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.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APFloat.h"
15#include "llvm/ADT/APSInt.h"
16#include "llvm/ADT/SetVector.h"
17#include "llvm/ADT/Statistic.h"
18#include "llvm/Analysis/CaptureTracking.h"
19#include "llvm/Analysis/CmpInstAnalysis.h"
20#include "llvm/Analysis/ConstantFolding.h"
21#include "llvm/Analysis/InstructionSimplify.h"
22#include "llvm/Analysis/Loads.h"
23#include "llvm/Analysis/Utils/Local.h"
24#include "llvm/Analysis/VectorUtils.h"
25#include "llvm/IR/ConstantRange.h"
26#include "llvm/IR/Constants.h"
27#include "llvm/IR/DataLayout.h"
28#include "llvm/IR/InstrTypes.h"
29#include "llvm/IR/Instructions.h"
30#include "llvm/IR/IntrinsicInst.h"
31#include "llvm/IR/PatternMatch.h"
32#include "llvm/Support/KnownBits.h"
33#include "llvm/Transforms/InstCombine/InstCombiner.h"
34#include <bitset>
35
36using namespace llvm;
37using namespace PatternMatch;
38
39#define DEBUG_TYPE "instcombine"
40
41// How many times is a select replaced by one of its operands?
42STATISTIC(NumSel, "Number of select opts");
43
44namespace llvm {
45extern cl::opt<bool> ProfcheckDisableMetadataFixes;
46}
47
48/// Compute Result = In1+In2, returning true if the result overflowed for this
49/// type.
50static bool addWithOverflow(APInt &Result, const APInt &In1, const APInt &In2,
51 bool IsSigned = false) {
52 bool Overflow;
53 if (IsSigned)
54 Result = In1.sadd_ov(RHS: In2, Overflow);
55 else
56 Result = In1.uadd_ov(RHS: In2, Overflow);
57
58 return Overflow;
59}
60
61/// Compute Result = In1-In2, returning true if the result overflowed for this
62/// type.
63static bool subWithOverflow(APInt &Result, const APInt &In1, const APInt &In2,
64 bool IsSigned = false) {
65 bool Overflow;
66 if (IsSigned)
67 Result = In1.ssub_ov(RHS: In2, Overflow);
68 else
69 Result = In1.usub_ov(RHS: In2, Overflow);
70
71 return Overflow;
72}
73
74/// Given an icmp instruction, return true if any use of this comparison is a
75/// branch on sign bit comparison.
76static bool hasBranchUse(ICmpInst &I) {
77 for (auto *U : I.users())
78 if (isa<CondBrInst>(Val: U))
79 return true;
80 return false;
81}
82
83/// Returns true if the exploded icmp can be expressed as a signed comparison
84/// to zero and updates the predicate accordingly.
85/// The signedness of the comparison is preserved.
86/// TODO: Refactor with decomposeBitTestICmp()?
87static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
88 if (!ICmpInst::isSigned(Pred))
89 return false;
90
91 if (C.isZero())
92 return ICmpInst::isRelational(P: Pred);
93
94 if (C.isOne()) {
95 if (Pred == ICmpInst::ICMP_SLT) {
96 Pred = ICmpInst::ICMP_SLE;
97 return true;
98 }
99 } else if (C.isAllOnes()) {
100 if (Pred == ICmpInst::ICMP_SGT) {
101 Pred = ICmpInst::ICMP_SGE;
102 return true;
103 }
104 }
105
106 return false;
107}
108
109/// This is called when we see this pattern:
110/// cmp pred (load (gep GV, ...)), cmpcst
111/// where GV is a global variable with a constant initializer. Try to simplify
112/// this into some simple computation that does not need the load. For example
113/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
114///
115/// If AndCst is non-null, then the loaded value is masked with that constant
116/// before doing the comparison. This handles cases like "A[i]&4 == 0".
117Instruction *InstCombinerImpl::foldCmpLoadFromIndexedGlobal(
118 LoadInst *LI, GetElementPtrInst *GEP, CmpInst &ICI, ConstantInt *AndCst) {
119 auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: GEP));
120 if (LI->isVolatile() || !GV || !GV->isConstant() ||
121 !GV->hasDefinitiveInitializer())
122 return nullptr;
123
124 Type *EltTy = LI->getType();
125 TypeSize EltSize = DL.getTypeStoreSize(Ty: EltTy);
126 if (EltSize.isScalable())
127 return nullptr;
128
129 LinearExpression Expr = decomposeLinearExpression(DL, Ptr: GEP);
130 if (!Expr.Index || Expr.BasePtr != GV || Expr.Offset.getBitWidth() > 64)
131 return nullptr;
132
133 Constant *Init = GV->getInitializer();
134 TypeSize GlobalSize = DL.getTypeAllocSize(Ty: Init->getType());
135
136 Value *Idx = Expr.Index;
137 const APInt &Stride = Expr.Scale;
138 const APInt &ConstOffset = Expr.Offset;
139
140 // Allow an additional context offset, but only within the stride.
141 if (!ConstOffset.ult(RHS: Stride))
142 return nullptr;
143
144 // Don't handle overlapping loads for now.
145 if (!Stride.uge(RHS: EltSize.getFixedValue()))
146 return nullptr;
147
148 // Don't blow up on huge arrays.
149 uint64_t ArrayElementCount =
150 divideCeil(Numerator: (GlobalSize.getFixedValue() - ConstOffset.getZExtValue()),
151 Denominator: Stride.getZExtValue());
152 if (ArrayElementCount > MaxArraySizeForCombine)
153 return nullptr;
154
155 enum { Overdefined = -3, Undefined = -2 };
156
157 // Variables for our state machines.
158
159 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
160 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
161 // and 87 is the second (and last) index. FirstTrueElement is -2 when
162 // undefined, otherwise set to the first true element. SecondTrueElement is
163 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
164 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
165
166 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
167 // form "i != 47 & i != 87". Same state transitions as for true elements.
168 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
169
170 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
171 /// define a state machine that triggers for ranges of values that the index
172 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
173 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
174 /// index in the range (inclusive). We use -2 for undefined here because we
175 /// use relative comparisons and don't want 0-1 to match -1.
176 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
177
178 // MagicBitvector - This is a magic bitvector where we set a bit if the
179 // comparison is true for element 'i'. If there are 64 elements or less in
180 // the array, this will fully represent all the comparison results.
181 uint64_t MagicBitvector = 0;
182
183 // Scan the array and see if one of our patterns matches.
184 Constant *CompareRHS = cast<Constant>(Val: ICI.getOperand(i_nocapture: 1));
185 APInt Offset = ConstOffset;
186 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i, Offset += Stride) {
187 Constant *Elt = ConstantFoldLoadFromConst(C: Init, Ty: EltTy, Offset, DL);
188 if (!Elt)
189 return nullptr;
190
191 // If the element is masked, handle it.
192 if (AndCst) {
193 Elt = ConstantFoldBinaryOpOperands(Opcode: Instruction::And, LHS: Elt, RHS: AndCst, DL);
194 if (!Elt)
195 return nullptr;
196 }
197
198 // Find out if the comparison would be true or false for the i'th element.
199 Constant *C = ConstantFoldCompareInstOperands(Predicate: ICI.getPredicate(), LHS: Elt,
200 RHS: CompareRHS, DL, TLI: &TLI);
201 if (!C)
202 return nullptr;
203
204 // If the result is undef for this element, ignore it.
205 if (isa<UndefValue>(Val: C)) {
206 // Extend range state machines to cover this element in case there is an
207 // undef in the middle of the range.
208 if (TrueRangeEnd == (int)i - 1)
209 TrueRangeEnd = i;
210 if (FalseRangeEnd == (int)i - 1)
211 FalseRangeEnd = i;
212 continue;
213 }
214
215 // If we can't compute the result for any of the elements, we have to give
216 // up evaluating the entire conditional.
217 if (!isa<ConstantInt>(Val: C))
218 return nullptr;
219
220 // Otherwise, we know if the comparison is true or false for this element,
221 // update our state machines.
222 bool IsTrueForElt = !cast<ConstantInt>(Val: C)->isZero();
223
224 // State machine for single/double/range index comparison.
225 if (IsTrueForElt) {
226 // Update the TrueElement state machine.
227 if (FirstTrueElement == Undefined)
228 FirstTrueElement = TrueRangeEnd = i; // First true element.
229 else {
230 // Update double-compare state machine.
231 if (SecondTrueElement == Undefined)
232 SecondTrueElement = i;
233 else
234 SecondTrueElement = Overdefined;
235
236 // Update range state machine.
237 if (TrueRangeEnd == (int)i - 1)
238 TrueRangeEnd = i;
239 else
240 TrueRangeEnd = Overdefined;
241 }
242 } else {
243 // Update the FalseElement state machine.
244 if (FirstFalseElement == Undefined)
245 FirstFalseElement = FalseRangeEnd = i; // First false element.
246 else {
247 // Update double-compare state machine.
248 if (SecondFalseElement == Undefined)
249 SecondFalseElement = i;
250 else
251 SecondFalseElement = Overdefined;
252
253 // Update range state machine.
254 if (FalseRangeEnd == (int)i - 1)
255 FalseRangeEnd = i;
256 else
257 FalseRangeEnd = Overdefined;
258 }
259 }
260
261 // If this element is in range, update our magic bitvector.
262 if (i < 64 && IsTrueForElt)
263 MagicBitvector |= 1ULL << i;
264
265 // If all of our states become overdefined, bail out early. Since the
266 // predicate is expensive, only check it every 8 elements. This is only
267 // really useful for really huge arrays.
268 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
269 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
270 FalseRangeEnd == Overdefined)
271 return nullptr;
272 }
273
274 // Now that we've scanned the entire array, emit our new comparison(s). We
275 // order the state machines in complexity of the generated code.
276
277 // If inbounds keyword is not present, Idx * Stride can overflow.
278 // Let's assume that Stride is 2 and the wanted value is at offset 0.
279 // Then, there are two possible values for Idx to match offset 0:
280 // 0x00..00, 0x80..00.
281 // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
282 // comparison is false if Idx was 0x80..00.
283 // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
284 auto MaskIdx = [&](Value *Idx) {
285 if (!Expr.Flags.isInBounds() && Stride.countr_zero() != 0) {
286 Value *Mask = Constant::getAllOnesValue(Ty: Idx->getType());
287 Mask = Builder.CreateLShr(LHS: Mask, RHS: Stride.countr_zero());
288 Idx = Builder.CreateAnd(LHS: Idx, RHS: Mask);
289 }
290 return Idx;
291 };
292
293 // If the comparison is only true for one or two elements, emit direct
294 // comparisons.
295 if (SecondTrueElement != Overdefined) {
296 Idx = MaskIdx(Idx);
297 // None true -> false.
298 if (FirstTrueElement == Undefined)
299 return replaceInstUsesWith(I&: ICI, V: Builder.getFalse());
300
301 Value *FirstTrueIdx = ConstantInt::get(Ty: Idx->getType(), V: FirstTrueElement);
302
303 // True for one element -> 'i == 47'.
304 if (SecondTrueElement == Undefined)
305 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
306
307 // True for two elements -> 'i == 47 | i == 72'.
308 Value *C1 = Builder.CreateICmpEQ(LHS: Idx, RHS: FirstTrueIdx);
309 Value *SecondTrueIdx = ConstantInt::get(Ty: Idx->getType(), V: SecondTrueElement);
310 Value *C2 = Builder.CreateICmpEQ(LHS: Idx, RHS: SecondTrueIdx);
311 return BinaryOperator::CreateOr(V1: C1, V2: C2);
312 }
313
314 // If the comparison is only false for one or two elements, emit direct
315 // comparisons.
316 if (SecondFalseElement != Overdefined) {
317 Idx = MaskIdx(Idx);
318 // None false -> true.
319 if (FirstFalseElement == Undefined)
320 return replaceInstUsesWith(I&: ICI, V: Builder.getTrue());
321
322 Value *FirstFalseIdx = ConstantInt::get(Ty: Idx->getType(), V: FirstFalseElement);
323
324 // False for one element -> 'i != 47'.
325 if (SecondFalseElement == Undefined)
326 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
327
328 // False for two elements -> 'i != 47 & i != 72'.
329 Value *C1 = Builder.CreateICmpNE(LHS: Idx, RHS: FirstFalseIdx);
330 Value *SecondFalseIdx =
331 ConstantInt::get(Ty: Idx->getType(), V: SecondFalseElement);
332 Value *C2 = Builder.CreateICmpNE(LHS: Idx, RHS: SecondFalseIdx);
333 return BinaryOperator::CreateAnd(V1: C1, V2: C2);
334 }
335
336 // If the comparison can be replaced with a range comparison for the elements
337 // where it is true, emit the range check.
338 if (TrueRangeEnd != Overdefined) {
339 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
340 Idx = MaskIdx(Idx);
341
342 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
343 if (FirstTrueElement) {
344 Value *Offs = ConstantInt::getSigned(Ty: Idx->getType(), V: -FirstTrueElement);
345 Idx = Builder.CreateAdd(LHS: Idx, RHS: Offs);
346 }
347
348 Value *End =
349 ConstantInt::get(Ty: Idx->getType(), V: TrueRangeEnd - FirstTrueElement + 1);
350 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
351 }
352
353 // False range check.
354 if (FalseRangeEnd != Overdefined) {
355 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
356 Idx = MaskIdx(Idx);
357 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
358 if (FirstFalseElement) {
359 Value *Offs = ConstantInt::getSigned(Ty: Idx->getType(), V: -FirstFalseElement);
360 Idx = Builder.CreateAdd(LHS: Idx, RHS: Offs);
361 }
362
363 Value *End =
364 ConstantInt::get(Ty: Idx->getType(), V: FalseRangeEnd - FirstFalseElement);
365 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
366 }
367
368 // If a magic bitvector captures the entire comparison state
369 // of this load, replace it with computation that does:
370 // ((magic_cst >> i) & 1) != 0
371 {
372 Type *Ty = nullptr;
373
374 // Look for an appropriate type:
375 // - The type of Idx if the magic fits
376 // - The smallest fitting legal type
377 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
378 Ty = Idx->getType();
379 else
380 Ty = DL.getSmallestLegalIntType(C&: Init->getContext(), Width: ArrayElementCount);
381
382 if (Ty) {
383 Idx = MaskIdx(Idx);
384 Value *V = Builder.CreateIntCast(V: Idx, DestTy: Ty, isSigned: false);
385 V = Builder.CreateLShr(LHS: ConstantInt::get(Ty, V: MagicBitvector), RHS: V);
386 V = Builder.CreateAnd(LHS: ConstantInt::get(Ty, V: 1), RHS: V);
387 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, V: 0));
388 }
389 }
390
391 return nullptr;
392}
393
394/// Returns true if we can rewrite Start as a GEP with pointer Base
395/// and some integer offset. The nodes that need to be re-written
396/// for this transformation will be added to Explored.
397static bool canRewriteGEPAsOffset(Value *Start, Value *Base, GEPNoWrapFlags &NW,
398 const DataLayout &DL,
399 SetVector<Value *> &Explored) {
400 SmallVector<Value *, 16> WorkList(1, Start);
401 Explored.insert(X: Base);
402
403 // The following traversal gives us an order which can be used
404 // when doing the final transformation. Since in the final
405 // transformation we create the PHI replacement instructions first,
406 // we don't have to get them in any particular order.
407 //
408 // However, for other instructions we will have to traverse the
409 // operands of an instruction first, which means that we have to
410 // do a post-order traversal.
411 while (!WorkList.empty()) {
412 SetVector<PHINode *> PHIs;
413
414 while (!WorkList.empty()) {
415 if (Explored.size() >= 100)
416 return false;
417
418 Value *V = WorkList.back();
419
420 if (Explored.contains(key: V)) {
421 WorkList.pop_back();
422 continue;
423 }
424
425 if (!isa<GetElementPtrInst>(Val: V) && !isa<PHINode>(Val: V))
426 // We've found some value that we can't explore which is different from
427 // the base. Therefore we can't do this transformation.
428 return false;
429
430 if (auto *GEP = dyn_cast<GEPOperator>(Val: V)) {
431 // Only allow inbounds GEPs with at most one variable offset.
432 auto IsNonConst = [](Value *V) { return !isa<ConstantInt>(Val: V); };
433 if (!GEP->isInBounds() || count_if(Range: GEP->indices(), P: IsNonConst) > 1)
434 return false;
435
436 NW = NW.intersectForOffsetAdd(Other: GEP->getNoWrapFlags());
437 if (!Explored.contains(key: GEP->getOperand(i_nocapture: 0)))
438 WorkList.push_back(Elt: GEP->getOperand(i_nocapture: 0));
439 }
440
441 if (WorkList.back() == V) {
442 WorkList.pop_back();
443 // We've finished visiting this node, mark it as such.
444 Explored.insert(X: V);
445 }
446
447 if (auto *PN = dyn_cast<PHINode>(Val: V)) {
448 // We cannot transform PHIs on unsplittable basic blocks.
449 if (isa<CatchSwitchInst>(Val: PN->getParent()->getTerminator()))
450 return false;
451 Explored.insert(X: PN);
452 PHIs.insert(X: PN);
453 }
454 }
455
456 // Explore the PHI nodes further.
457 for (auto *PN : PHIs)
458 for (Value *Op : PN->incoming_values())
459 if (!Explored.contains(key: Op))
460 WorkList.push_back(Elt: Op);
461 }
462
463 // Make sure that we can do this. Since we can't insert GEPs in a basic
464 // block before a PHI node, we can't easily do this transformation if
465 // we have PHI node users of transformed instructions.
466 for (Value *Val : Explored) {
467 for (Value *Use : Val->uses()) {
468
469 auto *PHI = dyn_cast<PHINode>(Val: Use);
470 auto *Inst = dyn_cast<Instruction>(Val);
471
472 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
473 !Explored.contains(key: PHI))
474 continue;
475
476 if (PHI->getParent() == Inst->getParent())
477 return false;
478 }
479 }
480 return true;
481}
482
483// Sets the appropriate insert point on Builder where we can add
484// a replacement Instruction for V (if that is possible).
485static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
486 bool Before = true) {
487 if (auto *PHI = dyn_cast<PHINode>(Val: V)) {
488 BasicBlock *Parent = PHI->getParent();
489 Builder.SetInsertPoint(TheBB: Parent, IP: Parent->getFirstInsertionPt());
490 return;
491 }
492 if (auto *I = dyn_cast<Instruction>(Val: V)) {
493 if (!Before)
494 I = &*std::next(x: I->getIterator());
495 Builder.SetInsertPoint(I);
496 return;
497 }
498 if (auto *A = dyn_cast<Argument>(Val: V)) {
499 // Set the insertion point in the entry block.
500 BasicBlock &Entry = A->getParent()->getEntryBlock();
501 Builder.SetInsertPoint(TheBB: &Entry, IP: Entry.getFirstInsertionPt());
502 return;
503 }
504 // Otherwise, this is a constant and we don't need to set a new
505 // insertion point.
506 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
507}
508
509/// Returns a re-written value of Start as an indexed GEP using Base as a
510/// pointer.
511static Value *rewriteGEPAsOffset(Value *Start, Value *Base, GEPNoWrapFlags NW,
512 const DataLayout &DL,
513 SetVector<Value *> &Explored,
514 InstCombiner &IC) {
515 // Perform all the substitutions. This is a bit tricky because we can
516 // have cycles in our use-def chains.
517 // 1. Create the PHI nodes without any incoming values.
518 // 2. Create all the other values.
519 // 3. Add the edges for the PHI nodes.
520 // 4. Emit GEPs to get the original pointers.
521 // 5. Remove the original instructions.
522 Type *IndexType = IntegerType::get(
523 C&: Base->getContext(), NumBits: DL.getIndexTypeSizeInBits(Ty: Start->getType()));
524
525 DenseMap<Value *, Value *> NewInsts;
526 NewInsts[Base] = ConstantInt::getNullValue(Ty: IndexType);
527
528 // Create the new PHI nodes, without adding any incoming values.
529 for (Value *Val : Explored) {
530 if (Val == Base)
531 continue;
532 // Create empty phi nodes. This avoids cyclic dependencies when creating
533 // the remaining instructions.
534 if (auto *PHI = dyn_cast<PHINode>(Val))
535 NewInsts[PHI] =
536 PHINode::Create(Ty: IndexType, NumReservedValues: PHI->getNumIncomingValues(),
537 NameStr: PHI->getName() + ".idx", InsertBefore: PHI->getIterator());
538 }
539 IRBuilder<> Builder(Base->getContext());
540
541 // Create all the other instructions.
542 for (Value *Val : Explored) {
543 if (NewInsts.contains(Val))
544 continue;
545
546 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
547 setInsertionPoint(Builder, V: GEP);
548 Value *Op = NewInsts[GEP->getOperand(i_nocapture: 0)];
549 Value *OffsetV = emitGEPOffset(Builder: &Builder, DL, GEP);
550 if (isa<ConstantInt>(Val: Op) && cast<ConstantInt>(Val: Op)->isZero())
551 NewInsts[GEP] = OffsetV;
552 else
553 NewInsts[GEP] = Builder.CreateAdd(
554 LHS: Op, RHS: OffsetV, Name: GEP->getOperand(i_nocapture: 0)->getName() + ".add",
555 /*NUW=*/HasNUW: NW.hasNoUnsignedWrap(),
556 /*NSW=*/HasNSW: NW.hasNoUnsignedSignedWrap());
557 continue;
558 }
559 if (isa<PHINode>(Val))
560 continue;
561
562 llvm_unreachable("Unexpected instruction type");
563 }
564
565 // Add the incoming values to the PHI nodes.
566 for (Value *Val : Explored) {
567 if (Val == Base)
568 continue;
569 // All the instructions have been created, we can now add edges to the
570 // phi nodes.
571 if (auto *PHI = dyn_cast<PHINode>(Val)) {
572 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
573 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
574 Value *NewIncoming = PHI->getIncomingValue(i: I);
575
576 auto It = NewInsts.find(Val: NewIncoming);
577 if (It != NewInsts.end())
578 NewIncoming = It->second;
579
580 NewPhi->addIncoming(V: NewIncoming, BB: PHI->getIncomingBlock(i: I));
581 }
582 }
583 }
584
585 for (Value *Val : Explored) {
586 if (Val == Base)
587 continue;
588
589 setInsertionPoint(Builder, V: Val, Before: false);
590 // Create GEP for external users.
591 Value *NewVal = Builder.CreateGEP(Ty: Builder.getInt8Ty(), Ptr: Base, IdxList: NewInsts[Val],
592 Name: Val->getName() + ".ptr", NW);
593 IC.replaceInstUsesWith(I&: *cast<Instruction>(Val), V: NewVal);
594 // Add old instruction to worklist for DCE. We don't directly remove it
595 // here because the original compare is one of the users.
596 IC.addToWorklist(I: cast<Instruction>(Val));
597 }
598
599 return NewInsts[Start];
600}
601
602/// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
603/// We can look through PHIs, GEPs and casts in order to determine a common base
604/// between GEPLHS and RHS.
605static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
606 CmpPredicate Cond,
607 const DataLayout &DL,
608 InstCombiner &IC) {
609 // FIXME: Support vector of pointers.
610 if (GEPLHS->getType()->isVectorTy())
611 return nullptr;
612
613 if (!GEPLHS->hasAllConstantIndices())
614 return nullptr;
615
616 APInt Offset(DL.getIndexTypeSizeInBits(Ty: GEPLHS->getType()), 0);
617 Value *PtrBase =
618 GEPLHS->stripAndAccumulateConstantOffsets(DL, Offset,
619 /*AllowNonInbounds*/ false);
620
621 // Bail if we looked through addrspacecast.
622 if (PtrBase->getType() != GEPLHS->getType())
623 return nullptr;
624
625 // The set of nodes that will take part in this transformation.
626 SetVector<Value *> Nodes;
627 GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags();
628 if (!canRewriteGEPAsOffset(Start: RHS, Base: PtrBase, NW, DL, Explored&: Nodes))
629 return nullptr;
630
631 // We know we can re-write this as
632 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
633 // Since we've only looked through inbouds GEPs we know that we
634 // can't have overflow on either side. We can therefore re-write
635 // this as:
636 // OFFSET1 cmp OFFSET2
637 Value *NewRHS = rewriteGEPAsOffset(Start: RHS, Base: PtrBase, NW, DL, Explored&: Nodes, IC);
638
639 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
640 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
641 // offset. Since Index is the offset of LHS to the base pointer, we will now
642 // compare the offsets instead of comparing the pointers.
