1 //===- InstCombineCompares.cpp --------------------------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements the visitICmp and visitFCmp functions.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APSInt.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/ConstantFolding.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/KnownBits.h"
29 using namespace PatternMatch;
31 #define DEBUG_TYPE "instcombine"
33 // How many times is a select replaced by one of its operands?
34 STATISTIC(NumSel, "Number of select opts");
37 /// Compute Result = In1+In2, returning true if the result overflowed for this
39 static bool addWithOverflow(APInt &Result, const APInt &In1,
40 const APInt &In2, bool IsSigned = false) {
43 Result = In1.sadd_ov(In2, Overflow);
45 Result = In1.uadd_ov(In2, Overflow);
50 /// Compute Result = In1-In2, returning true if the result overflowed for this
52 static bool subWithOverflow(APInt &Result, const APInt &In1,
53 const APInt &In2, bool IsSigned = false) {
56 Result = In1.ssub_ov(In2, Overflow);
58 Result = In1.usub_ov(In2, Overflow);
63 /// Given an icmp instruction, return true if any use of this comparison is a
64 /// branch on sign bit comparison.
65 static bool hasBranchUse(ICmpInst &I) {
66 for (auto *U : I.users())
67 if (isa<BranchInst>(U))
72 /// Returns true if the exploded icmp can be expressed as a signed comparison
73 /// to zero and updates the predicate accordingly.
74 /// The signedness of the comparison is preserved.
75 /// TODO: Refactor with decomposeBitTestICmp()?
76 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
77 if (!ICmpInst::isSigned(Pred))
81 return ICmpInst::isRelational(Pred);
84 if (Pred == ICmpInst::ICMP_SLT) {
85 Pred = ICmpInst::ICMP_SLE;
88 } else if (C.isAllOnesValue()) {
89 if (Pred == ICmpInst::ICMP_SGT) {
90 Pred = ICmpInst::ICMP_SGE;
98 /// Given a signed integer type and a set of known zero and one bits, compute
99 /// the maximum and minimum values that could have the specified known zero and
100 /// known one bits, returning them in Min/Max.
101 /// TODO: Move to method on KnownBits struct?
102 static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
103 APInt &Min, APInt &Max) {
104 assert(Known.getBitWidth() == Min.getBitWidth() &&
105 Known.getBitWidth() == Max.getBitWidth() &&
106 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
107 APInt UnknownBits = ~(Known.Zero|Known.One);
109 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
110 // bit if it is unknown.
112 Max = Known.One|UnknownBits;
114 if (UnknownBits.isNegative()) { // Sign bit is unknown
120 /// Given an unsigned integer type and a set of known zero and one bits, compute
121 /// the maximum and minimum values that could have the specified known zero and
122 /// known one bits, returning them in Min/Max.
123 /// TODO: Move to method on KnownBits struct?
124 static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
125 APInt &Min, APInt &Max) {
126 assert(Known.getBitWidth() == Min.getBitWidth() &&
127 Known.getBitWidth() == Max.getBitWidth() &&
128 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
129 APInt UnknownBits = ~(Known.Zero|Known.One);
131 // The minimum value is when the unknown bits are all zeros.
133 // The maximum value is when the unknown bits are all ones.
134 Max = Known.One|UnknownBits;
137 /// This is called when we see this pattern:
138 /// cmp pred (load (gep GV, ...)), cmpcst
139 /// where GV is a global variable with a constant initializer. Try to simplify
140 /// this into some simple computation that does not need the load. For example
141 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
143 /// If AndCst is non-null, then the loaded value is masked with that constant
144 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
145 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
148 ConstantInt *AndCst) {
149 Constant *Init = GV->getInitializer();
150 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
153 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
154 // Don't blow up on huge arrays.
155 if (ArrayElementCount > MaxArraySizeForCombine)
158 // There are many forms of this optimization we can handle, for now, just do
159 // the simple index into a single-dimensional array.
161 // Require: GEP GV, 0, i {{, constant indices}}
162 if (GEP->getNumOperands() < 3 ||
163 !isa<ConstantInt>(GEP->getOperand(1)) ||
164 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
165 isa<Constant>(GEP->getOperand(2)))
168 // Check that indices after the variable are constants and in-range for the
169 // type they index. Collect the indices. This is typically for arrays of
171 SmallVector<unsigned, 4> LaterIndices;
173 Type *EltTy = Init->getType()->getArrayElementType();
174 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
175 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
176 if (!Idx) return nullptr; // Variable index.
178 uint64_t IdxVal = Idx->getZExtValue();
179 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
181 if (StructType *STy = dyn_cast<StructType>(EltTy))
182 EltTy = STy->getElementType(IdxVal);
183 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
184 if (IdxVal >= ATy->getNumElements()) return nullptr;
185 EltTy = ATy->getElementType();
187 return nullptr; // Unknown type.
190 LaterIndices.push_back(IdxVal);
193 enum { Overdefined = -3, Undefined = -2 };
195 // Variables for our state machines.
197 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
198 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
199 // and 87 is the second (and last) index. FirstTrueElement is -2 when
200 // undefined, otherwise set to the first true element. SecondTrueElement is
201 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
202 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
204 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
205 // form "i != 47 & i != 87". Same state transitions as for true elements.
206 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
208 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
209 /// define a state machine that triggers for ranges of values that the index
210 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
211 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
212 /// index in the range (inclusive). We use -2 for undefined here because we
213 /// use relative comparisons and don't want 0-1 to match -1.
214 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
216 // MagicBitvector - This is a magic bitvector where we set a bit if the
217 // comparison is true for element 'i'. If there are 64 elements or less in
218 // the array, this will fully represent all the comparison results.
219 uint64_t MagicBitvector = 0;
221 // Scan the array and see if one of our patterns matches.
222 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
223 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
224 Constant *Elt = Init->getAggregateElement(i);
225 if (!Elt) return nullptr;
227 // If this is indexing an array of structures, get the structure element.
228 if (!LaterIndices.empty())
229 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
231 // If the element is masked, handle it.
232 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
234 // Find out if the comparison would be true or false for the i'th element.
235 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
236 CompareRHS, DL, &TLI);
237 // If the result is undef for this element, ignore it.
238 if (isa<UndefValue>(C)) {
239 // Extend range state machines to cover this element in case there is an
240 // undef in the middle of the range.
241 if (TrueRangeEnd == (int)i-1)
243 if (FalseRangeEnd == (int)i-1)
248 // If we can't compute the result for any of the elements, we have to give
249 // up evaluating the entire conditional.
250 if (!isa<ConstantInt>(C)) return nullptr;
252 // Otherwise, we know if the comparison is true or false for this element,
253 // update our state machines.
254 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
256 // State machine for single/double/range index comparison.
258 // Update the TrueElement state machine.
259 if (FirstTrueElement == Undefined)
260 FirstTrueElement = TrueRangeEnd = i; // First true element.
262 // Update double-compare state machine.
263 if (SecondTrueElement == Undefined)
264 SecondTrueElement = i;
266 SecondTrueElement = Overdefined;
268 // Update range state machine.
269 if (TrueRangeEnd == (int)i-1)
272 TrueRangeEnd = Overdefined;
275 // Update the FalseElement state machine.
276 if (FirstFalseElement == Undefined)
277 FirstFalseElement = FalseRangeEnd = i; // First false element.
279 // Update double-compare state machine.
280 if (SecondFalseElement == Undefined)
281 SecondFalseElement = i;
283 SecondFalseElement = Overdefined;
285 // Update range state machine.
286 if (FalseRangeEnd == (int)i-1)
289 FalseRangeEnd = Overdefined;
293 // If this element is in range, update our magic bitvector.
294 if (i < 64 && IsTrueForElt)
295 MagicBitvector |= 1ULL << i;
297 // If all of our states become overdefined, bail out early. Since the
298 // predicate is expensive, only check it every 8 elements. This is only
299 // really useful for really huge arrays.
300 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
301 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
302 FalseRangeEnd == Overdefined)
306 // Now that we've scanned the entire array, emit our new comparison(s). We
307 // order the state machines in complexity of the generated code.
308 Value *Idx = GEP->getOperand(2);
310 // If the index is larger than the pointer size of the target, truncate the
311 // index down like the GEP would do implicitly. We don't have to do this for
312 // an inbounds GEP because the index can't be out of range.
313 if (!GEP->isInBounds()) {
314 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
315 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
316 if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
317 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
320 // If the comparison is only true for one or two elements, emit direct
322 if (SecondTrueElement != Overdefined) {
323 // None true -> false.
324 if (FirstTrueElement == Undefined)
325 return replaceInstUsesWith(ICI, Builder.getFalse());
327 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
329 // True for one element -> 'i == 47'.
330 if (SecondTrueElement == Undefined)
331 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
333 // True for two elements -> 'i == 47 | i == 72'.
334 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
335 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
336 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
337 return BinaryOperator::CreateOr(C1, C2);
340 // If the comparison is only false for one or two elements, emit direct
342 if (SecondFalseElement != Overdefined) {
343 // None false -> true.
344 if (FirstFalseElement == Undefined)
345 return replaceInstUsesWith(ICI, Builder.getTrue());
347 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
349 // False for one element -> 'i != 47'.
350 if (SecondFalseElement == Undefined)
351 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
353 // False for two elements -> 'i != 47 & i != 72'.
354 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
355 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
356 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
357 return BinaryOperator::CreateAnd(C1, C2);
360 // If the comparison can be replaced with a range comparison for the elements
361 // where it is true, emit the range check.
362 if (TrueRangeEnd != Overdefined) {
363 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
365 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
366 if (FirstTrueElement) {
367 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
368 Idx = Builder.CreateAdd(Idx, Offs);
371 Value *End = ConstantInt::get(Idx->getType(),
372 TrueRangeEnd-FirstTrueElement+1);
373 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
376 // False range check.
377 if (FalseRangeEnd != Overdefined) {
378 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
379 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
380 if (FirstFalseElement) {
381 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
382 Idx = Builder.CreateAdd(Idx, Offs);
385 Value *End = ConstantInt::get(Idx->getType(),
386 FalseRangeEnd-FirstFalseElement);
387 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
390 // If a magic bitvector captures the entire comparison state
391 // of this load, replace it with computation that does:
392 // ((magic_cst >> i) & 1) != 0
396 // Look for an appropriate type:
397 // - The type of Idx if the magic fits
398 // - The smallest fitting legal type
399 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
402 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
405 Value *V = Builder.CreateIntCast(Idx, Ty, false);
406 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
407 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
408 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
415 /// Return a value that can be used to compare the *offset* implied by a GEP to
416 /// zero. For example, if we have &A[i], we want to return 'i' for
417 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
418 /// are involved. The above expression would also be legal to codegen as
419 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
420 /// This latter form is less amenable to optimization though, and we are allowed
421 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
423 /// If we can't emit an optimized form for this expression, this returns null.
425 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC,
426 const DataLayout &DL) {
427 gep_type_iterator GTI = gep_type_begin(GEP);
429 // Check to see if this gep only has a single variable index. If so, and if
430 // any constant indices are a multiple of its scale, then we can compute this
431 // in terms of the scale of the variable index. For example, if the GEP
432 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
433 // because the expression will cross zero at the same point.
434 unsigned i, e = GEP->getNumOperands();
436 for (i = 1; i != e; ++i, ++GTI) {
437 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
438 // Compute the aggregate offset of constant indices.
439 if (CI->isZero()) continue;
441 // Handle a struct index, which adds its field offset to the pointer.
442 if (StructType *STy = GTI.getStructTypeOrNull()) {
443 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
445 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
446 Offset += Size*CI->getSExtValue();
449 // Found our variable index.
454 // If there are no variable indices, we must have a constant offset, just
455 // evaluate it the general way.
456 if (i == e) return nullptr;
458 Value *VariableIdx = GEP->getOperand(i);
459 // Determine the scale factor of the variable element. For example, this is
460 // 4 if the variable index is into an array of i32.
461 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
463 // Verify that there are no other variable indices. If so, emit the hard way.
464 for (++i, ++GTI; i != e; ++i, ++GTI) {
465 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
466 if (!CI) return nullptr;
468 // Compute the aggregate offset of constant indices.
469 if (CI->isZero()) continue;
471 // Handle a struct index, which adds its field offset to the pointer.
472 if (StructType *STy = GTI.getStructTypeOrNull()) {
473 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
475 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
476 Offset += Size*CI->getSExtValue();
480 // Okay, we know we have a single variable index, which must be a
481 // pointer/array/vector index. If there is no offset, life is simple, return
483 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
484 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
486 // Cast to intptrty in case a truncation occurs. If an extension is needed,
487 // we don't need to bother extending: the extension won't affect where the
488 // computation crosses zero.
489 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
490 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
495 // Otherwise, there is an index. The computation we will do will be modulo
497 Offset = SignExtend64(Offset, IntPtrWidth);
498 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
500 // To do this transformation, any constant index must be a multiple of the
501 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
502 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
503 // multiple of the variable scale.
504 int64_t NewOffs = Offset / (int64_t)VariableScale;
505 if (Offset != NewOffs*(int64_t)VariableScale)
508 // Okay, we can do this evaluation. Start by converting the index to intptr.
509 if (VariableIdx->getType() != IntPtrTy)
510 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
512 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
513 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
516 /// Returns true if we can rewrite Start as a GEP with pointer Base
517 /// and some integer offset. The nodes that need to be re-written
518 /// for this transformation will be added to Explored.
519 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
520 const DataLayout &DL,
521 SetVector<Value *> &Explored) {
522 SmallVector<Value *, 16> WorkList(1, Start);
523 Explored.insert(Base);
525 // The following traversal gives us an order which can be used
526 // when doing the final transformation. Since in the final
527 // transformation we create the PHI replacement instructions first,
528 // we don't have to get them in any particular order.
530 // However, for other instructions we will have to traverse the
531 // operands of an instruction first, which means that we have to
532 // do a post-order traversal.
533 while (!WorkList.empty()) {
534 SetVector<PHINode *> PHIs;
536 while (!WorkList.empty()) {
537 if (Explored.size() >= 100)
540 Value *V = WorkList.back();
542 if (Explored.count(V) != 0) {
547 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
548 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
549 // We've found some value that we can't explore which is different from
550 // the base. Therefore we can't do this transformation.
553 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
554 auto *CI = dyn_cast<CastInst>(V);
555 if (!CI->isNoopCast(DL))
558 if (Explored.count(CI->getOperand(0)) == 0)
559 WorkList.push_back(CI->getOperand(0));
562 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
563 // We're limiting the GEP to having one index. This will preserve
564 // the original pointer type. We could handle more cases in the
566 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
567 GEP->getType() != Start->getType())
570 if (Explored.count(GEP->getOperand(0)) == 0)
571 WorkList.push_back(GEP->getOperand(0));
574 if (WorkList.back() == V) {
576 // We've finished visiting this node, mark it as such.
580 if (auto *PN = dyn_cast<PHINode>(V)) {
581 // We cannot transform PHIs on unsplittable basic blocks.
582 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
589 // Explore the PHI nodes further.
590 for (auto *PN : PHIs)
591 for (Value *Op : PN->incoming_values())
592 if (Explored.count(Op) == 0)
593 WorkList.push_back(Op);
596 // Make sure that we can do this. Since we can't insert GEPs in a basic
597 // block before a PHI node, we can't easily do this transformation if
598 // we have PHI node users of transformed instructions.
599 for (Value *Val : Explored) {
600 for (Value *Use : Val->uses()) {
602 auto *PHI = dyn_cast<PHINode>(Use);
603 auto *Inst = dyn_cast<Instruction>(Val);
605 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
606 Explored.count(PHI) == 0)
609 if (PHI->getParent() == Inst->getParent())
616 // Sets the appropriate insert point on Builder where we can add
617 // a replacement Instruction for V (if that is possible).
618 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
619 bool Before = true) {
620 if (auto *PHI = dyn_cast<PHINode>(V)) {
621 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
624 if (auto *I = dyn_cast<Instruction>(V)) {
626 I = &*std::next(I->getIterator());
627 Builder.SetInsertPoint(I);
630 if (auto *A = dyn_cast<Argument>(V)) {
631 // Set the insertion point in the entry block.
632 BasicBlock &Entry = A->getParent()->getEntryBlock();
633 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
636 // Otherwise, this is a constant and we don't need to set a new
638 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
641 /// Returns a re-written value of Start as an indexed GEP using Base as a
643 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
644 const DataLayout &DL,
645 SetVector<Value *> &Explored) {
646 // Perform all the substitutions. This is a bit tricky because we can
647 // have cycles in our use-def chains.
648 // 1. Create the PHI nodes without any incoming values.
649 // 2. Create all the other values.
650 // 3. Add the edges for the PHI nodes.
651 // 4. Emit GEPs to get the original pointers.
652 // 5. Remove the original instructions.
653 Type *IndexType = IntegerType::get(
654 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
656 DenseMap<Value *, Value *> NewInsts;
657 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
659 // Create the new PHI nodes, without adding any incoming values.
660 for (Value *Val : Explored) {
663 // Create empty phi nodes. This avoids cyclic dependencies when creating
664 // the remaining instructions.
665 if (auto *PHI = dyn_cast<PHINode>(Val))
666 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
667 PHI->getName() + ".idx", PHI);
669 IRBuilder<> Builder(Base->getContext());
671 // Create all the other instructions.
672 for (Value *Val : Explored) {
674 if (NewInsts.find(Val) != NewInsts.end())
677 if (auto *CI = dyn_cast<CastInst>(Val)) {
678 // Don't get rid of the intermediate variable here; the store can grow
679 // the map which will invalidate the reference to the input value.
680 Value *V = NewInsts[CI->getOperand(0)];
684 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
685 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
686 : GEP->getOperand(1);
687 setInsertionPoint(Builder, GEP);
688 // Indices might need to be sign extended. GEPs will magically do
689 // this, but we need to do it ourselves here.
690 if (Index->getType()->getScalarSizeInBits() !=
691 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
692 Index = Builder.CreateSExtOrTrunc(
693 Index, NewInsts[GEP->getOperand(0)]->getType(),
694 GEP->getOperand(0)->getName() + ".sext");
697 auto *Op = NewInsts[GEP->getOperand(0)];
698 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
699 NewInsts[GEP] = Index;
701 NewInsts[GEP] = Builder.CreateNSWAdd(
702 Op, Index, GEP->getOperand(0)->getName() + ".add");
705 if (isa<PHINode>(Val))
708 llvm_unreachable("Unexpected instruction type");
711 // Add the incoming values to the PHI nodes.
712 for (Value *Val : Explored) {
715 // All the instructions have been created, we can now add edges to the
717 if (auto *PHI = dyn_cast<PHINode>(Val)) {
718 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
719 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
720 Value *NewIncoming = PHI->getIncomingValue(I);
722 if (NewInsts.find(NewIncoming) != NewInsts.end())
723 NewIncoming = NewInsts[NewIncoming];
725 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
730 for (Value *Val : Explored) {
734 // Depending on the type, for external users we have to emit
735 // a GEP or a GEP + ptrtoint.
736 setInsertionPoint(Builder, Val, false);
738 // If required, create an inttoptr instruction for Base.
739 Value *NewBase = Base;
740 if (!Base->getType()->isPointerTy())
741 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
742 Start->getName() + "to.ptr");
744 Value *GEP = Builder.CreateInBoundsGEP(
745 Start->getType()->getPointerElementType(), NewBase,
746 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
748 if (!Val->getType()->isPointerTy()) {
749 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
750 Val->getName() + ".conv");
753 Val->replaceAllUsesWith(GEP);
756 return NewInsts[Start];
759 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
760 /// the input Value as a constant indexed GEP. Returns a pair containing
761 /// the GEPs Pointer and Index.
