) 4434 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4435 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B) 4436 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); 4437 break; 4438 case ICmpInst::ICMP_SLE: 4439 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"); 4440 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B) 4441 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4442 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B) 4443 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4444 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B) 4445 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); 4446 break; 4447 case ICmpInst::ICMP_UGE: 4448 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"); 4449 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B) 4450 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4451 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B) 4452 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4453 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B) 4454 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); 4455 break; 4456 case ICmpInst::ICMP_ULE: 4457 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"); 4458 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B) 4459 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType())); 4460 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B) 4461 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType())); 4462 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B) 4463 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); 4464 break; 4465 } 4466 4467 // Turn a signed comparison into an unsigned one if both operands are known to 4468 // have the same sign. 4469 if (I.isSigned() && 4470 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) || 4471 (Op0Known.One.isNegative() && Op1Known.One.isNegative()))) 4472 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1); 4473 4474 return nullptr; 4475 } 4476 4477 /// If we have an icmp le or icmp ge instruction with a constant operand, turn 4478 /// it into the appropriate icmp lt or icmp gt instruction. This transform 4479 /// allows them to be folded in visitICmpInst. 4480 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) { 4481 ICmpInst::Predicate Pred = I.getPredicate(); 4482 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE && 4483 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE) 4484 return nullptr; 4485 4486 Value *Op0 = I.getOperand(0); 4487 Value *Op1 = I.getOperand(1); 4488 auto *Op1C = dyn_cast<Constant>(Op1); 4489 if (!Op1C) 4490 return nullptr; 4491 4492 // Check if the constant operand can be safely incremented/decremented without 4493 // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled 4494 // the edge cases for us, so we just assert on them. For vectors, we must 4495 // handle the edge cases. 4496 Type *Op1Type = Op1->getType(); 4497 bool IsSigned = I.isSigned(); 4498 bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE); 4499 auto *CI = dyn_cast<ConstantInt>(Op1C); 4500 if (CI) { 4501 // A <= MAX -> TRUE ; A >= MIN -> TRUE 4502 assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned)); 4503 } else if (Op1Type->isVectorTy()) { 4504 // TODO? If the edge cases for vectors were guaranteed to be handled as they 4505 // are for scalar, we could remove the min/max checks. However, to do that, 4506 // we would have to use insertelement/shufflevector to replace edge values. 4507 unsigned NumElts = Op1Type->getVectorNumElements(); 4508 for (unsigned i = 0; i != NumElts; ++i) { 4509 Constant *Elt = Op1C->getAggregateElement(i); 4510 if (!Elt) 4511 return nullptr; 4512 4513 if (isa<UndefValue>(Elt)) 4514 continue; 4515 4516 // Bail out if we can't determine if this constant is min/max or if we 4517 // know that this constant is min/max. 4518 auto *CI = dyn_cast<ConstantInt>(Elt); 4519 if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned))) 4520 return nullptr; 4521 } 4522 } else { 4523 // ConstantExpr? 