[TySan] Don't report globals with incomplete types. (#121922)
[llvm-project.git] / llvm / lib / Transforms / InstCombine / InstCombineCompares.cpp
blobd764f845ffd430cc3abf26aa1199ed60a07d610b
1 //===- InstCombineCompares.cpp --------------------------------------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
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/CaptureTracking.h"
18 #include "llvm/Analysis/CmpInstAnalysis.h"
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/Utils/Local.h"
22 #include "llvm/Analysis/VectorUtils.h"
23 #include "llvm/IR/ConstantRange.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/InstrTypes.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/Support/KnownBits.h"
29 #include "llvm/Transforms/InstCombine/InstCombiner.h"
30 #include <bitset>
32 using namespace llvm;
33 using namespace PatternMatch;
35 #define DEBUG_TYPE "instcombine"
37 // How many times is a select replaced by one of its operands?
38 STATISTIC(NumSel, "Number of select opts");
41 /// Compute Result = In1+In2, returning true if the result overflowed for this
42 /// type.
43 static bool addWithOverflow(APInt &Result, const APInt &In1,
44 const APInt &In2, bool IsSigned = false) {
45 bool Overflow;
46 if (IsSigned)
47 Result = In1.sadd_ov(In2, Overflow);
48 else
49 Result = In1.uadd_ov(In2, Overflow);
51 return Overflow;
54 /// Compute Result = In1-In2, returning true if the result overflowed for this
55 /// type.
56 static bool subWithOverflow(APInt &Result, const APInt &In1,
57 const APInt &In2, bool IsSigned = false) {
58 bool Overflow;
59 if (IsSigned)
60 Result = In1.ssub_ov(In2, Overflow);
61 else
62 Result = In1.usub_ov(In2, Overflow);
64 return Overflow;
67 /// Given an icmp instruction, return true if any use of this comparison is a
68 /// branch on sign bit comparison.
69 static bool hasBranchUse(ICmpInst &I) {
70 for (auto *U : I.users())
71 if (isa<BranchInst>(U))
72 return true;
73 return false;
76 /// Returns true if the exploded icmp can be expressed as a signed comparison
77 /// to zero and updates the predicate accordingly.
78 /// The signedness of the comparison is preserved.
79 /// TODO: Refactor with decomposeBitTestICmp()?
80 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
81 if (!ICmpInst::isSigned(Pred))
82 return false;
84 if (C.isZero())
85 return ICmpInst::isRelational(Pred);
87 if (C.isOne()) {
88 if (Pred == ICmpInst::ICMP_SLT) {
89 Pred = ICmpInst::ICMP_SLE;
90 return true;
92 } else if (C.isAllOnes()) {
93 if (Pred == ICmpInst::ICMP_SGT) {
94 Pred = ICmpInst::ICMP_SGE;
95 return true;
99 return false;
102 /// This is called when we see this pattern:
103 /// cmp pred (load (gep GV, ...)), cmpcst
104 /// where GV is a global variable with a constant initializer. Try to simplify
105 /// this into some simple computation that does not need the load. For example
106 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
108 /// If AndCst is non-null, then the loaded value is masked with that constant
109 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
110 Instruction *InstCombinerImpl::foldCmpLoadFromIndexedGlobal(
111 LoadInst *LI, GetElementPtrInst *GEP, GlobalVariable *GV, CmpInst &ICI,
112 ConstantInt *AndCst) {
113 if (LI->isVolatile() || LI->getType() != GEP->getResultElementType() ||
114 GV->getValueType() != GEP->getSourceElementType() || !GV->isConstant() ||
115 !GV->hasDefinitiveInitializer())
116 return nullptr;
118 Constant *Init = GV->getInitializer();
119 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
120 return nullptr;
122 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
123 // Don't blow up on huge arrays.
124 if (ArrayElementCount > MaxArraySizeForCombine)
125 return nullptr;
127 // There are many forms of this optimization we can handle, for now, just do
128 // the simple index into a single-dimensional array.
130 // Require: GEP GV, 0, i {{, constant indices}}
131 if (GEP->getNumOperands() < 3 || !isa<ConstantInt>(GEP->getOperand(1)) ||
132 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
133 isa<Constant>(GEP->getOperand(2)))
134 return nullptr;
136 // Check that indices after the variable are constants and in-range for the
137 // type they index. Collect the indices. This is typically for arrays of
138 // structs.
139 SmallVector<unsigned, 4> LaterIndices;
141 Type *EltTy = Init->getType()->getArrayElementType();
142 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
143 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
144 if (!Idx)
145 return nullptr; // Variable index.
147 uint64_t IdxVal = Idx->getZExtValue();
148 if ((unsigned)IdxVal != IdxVal)
149 return nullptr; // Too large array index.
151 if (StructType *STy = dyn_cast<StructType>(EltTy))
152 EltTy = STy->getElementType(IdxVal);
153 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
154 if (IdxVal >= ATy->getNumElements())
155 return nullptr;
156 EltTy = ATy->getElementType();
157 } else {
158 return nullptr; // Unknown type.
161 LaterIndices.push_back(IdxVal);
164 enum { Overdefined = -3, Undefined = -2 };
166 // Variables for our state machines.
168 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
169 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
170 // and 87 is the second (and last) index. FirstTrueElement is -2 when
171 // undefined, otherwise set to the first true element. SecondTrueElement is
172 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
173 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
175 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
176 // form "i != 47 & i != 87". Same state transitions as for true elements.
177 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
179 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
180 /// define a state machine that triggers for ranges of values that the index
181 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
182 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
183 /// index in the range (inclusive). We use -2 for undefined here because we
184 /// use relative comparisons and don't want 0-1 to match -1.
185 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
187 // MagicBitvector - This is a magic bitvector where we set a bit if the
188 // comparison is true for element 'i'. If there are 64 elements or less in
189 // the array, this will fully represent all the comparison results.
190 uint64_t MagicBitvector = 0;
192 // Scan the array and see if one of our patterns matches.
193 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
194 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
195 Constant *Elt = Init->getAggregateElement(i);
196 if (!Elt)
197 return nullptr;
199 // If this is indexing an array of structures, get the structure element.
200 if (!LaterIndices.empty()) {
201 Elt = ConstantFoldExtractValueInstruction(Elt, LaterIndices);
202 if (!Elt)
203 return nullptr;
206 // If the element is masked, handle it.
207 if (AndCst) {
208 Elt = ConstantFoldBinaryOpOperands(Instruction::And, Elt, AndCst, DL);
209 if (!Elt)
210 return nullptr;
213 // Find out if the comparison would be true or false for the i'th element.
214 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
215 CompareRHS, DL, &TLI);
216 if (!C)
217 return nullptr;
219 // If the result is undef for this element, ignore it.
220 if (isa<UndefValue>(C)) {
221 // Extend range state machines to cover this element in case there is an
222 // undef in the middle of the range.
223 if (TrueRangeEnd == (int)i - 1)
224 TrueRangeEnd = i;
225 if (FalseRangeEnd == (int)i - 1)
226 FalseRangeEnd = i;
227 continue;
230 // If we can't compute the result for any of the elements, we have to give
231 // up evaluating the entire conditional.
232 if (!isa<ConstantInt>(C))
233 return nullptr;
235 // Otherwise, we know if the comparison is true or false for this element,
236 // update our state machines.
237 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
239 // State machine for single/double/range index comparison.
240 if (IsTrueForElt) {
241 // Update the TrueElement state machine.
242 if (FirstTrueElement == Undefined)
243 FirstTrueElement = TrueRangeEnd = i; // First true element.
244 else {
245 // Update double-compare state machine.
246 if (SecondTrueElement == Undefined)
247 SecondTrueElement = i;
248 else
249 SecondTrueElement = Overdefined;
251 // Update range state machine.
252 if (TrueRangeEnd == (int)i - 1)
253 TrueRangeEnd = i;
254 else
255 TrueRangeEnd = Overdefined;
257 } else {
258 // Update the FalseElement state machine.
259 if (FirstFalseElement == Undefined)
260 FirstFalseElement = FalseRangeEnd = i; // First false element.
261 else {
262 // Update double-compare state machine.
263 if (SecondFalseElement == Undefined)
264 SecondFalseElement = i;
265 else
266 SecondFalseElement = Overdefined;
268 // Update range state machine.
269 if (FalseRangeEnd == (int)i - 1)
270 FalseRangeEnd = i;
271 else
272 FalseRangeEnd = Overdefined;
276 // If this element is in range, update our magic bitvector.
277 if (i < 64 && IsTrueForElt)
278 MagicBitvector |= 1ULL << i;
280 // If all of our states become overdefined, bail out early. Since the
281 // predicate is expensive, only check it every 8 elements. This is only
282 // really useful for really huge arrays.
283 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
284 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
285 FalseRangeEnd == Overdefined)
286 return nullptr;
289 // Now that we've scanned the entire array, emit our new comparison(s). We
290 // order the state machines in complexity of the generated code.
291 Value *Idx = GEP->getOperand(2);
293 // If the index is larger than the pointer offset size of the target, truncate
294 // the index down like the GEP would do implicitly. We don't have to do this
295 // for an inbounds GEP because the index can't be out of range.
296 if (!GEP->isInBounds()) {
297 Type *PtrIdxTy = DL.getIndexType(GEP->getType());
298 unsigned OffsetSize = PtrIdxTy->getIntegerBitWidth();
299 if (Idx->getType()->getPrimitiveSizeInBits().getFixedValue() > OffsetSize)
300 Idx = Builder.CreateTrunc(Idx, PtrIdxTy);
303 // If inbounds keyword is not present, Idx * ElementSize can overflow.
304 // Let's assume that ElementSize is 2 and the wanted value is at offset 0.
305 // Then, there are two possible values for Idx to match offset 0:
306 // 0x00..00, 0x80..00.
307 // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
308 // comparison is false if Idx was 0x80..00.
309 // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
310 unsigned ElementSize =
311 DL.getTypeAllocSize(Init->getType()->getArrayElementType());
312 auto MaskIdx = [&](Value *Idx) {
313 if (!GEP->isInBounds() && llvm::countr_zero(ElementSize) != 0) {
314 Value *Mask = Constant::getAllOnesValue(Idx->getType());
315 Mask = Builder.CreateLShr(Mask, llvm::countr_zero(ElementSize));
316 Idx = Builder.CreateAnd(Idx, Mask);
318 return Idx;
321 // If the comparison is only true for one or two elements, emit direct
322 // comparisons.
323 if (SecondTrueElement != Overdefined) {
324 Idx = MaskIdx(Idx);
325 // None true -> false.
326 if (FirstTrueElement == Undefined)
327 return replaceInstUsesWith(ICI, Builder.getFalse());
329 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
331 // True for one element -> 'i == 47'.
332 if (SecondTrueElement == Undefined)
333 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
335 // True for two elements -> 'i == 47 | i == 72'.
336 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
337 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
338 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
339 return BinaryOperator::CreateOr(C1, C2);
342 // If the comparison is only false for one or two elements, emit direct
343 // comparisons.
344 if (SecondFalseElement != Overdefined) {
345 Idx = MaskIdx(Idx);
346 // None false -> true.
347 if (FirstFalseElement == Undefined)
348 return replaceInstUsesWith(ICI, Builder.getTrue());
350 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
352 // False for one element -> 'i != 47'.
353 if (SecondFalseElement == Undefined)
354 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
356 // False for two elements -> 'i != 47 & i != 72'.
357 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
358 Value *SecondFalseIdx =
359 ConstantInt::get(Idx->getType(), SecondFalseElement);
360 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
361 return BinaryOperator::CreateAnd(C1, C2);
364 // If the comparison can be replaced with a range comparison for the elements
365 // where it is true, emit the range check.
366 if (TrueRangeEnd != Overdefined) {
367 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
368 Idx = MaskIdx(Idx);
370 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
371 if (FirstTrueElement) {
372 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
373 Idx = Builder.CreateAdd(Idx, Offs);
376 Value *End =
377 ConstantInt::get(Idx->getType(), TrueRangeEnd - FirstTrueElement + 1);
378 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
381 // False range check.
382 if (FalseRangeEnd != Overdefined) {
383 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
384 Idx = MaskIdx(Idx);
385 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
386 if (FirstFalseElement) {
387 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
388 Idx = Builder.CreateAdd(Idx, Offs);
391 Value *End =
392 ConstantInt::get(Idx->getType(), FalseRangeEnd - FirstFalseElement);
393 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
396 // If a magic bitvector captures the entire comparison state
397 // of this load, replace it with computation that does:
398 // ((magic_cst >> i) & 1) != 0
400 Type *Ty = nullptr;
402 // Look for an appropriate type:
403 // - The type of Idx if the magic fits
404 // - The smallest fitting legal type
405 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
406 Ty = Idx->getType();
407 else
408 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
410 if (Ty) {
411 Idx = MaskIdx(Idx);
412 Value *V = Builder.CreateIntCast(Idx, Ty, false);
413 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
414 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
415 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
419 return nullptr;
422 /// Returns true if we can rewrite Start as a GEP with pointer Base
423 /// and some integer offset. The nodes that need to be re-written
424 /// for this transformation will be added to Explored.
425 static bool canRewriteGEPAsOffset(Value *Start, Value *Base, GEPNoWrapFlags &NW,
426 const DataLayout &DL,
427 SetVector<Value *> &Explored) {
428 SmallVector<Value *, 16> WorkList(1, Start);
429 Explored.insert(Base);
431 // The following traversal gives us an order which can be used
432 // when doing the final transformation. Since in the final
433 // transformation we create the PHI replacement instructions first,
434 // we don't have to get them in any particular order.
436 // However, for other instructions we will have to traverse the
437 // operands of an instruction first, which means that we have to
438 // do a post-order traversal.
439 while (!WorkList.empty()) {
440 SetVector<PHINode *> PHIs;
442 while (!WorkList.empty()) {
443 if (Explored.size() >= 100)
444 return false;
446 Value *V = WorkList.back();
448 if (Explored.contains(V)) {
449 WorkList.pop_back();
450 continue;
453 if (!isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
454 // We've found some value that we can't explore which is different from
455 // the base. Therefore we can't do this transformation.
456 return false;
458 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
459 // Only allow inbounds GEPs with at most one variable offset.
460 auto IsNonConst = [](Value *V) { return !isa<ConstantInt>(V); };
461 if (!GEP->isInBounds() || count_if(GEP->indices(), IsNonConst) > 1)
462 return false;
464 NW = NW.intersectForOffsetAdd(GEP->getNoWrapFlags());
465 if (!Explored.contains(GEP->getOperand(0)))
466 WorkList.push_back(GEP->getOperand(0));
469 if (WorkList.back() == V) {
470 WorkList.pop_back();
471 // We've finished visiting this node, mark it as such.
472 Explored.insert(V);
475 if (auto *PN = dyn_cast<PHINode>(V)) {
476 // We cannot transform PHIs on unsplittable basic blocks.
477 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
478 return false;
479 Explored.insert(PN);
480 PHIs.insert(PN);
484 // Explore the PHI nodes further.
485 for (auto *PN : PHIs)
486 for (Value *Op : PN->incoming_values())
487 if (!Explored.contains(Op))
488 WorkList.push_back(Op);
491 // Make sure that we can do this. Since we can't insert GEPs in a basic
492 // block before a PHI node, we can't easily do this transformation if
493 // we have PHI node users of transformed instructions.
494 for (Value *Val : Explored) {
495 for (Value *Use : Val->uses()) {
497 auto *PHI = dyn_cast<PHINode>(Use);
498 auto *Inst = dyn_cast<Instruction>(Val);
500 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
501 !Explored.contains(PHI))
502 continue;
504 if (PHI->getParent() == Inst->getParent())
505 return false;
508 return true;
511 // Sets the appropriate insert point on Builder where we can add
512 // a replacement Instruction for V (if that is possible).
513 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
514 bool Before = true) {
515 if (auto *PHI = dyn_cast<PHINode>(V)) {
516 BasicBlock *Parent = PHI->getParent();
517 Builder.SetInsertPoint(Parent, Parent->getFirstInsertionPt());
518 return;
520 if (auto *I = dyn_cast<Instruction>(V)) {
521 if (!Before)
522 I = &*std::next(I->getIterator());
523 Builder.SetInsertPoint(I);
524 return;
526 if (auto *A = dyn_cast<Argument>(V)) {
527 // Set the insertion point in the entry block.
528 BasicBlock &Entry = A->getParent()->getEntryBlock();
529 Builder.SetInsertPoint(&Entry, Entry.getFirstInsertionPt());
530 return;
532 // Otherwise, this is a constant and we don't need to set a new
533 // insertion point.
534 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
537 /// Returns a re-written value of Start as an indexed GEP using Base as a
538 /// pointer.
539 static Value *rewriteGEPAsOffset(Value *Start, Value *Base, GEPNoWrapFlags NW,
540 const DataLayout &DL,
541 SetVector<Value *> &Explored,
542 InstCombiner &IC) {
543 // Perform all the substitutions. This is a bit tricky because we can
544 // have cycles in our use-def chains.
545 // 1. Create the PHI nodes without any incoming values.
546 // 2. Create all the other values.
547 // 3. Add the edges for the PHI nodes.
548 // 4. Emit GEPs to get the original pointers.
549 // 5. Remove the original instructions.
550 Type *IndexType = IntegerType::get(
551 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
553 DenseMap<Value *, Value *> NewInsts;
554 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
556 // Create the new PHI nodes, without adding any incoming values.
557 for (Value *Val : Explored) {
558 if (Val == Base)
559 continue;
560 // Create empty phi nodes. This avoids cyclic dependencies when creating
561 // the remaining instructions.
562 if (auto *PHI = dyn_cast<PHINode>(Val))
563 NewInsts[PHI] =
564 PHINode::Create(IndexType, PHI->getNumIncomingValues(),
565 PHI->getName() + ".idx", PHI->getIterator());
567 IRBuilder<> Builder(Base->getContext());
569 // Create all the other instructions.
570 for (Value *Val : Explored) {
571 if (NewInsts.contains(Val))
572 continue;
574 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
575 setInsertionPoint(Builder, GEP);
576 Value *Op = NewInsts[GEP->getOperand(0)];
577 Value *OffsetV = emitGEPOffset(&Builder, DL, GEP);
578 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
579 NewInsts[GEP] = OffsetV;
580 else
581 NewInsts[GEP] = Builder.CreateAdd(
582 Op, OffsetV, GEP->getOperand(0)->getName() + ".add",
583 /*NUW=*/NW.hasNoUnsignedWrap(),
584 /*NSW=*/NW.hasNoUnsignedSignedWrap());
585 continue;
587 if (isa<PHINode>(Val))
588 continue;
590 llvm_unreachable("Unexpected instruction type");
593 // Add the incoming values to the PHI nodes.
594 for (Value *Val : Explored) {
595 if (Val == Base)
596 continue;
597 // All the instructions have been created, we can now add edges to the
598 // phi nodes.
599 if (auto *PHI = dyn_cast<PHINode>(Val)) {
600 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
601 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
602 Value *NewIncoming = PHI->getIncomingValue(I);
604 auto It = NewInsts.find(NewIncoming);
605 if (It != NewInsts.end())
606 NewIncoming = It->second;
608 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
613 for (Value *Val : Explored) {
614 if (Val == Base)
615 continue;
617 setInsertionPoint(Builder, Val, false);
618 // Create GEP for external users.
619 Value *NewVal = Builder.CreateGEP(Builder.getInt8Ty(), Base, NewInsts[Val],
620 Val->getName() + ".ptr", NW);
621 IC.replaceInstUsesWith(*cast<Instruction>(Val), NewVal);
622 // Add old instruction to worklist for DCE. We don't directly remove it
623 // here because the original compare is one of the users.
624 IC.addToWorklist(cast<Instruction>(Val));
627 return NewInsts[Start];
630 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
631 /// We can look through PHIs, GEPs and casts in order to determine a common base
632 /// between GEPLHS and RHS.
633 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
634 CmpPredicate Cond,
635 const DataLayout &DL,
636 InstCombiner &IC) {
637 // FIXME: Support vector of pointers.
638 if (GEPLHS->getType()->isVectorTy())
639 return nullptr;
641 if (!GEPLHS->hasAllConstantIndices())
642 return nullptr;
644 APInt Offset(DL.getIndexTypeSizeInBits(GEPLHS->getType()), 0);
645 Value *PtrBase =
646 GEPLHS->stripAndAccumulateConstantOffsets(DL, Offset,
647 /*AllowNonInbounds*/ false);
649 // Bail if we looked through addrspacecast.
650 if (PtrBase->getType() != GEPLHS->getType())
651 return nullptr;
653 // The set of nodes that will take part in this transformation.
654 SetVector<Value *> Nodes;
655 GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags();
656 if (!canRewriteGEPAsOffset(RHS, PtrBase, NW, DL, Nodes))
657 return nullptr;
659 // We know we can re-write this as
660 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
661 // Since we've only looked through inbouds GEPs we know that we
662 // can't have overflow on either side. We can therefore re-write
663 // this as:
664 // OFFSET1 cmp OFFSET2
665 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, NW, DL, Nodes, IC);
667 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
668 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
669 // offset. Since Index is the offset of LHS to the base pointer, we will now
670 // compare the offsets instead of comparing the pointers.
671 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
672 IC.Builder.getInt(Offset), NewRHS);
675 /// Fold comparisons between a GEP instruction and something else. At this point
676 /// we know that the GEP is on the LHS of the comparison.
677 Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
678 CmpPredicate Cond, Instruction &I) {
679 // Don't transform signed compares of GEPs into index compares. Even if the
680 // GEP is inbounds, the final add of the base pointer can have signed overflow
681 // and would change the result of the icmp.
682 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
683 // the maximum signed value for the pointer type.
684 if (ICmpInst::isSigned(Cond))
685 return nullptr;
687 // Look through bitcasts and addrspacecasts. We do not however want to remove
688 // 0 GEPs.
689 if (!isa<GetElementPtrInst>(RHS))
690 RHS = RHS->stripPointerCasts();
692 auto CanFold = [Cond](GEPNoWrapFlags NW) {
693 if (ICmpInst::isEquality(Cond))
694 return true;
696 // Unsigned predicates can be folded if the GEPs have *any* nowrap flags.
697 assert(ICmpInst::isUnsigned(Cond));
698 return NW != GEPNoWrapFlags::none();
701 auto NewICmp = [Cond](GEPNoWrapFlags NW, Value *Op1, Value *Op2) {
702 if (!NW.hasNoUnsignedWrap()) {
703 // Convert signed to unsigned comparison.
704 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Op1, Op2);
707 auto *I = new ICmpInst(Cond, Op1, Op2);
708 I->setSameSign(NW.hasNoUnsignedSignedWrap());
709 return I;
712 Value *PtrBase = GEPLHS->getOperand(0);
713 if (PtrBase == RHS && CanFold(GEPLHS->getNoWrapFlags())) {
714 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
715 Value *Offset = EmitGEPOffset(GEPLHS);
716 return NewICmp(GEPLHS->getNoWrapFlags(), Offset,
717 Constant::getNullValue(Offset->getType()));
720 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
721 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
722 !NullPointerIsDefined(I.getFunction(),
723 RHS->getType()->getPointerAddressSpace())) {
724 // For most address spaces, an allocation can't be placed at null, but null
725 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
726 // the only valid inbounds address derived from null, is null itself.
727 // Thus, we have four cases to consider:
728 // 1) Base == nullptr, Offset == 0 -> inbounds, null
729 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
730 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
731 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
733 // (Note if we're indexing a type of size 0, that simply collapses into one
734 // of the buckets above.)
736 // In general, we're allowed to make values less poison (i.e. remove
737 // sources of full UB), so in this case, we just select between the two
738 // non-poison cases (1 and 4 above).
740 // For vectors, we apply the same reasoning on a per-lane basis.
741 auto *Base = GEPLHS->getPointerOperand();
742 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
743 auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
744 Base = Builder.CreateVectorSplat(EC, Base);
746 return new ICmpInst(Cond, Base,
747 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
748 cast<Constant>(RHS), Base->getType()));
749 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
750 GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags() & GEPRHS->getNoWrapFlags();
752 // If the base pointers are different, but the indices are the same, just
753 // compare the base pointer.
754 if (PtrBase != GEPRHS->getOperand(0)) {
755 bool IndicesTheSame =
756 GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
757 GEPLHS->getPointerOperand()->getType() ==
758 GEPRHS->getPointerOperand()->getType() &&
759 GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType();
760 if (IndicesTheSame)
761 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
762 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
763 IndicesTheSame = false;
764 break;
767 // If all indices are the same, just compare the base pointers.
768 Type *BaseType = GEPLHS->getOperand(0)->getType();
769 if (IndicesTheSame &&
770 CmpInst::makeCmpResultType(BaseType) == I.getType() && CanFold(NW))
771 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
773 // If we're comparing GEPs with two base pointers that only differ in type
774 // and both GEPs have only constant indices or just one use, then fold
775 // the compare with the adjusted indices.
776 // FIXME: Support vector of pointers.
777 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
778 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
779 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
780 PtrBase->stripPointerCasts() ==
781 GEPRHS->getOperand(0)->stripPointerCasts() &&
782 !GEPLHS->getType()->isVectorTy()) {
783 Value *LOffset = EmitGEPOffset(GEPLHS);
784 Value *ROffset = EmitGEPOffset(GEPRHS);
786 // If we looked through an addrspacecast between different sized address
787 // spaces, the LHS and RHS pointers are different sized
788 // integers. Truncate to the smaller one.
789 Type *LHSIndexTy = LOffset->getType();
790 Type *RHSIndexTy = ROffset->getType();
791 if (LHSIndexTy != RHSIndexTy) {
792 if (LHSIndexTy->getPrimitiveSizeInBits().getFixedValue() <
793 RHSIndexTy->getPrimitiveSizeInBits().getFixedValue()) {
794 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
795 } else
796 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
799 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
800 LOffset, ROffset);
801 return replaceInstUsesWith(I, Cmp);
804 // Otherwise, the base pointers are different and the indices are
805 // different. Try convert this to an indexed compare by looking through
806 // PHIs/casts.
807 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL, *this);
810 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
811 GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType()) {
812 // If the GEPs only differ by one index, compare it.
813 unsigned NumDifferences = 0; // Keep track of # differences.
814 unsigned DiffOperand = 0; // The operand that differs.
815 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
816 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
817 Type *LHSType = GEPLHS->getOperand(i)->getType();
818 Type *RHSType = GEPRHS->getOperand(i)->getType();
819 // FIXME: Better support for vector of pointers.
820 if (LHSType->getPrimitiveSizeInBits() !=
821 RHSType->getPrimitiveSizeInBits() ||
822 (GEPLHS->getType()->isVectorTy() &&
823 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
824 // Irreconcilable differences.
825 NumDifferences = 2;
826 break;
829 if (NumDifferences++) break;
830 DiffOperand = i;
833 if (NumDifferences == 0) // SAME GEP?
834 return replaceInstUsesWith(I, // No comparison is needed here.
835 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
837 else if (NumDifferences == 1 && CanFold(NW)) {
838 Value *LHSV = GEPLHS->getOperand(DiffOperand);
839 Value *RHSV = GEPRHS->getOperand(DiffOperand);
840 return NewICmp(NW, LHSV, RHSV);
844 if (CanFold(NW)) {
845 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
846 Value *L = EmitGEPOffset(GEPLHS, /*RewriteGEP=*/true);
847 Value *R = EmitGEPOffset(GEPRHS, /*RewriteGEP=*/true);
848 return NewICmp(NW, L, R);
852 // Try convert this to an indexed compare by looking through PHIs/casts as a
853 // last resort.
854 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL, *this);
857 bool InstCombinerImpl::foldAllocaCmp(AllocaInst *Alloca) {
858 // It would be tempting to fold away comparisons between allocas and any
859 // pointer not based on that alloca (e.g. an argument). However, even
860 // though such pointers cannot alias, they can still compare equal.
862 // But LLVM doesn't specify where allocas get their memory, so if the alloca
863 // doesn't escape we can argue that it's impossible to guess its value, and we
864 // can therefore act as if any such guesses are wrong.
866 // However, we need to ensure that this folding is consistent: We can't fold
867 // one comparison to false, and then leave a different comparison against the
868 // same value alone (as it might evaluate to true at runtime, leading to a
869 // contradiction). As such, this code ensures that all comparisons are folded
870 // at the same time, and there are no other escapes.
872 struct CmpCaptureTracker : public CaptureTracker {
873 AllocaInst *Alloca;
874 bool Captured = false;
875 /// The value of the map is a bit mask of which icmp operands the alloca is
876 /// used in.
877 SmallMapVector<ICmpInst *, unsigned, 4> ICmps;
879 CmpCaptureTracker(AllocaInst *Alloca) : Alloca(Alloca) {}
881 void tooManyUses() override { Captured = true; }
883 bool captured(const Use *U) override {
884 auto *ICmp = dyn_cast<ICmpInst>(U->getUser());
885 // We need to check that U is based *only* on the alloca, and doesn't
886 // have other contributions from a select/phi operand.
887 // TODO: We could check whether getUnderlyingObjects() reduces to one
888 // object, which would allow looking through phi nodes.
889 if (ICmp && ICmp->isEquality() && getUnderlyingObject(*U) == Alloca) {
890 // Collect equality icmps of the alloca, and don't treat them as
891 // captures.
892 ICmps[ICmp] |= 1u << U->getOperandNo();
893 return false;
896 Captured = true;
897 return true;
901 CmpCaptureTracker Tracker(Alloca);
902 PointerMayBeCaptured(Alloca, &Tracker);
903 if (Tracker.Captured)
904 return false;
906 bool Changed = false;
907 for (auto [ICmp, Operands] : Tracker.ICmps) {
908 switch (Operands) {
909 case 1:
910 case 2: {
911 // The alloca is only used in one icmp operand. Assume that the
912 // equality is false.
913 auto *Res = ConstantInt::get(
914 ICmp->getType(), ICmp->getPredicate() == ICmpInst::ICMP_NE);
915 replaceInstUsesWith(*ICmp, Res);
916 eraseInstFromFunction(*ICmp);
917 Changed = true;
918 break;
920 case 3:
921 // Both icmp operands are based on the alloca, so this is comparing
922 // pointer offsets, without leaking any information about the address
923 // of the alloca. Ignore such comparisons.
924 break;
925 default:
926 llvm_unreachable("Cannot happen");
930 return Changed;
933 /// Fold "icmp pred (X+C), X".
934 Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
935 CmpPredicate Pred) {
936 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
937 // so the values can never be equal. Similarly for all other "or equals"
938 // operators.
939 assert(!!C && "C should not be zero!");
941 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
942 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
943 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
944 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
945 Constant *R = ConstantInt::get(X->getType(),
946 APInt::getMaxValue(C.getBitWidth()) - C);
947 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
950 // (X+1) >u X --> X <u (0-1) --> X != 255
951 // (X+2) >u X --> X <u (0-2) --> X <u 254
952 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
953 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
954 return new ICmpInst(ICmpInst::ICMP_ULT, X,
955 ConstantInt::get(X->getType(), -C));
957 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
959 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
960 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
961 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
962 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
963 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
964 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
965 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
966 return new ICmpInst(ICmpInst::ICMP_SGT, X,
967 ConstantInt::get(X->getType(), SMax - C));
969 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
970 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
971 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
972 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
973 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
974 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
976 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
977 return new ICmpInst(ICmpInst::ICMP_SLT, X,
978 ConstantInt::get(X->getType(), SMax - (C - 1)));
981 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
982 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
983 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
984 Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
985 const APInt &AP1,
986 const APInt &AP2) {
987 assert(I.isEquality() && "Cannot fold icmp gt/lt");
989 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
990 if (I.getPredicate() == I.ICMP_NE)
991 Pred = CmpInst::getInversePredicate(Pred);
992 return new ICmpInst(Pred, LHS, RHS);
995 // Don't bother doing any work for cases which InstSimplify handles.
996 if (AP2.isZero())
997 return nullptr;
999 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1000 if (IsAShr) {
1001 if (AP2.isAllOnes())
1002 return nullptr;
1003 if (AP2.isNegative() != AP1.isNegative())
1004 return nullptr;
1005 if (AP2.sgt(AP1))
1006 return nullptr;
1009 if (!AP1)
1010 // 'A' must be large enough to shift out the highest set bit.
1011 return getICmp(I.ICMP_UGT, A,
1012 ConstantInt::get(A->getType(), AP2.logBase2()));
1014 if (AP1 == AP2)
1015 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1017 int Shift;
1018 if (IsAShr && AP1.isNegative())
1019 Shift = AP1.countl_one() - AP2.countl_one();
1020 else
1021 Shift = AP1.countl_zero() - AP2.countl_zero();
1023 if (Shift > 0) {
1024 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1025 // There are multiple solutions if we are comparing against -1 and the LHS
1026 // of the ashr is not a power of two.
1027 if (AP1.isAllOnes() && !AP2.isPowerOf2())
1028 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1029 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1030 } else if (AP1 == AP2.lshr(Shift)) {
1031 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1035 // Shifting const2 will never be equal to const1.
1036 // FIXME: This should always be handled by InstSimplify?
1037 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1038 return replaceInstUsesWith(I, TorF);
1041 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1042 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1043 Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1044 const APInt &AP1,
1045 const APInt &AP2) {
1046 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1048 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1049 if (I.getPredicate() == I.ICMP_NE)
1050 Pred = CmpInst::getInversePredicate(Pred);
1051 return new ICmpInst(Pred, LHS, RHS);
1054 // Don't bother doing any work for cases which InstSimplify handles.
1055 if (AP2.isZero())
1056 return nullptr;
1058 unsigned AP2TrailingZeros = AP2.countr_zero();
1060 if (!AP1 && AP2TrailingZeros != 0)
1061 return getICmp(
1062 I.ICMP_UGE, A,
1063 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1065 if (AP1 == AP2)
1066 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1068 // Get the distance between the lowest bits that are set.
1069 int Shift = AP1.countr_zero() - AP2TrailingZeros;
1071 if (Shift > 0 && AP2.shl(Shift) == AP1)
1072 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1074 // Shifting const2 will never be equal to const1.
1075 // FIXME: This should always be handled by InstSimplify?
1076 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1077 return replaceInstUsesWith(I, TorF);
1080 /// The caller has matched a pattern of the form:
1081 /// I = icmp ugt (add (add A, B), CI2), CI1
1082 /// If this is of the form:
1083 /// sum = a + b
1084 /// if (sum+128 >u 255)
1085 /// Then replace it with llvm.sadd.with.overflow.i8.
1087 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1088 ConstantInt *CI2, ConstantInt *CI1,
1089 InstCombinerImpl &IC) {
1090 // The transformation we're trying to do here is to transform this into an
1091 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1092 // with a narrower add, and discard the add-with-constant that is part of the
1093 // range check (if we can't eliminate it, this isn't profitable).
1095 // In order to eliminate the add-with-constant, the compare can be its only
1096 // use.
1097 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1098 if (!AddWithCst->hasOneUse())
1099 return nullptr;
1101 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1102 if (!CI2->getValue().isPowerOf2())
1103 return nullptr;
1104 unsigned NewWidth = CI2->getValue().countr_zero();
1105 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1106 return nullptr;
1108 // The width of the new add formed is 1 more than the bias.
1109 ++NewWidth;
1111 // Check to see that CI1 is an all-ones value with NewWidth bits.
1112 if (CI1->getBitWidth() == NewWidth ||
1113 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1114 return nullptr;
1116 // This is only really a signed overflow check if the inputs have been
1117 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1118 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1119 if (IC.ComputeMaxSignificantBits(A, 0, &I) > NewWidth ||
1120 IC.ComputeMaxSignificantBits(B, 0, &I) > NewWidth)
1121 return nullptr;
1123 // In order to replace the original add with a narrower
1124 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1125 // and truncates that discard the high bits of the add. Verify that this is
1126 // the case.
1127 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1128 for (User *U : OrigAdd->users()) {
1129 if (U == AddWithCst)
1130 continue;
1132 // Only accept truncates for now. We would really like a nice recursive
1133 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1134 // chain to see which bits of a value are actually demanded. If the
1135 // original add had another add which was then immediately truncated, we
1136 // could still do the transformation.
1137 TruncInst *TI = dyn_cast<TruncInst>(U);
1138 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1139 return nullptr;
1142 // If the pattern matches, truncate the inputs to the narrower type and
1143 // use the sadd_with_overflow intrinsic to efficiently compute both the
1144 // result and the overflow bit.
1145 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1146 Function *F = Intrinsic::getOrInsertDeclaration(
1147 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1149 InstCombiner::BuilderTy &Builder = IC.Builder;
1151 // Put the new code above the original add, in case there are any uses of the
1152 // add between the add and the compare.
1153 Builder.SetInsertPoint(OrigAdd);
1155 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1156 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1157 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1158 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1159 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1161 // The inner add was the result of the narrow add, zero extended to the
1162 // wider type. Replace it with the result computed by the intrinsic.
1163 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1164 IC.eraseInstFromFunction(*OrigAdd);
1166 // The original icmp gets replaced with the overflow value.
1167 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1170 /// If we have:
1171 /// icmp eq/ne (urem/srem %x, %y), 0
1172 /// iff %y is a power-of-two, we can replace this with a bit test:
1173 /// icmp eq/ne (and %x, (add %y, -1)), 0
1174 Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1175 // This fold is only valid for equality predicates.
1176 if (!I.isEquality())
1177 return nullptr;
1178 CmpPredicate Pred;
1179 Value *X, *Y, *Zero;
1180 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1181 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1182 return nullptr;
1183 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1184 return nullptr;
1185 // This may increase instruction count, we don't enforce that Y is a constant.
1186 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1187 Value *Masked = Builder.CreateAnd(X, Mask);
1188 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1191 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1192 /// by one-less-than-bitwidth into a sign test on the original value.
1193 Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1194 Instruction *Val;
1195 CmpPredicate Pred;
1196 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1197 return nullptr;
1199 Value *X;
1200 Type *XTy;
1202 Constant *C;
1203 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1204 XTy = X->getType();
1205 unsigned XBitWidth = XTy->getScalarSizeInBits();
1206 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1207 APInt(XBitWidth, XBitWidth - 1))))
1208 return nullptr;
1209 } else if (isa<BinaryOperator>(Val) &&
1210 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1211 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1212 /*AnalyzeForSignBitExtraction=*/true))) {
1213 XTy = X->getType();
1214 } else
1215 return nullptr;
1217 return ICmpInst::Create(Instruction::ICmp,
1218 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1219 : ICmpInst::ICMP_SLT,
1220 X, ConstantInt::getNullValue(XTy));
1223 // Handle icmp pred X, 0
1224 Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1225 CmpInst::Predicate Pred = Cmp.getPredicate();
1226 if (!match(Cmp.getOperand(1), m_Zero()))
1227 return nullptr;
1229 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1230 if (Pred == ICmpInst::ICMP_SGT) {
1231 Value *A, *B;
1232 if (match(Cmp.getOperand(0), m_SMin(m_Value(A), m_Value(B)))) {
1233 if (isKnownPositive(A, SQ.getWithInstruction(&Cmp)))
1234 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1235 if (isKnownPositive(B, SQ.getWithInstruction(&Cmp)))
1236 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1240 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1241 return New;
1243 // Given:
1244 // icmp eq/ne (urem %x, %y), 0
1245 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1246 // icmp eq/ne %x, 0
1247 Value *X, *Y;
1248 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1249 ICmpInst::isEquality(Pred)) {
1250 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1251 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1252 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1253 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1256 // (icmp eq/ne (mul X Y)) -> (icmp eq/ne X/Y) if we know about whether X/Y are
1257 // odd/non-zero/there is no overflow.
1258 if (match(Cmp.getOperand(0), m_Mul(m_Value(X), m_Value(Y))) &&
1259 ICmpInst::isEquality(Pred)) {
1261 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1262 // if X % 2 != 0
1263 // (icmp eq/ne Y)
1264 if (XKnown.countMaxTrailingZeros() == 0)
1265 return new ICmpInst(Pred, Y, Cmp.getOperand(1));
1267 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1268 // if Y % 2 != 0
1269 // (icmp eq/ne X)
1270 if (YKnown.countMaxTrailingZeros() == 0)
1271 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1273 auto *BO0 = cast<OverflowingBinaryOperator>(Cmp.getOperand(0));
1274 if (BO0->hasNoUnsignedWrap() || BO0->hasNoSignedWrap()) {
1275 const SimplifyQuery Q = SQ.getWithInstruction(&Cmp);
1276 // `isKnownNonZero` does more analysis than just `!KnownBits.One.isZero()`
1277 // but to avoid unnecessary work, first just if this is an obvious case.