762 static std::pair<Value *, Value *>
763 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
764 Type *IndexType = IntegerType::get(V->getContext(),
765 DL.getIndexTypeSizeInBits(V->getType()));
767 Constant *Index = ConstantInt::getNullValue(IndexType);
769 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
770 // We accept only inbouds GEPs here to exclude the possibility of
772 if (!GEP->isInBounds())
774 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
775 GEP->getType() == V->getType()) {
776 V = GEP->getOperand(0);
777 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
778 Index = ConstantExpr::getAdd(
779 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
784 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
785 if (!CI->isNoopCast(DL))
787 V = CI->getOperand(0);
790 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
791 if (!CI->isNoopCast(DL))
793 V = CI->getOperand(0);
801 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
802 /// We can look through PHIs, GEPs and casts in order to determine a common base
803 /// between GEPLHS and RHS.
804 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
805 ICmpInst::Predicate Cond,
806 const DataLayout &DL) {
807 // FIXME: Support vector of pointers.
808 if (GEPLHS->getType()->isVectorTy())
811 if (!GEPLHS->hasAllConstantIndices())
814 // Make sure the pointers have the same type.
815 if (GEPLHS->getType() != RHS->getType())
818 Value *PtrBase, *Index;
819 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
821 // The set of nodes that will take part in this transformation.
822 SetVector<Value *> Nodes;
824 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
827 // We know we can re-write this as
828 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
829 // Since we've only looked through inbouds GEPs we know that we
830 // can't have overflow on either side. We can therefore re-write
832 // OFFSET1 cmp OFFSET2
833 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
835 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
836 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
837 // offset. Since Index is the offset of LHS to the base pointer, we will now
838 // compare the offsets instead of comparing the pointers.
839 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
842 /// Fold comparisons between a GEP instruction and something else. At this point
843 /// we know that the GEP is on the LHS of the comparison.
844 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
845 ICmpInst::Predicate Cond,
847 // Don't transform signed compares of GEPs into index compares. Even if the
848 // GEP is inbounds, the final add of the base pointer can have signed overflow
849 // and would change the result of the icmp.
850 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
851 // the maximum signed value for the pointer type.
852 if (ICmpInst::isSigned(Cond))
855 // Look through bitcasts and addrspacecasts. We do not however want to remove
857 if (!isa<GetElementPtrInst>(RHS))
858 RHS = RHS->stripPointerCasts();
860 Value *PtrBase = GEPLHS->getOperand(0);
861 // FIXME: Support vector pointer GEPs.
862 if (PtrBase == RHS && GEPLHS->isInBounds() &&
863 !GEPLHS->getType()->isVectorTy()) {
864 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
865 // This transformation (ignoring the base and scales) is valid because we
866 // know pointers can't overflow since the gep is inbounds. See if we can
867 // output an optimized form.
868 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
870 // If not, synthesize the offset the hard way.
872 Offset = EmitGEPOffset(GEPLHS);
873 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
874 Constant::getNullValue(Offset->getType()));
877 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
878 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
879 !NullPointerIsDefined(I.getFunction(),
880 RHS->getType()->getPointerAddressSpace())) {
881 // For most address spaces, an allocation can't be placed at null, but null
882 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
883 // the only valid inbounds address derived from null, is null itself.
884 // Thus, we have four cases to consider:
885 // 1) Base == nullptr, Offset == 0 -> inbounds, null
886 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
887 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
888 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
890 // (Note if we're indexing a type of size 0, that simply collapses into one
891 // of the buckets above.)
893 // In general, we're allowed to make values less poison (i.e. remove
894 // sources of full UB), so in this case, we just select between the two
895 // non-poison cases (1 and 4 above).
897 // For vectors, we apply the same reasoning on a per-lane basis.
898 auto *Base = GEPLHS->getPointerOperand();
899 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
900 int NumElts = GEPLHS->getType()->getVectorNumElements();
901 Base = Builder.CreateVectorSplat(NumElts, Base);
903 return new ICmpInst(Cond, Base,
904 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
905 cast<Constant>(RHS), Base->getType()));
906 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
907 // If the base pointers are different, but the indices are the same, just
908 // compare the base pointer.
909 if (PtrBase != GEPRHS->getOperand(0)) {
910 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
911 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
912 GEPRHS->getOperand(0)->getType();
914 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
915 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
916 IndicesTheSame = false;
920 // If all indices are the same, just compare the base pointers.
921 Type *BaseType = GEPLHS->getOperand(0)->getType();
922 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
923 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
925 // If we're comparing GEPs with two base pointers that only differ in type
926 // and both GEPs have only constant indices or just one use, then fold
927 // the compare with the adjusted indices.
928 // FIXME: Support vector of pointers.
929 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
930 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
931 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
932 PtrBase->stripPointerCasts() ==
933 GEPRHS->getOperand(0)->stripPointerCasts() &&
934 !GEPLHS->getType()->isVectorTy()) {
935 Value *LOffset = EmitGEPOffset(GEPLHS);
936 Value *ROffset = EmitGEPOffset(GEPRHS);
938 // If we looked through an addrspacecast between different sized address
939 // spaces, the LHS and RHS pointers are different sized
940 // integers. Truncate to the smaller one.
941 Type *LHSIndexTy = LOffset->getType();
942 Type *RHSIndexTy = ROffset->getType();
943 if (LHSIndexTy != RHSIndexTy) {
944 if (LHSIndexTy->getPrimitiveSizeInBits() <
945 RHSIndexTy->getPrimitiveSizeInBits()) {
946 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
948 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
951 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
953 return replaceInstUsesWith(I, Cmp);
956 // Otherwise, the base pointers are different and the indices are
957 // different. Try convert this to an indexed compare by looking through
959 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
962 // If one of the GEPs has all zero indices, recurse.
963 // FIXME: Handle vector of pointers.
964 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
965 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
966 ICmpInst::getSwappedPredicate(Cond), I);
968 // If the other GEP has all zero indices, recurse.
969 // FIXME: Handle vector of pointers.
970 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
971 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
973 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
974 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
975 // If the GEPs only differ by one index, compare it.
976 unsigned NumDifferences = 0; // Keep track of # differences.
977 unsigned DiffOperand = 0; // The operand that differs.
978 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
979 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
980 Type *LHSType = GEPLHS->getOperand(i)->getType();
981 Type *RHSType = GEPRHS->getOperand(i)->getType();
982 // FIXME: Better support for vector of pointers.
983 if (LHSType->getPrimitiveSizeInBits() !=
984 RHSType->getPrimitiveSizeInBits() ||
985 (GEPLHS->getType()->isVectorTy() &&
986 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
987 // Irreconcilable differences.
992 if (NumDifferences++) break;
996 if (NumDifferences == 0) // SAME GEP?
997 return replaceInstUsesWith(I, // No comparison is needed here.
998 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
1000 else if (NumDifferences == 1 && GEPsInBounds) {
1001 Value *LHSV = GEPLHS->getOperand(DiffOperand);
1002 Value *RHSV = GEPRHS->getOperand(DiffOperand);
1003 // Make sure we do a signed comparison here.
1004 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1008 // Only lower this if the icmp is the only user of the GEP or if we expect
1009 // the result to fold to a constant!
1010 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1011 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1012 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1013 Value *L = EmitGEPOffset(GEPLHS);
1014 Value *R = EmitGEPOffset(GEPRHS);
1015 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1019 // Try convert this to an indexed compare by looking through PHIs/casts as a
1021 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1024 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1025 const AllocaInst *Alloca,
1026 const Value *Other) {
1027 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1029 // It would be tempting to fold away comparisons between allocas and any
1030 // pointer not based on that alloca (e.g. an argument). However, even
1031 // though such pointers cannot alias, they can still compare equal.
1033 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1034 // doesn't escape we can argue that it's impossible to guess its value, and we
1035 // can therefore act as if any such guesses are wrong.
1037 // The code below checks that the alloca doesn't escape, and that it's only
1038 // used in a comparison once (the current instruction). The
1039 // single-comparison-use condition ensures that we're trivially folding all
1040 // comparisons against the alloca consistently, and avoids the risk of
1041 // erroneously folding a comparison of the pointer with itself.
1043 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1045 SmallVector<const Use *, 32> Worklist;
1046 for (const Use &U : Alloca->uses()) {
1047 if (Worklist.size() >= MaxIter)
1049 Worklist.push_back(&U);
1052 unsigned NumCmps = 0;
1053 while (!Worklist.empty()) {
1054 assert(Worklist.size() <= MaxIter);
1055 const Use *U = Worklist.pop_back_val();
1056 const Value *V = U->getUser();
1059 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1060 isa<SelectInst>(V)) {
1062 } else if (isa<LoadInst>(V)) {
1063 // Loading from the pointer doesn't escape it.
1065 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1066 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1067 if (SI->getValueOperand() == U->get())
1070 } else if (isa<ICmpInst>(V)) {
1072 return nullptr; // Found more than one cmp.
1074 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1075 switch (Intrin->getIntrinsicID()) {
1076 // These intrinsics don't escape or compare the pointer. Memset is safe
1077 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1078 // we don't allow stores, so src cannot point to V.
1079 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1080 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1088 for (const Use &U : V->uses()) {
1089 if (Worklist.size() >= MaxIter)
1091 Worklist.push_back(&U);
1095 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1096 return replaceInstUsesWith(
1098 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1101 /// Fold "icmp pred (X+C), X".
1102 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1103 ICmpInst::Predicate Pred) {
1104 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1105 // so the values can never be equal. Similarly for all other "or equals"
1107 assert(!!C && "C should not be zero!");
1109 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1110 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1111 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1112 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1113 Constant *R = ConstantInt::get(X->getType(),
1114 APInt::getMaxValue(C.getBitWidth()) - C);
1115 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1118 // (X+1) >u X --> X <u (0-1) --> X != 255
1119 // (X+2) >u X --> X <u (0-2) --> X <u 254
1120 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1121 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1122 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1123 ConstantInt::get(X->getType(), -C));
1125 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1127 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1128 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1129 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1130 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1131 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1132 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1133 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1134 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1135 ConstantInt::get(X->getType(), SMax - C));
1137 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1138 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1139 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1140 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1141 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1142 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1144 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1145 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1146 ConstantInt::get(X->getType(), SMax - (C - 1)));
1149 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1150 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1151 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1152 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1155 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1157 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1158 if (I.getPredicate() == I.ICMP_NE)
1159 Pred = CmpInst::getInversePredicate(Pred);
1160 return new ICmpInst(Pred, LHS, RHS);
1163 // Don't bother doing any work for cases which InstSimplify handles.
1164 if (AP2.isNullValue())
1167 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1169 if (AP2.isAllOnesValue())
1171 if (AP2.isNegative() != AP1.isNegative())
1178 // 'A' must be large enough to shift out the highest set bit.
1179 return getICmp(I.ICMP_UGT, A,
1180 ConstantInt::get(A->getType(), AP2.logBase2()));
1183 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1186 if (IsAShr && AP1.isNegative())
1187 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1189 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1192 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1193 // There are multiple solutions if we are comparing against -1 and the LHS
1194 // of the ashr is not a power of two.
1195 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1196 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1197 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1198 } else if (AP1 == AP2.lshr(Shift)) {
1199 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1203 // Shifting const2 will never be equal to const1.
1204 // FIXME: This should always be handled by InstSimplify?
1205 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1206 return replaceInstUsesWith(I, TorF);
1209 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1210 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1211 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1214 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1216 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1217 if (I.getPredicate() == I.ICMP_NE)
1218 Pred = CmpInst::getInversePredicate(Pred);
1219 return new ICmpInst(Pred, LHS, RHS);
1222 // Don't bother doing any work for cases which InstSimplify handles.
1223 if (AP2.isNullValue())
1226 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1228 if (!AP1 && AP2TrailingZeros != 0)
1231 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1234 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1236 // Get the distance between the lowest bits that are set.
1237 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1239 if (Shift > 0 && AP2.shl(Shift) == AP1)
1240 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1242 // Shifting const2 will never be equal to const1.
1243 // FIXME: This should always be handled by InstSimplify?
1244 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1245 return replaceInstUsesWith(I, TorF);
1248 /// The caller has matched a pattern of the form:
1249 /// I = icmp ugt (add (add A, B), CI2), CI1
1250 /// If this is of the form:
1252 /// if (sum+128 >u 255)
1253 /// Then replace it with llvm.sadd.with.overflow.i8.
1255 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1256 ConstantInt *CI2, ConstantInt *CI1,
1258 // The transformation we're trying to do here is to transform this into an
1259 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1260 // with a narrower add, and discard the add-with-constant that is part of the
1261 // range check (if we can't eliminate it, this isn't profitable).
1263 // In order to eliminate the add-with-constant, the compare can be its only
1265 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1266 if (!AddWithCst->hasOneUse())
1269 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1270 if (!CI2->getValue().isPowerOf2())
1272 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1273 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1276 // The width of the new add formed is 1 more than the bias.
1279 // Check to see that CI1 is an all-ones value with NewWidth bits.
1280 if (CI1->getBitWidth() == NewWidth ||
1281 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1284 // This is only really a signed overflow check if the inputs have been
1285 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1286 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1287 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1288 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1289 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1292 // In order to replace the original add with a narrower
1293 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1294 // and truncates that discard the high bits of the add. Verify that this is
1296 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1297 for (User *U : OrigAdd->users()) {
1298 if (U == AddWithCst)
1301 // Only accept truncates for now. We would really like a nice recursive
1302 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1303 // chain to see which bits of a value are actually demanded. If the
1304 // original add had another add which was then immediately truncated, we
1305 // could still do the transformation.
1306 TruncInst *TI = dyn_cast<TruncInst>(U);
1307 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1311 // If the pattern matches, truncate the inputs to the narrower type and
1312 // use the sadd_with_overflow intrinsic to efficiently compute both the
1313 // result and the overflow bit.
1314 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1315 Function *F = Intrinsic::getDeclaration(
1316 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1318 InstCombiner::BuilderTy &Builder = IC.Builder;
1320 // Put the new code above the original add, in case there are any uses of the
1321 // add between the add and the compare.
1322 Builder.SetInsertPoint(OrigAdd);
1324 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1325 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1326 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1327 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1328 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1330 // The inner add was the result of the narrow add, zero extended to the
1331 // wider type. Replace it with the result computed by the intrinsic.
1332 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1334 // The original icmp gets replaced with the overflow value.
1335 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1339 /// icmp eq/ne (urem/srem %x, %y), 0
1340 /// iff %y is a power-of-two, we can replace this with a bit test:
1341 /// icmp eq/ne (and %x, (add %y, -1)), 0
1342 Instruction *InstCombiner::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1343 // This fold is only valid for equality predicates.
1344 if (!I.isEquality())
1346 ICmpInst::Predicate Pred;
1347 Value *X, *Y, *Zero;
1348 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1349 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1351 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1353 // This may increase instruction count, we don't enforce that Y is a constant.
1354 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1355 Value *Masked = Builder.CreateAnd(X, Mask);
1356 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1359 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1360 /// by one-less-than-bitwidth into a sign test on the original value.
1361 Instruction *InstCombiner::foldSignBitTest(ICmpInst &I) {
1363 ICmpInst::Predicate Pred;
1364 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1371 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1373 unsigned XBitWidth = XTy->getScalarSizeInBits();
1374 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1375 APInt(XBitWidth, XBitWidth - 1))))
1377 } else if (isa<BinaryOperator>(Val) &&
1378 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1379 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1380 /*AnalyzeForSignBitExtraction=*/true))) {
1385 return ICmpInst::Create(Instruction::ICmp,
1386 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1387 : ICmpInst::ICMP_SLT,
1388 X, ConstantInt::getNullValue(XTy));
1391 // Handle icmp pred X, 0
1392 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1393 CmpInst::Predicate Pred = Cmp.getPredicate();
1394 if (!match(Cmp.getOperand(1), m_Zero()))
1397 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1398 if (Pred == ICmpInst::ICMP_SGT) {
1400 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1401 if (SPR.Flavor == SPF_SMIN) {
1402 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1403 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1404 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1405 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1409 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1413 // icmp eq/ne (urem %x, %y), 0
1414 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1417 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1418 ICmpInst::isEquality(Pred)) {
1419 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1420 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1421 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1422 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1428 /// Fold icmp Pred X, C.
1429 /// TODO: This code structure does not make sense. The saturating add fold
1430 /// should be moved to some other helper and extended as noted below (it is also
1431 /// possible that code has been made unnecessary - do we canonicalize IR to
1432 /// overflow/saturating intrinsics or not?).
1433 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1434 // Match the following pattern, which is a common idiom when writing
1435 // overflow-safe integer arithmetic functions. The source performs an addition
1436 // in wider type and explicitly checks for overflow using comparisons against
1437 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1439 // TODO: This could probably be generalized to handle other overflow-safe
1440 // operations if we worked out the formulas to compute the appropriate magic
1444 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1445 CmpInst::Predicate Pred = Cmp.getPredicate();
1446 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1448 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1449 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1450 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1451 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1457 /// Canonicalize icmp instructions based on dominating conditions.
1458 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1459 // This is a cheap/incomplete check for dominance - just match a single
1460 // predecessor with a conditional branch.
1461 BasicBlock *CmpBB = Cmp.getParent();
1462 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1467 BasicBlock *TrueBB, *FalseBB;
1468 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1471 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1472 "Predecessor block does not point to successor?");
1474 // The branch should get simplified. Don't bother simplifying this condition.
1475 if (TrueBB == FalseBB)
1478 // Try to simplify this compare to T/F based on the dominating condition.
1479 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1481 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1483 CmpInst::Predicate Pred = Cmp.getPredicate();
1484 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1485 ICmpInst::Predicate DomPred;
1486 const APInt *C, *DomC;
1487 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1488 match(Y, m_APInt(C))) {
1489 // We have 2 compares of a variable with constants. Calculate the constant
1490 // ranges of those compares to see if we can transform the 2nd compare:
1492 // DomCond = icmp DomPred X, DomC
1493 // br DomCond, CmpBB, FalseBB
1495 // Cmp = icmp Pred X, C
1496 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1497 ConstantRange DominatingCR =
1498 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1499 : ConstantRange::makeExactICmpRegion(
1500 CmpInst::getInversePredicate(DomPred), *DomC);
1501 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1502 ConstantRange Difference = DominatingCR.difference(CR);
1503 if (Intersection.isEmptySet())
1504 return replaceInstUsesWith(Cmp, Builder.getFalse());
1505 if (Difference.isEmptySet())
1506 return replaceInstUsesWith(Cmp, Builder.getTrue());
1508 // Canonicalizing a sign bit comparison that gets used in a branch,
1509 // pessimizes codegen by generating branch on zero instruction instead
1510 // of a test and branch. So we avoid canonicalizing in such situations
1511 // because test and branch instruction has better branch displacement
1512 // than compare and branch instruction.
1514 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1515 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1518 if (const APInt *EqC = Intersection.getSingleElement())
1519 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1520 if (const APInt *NeC = Difference.getSingleElement())
1521 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1527 /// Fold icmp (trunc X, Y), C.
1528 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1531 ICmpInst::Predicate Pred = Cmp.getPredicate();
1532 Value *X = Trunc->getOperand(0);
1533 if (C.isOneValue() && C.getBitWidth() > 1) {
1534 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1536 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1537 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1538 ConstantInt::get(V->getType(), 1));
1541 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1542 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1543 // of the high bits truncated out of x are known.
1544 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1545 SrcBits = X->getType()->getScalarSizeInBits();
1546 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1548 // If all the high bits are known, we can do this xform.
1549 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1550 // Pull in the high bits from known-ones set.