4524 return nullptr; 4525 } 4526 4527 // Increment or decrement the constant and set the new comparison predicate: 4528 // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT 4529 Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true); 4530 CmpInst::Predicate NewPred = IsLE ? ICmpInst::ICMP_ULT: ICmpInst::ICMP_UGT; 4531 NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred; 4532 return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne)); 4533 } 4534 4535 /// Integer compare with boolean values can always be turned into bitwise ops. 4536 static Instruction *canonicalizeICmpBool(ICmpInst &I, 4537 InstCombiner::BuilderTy &Builder) { 4538 Value *A = I.getOperand(0), *B = I.getOperand(1); 4539 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only"); 4540 4541 // A boolean compared to true/false can be simplified to Op0/true/false in 4542 // 14 out of the 20 (10 predicates * 2 constants) possible combinations. 4543 // Cases not handled by InstSimplify are always 'not' of Op0. 4544 if (match(B, m_Zero())) { 4545 switch (I.getPredicate()) { 4546 case CmpInst::ICMP_EQ: // A == 0 -> !A 4547 case CmpInst::ICMP_ULE: // A <=u 0 -> !A 4548 case CmpInst::ICMP_SGE: // A >=s 0 -> !A 4549 return BinaryOperator::CreateNot(A); 4550 default: 4551 llvm_unreachable("ICmp i1 X, C not simplified as expected."); 4552 } 4553 } else if (match(B, m_One())) { 4554 switch (I.getPredicate()) { 4555 case CmpInst::ICMP_NE: // A != 1 -> !A 4556 case CmpInst::ICMP_ULT: // A <u 1 -> !A 4557 case CmpInst::ICMP_SGT: // A >s -1 -> !A 4558 return BinaryOperator::CreateNot(A); 4559 default: 4560 llvm_unreachable("ICmp i1 X, C not simplified as expected."); 4561 } 4562 } 4563 4564 switch (I.getPredicate()) { 4565 default: 4566 llvm_unreachable("Invalid icmp instruction!"); 4567 case ICmpInst::ICMP_EQ: 4568 // icmp eq i1 A, B -> ~(A ^ B) 4569 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 4570 4571 case ICmpInst::ICMP_NE: 4572 // icmp ne i1 A, B -> A ^ B 4573 return BinaryOperator::CreateXor(A, B); 4574 4575 case ICmpInst::ICMP_UGT: 4576 // icmp ugt -> icmp ult 4577 std::swap(A, B); 4578 LLVM_FALLTHROUGH; 4579 case ICmpInst::ICMP_ULT: 4580 // icmp ult i1 A, B -> ~A & B 4581 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B); 4582 4583 case ICmpInst::ICMP_SGT: 4584 // icmp sgt -> icmp slt 4585 std::swap(A, B); 4586 LLVM_FALLTHROUGH; 4587 case ICmpInst::ICMP_SLT: 4588 // icmp slt i1 A, B -> A & ~B 4589 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A); 4590 4591 case ICmpInst::ICMP_UGE: 4592 // icmp uge -> icmp ule 4593 std::swap(A, B); 4594 LLVM_FALLTHROUGH; 4595 case ICmpInst::ICMP_ULE: 4596 // icmp ule i1 A, B -> ~A | B 4597 return BinaryOperator::CreateOr(Builder.CreateNot(A), B); 4598 4599 case ICmpInst::ICMP_SGE: 4600 // icmp sge -> icmp sle 4601 std::swap(A, B); 4602 LLVM_FALLTHROUGH; 4603 case ICmpInst::ICMP_SLE: 4604 // icmp sle i1 A, B -> A | ~B 4605 return BinaryOperator::CreateOr(Builder.CreateNot(B), A); 4606 } 4607 } 4608 4609 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) { 4610 bool Changed = false; 4611 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 4612 unsigned Op0Cplxity = getComplexity(Op0); 4613 unsigned Op1Cplxity = getComplexity(Op1); 4614 4615 /// Orders the operands of the compare so that they are listed from most 4616 /// complex to least complex. This puts constants before unary operators, 4617 /// before binary operators. 4618 if (Op0Cplxity < Op1Cplxity || 4619 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) { 4620 I.swapOperands(); 4621 std::swap(Op0, Op1); 4622 Changed = true; 4623 } 4624 4625 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, 4626 SQ.getWithInstruction(&I))) 4627 return replaceInstUsesWith(I, V); 4628 4629 // Comparing -val or val with non-zero is the same as just comparing val 4630 // ie, abs(val) != 0 -> val != 0 4631 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) { 4632 Value *Cond, *SelectTrue, *SelectFalse; 4633 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue), 4634 m_Value(SelectFalse)))) { 4635 if (Value *V = dyn_castNegVal(SelectTrue)) { 4636 if (V == SelectFalse) 4637 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 4638 } 4639 else if (Value *V = dyn_castNegVal(SelectFalse)) { 4640 if (V == SelectTrue) 4641 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1); 4642 } 4643 } 4644 } 4645 4646 if (Op0->getType()->isIntOrIntVectorTy(1)) 4647 if (Instruction *Res = canonicalizeICmpBool(I, Builder)) 4648 return Res; 4649 4650 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I)) 