1279 // if X non-zero and NoOverflow(X * Y)
1280 // (icmp eq/ne Y)
1281 if (!XKnown.One.isZero() || isKnownNonZero(X, Q))
1282 return new ICmpInst(Pred, Y, Cmp.getOperand(1));
1284 // if Y non-zero and NoOverflow(X * Y)
1285 // (icmp eq/ne X)
1286 if (!YKnown.One.isZero() || isKnownNonZero(Y, Q))
1287 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1289 // Note, we are skipping cases:
1290 // if Y % 2 != 0 AND X % 2 != 0
1291 // (false/true)
1292 // if X non-zero and Y non-zero and NoOverflow(X * Y)
1293 // (false/true)
1294 // Those can be simplified later as we would have already replaced the (icmp
1295 // eq/ne (mul X, Y)) with (icmp eq/ne X/Y) and if X/Y is known non-zero that
1296 // will fold to a constant elsewhere.
1298 return nullptr;
1301 /// Fold icmp Pred X, C.
1302 /// TODO: This code structure does not make sense. The saturating add fold
1303 /// should be moved to some other helper and extended as noted below (it is also
1304 /// possible that code has been made unnecessary - do we canonicalize IR to
1305 /// overflow/saturating intrinsics or not?).
1306 Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1307 // Match the following pattern, which is a common idiom when writing
1308 // overflow-safe integer arithmetic functions. The source performs an addition
1309 // in wider type and explicitly checks for overflow using comparisons against
1310 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1312 // TODO: This could probably be generalized to handle other overflow-safe
1313 // operations if we worked out the formulas to compute the appropriate magic
1314 // constants.
1316 // sum = a + b
1317 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1318 CmpInst::Predicate Pred = Cmp.getPredicate();
1319 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1320 Value *A, *B;
1321 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1322 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1323 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1324 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1325 return Res;
1327 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1328 Constant *C = dyn_cast<Constant>(Op1);
1329 if (!C)
1330 return nullptr;
1332 if (auto *Phi = dyn_cast<PHINode>(Op0))
1333 if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1334 SmallVector<Constant *> Ops;
1335 for (Value *V : Phi->incoming_values()) {
1336 Constant *Res =
1337 ConstantFoldCompareInstOperands(Pred, cast<Constant>(V), C, DL);
1338 if (!Res)
1339 return nullptr;
1340 Ops.push_back(Res);
1342 Builder.SetInsertPoint(Phi);
1343 PHINode *NewPhi = Builder.CreatePHI(Cmp.getType(), Phi->getNumOperands());
1344 for (auto [V, Pred] : zip(Ops, Phi->blocks()))
1345 NewPhi->addIncoming(V, Pred);
1346 return replaceInstUsesWith(Cmp, NewPhi);
1349 if (Instruction *R = tryFoldInstWithCtpopWithNot(&Cmp))
1350 return R;
1352 return nullptr;
1355 /// Canonicalize icmp instructions based on dominating conditions.
1356 Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1357 // We already checked simple implication in InstSimplify, only handle complex
1358 // cases here.
1359 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1360 const APInt *C;
1361 if (!match(Y, m_APInt(C)))
1362 return nullptr;
1364 CmpInst::Predicate Pred = Cmp.getPredicate();
1365 ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred, *C);
1367 auto handleDomCond = [&](ICmpInst::Predicate DomPred,
1368 const APInt *DomC) -> Instruction * {
1369 // We have 2 compares of a variable with constants. Calculate the constant
1370 // ranges of those compares to see if we can transform the 2nd compare:
1371 // DomBB:
1372 // DomCond = icmp DomPred X, DomC
1373 // br DomCond, CmpBB, FalseBB
1374 // CmpBB:
1375 // Cmp = icmp Pred X, C
1376 ConstantRange DominatingCR =
1377 ConstantRange::makeExactICmpRegion(DomPred, *DomC);
1378 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1379 ConstantRange Difference = DominatingCR.difference(CR);
1380 if (Intersection.isEmptySet())
1381 return replaceInstUsesWith(Cmp, Builder.getFalse());
1382 if (Difference.isEmptySet())
1383 return replaceInstUsesWith(Cmp, Builder.getTrue());
1385 // Canonicalizing a sign bit comparison that gets used in a branch,
1386 // pessimizes codegen by generating branch on zero instruction instead
1387 // of a test and branch. So we avoid canonicalizing in such situations
1388 // because test and branch instruction has better branch displacement
1389 // than compare and branch instruction.
1390 bool UnusedBit;
1391 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1392 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1393 return nullptr;
1395 // Avoid an infinite loop with min/max canonicalization.
1396 // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1397 if (Cmp.hasOneUse() &&
1398 match(Cmp.user_back(), m_MaxOrMin(m_Value(), m_Value())))
1399 return nullptr;
1401 if (const APInt *EqC = Intersection.getSingleElement())
1402 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1403 if (const APInt *NeC = Difference.getSingleElement())
1404 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1405 return nullptr;
1408 for (BranchInst *BI : DC.conditionsFor(X)) {
1409 CmpPredicate DomPred;
1410 const APInt *DomC;
1411 if (!match(BI->getCondition(),
1412 m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))))
1413 continue;
1415 BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(0));
1416 if (DT.dominates(Edge0, Cmp.getParent())) {
1417 if (auto *V = handleDomCond(DomPred, DomC))
1418 return V;
1419 } else {
1420 BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(1));
1421 if (DT.dominates(Edge1, Cmp.getParent()))
1422 if (auto *V =
1423 handleDomCond(CmpInst::getInversePredicate(DomPred), DomC))
1424 return V;
1428 return nullptr;
1431 /// Fold icmp (trunc X), C.
1432 Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1433 TruncInst *Trunc,
1434 const APInt &C) {
1435 ICmpInst::Predicate Pred = Cmp.getPredicate();
1436 Value *X = Trunc->getOperand(0);
1437 Type *SrcTy = X->getType();
1438 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1439 SrcBits = SrcTy->getScalarSizeInBits();
1441 // Match (icmp pred (trunc nuw/nsw X), C)
1442 // Which we can convert to (icmp pred X, (sext/zext C))
1443 if (shouldChangeType(Trunc->getType(), SrcTy)) {
1444 if (Trunc->hasNoSignedWrap())
1445 return new ICmpInst(Pred, X, ConstantInt::get(SrcTy, C.sext(SrcBits)));
1446 if (!Cmp.isSigned() && Trunc->hasNoUnsignedWrap())
1447 return new ICmpInst(Pred, X, ConstantInt::get(SrcTy, C.zext(SrcBits)));
1450 if (C.isOne() && C.getBitWidth() > 1) {
1451 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1452 Value *V = nullptr;
1453 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1454 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1455 ConstantInt::get(V->getType(), 1));
1458 // TODO: Handle any shifted constant by subtracting trailing zeros.
1459 // TODO: Handle non-equality predicates.
1460 Value *Y;
1461 if (Cmp.isEquality() && match(X, m_Shl(m_One(), m_Value(Y)))) {
1462 // (trunc (1 << Y) to iN) == 0 --> Y u>= N
1463 // (trunc (1 << Y) to iN) != 0 --> Y u< N
1464 if (C.isZero()) {
1465 auto NewPred = (Pred == Cmp.ICMP_EQ) ? Cmp.ICMP_UGE : Cmp.ICMP_ULT;
1466 return new ICmpInst(NewPred, Y, ConstantInt::get(SrcTy, DstBits));
1468 // (trunc (1 << Y) to iN) == 2**C --> Y == C
1469 // (trunc (1 << Y) to iN) != 2**C --> Y != C
1470 if (C.isPowerOf2())
1471 return new ICmpInst(Pred, Y, ConstantInt::get(SrcTy, C.logBase2()));
1474 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1475 // Canonicalize to a mask and wider compare if the wide type is suitable:
1476 // (trunc X to i8) == C --> (X & 0xff) == (zext C)
1477 if (!SrcTy->isVectorTy() && shouldChangeType(DstBits, SrcBits)) {
1478 Constant *Mask =
1479 ConstantInt::get(SrcTy, APInt::getLowBitsSet(SrcBits, DstBits));
1480 Value *And = Builder.CreateAnd(X, Mask);
1481 Constant *WideC = ConstantInt::get(SrcTy, C.zext(SrcBits));
1482 return new ICmpInst(Pred, And, WideC);
1485 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1486 // of the high bits truncated out of x are known.
1487 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1489 // If all the high bits are known, we can do this xform.
1490 if ((Known.Zero | Known.One).countl_one() >= SrcBits - DstBits) {
1491 // Pull in the high bits from known-ones set.
1492 APInt NewRHS = C.zext(SrcBits);
1493 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1494 return new ICmpInst(Pred, X, ConstantInt::get(SrcTy, NewRHS));
1498 // Look through truncated right-shift of the sign-bit for a sign-bit check:
1499 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1500 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1501 Value *ShOp;
1502 const APInt *ShAmtC;
1503 bool TrueIfSigned;
1504 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1505 match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1506 DstBits == SrcBits - ShAmtC->getZExtValue()) {
1507 return TrueIfSigned ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1508 ConstantInt::getNullValue(SrcTy))
1509 : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1510 ConstantInt::getAllOnesValue(SrcTy));
1513 return nullptr;
1516 /// Fold icmp (trunc nuw/nsw X), (trunc nuw/nsw Y).
1517 /// Fold icmp (trunc nuw/nsw X), (zext/sext Y).
1518 Instruction *
1519 InstCombinerImpl::foldICmpTruncWithTruncOrExt(ICmpInst &Cmp,
1520 const SimplifyQuery &Q) {
1521 Value *X, *Y;
1522 CmpPredicate Pred;
1523 bool YIsSExt = false;
1524 // Try to match icmp (trunc X), (trunc Y)
1525 if (match(&Cmp, m_ICmp(Pred, m_Trunc(m_Value(X)), m_Trunc(m_Value(Y))))) {
1526 unsigned NoWrapFlags = cast<TruncInst>(Cmp.getOperand(0))->getNoWrapKind() &
1527 cast<TruncInst>(Cmp.getOperand(1))->getNoWrapKind();
1528 if (Cmp.isSigned()) {
1529 // For signed comparisons, both truncs must be nsw.
1530 if (!(NoWrapFlags & TruncInst::NoSignedWrap))
1531 return nullptr;
1532 } else {
1533 // For unsigned and equality comparisons, either both must be nuw or
1534 // both must be nsw, we don't care which.
1535 if (!NoWrapFlags)
1536 return nullptr;
1539 if (X->getType() != Y->getType() &&
1540 (!Cmp.getOperand(0)->hasOneUse() || !Cmp.getOperand(1)->hasOneUse()))
1541 return nullptr;
1542 if (!isDesirableIntType(X->getType()->getScalarSizeInBits()) &&
1543 isDesirableIntType(Y->getType()->getScalarSizeInBits())) {
1544 std::swap(X, Y);
1545 Pred = Cmp.getSwappedPredicate(Pred);
1547 YIsSExt = !(NoWrapFlags & TruncInst::NoUnsignedWrap);
1549 // Try to match icmp (trunc nuw X), (zext Y)
1550 else if (!Cmp.isSigned() &&
1551 match(&Cmp, m_c_ICmp(Pred, m_NUWTrunc(m_Value(X)),
1552 m_OneUse(m_ZExt(m_Value(Y)))))) {
1553 // Can fold trunc nuw + zext for unsigned and equality predicates.
1555 // Try to match icmp (trunc nsw X), (sext Y)
1556 else if (match(&Cmp, m_c_ICmp(Pred, m_NSWTrunc(m_Value(X)),
1557 m_OneUse(m_ZExtOrSExt(m_Value(Y)))))) {
1558 // Can fold trunc nsw + zext/sext for all predicates.
1559 YIsSExt =
1560 isa<SExtInst>(Cmp.getOperand(0)) || isa<SExtInst>(Cmp.getOperand(1));
1561 } else
1562 return nullptr;
1564 Type *TruncTy = Cmp.getOperand(0)->getType();
1565 unsigned TruncBits = TruncTy->getScalarSizeInBits();
1567 // If this transform will end up changing from desirable types -> undesirable
1568 // types skip it.
1569 if (isDesirableIntType(TruncBits) &&
1570 !isDesirableIntType(X->getType()->getScalarSizeInBits()))
1571 return nullptr;
1573 Value *NewY = Builder.CreateIntCast(Y, X->getType(), YIsSExt);
1574 return new ICmpInst(Pred, X, NewY);
1577 /// Fold icmp (xor X, Y), C.
1578 Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1579 BinaryOperator *Xor,
1580 const APInt &C) {
1581 if (Instruction *I = foldICmpXorShiftConst(Cmp, Xor, C))
1582 return I;
1584 Value *X = Xor->getOperand(0);
1585 Value *Y = Xor->getOperand(1);
1586 const APInt *XorC;
1587 if (!match(Y, m_APInt(XorC)))
1588 return nullptr;
1590 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1591 // fold the xor.
1592 ICmpInst::Predicate Pred = Cmp.getPredicate();
1593 bool TrueIfSigned = false;
1594 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1596 // If the sign bit of the XorCst is not set, there is no change to
1597 // the operation, just stop using the Xor.
1598 if (!XorC->isNegative())
1599 return replaceOperand(Cmp, 0, X);
1601 // Emit the opposite comparison.
1602 if (TrueIfSigned)
1603 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1604 ConstantInt::getAllOnesValue(X->getType()));
1605 else
1606 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1607 ConstantInt::getNullValue(X->getType()));
1610 if (Xor->hasOneUse()) {
1611 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1612 if (!Cmp.isEquality() && XorC->isSignMask()) {
1613 Pred = Cmp.getFlippedSignednessPredicate();
1614 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1617 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1618 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1619 Pred = Cmp.getFlippedSignednessPredicate();
1620 Pred = Cmp.getSwappedPredicate(Pred);
1621 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1625 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1626 if (Pred == ICmpInst::ICMP_UGT) {
1627 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1628 if (*XorC == ~C && (C + 1).isPowerOf2())
1629 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1630 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1631 if (*XorC == C && (C + 1).isPowerOf2())
1632 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1634 if (Pred == ICmpInst::ICMP_ULT) {
1635 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1636 if (*XorC == -C && C.isPowerOf2())
1637 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1638 ConstantInt::get(X->getType(), ~C));
1639 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1640 if (*XorC == C && (-C).isPowerOf2())
1641 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1642 ConstantInt::get(X->getType(), ~C));
1644 return nullptr;
1647 /// For power-of-2 C:
1648 /// ((X s>> ShiftC) ^ X) u< C --> (X + C) u< (C << 1)
1649 /// ((X s>> ShiftC) ^ X) u> (C - 1) --> (X + C) u> ((C << 1) - 1)
1650 Instruction *InstCombinerImpl::foldICmpXorShiftConst(ICmpInst &Cmp,
1651 BinaryOperator *Xor,
1652 const APInt &C) {
1653 CmpInst::Predicate Pred = Cmp.getPredicate();
1654 APInt PowerOf2;
1655 if (Pred == ICmpInst::ICMP_ULT)
1656 PowerOf2 = C;
1657 else if (Pred == ICmpInst::ICMP_UGT && !C.isMaxValue())
1658 PowerOf2 = C + 1;
1659 else
1660 return nullptr;
1661 if (!PowerOf2.isPowerOf2())
1662 return nullptr;
1663 Value *X;
1664 const APInt *ShiftC;
1665 if (!match(Xor, m_OneUse(m_c_Xor(m_Value(X),
1666 m_AShr(m_Deferred(X), m_APInt(ShiftC))))))
1667 return nullptr;
1668 uint64_t Shift = ShiftC->getLimitedValue();
1669 Type *XType = X->getType();
1670 if (Shift == 0 || PowerOf2.isMinSignedValue())
1671 return nullptr;
1672 Value *Add = Builder.CreateAdd(X, ConstantInt::get(XType, PowerOf2));
1673 APInt Bound =
1674 Pred == ICmpInst::ICMP_ULT ? PowerOf2 << 1 : ((PowerOf2 << 1) - 1);
1675 return new ICmpInst(Pred, Add, ConstantInt::get(XType, Bound));
1678 /// Fold icmp (and (sh X, Y), C2), C1.
1679 Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1680 BinaryOperator *And,
1681 const APInt &C1,
1682 const APInt &C2) {
1683 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1684 if (!Shift || !Shift->isShift())
1685 return nullptr;
1687 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1688 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1689 // code produced by the clang front-end, for bitfield access.
1690 // This seemingly simple opportunity to fold away a shift turns out to be
1691 // rather complicated. See PR17827 for details.
1692 unsigned ShiftOpcode = Shift->getOpcode();
1693 bool IsShl = ShiftOpcode == Instruction::Shl;
1694 const APInt *C3;
1695 if (match(Shift->getOperand(1), m_APInt(C3))) {
1696 APInt NewAndCst, NewCmpCst;
1697 bool AnyCmpCstBitsShiftedOut;
1698 if (ShiftOpcode == Instruction::Shl) {
1699 // For a left shift, we can fold if the comparison is not signed. We can
1700 // also fold a signed comparison if the mask value and comparison value
1701 // are not negative. These constraints may not be obvious, but we can
1702 // prove that they are correct using an SMT solver.
1703 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1704 return nullptr;
1706 NewCmpCst = C1.lshr(*C3);
1707 NewAndCst = C2.lshr(*C3);
1708 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1709 } else if (ShiftOpcode == Instruction::LShr) {
1710 // For a logical right shift, we can fold if the comparison is not signed.
1711 // We can also fold a signed comparison if the shifted mask value and the
1712 // shifted comparison value are not negative. These constraints may not be
1713 // obvious, but we can prove that they are correct using an SMT solver.
1714 NewCmpCst = C1.shl(*C3);
1715 NewAndCst = C2.shl(*C3);
1716 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1717 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1718 return nullptr;
1719 } else {
1720 // For an arithmetic shift, check that both constants don't use (in a
1721 // signed sense) the top bits being shifted out.
1722 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1723 NewCmpCst = C1.shl(*C3);
1724 NewAndCst = C2.shl(*C3);
1725 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1726 if (NewAndCst.ashr(*C3) != C2)
1727 return nullptr;
1730 if (AnyCmpCstBitsShiftedOut) {
1731 // If we shifted bits out, the fold is not going to work out. As a
1732 // special case, check to see if this means that the result is always
1733 // true or false now.
1734 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1735 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1736 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1737 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1738 } else {
1739 Value *NewAnd = Builder.CreateAnd(
1740 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1741 return new ICmpInst(Cmp.getPredicate(),
1742 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1746 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1747 // preferable because it allows the C2 << Y expression to be hoisted out of a
1748 // loop if Y is invariant and X is not.
1749 if (Shift->hasOneUse() && C1.isZero() && Cmp.isEquality() &&
1750 !Shift->isArithmeticShift() &&
1751 ((!IsShl && C2.isOne()) || !isa<Constant>(Shift->getOperand(0)))) {
1752 // Compute C2 << Y.
1753 Value *NewShift =
1754 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1755 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1757 // Compute X & (C2 << Y).
1758 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1759 return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(1));
1762 return nullptr;
1765 /// Fold icmp (and X, C2), C1.
1766 Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1767 BinaryOperator *And,
1768 const APInt &C1) {
1769 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1771 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1772 // TODO: We canonicalize to the longer form for scalars because we have
1773 // better analysis/folds for icmp, and codegen may be better with icmp.
1774 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isZero() &&
1775 match(And->getOperand(1), m_One()))
1776 return new TruncInst(And->getOperand(0), Cmp.getType());
1778 const APInt *C2;
1779 Value *X;
1780 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1781 return nullptr;
1783 // (and X, highmask) s> [0, ~highmask] --> X s> ~highmask
1784 if (Cmp.getPredicate() == ICmpInst::ICMP_SGT && C1.ule(~*C2) &&
1785 C2->isNegatedPowerOf2())
1786 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1787 ConstantInt::get(X->getType(), ~*C2));
1788 // (and X, highmask) s< [1, -highmask] --> X s< -highmask
1789 if (Cmp.getPredicate() == ICmpInst::ICMP_SLT && !C1.isSignMask() &&
1790 (C1 - 1).ule(~*C2) && C2->isNegatedPowerOf2() && !C2->isSignMask())
1791 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1792 ConstantInt::get(X->getType(), -*C2));
1794 // Don't perform the following transforms if the AND has multiple uses
1795 if (!And->hasOneUse())
1796 return nullptr;
1798 if (Cmp.isEquality() && C1.isZero()) {
1799 // Restrict this fold to single-use 'and' (PR10267).
1800 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1801 if (C2->isSignMask()) {
1802 Constant *Zero = Constant::getNullValue(X->getType());
1803 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1804 return new ICmpInst(NewPred, X, Zero);
1807 APInt NewC2 = *C2;
1808 KnownBits Know = computeKnownBits(And->getOperand(0), 0, And);
1809 // Set high zeros of C2 to allow matching negated power-of-2.
1810 NewC2 = *C2 | APInt::getHighBitsSet(C2->getBitWidth(),
1811 Know.countMinLeadingZeros());
1813 // Restrict this fold only for single-use 'and' (PR10267).
1814 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1815 if (NewC2.isNegatedPowerOf2()) {
1816 Constant *NegBOC = ConstantInt::get(And->getType(), -NewC2);
1817 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1818 return new ICmpInst(NewPred, X, NegBOC);
1822 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1823 // the input width without changing the value produced, eliminate the cast:
1825 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1827 // We can do this transformation if the constants do not have their sign bits
1828 // set or if it is an equality comparison. Extending a relational comparison
1829 // when we're checking the sign bit would not work.
1830 Value *W;
1831 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1832 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1833 // TODO: Is this a good transform for vectors? Wider types may reduce
1834 // throughput. Should this transform be limited (even for scalars) by using
1835 // shouldChangeType()?
1836 if (!Cmp.getType()->isVectorTy()) {
1837 Type *WideType = W->getType();
1838 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1839 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1840 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1841 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1842 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1846 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1847 return I;
1849 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1850 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1852 // iff pred isn't signed
1853 if (!Cmp.isSigned() && C1.isZero() && And->getOperand(0)->hasOneUse() &&
1854 match(And->getOperand(1), m_One())) {
1855 Constant *One = cast<Constant>(And->getOperand(1));
1856 Value *Or = And->getOperand(0);
1857 Value *A, *B, *LShr;
1858 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1859 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1860 unsigned UsesRemoved = 0;
1861 if (And->hasOneUse())
1862 ++UsesRemoved;
1863 if (Or->hasOneUse())
1864 ++UsesRemoved;
1865 if (LShr->hasOneUse())
1866 ++UsesRemoved;
1868 // Compute A & ((1 << B) | 1)
1869 unsigned RequireUsesRemoved = match(B, m_ImmConstant()) ? 1 : 3;
1870 if (UsesRemoved >= RequireUsesRemoved) {
1871 Value *NewOr =
1872 Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1873 /*HasNUW=*/true),
1874 One, Or->getName());
1875 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1876 return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(1));
1881 // (icmp eq (and (bitcast X to int), ExponentMask), ExponentMask) -->
1882 // llvm.is.fpclass(X, fcInf|fcNan)
1883 // (icmp ne (and (bitcast X to int), ExponentMask), ExponentMask) -->
1884 // llvm.is.fpclass(X, ~(fcInf|fcNan))
1885 Value *V;
1886 if (!Cmp.getParent()->getParent()->hasFnAttribute(
1887 Attribute::NoImplicitFloat) &&
1888 Cmp.isEquality() &&
1889 match(X, m_OneUse(m_ElementWiseBitCast(m_Value(V))))) {
1890 Type *FPType = V->getType()->getScalarType();
1891 if (FPType->isIEEELikeFPTy() && C1 == *C2) {
1892 APInt ExponentMask =
1893 APFloat::getInf(FPType->getFltSemantics()).bitcastToAPInt();
1894 if (C1 == ExponentMask) {
1895 unsigned Mask = FPClassTest::fcNan | FPClassTest::fcInf;
1896 if (isICMP_NE)
1897 Mask = ~Mask & fcAllFlags;
1898 return replaceInstUsesWith(Cmp, Builder.createIsFPClass(V, Mask));
1903 return nullptr;
1906 /// Fold icmp (and X, Y), C.
1907 Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1908 BinaryOperator *And,
1909 const APInt &C) {
1910 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1911 return I;
1913 const ICmpInst::Predicate Pred = Cmp.getPredicate();
1914 bool TrueIfNeg;
1915 if (isSignBitCheck(Pred, C, TrueIfNeg)) {
1916 // ((X - 1) & ~X) < 0 --> X == 0
1917 // ((X - 1) & ~X) >= 0 --> X != 0
1918 Value *X;
1919 if (match(And->getOperand(0), m_Add(m_Value(X), m_AllOnes())) &&
1920 match(And->getOperand(1), m_Not(m_Specific(X)))) {
1921 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1922 return new ICmpInst(NewPred, X, ConstantInt::getNullValue(X->getType()));
1924 // (X & -X) < 0 --> X == MinSignedC
1925 // (X & -X) > -1 --> X != MinSignedC
1926 if (match(And, m_c_And(m_Neg(m_Value(X)), m_Deferred(X)))) {
1927 Constant *MinSignedC = ConstantInt::get(
1928 X->getType(),
1929 APInt::getSignedMinValue(X->getType()->getScalarSizeInBits()));
1930 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1931 return new ICmpInst(NewPred, X, MinSignedC);
1935 // TODO: These all require that Y is constant too, so refactor with the above.
1937 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1938 Value *X = And->getOperand(0);
1939 Value *Y = And->getOperand(1);
1940 if (auto *C2 = dyn_cast<ConstantInt>(Y))
1941 if (auto *LI = dyn_cast<LoadInst>(X))
1942 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1943 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1944 if (Instruction *Res =
1945 foldCmpLoadFromIndexedGlobal(LI, GEP, GV, Cmp, C2))
1946 return Res;
1948 if (!Cmp.isEquality())
1949 return nullptr;
1951 // X & -C == -C -> X > u ~C
1952 // X & -C != -C -> X <= u ~C
1953 // iff C is a power of 2
1954 if (Cmp.getOperand(1) == Y && C.isNegatedPowerOf2()) {
1955 auto NewPred =
1956 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1957 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1960 // If we are testing the intersection of 2 select-of-nonzero-constants with no
1961 // common bits set, it's the same as checking if exactly one select condition
1962 // is set:
1963 // ((A ? TC : FC) & (B ? TC : FC)) == 0 --> xor A, B
1964 // ((A ? TC : FC) & (B ? TC : FC)) != 0 --> not(xor A, B)
1965 // TODO: Generalize for non-constant values.
1966 // TODO: Handle signed/unsigned predicates.
1967 // TODO: Handle other bitwise logic connectors.
1968 // TODO: Extend to handle a non-zero compare constant.
1969 if (C.isZero() && (Pred == CmpInst::ICMP_EQ || And->hasOneUse())) {
1970 assert(Cmp.isEquality() && "Not expecting non-equality predicates");
1971 Value *A, *B;
1972 const APInt *TC, *FC;
1973 if (match(X, m_Select(m_Value(A), m_APInt(TC), m_APInt(FC))) &&
1974 match(Y,
1975 m_Select(m_Value(B), m_SpecificInt(*TC), m_SpecificInt(*FC))) &&
1976 !TC->isZero() && !FC->isZero() && !TC->intersects(*FC)) {
1977 Value *R = Builder.CreateXor(A, B);
1978 if (Pred == CmpInst::ICMP_NE)
1979 R = Builder.CreateNot(R);
1980 return replaceInstUsesWith(Cmp, R);
1984 // ((zext i1 X) & Y) == 0 --> !((trunc Y) & X)
1985 // ((zext i1 X) & Y) != 0 --> ((trunc Y) & X)
1986 // ((zext i1 X) & Y) == 1 --> ((trunc Y) & X)
1987 // ((zext i1 X) & Y) != 1 --> !((trunc Y) & X)
1988 if (match(And, m_OneUse(m_c_And(m_OneUse(m_ZExt(m_Value(X))), m_Value(Y)))) &&
1989 X->getType()->isIntOrIntVectorTy(1) && (C.isZero() || C.isOne())) {
1990 Value *TruncY = Builder.CreateTrunc(Y, X->getType());
1991 if (C.isZero() ^ (Pred == CmpInst::ICMP_NE)) {
1992 Value *And = Builder.CreateAnd(TruncY, X);
1993 return BinaryOperator::CreateNot(And);
1995 return BinaryOperator::CreateAnd(TruncY, X);
1998 // (icmp eq/ne (and (shl -1, X), Y), 0)
1999 // -> (icmp eq/ne (lshr Y, X), 0)
2000 // We could technically handle any C == 0 or (C < 0 && isOdd(C)) but it seems
2001 // highly unlikely the non-zero case will ever show up in code.
2002 if (C.isZero() &&
2003 match(And, m_OneUse(m_c_And(m_OneUse(m_Shl(m_AllOnes(), m_Value(X))),
2004 m_Value(Y))))) {
2005 Value *LShr = Builder.CreateLShr(Y, X);
2006 return new ICmpInst(Pred, LShr, Constant::getNullValue(LShr->getType()));
2009 // (icmp eq/ne (and (add A, Addend), Msk), C)
2010 // -> (icmp eq/ne (and A, Msk), (and (sub C, Addend), Msk))
2012 Value *A;
2013 const APInt *Addend, *Msk;
2014 if (match(And, m_And(m_OneUse(m_Add(m_Value(A), m_APInt(Addend))),
2015 m_APInt(Msk))) &&
2016 Msk->isMask() && C.ule(*Msk)) {
2017 APInt NewComperand = (C - *Addend) & *Msk;
2018 Value* MaskA = Builder.CreateAnd(A, ConstantInt::get(A->getType(), *Msk));
2019 return new ICmpInst(
2020 Pred, MaskA,
2021 Constant::getIntegerValue(MaskA->getType(), NewComperand));
2025 return nullptr;
2028 /// Fold icmp eq/ne (or (xor/sub (X1, X2), xor/sub (X3, X4))), 0.
2029 static Value *foldICmpOrXorSubChain(ICmpInst &Cmp, BinaryOperator *Or,
2030 InstCombiner::BuilderTy &Builder) {
2031 // Are we using xors or subs to bitwise check for a pair or pairs of
2032 // (in)equalities? Convert to a shorter form that has more potential to be
2033 // folded even further.
2034 // ((X1 ^/- X2) || (X3 ^/- X4)) == 0 --> (X1 == X2) && (X3 == X4)
2035 // ((X1 ^/- X2) || (X3 ^/- X4)) != 0 --> (X1 != X2) || (X3 != X4)
2036 // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) == 0 -->
2037 // (X1 == X2) && (X3 == X4) && (X5 == X6)
2038 // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) != 0 -->
2039 // (X1 != X2) || (X3 != X4) || (X5 != X6)
2040 SmallVector<std::pair<Value *, Value *>, 2> CmpValues;
2041 SmallVector<Value *, 16> WorkList(1, Or);
2043 while (!WorkList.empty()) {
2044 auto MatchOrOperatorArgument = [&](Value *OrOperatorArgument) {
2045 Value *Lhs, *Rhs;
2047 if (match(OrOperatorArgument,
2048 m_OneUse(m_Xor(m_Value(Lhs), m_Value(Rhs))))) {
2049 CmpValues.emplace_back(Lhs, Rhs);
2050 return;
2053 if (match(OrOperatorArgument,
2054 m_OneUse(m_Sub(m_Value(Lhs), m_Value(Rhs))))) {
2055 CmpValues.emplace_back(Lhs, Rhs);
2056 return;
2059 WorkList.push_back(OrOperatorArgument);
2062 Value *CurrentValue = WorkList.pop_back_val();
2063 Value *OrOperatorLhs, *OrOperatorRhs;
2065 if (!match(CurrentValue,
2066 m_Or(m_Value(OrOperatorLhs), m_Value(OrOperatorRhs)))) {
2067 return nullptr;
2070 MatchOrOperatorArgument(OrOperatorRhs);
2071 MatchOrOperatorArgument(OrOperatorLhs);
2074 ICmpInst::Predicate Pred = Cmp.getPredicate();
2075 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2076 Value *LhsCmp = Builder.CreateICmp(Pred, CmpValues.rbegin()->first,
2077 CmpValues.rbegin()->second);
2079 for (auto It = CmpValues.rbegin() + 1; It != CmpValues.rend(); ++It) {
2080 Value *RhsCmp = Builder.CreateICmp(Pred, It->first, It->second);
2081 LhsCmp = Builder.CreateBinOp(BOpc, LhsCmp, RhsCmp);
2084 return LhsCmp;
2087 /// Fold icmp (or X, Y), C.
2088 Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
2089 BinaryOperator *Or,
2090 const APInt &C) {
2091 ICmpInst::Predicate Pred = Cmp.getPredicate();
2092 if (C.isOne()) {
2093 // icmp slt signum(V) 1 --> icmp slt V, 1
2094 Value *V = nullptr;
2095 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
2096 return new ICmpInst(ICmpInst::ICMP_SLT, V,
2097 ConstantInt::get(V->getType(), 1));
2100 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
2102 // (icmp eq/ne (or disjoint x, C0), C1)
2103 // -> (icmp eq/ne x, C0^C1)
2104 if (Cmp.isEquality() && match(OrOp1, m_ImmConstant()) &&
2105 cast<PossiblyDisjointInst>(Or)->isDisjoint()) {
2106 Value *NewC =
2107 Builder.CreateXor(OrOp1, ConstantInt::get(OrOp1->getType(), C));
2108 return new ICmpInst(Pred, OrOp0, NewC);
2111 const APInt *MaskC;
2112 if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
2113 if (*MaskC == C && (C + 1).isPowerOf2()) {
2114 // X | C == C --> X <=u C
2115 // X | C != C --> X >u C
2116 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
2117 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
2118 return new ICmpInst(Pred, OrOp0, OrOp1);
2121 // More general: canonicalize 'equality with set bits mask' to
2122 // 'equality with clear bits mask'.
2123 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
2124 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
2125 if (Or->hasOneUse()) {
2126 Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
2127 Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
2128 return new ICmpInst(Pred, And, NewC);
2132 // (X | (X-1)) s< 0 --> X s< 1
2133 // (X | (X-1)) s> -1 --> X s> 0
2134 Value *X;
2135 bool TrueIfSigned;
2136 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
2137 match(Or, m_c_Or(m_Add(m_Value(X), m_AllOnes()), m_Deferred(X)))) {
2138 auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT;
2139 Constant *NewC = ConstantInt::get(X->getType(), TrueIfSigned ? 1 : 0);
2140 return new ICmpInst(NewPred, X, NewC);
2143 const APInt *OrC;
2144 // icmp(X | OrC, C) --> icmp(X, 0)
2145 if (C.isNonNegative() && match(Or, m_Or(m_Value(X), m_APInt(OrC)))) {
2146 switch (Pred) {
2147 // X | OrC s< C --> X s< 0 iff OrC s>= C s>= 0
2148 case ICmpInst::ICMP_SLT:
2149 // X | OrC s>= C --> X s>= 0 iff OrC s>= C s>= 0
2150 case ICmpInst::ICMP_SGE:
2151 if (OrC->sge(C))
2152 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2153 break;
2154 // X | OrC s<= C --> X s< 0 iff OrC s> C s>= 0
2155 case ICmpInst::ICMP_SLE:
2156 // X | OrC s> C --> X s>= 0 iff OrC s> C s>= 0
2157 case ICmpInst::ICMP_SGT:
2158 if (OrC->sgt(C))
2159 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(Pred), X,
2160 ConstantInt::getNullValue(X->getType()));
2161 break;
2162 default:
2163 break;
2167 if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse())
2168 return nullptr;
2170 Value *P, *Q;
2171 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
2172 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
2173 // -> and (icmp eq P, null), (icmp eq Q, null).
2174 Value *CmpP =
2175 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
2176 Value *CmpQ =
2177 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
2178 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2179 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
2182 if (Value *V = foldICmpOrXorSubChain(Cmp, Or, Builder))
2183 return replaceInstUsesWith(Cmp, V);
2185 return nullptr;
2188 /// Fold icmp (mul X, Y), C.
2189 Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
2190 BinaryOperator *Mul,
2191 const APInt &C) {
2192 ICmpInst::Predicate Pred = Cmp.getPredicate();
2193 Type *MulTy = Mul->getType();
2194 Value *X = Mul->getOperand(0);
2196 // If there's no overflow:
2197 // X * X == 0 --> X == 0
2198 // X * X != 0 --> X != 0
2199 if (Cmp.isEquality() && C.isZero() && X == Mul->getOperand(1) &&
2200 (Mul->hasNoUnsignedWrap() || Mul->hasNoSignedWrap()))
2201 return new ICmpInst(Pred, X, ConstantInt::getNullValue(MulTy));
2203 const APInt *MulC;
2204 if (!match(Mul->getOperand(1), m_APInt(MulC)))
2205 return nullptr;
2207 // If this is a test of the sign bit and the multiply is sign-preserving with
2208 // a constant operand, use the multiply LHS operand instead:
2209 // (X * +MulC) < 0 --> X < 0
2210 // (X * -MulC) < 0 --> X > 0
2211 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
2212 if (MulC->isNegative())
2213 Pred = ICmpInst::getSwappedPredicate(Pred);
2214 return new ICmpInst(Pred, X, ConstantInt::getNullValue(MulTy));
2217 if (MulC->isZero())
2218 return nullptr;
2220 // If the multiply does not wrap or the constant is odd, try to divide the
2221 // compare constant by the multiplication factor.
2222 if (Cmp.isEquality()) {
2223 // (mul nsw X, MulC) eq/ne C --> X eq/ne C /s MulC
2224 if (Mul->hasNoSignedWrap() && C.srem(*MulC).isZero()) {
2225 Constant *NewC = ConstantInt::get(MulTy, C.sdiv(*MulC));
2226 return new ICmpInst(Pred, X, NewC);
2229 // C % MulC == 0 is weaker than we could use if MulC is odd because it
2230 // correct to transform if MulC * N == C including overflow. I.e with i8
2231 // (icmp eq (mul X, 5), 101) -> (icmp eq X, 225) but since 101 % 5 != 0, we
2232 // miss that case.
2233 if (C.urem(*MulC).isZero()) {
2234 // (mul nuw X, MulC) eq/ne C --> X eq/ne C /u MulC
2235 // (mul X, OddC) eq/ne N * C --> X eq/ne N
2236 if ((*MulC & 1).isOne() || Mul->hasNoUnsignedWrap()) {
2237 Constant *NewC = ConstantInt::get(MulTy, C.udiv(*MulC));
2238 return new ICmpInst(Pred, X, NewC);
2243 // With a matching no-overflow guarantee, fold the constants:
2244 // (X * MulC) < C --> X < (C / MulC)
2245 // (X * MulC) > C --> X > (C / MulC)
2246 // TODO: Assert that Pred is not equal to SGE, SLE, UGE, ULE?
2247 Constant *NewC = nullptr;
2248 if (Mul->hasNoSignedWrap() && ICmpInst::isSigned(Pred)) {
2249 // MININT / -1 --> overflow.
2250 if (C.isMinSignedValue() && MulC->isAllOnes())
2251 return nullptr;
2252 if (MulC->isNegative())
2253 Pred = ICmpInst::getSwappedPredicate(Pred);
2255 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2256 NewC = ConstantInt::get(
2257 MulTy, APIntOps::RoundingSDiv(C, *MulC, APInt::Rounding::UP));
2258 } else {
2259 assert((Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_SGT) &&
2260 "Unexpected predicate");
2261 NewC = ConstantInt::get(
2262 MulTy, APIntOps::RoundingSDiv(C, *MulC, APInt::Rounding::DOWN));
2264 } else if (Mul->hasNoUnsignedWrap() && ICmpInst::isUnsigned(Pred)) {
2265 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) {
2266 NewC = ConstantInt::get(
2267 MulTy, APIntOps::RoundingUDiv(C, *MulC, APInt::Rounding::UP));
2268 } else {
2269 assert((Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
2270 "Unexpected predicate");
2271 NewC = ConstantInt::get(
2272 MulTy, APIntOps::RoundingUDiv(C, *MulC, APInt::Rounding::DOWN));
2276 return NewC ? new ICmpInst(Pred, X, NewC) : nullptr;
2279 /// Fold icmp (shl nuw C2, Y), C.
2280 static Instruction *foldICmpShlLHSC(ICmpInst &Cmp, Instruction *Shl,
2281 const APInt &C) {
2282 Value *Y;
2283 const APInt *C2;
2284 if (!match(Shl, m_NUWShl(m_APInt(C2), m_Value(Y))))
2285 return nullptr;
2287 Type *ShiftType = Shl->getType();
2288 unsigned TypeBits = C.getBitWidth();
2289 ICmpInst::Predicate Pred = Cmp.getPredicate();
2290 if (Cmp.isUnsigned()) {
2291 if (C2->isZero() || C2->ugt(C))
2292 return nullptr;
2293 APInt Div, Rem;
2294 APInt::udivrem(C, *C2, Div, Rem);
2295 bool CIsPowerOf2 = Rem.isZero() && Div.isPowerOf2();
2297 // (1 << Y) pred C -> Y pred Log2(C)
2298 if (!CIsPowerOf2) {
2299 // (1 << Y) < 30 -> Y <= 4
2300 // (1 << Y) <= 30 -> Y <= 4
2301 // (1 << Y) >= 30 -> Y > 4
2302 // (1 << Y) > 30 -> Y > 4
2303 if (Pred == ICmpInst::ICMP_ULT)
2304 Pred = ICmpInst::ICMP_ULE;
2305 else if (Pred == ICmpInst::ICMP_UGE)
2306 Pred = ICmpInst::ICMP_UGT;
2309 unsigned CLog2 = Div.logBase2();
2310 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2311 } else if (Cmp.isSigned() && C2->isOne()) {
2312 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2313 // (1 << Y) > 0 -> Y != 31
2314 // (1 << Y) > C -> Y != 31 if C is negative.