1551 APInt NewRHS = C.zext(SrcBits);
1552 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1553 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1560 /// Fold icmp (xor X, Y), C.
1561 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1562 BinaryOperator *Xor,
1564 Value *X = Xor->getOperand(0);
1565 Value *Y = Xor->getOperand(1);
1567 if (!match(Y, m_APInt(XorC)))
1570 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1572 ICmpInst::Predicate Pred = Cmp.getPredicate();
1573 bool TrueIfSigned = false;
1574 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1576 // If the sign bit of the XorCst is not set, there is no change to
1577 // the operation, just stop using the Xor.
1578 if (!XorC->isNegative()) {
1579 Cmp.setOperand(0, X);
1584 // Emit the opposite comparison.
1586 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1587 ConstantInt::getAllOnesValue(X->getType()));
1589 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1590 ConstantInt::getNullValue(X->getType()));
1593 if (Xor->hasOneUse()) {
1594 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1595 if (!Cmp.isEquality() && XorC->isSignMask()) {
1596 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1597 : Cmp.getSignedPredicate();
1598 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1601 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1602 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1603 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1604 : Cmp.getSignedPredicate();
1605 Pred = Cmp.getSwappedPredicate(Pred);
1606 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1610 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1611 if (Pred == ICmpInst::ICMP_UGT) {
1612 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1613 if (*XorC == ~C && (C + 1).isPowerOf2())
1614 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1615 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1616 if (*XorC == C && (C + 1).isPowerOf2())
1617 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1619 if (Pred == ICmpInst::ICMP_ULT) {
1620 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1621 if (*XorC == -C && C.isPowerOf2())
1622 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1623 ConstantInt::get(X->getType(), ~C));
1624 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1625 if (*XorC == C && (-C).isPowerOf2())
1626 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1627 ConstantInt::get(X->getType(), ~C));
1632 /// Fold icmp (and (sh X, Y), C2), C1.
1633 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1634 const APInt &C1, const APInt &C2) {
1635 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1636 if (!Shift || !Shift->isShift())
1639 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1640 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1641 // code produced by the clang front-end, for bitfield access.
1642 // This seemingly simple opportunity to fold away a shift turns out to be
1643 // rather complicated. See PR17827 for details.
1644 unsigned ShiftOpcode = Shift->getOpcode();
1645 bool IsShl = ShiftOpcode == Instruction::Shl;
1647 if (match(Shift->getOperand(1), m_APInt(C3))) {
1648 bool CanFold = false;
1649 if (ShiftOpcode == Instruction::Shl) {
1650 // For a left shift, we can fold if the comparison is not signed. We can
1651 // also fold a signed comparison if the mask value and comparison value
1652 // are not negative. These constraints may not be obvious, but we can
1653 // prove that they are correct using an SMT solver.
1654 if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1657 bool IsAshr = ShiftOpcode == Instruction::AShr;
1658 // For a logical right shift, we can fold if the comparison is not signed.
1659 // We can also fold a signed comparison if the shifted mask value and the
1660 // shifted comparison value are not negative. These constraints may not be
1661 // obvious, but we can prove that they are correct using an SMT solver.
1662 // For an arithmetic shift right we can do the same, if we ensure
1663 // the And doesn't use any bits being shifted in. Normally these would
1664 // be turned into lshr by SimplifyDemandedBits, but not if there is an
1666 if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1667 if (!Cmp.isSigned() ||
1668 (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1674 APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1675 APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1676 // Check to see if we are shifting out any of the bits being compared.
1677 if (SameAsC1 != C1) {
1678 // If we shifted bits out, the fold is not going to work out. As a
1679 // special case, check to see if this means that the result is always
1680 // true or false now.
1681 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1682 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1683 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1684 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1686 Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1687 APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1688 And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1689 And->setOperand(0, Shift->getOperand(0));
1690 Worklist.Add(Shift); // Shift is dead.
1696 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1697 // preferable because it allows the C2 << Y expression to be hoisted out of a
1698 // loop if Y is invariant and X is not.
1699 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1700 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1703 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1704 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1706 // Compute X & (C2 << Y).
1707 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1708 Cmp.setOperand(0, NewAnd);
1715 /// Fold icmp (and X, C2), C1.
1716 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1717 BinaryOperator *And,
1719 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1721 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1722 // TODO: We canonicalize to the longer form for scalars because we have
1723 // better analysis/folds for icmp, and codegen may be better with icmp.
1724 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1725 match(And->getOperand(1), m_One()))
1726 return new TruncInst(And->getOperand(0), Cmp.getType());
1730 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1733 // Don't perform the following transforms if the AND has multiple uses
1734 if (!And->hasOneUse())
1737 if (Cmp.isEquality() && C1.isNullValue()) {
1738 // Restrict this fold to single-use 'and' (PR10267).
1739 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1740 if (C2->isSignMask()) {
1741 Constant *Zero = Constant::getNullValue(X->getType());
1742 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1743 return new ICmpInst(NewPred, X, Zero);
1746 // Restrict this fold only for single-use 'and' (PR10267).
1747 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1748 if ((~(*C2) + 1).isPowerOf2()) {
1750 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1751 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1752 return new ICmpInst(NewPred, X, NegBOC);
1756 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1757 // the input width without changing the value produced, eliminate the cast:
1759 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1761 // We can do this transformation if the constants do not have their sign bits
1762 // set or if it is an equality comparison. Extending a relational comparison
1763 // when we're checking the sign bit would not work.
1765 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1766 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1767 // TODO: Is this a good transform for vectors? Wider types may reduce
1768 // throughput. Should this transform be limited (even for scalars) by using
1769 // shouldChangeType()?
1770 if (!Cmp.getType()->isVectorTy()) {
1771 Type *WideType = W->getType();
1772 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1773 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1774 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1775 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1776 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1780 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1783 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1784 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1786 // iff pred isn't signed
1787 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1788 match(And->getOperand(1), m_One())) {
1789 Constant *One = cast<Constant>(And->getOperand(1));
1790 Value *Or = And->getOperand(0);
1791 Value *A, *B, *LShr;
1792 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1793 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1794 unsigned UsesRemoved = 0;
1795 if (And->hasOneUse())
1797 if (Or->hasOneUse())
1799 if (LShr->hasOneUse())
1802 // Compute A & ((1 << B) | 1)
1803 Value *NewOr = nullptr;
1804 if (auto *C = dyn_cast<Constant>(B)) {
1805 if (UsesRemoved >= 1)
1806 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1808 if (UsesRemoved >= 3)
1809 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1811 One, Or->getName());
1814 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1815 Cmp.setOperand(0, NewAnd);
1824 /// Fold icmp (and X, Y), C.
1825 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1826 BinaryOperator *And,
1828 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1831 // TODO: These all require that Y is constant too, so refactor with the above.
1833 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1834 Value *X = And->getOperand(0);
1835 Value *Y = And->getOperand(1);
1836 if (auto *LI = dyn_cast<LoadInst>(X))
1837 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1838 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1839 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1840 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1841 ConstantInt *C2 = cast<ConstantInt>(Y);
1842 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1846 if (!Cmp.isEquality())
1849 // X & -C == -C -> X > u ~C
1850 // X & -C != -C -> X <= u ~C
1851 // iff C is a power of 2
1852 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1853 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1854 : CmpInst::ICMP_ULE;
1855 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1858 // (X & C2) == 0 -> (trunc X) >= 0
1859 // (X & C2) != 0 -> (trunc X) < 0
1860 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1862 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1863 int32_t ExactLogBase2 = C2->exactLogBase2();
1864 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1865 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1866 if (And->getType()->isVectorTy())
1867 NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1868 Value *Trunc = Builder.CreateTrunc(X, NTy);
1869 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1870 : CmpInst::ICMP_SLT;
1871 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1878 /// Fold icmp (or X, Y), C.
1879 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1881 ICmpInst::Predicate Pred = Cmp.getPredicate();
1882 if (C.isOneValue()) {
1883 // icmp slt signum(V) 1 --> icmp slt V, 1
1885 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1886 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1887 ConstantInt::get(V->getType(), 1));
1890 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1891 if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1892 // X | C == C --> X <=u C
1893 // X | C != C --> X >u C
1894 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1895 if ((C + 1).isPowerOf2()) {
1896 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1897 return new ICmpInst(Pred, OrOp0, OrOp1);
1899 // More general: are all bits outside of a mask constant set or not set?
1900 // X | C == C --> (X & ~C) == 0
1901 // X | C != C --> (X & ~C) != 0
1902 if (Or->hasOneUse()) {
1903 Value *A = Builder.CreateAnd(OrOp0, ~C);
1904 return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1908 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1912 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1913 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1914 // -> and (icmp eq P, null), (icmp eq Q, null).
1916 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1918 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1919 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1920 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1923 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1924 // a shorter form that has more potential to be folded even further.
1925 Value *X1, *X2, *X3, *X4;
1926 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1927 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1928 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1929 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1930 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1931 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1932 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1933 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1939 /// Fold icmp (mul X, Y), C.
1940 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1941 BinaryOperator *Mul,
1944 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1947 // If this is a test of the sign bit and the multiply is sign-preserving with
1948 // a constant operand, use the multiply LHS operand instead.
1949 ICmpInst::Predicate Pred = Cmp.getPredicate();
1950 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1951 if (MulC->isNegative())
1952 Pred = ICmpInst::getSwappedPredicate(Pred);
1953 return new ICmpInst(Pred, Mul->getOperand(0),
1954 Constant::getNullValue(Mul->getType()));
1960 /// Fold icmp (shl 1, Y), C.
1961 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
1964 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1967 Type *ShiftType = Shl->getType();
1968 unsigned TypeBits = C.getBitWidth();
1969 bool CIsPowerOf2 = C.isPowerOf2();
1970 ICmpInst::Predicate Pred = Cmp.getPredicate();
1971 if (Cmp.isUnsigned()) {
1972 // (1 << Y) pred C -> Y pred Log2(C)
1974 // (1 << Y) < 30 -> Y <= 4
1975 // (1 << Y) <= 30 -> Y <= 4
1976 // (1 << Y) >= 30 -> Y > 4
1977 // (1 << Y) > 30 -> Y > 4
1978 if (Pred == ICmpInst::ICMP_ULT)
1979 Pred = ICmpInst::ICMP_ULE;
1980 else if (Pred == ICmpInst::ICMP_UGE)
1981 Pred = ICmpInst::ICMP_UGT;
1984 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1985 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1986 unsigned CLog2 = C.logBase2();
1987 if (CLog2 == TypeBits - 1) {
1988 if (Pred == ICmpInst::ICMP_UGE)
1989 Pred = ICmpInst::ICMP_EQ;
1990 else if (Pred == ICmpInst::ICMP_ULT)
1991 Pred = ICmpInst::ICMP_NE;
1993 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1994 } else if (Cmp.isSigned()) {
1995 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1996 if (C.isAllOnesValue()) {
1997 // (1 << Y) <= -1 -> Y == 31
1998 if (Pred == ICmpInst::ICMP_SLE)
1999 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2001 // (1 << Y) > -1 -> Y != 31
2002 if (Pred == ICmpInst::ICMP_SGT)
2003 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2005 // (1 << Y) < 0 -> Y == 31
2006 // (1 << Y) <= 0 -> Y == 31
2007 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2008 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2010 // (1 << Y) >= 0 -> Y != 31
2011 // (1 << Y) > 0 -> Y != 31
2012 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2013 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2015 } else if (Cmp.isEquality() && CIsPowerOf2) {
2016 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2022 /// Fold icmp (shl X, Y), C.
2023 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
2024 BinaryOperator *Shl,
2026 const APInt *ShiftVal;
2027 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2028 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2030 const APInt *ShiftAmt;
2031 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2032 return foldICmpShlOne(Cmp, Shl, C);
2034 // Check that the shift amount is in range. If not, don't perform undefined
2035 // shifts. When the shift is visited, it will be simplified.
2036 unsigned TypeBits = C.getBitWidth();
2037 if (ShiftAmt->uge(TypeBits))
2040 ICmpInst::Predicate Pred = Cmp.getPredicate();
2041 Value *X = Shl->getOperand(0);
2042 Type *ShType = Shl->getType();
2044 // NSW guarantees that we are only shifting out sign bits from the high bits,
2045 // so we can ASHR the compare constant without needing a mask and eliminate
2047 if (Shl->hasNoSignedWrap()) {
2048 if (Pred == ICmpInst::ICMP_SGT) {
2049 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2050 APInt ShiftedC = C.ashr(*ShiftAmt);
2051 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2053 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2054 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2055 APInt ShiftedC = C.ashr(*ShiftAmt);
2056 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2058 if (Pred == ICmpInst::ICMP_SLT) {
2059 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2060 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2061 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2062 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2063 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2064 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2065 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2067 // If this is a signed comparison to 0 and the shift is sign preserving,
2068 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2069 // do that if we're sure to not continue on in this function.
2070 if (isSignTest(Pred, C))
2071 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2074 // NUW guarantees that we are only shifting out zero bits from the high bits,
2075 // so we can LSHR the compare constant without needing a mask and eliminate
2077 if (Shl->hasNoUnsignedWrap()) {
2078 if (Pred == ICmpInst::ICMP_UGT) {
2079 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2080 APInt ShiftedC = C.lshr(*ShiftAmt);
2081 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2083 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2084 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2085 APInt ShiftedC = C.lshr(*ShiftAmt);
2086 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2088 if (Pred == ICmpInst::ICMP_ULT) {
2089 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2090 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2091 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2092 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2093 assert(C.ugt(0) && "ult 0 should have been eliminated");
2094 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2095 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2099 if (Cmp.isEquality() && Shl->hasOneUse()) {
2100 // Strength-reduce the shift into an 'and'.
2101 Constant *Mask = ConstantInt::get(
2103 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2104 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2105 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2106 return new ICmpInst(Pred, And, LShrC);
2109 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2110 bool TrueIfSigned = false;
2111 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2112 // (X << 31) <s 0 --> (X & 1) != 0
2113 Constant *Mask = ConstantInt::get(
2115 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2116 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2117 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2118 And, Constant::getNullValue(ShType));
2121 // Simplify 'shl' inequality test into 'and' equality test.
2122 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2123 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2124 if ((C + 1).isPowerOf2() &&
2125 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2126 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2127 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2128 : ICmpInst::ICMP_NE,
2129 And, Constant::getNullValue(ShType));
2131 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2132 if (C.isPowerOf2() &&
2133 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2135 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2136 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2137 : ICmpInst::ICMP_NE,
2138 And, Constant::getNullValue(ShType));
2142 // Transform (icmp pred iM (shl iM %v, N), C)
2143 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2144 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2145 // This enables us to get rid of the shift in favor of a trunc that may be
2146 // free on the target. It has the additional benefit of comparing to a
2147 // smaller constant that may be more target-friendly.
2148 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2149 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2150 DL.isLegalInteger(TypeBits - Amt)) {
2151 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2152 if (ShType->isVectorTy())
2153 TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2155 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2156 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2162 /// Fold icmp ({al}shr X, Y), C.
2163 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2164 BinaryOperator *Shr,
2166 // An exact shr only shifts out zero bits, so:
2167 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2168 Value *X = Shr->getOperand(0);
2169 CmpInst::Predicate Pred = Cmp.getPredicate();
2170 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2172 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2174 const APInt *ShiftVal;
2175 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2176 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2178 const APInt *ShiftAmt;
2179 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2182 // Check that the shift amount is in range. If not, don't perform undefined
2183 // shifts. When the shift is visited it will be simplified.
2184 unsigned TypeBits = C.getBitWidth();
2185 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2186 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2189 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2190 bool IsExact = Shr->isExact();
2191 Type *ShrTy = Shr->getType();
2192 // TODO: If we could guarantee that InstSimplify would handle all of the
2193 // constant-value-based preconditions in the folds below, then we could assert
2194 // those conditions rather than checking them. This is difficult because of
2195 // undef/poison (PR34838).
2197 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2198 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2199 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2200 APInt ShiftedC = C.shl(ShAmtVal);
2201 if (ShiftedC.ashr(ShAmtVal) == C)
2202 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2204 if (Pred == CmpInst::ICMP_SGT) {
2205 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2206 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2207 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2208 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2209 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2212 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2213 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2214 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2215 APInt ShiftedC = C.shl(ShAmtVal);
2216 if (ShiftedC.lshr(ShAmtVal) == C)
2217 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2219 if (Pred == CmpInst::ICMP_UGT) {
2220 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2221 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2222 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2223 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2227 if (!Cmp.isEquality())
2230 // Handle equality comparisons of shift-by-constant.
2232 // If the comparison constant changes with the shift, the comparison cannot
2233 // succeed (bits of the comparison constant cannot match the shifted value).
2234 // This should be known by InstSimplify and already be folded to true/false.
2235 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2236 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2237 "Expected icmp+shr simplify did not occur.");
2239 // If the bits shifted out are known zero, compare the unshifted value:
2240 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2242 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2244 if (Shr->hasOneUse()) {
2245 // Canonicalize the shift into an 'and':
2246 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2247 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2248 Constant *Mask = ConstantInt::get(ShrTy, Val);
2249 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2250 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2256 Instruction *InstCombiner::foldICmpSRemConstant(ICmpInst &Cmp,
2257 BinaryOperator *SRem,
2259 // Match an 'is positive' or 'is negative' comparison of remainder by a
2260 // constant power-of-2 value:
2261 // (X % pow2C) sgt/slt 0
2262 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2263 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2266 // TODO: The one-use check is standard because we do not typically want to
2267 // create longer instruction sequences, but this might be a special-case
2268 // because srem is not good for analysis or codegen.
2269 if (!SRem->hasOneUse())
2272 const APInt *DivisorC;
2273 if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2276 // Mask off the sign bit and the modulo bits (low-bits).
2277 Type *Ty = SRem->getType();
2278 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2279 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2280 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2282 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2283 // bit is set. Example:
2284 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2285 if (Pred == ICmpInst::ICMP_SGT)
2286 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2288 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2289 // bit is set. Example:
2290 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2291 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2294 /// Fold icmp (udiv X, Y), C.
2295 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2296 BinaryOperator *UDiv,
2299 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2302 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2304 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2305 Value *Y = UDiv->getOperand(1);
2306 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2307 assert(!C.isMaxValue() &&
2308 "icmp ugt X, UINT_MAX should have been simplified already.");
2309 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2310 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2313 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2314 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2315 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2316 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2317 ConstantInt::get(Y->getType(), C2->udiv(C)));
2323 /// Fold icmp ({su}div X, Y), C.
2324 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2325 BinaryOperator *Div,
2327 // Fold: icmp pred ([us]div X, C2), C -> range test
2328 // Fold this div into the comparison, producing a range check.
2329 // Determine, based on the divide type, what the range is being
2330 // checked. If there is an overflow on the low or high side, remember
2331 // it, otherwise compute the range [low, hi) bounding the new value.
2332 // See: InsertRangeTest above for the kinds of replacements possible.
2334 if (!match(Div->getOperand(1), m_APInt(C2)))
2337 // FIXME: If the operand types don't match the type of the divide
2338 // then don't attempt this transform. The code below doesn't have the
2339 // logic to deal with a signed divide and an unsigned compare (and
2340 // vice versa). This is because (x /s C2) <s C produces different
2341 // results than (x /s C2) <u C or (x /u C2) <s C or even
2342 // (x /u C2) <u C. Simply casting the operands and result won't
2343 // work. :( The if statement below tests that condition and bails
2345 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2346 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2349 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2350 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2351 // division-by-constant cases should be present, we can not assert that they
2352 // have happened before we reach this icmp instruction.