4651 return NewICmp; 4652 4653 if (Instruction *Res = foldICmpWithConstant(I)) 4654 return Res; 4655 4656 if (Instruction *Res = foldICmpUsingKnownBits(I)) 4657 return Res; 4658 4659 // Test if the ICmpInst instruction is used exclusively by a select as 4660 // part of a minimum or maximum operation. If so, refrain from doing 4661 // any other folding. This helps out other analyses which understand 4662 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 4663 // and CodeGen. And in this case, at least one of the comparison 4664 // operands has at least one user besides the compare (the select), 4665 // which would often largely negate the benefit of folding anyway. 4666 // 4667 // Do the same for the other patterns recognized by matchSelectPattern. 4668 if (I.hasOneUse()) 4669 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) { 4670 Value *A, *B; 4671 SelectPatternResult SPR = matchSelectPattern(SI, A, B); 4672 if (SPR.Flavor != SPF_UNKNOWN) 4673 return nullptr; 4674 } 4675 4676 // Do this after checking for min/max to prevent infinite looping. 4677 if (Instruction *Res = foldICmpWithZero(I)) 4678 return Res; 4679 4680 // FIXME: We only do this after checking for min/max to prevent infinite 4681 // looping caused by a reverse canonicalization of these patterns for min/max. 4682 // FIXME: The organization of folds is a mess. These would naturally go into 4683 // canonicalizeCmpWithConstant(), but we can't move all of the above folds 4684 // down here after the min/max restriction. 4685 ICmpInst::Predicate Pred = I.getPredicate(); 4686 const APInt *C; 4687 if (match(Op1, m_APInt(C))) { 4688 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set 4689 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) { 4690 Constant *Zero = Constant::getNullValue(Op0->getType()); 4691 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero); 4692 } 4693 4694 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear 4695 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) { 4696 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType()); 4697 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes); 4698 } 4699 } 4700 4701 if (Instruction *Res = foldICmpInstWithConstant(I)) 4702 return Res; 4703 4704 if (Instruction *Res = foldICmpInstWithConstantNotInt(I)) 4705 return Res; 4706 4707 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now. 4708 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0)) 4709 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I)) 4710 return NI; 4711 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) 4712 if (Instruction *NI = foldGEPICmp(GEP, Op0, 4713 ICmpInst::getSwappedPredicate(I.getPredicate()), I)) 4714 return NI; 4715 4716 // Try to optimize equality comparisons against alloca-based pointers. 4717 if (Op0->getType()->isPointerTy() && I.isEquality()) { 4718 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?"); 4719 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL))) 4720 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1)) 4721 return New; 4722 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL))) 4723 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0)) 4724 return New; 4725 } 4726 4727 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast. 4728 Value *X; 4729 if (match(Op0, m_BitCast(m_SIToFP(m_Value(X))))) { 4730 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0 4731 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0 4732 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0 4733 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0 4734 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT || 4735 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) && 4736 match(Op1, m_Zero())) 4737 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType())); 4738 4739 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1 4740 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One())) 4741 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1)); 4742 4743 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1 4744 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes())) 4745 return new ICmpInst(Pred, X, ConstantInt::getAllOnesValue(X->getType())); 4746 } 4747 4748 // Zero-equality checks are preserved through unsigned