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/CmpInstAnalysis.h"
18 #include "llvm/Analysis/ConstantFolding.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/GetElementPtrTypeIterator.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/KnownBits.h"
28 #include "llvm/Transforms/InstCombine/InstCombiner.h"
31 using namespace PatternMatch;
33 #define DEBUG_TYPE "instcombine"
35 // How many times is a select replaced by one of its operands?
36 STATISTIC(NumSel, "Number of select opts");
39 /// Compute Result = In1+In2, returning true if the result overflowed for this
41 static bool addWithOverflow(APInt &Result, const APInt &In1,
42 const APInt &In2, bool IsSigned = false) {
45 Result = In1.sadd_ov(In2, Overflow);
47 Result = In1.uadd_ov(In2, Overflow);
52 /// Compute Result = In1-In2, returning true if the result overflowed for this
54 static bool subWithOverflow(APInt &Result, const APInt &In1,
55 const APInt &In2, bool IsSigned = false) {
58 Result = In1.ssub_ov(In2, Overflow);
60 Result = In1.usub_ov(In2, Overflow);
65 /// Given an icmp instruction, return true if any use of this comparison is a
66 /// branch on sign bit comparison.
67 static bool hasBranchUse(ICmpInst &I) {
68 for (auto *U : I.users())
69 if (isa<BranchInst>(U))
74 /// Returns true if the exploded icmp can be expressed as a signed comparison
75 /// to zero and updates the predicate accordingly.
76 /// The signedness of the comparison is preserved.
77 /// TODO: Refactor with decomposeBitTestICmp()?
78 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
79 if (!ICmpInst::isSigned(Pred))
83 return ICmpInst::isRelational(Pred);
86 if (Pred == ICmpInst::ICMP_SLT) {
87 Pred = ICmpInst::ICMP_SLE;
90 } else if (C.isAllOnes()) {
91 if (Pred == ICmpInst::ICMP_SGT) {
92 Pred = ICmpInst::ICMP_SGE;
100 /// This is called when we see this pattern:
101 /// cmp pred (load (gep GV, ...)), cmpcst
102 /// where GV is a global variable with a constant initializer. Try to simplify
103 /// this into some simple computation that does not need the load. For example
104 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
106 /// If AndCst is non-null, then the loaded value is masked with that constant
107 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
109 InstCombinerImpl::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
110 GlobalVariable *GV, CmpInst &ICI,
111 ConstantInt *AndCst) {
112 Constant *Init = GV->getInitializer();
113 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
116 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
117 // Don't blow up on huge arrays.
118 if (ArrayElementCount > MaxArraySizeForCombine)
121 // There are many forms of this optimization we can handle, for now, just do
122 // the simple index into a single-dimensional array.
124 // Require: GEP GV, 0, i {{, constant indices}}
125 if (GEP->getNumOperands() < 3 ||
126 !isa<ConstantInt>(GEP->getOperand(1)) ||
127 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
128 isa<Constant>(GEP->getOperand(2)))
131 // Check that indices after the variable are constants and in-range for the
132 // type they index. Collect the indices. This is typically for arrays of
134 SmallVector<unsigned, 4> LaterIndices;
136 Type *EltTy = Init->getType()->getArrayElementType();
137 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
138 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
139 if (!Idx) return nullptr; // Variable index.
141 uint64_t IdxVal = Idx->getZExtValue();
142 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
144 if (StructType *STy = dyn_cast<StructType>(EltTy))
145 EltTy = STy->getElementType(IdxVal);
146 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
147 if (IdxVal >= ATy->getNumElements()) return nullptr;
148 EltTy = ATy->getElementType();
150 return nullptr; // Unknown type.
153 LaterIndices.push_back(IdxVal);
156 enum { Overdefined = -3, Undefined = -2 };
158 // Variables for our state machines.
160 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
161 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
162 // and 87 is the second (and last) index. FirstTrueElement is -2 when
163 // undefined, otherwise set to the first true element. SecondTrueElement is
164 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
165 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
167 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
168 // form "i != 47 & i != 87". Same state transitions as for true elements.
169 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
171 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
172 /// define a state machine that triggers for ranges of values that the index
173 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
174 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
175 /// index in the range (inclusive). We use -2 for undefined here because we
176 /// use relative comparisons and don't want 0-1 to match -1.
177 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
179 // MagicBitvector - This is a magic bitvector where we set a bit if the
180 // comparison is true for element 'i'. If there are 64 elements or less in
181 // the array, this will fully represent all the comparison results.
182 uint64_t MagicBitvector = 0;
184 // Scan the array and see if one of our patterns matches.
185 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
186 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
187 Constant *Elt = Init->getAggregateElement(i);
188 if (!Elt) return nullptr;
190 // If this is indexing an array of structures, get the structure element.
191 if (!LaterIndices.empty())
192 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
194 // If the element is masked, handle it.
195 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
197 // Find out if the comparison would be true or false for the i'th element.
198 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
199 CompareRHS, DL, &TLI);
200 // If the result is undef for this element, ignore it.
201 if (isa<UndefValue>(C)) {
202 // Extend range state machines to cover this element in case there is an
203 // undef in the middle of the range.
204 if (TrueRangeEnd == (int)i-1)
206 if (FalseRangeEnd == (int)i-1)
211 // If we can't compute the result for any of the elements, we have to give
212 // up evaluating the entire conditional.
213 if (!isa<ConstantInt>(C)) return nullptr;
215 // Otherwise, we know if the comparison is true or false for this element,
216 // update our state machines.
217 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
219 // State machine for single/double/range index comparison.
221 // Update the TrueElement state machine.
222 if (FirstTrueElement == Undefined)
223 FirstTrueElement = TrueRangeEnd = i; // First true element.
225 // Update double-compare state machine.
226 if (SecondTrueElement == Undefined)
227 SecondTrueElement = i;
229 SecondTrueElement = Overdefined;
231 // Update range state machine.
232 if (TrueRangeEnd == (int)i-1)
235 TrueRangeEnd = Overdefined;
238 // Update the FalseElement state machine.
239 if (FirstFalseElement == Undefined)
240 FirstFalseElement = FalseRangeEnd = i; // First false element.
242 // Update double-compare state machine.
243 if (SecondFalseElement == Undefined)
244 SecondFalseElement = i;
246 SecondFalseElement = Overdefined;
248 // Update range state machine.
249 if (FalseRangeEnd == (int)i-1)
252 FalseRangeEnd = Overdefined;
256 // If this element is in range, update our magic bitvector.
257 if (i < 64 && IsTrueForElt)
258 MagicBitvector |= 1ULL << i;
260 // If all of our states become overdefined, bail out early. Since the
261 // predicate is expensive, only check it every 8 elements. This is only
262 // really useful for really huge arrays.
263 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
264 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
265 FalseRangeEnd == Overdefined)
269 // Now that we've scanned the entire array, emit our new comparison(s). We
270 // order the state machines in complexity of the generated code.
271 Value *Idx = GEP->getOperand(2);
273 // If the index is larger than the pointer size of the target, truncate the
274 // index down like the GEP would do implicitly. We don't have to do this for
275 // an inbounds GEP because the index can't be out of range.
276 if (!GEP->isInBounds()) {
277 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
278 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
279 if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
280 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
283 // If inbounds keyword is not present, Idx * ElementSize can overflow.
284 // Let's assume that ElementSize is 2 and the wanted value is at offset 0.
285 // Then, there are two possible values for Idx to match offset 0:
286 // 0x00..00, 0x80..00.
287 // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
288 // comparison is false if Idx was 0x80..00.
289 // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
290 unsigned ElementSize =
291 DL.getTypeAllocSize(Init->getType()->getArrayElementType());
292 auto MaskIdx = [&](Value* Idx){
293 if (!GEP->isInBounds() && countTrailingZeros(ElementSize) != 0) {
294 Value *Mask = ConstantInt::get(Idx->getType(), -1);
295 Mask = Builder.CreateLShr(Mask, countTrailingZeros(ElementSize));
296 Idx = Builder.CreateAnd(Idx, Mask);
301 // If the comparison is only true for one or two elements, emit direct
303 if (SecondTrueElement != Overdefined) {
305 // None true -> false.
306 if (FirstTrueElement == Undefined)
307 return replaceInstUsesWith(ICI, Builder.getFalse());
309 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
311 // True for one element -> 'i == 47'.
312 if (SecondTrueElement == Undefined)
313 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
315 // True for two elements -> 'i == 47 | i == 72'.
316 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
317 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
318 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
319 return BinaryOperator::CreateOr(C1, C2);
322 // If the comparison is only false for one or two elements, emit direct
324 if (SecondFalseElement != Overdefined) {
326 // None false -> true.
327 if (FirstFalseElement == Undefined)
328 return replaceInstUsesWith(ICI, Builder.getTrue());
330 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
332 // False for one element -> 'i != 47'.
333 if (SecondFalseElement == Undefined)
334 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
336 // False for two elements -> 'i != 47 & i != 72'.
337 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
338 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
339 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
340 return BinaryOperator::CreateAnd(C1, C2);
343 // If the comparison can be replaced with a range comparison for the elements
344 // where it is true, emit the range check.
345 if (TrueRangeEnd != Overdefined) {
346 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
349 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
350 if (FirstTrueElement) {
351 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
352 Idx = Builder.CreateAdd(Idx, Offs);
355 Value *End = ConstantInt::get(Idx->getType(),
356 TrueRangeEnd-FirstTrueElement+1);
357 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
360 // False range check.
361 if (FalseRangeEnd != Overdefined) {
362 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
364 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
365 if (FirstFalseElement) {
366 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
367 Idx = Builder.CreateAdd(Idx, Offs);
370 Value *End = ConstantInt::get(Idx->getType(),
371 FalseRangeEnd-FirstFalseElement);
372 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
375 // If a magic bitvector captures the entire comparison state
376 // of this load, replace it with computation that does:
377 // ((magic_cst >> i) & 1) != 0
381 // Look for an appropriate type:
382 // - The type of Idx if the magic fits
383 // - The smallest fitting legal type
384 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
387 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
391 Value *V = Builder.CreateIntCast(Idx, Ty, false);
392 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
393 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
394 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
401 /// Return a value that can be used to compare the *offset* implied by a GEP to
402 /// zero. For example, if we have &A[i], we want to return 'i' for
403 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
404 /// are involved. The above expression would also be legal to codegen as
405 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
406 /// This latter form is less amenable to optimization though, and we are allowed
407 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
409 /// If we can't emit an optimized form for this expression, this returns null.
411 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombinerImpl &IC,
412 const DataLayout &DL) {
413 gep_type_iterator GTI = gep_type_begin(GEP);
415 // Check to see if this gep only has a single variable index. If so, and if
416 // any constant indices are a multiple of its scale, then we can compute this
417 // in terms of the scale of the variable index. For example, if the GEP
418 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
419 // because the expression will cross zero at the same point.
420 unsigned i, e = GEP->getNumOperands();
422 for (i = 1; i != e; ++i, ++GTI) {
423 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
424 // Compute the aggregate offset of constant indices.
425 if (CI->isZero()) continue;
427 // Handle a struct index, which adds its field offset to the pointer.
428 if (StructType *STy = GTI.getStructTypeOrNull()) {
429 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
431 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
432 Offset += Size*CI->getSExtValue();
435 // Found our variable index.
440 // If there are no variable indices, we must have a constant offset, just
441 // evaluate it the general way.
442 if (i == e) return nullptr;
444 Value *VariableIdx = GEP->getOperand(i);
445 // Determine the scale factor of the variable element. For example, this is
446 // 4 if the variable index is into an array of i32.
447 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
449 // Verify that there are no other variable indices. If so, emit the hard way.
450 for (++i, ++GTI; i != e; ++i, ++GTI) {
451 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
452 if (!CI) return nullptr;
454 // Compute the aggregate offset of constant indices.
455 if (CI->isZero()) continue;
457 // Handle a struct index, which adds its field offset to the pointer.
458 if (StructType *STy = GTI.getStructTypeOrNull()) {
459 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
461 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
462 Offset += Size*CI->getSExtValue();
466 // Okay, we know we have a single variable index, which must be a
467 // pointer/array/vector index. If there is no offset, life is simple, return
469 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
470 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
472 // Cast to intptrty in case a truncation occurs. If an extension is needed,
473 // we don't need to bother extending: the extension won't affect where the
474 // computation crosses zero.
475 if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
477 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
482 // Otherwise, there is an index. The computation we will do will be modulo
484 Offset = SignExtend64(Offset, IntPtrWidth);
485 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
487 // To do this transformation, any constant index must be a multiple of the
488 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
489 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
490 // multiple of the variable scale.
491 int64_t NewOffs = Offset / (int64_t)VariableScale;
492 if (Offset != NewOffs*(int64_t)VariableScale)
495 // Okay, we can do this evaluation. Start by converting the index to intptr.
496 if (VariableIdx->getType() != IntPtrTy)
497 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
499 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
500 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
503 /// Returns true if we can rewrite Start as a GEP with pointer Base
504 /// and some integer offset. The nodes that need to be re-written
505 /// for this transformation will be added to Explored.
506 static bool canRewriteGEPAsOffset(Type *ElemTy, Value *Start, Value *Base,
507 const DataLayout &DL,
508 SetVector<Value *> &Explored) {
509 SmallVector<Value *, 16> WorkList(1, Start);
510 Explored.insert(Base);
512 // The following traversal gives us an order which can be used
513 // when doing the final transformation. Since in the final
514 // transformation we create the PHI replacement instructions first,
515 // we don't have to get them in any particular order.
517 // However, for other instructions we will have to traverse the
518 // operands of an instruction first, which means that we have to
519 // do a post-order traversal.
520 while (!WorkList.empty()) {
521 SetVector<PHINode *> PHIs;
523 while (!WorkList.empty()) {
524 if (Explored.size() >= 100)
527 Value *V = WorkList.back();
529 if (Explored.contains(V)) {
534 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
535 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
536 // We've found some value that we can't explore which is different from
537 // the base. Therefore we can't do this transformation.
540 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
541 auto *CI = cast<CastInst>(V);
542 if (!CI->isNoopCast(DL))
545 if (!Explored.contains(CI->getOperand(0)))
546 WorkList.push_back(CI->getOperand(0));
549 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
550 // We're limiting the GEP to having one index. This will preserve
551 // the original pointer type. We could handle more cases in the
553 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
554 GEP->getSourceElementType() != ElemTy)
557 if (!Explored.contains(GEP->getOperand(0)))
558 WorkList.push_back(GEP->getOperand(0));
561 if (WorkList.back() == V) {
563 // We've finished visiting this node, mark it as such.
567 if (auto *PN = dyn_cast<PHINode>(V)) {
568 // We cannot transform PHIs on unsplittable basic blocks.
569 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
576 // Explore the PHI nodes further.
577 for (auto *PN : PHIs)
578 for (Value *Op : PN->incoming_values())
579 if (!Explored.contains(Op))
580 WorkList.push_back(Op);
583 // Make sure that we can do this. Since we can't insert GEPs in a basic
584 // block before a PHI node, we can't easily do this transformation if
585 // we have PHI node users of transformed instructions.
586 for (Value *Val : Explored) {
587 for (Value *Use : Val->uses()) {
589 auto *PHI = dyn_cast<PHINode>(Use);
590 auto *Inst = dyn_cast<Instruction>(Val);
592 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
593 !Explored.contains(PHI))
596 if (PHI->getParent() == Inst->getParent())
603 // Sets the appropriate insert point on Builder where we can add
604 // a replacement Instruction for V (if that is possible).
605 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
606 bool Before = true) {
607 if (auto *PHI = dyn_cast<PHINode>(V)) {
608 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
611 if (auto *I = dyn_cast<Instruction>(V)) {
613 I = &*std::next(I->getIterator());
614 Builder.SetInsertPoint(I);
617 if (auto *A = dyn_cast<Argument>(V)) {
618 // Set the insertion point in the entry block.
619 BasicBlock &Entry = A->getParent()->getEntryBlock();
620 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
623 // Otherwise, this is a constant and we don't need to set a new
625 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
628 /// Returns a re-written value of Start as an indexed GEP using Base as a
630 static Value *rewriteGEPAsOffset(Type *ElemTy, Value *Start, Value *Base,
631 const DataLayout &DL,
632 SetVector<Value *> &Explored) {
633 // Perform all the substitutions. This is a bit tricky because we can
634 // have cycles in our use-def chains.
635 // 1. Create the PHI nodes without any incoming values.
636 // 2. Create all the other values.
637 // 3. Add the edges for the PHI nodes.
638 // 4. Emit GEPs to get the original pointers.
639 // 5. Remove the original instructions.
640 Type *IndexType = IntegerType::get(
641 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
643 DenseMap<Value *, Value *> NewInsts;
644 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
646 // Create the new PHI nodes, without adding any incoming values.
647 for (Value *Val : Explored) {
650 // Create empty phi nodes. This avoids cyclic dependencies when creating
651 // the remaining instructions.
652 if (auto *PHI = dyn_cast<PHINode>(Val))
653 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
654 PHI->getName() + ".idx", PHI);
656 IRBuilder<> Builder(Base->getContext());
658 // Create all the other instructions.
659 for (Value *Val : Explored) {
661 if (NewInsts.find(Val) != NewInsts.end())
664 if (auto *CI = dyn_cast<CastInst>(Val)) {
665 // Don't get rid of the intermediate variable here; the store can grow
666 // the map which will invalidate the reference to the input value.
667 Value *V = NewInsts[CI->getOperand(0)];
671 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
672 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
673 : GEP->getOperand(1);
674 setInsertionPoint(Builder, GEP);
675 // Indices might need to be sign extended. GEPs will magically do
676 // this, but we need to do it ourselves here.
677 if (Index->getType()->getScalarSizeInBits() !=
678 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
679 Index = Builder.CreateSExtOrTrunc(
680 Index, NewInsts[GEP->getOperand(0)]->getType(),
681 GEP->getOperand(0)->getName() + ".sext");
684 auto *Op = NewInsts[GEP->getOperand(0)];
685 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
686 NewInsts[GEP] = Index;
688 NewInsts[GEP] = Builder.CreateNSWAdd(
689 Op, Index, GEP->getOperand(0)->getName() + ".add");
692 if (isa<PHINode>(Val))
695 llvm_unreachable("Unexpected instruction type");
698 // Add the incoming values to the PHI nodes.
699 for (Value *Val : Explored) {
702 // All the instructions have been created, we can now add edges to the
704 if (auto *PHI = dyn_cast<PHINode>(Val)) {
705 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
706 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
707 Value *NewIncoming = PHI->getIncomingValue(I);
709 if (NewInsts.find(NewIncoming) != NewInsts.end())
710 NewIncoming = NewInsts[NewIncoming];
712 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
718 ElemTy->getPointerTo(Start->getType()->getPointerAddressSpace());
719 for (Value *Val : Explored) {
723 // Depending on the type, for external users we have to emit
724 // a GEP or a GEP + ptrtoint.
725 setInsertionPoint(Builder, Val, false);
727 // Cast base to the expected type.
728 Value *NewVal = Builder.CreateBitOrPointerCast(
729 Base, PtrTy, Start->getName() + "to.ptr");
730 NewVal = Builder.CreateInBoundsGEP(
731 ElemTy, NewVal, makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
732 NewVal = Builder.CreateBitOrPointerCast(
733 NewVal, Val->getType(), Val->getName() + ".conv");
734 Val->replaceAllUsesWith(NewVal);
737 return NewInsts[Start];
740 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
741 /// the input Value as a constant indexed GEP. Returns a pair containing
742 /// the GEPs Pointer and Index.
743 static std::pair<Value *, Value *>
744 getAsConstantIndexedAddress(Type *ElemTy, Value *V, const DataLayout &DL) {
745 Type *IndexType = IntegerType::get(V->getContext(),
746 DL.getIndexTypeSizeInBits(V->getType()));
748 Constant *Index = ConstantInt::getNullValue(IndexType);
750 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
751 // We accept only inbouds GEPs here to exclude the possibility of
753 if (!GEP->isInBounds())
755 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
756 GEP->getSourceElementType() == ElemTy) {
757 V = GEP->getOperand(0);
758 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
759 Index = ConstantExpr::getAdd(
760 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
765 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
766 if (!CI->isNoopCast(DL))
768 V = CI->getOperand(0);
771 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
772 if (!CI->isNoopCast(DL))
774 V = CI->getOperand(0);
782 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
783 /// We can look through PHIs, GEPs and casts in order to determine a common base
784 /// between GEPLHS and RHS.
785 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
786 ICmpInst::Predicate Cond,
787 const DataLayout &DL) {
788 // FIXME: Support vector of pointers.
789 if (GEPLHS->getType()->isVectorTy())
792 if (!GEPLHS->hasAllConstantIndices())
795 Type *ElemTy = GEPLHS->getSourceElementType();
796 Value *PtrBase, *Index;
797 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(ElemTy, GEPLHS, DL);
799 // The set of nodes that will take part in this transformation.
800 SetVector<Value *> Nodes;
802 if (!canRewriteGEPAsOffset(ElemTy, RHS, PtrBase, DL, Nodes))
805 // We know we can re-write this as
806 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
807 // Since we've only looked through inbouds GEPs we know that we
808 // can't have overflow on either side. We can therefore re-write
810 // OFFSET1 cmp OFFSET2
811 Value *NewRHS = rewriteGEPAsOffset(ElemTy, RHS, PtrBase, DL, Nodes);
813 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
814 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
815 // offset. Since Index is the offset of LHS to the base pointer, we will now
816 // compare the offsets instead of comparing the pointers.
817 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
820 /// Fold comparisons between a GEP instruction and something else. At this point
821 /// we know that the GEP is on the LHS of the comparison.
822 Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
823 ICmpInst::Predicate Cond,
825 // Don't transform signed compares of GEPs into index compares. Even if the
826 // GEP is inbounds, the final add of the base pointer can have signed overflow
827 // and would change the result of the icmp.
828 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
829 // the maximum signed value for the pointer type.
830 if (ICmpInst::isSigned(Cond))
833 // Look through bitcasts and addrspacecasts. We do not however want to remove
835 if (!isa<GetElementPtrInst>(RHS))
836 RHS = RHS->stripPointerCasts();
838 Value *PtrBase = GEPLHS->getOperand(0);
839 // FIXME: Support vector pointer GEPs.
840 if (PtrBase == RHS && GEPLHS->isInBounds() &&
841 !GEPLHS->getType()->isVectorTy()) {
842 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
843 // This transformation (ignoring the base and scales) is valid because we
844 // know pointers can't overflow since the gep is inbounds. See if we can
845 // output an optimized form.