643 return new ICmpInst(ICmpInst::getSignedPredicate(Pred: Cond),
644 IC.Builder.getInt(AI: Offset), NewRHS);
645}
646
647/// Fold comparisons between a GEP instruction and something else. At this point
648/// we know that the GEP is on the LHS of the comparison.
649Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
650 CmpPredicate Cond, Instruction &I) {
651 // Don't transform signed compares of GEPs into index compares. Even if the
652 // GEP is inbounds, the final add of the base pointer can have signed overflow
653 // and would change the result of the icmp.
654 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
655 // the maximum signed value for the pointer type.
656 if (ICmpInst::isSigned(Pred: Cond))
657 return nullptr;
658
659 // Look through bitcasts and addrspacecasts. We do not however want to remove
660 // 0 GEPs.
661 if (!isa<GetElementPtrInst>(Val: RHS))
662 RHS = RHS->stripPointerCasts();
663
664 auto CanFold = [Cond](GEPNoWrapFlags NW) {
665 if (ICmpInst::isEquality(P: Cond))
666 return true;
667
668 // Unsigned predicates can be folded if the GEPs have *any* nowrap flags.
669 assert(ICmpInst::isUnsigned(Cond));
670 return NW != GEPNoWrapFlags::none();
671 };
672
673 auto NewICmp = [Cond](GEPNoWrapFlags NW, Value *Op1, Value *Op2) {
674 if (!NW.hasNoUnsignedWrap()) {
675 // Convert signed to unsigned comparison.
676 return new ICmpInst(ICmpInst::getSignedPredicate(Pred: Cond), Op1, Op2);
677 }
678
679 auto *I = new ICmpInst(Cond, Op1, Op2);
680 I->setSameSign(NW.hasNoUnsignedSignedWrap());
681 return I;
682 };
683
684 CommonPointerBase Base = CommonPointerBase::compute(LHS: GEPLHS, RHS);
685 if (Base.Ptr == RHS && CanFold(Base.LHSNW) && !Base.isExpensive()) {
686 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
687 Type *IdxTy = DL.getIndexType(PtrTy: GEPLHS->getType());
688 Value *Offset =
689 EmitGEPOffsets(GEPs: Base.LHSGEPs, NW: Base.LHSNW, IdxTy, /*RewriteGEPs=*/true);
690 return NewICmp(Base.LHSNW, Offset,
691 Constant::getNullValue(Ty: Offset->getType()));
692 }
693
694 if (GEPLHS->isInBounds() && ICmpInst::isEquality(P: Cond) &&
695 isa<Constant>(Val: RHS) && cast<Constant>(Val: RHS)->isNullValue() &&
696 !NullPointerIsDefined(F: I.getFunction(),
697 AS: RHS->getType()->getPointerAddressSpace())) {
698 // For most address spaces, an allocation can't be placed at null, but null
699 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
700 // the only valid inbounds address derived from null, is null itself.
701 // Thus, we have four cases to consider:
702 // 1) Base == nullptr, Offset == 0 -> inbounds, null
703 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
704 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
705 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
706 //
707 // (Note if we're indexing a type of size 0, that simply collapses into one
708 // of the buckets above.)
709 //
710 // In general, we're allowed to make values less poison (i.e. remove
711 // sources of full UB), so in this case, we just select between the two
712 // non-poison cases (1 and 4 above).
713 //
714 // For vectors, we apply the same reasoning on a per-lane basis.
715 auto *Base = GEPLHS->getPointerOperand();
716 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
717 auto EC = cast<VectorType>(Val: GEPLHS->getType())->getElementCount();
718 Base = Builder.CreateVectorSplat(EC, V: Base);
719 }
720 return new ICmpInst(Cond, Base,
721 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
722 C: cast<Constant>(Val: RHS), Ty: Base->getType()));
723 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(Val: RHS)) {
724 GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags() & GEPRHS->getNoWrapFlags();
725
726 // If the base pointers are different, but the indices are the same, just
727 // compare the base pointer.
728 if (GEPLHS->getOperand(i_nocapture: 0) != GEPRHS->getOperand(i_nocapture: 0)) {
729 bool IndicesTheSame =
730 GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
731 GEPLHS->getPointerOperand()->getType() ==
732 GEPRHS->getPointerOperand()->getType() &&
733 GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType();
734 if (IndicesTheSame)
735 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
736 if (GEPLHS->getOperand(i_nocapture: i) != GEPRHS->getOperand(i_nocapture: i)) {
737 IndicesTheSame = false;
738 break;
739 }
740
741 // If all indices are the same, just compare the base pointers.
742 Type *BaseType = GEPLHS->getOperand(i_nocapture: 0)->getType();
743 if (IndicesTheSame &&
744 CmpInst::makeCmpResultType(opnd_type: BaseType) == I.getType() && CanFold(NW))
745 return new ICmpInst(Cond, GEPLHS->getOperand(i_nocapture: 0), GEPRHS->getOperand(i_nocapture: 0));
746
747 // If we're comparing GEPs with two base pointers that only differ in type
748 // and both GEPs have only constant indices or just one use, then fold
749 // the compare with the adjusted indices.
750 // FIXME: Support vector of pointers.
751 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
752 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
753 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
754 GEPLHS->getOperand(i_nocapture: 0)->stripPointerCasts() ==
755 GEPRHS->getOperand(i_nocapture: 0)->stripPointerCasts() &&
756 !GEPLHS->getType()->isVectorTy()) {
757 Value *LOffset = EmitGEPOffset(GEP: GEPLHS);
758 Value *ROffset = EmitGEPOffset(GEP: GEPRHS);
759
760 // If we looked through an addrspacecast between different sized address
761 // spaces, the LHS and RHS pointers are different sized
762 // integers. Truncate to the smaller one.
763 Type *LHSIndexTy = LOffset->getType();
764 Type *RHSIndexTy = ROffset->getType();
765 if (LHSIndexTy != RHSIndexTy) {
766 if (LHSIndexTy->getPrimitiveSizeInBits().getFixedValue() <
767 RHSIndexTy->getPrimitiveSizeInBits().getFixedValue()) {
768 ROffset = Builder.CreateTrunc(V: ROffset, DestTy: LHSIndexTy);
769 } else
770 LOffset = Builder.CreateTrunc(V: LOffset, DestTy: RHSIndexTy);
771 }
772
773 Value *Cmp = Builder.CreateICmp(P: ICmpInst::getSignedPredicate(Pred: Cond),
774 LHS: LOffset, RHS: ROffset);
775 return replaceInstUsesWith(I, V: Cmp);
776 }
777 }
778
779 if (GEPLHS->getOperand(i_nocapture: 0) == GEPRHS->getOperand(i_nocapture: 0) &&
780 GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
781 GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType()) {
782 // If the GEPs only differ by one index, compare it.
783 unsigned NumDifferences = 0; // Keep track of # differences.
784 unsigned DiffOperand = 0; // The operand that differs.
785 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
786 if (GEPLHS->getOperand(i_nocapture: i) != GEPRHS->getOperand(i_nocapture: i)) {
787 Type *LHSType = GEPLHS->getOperand(i_nocapture: i)->getType();
788 Type *RHSType = GEPRHS->getOperand(i_nocapture: i)->getType();
789 // FIXME: Better support for vector of pointers.
790 if (LHSType->getPrimitiveSizeInBits() !=
791 RHSType->getPrimitiveSizeInBits() ||
792 (GEPLHS->getType()->isVectorTy() &&
793 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
794 // Irreconcilable differences.
795 NumDifferences = 2;
796 break;
797 }
798
799 if (NumDifferences++)
800 break;
801 DiffOperand = i;
802 }
803
804 if (NumDifferences == 0) // SAME GEP?
805 return replaceInstUsesWith(
806 I, // No comparison is needed here.
807 V: ConstantInt::get(Ty: I.getType(), V: ICmpInst::isTrueWhenEqual(predicate: Cond)));
808 // If two GEPs only differ by an index, compare them.
809 // Note that nowrap flags are always needed when comparing two indices.
810 else if (NumDifferences == 1 && NW != GEPNoWrapFlags::none()) {
811 Value *LHSV = GEPLHS->getOperand(i_nocapture: DiffOperand);
812 Value *RHSV = GEPRHS->getOperand(i_nocapture: DiffOperand);
813 return NewICmp(NW, LHSV, RHSV);
814 }
815 }
816
817 if (Base.Ptr && CanFold(Base.LHSNW & Base.RHSNW) && !Base.isExpensive()) {
818 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
819 Type *IdxTy = DL.getIndexType(PtrTy: GEPLHS->getType());
820 Value *L =
821 EmitGEPOffsets(GEPs: Base.LHSGEPs, NW: Base.LHSNW, IdxTy, /*RewriteGEP=*/RewriteGEPs: true);
822 Value *R =
823 EmitGEPOffsets(GEPs: Base.RHSGEPs, NW: Base.RHSNW, IdxTy, /*RewriteGEP=*/RewriteGEPs: true);
824 return NewICmp(Base.LHSNW & Base.RHSNW, L, R);
825 }
826 }
827
828 // Try convert this to an indexed compare by looking through PHIs/casts as a
829 // last resort.
830 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL, IC&: *this);
831}
832
833bool InstCombinerImpl::foldAllocaCmp(AllocaInst *Alloca) {
834 // It would be tempting to fold away comparisons between allocas and any
835 // pointer not based on that alloca (e.g. an argument). However, even
836 // though such pointers cannot alias, they can still compare equal.
837 //
838 // But LLVM doesn't specify where allocas get their memory, so if the alloca
839 // doesn't escape we can argue that it's impossible to guess its value, and we
840 // can therefore act as if any such guesses are wrong.
841 //
842 // However, we need to ensure that this folding is consistent: We can't fold
843 // one comparison to false, and then leave a different comparison against the
844 // same value alone (as it might evaluate to true at runtime, leading to a
845 // contradiction). As such, this code ensures that all comparisons are folded
846 // at the same time, and there are no other escapes.
847
848 struct CmpCaptureTracker : public CaptureTracker {
849 AllocaInst *Alloca;
850 bool Captured = false;
851 /// The value of the map is a bit mask of which icmp operands the alloca is
852 /// used in.
853 SmallMapVector<ICmpInst *, unsigned, 4> ICmps;
854
855 CmpCaptureTracker(AllocaInst *Alloca) : Alloca(Alloca) {}
856
857 void tooManyUses() override { Captured = true; }
858
859 Action captured(const Use *U, UseCaptureInfo CI) override {
860 // TODO(captures): Use UseCaptureInfo.
861 auto *ICmp = dyn_cast<ICmpInst>(Val: U->getUser());
862 // We need to check that U is based *only* on the alloca, and doesn't
863 // have other contributions from a select/phi operand.
864 // TODO: We could check whether getUnderlyingObjects() reduces to one
865 // object, which would allow looking through phi nodes.
866 if (ICmp && ICmp->isEquality() && getUnderlyingObject(V: *U) == Alloca) {
867 // Collect equality icmps of the alloca, and don't treat them as
868 // captures.
869 ICmps[ICmp] |= 1u << U->getOperandNo();
870 return Continue;
871 }
872
873 Captured = true;
874 return Stop;
875 }
876 };
877
878 CmpCaptureTracker Tracker(Alloca);
879 PointerMayBeCaptured(V: Alloca, Tracker: &Tracker);
880 if (Tracker.Captured)
881 return false;
882
883 bool Changed = false;
884 for (auto [ICmp, Operands] : Tracker.ICmps) {
885 switch (Operands) {
886 case 1:
887 case 2: {
888 // The alloca is only used in one icmp operand. Assume that the
889 // equality is false.
890 auto *Res = ConstantInt::get(Ty: ICmp->getType(),
891 V: ICmp->getPredicate() == ICmpInst::ICMP_NE);
892 replaceInstUsesWith(I&: *ICmp, V: Res);
893 eraseInstFromFunction(I&: *ICmp);
894 Changed = true;
895 break;
896 }
897 case 3:
898 // Both icmp operands are based on the alloca, so this is comparing
899 // pointer offsets, without leaking any information about the address
900 // of the alloca. Ignore such comparisons.
901 break;
902 default:
903 llvm_unreachable("Cannot happen");
904 }
905 }
906
907 return Changed;
908}
909
910/// Fold "icmp pred (X+C), X".
911Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
912 CmpPredicate Pred) {
913 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
914 // so the values can never be equal. Similarly for all other "or equals"
915 // operators.
916 assert(!!C && "C should not be zero!");
917
918 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
919 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
920 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
921 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
922 Constant *R =
923 ConstantInt::get(Ty: X->getType(), V: APInt::getMaxValue(numBits: C.getBitWidth()) - C);
924 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
925 }
926
927 // (X+1) >u X --> X <u (0-1) --> X != 255
928 // (X+2) >u X --> X <u (0-2) --> X <u 254
929 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
930 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
931 return new ICmpInst(ICmpInst::ICMP_ULT, X,
932 ConstantInt::get(Ty: X->getType(), V: -C));
933
934 APInt SMax = APInt::getSignedMaxValue(numBits: C.getBitWidth());
935
936 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
937 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
938 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
939 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
940 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
941 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
942 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
943 return new ICmpInst(ICmpInst::ICMP_SGT, X,
944 ConstantInt::get(Ty: X->getType(), V: SMax - C));
945
946 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
947 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
948 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
949 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
950 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
951 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
952
953 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
954 return new ICmpInst(ICmpInst::ICMP_SLT, X,
955 ConstantInt::get(Ty: X->getType(), V: SMax - (C - 1)));
956}
957
958/// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
959/// (icmp eq/ne A, Log2(AP2/AP1)) ->
960/// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
961Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
962 const APInt &AP1,
963 const APInt &AP2) {
964 assert(I.isEquality() && "Cannot fold icmp gt/lt");
965
966 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
967 if (I.getPredicate() == I.ICMP_NE)
968 Pred = CmpInst::getInversePredicate(pred: Pred);
969 return new ICmpInst(Pred, LHS, RHS);
970 };
971
972 // Don't bother doing any work for cases which InstSimplify handles.
973 if (AP2.isZero())
974 return nullptr;
975
976 bool IsAShr = isa<AShrOperator>(Val: I.getOperand(i_nocapture: 0));
977 if (IsAShr) {
978 if (AP2.isAllOnes())
979 return nullptr;
980 if (AP2.isNegative() != AP1.isNegative())
981 return nullptr;
982 if (AP2.sgt(RHS: AP1))
983 return nullptr;
984 }
985
986 if (!AP1)
987 // 'A' must be large enough to shift out the highest set bit.
988 return getICmp(I.ICMP_UGT, A,
989 ConstantInt::get(Ty: A->getType(), V: AP2.logBase2()));
990
991 if (AP1 == AP2)
992 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(Ty: A->getType()));
993
994 int Shift;
995 if (IsAShr && AP1.isNegative())
996 Shift = AP1.countl_one() - AP2.countl_one();
997 else
998 Shift = AP1.countl_zero() - AP2.countl_zero();
999
1000 if (Shift > 0) {
1001 if (IsAShr && AP1 == AP2.ashr(ShiftAmt: Shift)) {
1002 // There are multiple solutions if we are comparing against -1 and the LHS
1003 // of the ashr is not a power of two.
1004 if (AP1.isAllOnes() && !AP2.isPowerOf2())
1005 return getICmp(I.ICMP_UGE, A, ConstantInt::get(Ty: A->getType(), V: Shift));
1006 return getICmp(I.ICMP_EQ, A, ConstantInt::get(Ty: A->getType(), V: Shift));
1007 } else if (AP1 == AP2.lshr(shiftAmt: Shift)) {
1008 return getICmp(I.ICMP_EQ, A, ConstantInt::get(Ty: A->getType(), V: Shift));
1009 }
1010 }
1011
1012 // Shifting const2 will never be equal to const1.
1013 // FIXME: This should always be handled by InstSimplify?
1014 auto *TorF = ConstantInt::get(Ty: I.getType(), V: I.getPredicate() == I.ICMP_NE);
1015 return replaceInstUsesWith(I, V: TorF);
1016}
1017
1018/// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1019/// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1020Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1021 const APInt &AP1,
1022 const APInt &AP2) {
1023 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1024
1025 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1026 if (I.getPredicate() == I.ICMP_NE)
1027 Pred = CmpInst::getInversePredicate(pred: Pred);
1028 return new ICmpInst(Pred, LHS, RHS);
1029 };
1030
1031 // Don't bother doing any work for cases which InstSimplify handles.
1032 if (AP2.isZero())
1033 return nullptr;
1034
1035 unsigned AP2TrailingZeros = AP2.countr_zero();
1036
1037 if (!AP1 && AP2TrailingZeros != 0)
1038 return getICmp(
1039 I.ICMP_UGE, A,
1040 ConstantInt::get(Ty: A->getType(), V: AP2.getBitWidth() - AP2TrailingZeros));
1041
1042 if (AP1 == AP2)
1043 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(Ty: A->getType()));
1044
1045 // Get the distance between the lowest bits that are set.
1046 int Shift = AP1.countr_zero() - AP2TrailingZeros;
1047
1048 if (Shift > 0 && AP2.shl(shiftAmt: Shift) == AP1)
1049 return getICmp(I.ICMP_EQ, A, ConstantInt::get(Ty: A->getType(), V: Shift));
1050
1051 // Shifting const2 will never be equal to const1.
1052 // FIXME: This should always be handled by InstSimplify?
1053 auto *TorF = ConstantInt::get(Ty: I.getType(), V: I.getPredicate() == I.ICMP_NE);
1054 return replaceInstUsesWith(I, V: TorF);
1055}
1056
1057/// The caller has matched a pattern of the form:
1058/// I = icmp ugt (add (add A, B), CI2), CI1
1059/// If this is of the form:
1060/// sum = a + b
1061/// if (sum+128 >u 255)
1062/// Then replace it with llvm.sadd.with.overflow.i8.
1063///
1064static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1065 ConstantInt *CI2, ConstantInt *CI1,
1066 InstCombinerImpl &IC) {
1067 // The transformation we're trying to do here is to transform this into an
1068 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1069 // with a narrower add, and discard the add-with-constant that is part of the
1070 // range check (if we can't eliminate it, this isn't profitable).
1071
1072 // In order to eliminate the add-with-constant, the compare can be its only
1073 // use.
1074 Instruction *AddWithCst = cast<Instruction>(Val: I.getOperand(i_nocapture: 0));
1075 if (!AddWithCst->hasOneUse())
1076 return nullptr;
1077
1078 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1079 if (!CI2->getValue().isPowerOf2())
1080 return nullptr;
1081 unsigned NewWidth = CI2->getValue().countr_zero();
1082 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1083 return nullptr;
1084
1085 // The width of the new add formed is 1 more than the bias.
1086 ++NewWidth;
1087
1088 // Check to see that CI1 is an all-ones value with NewWidth bits.
1089 if (CI1->getBitWidth() == NewWidth ||
1090 CI1->getValue() != APInt::getLowBitsSet(numBits: CI1->getBitWidth(), loBitsSet: NewWidth))
1091 return nullptr;
1092
1093 // This is only really a signed overflow check if the inputs have been
1094 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1095 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1096 if (IC.ComputeMaxSignificantBits(Op: A, CxtI: &I) > NewWidth ||
1097 IC.ComputeMaxSignificantBits(Op: B, CxtI: &I) > NewWidth)
1098 return nullptr;
1099
1100 // In order to replace the original add with a narrower
1101 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1102 // and truncates that discard the high bits of the add. Verify that this is
1103 // the case.
1104 Instruction *OrigAdd = cast<Instruction>(Val: AddWithCst->getOperand(i: 0));
1105 for (User *U : OrigAdd->users()) {
1106 if (U == AddWithCst)
1107 continue;
1108
1109 // Only accept truncates for now. We would really like a nice recursive
1110 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1111 // chain to see which bits of a value are actually demanded. If the
1112 // original add had another add which was then immediately truncated, we
1113 // could still do the transformation.
1114 TruncInst *TI = dyn_cast<TruncInst>(Val: U);
1115 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1116 return nullptr;
1117 }
1118
1119 // If the pattern matches, truncate the inputs to the narrower type and
1120 // use the sadd_with_overflow intrinsic to efficiently compute both the
1121 // result and the overflow bit.
1122 Type *NewType = IntegerType::get(C&: OrigAdd->getContext(), NumBits: NewWidth);
1123 Function *F = Intrinsic::getOrInsertDeclaration(
1124 M: I.getModule(), id: Intrinsic::sadd_with_overflow, Tys: NewType);
1125
1126 InstCombiner::BuilderTy &Builder = IC.Builder;
1127
1128 // Put the new code above the original add, in case there are any uses of the
1129 // add between the add and the compare.
1130 Builder.SetInsertPoint(OrigAdd);
1131
1132 Value *TruncA = Builder.CreateTrunc(V: A, DestTy: NewType, Name: A->getName() + ".trunc");
1133 Value *TruncB = Builder.CreateTrunc(V: B, DestTy: NewType, Name: B->getName() + ".trunc");
1134 CallInst *Call = Builder.CreateCall(Callee: F, Args: {TruncA, TruncB}, Name: "sadd");
1135 Value *Add = Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "sadd.result");
1136 Value *ZExt = Builder.CreateZExt(V: Add, DestTy: OrigAdd->getType());
1137
1138 // The inner add was the result of the narrow add, zero extended to the
1139 // wider type. Replace it with the result computed by the intrinsic.
1140 IC.replaceInstUsesWith(I&: *OrigAdd, V: ZExt);
1141 IC.eraseInstFromFunction(I&: *OrigAdd);
1142
1143 // The original icmp gets replaced with the overflow value.
1144 return ExtractValueInst::Create(Agg: Call, Idxs: 1, NameStr: "sadd.overflow");
1145}
1146
1147/// If we have:
1148/// icmp eq/ne (urem/srem %x, %y), 0
1149/// iff %y is a power-of-two, we can replace this with a bit test:
1150/// icmp eq/ne (and %x, (add %y, -1)), 0
1151Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1152 // This fold is only valid for equality predicates.
1153 if (!I.isEquality())
1154 return nullptr;
1155 CmpPredicate Pred;
1156 Value *X, *Y, *Zero;
1157 if (!match(V: &I, P: m_ICmp(Pred, L: m_OneUse(SubPattern: m_IRem(L: m_Value(V&: X), R: m_Value(V&: Y))),
1158 R: m_CombineAnd(L: m_Zero(), R: m_Value(V&: Zero)))))
1159 return nullptr;
1160 if (!isKnownToBeAPowerOfTwo(V: Y, /*OrZero*/ true, CxtI: &I))
1161 return nullptr;
1162 // This may increase instruction count, we don't enforce that Y is a constant.
1163 Value *Mask = Builder.CreateAdd(LHS: Y, RHS: Constant::getAllOnesValue(Ty: Y->getType()));
1164 Value *Masked = Builder.CreateAnd(LHS: X, RHS: Mask);
1165 return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: Masked, S2: Zero);
1166}
1167
1168/// Fold equality-comparison between zero and any (maybe truncated) right-shift
1169/// by one-less-than-bitwidth into a sign test on the original value.
1170Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1171 Instruction *Val;
1172 CmpPredicate Pred;
1173 if (!I.isEquality() || !match(V: &I, P: m_ICmp(Pred, L: m_Instruction(I&: Val), R: m_Zero())))
1174 return nullptr;
1175
1176 Value *X;
1177 Type *XTy;
1178
1179 Constant *C;
1180 if (match(V: Val, P: m_TruncOrSelf(Op: m_Shr(L: m_Value(V&: X), R: m_Constant(C))))) {
1181 XTy = X->getType();
1182 unsigned XBitWidth = XTy->getScalarSizeInBits();
1183 if (!match(V: C, P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_EQ,
1184 Threshold: APInt(XBitWidth, XBitWidth - 1))))
1185 return nullptr;
1186 } else if (isa<BinaryOperator>(Val) &&
1187 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1188 Sh0: cast<BinaryOperator>(Val), SQ: SQ.getWithInstruction(I: Val),
1189 /*AnalyzeForSignBitExtraction=*/true))) {
1190 XTy = X->getType();
1191 } else
1192 return nullptr;
1193
1194 return ICmpInst::Create(Op: Instruction::ICmp,
1195 Pred: Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1196 : ICmpInst::ICMP_SLT,
1197 S1: X, S2: ConstantInt::getNullValue(Ty: XTy));
1198}
1199
1200// Handle icmp pred X, 0
1201Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1202 CmpInst::Predicate Pred = Cmp.getPredicate();
1203 if (!match(V: Cmp.getOperand(i_nocapture: 1), P: m_Zero()))
1204 return nullptr;
1205
1206 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1207 if (Pred == ICmpInst::ICMP_SGT) {
1208 Value *A, *B;
1209 if (match(V: Cmp.getOperand(i_nocapture: 0), P: m_SMin(L: m_Value(V&: A), R: m_Value(V&: B)))) {
1210 if (isKnownPositive(V: A, SQ: SQ.getWithInstruction(I: &Cmp)))
1211 return new ICmpInst(Pred, B, Cmp.getOperand(i_nocapture: 1));
1212 if (isKnownPositive(V: B, SQ: SQ.getWithInstruction(I: &Cmp)))
1213 return new ICmpInst(Pred, A, Cmp.getOperand(i_nocapture: 1));
1214 }
1215 }
1216
1217 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(I&: Cmp))
1218 return New;
1219
1220 // Given:
1221 // icmp eq/ne (urem %x, %y), 0
1222 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1223 // icmp eq/ne %x, 0
1224 Value *X, *Y;
1225 if (match(V: Cmp.getOperand(i_nocapture: 0), P: m_URem(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
1226 ICmpInst::isEquality(P: Pred)) {
1227 KnownBits XKnown = computeKnownBits(V: X, CxtI: &Cmp);
1228 KnownBits YKnown = computeKnownBits(V: Y, CxtI: &Cmp);
1229 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1230 return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1));
1231 }
1232
1233 // (icmp eq/ne (mul X Y)) -> (icmp eq/ne X/Y) if we know about whether X/Y are
1234 // odd/non-zero/there is no overflow.