floating-point casts: 4749 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0 4750 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0 4751 if (match(Op0, m_BitCast(m_UIToFP(m_Value(X))))) 4752 if (I.isEquality() && match(Op1, m_Zero())) 4753 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType())); 4754 4755 // Test to see if the operands of the icmp are casted versions of other 4756 // values. If the ptr->ptr cast can be stripped off both arguments, we do so 4757 // now. 4758 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) { 4759 if (Op0->getType()->isPointerTy() && 4760 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) { 4761 // We keep moving the cast from the left operand over to the right 4762 // operand, where it can often be eliminated completely. 4763 Op0 = CI->getOperand(0); 4764 4765 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast 4766 // so eliminate it as well. 4767 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1)) 4768 Op1 = CI2->getOperand(0); 4769 4770 // If Op1 is a constant, we can fold the cast into the constant. 4771 if (Op0->getType() != Op1->getType()) { 4772 if (Constant *Op1C = dyn_cast<Constant>(Op1)) { 4773 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType()); 4774 } else { 4775 // Otherwise, cast the RHS right before the icmp 4776 Op1 = Builder.CreateBitCast(Op1, Op0->getType()); 4777 } 4778 } 4779 return new ICmpInst(I.getPredicate(), Op0, Op1); 4780 } 4781 } 4782 4783 if (isa<CastInst>(Op0)) { 4784 // Handle the special case of: icmp (cast bool to X), <cst> 4785 // This comes up when you have code like 4786 // int X = A < B; 4787 // if (X) ... 4788 // For generality, we handle any zero-extension of any operand comparison 4789 // with a constant or another cast from the same type. 4790 if (isa<Constant>(Op1) || isa<CastInst>(Op1)) 4791 if (Instruction *R = foldICmpWithCastAndCast(I)) 4792 return R; 4793 } 4794 4795 if (Instruction *Res = foldICmpBinOp(I)) 4796 return Res; 4797 4798 if (Instruction *Res = foldICmpWithMinMax(I)) 4799 return Res; 4800 4801 { 4802 Value *A, *B; 4803 // Transform (A & ~B) == 0 --> (A & B) != 0 4804 // and (A & ~B) != 0 --> (A & B) == 0 4805 // if A is a power of 2. 4806 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) && 4807 match(Op1, m_Zero()) && 4808 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality()) 4809 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B), 4810 Op1); 4811 4812 // ~X < ~Y --> Y < X 4813 // ~X < C --> X > ~C 4814 if (match(Op0, m_Not(m_Value(A)))) { 4815 if (match(Op1, m_Not(m_Value(B)))) 4816 return new ICmpInst(I.getPredicate(), B, A); 4817 4818 const APInt *C; 4819 if (match(Op1, m_APInt(C))) 4820 return new ICmpInst(I.getSwappedPredicate(), A, 4821 ConstantInt::get(Op1->getType(), ~(*C))); 4822 } 4823 4824 Instruction *AddI = nullptr; 4825 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B), 4826 m_Instruction(AddI))) && 4827 isa<IntegerType>(A->getType())) { 4828 Value *Result; 4829 Constant *Overflow; 4830 if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result, 4831 Overflow)) { 4832 replaceInstUsesWith(*AddI, Result); 4833 return replaceInstUsesWith(I, Overflow); 4834 } 4835 } 4836 4837 // (zext a) * (zext b) --> llvm.umul.with.overflow. 4838 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) { 4839 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this)) 4840 return R; 4841 } 4842 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) { 4843 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this)) 4844 return R; 4845 } 4846 } 4847 4848 if (Instruction *Res = foldICmpEquality(I)) 4849 return Res; 4850 4851 // The 'cmpxchg' instruction returns an aggregate containing the old value and 4852 // an i1 which indicates whether or not we successfully did the swap. 4853 // 4854 // Replace comparisons between the old value and the expected value with the 4855 // indicator that 'cmpxchg' returns. 4856 // 4857 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to 4858 // spuriously fail. In those cases, the old value may equal the expected 4859 // value but it is possible for the swap to not occur. 