2315 if (Pred == ICmpInst::ICMP_SGT && C.sle(0))
2316 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2318 // (1 << Y) < 0 -> Y == 31
2319 // (1 << Y) < 1 -> Y == 31
2320 // (1 << Y) < C -> Y == 31 if C is negative and not signed min.
2321 // Exclude signed min by subtracting 1 and lower the upper bound to 0.
2322 if (Pred == ICmpInst::ICMP_SLT && (C-1).sle(0))
2323 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2326 return nullptr;
2329 /// Fold icmp (shl X, Y), C.
2330 Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2331 BinaryOperator *Shl,
2332 const APInt &C) {
2333 const APInt *ShiftVal;
2334 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2335 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2337 ICmpInst::Predicate Pred = Cmp.getPredicate();
2338 // (icmp pred (shl nuw&nsw X, Y), Csle0)
2339 // -> (icmp pred X, Csle0)
2341 // The idea is the nuw/nsw essentially freeze the sign bit for the shift op
2342 // so X's must be what is used.
2343 if (C.sle(0) && Shl->hasNoUnsignedWrap() && Shl->hasNoSignedWrap())
2344 return new ICmpInst(Pred, Shl->getOperand(0), Cmp.getOperand(1));
2346 // (icmp eq/ne (shl nuw|nsw X, Y), 0)
2347 // -> (icmp eq/ne X, 0)
2348 if (ICmpInst::isEquality(Pred) && C.isZero() &&
2349 (Shl->hasNoUnsignedWrap() || Shl->hasNoSignedWrap()))
2350 return new ICmpInst(Pred, Shl->getOperand(0), Cmp.getOperand(1));
2352 // (icmp slt (shl nsw X, Y), 0/1)
2353 // -> (icmp slt X, 0/1)
2354 // (icmp sgt (shl nsw X, Y), 0/-1)
2355 // -> (icmp sgt X, 0/-1)
2357 // NB: sge/sle with a constant will canonicalize to sgt/slt.
2358 if (Shl->hasNoSignedWrap() &&
2359 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT))
2360 if (C.isZero() || (Pred == ICmpInst::ICMP_SGT ? C.isAllOnes() : C.isOne()))
2361 return new ICmpInst(Pred, Shl->getOperand(0), Cmp.getOperand(1));
2363 const APInt *ShiftAmt;
2364 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2365 return foldICmpShlLHSC(Cmp, Shl, C);
2367 // Check that the shift amount is in range. If not, don't perform undefined
2368 // shifts. When the shift is visited, it will be simplified.
2369 unsigned TypeBits = C.getBitWidth();
2370 if (ShiftAmt->uge(TypeBits))
2371 return nullptr;
2373 Value *X = Shl->getOperand(0);
2374 Type *ShType = Shl->getType();
2376 // NSW guarantees that we are only shifting out sign bits from the high bits,
2377 // so we can ASHR the compare constant without needing a mask and eliminate
2378 // the shift.
2379 if (Shl->hasNoSignedWrap()) {
2380 if (Pred == ICmpInst::ICMP_SGT) {
2381 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2382 APInt ShiftedC = C.ashr(*ShiftAmt);
2383 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2385 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2386 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2387 APInt ShiftedC = C.ashr(*ShiftAmt);
2388 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2390 if (Pred == ICmpInst::ICMP_SLT) {
2391 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2392 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2393 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2394 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2395 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2396 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2397 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2401 // NUW guarantees that we are only shifting out zero bits from the high bits,
2402 // so we can LSHR the compare constant without needing a mask and eliminate
2403 // the shift.
2404 if (Shl->hasNoUnsignedWrap()) {
2405 if (Pred == ICmpInst::ICMP_UGT) {
2406 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2407 APInt ShiftedC = C.lshr(*ShiftAmt);
2408 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2410 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2411 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2412 APInt ShiftedC = C.lshr(*ShiftAmt);
2413 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2415 if (Pred == ICmpInst::ICMP_ULT) {
2416 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2417 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2418 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2419 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2420 assert(C.ugt(0) && "ult 0 should have been eliminated");
2421 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2422 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2426 if (Cmp.isEquality() && Shl->hasOneUse()) {
2427 // Strength-reduce the shift into an 'and'.
2428 Constant *Mask = ConstantInt::get(
2429 ShType,
2430 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2431 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2432 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2433 return new ICmpInst(Pred, And, LShrC);
2436 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2437 bool TrueIfSigned = false;
2438 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2439 // (X << 31) <s 0 --> (X & 1) != 0
2440 Constant *Mask = ConstantInt::get(
2441 ShType,
2442 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2443 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2444 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2445 And, Constant::getNullValue(ShType));
2448 // Simplify 'shl' inequality test into 'and' equality test.
2449 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2450 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2451 if ((C + 1).isPowerOf2() &&
2452 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2453 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2454 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2455 : ICmpInst::ICMP_NE,
2456 And, Constant::getNullValue(ShType));
2458 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2459 if (C.isPowerOf2() &&
2460 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2461 Value *And =
2462 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2463 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2464 : ICmpInst::ICMP_NE,
2465 And, Constant::getNullValue(ShType));
2469 // Transform (icmp pred iM (shl iM %v, N), C)
2470 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2471 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2472 // This enables us to get rid of the shift in favor of a trunc that may be
2473 // free on the target. It has the additional benefit of comparing to a
2474 // smaller constant that may be more target-friendly.
2475 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2476 if (Shl->hasOneUse() && Amt != 0 &&
2477 shouldChangeType(ShType->getScalarSizeInBits(), TypeBits - Amt)) {
2478 ICmpInst::Predicate CmpPred = Pred;
2479 APInt RHSC = C;
2481 if (RHSC.countr_zero() < Amt && ICmpInst::isStrictPredicate(CmpPred)) {
2482 // Try the flipped strictness predicate.
2483 // e.g.:
2484 // icmp ult i64 (shl X, 32), 8589934593 ->
2485 // icmp ule i64 (shl X, 32), 8589934592 ->
2486 // icmp ule i32 (trunc X, i32), 2 ->
2487 // icmp ult i32 (trunc X, i32), 3
2488 if (auto FlippedStrictness =
2489 InstCombiner::getFlippedStrictnessPredicateAndConstant(
2490 Pred, ConstantInt::get(ShType->getContext(), C))) {
2491 CmpPred = FlippedStrictness->first;
2492 RHSC = cast<ConstantInt>(FlippedStrictness->second)->getValue();
2496 if (RHSC.countr_zero() >= Amt) {
2497 Type *TruncTy = ShType->getWithNewBitWidth(TypeBits - Amt);
2498 Constant *NewC =
2499 ConstantInt::get(TruncTy, RHSC.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2500 return new ICmpInst(CmpPred,
2501 Builder.CreateTrunc(X, TruncTy, "", /*IsNUW=*/false,
2502 Shl->hasNoSignedWrap()),
2503 NewC);
2507 return nullptr;
2510 /// Fold icmp ({al}shr X, Y), C.
2511 Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2512 BinaryOperator *Shr,
2513 const APInt &C) {
2514 // An exact shr only shifts out zero bits, so:
2515 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2516 Value *X = Shr->getOperand(0);
2517 CmpInst::Predicate Pred = Cmp.getPredicate();
2518 if (Cmp.isEquality() && Shr->isExact() && C.isZero())
2519 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2521 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2522 const APInt *ShiftValC;
2523 if (match(X, m_APInt(ShiftValC))) {
2524 if (Cmp.isEquality())
2525 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftValC);
2527 // (ShiftValC >> Y) >s -1 --> Y != 0 with ShiftValC < 0
2528 // (ShiftValC >> Y) <s 0 --> Y == 0 with ShiftValC < 0
2529 bool TrueIfSigned;
2530 if (!IsAShr && ShiftValC->isNegative() &&
2531 isSignBitCheck(Pred, C, TrueIfSigned))
2532 return new ICmpInst(TrueIfSigned ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE,
2533 Shr->getOperand(1),
2534 ConstantInt::getNullValue(X->getType()));
2536 // If the shifted constant is a power-of-2, test the shift amount directly:
2537 // (ShiftValC >> Y) >u C --> X <u (LZ(C) - LZ(ShiftValC))
2538 // (ShiftValC >> Y) <u C --> X >=u (LZ(C-1) - LZ(ShiftValC))
2539 if (!IsAShr && ShiftValC->isPowerOf2() &&
2540 (Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_ULT)) {
2541 bool IsUGT = Pred == CmpInst::ICMP_UGT;
2542 assert(ShiftValC->uge(C) && "Expected simplify of compare");
2543 assert((IsUGT || !C.isZero()) && "Expected X u< 0 to simplify");
2545 unsigned CmpLZ = IsUGT ? C.countl_zero() : (C - 1).countl_zero();
2546 unsigned ShiftLZ = ShiftValC->countl_zero();
2547 Constant *NewC = ConstantInt::get(Shr->getType(), CmpLZ - ShiftLZ);
2548 auto NewPred = IsUGT ? CmpInst::ICMP_ULT : CmpInst::ICMP_UGE;
2549 return new ICmpInst(NewPred, Shr->getOperand(1), NewC);
2553 const APInt *ShiftAmtC;
2554 if (!match(Shr->getOperand(1), m_APInt(ShiftAmtC)))
2555 return nullptr;
2557 // Check that the shift amount is in range. If not, don't perform undefined
2558 // shifts. When the shift is visited it will be simplified.
2559 unsigned TypeBits = C.getBitWidth();
2560 unsigned ShAmtVal = ShiftAmtC->getLimitedValue(TypeBits);
2561 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2562 return nullptr;
2564 bool IsExact = Shr->isExact();
2565 Type *ShrTy = Shr->getType();
2566 // TODO: If we could guarantee that InstSimplify would handle all of the
2567 // constant-value-based preconditions in the folds below, then we could assert
2568 // those conditions rather than checking them. This is difficult because of
2569 // undef/poison (PR34838).
2570 if (IsAShr && Shr->hasOneUse()) {
2571 if (IsExact && (Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) &&
2572 (C - 1).isPowerOf2() && C.countLeadingZeros() > ShAmtVal) {
2573 // When C - 1 is a power of two and the transform can be legally
2574 // performed, prefer this form so the produced constant is close to a
2575 // power of two.
2576 // icmp slt/ult (ashr exact X, ShAmtC), C
2577 // --> icmp slt/ult X, (C - 1) << ShAmtC) + 1
2578 APInt ShiftedC = (C - 1).shl(ShAmtVal) + 1;
2579 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2581 if (IsExact || Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) {
2582 // When ShAmtC can be shifted losslessly:
2583 // icmp PRED (ashr exact X, ShAmtC), C --> icmp PRED X, (C << ShAmtC)
2584 // icmp slt/ult (ashr X, ShAmtC), C --> icmp slt/ult X, (C << ShAmtC)
2585 APInt ShiftedC = C.shl(ShAmtVal);
2586 if (ShiftedC.ashr(ShAmtVal) == C)
2587 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2589 if (Pred == CmpInst::ICMP_SGT) {
2590 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2591 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2592 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2593 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2594 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2596 if (Pred == CmpInst::ICMP_UGT) {
2597 // icmp ugt (ashr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2598 // 'C + 1 << ShAmtC' can overflow as a signed number, so the 2nd
2599 // clause accounts for that pattern.
2600 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2601 if ((ShiftedC + 1).ashr(ShAmtVal) == (C + 1) ||
2602 (C + 1).shl(ShAmtVal).isMinSignedValue())
2603 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2606 // If the compare constant has significant bits above the lowest sign-bit,
2607 // then convert an unsigned cmp to a test of the sign-bit:
2608 // (ashr X, ShiftC) u> C --> X s< 0
2609 // (ashr X, ShiftC) u< C --> X s> -1
2610 if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2611 if (Pred == CmpInst::ICMP_UGT) {
2612 return new ICmpInst(CmpInst::ICMP_SLT, X,
2613 ConstantInt::getNullValue(ShrTy));
2615 if (Pred == CmpInst::ICMP_ULT) {
2616 return new ICmpInst(CmpInst::ICMP_SGT, X,
2617 ConstantInt::getAllOnesValue(ShrTy));
2620 } else if (!IsAShr) {
2621 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2622 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2623 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2624 APInt ShiftedC = C.shl(ShAmtVal);
2625 if (ShiftedC.lshr(ShAmtVal) == C)
2626 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2628 if (Pred == CmpInst::ICMP_UGT) {
2629 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2630 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2631 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2632 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2636 if (!Cmp.isEquality())
2637 return nullptr;
2639 // Handle equality comparisons of shift-by-constant.
2641 // If the comparison constant changes with the shift, the comparison cannot
2642 // succeed (bits of the comparison constant cannot match the shifted value).
2643 // This should be known by InstSimplify and already be folded to true/false.
2644 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2645 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2646 "Expected icmp+shr simplify did not occur.");
2648 // If the bits shifted out are known zero, compare the unshifted value:
2649 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2650 if (Shr->isExact())
2651 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2653 if (C.isZero()) {
2654 // == 0 is u< 1.
2655 if (Pred == CmpInst::ICMP_EQ)
2656 return new ICmpInst(CmpInst::ICMP_ULT, X,
2657 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2658 else
2659 return new ICmpInst(CmpInst::ICMP_UGT, X,
2660 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2663 if (Shr->hasOneUse()) {
2664 // Canonicalize the shift into an 'and':
2665 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2666 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2667 Constant *Mask = ConstantInt::get(ShrTy, Val);
2668 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2669 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2672 return nullptr;
2675 Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2676 BinaryOperator *SRem,
2677 const APInt &C) {
2678 // Match an 'is positive' or 'is negative' comparison of remainder by a
2679 // constant power-of-2 value:
2680 // (X % pow2C) sgt/slt 0
2681 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2682 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT &&
2683 Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2684 return nullptr;
2686 // TODO: The one-use check is standard because we do not typically want to
2687 // create longer instruction sequences, but this might be a special-case
2688 // because srem is not good for analysis or codegen.
2689 if (!SRem->hasOneUse())
2690 return nullptr;
2692 const APInt *DivisorC;
2693 if (!match(SRem->getOperand(1), m_Power2(DivisorC)))
2694 return nullptr;
2696 // For cmp_sgt/cmp_slt only zero valued C is handled.
2697 // For cmp_eq/cmp_ne only positive valued C is handled.
2698 if (((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT) &&
2699 !C.isZero()) ||
2700 ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2701 !C.isStrictlyPositive()))
2702 return nullptr;
2704 // Mask off the sign bit and the modulo bits (low-bits).
2705 Type *Ty = SRem->getType();
2706 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2707 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2708 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2710 if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)
2711 return new ICmpInst(Pred, And, ConstantInt::get(Ty, C));
2713 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2714 // bit is set. Example:
2715 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2716 if (Pred == ICmpInst::ICMP_SGT)
2717 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2719 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2720 // bit is set. Example:
2721 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2722 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2725 /// Fold icmp (udiv X, Y), C.
2726 Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2727 BinaryOperator *UDiv,
2728 const APInt &C) {
2729 ICmpInst::Predicate Pred = Cmp.getPredicate();
2730 Value *X = UDiv->getOperand(0);
2731 Value *Y = UDiv->getOperand(1);
2732 Type *Ty = UDiv->getType();
2734 const APInt *C2;
2735 if (!match(X, m_APInt(C2)))
2736 return nullptr;
2738 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2740 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2741 if (Pred == ICmpInst::ICMP_UGT) {
2742 assert(!C.isMaxValue() &&
2743 "icmp ugt X, UINT_MAX should have been simplified already.");
2744 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2745 ConstantInt::get(Ty, C2->udiv(C + 1)));
2748 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2749 if (Pred == ICmpInst::ICMP_ULT) {
2750 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2751 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2752 ConstantInt::get(Ty, C2->udiv(C)));
2755 return nullptr;
2758 /// Fold icmp ({su}div X, Y), C.
2759 Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2760 BinaryOperator *Div,
2761 const APInt &C) {
2762 ICmpInst::Predicate Pred = Cmp.getPredicate();
2763 Value *X = Div->getOperand(0);
2764 Value *Y = Div->getOperand(1);
2765 Type *Ty = Div->getType();
2766 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2768 // If unsigned division and the compare constant is bigger than
2769 // UMAX/2 (negative), there's only one pair of values that satisfies an
2770 // equality check, so eliminate the division:
2771 // (X u/ Y) == C --> (X == C) && (Y == 1)
2772 // (X u/ Y) != C --> (X != C) || (Y != 1)
2773 // Similarly, if signed division and the compare constant is exactly SMIN:
2774 // (X s/ Y) == SMIN --> (X == SMIN) && (Y == 1)
2775 // (X s/ Y) != SMIN --> (X != SMIN) || (Y != 1)
2776 if (Cmp.isEquality() && Div->hasOneUse() && C.isSignBitSet() &&
2777 (!DivIsSigned || C.isMinSignedValue())) {
2778 Value *XBig = Builder.CreateICmp(Pred, X, ConstantInt::get(Ty, C));
2779 Value *YOne = Builder.CreateICmp(Pred, Y, ConstantInt::get(Ty, 1));
2780 auto Logic = Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2781 return BinaryOperator::Create(Logic, XBig, YOne);
2784 // Fold: icmp pred ([us]div X, C2), C -> range test
2785 // Fold this div into the comparison, producing a range check.
2786 // Determine, based on the divide type, what the range is being
2787 // checked. If there is an overflow on the low or high side, remember
2788 // it, otherwise compute the range [low, hi) bounding the new value.
2789 // See: InsertRangeTest above for the kinds of replacements possible.
2790 const APInt *C2;
2791 if (!match(Y, m_APInt(C2)))
2792 return nullptr;
2794 // FIXME: If the operand types don't match the type of the divide
2795 // then don't attempt this transform. The code below doesn't have the
2796 // logic to deal with a signed divide and an unsigned compare (and
2797 // vice versa). This is because (x /s C2) <s C produces different
2798 // results than (x /s C2) <u C or (x /u C2) <s C or even
2799 // (x /u C2) <u C. Simply casting the operands and result won't
2800 // work. :( The if statement below tests that condition and bails
2801 // if it finds it.
2802 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2803 return nullptr;
2805 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2806 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2807 // division-by-constant cases should be present, we can not assert that they
2808 // have happened before we reach this icmp instruction.
2809 if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes()))
2810 return nullptr;
2812 // Compute Prod = C * C2. We are essentially solving an equation of
2813 // form X / C2 = C. We solve for X by multiplying C2 and C.
2814 // By solving for X, we can turn this into a range check instead of computing
2815 // a divide.
2816 APInt Prod = C * *C2;
2818 // Determine if the product overflows by seeing if the product is not equal to
2819 // the divide. Make sure we do the same kind of divide as in the LHS
2820 // instruction that we're folding.
2821 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2823 // If the division is known to be exact, then there is no remainder from the
2824 // divide, so the covered range size is unit, otherwise it is the divisor.
2825 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2827 // Figure out the interval that is being checked. For example, a comparison
2828 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2829 // Compute this interval based on the constants involved and the signedness of
2830 // the compare/divide. This computes a half-open interval, keeping track of
2831 // whether either value in the interval overflows. After analysis each
2832 // overflow variable is set to 0 if it's corresponding bound variable is valid
2833 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2834 int LoOverflow = 0, HiOverflow = 0;
2835 APInt LoBound, HiBound;
2837 if (!DivIsSigned) { // udiv
2838 // e.g. X/5 op 3 --> [15, 20)
2839 LoBound = Prod;
2840 HiOverflow = LoOverflow = ProdOV;
2841 if (!HiOverflow) {
2842 // If this is not an exact divide, then many values in the range collapse
2843 // to the same result value.
2844 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2846 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2847 if (C.isZero()) { // (X / pos) op 0
2848 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2849 LoBound = -(RangeSize - 1);
2850 HiBound = RangeSize;
2851 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2852 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2853 HiOverflow = LoOverflow = ProdOV;
2854 if (!HiOverflow)
2855 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2856 } else { // (X / pos) op neg
2857 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2858 HiBound = Prod + 1;
2859 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2860 if (!LoOverflow) {
2861 APInt DivNeg = -RangeSize;
2862 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2865 } else if (C2->isNegative()) { // Divisor is < 0.
2866 if (Div->isExact())
2867 RangeSize.negate();
2868 if (C.isZero()) { // (X / neg) op 0
2869 // e.g. X/-5 op 0 --> [-4, 5)
2870 LoBound = RangeSize + 1;
2871 HiBound = -RangeSize;
2872 if (HiBound == *C2) { // -INTMIN = INTMIN
2873 HiOverflow = 1; // [INTMIN+1, overflow)
2874 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2876 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2877 // e.g. X/-5 op 3 --> [-19, -14)
2878 HiBound = Prod + 1;
2879 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2880 if (!LoOverflow)
2881 LoOverflow =
2882 addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1 : 0;
2883 } else { // (X / neg) op neg
2884 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2885 LoOverflow = HiOverflow = ProdOV;
2886 if (!HiOverflow)
2887 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2890 // Dividing by a negative swaps the condition. LT <-> GT
2891 Pred = ICmpInst::getSwappedPredicate(Pred);
2894 switch (Pred) {
2895 default:
2896 llvm_unreachable("Unhandled icmp predicate!");
2897 case ICmpInst::ICMP_EQ:
2898 if (LoOverflow && HiOverflow)
2899 return replaceInstUsesWith(Cmp, Builder.getFalse());
2900 if (HiOverflow)
2901 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
2902 X, ConstantInt::get(Ty, LoBound));
2903 if (LoOverflow)
2904 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
2905 X, ConstantInt::get(Ty, HiBound));
2906 return replaceInstUsesWith(
2907 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2908 case ICmpInst::ICMP_NE:
2909 if (LoOverflow && HiOverflow)
2910 return replaceInstUsesWith(Cmp, Builder.getTrue());
2911 if (HiOverflow)
2912 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
2913 X, ConstantInt::get(Ty, LoBound));
2914 if (LoOverflow)
2915 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
2916 X, ConstantInt::get(Ty, HiBound));
2917 return replaceInstUsesWith(
2918 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, false));
2919 case ICmpInst::ICMP_ULT:
2920 case ICmpInst::ICMP_SLT:
2921 if (LoOverflow == +1) // Low bound is greater than input range.
2922 return replaceInstUsesWith(Cmp, Builder.getTrue());
2923 if (LoOverflow == -1) // Low bound is less than input range.
2924 return replaceInstUsesWith(Cmp, Builder.getFalse());
2925 return new ICmpInst(Pred, X, ConstantInt::get(Ty, LoBound));
2926 case ICmpInst::ICMP_UGT:
2927 case ICmpInst::ICMP_SGT:
2928 if (HiOverflow == +1) // High bound greater than input range.
2929 return replaceInstUsesWith(Cmp, Builder.getFalse());
2930 if (HiOverflow == -1) // High bound less than input range.
2931 return replaceInstUsesWith(Cmp, Builder.getTrue());
2932 if (Pred == ICmpInst::ICMP_UGT)
2933 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, HiBound));
2934 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, HiBound));
2937 return nullptr;
2940 /// Fold icmp (sub X, Y), C.
2941 Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2942 BinaryOperator *Sub,
2943 const APInt &C) {
2944 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2945 ICmpInst::Predicate Pred = Cmp.getPredicate();
2946 Type *Ty = Sub->getType();
2948 // (SubC - Y) == C) --> Y == (SubC - C)
2949 // (SubC - Y) != C) --> Y != (SubC - C)
2950 Constant *SubC;
2951 if (Cmp.isEquality() && match(X, m_ImmConstant(SubC))) {
2952 return new ICmpInst(Pred, Y,
2953 ConstantExpr::getSub(SubC, ConstantInt::get(Ty, C)));
2956 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2957 const APInt *C2;
2958 APInt SubResult;
2959 ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate();
2960 bool HasNSW = Sub->hasNoSignedWrap();
2961 bool HasNUW = Sub->hasNoUnsignedWrap();
2962 if (match(X, m_APInt(C2)) &&
2963 ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) &&
2964 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2965 return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, SubResult));
2967 // X - Y == 0 --> X == Y.
2968 // X - Y != 0 --> X != Y.
2969 // TODO: We allow this with multiple uses as long as the other uses are not
2970 // in phis. The phi use check is guarding against a codegen regression
2971 // for a loop test. If the backend could undo this (and possibly
2972 // subsequent transforms), we would not need this hack.
2973 if (Cmp.isEquality() && C.isZero() &&
2974 none_of((Sub->users()), [](const User *U) { return isa<PHINode>(U); }))
2975 return new ICmpInst(Pred, X, Y);
2977 // The following transforms are only worth it if the only user of the subtract
2978 // is the icmp.
2979 // TODO: This is an artificial restriction for all of the transforms below
2980 // that only need a single replacement icmp. Can these use the phi test
2981 // like the transform above here?
2982 if (!Sub->hasOneUse())
2983 return nullptr;
2985 if (Sub->hasNoSignedWrap()) {
2986 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2987 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
2988 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2990 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2991 if (Pred == ICmpInst::ICMP_SGT && C.isZero())
2992 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2994 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2995 if (Pred == ICmpInst::ICMP_SLT && C.isZero())
2996 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2998 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2999 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
3000 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3003 if (!match(X, m_APInt(C2)))
3004 return nullptr;
3006 // C2 - Y <u C -> (Y | (C - 1)) == C2
3007 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
3008 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
3009 (*C2 & (C - 1)) == (C - 1))
3010 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
3012 // C2 - Y >u C -> (Y | C) != C2
3013 // iff C2 & C == C and C + 1 is a power of 2
3014 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
3015 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
3017 // We have handled special cases that reduce.
3018 // Canonicalize any remaining sub to add as:
3019 // (C2 - Y) > C --> (Y + ~C2) < ~C
3020 Value *Add = Builder.CreateAdd(Y, ConstantInt::get(Ty, ~(*C2)), "notsub",
3021 HasNUW, HasNSW);
3022 return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, ~C));
3025 static Value *createLogicFromTable(const std::bitset<4> &Table, Value *Op0,
3026 Value *Op1, IRBuilderBase &Builder,
3027 bool HasOneUse) {
3028 auto FoldConstant = [&](bool Val) {
3029 Constant *Res = Val ? Builder.getTrue() : Builder.getFalse();
3030 if (Op0->getType()->isVectorTy())
3031 Res = ConstantVector::getSplat(
3032 cast<VectorType>(Op0->getType())->getElementCount(), Res);
3033 return Res;
3036 switch (Table.to_ulong()) {
3037 case 0: // 0 0 0 0
3038 return FoldConstant(false);
3039 case 1: // 0 0 0 1
3040 return HasOneUse ? Builder.CreateNot(Builder.CreateOr(Op0, Op1)) : nullptr;
3041 case 2: // 0 0 1 0
3042 return HasOneUse ? Builder.CreateAnd(Builder.CreateNot(Op0), Op1) : nullptr;
3043 case 3: // 0 0 1 1
3044 return Builder.CreateNot(Op0);
3045 case 4: // 0 1 0 0
3046 return HasOneUse ? Builder.CreateAnd(Op0, Builder.CreateNot(Op1)) : nullptr;
3047 case 5: // 0 1 0 1
3048 return Builder.CreateNot(Op1);
3049 case 6: // 0 1 1 0
3050 return Builder.CreateXor(Op0, Op1);
3051 case 7: // 0 1 1 1
3052 return HasOneUse ? Builder.CreateNot(Builder.CreateAnd(Op0, Op1)) : nullptr;
3053 case 8: // 1 0 0 0
3054 return Builder.CreateAnd(Op0, Op1);
3055 case 9: // 1 0 0 1
3056 return HasOneUse ? Builder.CreateNot(Builder.CreateXor(Op0, Op1)) : nullptr;
3057 case 10: // 1 0 1 0
3058 return Op1;
3059 case 11: // 1 0 1 1
3060 return HasOneUse ? Builder.CreateOr(Builder.CreateNot(Op0), Op1) : nullptr;
3061 case 12: // 1 1 0 0
3062 return Op0;
3063 case 13: // 1 1 0 1
3064 return HasOneUse ? Builder.CreateOr(Op0, Builder.CreateNot(Op1)) : nullptr;
3065 case 14: // 1 1 1 0
3066 return Builder.CreateOr(Op0, Op1);
3067 case 15: // 1 1 1 1
3068 return FoldConstant(true);
3069 default:
3070 llvm_unreachable("Invalid Operation");
3072 return nullptr;
3075 /// Fold icmp (add X, Y), C.
3076 Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
3077 BinaryOperator *Add,
3078 const APInt &C) {
3079 Value *Y = Add->getOperand(1);
3080 Value *X = Add->getOperand(0);
3082 Value *Op0, *Op1;
3083 Instruction *Ext0, *Ext1;
3084 const CmpInst::Predicate Pred = Cmp.getPredicate();
3085 if (match(Add,
3086 m_Add(m_CombineAnd(m_Instruction(Ext0), m_ZExtOrSExt(m_Value(Op0))),
3087 m_CombineAnd(m_Instruction(Ext1),
3088 m_ZExtOrSExt(m_Value(Op1))))) &&
3089 Op0->getType()->isIntOrIntVectorTy(1) &&
3090 Op1->getType()->isIntOrIntVectorTy(1)) {
3091 unsigned BW = C.getBitWidth();
3092 std::bitset<4> Table;
3093 auto ComputeTable = [&](bool Op0Val, bool Op1Val) {
3094 APInt Res(BW, 0);
3095 if (Op0Val)
3096 Res += APInt(BW, isa<ZExtInst>(Ext0) ? 1 : -1, /*isSigned=*/true);
3097 if (Op1Val)
3098 Res += APInt(BW, isa<ZExtInst>(Ext1) ? 1 : -1, /*isSigned=*/true);
3099 return ICmpInst::compare(Res, C, Pred);
3102 Table[0] = ComputeTable(false, false);
3103 Table[1] = ComputeTable(false, true);
3104 Table[2] = ComputeTable(true, false);
3105 Table[3] = ComputeTable(true, true);
3106 if (auto *Cond =
3107 createLogicFromTable(Table, Op0, Op1, Builder, Add->hasOneUse()))
3108 return replaceInstUsesWith(Cmp, Cond);
3110 const APInt *C2;
3111 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
3112 return nullptr;
3114 // Fold icmp pred (add X, C2), C.
3115 Type *Ty = Add->getType();
3117 // If the add does not wrap, we can always adjust the compare by subtracting
3118 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
3119 // are canonicalized to SGT/SLT/UGT/ULT.
3120 if ((Add->hasNoSignedWrap() &&
3121 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
3122 (Add->hasNoUnsignedWrap() &&
3123 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
3124 bool Overflow;
3125 APInt NewC =
3126 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
3127 // If there is overflow, the result must be true or false.
3128 // TODO: Can we assert there is no overflow because InstSimplify always
3129 // handles those cases?
3130 if (!Overflow)
3131 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
3132 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
3135 if (ICmpInst::isUnsigned(Pred) && Add->hasNoSignedWrap() &&
3136 C.isNonNegative() && (C - *C2).isNonNegative() &&
3137 computeConstantRange(X, /*ForSigned=*/true).add(*C2).isAllNonNegative())
3138 return new ICmpInst(ICmpInst::getSignedPredicate(Pred), X,
3139 ConstantInt::get(Ty, C - *C2));
3141 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
3142 const APInt &Upper = CR.getUpper();
3143 const APInt &Lower = CR.getLower();
3144 if (Cmp.isSigned()) {
3145 if (Lower.isSignMask())
3146 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
3147 if (Upper.isSignMask())
3148 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
3149 } else {
3150 if (Lower.isMinValue())
3151 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
3152 if (Upper.isMinValue())
3153 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
3156 // This set of folds is intentionally placed after folds that use no-wrapping
3157 // flags because those folds are likely better for later analysis/codegen.
3158 const APInt SMax = APInt::getSignedMaxValue(Ty->getScalarSizeInBits());
3159 const APInt SMin = APInt::getSignedMinValue(Ty->getScalarSizeInBits());
3161 // Fold compare with offset to opposite sign compare if it eliminates offset:
3162 // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
3163 if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
3164 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, -(*C2)));
3166 // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
3167 if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
3168 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, ~(*C2)));
3170 // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
3171 if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
3172 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, SMax - C));
3174 // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
3175 if (Pred == CmpInst::ICMP_SLT && C == *C2)
3176 return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, C ^ SMax));
3178 // (X + -1) <u C --> X <=u C (if X is never null)
3179 if (Pred == CmpInst::ICMP_ULT && C2->isAllOnes()) {
3180 const SimplifyQuery Q = SQ.getWithInstruction(&Cmp);
3181 if (llvm::isKnownNonZero(X, Q))
3182 return new ICmpInst(ICmpInst::ICMP_ULE, X, ConstantInt::get(Ty, C));
3185 if (!Add->hasOneUse())
3186 return nullptr;
3188 // X+C <u C2 -> (X & -C2) == C
3189 // iff C & (C2-1) == 0
3190 // C2 is a power of 2
3191 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
3192 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
3193 ConstantExpr::getNeg(cast<Constant>(Y)));
3195 // X+C2 <u C -> (X & C) == 2C
3196 // iff C == -(C2)
3197 // C2 is a power of 2
3198 if (Pred == ICmpInst::ICMP_ULT && C2->isPowerOf2() && C == -*C2)
3199 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, C),
3200 ConstantInt::get(Ty, C * 2));
3202 // X+C >u C2 -> (X & ~C2) != C
3203 // iff C & C2 == 0
3204 // C2+1 is a power of 2
3205 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
3206 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
3207 ConstantExpr::getNeg(cast<Constant>(Y)));
3209 // The range test idiom can use either ult or ugt. Arbitrarily canonicalize
3210 // to the ult form.
3211 // X+C2 >u C -> X+(C2-C-1) <u ~C
3212 if (Pred == ICmpInst::ICMP_UGT)
3213 return new ICmpInst(ICmpInst::ICMP_ULT,
3214 Builder.CreateAdd(X, ConstantInt::get(Ty, *C2 - C - 1)),
3215 ConstantInt::get(Ty, ~C));
3217 // zext(V) + C2 pred C -> V + C3 pred' C4
3218 Value *V;
3219 if (match(X, m_ZExt(m_Value(V)))) {
3220 Type *NewCmpTy = V->getType();
3221 unsigned NewCmpBW = NewCmpTy->getScalarSizeInBits();
3222 if (shouldChangeType(Ty, NewCmpTy)) {
3223 if (CR.getActiveBits() <= NewCmpBW) {
3224 ConstantRange SrcCR = CR.truncate(NewCmpBW);
3225 CmpInst::Predicate EquivPred;
3226 APInt EquivInt;
3227 APInt EquivOffset;
3229 SrcCR.getEquivalentICmp(EquivPred, EquivInt, EquivOffset);
3230 return new ICmpInst(
3231 EquivPred,
3232 EquivOffset.isZero()
3234 : Builder.CreateAdd(V, ConstantInt::get(NewCmpTy, EquivOffset)),
3235 ConstantInt::get(NewCmpTy, EquivInt));
3240 return nullptr;
3243 bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
3244 Value *&RHS, ConstantInt *&Less,
3245 ConstantInt *&Equal,
3246 ConstantInt *&Greater) {
3247 // TODO: Generalize this to work with other comparison idioms or ensure
3248 // they get canonicalized into this form.
3250 // select i1 (a == b),
3251 // i32 Equal,
3252 // i32 (select i1 (a < b), i32 Less, i32 Greater)
3253 // where Equal, Less and Greater are placeholders for any three constants.
3254 CmpPredicate PredA;
3255 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
3256 !ICmpInst::isEquality(PredA))
3257 return false;
3258 Value *EqualVal = SI->getTrueValue();
3259 Value *UnequalVal = SI->getFalseValue();
3260 // We still can get non-canonical predicate here, so canonicalize.
3261 if (PredA == ICmpInst::ICMP_NE)
3262 std::swap(EqualVal, UnequalVal);
3263 if (!match(EqualVal, m_ConstantInt(Equal)))
3264 return false;
3265 CmpPredicate PredB;
3266 Value *LHS2, *RHS2;
3267 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
3268 m_ConstantInt(Less), m_ConstantInt(Greater))))
3269 return false;
3270 // We can get predicate mismatch here, so canonicalize if possible:
3271 // First, ensure that 'LHS' match.
3272 if (LHS2 != LHS) {
3273 // x sgt y <--> y slt x
3274 std::swap(LHS2, RHS2);
3275 PredB = ICmpInst::getSwappedPredicate(PredB);
3277 if (LHS2 != LHS)
3278 return false;
3279 // We also need to canonicalize 'RHS'.
3280 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
3281 // x sgt C-1 <--> x sge C <--> not(x slt C)
3282 auto FlippedStrictness =
3283 InstCombiner::getFlippedStrictnessPredicateAndConstant(
3284 PredB, cast<Constant>(RHS2));
3285 if (!FlippedStrictness)
3286 return false;
3287 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE &&
3288 "basic correctness failure");
3289 RHS2 = FlippedStrictness->second;
3290 // And kind-of perform the result swap.
3291 std::swap(Less, Greater);
3292 PredB = ICmpInst::ICMP_SLT;
3294 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
3297 Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
3298 SelectInst *Select,
3299 ConstantInt *C) {
3301 assert(C && "Cmp RHS should be a constant int!");
3302 // If we're testing a constant value against the result of a three way
3303 // comparison, the result can be expressed directly in terms of the
3304 // original values being compared. Note: We could possibly be more
3305 // aggressive here and remove the hasOneUse test. The original select is
3306 // really likely to simplify or sink when we remove a test of the result.
3307 Value *OrigLHS, *OrigRHS;
3308 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
3309 if (Cmp.hasOneUse() &&
3310 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
3311 C3GreaterThan)) {
3312 assert(C1LessThan && C2Equal && C3GreaterThan);
3314 bool TrueWhenLessThan = ICmpInst::compare(
3315 C1LessThan->getValue(), C->getValue(), Cmp.getPredicate());
3316 bool TrueWhenEqual = ICmpInst::compare(C2Equal->getValue(), C->getValue(),
3317 Cmp.getPredicate());
3318 bool TrueWhenGreaterThan = ICmpInst::compare(
3319 C3GreaterThan->getValue(), C->getValue(), Cmp.getPredicate());
3321 // This generates the new instruction that will replace the original Cmp
3322 // Instruction. Instead of enumerating the various combinations when
3323 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
3324 // false, we rely on chaining of ORs and future passes of InstCombine to
3325 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
3327 // When none of the three constants satisfy the predicate for the RHS (C),
3328 // the entire original Cmp can be simplified to a false.
3329 Value *Cond = Builder.getFalse();
3330 if (TrueWhenLessThan)
3331 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
3332 OrigLHS, OrigRHS));
3333 if (TrueWhenEqual)
3334 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
3335 OrigLHS, OrigRHS));
3336 if (TrueWhenGreaterThan)
3337 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
3338 OrigLHS, OrigRHS));
3340 return replaceInstUsesWith(Cmp, Cond);
3342 return nullptr;
3345 Instruction *InstCombinerImpl::foldICmpBitCast(ICmpInst &Cmp) {
3346 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
3347 if (!Bitcast)
3348 return nullptr;
3350 ICmpInst::Predicate Pred = Cmp.getPredicate();
3351 Value *Op1 = Cmp.getOperand(1);
3352 Value *BCSrcOp = Bitcast->getOperand(0);
3353 Type *SrcType = Bitcast->getSrcTy();
3354 Type *DstType = Bitcast->getType();
3356 // Make sure the bitcast doesn't change between scalar and vector and
3357 // doesn't change the number of vector elements.
3358 if (SrcType->isVectorTy() == DstType->isVectorTy() &&
3359 SrcType->getScalarSizeInBits() == DstType->getScalarSizeInBits()) {
3360 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
3361 Value *X;
3362 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
3363 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
3364 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
3365 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
3366 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
3367 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
3368 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
3369 match(Op1, m_Zero()))
3370 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
3372 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
3373 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
3374 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
3376 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
3377 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
3378 return new ICmpInst(Pred, X,
3379 ConstantInt::getAllOnesValue(X->getType()));
3382 // Zero-equality checks are preserved through unsigned floating-point casts:
3383 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
3384 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
3385 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
3386 if (Cmp.isEquality() && match(Op1, m_Zero()))
3387 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
3389 const APInt *C;
3390 bool TrueIfSigned;
3391 if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse()) {
3392 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
3393 // the FP extend/truncate because that cast does not change the sign-bit.
3394 // This is true for all standard IEEE-754 types and the X86 80-bit type.
3395 // The sign-bit is always the most significant bit in those types.