846 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
848 // If not, synthesize the offset the hard way.
850 Offset = EmitGEPOffset(GEPLHS);
851 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
852 Constant::getNullValue(Offset->getType()));
855 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
856 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
857 !NullPointerIsDefined(I.getFunction(),
858 RHS->getType()->getPointerAddressSpace())) {
859 // For most address spaces, an allocation can't be placed at null, but null
860 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
861 // the only valid inbounds address derived from null, is null itself.
862 // Thus, we have four cases to consider:
863 // 1) Base == nullptr, Offset == 0 -> inbounds, null
864 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
865 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
866 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
868 // (Note if we're indexing a type of size 0, that simply collapses into one
869 // of the buckets above.)
871 // In general, we're allowed to make values less poison (i.e. remove
872 // sources of full UB), so in this case, we just select between the two
873 // non-poison cases (1 and 4 above).
875 // For vectors, we apply the same reasoning on a per-lane basis.
876 auto *Base = GEPLHS->getPointerOperand();
877 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
878 auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
879 Base = Builder.CreateVectorSplat(EC, Base);
881 return new ICmpInst(Cond, Base,
882 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
883 cast<Constant>(RHS), Base->getType()));
884 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
885 // If the base pointers are different, but the indices are the same, just
886 // compare the base pointer.
887 if (PtrBase != GEPRHS->getOperand(0)) {
888 bool IndicesTheSame =
889 GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
890 GEPLHS->getType() == GEPRHS->getType() &&
891 GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType();
893 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
894 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
895 IndicesTheSame = false;
899 // If all indices are the same, just compare the base pointers.
900 Type *BaseType = GEPLHS->getOperand(0)->getType();
901 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
902 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
904 // If we're comparing GEPs with two base pointers that only differ in type
905 // and both GEPs have only constant indices or just one use, then fold
906 // the compare with the adjusted indices.
907 // FIXME: Support vector of pointers.
908 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
909 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
910 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
911 PtrBase->stripPointerCasts() ==
912 GEPRHS->getOperand(0)->stripPointerCasts() &&
913 !GEPLHS->getType()->isVectorTy()) {
914 Value *LOffset = EmitGEPOffset(GEPLHS);
915 Value *ROffset = EmitGEPOffset(GEPRHS);
917 // If we looked through an addrspacecast between different sized address
918 // spaces, the LHS and RHS pointers are different sized
919 // integers. Truncate to the smaller one.
920 Type *LHSIndexTy = LOffset->getType();
921 Type *RHSIndexTy = ROffset->getType();
922 if (LHSIndexTy != RHSIndexTy) {
923 if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
924 RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
925 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
927 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
930 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
932 return replaceInstUsesWith(I, Cmp);
935 // Otherwise, the base pointers are different and the indices are
936 // different. Try convert this to an indexed compare by looking through
938 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
941 // If one of the GEPs has all zero indices, recurse.
942 // FIXME: Handle vector of pointers.
943 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
944 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
945 ICmpInst::getSwappedPredicate(Cond), I);
947 // If the other GEP has all zero indices, recurse.
948 // FIXME: Handle vector of pointers.
949 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
950 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
952 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
953 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
954 // If the GEPs only differ by one index, compare it.
955 unsigned NumDifferences = 0; // Keep track of # differences.
956 unsigned DiffOperand = 0; // The operand that differs.
957 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
958 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
959 Type *LHSType = GEPLHS->getOperand(i)->getType();
960 Type *RHSType = GEPRHS->getOperand(i)->getType();
961 // FIXME: Better support for vector of pointers.
962 if (LHSType->getPrimitiveSizeInBits() !=
963 RHSType->getPrimitiveSizeInBits() ||
964 (GEPLHS->getType()->isVectorTy() &&
965 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
966 // Irreconcilable differences.
971 if (NumDifferences++) break;
975 if (NumDifferences == 0) // SAME GEP?
976 return replaceInstUsesWith(I, // No comparison is needed here.
977 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
979 else if (NumDifferences == 1 && GEPsInBounds) {
980 Value *LHSV = GEPLHS->getOperand(DiffOperand);
981 Value *RHSV = GEPRHS->getOperand(DiffOperand);
982 // Make sure we do a signed comparison here.
983 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
987 // Only lower this if the icmp is the only user of the GEP or if we expect
988 // the result to fold to a constant!
989 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
990 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
991 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
992 Value *L = EmitGEPOffset(GEPLHS);
993 Value *R = EmitGEPOffset(GEPRHS);
994 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
998 // Try convert this to an indexed compare by looking through PHIs/casts as a
1000 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1003 Instruction *InstCombinerImpl::foldAllocaCmp(ICmpInst &ICI,
1004 const AllocaInst *Alloca,
1005 const Value *Other) {
1006 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1008 // It would be tempting to fold away comparisons between allocas and any
1009 // pointer not based on that alloca (e.g. an argument). However, even
1010 // though such pointers cannot alias, they can still compare equal.
1012 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1013 // doesn't escape we can argue that it's impossible to guess its value, and we
1014 // can therefore act as if any such guesses are wrong.
1016 // The code below checks that the alloca doesn't escape, and that it's only
1017 // used in a comparison once (the current instruction). The
1018 // single-comparison-use condition ensures that we're trivially folding all
1019 // comparisons against the alloca consistently, and avoids the risk of
1020 // erroneously folding a comparison of the pointer with itself.
1022 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1024 SmallVector<const Use *, 32> Worklist;
1025 for (const Use &U : Alloca->uses()) {
1026 if (Worklist.size() >= MaxIter)
1028 Worklist.push_back(&U);
1031 unsigned NumCmps = 0;
1032 while (!Worklist.empty()) {
1033 assert(Worklist.size() <= MaxIter);
1034 const Use *U = Worklist.pop_back_val();
1035 const Value *V = U->getUser();
1038 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1039 isa<SelectInst>(V)) {
1041 } else if (isa<LoadInst>(V)) {
1042 // Loading from the pointer doesn't escape it.
1044 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1045 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1046 if (SI->getValueOperand() == U->get())
1049 } else if (isa<ICmpInst>(V)) {
1051 return nullptr; // Found more than one cmp.
1053 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1054 switch (Intrin->getIntrinsicID()) {
1055 // These intrinsics don't escape or compare the pointer. Memset is safe
1056 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1057 // we don't allow stores, so src cannot point to V.
1058 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1059 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1067 for (const Use &U : V->uses()) {
1068 if (Worklist.size() >= MaxIter)
1070 Worklist.push_back(&U);
1074 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1075 return replaceInstUsesWith(
1077 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1080 /// Fold "icmp pred (X+C), X".
1081 Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
1082 ICmpInst::Predicate Pred) {
1083 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1084 // so the values can never be equal. Similarly for all other "or equals"
1086 assert(!!C && "C should not be zero!");
1088 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1089 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1090 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1091 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1092 Constant *R = ConstantInt::get(X->getType(),
1093 APInt::getMaxValue(C.getBitWidth()) - C);
1094 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1097 // (X+1) >u X --> X <u (0-1) --> X != 255
1098 // (X+2) >u X --> X <u (0-2) --> X <u 254
1099 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1100 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1101 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1102 ConstantInt::get(X->getType(), -C));
1104 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1106 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1107 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1108 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1109 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1110 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1111 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1112 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1113 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1114 ConstantInt::get(X->getType(), SMax - C));
1116 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1117 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1118 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1119 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1120 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1121 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1123 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1124 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1125 ConstantInt::get(X->getType(), SMax - (C - 1)));
1128 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1129 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1130 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1131 Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
1134 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1136 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1137 if (I.getPredicate() == I.ICMP_NE)
1138 Pred = CmpInst::getInversePredicate(Pred);
1139 return new ICmpInst(Pred, LHS, RHS);
1142 // Don't bother doing any work for cases which InstSimplify handles.
1146 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1148 if (AP2.isAllOnes())
1150 if (AP2.isNegative() != AP1.isNegative())
1157 // 'A' must be large enough to shift out the highest set bit.
1158 return getICmp(I.ICMP_UGT, A,
1159 ConstantInt::get(A->getType(), AP2.logBase2()));
1162 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1165 if (IsAShr && AP1.isNegative())
1166 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1168 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1171 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1172 // There are multiple solutions if we are comparing against -1 and the LHS
1173 // of the ashr is not a power of two.
1174 if (AP1.isAllOnes() && !AP2.isPowerOf2())
1175 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1176 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1177 } else if (AP1 == AP2.lshr(Shift)) {
1178 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1182 // Shifting const2 will never be equal to const1.
1183 // FIXME: This should always be handled by InstSimplify?
1184 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1185 return replaceInstUsesWith(I, TorF);
1188 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1189 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1190 Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1193 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1195 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1196 if (I.getPredicate() == I.ICMP_NE)
1197 Pred = CmpInst::getInversePredicate(Pred);
1198 return new ICmpInst(Pred, LHS, RHS);
1201 // Don't bother doing any work for cases which InstSimplify handles.
1205 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1207 if (!AP1 && AP2TrailingZeros != 0)
1210 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1213 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1215 // Get the distance between the lowest bits that are set.
1216 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1218 if (Shift > 0 && AP2.shl(Shift) == AP1)
1219 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1221 // Shifting const2 will never be equal to const1.
1222 // FIXME: This should always be handled by InstSimplify?
1223 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1224 return replaceInstUsesWith(I, TorF);
1227 /// The caller has matched a pattern of the form:
1228 /// I = icmp ugt (add (add A, B), CI2), CI1
1229 /// If this is of the form:
1231 /// if (sum+128 >u 255)
1232 /// Then replace it with llvm.sadd.with.overflow.i8.
1234 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1235 ConstantInt *CI2, ConstantInt *CI1,
1236 InstCombinerImpl &IC) {
1237 // The transformation we're trying to do here is to transform this into an
1238 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1239 // with a narrower add, and discard the add-with-constant that is part of the
1240 // range check (if we can't eliminate it, this isn't profitable).
1242 // In order to eliminate the add-with-constant, the compare can be its only
1244 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1245 if (!AddWithCst->hasOneUse())
1248 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1249 if (!CI2->getValue().isPowerOf2())
1251 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1252 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1255 // The width of the new add formed is 1 more than the bias.
1258 // Check to see that CI1 is an all-ones value with NewWidth bits.
1259 if (CI1->getBitWidth() == NewWidth ||
1260 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1263 // This is only really a signed overflow check if the inputs have been
1264 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1265 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1266 if (IC.ComputeMaxSignificantBits(A, 0, &I) > NewWidth ||
1267 IC.ComputeMaxSignificantBits(B, 0, &I) > NewWidth)
1270 // In order to replace the original add with a narrower
1271 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1272 // and truncates that discard the high bits of the add. Verify that this is
1274 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1275 for (User *U : OrigAdd->users()) {
1276 if (U == AddWithCst)
1279 // Only accept truncates for now. We would really like a nice recursive
1280 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1281 // chain to see which bits of a value are actually demanded. If the
1282 // original add had another add which was then immediately truncated, we
1283 // could still do the transformation.
1284 TruncInst *TI = dyn_cast<TruncInst>(U);
1285 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1289 // If the pattern matches, truncate the inputs to the narrower type and
1290 // use the sadd_with_overflow intrinsic to efficiently compute both the
1291 // result and the overflow bit.
1292 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1293 Function *F = Intrinsic::getDeclaration(
1294 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1296 InstCombiner::BuilderTy &Builder = IC.Builder;
1298 // Put the new code above the original add, in case there are any uses of the
1299 // add between the add and the compare.
1300 Builder.SetInsertPoint(OrigAdd);
1302 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1303 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1304 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1305 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1306 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1308 // The inner add was the result of the narrow add, zero extended to the
1309 // wider type. Replace it with the result computed by the intrinsic.
1310 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1311 IC.eraseInstFromFunction(*OrigAdd);
1313 // The original icmp gets replaced with the overflow value.
1314 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1318 /// icmp eq/ne (urem/srem %x, %y), 0
1319 /// iff %y is a power-of-two, we can replace this with a bit test:
1320 /// icmp eq/ne (and %x, (add %y, -1)), 0
1321 Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1322 // This fold is only valid for equality predicates.
1323 if (!I.isEquality())
1325 ICmpInst::Predicate Pred;
1326 Value *X, *Y, *Zero;
1327 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1328 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1330 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1332 // This may increase instruction count, we don't enforce that Y is a constant.
1333 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1334 Value *Masked = Builder.CreateAnd(X, Mask);
1335 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1338 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1339 /// by one-less-than-bitwidth into a sign test on the original value.
1340 Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1342 ICmpInst::Predicate Pred;
1343 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1350 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1352 unsigned XBitWidth = XTy->getScalarSizeInBits();
1353 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1354 APInt(XBitWidth, XBitWidth - 1))))
1356 } else if (isa<BinaryOperator>(Val) &&
1357 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1358 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1359 /*AnalyzeForSignBitExtraction=*/true))) {
1364 return ICmpInst::Create(Instruction::ICmp,
1365 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1366 : ICmpInst::ICMP_SLT,
1367 X, ConstantInt::getNullValue(XTy));
1370 // Handle icmp pred X, 0
1371 Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1372 CmpInst::Predicate Pred = Cmp.getPredicate();
1373 if (!match(Cmp.getOperand(1), m_Zero()))
1376 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1377 if (Pred == ICmpInst::ICMP_SGT) {
1379 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1380 if (SPR.Flavor == SPF_SMIN) {
1381 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1382 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1383 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1384 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1388 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1392 // icmp eq/ne (urem %x, %y), 0
1393 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1396 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1397 ICmpInst::isEquality(Pred)) {
1398 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1399 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1400 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1401 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1407 /// Fold icmp Pred X, C.
1408 /// TODO: This code structure does not make sense. The saturating add fold
1409 /// should be moved to some other helper and extended as noted below (it is also
1410 /// possible that code has been made unnecessary - do we canonicalize IR to
1411 /// overflow/saturating intrinsics or not?).
1412 Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1413 // Match the following pattern, which is a common idiom when writing
1414 // overflow-safe integer arithmetic functions. The source performs an addition
1415 // in wider type and explicitly checks for overflow using comparisons against
1416 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1418 // TODO: This could probably be generalized to handle other overflow-safe
1419 // operations if we worked out the formulas to compute the appropriate magic
1423 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1424 CmpInst::Predicate Pred = Cmp.getPredicate();
1425 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1427 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1428 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1429 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1430 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1433 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1434 Constant *C = dyn_cast<Constant>(Op1);
1435 if (!C || C->canTrap())
1438 if (auto *Phi = dyn_cast<PHINode>(Op0))
1439 if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1440 Type *Ty = Cmp.getType();
1441 Builder.SetInsertPoint(Phi);
1443 Builder.CreatePHI(Ty, Phi->getNumOperands());
1444 for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1446 cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1447 auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1448 NewPhi->addIncoming(BoolInput, Predecessor);
1450 NewPhi->takeName(&Cmp);
1451 return replaceInstUsesWith(Cmp, NewPhi);
1457 /// Canonicalize icmp instructions based on dominating conditions.
1458 Instruction *InstCombinerImpl::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::makeExactICmpRegion(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 // Avoid an infinite loop with min/max canonicalization.
1519 // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1520 if (Cmp.hasOneUse() &&
1521 match(Cmp.user_back(), m_MaxOrMin(m_Value(), m_Value())))
1524 if (const APInt *EqC = Intersection.getSingleElement())
1525 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1526 if (const APInt *NeC = Difference.getSingleElement())
1527 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1533 /// Fold icmp (trunc X, Y), C.
1534 Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1537 ICmpInst::Predicate Pred = Cmp.getPredicate();
1538 Value *X = Trunc->getOperand(0);
1539 if (C.isOne() && C.getBitWidth() > 1) {
1540 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1542 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1543 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1544 ConstantInt::get(V->getType(), 1));
1547 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1548 SrcBits = X->getType()->getScalarSizeInBits();
1549 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1550 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1551 // of the high bits truncated out of x are known.
1552 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1554 // If all the high bits are known, we can do this xform.
1555 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1556 // Pull in the high bits from known-ones set.
1557 APInt NewRHS = C.zext(SrcBits);
1558 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1559 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1563 // Look through truncated right-shift of the sign-bit for a sign-bit check:
1564 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1565 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1567 const APInt *ShAmtC;
1569 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1570 match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1571 DstBits == SrcBits - ShAmtC->getZExtValue()) {
1573 ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1574 ConstantInt::getNullValue(X->getType()))
1575 : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1576 ConstantInt::getAllOnesValue(X->getType()));
1582 /// Fold icmp (xor X, Y), C.
1583 Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1584 BinaryOperator *Xor,
1586 Value *X = Xor->getOperand(0);
1587 Value *Y = Xor->getOperand(1);
1589 if (!match(Y, m_APInt(XorC)))
1592 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1594 ICmpInst::Predicate Pred = Cmp.getPredicate();
1595 bool TrueIfSigned = false;
1596 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1598 // If the sign bit of the XorCst is not set, there is no change to
1599 // the operation, just stop using the Xor.
1600 if (!XorC->isNegative())
1601 return replaceOperand(Cmp, 0, X);
1603 // Emit the opposite comparison.
1605 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1606 ConstantInt::getAllOnesValue(X->getType()));
1608 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1609 ConstantInt::getNullValue(X->getType()));
1612 if (Xor->hasOneUse()) {
1613 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1614 if (!Cmp.isEquality() && XorC->isSignMask()) {
1615 Pred = Cmp.getFlippedSignednessPredicate();
1616 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1619 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1620 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1621 Pred = Cmp.getFlippedSignednessPredicate();
1622 Pred = Cmp.getSwappedPredicate(Pred);
1623 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1627 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1628 if (Pred == ICmpInst::ICMP_UGT) {
1629 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1630 if (*XorC == ~C && (C + 1).isPowerOf2())
1631 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1632 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1633 if (*XorC == C && (C + 1).isPowerOf2())
1634 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1636 if (Pred == ICmpInst::ICMP_ULT) {
1637 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1638 if (*XorC == -C && C.isPowerOf2())
1639 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1640 ConstantInt::get(X->getType(), ~C));
1641 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1642 if (*XorC == C && (-C).isPowerOf2())
1643 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1644 ConstantInt::get(X->getType(), ~C));
1649 /// Fold icmp (and (sh X, Y), C2), C1.
1650 Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1651 BinaryOperator *And,
1654 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1655 if (!Shift || !Shift->isShift())
1658 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1659 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1660 // code produced by the clang front-end, for bitfield access.
1661 // This seemingly simple opportunity to fold away a shift turns out to be
1662 // rather complicated. See PR17827 for details.
1663 unsigned ShiftOpcode = Shift->getOpcode();
1664 bool IsShl = ShiftOpcode == Instruction::Shl;
1666 if (match(Shift->getOperand(1), m_APInt(C3))) {
1667 APInt NewAndCst, NewCmpCst;
1668 bool AnyCmpCstBitsShiftedOut;
1669 if (ShiftOpcode == Instruction::Shl) {
1670 // For a left shift, we can fold if the comparison is not signed. We can
1671 // also fold a signed comparison if the mask value and comparison value
1672 // are not negative. These constraints may not be obvious, but we can
1673 // prove that they are correct using an SMT solver.
1674 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1677 NewCmpCst = C1.lshr(*C3);
1678 NewAndCst = C2.lshr(*C3);
1679 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1680 } else if (ShiftOpcode == Instruction::LShr) {
1681 // For a logical right shift, we can fold if the comparison is not signed.
1682 // We can also fold a signed comparison if the shifted mask value and the
1683 // shifted comparison value are not negative. These constraints may not be
1684 // obvious, but we can prove that they are correct using an SMT solver.
1685 NewCmpCst = C1.shl(*C3);
1686 NewAndCst = C2.shl(*C3);
1687 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1688 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1691 // For an arithmetic shift, check that both constants don't use (in a
1692 // signed sense) the top bits being shifted out.
1693 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1694 NewCmpCst = C1.shl(*C3);
1695 NewAndCst = C2.shl(*C3);
1696 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1697 if (NewAndCst.ashr(*C3) != C2)
1701 if (AnyCmpCstBitsShiftedOut) {
1702 // If we shifted bits out, the fold is not going to work out. As a
1703 // special case, check to see if this means that the result is always
1704 // true or false now.
1705 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1706 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1707 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1708 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1710 Value *NewAnd = Builder.CreateAnd(
1711 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1712 return new ICmpInst(Cmp.getPredicate(),
1713 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1717 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1718 // preferable because it allows the C2 << Y expression to be hoisted out of a
1719 // loop if Y is invariant and X is not.
1720 if (Shift->hasOneUse() && C1.isZero() && Cmp.isEquality() &&
1721 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1724 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1725 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1727 // Compute X & (C2 << Y).
1728 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1729 return replaceOperand(Cmp, 0, NewAnd);
1735 /// Fold icmp (and X, C2), C1.
1736 Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1737 BinaryOperator *And,
1739 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1741 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1742 // TODO: We canonicalize to the longer form for scalars because we have
1743 // better analysis/folds for icmp, and codegen may be better with icmp.
1744 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isZero() &&
1745 match(And->getOperand(1), m_One()))
1746 return new TruncInst(And->getOperand(0), Cmp.getType());
1750 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1753 // Don't perform the following transforms if the AND has multiple uses
1754 if (!And->hasOneUse())
1757 if (Cmp.isEquality() && C1.isZero()) {
1758 // Restrict this fold to single-use 'and' (PR10267).
1759 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1760 if (C2->isSignMask()) {
1761 Constant *Zero = Constant::getNullValue(X->getType());
1762 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1763 return new ICmpInst(NewPred, X, Zero);
1766 // Restrict this fold only for single-use 'and' (PR10267).
1767 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1768 if ((~(*C2) + 1).isPowerOf2()) {
1770 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1771 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1772 return new ICmpInst(NewPred, X, NegBOC);
1776 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1777 // the input width without changing the value produced, eliminate the cast:
1779 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1781 // We can do this transformation if the constants do not have their sign bits
1782 // set or if it is an equality comparison. Extending a relational comparison
1783 // when we're checking the sign bit would not work.
1785 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1786 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1787 // TODO: Is this a good transform for vectors? Wider types may reduce
1788 // throughput. Should this transform be limited (even for scalars) by using
1789 // shouldChangeType()?