4860 if (I.getPredicate() == ICmpInst::ICMP_EQ) 4861 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0)) 4862 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand())) 4863 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 && 4864 !ACXI->isWeak()) 4865 return ExtractValueInst::Create(ACXI, 1); 4866 4867 { 4868 Value *X; ConstantInt *Cst; 4869 // icmp X+Cst, X 4870 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X) 4871 return foldICmpAddOpConst(X, Cst, I.getPredicate()); 4872 4873 // icmp X, X+Cst 4874 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X) 4875 return foldICmpAddOpConst(X, Cst, I.getSwappedPredicate()); 4876 } 4877 4878 return Changed ? &I : nullptr; 4879 } 4880 4881 /// Fold fcmp ([us]itofp x, cst) if possible. 4882 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI, 4883 Constant *RHSC) { 4884 if (!isa<ConstantFP>(RHSC)) return nullptr; 4885 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF(); 4886 4887 // Get the width of the mantissa. We don't want to hack on conversions that 4888 // might lose information from the integer, e.g. "i64 -> float" 4889 int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); 4890 if (MantissaWidth == -1) return nullptr; // Unknown. 4891 4892 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType()); 4893 4894 bool LHSUnsigned = isa<UIToFPInst>(LHSI); 4895 4896 if (I.isEquality()) { 4897 FCmpInst::Predicate P = I.getPredicate(); 4898 bool IsExact = false; 4899 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned); 4900 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact); 4901 4902 // If the floating point constant isn't an integer value, we know if we will 4903 // ever compare equal / not equal to it. 4904 if (!IsExact) { 4905 // TODO: Can never be -0.0 and other non-representable values 4906 APFloat RHSRoundInt(RHS); 4907 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven); 4908 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) { 4909 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ) 4910 return replaceInstUsesWith(I, Builder.getFalse()); 4911 4912 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE); 4913 return replaceInstUsesWith(I, Builder.getTrue()); 4914 } 4915 } 4916 4917 // TODO: If the constant is exactly representable, is it always OK to do 4918 // equality compares as integer? 4919 } 4920 4921 // Check to see that the input is converted from an integer type that is small 4922 // enough that preserves all bits. TODO: check here for "known" sign bits. 4923 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. 4924 unsigned InputSize = IntTy->getScalarSizeInBits(); 4925 4926 // Following test does NOT adjust InputSize downwards for signed inputs, 4927 // because the most negative value still requires all the mantissa bits 4928 // to distinguish it from one less than that value. 4929 if ((int)InputSize > MantissaWidth) { 4930 // Conversion would lose accuracy. Check if loss can impact comparison. 4931 int Exp = ilogb(RHS); 4932 if (Exp == APFloat::IEK_Inf) { 4933 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics())); 4934 if (MaxExponent < (int)InputSize - !LHSUnsigned) 4935 // Conversion could create infinity. 4936 return nullptr; 4937 } else { 4938 // Note that if RHS is zero or NaN, then Exp is negative 4939 // and first condition is trivially false. 4940 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned) 4941 // Conversion could affect comparison. 4942 return nullptr; 4943 } 4944 } 4945 4946 // Otherwise, we can potentially simplify the comparison. We know that it 4947 // will always come through as an integer value and we know the constant is 4948 // not a NAN (it would have been previously simplified). 4949 assert(!RHS.isNaN() && "NaN comparison not already folded!"); 4950 4951 ICmpInst::Predicate Pred; 4952 switch (I.getPredicate()) { 4953 default: llvm_unreachable("Unexpected predicate!"); 4954 case FCmpInst::FCMP_UEQ: 4955 case FCmpInst::FCMP_OEQ: 4956 Pred = ICmpInst::ICMP_EQ; 4957 break; 4958 case FCmpInst::FCMP_UGT: 4959 case FCmpInst::FCMP_OGT: 4960 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; 4961 break; 4962 case FCmpInst::FCMP_UGE: 4963 case FCmpInst::FCMP_OGE: 4964 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; 4965 break; 4966 case FCmpInst::FCMP_ULT: 4967 case FCmpInst::FCMP_OLT: 4968 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; 4969 break; 4970 case FCmpInst::FCMP_ULE: 4971 case FCmpInst::FCMP_OLE: 4972 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; 4973 break; 4974 case FCmpInst::FCMP_UNE: 4975 case FCmpInst::FCMP_ONE: 4976 Pred = ICmpInst::ICMP_NE; 4977 break; 4978 case FCmpInst::FCMP_ORD: 4979 return replaceInstUsesWith(I, Builder.getTrue()); 4980 case FCmpInst::FCMP_UNO: 4981 return replaceInstUsesWith(I, Builder.getFalse()); 4982 } 4983 4984 // Now we know that the APFloat is a normal number, zero or inf. 4985 4986 // See if the FP constant is too large for the integer. For example, 4987 // comparing an i8 to 300.0. 