3396 if (isSignBitCheck(Pred, *C, TrueIfSigned) &&
3397 (match(BCSrcOp, m_FPExt(m_Value(X))) ||
3398 match(BCSrcOp, m_FPTrunc(m_Value(X))))) {
3399 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
3400 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
3401 Type *XType = X->getType();
3403 // We can't currently handle Power style floating point operations here.
3404 if (!(XType->isPPC_FP128Ty() || SrcType->isPPC_FP128Ty())) {
3405 Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
3406 if (auto *XVTy = dyn_cast<VectorType>(XType))
3407 NewType = VectorType::get(NewType, XVTy->getElementCount());
3408 Value *NewBitcast = Builder.CreateBitCast(X, NewType);
3409 if (TrueIfSigned)
3410 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
3411 ConstantInt::getNullValue(NewType));
3412 else
3413 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
3414 ConstantInt::getAllOnesValue(NewType));
3418 // icmp eq/ne (bitcast X to int), special fp -> llvm.is.fpclass(X, class)
3419 Type *FPType = SrcType->getScalarType();
3420 if (!Cmp.getParent()->getParent()->hasFnAttribute(
3421 Attribute::NoImplicitFloat) &&
3422 Cmp.isEquality() && FPType->isIEEELikeFPTy()) {
3423 FPClassTest Mask = APFloat(FPType->getFltSemantics(), *C).classify();
3424 if (Mask & (fcInf | fcZero)) {
3425 if (Pred == ICmpInst::ICMP_NE)
3426 Mask = ~Mask;
3427 return replaceInstUsesWith(Cmp,
3428 Builder.createIsFPClass(BCSrcOp, Mask));
3434 const APInt *C;
3435 if (!match(Cmp.getOperand(1), m_APInt(C)) || !DstType->isIntegerTy() ||
3436 !SrcType->isIntOrIntVectorTy())
3437 return nullptr;
3439 // If this is checking if all elements of a vector compare are set or not,
3440 // invert the casted vector equality compare and test if all compare
3441 // elements are clear or not. Compare against zero is generally easier for
3442 // analysis and codegen.
3443 // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
3444 // Example: are all elements equal? --> are zero elements not equal?
3445 // TODO: Try harder to reduce compare of 2 freely invertible operands?
3446 if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse()) {
3447 if (Value *NotBCSrcOp =
3448 getFreelyInverted(BCSrcOp, BCSrcOp->hasOneUse(), &Builder)) {
3449 Value *Cast = Builder.CreateBitCast(NotBCSrcOp, DstType);
3450 return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(DstType));
3454 // If this is checking if all elements of an extended vector are clear or not,
3455 // compare in a narrow type to eliminate the extend:
3456 // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
3457 Value *X;
3458 if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() &&
3459 match(BCSrcOp, m_ZExtOrSExt(m_Value(X)))) {
3460 if (auto *VecTy = dyn_cast<FixedVectorType>(X->getType())) {
3461 Type *NewType = Builder.getIntNTy(VecTy->getPrimitiveSizeInBits());
3462 Value *NewCast = Builder.CreateBitCast(X, NewType);
3463 return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(NewType));
3467 // Folding: icmp <pred> iN X, C
3468 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
3469 // and C is a splat of a K-bit pattern
3470 // and SC is a constant vector = <C', C', C', ..., C'>
3471 // Into:
3472 // %E = extractelement <M x iK> %vec, i32 C'
3473 // icmp <pred> iK %E, trunc(C)
3474 Value *Vec;
3475 ArrayRef<int> Mask;
3476 if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
3477 // Check whether every element of Mask is the same constant
3478 if (all_equal(Mask)) {
3479 auto *VecTy = cast<VectorType>(SrcType);
3480 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
3481 if (C->isSplat(EltTy->getBitWidth())) {
3482 // Fold the icmp based on the value of C
3483 // If C is M copies of an iK sized bit pattern,
3484 // then:
3485 // => %E = extractelement <N x iK> %vec, i32 Elem
3486 // icmp <pred> iK %SplatVal, <pattern>
3487 Value *Elem = Builder.getInt32(Mask[0]);
3488 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
3489 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
3490 return new ICmpInst(Pred, Extract, NewC);
3494 return nullptr;
3497 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3498 /// where X is some kind of instruction.
3499 Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
3500 const APInt *C;
3502 if (match(Cmp.getOperand(1), m_APInt(C))) {
3503 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0)))
3504 if (Instruction *I = foldICmpBinOpWithConstant(Cmp, BO, *C))
3505 return I;
3507 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0)))
3508 // For now, we only support constant integers while folding the
3509 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3510 // similar to the cases handled by binary ops above.
3511 if (auto *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
3512 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
3513 return I;
3515 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0)))
3516 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
3517 return I;
3519 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
3520 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
3521 return I;
3523 // (extractval ([s/u]subo X, Y), 0) == 0 --> X == Y
3524 // (extractval ([s/u]subo X, Y), 0) != 0 --> X != Y
3525 // TODO: This checks one-use, but that is not strictly necessary.
3526 Value *Cmp0 = Cmp.getOperand(0);
3527 Value *X, *Y;
3528 if (C->isZero() && Cmp.isEquality() && Cmp0->hasOneUse() &&
3529 (match(Cmp0,
3530 m_ExtractValue<0>(m_Intrinsic<Intrinsic::ssub_with_overflow>(
3531 m_Value(X), m_Value(Y)))) ||
3532 match(Cmp0,
3533 m_ExtractValue<0>(m_Intrinsic<Intrinsic::usub_with_overflow>(
3534 m_Value(X), m_Value(Y))))))
3535 return new ICmpInst(Cmp.getPredicate(), X, Y);
3538 if (match(Cmp.getOperand(1), m_APIntAllowPoison(C)))
3539 return foldICmpInstWithConstantAllowPoison(Cmp, *C);
3541 return nullptr;
3544 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3545 /// icmp eq/ne BO, C.
3546 Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
3547 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3548 // TODO: Some of these folds could work with arbitrary constants, but this
3549 // function is limited to scalar and vector splat constants.
3550 if (!Cmp.isEquality())
3551 return nullptr;
3553 ICmpInst::Predicate Pred = Cmp.getPredicate();
3554 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3555 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
3556 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3558 switch (BO->getOpcode()) {
3559 case Instruction::SRem:
3560 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3561 if (C.isZero() && BO->hasOneUse()) {
3562 const APInt *BOC;
3563 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3564 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3565 return new ICmpInst(Pred, NewRem,
3566 Constant::getNullValue(BO->getType()));
3569 break;
3570 case Instruction::Add: {
3571 // (A + C2) == C --> A == (C - C2)
3572 // (A + C2) != C --> A != (C - C2)
3573 // TODO: Remove the one-use limitation? See discussion in D58633.
3574 if (Constant *C2 = dyn_cast<Constant>(BOp1)) {
3575 if (BO->hasOneUse())
3576 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, C2));
3577 } else if (C.isZero()) {
3578 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3579 // efficiently invertible, or if the add has just this one use.
3580 if (Value *NegVal = dyn_castNegVal(BOp1))
3581 return new ICmpInst(Pred, BOp0, NegVal);
3582 if (Value *NegVal = dyn_castNegVal(BOp0))
3583 return new ICmpInst(Pred, NegVal, BOp1);
3584 if (BO->hasOneUse()) {
3585 // (add nuw A, B) != 0 -> (or A, B) != 0
3586 if (match(BO, m_NUWAdd(m_Value(), m_Value()))) {
3587 Value *Or = Builder.CreateOr(BOp0, BOp1);
3588 return new ICmpInst(Pred, Or, Constant::getNullValue(BO->getType()));
3590 Value *Neg = Builder.CreateNeg(BOp1);
3591 Neg->takeName(BO);
3592 return new ICmpInst(Pred, BOp0, Neg);
3595 break;
3597 case Instruction::Xor:
3598 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3599 // For the xor case, we can xor two constants together, eliminating
3600 // the explicit xor.
3601 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3602 } else if (C.isZero()) {
3603 // Replace ((xor A, B) != 0) with (A != B)
3604 return new ICmpInst(Pred, BOp0, BOp1);
3606 break;
3607 case Instruction::Or: {
3608 const APInt *BOC;
3609 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3610 // Comparing if all bits outside of a constant mask are set?
3611 // Replace (X | C) == -1 with (X & ~C) == ~C.
3612 // This removes the -1 constant.
3613 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3614 Value *And = Builder.CreateAnd(BOp0, NotBOC);
3615 return new ICmpInst(Pred, And, NotBOC);
3617 // (icmp eq (or (select cond, 0, NonZero), Other), 0)
3618 // -> (and cond, (icmp eq Other, 0))
3619 // (icmp ne (or (select cond, NonZero, 0), Other), 0)
3620 // -> (or cond, (icmp ne Other, 0))
3621 Value *Cond, *TV, *FV, *Other, *Sel;
3622 if (C.isZero() &&
3623 match(BO,
3624 m_OneUse(m_c_Or(m_CombineAnd(m_Value(Sel),
3625 m_Select(m_Value(Cond), m_Value(TV),
3626 m_Value(FV))),
3627 m_Value(Other)))) &&
3628 Cond->getType() == Cmp.getType()) {
3629 const SimplifyQuery Q = SQ.getWithInstruction(&Cmp);
3630 // Easy case is if eq/ne matches whether 0 is trueval/falseval.
3631 if (Pred == ICmpInst::ICMP_EQ
3632 ? (match(TV, m_Zero()) && isKnownNonZero(FV, Q))
3633 : (match(FV, m_Zero()) && isKnownNonZero(TV, Q))) {
3634 Value *Cmp = Builder.CreateICmp(
3635 Pred, Other, Constant::getNullValue(Other->getType()));
3636 return BinaryOperator::Create(
3637 Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, Cmp,
3638 Cond);
3640 // Harder case is if eq/ne matches whether 0 is falseval/trueval. In this
3641 // case we need to invert the select condition so we need to be careful to
3642 // avoid creating extra instructions.
3643 // (icmp ne (or (select cond, 0, NonZero), Other), 0)
3644 // -> (or (not cond), (icmp ne Other, 0))
3645 // (icmp eq (or (select cond, NonZero, 0), Other), 0)
3646 // -> (and (not cond), (icmp eq Other, 0))
3648 // Only do this if the inner select has one use, in which case we are
3649 // replacing `select` with `(not cond)`. Otherwise, we will create more
3650 // uses. NB: Trying to freely invert cond doesn't make sense here, as if
3651 // cond was freely invertable, the select arms would have been inverted.
3652 if (Sel->hasOneUse() &&
3653 (Pred == ICmpInst::ICMP_EQ
3654 ? (match(FV, m_Zero()) && isKnownNonZero(TV, Q))
3655 : (match(TV, m_Zero()) && isKnownNonZero(FV, Q)))) {
3656 Value *NotCond = Builder.CreateNot(Cond);
3657 Value *Cmp = Builder.CreateICmp(
3658 Pred, Other, Constant::getNullValue(Other->getType()));
3659 return BinaryOperator::Create(
3660 Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, Cmp,
3661 NotCond);
3664 break;
3666 case Instruction::UDiv:
3667 case Instruction::SDiv:
3668 if (BO->isExact()) {
3669 // div exact X, Y eq/ne 0 -> X eq/ne 0
3670 // div exact X, Y eq/ne 1 -> X eq/ne Y
3671 // div exact X, Y eq/ne C ->
3672 // if Y * C never-overflow && OneUse:
3673 // -> Y * C eq/ne X
3674 if (C.isZero())
3675 return new ICmpInst(Pred, BOp0, Constant::getNullValue(BO->getType()));
3676 else if (C.isOne())
3677 return new ICmpInst(Pred, BOp0, BOp1);
3678 else if (BO->hasOneUse()) {
3679 OverflowResult OR = computeOverflow(
3680 Instruction::Mul, BO->getOpcode() == Instruction::SDiv, BOp1,
3681 Cmp.getOperand(1), BO);
3682 if (OR == OverflowResult::NeverOverflows) {
3683 Value *YC =
3684 Builder.CreateMul(BOp1, ConstantInt::get(BO->getType(), C));
3685 return new ICmpInst(Pred, YC, BOp0);
3689 if (BO->getOpcode() == Instruction::UDiv && C.isZero()) {
3690 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3691 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3692 return new ICmpInst(NewPred, BOp1, BOp0);
3694 break;
3695 default:
3696 break;
3698 return nullptr;
3701 static Instruction *foldCtpopPow2Test(ICmpInst &I, IntrinsicInst *CtpopLhs,
3702 const APInt &CRhs,
3703 InstCombiner::BuilderTy &Builder,
3704 const SimplifyQuery &Q) {
3705 assert(CtpopLhs->getIntrinsicID() == Intrinsic::ctpop &&
3706 "Non-ctpop intrin in ctpop fold");
3707 if (!CtpopLhs->hasOneUse())
3708 return nullptr;
3710 // Power of 2 test:
3711 // isPow2OrZero : ctpop(X) u< 2
3712 // isPow2 : ctpop(X) == 1
3713 // NotPow2OrZero: ctpop(X) u> 1
3714 // NotPow2 : ctpop(X) != 1
3715 // If we know any bit of X can be folded to:
3716 // IsPow2 : X & (~Bit) == 0
3717 // NotPow2 : X & (~Bit) != 0
3718 const ICmpInst::Predicate Pred = I.getPredicate();
3719 if (((I.isEquality() || Pred == ICmpInst::ICMP_UGT) && CRhs == 1) ||
3720 (Pred == ICmpInst::ICMP_ULT && CRhs == 2)) {
3721 Value *Op = CtpopLhs->getArgOperand(0);
3722 KnownBits OpKnown = computeKnownBits(Op, Q.DL,
3723 /*Depth*/ 0, Q.AC, Q.CxtI, Q.DT);
3724 // No need to check for count > 1, that should be already constant folded.
3725 if (OpKnown.countMinPopulation() == 1) {
3726 Value *And = Builder.CreateAnd(
3727 Op, Constant::getIntegerValue(Op->getType(), ~(OpKnown.One)));
3728 return new ICmpInst(
3729 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_ULT)
3730 ? ICmpInst::ICMP_EQ
3731 : ICmpInst::ICMP_NE,
3732 And, Constant::getNullValue(Op->getType()));
3736 return nullptr;
3739 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3740 Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3741 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3742 Type *Ty = II->getType();
3743 unsigned BitWidth = C.getBitWidth();
3744 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3746 switch (II->getIntrinsicID()) {
3747 case Intrinsic::abs:
3748 // abs(A) == 0 -> A == 0
3749 // abs(A) == INT_MIN -> A == INT_MIN
3750 if (C.isZero() || C.isMinSignedValue())
3751 return new ICmpInst(Pred, II->getArgOperand(0), ConstantInt::get(Ty, C));
3752 break;
3754 case Intrinsic::bswap:
3755 // bswap(A) == C -> A == bswap(C)
3756 return new ICmpInst(Pred, II->getArgOperand(0),
3757 ConstantInt::get(Ty, C.byteSwap()));
3759 case Intrinsic::bitreverse:
3760 // bitreverse(A) == C -> A == bitreverse(C)
3761 return new ICmpInst(Pred, II->getArgOperand(0),
3762 ConstantInt::get(Ty, C.reverseBits()));
3764 case Intrinsic::ctlz:
3765 case Intrinsic::cttz: {
3766 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3767 if (C == BitWidth)
3768 return new ICmpInst(Pred, II->getArgOperand(0),
3769 ConstantInt::getNullValue(Ty));
3771 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3772 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3773 // Limit to one use to ensure we don't increase instruction count.
3774 unsigned Num = C.getLimitedValue(BitWidth);
3775 if (Num != BitWidth && II->hasOneUse()) {
3776 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3777 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3778 : APInt::getHighBitsSet(BitWidth, Num + 1);
3779 APInt Mask2 = IsTrailing
3780 ? APInt::getOneBitSet(BitWidth, Num)
3781 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3782 return new ICmpInst(Pred, Builder.CreateAnd(II->getArgOperand(0), Mask1),
3783 ConstantInt::get(Ty, Mask2));
3785 break;
3788 case Intrinsic::ctpop: {
3789 // popcount(A) == 0 -> A == 0 and likewise for !=
3790 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3791 bool IsZero = C.isZero();
3792 if (IsZero || C == BitWidth)
3793 return new ICmpInst(Pred, II->getArgOperand(0),
3794 IsZero ? Constant::getNullValue(Ty)
3795 : Constant::getAllOnesValue(Ty));
3797 break;
3800 case Intrinsic::fshl:
3801 case Intrinsic::fshr:
3802 if (II->getArgOperand(0) == II->getArgOperand(1)) {
3803 const APInt *RotAmtC;
3804 // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3805 // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3806 if (match(II->getArgOperand(2), m_APInt(RotAmtC)))
3807 return new ICmpInst(Pred, II->getArgOperand(0),
3808 II->getIntrinsicID() == Intrinsic::fshl
3809 ? ConstantInt::get(Ty, C.rotr(*RotAmtC))
3810 : ConstantInt::get(Ty, C.rotl(*RotAmtC)));
3812 break;
3814 case Intrinsic::umax:
3815 case Intrinsic::uadd_sat: {
3816 // uadd.sat(a, b) == 0 -> (a | b) == 0
3817 // umax(a, b) == 0 -> (a | b) == 0
3818 if (C.isZero() && II->hasOneUse()) {
3819 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3820 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3822 break;
3825 case Intrinsic::ssub_sat:
3826 // ssub.sat(a, b) == 0 -> a == b
3827 if (C.isZero())
3828 return new ICmpInst(Pred, II->getArgOperand(0), II->getArgOperand(1));
3829 break;
3830 case Intrinsic::usub_sat: {
3831 // usub.sat(a, b) == 0 -> a <= b
3832 if (C.isZero()) {
3833 ICmpInst::Predicate NewPred =
3834 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3835 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3837 break;
3839 default:
3840 break;
3843 return nullptr;
3846 /// Fold an icmp with LLVM intrinsics
3847 static Instruction *
3848 foldICmpIntrinsicWithIntrinsic(ICmpInst &Cmp,
3849 InstCombiner::BuilderTy &Builder) {
3850 assert(Cmp.isEquality());
3852 ICmpInst::Predicate Pred = Cmp.getPredicate();
3853 Value *Op0 = Cmp.getOperand(0);
3854 Value *Op1 = Cmp.getOperand(1);
3855 const auto *IIOp0 = dyn_cast<IntrinsicInst>(Op0);
3856 const auto *IIOp1 = dyn_cast<IntrinsicInst>(Op1);
3857 if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3858 return nullptr;
3860 switch (IIOp0->getIntrinsicID()) {
3861 case Intrinsic::bswap:
3862 case Intrinsic::bitreverse:
3863 // If both operands are byte-swapped or bit-reversed, just compare the
3864 // original values.
3865 return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3866 case Intrinsic::fshl:
3867 case Intrinsic::fshr: {
3868 // If both operands are rotated by same amount, just compare the
3869 // original values.
3870 if (IIOp0->getOperand(0) != IIOp0->getOperand(1))
3871 break;
3872 if (IIOp1->getOperand(0) != IIOp1->getOperand(1))
3873 break;
3874 if (IIOp0->getOperand(2) == IIOp1->getOperand(2))
3875 return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3877 // rotate(X, AmtX) == rotate(Y, AmtY)
3878 // -> rotate(X, AmtX - AmtY) == Y
3879 // Do this if either both rotates have one use or if only one has one use
3880 // and AmtX/AmtY are constants.
3881 unsigned OneUses = IIOp0->hasOneUse() + IIOp1->hasOneUse();
3882 if (OneUses == 2 ||
3883 (OneUses == 1 && match(IIOp0->getOperand(2), m_ImmConstant()) &&
3884 match(IIOp1->getOperand(2), m_ImmConstant()))) {
3885 Value *SubAmt =
3886 Builder.CreateSub(IIOp0->getOperand(2), IIOp1->getOperand(2));
3887 Value *CombinedRotate = Builder.CreateIntrinsic(
3888 Op0->getType(), IIOp0->getIntrinsicID(),
3889 {IIOp0->getOperand(0), IIOp0->getOperand(0), SubAmt});
3890 return new ICmpInst(Pred, IIOp1->getOperand(0), CombinedRotate);
3892 } break;
3893 default:
3894 break;
3897 return nullptr;
3900 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3901 /// where X is some kind of instruction and C is AllowPoison.
3902 /// TODO: Move more folds which allow poison to this function.
3903 Instruction *
3904 InstCombinerImpl::foldICmpInstWithConstantAllowPoison(ICmpInst &Cmp,
3905 const APInt &C) {
3906 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3907 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0))) {
3908 switch (II->getIntrinsicID()) {
3909 default:
3910 break;
3911 case Intrinsic::fshl:
3912 case Intrinsic::fshr:
3913 if (Cmp.isEquality() && II->getArgOperand(0) == II->getArgOperand(1)) {
3914 // (rot X, ?) == 0/-1 --> X == 0/-1
3915 if (C.isZero() || C.isAllOnes())
3916 return new ICmpInst(Pred, II->getArgOperand(0), Cmp.getOperand(1));
3918 break;
3922 return nullptr;
3925 /// Fold an icmp with BinaryOp and constant operand: icmp Pred BO, C.
3926 Instruction *InstCombinerImpl::foldICmpBinOpWithConstant(ICmpInst &Cmp,
3927 BinaryOperator *BO,
3928 const APInt &C) {
3929 switch (BO->getOpcode()) {
3930 case Instruction::Xor:
3931 if (Instruction *I = foldICmpXorConstant(Cmp, BO, C))
3932 return I;
3933 break;
3934 case Instruction::And:
3935 if (Instruction *I = foldICmpAndConstant(Cmp, BO, C))
3936 return I;
3937 break;
3938 case Instruction::Or:
3939 if (Instruction *I = foldICmpOrConstant(Cmp, BO, C))
3940 return I;
3941 break;
3942 case Instruction::Mul:
3943 if (Instruction *I = foldICmpMulConstant(Cmp, BO, C))
3944 return I;
3945 break;
3946 case Instruction::Shl:
3947 if (Instruction *I = foldICmpShlConstant(Cmp, BO, C))
3948 return I;
3949 break;
3950 case Instruction::LShr:
3951 case Instruction::AShr:
3952 if (Instruction *I = foldICmpShrConstant(Cmp, BO, C))
3953 return I;
3954 break;
3955 case Instruction::SRem:
3956 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, C))
3957 return I;
3958 break;
3959 case Instruction::UDiv:
3960 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, C))
3961 return I;
3962 [[fallthrough]];
3963 case Instruction::SDiv:
3964 if (Instruction *I = foldICmpDivConstant(Cmp, BO, C))
3965 return I;
3966 break;
3967 case Instruction::Sub:
3968 if (Instruction *I = foldICmpSubConstant(Cmp, BO, C))
3969 return I;
3970 break;
3971 case Instruction::Add:
3972 if (Instruction *I = foldICmpAddConstant(Cmp, BO, C))
3973 return I;
3974 break;
3975 default:
3976 break;
3979 // TODO: These folds could be refactored to be part of the above calls.
3980 return foldICmpBinOpEqualityWithConstant(Cmp, BO, C);
3983 static Instruction *
3984 foldICmpUSubSatOrUAddSatWithConstant(CmpPredicate Pred, SaturatingInst *II,
3985 const APInt &C,
3986 InstCombiner::BuilderTy &Builder) {
3987 // This transform may end up producing more than one instruction for the
3988 // intrinsic, so limit it to one user of the intrinsic.
3989 if (!II->hasOneUse())
3990 return nullptr;
3992 // Let Y = [add/sub]_sat(X, C) pred C2
3993 // SatVal = The saturating value for the operation
3994 // WillWrap = Whether or not the operation will underflow / overflow
3995 // => Y = (WillWrap ? SatVal : (X binop C)) pred C2
3996 // => Y = WillWrap ? (SatVal pred C2) : ((X binop C) pred C2)
3998 // When (SatVal pred C2) is true, then
3999 // Y = WillWrap ? true : ((X binop C) pred C2)
4000 // => Y = WillWrap || ((X binop C) pred C2)
4001 // else
4002 // Y = WillWrap ? false : ((X binop C) pred C2)
4003 // => Y = !WillWrap ? ((X binop C) pred C2) : false
4004 // => Y = !WillWrap && ((X binop C) pred C2)
4005 Value *Op0 = II->getOperand(0);
4006 Value *Op1 = II->getOperand(1);
4008 const APInt *COp1;
4009 // This transform only works when the intrinsic has an integral constant or
4010 // splat vector as the second operand.
4011 if (!match(Op1, m_APInt(COp1)))
4012 return nullptr;
4014 APInt SatVal;
4015 switch (II->getIntrinsicID()) {
4016 default:
4017 llvm_unreachable(
4018 "This function only works with usub_sat and uadd_sat for now!");
4019 case Intrinsic::uadd_sat:
4020 SatVal = APInt::getAllOnes(C.getBitWidth());
4021 break;
4022 case Intrinsic::usub_sat:
4023 SatVal = APInt::getZero(C.getBitWidth());
4024 break;
4027 // Check (SatVal pred C2)
4028 bool SatValCheck = ICmpInst::compare(SatVal, C, Pred);
4030 // !WillWrap.
4031 ConstantRange C1 = ConstantRange::makeExactNoWrapRegion(
4032 II->getBinaryOp(), *COp1, II->getNoWrapKind());
4034 // WillWrap.
4035 if (SatValCheck)
4036 C1 = C1.inverse();
4038 ConstantRange C2 = ConstantRange::makeExactICmpRegion(Pred, C);
4039 if (II->getBinaryOp() == Instruction::Add)
4040 C2 = C2.sub(*COp1);
4041 else
4042 C2 = C2.add(*COp1);
4044 Instruction::BinaryOps CombiningOp =
4045 SatValCheck ? Instruction::BinaryOps::Or : Instruction::BinaryOps::And;
4047 std::optional<ConstantRange> Combination;
4048 if (CombiningOp == Instruction::BinaryOps::Or)
4049 Combination = C1.exactUnionWith(C2);
4050 else /* CombiningOp == Instruction::BinaryOps::And */
4051 Combination = C1.exactIntersectWith(C2);
4053 if (!Combination)
4054 return nullptr;
4056 CmpInst::Predicate EquivPred;
4057 APInt EquivInt;
4058 APInt EquivOffset;
4060 Combination->getEquivalentICmp(EquivPred, EquivInt, EquivOffset);
4062 return new ICmpInst(
4063 EquivPred,
4064 Builder.CreateAdd(Op0, ConstantInt::get(Op1->getType(), EquivOffset)),
4065 ConstantInt::get(Op1->getType(), EquivInt));
4068 static Instruction *
4069 foldICmpOfCmpIntrinsicWithConstant(CmpPredicate Pred, IntrinsicInst *I,
4070 const APInt &C,
4071 InstCombiner::BuilderTy &Builder) {
4072 std::optional<ICmpInst::Predicate> NewPredicate = std::nullopt;
4073 switch (Pred) {
4074 case ICmpInst::ICMP_EQ:
4075 case ICmpInst::ICMP_NE:
4076 if (C.isZero())
4077 NewPredicate = Pred;
4078 else if (C.isOne())
4079 NewPredicate =
4080 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE;
4081 else if (C.isAllOnes())
4082 NewPredicate =
4083 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
4084 break;
4086 case ICmpInst::ICMP_SGT:
4087 if (C.isAllOnes())
4088 NewPredicate = ICmpInst::ICMP_UGE;
4089 else if (C.isZero())
4090 NewPredicate = ICmpInst::ICMP_UGT;
4091 break;
4093 case ICmpInst::ICMP_SLT:
4094 if (C.isZero())
4095 NewPredicate = ICmpInst::ICMP_ULT;
4096 else if (C.isOne())
4097 NewPredicate = ICmpInst::ICMP_ULE;
4098 break;
4100 case ICmpInst::ICMP_ULT:
4101 if (C.ugt(1))
4102 NewPredicate = ICmpInst::ICMP_UGE;
4103 break;
4105 case ICmpInst::ICMP_UGT:
4106 if (!C.isZero() && !C.isAllOnes())
4107 NewPredicate = ICmpInst::ICMP_ULT;
4108 break;
4110 default:
4111 break;
4114 if (!NewPredicate)
4115 return nullptr;
4117 if (I->getIntrinsicID() == Intrinsic::scmp)
4118 NewPredicate = ICmpInst::getSignedPredicate(*NewPredicate);
4119 Value *LHS = I->getOperand(0);
4120 Value *RHS = I->getOperand(1);
4121 return new ICmpInst(*NewPredicate, LHS, RHS);
4124 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
4125 Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
4126 IntrinsicInst *II,
4127 const APInt &C) {
4128 ICmpInst::Predicate Pred = Cmp.getPredicate();
4130 // Handle folds that apply for any kind of icmp.
4131 switch (II->getIntrinsicID()) {
4132 default:
4133 break;
4134 case Intrinsic::uadd_sat:
4135 case Intrinsic::usub_sat:
4136 if (auto *Folded = foldICmpUSubSatOrUAddSatWithConstant(
4137 Pred, cast<SaturatingInst>(II), C, Builder))
4138 return Folded;
4139 break;
4140 case Intrinsic::ctpop: {
4141 const SimplifyQuery Q = SQ.getWithInstruction(&Cmp);
4142 if (Instruction *R = foldCtpopPow2Test(Cmp, II, C, Builder, Q))
4143 return R;
4144 } break;
4145 case Intrinsic::scmp:
4146 case Intrinsic::ucmp:
4147 if (auto *Folded = foldICmpOfCmpIntrinsicWithConstant(Pred, II, C, Builder))
4148 return Folded;
4149 break;
4152 if (Cmp.isEquality())
4153 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
4155 Type *Ty = II->getType();
4156 unsigned BitWidth = C.getBitWidth();
4157 switch (II->getIntrinsicID()) {
4158 case Intrinsic::ctpop: {
4159 // (ctpop X > BitWidth - 1) --> X == -1
4160 Value *X = II->getArgOperand(0);
4161 if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
4162 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
4163 ConstantInt::getAllOnesValue(Ty));
4164 // (ctpop X < BitWidth) --> X != -1
4165 if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
4166 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
4167 ConstantInt::getAllOnesValue(Ty));
4168 break;
4170 case Intrinsic::ctlz: {
4171 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
4172 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
4173 unsigned Num = C.getLimitedValue();
4174 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
4175 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
4176 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
4179 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
4180 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
4181 unsigned Num = C.getLimitedValue();
4182 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
4183 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
4184 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
4186 break;
4188 case Intrinsic::cttz: {
4189 // Limit to one use to ensure we don't increase instruction count.
4190 if (!II->hasOneUse())
4191 return nullptr;
4193 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
4194 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
4195 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
4196 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
4197 Builder.CreateAnd(II->getArgOperand(0), Mask),
4198 ConstantInt::getNullValue(Ty));
4201 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
4202 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
4203 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
4204 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
4205 Builder.CreateAnd(II->getArgOperand(0), Mask),
4206 ConstantInt::getNullValue(Ty));
4208 break;
4210 case Intrinsic::ssub_sat:
4211 // ssub.sat(a, b) spred 0 -> a spred b
4212 if (ICmpInst::isSigned(Pred)) {
4213 if (C.isZero())
4214 return new ICmpInst(Pred, II->getArgOperand(0), II->getArgOperand(1));
4215 // X s<= 0 is cannonicalized to X s< 1
4216 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
4217 return new ICmpInst(ICmpInst::ICMP_SLE, II->getArgOperand(0),
4218 II->getArgOperand(1));
4219 // X s>= 0 is cannonicalized to X s> -1
4220 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
4221 return new ICmpInst(ICmpInst::ICMP_SGE, II->getArgOperand(0),
4222 II->getArgOperand(1));
4224 break;
4225 default:
4226 break;
4229 return nullptr;
4232 /// Handle icmp with constant (but not simple integer constant) RHS.
4233 Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
4234 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4235 Constant *RHSC = dyn_cast<Constant>(Op1);
4236 Instruction *LHSI = dyn_cast<Instruction>(Op0);
4237 if (!RHSC || !LHSI)
4238 return nullptr;
4240 switch (LHSI->getOpcode()) {
4241 case Instruction::PHI:
4242 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
4243 return NV;
4244 break;
4245 case Instruction::IntToPtr:
4246 // icmp pred inttoptr(X), null -> icmp pred X, 0
4247 if (RHSC->isNullValue() &&
4248 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
4249 return new ICmpInst(
4250 I.getPredicate(), LHSI->getOperand(0),
4251 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4252 break;
4254 case Instruction::Load:
4255 // Try to optimize things like "A[i] > 4" to index computations.
4256 if (GetElementPtrInst *GEP =
4257 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
4258 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
4259 if (Instruction *Res =
4260 foldCmpLoadFromIndexedGlobal(cast<LoadInst>(LHSI), GEP, GV, I))
4261 return Res;
4262 break;
4265 return nullptr;
4268 Instruction *InstCombinerImpl::foldSelectICmp(CmpPredicate Pred, SelectInst *SI,
4269 Value *RHS, const ICmpInst &I) {
4270 // Try to fold the comparison into the select arms, which will cause the
4271 // select to be converted into a logical and/or.
4272 auto SimplifyOp = [&](Value *Op, bool SelectCondIsTrue) -> Value * {
4273 if (Value *Res = simplifyICmpInst(Pred, Op, RHS, SQ))
4274 return Res;
4275 if (std::optional<bool> Impl = isImpliedCondition(
4276 SI->getCondition(), Pred, Op, RHS, DL, SelectCondIsTrue))
4277 return ConstantInt::get(I.getType(), *Impl);
4278 return nullptr;
4281 ConstantInt *CI = nullptr;
4282 Value *Op1 = SimplifyOp(SI->getOperand(1), true);
4283 if (Op1)
4284 CI = dyn_cast<ConstantInt>(Op1);
4286 Value *Op2 = SimplifyOp(SI->getOperand(2), false);
4287 if (Op2)
4288 CI = dyn_cast<ConstantInt>(Op2);
4290 auto Simplifies = [&](Value *Op, unsigned Idx) {
4291 // A comparison of ucmp/scmp with a constant will fold into an icmp.
4292 const APInt *Dummy;
4293 return Op ||
4294 (isa<CmpIntrinsic>(SI->getOperand(Idx)) &&
4295 SI->getOperand(Idx)->hasOneUse() && match(RHS, m_APInt(Dummy)));
4298 // We only want to perform this transformation if it will not lead to
4299 // additional code. This is true if either both sides of the select
4300 // fold to a constant (in which case the icmp is replaced with a select
4301 // which will usually simplify) or this is the only user of the
4302 // select (in which case we are trading a select+icmp for a simpler
4303 // select+icmp) or all uses of the select can be replaced based on
4304 // dominance information ("Global cases").
4305 bool Transform = false;
4306 if (Op1 && Op2)
4307 Transform = true;
4308 else if (Simplifies(Op1, 1) || Simplifies(Op2, 2)) {
4309 // Local case
4310 if (SI->hasOneUse())
4311 Transform = true;
4312 // Global cases
4313 else if (CI && !CI->isZero())
4314 // When Op1 is constant try replacing select with second operand.
4315 // Otherwise Op2 is constant and try replacing select with first
4316 // operand.
4317 Transform = replacedSelectWithOperand(SI, &I, Op1 ? 2 : 1);
4319 if (Transform) {
4320 if (!Op1)
4321 Op1 = Builder.CreateICmp(Pred, SI->getOperand(1), RHS, I.getName());
4322 if (!Op2)
4323 Op2 = Builder.CreateICmp(Pred, SI->getOperand(2), RHS, I.getName());
4324 return SelectInst::Create(SI->getOperand(0), Op1, Op2);
4327 return nullptr;
4330 // Returns whether V is a Mask ((X + 1) & X == 0) or ~Mask (-Pow2OrZero)
4331 static bool isMaskOrZero(const Value *V, bool Not, const SimplifyQuery &Q,
4332 unsigned Depth = 0) {
4333 if (Not ? match(V, m_NegatedPower2OrZero()) : match(V, m_LowBitMaskOrZero()))
4334 return true;
4335 if (V->getType()->getScalarSizeInBits() == 1)
4336 return true;
4337 if (Depth++ >= MaxAnalysisRecursionDepth)
4338 return false;
4339 Value *X;
4340 const Instruction *I = dyn_cast<Instruction>(V);
4341 if (!I)
4342 return false;
4343 switch (I->getOpcode()) {
4344 case Instruction::ZExt:
4345 // ZExt(Mask) is a Mask.
4346 return !Not && isMaskOrZero(I->getOperand(0), Not, Q, Depth);
4347 case Instruction::SExt:
4348 // SExt(Mask) is a Mask.
4349 // SExt(~Mask) is a ~Mask.
4350 return isMaskOrZero(I->getOperand(0), Not, Q, Depth);
4351 case Instruction::And:
4352 case Instruction::Or:
4353 // Mask0 | Mask1 is a Mask.
4354 // Mask0 & Mask1 is a Mask.
4355 // ~Mask0 | ~Mask1 is a ~Mask.
4356 // ~Mask0 & ~Mask1 is a ~Mask.
4357 return isMaskOrZero(I->getOperand(1), Not, Q, Depth) &&
4358 isMaskOrZero(I->getOperand(0), Not, Q, Depth);
4359 case Instruction::Xor:
4360 if (match(V, m_Not(m_Value(X))))
4361 return isMaskOrZero(X, !Not, Q, Depth);
4363 // (X ^ -X) is a ~Mask
4364 if (Not)
4365 return match(V, m_c_Xor(m_Value(X), m_Neg(m_Deferred(X))));
4366 // (X ^ (X - 1)) is a Mask
4367 else
4368 return match(V, m_c_Xor(m_Value(X), m_Add(m_Deferred(X), m_AllOnes())));
4369 case Instruction::Select:
4370 // c ? Mask0 : Mask1 is a Mask.
4371 return isMaskOrZero(I->getOperand(1), Not, Q, Depth) &&
4372 isMaskOrZero(I->getOperand(2), Not, Q, Depth);
4373 case Instruction::Shl:
4374 // (~Mask) << X is a ~Mask.
4375 return Not && isMaskOrZero(I->getOperand(0), Not, Q, Depth);
4376 case Instruction::LShr:
4377 // Mask >> X is a Mask.
4378 return !Not && isMaskOrZero(I->getOperand(0), Not, Q, Depth);
4379 case Instruction::AShr:
4380 // Mask s>> X is a Mask.
4381 // ~Mask s>> X is a ~Mask.
4382 return isMaskOrZero(I->getOperand(0), Not, Q, Depth);
4383 case Instruction::Add:
4384 // Pow2 - 1 is a Mask.
4385 if (!Not && match(I->getOperand(1), m_AllOnes()))
4386 return isKnownToBeAPowerOfTwo(I->getOperand(0), Q.DL, /*OrZero*/ true,
4387 Depth, Q.AC, Q.CxtI, Q.DT);
4388 break;
4389 case Instruction::Sub:
4390 // -Pow2 is a ~Mask.
4391 if (Not && match(I->getOperand(0), m_Zero()))
4392 return isKnownToBeAPowerOfTwo(I->getOperand(1), Q.DL, /*OrZero*/ true,
4393 Depth, Q.AC, Q.CxtI, Q.DT);
4394 break;
4395 case Instruction::Call: {
4396 if (auto *II = dyn_cast<IntrinsicInst>(I)) {
4397 switch (II->getIntrinsicID()) {
4398 // min/max(Mask0, Mask1) is a Mask.
4399 // min/max(~Mask0, ~Mask1) is a ~Mask.
4400 case Intrinsic::umax:
4401 case Intrinsic::smax:
4402 case Intrinsic::umin:
4403 case Intrinsic::smin:
4404 return isMaskOrZero(II->getArgOperand(1), Not, Q, Depth) &&
4405 isMaskOrZero(II->getArgOperand(0), Not, Q, Depth);
4407 // In the context of masks, bitreverse(Mask) == ~Mask
4408 case Intrinsic::bitreverse:
4409 return isMaskOrZero(II->getArgOperand(0), !Not, Q, Depth);
4410 default:
4411 break;
4414 break;
4416 default:
4417 break;
4419 return false;
4422 /// Some comparisons can be simplified.
4423 /// In this case, we are looking for comparisons that look like
4424 /// a check for a lossy truncation.
4425 /// Folds:
4426 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
4427 /// icmp SrcPred (x & ~Mask), ~Mask to icmp DstPred x, ~Mask
4428 /// icmp eq/ne (x & ~Mask), 0 to icmp DstPred x, Mask
4429 /// icmp eq/ne (~x | Mask), -1 to icmp DstPred x, Mask
4430 /// Where Mask is some pattern that produces all-ones in low bits:
4431 /// (-1 >> y)
4432 /// ((-1 << y) >> y) <- non-canonical, has extra uses
4433 /// ~(-1 << y)
4434 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
4435 /// The Mask can be a constant, too.
4436 /// For some predicates, the operands are commutative.
4437 /// For others, x can only be on a specific side.