1790 if (!Cmp.getType()->isVectorTy()) {
1791 Type *WideType = W->getType();
1792 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1793 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1794 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1795 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1796 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1800 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1803 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1804 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1806 // iff pred isn't signed
1807 if (!Cmp.isSigned() && C1.isZero() && And->getOperand(0)->hasOneUse() &&
1808 match(And->getOperand(1), m_One())) {
1809 Constant *One = cast<Constant>(And->getOperand(1));
1810 Value *Or = And->getOperand(0);
1811 Value *A, *B, *LShr;
1812 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1813 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1814 unsigned UsesRemoved = 0;
1815 if (And->hasOneUse())
1817 if (Or->hasOneUse())
1819 if (LShr->hasOneUse())
1822 // Compute A & ((1 << B) | 1)
1823 Value *NewOr = nullptr;
1824 if (auto *C = dyn_cast<Constant>(B)) {
1825 if (UsesRemoved >= 1)
1826 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1828 if (UsesRemoved >= 3)
1829 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1831 One, Or->getName());
1834 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1835 return replaceOperand(Cmp, 0, NewAnd);
1843 /// Fold icmp (and X, Y), C.
1844 Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1845 BinaryOperator *And,
1847 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1850 const ICmpInst::Predicate Pred = Cmp.getPredicate();
1852 if (isSignBitCheck(Pred, C, TrueIfNeg)) {
1853 // ((X - 1) & ~X) < 0 --> X == 0
1854 // ((X - 1) & ~X) >= 0 --> X != 0
1856 if (match(And->getOperand(0), m_Add(m_Value(X), m_AllOnes())) &&
1857 match(And->getOperand(1), m_Not(m_Specific(X)))) {
1858 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1859 return new ICmpInst(NewPred, X, ConstantInt::getNullValue(X->getType()));
1863 // TODO: These all require that Y is constant too, so refactor with the above.
1865 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1866 Value *X = And->getOperand(0);
1867 Value *Y = And->getOperand(1);
1868 if (auto *LI = dyn_cast<LoadInst>(X))
1869 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1870 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1871 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1872 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1873 ConstantInt *C2 = cast<ConstantInt>(Y);
1874 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1878 if (!Cmp.isEquality())
1881 // X & -C == -C -> X > u ~C
1882 // X & -C != -C -> X <= u ~C
1883 // iff C is a power of 2
1884 if (Cmp.getOperand(1) == Y && C.isNegatedPowerOf2()) {
1886 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1887 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1893 /// Fold icmp (or X, Y), C.
1894 Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
1897 ICmpInst::Predicate Pred = Cmp.getPredicate();
1899 // icmp slt signum(V) 1 --> icmp slt V, 1
1901 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1902 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1903 ConstantInt::get(V->getType(), 1));
1906 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1908 if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1909 if (*MaskC == C && (C + 1).isPowerOf2()) {
1910 // X | C == C --> X <=u C
1911 // X | C != C --> X >u C
1912 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1913 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1914 return new ICmpInst(Pred, OrOp0, OrOp1);
1917 // More general: canonicalize 'equality with set bits mask' to
1918 // 'equality with clear bits mask'.
1919 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1920 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1921 if (Or->hasOneUse()) {
1922 Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1923 Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1924 return new ICmpInst(Pred, And, NewC);
1928 // (X | (X-1)) s< 0 --> X s< 1
1929 // (X | (X-1)) s> -1 --> X s> 0
1932 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1933 match(Or, m_c_Or(m_Add(m_Value(X), m_AllOnes()), m_Deferred(X)))) {
1934 auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT;
1935 Constant *NewC = ConstantInt::get(X->getType(), TrueIfSigned ? 1 : 0);
1936 return new ICmpInst(NewPred, X, NewC);
1939 if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse())
1943 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1944 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1945 // -> and (icmp eq P, null), (icmp eq Q, null).
1947 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1949 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1950 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1951 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1954 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1955 // a shorter form that has more potential to be folded even further.
1956 Value *X1, *X2, *X3, *X4;
1957 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1958 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1959 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1960 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1961 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1962 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1963 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1964 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1970 /// Fold icmp (mul X, Y), C.
1971 Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
1972 BinaryOperator *Mul,
1975 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1978 // If this is a test of the sign bit and the multiply is sign-preserving with
1979 // a constant operand, use the multiply LHS operand instead.
1980 ICmpInst::Predicate Pred = Cmp.getPredicate();
1981 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1982 if (MulC->isNegative())
1983 Pred = ICmpInst::getSwappedPredicate(Pred);
1984 return new ICmpInst(Pred, Mul->getOperand(0),
1985 Constant::getNullValue(Mul->getType()));
1988 // If the multiply does not wrap, try to divide the compare constant by the
1989 // multiplication factor.
1990 if (Cmp.isEquality() && !MulC->isZero()) {
1991 // (mul nsw X, MulC) == C --> X == C /s MulC
1992 if (Mul->hasNoSignedWrap() && C.srem(*MulC).isZero()) {
1993 Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
1994 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1996 // (mul nuw X, MulC) == C --> X == C /u MulC
1997 if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isZero()) {
1998 Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
1999 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2006 /// Fold icmp (shl 1, Y), C.
2007 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
2010 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2013 Type *ShiftType = Shl->getType();
2014 unsigned TypeBits = C.getBitWidth();
2015 bool CIsPowerOf2 = C.isPowerOf2();
2016 ICmpInst::Predicate Pred = Cmp.getPredicate();
2017 if (Cmp.isUnsigned()) {
2018 // (1 << Y) pred C -> Y pred Log2(C)
2020 // (1 << Y) < 30 -> Y <= 4
2021 // (1 << Y) <= 30 -> Y <= 4
2022 // (1 << Y) >= 30 -> Y > 4
2023 // (1 << Y) > 30 -> Y > 4
2024 if (Pred == ICmpInst::ICMP_ULT)
2025 Pred = ICmpInst::ICMP_ULE;
2026 else if (Pred == ICmpInst::ICMP_UGE)
2027 Pred = ICmpInst::ICMP_UGT;
2030 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2031 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2032 unsigned CLog2 = C.logBase2();
2033 if (CLog2 == TypeBits - 1) {
2034 if (Pred == ICmpInst::ICMP_UGE)
2035 Pred = ICmpInst::ICMP_EQ;
2036 else if (Pred == ICmpInst::ICMP_ULT)
2037 Pred = ICmpInst::ICMP_NE;
2039 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2040 } else if (Cmp.isSigned()) {
2041 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2042 if (C.isAllOnes()) {
2043 // (1 << Y) <= -1 -> Y == 31
2044 if (Pred == ICmpInst::ICMP_SLE)
2045 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2047 // (1 << Y) > -1 -> Y != 31
2048 if (Pred == ICmpInst::ICMP_SGT)
2049 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2051 // (1 << Y) < 0 -> Y == 31
2052 // (1 << Y) <= 0 -> Y == 31
2053 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2054 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2056 // (1 << Y) >= 0 -> Y != 31
2057 // (1 << Y) > 0 -> Y != 31
2058 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2059 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2061 } else if (Cmp.isEquality() && CIsPowerOf2) {
2062 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2068 /// Fold icmp (shl X, Y), C.
2069 Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2070 BinaryOperator *Shl,
2072 const APInt *ShiftVal;
2073 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2074 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2076 const APInt *ShiftAmt;
2077 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2078 return foldICmpShlOne(Cmp, Shl, C);
2080 // Check that the shift amount is in range. If not, don't perform undefined
2081 // shifts. When the shift is visited, it will be simplified.
2082 unsigned TypeBits = C.getBitWidth();
2083 if (ShiftAmt->uge(TypeBits))
2086 ICmpInst::Predicate Pred = Cmp.getPredicate();
2087 Value *X = Shl->getOperand(0);
2088 Type *ShType = Shl->getType();
2090 // NSW guarantees that we are only shifting out sign bits from the high bits,
2091 // so we can ASHR the compare constant without needing a mask and eliminate
2093 if (Shl->hasNoSignedWrap()) {
2094 if (Pred == ICmpInst::ICMP_SGT) {
2095 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2096 APInt ShiftedC = C.ashr(*ShiftAmt);
2097 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2099 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2100 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2101 APInt ShiftedC = C.ashr(*ShiftAmt);
2102 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2104 if (Pred == ICmpInst::ICMP_SLT) {
2105 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2106 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2107 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2108 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2109 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2110 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2111 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2113 // If this is a signed comparison to 0 and the shift is sign preserving,
2114 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2115 // do that if we're sure to not continue on in this function.
2116 if (isSignTest(Pred, C))
2117 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2120 // NUW guarantees that we are only shifting out zero bits from the high bits,
2121 // so we can LSHR the compare constant without needing a mask and eliminate
2123 if (Shl->hasNoUnsignedWrap()) {
2124 if (Pred == ICmpInst::ICMP_UGT) {
2125 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2126 APInt ShiftedC = C.lshr(*ShiftAmt);
2127 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2129 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2130 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2131 APInt ShiftedC = C.lshr(*ShiftAmt);
2132 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2134 if (Pred == ICmpInst::ICMP_ULT) {
2135 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2136 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2137 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2138 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2139 assert(C.ugt(0) && "ult 0 should have been eliminated");
2140 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2141 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2145 if (Cmp.isEquality() && Shl->hasOneUse()) {
2146 // Strength-reduce the shift into an 'and'.
2147 Constant *Mask = ConstantInt::get(
2149 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2150 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2151 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2152 return new ICmpInst(Pred, And, LShrC);
2155 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2156 bool TrueIfSigned = false;
2157 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2158 // (X << 31) <s 0 --> (X & 1) != 0
2159 Constant *Mask = ConstantInt::get(
2161 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2162 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2163 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2164 And, Constant::getNullValue(ShType));
2167 // Simplify 'shl' inequality test into 'and' equality test.
2168 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2169 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2170 if ((C + 1).isPowerOf2() &&
2171 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2172 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2173 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2174 : ICmpInst::ICMP_NE,
2175 And, Constant::getNullValue(ShType));
2177 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2178 if (C.isPowerOf2() &&
2179 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2181 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2182 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2183 : ICmpInst::ICMP_NE,
2184 And, Constant::getNullValue(ShType));
2188 // Transform (icmp pred iM (shl iM %v, N), C)
2189 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2190 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2191 // This enables us to get rid of the shift in favor of a trunc that may be
2192 // free on the target. It has the additional benefit of comparing to a
2193 // smaller constant that may be more target-friendly.
2194 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2195 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2196 DL.isLegalInteger(TypeBits - Amt)) {
2197 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2198 if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2199 TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
2201 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2202 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2208 /// Fold icmp ({al}shr X, Y), C.
2209 Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2210 BinaryOperator *Shr,
2212 // An exact shr only shifts out zero bits, so:
2213 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2214 Value *X = Shr->getOperand(0);
2215 CmpInst::Predicate Pred = Cmp.getPredicate();
2216 if (Cmp.isEquality() && Shr->isExact() && C.isZero())
2217 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2219 const APInt *ShiftVal;
2220 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2221 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2223 const APInt *ShiftAmt;
2224 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2227 // Check that the shift amount is in range. If not, don't perform undefined
2228 // shifts. When the shift is visited it will be simplified.
2229 unsigned TypeBits = C.getBitWidth();
2230 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2231 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2234 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2235 bool IsExact = Shr->isExact();
2236 Type *ShrTy = Shr->getType();
2237 // TODO: If we could guarantee that InstSimplify would handle all of the
2238 // constant-value-based preconditions in the folds below, then we could assert
2239 // those conditions rather than checking them. This is difficult because of
2240 // undef/poison (PR34838).
2242 if (IsExact || Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) {
2243 // When ShAmtC can be shifted losslessly:
2244 // icmp PRED (ashr exact X, ShAmtC), C --> icmp PRED X, (C << ShAmtC)
2245 // icmp slt/ult (ashr X, ShAmtC), C --> icmp slt/ult X, (C << ShAmtC)
2246 APInt ShiftedC = C.shl(ShAmtVal);
2247 if (ShiftedC.ashr(ShAmtVal) == C)
2248 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2250 if (Pred == CmpInst::ICMP_SGT) {
2251 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2252 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2253 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2254 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2255 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2257 if (Pred == CmpInst::ICMP_UGT) {
2258 // icmp ugt (ashr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2259 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2260 if ((ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2261 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2264 // If the compare constant has significant bits above the lowest sign-bit,
2265 // then convert an unsigned cmp to a test of the sign-bit:
2266 // (ashr X, ShiftC) u> C --> X s< 0
2267 // (ashr X, ShiftC) u< C --> X s> -1
2268 if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2269 if (Pred == CmpInst::ICMP_UGT) {
2270 return new ICmpInst(CmpInst::ICMP_SLT, X,
2271 ConstantInt::getNullValue(ShrTy));
2273 if (Pred == CmpInst::ICMP_ULT) {
2274 return new ICmpInst(CmpInst::ICMP_SGT, X,
2275 ConstantInt::getAllOnesValue(ShrTy));
2279 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2280 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2281 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2282 APInt ShiftedC = C.shl(ShAmtVal);
2283 if (ShiftedC.lshr(ShAmtVal) == C)
2284 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2286 if (Pred == CmpInst::ICMP_UGT) {
2287 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2288 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2289 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2290 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2294 if (!Cmp.isEquality())
2297 // Handle equality comparisons of shift-by-constant.
2299 // If the comparison constant changes with the shift, the comparison cannot
2300 // succeed (bits of the comparison constant cannot match the shifted value).
2301 // This should be known by InstSimplify and already be folded to true/false.
2302 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2303 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2304 "Expected icmp+shr simplify did not occur.");
2306 // If the bits shifted out are known zero, compare the unshifted value:
2307 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2309 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2313 if (Pred == CmpInst::ICMP_EQ)
2314 return new ICmpInst(CmpInst::ICMP_ULT, X,
2315 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2317 return new ICmpInst(CmpInst::ICMP_UGT, X,
2318 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2321 if (Shr->hasOneUse()) {
2322 // Canonicalize the shift into an 'and':
2323 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2324 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2325 Constant *Mask = ConstantInt::get(ShrTy, Val);
2326 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2327 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2333 Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2334 BinaryOperator *SRem,
2336 // Match an 'is positive' or 'is negative' comparison of remainder by a
2337 // constant power-of-2 value:
2338 // (X % pow2C) sgt/slt 0
2339 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2340 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2343 // TODO: The one-use check is standard because we do not typically want to
2344 // create longer instruction sequences, but this might be a special-case
2345 // because srem is not good for analysis or codegen.
2346 if (!SRem->hasOneUse())
2349 const APInt *DivisorC;
2350 if (!C.isZero() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2353 // Mask off the sign bit and the modulo bits (low-bits).
2354 Type *Ty = SRem->getType();
2355 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2356 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2357 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2359 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2360 // bit is set. Example:
2361 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2362 if (Pred == ICmpInst::ICMP_SGT)
2363 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2365 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2366 // bit is set. Example:
2367 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2368 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2371 /// Fold icmp (udiv X, Y), C.
2372 Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2373 BinaryOperator *UDiv,
2376 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2379 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2381 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2382 Value *Y = UDiv->getOperand(1);
2383 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2384 assert(!C.isMaxValue() &&
2385 "icmp ugt X, UINT_MAX should have been simplified already.");
2386 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2387 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2390 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2391 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2392 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2393 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2394 ConstantInt::get(Y->getType(), C2->udiv(C)));
2400 /// Fold icmp ({su}div X, Y), C.
2401 Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2402 BinaryOperator *Div,
2404 // Fold: icmp pred ([us]div X, C2), C -> range test
2405 // Fold this div into the comparison, producing a range check.
2406 // Determine, based on the divide type, what the range is being
2407 // checked. If there is an overflow on the low or high side, remember
2408 // it, otherwise compute the range [low, hi) bounding the new value.
2409 // See: InsertRangeTest above for the kinds of replacements possible.
2411 if (!match(Div->getOperand(1), m_APInt(C2)))
2414 // FIXME: If the operand types don't match the type of the divide
2415 // then don't attempt this transform. The code below doesn't have the
2416 // logic to deal with a signed divide and an unsigned compare (and
2417 // vice versa). This is because (x /s C2) <s C produces different
2418 // results than (x /s C2) <u C or (x /u C2) <s C or even
2419 // (x /u C2) <u C. Simply casting the operands and result won't
2420 // work. :( The if statement below tests that condition and bails
2422 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2423 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2426 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2427 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2428 // division-by-constant cases should be present, we can not assert that they
2429 // have happened before we reach this icmp instruction.
2430 if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes()))
2433 // Compute Prod = C * C2. We are essentially solving an equation of
2434 // form X / C2 = C. We solve for X by multiplying C2 and C.
2435 // By solving for X, we can turn this into a range check instead of computing
2437 APInt Prod = C * *C2;
2439 // Determine if the product overflows by seeing if the product is not equal to
2440 // the divide. Make sure we do the same kind of divide as in the LHS
2441 // instruction that we're folding.
2442 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2444 ICmpInst::Predicate Pred = Cmp.getPredicate();
2446 // If the division is known to be exact, then there is no remainder from the
2447 // divide, so the covered range size is unit, otherwise it is the divisor.
2448 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2450 // Figure out the interval that is being checked. For example, a comparison
2451 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2452 // Compute this interval based on the constants involved and the signedness of
2453 // the compare/divide. This computes a half-open interval, keeping track of
2454 // whether either value in the interval overflows. After analysis each
2455 // overflow variable is set to 0 if it's corresponding bound variable is valid
2456 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2457 int LoOverflow = 0, HiOverflow = 0;
2458 APInt LoBound, HiBound;
2460 if (!DivIsSigned) { // udiv
2461 // e.g. X/5 op 3 --> [15, 20)
2463 HiOverflow = LoOverflow = ProdOV;
2465 // If this is not an exact divide, then many values in the range collapse
2466 // to the same result value.
2467 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2469 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2470 if (C.isZero()) { // (X / pos) op 0
2471 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2472 LoBound = -(RangeSize - 1);
2473 HiBound = RangeSize;
2474 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2475 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2476 HiOverflow = LoOverflow = ProdOV;
2478 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2479 } else { // (X / pos) op neg
2480 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2482 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2484 APInt DivNeg = -RangeSize;
2485 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2488 } else if (C2->isNegative()) { // Divisor is < 0.
2491 if (C.isZero()) { // (X / neg) op 0
2492 // e.g. X/-5 op 0 --> [-4, 5)
2493 LoBound = RangeSize + 1;
2494 HiBound = -RangeSize;
2495 if (HiBound == *C2) { // -INTMIN = INTMIN
2496 HiOverflow = 1; // [INTMIN+1, overflow)
2497 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2499 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2500 // e.g. X/-5 op 3 --> [-19, -14)
2502 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2504 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2505 } else { // (X / neg) op neg
2506 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2507 LoOverflow = HiOverflow = ProdOV;
2509 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2512 // Dividing by a negative swaps the condition. LT <-> GT
2513 Pred = ICmpInst::getSwappedPredicate(Pred);
2516 Value *X = Div->getOperand(0);
2518 default: llvm_unreachable("Unhandled icmp opcode!");
2519 case ICmpInst::ICMP_EQ:
2520 if (LoOverflow && HiOverflow)
2521 return replaceInstUsesWith(Cmp, Builder.getFalse());
2523 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2524 ICmpInst::ICMP_UGE, X,
2525 ConstantInt::get(Div->getType(), LoBound));
2527 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2528 ICmpInst::ICMP_ULT, X,
2529 ConstantInt::get(Div->getType(), HiBound));
2530 return replaceInstUsesWith(
2531 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2532 case ICmpInst::ICMP_NE:
2533 if (LoOverflow && HiOverflow)
2534 return replaceInstUsesWith(Cmp, Builder.getTrue());
2536 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2537 ICmpInst::ICMP_ULT, X,
2538 ConstantInt::get(Div->getType(), LoBound));
2540 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2541 ICmpInst::ICMP_UGE, X,
2542 ConstantInt::get(Div->getType(), HiBound));
2543 return replaceInstUsesWith(Cmp,
2544 insertRangeTest(X, LoBound, HiBound,
2545 DivIsSigned, false));
2546 case ICmpInst::ICMP_ULT:
2547 case ICmpInst::ICMP_SLT:
2548 if (LoOverflow == +1) // Low bound is greater than input range.
2549 return replaceInstUsesWith(Cmp, Builder.getTrue());
2550 if (LoOverflow == -1) // Low bound is less than input range.
2551 return replaceInstUsesWith(Cmp, Builder.getFalse());
2552 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2553 case ICmpInst::ICMP_UGT:
2554 case ICmpInst::ICMP_SGT:
2555 if (HiOverflow == +1) // High bound greater than input range.
2556 return replaceInstUsesWith(Cmp, Builder.getFalse());
2557 if (HiOverflow == -1) // High bound less than input range.
2558 return replaceInstUsesWith(Cmp, Builder.getTrue());
2559 if (Pred == ICmpInst::ICMP_UGT)
2560 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2561 ConstantInt::get(Div->getType(), HiBound));
2562 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2563 ConstantInt::get(Div->getType(), HiBound));
2569 /// Fold icmp (sub X, Y), C.
2570 Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2571 BinaryOperator *Sub,
2573 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2574 ICmpInst::Predicate Pred = Cmp.getPredicate();
2575 Type *Ty = Sub->getType();
2577 // (SubC - Y) == C) --> Y == (SubC - C)
2578 // (SubC - Y) != C) --> Y != (SubC - C)
2580 if (Cmp.isEquality() && match(X, m_ImmConstant(SubC))) {
2581 return new ICmpInst(Pred, Y,
2582 ConstantExpr::getSub(SubC, ConstantInt::get(Ty, C)));
2585 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2588 ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate();
2589 bool HasNSW = Sub->hasNoSignedWrap();
2590 bool HasNUW = Sub->hasNoUnsignedWrap();
2591 if (match(X, m_APInt(C2)) &&
2592 ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) &&
2593 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2594 return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, SubResult));
2596 // The following transforms are only worth it if the only user of the subtract
2598 // TODO: This is an artificial restriction for all of the transforms below
2599 // that only need a single replacement icmp.
2600 if (!Sub->hasOneUse())
2603 // X - Y == 0 --> X == Y.
2604 // X - Y != 0 --> X != Y.
2605 if (Cmp.isEquality() && C.isZero())
2606 return new ICmpInst(Pred, X, Y);
2608 if (Sub->hasNoSignedWrap()) {
2609 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2610 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
2611 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2613 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2614 if (Pred == ICmpInst::ICMP_SGT && C.isZero())
2615 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2617 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2618 if (Pred == ICmpInst::ICMP_SLT && C.isZero())
2619 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2621 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2622 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
2623 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2626 if (!match(X, m_APInt(C2)))
2629 // C2 - Y <u C -> (Y | (C - 1)) == C2
2630 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2631 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2632 (*C2 & (C - 1)) == (C - 1))
2633 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2635 // C2 - Y >u C -> (Y | C) != C2
2636 // iff C2 & C == C and C + 1 is a power of 2
2637 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2638 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2640 // We have handled special cases that reduce.