4988 unsigned IntWidth = IntTy->getScalarSizeInBits(); 4989 4990 if (!LHSUnsigned) { 4991 // If the RHS value is > SignedMax, fold the comparison. This handles +INF 4992 // and large values. 4993 APFloat SMax(RHS.getSemantics()); 4994 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true, 4995 APFloat::rmNearestTiesToEven); 4996 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0 4997 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || 4998 Pred == ICmpInst::ICMP_SLE) 4999 return replaceInstUsesWith(I, Builder.getTrue()); 5000 return replaceInstUsesWith(I, Builder.getFalse()); 5001 } 5002 } else { 5003 // If the RHS value is > UnsignedMax, fold the comparison. This handles 5004 // +INF and large values. 5005 APFloat UMax(RHS.getSemantics()); 5006 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false, 5007 APFloat::rmNearestTiesToEven); 5008 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0 5009 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || 5010 Pred == ICmpInst::ICMP_ULE) 5011 return replaceInstUsesWith(I, Builder.getTrue()); 5012 return replaceInstUsesWith(I, Builder.getFalse()); 5013 } 5014 } 5015 5016 if (!LHSUnsigned) { 5017 // See if the RHS value is < SignedMin. 5018 APFloat SMin(RHS.getSemantics()); 5019 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true, 5020 APFloat::rmNearestTiesToEven); 5021 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0 5022 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || 5023 Pred == ICmpInst::ICMP_SGE) 5024 return replaceInstUsesWith(I, Builder.getTrue()); 5025 return replaceInstUsesWith(I, Builder.getFalse()); 5026 } 5027 } else { 5028 // See if the RHS value is < UnsignedMin. 5029 APFloat SMin(RHS.getSemantics()); 5030 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true, 5031 APFloat::rmNearestTiesToEven); 5032 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0 5033 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || 5034 Pred == ICmpInst::ICMP_UGE) 5035 return replaceInstUsesWith(I, Builder.getTrue()); 5036 return replaceInstUsesWith(I, Builder.getFalse()); 5037 } 5038 } 5039 5040 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or 5041 // [0, UMAX], but it may still be fractional. See if it is fractional by 5042 // casting the FP value to the integer value and back, checking for equality. 5043 // Don't do this for zero, because -0.0 is not fractional. 5044 Constant *RHSInt = LHSUnsigned 5045 ? ConstantExpr::getFPToUI(RHSC, IntTy) 5046 : ConstantExpr::getFPToSI(RHSC, IntTy); 5047 if (!RHS.isZero()) { 5048 bool Equal = LHSUnsigned 5049 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC 5050 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC; 5051 if (!Equal) { 5052 // If we had a comparison against a fractional value, we have to adjust 5053 // the compare predicate and sometimes the value. RHSC is rounded towards 5054 // zero at this point. 5055 switch (Pred) { 5056 default: llvm_unreachable("Unexpected integer comparison!"); 5057 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true 5058 return replaceInstUsesWith(I, Builder.getTrue()); 5059 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false 5060 return replaceInstUsesWith(I, Builder.getFalse()); 5061 case ICmpInst::ICMP_ULE: 5062 // (float)int <= 4.4 --> int <= 4 5063 // (float)int <= -4.4 --> false 5064 if (RHS.isNegative()) 5065 return replaceInstUsesWith(I, Builder.getFalse()); 5066 break; 5067 case ICmpInst::ICMP_SLE: 5068 // (float)int <= 4.4 --> int <= 4 5069 // (float)int <= -4.4 --> int < -4 5070 if (RHS.isNegative()) 5071 Pred = ICmpInst::ICMP_SLT; 5072 break; 5073 case ICmpInst::ICMP_ULT: 5074 // (float)int < -4.4 --> false 5075 // (float)int < 