1235 if (match(V: Cmp.getOperand(i_nocapture: 0), P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
1236 ICmpInst::isEquality(P: Pred)) {
1237
1238 KnownBits XKnown = computeKnownBits(V: X, CxtI: &Cmp);
1239 // if X % 2 != 0
1240 // (icmp eq/ne Y)
1241 if (XKnown.countMaxTrailingZeros() == 0)
1242 return new ICmpInst(Pred, Y, Cmp.getOperand(i_nocapture: 1));
1243
1244 KnownBits YKnown = computeKnownBits(V: Y, CxtI: &Cmp);
1245 // if Y % 2 != 0
1246 // (icmp eq/ne X)
1247 if (YKnown.countMaxTrailingZeros() == 0)
1248 return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1));
1249
1250 auto *BO0 = cast<OverflowingBinaryOperator>(Val: Cmp.getOperand(i_nocapture: 0));
1251 if (BO0->hasNoUnsignedWrap() || BO0->hasNoSignedWrap()) {
1252 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
1253 // `isKnownNonZero` does more analysis than just `!KnownBits.One.isZero()`
1254 // but to avoid unnecessary work, first just if this is an obvious case.
1255
1256 // if X non-zero and NoOverflow(X * Y)
1257 // (icmp eq/ne Y)
1258 if (!XKnown.One.isZero() || isKnownNonZero(V: X, Q))
1259 return new ICmpInst(Pred, Y, Cmp.getOperand(i_nocapture: 1));
1260
1261 // if Y non-zero and NoOverflow(X * Y)
1262 // (icmp eq/ne X)
1263 if (!YKnown.One.isZero() || isKnownNonZero(V: Y, Q))
1264 return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1));
1265 }
1266 // Note, we are skipping cases:
1267 // if Y % 2 != 0 AND X % 2 != 0
1268 // (false/true)
1269 // if X non-zero and Y non-zero and NoOverflow(X * Y)
1270 // (false/true)
1271 // Those can be simplified later as we would have already replaced the (icmp
1272 // eq/ne (mul X, Y)) with (icmp eq/ne X/Y) and if X/Y is known non-zero that
1273 // will fold to a constant elsewhere.
1274 }
1275
1276 // (icmp eq/ne f(X), 0) -> (icmp eq/ne X, 0)
1277 // where f(X) == 0 if and only if X == 0
1278 if (ICmpInst::isEquality(P: Pred))
1279 if (Value *Stripped = stripNullTest(V: Cmp.getOperand(i_nocapture: 0)))
1280 return new ICmpInst(Pred, Stripped,
1281 Constant::getNullValue(Ty: Stripped->getType()));
1282
1283 return nullptr;
1284}
1285
1286/// Fold icmp eq (num + mask) & ~mask, num
1287/// to
1288/// icmp eq (and num, mask), 0
1289/// Where mask is a low bit mask.
1290Instruction *InstCombinerImpl::foldIsMultipleOfAPowerOfTwo(ICmpInst &Cmp) {
1291 Value *Num;
1292 CmpPredicate Pred;
1293 const APInt *Mask, *Neg;
1294
1295 if (!match(V: &Cmp,
1296 P: m_c_ICmp(Pred, L: m_Value(V&: Num),
1297 R: m_OneUse(SubPattern: m_c_And(L: m_OneUse(SubPattern: m_c_Add(L: m_Deferred(V: Num),
1298 R: m_LowBitMask(V&: Mask))),
1299 R: m_APInt(Res&: Neg))))))
1300 return nullptr;
1301
1302 if (*Neg != ~*Mask)
1303 return nullptr;
1304
1305 if (!ICmpInst::isEquality(P: Pred))
1306 return nullptr;
1307
1308 // Create new icmp eq (num & mask), 0
1309 auto *NewAnd = Builder.CreateAnd(LHS: Num, RHS: *Mask);
1310 auto *Zero = Constant::getNullValue(Ty: Num->getType());
1311
1312 return new ICmpInst(Pred, NewAnd, Zero);
1313}
1314
1315/// Fold icmp Pred X, C.
1316/// TODO: This code structure does not make sense. The saturating add fold
1317/// should be moved to some other helper and extended as noted below (it is also
1318/// possible that code has been made unnecessary - do we canonicalize IR to
1319/// overflow/saturating intrinsics or not?).
1320Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1321 // Match the following pattern, which is a common idiom when writing
1322 // overflow-safe integer arithmetic functions. The source performs an addition
1323 // in wider type and explicitly checks for overflow using comparisons against
1324 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1325 //
1326 // TODO: This could probably be generalized to handle other overflow-safe
1327 // operations if we worked out the formulas to compute the appropriate magic
1328 // constants.
1329 //
1330 // sum = a + b
1331 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1332 CmpInst::Predicate Pred = Cmp.getPredicate();
1333 Value *Op0 = Cmp.getOperand(i_nocapture: 0), *Op1 = Cmp.getOperand(i_nocapture: 1);
1334 Value *A, *B;
1335 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1336 if (Pred == ICmpInst::ICMP_UGT && match(V: Op1, P: m_ConstantInt(CI)) &&
1337 match(V: Op0, P: m_Add(L: m_Add(L: m_Value(V&: A), R: m_Value(V&: B)), R: m_ConstantInt(CI&: CI2))))
1338 if (Instruction *Res = processUGT_ADDCST_ADD(I&: Cmp, A, B, CI2, CI1: CI, IC&: *this))
1339 return Res;
1340
1341 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1342 Constant *C = dyn_cast<Constant>(Val: Op1);
1343 if (!C)
1344 return nullptr;
1345
1346 if (auto *Phi = dyn_cast<PHINode>(Val: Op0))
1347 if (all_of(Range: Phi->operands(), P: IsaPred<Constant>)) {
1348 SmallVector<Constant *> Ops;
1349 for (Value *V : Phi->incoming_values()) {
1350 Constant *Res =
1351 ConstantFoldCompareInstOperands(Predicate: Pred, LHS: cast<Constant>(Val: V), RHS: C, DL);
1352 if (!Res)
1353 return nullptr;
1354 Ops.push_back(Elt: Res);
1355 }
1356 Builder.SetInsertPoint(Phi);
1357 PHINode *NewPhi = Builder.CreatePHI(Ty: Cmp.getType(), NumReservedValues: Phi->getNumOperands());
1358 for (auto [V, Pred] : zip(t&: Ops, u: Phi->blocks()))
1359 NewPhi->addIncoming(V, BB: Pred);
1360 return replaceInstUsesWith(I&: Cmp, V: NewPhi);
1361 }
1362
1363 if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &Cmp))
1364 return R;
1365
1366 return nullptr;
1367}
1368
1369/// Canonicalize icmp instructions based on dominating conditions.
1370Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1371 // We already checked simple implication in InstSimplify, only handle complex
1372 // cases here.
1373 Value *X = Cmp.getOperand(i_nocapture: 0), *Y = Cmp.getOperand(i_nocapture: 1);
1374 const APInt *C;
1375 if (!match(V: Y, P: m_APInt(Res&: C)))
1376 return nullptr;
1377
1378 CmpInst::Predicate Pred = Cmp.getPredicate();
1379 ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred, Other: *C);
1380
1381 auto handleDomCond = [&](ICmpInst::Predicate DomPred,
1382 const APInt *DomC) -> Instruction * {
1383 // We have 2 compares of a variable with constants. Calculate the constant
1384 // ranges of those compares to see if we can transform the 2nd compare:
1385 // DomBB:
1386 // DomCond = icmp DomPred X, DomC
1387 // br DomCond, CmpBB, FalseBB
1388 // CmpBB:
1389 // Cmp = icmp Pred X, C
1390 ConstantRange DominatingCR =
1391 ConstantRange::makeExactICmpRegion(Pred: DomPred, Other: *DomC);
1392 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1393 ConstantRange Difference = DominatingCR.difference(CR);
1394 if (Intersection.isEmptySet())
1395 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
1396 if (Difference.isEmptySet())
1397 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
1398
1399 // Canonicalizing a sign bit comparison that gets used in a branch,
1400 // pessimizes codegen by generating branch on zero instruction instead
1401 // of a test and branch. So we avoid canonicalizing in such situations
1402 // because test and branch instruction has better branch displacement
1403 // than compare and branch instruction.
1404 bool UnusedBit;
1405 bool IsSignBit = isSignBitCheck(Pred, RHS: *C, TrueIfSigned&: UnusedBit);
1406 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(I&: Cmp)))
1407 return nullptr;
1408
1409 // Avoid an infinite loop with min/max canonicalization.
1410 // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1411 if (Cmp.hasOneUse() &&
1412 match(V: Cmp.user_back(), P: m_MaxOrMin(L: m_Value(), R: m_Value())))
1413 return nullptr;
1414
1415 if (const APInt *EqC = Intersection.getSingleElement())
1416 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(AI: *EqC));
1417 if (const APInt *NeC = Difference.getSingleElement())
1418 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(AI: *NeC));
1419 return nullptr;
1420 };
1421
1422 for (CondBrInst *BI : DC.conditionsFor(V: X)) {
1423 CmpPredicate DomPred;
1424 const APInt *DomC;
1425 if (!match(V: BI->getCondition(),
1426 P: m_ICmp(Pred&: DomPred, L: m_Specific(V: X), R: m_APInt(Res&: DomC))))
1427 continue;
1428
1429 BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0));
1430 if (DT.dominates(BBE: Edge0, BB: Cmp.getParent())) {
1431 if (auto *V = handleDomCond(DomPred, DomC))
1432 return V;
1433 } else {
1434 BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1));
1435 if (DT.dominates(BBE: Edge1, BB: Cmp.getParent()))
1436 if (auto *V =
1437 handleDomCond(CmpInst::getInversePredicate(pred: DomPred), DomC))
1438 return V;
1439 }
1440 }
1441
1442 return nullptr;
1443}
1444
1445/// Fold icmp (trunc X), C.
1446Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1447 TruncInst *Trunc,
1448 const APInt &C) {
1449 ICmpInst::Predicate Pred = Cmp.getPredicate();
1450 Value *X = Trunc->getOperand(i_nocapture: 0);
1451 Type *SrcTy = X->getType();
1452 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1453 SrcBits = SrcTy->getScalarSizeInBits();
1454
1455 // Match (icmp pred (trunc nuw/nsw X), C)
1456 // Which we can convert to (icmp pred X, (sext/zext C))
1457 if (shouldChangeType(From: Trunc->getType(), To: SrcTy)) {
1458 if (Trunc->hasNoSignedWrap())
1459 return new ICmpInst(Pred, X, ConstantInt::get(Ty: SrcTy, V: C.sext(width: SrcBits)));
1460 if (!Cmp.isSigned() && Trunc->hasNoUnsignedWrap())
1461 return new ICmpInst(Pred, X, ConstantInt::get(Ty: SrcTy, V: C.zext(width: SrcBits)));
1462 }
1463
1464 if (C.isOne() && C.getBitWidth() > 1) {
1465 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1466 Value *V = nullptr;
1467 if (Pred == ICmpInst::ICMP_SLT && match(V: X, P: m_Signum(V: m_Value(V))))
1468 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1469 ConstantInt::get(Ty: V->getType(), V: 1));
1470 }
1471
1472 // TODO: Handle non-equality predicates.
1473 Value *Y;
1474 const APInt *Pow2;
1475 if (Cmp.isEquality() && match(V: X, P: m_Shl(L: m_Power2(V&: Pow2), R: m_Value(V&: Y))) &&
1476 DstBits > Pow2->logBase2()) {
1477 // (trunc (Pow2 << Y) to iN) == 0 --> Y u>= N - log2(Pow2)
1478 // (trunc (Pow2 << Y) to iN) != 0 --> Y u< N - log2(Pow2)
1479 // iff N > log2(Pow2)
1480 if (C.isZero()) {
1481 auto NewPred = (Pred == Cmp.ICMP_EQ) ? Cmp.ICMP_UGE : Cmp.ICMP_ULT;
1482 return new ICmpInst(NewPred, Y,
1483 ConstantInt::get(Ty: SrcTy, V: DstBits - Pow2->logBase2()));
1484 }
1485 // (trunc (Pow2 << Y) to iN) == 2**C --> Y == C - log2(Pow2)
1486 // (trunc (Pow2 << Y) to iN) != 2**C --> Y != C - log2(Pow2)
1487 if (C.isPowerOf2())
1488 return new ICmpInst(
1489 Pred, Y, ConstantInt::get(Ty: SrcTy, V: C.logBase2() - Pow2->logBase2()));
1490 }
1491
1492 if (Cmp.isEquality() && (Trunc->hasOneUse() || Trunc->hasNoUnsignedWrap())) {
1493 // Canonicalize to a mask and wider compare if the wide type is suitable:
1494 // (trunc X to i8) == C --> (X & 0xff) == (zext C)
1495 if (!SrcTy->isVectorTy() && shouldChangeType(FromBitWidth: DstBits, ToBitWidth: SrcBits)) {
1496 Constant *Mask =
1497 ConstantInt::get(Ty: SrcTy, V: APInt::getLowBitsSet(numBits: SrcBits, loBitsSet: DstBits));
1498 Value *And = Trunc->hasNoUnsignedWrap() ? X : Builder.CreateAnd(LHS: X, RHS: Mask);
1499 Constant *WideC = ConstantInt::get(Ty: SrcTy, V: C.zext(width: SrcBits));
1500 return new ICmpInst(Pred, And, WideC);
1501 }
1502
1503 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1504 // of the high bits truncated out of x are known.
1505 KnownBits Known = computeKnownBits(V: X, CxtI: &Cmp);
1506
1507 // If all the high bits are known, we can do this xform.
1508 if ((Known.Zero | Known.One).countl_one() >= SrcBits - DstBits) {
1509 // Pull in the high bits from known-ones set.
1510 APInt NewRHS = C.zext(width: SrcBits);
1511 NewRHS |= Known.One & APInt::getHighBitsSet(numBits: SrcBits, hiBitsSet: SrcBits - DstBits);
1512 return new ICmpInst(Pred, X, ConstantInt::get(Ty: SrcTy, V: NewRHS));
1513 }
1514 }
1515
1516 // Look through truncated right-shift of the sign-bit for a sign-bit check:
1517 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1518 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1519 Value *ShOp;
1520 uint64_t ShAmt;
1521 bool TrueIfSigned;
1522 if (isSignBitCheck(Pred, RHS: C, TrueIfSigned) &&
1523 match(V: X, P: m_Shr(L: m_Value(V&: ShOp), R: m_ConstantInt(V&: ShAmt))) &&
1524 DstBits == SrcBits - ShAmt) {
1525 return TrueIfSigned ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1526 ConstantInt::getNullValue(Ty: SrcTy))
1527 : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1528 ConstantInt::getAllOnesValue(Ty: SrcTy));
1529 }
1530
1531 return nullptr;
1532}
1533
1534/// Fold icmp (trunc nuw/nsw X), (trunc nuw/nsw Y).
1535/// Fold icmp (trunc nuw/nsw X), (zext/sext Y).
1536Instruction *
1537InstCombinerImpl::foldICmpTruncWithTruncOrExt(ICmpInst &Cmp,
1538 const SimplifyQuery &Q) {
1539 Value *X, *Y;
1540 CmpPredicate Pred;
1541 bool YIsSExt = false;
1542 // Try to match icmp (trunc X), (trunc Y)
1543 if (match(V: &Cmp, P: m_ICmp(Pred, L: m_Trunc(Op: m_Value(V&: X)), R: m_Trunc(Op: m_Value(V&: Y))))) {
1544 unsigned NoWrapFlags = cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 0))->getNoWrapKind() &
1545 cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 1))->getNoWrapKind();
1546 if (Cmp.isSigned()) {
1547 // For signed comparisons, both truncs must be nsw.
1548 if (!(NoWrapFlags & TruncInst::NoSignedWrap))
1549 return nullptr;
1550 } else {
1551 // For unsigned and equality comparisons, either both must be nuw or
1552 // both must be nsw, we don't care which.
1553 if (!NoWrapFlags)
1554 return nullptr;
1555 }
1556
1557 if (X->getType() != Y->getType() &&
1558 (!Cmp.getOperand(i_nocapture: 0)->hasOneUse() || !Cmp.getOperand(i_nocapture: 1)->hasOneUse()))
1559 return nullptr;
1560 if (!isDesirableIntType(BitWidth: X->getType()->getScalarSizeInBits()) &&
1561 isDesirableIntType(BitWidth: Y->getType()->getScalarSizeInBits())) {
1562 std::swap(a&: X, b&: Y);
1563 Pred = Cmp.getSwappedPredicate(pred: Pred);
1564 }
1565 YIsSExt = !(NoWrapFlags & TruncInst::NoUnsignedWrap);
1566 }
1567 // Try to match icmp (trunc nuw X), (zext Y)
1568 else if (!Cmp.isSigned() &&
1569 match(V: &Cmp, P: m_c_ICmp(Pred, L: m_NUWTrunc(Op: m_Value(V&: X)),
1570 R: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: Y)))))) {
1571 // Can fold trunc nuw + zext for unsigned and equality predicates.
1572 }
1573 // Try to match icmp (trunc nsw X), (sext Y)
1574 else if (match(V: &Cmp, P: m_c_ICmp(Pred, L: m_NSWTrunc(Op: m_Value(V&: X)),
1575 R: m_OneUse(SubPattern: m_ZExtOrSExt(Op: m_Value(V&: Y)))))) {
1576 // Can fold trunc nsw + zext/sext for all predicates.
1577 YIsSExt =
1578 isa<SExtInst>(Val: Cmp.getOperand(i_nocapture: 0)) || isa<SExtInst>(Val: Cmp.getOperand(i_nocapture: 1));
1579 } else
1580 return nullptr;
1581
1582 Type *TruncTy = Cmp.getOperand(i_nocapture: 0)->getType();
1583 unsigned TruncBits = TruncTy->getScalarSizeInBits();
1584
1585 // If this transform will end up changing from desirable types -> undesirable
1586 // types skip it.
1587 if (isDesirableIntType(BitWidth: TruncBits) &&
1588 !isDesirableIntType(BitWidth: X->getType()->getScalarSizeInBits()))
1589 return nullptr;
1590
1591 Value *NewY = Builder.CreateIntCast(V: Y, DestTy: X->getType(), isSigned: YIsSExt);
1592 return new ICmpInst(Pred, X, NewY);
1593}
1594
1595/// Fold icmp (xor X, Y), C.
1596Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1597 BinaryOperator *Xor,
1598 const APInt &C) {
1599 if (Instruction *I = foldICmpXorShiftConst(Cmp, Xor, C))
1600 return I;
1601
1602 Value *X = Xor->getOperand(i_nocapture: 0);
1603 Value *Y = Xor->getOperand(i_nocapture: 1);
1604 const APInt *XorC;
1605 if (!match(V: Y, P: m_APInt(Res&: XorC)))
1606 return nullptr;
1607
1608 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1609 // fold the xor.
1610 ICmpInst::Predicate Pred = Cmp.getPredicate();
1611 bool TrueIfSigned = false;
1612 if (isSignBitCheck(Pred: Cmp.getPredicate(), RHS: C, TrueIfSigned)) {
1613
1614 // If the sign bit of the XorCst is not set, there is no change to
1615 // the operation, just stop using the Xor.
1616 if (!XorC->isNegative())
1617 return replaceOperand(I&: Cmp, OpNum: 0, V: X);
1618
1619 // Emit the opposite comparison.
1620 if (TrueIfSigned)
1621 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1622 ConstantInt::getAllOnesValue(Ty: X->getType()));
1623 else
1624 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1625 ConstantInt::getNullValue(Ty: X->getType()));
1626 }
1627
1628 if (Xor->hasOneUse()) {
1629 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1630 if (!Cmp.isEquality() && XorC->isSignMask()) {
1631 Pred = Cmp.getFlippedSignednessPredicate();
1632 return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: C ^ *XorC));
1633 }
1634
1635 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1636 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1637 Pred = Cmp.getFlippedSignednessPredicate();
1638 Pred = Cmp.getSwappedPredicate(pred: Pred);
1639 return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: C ^ *XorC));
1640 }
1641 }
1642
1643 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1644 if (Pred == ICmpInst::ICMP_UGT) {
1645 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1646 if (*XorC == ~C && (C + 1).isPowerOf2())
1647 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1648 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1649 if (*XorC == C && (C + 1).isPowerOf2())
1650 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1651 }
1652 if (Pred == ICmpInst::ICMP_ULT) {
1653 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1654 if (*XorC == -C && C.isPowerOf2())
1655 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1656 ConstantInt::get(Ty: X->getType(), V: ~C));
1657 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1658 if (*XorC == C && (-C).isPowerOf2())
1659 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1660 ConstantInt::get(Ty: X->getType(), V: ~C));
1661 }
1662 return nullptr;
1663}
1664
1665/// For power-of-2 C:
1666/// ((X s>> ShiftC) ^ X) u< C --> (X + C) u< (C << 1)
1667/// ((X s>> ShiftC) ^ X) u> (C - 1) --> (X + C) u> ((C << 1) - 1)
1668Instruction *InstCombinerImpl::foldICmpXorShiftConst(ICmpInst &Cmp,
1669 BinaryOperator *Xor,
1670 const APInt &C) {
1671 CmpInst::Predicate Pred = Cmp.getPredicate();
1672 APInt PowerOf2;
1673 if (Pred == ICmpInst::ICMP_ULT)
1674 PowerOf2 = C;
1675 else if (Pred == ICmpInst::ICMP_UGT && !C.isMaxValue())
1676 PowerOf2 = C + 1;
1677 else
1678 return nullptr;
1679 if (!PowerOf2.isPowerOf2())
1680 return nullptr;
1681 Value *X;
1682 const APInt *ShiftC;
1683 if (!match(V: Xor, P: m_OneUse(SubPattern: m_c_Xor(L: m_Value(V&: X),
1684 R: m_AShr(L: m_Deferred(V: X), R: m_APInt(Res&: ShiftC))))))
1685 return nullptr;
1686 uint64_t Shift = ShiftC->getLimitedValue();
1687 Type *XType = X->getType();
1688 if (Shift == 0 || PowerOf2.isMinSignedValue())
1689 return nullptr;
1690 Value *Add = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: XType, V: PowerOf2));
1691 APInt Bound =
1692 Pred == ICmpInst::ICMP_ULT ? PowerOf2 << 1 : ((PowerOf2 << 1) - 1);
1693 return new ICmpInst(Pred, Add, ConstantInt::get(Ty: XType, V: Bound));
1694}
1695
1696/// Fold icmp (and (sh X, Y), C2), C1.
1697Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1698 BinaryOperator *And,
1699 const APInt &C1,
1700 const APInt &C2) {
1701 BinaryOperator *Shift = dyn_cast<BinaryOperator>(Val: And->getOperand(i_nocapture: 0));
1702 if (!Shift || !Shift->isShift())
1703 return nullptr;
1704
1705 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1706 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1707 // code produced by the clang front-end, for bitfield access.
1708 // This seemingly simple opportunity to fold away a shift turns out to be
1709 // rather complicated. See PR17827 for details.