2641 // Canonicalize any remaining sub to add as:
2642 // (C2 - Y) > C --> (Y + ~C2) < ~C
2643 Value *Add = Builder.CreateAdd(Y, ConstantInt::get(Ty, ~(*C2)), "notsub",
2645 return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, ~C));
2648 /// Fold icmp (add X, Y), C.
2649 Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
2650 BinaryOperator *Add,
2652 Value *Y = Add->getOperand(1);
2654 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2657 // Fold icmp pred (add X, C2), C.
2658 Value *X = Add->getOperand(0);
2659 Type *Ty = Add->getType();
2660 const CmpInst::Predicate Pred = Cmp.getPredicate();
2662 // If the add does not wrap, we can always adjust the compare by subtracting
2663 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2664 // are canonicalized to SGT/SLT/UGT/ULT.
2665 if ((Add->hasNoSignedWrap() &&
2666 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2667 (Add->hasNoUnsignedWrap() &&
2668 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2671 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2672 // If there is overflow, the result must be true or false.
2673 // TODO: Can we assert there is no overflow because InstSimplify always
2674 // handles those cases?
2676 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2677 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2680 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2681 const APInt &Upper = CR.getUpper();
2682 const APInt &Lower = CR.getLower();
2683 if (Cmp.isSigned()) {
2684 if (Lower.isSignMask())
2685 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2686 if (Upper.isSignMask())
2687 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2689 if (Lower.isMinValue())
2690 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2691 if (Upper.isMinValue())
2692 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2695 // This set of folds is intentionally placed after folds that use no-wrapping
2696 // flags because those folds are likely better for later analysis/codegen.
2697 const APInt SMax = APInt::getSignedMaxValue(Ty->getScalarSizeInBits());
2698 const APInt SMin = APInt::getSignedMinValue(Ty->getScalarSizeInBits());
2700 // Fold compare with offset to opposite sign compare if it eliminates offset:
2701 // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
2702 if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
2703 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, -(*C2)));
2705 // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
2706 if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
2707 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, ~(*C2)));
2709 // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
2710 if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
2711 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, SMax - C));
2713 // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
2714 if (Pred == CmpInst::ICMP_SLT && C == *C2)
2715 return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, C ^ SMax));
2717 if (!Add->hasOneUse())
2720 // X+C <u C2 -> (X & -C2) == C
2721 // iff C & (C2-1) == 0
2722 // C2 is a power of 2
2723 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2724 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2725 ConstantExpr::getNeg(cast<Constant>(Y)));
2727 // X+C >u C2 -> (X & ~C2) != C
2729 // C2+1 is a power of 2
2730 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2731 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2732 ConstantExpr::getNeg(cast<Constant>(Y)));
2734 // The range test idiom can use either ult or ugt. Arbitrarily canonicalize
2736 // X+C2 >u C -> X+(C2-C-1) <u ~C
2737 if (Pred == ICmpInst::ICMP_UGT)
2738 return new ICmpInst(ICmpInst::ICMP_ULT,
2739 Builder.CreateAdd(X, ConstantInt::get(Ty, *C2 - C - 1)),
2740 ConstantInt::get(Ty, ~C));
2745 bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2746 Value *&RHS, ConstantInt *&Less,
2747 ConstantInt *&Equal,
2748 ConstantInt *&Greater) {
2749 // TODO: Generalize this to work with other comparison idioms or ensure
2750 // they get canonicalized into this form.
2752 // select i1 (a == b),
2754 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2755 // where Equal, Less and Greater are placeholders for any three constants.
2756 ICmpInst::Predicate PredA;
2757 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2758 !ICmpInst::isEquality(PredA))
2760 Value *EqualVal = SI->getTrueValue();
2761 Value *UnequalVal = SI->getFalseValue();
2762 // We still can get non-canonical predicate here, so canonicalize.
2763 if (PredA == ICmpInst::ICMP_NE)
2764 std::swap(EqualVal, UnequalVal);
2765 if (!match(EqualVal, m_ConstantInt(Equal)))
2767 ICmpInst::Predicate PredB;
2769 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2770 m_ConstantInt(Less), m_ConstantInt(Greater))))
2772 // We can get predicate mismatch here, so canonicalize if possible:
2773 // First, ensure that 'LHS' match.
2775 // x sgt y <--> y slt x
2776 std::swap(LHS2, RHS2);
2777 PredB = ICmpInst::getSwappedPredicate(PredB);
2781 // We also need to canonicalize 'RHS'.
2782 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2783 // x sgt C-1 <--> x sge C <--> not(x slt C)
2784 auto FlippedStrictness =
2785 InstCombiner::getFlippedStrictnessPredicateAndConstant(
2786 PredB, cast<Constant>(RHS2));
2787 if (!FlippedStrictness)
2789 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE &&
2790 "basic correctness failure");
2791 RHS2 = FlippedStrictness->second;
2792 // And kind-of perform the result swap.
2793 std::swap(Less, Greater);
2794 PredB = ICmpInst::ICMP_SLT;
2796 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2799 Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
2803 assert(C && "Cmp RHS should be a constant int!");
2804 // If we're testing a constant value against the result of a three way
2805 // comparison, the result can be expressed directly in terms of the
2806 // original values being compared. Note: We could possibly be more
2807 // aggressive here and remove the hasOneUse test. The original select is
2808 // really likely to simplify or sink when we remove a test of the result.
2809 Value *OrigLHS, *OrigRHS;
2810 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2811 if (Cmp.hasOneUse() &&
2812 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2814 assert(C1LessThan && C2Equal && C3GreaterThan);
2816 bool TrueWhenLessThan =
2817 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2819 bool TrueWhenEqual =
2820 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2822 bool TrueWhenGreaterThan =
2823 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2826 // This generates the new instruction that will replace the original Cmp
2827 // Instruction. Instead of enumerating the various combinations when
2828 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2829 // false, we rely on chaining of ORs and future passes of InstCombine to
2830 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2832 // When none of the three constants satisfy the predicate for the RHS (C),
2833 // the entire original Cmp can be simplified to a false.
2834 Value *Cond = Builder.getFalse();
2835 if (TrueWhenLessThan)
2836 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2839 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2841 if (TrueWhenGreaterThan)
2842 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2845 return replaceInstUsesWith(Cmp, Cond);
2850 Instruction *InstCombinerImpl::foldICmpBitCast(ICmpInst &Cmp) {
2851 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2855 ICmpInst::Predicate Pred = Cmp.getPredicate();
2856 Value *Op1 = Cmp.getOperand(1);
2857 Value *BCSrcOp = Bitcast->getOperand(0);
2859 // Make sure the bitcast doesn't change the number of vector elements.
2860 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2861 Bitcast->getDestTy()->getScalarSizeInBits()) {
2862 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2864 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2865 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2866 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2867 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2868 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2869 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2870 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2871 match(Op1, m_Zero()))
2872 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2874 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2875 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2876 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2878 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2879 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2880 return new ICmpInst(Pred, X,
2881 ConstantInt::getAllOnesValue(X->getType()));
2884 // Zero-equality checks are preserved through unsigned floating-point casts:
2885 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2886 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2887 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2888 if (Cmp.isEquality() && match(Op1, m_Zero()))
2889 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2891 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2892 // the FP extend/truncate because that cast does not change the sign-bit.
2893 // This is true for all standard IEEE-754 types and the X86 80-bit type.
2894 // The sign-bit is always the most significant bit in those types.
2897 if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2898 InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2899 if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2900 match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2901 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2902 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2903 Type *XType = X->getType();
2905 // We can't currently handle Power style floating point operations here.
2906 if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2908 Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2909 if (auto *XVTy = dyn_cast<VectorType>(XType))
2910 NewType = VectorType::get(NewType, XVTy->getElementCount());
2911 Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2913 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2914 ConstantInt::getNullValue(NewType));
2916 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2917 ConstantInt::getAllOnesValue(NewType));
2923 // Test to see if the operands of the icmp are casted versions of other
2924 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2925 if (Bitcast->getType()->isPointerTy() &&
2926 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2927 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2928 // so eliminate it as well.
2929 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2930 Op1 = BC2->getOperand(0);
2932 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2933 return new ICmpInst(Pred, BCSrcOp, Op1);
2937 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2938 !Bitcast->getType()->isIntegerTy() ||
2939 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2942 // If this is checking if all elements of a vector compare are set or not,
2943 // invert the casted vector equality compare and test if all compare
2944 // elements are clear or not. Compare against zero is generally easier for
2945 // analysis and codegen.
2946 // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
2947 // Example: are all elements equal? --> are zero elements not equal?
2948 // TODO: Try harder to reduce compare of 2 freely invertible operands?
2949 if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse() &&
2950 isFreeToInvert(BCSrcOp, BCSrcOp->hasOneUse())) {
2951 Type *ScalarTy = Bitcast->getType();
2952 Value *Cast = Builder.CreateBitCast(Builder.CreateNot(BCSrcOp), ScalarTy);
2953 return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(ScalarTy));
2956 // If this is checking if all elements of an extended vector are clear or not,
2957 // compare in a narrow type to eliminate the extend:
2958 // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
2960 if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() &&
2961 match(BCSrcOp, m_ZExtOrSExt(m_Value(X)))) {
2962 if (auto *VecTy = dyn_cast<FixedVectorType>(X->getType())) {
2963 Type *NewType = Builder.getIntNTy(VecTy->getPrimitiveSizeInBits());
2964 Value *NewCast = Builder.CreateBitCast(X, NewType);
2965 return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(NewType));
2969 // Folding: icmp <pred> iN X, C
2970 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2971 // and C is a splat of a K-bit pattern
2972 // and SC is a constant vector = <C', C', C', ..., C'>
2974 // %E = extractelement <M x iK> %vec, i32 C'
2975 // icmp <pred> iK %E, trunc(C)
2978 if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2979 // Check whether every element of Mask is the same constant
2980 if (is_splat(Mask)) {
2981 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2982 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2983 if (C->isSplat(EltTy->getBitWidth())) {
2984 // Fold the icmp based on the value of C
2985 // If C is M copies of an iK sized bit pattern,
2987 // => %E = extractelement <N x iK> %vec, i32 Elem
2988 // icmp <pred> iK %SplatVal, <pattern>
2989 Value *Elem = Builder.getInt32(Mask[0]);
2990 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2991 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2992 return new ICmpInst(Pred, Extract, NewC);
2999 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3000 /// where X is some kind of instruction.
3001 Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
3003 if (!match(Cmp.getOperand(1), m_APInt(C)))
3006 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
3007 switch (BO->getOpcode()) {
3008 case Instruction::Xor:
3009 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
3012 case Instruction::And:
3013 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
3016 case Instruction::Or:
3017 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
3020 case Instruction::Mul:
3021 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
3024 case Instruction::Shl:
3025 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
3028 case Instruction::LShr:
3029 case Instruction::AShr:
3030 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
3033 case Instruction::SRem:
3034 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
3037 case Instruction::UDiv:
3038 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
3041 case Instruction::SDiv:
3042 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
3045 case Instruction::Sub:
3046 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
3049 case Instruction::Add:
3050 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
3056 // TODO: These folds could be refactored to be part of the above calls.
3057 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
3061 // Match against CmpInst LHS being instructions other than binary operators.
3063 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
3064 // For now, we only support constant integers while folding the
3065 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3066 // similar to the cases handled by binary ops above.
3067 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
3068 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
3072 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
3073 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
3077 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
3078 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
3084 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3085 /// icmp eq/ne BO, C.
3086 Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
3087 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3088 // TODO: Some of these folds could work with arbitrary constants, but this
3089 // function is limited to scalar and vector splat constants.
3090 if (!Cmp.isEquality())
3093 ICmpInst::Predicate Pred = Cmp.getPredicate();
3094 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3095 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
3096 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3098 switch (BO->getOpcode()) {
3099 case Instruction::SRem:
3100 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3101 if (C.isZero() && BO->hasOneUse()) {
3103 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3104 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3105 return new ICmpInst(Pred, NewRem,
3106 Constant::getNullValue(BO->getType()));
3110 case Instruction::Add: {
3111 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3112 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3113 if (BO->hasOneUse())
3114 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3115 } else if (C.isZero()) {
3116 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3117 // efficiently invertible, or if the add has just this one use.
3118 if (Value *NegVal = dyn_castNegVal(BOp1))
3119 return new ICmpInst(Pred, BOp0, NegVal);
3120 if (Value *NegVal = dyn_castNegVal(BOp0))
3121 return new ICmpInst(Pred, NegVal, BOp1);
3122 if (BO->hasOneUse()) {
3123 Value *Neg = Builder.CreateNeg(BOp1);
3125 return new ICmpInst(Pred, BOp0, Neg);
3130 case Instruction::Xor:
3131 if (BO->hasOneUse()) {
3132 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3133 // For the xor case, we can xor two constants together, eliminating
3134 // the explicit xor.
3135 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3136 } else if (C.isZero()) {
3137 // Replace ((xor A, B) != 0) with (A != B)
3138 return new ICmpInst(Pred, BOp0, BOp1);
3142 case Instruction::Or: {
3144 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3145 // Comparing if all bits outside of a constant mask are set?
3146 // Replace (X | C) == -1 with (X & ~C) == ~C.
3147 // This removes the -1 constant.
3148 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3149 Value *And = Builder.CreateAnd(BOp0, NotBOC);
3150 return new ICmpInst(Pred, And, NotBOC);
3154 case Instruction::And: {
3156 if (match(BOp1, m_APInt(BOC))) {
3157 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3158 if (C == *BOC && C.isPowerOf2())
3159 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3160 BO, Constant::getNullValue(RHS->getType()));
3164 case Instruction::UDiv:
3166 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3167 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3168 return new ICmpInst(NewPred, BOp1, BOp0);
3177 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3178 Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3179 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3180 Type *Ty = II->getType();
3181 unsigned BitWidth = C.getBitWidth();
3182 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3184 switch (II->getIntrinsicID()) {
3185 case Intrinsic::abs:
3186 // abs(A) == 0 -> A == 0
3187 // abs(A) == INT_MIN -> A == INT_MIN
3188 if (C.isZero() || C.isMinSignedValue())
3189 return new ICmpInst(Pred, II->getArgOperand(0), ConstantInt::get(Ty, C));
3192 case Intrinsic::bswap:
3193 // bswap(A) == C -> A == bswap(C)
3194 return new ICmpInst(Pred, II->getArgOperand(0),
3195 ConstantInt::get(Ty, C.byteSwap()));
3197 case Intrinsic::ctlz:
3198 case Intrinsic::cttz: {
3199 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3201 return new ICmpInst(Pred, II->getArgOperand(0),
3202 ConstantInt::getNullValue(Ty));
3204 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3205 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3206 // Limit to one use to ensure we don't increase instruction count.
3207 unsigned Num = C.getLimitedValue(BitWidth);
3208 if (Num != BitWidth && II->hasOneUse()) {
3209 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3210 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3211 : APInt::getHighBitsSet(BitWidth, Num + 1);
3212 APInt Mask2 = IsTrailing
3213 ? APInt::getOneBitSet(BitWidth, Num)
3214 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3215 return new ICmpInst(Pred, Builder.CreateAnd(II->getArgOperand(0), Mask1),
3216 ConstantInt::get(Ty, Mask2));
3221 case Intrinsic::ctpop: {
3222 // popcount(A) == 0 -> A == 0 and likewise for !=
3223 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3224 bool IsZero = C.isZero();
3225 if (IsZero || C == BitWidth)
3226 return new ICmpInst(Pred, II->getArgOperand(0),
3227 IsZero ? Constant::getNullValue(Ty)
3228 : Constant::getAllOnesValue(Ty));
3233 case Intrinsic::fshl:
3234 case Intrinsic::fshr:
3235 if (II->getArgOperand(0) == II->getArgOperand(1)) {
3236 // (rot X, ?) == 0/-1 --> X == 0/-1
3237 // TODO: This transform is safe to re-use undef elts in a vector, but
3238 // the constant value passed in by the caller doesn't allow that.
3239 if (C.isZero() || C.isAllOnes())
3240 return new ICmpInst(Pred, II->getArgOperand(0), Cmp.getOperand(1));
3242 const APInt *RotAmtC;
3243 // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3244 // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3245 if (match(II->getArgOperand(2), m_APInt(RotAmtC)))
3246 return new ICmpInst(Pred, II->getArgOperand(0),
3247 II->getIntrinsicID() == Intrinsic::fshl
3248 ? ConstantInt::get(Ty, C.rotr(*RotAmtC))
3249 : ConstantInt::get(Ty, C.rotl(*RotAmtC)));
3253 case Intrinsic::uadd_sat: {
3254 // uadd.sat(a, b) == 0 -> (a | b) == 0
3256 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3257 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3262 case Intrinsic::usub_sat: {
3263 // usub.sat(a, b) == 0 -> a <= b
3265 ICmpInst::Predicate NewPred =
3266 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3267 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3278 /// Fold an icmp with LLVM intrinsics
3279 static Instruction *foldICmpIntrinsicWithIntrinsic(ICmpInst &Cmp) {
3280 assert(Cmp.isEquality());
3282 ICmpInst::Predicate Pred = Cmp.getPredicate();
3283 Value *Op0 = Cmp.getOperand(0);
3284 Value *Op1 = Cmp.getOperand(1);
3285 const auto *IIOp0 = dyn_cast<IntrinsicInst>(Op0);
3286 const auto *IIOp1 = dyn_cast<IntrinsicInst>(Op1);
3287 if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3290 switch (IIOp0->getIntrinsicID()) {
3291 case Intrinsic::bswap:
3292 case Intrinsic::bitreverse:
3293 // If both operands are byte-swapped or bit-reversed, just compare the
3295 return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3296 case Intrinsic::fshl:
3297 case Intrinsic::fshr:
3298 // If both operands are rotated by same amount, just compare the
3300 if (IIOp0->getOperand(0) != IIOp0->getOperand(1))
3302 if (IIOp1->getOperand(0) != IIOp1->getOperand(1))
3304 if (IIOp0->getOperand(2) != IIOp1->getOperand(2))
3306 return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3314 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3315 Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3318 if (Cmp.isEquality())
3319 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3321 Type *Ty = II->getType();
3322 unsigned BitWidth = C.getBitWidth();
3323 ICmpInst::Predicate Pred = Cmp.getPredicate();
3324 switch (II->getIntrinsicID()) {
3325 case Intrinsic::ctpop: {
3326 // (ctpop X > BitWidth - 1) --> X == -1
3327 Value *X = II->getArgOperand(0);
3328 if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
3329 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
3330 ConstantInt::getAllOnesValue(Ty));
3331 // (ctpop X < BitWidth) --> X != -1
3332 if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
3333 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
3334 ConstantInt::getAllOnesValue(Ty));
3337 case Intrinsic::ctlz: {
3338 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3339 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3340 unsigned Num = C.getLimitedValue();
3341 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3342 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3343 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3346 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3347 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3348 unsigned Num = C.getLimitedValue();
3349 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3350 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3351 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3355 case Intrinsic::cttz: {
3356 // Limit to one use to ensure we don't increase instruction count.
3357 if (!II->hasOneUse())
3360 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3361 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3362 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3363 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3364 Builder.CreateAnd(II->getArgOperand(0), Mask),
3365 ConstantInt::getNullValue(Ty));
3368 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3369 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3370 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3371 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3372 Builder.CreateAnd(II->getArgOperand(0), Mask),
3373 ConstantInt::getNullValue(Ty));
3384 /// Handle icmp with constant (but not simple integer constant) RHS.
3385 Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3386 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3387 Constant *RHSC = dyn_cast<Constant>(Op1);
3388 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3392 switch (LHSI->getOpcode()) {
3393 case Instruction::GetElementPtr:
3394 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3395 if (RHSC->isNullValue() &&
3396 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3397 return new ICmpInst(
3398 I.getPredicate(), LHSI->getOperand(0),
3399 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3401 case Instruction::PHI:
3402 // Only fold icmp into the PHI if the phi and icmp are in the same
3403 // block. If in the same block, we're encouraging jump threading. If
3404 // not, we are just pessimizing the code by making an i1 phi.
3405 if (LHSI->getParent() == I.getParent())
3406 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3409 case Instruction::Select: {
3410 // If either operand of the select is a constant, we can fold the
3411 // comparison into the select arms, which will cause one to be
3412 // constant folded and the select turned into a bitwise or.
3413 Value *Op1 = nullptr, *Op2 = nullptr;
3414 ConstantInt *CI = nullptr;
3416 auto SimplifyOp = [&](Value *V) {
3417 Value *Op = nullptr;
3418 if (Constant *C = dyn_cast<Constant>(V)) {
3419 Op = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3420 } else if (RHSC->isNullValue()) {
3421 // If null is being compared, check if it can be further simplified.
3422 Op = SimplifyICmpInst(I.getPredicate(), V, RHSC, SQ);
3426 Op1 = SimplifyOp(LHSI->getOperand(1));
3428 CI = dyn_cast<ConstantInt>(Op1);
3430 Op2 = SimplifyOp(LHSI->getOperand(2));
3432 CI = dyn_cast<ConstantInt>(Op2);
3434 // We only want to perform this transformation if it will not lead to
3435 // additional code. This is true if either both sides of the select
3436 // fold to a constant (in which case the icmp is replaced with a select
3437 // which will usually simplify) or this is the only user of the
3438 // select (in which case we are trading a select+icmp for a simpler
3439 // select+icmp) or all uses of the select can be replaced based on
3440 // dominance information ("Global cases").
3441 bool Transform = false;
3444 else if (Op1 || Op2) {
3446 if (LHSI->hasOneUse())
3449 else if (CI && !CI->isZero())
3450 // When Op1 is constant try replacing select with second operand.
3451 // Otherwise Op2 is constant and try replacing select with first
3454 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3458 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3461 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3463 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3467 case Instruction::IntToPtr:
3468 // icmp pred inttoptr(X), null -> icmp pred X, 0
3469 if (RHSC->isNullValue() &&
3470 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3471 return new ICmpInst(
3472 I.getPredicate(), LHSI->getOperand(0),
3473 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3476 case Instruction::Load:
3477 // Try to optimize things like "A[i] > 4" to index computations.
3478 if (GetElementPtrInst *GEP =
3479 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3480 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3481 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3482 !cast<LoadInst>(LHSI)->isVolatile())
3483 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3492 /// Some comparisons can be simplified.
3493 /// In this case, we are looking for comparisons that look like
3494 /// a check for a lossy truncation.