4438 static Value *foldICmpWithLowBitMaskedVal(CmpPredicate Pred, Value *Op0,
4439 Value *Op1, const SimplifyQuery &Q,
4440 InstCombiner &IC) {
4442 ICmpInst::Predicate DstPred;
4443 switch (Pred) {
4444 case ICmpInst::Predicate::ICMP_EQ:
4445 // x & Mask == x
4446 // x & ~Mask == 0
4447 // ~x | Mask == -1
4448 // -> x u<= Mask
4449 // x & ~Mask == ~Mask
4450 // -> ~Mask u<= x
4451 DstPred = ICmpInst::Predicate::ICMP_ULE;
4452 break;
4453 case ICmpInst::Predicate::ICMP_NE:
4454 // x & Mask != x
4455 // x & ~Mask != 0
4456 // ~x | Mask != -1
4457 // -> x u> Mask
4458 // x & ~Mask != ~Mask
4459 // -> ~Mask u> x
4460 DstPred = ICmpInst::Predicate::ICMP_UGT;
4461 break;
4462 case ICmpInst::Predicate::ICMP_ULT:
4463 // x & Mask u< x
4464 // -> x u> Mask
4465 // x & ~Mask u< ~Mask
4466 // -> ~Mask u> x
4467 DstPred = ICmpInst::Predicate::ICMP_UGT;
4468 break;
4469 case ICmpInst::Predicate::ICMP_UGE:
4470 // x & Mask u>= x
4471 // -> x u<= Mask
4472 // x & ~Mask u>= ~Mask
4473 // -> ~Mask u<= x
4474 DstPred = ICmpInst::Predicate::ICMP_ULE;
4475 break;
4476 case ICmpInst::Predicate::ICMP_SLT:
4477 // x & Mask s< x [iff Mask s>= 0]
4478 // -> x s> Mask
4479 // x & ~Mask s< ~Mask [iff ~Mask != 0]
4480 // -> ~Mask s> x
4481 DstPred = ICmpInst::Predicate::ICMP_SGT;
4482 break;
4483 case ICmpInst::Predicate::ICMP_SGE:
4484 // x & Mask s>= x [iff Mask s>= 0]
4485 // -> x s<= Mask
4486 // x & ~Mask s>= ~Mask [iff ~Mask != 0]
4487 // -> ~Mask s<= x
4488 DstPred = ICmpInst::Predicate::ICMP_SLE;
4489 break;
4490 default:
4491 // We don't support sgt,sle
4492 // ult/ugt are simplified to true/false respectively.
4493 return nullptr;
4496 Value *X, *M;
4497 // Put search code in lambda for early positive returns.
4498 auto IsLowBitMask = [&]() {
4499 if (match(Op0, m_c_And(m_Specific(Op1), m_Value(M)))) {
4500 X = Op1;
4501 // Look for: x & Mask pred x
4502 if (isMaskOrZero(M, /*Not=*/false, Q)) {
4503 return !ICmpInst::isSigned(Pred) ||
4504 (match(M, m_NonNegative()) || isKnownNonNegative(M, Q));
4507 // Look for: x & ~Mask pred ~Mask
4508 if (isMaskOrZero(X, /*Not=*/true, Q)) {
4509 return !ICmpInst::isSigned(Pred) || isKnownNonZero(X, Q);
4511 return false;
4513 if (ICmpInst::isEquality(Pred) && match(Op1, m_AllOnes()) &&
4514 match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(M))))) {
4516 auto Check = [&]() {
4517 // Look for: ~x | Mask == -1
4518 if (isMaskOrZero(M, /*Not=*/false, Q)) {
4519 if (Value *NotX =
4520 IC.getFreelyInverted(X, X->hasOneUse(), &IC.Builder)) {
4521 X = NotX;
4522 return true;
4525 return false;
4527 if (Check())
4528 return true;
4529 std::swap(X, M);
4530 return Check();
4532 if (ICmpInst::isEquality(Pred) && match(Op1, m_Zero()) &&
4533 match(Op0, m_OneUse(m_And(m_Value(X), m_Value(M))))) {
4534 auto Check = [&]() {
4535 // Look for: x & ~Mask == 0
4536 if (isMaskOrZero(M, /*Not=*/true, Q)) {
4537 if (Value *NotM =
4538 IC.getFreelyInverted(M, M->hasOneUse(), &IC.Builder)) {
4539 M = NotM;
4540 return true;
4543 return false;
4545 if (Check())
4546 return true;
4547 std::swap(X, M);
4548 return Check();
4550 return false;
4553 if (!IsLowBitMask())
4554 return nullptr;
4556 return IC.Builder.CreateICmp(DstPred, X, M);
4559 /// Some comparisons can be simplified.
4560 /// In this case, we are looking for comparisons that look like
4561 /// a check for a lossy signed truncation.
4562 /// Folds: (MaskedBits is a constant.)
4563 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
4564 /// Into:
4565 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
4566 /// Where KeptBits = bitwidth(%x) - MaskedBits
4567 static Value *
4568 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
4569 InstCombiner::BuilderTy &Builder) {
4570 CmpPredicate SrcPred;
4571 Value *X;
4572 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
4573 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
4574 if (!match(&I, m_c_ICmp(SrcPred,
4575 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
4576 m_APInt(C1))),
4577 m_Deferred(X))))
4578 return nullptr;
4580 // Potential handling of non-splats: for each element:
4581 // * if both are undef, replace with constant 0.
4582 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
4583 // * if both are not undef, and are different, bailout.
4584 // * else, only one is undef, then pick the non-undef one.
4586 // The shift amount must be equal.
4587 if (*C0 != *C1)
4588 return nullptr;
4589 const APInt &MaskedBits = *C0;
4590 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
4592 ICmpInst::Predicate DstPred;
4593 switch (SrcPred) {
4594 case ICmpInst::Predicate::ICMP_EQ:
4595 // ((%x << MaskedBits) a>> MaskedBits) == %x
4596 // =>
4597 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
4598 DstPred = ICmpInst::Predicate::ICMP_ULT;
4599 break;
4600 case ICmpInst::Predicate::ICMP_NE:
4601 // ((%x << MaskedBits) a>> MaskedBits) != %x
4602 // =>
4603 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
4604 DstPred = ICmpInst::Predicate::ICMP_UGE;
4605 break;
4606 // FIXME: are more folds possible?
4607 default:
4608 return nullptr;
4611 auto *XType = X->getType();
4612 const unsigned XBitWidth = XType->getScalarSizeInBits();
4613 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
4614 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
4616 // KeptBits = bitwidth(%x) - MaskedBits
4617 const APInt KeptBits = BitWidth - MaskedBits;
4618 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
4619 // ICmpCst = (1 << KeptBits)
4620 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
4621 assert(ICmpCst.isPowerOf2());
4622 // AddCst = (1 << (KeptBits-1))
4623 const APInt AddCst = ICmpCst.lshr(1);
4624 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
4626 // T0 = add %x, AddCst
4627 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
4628 // T1 = T0 DstPred ICmpCst
4629 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
4631 return T1;
4634 // Given pattern:
4635 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
4636 // we should move shifts to the same hand of 'and', i.e. rewrite as
4637 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
4638 // We are only interested in opposite logical shifts here.
4639 // One of the shifts can be truncated.
4640 // If we can, we want to end up creating 'lshr' shift.
4641 static Value *
4642 foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
4643 InstCombiner::BuilderTy &Builder) {
4644 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
4645 !I.getOperand(0)->hasOneUse())
4646 return nullptr;
4648 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
4650 // Look for an 'and' of two logical shifts, one of which may be truncated.
4651 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
4652 Instruction *XShift, *MaybeTruncation, *YShift;
4653 if (!match(
4654 I.getOperand(0),
4655 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
4656 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
4657 m_AnyLogicalShift, m_Instruction(YShift))),
4658 m_Instruction(MaybeTruncation)))))
4659 return nullptr;
4661 // We potentially looked past 'trunc', but only when matching YShift,
4662 // therefore YShift must have the widest type.
4663 Instruction *WidestShift = YShift;
4664 // Therefore XShift must have the shallowest type.
4665 // Or they both have identical types if there was no truncation.
4666 Instruction *NarrowestShift = XShift;
4668 Type *WidestTy = WidestShift->getType();
4669 Type *NarrowestTy = NarrowestShift->getType();
4670 assert(NarrowestTy == I.getOperand(0)->getType() &&
4671 "We did not look past any shifts while matching XShift though.");
4672 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
4674 // If YShift is a 'lshr', swap the shifts around.
4675 if (match(YShift, m_LShr(m_Value(), m_Value())))
4676 std::swap(XShift, YShift);
4678 // The shifts must be in opposite directions.
4679 auto XShiftOpcode = XShift->getOpcode();
4680 if (XShiftOpcode == YShift->getOpcode())
4681 return nullptr; // Do not care about same-direction shifts here.
4683 Value *X, *XShAmt, *Y, *YShAmt;
4684 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
4685 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
4687 // If one of the values being shifted is a constant, then we will end with
4688 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
4689 // however, we will need to ensure that we won't increase instruction count.
4690 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
4691 // At least one of the hands of the 'and' should be one-use shift.
4692 if (!match(I.getOperand(0),
4693 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
4694 return nullptr;
4695 if (HadTrunc) {
4696 // Due to the 'trunc', we will need to widen X. For that either the old
4697 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
4698 if (!MaybeTruncation->hasOneUse() &&
4699 !NarrowestShift->getOperand(1)->hasOneUse())
4700 return nullptr;
4704 // We have two shift amounts from two different shifts. The types of those
4705 // shift amounts may not match. If that's the case let's bailout now.
4706 if (XShAmt->getType() != YShAmt->getType())
4707 return nullptr;
4709 // As input, we have the following pattern:
4710 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
4711 // We want to rewrite that as:
4712 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
4713 // While we know that originally (Q+K) would not overflow
4714 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
4715 // shift amounts. so it may now overflow in smaller bitwidth.
4716 // To ensure that does not happen, we need to ensure that the total maximal
4717 // shift amount is still representable in that smaller bit width.
4718 unsigned MaximalPossibleTotalShiftAmount =
4719 (WidestTy->getScalarSizeInBits() - 1) +
4720 (NarrowestTy->getScalarSizeInBits() - 1);
4721 APInt MaximalRepresentableShiftAmount =
4722 APInt::getAllOnes(XShAmt->getType()->getScalarSizeInBits());
4723 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
4724 return nullptr;
4726 // Can we fold (XShAmt+YShAmt) ?
4727 auto *NewShAmt = dyn_cast_or_null<Constant>(
4728 simplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
4729 /*isNUW=*/false, SQ.getWithInstruction(&I)));
4730 if (!NewShAmt)
4731 return nullptr;
4732 if (NewShAmt->getType() != WidestTy) {
4733 NewShAmt =
4734 ConstantFoldCastOperand(Instruction::ZExt, NewShAmt, WidestTy, SQ.DL);
4735 if (!NewShAmt)
4736 return nullptr;
4738 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
4740 // Is the new shift amount smaller than the bit width?
4741 // FIXME: could also rely on ConstantRange.
4742 if (!match(NewShAmt,
4743 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
4744 APInt(WidestBitWidth, WidestBitWidth))))
4745 return nullptr;
4747 // An extra legality check is needed if we had trunc-of-lshr.
4748 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
4749 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
4750 WidestShift]() {
4751 // It isn't obvious whether it's worth it to analyze non-constants here.
4752 // Also, let's basically give up on non-splat cases, pessimizing vectors.
4753 // If *any* of these preconditions matches we can perform the fold.
4754 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
4755 ? NewShAmt->getSplatValue()
4756 : NewShAmt;
4757 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
4758 if (NewShAmtSplat &&
4759 (NewShAmtSplat->isNullValue() ||
4760 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
4761 return true;
4762 // We consider *min* leading zeros so a single outlier
4763 // blocks the transform as opposed to allowing it.
4764 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
4765 KnownBits Known = computeKnownBits(C, SQ.DL);
4766 unsigned MinLeadZero = Known.countMinLeadingZeros();
4767 // If the value being shifted has at most lowest bit set we can fold.
4768 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
4769 if (MaxActiveBits <= 1)
4770 return true;
4771 // Precondition: NewShAmt u<= countLeadingZeros(C)
4772 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
4773 return true;
4775 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
4776 KnownBits Known = computeKnownBits(C, SQ.DL);
4777 unsigned MinLeadZero = Known.countMinLeadingZeros();
4778 // If the value being shifted has at most lowest bit set we can fold.
4779 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
4780 if (MaxActiveBits <= 1)
4781 return true;
4782 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
4783 if (NewShAmtSplat) {
4784 APInt AdjNewShAmt =
4785 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
4786 if (AdjNewShAmt.ule(MinLeadZero))
4787 return true;
4790 return false; // Can't tell if it's ok.
4792 if (!CanFold())
4793 return nullptr;
4796 // All good, we can do this fold.
4797 X = Builder.CreateZExt(X, WidestTy);
4798 Y = Builder.CreateZExt(Y, WidestTy);
4799 // The shift is the same that was for X.
4800 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
4801 ? Builder.CreateLShr(X, NewShAmt)
4802 : Builder.CreateShl(X, NewShAmt);
4803 Value *T1 = Builder.CreateAnd(T0, Y);
4804 return Builder.CreateICmp(I.getPredicate(), T1,
4805 Constant::getNullValue(WidestTy));
4808 /// Fold
4809 /// (-1 u/ x) u< y
4810 /// ((x * y) ?/ x) != y
4811 /// to
4812 /// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
4813 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
4814 /// will mean that we are looking for the opposite answer.
4815 Value *InstCombinerImpl::foldMultiplicationOverflowCheck(ICmpInst &I) {
4816 CmpPredicate Pred;
4817 Value *X, *Y;
4818 Instruction *Mul;
4819 Instruction *Div;
4820 bool NeedNegation;
4821 // Look for: (-1 u/ x) u</u>= y
4822 if (!I.isEquality() &&
4823 match(&I, m_c_ICmp(Pred,
4824 m_CombineAnd(m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
4825 m_Instruction(Div)),
4826 m_Value(Y)))) {
4827 Mul = nullptr;
4829 // Are we checking that overflow does not happen, or does happen?
4830 switch (Pred) {
4831 case ICmpInst::Predicate::ICMP_ULT:
4832 NeedNegation = false;
4833 break; // OK
4834 case ICmpInst::Predicate::ICMP_UGE:
4835 NeedNegation = true;
4836 break; // OK
4837 default:
4838 return nullptr; // Wrong predicate.
4840 } else // Look for: ((x * y) / x) !=/== y
4841 if (I.isEquality() &&
4842 match(&I,
4843 m_c_ICmp(Pred, m_Value(Y),
4844 m_CombineAnd(
4845 m_OneUse(m_IDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
4846 m_Value(X)),
4847 m_Instruction(Mul)),
4848 m_Deferred(X))),
4849 m_Instruction(Div))))) {
4850 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
4851 } else
4852 return nullptr;
4854 BuilderTy::InsertPointGuard Guard(Builder);
4855 // If the pattern included (x * y), we'll want to insert new instructions
4856 // right before that original multiplication so that we can replace it.
4857 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
4858 if (MulHadOtherUses)
4859 Builder.SetInsertPoint(Mul);
4861 CallInst *Call = Builder.CreateIntrinsic(
4862 Div->getOpcode() == Instruction::UDiv ? Intrinsic::umul_with_overflow
4863 : Intrinsic::smul_with_overflow,
4864 X->getType(), {X, Y}, /*FMFSource=*/nullptr, "mul");
4866 // If the multiplication was used elsewhere, to ensure that we don't leave
4867 // "duplicate" instructions, replace uses of that original multiplication
4868 // with the multiplication result from the with.overflow intrinsic.
4869 if (MulHadOtherUses)
4870 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "mul.val"));
4872 Value *Res = Builder.CreateExtractValue(Call, 1, "mul.ov");
4873 if (NeedNegation) // This technically increases instruction count.
4874 Res = Builder.CreateNot(Res, "mul.not.ov");
4876 // If we replaced the mul, erase it. Do this after all uses of Builder,
4877 // as the mul is used as insertion point.
4878 if (MulHadOtherUses)
4879 eraseInstFromFunction(*Mul);
4881 return Res;
4884 static Instruction *foldICmpXNegX(ICmpInst &I,
4885 InstCombiner::BuilderTy &Builder) {
4886 CmpPredicate Pred;
4887 Value *X;
4888 if (match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X)))) {
4890 if (ICmpInst::isSigned(Pred))
4891 Pred = ICmpInst::getSwappedPredicate(Pred);
4892 else if (ICmpInst::isUnsigned(Pred))
4893 Pred = ICmpInst::getSignedPredicate(Pred);
4894 // else for equality-comparisons just keep the predicate.
4896 return ICmpInst::Create(Instruction::ICmp, Pred, X,
4897 Constant::getNullValue(X->getType()), I.getName());
4900 // A value is not equal to its negation unless that value is 0 or
4901 // MinSignedValue, ie: a != -a --> (a & MaxSignedVal) != 0
4902 if (match(&I, m_c_ICmp(Pred, m_OneUse(m_Neg(m_Value(X))), m_Deferred(X))) &&
4903 ICmpInst::isEquality(Pred)) {
4904 Type *Ty = X->getType();
4905 uint32_t BitWidth = Ty->getScalarSizeInBits();
4906 Constant *MaxSignedVal =
4907 ConstantInt::get(Ty, APInt::getSignedMaxValue(BitWidth));
4908 Value *And = Builder.CreateAnd(X, MaxSignedVal);
4909 Constant *Zero = Constant::getNullValue(Ty);
4910 return CmpInst::Create(Instruction::ICmp, Pred, And, Zero);
4913 return nullptr;
4916 static Instruction *foldICmpAndXX(ICmpInst &I, const SimplifyQuery &Q,
4917 InstCombinerImpl &IC) {
4918 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1), *A;
4919 // Normalize and operand as operand 0.
4920 CmpInst::Predicate Pred = I.getPredicate();
4921 if (match(Op1, m_c_And(m_Specific(Op0), m_Value()))) {
4922 std::swap(Op0, Op1);
4923 Pred = ICmpInst::getSwappedPredicate(Pred);
4926 if (!match(Op0, m_c_And(m_Specific(Op1), m_Value(A))))
4927 return nullptr;
4929 // (icmp (X & Y) u< X --> (X & Y) != X
4930 if (Pred == ICmpInst::ICMP_ULT)
4931 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4933 // (icmp (X & Y) u>= X --> (X & Y) == X
4934 if (Pred == ICmpInst::ICMP_UGE)
4935 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4937 if (ICmpInst::isEquality(Pred) && Op0->hasOneUse()) {
4938 // icmp (X & Y) eq/ne Y --> (X | ~Y) eq/ne -1 if Y is freely invertible and
4939 // Y is non-constant. If Y is constant the `X & C == C` form is preferable
4940 // so don't do this fold.
4941 if (!match(Op1, m_ImmConstant()))
4942 if (auto *NotOp1 =
4943 IC.getFreelyInverted(Op1, !Op1->hasNUsesOrMore(3), &IC.Builder))
4944 return new ICmpInst(Pred, IC.Builder.CreateOr(A, NotOp1),
4945 Constant::getAllOnesValue(Op1->getType()));
4946 // icmp (X & Y) eq/ne Y --> (~X & Y) eq/ne 0 if X is freely invertible.
4947 if (auto *NotA = IC.getFreelyInverted(A, A->hasOneUse(), &IC.Builder))
4948 return new ICmpInst(Pred, IC.Builder.CreateAnd(Op1, NotA),
4949 Constant::getNullValue(Op1->getType()));
4952 if (!ICmpInst::isSigned(Pred))
4953 return nullptr;
4955 KnownBits KnownY = IC.computeKnownBits(A, /*Depth=*/0, &I);
4956 // (X & NegY) spred X --> (X & NegY) upred X
4957 if (KnownY.isNegative())
4958 return new ICmpInst(ICmpInst::getUnsignedPredicate(Pred), Op0, Op1);
4960 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGT)
4961 return nullptr;
4963 if (KnownY.isNonNegative())
4964 // (X & PosY) s<= X --> X s>= 0
4965 // (X & PosY) s> X --> X s< 0
4966 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), Op1,
4967 Constant::getNullValue(Op1->getType()));
4969 if (isKnownNegative(Op1, IC.getSimplifyQuery().getWithInstruction(&I)))
4970 // (NegX & Y) s<= NegX --> Y s< 0
4971 // (NegX & Y) s> NegX --> Y s>= 0
4972 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(Pred), A,
4973 Constant::getNullValue(A->getType()));
4975 return nullptr;
4978 static Instruction *foldICmpOrXX(ICmpInst &I, const SimplifyQuery &Q,
4979 InstCombinerImpl &IC) {
4980 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1), *A;
4982 // Normalize or operand as operand 0.
4983 CmpInst::Predicate Pred = I.getPredicate();
4984 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value(A)))) {
4985 std::swap(Op0, Op1);
4986 Pred = ICmpInst::getSwappedPredicate(Pred);
4987 } else if (!match(Op0, m_c_Or(m_Specific(Op1), m_Value(A)))) {
4988 return nullptr;
4991 // icmp (X | Y) u<= X --> (X | Y) == X
4992 if (Pred == ICmpInst::ICMP_ULE)
4993 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4995 // icmp (X | Y) u> X --> (X | Y) != X
4996 if (Pred == ICmpInst::ICMP_UGT)
4997 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4999 if (ICmpInst::isEquality(Pred) && Op0->hasOneUse()) {
5000 // icmp (X | Y) eq/ne Y --> (X & ~Y) eq/ne 0 if Y is freely invertible
5001 if (Value *NotOp1 =
5002 IC.getFreelyInverted(Op1, !Op1->hasNUsesOrMore(3), &IC.Builder))
5003 return new ICmpInst(Pred, IC.Builder.CreateAnd(A, NotOp1),
5004 Constant::getNullValue(Op1->getType()));
5005 // icmp (X | Y) eq/ne Y --> (~X | Y) eq/ne -1 if X is freely invertible.
5006 if (Value *NotA = IC.getFreelyInverted(A, A->hasOneUse(), &IC.Builder))
5007 return new ICmpInst(Pred, IC.Builder.CreateOr(Op1, NotA),
5008 Constant::getAllOnesValue(Op1->getType()));
5010 return nullptr;
5013 static Instruction *foldICmpXorXX(ICmpInst &I, const SimplifyQuery &Q,
5014 InstCombinerImpl &IC) {
5015 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1), *A;
5016 // Normalize xor operand as operand 0.
5017 CmpInst::Predicate Pred = I.getPredicate();
5018 if (match(Op1, m_c_Xor(m_Specific(Op0), m_Value()))) {
5019 std::swap(Op0, Op1);
5020 Pred = ICmpInst::getSwappedPredicate(Pred);
5022 if (!match(Op0, m_c_Xor(m_Specific(Op1), m_Value(A))))
5023 return nullptr;
5025 // icmp (X ^ Y_NonZero) u>= X --> icmp (X ^ Y_NonZero) u> X
5026 // icmp (X ^ Y_NonZero) u<= X --> icmp (X ^ Y_NonZero) u< X
5027 // icmp (X ^ Y_NonZero) s>= X --> icmp (X ^ Y_NonZero) s> X
5028 // icmp (X ^ Y_NonZero) s<= X --> icmp (X ^ Y_NonZero) s< X
5029 CmpInst::Predicate PredOut = CmpInst::getStrictPredicate(Pred);
5030 if (PredOut != Pred && isKnownNonZero(A, Q))
5031 return new ICmpInst(PredOut, Op0, Op1);
5033 return nullptr;
5036 /// Try to fold icmp (binop), X or icmp X, (binop).
5037 /// TODO: A large part of this logic is duplicated in InstSimplify's
5038 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
5039 /// duplication.
5040 Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
5041 const SimplifyQuery &SQ) {
5042 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5043 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5045 // Special logic for binary operators.
5046 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
5047 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
5048 if (!BO0 && !BO1)
5049 return nullptr;
5051 if (Instruction *NewICmp = foldICmpXNegX(I, Builder))
5052 return NewICmp;
5054 const CmpInst::Predicate Pred = I.getPredicate();
5055 Value *X;
5057 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
5058 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
5059 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
5060 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
5061 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
5062 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
5063 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
5064 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
5065 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
5068 // (Op1 + X) + C u</u>= Op1 --> ~C - X u</u>= Op1
5069 Constant *C;
5070 if (match(Op0, m_OneUse(m_Add(m_c_Add(m_Specific(Op1), m_Value(X)),
5071 m_ImmConstant(C)))) &&
5072 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
5073 Constant *C2 = ConstantExpr::getNot(C);
5074 return new ICmpInst(Pred, Builder.CreateSub(C2, X), Op1);
5076 // Op0 u>/u<= (Op0 + X) + C --> Op0 u>/u<= ~C - X
5077 if (match(Op1, m_OneUse(m_Add(m_c_Add(m_Specific(Op0), m_Value(X)),
5078 m_ImmConstant(C)))) &&
5079 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE)) {
5080 Constant *C2 = ConstantExpr::getNot(C);
5081 return new ICmpInst(Pred, Op0, Builder.CreateSub(C2, X));
5085 // (icmp eq/ne (X, -P2), INT_MIN)
5086 // -> (icmp slt/sge X, INT_MIN + P2)
5087 if (ICmpInst::isEquality(Pred) && BO0 &&
5088 match(I.getOperand(1), m_SignMask()) &&
5089 match(BO0, m_And(m_Value(), m_NegatedPower2OrZero()))) {
5090 // Will Constant fold.
5091 Value *NewC = Builder.CreateSub(I.getOperand(1), BO0->getOperand(1));
5092 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SLT
5093 : ICmpInst::ICMP_SGE,
5094 BO0->getOperand(0), NewC);
5098 // Similar to above: an unsigned overflow comparison may use offset + mask:
5099 // ((Op1 + C) & C) u< Op1 --> Op1 != 0
5100 // ((Op1 + C) & C) u>= Op1 --> Op1 == 0
5101 // Op0 u> ((Op0 + C) & C) --> Op0 != 0
5102 // Op0 u<= ((Op0 + C) & C) --> Op0 == 0
5103 BinaryOperator *BO;
5104 const APInt *C;
5105 if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) &&
5106 match(Op0, m_And(m_BinOp(BO), m_LowBitMask(C))) &&
5107 match(BO, m_Add(m_Specific(Op1), m_SpecificIntAllowPoison(*C)))) {
5108 CmpInst::Predicate NewPred =
5109 Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
5110 Constant *Zero = ConstantInt::getNullValue(Op1->getType());
5111 return new ICmpInst(NewPred, Op1, Zero);
5114 if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
5115 match(Op1, m_And(m_BinOp(BO), m_LowBitMask(C))) &&
5116 match(BO, m_Add(m_Specific(Op0), m_SpecificIntAllowPoison(*C)))) {
5117 CmpInst::Predicate NewPred =
5118 Pred == ICmpInst::ICMP_UGT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
5119 Constant *Zero = ConstantInt::getNullValue(Op1->getType());
5120 return new ICmpInst(NewPred, Op0, Zero);
5124 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
5125 bool Op0HasNUW = false, Op1HasNUW = false;
5126 bool Op0HasNSW = false, Op1HasNSW = false;
5127 // Analyze the case when either Op0 or Op1 is an add instruction.
5128 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
5129 auto hasNoWrapProblem = [](const BinaryOperator &BO, CmpInst::Predicate Pred,
5130 bool &HasNSW, bool &HasNUW) -> bool {
5131 if (isa<OverflowingBinaryOperator>(BO)) {
5132 HasNUW = BO.hasNoUnsignedWrap();
5133 HasNSW = BO.hasNoSignedWrap();
5134 return ICmpInst::isEquality(Pred) ||
5135 (CmpInst::isUnsigned(Pred) && HasNUW) ||
5136 (CmpInst::isSigned(Pred) && HasNSW);
5137 } else if (BO.getOpcode() == Instruction::Or) {
5138 HasNUW = true;
5139 HasNSW = true;
5140 return true;
5141 } else {
5142 return false;
5145 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
5147 if (BO0) {
5148 match(BO0, m_AddLike(m_Value(A), m_Value(B)));
5149 NoOp0WrapProblem = hasNoWrapProblem(*BO0, Pred, Op0HasNSW, Op0HasNUW);
5151 if (BO1) {
5152 match(BO1, m_AddLike(m_Value(C), m_Value(D)));
5153 NoOp1WrapProblem = hasNoWrapProblem(*BO1, Pred, Op1HasNSW, Op1HasNUW);
5156 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
5157 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
5158 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
5159 return new ICmpInst(Pred, A == Op1 ? B : A,
5160 Constant::getNullValue(Op1->getType()));
5162 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
5163 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
5164 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
5165 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
5166 C == Op0 ? D : C);
5168 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
5169 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
5170 NoOp1WrapProblem) {
5171 // Determine Y and Z in the form icmp (X+Y), (X+Z).
5172 Value *Y, *Z;
5173 if (A == C) {
5174 // C + B == C + D -> B == D
5175 Y = B;
5176 Z = D;
5177 } else if (A == D) {
5178 // D + B == C + D -> B == C
5179 Y = B;
5180 Z = C;
5181 } else if (B == C) {
5182 // A + C == C + D -> A == D
5183 Y = A;
5184 Z = D;
5185 } else {
5186 assert(B == D);
5187 // A + D == C + D -> A == C
5188 Y = A;
5189 Z = C;
5191 return new ICmpInst(Pred, Y, Z);
5194 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
5195 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
5196 match(B, m_AllOnes()))
5197 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
5199 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
5200 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
5201 match(B, m_AllOnes()))
5202 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
5204 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
5205 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
5206 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
5208 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
5209 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
5210 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
5212 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
5213 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
5214 match(D, m_AllOnes()))
5215 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
5217 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
5218 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
5219 match(D, m_AllOnes()))
5220 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
5222 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
5223 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
5224 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
5226 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
5227 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
5228 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
5230 // TODO: The subtraction-related identities shown below also hold, but
5231 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
5232 // wouldn't happen even if they were implemented.
5234 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
5235 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
5236 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
5237 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
5239 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
5240 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
5241 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
5243 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
5244 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
5245 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
5247 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
5248 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
5249 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
5251 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
5252 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
5253 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
5255 // if C1 has greater magnitude than C2:
5256 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
5257 // s.t. C3 = C1 - C2
5259 // if C2 has greater magnitude than C1:
5260 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
5261 // s.t. C3 = C2 - C1
5262 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
5263 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) {
5264 const APInt *AP1, *AP2;
5265 // TODO: Support non-uniform vectors.
5266 // TODO: Allow poison passthrough if B or D's element is poison.
5267 if (match(B, m_APIntAllowPoison(AP1)) &&
5268 match(D, m_APIntAllowPoison(AP2)) &&
5269 AP1->isNegative() == AP2->isNegative()) {
5270 APInt AP1Abs = AP1->abs();
5271 APInt AP2Abs = AP2->abs();
5272 if (AP1Abs.uge(AP2Abs)) {
5273 APInt Diff = *AP1 - *AP2;
5274 Constant *C3 = Constant::getIntegerValue(BO0->getType(), Diff);
5275 Value *NewAdd = Builder.CreateAdd(
5276 A, C3, "", Op0HasNUW && Diff.ule(*AP1), Op0HasNSW);
5277 return new ICmpInst(Pred, NewAdd, C);
5278 } else {
5279 APInt Diff = *AP2 - *AP1;
5280 Constant *C3 = Constant::getIntegerValue(BO0->getType(), Diff);
5281 Value *NewAdd = Builder.CreateAdd(
5282 C, C3, "", Op1HasNUW && Diff.ule(*AP2), Op1HasNSW);
5283 return new ICmpInst(Pred, A, NewAdd);
5286 Constant *Cst1, *Cst2;
5287 if (match(B, m_ImmConstant(Cst1)) && match(D, m_ImmConstant(Cst2)) &&
5288 ICmpInst::isEquality(Pred)) {
5289 Constant *Diff = ConstantExpr::getSub(Cst2, Cst1);
5290 Value *NewAdd = Builder.CreateAdd(C, Diff);
5291 return new ICmpInst(Pred, A, NewAdd);
5295 // Analyze the case when either Op0 or Op1 is a sub instruction.
5296 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
5297 A = nullptr;
5298 B = nullptr;
5299 C = nullptr;
5300 D = nullptr;
5301 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
5302 A = BO0->getOperand(0);
5303 B = BO0->getOperand(1);
5305 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
5306 C = BO1->getOperand(0);
5307 D = BO1->getOperand(1);
5310 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
5311 if (A == Op1 && NoOp0WrapProblem)
5312 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
5313 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
5314 if (C == Op0 && NoOp1WrapProblem)
5315 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
5317 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
5318 // (A - B) u>/u<= A --> B u>/u<= A
5319 if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
5320 return new ICmpInst(Pred, B, A);
5321 // C u</u>= (C - D) --> C u</u>= D
5322 if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
5323 return new ICmpInst(Pred, C, D);
5324 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
5325 if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
5326 isKnownNonZero(B, Q))
5327 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
5328 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
5329 if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
5330 isKnownNonZero(D, Q))
5331 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
5333 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
5334 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
5335 return new ICmpInst(Pred, A, C);
5337 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
5338 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
5339 return new ICmpInst(Pred, D, B);
5341 // icmp (0-X) < cst --> x > -cst
5342 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
5343 Value *X;
5344 if (match(BO0, m_Neg(m_Value(X))))
5345 if (Constant *RHSC = dyn_cast<Constant>(Op1))
5346 if (RHSC->isNotMinSignedValue())
5347 return new ICmpInst(I.getSwappedPredicate(), X,
5348 ConstantExpr::getNeg(RHSC));
5351 if (Instruction * R = foldICmpXorXX(I, Q, *this))
5352 return R;
5353 if (Instruction *R = foldICmpOrXX(I, Q, *this))
5354 return R;
5357 // Try to remove shared multiplier from comparison:
5358 // X * Z pred Y * Z
5359 Value *X, *Y, *Z;
5360 if ((match(Op0, m_Mul(m_Value(X), m_Value(Z))) &&
5361 match(Op1, m_c_Mul(m_Specific(Z), m_Value(Y)))) ||
5362 (match(Op0, m_Mul(m_Value(Z), m_Value(X))) &&
5363 match(Op1, m_c_Mul(m_Specific(Z), m_Value(Y))))) {
5364 if (ICmpInst::isSigned(Pred)) {
5365 if (Op0HasNSW && Op1HasNSW) {
5366 KnownBits ZKnown = computeKnownBits(Z, 0, &I);
5367 if (ZKnown.isStrictlyPositive())
5368 return new ICmpInst(Pred, X, Y);
5369 if (ZKnown.isNegative())
5370 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), X, Y);
5371 Value *LessThan = simplifyICmpInst(ICmpInst::ICMP_SLT, X, Y,
5372 SQ.getWithInstruction(&I));
5373 if (LessThan && match(LessThan, m_One()))
5374 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), Z,
5375 Constant::getNullValue(Z->getType()));
5376 Value *GreaterThan = simplifyICmpInst(ICmpInst::ICMP_SGT, X, Y,
5377 SQ.getWithInstruction(&I));
5378 if (GreaterThan && match(GreaterThan, m_One()))
5379 return new ICmpInst(Pred, Z, Constant::getNullValue(Z->getType()));
5381 } else {
5382 bool NonZero;
5383 if (ICmpInst::isEquality(Pred)) {
5384 // If X != Y, fold (X *nw Z) eq/ne (Y *nw Z) -> Z eq/ne 0
5385 if (((Op0HasNSW && Op1HasNSW) || (Op0HasNUW && Op1HasNUW)) &&
5386 isKnownNonEqual(X, Y, DL, &AC, &I, &DT))
5387 return new ICmpInst(Pred, Z, Constant::getNullValue(Z->getType()));
5389 KnownBits ZKnown = computeKnownBits(Z, 0, &I);
5390 // if Z % 2 != 0
5391 // X * Z eq/ne Y * Z -> X eq/ne Y
5392 if (ZKnown.countMaxTrailingZeros() == 0)
5393 return new ICmpInst(Pred, X, Y);
5394 NonZero = !ZKnown.One.isZero() || isKnownNonZero(Z, Q);
5395 // if Z != 0 and nsw(X * Z) and nsw(Y * Z)
5396 // X * Z eq/ne Y * Z -> X eq/ne Y
5397 if (NonZero && BO0 && BO1 && Op0HasNSW && Op1HasNSW)
5398 return new ICmpInst(Pred, X, Y);
5399 } else
5400 NonZero = isKnownNonZero(Z, Q);
5402 // If Z != 0 and nuw(X * Z) and nuw(Y * Z)
5403 // X * Z u{lt/le/gt/ge}/eq/ne Y * Z -> X u{lt/le/gt/ge}/eq/ne Y
5404 if (NonZero && BO0 && BO1 && Op0HasNUW && Op1HasNUW)
5405 return new ICmpInst(Pred, X, Y);
5410 BinaryOperator *SRem = nullptr;
5411 // icmp (srem X, Y), Y
5412 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
5413 SRem = BO0;
5414 // icmp Y, (srem X, Y)
5415 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
5416 Op0 == BO1->getOperand(1))
5417 SRem = BO1;
5418 if (SRem) {
5419 // We don't check hasOneUse to avoid increasing register pressure because
5420 // the value we use is the same value this instruction was already using.
5421 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
5422 default:
5423 break;
5424 case ICmpInst::ICMP_EQ:
5425 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5426 case ICmpInst::ICMP_NE:
5427 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5428 case ICmpInst::ICMP_SGT:
5429 case ICmpInst::ICMP_SGE:
5430 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
5431 Constant::getAllOnesValue(SRem->getType()));
5432 case ICmpInst::ICMP_SLT:
5433 case ICmpInst::ICMP_SLE:
5434 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
5435 Constant::getNullValue(SRem->getType()));
5439 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
5440 (BO0->hasOneUse() || BO1->hasOneUse()) &&
5441 BO0->getOperand(1) == BO1->getOperand(1)) {
5442 switch (BO0->getOpcode()) {
5443 default:
5444 break;
5445 case Instruction::Add:
5446 case Instruction::Sub:
5447 case Instruction::Xor: {
5448 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
5449 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
5451 const APInt *C;
5452 if (match(BO0->getOperand(1), m_APInt(C))) {
5453 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
5454 if (C->isSignMask()) {
5455 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
5456 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
5459 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
5460 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
5461 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
5462 NewPred = I.getSwappedPredicate(NewPred);
5463 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
5466 break;
5468 case Instruction::Mul: {
5469 if (!I.isEquality())
5470 break;
5472 const APInt *C;
5473 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isZero() &&
5474 !C->isOne()) {
5475 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
5476 // Mask = -1 >> count-trailing-zeros(C).
5477 if (unsigned TZs = C->countr_zero()) {
5478 Constant *Mask = ConstantInt::get(
5479 BO0->getType(),
5480 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
5481 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
5482 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
5483 return new ICmpInst(Pred, And1, And2);
5486 break;
5488 case Instruction::UDiv:
5489 case Instruction::LShr:
5490 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
5491 break;
5492 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
5494 case Instruction::SDiv:
5495 if (!(I.isEquality() || match(BO0->getOperand(1), m_NonNegative())) ||
5496 !BO0->isExact() || !BO1->isExact())
5497 break;
5498 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
5500 case Instruction::AShr:
5501 if (!BO0->isExact() || !BO1->isExact())
5502 break;
5503 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
5505 case Instruction::Shl: {
5506 bool NUW = Op0HasNUW && Op1HasNUW;
5507 bool NSW = Op0HasNSW && Op1HasNSW;
5508 if (!NUW && !NSW)
5509 break;
5510 if (!NSW && I.isSigned())
5511 break;
5512 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
5517 if (BO0) {
5518 // Transform A & (L - 1) `ult` L --> L != 0
5519 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
5520 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
5522 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
5523 auto *Zero = Constant::getNullValue(BO0->getType());
5524 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
5528 // For unsigned predicates / eq / ne:
5529 // icmp pred (x << 1), x --> icmp getSignedPredicate(pred) x, 0
5530 // icmp pred x, (x << 1) --> icmp getSignedPredicate(pred) 0, x
5531 if (!ICmpInst::isSigned(Pred)) {
5532 if (match(Op0, m_Shl(m_Specific(Op1), m_One())))
5533 return new ICmpInst(ICmpInst::getSignedPredicate(Pred), Op1,
5534 Constant::getNullValue(Op1->getType()));
5535 else if (match(Op1, m_Shl(m_Specific(Op0), m_One())))
5536 return new ICmpInst(ICmpInst::getSignedPredicate(Pred),
5537 Constant::getNullValue(Op0->getType()), Op0);
5540 if (Value *V = foldMultiplicationOverflowCheck(I))
5541 return replaceInstUsesWith(I, V);
5543 if (Instruction *R = foldICmpAndXX(I, Q, *this))
5544 return R;
5546 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
5547 return replaceInstUsesWith(I, V);
5549 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
5550 return replaceInstUsesWith(I, V);
5552 return nullptr;
5555 /// Fold icmp Pred min|max(X, Y), Z.