2353 if (C2->isNullValue() || C2->isOneValue() ||
2354 (DivIsSigned && C2->isAllOnesValue()))
2357 // Compute Prod = C * C2. We are essentially solving an equation of
2358 // form X / C2 = C. We solve for X by multiplying C2 and C.
2359 // By solving for X, we can turn this into a range check instead of computing
2361 APInt Prod = C * *C2;
2363 // Determine if the product overflows by seeing if the product is not equal to
2364 // the divide. Make sure we do the same kind of divide as in the LHS
2365 // instruction that we're folding.
2366 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2368 ICmpInst::Predicate Pred = Cmp.getPredicate();
2370 // If the division is known to be exact, then there is no remainder from the
2371 // divide, so the covered range size is unit, otherwise it is the divisor.
2372 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2374 // Figure out the interval that is being checked. For example, a comparison
2375 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2376 // Compute this interval based on the constants involved and the signedness of
2377 // the compare/divide. This computes a half-open interval, keeping track of
2378 // whether either value in the interval overflows. After analysis each
2379 // overflow variable is set to 0 if it's corresponding bound variable is valid
2380 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2381 int LoOverflow = 0, HiOverflow = 0;
2382 APInt LoBound, HiBound;
2384 if (!DivIsSigned) { // udiv
2385 // e.g. X/5 op 3 --> [15, 20)
2387 HiOverflow = LoOverflow = ProdOV;
2389 // If this is not an exact divide, then many values in the range collapse
2390 // to the same result value.
2391 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2393 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2394 if (C.isNullValue()) { // (X / pos) op 0
2395 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2396 LoBound = -(RangeSize - 1);
2397 HiBound = RangeSize;
2398 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2399 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2400 HiOverflow = LoOverflow = ProdOV;
2402 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2403 } else { // (X / pos) op neg
2404 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2406 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2408 APInt DivNeg = -RangeSize;
2409 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2412 } else if (C2->isNegative()) { // Divisor is < 0.
2415 if (C.isNullValue()) { // (X / neg) op 0
2416 // e.g. X/-5 op 0 --> [-4, 5)
2417 LoBound = RangeSize + 1;
2418 HiBound = -RangeSize;
2419 if (HiBound == *C2) { // -INTMIN = INTMIN
2420 HiOverflow = 1; // [INTMIN+1, overflow)
2421 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2423 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2424 // e.g. X/-5 op 3 --> [-19, -14)
2426 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2428 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2429 } else { // (X / neg) op neg
2430 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2431 LoOverflow = HiOverflow = ProdOV;
2433 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2436 // Dividing by a negative swaps the condition. LT <-> GT
2437 Pred = ICmpInst::getSwappedPredicate(Pred);
2440 Value *X = Div->getOperand(0);
2442 default: llvm_unreachable("Unhandled icmp opcode!");
2443 case ICmpInst::ICMP_EQ:
2444 if (LoOverflow && HiOverflow)
2445 return replaceInstUsesWith(Cmp, Builder.getFalse());
2447 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2448 ICmpInst::ICMP_UGE, X,
2449 ConstantInt::get(Div->getType(), LoBound));
2451 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2452 ICmpInst::ICMP_ULT, X,
2453 ConstantInt::get(Div->getType(), HiBound));
2454 return replaceInstUsesWith(
2455 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2456 case ICmpInst::ICMP_NE:
2457 if (LoOverflow && HiOverflow)
2458 return replaceInstUsesWith(Cmp, Builder.getTrue());
2460 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2461 ICmpInst::ICMP_ULT, X,
2462 ConstantInt::get(Div->getType(), LoBound));
2464 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2465 ICmpInst::ICMP_UGE, X,
2466 ConstantInt::get(Div->getType(), HiBound));
2467 return replaceInstUsesWith(Cmp,
2468 insertRangeTest(X, LoBound, HiBound,
2469 DivIsSigned, false));
2470 case ICmpInst::ICMP_ULT:
2471 case ICmpInst::ICMP_SLT:
2472 if (LoOverflow == +1) // Low bound is greater than input range.
2473 return replaceInstUsesWith(Cmp, Builder.getTrue());
2474 if (LoOverflow == -1) // Low bound is less than input range.
2475 return replaceInstUsesWith(Cmp, Builder.getFalse());
2476 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2477 case ICmpInst::ICMP_UGT:
2478 case ICmpInst::ICMP_SGT:
2479 if (HiOverflow == +1) // High bound greater than input range.
2480 return replaceInstUsesWith(Cmp, Builder.getFalse());
2481 if (HiOverflow == -1) // High bound less than input range.
2482 return replaceInstUsesWith(Cmp, Builder.getTrue());
2483 if (Pred == ICmpInst::ICMP_UGT)
2484 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2485 ConstantInt::get(Div->getType(), HiBound));
2486 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2487 ConstantInt::get(Div->getType(), HiBound));
2493 /// Fold icmp (sub X, Y), C.
2494 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2495 BinaryOperator *Sub,
2497 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2498 ICmpInst::Predicate Pred = Cmp.getPredicate();
2502 // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2503 if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2504 return new ICmpInst(Cmp.getPredicate(), Y,
2505 ConstantInt::get(Y->getType(), 0));
2507 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2508 if (match(X, m_APInt(C2)) &&
2509 ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2510 (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2511 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2512 return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2513 ConstantInt::get(Y->getType(), SubResult));
2515 // The following transforms are only worth it if the only user of the subtract
2517 if (!Sub->hasOneUse())
2520 if (Sub->hasNoSignedWrap()) {
2521 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2522 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2523 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2525 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2526 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2527 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2529 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2530 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2531 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2533 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2534 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2535 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2538 if (!match(X, m_APInt(C2)))
2541 // C2 - Y <u C -> (Y | (C - 1)) == C2
2542 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2543 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2544 (*C2 & (C - 1)) == (C - 1))
2545 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2547 // C2 - Y >u C -> (Y | C) != C2
2548 // iff C2 & C == C and C + 1 is a power of 2
2549 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2550 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2555 /// Fold icmp (add X, Y), C.
2556 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2557 BinaryOperator *Add,
2559 Value *Y = Add->getOperand(1);
2561 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2564 // Fold icmp pred (add X, C2), C.
2565 Value *X = Add->getOperand(0);
2566 Type *Ty = Add->getType();
2567 CmpInst::Predicate Pred = Cmp.getPredicate();
2569 // If the add does not wrap, we can always adjust the compare by subtracting
2570 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2571 // are canonicalized to SGT/SLT/UGT/ULT.
2572 if ((Add->hasNoSignedWrap() &&
2573 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2574 (Add->hasNoUnsignedWrap() &&
2575 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2578 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2579 // If there is overflow, the result must be true or false.
2580 // TODO: Can we assert there is no overflow because InstSimplify always
2581 // handles those cases?
2583 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2584 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2587 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2588 const APInt &Upper = CR.getUpper();
2589 const APInt &Lower = CR.getLower();
2590 if (Cmp.isSigned()) {
2591 if (Lower.isSignMask())
2592 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2593 if (Upper.isSignMask())
2594 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2596 if (Lower.isMinValue())
2597 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2598 if (Upper.isMinValue())
2599 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2602 if (!Add->hasOneUse())
2605 // X+C <u C2 -> (X & -C2) == C
2606 // iff C & (C2-1) == 0
2607 // C2 is a power of 2
2608 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2609 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2610 ConstantExpr::getNeg(cast<Constant>(Y)));
2612 // X+C >u C2 -> (X & ~C2) != C
2614 // C2+1 is a power of 2
2615 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2616 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2617 ConstantExpr::getNeg(cast<Constant>(Y)));
2622 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2623 Value *&RHS, ConstantInt *&Less,
2624 ConstantInt *&Equal,
2625 ConstantInt *&Greater) {
2626 // TODO: Generalize this to work with other comparison idioms or ensure
2627 // they get canonicalized into this form.
2629 // select i1 (a == b),
2631 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2632 // where Equal, Less and Greater are placeholders for any three constants.
2633 ICmpInst::Predicate PredA;
2634 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2635 !ICmpInst::isEquality(PredA))
2637 Value *EqualVal = SI->getTrueValue();
2638 Value *UnequalVal = SI->getFalseValue();
2639 // We still can get non-canonical predicate here, so canonicalize.
2640 if (PredA == ICmpInst::ICMP_NE)
2641 std::swap(EqualVal, UnequalVal);
2642 if (!match(EqualVal, m_ConstantInt(Equal)))
2644 ICmpInst::Predicate PredB;
2646 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2647 m_ConstantInt(Less), m_ConstantInt(Greater))))
2649 // We can get predicate mismatch here, so canonicalize if possible:
2650 // First, ensure that 'LHS' match.
2652 // x sgt y <--> y slt x
2653 std::swap(LHS2, RHS2);
2654 PredB = ICmpInst::getSwappedPredicate(PredB);
2658 // We also need to canonicalize 'RHS'.
2659 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2660 // x sgt C-1 <--> x sge C <--> not(x slt C)
2661 auto FlippedStrictness =
2662 getFlippedStrictnessPredicateAndConstant(PredB, cast<Constant>(RHS2));
2663 if (!FlippedStrictness)
2665 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check");
2666 RHS2 = FlippedStrictness->second;
2667 // And kind-of perform the result swap.
2668 std::swap(Less, Greater);
2669 PredB = ICmpInst::ICMP_SLT;
2671 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2674 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2678 assert(C && "Cmp RHS should be a constant int!");
2679 // If we're testing a constant value against the result of a three way
2680 // comparison, the result can be expressed directly in terms of the
2681 // original values being compared. Note: We could possibly be more
2682 // aggressive here and remove the hasOneUse test. The original select is
2683 // really likely to simplify or sink when we remove a test of the result.
2684 Value *OrigLHS, *OrigRHS;
2685 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2686 if (Cmp.hasOneUse() &&
2687 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2689 assert(C1LessThan && C2Equal && C3GreaterThan);
2691 bool TrueWhenLessThan =
2692 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2694 bool TrueWhenEqual =
2695 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2697 bool TrueWhenGreaterThan =
2698 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2701 // This generates the new instruction that will replace the original Cmp
2702 // Instruction. Instead of enumerating the various combinations when
2703 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2704 // false, we rely on chaining of ORs and future passes of InstCombine to
2705 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2707 // When none of the three constants satisfy the predicate for the RHS (C),
2708 // the entire original Cmp can be simplified to a false.
2709 Value *Cond = Builder.getFalse();
2710 if (TrueWhenLessThan)
2711 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2714 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2716 if (TrueWhenGreaterThan)
2717 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2720 return replaceInstUsesWith(Cmp, Cond);
2725 static Instruction *foldICmpBitCast(ICmpInst &Cmp,
2726 InstCombiner::BuilderTy &Builder) {
2727 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2731 ICmpInst::Predicate Pred = Cmp.getPredicate();
2732 Value *Op1 = Cmp.getOperand(1);
2733 Value *BCSrcOp = Bitcast->getOperand(0);
2735 // Make sure the bitcast doesn't change the number of vector elements.
2736 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2737 Bitcast->getDestTy()->getScalarSizeInBits()) {
2738 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2740 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2741 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2742 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2743 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2744 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2745 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2746 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2747 match(Op1, m_Zero()))
2748 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2750 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2751 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2752 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2754 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2755 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2756 return new ICmpInst(Pred, X,
2757 ConstantInt::getAllOnesValue(X->getType()));
2760 // Zero-equality checks are preserved through unsigned floating-point casts:
2761 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2762 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2763 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2764 if (Cmp.isEquality() && match(Op1, m_Zero()))
2765 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2768 // Test to see if the operands of the icmp are casted versions of other
2769 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2770 if (Bitcast->getType()->isPointerTy() &&
2771 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2772 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2773 // so eliminate it as well.
2774 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2775 Op1 = BC2->getOperand(0);
2777 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2778 return new ICmpInst(Pred, BCSrcOp, Op1);
2781 // Folding: icmp <pred> iN X, C
2782 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2783 // and C is a splat of a K-bit pattern
2784 // and SC is a constant vector = <C', C', C', ..., C'>
2786 // %E = extractelement <M x iK> %vec, i32 C'
2787 // icmp <pred> iK %E, trunc(C)
2789 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2790 !Bitcast->getType()->isIntegerTy() ||
2791 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2797 m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2798 // Check whether every element of Mask is the same constant
2799 if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2800 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2801 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2802 if (C->isSplat(EltTy->getBitWidth())) {
2803 // Fold the icmp based on the value of C
2804 // If C is M copies of an iK sized bit pattern,
2806 // => %E = extractelement <N x iK> %vec, i32 Elem
2807 // icmp <pred> iK %SplatVal, <pattern>
2808 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2809 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2810 return new ICmpInst(Pred, Extract, NewC);
2817 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2818 /// where X is some kind of instruction.
2819 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2821 if (!match(Cmp.getOperand(1), m_APInt(C)))
2824 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2825 switch (BO->getOpcode()) {
2826 case Instruction::Xor:
2827 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2830 case Instruction::And:
2831 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2834 case Instruction::Or:
2835 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2838 case Instruction::Mul:
2839 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2842 case Instruction::Shl:
2843 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2846 case Instruction::LShr:
2847 case Instruction::AShr:
2848 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2851 case Instruction::SRem:
2852 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2855 case Instruction::UDiv:
2856 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2859 case Instruction::SDiv:
2860 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2863 case Instruction::Sub:
2864 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2867 case Instruction::Add:
2868 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2874 // TODO: These folds could be refactored to be part of the above calls.
2875 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2879 // Match against CmpInst LHS being instructions other than binary operators.
2881 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2882 // For now, we only support constant integers while folding the
2883 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2884 // similar to the cases handled by binary ops above.
2885 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2886 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2890 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2891 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2895 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2896 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2902 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2903 /// icmp eq/ne BO, C.
2904 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2907 // TODO: Some of these folds could work with arbitrary constants, but this
2908 // function is limited to scalar and vector splat constants.
2909 if (!Cmp.isEquality())
2912 ICmpInst::Predicate Pred = Cmp.getPredicate();
2913 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2914 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2915 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2917 switch (BO->getOpcode()) {
2918 case Instruction::SRem:
2919 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2920 if (C.isNullValue() && BO->hasOneUse()) {
2922 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2923 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2924 return new ICmpInst(Pred, NewRem,
2925 Constant::getNullValue(BO->getType()));
2929 case Instruction::Add: {
2930 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2932 if (match(BOp1, m_APInt(BOC))) {
2933 if (BO->hasOneUse()) {
2934 Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2935 return new ICmpInst(Pred, BOp0, SubC);
2937 } else if (C.isNullValue()) {
2938 // Replace ((add A, B) != 0) with (A != -B) if A or B is
2939 // efficiently invertible, or if the add has just this one use.
2940 if (Value *NegVal = dyn_castNegVal(BOp1))
2941 return new ICmpInst(Pred, BOp0, NegVal);
2942 if (Value *NegVal = dyn_castNegVal(BOp0))
2943 return new ICmpInst(Pred, NegVal, BOp1);
2944 if (BO->hasOneUse()) {
2945 Value *Neg = Builder.CreateNeg(BOp1);
2947 return new ICmpInst(Pred, BOp0, Neg);
2952 case Instruction::Xor:
2953 if (BO->hasOneUse()) {
2954 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2955 // For the xor case, we can xor two constants together, eliminating
2956 // the explicit xor.
2957 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2958 } else if (C.isNullValue()) {
2959 // Replace ((xor A, B) != 0) with (A != B)
2960 return new ICmpInst(Pred, BOp0, BOp1);
2964 case Instruction::Sub:
2965 if (BO->hasOneUse()) {
2967 if (match(BOp0, m_APInt(BOC))) {
2968 // Replace ((sub BOC, B) != C) with (B != BOC-C).
2969 Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2970 return new ICmpInst(Pred, BOp1, SubC);
2971 } else if (C.isNullValue()) {
2972 // Replace ((sub A, B) != 0) with (A != B).
2973 return new ICmpInst(Pred, BOp0, BOp1);
2977 case Instruction::Or: {
2979 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2980 // Comparing if all bits outside of a constant mask are set?
2981 // Replace (X | C) == -1 with (X & ~C) == ~C.
2982 // This removes the -1 constant.
2983 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2984 Value *And = Builder.CreateAnd(BOp0, NotBOC);
2985 return new ICmpInst(Pred, And, NotBOC);
2989 case Instruction::And: {
2991 if (match(BOp1, m_APInt(BOC))) {
2992 // If we have ((X & C) == C), turn it into ((X & C) != 0).
2993 if (C == *BOC && C.isPowerOf2())
2994 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2995 BO, Constant::getNullValue(RHS->getType()));
2999 case Instruction::Mul:
3000 if (C.isNullValue() && BO->hasNoSignedWrap()) {
3002 if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
3003 // The trivial case (mul X, 0) is handled by InstSimplify.
3004 // General case : (mul X, C) != 0 iff X != 0
3005 // (mul X, C) == 0 iff X == 0
3006 return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
3010 case Instruction::UDiv:
3011 if (C.isNullValue()) {
3012 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3013 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3014 return new ICmpInst(NewPred, BOp1, BOp0);
3023 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3024 Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
3027 Type *Ty = II->getType();
3028 unsigned BitWidth = C.getBitWidth();
3029 switch (II->getIntrinsicID()) {
3030 case Intrinsic::bswap:
3032 Cmp.setOperand(0, II->getArgOperand(0));
3033 Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
3036 case Intrinsic::ctlz:
3037 case Intrinsic::cttz: {
3038 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3039 if (C == BitWidth) {
3041 Cmp.setOperand(0, II->getArgOperand(0));
3042 Cmp.setOperand(1, ConstantInt::getNullValue(Ty));
3046 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3047 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3048 // Limit to one use to ensure we don't increase instruction count.
3049 unsigned Num = C.getLimitedValue(BitWidth);
3050 if (Num != BitWidth && II->hasOneUse()) {
3051 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3052 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3053 : APInt::getHighBitsSet(BitWidth, Num + 1);
3054 APInt Mask2 = IsTrailing
3055 ? APInt::getOneBitSet(BitWidth, Num)
3056 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3057 Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
3058 Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
3065 case Intrinsic::ctpop: {
3066 // popcount(A) == 0 -> A == 0 and likewise for !=
3067 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3068 bool IsZero = C.isNullValue();
3069 if (IsZero || C == BitWidth) {
3071 Cmp.setOperand(0, II->getArgOperand(0));
3073 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty);
3074 Cmp.setOperand(1, NewOp);
3080 case Intrinsic::uadd_sat: {
3081 // uadd.sat(a, b) == 0 -> (a | b) == 0
3082 if (C.isNullValue()) {
3083 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3084 return replaceInstUsesWith(Cmp, Builder.CreateICmp(
3085 Cmp.getPredicate(), Or, Constant::getNullValue(Ty)));
3091 case Intrinsic::usub_sat: {
3092 // usub.sat(a, b) == 0 -> a <= b
3093 if (C.isNullValue()) {
3094 ICmpInst::Predicate NewPred = Cmp.getPredicate() == ICmpInst::ICMP_EQ
3095 ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3096 return ICmpInst::Create(Instruction::ICmp, NewPred,
3097 II->getArgOperand(0), II->getArgOperand(1));
3108 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3109 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3112 if (Cmp.isEquality())
3113 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3115 Type *Ty = II->getType();
3116 unsigned BitWidth = C.getBitWidth();
3117 switch (II->getIntrinsicID()) {
3118 case Intrinsic::ctlz: {
3119 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3120 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3121 unsigned Num = C.getLimitedValue();
3122 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3123 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3124 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3127 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3128 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3129 C.uge(1) && C.ule(BitWidth)) {
3130 unsigned Num = C.getLimitedValue();
3131 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3132 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3133 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3137 case Intrinsic::cttz: {
3138 // Limit to one use to ensure we don't increase instruction count.