4.4 --> int <= 4 5076 if (RHS.isNegative()) 5077 return replaceInstUsesWith(I, Builder.getFalse()); 5078 Pred = ICmpInst::ICMP_ULE; 5079 break; 5080 case ICmpInst::ICMP_SLT: 5081 // (float)int < -4.4 --> int < -4 5082 // (float)int < 4.4 --> int <= 4 5083 if (!RHS.isNegative()) 5084 Pred = ICmpInst::ICMP_SLE; 5085 break; 5086 case ICmpInst::ICMP_UGT: 5087 // (float)int > 4.4 --> int > 4 5088 // (float)int > -4.4 --> true 5089 if (RHS.isNegative()) 5090 return replaceInstUsesWith(I, Builder.getTrue()); 5091 break; 5092 case ICmpInst::ICMP_SGT: 5093 // (float)int > 4.4 --> int > 4 5094 // (float)int > -4.4 --> int >= -4 5095 if (RHS.isNegative()) 5096 Pred = ICmpInst::ICMP_SGE; 5097 break; 5098 case ICmpInst::ICMP_UGE: 5099 // (float)int >= -4.4 --> true 5100 // (float)int >= 4.4 --> int > 4 5101 if (RHS.isNegative()) 5102 return replaceInstUsesWith(I, Builder.getTrue()); 5103 Pred = ICmpInst::ICMP_UGT; 5104 break; 5105 case ICmpInst::ICMP_SGE: 5106 // (float)int >= -4.4 --> int >= -4 5107 // (float)int >= 4.4 --> int > 4 5108 if (!RHS.isNegative()) 5109 Pred = ICmpInst::ICMP_SGT; 5110 break; 5111 } 5112 } 5113 } 5114 5115 // Lower this FP comparison into an appropriate integer version of the 5116 // comparison. 5117 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt); 5118 } 5119 5120 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) { 5121 bool Changed = false; 5122 5123 /// Orders the operands of the compare so that they are listed from most 5124 /// complex to least complex. This puts constants before unary operators, 5125 /// before binary operators. 5126 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) { 5127 I.swapOperands(); 5128 Changed = true; 5129 } 5130 5131 const CmpInst::Predicate Pred = I.getPredicate(); 5132 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 5133 if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(), 5134 SQ.getWithInstruction(&I))) 5135 return replaceInstUsesWith(I, V); 5136 5137 // Simplify 'fcmp pred X, X' 5138 if (Op0 == Op1) { 5139 switch (Pred) { 5140 default: break; 5141 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) 5142 case FCmpInst::FCMP_ULT: // True if unordered or less than 5143 case FCmpInst::FCMP_UGT: // True if unordered or greater than 5144 case FCmpInst::FCMP_UNE: // True if unordered or not equal 5145 // Canonicalize these to be 'fcmp uno %X, 0.0'. 5146 I.setPredicate(FCmpInst::FCMP_UNO); 5147 I.setOperand(1, Constant::getNullValue(Op0->getType())); 5148 return &I; 5149 5150 case FCmpInst::FCMP_ORD: // True if ordered (no nans) 5151 case FCmpInst::FCMP_OEQ: // True if ordered and equal 5152 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal 5153 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal 5154 // Canonicalize these to be 'fcmp ord %X, 0.0'. 5155 I.setPredicate(FCmpInst::FCMP_ORD); 5156 I.setOperand(1, Constant::getNullValue(Op0->getType())); 5157 return &I; 5158 } 5159 } 5160 5161 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand, 5162 // then canonicalize the operand to 0.0. 5163 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) { 5164 if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0)) { 5165 I.setOperand(0, ConstantFP::getNullValue(Op0->getType())); 5166 return &I; 5167 } 5168 if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1)) { 5169 I.setOperand(1, ConstantFP::getNullValue(Op0->getType())); 5170 return &I; 5171 } 5172 } 5173 5174 // Test if the FCmpInst instruction is used exclusively by a select as 5175 // part of a minimum or maximum operation. If so, refrain from doing 5176 // any other folding. This helps out other analyses which understand 5177 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution 5178 // and CodeGen. And in this case, at least one of the comparison 5179 // operands has at least one user besides the compare (the select), 5180 // which would often largely negate the benefit of folding anyway. 