3496 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3497 /// Where Mask is some pattern that produces all-ones in low bits:
3499 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3501 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3502 /// The Mask can be a constant, too.
3503 /// For some predicates, the operands are commutative.
3504 /// For others, x can only be on a specific side.
3505 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3506 InstCombiner::BuilderTy &Builder) {
3507 ICmpInst::Predicate SrcPred;
3509 auto m_VariableMask = m_CombineOr(
3510 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3511 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3512 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3513 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3514 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3515 if (!match(&I, m_c_ICmp(SrcPred,
3516 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3520 ICmpInst::Predicate DstPred;
3522 case ICmpInst::Predicate::ICMP_EQ:
3523 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3524 DstPred = ICmpInst::Predicate::ICMP_ULE;
3526 case ICmpInst::Predicate::ICMP_NE:
3527 // x & (-1 >> y) != x -> x u> (-1 >> y)
3528 DstPred = ICmpInst::Predicate::ICMP_UGT;
3530 case ICmpInst::Predicate::ICMP_ULT:
3531 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3532 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3533 DstPred = ICmpInst::Predicate::ICMP_UGT;
3535 case ICmpInst::Predicate::ICMP_UGE:
3536 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3537 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3538 DstPred = ICmpInst::Predicate::ICMP_ULE;
3540 case ICmpInst::Predicate::ICMP_SLT:
3541 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3542 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3543 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3545 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3547 DstPred = ICmpInst::Predicate::ICMP_SGT;
3549 case ICmpInst::Predicate::ICMP_SGE:
3550 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3551 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3552 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3554 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3556 DstPred = ICmpInst::Predicate::ICMP_SLE;
3558 case ICmpInst::Predicate::ICMP_SGT:
3559 case ICmpInst::Predicate::ICMP_SLE:
3561 case ICmpInst::Predicate::ICMP_UGT:
3562 case ICmpInst::Predicate::ICMP_ULE:
3563 llvm_unreachable("Instsimplify took care of commut. variant");
3566 llvm_unreachable("All possible folds are handled.");
3569 // The mask value may be a vector constant that has undefined elements. But it
3570 // may not be safe to propagate those undefs into the new compare, so replace
3571 // those elements by copying an existing, defined, and safe scalar constant.
3572 Type *OpTy = M->getType();
3573 auto *VecC = dyn_cast<Constant>(M);
3574 auto *OpVTy = dyn_cast<FixedVectorType>(OpTy);
3575 if (OpVTy && VecC && VecC->containsUndefOrPoisonElement()) {
3576 Constant *SafeReplacementConstant = nullptr;
3577 for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3578 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3579 SafeReplacementConstant = VecC->getAggregateElement(i);
3583 assert(SafeReplacementConstant && "Failed to find undef replacement");
3584 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3587 return Builder.CreateICmp(DstPred, X, M);
3590 /// Some comparisons can be simplified.
3591 /// In this case, we are looking for comparisons that look like
3592 /// a check for a lossy signed truncation.
3593 /// Folds: (MaskedBits is a constant.)
3594 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3596 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3597 /// Where KeptBits = bitwidth(%x) - MaskedBits
3599 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3600 InstCombiner::BuilderTy &Builder) {
3601 ICmpInst::Predicate SrcPred;
3603 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3604 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3605 if (!match(&I, m_c_ICmp(SrcPred,
3606 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3611 // Potential handling of non-splats: for each element:
3612 // * if both are undef, replace with constant 0.
3613 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3614 // * if both are not undef, and are different, bailout.
3615 // * else, only one is undef, then pick the non-undef one.
3617 // The shift amount must be equal.
3620 const APInt &MaskedBits = *C0;
3621 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3623 ICmpInst::Predicate DstPred;
3625 case ICmpInst::Predicate::ICMP_EQ:
3626 // ((%x << MaskedBits) a>> MaskedBits) == %x
3628 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3629 DstPred = ICmpInst::Predicate::ICMP_ULT;
3631 case ICmpInst::Predicate::ICMP_NE:
3632 // ((%x << MaskedBits) a>> MaskedBits) != %x
3634 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3635 DstPred = ICmpInst::Predicate::ICMP_UGE;
3637 // FIXME: are more folds possible?
3642 auto *XType = X->getType();
3643 const unsigned XBitWidth = XType->getScalarSizeInBits();
3644 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3645 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3647 // KeptBits = bitwidth(%x) - MaskedBits
3648 const APInt KeptBits = BitWidth - MaskedBits;
3649 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3650 // ICmpCst = (1 << KeptBits)
3651 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3652 assert(ICmpCst.isPowerOf2());
3653 // AddCst = (1 << (KeptBits-1))
3654 const APInt AddCst = ICmpCst.lshr(1);
3655 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3657 // T0 = add %x, AddCst
3658 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3659 // T1 = T0 DstPred ICmpCst
3660 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3666 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3667 // we should move shifts to the same hand of 'and', i.e. rewrite as
3668 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3669 // We are only interested in opposite logical shifts here.
3670 // One of the shifts can be truncated.
3671 // If we can, we want to end up creating 'lshr' shift.
3673 foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3674 InstCombiner::BuilderTy &Builder) {
3675 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3676 !I.getOperand(0)->hasOneUse())
3679 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3681 // Look for an 'and' of two logical shifts, one of which may be truncated.
3682 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3683 Instruction *XShift, *MaybeTruncation, *YShift;
3686 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3687 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3688 m_AnyLogicalShift, m_Instruction(YShift))),
3689 m_Instruction(MaybeTruncation)))))
3692 // We potentially looked past 'trunc', but only when matching YShift,
3693 // therefore YShift must have the widest type.
3694 Instruction *WidestShift = YShift;
3695 // Therefore XShift must have the shallowest type.
3696 // Or they both have identical types if there was no truncation.
3697 Instruction *NarrowestShift = XShift;
3699 Type *WidestTy = WidestShift->getType();
3700 Type *NarrowestTy = NarrowestShift->getType();
3701 assert(NarrowestTy == I.getOperand(0)->getType() &&
3702 "We did not look past any shifts while matching XShift though.");
3703 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3705 // If YShift is a 'lshr', swap the shifts around.
3706 if (match(YShift, m_LShr(m_Value(), m_Value())))
3707 std::swap(XShift, YShift);
3709 // The shifts must be in opposite directions.
3710 auto XShiftOpcode = XShift->getOpcode();
3711 if (XShiftOpcode == YShift->getOpcode())
3712 return nullptr; // Do not care about same-direction shifts here.
3714 Value *X, *XShAmt, *Y, *YShAmt;
3715 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3716 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3718 // If one of the values being shifted is a constant, then we will end with
3719 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3720 // however, we will need to ensure that we won't increase instruction count.
3721 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3722 // At least one of the hands of the 'and' should be one-use shift.
3723 if (!match(I.getOperand(0),
3724 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3727 // Due to the 'trunc', we will need to widen X. For that either the old
3728 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3729 if (!MaybeTruncation->hasOneUse() &&
3730 !NarrowestShift->getOperand(1)->hasOneUse())
3735 // We have two shift amounts from two different shifts. The types of those
3736 // shift amounts may not match. If that's the case let's bailout now.
3737 if (XShAmt->getType() != YShAmt->getType())
3740 // As input, we have the following pattern:
3741 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3742 // We want to rewrite that as:
3743 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3744 // While we know that originally (Q+K) would not overflow
3745 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3746 // shift amounts. so it may now overflow in smaller bitwidth.
3747 // To ensure that does not happen, we need to ensure that the total maximal
3748 // shift amount is still representable in that smaller bit width.
3749 unsigned MaximalPossibleTotalShiftAmount =
3750 (WidestTy->getScalarSizeInBits() - 1) +
3751 (NarrowestTy->getScalarSizeInBits() - 1);
3752 APInt MaximalRepresentableShiftAmount =
3753 APInt::getAllOnes(XShAmt->getType()->getScalarSizeInBits());
3754 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3757 // Can we fold (XShAmt+YShAmt) ?
3758 auto *NewShAmt = dyn_cast_or_null<Constant>(
3759 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3760 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3763 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3764 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3766 // Is the new shift amount smaller than the bit width?
3767 // FIXME: could also rely on ConstantRange.
3768 if (!match(NewShAmt,
3769 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3770 APInt(WidestBitWidth, WidestBitWidth))))
3773 // An extra legality check is needed if we had trunc-of-lshr.
3774 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3775 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3777 // It isn't obvious whether it's worth it to analyze non-constants here.
3778 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3779 // If *any* of these preconditions matches we can perform the fold.
3780 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3781 ? NewShAmt->getSplatValue()
3783 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3784 if (NewShAmtSplat &&
3785 (NewShAmtSplat->isNullValue() ||
3786 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3788 // We consider *min* leading zeros so a single outlier
3789 // blocks the transform as opposed to allowing it.
3790 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3791 KnownBits Known = computeKnownBits(C, SQ.DL);
3792 unsigned MinLeadZero = Known.countMinLeadingZeros();
3793 // If the value being shifted has at most lowest bit set we can fold.
3794 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3795 if (MaxActiveBits <= 1)
3797 // Precondition: NewShAmt u<= countLeadingZeros(C)
3798 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3801 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3802 KnownBits Known = computeKnownBits(C, SQ.DL);
3803 unsigned MinLeadZero = Known.countMinLeadingZeros();
3804 // If the value being shifted has at most lowest bit set we can fold.
3805 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3806 if (MaxActiveBits <= 1)
3808 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3809 if (NewShAmtSplat) {
3811 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3812 if (AdjNewShAmt.ule(MinLeadZero))
3816 return false; // Can't tell if it's ok.
3822 // All good, we can do this fold.
3823 X = Builder.CreateZExt(X, WidestTy);
3824 Y = Builder.CreateZExt(Y, WidestTy);
3825 // The shift is the same that was for X.
3826 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3827 ? Builder.CreateLShr(X, NewShAmt)
3828 : Builder.CreateShl(X, NewShAmt);
3829 Value *T1 = Builder.CreateAnd(T0, Y);
3830 return Builder.CreateICmp(I.getPredicate(), T1,
3831 Constant::getNullValue(WidestTy));
3836 /// ((x * y) ?/ x) != y
3838 /// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
3839 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3840 /// will mean that we are looking for the opposite answer.
3841 Value *InstCombinerImpl::foldMultiplicationOverflowCheck(ICmpInst &I) {
3842 ICmpInst::Predicate Pred;
3847 // Look for: (-1 u/ x) u</u>= y
3848 if (!I.isEquality() &&
3849 match(&I, m_c_ICmp(Pred,
3850 m_CombineAnd(m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3851 m_Instruction(Div)),
3855 // Are we checking that overflow does not happen, or does happen?
3857 case ICmpInst::Predicate::ICMP_ULT:
3858 NeedNegation = false;
3860 case ICmpInst::Predicate::ICMP_UGE:
3861 NeedNegation = true;
3864 return nullptr; // Wrong predicate.
3866 } else // Look for: ((x * y) / x) !=/== y
3867 if (I.isEquality() &&
3869 m_c_ICmp(Pred, m_Value(Y),
3871 m_OneUse(m_IDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3873 m_Instruction(Mul)),
3875 m_Instruction(Div))))) {
3876 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3880 BuilderTy::InsertPointGuard Guard(Builder);
3881 // If the pattern included (x * y), we'll want to insert new instructions
3882 // right before that original multiplication so that we can replace it.
3883 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3884 if (MulHadOtherUses)
3885 Builder.SetInsertPoint(Mul);
3887 Function *F = Intrinsic::getDeclaration(I.getModule(),
3888 Div->getOpcode() == Instruction::UDiv
3889 ? Intrinsic::umul_with_overflow
3890 : Intrinsic::smul_with_overflow,
3892 CallInst *Call = Builder.CreateCall(F, {X, Y}, "mul");
3894 // If the multiplication was used elsewhere, to ensure that we don't leave
3895 // "duplicate" instructions, replace uses of that original multiplication
3896 // with the multiplication result from the with.overflow intrinsic.
3897 if (MulHadOtherUses)
3898 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "mul.val"));
3900 Value *Res = Builder.CreateExtractValue(Call, 1, "mul.ov");
3901 if (NeedNegation) // This technically increases instruction count.
3902 Res = Builder.CreateNot(Res, "mul.not.ov");
3904 // If we replaced the mul, erase it. Do this after all uses of Builder,
3905 // as the mul is used as insertion point.
3906 if (MulHadOtherUses)
3907 eraseInstFromFunction(*Mul);
3912 static Instruction *foldICmpXNegX(ICmpInst &I) {
3913 CmpInst::Predicate Pred;
3915 if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3918 if (ICmpInst::isSigned(Pred))
3919 Pred = ICmpInst::getSwappedPredicate(Pred);
3920 else if (ICmpInst::isUnsigned(Pred))
3921 Pred = ICmpInst::getSignedPredicate(Pred);
3922 // else for equality-comparisons just keep the predicate.
3924 return ICmpInst::Create(Instruction::ICmp, Pred, X,
3925 Constant::getNullValue(X->getType()), I.getName());
3928 /// Try to fold icmp (binop), X or icmp X, (binop).
3929 /// TODO: A large part of this logic is duplicated in InstSimplify's
3930 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3932 Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
3933 const SimplifyQuery &SQ) {
3934 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3935 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3937 // Special logic for binary operators.
3938 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3939 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3943 if (Instruction *NewICmp = foldICmpXNegX(I))
3946 const CmpInst::Predicate Pred = I.getPredicate();
3949 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3950 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3951 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3952 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3953 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3954 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3955 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3956 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3957 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3960 // Similar to above: an unsigned overflow comparison may use offset + mask:
3961 // ((Op1 + C) & C) u< Op1 --> Op1 != 0
3962 // ((Op1 + C) & C) u>= Op1 --> Op1 == 0
3963 // Op0 u> ((Op0 + C) & C) --> Op0 != 0
3964 // Op0 u<= ((Op0 + C) & C) --> Op0 == 0
3967 if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) &&
3968 match(Op0, m_And(m_BinOp(BO), m_LowBitMask(C))) &&
3969 match(BO, m_Add(m_Specific(Op1), m_SpecificIntAllowUndef(*C)))) {
3970 CmpInst::Predicate NewPred =
3971 Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
3972 Constant *Zero = ConstantInt::getNullValue(Op1->getType());
3973 return new ICmpInst(NewPred, Op1, Zero);
3976 if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
3977 match(Op1, m_And(m_BinOp(BO), m_LowBitMask(C))) &&
3978 match(BO, m_Add(m_Specific(Op0), m_SpecificIntAllowUndef(*C)))) {
3979 CmpInst::Predicate NewPred =
3980 Pred == ICmpInst::ICMP_UGT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
3981 Constant *Zero = ConstantInt::getNullValue(Op1->getType());
3982 return new ICmpInst(NewPred, Op0, Zero);
3986 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3987 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3989 ICmpInst::isEquality(Pred) ||
3990 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3991 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3992 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3994 ICmpInst::isEquality(Pred) ||
3995 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3996 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3998 // Analyze the case when either Op0 or Op1 is an add instruction.
3999 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
4000 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
4001 if (BO0 && BO0->getOpcode() == Instruction::Add) {
4002 A = BO0->getOperand(0);
4003 B = BO0->getOperand(1);
4005 if (BO1 && BO1->getOpcode() == Instruction::Add) {
4006 C = BO1->getOperand(0);
4007 D = BO1->getOperand(1);
4010 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
4011 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
4012 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
4013 return new ICmpInst(Pred, A == Op1 ? B : A,
4014 Constant::getNullValue(Op1->getType()));
4016 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
4017 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
4018 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
4019 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
4022 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
4023 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
4025 // Determine Y and Z in the form icmp (X+Y), (X+Z).
4028 // C + B == C + D -> B == D
4031 } else if (A == D) {
4032 // D + B == C + D -> B == C
4035 } else if (B == C) {
4036 // A + C == C + D -> A == D
4041 // A + D == C + D -> A == C
4045 return new ICmpInst(Pred, Y, Z);
4048 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
4049 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
4050 match(B, m_AllOnes()))
4051 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
4053 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
4054 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
4055 match(B, m_AllOnes()))
4056 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
4058 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
4059 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
4060 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
4062 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
4063 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
4064 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
4066 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
4067 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
4068 match(D, m_AllOnes()))
4069 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
4071 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
4072 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
4073 match(D, m_AllOnes()))
4074 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
4076 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
4077 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
4078 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
4080 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
4081 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
4082 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
4084 // TODO: The subtraction-related identities shown below also hold, but
4085 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
4086 // wouldn't happen even if they were implemented.
4088 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
4089 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
4090 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
4091 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
4093 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
4094 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
4095 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
4097 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
4098 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
4099 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
4101 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
4102 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
4103 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
4105 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
4106 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
4107 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
4109 // if C1 has greater magnitude than C2:
4110 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
4111 // s.t. C3 = C1 - C2
4113 // if C2 has greater magnitude than C1:
4114 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
4115 // s.t. C3 = C2 - C1
4116 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
4117 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
4118 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
4119 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
4120 const APInt &AP1 = C1->getValue();
4121 const APInt &AP2 = C2->getValue();
4122 if (AP1.isNegative() == AP2.isNegative()) {
4123 APInt AP1Abs = C1->getValue().abs();
4124 APInt AP2Abs = C2->getValue().abs();
4125 if (AP1Abs.uge(AP2Abs)) {
4126 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
4127 bool HasNUW = BO0->hasNoUnsignedWrap() && C3->getValue().ule(AP1);
4128 bool HasNSW = BO0->hasNoSignedWrap();
4129 Value *NewAdd = Builder.CreateAdd(A, C3, "", HasNUW, HasNSW);
4130 return new ICmpInst(Pred, NewAdd, C);
4132 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
4133 bool HasNUW = BO1->hasNoUnsignedWrap() && C3->getValue().ule(AP2);
4134 bool HasNSW = BO1->hasNoSignedWrap();
4135 Value *NewAdd = Builder.CreateAdd(C, C3, "", HasNUW, HasNSW);
4136 return new ICmpInst(Pred, A, NewAdd);
4141 // Analyze the case when either Op0 or Op1 is a sub instruction.
4142 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
4147 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
4148 A = BO0->getOperand(0);
4149 B = BO0->getOperand(1);
4151 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
4152 C = BO1->getOperand(0);
4153 D = BO1->getOperand(1);
4156 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
4157 if (A == Op1 && NoOp0WrapProblem)
4158 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
4159 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
4160 if (C == Op0 && NoOp1WrapProblem)
4161 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
4163 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
4164 // (A - B) u>/u<= A --> B u>/u<= A
4165 if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
4166 return new ICmpInst(Pred, B, A);
4167 // C u</u>= (C - D) --> C u</u>= D
4168 if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
4169 return new ICmpInst(Pred, C, D);
4170 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
4171 if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
4172 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4173 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
4174 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
4175 if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
4176 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4177 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
4179 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
4180 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
4181 return new ICmpInst(Pred, A, C);
4183 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
4184 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
4185 return new ICmpInst(Pred, D, B);
4187 // icmp (0-X) < cst --> x > -cst
4188 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
4190 if (match(BO0, m_Neg(m_Value(X))))
4191 if (Constant *RHSC = dyn_cast<Constant>(Op1))
4192 if (RHSC->isNotMinSignedValue())
4193 return new ICmpInst(I.getSwappedPredicate(), X,
4194 ConstantExpr::getNeg(RHSC));
4198 // Try to remove shared constant multiplier from equality comparison:
4199 // X * C == Y * C (with no overflowing/aliasing) --> X == Y
4202 if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
4203 match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
4204 if (!C->countTrailingZeros() ||
4205 (BO0 && BO1 && BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
4206 (BO0 && BO1 && BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
4207 return new ICmpInst(Pred, X, Y);
4210 BinaryOperator *SRem = nullptr;
4211 // icmp (srem X, Y), Y
4212 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
4214 // icmp Y, (srem X, Y)
4215 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
4216 Op0 == BO1->getOperand(1))
4219 // We don't check hasOneUse to avoid increasing register pressure because
4220 // the value we use is the same value this instruction was already using.
4221 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4224 case ICmpInst::ICMP_EQ:
4225 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4226 case ICmpInst::ICMP_NE:
4227 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4228 case ICmpInst::ICMP_SGT:
4229 case ICmpInst::ICMP_SGE:
4230 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4231 Constant::getAllOnesValue(SRem->getType()));
4232 case ICmpInst::ICMP_SLT:
4233 case ICmpInst::ICMP_SLE:
4234 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4235 Constant::getNullValue(SRem->getType()));
4239 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4240 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4241 switch (BO0->getOpcode()) {
4244 case Instruction::Add:
4245 case Instruction::Sub:
4246 case Instruction::Xor: {
4247 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4248 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4251 if (match(BO0->getOperand(1), m_APInt(C))) {
4252 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4253 if (C->isSignMask()) {
4254 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4255 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4258 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4259 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4260 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4261 NewPred = I.getSwappedPredicate(NewPred);
4262 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4267 case Instruction::Mul: {
4268 if (!I.isEquality())
4272 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isZero() &&
4274 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4275 // Mask = -1 >> count-trailing-zeros(C).
4276 if (unsigned TZs = C->countTrailingZeros()) {
4277 Constant *Mask = ConstantInt::get(
4279 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4280 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4281 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4282 return new ICmpInst(Pred, And1, And2);
4287 case Instruction::UDiv:
4288 case Instruction::LShr:
4289 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4291 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4293 case Instruction::SDiv:
4294 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4296 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4298 case Instruction::AShr:
4299 if (!BO0->isExact() || !BO1->isExact())
4301 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4303 case Instruction::Shl: {
4304 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4305 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4308 if (!NSW && I.isSigned())
4310 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4316 // Transform A & (L - 1) `ult` L --> L != 0
4317 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4318 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4320 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4321 auto *Zero = Constant::getNullValue(BO0->getType());
4322 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4326 if (Value *V = foldMultiplicationOverflowCheck(I))
4327 return replaceInstUsesWith(I, V);
4329 if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
4330 return replaceInstUsesWith(I, V);
4332 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
4333 return replaceInstUsesWith(I, V);
4335 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
4336 return replaceInstUsesWith(I, V);
4341 /// Fold icmp Pred min|max(X, Y), X.
4342 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4343 ICmpInst::Predicate Pred = Cmp.getPredicate();
4344 Value *Op0 = Cmp.getOperand(0);
4345 Value *X = Cmp.getOperand(1);
4347 // Canonicalize minimum or maximum operand to LHS of the icmp.