3139 if (!II->hasOneUse())
3142 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3143 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3144 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3145 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3146 Builder.CreateAnd(II->getArgOperand(0), Mask),
3147 ConstantInt::getNullValue(Ty));
3150 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3151 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3152 C.uge(1) && C.ule(BitWidth)) {
3153 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3154 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3155 Builder.CreateAnd(II->getArgOperand(0), Mask),
3156 ConstantInt::getNullValue(Ty));
3167 /// Handle icmp with constant (but not simple integer constant) RHS.
3168 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3169 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3170 Constant *RHSC = dyn_cast<Constant>(Op1);
3171 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3175 switch (LHSI->getOpcode()) {
3176 case Instruction::GetElementPtr:
3177 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3178 if (RHSC->isNullValue() &&
3179 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3180 return new ICmpInst(
3181 I.getPredicate(), LHSI->getOperand(0),
3182 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3184 case Instruction::PHI:
3185 // Only fold icmp into the PHI if the phi and icmp are in the same
3186 // block. If in the same block, we're encouraging jump threading. If
3187 // not, we are just pessimizing the code by making an i1 phi.
3188 if (LHSI->getParent() == I.getParent())
3189 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3192 case Instruction::Select: {
3193 // If either operand of the select is a constant, we can fold the
3194 // comparison into the select arms, which will cause one to be
3195 // constant folded and the select turned into a bitwise or.
3196 Value *Op1 = nullptr, *Op2 = nullptr;
3197 ConstantInt *CI = nullptr;
3198 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3199 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3200 CI = dyn_cast<ConstantInt>(Op1);
3202 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3203 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3204 CI = dyn_cast<ConstantInt>(Op2);
3207 // We only want to perform this transformation if it will not lead to
3208 // additional code. This is true if either both sides of the select
3209 // fold to a constant (in which case the icmp is replaced with a select
3210 // which will usually simplify) or this is the only user of the
3211 // select (in which case we are trading a select+icmp for a simpler
3212 // select+icmp) or all uses of the select can be replaced based on
3213 // dominance information ("Global cases").
3214 bool Transform = false;
3217 else if (Op1 || Op2) {
3219 if (LHSI->hasOneUse())
3222 else if (CI && !CI->isZero())
3223 // When Op1 is constant try replacing select with second operand.
3224 // Otherwise Op2 is constant and try replacing select with first
3227 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3231 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3234 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3236 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3240 case Instruction::IntToPtr:
3241 // icmp pred inttoptr(X), null -> icmp pred X, 0
3242 if (RHSC->isNullValue() &&
3243 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3244 return new ICmpInst(
3245 I.getPredicate(), LHSI->getOperand(0),
3246 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3249 case Instruction::Load:
3250 // Try to optimize things like "A[i] > 4" to index computations.
3251 if (GetElementPtrInst *GEP =
3252 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3253 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3254 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3255 !cast<LoadInst>(LHSI)->isVolatile())
3256 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3265 /// Some comparisons can be simplified.
3266 /// In this case, we are looking for comparisons that look like
3267 /// a check for a lossy truncation.
3269 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3270 /// Where Mask is some pattern that produces all-ones in low bits:
3272 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3274 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3275 /// The Mask can be a constant, too.
3276 /// For some predicates, the operands are commutative.
3277 /// For others, x can only be on a specific side.
3278 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3279 InstCombiner::BuilderTy &Builder) {
3280 ICmpInst::Predicate SrcPred;
3282 auto m_VariableMask = m_CombineOr(
3283 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3284 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3285 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3286 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3287 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3288 if (!match(&I, m_c_ICmp(SrcPred,
3289 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3293 ICmpInst::Predicate DstPred;
3295 case ICmpInst::Predicate::ICMP_EQ:
3296 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3297 DstPred = ICmpInst::Predicate::ICMP_ULE;
3299 case ICmpInst::Predicate::ICMP_NE:
3300 // x & (-1 >> y) != x -> x u> (-1 >> y)
3301 DstPred = ICmpInst::Predicate::ICMP_UGT;
3303 case ICmpInst::Predicate::ICMP_UGT:
3304 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3305 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3306 DstPred = ICmpInst::Predicate::ICMP_UGT;
3308 case ICmpInst::Predicate::ICMP_UGE:
3309 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3310 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3311 DstPred = ICmpInst::Predicate::ICMP_ULE;
3313 case ICmpInst::Predicate::ICMP_ULT:
3314 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3315 assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3316 DstPred = ICmpInst::Predicate::ICMP_UGT;
3318 case ICmpInst::Predicate::ICMP_ULE:
3319 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3320 assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3321 DstPred = ICmpInst::Predicate::ICMP_ULE;
3323 case ICmpInst::Predicate::ICMP_SGT:
3324 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3325 if (X != I.getOperand(0)) // X must be on LHS of comparison!
3326 return nullptr; // Ignore the other case.
3327 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3329 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3331 DstPred = ICmpInst::Predicate::ICMP_SGT;
3333 case ICmpInst::Predicate::ICMP_SGE:
3334 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3335 if (X != I.getOperand(1)) // X must be on RHS of comparison!
3336 return nullptr; // Ignore the other case.
3337 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3339 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3341 DstPred = ICmpInst::Predicate::ICMP_SLE;
3343 case ICmpInst::Predicate::ICMP_SLT:
3344 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3345 if (X != I.getOperand(1)) // X must be on RHS of comparison!
3346 return nullptr; // Ignore the other case.
3347 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3349 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3351 DstPred = ICmpInst::Predicate::ICMP_SGT;
3353 case ICmpInst::Predicate::ICMP_SLE:
3354 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3355 if (X != I.getOperand(0)) // X must be on LHS of comparison!
3356 return nullptr; // Ignore the other case.
3357 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3359 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3361 DstPred = ICmpInst::Predicate::ICMP_SLE;
3364 llvm_unreachable("All possible folds are handled.");
3367 // The mask value may be a vector constant that has undefined elements. But it
3368 // may not be safe to propagate those undefs into the new compare, so replace
3369 // those elements by copying an existing, defined, and safe scalar constant.
3370 Type *OpTy = M->getType();
3371 auto *VecC = dyn_cast<Constant>(M);
3372 if (OpTy->isVectorTy() && VecC && VecC->containsUndefElement()) {
3373 Constant *SafeReplacementConstant = nullptr;
3374 for (unsigned i = 0, e = OpTy->getVectorNumElements(); i != e; ++i) {
3375 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3376 SafeReplacementConstant = VecC->getAggregateElement(i);
3380 assert(SafeReplacementConstant && "Failed to find undef replacement");
3381 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3384 return Builder.CreateICmp(DstPred, X, M);
3387 /// Some comparisons can be simplified.
3388 /// In this case, we are looking for comparisons that look like
3389 /// a check for a lossy signed truncation.
3390 /// Folds: (MaskedBits is a constant.)
3391 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3393 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3394 /// Where KeptBits = bitwidth(%x) - MaskedBits
3396 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3397 InstCombiner::BuilderTy &Builder) {
3398 ICmpInst::Predicate SrcPred;
3400 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3401 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3402 if (!match(&I, m_c_ICmp(SrcPred,
3403 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3408 // Potential handling of non-splats: for each element:
3409 // * if both are undef, replace with constant 0.
3410 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3411 // * if both are not undef, and are different, bailout.
3412 // * else, only one is undef, then pick the non-undef one.
3414 // The shift amount must be equal.
3417 const APInt &MaskedBits = *C0;
3418 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3420 ICmpInst::Predicate DstPred;
3422 case ICmpInst::Predicate::ICMP_EQ:
3423 // ((%x << MaskedBits) a>> MaskedBits) == %x
3425 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3426 DstPred = ICmpInst::Predicate::ICMP_ULT;
3428 case ICmpInst::Predicate::ICMP_NE:
3429 // ((%x << MaskedBits) a>> MaskedBits) != %x
3431 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3432 DstPred = ICmpInst::Predicate::ICMP_UGE;
3434 // FIXME: are more folds possible?
3439 auto *XType = X->getType();
3440 const unsigned XBitWidth = XType->getScalarSizeInBits();
3441 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3442 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3444 // KeptBits = bitwidth(%x) - MaskedBits
3445 const APInt KeptBits = BitWidth - MaskedBits;
3446 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3447 // ICmpCst = (1 << KeptBits)
3448 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3449 assert(ICmpCst.isPowerOf2());
3450 // AddCst = (1 << (KeptBits-1))
3451 const APInt AddCst = ICmpCst.lshr(1);
3452 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3454 // T0 = add %x, AddCst
3455 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3456 // T1 = T0 DstPred ICmpCst
3457 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3463 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3464 // we should move shifts to the same hand of 'and', i.e. rewrite as
3465 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3466 // We are only interested in opposite logical shifts here.
3467 // One of the shifts can be truncated.
3468 // If we can, we want to end up creating 'lshr' shift.
3470 foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3471 InstCombiner::BuilderTy &Builder) {
3472 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3473 !I.getOperand(0)->hasOneUse())
3476 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3478 // Look for an 'and' of two logical shifts, one of which may be truncated.
3479 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3480 Instruction *XShift, *MaybeTruncation, *YShift;
3483 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3484 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3485 m_AnyLogicalShift, m_Instruction(YShift))),
3486 m_Instruction(MaybeTruncation)))))
3489 // We potentially looked past 'trunc', but only when matching YShift,
3490 // therefore YShift must have the widest type.
3491 Instruction *WidestShift = YShift;
3492 // Therefore XShift must have the shallowest type.
3493 // Or they both have identical types if there was no truncation.
3494 Instruction *NarrowestShift = XShift;
3496 Type *WidestTy = WidestShift->getType();
3497 Type *NarrowestTy = NarrowestShift->getType();
3498 assert(NarrowestTy == I.getOperand(0)->getType() &&
3499 "We did not look past any shifts while matching XShift though.");
3500 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3502 // If YShift is a 'lshr', swap the shifts around.
3503 if (match(YShift, m_LShr(m_Value(), m_Value())))
3504 std::swap(XShift, YShift);
3506 // The shifts must be in opposite directions.
3507 auto XShiftOpcode = XShift->getOpcode();
3508 if (XShiftOpcode == YShift->getOpcode())
3509 return nullptr; // Do not care about same-direction shifts here.
3511 Value *X, *XShAmt, *Y, *YShAmt;
3512 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3513 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3515 // If one of the values being shifted is a constant, then we will end with
3516 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3517 // however, we will need to ensure that we won't increase instruction count.
3518 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3519 // At least one of the hands of the 'and' should be one-use shift.
3520 if (!match(I.getOperand(0),
3521 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3524 // Due to the 'trunc', we will need to widen X. For that either the old
3525 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3526 if (!MaybeTruncation->hasOneUse() &&
3527 !NarrowestShift->getOperand(1)->hasOneUse())
3532 // We have two shift amounts from two different shifts. The types of those
3533 // shift amounts may not match. If that's the case let's bailout now.
3534 if (XShAmt->getType() != YShAmt->getType())
3537 // As input, we have the following pattern:
3538 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3539 // We want to rewrite that as:
3540 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3541 // While we know that originally (Q+K) would not overflow
3542 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3543 // shift amounts. so it may now overflow in smaller bitwidth.
3544 // To ensure that does not happen, we need to ensure that the total maximal
3545 // shift amount is still representable in that smaller bit width.
3546 unsigned MaximalPossibleTotalShiftAmount =
3547 (WidestTy->getScalarSizeInBits() - 1) +
3548 (NarrowestTy->getScalarSizeInBits() - 1);
3549 APInt MaximalRepresentableShiftAmount =
3550 APInt::getAllOnesValue(XShAmt->getType()->getScalarSizeInBits());
3551 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3554 // Can we fold (XShAmt+YShAmt) ?
3555 auto *NewShAmt = dyn_cast_or_null<Constant>(
3556 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3557 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3560 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3561 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3563 // Is the new shift amount smaller than the bit width?
3564 // FIXME: could also rely on ConstantRange.
3565 if (!match(NewShAmt,
3566 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3567 APInt(WidestBitWidth, WidestBitWidth))))
3570 // An extra legality check is needed if we had trunc-of-lshr.
3571 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3572 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3574 // It isn't obvious whether it's worth it to analyze non-constants here.
3575 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3576 // If *any* of these preconditions matches we can perform the fold.
3577 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3578 ? NewShAmt->getSplatValue()
3580 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3581 if (NewShAmtSplat &&
3582 (NewShAmtSplat->isNullValue() ||
3583 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3585 // We consider *min* leading zeros so a single outlier
3586 // blocks the transform as opposed to allowing it.
3587 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3588 KnownBits Known = computeKnownBits(C, SQ.DL);
3589 unsigned MinLeadZero = Known.countMinLeadingZeros();
3590 // If the value being shifted has at most lowest bit set we can fold.
3591 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3592 if (MaxActiveBits <= 1)
3594 // Precondition: NewShAmt u<= countLeadingZeros(C)
3595 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3598 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3599 KnownBits Known = computeKnownBits(C, SQ.DL);
3600 unsigned MinLeadZero = Known.countMinLeadingZeros();
3601 // If the value being shifted has at most lowest bit set we can fold.
3602 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3603 if (MaxActiveBits <= 1)
3605 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3606 if (NewShAmtSplat) {
3608 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3609 if (AdjNewShAmt.ule(MinLeadZero))
3613 return false; // Can't tell if it's ok.
3619 // All good, we can do this fold.
3620 X = Builder.CreateZExt(X, WidestTy);
3621 Y = Builder.CreateZExt(Y, WidestTy);
3622 // The shift is the same that was for X.
3623 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3624 ? Builder.CreateLShr(X, NewShAmt)
3625 : Builder.CreateShl(X, NewShAmt);
3626 Value *T1 = Builder.CreateAnd(T0, Y);
3627 return Builder.CreateICmp(I.getPredicate(), T1,
3628 Constant::getNullValue(WidestTy));
3633 /// ((x * y) u/ x) != y
3635 /// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3636 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3637 /// will mean that we are looking for the opposite answer.
3638 Value *InstCombiner::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3639 ICmpInst::Predicate Pred;
3643 // Look for: (-1 u/ x) u</u>= y
3644 if (!I.isEquality() &&
3645 match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3648 // Canonicalize as-if y was on RHS.
3649 if (I.getOperand(1) != Y)
3650 Pred = I.getSwappedPredicate();
3652 // Are we checking that overflow does not happen, or does happen?
3654 case ICmpInst::Predicate::ICMP_ULT:
3655 NeedNegation = false;
3657 case ICmpInst::Predicate::ICMP_UGE:
3658 NeedNegation = true;
3661 return nullptr; // Wrong predicate.
3663 } else // Look for: ((x * y) u/ x) !=/== y
3664 if (I.isEquality() &&
3665 match(&I, m_c_ICmp(Pred, m_Value(Y),
3666 m_OneUse(m_UDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3668 m_Instruction(Mul)),
3669 m_Deferred(X)))))) {
3670 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3674 BuilderTy::InsertPointGuard Guard(Builder);
3675 // If the pattern included (x * y), we'll want to insert new instructions
3676 // right before that original multiplication so that we can replace it.
3677 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3678 if (MulHadOtherUses)
3679 Builder.SetInsertPoint(Mul);
3681 Function *F = Intrinsic::getDeclaration(
3682 I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3683 CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3685 // If the multiplication was used elsewhere, to ensure that we don't leave
3686 // "duplicate" instructions, replace uses of that original multiplication
3687 // with the multiplication result from the with.overflow intrinsic.
3688 if (MulHadOtherUses)
3689 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3691 Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3692 if (NeedNegation) // This technically increases instruction count.
3693 Res = Builder.CreateNot(Res, "umul.not.ov");
3698 /// Try to fold icmp (binop), X or icmp X, (binop).
3699 /// TODO: A large part of this logic is duplicated in InstSimplify's
3700 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3702 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I, const SimplifyQuery &SQ) {
3703 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3704 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3706 // Special logic for binary operators.
3707 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3708 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3712 const CmpInst::Predicate Pred = I.getPredicate();
3715 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3716 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3717 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3718 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3719 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3720 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3721 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3722 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3723 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3725 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3726 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3728 ICmpInst::isEquality(Pred) ||
3729 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3730 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3731 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3733 ICmpInst::isEquality(Pred) ||
3734 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3735 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3737 // Analyze the case when either Op0 or Op1 is an add instruction.
3738 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3739 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3740 if (BO0 && BO0->getOpcode() == Instruction::Add) {
3741 A = BO0->getOperand(0);
3742 B = BO0->getOperand(1);
3744 if (BO1 && BO1->getOpcode() == Instruction::Add) {
3745 C = BO1->getOperand(0);
3746 D = BO1->getOperand(1);
3749 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3750 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3751 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3752 return new ICmpInst(Pred, A == Op1 ? B : A,
3753 Constant::getNullValue(Op1->getType()));
3755 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3756 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3757 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3758 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3761 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3762 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3764 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3767 // C + B == C + D -> B == D
3770 } else if (A == D) {
3771 // D + B == C + D -> B == C
3774 } else if (B == C) {
3775 // A + C == C + D -> A == D
3780 // A + D == C + D -> A == C
3784 return new ICmpInst(Pred, Y, Z);
3787 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3788 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3789 match(B, m_AllOnes()))
3790 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3792 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3793 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3794 match(B, m_AllOnes()))
3795 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3797 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3798 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3799 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3801 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3802 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3803 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3805 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3806 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3807 match(D, m_AllOnes()))
3808 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3810 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3811 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3812 match(D, m_AllOnes()))
3813 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3815 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3816 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3817 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3819 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3820 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3821 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3823 // TODO: The subtraction-related identities shown below also hold, but
3824 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3825 // wouldn't happen even if they were implemented.
3827 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3828 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3829 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3830 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3832 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3833 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3834 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3836 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3837 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3838 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3840 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3841 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3842 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3844 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3845 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3846 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3848 // if C1 has greater magnitude than C2:
3849 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3850 // s.t. C3 = C1 - C2
3852 // if C2 has greater magnitude than C1:
3853 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3854 // s.t. C3 = C2 - C1
3855 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3856 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3857 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3858 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3859 const APInt &AP1 = C1->getValue();
3860 const APInt &AP2 = C2->getValue();
3861 if (AP1.isNegative() == AP2.isNegative()) {
3862 APInt AP1Abs = C1->getValue().abs();
3863 APInt AP2Abs = C2->getValue().abs();
3864 if (AP1Abs.uge(AP2Abs)) {
3865 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3866 Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3867 return new ICmpInst(Pred, NewAdd, C);
3869 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3870 Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3871 return new ICmpInst(Pred, A, NewAdd);
3876 // Analyze the case when either Op0 or Op1 is a sub instruction.
3877 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3882 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3883 A = BO0->getOperand(0);
3884 B = BO0->getOperand(1);
3886 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3887 C = BO1->getOperand(0);
3888 D = BO1->getOperand(1);
3891 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3892 if (A == Op1 && NoOp0WrapProblem)
3893 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3894 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
3895 if (C == Op0 && NoOp1WrapProblem)
3896 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3898 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3899 // (A - B) u>/u<= A --> B u>/u<= A
3900 if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3901 return new ICmpInst(Pred, B, A);
3902 // C u</u>= (C - D) --> C u</u>= D
3903 if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3904 return new ICmpInst(Pred, C, D);
3905 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
3906 if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
3907 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3908 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
3909 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
3910 if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
3911 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3912 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
3914 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
3915 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
3916 return new ICmpInst(Pred, A, C);
3918 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
3919 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
3920 return new ICmpInst(Pred, D, B);
3922 // icmp (0-X) < cst --> x > -cst
3923 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3925 if (match(BO0, m_Neg(m_Value(X))))
3926 if (Constant *RHSC = dyn_cast<Constant>(Op1))
3927 if (RHSC->isNotMinSignedValue())
3928 return new ICmpInst(I.getSwappedPredicate(), X,
3929 ConstantExpr::getNeg(RHSC));
3932 BinaryOperator *SRem = nullptr;
3933 // icmp (srem X, Y), Y
3934 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3936 // icmp Y, (srem X, Y)
3937 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3938 Op0 == BO1->getOperand(1))
3941 // We don't check hasOneUse to avoid increasing register pressure because
3942 // the value we use is the same value this instruction was already using.