5181 if (I.hasOneUse()) 5182 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) { 5183 Value *A, *B; 5184 SelectPatternResult SPR = matchSelectPattern(SI, A, B); 5185 if (SPR.Flavor != SPF_UNKNOWN) 5186 return nullptr; 5187 } 5188 5189 // Handle fcmp with constant RHS 5190 if (Constant *RHSC = dyn_cast<Constant>(Op1)) { 5191 if (Instruction *LHSI = dyn_cast<Instruction>(Op0)) 5192 switch (LHSI->getOpcode()) { 5193 case Instruction::FPExt: { 5194 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless 5195 FPExtInst *LHSExt = cast<FPExtInst>(LHSI); 5196 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC); 5197 if (!RHSF) 5198 break; 5199 5200 const fltSemantics *Sem; 5201 // FIXME: This shouldn't be here. 5202 if (LHSExt->getSrcTy()->isHalfTy()) 5203 Sem = &APFloat::IEEEhalf(); 5204 else if (LHSExt->getSrcTy()->isFloatTy()) 5205 Sem = &APFloat::IEEEsingle(); 5206 else if (LHSExt->getSrcTy()->isDoubleTy()) 5207 Sem = &APFloat::IEEEdouble(); 5208 else if (LHSExt->getSrcTy()->isFP128Ty()) 5209 Sem = &APFloat::IEEEquad(); 5210 else if (LHSExt->getSrcTy()->isX86_FP80Ty()) 5211 Sem = &APFloat::x87DoubleExtended(); 5212 else if (LHSExt->getSrcTy()->isPPC_FP128Ty()) 5213 Sem = &APFloat::PPCDoubleDouble(); 5214 else 5215 break; 5216 5217 bool Lossy; 5218 APFloat F = RHSF->getValueAPF(); 5219 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy); 5220 5221 // Avoid lossy conversions and denormals. Zero is a special case 5222 // that's OK to convert. 5223 APFloat Fabs = F; 5224 Fabs.clearSign(); 5225 if (!Lossy && 5226 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) != 5227 APFloat::cmpLessThan) || Fabs.isZero())) 5228 5229 return new FCmpInst(Pred, LHSExt->getOperand(0), 5230 ConstantFP::get(RHSC->getContext(), F)); 5231 break; 5232 } 5233 case Instruction::PHI: 5234 // Only fold fcmp into the PHI if the phi and fcmp are in the same 5235 // block. If in the same block, we're encouraging jump threading. If 5236 // not, we are just pessimizing the code by making an i1 phi. 5237 if (LHSI->getParent() == I.getParent()) 5238 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI))) 5239 return NV; 5240 break; 5241 case Instruction::SIToFP: 5242 case Instruction::UIToFP: 5243 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC)) 5244 return NV; 5245 break; 5246 case Instruction::FSub: { 5247 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C 5248 Value *Op; 5249 if (match(LHSI, m_FNeg(m_Value(Op)))) 5250 return new FCmpInst(I.getSwappedPredicate(), Op, 5251 ConstantExpr::getFNeg(RHSC)); 5252 break; 5253 } 5254 case Instruction::Load: 5255 if (GetElementPtrInst *GEP = 5256 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) { 5257 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0))) 5258 if (GV->isConstant() && GV->hasDefinitiveInitializer() && 5259 !cast<LoadInst>(LHSI)->isVolatile()) 5260 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I)) 5261 return Res; 5262 } 5263 break; 5264 case Instruction::Call: { 5265 if (!RHSC->isNullValue()) 5266 break; 5267 5268 CallInst *CI = cast<CallInst>(LHSI); 5269 Intrinsic::ID IID = getIntrinsicForCallSite(CI, &TLI); 5270 if (IID != Intrinsic::fabs) 5271 break; 5272 5273 // Various optimization for fabs compared with zero. 5274 switch (Pred) { 5275 default: 5276 break; 5277 // fabs(x) < 0 --> false 5278 case FCmpInst::FCMP_OLT: 5279 llvm_unreachable("handled by SimplifyFCmpInst"); 5280 // fabs(x) > 0 --> x != 0 5281 case FCmpInst::FCMP_OGT: 5282 return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0), RHSC); 5283 // fabs(x) <= 0 --> x == 0 5284 case FCmpInst::FCMP_OLE: 5285 return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0), RHSC); 5286 // fabs(x) >= 0 --> !isnan(x) 5287 case FCmpInst::FCMP_OGE: 5288 return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0), RHSC); 5289 // fabs(x) == 0 --> x == 0 5290 // fabs(x) != 0 --> x != 0 5291 case FCmpInst::FCMP_OEQ: 5292 case FCmpInst::FCMP_UEQ: 5293 case FCmpInst::FCMP_ONE: 5294 case FCmpInst::FCMP_UNE: 5295 return new FCmpInst(Pred, CI->getArgOperand(0), RHSC); 5296 } 5297 } 5298 } 5299 } 5300 5301 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y 5302 Value *X, *Y; 5303 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) 5304 return new FCmpInst(I.getSwappedPredicate(), X, Y); 5305 5306 // fcmp (fpext x), (fpext y) -> fcmp x, y 5307 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0)) 5308 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1)) 5309 if (LHSExt->getSrcTy() == RHSExt->getSrcTy()) 5310 return new FCmpInst(Pred, LHSExt->getOperand(0), RHSExt->getOperand(0)); 5311 5312 return Changed ? &I : nullptr; 5313 } 5314