4348 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4349 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4350 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4351 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4353 Pred = Cmp.getSwappedPredicate();
4357 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4358 // smin(X, Y) == X --> X s<= Y
4359 // smin(X, Y) s>= X --> X s<= Y
4360 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4361 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4363 // smin(X, Y) != X --> X s> Y
4364 // smin(X, Y) s< X --> X s> Y
4365 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4366 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4368 // These cases should be handled in InstSimplify:
4369 // smin(X, Y) s<= X --> true
4370 // smin(X, Y) s> X --> false
4374 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4375 // smax(X, Y) == X --> X s>= Y
4376 // smax(X, Y) s<= X --> X s>= Y
4377 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4378 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4380 // smax(X, Y) != X --> X s< Y
4381 // smax(X, Y) s> X --> X s< Y
4382 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4383 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4385 // These cases should be handled in InstSimplify:
4386 // smax(X, Y) s>= X --> true
4387 // smax(X, Y) s< X --> false
4391 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4392 // umin(X, Y) == X --> X u<= Y
4393 // umin(X, Y) u>= X --> X u<= Y
4394 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4395 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4397 // umin(X, Y) != X --> X u> Y
4398 // umin(X, Y) u< X --> X u> Y
4399 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4400 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4402 // These cases should be handled in InstSimplify:
4403 // umin(X, Y) u<= X --> true
4404 // umin(X, Y) u> X --> false
4408 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4409 // umax(X, Y) == X --> X u>= Y
4410 // umax(X, Y) u<= X --> X u>= Y
4411 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4412 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4414 // umax(X, Y) != X --> X u< Y
4415 // umax(X, Y) u> X --> X u< Y
4416 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4417 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4419 // These cases should be handled in InstSimplify:
4420 // umax(X, Y) u>= X --> true
4421 // umax(X, Y) u< X --> false
4428 Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
4429 if (!I.isEquality())
4432 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4433 const CmpInst::Predicate Pred = I.getPredicate();
4434 Value *A, *B, *C, *D;
4435 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4436 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4437 Value *OtherVal = A == Op1 ? B : A;
4438 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4441 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4442 // A^c1 == C^c2 --> A == C^(c1^c2)
4443 ConstantInt *C1, *C2;
4444 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4446 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4447 Value *Xor = Builder.CreateXor(C, NC);
4448 return new ICmpInst(Pred, A, Xor);
4451 // A^B == A^D -> B == D
4453 return new ICmpInst(Pred, B, D);
4455 return new ICmpInst(Pred, B, C);
4457 return new ICmpInst(Pred, A, D);
4459 return new ICmpInst(Pred, A, C);
4463 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4464 // A == (A^B) -> B == 0
4465 Value *OtherVal = A == Op0 ? B : A;
4466 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4469 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4470 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4471 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4472 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4478 } else if (A == D) {
4482 } else if (B == C) {
4486 } else if (B == D) {
4492 if (X) { // Build (X^Y) & Z
4493 Op1 = Builder.CreateXor(X, Y);
4494 Op1 = Builder.CreateAnd(Op1, Z);
4495 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4500 // Similar to above, but specialized for constant because invert is needed:
4501 // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
4504 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_Constant(C)))) &&
4505 match(Op1, m_OneUse(m_Or(m_Value(Y), m_Specific(C))))) {
4506 Value *Xor = Builder.CreateXor(X, Y);
4507 Value *And = Builder.CreateAnd(Xor, ConstantExpr::getNot(C));
4508 return new ICmpInst(Pred, And, Constant::getNullValue(And->getType()));
4512 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4513 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4515 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4516 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4517 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4518 match(Op1, m_ZExt(m_Value(A))))) {
4519 APInt Pow2 = Cst1->getValue() + 1;
4520 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4521 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4522 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4525 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4526 // For lshr and ashr pairs.
4527 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4528 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4529 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4530 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4531 unsigned TypeBits = Cst1->getBitWidth();
4532 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4533 if (ShAmt < TypeBits && ShAmt != 0) {
4534 ICmpInst::Predicate NewPred =
4535 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4536 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4537 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4538 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4542 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4543 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4544 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4545 unsigned TypeBits = Cst1->getBitWidth();
4546 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4547 if (ShAmt < TypeBits && ShAmt != 0) {
4548 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4549 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4550 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4551 I.getName() + ".mask");
4552 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4556 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4557 // "icmp (and X, mask), cst"
4559 if (Op0->hasOneUse() &&
4560 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4561 match(Op1, m_ConstantInt(Cst1)) &&
4562 // Only do this when A has multiple uses. This is most important to do
4563 // when it exposes other optimizations.
4565 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4567 if (ShAmt < ASize) {
4569 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4572 APInt CmpV = Cst1->getValue().zext(ASize);
4575 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4576 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4580 if (Instruction *ICmp = foldICmpIntrinsicWithIntrinsic(I))
4583 // Canonicalize checking for a power-of-2-or-zero value:
4584 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4585 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4586 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4588 !match(Op1, m_ZeroInt()))
4591 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4592 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4593 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4596 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4600 Type *Ty = A->getType();
4601 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4602 return Pred == ICmpInst::ICMP_EQ
4603 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4604 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4607 // Match icmp eq (trunc (lshr A, BW), (ashr (trunc A), BW-1)), which checks the
4608 // top BW/2 + 1 bits are all the same. Create "A >=s INT_MIN && A <=s INT_MAX",
4609 // which we generate as "icmp ult (add A, 2^(BW-1)), 2^BW" to skip a few steps
4611 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
4612 if (match(Op0, m_AShr(m_Trunc(m_Value(A)), m_SpecificInt(BitWidth - 1))) &&
4613 match(Op1, m_Trunc(m_LShr(m_Specific(A), m_SpecificInt(BitWidth)))) &&
4614 A->getType()->getScalarSizeInBits() == BitWidth * 2 &&
4615 (I.getOperand(0)->hasOneUse() || I.getOperand(1)->hasOneUse())) {
4616 APInt C = APInt::getOneBitSet(BitWidth * 2, BitWidth - 1);
4617 Value *Add = Builder.CreateAdd(A, ConstantInt::get(A->getType(), C));
4618 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT
4619 : ICmpInst::ICMP_UGE,
4620 Add, ConstantInt::get(A->getType(), C.shl(1)));
4626 static Instruction *foldICmpWithTrunc(ICmpInst &ICmp,
4627 InstCombiner::BuilderTy &Builder) {
4628 ICmpInst::Predicate Pred = ICmp.getPredicate();
4629 Value *Op0 = ICmp.getOperand(0), *Op1 = ICmp.getOperand(1);
4631 // Try to canonicalize trunc + compare-to-constant into a mask + cmp.
4632 // The trunc masks high bits while the compare may effectively mask low bits.
4635 if (!match(Op0, m_OneUse(m_Trunc(m_Value(X)))) || !match(Op1, m_APInt(C)))
4638 // This matches patterns corresponding to tests of the signbit as well as:
4639 // (trunc X) u< C --> (X & -C) == 0 (are all masked-high-bits clear?)
4640 // (trunc X) u> C --> (X & ~C) != 0 (are any masked-high-bits set?)
4642 if (decomposeBitTestICmp(Op0, Op1, Pred, X, Mask, true /* WithTrunc */)) {
4643 Value *And = Builder.CreateAnd(X, Mask);
4644 Constant *Zero = ConstantInt::getNullValue(X->getType());
4645 return new ICmpInst(Pred, And, Zero);
4648 unsigned SrcBits = X->getType()->getScalarSizeInBits();
4649 if (Pred == ICmpInst::ICMP_ULT && C->isNegatedPowerOf2()) {
4650 // If C is a negative power-of-2 (high-bit mask):
4651 // (trunc X) u< C --> (X & C) != C (are any masked-high-bits clear?)
4652 Constant *MaskC = ConstantInt::get(X->getType(), C->zext(SrcBits));
4653 Value *And = Builder.CreateAnd(X, MaskC);
4654 return new ICmpInst(ICmpInst::ICMP_NE, And, MaskC);
4657 if (Pred == ICmpInst::ICMP_UGT && (~*C).isPowerOf2()) {
4658 // If C is not-of-power-of-2 (one clear bit):
4659 // (trunc X) u> C --> (X & (C+1)) == C+1 (are all masked-high-bits set?)
4660 Constant *MaskC = ConstantInt::get(X->getType(), (*C + 1).zext(SrcBits));
4661 Value *And = Builder.CreateAnd(X, MaskC);
4662 return new ICmpInst(ICmpInst::ICMP_EQ, And, MaskC);
4668 static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4669 InstCombiner::BuilderTy &Builder) {
4670 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
4671 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4673 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4676 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4677 bool IsSignedCmp = ICmp.isSigned();
4678 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4679 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4680 // and the other is a zext), then we can't handle this.
4681 // TODO: This is too strict. We can handle some predicates (equality?).
4682 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4685 // Not an extension from the same type?
4686 Value *Y = CastOp1->getOperand(0);
4687 Type *XTy = X->getType(), *YTy = Y->getType();
4689 // One of the casts must have one use because we are creating a new cast.
4690 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4692 // Extend the narrower operand to the type of the wider operand.
4693 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4694 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4695 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4696 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4701 // (zext X) == (zext Y) --> X == Y
4702 // (sext X) == (sext Y) --> X == Y
4703 if (ICmp.isEquality())
4704 return new ICmpInst(ICmp.getPredicate(), X, Y);
4706 // A signed comparison of sign extended values simplifies into a
4707 // signed comparison.
4708 if (IsSignedCmp && IsSignedExt)
4709 return new ICmpInst(ICmp.getPredicate(), X, Y);
4711 // The other three cases all fold into an unsigned comparison.
4712 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4715 // Below here, we are only folding a compare with constant.
4716 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4720 // Compute the constant that would happen if we truncated to SrcTy then
4721 // re-extended to DestTy.
4722 Type *SrcTy = CastOp0->getSrcTy();
4723 Type *DestTy = CastOp0->getDestTy();
4724 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4725 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4727 // If the re-extended constant didn't change...
4729 if (ICmp.isEquality())
4730 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4732 // A signed comparison of sign extended values simplifies into a
4733 // signed comparison.
4734 if (IsSignedExt && IsSignedCmp)
4735 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4737 // The other three cases all fold into an unsigned comparison.
4738 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4741 // The re-extended constant changed, partly changed (in the case of a vector),
4742 // or could not be determined to be equal (in the case of a constant
4743 // expression), so the constant cannot be represented in the shorter type.
4744 // All the cases that fold to true or false will have already been handled
4745 // by SimplifyICmpInst, so only deal with the tricky case.
4746 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4749 // Is source op positive?
4750 // icmp ult (sext X), C --> icmp sgt X, -1
4751 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4752 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4754 // Is source op negative?
4755 // icmp ugt (sext X), C --> icmp slt X, 0
4756 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4757 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4760 /// Handle icmp (cast x), (cast or constant).
4761 Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
4762 // If any operand of ICmp is a inttoptr roundtrip cast then remove it as
4763 // icmp compares only pointer's value.
4764 // icmp (inttoptr (ptrtoint p1)), p2 --> icmp p1, p2.
4765 Value *SimplifiedOp0 = simplifyIntToPtrRoundTripCast(ICmp.getOperand(0));
4766 Value *SimplifiedOp1 = simplifyIntToPtrRoundTripCast(ICmp.getOperand(1));
4767 if (SimplifiedOp0 || SimplifiedOp1)
4768 return new ICmpInst(ICmp.getPredicate(),
4769 SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(0),
4770 SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(1));
4772 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4775 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4778 Value *Op0Src = CastOp0->getOperand(0);
4779 Type *SrcTy = CastOp0->getSrcTy();
4780 Type *DestTy = CastOp0->getDestTy();
4782 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4783 // integer type is the same size as the pointer type.
4784 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4785 if (isa<VectorType>(SrcTy)) {
4786 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4787 DestTy = cast<VectorType>(DestTy)->getElementType();
4789 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4791 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4792 CompatibleSizes(SrcTy, DestTy)) {
4793 Value *NewOp1 = nullptr;
4794 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4795 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4796 if (PtrSrc->getType()->getPointerAddressSpace() ==
4797 Op0Src->getType()->getPointerAddressSpace()) {
4798 NewOp1 = PtrToIntOp1->getOperand(0);
4799 // If the pointer types don't match, insert a bitcast.
4800 if (Op0Src->getType() != NewOp1->getType())
4801 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4803 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4804 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4808 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4811 if (Instruction *R = foldICmpWithTrunc(ICmp, Builder))
4814 return foldICmpWithZextOrSext(ICmp, Builder);
4817 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4820 llvm_unreachable("Unsupported binary op");
4821 case Instruction::Add:
4822 case Instruction::Sub:
4823 return match(RHS, m_Zero());
4824 case Instruction::Mul:
4825 return match(RHS, m_One());
4830 InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
4831 bool IsSigned, Value *LHS, Value *RHS,
4832 Instruction *CxtI) const {
4835 llvm_unreachable("Unsupported binary op");
4836 case Instruction::Add:
4838 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4840 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4841 case Instruction::Sub:
4843 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4845 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4846 case Instruction::Mul:
4848 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4850 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4854 bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
4855 bool IsSigned, Value *LHS,
4856 Value *RHS, Instruction &OrigI,
4858 Constant *&Overflow) {
4859 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4860 std::swap(LHS, RHS);
4862 // If the overflow check was an add followed by a compare, the insertion point
4863 // may be pointing to the compare. We want to insert the new instructions
4864 // before the add in case there are uses of the add between the add and the
4866 Builder.SetInsertPoint(&OrigI);
4868 Type *OverflowTy = Type::getInt1Ty(LHS->getContext());
4869 if (auto *LHSTy = dyn_cast<VectorType>(LHS->getType()))
4870 OverflowTy = VectorType::get(OverflowTy, LHSTy->getElementCount());
4872 if (isNeutralValue(BinaryOp, RHS)) {
4874 Overflow = ConstantInt::getFalse(OverflowTy);
4878 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4879 case OverflowResult::MayOverflow:
4881 case OverflowResult::AlwaysOverflowsLow:
4882 case OverflowResult::AlwaysOverflowsHigh:
4883 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4884 Result->takeName(&OrigI);
4885 Overflow = ConstantInt::getTrue(OverflowTy);
4887 case OverflowResult::NeverOverflows:
4888 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4889 Result->takeName(&OrigI);
4890 Overflow = ConstantInt::getFalse(OverflowTy);
4891 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4893 Inst->setHasNoSignedWrap();
4895 Inst->setHasNoUnsignedWrap();
4900 llvm_unreachable("Unexpected overflow result");
4903 /// Recognize and process idiom involving test for multiplication
4906 /// The caller has matched a pattern of the form:
4907 /// I = cmp u (mul(zext A, zext B), V
4908 /// The function checks if this is a test for overflow and if so replaces
4909 /// multiplication with call to 'mul.with.overflow' intrinsic.
4911 /// \param I Compare instruction.
4912 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4913 /// the compare instruction. Must be of integer type.
4914 /// \param OtherVal The other argument of compare instruction.
4915 /// \returns Instruction which must replace the compare instruction, NULL if no
4916 /// replacement required.
4917 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4919 InstCombinerImpl &IC) {
4920 // Don't bother doing this transformation for pointers, don't do it for
4922 if (!isa<IntegerType>(MulVal->getType()))
4925 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4926 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4927 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4930 assert(MulInstr->getOpcode() == Instruction::Mul);
4932 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4933 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4934 assert(LHS->getOpcode() == Instruction::ZExt);
4935 assert(RHS->getOpcode() == Instruction::ZExt);
4936 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4938 // Calculate type and width of the result produced by mul.with.overflow.
4939 Type *TyA = A->getType(), *TyB = B->getType();
4940 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4941 WidthB = TyB->getPrimitiveSizeInBits();
4944 if (WidthB > WidthA) {
4952 // In order to replace the original mul with a narrower mul.with.overflow,
4953 // all uses must ignore upper bits of the product. The number of used low
4954 // bits must be not greater than the width of mul.with.overflow.
4955 if (MulVal->hasNUsesOrMore(2))
4956 for (User *U : MulVal->users()) {
4959 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4960 // Check if truncation ignores bits above MulWidth.
4961 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4962 if (TruncWidth > MulWidth)
4964 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4965 // Check if AND ignores bits above MulWidth.
4966 if (BO->getOpcode() != Instruction::And)
4968 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4969 const APInt &CVal = CI->getValue();
4970 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4973 // In this case we could have the operand of the binary operation
4974 // being defined in another block, and performing the replacement
4975 // could break the dominance relation.
4979 // Other uses prohibit this transformation.
4984 // Recognize patterns
4985 switch (I.getPredicate()) {
4986 case ICmpInst::ICMP_EQ:
4987 case ICmpInst::ICMP_NE:
4988 // Recognize pattern:
4989 // mulval = mul(zext A, zext B)
4990 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4993 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4994 if (ValToMask != MulVal)
4996 const APInt &CVal = CI->getValue() + 1;
4997 if (CVal.isPowerOf2()) {
4998 unsigned MaskWidth = CVal.logBase2();
4999 if (MaskWidth == MulWidth)
5000 break; // Recognized
5005 case ICmpInst::ICMP_UGT:
5006 // Recognize pattern:
5007 // mulval = mul(zext A, zext B)
5008 // cmp ugt mulval, max
5009 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
5010 APInt MaxVal = APInt::getMaxValue(MulWidth);
5011 MaxVal = MaxVal.zext(CI->getBitWidth());
5012 if (MaxVal.eq(CI->getValue()))
5013 break; // Recognized
5017 case ICmpInst::ICMP_UGE:
5018 // Recognize pattern:
5019 // mulval = mul(zext A, zext B)
5020 // cmp uge mulval, max+1
5021 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
5022 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
5023 if (MaxVal.eq(CI->getValue()))
5024 break; // Recognized
5028 case ICmpInst::ICMP_ULE:
5029 // Recognize pattern:
5030 // mulval = mul(zext A, zext B)
5031 // cmp ule mulval, max
5032 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
5033 APInt MaxVal = APInt::getMaxValue(MulWidth);
5034 MaxVal = MaxVal.zext(CI->getBitWidth());
5035 if (MaxVal.eq(CI->getValue()))
5036 break; // Recognized
5040 case ICmpInst::ICMP_ULT:
5041 // Recognize pattern:
5042 // mulval = mul(zext A, zext B)
5043 // cmp ule mulval, max + 1
5044 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
5045 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
5046 if (MaxVal.eq(CI->getValue()))
5047 break; // Recognized
5055 InstCombiner::BuilderTy &Builder = IC.Builder;
5056 Builder.SetInsertPoint(MulInstr);
5058 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
5059 Value *MulA = A, *MulB = B;
5060 if (WidthA < MulWidth)
5061 MulA = Builder.CreateZExt(A, MulType);
5062 if (WidthB < MulWidth)
5063 MulB = Builder.CreateZExt(B, MulType);
5064 Function *F = Intrinsic::getDeclaration(
5065 I.getModule(), Intrinsic::umul_with_overflow, MulType);
5066 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
5067 IC.addToWorklist(MulInstr);
5069 // If there are uses of mul result other than the comparison, we know that
5070 // they are truncation or binary AND. Change them to use result of
5071 // mul.with.overflow and adjust properly mask/size.
5072 if (MulVal->hasNUsesOrMore(2)) {
5073 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
5074 for (User *U : make_early_inc_range(MulVal->users())) {
5075 if (U == &I || U == OtherVal)
5077 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
5078 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
5079 IC.replaceInstUsesWith(*TI, Mul);
5081 TI->setOperand(0, Mul);
5082 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
5083 assert(BO->getOpcode() == Instruction::And);
5084 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
5085 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
5086 APInt ShortMask = CI->getValue().trunc(MulWidth);
5087 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
5088 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
5089 IC.replaceInstUsesWith(*BO, Zext);
5091 llvm_unreachable("Unexpected Binary operation");
5093 IC.addToWorklist(cast<Instruction>(U));
5096 if (isa<Instruction>(OtherVal))
5097 IC.addToWorklist(cast<Instruction>(OtherVal));
5099 // The original icmp gets replaced with the overflow value, maybe inverted
5100 // depending on predicate.
5101 bool Inverse = false;
5102 switch (I.getPredicate()) {
5103 case ICmpInst::ICMP_NE:
5105 case ICmpInst::ICMP_EQ:
5108 case ICmpInst::ICMP_UGT:
5109 case ICmpInst::ICMP_UGE:
5110 if (I.getOperand(0) == MulVal)
5114 case ICmpInst::ICMP_ULT:
5115 case ICmpInst::ICMP_ULE:
5116 if (I.getOperand(1) == MulVal)
5121 llvm_unreachable("Unexpected predicate");
5124 Value *Res = Builder.CreateExtractValue(Call, 1);
5125 return BinaryOperator::CreateNot(Res);
5128 return ExtractValueInst::Create(Call, 1);
5131 /// When performing a comparison against a constant, it is possible that not all
5132 /// the bits in the LHS are demanded. This helper method computes the mask that
5134 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
5136 if (!match(I.getOperand(1), m_APInt(RHS)))
5137 return APInt::getAllOnes(BitWidth);
5139 // If this is a normal comparison, it demands all bits. If it is a sign bit
5140 // comparison, it only demands the sign bit.
5142 if (InstCombiner::isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
5143 return APInt::getSignMask(BitWidth);
5145 switch (I.getPredicate()) {
5146 // For a UGT comparison, we don't care about any bits that
5147 // correspond to the trailing ones of the comparand. The value of these
5148 // bits doesn't impact the outcome of the comparison, because any value
5149 // greater than the RHS must differ in a bit higher than these due to carry.
5150 case ICmpInst::ICMP_UGT:
5151 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
5153 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
5154 // Any value less than the RHS must differ in a higher bit because of carries.
5155 case ICmpInst::ICMP_ULT:
5156 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
5159 return APInt::getAllOnes(BitWidth);
5163 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
5164 /// should be swapped.
5165 /// The decision is based on how many times these two operands are reused
5166 /// as subtract operands and their positions in those instructions.
5167 /// The rationale is that several architectures use the same instruction for
5168 /// both subtract and cmp. Thus, it is better if the order of those operands
5170 /// \return true if Op0 and Op1 should be swapped.
5171 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
5172 // Filter out pointer values as those cannot appear directly in subtract.
5173 // FIXME: we may want to go through inttoptrs or bitcasts.
5174 if (Op0->getType()->isPointerTy())
5176 // If a subtract already has the same operands as a compare, swapping would be
5177 // bad. If a subtract has the same operands as a compare but in reverse order,
5178 // then swapping is good.
5180 for (const User *U : Op0->users()) {
5181 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
5183 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
5186 return GoodToSwap > 0;
5189 /// Check that one use is in the same block as the definition and all
5190 /// other uses are in blocks dominated by a given block.