1710 unsigned ShiftOpcode = Shift->getOpcode();
1711 bool IsShl = ShiftOpcode == Instruction::Shl;
1712 const APInt *C3;
1713 if (match(V: Shift->getOperand(i_nocapture: 1), P: m_APInt(Res&: C3))) {
1714 APInt NewAndCst, NewCmpCst;
1715 bool AnyCmpCstBitsShiftedOut;
1716 if (ShiftOpcode == Instruction::Shl) {
1717 // For a left shift, we can fold if the comparison is not signed. We can
1718 // also fold a signed comparison if the mask value and comparison value
1719 // are not negative. These constraints may not be obvious, but we can
1720 // prove that they are correct using an SMT solver.
1721 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1722 return nullptr;
1723
1724 NewCmpCst = C1.lshr(ShiftAmt: *C3);
1725 NewAndCst = C2.lshr(ShiftAmt: *C3);
1726 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(ShiftAmt: *C3) != C1;
1727 } else if (ShiftOpcode == Instruction::LShr) {
1728 // For a logical right shift, we can fold if the comparison is not signed.
1729 // We can also fold a signed comparison if the shifted mask value and the
1730 // shifted comparison value are not negative. These constraints may not be
1731 // obvious, but we can prove that they are correct using an SMT solver.
1732 NewCmpCst = C1.shl(ShiftAmt: *C3);
1733 NewAndCst = C2.shl(ShiftAmt: *C3);
1734 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(ShiftAmt: *C3) != C1;
1735 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1736 return nullptr;
1737 } else {
1738 // For an arithmetic shift, check that both constants don't use (in a
1739 // signed sense) the top bits being shifted out.
1740 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1741 NewCmpCst = C1.shl(ShiftAmt: *C3);
1742 NewAndCst = C2.shl(ShiftAmt: *C3);
1743 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(ShiftAmt: *C3) != C1;
1744 if (NewAndCst.ashr(ShiftAmt: *C3) != C2)
1745 return nullptr;
1746 }
1747
1748 if (AnyCmpCstBitsShiftedOut) {
1749 // If we shifted bits out, the fold is not going to work out. As a
1750 // special case, check to see if this means that the result is always
1751 // true or false now.
1752 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1753 return replaceInstUsesWith(I&: Cmp, V: ConstantInt::getFalse(Ty: Cmp.getType()));
1754 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1755 return replaceInstUsesWith(I&: Cmp, V: ConstantInt::getTrue(Ty: Cmp.getType()));
1756 } else {
1757 Value *NewAnd = Builder.CreateAnd(
1758 LHS: Shift->getOperand(i_nocapture: 0), RHS: ConstantInt::get(Ty: And->getType(), V: NewAndCst));
1759 return new ICmpInst(Cmp.getPredicate(), NewAnd,
1760 ConstantInt::get(Ty: And->getType(), V: NewCmpCst));
1761 }
1762 }
1763
1764 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1765 // preferable because it allows the C2 << Y expression to be hoisted out of a
1766 // loop if Y is invariant and X is not.
1767 if (Shift->hasOneUse() && C1.isZero() && Cmp.isEquality() &&
1768 !Shift->isArithmeticShift() &&
1769 ((!IsShl && C2.isOne()) || !isa<Constant>(Val: Shift->getOperand(i_nocapture: 0)))) {
1770 // Compute C2 << Y.
1771 Value *NewShift =
1772 IsShl ? Builder.CreateLShr(LHS: And->getOperand(i_nocapture: 1), RHS: Shift->getOperand(i_nocapture: 1))
1773 : Builder.CreateShl(LHS: And->getOperand(i_nocapture: 1), RHS: Shift->getOperand(i_nocapture: 1));
1774
1775 // Compute X & (C2 << Y).
1776 Value *NewAnd = Builder.CreateAnd(LHS: Shift->getOperand(i_nocapture: 0), RHS: NewShift);
1777 return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(i_nocapture: 1));
1778 }
1779
1780 return nullptr;
1781}
1782
1783/// Fold icmp (and X, C2), C1.
1784Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1785 BinaryOperator *And,
1786 const APInt &C1) {
1787 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1788
1789 // icmp ne (and X, 1), 0 --> trunc X to i1
1790 if (isICMP_NE && C1.isZero() && match(V: And->getOperand(i_nocapture: 1), P: m_One()))
1791 return new TruncInst(And->getOperand(i_nocapture: 0), Cmp.getType());
1792
1793 const APInt *C2;
1794 Value *X;
1795 if (!match(V: And, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: C2))))
1796 return nullptr;
1797
1798 // (and X, highmask) s> [0, ~highmask] --> X s> ~highmask
1799 if (Cmp.getPredicate() == ICmpInst::ICMP_SGT && C1.ule(RHS: ~*C2) &&
1800 C2->isNegatedPowerOf2())
1801 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1802 ConstantInt::get(Ty: X->getType(), V: ~*C2));
1803 // (and X, highmask) s< [1, -highmask] --> X s< -highmask
1804 if (Cmp.getPredicate() == ICmpInst::ICMP_SLT && !C1.isSignMask() &&
1805 (C1 - 1).ule(RHS: ~*C2) && C2->isNegatedPowerOf2() && !C2->isSignMask())
1806 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1807 ConstantInt::get(Ty: X->getType(), V: -*C2));
1808
1809 // Don't perform the following transforms if the AND has multiple uses
1810 if (!And->hasOneUse())
1811 return nullptr;
1812
1813 if (Cmp.isEquality() && C1.isZero()) {
1814 // Restrict this fold to single-use 'and' (PR10267).
1815 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1816 if (C2->isSignMask()) {
1817 Constant *Zero = Constant::getNullValue(Ty: X->getType());
1818 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1819 return new ICmpInst(NewPred, X, Zero);
1820 }
1821
1822 APInt NewC2 = *C2;
1823 KnownBits Know = computeKnownBits(V: And->getOperand(i_nocapture: 0), CxtI: And);
1824 // Set high zeros of C2 to allow matching negated power-of-2.
1825 NewC2 = *C2 | APInt::getHighBitsSet(numBits: C2->getBitWidth(),
1826 hiBitsSet: Know.countMinLeadingZeros());
1827
1828 // Restrict this fold only for single-use 'and' (PR10267).
1829 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1830 if (NewC2.isNegatedPowerOf2()) {
1831 Constant *NegBOC = ConstantInt::get(Ty: And->getType(), V: -NewC2);
1832 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1833 return new ICmpInst(NewPred, X, NegBOC);
1834 }
1835 }
1836
1837 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1838 // the input width without changing the value produced, eliminate the cast:
1839 //
1840 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1841 //
1842 // We can do this transformation if the constants do not have their sign bits
1843 // set or if it is an equality comparison. Extending a relational comparison
1844 // when we're checking the sign bit would not work.
1845 Value *W;
1846 if (match(V: And->getOperand(i_nocapture: 0), P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: W)))) &&
1847 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1848 // TODO: Is this a good transform for vectors? Wider types may reduce
1849 // throughput. Should this transform be limited (even for scalars) by using
1850 // shouldChangeType()?
1851 if (!Cmp.getType()->isVectorTy()) {
1852 Type *WideType = W->getType();
1853 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1854 Constant *ZextC1 = ConstantInt::get(Ty: WideType, V: C1.zext(width: WideScalarBits));
1855 Constant *ZextC2 = ConstantInt::get(Ty: WideType, V: C2->zext(width: WideScalarBits));
1856 Value *NewAnd = Builder.CreateAnd(LHS: W, RHS: ZextC2, Name: And->getName());
1857 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1858 }
1859 }
1860
1861 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2: *C2))
1862 return I;
1863
1864 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1865 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1866 //
1867 // iff pred isn't signed
1868 if (!Cmp.isSigned() && C1.isZero() && And->getOperand(i_nocapture: 0)->hasOneUse() &&
1869 match(V: And->getOperand(i_nocapture: 1), P: m_One())) {
1870 Constant *One = cast<Constant>(Val: And->getOperand(i_nocapture: 1));
1871 Value *Or = And->getOperand(i_nocapture: 0);
1872 Value *A, *B, *LShr;
1873 if (match(V: Or, P: m_Or(L: m_Value(V&: LShr), R: m_Value(V&: A))) &&
1874 match(V: LShr, P: m_LShr(L: m_Specific(V: A), R: m_Value(V&: B)))) {
1875 unsigned UsesRemoved = 0;
1876 if (And->hasOneUse())
1877 ++UsesRemoved;
1878 if (Or->hasOneUse())
1879 ++UsesRemoved;
1880 if (LShr->hasOneUse())
1881 ++UsesRemoved;
1882
1883 // Compute A & ((1 << B) | 1)
1884 unsigned RequireUsesRemoved = match(V: B, P: m_ImmConstant()) ? 1 : 3;
1885 if (UsesRemoved >= RequireUsesRemoved) {
1886 Value *NewOr =
1887 Builder.CreateOr(LHS: Builder.CreateShl(LHS: One, RHS: B, Name: LShr->getName(),
1888 /*HasNUW=*/true),
1889 RHS: One, Name: Or->getName());
1890 Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: NewOr, Name: And->getName());
1891 return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(i_nocapture: 1));
1892 }
1893 }
1894 }
1895
1896 // (icmp eq (and (bitcast X to int), ExponentMask), ExponentMask) -->
1897 // llvm.is.fpclass(X, fcInf|fcNan)
1898 // (icmp ne (and (bitcast X to int), ExponentMask), ExponentMask) -->
1899 // llvm.is.fpclass(X, ~(fcInf|fcNan))
1900 // (icmp eq (and (bitcast X to int), ExponentMask), 0) -->
1901 // llvm.is.fpclass(X, fcSubnormal|fcZero)
1902 // (icmp ne (and (bitcast X to int), ExponentMask), 0) -->
1903 // llvm.is.fpclass(X, ~(fcSubnormal|fcZero))
1904 Value *V;
1905 if (!Cmp.getParent()->getParent()->hasFnAttribute(
1906 Kind: Attribute::NoImplicitFloat) &&
1907 Cmp.isEquality() &&
1908 match(V: X, P: m_OneUse(SubPattern: m_ElementWiseBitCast(Op: m_Value(V))))) {
1909 Type *FPType = V->getType()->getScalarType();
1910 if (FPType->isIEEELikeFPTy() && (C1.isZero() || C1 == *C2)) {
1911 APInt ExponentMask =
1912 APFloat::getInf(Sem: FPType->getFltSemantics()).bitcastToAPInt();
1913 if (*C2 == ExponentMask) {
1914 unsigned Mask = C1.isZero()
1915 ? FPClassTest::fcZero | FPClassTest::fcSubnormal
1916 : FPClassTest::fcNan | FPClassTest::fcInf;
1917 if (isICMP_NE)
1918 Mask = ~Mask & fcAllFlags;
1919 return replaceInstUsesWith(I&: Cmp, V: Builder.createIsFPClass(FPNum: V, Test: Mask));
1920 }
1921 }
1922 }
1923
1924 return nullptr;
1925}
1926
1927/// Fold icmp (and X, Y), C.
1928Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1929 BinaryOperator *And,
1930 const APInt &C) {
1931 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C1: C))
1932 return I;
1933
1934 const ICmpInst::Predicate Pred = Cmp.getPredicate();
1935 bool TrueIfNeg;
1936 if (isSignBitCheck(Pred, RHS: C, TrueIfSigned&: TrueIfNeg)) {
1937 // ((X - 1) & ~X) < 0 --> X == 0
1938 // ((X - 1) & ~X) >= 0 --> X != 0
1939 Value *X;
1940 if (match(V: And->getOperand(i_nocapture: 0), P: m_Add(L: m_Value(V&: X), R: m_AllOnes())) &&
1941 match(V: And->getOperand(i_nocapture: 1), P: m_Not(V: m_Specific(V: X)))) {
1942 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1943 return new ICmpInst(NewPred, X, ConstantInt::getNullValue(Ty: X->getType()));
1944 }
1945 // (X & -X) < 0 --> X == MinSignedC
1946 // (X & -X) > -1 --> X != MinSignedC
1947 if (match(V: And, P: m_c_And(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) {
1948 Constant *MinSignedC = ConstantInt::get(
1949 Ty: X->getType(),
1950 V: APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits()));
1951 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1952 return new ICmpInst(NewPred, X, MinSignedC);
1953 }
1954 }
1955
1956 // TODO: These all require that Y is constant too, so refactor with the above.
1957
1958 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1959 Value *X = And->getOperand(i_nocapture: 0);
1960 Value *Y = And->getOperand(i_nocapture: 1);
1961 if (auto *C2 = dyn_cast<ConstantInt>(Val: Y))
1962 if (auto *LI = dyn_cast<LoadInst>(Val: X))
1963 if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: LI->getOperand(i_nocapture: 0)))
1964 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(LI, GEP, ICI&: Cmp, AndCst: C2))
1965 return Res;
1966
1967 if (!Cmp.isEquality())
1968 return nullptr;
1969
1970 // X & -C == -C -> X > u ~C
1971 // X & -C != -C -> X <= u ~C
1972 // iff C is a power of 2
1973 if (Cmp.getOperand(i_nocapture: 1) == Y && C.isNegatedPowerOf2()) {
1974 auto NewPred =
1975 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1976 return new ICmpInst(NewPred, X, SubOne(C: cast<Constant>(Val: Cmp.getOperand(i_nocapture: 1))));
1977 }
1978
1979 // ((zext i1 X) & Y) == 0 --> !((trunc Y) & X)
1980 // ((zext i1 X) & Y) != 0 --> ((trunc Y) & X)
1981 // ((zext i1 X) & Y) == 1 --> ((trunc Y) & X)
1982 // ((zext i1 X) & Y) != 1 --> !((trunc Y) & X)
1983 if (match(V: And, P: m_OneUse(SubPattern: m_c_And(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))), R: m_Value(V&: Y)))) &&
1984 X->getType()->isIntOrIntVectorTy(BitWidth: 1) && (C.isZero() || C.isOne())) {
1985 Value *TruncY = Builder.CreateTrunc(V: Y, DestTy: X->getType());
1986 if (C.isZero() ^ (Pred == CmpInst::ICMP_NE)) {
1987 Value *And = Builder.CreateAnd(LHS: TruncY, RHS: X);
1988 return BinaryOperator::CreateNot(Op: And);
1989 }
1990 return BinaryOperator::CreateAnd(V1: TruncY, V2: X);
1991 }
1992
1993 // (icmp eq/ne (and (shl -1, X), Y), 0)
1994 // -> (icmp eq/ne (lshr Y, X), 0)
1995 // We could technically handle any C == 0 or (C < 0 && isOdd(C)) but it seems
1996 // highly unlikely the non-zero case will ever show up in code.
1997 if (C.isZero() &&
1998 match(V: And, P: m_OneUse(SubPattern: m_c_And(L: m_OneUse(SubPattern: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))),
1999 R: m_Value(V&: Y))))) {
2000 Value *LShr = Builder.CreateLShr(LHS: Y, RHS: X);
2001 return new ICmpInst(Pred, LShr, Constant::getNullValue(Ty: LShr->getType()));
2002 }
2003
2004 // (icmp eq/ne (and (add A, Addend), Msk), C)
2005 // -> (icmp eq/ne (and A, Msk), (and (sub C, Addend), Msk))
2006 {
2007 Value *A;
2008 const APInt *Addend, *Msk;
2009 if (match(V: And, P: m_OneUse(SubPattern: m_And(L: m_OneUse(SubPattern: m_Add(L: m_Value(V&: A), R: m_APInt(Res&: Addend))),
2010 R: m_LowBitMask(V&: Msk)))) &&
2011 C.ule(RHS: *Msk)) {
2012 APInt NewComperand = (C - *Addend) & *Msk;
2013 Value *MaskA = Builder.CreateAnd(LHS: A, RHS: ConstantInt::get(Ty: A->getType(), V: *Msk));
2014 return new ICmpInst(Pred, MaskA,
2015 ConstantInt::get(Ty: MaskA->getType(), V: NewComperand));
2016 }
2017 }
2018
2019 return nullptr;
2020}
2021
2022/// Fold icmp eq/ne (or (xor/sub (X1, X2), xor/sub (X3, X4))), 0.
2023static Value *foldICmpOrXorSubChain(ICmpInst &Cmp, BinaryOperator *Or,
2024 InstCombiner::BuilderTy &Builder) {
2025 // Are we using xors or subs to bitwise check for a pair or pairs of
2026 // (in)equalities? Convert to a shorter form that has more potential to be
2027 // folded even further.
2028 // ((X1 ^/- X2) || (X3 ^/- X4)) == 0 --> (X1 == X2) && (X3 == X4)
2029 // ((X1 ^/- X2) || (X3 ^/- X4)) != 0 --> (X1 != X2) || (X3 != X4)
2030 // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) == 0 -->
2031 // (X1 == X2) && (X3 == X4) && (X5 == X6)
2032 // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) != 0 -->
2033 // (X1 != X2) || (X3 != X4) || (X5 != X6)
2034 SmallVector<std::pair<Value *, Value *>, 2> CmpValues;
2035 SmallVector<Value *, 16> WorkList(1, Or);
2036
2037 while (!WorkList.empty()) {
2038 auto MatchOrOperatorArgument = [&](Value *OrOperatorArgument) {
2039 Value *Lhs, *Rhs;
2040
2041 if (match(V: OrOperatorArgument,
2042 P: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: Lhs), R: m_Value(V&: Rhs))))) {
2043 CmpValues.emplace_back(Args&: Lhs, Args&: Rhs);
2044 return;
2045 }
2046
2047 if (match(V: OrOperatorArgument,
2048 P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: Lhs), R: m_Value(V&: Rhs))))) {
2049 CmpValues.emplace_back(Args&: Lhs, Args&: Rhs);
2050 return;
2051 }
2052
2053 WorkList.push_back(Elt: OrOperatorArgument);
2054 };
2055
2056 Value *CurrentValue = WorkList.pop_back_val();
2057 Value *OrOperatorLhs, *OrOperatorRhs;
2058
2059 if (!match(V: CurrentValue,
2060 P: m_Or(L: m_Value(V&: OrOperatorLhs), R: m_Value(V&: OrOperatorRhs)))) {
2061 return nullptr;
2062 }
2063
2064 MatchOrOperatorArgument(OrOperatorRhs);
2065 MatchOrOperatorArgument(OrOperatorLhs);
2066 }
2067
2068 ICmpInst::Predicate Pred = Cmp.getPredicate();
2069 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2070 Value *LhsCmp = Builder.CreateICmp(P: Pred, LHS: CmpValues.rbegin()->first,
2071 RHS: CmpValues.rbegin()->second);
2072
2073 for (auto It = CmpValues.rbegin() + 1; It != CmpValues.rend(); ++It) {
2074 Value *RhsCmp = Builder.CreateICmp(P: Pred, LHS: It->first, RHS: It->second);
2075 LhsCmp = Builder.CreateBinOp(Opc: BOpc, LHS: LhsCmp, RHS: RhsCmp);
2076 }
2077
2078 return LhsCmp;
2079}
2080
2081/// Fold icmp (or X, Y), C.
2082Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
2083 BinaryOperator *Or,
2084 const APInt &C) {
2085 ICmpInst::Predicate Pred = Cmp.getPredicate();
2086 if (C.isOne()) {
2087 // icmp slt signum(V) 1 --> icmp slt V, 1
2088 Value *V = nullptr;
2089 if (Pred == ICmpInst::ICMP_SLT && match(V: Or, P: m_Signum(V: m_Value(V))))
2090 return new ICmpInst(ICmpInst::ICMP_SLT, V,
2091 ConstantInt::get(Ty: V->getType(), V: 1));
2092 }
2093
2094 Value *OrOp0 = Or->getOperand(i_nocapture: 0), *OrOp1 = Or->getOperand(i_nocapture: 1);
2095
2096 // (icmp eq/ne (or disjoint x, C0), C1)
2097 // -> (icmp eq/ne x, C0^C1)
2098 if (Cmp.isEquality() && match(V: OrOp1, P: m_ImmConstant()) &&
2099 cast<PossiblyDisjointInst>(Val: Or)->isDisjoint()) {
2100 Value *NewC =
2101 Builder.CreateXor(LHS: OrOp1, RHS: ConstantInt::get(Ty: OrOp1->getType(), V: C));
2102 return new ICmpInst(Pred, OrOp0, NewC);
2103 }
2104
2105 const APInt *MaskC;
2106 if (match(V: OrOp1, P: m_APInt(Res&: MaskC)) && Cmp.isEquality()) {
2107 if (*MaskC == C && (C + 1).isPowerOf2()) {
2108 // X | C == C --> X <=u C
2109 // X | C != C --> X >u C
2110 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
2111 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
2112 return new ICmpInst(Pred, OrOp0, OrOp1);
2113 }
2114
2115 // More general: canonicalize 'equality with set bits mask' to
2116 // 'equality with clear bits mask'.
2117 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
2118 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
2119 if (Or->hasOneUse()) {
2120 Value *And = Builder.CreateAnd(LHS: OrOp0, RHS: ~(*MaskC));
2121 Constant *NewC = ConstantInt::get(Ty: Or->getType(), V: C ^ (*MaskC));
2122 return new ICmpInst(Pred, And, NewC);
2123 }
2124 }
2125
2126 // (X | (X-1)) s< 0 --> X s< 1
2127 // (X | (X-1)) s> -1 --> X s> 0
2128 Value *X;
2129 bool TrueIfSigned;
2130 if (isSignBitCheck(Pred, RHS: C, TrueIfSigned) &&
2131 match(V: Or, P: m_c_Or(L: m_Add(L: m_Value(V&: X), R: m_AllOnes()), R: m_Deferred(V: X)))) {
2132 auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT;
2133 Constant *NewC = ConstantInt::get(Ty: X->getType(), V: TrueIfSigned ? 1 : 0);
2134 return new ICmpInst(NewPred, X, NewC);
2135 }
2136
2137 const APInt *OrC;
2138 // icmp(X | OrC, C) --> icmp(X, 0)
2139 if (C.isNonNegative() && match(V: Or, P: m_Or(L: m_Value(V&: X), R: m_APInt(Res&: OrC)))) {
2140 switch (Pred) {
2141 // X | OrC s< C --> X s< 0 iff OrC s>= C s>= 0
2142 case ICmpInst::ICMP_SLT:
2143 // X | OrC s>= C --> X s>= 0 iff OrC s>= C s>= 0
2144 case ICmpInst::ICMP_SGE:
2145 if (OrC->sge(RHS: C))
2146 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType()));
2147 break;
2148 // X | OrC s<= C --> X s< 0 iff OrC s> C s>= 0
2149 case ICmpInst::ICMP_SLE:
2150 // X | OrC s> C --> X s>= 0 iff OrC s> C s>= 0
2151 case ICmpInst::ICMP_SGT:
2152 if (OrC->sgt(RHS: C))
2153 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), X,
2154 ConstantInt::getNullValue(Ty: X->getType()));
2155 break;
2156 default:
2157 break;
2158 }
2159 }
2160
2161 if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse())
2162 return nullptr;
2163
2164 Value *P, *Q;
2165 if (match(V: Or, P: m_Or(L: m_PtrToInt(Op: m_Value(V&: P)), R: m_PtrToInt(Op: m_Value(V&: Q))))) {
2166 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
2167 // -> and (icmp eq P, null), (icmp eq Q, null).
2168 Value *CmpP =
2169 Builder.CreateICmp(P: Pred, LHS: P, RHS: ConstantInt::getNullValue(Ty: P->getType()));
2170 Value *CmpQ =
2171 Builder.CreateICmp(P: Pred, LHS: Q, RHS: ConstantInt::getNullValue(Ty: Q->getType()));
2172 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2173 return BinaryOperator::Create(Op: BOpc, S1: CmpP, S2: CmpQ);
2174 }
2175
2176 if (Value *V = foldICmpOrXorSubChain(Cmp, Or, Builder))
2177 return replaceInstUsesWith(I&: Cmp, V);
2178
2179 return nullptr;
2180}
2181
2182/// Fold icmp (mul X, Y), C.
2183Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
2184 BinaryOperator *Mul,
2185 const APInt &C) {
2186 ICmpInst::Predicate Pred = Cmp.getPredicate();
2187 Type *MulTy = Mul->getType();
2188 Value *X = Mul->getOperand(i_nocapture: 0);
2189
2190 // If there's no overflow:
2191 // X * X == 0 --> X == 0
2192 // X * X != 0 --> X != 0
2193 if (Cmp.isEquality() && C.isZero() && X == Mul->getOperand(i_nocapture: 1) &&
2194 (Mul->hasNoUnsignedWrap() || Mul->hasNoSignedWrap()))
2195 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: MulTy));
2196
2197 const APInt *MulC;
2198 if (!match(V: Mul->getOperand(i_nocapture: 1), P: m_APInt(Res&: MulC)))
2199 return nullptr;
2200
2201 // If this is a test of the sign bit and the multiply is sign-preserving with
2202 // a constant operand, use the multiply LHS operand instead:
2203 // (X * +MulC) < 0 --> X < 0
2204 // (X * -MulC) < 0 --> X > 0
2205 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
2206 if (MulC->isNegative())
2207 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2208 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: MulTy));
2209 }
2210
2211 if (MulC->isZero())
2212 return nullptr;
2213
2214 // If the multiply does not wrap or the constant is odd, try to divide the
2215 // compare constant by the multiplication factor.