5556 Instruction *InstCombinerImpl::foldICmpWithMinMax(Instruction &I,
5557 MinMaxIntrinsic *MinMax,
5558 Value *Z, CmpPredicate Pred) {
5559 Value *X = MinMax->getLHS();
5560 Value *Y = MinMax->getRHS();
5561 if (ICmpInst::isSigned(Pred) && !MinMax->isSigned())
5562 return nullptr;
5563 if (ICmpInst::isUnsigned(Pred) && MinMax->isSigned()) {
5564 // Revert the transform signed pred -> unsigned pred
5565 // TODO: We can flip the signedness of predicate if both operands of icmp
5566 // are negative.
5567 if (isKnownNonNegative(Z, SQ.getWithInstruction(&I)) &&
5568 isKnownNonNegative(MinMax, SQ.getWithInstruction(&I))) {
5569 Pred = ICmpInst::getFlippedSignednessPredicate(Pred);
5570 } else
5571 return nullptr;
5573 SimplifyQuery Q = SQ.getWithInstruction(&I);
5574 auto IsCondKnownTrue = [](Value *Val) -> std::optional<bool> {
5575 if (!Val)
5576 return std::nullopt;
5577 if (match(Val, m_One()))
5578 return true;
5579 if (match(Val, m_Zero()))
5580 return false;
5581 return std::nullopt;
5583 auto CmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred, X, Z, Q));
5584 auto CmpYZ = IsCondKnownTrue(simplifyICmpInst(Pred, Y, Z, Q));
5585 if (!CmpXZ.has_value() && !CmpYZ.has_value())
5586 return nullptr;
5587 if (!CmpXZ.has_value()) {
5588 std::swap(X, Y);
5589 std::swap(CmpXZ, CmpYZ);
5592 auto FoldIntoCmpYZ = [&]() -> Instruction * {
5593 if (CmpYZ.has_value())
5594 return replaceInstUsesWith(I, ConstantInt::getBool(I.getType(), *CmpYZ));
5595 return ICmpInst::Create(Instruction::ICmp, Pred, Y, Z);
5598 switch (Pred) {
5599 case ICmpInst::ICMP_EQ:
5600 case ICmpInst::ICMP_NE: {
5601 // If X == Z:
5602 // Expr Result
5603 // min(X, Y) == Z X <= Y
5604 // max(X, Y) == Z X >= Y
5605 // min(X, Y) != Z X > Y
5606 // max(X, Y) != Z X < Y
5607 if ((Pred == ICmpInst::ICMP_EQ) == *CmpXZ) {
5608 ICmpInst::Predicate NewPred =
5609 ICmpInst::getNonStrictPredicate(MinMax->getPredicate());
5610 if (Pred == ICmpInst::ICMP_NE)
5611 NewPred = ICmpInst::getInversePredicate(NewPred);
5612 return ICmpInst::Create(Instruction::ICmp, NewPred, X, Y);
5614 // Otherwise (X != Z):
5615 ICmpInst::Predicate NewPred = MinMax->getPredicate();
5616 auto MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(NewPred, X, Z, Q));
5617 if (!MinMaxCmpXZ.has_value()) {
5618 std::swap(X, Y);
5619 std::swap(CmpXZ, CmpYZ);
5620 // Re-check pre-condition X != Z
5621 if (!CmpXZ.has_value() || (Pred == ICmpInst::ICMP_EQ) == *CmpXZ)
5622 break;
5623 MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(NewPred, X, Z, Q));
5625 if (!MinMaxCmpXZ.has_value())
5626 break;
5627 if (*MinMaxCmpXZ) {
5628 // Expr Fact Result
5629 // min(X, Y) == Z X < Z false
5630 // max(X, Y) == Z X > Z false
5631 // min(X, Y) != Z X < Z true
5632 // max(X, Y) != Z X > Z true
5633 return replaceInstUsesWith(
5634 I, ConstantInt::getBool(I.getType(), Pred == ICmpInst::ICMP_NE));
5635 } else {
5636 // Expr Fact Result
5637 // min(X, Y) == Z X > Z Y == Z
5638 // max(X, Y) == Z X < Z Y == Z
5639 // min(X, Y) != Z X > Z Y != Z
5640 // max(X, Y) != Z X < Z Y != Z
5641 return FoldIntoCmpYZ();
5643 break;
5645 case ICmpInst::ICMP_SLT:
5646 case ICmpInst::ICMP_ULT:
5647 case ICmpInst::ICMP_SLE:
5648 case ICmpInst::ICMP_ULE:
5649 case ICmpInst::ICMP_SGT:
5650 case ICmpInst::ICMP_UGT:
5651 case ICmpInst::ICMP_SGE:
5652 case ICmpInst::ICMP_UGE: {
5653 bool IsSame = MinMax->getPredicate() == ICmpInst::getStrictPredicate(Pred);
5654 if (*CmpXZ) {
5655 if (IsSame) {
5656 // Expr Fact Result
5657 // min(X, Y) < Z X < Z true
5658 // min(X, Y) <= Z X <= Z true
5659 // max(X, Y) > Z X > Z true
5660 // max(X, Y) >= Z X >= Z true
5661 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5662 } else {
5663 // Expr Fact Result
5664 // max(X, Y) < Z X < Z Y < Z
5665 // max(X, Y) <= Z X <= Z Y <= Z
5666 // min(X, Y) > Z X > Z Y > Z
5667 // min(X, Y) >= Z X >= Z Y >= Z
5668 return FoldIntoCmpYZ();
5670 } else {
5671 if (IsSame) {
5672 // Expr Fact Result
5673 // min(X, Y) < Z X >= Z Y < Z
5674 // min(X, Y) <= Z X > Z Y <= Z
5675 // max(X, Y) > Z X <= Z Y > Z
5676 // max(X, Y) >= Z X < Z Y >= Z
5677 return FoldIntoCmpYZ();
5678 } else {
5679 // Expr Fact Result
5680 // max(X, Y) < Z X >= Z false
5681 // max(X, Y) <= Z X > Z false
5682 // min(X, Y) > Z X <= Z false
5683 // min(X, Y) >= Z X < Z false
5684 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5687 break;
5689 default:
5690 break;
5693 return nullptr;
5696 // Canonicalize checking for a power-of-2-or-zero value:
5697 static Instruction *foldICmpPow2Test(ICmpInst &I,
5698 InstCombiner::BuilderTy &Builder) {
5699 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5700 const CmpInst::Predicate Pred = I.getPredicate();
5701 Value *A = nullptr;
5702 bool CheckIs;
5703 if (I.isEquality()) {
5704 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
5705 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
5706 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
5707 m_Deferred(A)))) ||
5708 !match(Op1, m_ZeroInt()))
5709 A = nullptr;
5711 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
5712 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
5713 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
5714 A = Op1;
5715 else if (match(Op1,
5716 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
5717 A = Op0;
5719 CheckIs = Pred == ICmpInst::ICMP_EQ;
5720 } else if (ICmpInst::isUnsigned(Pred)) {
5721 // (A ^ (A-1)) u>= A --> ctpop(A) < 2 (two commuted variants)
5722 // ((A-1) ^ A) u< A --> ctpop(A) > 1 (two commuted variants)
5724 if ((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
5725 match(Op0, m_OneUse(m_c_Xor(m_Add(m_Specific(Op1), m_AllOnes()),
5726 m_Specific(Op1))))) {
5727 A = Op1;
5728 CheckIs = Pred == ICmpInst::ICMP_UGE;
5729 } else if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
5730 match(Op1, m_OneUse(m_c_Xor(m_Add(m_Specific(Op0), m_AllOnes()),
5731 m_Specific(Op0))))) {
5732 A = Op0;
5733 CheckIs = Pred == ICmpInst::ICMP_ULE;
5737 if (A) {
5738 Type *Ty = A->getType();
5739 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
5740 return CheckIs ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop,
5741 ConstantInt::get(Ty, 2))
5742 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop,
5743 ConstantInt::get(Ty, 1));
5746 return nullptr;
5749 Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
5750 if (!I.isEquality())
5751 return nullptr;
5753 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5754 const CmpInst::Predicate Pred = I.getPredicate();
5755 Value *A, *B, *C, *D;
5756 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5757 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5758 Value *OtherVal = A == Op1 ? B : A;
5759 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
5762 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5763 // A^c1 == C^c2 --> A == C^(c1^c2)
5764 ConstantInt *C1, *C2;
5765 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
5766 Op1->hasOneUse()) {
5767 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
5768 Value *Xor = Builder.CreateXor(C, NC);
5769 return new ICmpInst(Pred, A, Xor);
5772 // A^B == A^D -> B == D
5773 if (A == C)
5774 return new ICmpInst(Pred, B, D);
5775 if (A == D)
5776 return new ICmpInst(Pred, B, C);
5777 if (B == C)
5778 return new ICmpInst(Pred, A, D);
5779 if (B == D)
5780 return new ICmpInst(Pred, A, C);
5784 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
5785 // A == (A^B) -> B == 0
5786 Value *OtherVal = A == Op0 ? B : A;
5787 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
5790 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5791 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
5792 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5793 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
5795 if (A == C) {
5796 X = B;
5797 Y = D;
5798 Z = A;
5799 } else if (A == D) {
5800 X = B;
5801 Y = C;
5802 Z = A;
5803 } else if (B == C) {
5804 X = A;
5805 Y = D;
5806 Z = B;
5807 } else if (B == D) {
5808 X = A;
5809 Y = C;
5810 Z = B;
5813 if (X) {
5814 // If X^Y is a negative power of two, then `icmp eq/ne (Z & NegP2), 0`
5815 // will fold to `icmp ult/uge Z, -NegP2` incurringb no additional
5816 // instructions.
5817 const APInt *C0, *C1;
5818 bool XorIsNegP2 = match(X, m_APInt(C0)) && match(Y, m_APInt(C1)) &&
5819 (*C0 ^ *C1).isNegatedPowerOf2();
5821 // If either Op0/Op1 are both one use or X^Y will constant fold and one of
5822 // Op0/Op1 are one use, proceed. In those cases we are instruction neutral
5823 // but `icmp eq/ne A, 0` is easier to analyze than `icmp eq/ne A, B`.
5824 int UseCnt =
5825 int(Op0->hasOneUse()) + int(Op1->hasOneUse()) +
5826 (int(match(X, m_ImmConstant()) && match(Y, m_ImmConstant())));
5827 if (XorIsNegP2 || UseCnt >= 2) {
5828 // Build (X^Y) & Z
5829 Op1 = Builder.CreateXor(X, Y);
5830 Op1 = Builder.CreateAnd(Op1, Z);
5831 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
5837 // Similar to above, but specialized for constant because invert is needed:
5838 // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
5839 Value *X, *Y;
5840 Constant *C;
5841 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_Constant(C)))) &&
5842 match(Op1, m_OneUse(m_Or(m_Value(Y), m_Specific(C))))) {
5843 Value *Xor = Builder.CreateXor(X, Y);
5844 Value *And = Builder.CreateAnd(Xor, ConstantExpr::getNot(C));
5845 return new ICmpInst(Pred, And, Constant::getNullValue(And->getType()));
5849 if (match(Op1, m_ZExt(m_Value(A))) &&
5850 (Op0->hasOneUse() || Op1->hasOneUse())) {
5851 // (B & (Pow2C-1)) == zext A --> A == trunc B
5852 // (B & (Pow2C-1)) != zext A --> A != trunc B
5853 const APInt *MaskC;
5854 if (match(Op0, m_And(m_Value(B), m_LowBitMask(MaskC))) &&
5855 MaskC->countr_one() == A->getType()->getScalarSizeInBits())
5856 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
5859 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
5860 // For lshr and ashr pairs.
5861 const APInt *AP1, *AP2;
5862 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_APIntAllowPoison(AP1)))) &&
5863 match(Op1, m_OneUse(m_LShr(m_Value(B), m_APIntAllowPoison(AP2))))) ||
5864 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_APIntAllowPoison(AP1)))) &&
5865 match(Op1, m_OneUse(m_AShr(m_Value(B), m_APIntAllowPoison(AP2)))))) {
5866 if (AP1 != AP2)
5867 return nullptr;
5868 unsigned TypeBits = AP1->getBitWidth();
5869 unsigned ShAmt = AP1->getLimitedValue(TypeBits);
5870 if (ShAmt < TypeBits && ShAmt != 0) {
5871 ICmpInst::Predicate NewPred =
5872 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5873 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
5874 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
5875 return new ICmpInst(NewPred, Xor, ConstantInt::get(A->getType(), CmpVal));
5879 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
5880 ConstantInt *Cst1;
5881 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
5882 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
5883 unsigned TypeBits = Cst1->getBitWidth();
5884 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
5885 if (ShAmt < TypeBits && ShAmt != 0) {
5886 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
5887 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
5888 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
5889 I.getName() + ".mask");
5890 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
5894 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
5895 // "icmp (and X, mask), cst"
5896 uint64_t ShAmt = 0;
5897 if (Op0->hasOneUse() &&
5898 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
5899 match(Op1, m_ConstantInt(Cst1)) &&
5900 // Only do this when A has multiple uses. This is most important to do
5901 // when it exposes other optimizations.
5902 !A->hasOneUse()) {
5903 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
5905 if (ShAmt < ASize) {
5906 APInt MaskV =
5907 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
5908 MaskV <<= ShAmt;
5910 APInt CmpV = Cst1->getValue().zext(ASize);
5911 CmpV <<= ShAmt;
5913 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
5914 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
5918 if (Instruction *ICmp = foldICmpIntrinsicWithIntrinsic(I, Builder))
5919 return ICmp;
5921 // Match icmp eq (trunc (lshr A, BW), (ashr (trunc A), BW-1)), which checks the
5922 // top BW/2 + 1 bits are all the same. Create "A >=s INT_MIN && A <=s INT_MAX",
5923 // which we generate as "icmp ult (add A, 2^(BW-1)), 2^BW" to skip a few steps
5924 // of instcombine.
5925 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
5926 if (match(Op0, m_AShr(m_Trunc(m_Value(A)), m_SpecificInt(BitWidth - 1))) &&
5927 match(Op1, m_Trunc(m_LShr(m_Specific(A), m_SpecificInt(BitWidth)))) &&
5928 A->getType()->getScalarSizeInBits() == BitWidth * 2 &&
5929 (I.getOperand(0)->hasOneUse() || I.getOperand(1)->hasOneUse())) {
5930 APInt C = APInt::getOneBitSet(BitWidth * 2, BitWidth - 1);
5931 Value *Add = Builder.CreateAdd(A, ConstantInt::get(A->getType(), C));
5932 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT
5933 : ICmpInst::ICMP_UGE,
5934 Add, ConstantInt::get(A->getType(), C.shl(1)));
5937 // Canonicalize:
5938 // Assume B_Pow2 != 0
5939 // 1. A & B_Pow2 != B_Pow2 -> A & B_Pow2 == 0
5940 // 2. A & B_Pow2 == B_Pow2 -> A & B_Pow2 != 0
5941 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())) &&
5942 isKnownToBeAPowerOfTwo(Op1, /* OrZero */ false, 0, &I))
5943 return new ICmpInst(CmpInst::getInversePredicate(Pred), Op0,
5944 ConstantInt::getNullValue(Op0->getType()));
5946 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())) &&
5947 isKnownToBeAPowerOfTwo(Op0, /* OrZero */ false, 0, &I))
5948 return new ICmpInst(CmpInst::getInversePredicate(Pred), Op1,
5949 ConstantInt::getNullValue(Op1->getType()));
5951 // Canonicalize:
5952 // icmp eq/ne X, OneUse(rotate-right(X))
5953 // -> icmp eq/ne X, rotate-left(X)
5954 // We generally try to convert rotate-right -> rotate-left, this just
5955 // canonicalizes another case.
5956 if (match(&I, m_c_ICmp(m_Value(A),
5957 m_OneUse(m_Intrinsic<Intrinsic::fshr>(
5958 m_Deferred(A), m_Deferred(A), m_Value(B))))))
5959 return new ICmpInst(
5960 Pred, A,
5961 Builder.CreateIntrinsic(Op0->getType(), Intrinsic::fshl, {A, A, B}));
5963 // Canonicalize:
5964 // icmp eq/ne OneUse(A ^ Cst), B --> icmp eq/ne (A ^ B), Cst
5965 Constant *Cst;
5966 if (match(&I, m_c_ICmp(m_OneUse(m_Xor(m_Value(A), m_ImmConstant(Cst))),
5967 m_CombineAnd(m_Value(B), m_Unless(m_ImmConstant())))))
5968 return new ICmpInst(Pred, Builder.CreateXor(A, B), Cst);
5971 // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
5972 auto m_Matcher =
5973 m_CombineOr(m_CombineOr(m_c_Add(m_Value(B), m_Deferred(A)),
5974 m_c_Xor(m_Value(B), m_Deferred(A))),
5975 m_Sub(m_Value(B), m_Deferred(A)));
5976 std::optional<bool> IsZero = std::nullopt;
5977 if (match(&I, m_c_ICmp(m_OneUse(m_c_And(m_Value(A), m_Matcher)),
5978 m_Deferred(A))))
5979 IsZero = false;
5980 // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
5981 else if (match(&I,
5982 m_ICmp(m_OneUse(m_c_And(m_Value(A), m_Matcher)), m_Zero())))
5983 IsZero = true;
5985 if (IsZero && isKnownToBeAPowerOfTwo(A, /* OrZero */ true, /*Depth*/ 0, &I))
5986 // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
5987 // -> (icmp eq/ne (and X, P2), 0)
5988 // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
5989 // -> (icmp eq/ne (and X, P2), P2)
5990 return new ICmpInst(Pred, Builder.CreateAnd(B, A),
5991 *IsZero ? A
5992 : ConstantInt::getNullValue(A->getType()));
5995 return nullptr;
5998 Instruction *InstCombinerImpl::foldICmpWithTrunc(ICmpInst &ICmp) {
5999 ICmpInst::Predicate Pred = ICmp.getPredicate();
6000 Value *Op0 = ICmp.getOperand(0), *Op1 = ICmp.getOperand(1);
6002 // Try to canonicalize trunc + compare-to-constant into a mask + cmp.
6003 // The trunc masks high bits while the compare may effectively mask low bits.
6004 Value *X;
6005 const APInt *C;
6006 if (!match(Op0, m_OneUse(m_Trunc(m_Value(X)))) || !match(Op1, m_APInt(C)))
6007 return nullptr;
6009 // This matches patterns corresponding to tests of the signbit as well as:
6010 // (trunc X) pred C2 --> (X & Mask) == C
6011 if (auto Res = decomposeBitTestICmp(Op0, Op1, Pred, /*WithTrunc=*/true,
6012 /*AllowNonZeroC=*/true)) {
6013 Value *And = Builder.CreateAnd(Res->X, Res->Mask);
6014 Constant *C = ConstantInt::get(Res->X->getType(), Res->C);
6015 return new ICmpInst(Res->Pred, And, C);
6018 unsigned SrcBits = X->getType()->getScalarSizeInBits();
6019 if (auto *II = dyn_cast<IntrinsicInst>(X)) {
6020 if (II->getIntrinsicID() == Intrinsic::cttz ||
6021 II->getIntrinsicID() == Intrinsic::ctlz) {
6022 unsigned MaxRet = SrcBits;
6023 // If the "is_zero_poison" argument is set, then we know at least
6024 // one bit is set in the input, so the result is always at least one
6025 // less than the full bitwidth of that input.
6026 if (match(II->getArgOperand(1), m_One()))
6027 MaxRet--;
6029 // Make sure the destination is wide enough to hold the largest output of
6030 // the intrinsic.
6031 if (llvm::Log2_32(MaxRet) + 1 <= Op0->getType()->getScalarSizeInBits())
6032 if (Instruction *I =
6033 foldICmpIntrinsicWithConstant(ICmp, II, C->zext(SrcBits)))
6034 return I;
6038 return nullptr;
6041 Instruction *InstCombinerImpl::foldICmpWithZextOrSext(ICmpInst &ICmp) {
6042 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
6043 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
6044 Value *X;
6045 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
6046 return nullptr;
6048 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
6049 bool IsSignedCmp = ICmp.isSigned();
6051 // icmp Pred (ext X), (ext Y)
6052 Value *Y;
6053 if (match(ICmp.getOperand(1), m_ZExtOrSExt(m_Value(Y)))) {
6054 bool IsZext0 = isa<ZExtInst>(ICmp.getOperand(0));
6055 bool IsZext1 = isa<ZExtInst>(ICmp.getOperand(1));
6057 if (IsZext0 != IsZext1) {
6058 // If X and Y and both i1
6059 // (icmp eq/ne (zext X) (sext Y))
6060 // eq -> (icmp eq (or X, Y), 0)
6061 // ne -> (icmp ne (or X, Y), 0)
6062 if (ICmp.isEquality() && X->getType()->isIntOrIntVectorTy(1) &&
6063 Y->getType()->isIntOrIntVectorTy(1))
6064 return new ICmpInst(ICmp.getPredicate(), Builder.CreateOr(X, Y),
6065 Constant::getNullValue(X->getType()));
6067 // If we have mismatched casts and zext has the nneg flag, we can
6068 // treat the "zext nneg" as "sext". Otherwise, we cannot fold and quit.
6070 auto *NonNegInst0 = dyn_cast<PossiblyNonNegInst>(ICmp.getOperand(0));
6071 auto *NonNegInst1 = dyn_cast<PossiblyNonNegInst>(ICmp.getOperand(1));
6073 bool IsNonNeg0 = NonNegInst0 && NonNegInst0->hasNonNeg();
6074 bool IsNonNeg1 = NonNegInst1 && NonNegInst1->hasNonNeg();
6076 if ((IsZext0 && IsNonNeg0) || (IsZext1 && IsNonNeg1))
6077 IsSignedExt = true;
6078 else
6079 return nullptr;
6082 // Not an extension from the same type?
6083 Type *XTy = X->getType(), *YTy = Y->getType();
6084 if (XTy != YTy) {
6085 // One of the casts must have one use because we are creating a new cast.
6086 if (!ICmp.getOperand(0)->hasOneUse() && !ICmp.getOperand(1)->hasOneUse())
6087 return nullptr;
6088 // Extend the narrower operand to the type of the wider operand.
6089 CastInst::CastOps CastOpcode =
6090 IsSignedExt ? Instruction::SExt : Instruction::ZExt;
6091 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
6092 X = Builder.CreateCast(CastOpcode, X, YTy);
6093 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
6094 Y = Builder.CreateCast(CastOpcode, Y, XTy);
6095 else
6096 return nullptr;
6099 // (zext X) == (zext Y) --> X == Y
6100 // (sext X) == (sext Y) --> X == Y
6101 if (ICmp.isEquality())
6102 return new ICmpInst(ICmp.getPredicate(), X, Y);
6104 // A signed comparison of sign extended values simplifies into a
6105 // signed comparison.
6106 if (IsSignedCmp && IsSignedExt)
6107 return new ICmpInst(ICmp.getPredicate(), X, Y);
6109 // The other three cases all fold into an unsigned comparison.
6110 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
6113 // Below here, we are only folding a compare with constant.
6114 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
6115 if (!C)
6116 return nullptr;
6118 // If a lossless truncate is possible...
6119 Type *SrcTy = CastOp0->getSrcTy();
6120 Constant *Res = getLosslessTrunc(C, SrcTy, CastOp0->getOpcode());
6121 if (Res) {
6122 if (ICmp.isEquality())
6123 return new ICmpInst(ICmp.getPredicate(), X, Res);
6125 // A signed comparison of sign extended values simplifies into a
6126 // signed comparison.
6127 if (IsSignedExt && IsSignedCmp)
6128 return new ICmpInst(ICmp.getPredicate(), X, Res);
6130 // The other three cases all fold into an unsigned comparison.
6131 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res);
6134 // The re-extended constant changed, partly changed (in the case of a vector),
6135 // or could not be determined to be equal (in the case of a constant
6136 // expression), so the constant cannot be represented in the shorter type.
6137 // All the cases that fold to true or false will have already been handled
6138 // by simplifyICmpInst, so only deal with the tricky case.
6139 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
6140 return nullptr;
6142 // Is source op positive?
6143 // icmp ult (sext X), C --> icmp sgt X, -1
6144 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
6145 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
6147 // Is source op negative?
6148 // icmp ugt (sext X), C --> icmp slt X, 0
6149 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
6150 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
6153 /// Handle icmp (cast x), (cast or constant).
6154 Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
6155 // If any operand of ICmp is a inttoptr roundtrip cast then remove it as
6156 // icmp compares only pointer's value.
6157 // icmp (inttoptr (ptrtoint p1)), p2 --> icmp p1, p2.
6158 Value *SimplifiedOp0 = simplifyIntToPtrRoundTripCast(ICmp.getOperand(0));
6159 Value *SimplifiedOp1 = simplifyIntToPtrRoundTripCast(ICmp.getOperand(1));
6160 if (SimplifiedOp0 || SimplifiedOp1)
6161 return new ICmpInst(ICmp.getPredicate(),
6162 SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(0),
6163 SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(1));
6165 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
6166 if (!CastOp0)
6167 return nullptr;
6168 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
6169 return nullptr;
6171 Value *Op0Src = CastOp0->getOperand(0);
6172 Type *SrcTy = CastOp0->getSrcTy();
6173 Type *DestTy = CastOp0->getDestTy();
6175 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6176 // integer type is the same size as the pointer type.
6177 auto CompatibleSizes = [&](Type *PtrTy, Type *IntTy) {
6178 if (isa<VectorType>(PtrTy)) {
6179 PtrTy = cast<VectorType>(PtrTy)->getElementType();
6180 IntTy = cast<VectorType>(IntTy)->getElementType();
6182 return DL.getPointerTypeSizeInBits(PtrTy) == IntTy->getIntegerBitWidth();
6184 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
6185 CompatibleSizes(SrcTy, DestTy)) {
6186 Value *NewOp1 = nullptr;
6187 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
6188 Value *PtrSrc = PtrToIntOp1->getOperand(0);
6189 if (PtrSrc->getType() == Op0Src->getType())
6190 NewOp1 = PtrToIntOp1->getOperand(0);
6191 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
6192 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6195 if (NewOp1)
6196 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
6199 // Do the same in the other direction for icmp (inttoptr x), (inttoptr/c).
6200 if (CastOp0->getOpcode() == Instruction::IntToPtr &&
6201 CompatibleSizes(DestTy, SrcTy)) {
6202 Value *NewOp1 = nullptr;
6203 if (auto *IntToPtrOp1 = dyn_cast<IntToPtrInst>(ICmp.getOperand(1))) {
6204 Value *IntSrc = IntToPtrOp1->getOperand(0);
6205 if (IntSrc->getType() == Op0Src->getType())
6206 NewOp1 = IntToPtrOp1->getOperand(0);
6207 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
6208 NewOp1 = ConstantFoldConstant(ConstantExpr::getPtrToInt(RHSC, SrcTy), DL);
6211 if (NewOp1)
6212 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
6215 if (Instruction *R = foldICmpWithTrunc(ICmp))
6216 return R;
6218 return foldICmpWithZextOrSext(ICmp);
6221 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS, bool IsSigned) {
6222 switch (BinaryOp) {
6223 default:
6224 llvm_unreachable("Unsupported binary op");
6225 case Instruction::Add:
6226 case Instruction::Sub:
6227 return match(RHS, m_Zero());
6228 case Instruction::Mul:
6229 return !(RHS->getType()->isIntOrIntVectorTy(1) && IsSigned) &&
6230 match(RHS, m_One());
6234 OverflowResult
6235 InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
6236 bool IsSigned, Value *LHS, Value *RHS,
6237 Instruction *CxtI) const {
6238 switch (BinaryOp) {
6239 default:
6240 llvm_unreachable("Unsupported binary op");
6241 case Instruction::Add:
6242 if (IsSigned)
6243 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
6244 else
6245 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
6246 case Instruction::Sub:
6247 if (IsSigned)
6248 return computeOverflowForSignedSub(LHS, RHS, CxtI);
6249 else
6250 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
6251 case Instruction::Mul:
6252 if (IsSigned)
6253 return computeOverflowForSignedMul(LHS, RHS, CxtI);
6254 else
6255 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
6259 bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
6260 bool IsSigned, Value *LHS,
6261 Value *RHS, Instruction &OrigI,
6262 Value *&Result,
6263 Constant *&Overflow) {
6264 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
6265 std::swap(LHS, RHS);
6267 // If the overflow check was an add followed by a compare, the insertion point
6268 // may be pointing to the compare. We want to insert the new instructions
6269 // before the add in case there are uses of the add between the add and the
6270 // compare.
6271 Builder.SetInsertPoint(&OrigI);
6273 Type *OverflowTy = Type::getInt1Ty(LHS->getContext());
6274 if (auto *LHSTy = dyn_cast<VectorType>(LHS->getType()))
6275 OverflowTy = VectorType::get(OverflowTy, LHSTy->getElementCount());
6277 if (isNeutralValue(BinaryOp, RHS, IsSigned)) {
6278 Result = LHS;
6279 Overflow = ConstantInt::getFalse(OverflowTy);
6280 return true;
6283 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
6284 case OverflowResult::MayOverflow:
6285 return false;
6286 case OverflowResult::AlwaysOverflowsLow:
6287 case OverflowResult::AlwaysOverflowsHigh:
6288 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
6289 Result->takeName(&OrigI);
6290 Overflow = ConstantInt::getTrue(OverflowTy);
6291 return true;
6292 case OverflowResult::NeverOverflows:
6293 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
6294 Result->takeName(&OrigI);
6295 Overflow = ConstantInt::getFalse(OverflowTy);
6296 if (auto *Inst = dyn_cast<Instruction>(Result)) {
6297 if (IsSigned)
6298 Inst->setHasNoSignedWrap();
6299 else
6300 Inst->setHasNoUnsignedWrap();
6302 return true;
6305 llvm_unreachable("Unexpected overflow result");
6308 /// Recognize and process idiom involving test for multiplication
6309 /// overflow.
6311 /// The caller has matched a pattern of the form:
6312 /// I = cmp u (mul(zext A, zext B), V
6313 /// The function checks if this is a test for overflow and if so replaces
6314 /// multiplication with call to 'mul.with.overflow' intrinsic.
6316 /// \param I Compare instruction.
6317 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
6318 /// the compare instruction. Must be of integer type.
6319 /// \param OtherVal The other argument of compare instruction.
6320 /// \returns Instruction which must replace the compare instruction, NULL if no
6321 /// replacement required.
6322 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
6323 const APInt *OtherVal,
6324 InstCombinerImpl &IC) {
6325 // Don't bother doing this transformation for pointers, don't do it for
6326 // vectors.
6327 if (!isa<IntegerType>(MulVal->getType()))
6328 return nullptr;
6330 auto *MulInstr = dyn_cast<Instruction>(MulVal);
6331 if (!MulInstr)
6332 return nullptr;
6333 assert(MulInstr->getOpcode() == Instruction::Mul);
6335 auto *LHS = cast<ZExtInst>(MulInstr->getOperand(0)),
6336 *RHS = cast<ZExtInst>(MulInstr->getOperand(1));
6337 assert(LHS->getOpcode() == Instruction::ZExt);
6338 assert(RHS->getOpcode() == Instruction::ZExt);
6339 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
6341 // Calculate type and width of the result produced by mul.with.overflow.
6342 Type *TyA = A->getType(), *TyB = B->getType();
6343 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
6344 WidthB = TyB->getPrimitiveSizeInBits();
6345 unsigned MulWidth;
6346 Type *MulType;
6347 if (WidthB > WidthA) {
6348 MulWidth = WidthB;
6349 MulType = TyB;
6350 } else {
6351 MulWidth = WidthA;
6352 MulType = TyA;
6355 // In order to replace the original mul with a narrower mul.with.overflow,
6356 // all uses must ignore upper bits of the product. The number of used low
6357 // bits must be not greater than the width of mul.with.overflow.
6358 if (MulVal->hasNUsesOrMore(2))
6359 for (User *U : MulVal->users()) {
6360 if (U == &I)
6361 continue;
6362 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
6363 // Check if truncation ignores bits above MulWidth.
6364 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
6365 if (TruncWidth > MulWidth)
6366 return nullptr;
6367 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
6368 // Check if AND ignores bits above MulWidth.
6369 if (BO->getOpcode() != Instruction::And)
6370 return nullptr;
6371 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6372 const APInt &CVal = CI->getValue();
6373 if (CVal.getBitWidth() - CVal.countl_zero() > MulWidth)
6374 return nullptr;
6375 } else {
6376 // In this case we could have the operand of the binary operation
6377 // being defined in another block, and performing the replacement
6378 // could break the dominance relation.
6379 return nullptr;
6381 } else {
6382 // Other uses prohibit this transformation.
6383 return nullptr;
6387 // Recognize patterns
6388 switch (I.getPredicate()) {
6389 case ICmpInst::ICMP_UGT: {
6390 // Recognize pattern:
6391 // mulval = mul(zext A, zext B)
6392 // cmp ugt mulval, max
6393 APInt MaxVal = APInt::getMaxValue(MulWidth);
6394 MaxVal = MaxVal.zext(OtherVal->getBitWidth());
6395 if (MaxVal.eq(*OtherVal))
6396 break; // Recognized
6397 return nullptr;
6400 case ICmpInst::ICMP_ULT: {
6401 // Recognize pattern:
6402 // mulval = mul(zext A, zext B)
6403 // cmp ule mulval, max + 1
6404 APInt MaxVal = APInt::getOneBitSet(OtherVal->getBitWidth(), MulWidth);
6405 if (MaxVal.eq(*OtherVal))
6406 break; // Recognized
6407 return nullptr;
6410 default:
6411 return nullptr;
6414 InstCombiner::BuilderTy &Builder = IC.Builder;
6415 Builder.SetInsertPoint(MulInstr);
6417 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
6418 Value *MulA = A, *MulB = B;
6419 if (WidthA < MulWidth)
6420 MulA = Builder.CreateZExt(A, MulType);
6421 if (WidthB < MulWidth)
6422 MulB = Builder.CreateZExt(B, MulType);
6423 CallInst *Call =
6424 Builder.CreateIntrinsic(Intrinsic::umul_with_overflow, MulType,
6425 {MulA, MulB}, /*FMFSource=*/nullptr, "umul");
6426 IC.addToWorklist(MulInstr);
6428 // If there are uses of mul result other than the comparison, we know that
6429 // they are truncation or binary AND. Change them to use result of
6430 // mul.with.overflow and adjust properly mask/size.
6431 if (MulVal->hasNUsesOrMore(2)) {
6432 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
6433 for (User *U : make_early_inc_range(MulVal->users())) {
6434 if (U == &I)
6435 continue;
6436 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
6437 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
6438 IC.replaceInstUsesWith(*TI, Mul);
6439 else
6440 TI->setOperand(0, Mul);
6441 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
6442 assert(BO->getOpcode() == Instruction::And);
6443 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
6444 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
6445 APInt ShortMask = CI->getValue().trunc(MulWidth);
6446 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
6447 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
6448 IC.replaceInstUsesWith(*BO, Zext);
6449 } else {
6450 llvm_unreachable("Unexpected Binary operation");
6452 IC.addToWorklist(cast<Instruction>(U));
6456 // The original icmp gets replaced with the overflow value, maybe inverted
6457 // depending on predicate.
6458 if (I.getPredicate() == ICmpInst::ICMP_ULT) {
6459 Value *Res = Builder.CreateExtractValue(Call, 1);
6460 return BinaryOperator::CreateNot(Res);
6463 return ExtractValueInst::Create(Call, 1);
6466 /// When performing a comparison against a constant, it is possible that not all
6467 /// the bits in the LHS are demanded. This helper method computes the mask that
6468 /// IS demanded.
6469 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
6470 const APInt *RHS;
6471 if (!match(I.getOperand(1), m_APInt(RHS)))
6472 return APInt::getAllOnes(BitWidth);
6474 // If this is a normal comparison, it demands all bits. If it is a sign bit
6475 // comparison, it only demands the sign bit.
6476 bool UnusedBit;
6477 if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
6478 return APInt::getSignMask(BitWidth);
6480 switch (I.getPredicate()) {
6481 // For a UGT comparison, we don't care about any bits that
6482 // correspond to the trailing ones of the comparand. The value of these
6483 // bits doesn't impact the outcome of the comparison, because any value
6484 // greater than the RHS must differ in a bit higher than these due to carry.
6485 case ICmpInst::ICMP_UGT:
6486 return APInt::getBitsSetFrom(BitWidth, RHS->countr_one());
6488 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
6489 // Any value less than the RHS must differ in a higher bit because of carries.
6490 case ICmpInst::ICMP_ULT:
6491 return APInt::getBitsSetFrom(BitWidth, RHS->countr_zero());
6493 default:
6494 return APInt::getAllOnes(BitWidth);
6498 /// Check that one use is in the same block as the definition and all
6499 /// other uses are in blocks dominated by a given block.
6501 /// \param DI Definition
6502 /// \param UI Use
6503 /// \param DB Block that must dominate all uses of \p DI outside
6504 /// the parent block
6505 /// \return true when \p UI is the only use of \p DI in the parent block
6506 /// and all other uses of \p DI are in blocks dominated by \p DB.
6508 bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
6509 const Instruction *UI,
6510 const BasicBlock *DB) const {
6511 assert(DI && UI && "Instruction not defined\n");
6512 // Ignore incomplete definitions.
6513 if (!DI->getParent())
6514 return false;
6515 // DI and UI must be in the same block.
6516 if (DI->getParent() != UI->getParent())
6517 return false;
6518 // Protect from self-referencing blocks.
6519 if (DI->getParent() == DB)
6520 return false;
6521 for (const User *U : DI->users()) {
6522 auto *Usr = cast<Instruction>(U);
6523 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
6524 return false;
6526 return true;
6529 /// Return true when the instruction sequence within a block is select-cmp-br.
6530 static bool isChainSelectCmpBranch(const SelectInst *SI) {
6531 const BasicBlock *BB = SI->getParent();
6532 if (!BB)
6533 return false;
6534 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
6535 if (!BI || BI->getNumSuccessors() != 2)
6536 return false;
6537 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
6538 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
6539 return false;
6540 return true;
6543 /// True when a select result is replaced by one of its operands
6544 /// in select-icmp sequence. This will eventually result in the elimination
6545 /// of the select.
6547 /// \param SI Select instruction
6548 /// \param Icmp Compare instruction
6549 /// \param SIOpd Operand that replaces the select
6551 /// Notes:
6552 /// - The replacement is global and requires dominator information
6553 /// - The caller is responsible for the actual replacement
6555 /// Example:
6557 /// entry:
6558 /// %4 = select i1 %3, %C* %0, %C* null
6559 /// %5 = icmp eq %C* %4, null
6560 /// br i1 %5, label %9, label %7
6561 /// ...
6562 /// ; <label>:7 ; preds = %entry
6563 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
6564 /// ...
6566 /// can be transformed to
6568 /// %5 = icmp eq %C* %0, null
6569 /// %6 = select i1 %3, i1 %5, i1 true
6570 /// br i1 %6, label %9, label %7
6571 /// ...
6572 /// ; <label>:7 ; preds = %entry
6573 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
6575 /// Similar when the first operand of the select is a constant or/and
6576 /// the compare is for not equal rather than equal.
6578 /// NOTE: The function is only called when the select and compare constants
6579 /// are equal, the optimization can work only for EQ predicates. This is not a
6580 /// major restriction since a NE compare should be 'normalized' to an equal
6581 /// compare, which usually happens in the combiner and test case
6582 /// select-cmp-br.ll checks for it.
6583 bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
6584 const ICmpInst *Icmp,
6585 const unsigned SIOpd) {
6586 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
6587 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
6588 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
6589 // The check for the single predecessor is not the best that can be
6590 // done. But it protects efficiently against cases like when SI's
6591 // home block has two successors, Succ and Succ1, and Succ1 predecessor
6592 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
6593 // replaced can be reached on either path. So the uniqueness check
6594 // guarantees that the path all uses of SI (outside SI's parent) are on
6595 // is disjoint from all other paths out of SI. But that information
6596 // is more expensive to compute, and the trade-off here is in favor
6597 // of compile-time. It should also be noticed that we check for a single
6598 // predecessor and not only uniqueness. This to handle the situation when
6599 // Succ and Succ1 points to the same basic block.
6600 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
6601 NumSel++;
6602 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
6603 return true;
6606 return false;
6609 /// Try to fold the comparison based on range information we can get by checking
6610 /// whether bits are known to be zero or one in the inputs.
6611 Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
6612 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6613 Type *Ty = Op0->getType();
6614 ICmpInst::Predicate Pred = I.getPredicate();
6616 // Get scalar or pointer size.
6617 unsigned BitWidth = Ty->isIntOrIntVectorTy()
6618 ? Ty->getScalarSizeInBits()
6619 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
6621 if (!BitWidth)
6622 return nullptr;
6624 KnownBits Op0Known(BitWidth);
6625 KnownBits Op1Known(BitWidth);
6628 // Don't use dominating conditions when folding icmp using known bits. This
6629 // may convert signed into unsigned predicates in ways that other passes
6630 // (especially IndVarSimplify) may not be able to reliably undo.