3943 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3946 case ICmpInst::ICMP_EQ:
3947 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3948 case ICmpInst::ICMP_NE:
3949 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3950 case ICmpInst::ICMP_SGT:
3951 case ICmpInst::ICMP_SGE:
3952 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3953 Constant::getAllOnesValue(SRem->getType()));
3954 case ICmpInst::ICMP_SLT:
3955 case ICmpInst::ICMP_SLE:
3956 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3957 Constant::getNullValue(SRem->getType()));
3961 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3962 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3963 switch (BO0->getOpcode()) {
3966 case Instruction::Add:
3967 case Instruction::Sub:
3968 case Instruction::Xor: {
3969 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3970 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3973 if (match(BO0->getOperand(1), m_APInt(C))) {
3974 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3975 if (C->isSignMask()) {
3976 ICmpInst::Predicate NewPred =
3977 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3978 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3981 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3982 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3983 ICmpInst::Predicate NewPred =
3984 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
3985 NewPred = I.getSwappedPredicate(NewPred);
3986 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3991 case Instruction::Mul: {
3992 if (!I.isEquality())
3996 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3998 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3999 // Mask = -1 >> count-trailing-zeros(C).
4000 if (unsigned TZs = C->countTrailingZeros()) {
4001 Constant *Mask = ConstantInt::get(
4003 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4004 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4005 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4006 return new ICmpInst(Pred, And1, And2);
4008 // If there are no trailing zeros in the multiplier, just eliminate
4009 // the multiplies (no masking is needed):
4010 // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
4011 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4015 case Instruction::UDiv:
4016 case Instruction::LShr:
4017 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4019 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4021 case Instruction::SDiv:
4022 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4024 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4026 case Instruction::AShr:
4027 if (!BO0->isExact() || !BO1->isExact())
4029 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4031 case Instruction::Shl: {
4032 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4033 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4036 if (!NSW && I.isSigned())
4038 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4044 // Transform A & (L - 1) `ult` L --> L != 0
4045 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4046 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4048 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4049 auto *Zero = Constant::getNullValue(BO0->getType());
4050 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4054 if (Value *V = foldUnsignedMultiplicationOverflowCheck(I))
4055 return replaceInstUsesWith(I, V);
4057 if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
4058 return replaceInstUsesWith(I, V);
4060 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
4061 return replaceInstUsesWith(I, V);
4063 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
4064 return replaceInstUsesWith(I, V);
4069 /// Fold icmp Pred min|max(X, Y), X.
4070 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4071 ICmpInst::Predicate Pred = Cmp.getPredicate();
4072 Value *Op0 = Cmp.getOperand(0);
4073 Value *X = Cmp.getOperand(1);
4075 // Canonicalize minimum or maximum operand to LHS of the icmp.
4076 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4077 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4078 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4079 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4081 Pred = Cmp.getSwappedPredicate();
4085 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4086 // smin(X, Y) == X --> X s<= Y
4087 // smin(X, Y) s>= X --> X s<= Y
4088 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4089 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4091 // smin(X, Y) != X --> X s> Y
4092 // smin(X, Y) s< X --> X s> Y
4093 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4094 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4096 // These cases should be handled in InstSimplify:
4097 // smin(X, Y) s<= X --> true
4098 // smin(X, Y) s> X --> false
4102 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4103 // smax(X, Y) == X --> X s>= Y
4104 // smax(X, Y) s<= X --> X s>= Y
4105 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4106 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4108 // smax(X, Y) != X --> X s< Y
4109 // smax(X, Y) s> X --> X s< Y
4110 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4111 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4113 // These cases should be handled in InstSimplify:
4114 // smax(X, Y) s>= X --> true
4115 // smax(X, Y) s< X --> false
4119 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4120 // umin(X, Y) == X --> X u<= Y
4121 // umin(X, Y) u>= X --> X u<= Y
4122 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4123 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4125 // umin(X, Y) != X --> X u> Y
4126 // umin(X, Y) u< X --> X u> Y
4127 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4128 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4130 // These cases should be handled in InstSimplify:
4131 // umin(X, Y) u<= X --> true
4132 // umin(X, Y) u> X --> false
4136 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4137 // umax(X, Y) == X --> X u>= Y
4138 // umax(X, Y) u<= X --> X u>= Y
4139 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4140 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4142 // umax(X, Y) != X --> X u< Y
4143 // umax(X, Y) u> X --> X u< Y
4144 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4145 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4147 // These cases should be handled in InstSimplify:
4148 // umax(X, Y) u>= X --> true
4149 // umax(X, Y) u< X --> false
4156 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
4157 if (!I.isEquality())
4160 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4161 const CmpInst::Predicate Pred = I.getPredicate();
4162 Value *A, *B, *C, *D;
4163 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4164 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4165 Value *OtherVal = A == Op1 ? B : A;
4166 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4169 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4170 // A^c1 == C^c2 --> A == C^(c1^c2)
4171 ConstantInt *C1, *C2;
4172 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4174 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4175 Value *Xor = Builder.CreateXor(C, NC);
4176 return new ICmpInst(Pred, A, Xor);
4179 // A^B == A^D -> B == D
4181 return new ICmpInst(Pred, B, D);
4183 return new ICmpInst(Pred, B, C);
4185 return new ICmpInst(Pred, A, D);
4187 return new ICmpInst(Pred, A, C);
4191 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4192 // A == (A^B) -> B == 0
4193 Value *OtherVal = A == Op0 ? B : A;
4194 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4197 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4198 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4199 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4200 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4206 } else if (A == D) {
4210 } else if (B == C) {
4214 } else if (B == D) {
4220 if (X) { // Build (X^Y) & Z
4221 Op1 = Builder.CreateXor(X, Y);
4222 Op1 = Builder.CreateAnd(Op1, Z);
4223 I.setOperand(0, Op1);
4224 I.setOperand(1, Constant::getNullValue(Op1->getType()));
4229 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4230 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4232 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4233 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4234 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4235 match(Op1, m_ZExt(m_Value(A))))) {
4236 APInt Pow2 = Cst1->getValue() + 1;
4237 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4238 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4239 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4242 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4243 // For lshr and ashr pairs.
4244 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4245 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4246 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4247 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4248 unsigned TypeBits = Cst1->getBitWidth();
4249 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4250 if (ShAmt < TypeBits && ShAmt != 0) {
4251 ICmpInst::Predicate NewPred =
4252 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4253 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4254 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4255 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4259 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4260 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4261 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4262 unsigned TypeBits = Cst1->getBitWidth();
4263 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4264 if (ShAmt < TypeBits && ShAmt != 0) {
4265 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4266 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4267 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4268 I.getName() + ".mask");
4269 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4273 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4274 // "icmp (and X, mask), cst"
4276 if (Op0->hasOneUse() &&
4277 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4278 match(Op1, m_ConstantInt(Cst1)) &&
4279 // Only do this when A has multiple uses. This is most important to do
4280 // when it exposes other optimizations.
4282 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4284 if (ShAmt < ASize) {
4286 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4289 APInt CmpV = Cst1->getValue().zext(ASize);
4292 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4293 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4297 // If both operands are byte-swapped or bit-reversed, just compare the
4299 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4300 // and handle more intrinsics.
4301 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4302 (match(Op0, m_BitReverse(m_Value(A))) &&
4303 match(Op1, m_BitReverse(m_Value(B)))))
4304 return new ICmpInst(Pred, A, B);
4306 // Canonicalize checking for a power-of-2-or-zero value:
4307 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4308 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4309 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4311 !match(Op1, m_ZeroInt()))
4314 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4315 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4316 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4319 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4323 Type *Ty = A->getType();
4324 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4325 return Pred == ICmpInst::ICMP_EQ
4326 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4327 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4333 static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4334 InstCombiner::BuilderTy &Builder) {
4335 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
4336 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4338 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4341 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4342 bool IsSignedCmp = ICmp.isSigned();
4343 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4344 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4345 // and the other is a zext), then we can't handle this.
4346 // TODO: This is too strict. We can handle some predicates (equality?).
4347 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4350 // Not an extension from the same type?
4351 Value *Y = CastOp1->getOperand(0);
4352 Type *XTy = X->getType(), *YTy = Y->getType();
4354 // One of the casts must have one use because we are creating a new cast.
4355 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4357 // Extend the narrower operand to the type of the wider operand.
4358 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4359 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4360 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4361 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4366 // (zext X) == (zext Y) --> X == Y
4367 // (sext X) == (sext Y) --> X == Y
4368 if (ICmp.isEquality())
4369 return new ICmpInst(ICmp.getPredicate(), X, Y);
4371 // A signed comparison of sign extended values simplifies into a
4372 // signed comparison.
4373 if (IsSignedCmp && IsSignedExt)
4374 return new ICmpInst(ICmp.getPredicate(), X, Y);
4376 // The other three cases all fold into an unsigned comparison.
4377 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4380 // Below here, we are only folding a compare with constant.
4381 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4385 // Compute the constant that would happen if we truncated to SrcTy then
4386 // re-extended to DestTy.
4387 Type *SrcTy = CastOp0->getSrcTy();
4388 Type *DestTy = CastOp0->getDestTy();
4389 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4390 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4392 // If the re-extended constant didn't change...
4394 if (ICmp.isEquality())
4395 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4397 // A signed comparison of sign extended values simplifies into a
4398 // signed comparison.
4399 if (IsSignedExt && IsSignedCmp)
4400 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4402 // The other three cases all fold into an unsigned comparison.
4403 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4406 // The re-extended constant changed, partly changed (in the case of a vector),
4407 // or could not be determined to be equal (in the case of a constant
4408 // expression), so the constant cannot be represented in the shorter type.
4409 // All the cases that fold to true or false will have already been handled
4410 // by SimplifyICmpInst, so only deal with the tricky case.
4411 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4414 // Is source op positive?
4415 // icmp ult (sext X), C --> icmp sgt X, -1
4416 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4417 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4419 // Is source op negative?
4420 // icmp ugt (sext X), C --> icmp slt X, 0
4421 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4422 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4425 /// Handle icmp (cast x), (cast or constant).
4426 Instruction *InstCombiner::foldICmpWithCastOp(ICmpInst &ICmp) {
4427 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4430 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4433 Value *Op0Src = CastOp0->getOperand(0);
4434 Type *SrcTy = CastOp0->getSrcTy();
4435 Type *DestTy = CastOp0->getDestTy();
4437 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4438 // integer type is the same size as the pointer type.
4439 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4440 if (isa<VectorType>(SrcTy)) {
4441 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4442 DestTy = cast<VectorType>(DestTy)->getElementType();
4444 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4446 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4447 CompatibleSizes(SrcTy, DestTy)) {
4448 Value *NewOp1 = nullptr;
4449 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4450 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4451 if (PtrSrc->getType()->getPointerAddressSpace() ==
4452 Op0Src->getType()->getPointerAddressSpace()) {
4453 NewOp1 = PtrToIntOp1->getOperand(0);
4454 // If the pointer types don't match, insert a bitcast.
4455 if (Op0Src->getType() != NewOp1->getType())
4456 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4458 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4459 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4463 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4466 return foldICmpWithZextOrSext(ICmp, Builder);
4469 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4472 llvm_unreachable("Unsupported binary op");
4473 case Instruction::Add:
4474 case Instruction::Sub:
4475 return match(RHS, m_Zero());
4476 case Instruction::Mul:
4477 return match(RHS, m_One());
4481 OverflowResult InstCombiner::computeOverflow(
4482 Instruction::BinaryOps BinaryOp, bool IsSigned,
4483 Value *LHS, Value *RHS, Instruction *CxtI) const {
4486 llvm_unreachable("Unsupported binary op");
4487 case Instruction::Add:
4489 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4491 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4492 case Instruction::Sub:
4494 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4496 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4497 case Instruction::Mul:
4499 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4501 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4505 bool InstCombiner::OptimizeOverflowCheck(
4506 Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS,
4507 Instruction &OrigI, Value *&Result, Constant *&Overflow) {
4508 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4509 std::swap(LHS, RHS);
4511 // If the overflow check was an add followed by a compare, the insertion point
4512 // may be pointing to the compare. We want to insert the new instructions
4513 // before the add in case there are uses of the add between the add and the
4515 Builder.SetInsertPoint(&OrigI);
4517 if (isNeutralValue(BinaryOp, RHS)) {
4519 Overflow = Builder.getFalse();
4523 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4524 case OverflowResult::MayOverflow:
4526 case OverflowResult::AlwaysOverflowsLow:
4527 case OverflowResult::AlwaysOverflowsHigh:
4528 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4529 Result->takeName(&OrigI);
4530 Overflow = Builder.getTrue();
4532 case OverflowResult::NeverOverflows:
4533 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4534 Result->takeName(&OrigI);
4535 Overflow = Builder.getFalse();
4536 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4538 Inst->setHasNoSignedWrap();
4540 Inst->setHasNoUnsignedWrap();
4545 llvm_unreachable("Unexpected overflow result");
4548 /// Recognize and process idiom involving test for multiplication
4551 /// The caller has matched a pattern of the form:
4552 /// I = cmp u (mul(zext A, zext B), V
4553 /// The function checks if this is a test for overflow and if so replaces
4554 /// multiplication with call to 'mul.with.overflow' intrinsic.
4556 /// \param I Compare instruction.
4557 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4558 /// the compare instruction. Must be of integer type.
4559 /// \param OtherVal The other argument of compare instruction.
4560 /// \returns Instruction which must replace the compare instruction, NULL if no
4561 /// replacement required.
4562 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4563 Value *OtherVal, InstCombiner &IC) {
4564 // Don't bother doing this transformation for pointers, don't do it for
4566 if (!isa<IntegerType>(MulVal->getType()))
4569 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4570 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4571 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4574 assert(MulInstr->getOpcode() == Instruction::Mul);
4576 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4577 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4578 assert(LHS->getOpcode() == Instruction::ZExt);
4579 assert(RHS->getOpcode() == Instruction::ZExt);
4580 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4582 // Calculate type and width of the result produced by mul.with.overflow.
4583 Type *TyA = A->getType(), *TyB = B->getType();
4584 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4585 WidthB = TyB->getPrimitiveSizeInBits();
4588 if (WidthB > WidthA) {
4596 // In order to replace the original mul with a narrower mul.with.overflow,
4597 // all uses must ignore upper bits of the product. The number of used low
4598 // bits must be not greater than the width of mul.with.overflow.
4599 if (MulVal->hasNUsesOrMore(2))
4600 for (User *U : MulVal->users()) {
4603 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4604 // Check if truncation ignores bits above MulWidth.
4605 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4606 if (TruncWidth > MulWidth)
4608 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4609 // Check if AND ignores bits above MulWidth.
4610 if (BO->getOpcode() != Instruction::And)
4612 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4613 const APInt &CVal = CI->getValue();
4614 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4617 // In this case we could have the operand of the binary operation
4618 // being defined in another block, and performing the replacement
4619 // could break the dominance relation.
4623 // Other uses prohibit this transformation.
4628 // Recognize patterns
4629 switch (I.getPredicate()) {
4630 case ICmpInst::ICMP_EQ:
4631 case ICmpInst::ICMP_NE:
4632 // Recognize pattern:
4633 // mulval = mul(zext A, zext B)
4634 // cmp eq/neq mulval, zext trunc mulval
4635 if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4636 if (Zext->hasOneUse()) {
4637 Value *ZextArg = Zext->getOperand(0);
4638 if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4639 if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4643 // Recognize pattern:
4644 // mulval = mul(zext A, zext B)
4645 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4648 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4649 if (ValToMask != MulVal)
4651 const APInt &CVal = CI->getValue() + 1;
4652 if (CVal.isPowerOf2()) {
4653 unsigned MaskWidth = CVal.logBase2();
4654 if (MaskWidth == MulWidth)
4655 break; // Recognized
4660 case ICmpInst::ICMP_UGT:
4661 // Recognize pattern:
4662 // mulval = mul(zext A, zext B)
4663 // cmp ugt mulval, max
4664 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4665 APInt MaxVal = APInt::getMaxValue(MulWidth);
4666 MaxVal = MaxVal.zext(CI->getBitWidth());
4667 if (MaxVal.eq(CI->getValue()))
4668 break; // Recognized
4672 case ICmpInst::ICMP_UGE:
4673 // Recognize pattern:
4674 // mulval = mul(zext A, zext B)
4675 // cmp uge mulval, max+1
4676 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4677 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4678 if (MaxVal.eq(CI->getValue()))
4679 break; // Recognized
4683 case ICmpInst::ICMP_ULE:
4684 // Recognize pattern:
4685 // mulval = mul(zext A, zext B)
4686 // cmp ule mulval, max
4687 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4688 APInt MaxVal = APInt::getMaxValue(MulWidth);
4689 MaxVal = MaxVal.zext(CI->getBitWidth());
4690 if (MaxVal.eq(CI->getValue()))
4691 break; // Recognized
4695 case ICmpInst::ICMP_ULT:
4696 // Recognize pattern:
4697 // mulval = mul(zext A, zext B)
4698 // cmp ule mulval, max + 1
4699 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4700 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4701 if (MaxVal.eq(CI->getValue()))
4702 break; // Recognized
4710 InstCombiner::BuilderTy &Builder = IC.Builder;
4711 Builder.SetInsertPoint(MulInstr);
4713 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4714 Value *MulA = A, *MulB = B;
4715 if (WidthA < MulWidth)
4716 MulA = Builder.CreateZExt(A, MulType);
4717 if (WidthB < MulWidth)
4718 MulB = Builder.CreateZExt(B, MulType);
4719 Function *F = Intrinsic::getDeclaration(
4720 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4721 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4722 IC.Worklist.Add(MulInstr);
4724 // If there are uses of mul result other than the comparison, we know that
4725 // they are truncation or binary AND. Change them to use result of
4726 // mul.with.overflow and adjust properly mask/size.
4727 if (MulVal->hasNUsesOrMore(2)) {
4728 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4729 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4731 if (U == &I || U == OtherVal)
4733 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4734 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4735 IC.replaceInstUsesWith(*TI, Mul);
4737 TI->setOperand(0, Mul);
4738 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4739 assert(BO->getOpcode() == Instruction::And);
4740 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4741 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4742 APInt ShortMask = CI->getValue().trunc(MulWidth);
4743 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4745 cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4746 IC.Worklist.Add(Zext);
4747 IC.replaceInstUsesWith(*BO, Zext);
4749 llvm_unreachable("Unexpected Binary operation");
4751 IC.Worklist.Add(cast<Instruction>(U));
4754 if (isa<Instruction>(OtherVal))
4755 IC.Worklist.Add(cast<Instruction>(OtherVal));
4757 // The original icmp gets replaced with the overflow value, maybe inverted
4758 // depending on predicate.