2216 if (Cmp.isEquality()) {
2217 // (mul nsw X, MulC) eq/ne C --> X eq/ne C /s MulC
2218 if (Mul->hasNoSignedWrap() && C.srem(RHS: *MulC).isZero()) {
2219 Constant *NewC = ConstantInt::get(Ty: MulTy, V: C.sdiv(RHS: *MulC));
2220 return new ICmpInst(Pred, X, NewC);
2221 }
2222
2223 // C % MulC == 0 is weaker than we could use if MulC is odd because it
2224 // correct to transform if MulC * N == C including overflow. I.e with i8
2225 // (icmp eq (mul X, 5), 101) -> (icmp eq X, 225) but since 101 % 5 != 0, we
2226 // miss that case.
2227 if (C.urem(RHS: *MulC).isZero()) {
2228 // (mul nuw X, MulC) eq/ne C --> X eq/ne C /u MulC
2229 // (mul X, OddC) eq/ne N * C --> X eq/ne N
2230 if ((*MulC & 1).isOne() || Mul->hasNoUnsignedWrap()) {
2231 Constant *NewC = ConstantInt::get(Ty: MulTy, V: C.udiv(RHS: *MulC));
2232 return new ICmpInst(Pred, X, NewC);
2233 }
2234 }
2235 }
2236
2237 // With a matching no-overflow guarantee, fold the constants:
2238 // (X * MulC) < C --> X < (C / MulC)
2239 // (X * MulC) > C --> X > (C / MulC)
2240 // TODO: Assert that Pred is not equal to SGE, SLE, UGE, ULE?
2241 Constant *NewC = nullptr;
2242 if (Mul->hasNoSignedWrap() && ICmpInst::isSigned(Pred)) {
2243 // MININT / -1 --> overflow.
2244 if (C.isMinSignedValue() && MulC->isAllOnes())
2245 return nullptr;
2246 if (MulC->isNegative())
2247 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2248
2249 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2250 NewC = ConstantInt::get(
2251 Ty: MulTy, V: APIntOps::RoundingSDiv(A: C, B: *MulC, RM: APInt::Rounding::UP));
2252 } else {
2253 assert((Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_SGT) &&
2254 "Unexpected predicate");
2255 NewC = ConstantInt::get(
2256 Ty: MulTy, V: APIntOps::RoundingSDiv(A: C, B: *MulC, RM: APInt::Rounding::DOWN));
2257 }
2258 } else if (Mul->hasNoUnsignedWrap() && ICmpInst::isUnsigned(Pred)) {
2259 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) {
2260 NewC = ConstantInt::get(
2261 Ty: MulTy, V: APIntOps::RoundingUDiv(A: C, B: *MulC, RM: APInt::Rounding::UP));
2262 } else {
2263 assert((Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
2264 "Unexpected predicate");
2265 NewC = ConstantInt::get(
2266 Ty: MulTy, V: APIntOps::RoundingUDiv(A: C, B: *MulC, RM: APInt::Rounding::DOWN));
2267 }
2268 }
2269
2270 return NewC ? new ICmpInst(Pred, X, NewC) : nullptr;
2271}
2272
2273/// Fold icmp (shl nuw C2, Y), C.
2274static Instruction *foldICmpShlLHSC(ICmpInst &Cmp, Instruction *Shl,
2275 const APInt &C) {
2276 Value *Y;
2277 const APInt *C2;
2278 if (!match(V: Shl, P: m_NUWShl(L: m_APInt(Res&: C2), R: m_Value(V&: Y))))
2279 return nullptr;
2280
2281 Type *ShiftType = Shl->getType();
2282 unsigned TypeBits = C.getBitWidth();
2283 ICmpInst::Predicate Pred = Cmp.getPredicate();
2284 if (Cmp.isUnsigned()) {
2285 if (C2->isZero() || C2->ugt(RHS: C))
2286 return nullptr;
2287 APInt Div, Rem;
2288 APInt::udivrem(LHS: C, RHS: *C2, Quotient&: Div, Remainder&: Rem);
2289 bool CIsPowerOf2 = Rem.isZero() && Div.isPowerOf2();
2290
2291 // (1 << Y) pred C -> Y pred Log2(C)
2292 if (!CIsPowerOf2) {
2293 // (1 << Y) < 30 -> Y <= 4
2294 // (1 << Y) <= 30 -> Y <= 4
2295 // (1 << Y) >= 30 -> Y > 4
2296 // (1 << Y) > 30 -> Y > 4
2297 if (Pred == ICmpInst::ICMP_ULT)
2298 Pred = ICmpInst::ICMP_ULE;
2299 else if (Pred == ICmpInst::ICMP_UGE)
2300 Pred = ICmpInst::ICMP_UGT;
2301 }
2302
2303 unsigned CLog2 = Div.logBase2();
2304 return new ICmpInst(Pred, Y, ConstantInt::get(Ty: ShiftType, V: CLog2));
2305 } else if (Cmp.isSigned() && C2->isOne()) {
2306 Constant *BitWidthMinusOne = ConstantInt::get(Ty: ShiftType, V: TypeBits - 1);
2307 // (1 << Y) > 0 -> Y != 31
2308 // (1 << Y) > C -> Y != 31 if C is negative.
2309 if (Pred == ICmpInst::ICMP_SGT && C.sle(RHS: 0))
2310 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2311
2312 // (1 << Y) < 0 -> Y == 31
2313 // (1 << Y) < 1 -> Y == 31
2314 // (1 << Y) < C -> Y == 31 if C is negative and not signed min.
2315 // Exclude signed min by subtracting 1 and lower the upper bound to 0.
2316 if (Pred == ICmpInst::ICMP_SLT && (C - 1).sle(RHS: 0))
2317 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2318 }
2319
2320 return nullptr;
2321}
2322
2323/// Fold icmp (shl X, Y), C.
2324Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2325 BinaryOperator *Shl,
2326 const APInt &C) {
2327 const APInt *ShiftVal;
2328 if (Cmp.isEquality() && match(V: Shl->getOperand(i_nocapture: 0), P: m_APInt(Res&: ShiftVal)))
2329 return foldICmpShlConstConst(I&: Cmp, A: Shl->getOperand(i_nocapture: 1), AP1: C, AP2: *ShiftVal);
2330
2331 ICmpInst::Predicate Pred = Cmp.getPredicate();
2332 // (icmp pred (shl nuw&nsw X, Y), Csle0)
2333 // -> (icmp pred X, Csle0)
2334 //
2335 // The idea is the nuw/nsw essentially freeze the sign bit for the shift op
2336 // so X's must be what is used.
2337 if (C.sle(RHS: 0) && Shl->hasNoUnsignedWrap() && Shl->hasNoSignedWrap())
2338 return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1));
2339
2340 // (icmp eq/ne (shl nuw|nsw X, Y), 0)
2341 // -> (icmp eq/ne X, 0)
2342 if (ICmpInst::isEquality(P: Pred) && C.isZero() &&
2343 (Shl->hasNoUnsignedWrap() || Shl->hasNoSignedWrap()))
2344 return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1));
2345
2346 // (icmp slt (shl nsw X, Y), 0/1)
2347 // -> (icmp slt X, 0/1)
2348 // (icmp sgt (shl nsw X, Y), 0/-1)
2349 // -> (icmp sgt X, 0/-1)
2350 //
2351 // NB: sge/sle with a constant will canonicalize to sgt/slt.
2352 if (Shl->hasNoSignedWrap() &&
2353 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT))
2354 if (C.isZero() || (Pred == ICmpInst::ICMP_SGT ? C.isAllOnes() : C.isOne()))
2355 return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1));
2356
2357 const APInt *ShiftAmt;
2358 if (!match(V: Shl->getOperand(i_nocapture: 1), P: m_APInt(Res&: ShiftAmt)))
2359 return foldICmpShlLHSC(Cmp, Shl, C);
2360
2361 // Check that the shift amount is in range. If not, don't perform undefined
2362 // shifts. When the shift is visited, it will be simplified.
2363 unsigned TypeBits = C.getBitWidth();
2364 if (ShiftAmt->uge(RHS: TypeBits))
2365 return nullptr;
2366
2367 Value *X = Shl->getOperand(i_nocapture: 0);
2368 Type *ShType = Shl->getType();
2369
2370 // NSW guarantees that we are only shifting out sign bits from the high bits,
2371 // so we can ASHR the compare constant without needing a mask and eliminate
2372 // the shift.
2373 if (Shl->hasNoSignedWrap()) {
2374 if (Pred == ICmpInst::ICMP_SGT) {
2375 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2376 APInt ShiftedC = C.ashr(ShiftAmt: *ShiftAmt);
2377 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2378 }
2379 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2380 C.ashr(ShiftAmt: *ShiftAmt).shl(ShiftAmt: *ShiftAmt) == C) {
2381 APInt ShiftedC = C.ashr(ShiftAmt: *ShiftAmt);
2382 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2383 }
2384 if (Pred == ICmpInst::ICMP_SLT) {
2385 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2386 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2387 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2388 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2389 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2390 APInt ShiftedC = (C - 1).ashr(ShiftAmt: *ShiftAmt) + 1;
2391 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2392 }
2393 }
2394
2395 // NUW guarantees that we are only shifting out zero bits from the high bits,
2396 // so we can LSHR the compare constant without needing a mask and eliminate
2397 // the shift.
2398 if (Shl->hasNoUnsignedWrap()) {
2399 if (Pred == ICmpInst::ICMP_UGT) {
2400 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2401 APInt ShiftedC = C.lshr(ShiftAmt: *ShiftAmt);
2402 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2403 }
2404 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2405 C.lshr(ShiftAmt: *ShiftAmt).shl(ShiftAmt: *ShiftAmt) == C) {
2406 APInt ShiftedC = C.lshr(ShiftAmt: *ShiftAmt);
2407 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2408 }
2409 if (Pred == ICmpInst::ICMP_ULT) {
2410 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2411 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2412 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2413 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2414 assert(C.ugt(0) && "ult 0 should have been eliminated");
2415 APInt ShiftedC = (C - 1).lshr(ShiftAmt: *ShiftAmt) + 1;
2416 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2417 }
2418 }
2419
2420 if (Cmp.isEquality() && Shl->hasOneUse()) {
2421 // Strength-reduce the shift into an 'and'.
2422 Constant *Mask = ConstantInt::get(
2423 Ty: ShType,
2424 V: APInt::getLowBitsSet(numBits: TypeBits, loBitsSet: TypeBits - ShiftAmt->getZExtValue()));
2425 Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shl->getName() + ".mask");
2426 Constant *LShrC = ConstantInt::get(Ty: ShType, V: C.lshr(ShiftAmt: *ShiftAmt));
2427 return new ICmpInst(Pred, And, LShrC);
2428 }
2429
2430 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2431 bool TrueIfSigned = false;
2432 if (Shl->hasOneUse() && isSignBitCheck(Pred, RHS: C, TrueIfSigned)) {
2433 // (X << 31) <s 0 --> (X & 1) != 0
2434 Constant *Mask = ConstantInt::get(
2435 Ty: ShType,
2436 V: APInt::getOneBitSet(numBits: TypeBits, BitNo: TypeBits - ShiftAmt->getZExtValue() - 1));
2437 Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shl->getName() + ".mask");
2438 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2439 And, Constant::getNullValue(Ty: ShType));
2440 }
2441
2442 // Simplify 'shl' inequality test into 'and' equality test.
2443 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2444 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2445 if ((C + 1).isPowerOf2() &&
2446 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2447 Value *And = Builder.CreateAnd(LHS: X, RHS: (~C).lshr(shiftAmt: ShiftAmt->getZExtValue()));
2448 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2449 : ICmpInst::ICMP_NE,
2450 And, Constant::getNullValue(Ty: ShType));
2451 }
2452 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2453 if (C.isPowerOf2() &&
2454 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2455 Value *And =
2456 Builder.CreateAnd(LHS: X, RHS: (~(C - 1)).lshr(shiftAmt: ShiftAmt->getZExtValue()));
2457 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2458 : ICmpInst::ICMP_NE,
2459 And, Constant::getNullValue(Ty: ShType));
2460 }
2461 }
2462
2463 // Transform (icmp pred iM (shl iM %v, N), C)
2464 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2465 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2466 // This enables us to get rid of the shift in favor of a trunc that may be
2467 // free on the target. It has the additional benefit of comparing to a
2468 // smaller constant that may be more target-friendly.
2469 unsigned Amt = ShiftAmt->getLimitedValue(Limit: TypeBits - 1);
2470 if (Shl->hasOneUse() && Amt != 0 &&
2471 shouldChangeType(FromBitWidth: ShType->getScalarSizeInBits(), ToBitWidth: TypeBits - Amt)) {
2472 ICmpInst::Predicate CmpPred = Pred;
2473 APInt RHSC = C;
2474
2475 if (RHSC.countr_zero() < Amt && ICmpInst::isStrictPredicate(predicate: CmpPred)) {
2476 // Try the flipped strictness predicate.
2477 // e.g.:
2478 // icmp ult i64 (shl X, 32), 8589934593 ->
2479 // icmp ule i64 (shl X, 32), 8589934592 ->
2480 // icmp ule i32 (trunc X, i32), 2 ->
2481 // icmp ult i32 (trunc X, i32), 3
2482 if (auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(
2483 Pred, C: ConstantInt::get(Context&: ShType->getContext(), V: C))) {
2484 CmpPred = FlippedStrictness->first;
2485 RHSC = cast<ConstantInt>(Val: FlippedStrictness->second)->getValue();
2486 }
2487 }
2488
2489 if (RHSC.countr_zero() >= Amt) {
2490 Type *TruncTy = ShType->getWithNewBitWidth(NewBitWidth: TypeBits - Amt);
2491 Constant *NewC =
2492 ConstantInt::get(Ty: TruncTy, V: RHSC.ashr(ShiftAmt: *ShiftAmt).trunc(width: TypeBits - Amt));
2493 return new ICmpInst(CmpPred,
2494 Builder.CreateTrunc(V: X, DestTy: TruncTy, Name: "", /*IsNUW=*/false,
2495 IsNSW: Shl->hasNoSignedWrap()),
2496 NewC);
2497 }
2498 }
2499
2500 return nullptr;
2501}
2502
2503/// Fold icmp ({al}shr X, Y), C.
2504Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2505 BinaryOperator *Shr,
2506 const APInt &C) {
2507 // An exact shr only shifts out zero bits, so:
2508 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2509 Value *X = Shr->getOperand(i_nocapture: 0);
2510 CmpInst::Predicate Pred = Cmp.getPredicate();
2511 if (Cmp.isEquality() && Shr->isExact() && C.isZero())
2512 return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1));
2513
2514 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2515 const APInt *ShiftValC;
2516 if (match(V: X, P: m_APInt(Res&: ShiftValC))) {
2517 if (Cmp.isEquality())
2518 return foldICmpShrConstConst(I&: Cmp, A: Shr->getOperand(i_nocapture: 1), AP1: C, AP2: *ShiftValC);
2519
2520 // (ShiftValC >> Y) >s -1 --> Y != 0 with ShiftValC < 0
2521 // (ShiftValC >> Y) <s 0 --> Y == 0 with ShiftValC < 0
2522 bool TrueIfSigned;
2523 if (!IsAShr && ShiftValC->isNegative() &&
2524 isSignBitCheck(Pred, RHS: C, TrueIfSigned))
2525 return new ICmpInst(TrueIfSigned ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE,
2526 Shr->getOperand(i_nocapture: 1),
2527 ConstantInt::getNullValue(Ty: X->getType()));
2528
2529 // If the shifted constant is a power-of-2, test the shift amount directly:
2530 // (ShiftValC >> Y) >u C --> X <u (LZ(C) - LZ(ShiftValC))
2531 // (ShiftValC >> Y) <u C --> X >=u (LZ(C-1) - LZ(ShiftValC))
2532 if (!IsAShr && ShiftValC->isPowerOf2() &&
2533 (Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_ULT)) {
2534 bool IsUGT = Pred == CmpInst::ICMP_UGT;
2535 assert(ShiftValC->uge(C) && "Expected simplify of compare");
2536 assert((IsUGT || !C.isZero()) && "Expected X u< 0 to simplify");
2537
2538 unsigned CmpLZ = IsUGT ? C.countl_zero() : (C - 1).countl_zero();
2539 unsigned ShiftLZ = ShiftValC->countl_zero();
2540 Constant *NewC = ConstantInt::get(Ty: Shr->getType(), V: CmpLZ - ShiftLZ);
2541 auto NewPred = IsUGT ? CmpInst::ICMP_ULT : CmpInst::ICMP_UGE;
2542 return new ICmpInst(NewPred, Shr->getOperand(i_nocapture: 1), NewC);
2543 }
2544 }
2545
2546 const APInt *ShiftAmtC;
2547 if (!match(V: Shr->getOperand(i_nocapture: 1), P: m_APInt(Res&: ShiftAmtC)))
2548 return nullptr;
2549
2550 // Check that the shift amount is in range. If not, don't perform undefined
2551 // shifts. When the shift is visited it will be simplified.
2552 unsigned TypeBits = C.getBitWidth();
2553 unsigned ShAmtVal = ShiftAmtC->getLimitedValue(Limit: TypeBits);
2554 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2555 return nullptr;
2556
2557 bool IsExact = Shr->isExact();
2558 Type *ShrTy = Shr->getType();
2559 // TODO: If we could guarantee that InstSimplify would handle all of the
2560 // constant-value-based preconditions in the folds below, then we could assert
2561 // those conditions rather than checking them. This is difficult because of
2562 // undef/poison (PR34838).
2563 if (IsAShr && Shr->hasOneUse()) {
2564 if (IsExact && (Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) &&
2565 (C - 1).isPowerOf2() && C.countLeadingZeros() > ShAmtVal) {
2566 // When C - 1 is a power of two and the transform can be legally
2567 // performed, prefer this form so the produced constant is close to a
2568 // power of two.
2569 // icmp slt/ult (ashr exact X, ShAmtC), C
2570 // --> icmp slt/ult X, (C - 1) << ShAmtC) + 1
2571 APInt ShiftedC = (C - 1).shl(shiftAmt: ShAmtVal) + 1;
2572 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2573 }
2574 if (IsExact || Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) {
2575 // When ShAmtC can be shifted losslessly:
2576 // icmp PRED (ashr exact X, ShAmtC), C --> icmp PRED X, (C << ShAmtC)
2577 // icmp slt/ult (ashr X, ShAmtC), C --> icmp slt/ult X, (C << ShAmtC)
2578 APInt ShiftedC = C.shl(shiftAmt: ShAmtVal);
2579 if (ShiftedC.ashr(ShiftAmt: ShAmtVal) == C)
2580 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2581 }
2582 if (Pred == CmpInst::ICMP_SGT) {
2583 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2584 APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1;
2585 if (!C.isMaxSignedValue() && !(C + 1).shl(shiftAmt: ShAmtVal).isMinSignedValue() &&
2586 (ShiftedC + 1).ashr(ShiftAmt: ShAmtVal) == (C + 1))
2587 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2588 }
2589 if (Pred == CmpInst::ICMP_UGT) {
2590 // icmp ugt (ashr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2591 // 'C + 1 << ShAmtC' can overflow as a signed number, so the 2nd
2592 // clause accounts for that pattern.
2593 APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1;
2594 if ((ShiftedC + 1).ashr(ShiftAmt: ShAmtVal) == (C + 1) ||
2595 (C + 1).shl(shiftAmt: ShAmtVal).isMinSignedValue())
2596 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2597 }
2598
2599 // If the compare constant has significant bits above the lowest sign-bit,
2600 // then convert an unsigned cmp to a test of the sign-bit:
2601 // (ashr X, ShiftC) u> C --> X s< 0
2602 // (ashr X, ShiftC) u< C --> X s> -1
2603 if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2604 if (Pred == CmpInst::ICMP_UGT) {
2605 return new ICmpInst(CmpInst::ICMP_SLT, X,
2606 ConstantInt::getNullValue(Ty: ShrTy));
2607 }
2608 if (Pred == CmpInst::ICMP_ULT) {
2609 return new ICmpInst(CmpInst::ICMP_SGT, X,
2610 ConstantInt::getAllOnesValue(Ty: ShrTy));
2611 }
2612 }
2613 } else if (!IsAShr) {
2614 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2615 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2616 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2617 APInt ShiftedC = C.shl(shiftAmt: ShAmtVal);
2618 if (ShiftedC.lshr(shiftAmt: ShAmtVal) == C)
2619 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2620 }
2621 if (Pred == CmpInst::ICMP_UGT) {
2622 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2623 APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1;
2624 if ((ShiftedC + 1).lshr(shiftAmt: ShAmtVal) == (C + 1))
2625 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2626 }
2627 }
2628
2629 if (!Cmp.isEquality())
2630 return nullptr;
2631
2632 // Handle equality comparisons of shift-by-constant.
2633
2634 // If the comparison constant changes with the shift, the comparison cannot
2635 // succeed (bits of the comparison constant cannot match the shifted value).
2636 // This should be known by InstSimplify and already be folded to true/false.
2637 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2638 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2639 "Expected icmp+shr simplify did not occur.");
2640
2641 // If the bits shifted out are known zero, compare the unshifted value:
2642 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2643 if (Shr->isExact())
2644 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: C << ShAmtVal));
2645
2646 if (Shr->hasOneUse()) {
2647 // Canonicalize the shift into an 'and':
2648 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2649 APInt Val(APInt::getHighBitsSet(numBits: TypeBits, hiBitsSet: TypeBits - ShAmtVal));
2650 Constant *Mask = ConstantInt::get(Ty: ShrTy, V: Val);
2651 Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shr->getName() + ".mask");
2652 return new ICmpInst(Pred, And, ConstantInt::get(Ty: ShrTy, V: C << ShAmtVal));
2653 }
2654
2655 return nullptr;
2656}
2657
2658Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2659 BinaryOperator *SRem,
2660 const APInt &C) {
2661 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2662 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT) {
2663 // Canonicalize unsigned predicates to signed:
2664 // (X s% DivisorC) u> C -> (X s% DivisorC) s< 0
2665 // iff (C s< 0 ? ~C : C) u>= abs(DivisorC)-1
2666 // (X s% DivisorC) u< C+1 -> (X s% DivisorC) s> -1
2667 // iff (C+1 s< 0 ? ~C : C) u>= abs(DivisorC)-1
2668
2669 const APInt *DivisorC;
2670 if (!match(V: SRem->getOperand(i_nocapture: 1), P: m_APInt(Res&: DivisorC)))
2671 return nullptr;
2672 if (DivisorC->isZero())
2673 return nullptr;
2674
2675 APInt NormalizedC = C;
2676 if (Pred == ICmpInst::ICMP_ULT) {
2677 assert(!NormalizedC.isZero() &&
2678 "ult X, 0 should have been simplified already.");
2679 --NormalizedC;
2680 }
2681 if (C.isNegative())
2682 NormalizedC.flipAllBits();
2683 if (!NormalizedC.uge(RHS: DivisorC->abs() - 1))
2684 return nullptr;
2685
2686 Type *Ty = SRem->getType();
2687 if (Pred == ICmpInst::ICMP_UGT)
2688 return new ICmpInst(ICmpInst::ICMP_SLT, SRem,
2689 ConstantInt::getNullValue(Ty));
2690 return new ICmpInst(ICmpInst::ICMP_SGT, SRem,
2691 ConstantInt::getAllOnesValue(Ty));
2692 }
2693 // Match an 'is positive' or 'is negative' comparison of remainder by a
2694 // constant power-of-2 value:
2695 // (X % pow2C) sgt/slt 0
2696 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT &&
2697 Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2698 return nullptr;
2699
2700 // TODO: The one-use check is standard because we do not typically want to
2701 // create longer instruction sequences, but this might be a special-case
2702 // because srem is not good for analysis or codegen.
2703 if (!SRem->hasOneUse())
2704 return nullptr;
2705
2706 const APInt *DivisorC;
2707 if (!match(V: SRem->getOperand(i_nocapture: 1), P: m_Power2(V&: DivisorC)))
2708 return nullptr;
2709
2710 // For cmp_sgt/cmp_slt only zero valued C is handled.