5192 /// \param DI Definition
5194 /// \param DB Block that must dominate all uses of \p DI outside
5195 /// the parent block
5196 /// \return true when \p UI is the only use of \p DI in the parent block
5197 /// and all other uses of \p DI are in blocks dominated by \p DB.
5199 bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
5200 const Instruction *UI,
5201 const BasicBlock *DB) const {
5202 assert(DI && UI && "Instruction not defined\n");
5203 // Ignore incomplete definitions.
5204 if (!DI->getParent())
5206 // DI and UI must be in the same block.
5207 if (DI->getParent() != UI->getParent())
5209 // Protect from self-referencing blocks.
5210 if (DI->getParent() == DB)
5212 for (const User *U : DI->users()) {
5213 auto *Usr = cast<Instruction>(U);
5214 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
5220 /// Return true when the instruction sequence within a block is select-cmp-br.
5221 static bool isChainSelectCmpBranch(const SelectInst *SI) {
5222 const BasicBlock *BB = SI->getParent();
5225 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
5226 if (!BI || BI->getNumSuccessors() != 2)
5228 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
5229 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
5234 /// True when a select result is replaced by one of its operands
5235 /// in select-icmp sequence. This will eventually result in the elimination
5238 /// \param SI Select instruction
5239 /// \param Icmp Compare instruction
5240 /// \param SIOpd Operand that replaces the select
5243 /// - The replacement is global and requires dominator information
5244 /// - The caller is responsible for the actual replacement
5249 /// %4 = select i1 %3, %C* %0, %C* null
5250 /// %5 = icmp eq %C* %4, null
5251 /// br i1 %5, label %9, label %7
5253 /// ; <label>:7 ; preds = %entry
5254 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
5257 /// can be transformed to
5259 /// %5 = icmp eq %C* %0, null
5260 /// %6 = select i1 %3, i1 %5, i1 true
5261 /// br i1 %6, label %9, label %7
5263 /// ; <label>:7 ; preds = %entry
5264 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
5266 /// Similar when the first operand of the select is a constant or/and
5267 /// the compare is for not equal rather than equal.
5269 /// NOTE: The function is only called when the select and compare constants
5270 /// are equal, the optimization can work only for EQ predicates. This is not a
5271 /// major restriction since a NE compare should be 'normalized' to an equal
5272 /// compare, which usually happens in the combiner and test case
5273 /// select-cmp-br.ll checks for it.
5274 bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
5275 const ICmpInst *Icmp,
5276 const unsigned SIOpd) {
5277 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
5278 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
5279 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
5280 // The check for the single predecessor is not the best that can be
5281 // done. But it protects efficiently against cases like when SI's
5282 // home block has two successors, Succ and Succ1, and Succ1 predecessor
5283 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
5284 // replaced can be reached on either path. So the uniqueness check
5285 // guarantees that the path all uses of SI (outside SI's parent) are on
5286 // is disjoint from all other paths out of SI. But that information
5287 // is more expensive to compute, and the trade-off here is in favor
5288 // of compile-time. It should also be noticed that we check for a single
5289 // predecessor and not only uniqueness. This to handle the situation when
5290 // Succ and Succ1 points to the same basic block.
5291 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
5293 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
5300 /// Try to fold the comparison based on range information we can get by checking
5301 /// whether bits are known to be zero or one in the inputs.
5302 Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
5303 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5304 Type *Ty = Op0->getType();
5305 ICmpInst::Predicate Pred = I.getPredicate();
5307 // Get scalar or pointer size.
5308 unsigned BitWidth = Ty->isIntOrIntVectorTy()
5309 ? Ty->getScalarSizeInBits()
5310 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
5315 KnownBits Op0Known(BitWidth);
5316 KnownBits Op1Known(BitWidth);
5318 if (SimplifyDemandedBits(&I, 0,
5319 getDemandedBitsLHSMask(I, BitWidth),
5323 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnes(BitWidth), Op1Known, 0))
5326 // Given the known and unknown bits, compute a range that the LHS could be
5327 // in. Compute the Min, Max and RHS values based on the known bits. For the
5328 // EQ and NE we use unsigned values.
5329 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
5330 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
5332 Op0Min = Op0Known.getSignedMinValue();
5333 Op0Max = Op0Known.getSignedMaxValue();
5334 Op1Min = Op1Known.getSignedMinValue();
5335 Op1Max = Op1Known.getSignedMaxValue();
5337 Op0Min = Op0Known.getMinValue();
5338 Op0Max = Op0Known.getMaxValue();
5339 Op1Min = Op1Known.getMinValue();
5340 Op1Max = Op1Known.getMaxValue();
5343 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
5344 // out that the LHS or RHS is a constant. Constant fold this now, so that
5345 // code below can assume that Min != Max.
5346 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
5347 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
5348 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
5349 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
5351 // Don't break up a clamp pattern -- (min(max X, Y), Z) -- by replacing a
5352 // min/max canonical compare with some other compare. That could lead to
5353 // conflict with select canonicalization and infinite looping.
5354 // FIXME: This constraint may go away if min/max intrinsics are canonical.
5355 auto isMinMaxCmp = [&](Instruction &Cmp) {
5356 if (!Cmp.hasOneUse())
5359 SelectPatternFlavor SPF = matchSelectPattern(Cmp.user_back(), A, B).Flavor;
5360 if (!SelectPatternResult::isMinOrMax(SPF))
5362 return match(Op0, m_MaxOrMin(m_Value(), m_Value())) ||
5363 match(Op1, m_MaxOrMin(m_Value(), m_Value()));
5365 if (!isMinMaxCmp(I)) {
5369 case ICmpInst::ICMP_ULT: {
5370 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5371 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5373 if (match(Op1, m_APInt(CmpC))) {
5374 // A <u C -> A == C-1 if min(A)+1 == C
5375 if (*CmpC == Op0Min + 1)
5376 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5377 ConstantInt::get(Op1->getType(), *CmpC - 1));
5378 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5379 // exceeds the log2 of C.
5380 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5381 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5382 Constant::getNullValue(Op1->getType()));
5386 case ICmpInst::ICMP_UGT: {
5387 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5388 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5390 if (match(Op1, m_APInt(CmpC))) {
5391 // A >u C -> A == C+1 if max(a)-1 == C
5392 if (*CmpC == Op0Max - 1)
5393 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5394 ConstantInt::get(Op1->getType(), *CmpC + 1));
5395 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5396 // exceeds the log2 of C.
5397 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5398 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5399 Constant::getNullValue(Op1->getType()));
5403 case ICmpInst::ICMP_SLT: {
5404 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5405 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5407 if (match(Op1, m_APInt(CmpC))) {
5408 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5409 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5410 ConstantInt::get(Op1->getType(), *CmpC - 1));
5414 case ICmpInst::ICMP_SGT: {
5415 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5416 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5418 if (match(Op1, m_APInt(CmpC))) {
5419 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5420 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5421 ConstantInt::get(Op1->getType(), *CmpC + 1));
5428 // Based on the range information we know about the LHS, see if we can
5429 // simplify this comparison. For example, (x&4) < 8 is always true.
5432 llvm_unreachable("Unknown icmp opcode!");
5433 case ICmpInst::ICMP_EQ:
5434 case ICmpInst::ICMP_NE: {
5435 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
5436 return replaceInstUsesWith(
5437 I, ConstantInt::getBool(I.getType(), Pred == CmpInst::ICMP_NE));
5439 // If all bits are known zero except for one, then we know at most one bit
5440 // is set. If the comparison is against zero, then this is a check to see if
5441 // *that* bit is set.
5442 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
5443 if (Op1Known.isZero()) {
5444 // If the LHS is an AND with the same constant, look through it.
5445 Value *LHS = nullptr;
5447 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
5448 *LHSC != Op0KnownZeroInverted)
5452 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
5453 APInt ValToCheck = Op0KnownZeroInverted;
5454 Type *XTy = X->getType();
5455 if (ValToCheck.isPowerOf2()) {
5456 // ((1 << X) & 8) == 0 -> X != 3
5457 // ((1 << X) & 8) != 0 -> X == 3
5458 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5459 auto NewPred = ICmpInst::getInversePredicate(Pred);
5460 return new ICmpInst(NewPred, X, CmpC);
5461 } else if ((++ValToCheck).isPowerOf2()) {
5462 // ((1 << X) & 7) == 0 -> X >= 3
5463 // ((1 << X) & 7) != 0 -> X < 3
5464 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5466 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
5467 return new ICmpInst(NewPred, X, CmpC);
5471 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5473 if (Op0KnownZeroInverted.isOne() &&
5474 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5475 // ((8 >>u X) & 1) == 0 -> X != 3
5476 // ((8 >>u X) & 1) != 0 -> X == 3
5477 unsigned CmpVal = CI->countTrailingZeros();
5478 auto NewPred = ICmpInst::getInversePredicate(Pred);
5479 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5484 case ICmpInst::ICMP_ULT: {
5485 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5486 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5487 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5488 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5491 case ICmpInst::ICMP_UGT: {
5492 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5493 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5494 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5495 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5498 case ICmpInst::ICMP_SLT: {
5499 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5500 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5501 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5502 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5505 case ICmpInst::ICMP_SGT: {
5506 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5507 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5508 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5509 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5512 case ICmpInst::ICMP_SGE:
5513 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
5514 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5515 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5516 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5517 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5518 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5519 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5521 case ICmpInst::ICMP_SLE:
5522 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
5523 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5524 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5525 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5526 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5527 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5528 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5530 case ICmpInst::ICMP_UGE:
5531 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
5532 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5533 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5534 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5535 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5536 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5537 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5539 case ICmpInst::ICMP_ULE:
5540 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
5541 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5542 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5543 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5544 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5545 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5546 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5550 // Turn a signed comparison into an unsigned one if both operands are known to
5551 // have the same sign.
5553 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5554 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5555 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5560 llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
5561 InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5563 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&
5564 "Only for relational integer predicates.");
5566 Type *Type = C->getType();
5567 bool IsSigned = ICmpInst::isSigned(Pred);
5569 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5570 bool WillIncrement =
5571 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5573 // Check if the constant operand can be safely incremented/decremented
5574 // without overflowing/underflowing.
5575 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5576 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5579 Constant *SafeReplacementConstant = nullptr;
5580 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5581 // Bail out if the constant can't be safely incremented/decremented.
5582 if (!ConstantIsOk(CI))
5584 } else if (auto *FVTy = dyn_cast<FixedVectorType>(Type)) {
5585 unsigned NumElts = FVTy->getNumElements();
5586 for (unsigned i = 0; i != NumElts; ++i) {
5587 Constant *Elt = C->getAggregateElement(i);
5591 if (isa<UndefValue>(Elt))
5594 // Bail out if we can't determine if this constant is min/max or if we
5595 // know that this constant is min/max.
5596 auto *CI = dyn_cast<ConstantInt>(Elt);
5597 if (!CI || !ConstantIsOk(CI))
5600 if (!SafeReplacementConstant)
5601 SafeReplacementConstant = CI;
5608 // It may not be safe to change a compare predicate in the presence of
5609 // undefined elements, so replace those elements with the first safe constant
5611 // TODO: in case of poison, it is safe; let's replace undefs only.
5612 if (C->containsUndefOrPoisonElement()) {
5613 assert(SafeReplacementConstant && "Replacement constant not set");
5614 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5617 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5619 // Increment or decrement the constant.
5620 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5621 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5623 return std::make_pair(NewPred, NewC);
5626 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
5627 /// it into the appropriate icmp lt or icmp gt instruction. This transform
5628 /// allows them to be folded in visitICmpInst.
5629 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5630 ICmpInst::Predicate Pred = I.getPredicate();
5631 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5632 InstCombiner::isCanonicalPredicate(Pred))
5635 Value *Op0 = I.getOperand(0);
5636 Value *Op1 = I.getOperand(1);
5637 auto *Op1C = dyn_cast<Constant>(Op1);
5641 auto FlippedStrictness =
5642 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5643 if (!FlippedStrictness)
5646 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5649 /// If we have a comparison with a non-canonical predicate, if we can update
5650 /// all the users, invert the predicate and adjust all the users.
5651 CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
5652 // Is the predicate already canonical?
5653 CmpInst::Predicate Pred = I.getPredicate();
5654 if (InstCombiner::isCanonicalPredicate(Pred))
5657 // Can all users be adjusted to predicate inversion?
5658 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
5661 // Ok, we can canonicalize comparison!
5662 // Let's first invert the comparison's predicate.
5663 I.setPredicate(CmpInst::getInversePredicate(Pred));
5664 I.setName(I.getName() + ".not");
5666 // And, adapt users.
5667 freelyInvertAllUsersOf(&I);
5672 /// Integer compare with boolean values can always be turned into bitwise ops.
5673 static Instruction *canonicalizeICmpBool(ICmpInst &I,
5674 InstCombiner::BuilderTy &Builder) {
5675 Value *A = I.getOperand(0), *B = I.getOperand(1);
5676 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
5678 // A boolean compared to true/false can be simplified to Op0/true/false in
5679 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5680 // Cases not handled by InstSimplify are always 'not' of Op0.
5681 if (match(B, m_Zero())) {
5682 switch (I.getPredicate()) {
5683 case CmpInst::ICMP_EQ: // A == 0 -> !A
5684 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5685 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5686 return BinaryOperator::CreateNot(A);
5688 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
5690 } else if (match(B, m_One())) {
5691 switch (I.getPredicate()) {
5692 case CmpInst::ICMP_NE: // A != 1 -> !A
5693 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5694 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5695 return BinaryOperator::CreateNot(A);
5697 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
5701 switch (I.getPredicate()) {
5703 llvm_unreachable("Invalid icmp instruction!");
5704 case ICmpInst::ICMP_EQ:
5705 // icmp eq i1 A, B -> ~(A ^ B)
5706 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5708 case ICmpInst::ICMP_NE:
5709 // icmp ne i1 A, B -> A ^ B
5710 return BinaryOperator::CreateXor(A, B);
5712 case ICmpInst::ICMP_UGT:
5713 // icmp ugt -> icmp ult
5716 case ICmpInst::ICMP_ULT:
5717 // icmp ult i1 A, B -> ~A & B
5718 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5720 case ICmpInst::ICMP_SGT:
5721 // icmp sgt -> icmp slt
5724 case ICmpInst::ICMP_SLT:
5725 // icmp slt i1 A, B -> A & ~B
5726 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5728 case ICmpInst::ICMP_UGE:
5729 // icmp uge -> icmp ule
5732 case ICmpInst::ICMP_ULE:
5733 // icmp ule i1 A, B -> ~A | B
5734 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5736 case ICmpInst::ICMP_SGE:
5737 // icmp sge -> icmp sle
5740 case ICmpInst::ICMP_SLE:
5741 // icmp sle i1 A, B -> A | ~B
5742 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5746 // Transform pattern like:
5747 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5748 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5752 static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5753 InstCombiner::BuilderTy &Builder) {
5754 ICmpInst::Predicate Pred, NewPred;
5757 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5759 case ICmpInst::ICMP_ULE:
5760 NewPred = ICmpInst::ICMP_NE;
5762 case ICmpInst::ICMP_UGT:
5763 NewPred = ICmpInst::ICMP_EQ;
5768 } else if (match(&Cmp, m_c_ICmp(Pred,
5769 m_OneUse(m_CombineOr(
5770 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5771 m_Add(m_Shl(m_One(), m_Value(Y)),
5774 // The variant with 'add' is not canonical, (the variant with 'not' is)
5775 // we only get it because it has extra uses, and can't be canonicalized,
5778 case ICmpInst::ICMP_ULT:
5779 NewPred = ICmpInst::ICMP_NE;
5781 case ICmpInst::ICMP_UGE:
5782 NewPred = ICmpInst::ICMP_EQ;
5790 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5791 Constant *Zero = Constant::getNullValue(NewX->getType());
5792 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5795 static Instruction *foldVectorCmp(CmpInst &Cmp,
5796 InstCombiner::BuilderTy &Builder) {
5797 const CmpInst::Predicate Pred = Cmp.getPredicate();
5798 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5801 if (!match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(M))))
5804 // If both arguments of the cmp are shuffles that use the same mask and
5805 // shuffle within a single vector, move the shuffle after the cmp:
5806 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5807 Type *V1Ty = V1->getType();
5808 if (match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(M))) &&
5809 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
5810 Value *NewCmp = Builder.CreateCmp(Pred, V1, V2);
5811 return new ShuffleVectorInst(NewCmp, M);
5814 // Try to canonicalize compare with splatted operand and splat constant.
5815 // TODO: We could generalize this for more than splats. See/use the code in
5816 // InstCombiner::foldVectorBinop().
5818 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
5821 // Length-changing splats are ok, so adjust the constants as needed:
5822 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
5823 Constant *ScalarC = C->getSplatValue(/* AllowUndefs */ true);
5825 if (ScalarC && match(M, m_SplatOrUndefMask(MaskSplatIndex))) {
5826 // We allow undefs in matching, but this transform removes those for safety.
5827 // Demanded elements analysis should be able to recover some/all of that.
5828 C = ConstantVector::getSplat(cast<VectorType>(V1Ty)->getElementCount(),
5830 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
5831 Value *NewCmp = Builder.CreateCmp(Pred, V1, C);
5832 return new ShuffleVectorInst(NewCmp, NewM);
5838 // extract(uadd.with.overflow(A, B), 0) ult A
5839 // -> extract(uadd.with.overflow(A, B), 1)
5840 static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5841 CmpInst::Predicate Pred = I.getPredicate();
5842 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5846 auto UAddOvResultPat = m_ExtractValue<0>(
5847 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5848 if (match(Op0, UAddOvResultPat) &&
5849 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5850 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5851 (match(A, m_One()) || match(B, m_One()))) ||
5852 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5853 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5854 // extract(uadd.with.overflow(A, B), 0) < A
5855 // extract(uadd.with.overflow(A, 1), 0) == 0
5856 // extract(uadd.with.overflow(A, -1), 0) != -1
5857 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5858 else if (match(Op1, UAddOvResultPat) &&
5859 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5860 // A > extract(uadd.with.overflow(A, B), 0)
5861 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5865 return ExtractValueInst::Create(UAddOv, 1);
5868 static Instruction *foldICmpInvariantGroup(ICmpInst &I) {
5869 if (!I.getOperand(0)->getType()->isPointerTy() ||
5870 NullPointerIsDefined(
5871 I.getParent()->getParent(),
5872 I.getOperand(0)->getType()->getPointerAddressSpace())) {
5876 if (match(I.getOperand(0), m_Instruction(Op)) &&
5877 match(I.getOperand(1), m_Zero()) &&
5878 Op->isLaunderOrStripInvariantGroup()) {
5879 return ICmpInst::Create(Instruction::ICmp, I.getPredicate(),
5880 Op->getOperand(0), I.getOperand(1));
5885 /// This function folds patterns produced by lowering of reduce idioms, such as
5886 /// llvm.vector.reduce.and which are lowered into instruction chains. This code
5887 /// attempts to generate fewer number of scalar comparisons instead of vector
5888 /// comparisons when possible.
5889 static Instruction *foldReductionIdiom(ICmpInst &I,
5890 InstCombiner::BuilderTy &Builder,
5891 const DataLayout &DL) {
5892 if (I.getType()->isVectorTy())
5894 ICmpInst::Predicate OuterPred, InnerPred;
5897 // Match lowering of @llvm.vector.reduce.and. Turn
5898 /// %vec_ne = icmp ne <8 x i8> %lhs, %rhs
5899 /// %scalar_ne = bitcast <8 x i1> %vec_ne to i8
5900 /// %res = icmp <pred> i8 %scalar_ne, 0
5904 /// %lhs.scalar = bitcast <8 x i8> %lhs to i64
5905 /// %rhs.scalar = bitcast <8 x i8> %rhs to i64
5906 /// %res = icmp <pred> i64 %lhs.scalar, %rhs.scalar
5908 /// for <pred> in {ne, eq}.
5909 if (!match(&I, m_ICmp(OuterPred,
5910 m_OneUse(m_BitCast(m_OneUse(
5911 m_ICmp(InnerPred, m_Value(LHS), m_Value(RHS))))),
5914 auto *LHSTy = dyn_cast<FixedVectorType>(LHS->getType());
5915 if (!LHSTy || !LHSTy->getElementType()->isIntegerTy())
5918 LHSTy->getNumElements() * LHSTy->getElementType()->getIntegerBitWidth();
5919 // TODO: Relax this to "not wider than max legal integer type"?
5920 if (!DL.isLegalInteger(NumBits))
5923 if (ICmpInst::isEquality(OuterPred) && InnerPred == ICmpInst::ICMP_NE) {
5924 auto *ScalarTy = Builder.getIntNTy(NumBits);
5925 LHS = Builder.CreateBitCast(LHS, ScalarTy, LHS->getName() + ".scalar");
5926 RHS = Builder.CreateBitCast(RHS, ScalarTy, RHS->getName() + ".scalar");
5927 return ICmpInst::Create(Instruction::ICmp, OuterPred, LHS, RHS,
5934 Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
5935 bool Changed = false;
5936 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5937 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5938 unsigned Op0Cplxity = getComplexity(Op0);
5939 unsigned Op1Cplxity = getComplexity(Op1);
5941 /// Orders the operands of the compare so that they are listed from most
5942 /// complex to least complex. This puts constants before unary operators,
5943 /// before binary operators.
5944 if (Op0Cplxity < Op1Cplxity ||
5945 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5947 std::swap(Op0, Op1);
5951 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, Q))
5952 return replaceInstUsesWith(I, V);
5954 // Comparing -val or val with non-zero is the same as just comparing val
5955 // ie, abs(val) != 0 -> val != 0
5956 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5957 Value *Cond, *SelectTrue, *SelectFalse;
5958 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5959 m_Value(SelectFalse)))) {
5960 if (Value *V = dyn_castNegVal(SelectTrue)) {
5961 if (V == SelectFalse)
5962 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5964 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5965 if (V == SelectTrue)
5966 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5971 if (Op0->getType()->isIntOrIntVectorTy(1))
5972 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
5975 if (Instruction *Res = canonicalizeCmpWithConstant(I))
5978 if (Instruction *Res = canonicalizeICmpPredicate(I))
5981 if (Instruction *Res = foldICmpWithConstant(I))
5984 if (Instruction *Res = foldICmpWithDominatingICmp(I))
5987 if (Instruction *Res = foldICmpUsingKnownBits(I))
5990 // Test if the ICmpInst instruction is used exclusively by a select as
5991 // part of a minimum or maximum operation. If so, refrain from doing
5992 // any other folding. This helps out other analyses which understand
5993 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5994 // and CodeGen. And in this case, at least one of the comparison
5995 // operands has at least one user besides the compare (the select),
5996 // which would often largely negate the benefit of folding anyway.