4759 bool Inverse = false;
4760 switch (I.getPredicate()) {
4761 case ICmpInst::ICMP_NE:
4763 case ICmpInst::ICMP_EQ:
4766 case ICmpInst::ICMP_UGT:
4767 case ICmpInst::ICMP_UGE:
4768 if (I.getOperand(0) == MulVal)
4772 case ICmpInst::ICMP_ULT:
4773 case ICmpInst::ICMP_ULE:
4774 if (I.getOperand(1) == MulVal)
4779 llvm_unreachable("Unexpected predicate");
4782 Value *Res = Builder.CreateExtractValue(Call, 1);
4783 return BinaryOperator::CreateNot(Res);
4786 return ExtractValueInst::Create(Call, 1);
4789 /// When performing a comparison against a constant, it is possible that not all
4790 /// the bits in the LHS are demanded. This helper method computes the mask that
4792 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4794 if (!match(I.getOperand(1), m_APInt(RHS)))
4795 return APInt::getAllOnesValue(BitWidth);
4797 // If this is a normal comparison, it demands all bits. If it is a sign bit
4798 // comparison, it only demands the sign bit.
4800 if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4801 return APInt::getSignMask(BitWidth);
4803 switch (I.getPredicate()) {
4804 // For a UGT comparison, we don't care about any bits that
4805 // correspond to the trailing ones of the comparand. The value of these
4806 // bits doesn't impact the outcome of the comparison, because any value
4807 // greater than the RHS must differ in a bit higher than these due to carry.
4808 case ICmpInst::ICMP_UGT:
4809 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4811 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4812 // Any value less than the RHS must differ in a higher bit because of carries.
4813 case ICmpInst::ICMP_ULT:
4814 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4817 return APInt::getAllOnesValue(BitWidth);
4821 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4822 /// should be swapped.
4823 /// The decision is based on how many times these two operands are reused
4824 /// as subtract operands and their positions in those instructions.
4825 /// The rationale is that several architectures use the same instruction for
4826 /// both subtract and cmp. Thus, it is better if the order of those operands
4828 /// \return true if Op0 and Op1 should be swapped.
4829 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4830 // Filter out pointer values as those cannot appear directly in subtract.
4831 // FIXME: we may want to go through inttoptrs or bitcasts.
4832 if (Op0->getType()->isPointerTy())
4834 // If a subtract already has the same operands as a compare, swapping would be
4835 // bad. If a subtract has the same operands as a compare but in reverse order,
4836 // then swapping is good.
4838 for (const User *U : Op0->users()) {
4839 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4841 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4844 return GoodToSwap > 0;
4847 /// Check that one use is in the same block as the definition and all
4848 /// other uses are in blocks dominated by a given block.
4850 /// \param DI Definition
4852 /// \param DB Block that must dominate all uses of \p DI outside
4853 /// the parent block
4854 /// \return true when \p UI is the only use of \p DI in the parent block
4855 /// and all other uses of \p DI are in blocks dominated by \p DB.
4857 bool InstCombiner::dominatesAllUses(const Instruction *DI,
4858 const Instruction *UI,
4859 const BasicBlock *DB) const {
4860 assert(DI && UI && "Instruction not defined\n");
4861 // Ignore incomplete definitions.
4862 if (!DI->getParent())
4864 // DI and UI must be in the same block.
4865 if (DI->getParent() != UI->getParent())
4867 // Protect from self-referencing blocks.
4868 if (DI->getParent() == DB)
4870 for (const User *U : DI->users()) {
4871 auto *Usr = cast<Instruction>(U);
4872 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4878 /// Return true when the instruction sequence within a block is select-cmp-br.
4879 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4880 const BasicBlock *BB = SI->getParent();
4883 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4884 if (!BI || BI->getNumSuccessors() != 2)
4886 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4887 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4892 /// True when a select result is replaced by one of its operands
4893 /// in select-icmp sequence. This will eventually result in the elimination
4896 /// \param SI Select instruction
4897 /// \param Icmp Compare instruction
4898 /// \param SIOpd Operand that replaces the select
4901 /// - The replacement is global and requires dominator information
4902 /// - The caller is responsible for the actual replacement
4907 /// %4 = select i1 %3, %C* %0, %C* null
4908 /// %5 = icmp eq %C* %4, null
4909 /// br i1 %5, label %9, label %7
4911 /// ; <label>:7 ; preds = %entry
4912 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4915 /// can be transformed to
4917 /// %5 = icmp eq %C* %0, null
4918 /// %6 = select i1 %3, i1 %5, i1 true
4919 /// br i1 %6, label %9, label %7
4921 /// ; <label>:7 ; preds = %entry
4922 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4924 /// Similar when the first operand of the select is a constant or/and
4925 /// the compare is for not equal rather than equal.
4927 /// NOTE: The function is only called when the select and compare constants
4928 /// are equal, the optimization can work only for EQ predicates. This is not a
4929 /// major restriction since a NE compare should be 'normalized' to an equal
4930 /// compare, which usually happens in the combiner and test case
4931 /// select-cmp-br.ll checks for it.
4932 bool InstCombiner::replacedSelectWithOperand(SelectInst *SI,
4933 const ICmpInst *Icmp,
4934 const unsigned SIOpd) {
4935 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4936 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4937 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4938 // The check for the single predecessor is not the best that can be
4939 // done. But it protects efficiently against cases like when SI's
4940 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4941 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4942 // replaced can be reached on either path. So the uniqueness check
4943 // guarantees that the path all uses of SI (outside SI's parent) are on
4944 // is disjoint from all other paths out of SI. But that information
4945 // is more expensive to compute, and the trade-off here is in favor
4946 // of compile-time. It should also be noticed that we check for a single
4947 // predecessor and not only uniqueness. This to handle the situation when
4948 // Succ and Succ1 points to the same basic block.
4949 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4951 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4958 /// Try to fold the comparison based on range information we can get by checking
4959 /// whether bits are known to be zero or one in the inputs.
4960 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4961 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4962 Type *Ty = Op0->getType();
4963 ICmpInst::Predicate Pred = I.getPredicate();
4965 // Get scalar or pointer size.
4966 unsigned BitWidth = Ty->isIntOrIntVectorTy()
4967 ? Ty->getScalarSizeInBits()
4968 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
4973 KnownBits Op0Known(BitWidth);
4974 KnownBits Op1Known(BitWidth);
4976 if (SimplifyDemandedBits(&I, 0,
4977 getDemandedBitsLHSMask(I, BitWidth),
4981 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4985 // Given the known and unknown bits, compute a range that the LHS could be
4986 // in. Compute the Min, Max and RHS values based on the known bits. For the
4987 // EQ and NE we use unsigned values.
4988 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4989 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4991 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4992 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4994 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4995 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4998 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4999 // out that the LHS or RHS is a constant. Constant fold this now, so that
5000 // code below can assume that Min != Max.
5001 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
5002 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
5003 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
5004 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
5006 // Based on the range information we know about the LHS, see if we can
5007 // simplify this comparison. For example, (x&4) < 8 is always true.
5010 llvm_unreachable("Unknown icmp opcode!");
5011 case ICmpInst::ICMP_EQ:
5012 case ICmpInst::ICMP_NE: {
5013 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
5014 return Pred == CmpInst::ICMP_EQ
5015 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
5016 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5019 // If all bits are known zero except for one, then we know at most one bit
5020 // is set. If the comparison is against zero, then this is a check to see if
5021 // *that* bit is set.
5022 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
5023 if (Op1Known.isZero()) {
5024 // If the LHS is an AND with the same constant, look through it.
5025 Value *LHS = nullptr;
5027 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
5028 *LHSC != Op0KnownZeroInverted)
5032 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
5033 APInt ValToCheck = Op0KnownZeroInverted;
5034 Type *XTy = X->getType();
5035 if (ValToCheck.isPowerOf2()) {
5036 // ((1 << X) & 8) == 0 -> X != 3
5037 // ((1 << X) & 8) != 0 -> X == 3
5038 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5039 auto NewPred = ICmpInst::getInversePredicate(Pred);
5040 return new ICmpInst(NewPred, X, CmpC);
5041 } else if ((++ValToCheck).isPowerOf2()) {
5042 // ((1 << X) & 7) == 0 -> X >= 3
5043 // ((1 << X) & 7) != 0 -> X < 3
5044 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5046 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
5047 return new ICmpInst(NewPred, X, CmpC);
5051 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5053 if (Op0KnownZeroInverted.isOneValue() &&
5054 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5055 // ((8 >>u X) & 1) == 0 -> X != 3
5056 // ((8 >>u X) & 1) != 0 -> X == 3
5057 unsigned CmpVal = CI->countTrailingZeros();
5058 auto NewPred = ICmpInst::getInversePredicate(Pred);
5059 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5064 case ICmpInst::ICMP_ULT: {
5065 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5066 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5067 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5068 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5069 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5070 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5073 if (match(Op1, m_APInt(CmpC))) {
5074 // A <u C -> A == C-1 if min(A)+1 == C
5075 if (*CmpC == Op0Min + 1)
5076 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5077 ConstantInt::get(Op1->getType(), *CmpC - 1));
5078 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5079 // exceeds the log2 of C.
5080 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5081 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5082 Constant::getNullValue(Op1->getType()));
5086 case ICmpInst::ICMP_UGT: {
5087 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5088 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5089 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5090 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5091 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5092 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5095 if (match(Op1, m_APInt(CmpC))) {
5096 // A >u C -> A == C+1 if max(a)-1 == C
5097 if (*CmpC == Op0Max - 1)
5098 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5099 ConstantInt::get(Op1->getType(), *CmpC + 1));
5100 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5101 // exceeds the log2 of C.
5102 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5103 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5104 Constant::getNullValue(Op1->getType()));
5108 case ICmpInst::ICMP_SLT: {
5109 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5110 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5111 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5112 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5113 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5114 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5116 if (match(Op1, m_APInt(CmpC))) {
5117 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5118 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5119 ConstantInt::get(Op1->getType(), *CmpC - 1));
5123 case ICmpInst::ICMP_SGT: {
5124 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5125 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5126 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5127 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5128 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5129 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5131 if (match(Op1, m_APInt(CmpC))) {
5132 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5133 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5134 ConstantInt::get(Op1->getType(), *CmpC + 1));
5138 case ICmpInst::ICMP_SGE:
5139 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
5140 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5141 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5142 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5143 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5144 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5145 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5147 case ICmpInst::ICMP_SLE:
5148 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
5149 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5150 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5151 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5152 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5153 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5154 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5156 case ICmpInst::ICMP_UGE:
5157 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
5158 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5159 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5160 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5161 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5162 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5163 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5165 case ICmpInst::ICMP_ULE:
5166 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
5167 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5168 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5169 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5170 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5171 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5172 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5176 // Turn a signed comparison into an unsigned one if both operands are known to
5177 // have the same sign.
5179 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5180 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5181 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5186 llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
5187 llvm::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5189 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&
5190 "Only for relational integer predicates.");
5192 Type *Type = C->getType();
5193 bool IsSigned = ICmpInst::isSigned(Pred);
5195 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5196 bool WillIncrement =
5197 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5199 // Check if the constant operand can be safely incremented/decremented
5200 // without overflowing/underflowing.
5201 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5202 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5205 Constant *SafeReplacementConstant = nullptr;
5206 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5207 // Bail out if the constant can't be safely incremented/decremented.
5208 if (!ConstantIsOk(CI))
5210 } else if (Type->isVectorTy()) {
5211 unsigned NumElts = Type->getVectorNumElements();
5212 for (unsigned i = 0; i != NumElts; ++i) {
5213 Constant *Elt = C->getAggregateElement(i);
5217 if (isa<UndefValue>(Elt))
5220 // Bail out if we can't determine if this constant is min/max or if we
5221 // know that this constant is min/max.
5222 auto *CI = dyn_cast<ConstantInt>(Elt);
5223 if (!CI || !ConstantIsOk(CI))
5226 if (!SafeReplacementConstant)
5227 SafeReplacementConstant = CI;
5234 // It may not be safe to change a compare predicate in the presence of
5235 // undefined elements, so replace those elements with the first safe constant
5237 if (C->containsUndefElement()) {
5238 assert(SafeReplacementConstant && "Replacement constant not set");
5239 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5242 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5244 // Increment or decrement the constant.
5245 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5246 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5248 return std::make_pair(NewPred, NewC);
5251 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
5252 /// it into the appropriate icmp lt or icmp gt instruction. This transform
5253 /// allows them to be folded in visitICmpInst.
5254 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5255 ICmpInst::Predicate Pred = I.getPredicate();
5256 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5257 isCanonicalPredicate(Pred))
5260 Value *Op0 = I.getOperand(0);
5261 Value *Op1 = I.getOperand(1);
5262 auto *Op1C = dyn_cast<Constant>(Op1);
5266 auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5267 if (!FlippedStrictness)
5270 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5273 /// Integer compare with boolean values can always be turned into bitwise ops.
5274 static Instruction *canonicalizeICmpBool(ICmpInst &I,
5275 InstCombiner::BuilderTy &Builder) {
5276 Value *A = I.getOperand(0), *B = I.getOperand(1);
5277 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
5279 // A boolean compared to true/false can be simplified to Op0/true/false in
5280 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5281 // Cases not handled by InstSimplify are always 'not' of Op0.
5282 if (match(B, m_Zero())) {
5283 switch (I.getPredicate()) {
5284 case CmpInst::ICMP_EQ: // A == 0 -> !A
5285 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5286 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5287 return BinaryOperator::CreateNot(A);
5289 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
5291 } else if (match(B, m_One())) {
5292 switch (I.getPredicate()) {
5293 case CmpInst::ICMP_NE: // A != 1 -> !A
5294 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5295 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5296 return BinaryOperator::CreateNot(A);
5298 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
5302 switch (I.getPredicate()) {
5304 llvm_unreachable("Invalid icmp instruction!");
5305 case ICmpInst::ICMP_EQ:
5306 // icmp eq i1 A, B -> ~(A ^ B)
5307 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5309 case ICmpInst::ICMP_NE:
5310 // icmp ne i1 A, B -> A ^ B
5311 return BinaryOperator::CreateXor(A, B);
5313 case ICmpInst::ICMP_UGT:
5314 // icmp ugt -> icmp ult
5317 case ICmpInst::ICMP_ULT:
5318 // icmp ult i1 A, B -> ~A & B
5319 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5321 case ICmpInst::ICMP_SGT:
5322 // icmp sgt -> icmp slt
5325 case ICmpInst::ICMP_SLT:
5326 // icmp slt i1 A, B -> A & ~B
5327 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5329 case ICmpInst::ICMP_UGE:
5330 // icmp uge -> icmp ule
5333 case ICmpInst::ICMP_ULE:
5334 // icmp ule i1 A, B -> ~A | B
5335 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5337 case ICmpInst::ICMP_SGE:
5338 // icmp sge -> icmp sle
5341 case ICmpInst::ICMP_SLE:
5342 // icmp sle i1 A, B -> A | ~B
5343 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5347 // Transform pattern like:
5348 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5349 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5353 static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5354 InstCombiner::BuilderTy &Builder) {
5355 ICmpInst::Predicate Pred, NewPred;
5358 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5359 // We want X to be the icmp's second operand, so swap predicate if it isn't.
5360 if (Cmp.getOperand(0) == X)
5361 Pred = Cmp.getSwappedPredicate();
5364 case ICmpInst::ICMP_ULE:
5365 NewPred = ICmpInst::ICMP_NE;
5367 case ICmpInst::ICMP_UGT:
5368 NewPred = ICmpInst::ICMP_EQ;
5373 } else if (match(&Cmp, m_c_ICmp(Pred,
5374 m_OneUse(m_CombineOr(
5375 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5376 m_Add(m_Shl(m_One(), m_Value(Y)),
5379 // The variant with 'add' is not canonical, (the variant with 'not' is)
5380 // we only get it because it has extra uses, and can't be canonicalized,
5382 // We want X to be the icmp's second operand, so swap predicate if it isn't.
5383 if (Cmp.getOperand(0) == X)
5384 Pred = Cmp.getSwappedPredicate();
5387 case ICmpInst::ICMP_ULT:
5388 NewPred = ICmpInst::ICMP_NE;
5390 case ICmpInst::ICMP_UGE:
5391 NewPred = ICmpInst::ICMP_EQ;
5399 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5400 Constant *Zero = Constant::getNullValue(NewX->getType());
5401 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5404 static Instruction *foldVectorCmp(CmpInst &Cmp,
5405 InstCombiner::BuilderTy &Builder) {
5406 // If both arguments of the cmp are shuffles that use the same mask and
5407 // shuffle within a single vector, move the shuffle after the cmp.
5408 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5411 if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))) &&
5412 match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
5413 V1->getType() == V2->getType() &&
5414 (LHS->hasOneUse() || RHS->hasOneUse())) {
5415 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5416 CmpInst::Predicate P = Cmp.getPredicate();
5417 Value *NewCmp = isa<ICmpInst>(Cmp) ? Builder.CreateICmp(P, V1, V2)
5418 : Builder.CreateFCmp(P, V1, V2);
5419 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5424 // extract(uadd.with.overflow(A, B), 0) ult A
5425 // -> extract(uadd.with.overflow(A, B), 1)
5426 static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5427 CmpInst::Predicate Pred = I.getPredicate();
5428 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5432 auto UAddOvResultPat = m_ExtractValue<0>(
5433 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5434 if (match(Op0, UAddOvResultPat) &&
5435 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5436 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5437 (match(A, m_One()) || match(B, m_One()))) ||
5438 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5439 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5440 // extract(uadd.with.overflow(A, B), 0) < A
5441 // extract(uadd.with.overflow(A, 1), 0) == 0
5442 // extract(uadd.with.overflow(A, -1), 0) != -1
5443 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5444 else if (match(Op1, UAddOvResultPat) &&
5445 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5446 // A > extract(uadd.with.overflow(A, B), 0)
5447 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5451 return ExtractValueInst::Create(UAddOv, 1);
5454 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5455 bool Changed = false;
5456 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5457 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5458 unsigned Op0Cplxity = getComplexity(Op0);
5459 unsigned Op1Cplxity = getComplexity(Op1);
5461 /// Orders the operands of the compare so that they are listed from most
5462 /// complex to least complex. This puts constants before unary operators,
5463 /// before binary operators.
5464 if (Op0Cplxity < Op1Cplxity ||
5465 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5467 std::swap(Op0, Op1);
5471 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, Q))
5472 return replaceInstUsesWith(I, V);
5474 // Comparing -val or val with non-zero is the same as just comparing val
5475 // ie, abs(val) != 0 -> val != 0
5476 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5477 Value *Cond, *SelectTrue, *SelectFalse;
5478 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5479 m_Value(SelectFalse)))) {
5480 if (Value *V = dyn_castNegVal(SelectTrue)) {
5481 if (V == SelectFalse)
5482 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5484 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5485 if (V == SelectTrue)
5486 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5491 if (Op0->getType()->isIntOrIntVectorTy(1))
5492 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
5495 if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
5498 if (Instruction *Res = foldICmpWithConstant(I))
5501 if (Instruction *Res = foldICmpWithDominatingICmp(I))
5504 if (Instruction *Res = foldICmpBinOp(I, Q))
5507 if (Instruction *Res = foldICmpUsingKnownBits(I))
5510 // Test if the ICmpInst instruction is used exclusively by a select as
5511 // part of a minimum or maximum operation. If so, refrain from doing
5512 // any other folding. This helps out other analyses which understand
5513 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5514 // and CodeGen. And in this case, at least one of the comparison
5515 // operands has at least one user besides the compare (the select),
5516 // which would often largely negate the benefit of folding anyway.
5518 // Do the same for the other patterns recognized by matchSelectPattern.
5520 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5522 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5523 if (SPR.Flavor != SPF_UNKNOWN)
5527 // Do this after checking for min/max to prevent infinite looping.
5528 if (Instruction *Res = foldICmpWithZero(I))
5531 // FIXME: We only do this after checking for min/max to prevent infinite
5532 // looping caused by a reverse canonicalization of these patterns for min/max.