2711 // For cmp_eq/cmp_ne only positive valued C is handled.
2712 if (((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT) &&
2713 !C.isZero()) ||
2714 ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2715 !C.isStrictlyPositive()))
2716 return nullptr;
2717
2718 // Mask off the sign bit and the modulo bits (low-bits).
2719 Type *Ty = SRem->getType();
2720 APInt SignMask = APInt::getSignMask(BitWidth: Ty->getScalarSizeInBits());
2721 Constant *MaskC = ConstantInt::get(Ty, V: SignMask | (*DivisorC - 1));
2722 Value *And = Builder.CreateAnd(LHS: SRem->getOperand(i_nocapture: 0), RHS: MaskC);
2723
2724 if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)
2725 return new ICmpInst(Pred, And, ConstantInt::get(Ty, V: C));
2726
2727 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2728 // bit is set. Example:
2729 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2730 if (Pred == ICmpInst::ICMP_SGT)
2731 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2732
2733 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2734 // bit is set. Example:
2735 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2736 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, V: SignMask));
2737}
2738
2739/// Fold icmp (udiv X, Y), C.
2740Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2741 BinaryOperator *UDiv,
2742 const APInt &C) {
2743 ICmpInst::Predicate Pred = Cmp.getPredicate();
2744 Value *X = UDiv->getOperand(i_nocapture: 0);
2745 Value *Y = UDiv->getOperand(i_nocapture: 1);
2746 Type *Ty = UDiv->getType();
2747
2748 const APInt *C2;
2749 if (!match(V: X, P: m_APInt(Res&: C2)))
2750 return nullptr;
2751
2752 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2753
2754 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2755 if (Pred == ICmpInst::ICMP_UGT) {
2756 assert(!C.isMaxValue() &&
2757 "icmp ugt X, UINT_MAX should have been simplified already.");
2758 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2759 ConstantInt::get(Ty, V: C2->udiv(RHS: C + 1)));
2760 }
2761
2762 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2763 if (Pred == ICmpInst::ICMP_ULT) {
2764 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2765 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2766 ConstantInt::get(Ty, V: C2->udiv(RHS: C)));
2767 }
2768
2769 return nullptr;
2770}
2771
2772/// Fold icmp ({su}div X, Y), C.
2773Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2774 BinaryOperator *Div,
2775 const APInt &C) {
2776 ICmpInst::Predicate Pred = Cmp.getPredicate();
2777 Value *X = Div->getOperand(i_nocapture: 0);
2778 Value *Y = Div->getOperand(i_nocapture: 1);
2779 Type *Ty = Div->getType();
2780 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2781
2782 // If unsigned division and the compare constant is bigger than
2783 // UMAX/2 (negative), there's only one pair of values that satisfies an
2784 // equality check, so eliminate the division:
2785 // (X u/ Y) == C --> (X == C) && (Y == 1)
2786 // (X u/ Y) != C --> (X != C) || (Y != 1)
2787 // Similarly, if signed division and the compare constant is exactly SMIN:
2788 // (X s/ Y) == SMIN --> (X == SMIN) && (Y == 1)
2789 // (X s/ Y) != SMIN --> (X != SMIN) || (Y != 1)
2790 if (Cmp.isEquality() && Div->hasOneUse() && C.isSignBitSet() &&
2791 (!DivIsSigned || C.isMinSignedValue())) {
2792 Value *XBig = Builder.CreateICmp(P: Pred, LHS: X, RHS: ConstantInt::get(Ty, V: C));
2793 Value *YOne = Builder.CreateICmp(P: Pred, LHS: Y, RHS: ConstantInt::get(Ty, V: 1));
2794 auto Logic = Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2795 return BinaryOperator::Create(Op: Logic, S1: XBig, S2: YOne);
2796 }
2797
2798 // Fold: icmp pred ([us]div X, C2), C -> range test
2799 // Fold this div into the comparison, producing a range check.
2800 // Determine, based on the divide type, what the range is being
2801 // checked. If there is an overflow on the low or high side, remember
2802 // it, otherwise compute the range [low, hi) bounding the new value.
2803 // See: InsertRangeTest above for the kinds of replacements possible.
2804 const APInt *C2;
2805 if (!match(V: Y, P: m_APInt(Res&: C2)))
2806 return nullptr;
2807
2808 // FIXME: If the operand types don't match the type of the divide
2809 // then don't attempt this transform. The code below doesn't have the
2810 // logic to deal with a signed divide and an unsigned compare (and
2811 // vice versa). This is because (x /s C2) <s C produces different
2812 // results than (x /s C2) <u C or (x /u C2) <s C or even
2813 // (x /u C2) <u C. Simply casting the operands and result won't
2814 // work. :( The if statement below tests that condition and bails
2815 // if it finds it.
2816 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2817 return nullptr;
2818
2819 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2820 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2821 // division-by-constant cases should be present, we can not assert that they
2822 // have happened before we reach this icmp instruction.
2823 if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes()))
2824 return nullptr;
2825
2826 // Compute Prod = C * C2. We are essentially solving an equation of
2827 // form X / C2 = C. We solve for X by multiplying C2 and C.
2828 // By solving for X, we can turn this into a range check instead of computing
2829 // a divide.
2830 APInt Prod = C * *C2;
2831
2832 // Determine if the product overflows by seeing if the product is not equal to
2833 // the divide. Make sure we do the same kind of divide as in the LHS
2834 // instruction that we're folding.
2835 bool ProdOV = (DivIsSigned ? Prod.sdiv(RHS: *C2) : Prod.udiv(RHS: *C2)) != C;
2836
2837 // If the division is known to be exact, then there is no remainder from the
2838 // divide, so the covered range size is unit, otherwise it is the divisor.
2839 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2840
2841 // Figure out the interval that is being checked. For example, a comparison
2842 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2843 // Compute this interval based on the constants involved and the signedness of
2844 // the compare/divide. This computes a half-open interval, keeping track of
2845 // whether either value in the interval overflows. After analysis each
2846 // overflow variable is set to 0 if it's corresponding bound variable is valid
2847 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2848 int LoOverflow = 0, HiOverflow = 0;
2849 APInt LoBound, HiBound;
2850
2851 if (!DivIsSigned) { // udiv
2852 // e.g. X/5 op 3 --> [15, 20)
2853 LoBound = Prod;
2854 HiOverflow = LoOverflow = ProdOV;
2855 if (!HiOverflow) {
2856 // If this is not an exact divide, then many values in the range collapse
2857 // to the same result value.
2858 HiOverflow = addWithOverflow(Result&: HiBound, In1: LoBound, In2: RangeSize, IsSigned: false);
2859 }
2860 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2861 if (C.isZero()) { // (X / pos) op 0
2862 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2863 LoBound = -(RangeSize - 1);
2864 HiBound = RangeSize;
2865 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2866 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2867 HiOverflow = LoOverflow = ProdOV;
2868 if (!HiOverflow)
2869 HiOverflow = addWithOverflow(Result&: HiBound, In1: Prod, In2: RangeSize, IsSigned: true);
2870 } else { // (X / pos) op neg
2871 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2872 HiBound = Prod + 1;
2873 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2874 if (!LoOverflow) {
2875 APInt DivNeg = -RangeSize;
2876 LoOverflow = addWithOverflow(Result&: LoBound, In1: HiBound, In2: DivNeg, IsSigned: true) ? -1 : 0;
2877 }
2878 }
2879 } else if (C2->isNegative()) { // Divisor is < 0.
2880 if (Div->isExact())
2881 RangeSize.negate();
2882 if (C.isZero()) { // (X / neg) op 0
2883 // e.g. X/-5 op 0 --> [-4, 5)
2884 LoBound = RangeSize + 1;
2885 HiBound = -RangeSize;
2886 if (HiBound == *C2) { // -INTMIN = INTMIN
2887 HiOverflow = 1; // [INTMIN+1, overflow)
2888 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2889 }
2890 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2891 // e.g. X/-5 op 3 --> [-19, -14)
2892 HiBound = Prod + 1;
2893 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2894 if (!LoOverflow)
2895 LoOverflow =
2896 addWithOverflow(Result&: LoBound, In1: HiBound, In2: RangeSize, IsSigned: true) ? -1 : 0;
2897 } else { // (X / neg) op neg
2898 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2899 LoOverflow = HiOverflow = ProdOV;
2900 if (!HiOverflow)
2901 HiOverflow = subWithOverflow(Result&: HiBound, In1: Prod, In2: RangeSize, IsSigned: true);
2902 }
2903
2904 // Dividing by a negative swaps the condition. LT <-> GT
2905 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2906 }
2907
2908 switch (Pred) {
2909 default:
2910 llvm_unreachable("Unhandled icmp predicate!");
2911 case ICmpInst::ICMP_EQ:
2912 if (LoOverflow && HiOverflow)
2913 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
2914 if (HiOverflow)
2915 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
2916 X, ConstantInt::get(Ty, V: LoBound));
2917 if (LoOverflow)
2918 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
2919 X, ConstantInt::get(Ty, V: HiBound));
2920 return replaceInstUsesWith(
2921 I&: Cmp, V: insertRangeTest(V: X, Lo: LoBound, Hi: HiBound, isSigned: DivIsSigned, Inside: true));
2922 case ICmpInst::ICMP_NE:
2923 if (LoOverflow && HiOverflow)
2924 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
2925 if (HiOverflow)
2926 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
2927 X, ConstantInt::get(Ty, V: LoBound));
2928 if (LoOverflow)
2929 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
2930 X, ConstantInt::get(Ty, V: HiBound));
2931 return replaceInstUsesWith(
2932 I&: Cmp, V: insertRangeTest(V: X, Lo: LoBound, Hi: HiBound, isSigned: DivIsSigned, Inside: false));
2933 case ICmpInst::ICMP_ULT:
2934 case ICmpInst::ICMP_SLT:
2935 if (LoOverflow == +1) // Low bound is greater than input range.
2936 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
2937 if (LoOverflow == -1) // Low bound is less than input range.
2938 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
2939 return new ICmpInst(Pred, X, ConstantInt::get(Ty, V: LoBound));
2940 case ICmpInst::ICMP_UGT:
2941 case ICmpInst::ICMP_SGT:
2942 if (HiOverflow == +1) // High bound greater than input range.
2943 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
2944 if (HiOverflow == -1) // High bound less than input range.
2945 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
2946 if (Pred == ICmpInst::ICMP_UGT)
2947 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, V: HiBound));
2948 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, V: HiBound));
2949 }
2950
2951 return nullptr;
2952}
2953
2954/// Fold icmp (sub X, Y), C.
2955Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2956 BinaryOperator *Sub,
2957 const APInt &C) {
2958 Value *X = Sub->getOperand(i_nocapture: 0), *Y = Sub->getOperand(i_nocapture: 1);
2959 ICmpInst::Predicate Pred = Cmp.getPredicate();
2960 Type *Ty = Sub->getType();
2961
2962 // (SubC - Y) == C) --> Y == (SubC - C)
2963 // (SubC - Y) != C) --> Y != (SubC - C)
2964 Constant *SubC;
2965 if (Cmp.isEquality() && match(V: X, P: m_ImmConstant(C&: SubC))) {
2966 return new ICmpInst(Pred, Y,
2967 ConstantExpr::getSub(C1: SubC, C2: ConstantInt::get(Ty, V: C)));
2968 }
2969
2970 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2971 const APInt *C2;
2972 APInt SubResult;
2973 ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate();
2974 bool HasNSW = Sub->hasNoSignedWrap();
2975 bool HasNUW = Sub->hasNoUnsignedWrap();
2976 if (match(V: X, P: m_APInt(Res&: C2)) &&
2977 ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) &&
2978 !subWithOverflow(Result&: SubResult, In1: *C2, In2: C, IsSigned: Cmp.isSigned()))
2979 return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, V: SubResult));
2980
2981 // X - Y == 0 --> X == Y.
2982 // X - Y != 0 --> X != Y.
2983 // TODO: We allow this with multiple uses as long as the other uses are not
2984 // in phis. The phi use check is guarding against a codegen regression
2985 // for a loop test. If the backend could undo this (and possibly
2986 // subsequent transforms), we would not need this hack.
2987 if (Cmp.isEquality() && C.isZero() &&
2988 none_of(Range: (Sub->users()), P: [](const User *U) { return isa<PHINode>(Val: U); }))
2989 return new ICmpInst(Pred, X, Y);
2990
2991 // The following transforms are only worth it if the only user of the subtract
2992 // is the icmp.
2993 // TODO: This is an artificial restriction for all of the transforms below
2994 // that only need a single replacement icmp. Can these use the phi test
2995 // like the transform above here?
2996 if (!Sub->hasOneUse())
2997 return nullptr;
2998
2999 if (Sub->hasNoSignedWrap()) {
3000 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
3001 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
3002 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3003
3004 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
3005 if (Pred == ICmpInst::ICMP_SGT && C.isZero())
3006 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3007
3008 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
3009 if (Pred == ICmpInst::ICMP_SLT && C.isZero())
3010 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3011
3012 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
3013 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
3014 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3015 }
3016
3017 if (!match(V: X, P: m_APInt(Res&: C2)))
3018 return nullptr;
3019
3020 // C2 - Y <u C -> (Y | (C - 1)) == C2
3021 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
3022 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
3023 (*C2 & (C - 1)) == (C - 1))
3024 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(LHS: Y, RHS: C - 1), X);
3025
3026 // C2 - Y >u C -> (Y | C) != C2
3027 // iff C2 & C == C and C + 1 is a power of 2
3028 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
3029 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(LHS: Y, RHS: C), X);
3030
3031 // We have handled special cases that reduce.
3032 // Canonicalize any remaining sub to add as:
3033 // (C2 - Y) > C --> (Y + ~C2) < ~C
3034 Value *Add = Builder.CreateAdd(LHS: Y, RHS: ConstantInt::get(Ty, V: ~(*C2)), Name: "notsub",
3035 HasNUW, HasNSW);
3036 return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, V: ~C));
3037}
3038
3039static Value *createLogicFromTable(const std::bitset<4> &Table, Value *Op0,
3040 Value *Op1, IRBuilderBase &Builder,
3041 bool HasOneUse) {
3042 auto FoldConstant = [&](bool Val) {
3043 Constant *Res = Val ? Builder.getTrue() : Builder.getFalse();
3044 if (Op0->getType()->isVectorTy())
3045 Res = ConstantVector::getSplat(
3046 EC: cast<VectorType>(Val: Op0->getType())->getElementCount(), Elt: Res);
3047 return Res;
3048 };
3049
3050 switch (Table.to_ulong()) {
3051 case 0: // 0 0 0 0
3052 return FoldConstant(false);
3053 case 1: // 0 0 0 1
3054 return HasOneUse ? Builder.CreateNot(V: Builder.CreateOr(LHS: Op0, RHS: Op1)) : nullptr;
3055 case 2: // 0 0 1 0
3056 return HasOneUse ? Builder.CreateAnd(LHS: Builder.CreateNot(V: Op0), RHS: Op1) : nullptr;
3057 case 3: // 0 0 1 1
3058 return Builder.CreateNot(V: Op0);
3059 case 4: // 0 1 0 0
3060 return HasOneUse ? Builder.CreateAnd(LHS: Op0, RHS: Builder.CreateNot(V: Op1)) : nullptr;
3061 case 5: // 0 1 0 1
3062 return Builder.CreateNot(V: Op1);
3063 case 6: // 0 1 1 0
3064 return Builder.CreateXor(LHS: Op0, RHS: Op1);
3065 case 7: // 0 1 1 1
3066 return HasOneUse ? Builder.CreateNot(V: Builder.CreateAnd(LHS: Op0, RHS: Op1)) : nullptr;
3067 case 8: // 1 0 0 0
3068 return Builder.CreateAnd(LHS: Op0, RHS: Op1);
3069 case 9: // 1 0 0 1
3070 return HasOneUse ? Builder.CreateNot(V: Builder.CreateXor(LHS: Op0, RHS: Op1)) : nullptr;
3071 case 10: // 1 0 1 0
3072 return Op1;
3073 case 11: // 1 0 1 1
3074 return HasOneUse ? Builder.CreateOr(LHS: Builder.CreateNot(V: Op0), RHS: Op1) : nullptr;
3075 case 12: // 1 1 0 0
3076 return Op0;
3077 case 13: // 1 1 0 1
3078 return HasOneUse ? Builder.CreateOr(LHS: Op0, RHS: Builder.CreateNot(V: Op1)) : nullptr;
3079 case 14: // 1 1 1 0
3080 return Builder.CreateOr(LHS: Op0, RHS: Op1);
3081 case 15: // 1 1 1 1
3082 return FoldConstant(true);
3083 default:
3084 llvm_unreachable("Invalid Operation");
3085 }
3086 return nullptr;
3087}
3088
3089Instruction *InstCombinerImpl::foldICmpBinOpWithConstantViaTruthTable(
3090 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3091 Value *A, *B;
3092 Constant *C1, *C2, *C3, *C4;
3093 if (!match(V: BO->getOperand(i_nocapture: 0),
3094 P: m_SelectLike(C: m_Value(V&: A), TrueC: m_Constant(C&: C1), FalseC: m_Constant(C&: C2))) ||
3095 !match(V: BO->getOperand(i_nocapture: 1),
3096 P: m_SelectLike(C: m_Value(V&: B), TrueC: m_Constant(C&: C3), FalseC: m_Constant(C&: C4))) ||
3097 Cmp.getType() != A->getType() || Cmp.getType() != B->getType())
3098 return nullptr;
3099
3100 std::bitset<4> Table;
3101 auto ComputeTable = [&](bool First, bool Second) -> std::optional<bool> {
3102 Constant *L = First ? C1 : C2;
3103 Constant *R = Second ? C3 : C4;
3104 if (auto *Res = ConstantFoldBinaryOpOperands(Opcode: BO->getOpcode(), LHS: L, RHS: R, DL)) {
3105 auto *Val = Res->getType()->isVectorTy() ? Res->getSplatValue() : Res;
3106 if (auto *CI = dyn_cast_or_null<ConstantInt>(Val))
3107 return ICmpInst::compare(LHS: CI->getValue(), RHS: C, Pred: Cmp.getPredicate());
3108 }
3109 return std::nullopt;
3110 };
3111
3112 for (unsigned I = 0; I < 4; ++I) {
3113 bool First = (I >> 1) & 1;
3114 bool Second = I & 1;
3115 if (auto Res = ComputeTable(First, Second))
3116 Table[I] = *Res;
3117 else
3118 return nullptr;
3119 }
3120
3121 // Synthesize optimal logic.
3122 if (auto *Cond = createLogicFromTable(Table, Op0: A, Op1: B, Builder, HasOneUse: BO->hasOneUse()))
3123 return replaceInstUsesWith(I&: Cmp, V: Cond);
3124 return nullptr;
3125}
3126
3127/// Fold icmp (add X, Y), C.
3128Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
3129 BinaryOperator *Add,
3130 const APInt &C) {
3131 Value *Y = Add->getOperand(i_nocapture: 1);
3132 Value *X = Add->getOperand(i_nocapture: 0);
3133 const CmpPredicate Pred = Cmp.getCmpPredicate();
3134
3135 // icmp ult (add nuw A, (lshr A, ShAmtC)), C --> icmp ult A, C
3136 // when C <= (1 << ShAmtC).
3137 const APInt *ShAmtC;
3138 Value *A;
3139 unsigned BitWidth = C.getBitWidth();
3140 if (Pred == ICmpInst::ICMP_ULT &&
3141 match(V: Add,
3142 P: m_c_NUWAdd(L: m_Value(V&: A), R: m_LShr(L: m_Deferred(V: A), R: m_APInt(Res&: ShAmtC)))) &&
3143 ShAmtC->ult(RHS: BitWidth) &&
3144 C.ule(RHS: APInt::getOneBitSet(numBits: BitWidth, BitNo: ShAmtC->getZExtValue())))
3145 return new ICmpInst(Pred, A, ConstantInt::get(Ty: A->getType(), V: C));
3146
3147 const APInt *C2;
3148 if (Cmp.isEquality() || !match(V: Y, P: m_APInt(Res&: C2)))
3149 return nullptr;
3150
3151 // Fold icmp pred (add X, C2), C.
3152 Type *Ty = Add->getType();
3153
3154 // If the add does not wrap, we can always adjust the compare by subtracting
3155 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
3156 // have been canonicalized to SGT/SLT/UGT/ULT.
3157 if (Add->hasNoUnsignedWrap() &&
3158 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT)) {
3159 bool Overflow;
3160 APInt NewC = C.usub_ov(RHS: *C2, Overflow);
3161 // If there is overflow, the result must be true or false.
3162 if (!Overflow)
3163 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
3164 return new ICmpInst(Pred, X, ConstantInt::get(Ty, V: NewC));
3165 }
3166
3167 CmpInst::Predicate ChosenPred = Pred.getPreferredSignedPredicate();
3168
3169 if (Add->hasNoSignedWrap() &&
3170 (ChosenPred == ICmpInst::ICMP_SGT || ChosenPred == ICmpInst::ICMP_SLT)) {
3171 bool Overflow;
3172 APInt NewC = C.ssub_ov(RHS: *C2, Overflow);
3173 if (!Overflow)
3174 // icmp samesign ugt/ult (add nsw X, C2), C
3175 // -> icmp sgt/slt X, (C - C2)
3176 return new ICmpInst(ChosenPred, X, ConstantInt::get(Ty, V: NewC));
3177 }
3178
3179 if (ICmpInst::isUnsigned(Pred) && Add->hasNoSignedWrap() &&
3180 C.isNonNegative() && (C - *C2).isNonNegative() &&
3181 computeConstantRange(V: X, /*ForSigned=*/true).add(Other: *C2).isAllNonNegative())
3182 return new ICmpInst(ICmpInst::getSignedPredicate(Pred), X,
3183 ConstantInt::get(Ty, V: C - *C2));
3184
3185 auto CR = ConstantRange::makeExactICmpRegion(Pred, Other: C).subtract(CI: *C2);
3186 const APInt &Upper = CR.getUpper();
3187 const APInt &Lower = CR.getLower();
3188 if (Cmp.isSigned()) {
3189 if (Lower.isSignMask())
3190 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, V: Upper));
3191 if (Upper.isSignMask())
3192 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, V: Lower));
3193 } else {
3194 if (Lower.isMinValue())
3195 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, V: Upper));
3196 if (Upper.isMinValue())
3197 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, V: Lower));
3198 }
3199
3200 // This set of folds is intentionally placed after folds that use no-wrapping
3201 // flags because those folds are likely better for later analysis/codegen.
3202 const APInt SMax = APInt::getSignedMaxValue(numBits: Ty->getScalarSizeInBits());
3203 const APInt SMin = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits());
3204
3205 // Fold compare with offset to opposite sign compare if it eliminates offset:
3206 // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
3207 if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
3208 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, V: -(*C2)));
3209
3210 // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
3211 if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
3212 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, V: ~(*C2)));
3213
3214 // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
3215 if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
3216 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, V: SMax - C));
3217
3218 // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
3219 if (Pred == CmpInst::ICMP_SLT && C == *C2)
3220 return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, V: C ^ SMax));
3221
3222 // (X + -1) <u C --> X <=u C (if X is never null)
3223 if (Pred == CmpInst::ICMP_ULT && C2->isAllOnes()) {
3224 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
3225 if (llvm::isKnownNonZero(V: X, Q))
3226 return new ICmpInst(ICmpInst::ICMP_ULE, X, ConstantInt::get(Ty, V: C));
3227 }
3228
3229 if (!Add->hasOneUse())
3230 return nullptr;
3231
3232 // X+C <u C2 -> (X & -C2) == C
3233 // iff C & (C2-1) == 0
3234 // C2 is a power of 2
3235 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
3236 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(LHS: X, RHS: -C),
3237 ConstantExpr::getNeg(C: cast<Constant>(Val: Y)));
3238
3239 // X+C2 <u C -> (X & C) == 2C
3240 // iff C == -(C2)
3241 // C2 is a power of 2
3242 if (Pred == ICmpInst::ICMP_ULT && C2->isPowerOf2() && C == -*C2)
3243 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(LHS: X, RHS: C),
3244 ConstantInt::get(Ty, V: C * 2));
3245
3246 // X+C >u C2 -> (X & ~C2) != C
3247 // iff C & C2 == 0
3248 // C2+1 is a power of 2
3249 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
3250 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(LHS: X, RHS: ~C),
3251 ConstantExpr::getNeg(C: cast<Constant>(Val: Y)));
3252
3253 // The range test idiom can use either ult or ugt. Arbitrarily canonicalize
3254 // to the ult form.