6631 SimplifyQuery Q = SQ.getWithoutDomCondCache().getWithInstruction(&I);
6632 if (SimplifyDemandedBits(&I, 0, getDemandedBitsLHSMask(I, BitWidth),
6633 Op0Known, /*Depth=*/0, Q))
6634 return &I;
6636 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnes(BitWidth), Op1Known,
6637 /*Depth=*/0, Q))
6638 return &I;
6641 if (!isa<Constant>(Op0) && Op0Known.isConstant())
6642 return new ICmpInst(
6643 Pred, ConstantExpr::getIntegerValue(Ty, Op0Known.getConstant()), Op1);
6644 if (!isa<Constant>(Op1) && Op1Known.isConstant())
6645 return new ICmpInst(
6646 Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Known.getConstant()));
6648 if (std::optional<bool> Res = ICmpInst::compare(Op0Known, Op1Known, Pred))
6649 return replaceInstUsesWith(I, ConstantInt::getBool(I.getType(), *Res));
6651 // Given the known and unknown bits, compute a range that the LHS could be
6652 // in. Compute the Min, Max and RHS values based on the known bits. For the
6653 // EQ and NE we use unsigned values.
6654 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6655 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6656 if (I.isSigned()) {
6657 Op0Min = Op0Known.getSignedMinValue();
6658 Op0Max = Op0Known.getSignedMaxValue();
6659 Op1Min = Op1Known.getSignedMinValue();
6660 Op1Max = Op1Known.getSignedMaxValue();
6661 } else {
6662 Op0Min = Op0Known.getMinValue();
6663 Op0Max = Op0Known.getMaxValue();
6664 Op1Min = Op1Known.getMinValue();
6665 Op1Max = Op1Known.getMaxValue();
6668 // Don't break up a clamp pattern -- (min(max X, Y), Z) -- by replacing a
6669 // min/max canonical compare with some other compare. That could lead to
6670 // conflict with select canonicalization and infinite looping.
6671 // FIXME: This constraint may go away if min/max intrinsics are canonical.
6672 auto isMinMaxCmp = [&](Instruction &Cmp) {
6673 if (!Cmp.hasOneUse())
6674 return false;
6675 Value *A, *B;
6676 SelectPatternFlavor SPF = matchSelectPattern(Cmp.user_back(), A, B).Flavor;
6677 if (!SelectPatternResult::isMinOrMax(SPF))
6678 return false;
6679 return match(Op0, m_MaxOrMin(m_Value(), m_Value())) ||
6680 match(Op1, m_MaxOrMin(m_Value(), m_Value()));
6682 if (!isMinMaxCmp(I)) {
6683 switch (Pred) {
6684 default:
6685 break;
6686 case ICmpInst::ICMP_ULT: {
6687 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6688 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6689 const APInt *CmpC;
6690 if (match(Op1, m_APInt(CmpC))) {
6691 // A <u C -> A == C-1 if min(A)+1 == C
6692 if (*CmpC == Op0Min + 1)
6693 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6694 ConstantInt::get(Op1->getType(), *CmpC - 1));
6695 // X <u C --> X == 0, if the number of zero bits in the bottom of X
6696 // exceeds the log2 of C.
6697 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
6698 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6699 Constant::getNullValue(Op1->getType()));
6701 break;
6703 case ICmpInst::ICMP_UGT: {
6704 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6705 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6706 const APInt *CmpC;
6707 if (match(Op1, m_APInt(CmpC))) {
6708 // A >u C -> A == C+1 if max(a)-1 == C
6709 if (*CmpC == Op0Max - 1)
6710 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6711 ConstantInt::get(Op1->getType(), *CmpC + 1));
6712 // X >u C --> X != 0, if the number of zero bits in the bottom of X
6713 // exceeds the log2 of C.
6714 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
6715 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
6716 Constant::getNullValue(Op1->getType()));
6718 break;
6720 case ICmpInst::ICMP_SLT: {
6721 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6722 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6723 const APInt *CmpC;
6724 if (match(Op1, m_APInt(CmpC))) {
6725 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
6726 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6727 ConstantInt::get(Op1->getType(), *CmpC - 1));
6729 break;
6731 case ICmpInst::ICMP_SGT: {
6732 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6733 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6734 const APInt *CmpC;
6735 if (match(Op1, m_APInt(CmpC))) {
6736 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
6737 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6738 ConstantInt::get(Op1->getType(), *CmpC + 1));
6740 break;
6745 // Based on the range information we know about the LHS, see if we can
6746 // simplify this comparison. For example, (x&4) < 8 is always true.
6747 switch (Pred) {
6748 default:
6749 break;
6750 case ICmpInst::ICMP_EQ:
6751 case ICmpInst::ICMP_NE: {
6752 // If all bits are known zero except for one, then we know at most one bit
6753 // is set. If the comparison is against zero, then this is a check to see if
6754 // *that* bit is set.
6755 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
6756 if (Op1Known.isZero()) {
6757 // If the LHS is an AND with the same constant, look through it.
6758 Value *LHS = nullptr;
6759 const APInt *LHSC;
6760 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
6761 *LHSC != Op0KnownZeroInverted)
6762 LHS = Op0;
6764 Value *X;
6765 const APInt *C1;
6766 if (match(LHS, m_Shl(m_Power2(C1), m_Value(X)))) {
6767 Type *XTy = X->getType();
6768 unsigned Log2C1 = C1->countr_zero();
6769 APInt C2 = Op0KnownZeroInverted;
6770 APInt C2Pow2 = (C2 & ~(*C1 - 1)) + *C1;
6771 if (C2Pow2.isPowerOf2()) {
6772 // iff (C1 is pow2) & ((C2 & ~(C1-1)) + C1) is pow2):
6773 // ((C1 << X) & C2) == 0 -> X >= (Log2(C2+C1) - Log2(C1))
6774 // ((C1 << X) & C2) != 0 -> X < (Log2(C2+C1) - Log2(C1))
6775 unsigned Log2C2 = C2Pow2.countr_zero();
6776 auto *CmpC = ConstantInt::get(XTy, Log2C2 - Log2C1);
6777 auto NewPred =
6778 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
6779 return new ICmpInst(NewPred, X, CmpC);
6784 // Op0 eq C_Pow2 -> Op0 ne 0 if Op0 is known to be C_Pow2 or zero.
6785 if (Op1Known.isConstant() && Op1Known.getConstant().isPowerOf2() &&
6786 (Op0Known & Op1Known) == Op0Known)
6787 return new ICmpInst(CmpInst::getInversePredicate(Pred), Op0,
6788 ConstantInt::getNullValue(Op1->getType()));
6789 break;
6791 case ICmpInst::ICMP_SGE:
6792 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
6793 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
6794 break;
6795 case ICmpInst::ICMP_SLE:
6796 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
6797 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
6798 break;
6799 case ICmpInst::ICMP_UGE:
6800 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
6801 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
6802 break;
6803 case ICmpInst::ICMP_ULE:
6804 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
6805 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
6806 break;
6809 // Turn a signed comparison into an unsigned one if both operands are known to
6810 // have the same sign. Set samesign if possible (except for equality
6811 // predicates).
6812 if ((I.isSigned() || (I.isUnsigned() && !I.hasSameSign())) &&
6813 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
6814 (Op0Known.One.isNegative() && Op1Known.One.isNegative()))) {
6815 I.setPredicate(I.getUnsignedPredicate());
6816 I.setSameSign();
6817 return &I;
6820 return nullptr;
6823 /// If one operand of an icmp is effectively a bool (value range of {0,1}),
6824 /// then try to reduce patterns based on that limit.
6825 Instruction *InstCombinerImpl::foldICmpUsingBoolRange(ICmpInst &I) {
6826 Value *X, *Y;
6827 CmpPredicate Pred;
6829 // X must be 0 and bool must be true for "ULT":
6830 // X <u (zext i1 Y) --> (X == 0) & Y
6831 if (match(&I, m_c_ICmp(Pred, m_Value(X), m_OneUse(m_ZExt(m_Value(Y))))) &&
6832 Y->getType()->isIntOrIntVectorTy(1) && Pred == ICmpInst::ICMP_ULT)
6833 return BinaryOperator::CreateAnd(Builder.CreateIsNull(X), Y);
6835 // X must be 0 or bool must be true for "ULE":
6836 // X <=u (sext i1 Y) --> (X == 0) | Y
6837 if (match(&I, m_c_ICmp(Pred, m_Value(X), m_OneUse(m_SExt(m_Value(Y))))) &&
6838 Y->getType()->isIntOrIntVectorTy(1) && Pred == ICmpInst::ICMP_ULE)
6839 return BinaryOperator::CreateOr(Builder.CreateIsNull(X), Y);
6841 // icmp eq/ne X, (zext/sext (icmp eq/ne X, C))
6842 CmpPredicate Pred1, Pred2;
6843 const APInt *C;
6844 Instruction *ExtI;
6845 if (match(&I, m_c_ICmp(Pred1, m_Value(X),
6846 m_CombineAnd(m_Instruction(ExtI),
6847 m_ZExtOrSExt(m_ICmp(Pred2, m_Deferred(X),
6848 m_APInt(C)))))) &&
6849 ICmpInst::isEquality(Pred1) && ICmpInst::isEquality(Pred2)) {
6850 bool IsSExt = ExtI->getOpcode() == Instruction::SExt;
6851 bool HasOneUse = ExtI->hasOneUse() && ExtI->getOperand(0)->hasOneUse();
6852 auto CreateRangeCheck = [&] {
6853 Value *CmpV1 =
6854 Builder.CreateICmp(Pred1, X, Constant::getNullValue(X->getType()));
6855 Value *CmpV2 = Builder.CreateICmp(
6856 Pred1, X, ConstantInt::getSigned(X->getType(), IsSExt ? -1 : 1));
6857 return BinaryOperator::Create(
6858 Pred1 == ICmpInst::ICMP_EQ ? Instruction::Or : Instruction::And,
6859 CmpV1, CmpV2);
6861 if (C->isZero()) {
6862 if (Pred2 == ICmpInst::ICMP_EQ) {
6863 // icmp eq X, (zext/sext (icmp eq X, 0)) --> false
6864 // icmp ne X, (zext/sext (icmp eq X, 0)) --> true
6865 return replaceInstUsesWith(
6866 I, ConstantInt::getBool(I.getType(), Pred1 == ICmpInst::ICMP_NE));
6867 } else if (!IsSExt || HasOneUse) {
6868 // icmp eq X, (zext (icmp ne X, 0)) --> X == 0 || X == 1
6869 // icmp ne X, (zext (icmp ne X, 0)) --> X != 0 && X != 1
6870 // icmp eq X, (sext (icmp ne X, 0)) --> X == 0 || X == -1
6871 // icmp ne X, (sext (icmp ne X, 0)) --> X != 0 && X == -1
6872 return CreateRangeCheck();
6874 } else if (IsSExt ? C->isAllOnes() : C->isOne()) {
6875 if (Pred2 == ICmpInst::ICMP_NE) {
6876 // icmp eq X, (zext (icmp ne X, 1)) --> false
6877 // icmp ne X, (zext (icmp ne X, 1)) --> true
6878 // icmp eq X, (sext (icmp ne X, -1)) --> false
6879 // icmp ne X, (sext (icmp ne X, -1)) --> true
6880 return replaceInstUsesWith(
6881 I, ConstantInt::getBool(I.getType(), Pred1 == ICmpInst::ICMP_NE));
6882 } else if (!IsSExt || HasOneUse) {
6883 // icmp eq X, (zext (icmp eq X, 1)) --> X == 0 || X == 1
6884 // icmp ne X, (zext (icmp eq X, 1)) --> X != 0 && X != 1
6885 // icmp eq X, (sext (icmp eq X, -1)) --> X == 0 || X == -1
6886 // icmp ne X, (sext (icmp eq X, -1)) --> X != 0 && X == -1
6887 return CreateRangeCheck();
6889 } else {
6890 // when C != 0 && C != 1:
6891 // icmp eq X, (zext (icmp eq X, C)) --> icmp eq X, 0
6892 // icmp eq X, (zext (icmp ne X, C)) --> icmp eq X, 1
6893 // icmp ne X, (zext (icmp eq X, C)) --> icmp ne X, 0
6894 // icmp ne X, (zext (icmp ne X, C)) --> icmp ne X, 1
6895 // when C != 0 && C != -1:
6896 // icmp eq X, (sext (icmp eq X, C)) --> icmp eq X, 0
6897 // icmp eq X, (sext (icmp ne X, C)) --> icmp eq X, -1
6898 // icmp ne X, (sext (icmp eq X, C)) --> icmp ne X, 0
6899 // icmp ne X, (sext (icmp ne X, C)) --> icmp ne X, -1
6900 return ICmpInst::Create(
6901 Instruction::ICmp, Pred1, X,
6902 ConstantInt::getSigned(X->getType(), Pred2 == ICmpInst::ICMP_NE
6903 ? (IsSExt ? -1 : 1)
6904 : 0));
6908 return nullptr;
6911 std::optional<std::pair<CmpPredicate, Constant *>>
6912 InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpPredicate Pred,
6913 Constant *C) {
6914 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&
6915 "Only for relational integer predicates.");
6917 Type *Type = C->getType();
6918 bool IsSigned = ICmpInst::isSigned(Pred);
6920 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
6921 bool WillIncrement =
6922 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
6924 // Check if the constant operand can be safely incremented/decremented
6925 // without overflowing/underflowing.
6926 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
6927 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
6930 Constant *SafeReplacementConstant = nullptr;
6931 if (auto *CI = dyn_cast<ConstantInt>(C)) {
6932 // Bail out if the constant can't be safely incremented/decremented.
6933 if (!ConstantIsOk(CI))
6934 return std::nullopt;
6935 } else if (auto *FVTy = dyn_cast<FixedVectorType>(Type)) {
6936 unsigned NumElts = FVTy->getNumElements();
6937 for (unsigned i = 0; i != NumElts; ++i) {
6938 Constant *Elt = C->getAggregateElement(i);
6939 if (!Elt)
6940 return std::nullopt;
6942 if (isa<UndefValue>(Elt))
6943 continue;
6945 // Bail out if we can't determine if this constant is min/max or if we
6946 // know that this constant is min/max.
6947 auto *CI = dyn_cast<ConstantInt>(Elt);
6948 if (!CI || !ConstantIsOk(CI))
6949 return std::nullopt;
6951 if (!SafeReplacementConstant)
6952 SafeReplacementConstant = CI;
6954 } else if (isa<VectorType>(C->getType())) {
6955 // Handle scalable splat
6956 Value *SplatC = C->getSplatValue();
6957 auto *CI = dyn_cast_or_null<ConstantInt>(SplatC);
6958 // Bail out if the constant can't be safely incremented/decremented.
6959 if (!CI || !ConstantIsOk(CI))
6960 return std::nullopt;
6961 } else {
6962 // ConstantExpr?
6963 return std::nullopt;
6966 // It may not be safe to change a compare predicate in the presence of
6967 // undefined elements, so replace those elements with the first safe constant
6968 // that we found.
6969 // TODO: in case of poison, it is safe; let's replace undefs only.
6970 if (C->containsUndefOrPoisonElement()) {
6971 assert(SafeReplacementConstant && "Replacement constant not set");
6972 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
6975 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
6977 // Increment or decrement the constant.
6978 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
6979 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
6981 return std::make_pair(NewPred, NewC);
6984 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
6985 /// it into the appropriate icmp lt or icmp gt instruction. This transform
6986 /// allows them to be folded in visitICmpInst.
6987 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
6988 ICmpInst::Predicate Pred = I.getPredicate();
6989 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
6990 InstCombiner::isCanonicalPredicate(Pred))
6991 return nullptr;
6993 Value *Op0 = I.getOperand(0);
6994 Value *Op1 = I.getOperand(1);
6995 auto *Op1C = dyn_cast<Constant>(Op1);
6996 if (!Op1C)
6997 return nullptr;
6999 auto FlippedStrictness =
7000 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
7001 if (!FlippedStrictness)
7002 return nullptr;
7004 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
7007 /// If we have a comparison with a non-canonical predicate, if we can update
7008 /// all the users, invert the predicate and adjust all the users.
7009 CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
7010 // Is the predicate already canonical?
7011 CmpInst::Predicate Pred = I.getPredicate();
7012 if (InstCombiner::isCanonicalPredicate(Pred))
7013 return nullptr;
7015 // Can all users be adjusted to predicate inversion?
7016 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
7017 return nullptr;
7019 // Ok, we can canonicalize comparison!
7020 // Let's first invert the comparison's predicate.
7021 I.setPredicate(CmpInst::getInversePredicate(Pred));
7022 I.setName(I.getName() + ".not");
7024 // And, adapt users.
7025 freelyInvertAllUsersOf(&I);
7027 return &I;
7030 /// Integer compare with boolean values can always be turned into bitwise ops.
7031 static Instruction *canonicalizeICmpBool(ICmpInst &I,
7032 InstCombiner::BuilderTy &Builder) {
7033 Value *A = I.getOperand(0), *B = I.getOperand(1);
7034 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
7036 // A boolean compared to true/false can be simplified to Op0/true/false in
7037 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
7038 // Cases not handled by InstSimplify are always 'not' of Op0.
7039 if (match(B, m_Zero())) {
7040 switch (I.getPredicate()) {
7041 case CmpInst::ICMP_EQ: // A == 0 -> !A
7042 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
7043 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
7044 return BinaryOperator::CreateNot(A);
7045 default:
7046 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
7048 } else if (match(B, m_One())) {
7049 switch (I.getPredicate()) {
7050 case CmpInst::ICMP_NE: // A != 1 -> !A
7051 case CmpInst::ICMP_ULT: // A <u 1 -> !A
7052 case CmpInst::ICMP_SGT: // A >s -1 -> !A
7053 return BinaryOperator::CreateNot(A);
7054 default:
7055 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
7059 switch (I.getPredicate()) {
7060 default:
7061 llvm_unreachable("Invalid icmp instruction!");
7062 case ICmpInst::ICMP_EQ:
7063 // icmp eq i1 A, B -> ~(A ^ B)
7064 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
7066 case ICmpInst::ICMP_NE:
7067 // icmp ne i1 A, B -> A ^ B
7068 return BinaryOperator::CreateXor(A, B);
7070 case ICmpInst::ICMP_UGT:
7071 // icmp ugt -> icmp ult
7072 std::swap(A, B);
7073 [[fallthrough]];
7074 case ICmpInst::ICMP_ULT:
7075 // icmp ult i1 A, B -> ~A & B
7076 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
7078 case ICmpInst::ICMP_SGT:
7079 // icmp sgt -> icmp slt
7080 std::swap(A, B);
7081 [[fallthrough]];
7082 case ICmpInst::ICMP_SLT:
7083 // icmp slt i1 A, B -> A & ~B
7084 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
7086 case ICmpInst::ICMP_UGE:
7087 // icmp uge -> icmp ule
7088 std::swap(A, B);
7089 [[fallthrough]];
7090 case ICmpInst::ICMP_ULE:
7091 // icmp ule i1 A, B -> ~A | B
7092 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
7094 case ICmpInst::ICMP_SGE:
7095 // icmp sge -> icmp sle
7096 std::swap(A, B);
7097 [[fallthrough]];
7098 case ICmpInst::ICMP_SLE:
7099 // icmp sle i1 A, B -> A | ~B
7100 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
7104 // Transform pattern like:
7105 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
7106 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
7107 // Into:
7108 // (X l>> Y) != 0
7109 // (X l>> Y) == 0
7110 static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
7111 InstCombiner::BuilderTy &Builder) {
7112 CmpPredicate Pred, NewPred;
7113 Value *X, *Y;
7114 if (match(&Cmp,
7115 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
7116 switch (Pred) {
7117 case ICmpInst::ICMP_ULE:
7118 NewPred = ICmpInst::ICMP_NE;
7119 break;
7120 case ICmpInst::ICMP_UGT:
7121 NewPred = ICmpInst::ICMP_EQ;
7122 break;
7123 default:
7124 return nullptr;
7126 } else if (match(&Cmp, m_c_ICmp(Pred,
7127 m_OneUse(m_CombineOr(
7128 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
7129 m_Add(m_Shl(m_One(), m_Value(Y)),
7130 m_AllOnes()))),
7131 m_Value(X)))) {
7132 // The variant with 'add' is not canonical, (the variant with 'not' is)
7133 // we only get it because it has extra uses, and can't be canonicalized,
7135 switch (Pred) {
7136 case ICmpInst::ICMP_ULT:
7137 NewPred = ICmpInst::ICMP_NE;
7138 break;
7139 case ICmpInst::ICMP_UGE:
7140 NewPred = ICmpInst::ICMP_EQ;
7141 break;
7142 default:
7143 return nullptr;
7145 } else
7146 return nullptr;
7148 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
7149 Constant *Zero = Constant::getNullValue(NewX->getType());
7150 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
7153 static Instruction *foldVectorCmp(CmpInst &Cmp,
7154 InstCombiner::BuilderTy &Builder) {
7155 const CmpInst::Predicate Pred = Cmp.getPredicate();
7156 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
7157 Value *V1, *V2;
7159 auto createCmpReverse = [&](CmpInst::Predicate Pred, Value *X, Value *Y) {
7160 Value *V = Builder.CreateCmp(Pred, X, Y, Cmp.getName());
7161 if (auto *I = dyn_cast<Instruction>(V))
7162 I->copyIRFlags(&Cmp);
7163 Module *M = Cmp.getModule();
7164 Function *F = Intrinsic::getOrInsertDeclaration(
7165 M, Intrinsic::vector_reverse, V->getType());
7166 return CallInst::Create(F, V);
7169 if (match(LHS, m_VecReverse(m_Value(V1)))) {
7170 // cmp Pred, rev(V1), rev(V2) --> rev(cmp Pred, V1, V2)
7171 if (match(RHS, m_VecReverse(m_Value(V2))) &&
7172 (LHS->hasOneUse() || RHS->hasOneUse()))
7173 return createCmpReverse(Pred, V1, V2);
7175 // cmp Pred, rev(V1), RHSSplat --> rev(cmp Pred, V1, RHSSplat)
7176 if (LHS->hasOneUse() && isSplatValue(RHS))
7177 return createCmpReverse(Pred, V1, RHS);
7179 // cmp Pred, LHSSplat, rev(V2) --> rev(cmp Pred, LHSSplat, V2)
7180 else if (isSplatValue(LHS) && match(RHS, m_OneUse(m_VecReverse(m_Value(V2)))))
7181 return createCmpReverse(Pred, LHS, V2);
7183 ArrayRef<int> M;
7184 if (!match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(M))))
7185 return nullptr;
7187 // If both arguments of the cmp are shuffles that use the same mask and
7188 // shuffle within a single vector, move the shuffle after the cmp:
7189 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
7190 Type *V1Ty = V1->getType();
7191 if (match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(M))) &&
7192 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
7193 Value *NewCmp = Builder.CreateCmp(Pred, V1, V2);
7194 return new ShuffleVectorInst(NewCmp, M);
7197 // Try to canonicalize compare with splatted operand and splat constant.
7198 // TODO: We could generalize this for more than splats. See/use the code in
7199 // InstCombiner::foldVectorBinop().
7200 Constant *C;
7201 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
7202 return nullptr;
7204 // Length-changing splats are ok, so adjust the constants as needed:
7205 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
7206 Constant *ScalarC = C->getSplatValue(/* AllowPoison */ true);
7207 int MaskSplatIndex;
7208 if (ScalarC && match(M, m_SplatOrPoisonMask(MaskSplatIndex))) {
7209 // We allow poison in matching, but this transform removes it for safety.
7210 // Demanded elements analysis should be able to recover some/all of that.
7211 C = ConstantVector::getSplat(cast<VectorType>(V1Ty)->getElementCount(),
7212 ScalarC);
7213 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
7214 Value *NewCmp = Builder.CreateCmp(Pred, V1, C);
7215 return new ShuffleVectorInst(NewCmp, NewM);
7218 return nullptr;
7221 // extract(uadd.with.overflow(A, B), 0) ult A
7222 // -> extract(uadd.with.overflow(A, B), 1)
7223 static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
7224 CmpInst::Predicate Pred = I.getPredicate();
7225 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7227 Value *UAddOv;
7228 Value *A, *B;
7229 auto UAddOvResultPat = m_ExtractValue<0>(
7230 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
7231 if (match(Op0, UAddOvResultPat) &&
7232 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
7233 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
7234 (match(A, m_One()) || match(B, m_One()))) ||
7235 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
7236 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
7237 // extract(uadd.with.overflow(A, B), 0) < A
7238 // extract(uadd.with.overflow(A, 1), 0) == 0
7239 // extract(uadd.with.overflow(A, -1), 0) != -1
7240 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
7241 else if (match(Op1, UAddOvResultPat) &&
7242 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
7243 // A > extract(uadd.with.overflow(A, B), 0)
7244 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
7245 else
7246 return nullptr;
7248 return ExtractValueInst::Create(UAddOv, 1);
7251 static Instruction *foldICmpInvariantGroup(ICmpInst &I) {
7252 if (!I.getOperand(0)->getType()->isPointerTy() ||
7253 NullPointerIsDefined(
7254 I.getParent()->getParent(),
7255 I.getOperand(0)->getType()->getPointerAddressSpace())) {
7256 return nullptr;
7258 Instruction *Op;
7259 if (match(I.getOperand(0), m_Instruction(Op)) &&
7260 match(I.getOperand(1), m_Zero()) &&
7261 Op->isLaunderOrStripInvariantGroup()) {
7262 return ICmpInst::Create(Instruction::ICmp, I.getPredicate(),
7263 Op->getOperand(0), I.getOperand(1));
7265 return nullptr;
7268 /// This function folds patterns produced by lowering of reduce idioms, such as
7269 /// llvm.vector.reduce.and which are lowered into instruction chains. This code
7270 /// attempts to generate fewer number of scalar comparisons instead of vector
7271 /// comparisons when possible.
7272 static Instruction *foldReductionIdiom(ICmpInst &I,
7273 InstCombiner::BuilderTy &Builder,
7274 const DataLayout &DL) {
7275 if (I.getType()->isVectorTy())
7276 return nullptr;
7277 CmpPredicate OuterPred, InnerPred;
7278 Value *LHS, *RHS;
7280 // Match lowering of @llvm.vector.reduce.and. Turn
7281 /// %vec_ne = icmp ne <8 x i8> %lhs, %rhs
7282 /// %scalar_ne = bitcast <8 x i1> %vec_ne to i8
7283 /// %res = icmp <pred> i8 %scalar_ne, 0
7285 /// into
7287 /// %lhs.scalar = bitcast <8 x i8> %lhs to i64
7288 /// %rhs.scalar = bitcast <8 x i8> %rhs to i64
7289 /// %res = icmp <pred> i64 %lhs.scalar, %rhs.scalar
7291 /// for <pred> in {ne, eq}.
7292 if (!match(&I, m_ICmp(OuterPred,
7293 m_OneUse(m_BitCast(m_OneUse(
7294 m_ICmp(InnerPred, m_Value(LHS), m_Value(RHS))))),
7295 m_Zero())))
7296 return nullptr;
7297 auto *LHSTy = dyn_cast<FixedVectorType>(LHS->getType());
7298 if (!LHSTy || !LHSTy->getElementType()->isIntegerTy())
7299 return nullptr;
7300 unsigned NumBits =
7301 LHSTy->getNumElements() * LHSTy->getElementType()->getIntegerBitWidth();
7302 // TODO: Relax this to "not wider than max legal integer type"?
7303 if (!DL.isLegalInteger(NumBits))
7304 return nullptr;
7306 if (ICmpInst::isEquality(OuterPred) && InnerPred == ICmpInst::ICMP_NE) {
7307 auto *ScalarTy = Builder.getIntNTy(NumBits);
7308 LHS = Builder.CreateBitCast(LHS, ScalarTy, LHS->getName() + ".scalar");
7309 RHS = Builder.CreateBitCast(RHS, ScalarTy, RHS->getName() + ".scalar");
7310 return ICmpInst::Create(Instruction::ICmp, OuterPred, LHS, RHS,
7311 I.getName());
7314 return nullptr;
7317 // This helper will be called with icmp operands in both orders.
7318 Instruction *InstCombinerImpl::foldICmpCommutative(CmpPredicate Pred,
7319 Value *Op0, Value *Op1,
7320 ICmpInst &CxtI) {
7321 // Try to optimize 'icmp GEP, P' or 'icmp P, GEP'.
7322 if (auto *GEP = dyn_cast<GEPOperator>(Op0))
7323 if (Instruction *NI = foldGEPICmp(GEP, Op1, Pred, CxtI))
7324 return NI;
7326 if (auto *SI = dyn_cast<SelectInst>(Op0))
7327 if (Instruction *NI = foldSelectICmp(Pred, SI, Op1, CxtI))
7328 return NI;
7330 if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op0))
7331 if (Instruction *Res = foldICmpWithMinMax(CxtI, MinMax, Op1, Pred))
7332 return Res;
7335 Value *X;
7336 const APInt *C;
7337 // icmp X+Cst, X
7338 if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
7339 return foldICmpAddOpConst(X, *C, Pred);
7342 // abs(X) >= X --> true
7343 // abs(X) u<= X --> true
7344 // abs(X) < X --> false
7345 // abs(X) u> X --> false
7346 // abs(X) u>= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7347 // abs(X) <= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7348 // abs(X) == X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7349 // abs(X) u< X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7350 // abs(X) > X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7351 // abs(X) != X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7353 Value *X;
7354 Constant *C;
7355 if (match(Op0, m_Intrinsic<Intrinsic::abs>(m_Value(X), m_Constant(C))) &&
7356 match(Op1, m_Specific(X))) {
7357 Value *NullValue = Constant::getNullValue(X->getType());
7358 Value *AllOnesValue = Constant::getAllOnesValue(X->getType());
7359 const APInt SMin =
7360 APInt::getSignedMinValue(X->getType()->getScalarSizeInBits());
7361 bool IsIntMinPosion = C->isAllOnesValue();
7362 switch (Pred) {
7363 case CmpInst::ICMP_ULE:
7364 case CmpInst::ICMP_SGE:
7365 return replaceInstUsesWith(CxtI, ConstantInt::getTrue(CxtI.getType()));
7366 case CmpInst::ICMP_UGT:
7367 case CmpInst::ICMP_SLT:
7368 return replaceInstUsesWith(CxtI, ConstantInt::getFalse(CxtI.getType()));
7369 case CmpInst::ICMP_UGE:
7370 case CmpInst::ICMP_SLE:
7371 case CmpInst::ICMP_EQ: {
7372 return replaceInstUsesWith(
7373 CxtI, IsIntMinPosion
7374 ? Builder.CreateICmpSGT(X, AllOnesValue)
7375 : Builder.CreateICmpULT(
7376 X, ConstantInt::get(X->getType(), SMin + 1)));
7378 case CmpInst::ICMP_ULT:
7379 case CmpInst::ICMP_SGT:
7380 case CmpInst::ICMP_NE: {
7381 return replaceInstUsesWith(
7382 CxtI, IsIntMinPosion
7383 ? Builder.CreateICmpSLT(X, NullValue)
7384 : Builder.CreateICmpUGT(
7385 X, ConstantInt::get(X->getType(), SMin)));
7387 default:
7388 llvm_unreachable("Invalid predicate!");
7393 const SimplifyQuery Q = SQ.getWithInstruction(&CxtI);
7394 if (Value *V = foldICmpWithLowBitMaskedVal(Pred, Op0, Op1, Q, *this))
7395 return replaceInstUsesWith(CxtI, V);
7397 // Folding (X / Y) pred X => X swap(pred) 0 for constant Y other than 0 or 1
7398 auto CheckUGT1 = [](const APInt &Divisor) { return Divisor.ugt(1); };
7400 if (match(Op0, m_UDiv(m_Specific(Op1), m_CheckedInt(CheckUGT1)))) {
7401 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), Op1,
7402 Constant::getNullValue(Op1->getType()));
7405 if (!ICmpInst::isUnsigned(Pred) &&
7406 match(Op0, m_SDiv(m_Specific(Op1), m_CheckedInt(CheckUGT1)))) {
7407 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), Op1,
7408 Constant::getNullValue(Op1->getType()));
7412 // Another case of this fold is (X >> Y) pred X => X swap(pred) 0 if Y != 0
7413 auto CheckNE0 = [](const APInt &Shift) { return !Shift.isZero(); };
7415 if (match(Op0, m_LShr(m_Specific(Op1), m_CheckedInt(CheckNE0)))) {
7416 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), Op1,
7417 Constant::getNullValue(Op1->getType()));
7420 if ((Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SGE) &&
7421 match(Op0, m_AShr(m_Specific(Op1), m_CheckedInt(CheckNE0)))) {
7422 return new ICmpInst(ICmpInst::getSwappedPredicate(Pred), Op1,
7423 Constant::getNullValue(Op1->getType()));
7427 return nullptr;
7430 Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
7431 bool Changed = false;
7432 const SimplifyQuery Q = SQ.getWithInstruction(&I);
7433 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7434 unsigned Op0Cplxity = getComplexity(Op0);
7435 unsigned Op1Cplxity = getComplexity(Op1);
7437 /// Orders the operands of the compare so that they are listed from most
7438 /// complex to least complex. This puts constants before unary operators,
7439 /// before binary operators.
7440 if (Op0Cplxity < Op1Cplxity) {
7441 I.swapOperands();
7442 std::swap(Op0, Op1);
7443 Changed = true;
7446 if (Value *V = simplifyICmpInst(I.getCmpPredicate(), Op0, Op1, Q))
7447 return replaceInstUsesWith(I, V);
7449 // Comparing -val or val with non-zero is the same as just comparing val
7450 // ie, abs(val) != 0 -> val != 0
7451 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
7452 Value *Cond, *SelectTrue, *SelectFalse;
7453 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
7454 m_Value(SelectFalse)))) {
7455 if (Value *V = dyn_castNegVal(SelectTrue)) {
7456 if (V == SelectFalse)
7457 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
7459 else if (Value *V = dyn_castNegVal(SelectFalse)) {
7460 if (V == SelectTrue)
7461 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
7466 if (Op0->getType()->isIntOrIntVectorTy(1))
7467 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
7468 return Res;
7470 if (Instruction *Res = canonicalizeCmpWithConstant(I))
7471 return Res;
7473 if (Instruction *Res = canonicalizeICmpPredicate(I))
7474 return Res;
7476 if (Instruction *Res = foldICmpWithConstant(I))
7477 return Res;
7479 if (Instruction *Res = foldICmpWithDominatingICmp(I))
7480 return Res;
7482 if (Instruction *Res = foldICmpUsingBoolRange(I))
7483 return Res;
7485 if (Instruction *Res = foldICmpUsingKnownBits(I))
7486 return Res;
7488 if (Instruction *Res = foldICmpTruncWithTruncOrExt(I, Q))
7489 return Res;
7491 // Test if the ICmpInst instruction is used exclusively by a select as
7492 // part of a minimum or maximum operation. If so, refrain from doing
7493 // any other folding. This helps out other analyses which understand
7494 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
7495 // and CodeGen. And in this case, at least one of the comparison
7496 // operands has at least one user besides the compare (the select),
7497 // which would often largely negate the benefit of folding anyway.
7499 // Do the same for the other patterns recognized by matchSelectPattern.
7500 if (I.hasOneUse())
7501 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
7502 Value *A, *B;
7503 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
7504 if (SPR.Flavor != SPF_UNKNOWN)
7505 return nullptr;
7508 // Do this after checking for min/max to prevent infinite looping.
7509 if (Instruction *Res = foldICmpWithZero(I))
7510 return Res;
7512 // FIXME: We only do this after checking for min/max to prevent infinite
7513 // looping caused by a reverse canonicalization of these patterns for min/max.
7514 // FIXME: The organization of folds is a mess. These would naturally go into
7515 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
7516 // down here after the min/max restriction.
7517 ICmpInst::Predicate Pred = I.getPredicate();
7518 const APInt *C;
7519 if (match(Op1, m_APInt(C))) {
7520 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
7521 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
7522 Constant *Zero = Constant::getNullValue(Op0->getType());
7523 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
7526 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
7527 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
7528 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
7529 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
7533 // The folds in here may rely on wrapping flags and special constants, so
7534 // they can break up min/max idioms in some cases but not seemingly similar
7535 // patterns.
7536 // FIXME: It may be possible to enhance select folding to make this
7537 // unnecessary. It may also be moot if we canonicalize to min/max
7538 // intrinsics.
7539 if (Instruction *Res = foldICmpBinOp(I, Q))
7540 return Res;
7542 if (Instruction *Res = foldICmpInstWithConstant(I))
7543 return Res;
7545 // Try to match comparison as a sign bit test. Intentionally do this after
7546 // foldICmpInstWithConstant() to potentially let other folds to happen first.
7547 if (Instruction *New = foldSignBitTest(I))
7548 return New;
7550 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
7551 return Res;
7553 if (Instruction *Res = foldICmpCommutative(I.getCmpPredicate(), Op0, Op1, I))
7554 return Res;
7555 if (Instruction *Res =
7556 foldICmpCommutative(I.getSwappedCmpPredicate(), Op1, Op0, I))
7557 return Res;
7559 if (I.isCommutative()) {
7560 if (auto Pair = matchSymmetricPair(I.getOperand(0), I.getOperand(1))) {
7561 replaceOperand(I, 0, Pair->first);
7562 replaceOperand(I, 1, Pair->second);
7563 return &I;
7567 // In case of a comparison with two select instructions having the same
7568 // condition, check whether one of the resulting branches can be simplified.
7569 // If so, just compare the other branch and select the appropriate result.
7570 // For example:
7571 // %tmp1 = select i1 %cmp, i32 %y, i32 %x
7572 // %tmp2 = select i1 %cmp, i32 %z, i32 %x
7573 // %cmp2 = icmp slt i32 %tmp2, %tmp1
7574 // The icmp will result false for the false value of selects and the result
7575 // will depend upon the comparison of true values of selects if %cmp is
7576 // true. Thus, transform this into:
7577 // %cmp = icmp slt i32 %y, %z
7578 // %sel = select i1 %cond, i1 %cmp, i1 false
7579 // This handles similar cases to transform.
7581 Value *Cond, *A, *B, *C, *D;
7582 if (match(Op0, m_Select(m_Value(Cond), m_Value(A), m_Value(B))) &&
7583 match(Op1, m_Select(m_Specific(Cond), m_Value(C), m_Value(D))) &&
7584 (Op0->hasOneUse() || Op1->hasOneUse())) {
7585 // Check whether comparison of TrueValues can be simplified
7586 if (Value *Res = simplifyICmpInst(Pred, A, C, SQ)) {
7587 Value *NewICMP = Builder.CreateICmp(Pred, B, D);
7588 return SelectInst::Create(Cond, Res, NewICMP);
7590 // Check whether comparison of FalseValues can be simplified
7591 if (Value *Res = simplifyICmpInst(Pred, B, D, SQ)) {
7592 Value *NewICMP = Builder.CreateICmp(Pred, A, C);
7593 return SelectInst::Create(Cond, NewICMP, Res);
7598 // Try to optimize equality comparisons against alloca-based pointers.
7599 if (Op0->getType()->isPointerTy() && I.isEquality()) {
7600 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
7601 if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(Op0)))
7602 if (foldAllocaCmp(Alloca))
7603 return nullptr;
7604 if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(Op1)))
7605 if (foldAllocaCmp(Alloca))
7606 return nullptr;
7609 if (Instruction *Res = foldICmpBitCast(I))
7610 return Res;
7612 // TODO: Hoist this above the min/max bailout.
7613 if (Instruction *R = foldICmpWithCastOp(I))
7614 return R;
7617 Value *X, *Y;
7618 // Transform (X & ~Y) == 0 --> (X & Y) != 0
7619 // and (X & ~Y) != 0 --> (X & Y) == 0
7620 // if A is a power of 2.
7621 if (match(Op0, m_And(m_Value(X), m_Not(m_Value(Y)))) &&
7622 match(Op1, m_Zero()) && isKnownToBeAPowerOfTwo(X, false, 0, &I) &&
7623 I.isEquality())
7624 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(X, Y),
7625 Op1);
7627 // Op0 pred Op1 -> ~Op1 pred ~Op0, if this allows us to drop an instruction.