5998 // Do the same for the other patterns recognized by matchSelectPattern.
6000 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
6002 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
6003 if (SPR.Flavor != SPF_UNKNOWN)
6007 // Do this after checking for min/max to prevent infinite looping.
6008 if (Instruction *Res = foldICmpWithZero(I))
6011 // FIXME: We only do this after checking for min/max to prevent infinite
6012 // looping caused by a reverse canonicalization of these patterns for min/max.
6013 // FIXME: The organization of folds is a mess. These would naturally go into
6014 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
6015 // down here after the min/max restriction.
6016 ICmpInst::Predicate Pred = I.getPredicate();
6018 if (match(Op1, m_APInt(C))) {
6019 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
6020 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
6021 Constant *Zero = Constant::getNullValue(Op0->getType());
6022 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
6025 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
6026 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
6027 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
6028 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
6032 // The folds in here may rely on wrapping flags and special constants, so
6033 // they can break up min/max idioms in some cases but not seemingly similar
6035 // FIXME: It may be possible to enhance select folding to make this
6036 // unnecessary. It may also be moot if we canonicalize to min/max
6038 if (Instruction *Res = foldICmpBinOp(I, Q))
6041 if (Instruction *Res = foldICmpInstWithConstant(I))
6044 // Try to match comparison as a sign bit test. Intentionally do this after
6045 // foldICmpInstWithConstant() to potentially let other folds to happen first.
6046 if (Instruction *New = foldSignBitTest(I))
6049 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
6052 // Try to optimize 'icmp GEP, P' or 'icmp P, GEP'.
6053 if (auto *GEP = dyn_cast<GEPOperator>(Op0))
6054 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
6056 if (auto *GEP = dyn_cast<GEPOperator>(Op1))
6057 if (Instruction *NI = foldGEPICmp(GEP, Op0, I.getSwappedPredicate(), I))
6060 // Try to optimize equality comparisons against alloca-based pointers.
6061 if (Op0->getType()->isPointerTy() && I.isEquality()) {
6062 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
6063 if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(Op0)))
6064 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
6066 if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(Op1)))
6067 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
6071 if (Instruction *Res = foldICmpBitCast(I))
6074 // TODO: Hoist this above the min/max bailout.
6075 if (Instruction *R = foldICmpWithCastOp(I))
6078 if (Instruction *Res = foldICmpWithMinMax(I))
6083 // Transform (A & ~B) == 0 --> (A & B) != 0
6084 // and (A & ~B) != 0 --> (A & B) == 0
6085 // if A is a power of 2.
6086 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
6087 match(Op1, m_Zero()) &&
6088 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
6089 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
6092 // ~X < ~Y --> Y < X
6093 // ~X < C --> X > ~C
6094 if (match(Op0, m_Not(m_Value(A)))) {
6095 if (match(Op1, m_Not(m_Value(B))))
6096 return new ICmpInst(I.getPredicate(), B, A);
6099 if (match(Op1, m_APInt(C)))
6100 return new ICmpInst(I.getSwappedPredicate(), A,
6101 ConstantInt::get(Op1->getType(), ~(*C)));
6104 Instruction *AddI = nullptr;
6105 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
6106 m_Instruction(AddI))) &&
6107 isa<IntegerType>(A->getType())) {
6110 // m_UAddWithOverflow can match patterns that do not include an explicit
6111 // "add" instruction, so check the opcode of the matched op.
6112 if (AddI->getOpcode() == Instruction::Add &&
6113 OptimizeOverflowCheck(Instruction::Add, /*Signed*/ false, A, B, *AddI,
6114 Result, Overflow)) {
6115 replaceInstUsesWith(*AddI, Result);
6116 eraseInstFromFunction(*AddI);
6117 return replaceInstUsesWith(I, Overflow);
6121 // (zext a) * (zext b) --> llvm.umul.with.overflow.
6122 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
6123 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
6126 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
6127 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
6132 if (Instruction *Res = foldICmpEquality(I))
6135 if (Instruction *Res = foldICmpOfUAddOv(I))
6138 // The 'cmpxchg' instruction returns an aggregate containing the old value and
6139 // an i1 which indicates whether or not we successfully did the swap.
6141 // Replace comparisons between the old value and the expected value with the
6142 // indicator that 'cmpxchg' returns.
6144 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
6145 // spuriously fail. In those cases, the old value may equal the expected
6146 // value but it is possible for the swap to not occur.
6147 if (I.getPredicate() == ICmpInst::ICMP_EQ)
6148 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
6149 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
6150 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
6152 return ExtractValueInst::Create(ACXI, 1);
6158 if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
6159 return foldICmpAddOpConst(X, *C, I.getPredicate());
6162 if (match(Op1, m_Add(m_Value(X), m_APInt(C))) && Op0 == X)
6163 return foldICmpAddOpConst(X, *C, I.getSwappedPredicate());
6166 if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
6169 if (I.getType()->isVectorTy())
6170 if (Instruction *Res = foldVectorCmp(I, Builder))
6173 if (Instruction *Res = foldICmpInvariantGroup(I))
6176 if (Instruction *Res = foldReductionIdiom(I, Builder, DL))
6179 return Changed ? &I : nullptr;
6182 /// Fold fcmp ([us]itofp x, cst) if possible.
6183 Instruction *InstCombinerImpl::foldFCmpIntToFPConst(FCmpInst &I,
6186 if (!isa<ConstantFP>(RHSC)) return nullptr;
6187 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
6189 // Get the width of the mantissa. We don't want to hack on conversions that
6190 // might lose information from the integer, e.g. "i64 -> float"
6191 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
6192 if (MantissaWidth == -1) return nullptr; // Unknown.
6194 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
6196 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
6198 if (I.isEquality()) {
6199 FCmpInst::Predicate P = I.getPredicate();
6200 bool IsExact = false;
6201 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
6202 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
6204 // If the floating point constant isn't an integer value, we know if we will
6205 // ever compare equal / not equal to it.
6207 // TODO: Can never be -0.0 and other non-representable values
6208 APFloat RHSRoundInt(RHS);
6209 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
6210 if (RHS != RHSRoundInt) {
6211 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
6212 return replaceInstUsesWith(I, Builder.getFalse());
6214 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
6215 return replaceInstUsesWith(I, Builder.getTrue());
6219 // TODO: If the constant is exactly representable, is it always OK to do
6220 // equality compares as integer?
6223 // Check to see that the input is converted from an integer type that is small
6224 // enough that preserves all bits. TODO: check here for "known" sign bits.
6225 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
6226 unsigned InputSize = IntTy->getScalarSizeInBits();
6228 // Following test does NOT adjust InputSize downwards for signed inputs,
6229 // because the most negative value still requires all the mantissa bits
6230 // to distinguish it from one less than that value.
6231 if ((int)InputSize > MantissaWidth) {
6232 // Conversion would lose accuracy. Check if loss can impact comparison.
6233 int Exp = ilogb(RHS);
6234 if (Exp == APFloat::IEK_Inf) {
6235 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
6236 if (MaxExponent < (int)InputSize - !LHSUnsigned)
6237 // Conversion could create infinity.
6240 // Note that if RHS is zero or NaN, then Exp is negative
6241 // and first condition is trivially false.
6242 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
6243 // Conversion could affect comparison.
6248 // Otherwise, we can potentially simplify the comparison. We know that it
6249 // will always come through as an integer value and we know the constant is
6250 // not a NAN (it would have been previously simplified).
6251 assert(!RHS.isNaN() && "NaN comparison not already folded!");
6253 ICmpInst::Predicate Pred;
6254 switch (I.getPredicate()) {
6255 default: llvm_unreachable("Unexpected predicate!");
6256 case FCmpInst::FCMP_UEQ:
6257 case FCmpInst::FCMP_OEQ:
6258 Pred = ICmpInst::ICMP_EQ;
6260 case FCmpInst::FCMP_UGT:
6261 case FCmpInst::FCMP_OGT:
6262 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
6264 case FCmpInst::FCMP_UGE:
6265 case FCmpInst::FCMP_OGE:
6266 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
6268 case FCmpInst::FCMP_ULT:
6269 case FCmpInst::FCMP_OLT:
6270 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
6272 case FCmpInst::FCMP_ULE:
6273 case FCmpInst::FCMP_OLE:
6274 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
6276 case FCmpInst::FCMP_UNE:
6277 case FCmpInst::FCMP_ONE:
6278 Pred = ICmpInst::ICMP_NE;
6280 case FCmpInst::FCMP_ORD:
6281 return replaceInstUsesWith(I, Builder.getTrue());
6282 case FCmpInst::FCMP_UNO:
6283 return replaceInstUsesWith(I, Builder.getFalse());
6286 // Now we know that the APFloat is a normal number, zero or inf.
6288 // See if the FP constant is too large for the integer. For example,
6289 // comparing an i8 to 300.0.
6290 unsigned IntWidth = IntTy->getScalarSizeInBits();
6293 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
6294 // and large values.
6295 APFloat SMax(RHS.getSemantics());
6296 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
6297 APFloat::rmNearestTiesToEven);
6298 if (SMax < RHS) { // smax < 13123.0
6299 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
6300 Pred == ICmpInst::ICMP_SLE)
6301 return replaceInstUsesWith(I, Builder.getTrue());
6302 return replaceInstUsesWith(I, Builder.getFalse());
6305 // If the RHS value is > UnsignedMax, fold the comparison. This handles
6306 // +INF and large values.
6307 APFloat UMax(RHS.getSemantics());
6308 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
6309 APFloat::rmNearestTiesToEven);
6310 if (UMax < RHS) { // umax < 13123.0
6311 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
6312 Pred == ICmpInst::ICMP_ULE)
6313 return replaceInstUsesWith(I, Builder.getTrue());
6314 return replaceInstUsesWith(I, Builder.getFalse());
6319 // See if the RHS value is < SignedMin.
6320 APFloat SMin(RHS.getSemantics());
6321 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
6322 APFloat::rmNearestTiesToEven);
6323 if (SMin > RHS) { // smin > 12312.0
6324 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
6325 Pred == ICmpInst::ICMP_SGE)
6326 return replaceInstUsesWith(I, Builder.getTrue());
6327 return replaceInstUsesWith(I, Builder.getFalse());
6330 // See if the RHS value is < UnsignedMin.
6331 APFloat UMin(RHS.getSemantics());
6332 UMin.convertFromAPInt(APInt::getMinValue(IntWidth), false,
6333 APFloat::rmNearestTiesToEven);
6334 if (UMin > RHS) { // umin > 12312.0
6335 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
6336 Pred == ICmpInst::ICMP_UGE)
6337 return replaceInstUsesWith(I, Builder.getTrue());
6338 return replaceInstUsesWith(I, Builder.getFalse());
6342 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
6343 // [0, UMAX], but it may still be fractional. See if it is fractional by
6344 // casting the FP value to the integer value and back, checking for equality.
6345 // Don't do this for zero, because -0.0 is not fractional.
6346 Constant *RHSInt = LHSUnsigned
6347 ? ConstantExpr::getFPToUI(RHSC, IntTy)
6348 : ConstantExpr::getFPToSI(RHSC, IntTy);
6349 if (!RHS.isZero()) {
6350 bool Equal = LHSUnsigned
6351 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
6352 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
6354 // If we had a comparison against a fractional value, we have to adjust
6355 // the compare predicate and sometimes the value. RHSC is rounded towards
6356 // zero at this point.
6358 default: llvm_unreachable("Unexpected integer comparison!");
6359 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
6360 return replaceInstUsesWith(I, Builder.getTrue());
6361 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
6362 return replaceInstUsesWith(I, Builder.getFalse());
6363 case ICmpInst::ICMP_ULE:
6364 // (float)int <= 4.4 --> int <= 4
6365 // (float)int <= -4.4 --> false
6366 if (RHS.isNegative())
6367 return replaceInstUsesWith(I, Builder.getFalse());
6369 case ICmpInst::ICMP_SLE:
6370 // (float)int <= 4.4 --> int <= 4
6371 // (float)int <= -4.4 --> int < -4
6372 if (RHS.isNegative())
6373 Pred = ICmpInst::ICMP_SLT;
6375 case ICmpInst::ICMP_ULT:
6376 // (float)int < -4.4 --> false
6377 // (float)int < 4.4 --> int <= 4
6378 if (RHS.isNegative())
6379 return replaceInstUsesWith(I, Builder.getFalse());
6380 Pred = ICmpInst::ICMP_ULE;
6382 case ICmpInst::ICMP_SLT:
6383 // (float)int < -4.4 --> int < -4
6384 // (float)int < 4.4 --> int <= 4
6385 if (!RHS.isNegative())
6386 Pred = ICmpInst::ICMP_SLE;
6388 case ICmpInst::ICMP_UGT:
6389 // (float)int > 4.4 --> int > 4
6390 // (float)int > -4.4 --> true
6391 if (RHS.isNegative())
6392 return replaceInstUsesWith(I, Builder.getTrue());
6394 case ICmpInst::ICMP_SGT:
6395 // (float)int > 4.4 --> int > 4
6396 // (float)int > -4.4 --> int >= -4
6397 if (RHS.isNegative())
6398 Pred = ICmpInst::ICMP_SGE;
6400 case ICmpInst::ICMP_UGE:
6401 // (float)int >= -4.4 --> true
6402 // (float)int >= 4.4 --> int > 4
6403 if (RHS.isNegative())
6404 return replaceInstUsesWith(I, Builder.getTrue());
6405 Pred = ICmpInst::ICMP_UGT;
6407 case ICmpInst::ICMP_SGE:
6408 // (float)int >= -4.4 --> int >= -4
6409 // (float)int >= 4.4 --> int > 4
6410 if (!RHS.isNegative())
6411 Pred = ICmpInst::ICMP_SGT;
6417 // Lower this FP comparison into an appropriate integer version of the
6419 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
6422 /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
6423 static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
6425 // When C is not 0.0 and infinities are not allowed:
6426 // (C / X) < 0.0 is a sign-bit test of X
6427 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
6428 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
6431 // Multiply (C / X) < 0.0 by X * X / C.
6432 // - X is non zero, if it is the flag 'ninf' is violated.
6433 // - C defines the sign of X * X * C. Thus it also defines whether to swap
6434 // the predicate. C is also non zero by definition.
6436 // Thus X * X / C is non zero and the transformation is valid. [qed]
6438 FCmpInst::Predicate Pred = I.getPredicate();
6440 // Check that predicates are valid.
6441 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
6442 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
6445 // Check that RHS operand is zero.
6446 if (!match(RHSC, m_AnyZeroFP()))
6449 // Check fastmath flags ('ninf').
6450 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
6453 // Check the properties of the dividend. It must not be zero to avoid a
6454 // division by zero (see Proof).
6456 if (!match(LHSI->getOperand(0), m_APFloat(C)))
6462 // Get swapped predicate if necessary.
6463 if (C->isNegative())
6464 Pred = I.getSwappedPredicate();
6466 return new FCmpInst(Pred, LHSI->getOperand(1), RHSC, "", &I);
6469 /// Optimize fabs(X) compared with zero.
6470 static Instruction *foldFabsWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
6472 if (!match(I.getOperand(0), m_FAbs(m_Value(X))) ||
6473 !match(I.getOperand(1), m_PosZeroFP()))
6476 auto replacePredAndOp0 = [&IC](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
6478 return IC.replaceOperand(*I, 0, X);
6481 switch (I.getPredicate()) {
6482 case FCmpInst::FCMP_UGE:
6483 case FCmpInst::FCMP_OLT:
6484 // fabs(X) >= 0.0 --> true
6485 // fabs(X) < 0.0 --> false
6486 llvm_unreachable("fcmp should have simplified");
6488 case FCmpInst::FCMP_OGT:
6489 // fabs(X) > 0.0 --> X != 0.0
6490 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
6492 case FCmpInst::FCMP_UGT:
6493 // fabs(X) u> 0.0 --> X u!= 0.0
6494 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
6496 case FCmpInst::FCMP_OLE:
6497 // fabs(X) <= 0.0 --> X == 0.0
6498 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
6500 case FCmpInst::FCMP_ULE:
6501 // fabs(X) u<= 0.0 --> X u== 0.0
6502 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
6504 case FCmpInst::FCMP_OGE:
6505 // fabs(X) >= 0.0 --> !isnan(X)
6506 assert(!I.hasNoNaNs() && "fcmp should have simplified");
6507 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
6509 case FCmpInst::FCMP_ULT:
6510 // fabs(X) u< 0.0 --> isnan(X)
6511 assert(!I.hasNoNaNs() && "fcmp should have simplified");
6512 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
6514 case FCmpInst::FCMP_OEQ:
6515 case FCmpInst::FCMP_UEQ:
6516 case FCmpInst::FCMP_ONE:
6517 case FCmpInst::FCMP_UNE:
6518 case FCmpInst::FCMP_ORD:
6519 case FCmpInst::FCMP_UNO:
6520 // Look through the fabs() because it doesn't change anything but the sign.
6521 // fabs(X) == 0.0 --> X == 0.0,
6522 // fabs(X) != 0.0 --> X != 0.0
6523 // isnan(fabs(X)) --> isnan(X)
6524 // !isnan(fabs(X) --> !isnan(X)
6525 return replacePredAndOp0(&I, I.getPredicate(), X);
6532 Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
6533 bool Changed = false;
6535 /// Orders the operands of the compare so that they are listed from most
6536 /// complex to least complex. This puts constants before unary operators,
6537 /// before binary operators.
6538 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6543 const CmpInst::Predicate Pred = I.getPredicate();
6544 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6545 if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
6546 SQ.getWithInstruction(&I)))
6547 return replaceInstUsesWith(I, V);
6549 // Simplify 'fcmp pred X, X'
6550 Type *OpType = Op0->getType();
6551 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
6555 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
6556 case FCmpInst::FCMP_ULT: // True if unordered or less than
6557 case FCmpInst::FCMP_UGT: // True if unordered or greater than
6558 case FCmpInst::FCMP_UNE: // True if unordered or not equal
6559 // Canonicalize these to be 'fcmp uno %X, 0.0'.
6560 I.setPredicate(FCmpInst::FCMP_UNO);
6561 I.setOperand(1, Constant::getNullValue(OpType));
6564 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
6565 case FCmpInst::FCMP_OEQ: // True if ordered and equal
6566 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
6567 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
6568 // Canonicalize these to be 'fcmp ord %X, 0.0'.
6569 I.setPredicate(FCmpInst::FCMP_ORD);
6570 I.setOperand(1, Constant::getNullValue(OpType));
6575 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
6576 // then canonicalize the operand to 0.0.
6577 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
6578 if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0, &TLI))
6579 return replaceOperand(I, 0, ConstantFP::getNullValue(OpType));
6581 if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1, &TLI))
6582 return replaceOperand(I, 1, ConstantFP::getNullValue(OpType));
6585 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
6587 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
6588 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
6590 // Test if the FCmpInst instruction is used exclusively by a select as
6591 // part of a minimum or maximum operation. If so, refrain from doing
6592 // any other folding. This helps out other analyses which understand
6593 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6594 // and CodeGen. And in this case, at least one of the comparison
6595 // operands has at least one user besides the compare (the select),
6596 // which would often largely negate the benefit of folding anyway.
6598 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
6600 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
6601 if (SPR.Flavor != SPF_UNKNOWN)
6605 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
6606 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
6607 if (match(Op1, m_AnyZeroFP()) && !match(Op1, m_PosZeroFP()))
6608 return replaceOperand(I, 1, ConstantFP::getNullValue(OpType));
6610 // Handle fcmp with instruction LHS and constant RHS.
6613 if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
6614 switch (LHSI->getOpcode()) {
6615 case Instruction::PHI:
6616 // Only fold fcmp into the PHI if the phi and fcmp are in the same
6617 // block. If in the same block, we're encouraging jump threading. If
6618 // not, we are just pessimizing the code by making an i1 phi.
6619 if (LHSI->getParent() == I.getParent())
6620 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
6623 case Instruction::SIToFP:
6624 case Instruction::UIToFP:
6625 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
6628 case Instruction::FDiv:
6629 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
6632 case Instruction::Load:
6633 if (auto *GEP = dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
6634 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
6635 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
6636 !cast<LoadInst>(LHSI)->isVolatile())
6637 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
6643 if (Instruction *R = foldFabsWithFcmpZero(I, *this))
6646 if (match(Op0, m_FNeg(m_Value(X)))) {
6647 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
6649 if (match(Op1, m_Constant(C))) {
6650 Constant *NegC = ConstantExpr::getFNeg(C);
6651 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
6655 if (match(Op0, m_FPExt(m_Value(X)))) {
6656 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
6657 if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
6658 return new FCmpInst(Pred, X, Y, "", &I);
6660 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
6662 if (match(Op1, m_APFloat(C))) {
6663 const fltSemantics &FPSem =
6664 X->getType()->getScalarType()->getFltSemantics();
6666 APFloat TruncC = *C;
6667 TruncC.convert(FPSem, APFloat::rmNearestTiesToEven, &Lossy);
6669 // Avoid lossy conversions and denormals.
6670 // Zero is a special case that's OK to convert.
6671 APFloat Fabs = TruncC;
6674 (!(Fabs < APFloat::getSmallestNormalized(FPSem)) || Fabs.isZero())) {
6675 Constant *NewC = ConstantFP::get(X->getType(), TruncC);
6676 return new FCmpInst(Pred, X, NewC, "", &I);
6681 // Convert a sign-bit test of an FP value into a cast and integer compare.
6682 // TODO: Simplify if the copysign constant is 0.0 or NaN.
6683 // TODO: Handle non-zero compare constants.
6684 // TODO: Handle other predicates.
6686 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::copysign>(m_APFloat(C),
6688 match(Op1, m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) {
6689 Type *IntType = Builder.getIntNTy(X->getType()->getScalarSizeInBits());
6690 if (auto *VecTy = dyn_cast<VectorType>(OpType))
6691 IntType = VectorType::get(IntType, VecTy->getElementCount());
6693 // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0
6694 if (Pred == FCmpInst::FCMP_OLT) {
6695 Value *IntX = Builder.CreateBitCast(X, IntType);
6696 return new ICmpInst(ICmpInst::ICMP_SLT, IntX,
6697 ConstantInt::getNullValue(IntType));
6701 if (I.getType()->isVectorTy())
6702 if (Instruction *Res = foldVectorCmp(I, Builder))
6705 return Changed ? &I : nullptr;