3255 // X+C2 >u C -> X+(C2-C-1) <u ~C
3256 if (Pred == ICmpInst::ICMP_UGT)
3257 return new ICmpInst(ICmpInst::ICMP_ULT,
3258 Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: *C2 - C - 1)),
3259 ConstantInt::get(Ty, V: ~C));
3260
3261 // zext(V) + C2 pred C -> V + C3 pred' C4
3262 Value *V;
3263 if (match(V: X, P: m_ZExt(Op: m_Value(V)))) {
3264 Type *NewCmpTy = V->getType();
3265 unsigned NewCmpBW = NewCmpTy->getScalarSizeInBits();
3266 if (shouldChangeType(From: Ty, To: NewCmpTy)) {
3267 ConstantRange SrcCR = CR.truncate(BitWidth: NewCmpBW, NoWrapKind: TruncInst::NoUnsignedWrap);
3268 CmpInst::Predicate EquivPred;
3269 APInt EquivInt;
3270 APInt EquivOffset;
3271
3272 SrcCR.getEquivalentICmp(Pred&: EquivPred, RHS&: EquivInt, Offset&: EquivOffset);
3273 return new ICmpInst(
3274 EquivPred,
3275 EquivOffset.isZero()
3276 ? V
3277 : Builder.CreateAdd(LHS: V, RHS: ConstantInt::get(Ty: NewCmpTy, V: EquivOffset)),
3278 ConstantInt::get(Ty: NewCmpTy, V: EquivInt));
3279 }
3280 }
3281
3282 return nullptr;
3283}
3284
3285bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
3286 Value *&RHS, ConstantInt *&Less,
3287 ConstantInt *&Equal,
3288 ConstantInt *&Greater) {
3289 // TODO: Generalize this to work with other comparison idioms or ensure
3290 // they get canonicalized into this form.
3291
3292 // select i1 (a == b),
3293 // i32 Equal,
3294 // i32 (select i1 (a < b), i32 Less, i32 Greater)
3295 // where Equal, Less and Greater are placeholders for any three constants.
3296 CmpPredicate PredA;
3297 if (!match(V: SI->getCondition(), P: m_ICmp(Pred&: PredA, L: m_Value(V&: LHS), R: m_Value(V&: RHS))) ||
3298 !ICmpInst::isEquality(P: PredA))
3299 return false;
3300 Value *EqualVal = SI->getTrueValue();
3301 Value *UnequalVal = SI->getFalseValue();
3302 // We still can get non-canonical predicate here, so canonicalize.
3303 if (PredA == ICmpInst::ICMP_NE)
3304 std::swap(a&: EqualVal, b&: UnequalVal);
3305 if (!match(V: EqualVal, P: m_ConstantInt(CI&: Equal)))
3306 return false;
3307 CmpPredicate PredB;
3308 Value *LHS2, *RHS2;
3309 if (!match(V: UnequalVal, P: m_Select(C: m_ICmp(Pred&: PredB, L: m_Value(V&: LHS2), R: m_Value(V&: RHS2)),
3310 L: m_ConstantInt(CI&: Less), R: m_ConstantInt(CI&: Greater))))
3311 return false;
3312 // We can get predicate mismatch here, so canonicalize if possible:
3313 // First, ensure that 'LHS' match.
3314 if (LHS2 != LHS) {
3315 // x sgt y <--> y slt x
3316 std::swap(a&: LHS2, b&: RHS2);
3317 PredB = ICmpInst::getSwappedPredicate(pred: PredB);
3318 }
3319 if (LHS2 != LHS)
3320 return false;
3321 // We also need to canonicalize 'RHS'.
3322 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(Val: RHS2)) {
3323 // x sgt C-1 <--> x sge C <--> not(x slt C)
3324 auto FlippedStrictness =
3325 getFlippedStrictnessPredicateAndConstant(Pred: PredB, C: cast<Constant>(Val: RHS2));
3326 if (!FlippedStrictness)
3327 return false;
3328 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE &&
3329 "basic correctness failure");
3330 RHS2 = FlippedStrictness->second;
3331 // And kind-of perform the result swap.
3332 std::swap(a&: Less, b&: Greater);
3333 PredB = ICmpInst::ICMP_SLT;
3334 }
3335 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
3336}
3337
3338Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
3339 SelectInst *Select,
3340 ConstantInt *C) {
3341
3342 assert(C && "Cmp RHS should be a constant int!");
3343 // If we're testing a constant value against the result of a three way
3344 // comparison, the result can be expressed directly in terms of the
3345 // original values being compared. Note: We could possibly be more
3346 // aggressive here and remove the hasOneUse test. The original select is
3347 // really likely to simplify or sink when we remove a test of the result.
3348 Value *OrigLHS, *OrigRHS;
3349 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
3350 if (Cmp.hasOneUse() &&
3351 matchThreeWayIntCompare(SI: Select, LHS&: OrigLHS, RHS&: OrigRHS, Less&: C1LessThan, Equal&: C2Equal,
3352 Greater&: C3GreaterThan)) {
3353 assert(C1LessThan && C2Equal && C3GreaterThan);
3354
3355 bool TrueWhenLessThan = ICmpInst::compare(
3356 LHS: C1LessThan->getValue(), RHS: C->getValue(), Pred: Cmp.getPredicate());
3357 bool TrueWhenEqual = ICmpInst::compare(LHS: C2Equal->getValue(), RHS: C->getValue(),
3358 Pred: Cmp.getPredicate());
3359 bool TrueWhenGreaterThan = ICmpInst::compare(
3360 LHS: C3GreaterThan->getValue(), RHS: C->getValue(), Pred: Cmp.getPredicate());
3361
3362 // This generates the new instruction that will replace the original Cmp
3363 // Instruction. Instead of enumerating the various combinations when
3364 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
3365 // false, we rely on chaining of ORs and future passes of InstCombine to
3366 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
3367
3368 // When none of the three constants satisfy the predicate for the RHS (C),
3369 // the entire original Cmp can be simplified to a false.
3370 Value *Cond = Builder.getFalse();
3371 if (TrueWhenLessThan)
3372 Cond = Builder.CreateOr(
3373 LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_SLT, LHS: OrigLHS, RHS: OrigRHS));
3374 if (TrueWhenEqual)
3375 Cond = Builder.CreateOr(
3376 LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_EQ, LHS: OrigLHS, RHS: OrigRHS));
3377 if (TrueWhenGreaterThan)
3378 Cond = Builder.CreateOr(
3379 LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_SGT, LHS: OrigLHS, RHS: OrigRHS));
3380
3381 return replaceInstUsesWith(I&: Cmp, V: Cond);
3382 }
3383 return nullptr;
3384}
3385
3386Instruction *InstCombinerImpl::foldICmpBitCast(ICmpInst &Cmp) {
3387 auto *Bitcast = dyn_cast<BitCastInst>(Val: Cmp.getOperand(i_nocapture: 0));
3388 if (!Bitcast)
3389 return nullptr;
3390
3391 ICmpInst::Predicate Pred = Cmp.getPredicate();
3392 Value *Op1 = Cmp.getOperand(i_nocapture: 1);
3393 Value *BCSrcOp = Bitcast->getOperand(i_nocapture: 0);
3394 Type *SrcType = Bitcast->getSrcTy();
3395 Type *DstType = Bitcast->getType();
3396
3397 // Make sure the bitcast doesn't change between scalar and vector and
3398 // doesn't change the number of vector elements.
3399 if (SrcType->isVectorTy() == DstType->isVectorTy() &&
3400 SrcType->getScalarSizeInBits() == DstType->getScalarSizeInBits()) {
3401 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
3402 Value *X;
3403 if (match(V: BCSrcOp, P: m_SIToFP(Op: m_Value(V&: X)))) {
3404 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
3405 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
3406 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
3407 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
3408 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
3409 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
3410 match(V: Op1, P: m_Zero()))
3411 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType()));
3412
3413 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
3414 if (Pred == ICmpInst::ICMP_SLT && match(V: Op1, P: m_One()))
3415 return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: 1));
3416
3417 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
3418 if (Pred == ICmpInst::ICMP_SGT && match(V: Op1, P: m_AllOnes()))
3419 return new ICmpInst(Pred, X,
3420 ConstantInt::getAllOnesValue(Ty: X->getType()));
3421 }
3422
3423 // Zero-equality checks are preserved through unsigned floating-point casts:
3424 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
3425 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
3426 if (match(V: BCSrcOp, P: m_UIToFP(Op: m_Value(V&: X))))
3427 if (Cmp.isEquality() && match(V: Op1, P: m_Zero()))
3428 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType()));
3429
3430 const APInt *C;
3431 bool TrueIfSigned;
3432 if (match(V: Op1, P: m_APInt(Res&: C)) && Bitcast->hasOneUse()) {
3433 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
3434 // the FP extend/truncate because that cast does not change the sign-bit.
3435 // This is true for all standard IEEE-754 types and the X86 80-bit type.
3436 // The sign-bit is always the most significant bit in those types.
3437 if (isSignBitCheck(Pred, RHS: *C, TrueIfSigned) &&
3438 (match(V: BCSrcOp, P: m_FPExt(Op: m_Value(V&: X))) ||
3439 match(V: BCSrcOp, P: m_FPTrunc(Op: m_Value(V&: X))))) {
3440 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
3441 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
3442 Type *XType = X->getType();
3443
3444 // We can't currently handle Power style floating point operations here.
3445 if (!(XType->isPPC_FP128Ty() || SrcType->isPPC_FP128Ty())) {
3446 Type *NewType = Builder.getIntNTy(N: XType->getScalarSizeInBits());
3447 if (auto *XVTy = dyn_cast<VectorType>(Val: XType))
3448 NewType = VectorType::get(ElementType: NewType, EC: XVTy->getElementCount());
3449 Value *NewBitcast = Builder.CreateBitCast(V: X, DestTy: NewType);
3450 if (TrueIfSigned)
3451 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
3452 ConstantInt::getNullValue(Ty: NewType));
3453 else
3454 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
3455 ConstantInt::getAllOnesValue(Ty: NewType));
3456 }
3457 }
3458
3459 // icmp eq/ne (bitcast X to int), special fp -> llvm.is.fpclass(X, class)
3460 Type *FPType = SrcType->getScalarType();
3461 if (!Cmp.getParent()->getParent()->hasFnAttribute(
3462 Kind: Attribute::NoImplicitFloat) &&
3463 Cmp.isEquality() && FPType->isIEEELikeFPTy()) {
3464 FPClassTest Mask = APFloat(FPType->getFltSemantics(), *C).classify();
3465 if (Mask & (fcInf | fcZero)) {
3466 if (Pred == ICmpInst::ICMP_NE)
3467 Mask = ~Mask;
3468 return replaceInstUsesWith(I&: Cmp,
3469 V: Builder.createIsFPClass(FPNum: BCSrcOp, Test: Mask));
3470 }
3471 }
3472 }
3473 }
3474
3475 const APInt *C;
3476 if (!match(V: Cmp.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) || !DstType->isIntegerTy() ||
3477 !SrcType->isIntOrIntVectorTy())
3478 return nullptr;
3479
3480 // If this is checking if all elements of a vector compare are set or not,
3481 // invert the casted vector equality compare and test if all compare
3482 // elements are clear or not. Compare against zero is generally easier for
3483 // analysis and codegen.
3484 // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
3485 // Example: are all elements equal? --> are zero elements not equal?
3486 // TODO: Try harder to reduce compare of 2 freely invertible operands?
3487 if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse()) {
3488 if (Value *NotBCSrcOp =
3489 getFreelyInverted(V: BCSrcOp, WillInvertAllUses: BCSrcOp->hasOneUse(), Builder: &Builder)) {
3490 Value *Cast = Builder.CreateBitCast(V: NotBCSrcOp, DestTy: DstType);
3491 return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(Ty: DstType));
3492 }
3493 }
3494
3495 // If this is checking if all elements of an extended vector are clear or not,
3496 // compare in a narrow type to eliminate the extend:
3497 // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
3498 Value *X;
3499 if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() &&
3500 match(V: BCSrcOp, P: m_ZExtOrSExt(Op: m_Value(V&: X)))) {
3501 if (auto *VecTy = dyn_cast<FixedVectorType>(Val: X->getType())) {
3502 Type *NewType = Builder.getIntNTy(N: VecTy->getPrimitiveSizeInBits());
3503 Value *NewCast = Builder.CreateBitCast(V: X, DestTy: NewType);
3504 return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(Ty: NewType));
3505 }
3506 }
3507
3508 // Folding: icmp <pred> iN X, C
3509 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
3510 // and C is a splat of a K-bit pattern
3511 // and SC is a constant vector = <C', C', C', ..., C'>
3512 // Into:
3513 // %E = extractelement <M x iK> %vec, i32 C'
3514 // icmp <pred> iK %E, trunc(C)
3515 Value *Vec;
3516 ArrayRef<int> Mask;
3517 if (match(V: BCSrcOp, P: m_Shuffle(v1: m_Value(V&: Vec), v2: m_Undef(), mask: m_Mask(Mask)))) {
3518 // Check whether every element of Mask is the same constant
3519 if (all_equal(Range&: Mask)) {
3520 auto *VecTy = cast<VectorType>(Val: SrcType);
3521 auto *EltTy = cast<IntegerType>(Val: VecTy->getElementType());
3522 if (C->isSplat(SplatSizeInBits: EltTy->getBitWidth())) {
3523 // Fold the icmp based on the value of C
3524 // If C is M copies of an iK sized bit pattern,
3525 // then:
3526 // => %E = extractelement <N x iK> %vec, i32 Elem
3527 // icmp <pred> iK %SplatVal, <pattern>
3528 Value *Elem = Builder.getInt32(C: Mask[0]);
3529 Value *Extract = Builder.CreateExtractElement(Vec, Idx: Elem);
3530 Value *NewC = ConstantInt::get(Ty: EltTy, V: C->trunc(width: EltTy->getBitWidth()));
3531 return new ICmpInst(Pred, Extract, NewC);
3532 }
3533 }
3534 }
3535 return nullptr;
3536}
3537
3538/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3539/// where X is some kind of instruction.
3540Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
3541 const APInt *C;
3542
3543 if (match(V: Cmp.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) {
3544 if (auto *BO = dyn_cast<BinaryOperator>(Val: Cmp.getOperand(i_nocapture: 0)))
3545 if (Instruction *I = foldICmpBinOpWithConstant(Cmp, BO, C: *C))
3546 return I;
3547
3548 if (auto *SI = dyn_cast<SelectInst>(Val: Cmp.getOperand(i_nocapture: 0)))
3549 // For now, we only support constant integers while folding the
3550 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3551 // similar to the cases handled by binary ops above.
3552 if (auto *ConstRHS = dyn_cast<ConstantInt>(Val: Cmp.getOperand(i_nocapture: 1)))
3553 if (Instruction *I = foldICmpSelectConstant(Cmp, Select: SI, C: ConstRHS))
3554 return I;
3555
3556 if (auto *TI = dyn_cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 0)))
3557 if (Instruction *I = foldICmpTruncConstant(Cmp, Trunc: TI, C: *C))
3558 return I;
3559
3560 if (auto *II = dyn_cast<IntrinsicInst>(Val: Cmp.getOperand(i_nocapture: 0)))
3561 if (Instruction *I = foldICmpIntrinsicWithConstant(ICI&: Cmp, II, C: *C))
3562 return I;
3563
3564 // (extractval ([s/u]subo X, Y), 0) == 0 --> X == Y
3565 // (extractval ([s/u]subo X, Y), 0) != 0 --> X != Y
3566 // TODO: This checks one-use, but that is not strictly necessary.
3567 Value *Cmp0 = Cmp.getOperand(i_nocapture: 0);
3568 Value *X, *Y;
3569 if (C->isZero() && Cmp.isEquality() && Cmp0->hasOneUse() &&
3570 (match(V: Cmp0,
3571 P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::ssub_with_overflow>(
3572 Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) ||
3573 match(V: Cmp0,
3574 P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::usub_with_overflow>(
3575 Op0: m_Value(V&: X), Op1: m_Value(V&: Y))))))
3576 return new ICmpInst(Cmp.getPredicate(), X, Y);
3577 }
3578
3579 if (match(V: Cmp.getOperand(i_nocapture: 1), P: m_APIntAllowPoison(Res&: C)))
3580 return foldICmpInstWithConstantAllowPoison(Cmp, C: *C);
3581
3582 return nullptr;
3583}
3584
3585/// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3586/// icmp eq/ne BO, C.
3587Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
3588 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3589 // TODO: Some of these folds could work with arbitrary constants, but this
3590 // function is limited to scalar and vector splat constants.
3591 if (!Cmp.isEquality())
3592 return nullptr;
3593
3594 ICmpInst::Predicate Pred = Cmp.getPredicate();
3595 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3596 Constant *RHS = cast<Constant>(Val: Cmp.getOperand(i_nocapture: 1));
3597 Value *BOp0 = BO->getOperand(i_nocapture: 0), *BOp1 = BO->getOperand(i_nocapture: 1);
3598
3599 switch (BO->getOpcode()) {
3600 case Instruction::SRem:
3601 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3602 if (C.isZero() && BO->hasOneUse()) {
3603 const APInt *BOC;
3604 if (match(V: BOp1, P: m_APInt(Res&: BOC)) && BOC->sgt(RHS: 1) && BOC->isPowerOf2()) {
3605 Value *NewRem = Builder.CreateURem(LHS: BOp0, RHS: BOp1, Name: BO->getName());
3606 return new ICmpInst(Pred, NewRem,
3607 Constant::getNullValue(Ty: BO->getType()));
3608 }
3609 }
3610 break;
3611 case Instruction::Add: {
3612 // (A + C2) == C --> A == (C - C2)
3613 // (A + C2) != C --> A != (C - C2)
3614 // TODO: Remove the one-use limitation? See discussion in D58633.
3615 if (Constant *C2 = dyn_cast<Constant>(Val: BOp1)) {
3616 if (BO->hasOneUse())
3617 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(C1: RHS, C2));
3618 } else if (C.isZero()) {
3619 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3620 // efficiently invertible, or if the add has just this one use.
3621 if (Value *NegVal = dyn_castNegVal(V: BOp1))
3622 return new ICmpInst(Pred, BOp0, NegVal);
3623 if (Value *NegVal = dyn_castNegVal(V: BOp0))
3624 return new ICmpInst(Pred, NegVal, BOp1);
3625 if (BO->hasOneUse()) {
3626 // (add nuw A, B) != 0 -> (or A, B) != 0
3627 if (match(V: BO, P: m_NUWAdd(L: m_Value(), R: m_Value()))) {
3628 Value *Or = Builder.CreateOr(LHS: BOp0, RHS: BOp1);
3629 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty: BO->getType()));
3630 }
3631 Value *Neg = Builder.CreateNeg(V: BOp1);
3632 Neg->takeName(V: BO);
3633 return new ICmpInst(Pred, BOp0, Neg);
3634 }
3635 }
3636 break;
3637 }
3638 case Instruction::Xor:
3639 if (Constant *BOC = dyn_cast<Constant>(Val: BOp1)) {
3640 // For the xor case, we can xor two constants together, eliminating
3641 // the explicit xor.
3642 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(C1: RHS, C2: BOC));
3643 } else if (C.isZero()) {
3644 // Replace ((xor A, B) != 0) with (A != B)
3645 return new ICmpInst(Pred, BOp0, BOp1);
3646 }
3647 break;
3648 case Instruction::Or: {
3649 const APInt *BOC;
3650 if (match(V: BOp1, P: m_APInt(Res&: BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3651 // Comparing if all bits outside of a constant mask are set?
3652 // Replace (X | C) == -1 with (X & ~C) == ~C.
3653 // This removes the -1 constant.
3654 Constant *NotBOC = ConstantExpr::getNot(C: cast<Constant>(Val: BOp1));
3655 Value *And = Builder.CreateAnd(LHS: BOp0, RHS: NotBOC);
3656 return new ICmpInst(Pred, And, NotBOC);
3657 }
3658 // (icmp eq (or (select cond, 0, NonZero), Other), 0)
3659 // -> (and cond, (icmp eq Other, 0))
3660 // (icmp ne (or (select cond, NonZero, 0), Other), 0)
3661 // -> (or cond, (icmp ne Other, 0))
3662 Value *Cond, *TV, *FV, *Other, *Sel;
3663 if (C.isZero() &&
3664 match(V: BO,
3665 P: m_OneUse(SubPattern: m_c_Or(L: m_CombineAnd(L: m_Value(V&: Sel),
3666 R: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TV),
3667 R: m_Value(V&: FV))),
3668 R: m_Value(V&: Other)))) &&
3669 Cond->getType() == Cmp.getType()) {
3670 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
3671 // Easy case is if eq/ne matches whether 0 is trueval/falseval.
3672 if (Pred == ICmpInst::ICMP_EQ
3673 ? (match(V: TV, P: m_Zero()) && isKnownNonZero(V: FV, Q))
3674 : (match(V: FV, P: m_Zero()) && isKnownNonZero(V: TV, Q))) {
3675 Value *Cmp = Builder.CreateICmp(
3676 P: Pred, LHS: Other, RHS: Constant::getNullValue(Ty: Other->getType()));
3677 return BinaryOperator::Create(
3678 Op: Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, S1: Cmp,
3679 S2: Cond);
3680 }
3681 // Harder case is if eq/ne matches whether 0 is falseval/trueval. In this
3682 // case we need to invert the select condition so we need to be careful to
3683 // avoid creating extra instructions.
3684 // (icmp ne (or (select cond, 0, NonZero), Other), 0)
3685 // -> (or (not cond), (icmp ne Other, 0))
3686 // (icmp eq (or (select cond, NonZero, 0), Other), 0)
3687 // -> (and (not cond), (icmp eq Other, 0))
3688 //
3689 // Only do this if the inner select has one use, in which case we are
3690 // replacing `select` with `(not cond)`. Otherwise, we will create more
3691 // uses. NB: Trying to freely invert cond doesn't make sense here, as if
3692 // cond was freely invertable, the select arms would have been inverted.
3693 if (Sel->hasOneUse() &&
3694 (Pred == ICmpInst::ICMP_EQ
3695 ? (match(V: FV, P: m_Zero()) && isKnownNonZero(V: TV, Q))
3696 : (match(V: TV, P: m_Zero()) && isKnownNonZero(V: FV, Q)))) {
3697 Value *NotCond = Builder.CreateNot(V: Cond);
3698 Value *Cmp = Builder.CreateICmp(
3699 P: Pred, LHS: Other, RHS: Constant::getNullValue(Ty: Other->getType()));
3700 return BinaryOperator::Create(
3701 Op: Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, S1: Cmp,
3702 S2: NotCond);
3703 }
3704 }
3705 break;
3706 }
3707 case Instruction::UDiv:
3708 case Instruction::SDiv:
3709 if (BO->isExact()) {
3710 // div exact X, Y eq/ne 0 -> X eq/ne 0
3711 // div exact X, Y eq/ne 1 -> X eq/ne Y
3712 // div exact X, Y eq/ne C ->
3713 // if Y * C never-overflow && OneUse:
3714 // -> Y * C eq/ne X
3715 if (C.isZero())
3716 return new ICmpInst(Pred, BOp0, Constant::getNullValue(Ty: BO->getType()));
3717 else if (C.isOne())
3718 return new ICmpInst(Pred, BOp0, BOp1);
3719 else if (BO->hasOneUse()) {
3720 OverflowResult OR = computeOverflow(
3721 BinaryOp: Instruction::Mul, IsSigned: BO->getOpcode() == Instruction::SDiv, LHS: BOp1,
3722 RHS: Cmp.getOperand(i_nocapture: 1), CxtI: BO);
3723 if (OR == OverflowResult::NeverOverflows) {
3724 Value *YC =
3725 Builder.CreateMul(LHS: BOp1, RHS: ConstantInt::get(Ty: BO->getType(), V: C));
3726 return new ICmpInst(Pred, YC, BOp0);
3727 }
3728 }
3729 }
3730 if (BO->getOpcode() == Instruction::UDiv && C.isZero()) {
3731 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3732 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3733 return new ICmpInst(NewPred, BOp1, BOp0);
3734 }
3735 break;
3736 default:
3737 break;
3738 }
3739 return nullptr;
3740}
3741
3742static Instruction *foldCtpopPow2Test(ICmpInst &I, IntrinsicInst *CtpopLhs,
3743 const APInt &CRhs,
3744 InstCombiner::BuilderTy &Builder,
3745 const SimplifyQuery &Q) {
3746 assert(CtpopLhs->getIntrinsicID() == Intrinsic::ctpop &&
3747 "Non-ctpop intrin in ctpop fold");
3748 if (!CtpopLhs->hasOneUse())
3749 return nullptr;
3750
3751 // Power of 2 test:
3752 // isPow2OrZero : ctpop(X) u< 2
3753 // isPow2 : ctpop(X) == 1
3754 // NotPow2OrZero: ctpop(X) u> 1
3755 // NotPow2 : ctpop(X) != 1
3756 // If we know any bit of X can be folded to:
3757 // IsPow2 : X & (~Bit) == 0
3758 // NotPow2 : X & (~Bit) != 0
3759 const ICmpInst::Predicate Pred = I.getPredicate();
3760 if (((I.isEquality() || Pred == ICmpInst::ICMP_UGT) && CRhs == 1) ||
3761 (Pred == ICmpInst::ICMP_ULT && CRhs == 2)) {
3762 Value *Op = CtpopLhs->getArgOperand(i: 0);
3763 KnownBits OpKnown = computeKnownBits(V: Op, DL: Q.DL, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT);
3764 // No need to check for count > 1, that should be already constant folded.