5533 // FIXME: The organization of folds is a mess. These would naturally go into
5534 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
5535 // down here after the min/max restriction.
5536 ICmpInst::Predicate Pred = I.getPredicate();
5538 if (match(Op1, m_APInt(C))) {
5539 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
5540 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
5541 Constant *Zero = Constant::getNullValue(Op0->getType());
5542 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
5545 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
5546 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
5547 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
5548 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
5552 if (Instruction *Res = foldICmpInstWithConstant(I))
5555 // Try to match comparison as a sign bit test. Intentionally do this after
5556 // foldICmpInstWithConstant() to potentially let other folds to happen first.
5557 if (Instruction *New = foldSignBitTest(I))
5560 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
5563 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5564 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
5565 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
5567 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
5568 if (Instruction *NI = foldGEPICmp(GEP, Op0,
5569 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5572 // Try to optimize equality comparisons against alloca-based pointers.
5573 if (Op0->getType()->isPointerTy() && I.isEquality()) {
5574 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
5575 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
5576 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
5578 if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
5579 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
5583 if (Instruction *Res = foldICmpBitCast(I, Builder))
5586 if (Instruction *R = foldICmpWithCastOp(I))
5589 if (Instruction *Res = foldICmpWithMinMax(I))
5594 // Transform (A & ~B) == 0 --> (A & B) != 0
5595 // and (A & ~B) != 0 --> (A & B) == 0
5596 // if A is a power of 2.
5597 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
5598 match(Op1, m_Zero()) &&
5599 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
5600 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
5603 // ~X < ~Y --> Y < X
5604 // ~X < C --> X > ~C
5605 if (match(Op0, m_Not(m_Value(A)))) {
5606 if (match(Op1, m_Not(m_Value(B))))
5607 return new ICmpInst(I.getPredicate(), B, A);
5610 if (match(Op1, m_APInt(C)))
5611 return new ICmpInst(I.getSwappedPredicate(), A,
5612 ConstantInt::get(Op1->getType(), ~(*C)));
5615 Instruction *AddI = nullptr;
5616 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
5617 m_Instruction(AddI))) &&
5618 isa<IntegerType>(A->getType())) {
5621 if (OptimizeOverflowCheck(Instruction::Add, /*Signed*/false, A, B,
5622 *AddI, Result, Overflow)) {
5623 replaceInstUsesWith(*AddI, Result);
5624 return replaceInstUsesWith(I, Overflow);
5628 // (zext a) * (zext b) --> llvm.umul.with.overflow.
5629 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5630 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
5633 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5634 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
5639 if (Instruction *Res = foldICmpEquality(I))
5642 if (Instruction *Res = foldICmpOfUAddOv(I))
5645 // The 'cmpxchg' instruction returns an aggregate containing the old value and
5646 // an i1 which indicates whether or not we successfully did the swap.
5648 // Replace comparisons between the old value and the expected value with the
5649 // indicator that 'cmpxchg' returns.
5651 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
5652 // spuriously fail. In those cases, the old value may equal the expected
5653 // value but it is possible for the swap to not occur.
5654 if (I.getPredicate() == ICmpInst::ICMP_EQ)
5655 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
5656 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
5657 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
5659 return ExtractValueInst::Create(ACXI, 1);
5665 if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
5666 return foldICmpAddOpConst(X, *C, I.getPredicate());
5669 if (match(Op1, m_Add(m_Value(X), m_APInt(C))) && Op0 == X)
5670 return foldICmpAddOpConst(X, *C, I.getSwappedPredicate());
5673 if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
5676 if (I.getType()->isVectorTy())
5677 if (Instruction *Res = foldVectorCmp(I, Builder))
5680 return Changed ? &I : nullptr;
5683 /// Fold fcmp ([us]itofp x, cst) if possible.
5684 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
5686 if (!isa<ConstantFP>(RHSC)) return nullptr;
5687 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5689 // Get the width of the mantissa. We don't want to hack on conversions that
5690 // might lose information from the integer, e.g. "i64 -> float"
5691 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5692 if (MantissaWidth == -1) return nullptr; // Unknown.
5694 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5696 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5698 if (I.isEquality()) {
5699 FCmpInst::Predicate P = I.getPredicate();
5700 bool IsExact = false;
5701 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
5702 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
5704 // If the floating point constant isn't an integer value, we know if we will
5705 // ever compare equal / not equal to it.
5707 // TODO: Can never be -0.0 and other non-representable values
5708 APFloat RHSRoundInt(RHS);
5709 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
5710 if (RHS.compare(RHSRoundInt) != APFloat::cmpEqual) {
5711 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
5712 return replaceInstUsesWith(I, Builder.getFalse());
5714 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
5715 return replaceInstUsesWith(I, Builder.getTrue());
5719 // TODO: If the constant is exactly representable, is it always OK to do
5720 // equality compares as integer?
5723 // Check to see that the input is converted from an integer type that is small
5724 // enough that preserves all bits. TODO: check here for "known" sign bits.
5725 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5726 unsigned InputSize = IntTy->getScalarSizeInBits();
5728 // Following test does NOT adjust InputSize downwards for signed inputs,
5729 // because the most negative value still requires all the mantissa bits
5730 // to distinguish it from one less than that value.
5731 if ((int)InputSize > MantissaWidth) {
5732 // Conversion would lose accuracy. Check if loss can impact comparison.
5733 int Exp = ilogb(RHS);
5734 if (Exp == APFloat::IEK_Inf) {
5735 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
5736 if (MaxExponent < (int)InputSize - !LHSUnsigned)
5737 // Conversion could create infinity.
5740 // Note that if RHS is zero or NaN, then Exp is negative
5741 // and first condition is trivially false.
5742 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
5743 // Conversion could affect comparison.
5748 // Otherwise, we can potentially simplify the comparison. We know that it
5749 // will always come through as an integer value and we know the constant is
5750 // not a NAN (it would have been previously simplified).
5751 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5753 ICmpInst::Predicate Pred;
5754 switch (I.getPredicate()) {
5755 default: llvm_unreachable("Unexpected predicate!");
5756 case FCmpInst::FCMP_UEQ:
5757 case FCmpInst::FCMP_OEQ:
5758 Pred = ICmpInst::ICMP_EQ;
5760 case FCmpInst::FCMP_UGT:
5761 case FCmpInst::FCMP_OGT:
5762 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5764 case FCmpInst::FCMP_UGE:
5765 case FCmpInst::FCMP_OGE:
5766 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5768 case FCmpInst::FCMP_ULT:
5769 case FCmpInst::FCMP_OLT:
5770 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5772 case FCmpInst::FCMP_ULE:
5773 case FCmpInst::FCMP_OLE:
5774 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5776 case FCmpInst::FCMP_UNE:
5777 case FCmpInst::FCMP_ONE:
5778 Pred = ICmpInst::ICMP_NE;
5780 case FCmpInst::FCMP_ORD:
5781 return replaceInstUsesWith(I, Builder.getTrue());
5782 case FCmpInst::FCMP_UNO:
5783 return replaceInstUsesWith(I, Builder.getFalse());
5786 // Now we know that the APFloat is a normal number, zero or inf.
5788 // See if the FP constant is too large for the integer. For example,
5789 // comparing an i8 to 300.0.
5790 unsigned IntWidth = IntTy->getScalarSizeInBits();
5793 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5794 // and large values.
5795 APFloat SMax(RHS.getSemantics());
5796 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5797 APFloat::rmNearestTiesToEven);
5798 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5799 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5800 Pred == ICmpInst::ICMP_SLE)
5801 return replaceInstUsesWith(I, Builder.getTrue());
5802 return replaceInstUsesWith(I, Builder.getFalse());
5805 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5806 // +INF and large values.
5807 APFloat UMax(RHS.getSemantics());
5808 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5809 APFloat::rmNearestTiesToEven);
5810 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5811 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5812 Pred == ICmpInst::ICMP_ULE)
5813 return replaceInstUsesWith(I, Builder.getTrue());
5814 return replaceInstUsesWith(I, Builder.getFalse());
5819 // See if the RHS value is < SignedMin.
5820 APFloat SMin(RHS.getSemantics());
5821 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5822 APFloat::rmNearestTiesToEven);
5823 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5824 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5825 Pred == ICmpInst::ICMP_SGE)
5826 return replaceInstUsesWith(I, Builder.getTrue());
5827 return replaceInstUsesWith(I, Builder.getFalse());
5830 // See if the RHS value is < UnsignedMin.
5831 APFloat SMin(RHS.getSemantics());
5832 SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
5833 APFloat::rmNearestTiesToEven);
5834 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
5835 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
5836 Pred == ICmpInst::ICMP_UGE)
5837 return replaceInstUsesWith(I, Builder.getTrue());
5838 return replaceInstUsesWith(I, Builder.getFalse());
5842 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5843 // [0, UMAX], but it may still be fractional. See if it is fractional by
5844 // casting the FP value to the integer value and back, checking for equality.
5845 // Don't do this for zero, because -0.0 is not fractional.
5846 Constant *RHSInt = LHSUnsigned
5847 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5848 : ConstantExpr::getFPToSI(RHSC, IntTy);
5849 if (!RHS.isZero()) {
5850 bool Equal = LHSUnsigned
5851 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5852 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5854 // If we had a comparison against a fractional value, we have to adjust
5855 // the compare predicate and sometimes the value. RHSC is rounded towards
5856 // zero at this point.
5858 default: llvm_unreachable("Unexpected integer comparison!");
5859 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5860 return replaceInstUsesWith(I, Builder.getTrue());
5861 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5862 return replaceInstUsesWith(I, Builder.getFalse());
5863 case ICmpInst::ICMP_ULE:
5864 // (float)int <= 4.4 --> int <= 4
5865 // (float)int <= -4.4 --> false
5866 if (RHS.isNegative())
5867 return replaceInstUsesWith(I, Builder.getFalse());
5869 case ICmpInst::ICMP_SLE:
5870 // (float)int <= 4.4 --> int <= 4
5871 // (float)int <= -4.4 --> int < -4
5872 if (RHS.isNegative())
5873 Pred = ICmpInst::ICMP_SLT;
5875 case ICmpInst::ICMP_ULT:
5876 // (float)int < -4.4 --> false
5877 // (float)int < 4.4 --> int <= 4
5878 if (RHS.isNegative())
5879 return replaceInstUsesWith(I, Builder.getFalse());
5880 Pred = ICmpInst::ICMP_ULE;
5882 case ICmpInst::ICMP_SLT:
5883 // (float)int < -4.4 --> int < -4
5884 // (float)int < 4.4 --> int <= 4
5885 if (!RHS.isNegative())
5886 Pred = ICmpInst::ICMP_SLE;
5888 case ICmpInst::ICMP_UGT:
5889 // (float)int > 4.4 --> int > 4
5890 // (float)int > -4.4 --> true
5891 if (RHS.isNegative())
5892 return replaceInstUsesWith(I, Builder.getTrue());
5894 case ICmpInst::ICMP_SGT:
5895 // (float)int > 4.4 --> int > 4
5896 // (float)int > -4.4 --> int >= -4
5897 if (RHS.isNegative())
5898 Pred = ICmpInst::ICMP_SGE;
5900 case ICmpInst::ICMP_UGE:
5901 // (float)int >= -4.4 --> true
5902 // (float)int >= 4.4 --> int > 4
5903 if (RHS.isNegative())
5904 return replaceInstUsesWith(I, Builder.getTrue());
5905 Pred = ICmpInst::ICMP_UGT;
5907 case ICmpInst::ICMP_SGE:
5908 // (float)int >= -4.4 --> int >= -4
5909 // (float)int >= 4.4 --> int > 4
5910 if (!RHS.isNegative())
5911 Pred = ICmpInst::ICMP_SGT;
5917 // Lower this FP comparison into an appropriate integer version of the
5919 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5922 /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
5923 static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
5925 // When C is not 0.0 and infinities are not allowed:
5926 // (C / X) < 0.0 is a sign-bit test of X
5927 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
5928 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
5931 // Multiply (C / X) < 0.0 by X * X / C.
5932 // - X is non zero, if it is the flag 'ninf' is violated.
5933 // - C defines the sign of X * X * C. Thus it also defines whether to swap
5934 // the predicate. C is also non zero by definition.
5936 // Thus X * X / C is non zero and the transformation is valid. [qed]
5938 FCmpInst::Predicate Pred = I.getPredicate();
5940 // Check that predicates are valid.
5941 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
5942 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
5945 // Check that RHS operand is zero.
5946 if (!match(RHSC, m_AnyZeroFP()))
5949 // Check fastmath flags ('ninf').
5950 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
5953 // Check the properties of the dividend. It must not be zero to avoid a
5954 // division by zero (see Proof).
5956 if (!match(LHSI->getOperand(0), m_APFloat(C)))
5962 // Get swapped predicate if necessary.
5963 if (C->isNegative())
5964 Pred = I.getSwappedPredicate();
5966 return new FCmpInst(Pred, LHSI->getOperand(1), RHSC, "", &I);
5969 /// Optimize fabs(X) compared with zero.
5970 static Instruction *foldFabsWithFcmpZero(FCmpInst &I) {
5972 if (!match(I.getOperand(0), m_Intrinsic<Intrinsic::fabs>(m_Value(X))) ||
5973 !match(I.getOperand(1), m_PosZeroFP()))
5976 auto replacePredAndOp0 = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
5978 I->setOperand(0, X);
5982 switch (I.getPredicate()) {
5983 case FCmpInst::FCMP_UGE:
5984 case FCmpInst::FCMP_OLT:
5985 // fabs(X) >= 0.0 --> true
5986 // fabs(X) < 0.0 --> false
5987 llvm_unreachable("fcmp should have simplified");
5989 case FCmpInst::FCMP_OGT:
5990 // fabs(X) > 0.0 --> X != 0.0
5991 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
5993 case FCmpInst::FCMP_UGT:
5994 // fabs(X) u> 0.0 --> X u!= 0.0
5995 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
5997 case FCmpInst::FCMP_OLE:
5998 // fabs(X) <= 0.0 --> X == 0.0
5999 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
6001 case FCmpInst::FCMP_ULE:
6002 // fabs(X) u<= 0.0 --> X u== 0.0
6003 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
6005 case FCmpInst::FCMP_OGE:
6006 // fabs(X) >= 0.0 --> !isnan(X)
6007 assert(!I.hasNoNaNs() && "fcmp should have simplified");
6008 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
6010 case FCmpInst::FCMP_ULT:
6011 // fabs(X) u< 0.0 --> isnan(X)
6012 assert(!I.hasNoNaNs() && "fcmp should have simplified");
6013 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
6015 case FCmpInst::FCMP_OEQ:
6016 case FCmpInst::FCMP_UEQ:
6017 case FCmpInst::FCMP_ONE:
6018 case FCmpInst::FCMP_UNE:
6019 case FCmpInst::FCMP_ORD:
6020 case FCmpInst::FCMP_UNO:
6021 // Look through the fabs() because it doesn't change anything but the sign.
6022 // fabs(X) == 0.0 --> X == 0.0,
6023 // fabs(X) != 0.0 --> X != 0.0
6024 // isnan(fabs(X)) --> isnan(X)
6025 // !isnan(fabs(X) --> !isnan(X)
6026 return replacePredAndOp0(&I, I.getPredicate(), X);
6033 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
6034 bool Changed = false;
6036 /// Orders the operands of the compare so that they are listed from most
6037 /// complex to least complex. This puts constants before unary operators,
6038 /// before binary operators.
6039 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6044 const CmpInst::Predicate Pred = I.getPredicate();
6045 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6046 if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
6047 SQ.getWithInstruction(&I)))
6048 return replaceInstUsesWith(I, V);
6050 // Simplify 'fcmp pred X, X'
6051 Type *OpType = Op0->getType();
6052 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
6056 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
6057 case FCmpInst::FCMP_ULT: // True if unordered or less than
6058 case FCmpInst::FCMP_UGT: // True if unordered or greater than
6059 case FCmpInst::FCMP_UNE: // True if unordered or not equal
6060 // Canonicalize these to be 'fcmp uno %X, 0.0'.
6061 I.setPredicate(FCmpInst::FCMP_UNO);
6062 I.setOperand(1, Constant::getNullValue(OpType));
6065 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
6066 case FCmpInst::FCMP_OEQ: // True if ordered and equal
6067 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
6068 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
6069 // Canonicalize these to be 'fcmp ord %X, 0.0'.
6070 I.setPredicate(FCmpInst::FCMP_ORD);
6071 I.setOperand(1, Constant::getNullValue(OpType));
6076 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
6077 // then canonicalize the operand to 0.0.
6078 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
6079 if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0, &TLI)) {
6080 I.setOperand(0, ConstantFP::getNullValue(OpType));
6083 if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1, &TLI)) {
6084 I.setOperand(1, ConstantFP::getNullValue(OpType));
6089 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
6091 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
6092 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
6094 // Test if the FCmpInst instruction is used exclusively by a select as
6095 // part of a minimum or maximum operation. If so, refrain from doing
6096 // any other folding. This helps out other analyses which understand
6097 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6098 // and CodeGen. And in this case, at least one of the comparison
6099 // operands has at least one user besides the compare (the select),
6100 // which would often largely negate the benefit of folding anyway.
6102 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
6104 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
6105 if (SPR.Flavor != SPF_UNKNOWN)
6109 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
6110 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
6111 if (match(Op1, m_AnyZeroFP()) && !match(Op1, m_PosZeroFP())) {
6112 I.setOperand(1, ConstantFP::getNullValue(OpType));
6116 // Handle fcmp with instruction LHS and constant RHS.
6119 if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
6120 switch (LHSI->getOpcode()) {
6121 case Instruction::PHI:
6122 // Only fold fcmp into the PHI if the phi and fcmp are in the same
6123 // block. If in the same block, we're encouraging jump threading. If
6124 // not, we are just pessimizing the code by making an i1 phi.
6125 if (LHSI->getParent() == I.getParent())
6126 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
6129 case Instruction::SIToFP:
6130 case Instruction::UIToFP:
6131 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
6134 case Instruction::FDiv:
6135 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
6138 case Instruction::Load:
6139 if (auto *GEP = dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
6140 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
6141 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
6142 !cast<LoadInst>(LHSI)->isVolatile())
6143 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
6149 if (Instruction *R = foldFabsWithFcmpZero(I))
6152 if (match(Op0, m_FNeg(m_Value(X)))) {
6153 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
6155 if (match(Op1, m_Constant(C))) {
6156 Constant *NegC = ConstantExpr::getFNeg(C);
6157 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
6161 if (match(Op0, m_FPExt(m_Value(X)))) {
6162 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
6163 if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
6164 return new FCmpInst(Pred, X, Y, "", &I);
6166 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
6168 if (match(Op1, m_APFloat(C))) {
6169 const fltSemantics &FPSem =
6170 X->getType()->getScalarType()->getFltSemantics();
6172 APFloat TruncC = *C;
6173 TruncC.convert(FPSem, APFloat::rmNearestTiesToEven, &Lossy);
6175 // Avoid lossy conversions and denormals.
6176 // Zero is a special case that's OK to convert.
6177 APFloat Fabs = TruncC;
6180 ((Fabs.compare(APFloat::getSmallestNormalized(FPSem)) !=
6181 APFloat::cmpLessThan) || Fabs.isZero())) {
6182 Constant *NewC = ConstantFP::get(X->getType(), TruncC);
6183 return new FCmpInst(Pred, X, NewC, "", &I);
6188 if (I.getType()->isVectorTy())
6189 if (Instruction *Res = foldVectorCmp(I, Builder))
6192 return Changed ? &I : nullptr;