7628 if (Op0->getType()->isIntOrIntVectorTy()) {
7629 bool ConsumesOp0, ConsumesOp1;
7630 if (isFreeToInvert(Op0, Op0->hasOneUse(), ConsumesOp0) &&
7631 isFreeToInvert(Op1, Op1->hasOneUse(), ConsumesOp1) &&
7632 (ConsumesOp0 || ConsumesOp1)) {
7633 Value *InvOp0 = getFreelyInverted(Op0, Op0->hasOneUse(), &Builder);
7634 Value *InvOp1 = getFreelyInverted(Op1, Op1->hasOneUse(), &Builder);
7635 assert(InvOp0 && InvOp1 &&
7636 "Mismatch between isFreeToInvert and getFreelyInverted");
7637 return new ICmpInst(I.getSwappedPredicate(), InvOp0, InvOp1);
7641 Instruction *AddI = nullptr;
7642 if (match(&I, m_UAddWithOverflow(m_Value(X), m_Value(Y),
7643 m_Instruction(AddI))) &&
7644 isa<IntegerType>(X->getType())) {
7645 Value *Result;
7646 Constant *Overflow;
7647 // m_UAddWithOverflow can match patterns that do not include an explicit
7648 // "add" instruction, so check the opcode of the matched op.
7649 if (AddI->getOpcode() == Instruction::Add &&
7650 OptimizeOverflowCheck(Instruction::Add, /*Signed*/ false, X, Y, *AddI,
7651 Result, Overflow)) {
7652 replaceInstUsesWith(*AddI, Result);
7653 eraseInstFromFunction(*AddI);
7654 return replaceInstUsesWith(I, Overflow);
7658 // (zext X) * (zext Y) --> llvm.umul.with.overflow.
7659 if (match(Op0, m_NUWMul(m_ZExt(m_Value(X)), m_ZExt(m_Value(Y)))) &&
7660 match(Op1, m_APInt(C))) {
7661 if (Instruction *R = processUMulZExtIdiom(I, Op0, C, *this))
7662 return R;
7665 // Signbit test folds
7666 // Fold (X u>> BitWidth - 1 Pred ZExt(i1)) --> X s< 0 Pred i1
7667 // Fold (X s>> BitWidth - 1 Pred SExt(i1)) --> X s< 0 Pred i1
7668 Instruction *ExtI;
7669 if ((I.isUnsigned() || I.isEquality()) &&
7670 match(Op1,
7671 m_CombineAnd(m_Instruction(ExtI), m_ZExtOrSExt(m_Value(Y)))) &&
7672 Y->getType()->getScalarSizeInBits() == 1 &&
7673 (Op0->hasOneUse() || Op1->hasOneUse())) {
7674 unsigned OpWidth = Op0->getType()->getScalarSizeInBits();
7675 Instruction *ShiftI;
7676 if (match(Op0, m_CombineAnd(m_Instruction(ShiftI),
7677 m_Shr(m_Value(X), m_SpecificIntAllowPoison(
7678 OpWidth - 1))))) {
7679 unsigned ExtOpc = ExtI->getOpcode();
7680 unsigned ShiftOpc = ShiftI->getOpcode();
7681 if ((ExtOpc == Instruction::ZExt && ShiftOpc == Instruction::LShr) ||
7682 (ExtOpc == Instruction::SExt && ShiftOpc == Instruction::AShr)) {
7683 Value *SLTZero =
7684 Builder.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
7685 Value *Cmp = Builder.CreateICmp(Pred, SLTZero, Y, I.getName());
7686 return replaceInstUsesWith(I, Cmp);
7692 if (Instruction *Res = foldICmpEquality(I))
7693 return Res;
7695 if (Instruction *Res = foldICmpPow2Test(I, Builder))
7696 return Res;
7698 if (Instruction *Res = foldICmpOfUAddOv(I))
7699 return Res;
7701 // The 'cmpxchg' instruction returns an aggregate containing the old value and
7702 // an i1 which indicates whether or not we successfully did the swap.
7704 // Replace comparisons between the old value and the expected value with the
7705 // indicator that 'cmpxchg' returns.
7707 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
7708 // spuriously fail. In those cases, the old value may equal the expected
7709 // value but it is possible for the swap to not occur.
7710 if (I.getPredicate() == ICmpInst::ICMP_EQ)
7711 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
7712 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
7713 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
7714 !ACXI->isWeak())
7715 return ExtractValueInst::Create(ACXI, 1);
7717 if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
7718 return Res;
7720 if (I.getType()->isVectorTy())
7721 if (Instruction *Res = foldVectorCmp(I, Builder))
7722 return Res;
7724 if (Instruction *Res = foldICmpInvariantGroup(I))
7725 return Res;
7727 if (Instruction *Res = foldReductionIdiom(I, Builder, DL))
7728 return Res;
7731 Value *A;
7732 const APInt *C1, *C2;
7733 ICmpInst::Predicate Pred = I.getPredicate();
7734 if (ICmpInst::isEquality(Pred)) {
7735 // sext(a) & c1 == c2 --> a & c3 == trunc(c2)
7736 // sext(a) & c1 != c2 --> a & c3 != trunc(c2)
7737 if (match(Op0, m_And(m_SExt(m_Value(A)), m_APInt(C1))) &&
7738 match(Op1, m_APInt(C2))) {
7739 Type *InputTy = A->getType();
7740 unsigned InputBitWidth = InputTy->getScalarSizeInBits();
7741 // c2 must be non-negative at the bitwidth of a.
7742 if (C2->getActiveBits() < InputBitWidth) {
7743 APInt TruncC1 = C1->trunc(InputBitWidth);
7744 // Check if there are 1s in C1 high bits of size InputBitWidth.
7745 if (C1->uge(APInt::getOneBitSet(C1->getBitWidth(), InputBitWidth)))
7746 TruncC1.setBit(InputBitWidth - 1);
7747 Value *AndInst = Builder.CreateAnd(A, TruncC1);
7748 return new ICmpInst(
7749 Pred, AndInst,
7750 ConstantInt::get(InputTy, C2->trunc(InputBitWidth)));
7756 return Changed ? &I : nullptr;
7759 /// Fold fcmp ([us]itofp x, cst) if possible.
7760 Instruction *InstCombinerImpl::foldFCmpIntToFPConst(FCmpInst &I,
7761 Instruction *LHSI,
7762 Constant *RHSC) {
7763 const APFloat *RHS;
7764 if (!match(RHSC, m_APFloat(RHS)))
7765 return nullptr;
7767 // Get the width of the mantissa. We don't want to hack on conversions that
7768 // might lose information from the integer, e.g. "i64 -> float"
7769 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
7770 if (MantissaWidth == -1) return nullptr; // Unknown.
7772 Type *IntTy = LHSI->getOperand(0)->getType();
7773 unsigned IntWidth = IntTy->getScalarSizeInBits();
7774 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
7776 if (I.isEquality()) {
7777 FCmpInst::Predicate P = I.getPredicate();
7778 bool IsExact = false;
7779 APSInt RHSCvt(IntWidth, LHSUnsigned);
7780 RHS->convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
7782 // If the floating point constant isn't an integer value, we know if we will
7783 // ever compare equal / not equal to it.
7784 if (!IsExact) {
7785 // TODO: Can never be -0.0 and other non-representable values
7786 APFloat RHSRoundInt(*RHS);
7787 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
7788 if (*RHS != RHSRoundInt) {
7789 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
7790 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7792 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
7793 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7797 // TODO: If the constant is exactly representable, is it always OK to do
7798 // equality compares as integer?
7801 // Check to see that the input is converted from an integer type that is small
7802 // enough that preserves all bits. TODO: check here for "known" sign bits.
7803 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
7805 // Following test does NOT adjust IntWidth downwards for signed inputs,
7806 // because the most negative value still requires all the mantissa bits
7807 // to distinguish it from one less than that value.
7808 if ((int)IntWidth > MantissaWidth) {
7809 // Conversion would lose accuracy. Check if loss can impact comparison.
7810 int Exp = ilogb(*RHS);
7811 if (Exp == APFloat::IEK_Inf) {
7812 int MaxExponent = ilogb(APFloat::getLargest(RHS->getSemantics()));
7813 if (MaxExponent < (int)IntWidth - !LHSUnsigned)
7814 // Conversion could create infinity.
7815 return nullptr;
7816 } else {
7817 // Note that if RHS is zero or NaN, then Exp is negative
7818 // and first condition is trivially false.
7819 if (MantissaWidth <= Exp && Exp <= (int)IntWidth - !LHSUnsigned)
7820 // Conversion could affect comparison.
7821 return nullptr;
7825 // Otherwise, we can potentially simplify the comparison. We know that it
7826 // will always come through as an integer value and we know the constant is
7827 // not a NAN (it would have been previously simplified).
7828 assert(!RHS->isNaN() && "NaN comparison not already folded!");
7830 ICmpInst::Predicate Pred;
7831 switch (I.getPredicate()) {
7832 default: llvm_unreachable("Unexpected predicate!");
7833 case FCmpInst::FCMP_UEQ:
7834 case FCmpInst::FCMP_OEQ:
7835 Pred = ICmpInst::ICMP_EQ;
7836 break;
7837 case FCmpInst::FCMP_UGT:
7838 case FCmpInst::FCMP_OGT:
7839 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
7840 break;
7841 case FCmpInst::FCMP_UGE:
7842 case FCmpInst::FCMP_OGE:
7843 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
7844 break;
7845 case FCmpInst::FCMP_ULT:
7846 case FCmpInst::FCMP_OLT:
7847 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
7848 break;
7849 case FCmpInst::FCMP_ULE:
7850 case FCmpInst::FCMP_OLE:
7851 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
7852 break;
7853 case FCmpInst::FCMP_UNE:
7854 case FCmpInst::FCMP_ONE:
7855 Pred = ICmpInst::ICMP_NE;
7856 break;
7857 case FCmpInst::FCMP_ORD:
7858 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7859 case FCmpInst::FCMP_UNO:
7860 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7863 // Now we know that the APFloat is a normal number, zero or inf.
7865 // See if the FP constant is too large for the integer. For example,
7866 // comparing an i8 to 300.0.
7867 if (!LHSUnsigned) {
7868 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
7869 // and large values.
7870 APFloat SMax(RHS->getSemantics());
7871 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
7872 APFloat::rmNearestTiesToEven);
7873 if (SMax < *RHS) { // smax < 13123.0
7874 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
7875 Pred == ICmpInst::ICMP_SLE)
7876 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7877 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7879 } else {
7880 // If the RHS value is > UnsignedMax, fold the comparison. This handles
7881 // +INF and large values.
7882 APFloat UMax(RHS->getSemantics());
7883 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
7884 APFloat::rmNearestTiesToEven);
7885 if (UMax < *RHS) { // umax < 13123.0
7886 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
7887 Pred == ICmpInst::ICMP_ULE)
7888 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7889 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7893 if (!LHSUnsigned) {
7894 // See if the RHS value is < SignedMin.
7895 APFloat SMin(RHS->getSemantics());
7896 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
7897 APFloat::rmNearestTiesToEven);
7898 if (SMin > *RHS) { // smin > 12312.0
7899 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
7900 Pred == ICmpInst::ICMP_SGE)
7901 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7902 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7904 } else {
7905 // See if the RHS value is < UnsignedMin.
7906 APFloat UMin(RHS->getSemantics());
7907 UMin.convertFromAPInt(APInt::getMinValue(IntWidth), false,
7908 APFloat::rmNearestTiesToEven);
7909 if (UMin > *RHS) { // umin > 12312.0
7910 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
7911 Pred == ICmpInst::ICMP_UGE)
7912 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7913 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7917 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
7918 // [0, UMAX], but it may still be fractional. Check whether this is the case
7919 // using the IsExact flag.
7920 // Don't do this for zero, because -0.0 is not fractional.
7921 APSInt RHSInt(IntWidth, LHSUnsigned);
7922 bool IsExact;
7923 RHS->convertToInteger(RHSInt, APFloat::rmTowardZero, &IsExact);
7924 if (!RHS->isZero()) {
7925 if (!IsExact) {
7926 // If we had a comparison against a fractional value, we have to adjust
7927 // the compare predicate and sometimes the value. RHSC is rounded towards
7928 // zero at this point.
7929 switch (Pred) {
7930 default: llvm_unreachable("Unexpected integer comparison!");
7931 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
7932 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7933 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
7934 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7935 case ICmpInst::ICMP_ULE:
7936 // (float)int <= 4.4 --> int <= 4
7937 // (float)int <= -4.4 --> false
7938 if (RHS->isNegative())
7939 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7940 break;
7941 case ICmpInst::ICMP_SLE:
7942 // (float)int <= 4.4 --> int <= 4
7943 // (float)int <= -4.4 --> int < -4
7944 if (RHS->isNegative())
7945 Pred = ICmpInst::ICMP_SLT;
7946 break;
7947 case ICmpInst::ICMP_ULT:
7948 // (float)int < -4.4 --> false
7949 // (float)int < 4.4 --> int <= 4
7950 if (RHS->isNegative())
7951 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
7952 Pred = ICmpInst::ICMP_ULE;
7953 break;
7954 case ICmpInst::ICMP_SLT:
7955 // (float)int < -4.4 --> int < -4
7956 // (float)int < 4.4 --> int <= 4
7957 if (!RHS->isNegative())
7958 Pred = ICmpInst::ICMP_SLE;
7959 break;
7960 case ICmpInst::ICMP_UGT:
7961 // (float)int > 4.4 --> int > 4
7962 // (float)int > -4.4 --> true
7963 if (RHS->isNegative())
7964 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7965 break;
7966 case ICmpInst::ICMP_SGT:
7967 // (float)int > 4.4 --> int > 4
7968 // (float)int > -4.4 --> int >= -4
7969 if (RHS->isNegative())
7970 Pred = ICmpInst::ICMP_SGE;
7971 break;
7972 case ICmpInst::ICMP_UGE:
7973 // (float)int >= -4.4 --> true
7974 // (float)int >= 4.4 --> int > 4
7975 if (RHS->isNegative())
7976 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
7977 Pred = ICmpInst::ICMP_UGT;
7978 break;
7979 case ICmpInst::ICMP_SGE:
7980 // (float)int >= -4.4 --> int >= -4
7981 // (float)int >= 4.4 --> int > 4
7982 if (!RHS->isNegative())
7983 Pred = ICmpInst::ICMP_SGT;
7984 break;
7989 // Lower this FP comparison into an appropriate integer version of the
7990 // comparison.
7991 return new ICmpInst(Pred, LHSI->getOperand(0),
7992 ConstantInt::get(LHSI->getOperand(0)->getType(), RHSInt));
7995 /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
7996 static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
7997 Constant *RHSC) {
7998 // When C is not 0.0 and infinities are not allowed:
7999 // (C / X) < 0.0 is a sign-bit test of X
8000 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
8001 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
8003 // Proof:
8004 // Multiply (C / X) < 0.0 by X * X / C.
8005 // - X is non zero, if it is the flag 'ninf' is violated.
8006 // - C defines the sign of X * X * C. Thus it also defines whether to swap
8007 // the predicate. C is also non zero by definition.
8009 // Thus X * X / C is non zero and the transformation is valid. [qed]
8011 FCmpInst::Predicate Pred = I.getPredicate();
8013 // Check that predicates are valid.
8014 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
8015 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
8016 return nullptr;
8018 // Check that RHS operand is zero.
8019 if (!match(RHSC, m_AnyZeroFP()))
8020 return nullptr;
8022 // Check fastmath flags ('ninf').
8023 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
8024 return nullptr;
8026 // Check the properties of the dividend. It must not be zero to avoid a
8027 // division by zero (see Proof).
8028 const APFloat *C;
8029 if (!match(LHSI->getOperand(0), m_APFloat(C)))
8030 return nullptr;
8032 if (C->isZero())
8033 return nullptr;
8035 // Get swapped predicate if necessary.
8036 if (C->isNegative())
8037 Pred = I.getSwappedPredicate();
8039 return new FCmpInst(Pred, LHSI->getOperand(1), RHSC, "", &I);
8042 /// Optimize fabs(X) compared with zero.
8043 static Instruction *foldFabsWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
8044 Value *X;
8045 if (!match(I.getOperand(0), m_FAbs(m_Value(X))))
8046 return nullptr;
8048 const APFloat *C;
8049 if (!match(I.getOperand(1), m_APFloat(C)))
8050 return nullptr;
8052 if (!C->isPosZero()) {
8053 if (!C->isSmallestNormalized())
8054 return nullptr;
8056 const Function *F = I.getFunction();
8057 DenormalMode Mode = F->getDenormalMode(C->getSemantics());
8058 if (Mode.Input == DenormalMode::PreserveSign ||
8059 Mode.Input == DenormalMode::PositiveZero) {
8061 auto replaceFCmp = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
8062 Constant *Zero = ConstantFP::getZero(X->getType());
8063 return new FCmpInst(P, X, Zero, "", I);
8066 switch (I.getPredicate()) {
8067 case FCmpInst::FCMP_OLT:
8068 // fcmp olt fabs(x), smallest_normalized_number -> fcmp oeq x, 0.0
8069 return replaceFCmp(&I, FCmpInst::FCMP_OEQ, X);
8070 case FCmpInst::FCMP_UGE:
8071 // fcmp uge fabs(x), smallest_normalized_number -> fcmp une x, 0.0
8072 return replaceFCmp(&I, FCmpInst::FCMP_UNE, X);
8073 case FCmpInst::FCMP_OGE:
8074 // fcmp oge fabs(x), smallest_normalized_number -> fcmp one x, 0.0
8075 return replaceFCmp(&I, FCmpInst::FCMP_ONE, X);
8076 case FCmpInst::FCMP_ULT:
8077 // fcmp ult fabs(x), smallest_normalized_number -> fcmp ueq x, 0.0
8078 return replaceFCmp(&I, FCmpInst::FCMP_UEQ, X);
8079 default:
8080 break;
8084 return nullptr;
8087 auto replacePredAndOp0 = [&IC](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
8088 I->setPredicate(P);
8089 return IC.replaceOperand(*I, 0, X);
8092 switch (I.getPredicate()) {
8093 case FCmpInst::FCMP_UGE:
8094 case FCmpInst::FCMP_OLT:
8095 // fabs(X) >= 0.0 --> true
8096 // fabs(X) < 0.0 --> false
8097 llvm_unreachable("fcmp should have simplified");
8099 case FCmpInst::FCMP_OGT:
8100 // fabs(X) > 0.0 --> X != 0.0
8101 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
8103 case FCmpInst::FCMP_UGT:
8104 // fabs(X) u> 0.0 --> X u!= 0.0
8105 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
8107 case FCmpInst::FCMP_OLE:
8108 // fabs(X) <= 0.0 --> X == 0.0
8109 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
8111 case FCmpInst::FCMP_ULE:
8112 // fabs(X) u<= 0.0 --> X u== 0.0
8113 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
8115 case FCmpInst::FCMP_OGE:
8116 // fabs(X) >= 0.0 --> !isnan(X)
8117 assert(!I.hasNoNaNs() && "fcmp should have simplified");
8118 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
8120 case FCmpInst::FCMP_ULT:
8121 // fabs(X) u< 0.0 --> isnan(X)
8122 assert(!I.hasNoNaNs() && "fcmp should have simplified");
8123 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
8125 case FCmpInst::FCMP_OEQ:
8126 case FCmpInst::FCMP_UEQ:
8127 case FCmpInst::FCMP_ONE:
8128 case FCmpInst::FCMP_UNE:
8129 case FCmpInst::FCMP_ORD:
8130 case FCmpInst::FCMP_UNO:
8131 // Look through the fabs() because it doesn't change anything but the sign.
8132 // fabs(X) == 0.0 --> X == 0.0,
8133 // fabs(X) != 0.0 --> X != 0.0
8134 // isnan(fabs(X)) --> isnan(X)
8135 // !isnan(fabs(X) --> !isnan(X)
8136 return replacePredAndOp0(&I, I.getPredicate(), X);
8138 default:
8139 return nullptr;
8143 /// Optimize sqrt(X) compared with zero.
8144 static Instruction *foldSqrtWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
8145 Value *X;
8146 if (!match(I.getOperand(0), m_Sqrt(m_Value(X))))
8147 return nullptr;
8149 if (!match(I.getOperand(1), m_PosZeroFP()))
8150 return nullptr;
8152 auto ReplacePredAndOp0 = [&](FCmpInst::Predicate P) {
8153 I.setPredicate(P);
8154 return IC.replaceOperand(I, 0, X);
8157 // Clear ninf flag if sqrt doesn't have it.
8158 if (!cast<Instruction>(I.getOperand(0))->hasNoInfs())
8159 I.setHasNoInfs(false);
8161 switch (I.getPredicate()) {
8162 case FCmpInst::FCMP_OLT:
8163 case FCmpInst::FCMP_UGE:
8164 // sqrt(X) < 0.0 --> false
8165 // sqrt(X) u>= 0.0 --> true
8166 llvm_unreachable("fcmp should have simplified");
8167 case FCmpInst::FCMP_ULT:
8168 case FCmpInst::FCMP_ULE:
8169 case FCmpInst::FCMP_OGT:
8170 case FCmpInst::FCMP_OGE:
8171 case FCmpInst::FCMP_OEQ:
8172 case FCmpInst::FCMP_UNE:
8173 // sqrt(X) u< 0.0 --> X u< 0.0
8174 // sqrt(X) u<= 0.0 --> X u<= 0.0
8175 // sqrt(X) > 0.0 --> X > 0.0
8176 // sqrt(X) >= 0.0 --> X >= 0.0
8177 // sqrt(X) == 0.0 --> X == 0.0
8178 // sqrt(X) u!= 0.0 --> X u!= 0.0
8179 return IC.replaceOperand(I, 0, X);
8181 case FCmpInst::FCMP_OLE:
8182 // sqrt(X) <= 0.0 --> X == 0.0
8183 return ReplacePredAndOp0(FCmpInst::FCMP_OEQ);
8184 case FCmpInst::FCMP_UGT:
8185 // sqrt(X) u> 0.0 --> X u!= 0.0
8186 return ReplacePredAndOp0(FCmpInst::FCMP_UNE);
8187 case FCmpInst::FCMP_UEQ:
8188 // sqrt(X) u== 0.0 --> X u<= 0.0
8189 return ReplacePredAndOp0(FCmpInst::FCMP_ULE);
8190 case FCmpInst::FCMP_ONE:
8191 // sqrt(X) != 0.0 --> X > 0.0
8192 return ReplacePredAndOp0(FCmpInst::FCMP_OGT);
8193 case FCmpInst::FCMP_ORD:
8194 // !isnan(sqrt(X)) --> X >= 0.0
8195 return ReplacePredAndOp0(FCmpInst::FCMP_OGE);
8196 case FCmpInst::FCMP_UNO:
8197 // isnan(sqrt(X)) --> X u< 0.0
8198 return ReplacePredAndOp0(FCmpInst::FCMP_ULT);
8199 default:
8200 llvm_unreachable("Unexpected predicate!");
8204 static Instruction *foldFCmpFNegCommonOp(FCmpInst &I) {
8205 CmpInst::Predicate Pred = I.getPredicate();
8206 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
8208 // Canonicalize fneg as Op1.
8209 if (match(Op0, m_FNeg(m_Value())) && !match(Op1, m_FNeg(m_Value()))) {
8210 std::swap(Op0, Op1);
8211 Pred = I.getSwappedPredicate();
8214 if (!match(Op1, m_FNeg(m_Specific(Op0))))
8215 return nullptr;
8217 // Replace the negated operand with 0.0:
8218 // fcmp Pred Op0, -Op0 --> fcmp Pred Op0, 0.0
8219 Constant *Zero = ConstantFP::getZero(Op0->getType());
8220 return new FCmpInst(Pred, Op0, Zero, "", &I);
8223 static Instruction *foldFCmpFSubIntoFCmp(FCmpInst &I, Instruction *LHSI,
8224 Constant *RHSC, InstCombinerImpl &CI) {
8225 const CmpInst::Predicate Pred = I.getPredicate();
8226 Value *X = LHSI->getOperand(0);
8227 Value *Y = LHSI->getOperand(1);
8228 switch (Pred) {
8229 default:
8230 break;
8231 case FCmpInst::FCMP_UGT:
8232 case FCmpInst::FCMP_ULT:
8233 case FCmpInst::FCMP_UNE:
8234 case FCmpInst::FCMP_OEQ:
8235 case FCmpInst::FCMP_OGE:
8236 case FCmpInst::FCMP_OLE:
8237 // The optimization is not valid if X and Y are infinities of the same
8238 // sign, i.e. the inf - inf = nan case. If the fsub has the ninf or nnan
8239 // flag then we can assume we do not have that case. Otherwise we might be
8240 // able to prove that either X or Y is not infinity.
8241 if (!LHSI->hasNoNaNs() && !LHSI->hasNoInfs() &&
8242 !isKnownNeverInfinity(Y, /*Depth=*/0,
8243 CI.getSimplifyQuery().getWithInstruction(&I)) &&
8244 !isKnownNeverInfinity(X, /*Depth=*/0,
8245 CI.getSimplifyQuery().getWithInstruction(&I)))
8246 break;
8248 [[fallthrough]];
8249 case FCmpInst::FCMP_OGT:
8250 case FCmpInst::FCMP_OLT:
8251 case FCmpInst::FCMP_ONE:
8252 case FCmpInst::FCMP_UEQ:
8253 case FCmpInst::FCMP_UGE:
8254 case FCmpInst::FCMP_ULE:
8255 // fcmp pred (x - y), 0 --> fcmp pred x, y
8256 if (match(RHSC, m_AnyZeroFP()) &&
8257 I.getFunction()->getDenormalMode(
8258 LHSI->getType()->getScalarType()->getFltSemantics()) ==
8259 DenormalMode::getIEEE()) {
8260 CI.replaceOperand(I, 0, X);
8261 CI.replaceOperand(I, 1, Y);
8262 return &I;
8264 break;
8267 return nullptr;
8270 static Instruction *foldFCmpWithFloorAndCeil(FCmpInst &I,
8271 InstCombinerImpl &IC) {
8272 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
8273 Type *OpType = LHS->getType();
8274 CmpInst::Predicate Pred = I.getPredicate();
8276 bool FloorX = match(LHS, m_Intrinsic<Intrinsic::floor>(m_Specific(RHS)));
8277 bool CeilX = match(LHS, m_Intrinsic<Intrinsic::ceil>(m_Specific(RHS)));
8279 if (!FloorX && !CeilX) {
8280 if ((FloorX = match(RHS, m_Intrinsic<Intrinsic::floor>(m_Specific(LHS)))) ||
8281 (CeilX = match(RHS, m_Intrinsic<Intrinsic::ceil>(m_Specific(LHS))))) {
8282 std::swap(LHS, RHS);
8283 Pred = I.getSwappedPredicate();
8287 switch (Pred) {
8288 case FCmpInst::FCMP_OLE:
8289 // fcmp ole floor(x), x => fcmp ord x, 0
8290 if (FloorX)
8291 return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(OpType),
8292 "", &I);
8293 break;
8294 case FCmpInst::FCMP_OGT:
8295 // fcmp ogt floor(x), x => false
8296 if (FloorX)
8297 return IC.replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
8298 break;
8299 case FCmpInst::FCMP_OGE:
8300 // fcmp oge ceil(x), x => fcmp ord x, 0
8301 if (CeilX)
8302 return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(OpType),
8303 "", &I);
8304 break;
8305 case FCmpInst::FCMP_OLT:
8306 // fcmp olt ceil(x), x => false
8307 if (CeilX)
8308 return IC.replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
8309 break;
8310 case FCmpInst::FCMP_ULE:
8311 // fcmp ule floor(x), x => true
8312 if (FloorX)
8313 return IC.replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
8314 break;
8315 case FCmpInst::FCMP_UGT:
8316 // fcmp ugt floor(x), x => fcmp uno x, 0
8317 if (FloorX)
8318 return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(OpType),
8319 "", &I);
8320 break;
8321 case FCmpInst::FCMP_UGE:
8322 // fcmp uge ceil(x), x => true
8323 if (CeilX)
8324 return IC.replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
8325 break;
8326 case FCmpInst::FCMP_ULT:
8327 // fcmp ult ceil(x), x => fcmp uno x, 0
8328 if (CeilX)
8329 return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(OpType),
8330 "", &I);
8331 break;
8332 default:
8333 break;
8336 return nullptr;
8339 Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
8340 bool Changed = false;
8342 /// Orders the operands of the compare so that they are listed from most
8343 /// complex to least complex. This puts constants before unary operators,
8344 /// before binary operators.
8345 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
8346 I.swapOperands();
8347 Changed = true;
8350 const CmpInst::Predicate Pred = I.getPredicate();
8351 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
8352 if (Value *V = simplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
8353 SQ.getWithInstruction(&I)))
8354 return replaceInstUsesWith(I, V);
8356 // Simplify 'fcmp pred X, X'
8357 Type *OpType = Op0->getType();
8358 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
8359 if (Op0 == Op1) {
8360 switch (Pred) {
8361 default: break;
8362 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
8363 case FCmpInst::FCMP_ULT: // True if unordered or less than
8364 case FCmpInst::FCMP_UGT: // True if unordered or greater than
8365 case FCmpInst::FCMP_UNE: // True if unordered or not equal
8366 // Canonicalize these to be 'fcmp uno %X, 0.0'.
8367 I.setPredicate(FCmpInst::FCMP_UNO);
8368 I.setOperand(1, Constant::getNullValue(OpType));
8369 return &I;
8371 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
8372 case FCmpInst::FCMP_OEQ: // True if ordered and equal
8373 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
8374 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
8375 // Canonicalize these to be 'fcmp ord %X, 0.0'.
8376 I.setPredicate(FCmpInst::FCMP_ORD);
8377 I.setOperand(1, Constant::getNullValue(OpType));
8378 return &I;
8382 if (I.isCommutative()) {
8383 if (auto Pair = matchSymmetricPair(I.getOperand(0), I.getOperand(1))) {
8384 replaceOperand(I, 0, Pair->first);
8385 replaceOperand(I, 1, Pair->second);
8386 return &I;
8390 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
8391 // then canonicalize the operand to 0.0.
8392 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
8393 if (!match(Op0, m_PosZeroFP()) &&
8394 isKnownNeverNaN(Op0, 0, getSimplifyQuery().getWithInstruction(&I)))
8395 return replaceOperand(I, 0, ConstantFP::getZero(OpType));
8397 if (!match(Op1, m_PosZeroFP()) &&
8398 isKnownNeverNaN(Op1, 0, getSimplifyQuery().getWithInstruction(&I)))
8399 return replaceOperand(I, 1, ConstantFP::getZero(OpType));
8402 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
8403 Value *X, *Y;
8404 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
8405 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
8407 if (Instruction *R = foldFCmpFNegCommonOp(I))
8408 return R;
8410 // Test if the FCmpInst instruction is used exclusively by a select as
8411 // part of a minimum or maximum operation. If so, refrain from doing
8412 // any other folding. This helps out other analyses which understand
8413 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
8414 // and CodeGen. And in this case, at least one of the comparison
8415 // operands has at least one user besides the compare (the select),
8416 // which would often largely negate the benefit of folding anyway.
8417 if (I.hasOneUse())
8418 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
8419 Value *A, *B;
8420 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
8421 if (SPR.Flavor != SPF_UNKNOWN)
8422 return nullptr;
8425 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
8426 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
8427 if (match(Op1, m_AnyZeroFP()) && !match(Op1, m_PosZeroFP()))
8428 return replaceOperand(I, 1, ConstantFP::getZero(OpType));
8430 // Canonicalize:
8431 // fcmp olt X, +inf -> fcmp one X, +inf
8432 // fcmp ole X, +inf -> fcmp ord X, 0
8433 // fcmp ogt X, +inf -> false
8434 // fcmp oge X, +inf -> fcmp oeq X, +inf
8435 // fcmp ult X, +inf -> fcmp une X, +inf
8436 // fcmp ule X, +inf -> true
8437 // fcmp ugt X, +inf -> fcmp uno X, 0
8438 // fcmp uge X, +inf -> fcmp ueq X, +inf
8439 // fcmp olt X, -inf -> false
8440 // fcmp ole X, -inf -> fcmp oeq X, -inf
8441 // fcmp ogt X, -inf -> fcmp one X, -inf
8442 // fcmp oge X, -inf -> fcmp ord X, 0
8443 // fcmp ult X, -inf -> fcmp uno X, 0
8444 // fcmp ule X, -inf -> fcmp ueq X, -inf
8445 // fcmp ugt X, -inf -> fcmp une X, -inf
8446 // fcmp uge X, -inf -> true
8447 const APFloat *C;
8448 if (match(Op1, m_APFloat(C)) && C->isInfinity()) {
8449 switch (C->isNegative() ? FCmpInst::getSwappedPredicate(Pred) : Pred) {
8450 default:
8451 break;
8452 case FCmpInst::FCMP_ORD:
8453 case FCmpInst::FCMP_UNO:
8454 case FCmpInst::FCMP_TRUE:
8455 case FCmpInst::FCMP_FALSE:
8456 case FCmpInst::FCMP_OGT:
8457 case FCmpInst::FCMP_ULE:
8458 llvm_unreachable("Should be simplified by InstSimplify");
8459 case FCmpInst::FCMP_OLT:
8460 return new FCmpInst(FCmpInst::FCMP_ONE, Op0, Op1, "", &I);
8461 case FCmpInst::FCMP_OLE:
8462 return new FCmpInst(FCmpInst::FCMP_ORD, Op0, ConstantFP::getZero(OpType),
8463 "", &I);
8464 case FCmpInst::FCMP_OGE:
8465 return new FCmpInst(FCmpInst::FCMP_OEQ, Op0, Op1, "", &I);
8466 case FCmpInst::FCMP_ULT:
8467 return new FCmpInst(FCmpInst::FCMP_UNE, Op0, Op1, "", &I);
8468 case FCmpInst::FCMP_UGT:
8469 return new FCmpInst(FCmpInst::FCMP_UNO, Op0, ConstantFP::getZero(OpType),
8470 "", &I);
8471 case FCmpInst::FCMP_UGE:
8472 return new FCmpInst(FCmpInst::FCMP_UEQ, Op0, Op1, "", &I);
8476 // Ignore signbit of bitcasted int when comparing equality to FP 0.0:
8477 // fcmp oeq/une (bitcast X), 0.0 --> (and X, SignMaskC) ==/!= 0
8478 if (match(Op1, m_PosZeroFP()) &&
8479 match(Op0, m_OneUse(m_ElementWiseBitCast(m_Value(X))))) {
8480 ICmpInst::Predicate IntPred = ICmpInst::BAD_ICMP_PREDICATE;
8481 if (Pred == FCmpInst::FCMP_OEQ)
8482 IntPred = ICmpInst::ICMP_EQ;
8483 else if (Pred == FCmpInst::FCMP_UNE)
8484 IntPred = ICmpInst::ICMP_NE;
8486 if (IntPred != ICmpInst::BAD_ICMP_PREDICATE) {
8487 Type *IntTy = X->getType();
8488 const APInt &SignMask = ~APInt::getSignMask(IntTy->getScalarSizeInBits());
8489 Value *MaskX = Builder.CreateAnd(X, ConstantInt::get(IntTy, SignMask));
8490 return new ICmpInst(IntPred, MaskX, ConstantInt::getNullValue(IntTy));
8494 // Handle fcmp with instruction LHS and constant RHS.
8495 Instruction *LHSI;
8496 Constant *RHSC;
8497 if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
8498 switch (LHSI->getOpcode()) {
8499 case Instruction::Select:
8500 // fcmp eq (cond ? x : -x), 0 --> fcmp eq x, 0
8501 if (FCmpInst::isEquality(Pred) && match(RHSC, m_AnyZeroFP()) &&
8502 match(LHSI, m_c_Select(m_FNeg(m_Value(X)), m_Deferred(X))))
8503 return replaceOperand(I, 0, X);
8504 if (Instruction *NV = FoldOpIntoSelect(I, cast<SelectInst>(LHSI)))
8505 return NV;
8506 break;
8507 case Instruction::FSub:
8508 if (LHSI->hasOneUse())
8509 if (Instruction *NV = foldFCmpFSubIntoFCmp(I, LHSI, RHSC, *this))
8510 return NV;
8511 break;
8512 case Instruction::PHI:
8513 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
8514 return NV;
8515 break;
8516 case Instruction::SIToFP:
8517 case Instruction::UIToFP:
8518 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
8519 return NV;
8520 break;
8521 case Instruction::FDiv:
8522 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
8523 return NV;
8524 break;
8525 case Instruction::Load:
8526 if (auto *GEP = dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
8527 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
8528 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(
8529 cast<LoadInst>(LHSI), GEP, GV, I))
8530 return Res;
8531 break;
8535 if (Instruction *R = foldFabsWithFcmpZero(I, *this))
8536 return R;
8538 if (Instruction *R = foldSqrtWithFcmpZero(I, *this))
8539 return R;
8541 if (Instruction *R = foldFCmpWithFloorAndCeil(I, *this))
8542 return R;
8544 if (match(Op0, m_FNeg(m_Value(X)))) {
8545 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
8546 Constant *C;
8547 if (match(Op1, m_Constant(C)))
8548 if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
8549 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
8552 // fcmp (fadd X, 0.0), Y --> fcmp X, Y
8553 if (match(Op0, m_FAdd(m_Value(X), m_AnyZeroFP())))
8554 return new FCmpInst(Pred, X, Op1, "", &I);
8556 // fcmp X, (fadd Y, 0.0) --> fcmp X, Y
8557 if (match(Op1, m_FAdd(m_Value(Y), m_AnyZeroFP())))
8558 return new FCmpInst(Pred, Op0, Y, "", &I);
8560 if (match(Op0, m_FPExt(m_Value(X)))) {
8561 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
8562 if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
8563 return new FCmpInst(Pred, X, Y, "", &I);
8565 const APFloat *C;
8566 if (match(Op1, m_APFloat(C))) {
8567 const fltSemantics &FPSem =
8568 X->getType()->getScalarType()->getFltSemantics();
8569 bool Lossy;
8570 APFloat TruncC = *C;
8571 TruncC.convert(FPSem, APFloat::rmNearestTiesToEven, &Lossy);
8573 if (Lossy) {
8574 // X can't possibly equal the higher-precision constant, so reduce any
8575 // equality comparison.
8576 // TODO: Other predicates can be handled via getFCmpCode().
8577 switch (Pred) {
8578 case FCmpInst::FCMP_OEQ:
8579 // X is ordered and equal to an impossible constant --> false
8580 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
8581 case FCmpInst::FCMP_ONE:
8582 // X is ordered and not equal to an impossible constant --> ordered
8583 return new FCmpInst(FCmpInst::FCMP_ORD, X,
8584 ConstantFP::getZero(X->getType()));
8585 case FCmpInst::FCMP_UEQ:
8586 // X is unordered or equal to an impossible constant --> unordered
8587 return new FCmpInst(FCmpInst::FCMP_UNO, X,
8588 ConstantFP::getZero(X->getType()));
8589 case FCmpInst::FCMP_UNE:
8590 // X is unordered or not equal to an impossible constant --> true
8591 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
8592 default:
8593 break;
8597 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
8598 // Avoid lossy conversions and denormals.
8599 // Zero is a special case that's OK to convert.
8600 APFloat Fabs = TruncC;
8601 Fabs.clearSign();
8602 if (!Lossy &&
8603 (Fabs.isZero() || !(Fabs < APFloat::getSmallestNormalized(FPSem)))) {
8604 Constant *NewC = ConstantFP::get(X->getType(), TruncC);
8605 return new FCmpInst(Pred, X, NewC, "", &I);
8610 // Convert a sign-bit test of an FP value into a cast and integer compare.
8611 // TODO: Simplify if the copysign constant is 0.0 or NaN.
8612 // TODO: Handle non-zero compare constants.
8613 // TODO: Handle other predicates.
8614 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::copysign>(m_APFloat(C),
8615 m_Value(X)))) &&
8616 match(Op1, m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) {
8617 Type *IntType = Builder.getIntNTy(X->getType()->getScalarSizeInBits());
8618 if (auto *VecTy = dyn_cast<VectorType>(OpType))
8619 IntType = VectorType::get(IntType, VecTy->getElementCount());
8621 // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0
8622 if (Pred == FCmpInst::FCMP_OLT) {
8623 Value *IntX = Builder.CreateBitCast(X, IntType);
8624 return new ICmpInst(ICmpInst::ICMP_SLT, IntX,
8625 ConstantInt::getNullValue(IntType));
8630 Value *CanonLHS = nullptr, *CanonRHS = nullptr;
8631 match(Op0, m_Intrinsic<Intrinsic::canonicalize>(m_Value(CanonLHS)));
8632 match(Op1, m_Intrinsic<Intrinsic::canonicalize>(m_Value(CanonRHS)));
8634 // (canonicalize(x) == x) => (x == x)
8635 if (CanonLHS == Op1)
8636 return new FCmpInst(Pred, Op1, Op1, "", &I);
8638 // (x == canonicalize(x)) => (x == x)
8639 if (CanonRHS == Op0)
8640 return new FCmpInst(Pred, Op0, Op0, "", &I);
8642 // (canonicalize(x) == canonicalize(y)) => (x == y)
8643 if (CanonLHS && CanonRHS)
8644 return new FCmpInst(Pred, CanonLHS, CanonRHS, "", &I);
8647 if (I.getType()->isVectorTy())
8648 if (Instruction *Res = foldVectorCmp(I, Builder))
8649 return Res;
8651 return Changed ? &I : nullptr;