3765 if (OpKnown.countMinPopulation() == 1) {
3766 Value *And = Builder.CreateAnd(
3767 LHS: Op, RHS: Constant::getIntegerValue(Ty: Op->getType(), V: ~(OpKnown.One)));
3768 return new ICmpInst(
3769 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_ULT)
3770 ? ICmpInst::ICMP_EQ
3771 : ICmpInst::ICMP_NE,
3772 And, Constant::getNullValue(Ty: Op->getType()));
3773 }
3774 }
3775
3776 return nullptr;
3777}
3778
3779/// Fold an equality icmp with LLVM intrinsic and constant operand.
3780Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3781 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3782 Type *Ty = II->getType();
3783 unsigned BitWidth = C.getBitWidth();
3784 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3785
3786 switch (II->getIntrinsicID()) {
3787 case Intrinsic::abs:
3788 // abs(A) == 0 -> A == 0
3789 // abs(A) == INT_MIN -> A == INT_MIN
3790 if (C.isZero() || C.isMinSignedValue())
3791 return new ICmpInst(Pred, II->getArgOperand(i: 0), ConstantInt::get(Ty, V: C));
3792 break;
3793
3794 case Intrinsic::bswap:
3795 // bswap(A) == C -> A == bswap(C)
3796 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3797 ConstantInt::get(Ty, V: C.byteSwap()));
3798
3799 case Intrinsic::bitreverse:
3800 // bitreverse(A) == C -> A == bitreverse(C)
3801 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3802 ConstantInt::get(Ty, V: C.reverseBits()));
3803
3804 case Intrinsic::ctlz:
3805 case Intrinsic::cttz: {
3806 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3807 if (C == BitWidth)
3808 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3809 ConstantInt::getNullValue(Ty));
3810
3811 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3812 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3813 // Limit to one use to ensure we don't increase instruction count.
3814 unsigned Num = C.getLimitedValue(Limit: BitWidth);
3815 if (Num != BitWidth && II->hasOneUse()) {
3816 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3817 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: Num + 1)
3818 : APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: Num + 1);
3819 APInt Mask2 = IsTrailing
3820 ? APInt::getOneBitSet(numBits: BitWidth, BitNo: Num)
3821 : APInt::getOneBitSet(numBits: BitWidth, BitNo: BitWidth - Num - 1);
3822 return new ICmpInst(Pred, Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask1),
3823 ConstantInt::get(Ty, V: Mask2));
3824 }
3825 break;
3826 }
3827
3828 case Intrinsic::ctpop: {
3829 // popcount(A) == 0 -> A == 0 and likewise for !=
3830 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3831 bool IsZero = C.isZero();
3832 if (IsZero || C == BitWidth)
3833 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3834 IsZero ? Constant::getNullValue(Ty)
3835 : Constant::getAllOnesValue(Ty));
3836
3837 break;
3838 }
3839
3840 case Intrinsic::fshl:
3841 case Intrinsic::fshr:
3842 if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) {
3843 const APInt *RotAmtC;
3844 // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3845 // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3846 if (match(V: II->getArgOperand(i: 2), P: m_APInt(Res&: RotAmtC)))
3847 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3848 II->getIntrinsicID() == Intrinsic::fshl
3849 ? ConstantInt::get(Ty, V: C.rotr(rotateAmt: *RotAmtC))
3850 : ConstantInt::get(Ty, V: C.rotl(rotateAmt: *RotAmtC)));
3851 }
3852 break;
3853
3854 case Intrinsic::umax:
3855 case Intrinsic::uadd_sat: {
3856 // uadd.sat(a, b) == 0 -> (a | b) == 0
3857 // umax(a, b) == 0 -> (a | b) == 0
3858 if (C.isZero() && II->hasOneUse()) {
3859 Value *Or = Builder.CreateOr(LHS: II->getArgOperand(i: 0), RHS: II->getArgOperand(i: 1));
3860 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3861 }
3862 break;
3863 }
3864
3865 case Intrinsic::ssub_sat:
3866 // ssub.sat(a, b) == 0 -> a == b
3867 //
3868 // Note this doesn't work for ssub.sat.i1 because ssub.sat.i1 0, -1 = 0
3869 // (because 1 saturates to 0). Just skip the optimization for i1.
3870 if (C.isZero() && II->getType()->getScalarSizeInBits() > 1)
3871 return new ICmpInst(Pred, II->getArgOperand(i: 0), II->getArgOperand(i: 1));
3872 break;
3873 case Intrinsic::usub_sat: {
3874 // usub.sat(a, b) == 0 -> a <= b
3875 if (C.isZero()) {
3876 ICmpInst::Predicate NewPred =
3877 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3878 return new ICmpInst(NewPred, II->getArgOperand(i: 0), II->getArgOperand(i: 1));
3879 }
3880 break;
3881 }
3882 default:
3883 break;
3884 }
3885
3886 return nullptr;
3887}
3888
3889/// Fold an icmp with LLVM intrinsics
3890static Instruction *
3891foldICmpIntrinsicWithIntrinsic(ICmpInst &Cmp,
3892 InstCombiner::BuilderTy &Builder) {
3893 assert(Cmp.isEquality());
3894
3895 ICmpInst::Predicate Pred = Cmp.getPredicate();
3896 Value *Op0 = Cmp.getOperand(i_nocapture: 0);
3897 Value *Op1 = Cmp.getOperand(i_nocapture: 1);
3898 const auto *IIOp0 = dyn_cast<IntrinsicInst>(Val: Op0);
3899 const auto *IIOp1 = dyn_cast<IntrinsicInst>(Val: Op1);
3900 if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3901 return nullptr;
3902
3903 switch (IIOp0->getIntrinsicID()) {
3904 case Intrinsic::bswap:
3905 case Intrinsic::bitreverse:
3906 // If both operands are byte-swapped or bit-reversed, just compare the
3907 // original values.
3908 return new ICmpInst(Pred, IIOp0->getOperand(i_nocapture: 0), IIOp1->getOperand(i_nocapture: 0));
3909 case Intrinsic::fshl:
3910 case Intrinsic::fshr: {
3911 // If both operands are rotated by same amount, just compare the
3912 // original values.
3913 if (IIOp0->getOperand(i_nocapture: 0) != IIOp0->getOperand(i_nocapture: 1))
3914 break;
3915 if (IIOp1->getOperand(i_nocapture: 0) != IIOp1->getOperand(i_nocapture: 1))
3916 break;
3917 if (IIOp0->getOperand(i_nocapture: 2) == IIOp1->getOperand(i_nocapture: 2))
3918 return new ICmpInst(Pred, IIOp0->getOperand(i_nocapture: 0), IIOp1->getOperand(i_nocapture: 0));
3919
3920 // rotate(X, AmtX) == rotate(Y, AmtY)
3921 // -> rotate(X, AmtX - AmtY) == Y
3922 // Do this if either both rotates have one use or if only one has one use
3923 // and AmtX/AmtY are constants.
3924 unsigned OneUses = IIOp0->hasOneUse() + IIOp1->hasOneUse();
3925 if (OneUses == 2 ||
3926 (OneUses == 1 && match(V: IIOp0->getOperand(i_nocapture: 2), P: m_ImmConstant()) &&
3927 match(V: IIOp1->getOperand(i_nocapture: 2), P: m_ImmConstant()))) {
3928 Value *SubAmt =
3929 Builder.CreateSub(LHS: IIOp0->getOperand(i_nocapture: 2), RHS: IIOp1->getOperand(i_nocapture: 2));
3930 Value *CombinedRotate = Builder.CreateIntrinsic(
3931 RetTy: Op0->getType(), ID: IIOp0->getIntrinsicID(),
3932 Args: {IIOp0->getOperand(i_nocapture: 0), IIOp0->getOperand(i_nocapture: 0), SubAmt});
3933 return new ICmpInst(Pred, IIOp1->getOperand(i_nocapture: 0), CombinedRotate);
3934 }
3935 } break;
3936 default:
3937 break;
3938 }
3939
3940 return nullptr;
3941}
3942
3943/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3944/// where X is some kind of instruction and C is AllowPoison.
3945/// TODO: Move more folds which allow poison to this function.
3946Instruction *
3947InstCombinerImpl::foldICmpInstWithConstantAllowPoison(ICmpInst &Cmp,
3948 const APInt &C) {
3949 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3950 if (auto *II = dyn_cast<IntrinsicInst>(Val: Cmp.getOperand(i_nocapture: 0))) {
3951 switch (II->getIntrinsicID()) {
3952 default:
3953 break;
3954 case Intrinsic::fshl:
3955 case Intrinsic::fshr:
3956 if (Cmp.isEquality() && II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) {
3957 // (rot X, ?) == 0/-1 --> X == 0/-1
3958 if (C.isZero() || C.isAllOnes())
3959 return new ICmpInst(Pred, II->getArgOperand(i: 0), Cmp.getOperand(i_nocapture: 1));
3960 }
3961 break;
3962 }
3963 }
3964
3965 return nullptr;
3966}
3967
3968/// Fold an icmp with BinaryOp and constant operand: icmp Pred BO, C.
3969Instruction *InstCombinerImpl::foldICmpBinOpWithConstant(ICmpInst &Cmp,
3970 BinaryOperator *BO,
3971 const APInt &C) {
3972 switch (BO->getOpcode()) {
3973 case Instruction::Xor:
3974 if (Instruction *I = foldICmpXorConstant(Cmp, Xor: BO, C))
3975 return I;
3976 break;
3977 case Instruction::And:
3978 if (Instruction *I = foldICmpAndConstant(Cmp, And: BO, C))
3979 return I;
3980 break;
3981 case Instruction::Or:
3982 if (Instruction *I = foldICmpOrConstant(Cmp, Or: BO, C))
3983 return I;
3984 break;
3985 case Instruction::Mul:
3986 if (Instruction *I = foldICmpMulConstant(Cmp, Mul: BO, C))
3987 return I;
3988 break;
3989 case Instruction::Shl:
3990 if (Instruction *I = foldICmpShlConstant(Cmp, Shl: BO, C))
3991 return I;
3992 break;
3993 case Instruction::LShr:
3994 case Instruction::AShr:
3995 if (Instruction *I = foldICmpShrConstant(Cmp, Shr: BO, C))
3996 return I;
3997 break;
3998 case Instruction::SRem:
3999 if (Instruction *I = foldICmpSRemConstant(Cmp, SRem: BO, C))
4000 return I;
4001 break;
4002 case Instruction::UDiv:
4003 if (Instruction *I = foldICmpUDivConstant(Cmp, UDiv: BO, C))
4004 return I;
4005 [[fallthrough]];
4006 case Instruction::SDiv:
4007 if (Instruction *I = foldICmpDivConstant(Cmp, Div: BO, C))
4008 return I;
4009 break;
4010 case Instruction::Sub:
4011 if (Instruction *I = foldICmpSubConstant(Cmp, Sub: BO, C))
4012 return I;
4013 break;
4014 case Instruction::Add:
4015 if (Instruction *I = foldICmpAddConstant(Cmp, Add: BO, C))
4016 return I;
4017 break;
4018 default:
4019 break;
4020 }
4021
4022 // TODO: These folds could be refactored to be part of the above calls.
4023 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C))
4024 return I;
4025
4026 // Fall back to handling `icmp pred (select A ? C1 : C2) binop (select B ? C3
4027 // : C4), C5` pattern, by computing a truth table of the four constant
4028 // variants.
4029 return foldICmpBinOpWithConstantViaTruthTable(Cmp, BO, C);
4030}
4031
4032static Instruction *
4033foldICmpUSubSatOrUAddSatWithConstant(CmpPredicate Pred, SaturatingInst *II,
4034 const APInt &C,
4035 InstCombiner::BuilderTy &Builder) {
4036 // This transform may end up producing more than one instruction for the
4037 // intrinsic, so limit it to one user of the intrinsic.
4038 if (!II->hasOneUse())
4039 return nullptr;
4040
4041 // Let Y = [add/sub]_sat(X, C) pred C2
4042 // SatVal = The saturating value for the operation
4043 // WillWrap = Whether or not the operation will underflow / overflow
4044 // => Y = (WillWrap ? SatVal : (X binop C)) pred C2
4045 // => Y = WillWrap ? (SatVal pred C2) : ((X binop C) pred C2)
4046 //
4047 // When (SatVal pred C2) is true, then
4048 // Y = WillWrap ? true : ((X binop C) pred C2)
4049 // => Y = WillWrap || ((X binop C) pred C2)
4050 // else
4051 // Y = WillWrap ? false : ((X binop C) pred C2)
4052 // => Y = !WillWrap ? ((X binop C) pred C2) : false
4053 // => Y = !WillWrap && ((X binop C) pred C2)
4054 Value *Op0 = II->getOperand(i_nocapture: 0);
4055 Value *Op1 = II->getOperand(i_nocapture: 1);
4056
4057 const APInt *COp1;
4058 // This transform only works when the intrinsic has an integral constant or
4059 // splat vector as the second operand.
4060 if (!match(V: Op1, P: m_APInt(Res&: COp1)))
4061 return nullptr;
4062
4063 APInt SatVal;
4064 switch (II->getIntrinsicID()) {
4065 default:
4066 llvm_unreachable(
4067 "This function only works with usub_sat and uadd_sat for now!");
4068 case Intrinsic::uadd_sat:
4069 SatVal = APInt::getAllOnes(numBits: C.getBitWidth());
4070 break;
4071 case Intrinsic::usub_sat:
4072 SatVal = APInt::getZero(numBits: C.getBitWidth());
4073 break;
4074 }
4075
4076 // Check (SatVal pred C2)
4077 bool SatValCheck = ICmpInst::compare(LHS: SatVal, RHS: C, Pred);
4078
4079 // !WillWrap.
4080 ConstantRange C1 = ConstantRange::makeExactNoWrapRegion(
4081 BinOp: II->getBinaryOp(), Other: *COp1, NoWrapKind: II->getNoWrapKind());
4082
4083 // WillWrap.
4084 if (SatValCheck)
4085 C1 = C1.inverse();
4086
4087 ConstantRange C2 = ConstantRange::makeExactICmpRegion(Pred, Other: C);
4088 if (II->getBinaryOp() == Instruction::Add)
4089 C2 = C2.sub(Other: *COp1);
4090 else
4091 C2 = C2.add(Other: *COp1);
4092
4093 Instruction::BinaryOps CombiningOp =
4094 SatValCheck ? Instruction::BinaryOps::Or : Instruction::BinaryOps::And;
4095
4096 std::optional<ConstantRange> Combination;
4097 if (CombiningOp == Instruction::BinaryOps::Or)
4098 Combination = C1.exactUnionWith(CR: C2);
4099 else /* CombiningOp == Instruction::BinaryOps::And */
4100 Combination = C1.exactIntersectWith(CR: C2);
4101
4102 if (!Combination)
4103 return nullptr;
4104
4105 CmpInst::Predicate EquivPred;
4106 APInt EquivInt;
4107 APInt EquivOffset;
4108
4109 Combination->getEquivalentICmp(Pred&: EquivPred, RHS&: EquivInt, Offset&: EquivOffset);
4110
4111 return new ICmpInst(
4112 EquivPred,
4113 Builder.CreateAdd(LHS: Op0, RHS: ConstantInt::get(Ty: Op1->getType(), V: EquivOffset)),
4114 ConstantInt::get(Ty: Op1->getType(), V: EquivInt));
4115}
4116
4117static Instruction *
4118foldICmpOfCmpIntrinsicWithConstant(CmpPredicate Pred, IntrinsicInst *I,
4119 const APInt &C,
4120 InstCombiner::BuilderTy &Builder) {
4121 std::optional<ICmpInst::Predicate> NewPredicate = std::nullopt;
4122 switch (Pred) {
4123 case ICmpInst::ICMP_EQ:
4124 case ICmpInst::ICMP_NE:
4125 if (C.isZero())
4126 NewPredicate = Pred;
4127 else if (C.isOne())
4128 NewPredicate =
4129 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE;
4130 else if (C.isAllOnes())
4131 NewPredicate =
4132 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
4133 break;
4134
4135 case ICmpInst::ICMP_SGT:
4136 if (C.isAllOnes())
4137 NewPredicate = ICmpInst::ICMP_UGE;
4138 else if (C.isZero())
4139 NewPredicate = ICmpInst::ICMP_UGT;
4140 break;
4141
4142 case ICmpInst::ICMP_SLT:
4143 if (C.isZero())
4144 NewPredicate = ICmpInst::ICMP_ULT;
4145 else if (C.isOne())
4146 NewPredicate = ICmpInst::ICMP_ULE;
4147 break;
4148
4149 case ICmpInst::ICMP_ULT:
4150 if (C.ugt(RHS: 1))
4151 NewPredicate = ICmpInst::ICMP_UGE;
4152 break;
4153
4154 case ICmpInst::ICMP_UGT:
4155 if (!C.isZero() && !C.isAllOnes())
4156 NewPredicate = ICmpInst::ICMP_ULT;
4157 break;
4158
4159 default:
4160 break;
4161 }
4162
4163 if (!NewPredicate)
4164 return nullptr;
4165
4166 if (I->getIntrinsicID() == Intrinsic::scmp)
4167 NewPredicate = ICmpInst::getSignedPredicate(Pred: *NewPredicate);
4168 Value *LHS = I->getOperand(i_nocapture: 0);
4169 Value *RHS = I->getOperand(i_nocapture: 1);
4170 return new ICmpInst(*NewPredicate, LHS, RHS);
4171}
4172
4173/// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
4174Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
4175 IntrinsicInst *II,
4176 const APInt &C) {
4177 ICmpInst::Predicate Pred = Cmp.getPredicate();
4178
4179 // Handle folds that apply for any kind of icmp.
4180 switch (II->getIntrinsicID()) {
4181 default:
4182 break;
4183 case Intrinsic::uadd_sat:
4184 case Intrinsic::usub_sat:
4185 if (auto *Folded = foldICmpUSubSatOrUAddSatWithConstant(
4186 Pred, II: cast<SaturatingInst>(Val: II), C, Builder))
4187 return Folded;
4188 break;
4189 case Intrinsic::ctpop: {
4190 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
4191 if (Instruction *R = foldCtpopPow2Test(I&: Cmp, CtpopLhs: II, CRhs: C, Builder, Q))
4192 return R;
4193 } break;
4194 case Intrinsic::scmp:
4195 case Intrinsic::ucmp:
4196 if (auto *Folded = foldICmpOfCmpIntrinsicWithConstant(Pred, I: II, C, Builder))
4197 return Folded;
4198 break;
4199 }
4200
4201 if (Cmp.isEquality())
4202 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
4203
4204 Type *Ty = II->getType();
4205 unsigned BitWidth = C.getBitWidth();
4206 switch (II->getIntrinsicID()) {
4207 case Intrinsic::ctpop: {
4208 // (ctpop X > BitWidth - 1) --> X == -1
4209 Value *X = II->getArgOperand(i: 0);
4210 if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
4211 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_EQ, S1: X,
4212 S2: ConstantInt::getAllOnesValue(Ty));
4213 // (ctpop X < BitWidth) --> X != -1
4214 if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
4215 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_NE, S1: X,
4216 S2: ConstantInt::getAllOnesValue(Ty));
4217 break;
4218 }
4219 case Intrinsic::ctlz: {
4220 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
4221 if (Pred == ICmpInst::ICMP_UGT && C.ult(RHS: BitWidth)) {
4222 unsigned Num = C.getLimitedValue();
4223 APInt Limit = APInt::getOneBitSet(numBits: BitWidth, BitNo: BitWidth - Num - 1);
4224 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_ULT,
4225 S1: II->getArgOperand(i: 0), S2: ConstantInt::get(Ty, V: Limit));
4226 }
4227
4228 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
4229 if (Pred == ICmpInst::ICMP_ULT && C.uge(RHS: 1) && C.ule(RHS: BitWidth)) {
4230 unsigned Num = C.getLimitedValue();
4231 APInt Limit = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - Num);
4232 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_UGT,
4233 S1: II->getArgOperand(i: 0), S2: ConstantInt::get(Ty, V: Limit));
4234 }
4235 break;
4236 }
4237 case Intrinsic::cttz: {
4238 // Limit to one use to ensure we don't increase instruction count.
4239 if (!II->hasOneUse())
4240 return nullptr;
4241
4242 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
4243 if (Pred == ICmpInst::ICMP_UGT && C.ult(RHS: BitWidth)) {
4244 APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: C.getLimitedValue() + 1);
4245 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_EQ,
4246 S1: Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask),
4247 S2: ConstantInt::getNullValue(Ty));
4248 }
4249
4250 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
4251 if (Pred == ICmpInst::ICMP_ULT && C.uge(RHS: 1) && C.ule(RHS: BitWidth)) {
4252 APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: C.getLimitedValue());
4253 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_NE,
4254 S1: Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask),
4255 S2: ConstantInt::getNullValue(Ty));
4256 }
4257 break;
4258 }
4259 case Intrinsic::ssub_sat:
4260 // ssub.sat(a, b) spred 0 -> a spred b
4261 //
4262 // Note this doesn't work for ssub.sat.i1 because ssub.sat.i1 0, -1 = 0
4263 // (because 1 saturates to 0). Just skip the optimization for i1.
4264 if (ICmpInst::isSigned(Pred) && C.getBitWidth() > 1) {
4265 if (C.isZero())
4266 return new ICmpInst(Pred, II->getArgOperand(i: 0), II->getArgOperand(i: 1));
4267 // X s<= 0 is cannonicalized to X s< 1
4268 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
4269 return new ICmpInst(ICmpInst::ICMP_SLE, II->getArgOperand(i: 0),
4270 II->getArgOperand(i: 1));
4271 // X s>= 0 is cannonicalized to X s> -1
4272 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
4273 return new ICmpInst(ICmpInst::ICMP_SGE, II->getArgOperand(i: 0),
4274 II->getArgOperand(i: 1));
4275 }
4276 break;
4277 case Intrinsic::abs: {
4278 if (!II->hasOneUse())
4279 return nullptr;
4280
4281 Value *X = II->getArgOperand(i: 0);
4282 bool IsIntMinPoison =
4283 cast<ConstantInt>(Val: II->getArgOperand(i: 1))->getValue().isOne();
4284
4285 // If C >= 0:
4286 // abs(X) u> C --> X + C u> 2 * C
4287 if (Pred == CmpInst::ICMP_UGT && C.isNonNegative()) {
4288 return new ICmpInst(ICmpInst::ICMP_UGT,
4289 Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: C)),
4290 ConstantInt::get(Ty, V: 2 * C));
4291 }
4292
4293 // If abs(INT_MIN) is poison and C >= 1:
4294 // abs(X) u< C --> X + (C - 1) u<= 2 * (C - 1)
4295 if (IsIntMinPoison && Pred == CmpInst::ICMP_ULT && C.sge(RHS: 1)) {
4296 return new ICmpInst(ICmpInst::ICMP_ULE,
4297 Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: C - 1)),
4298 ConstantInt::get(Ty, V: 2 * (C - 1)));
4299 }
4300
4301 break;
4302 }
4303 default:
4304 break;
4305 }
4306
4307 return nullptr;
4308}
4309
4310/// Handle icmp with constant (but not simple integer constant) RHS.
4311Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
4312 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
4313 Constant *RHSC = dyn_cast<Constant>(Val: Op1);
4314 Instruction *LHSI = dyn_cast<Instruction>(Val: Op0);
4315 if (!RHSC || !LHSI)
4316 return nullptr;
4317
4318 switch (LHSI->getOpcode()) {
4319 case Instruction::IntToPtr:
4320 // icmp pred inttoptr(X), null -> icmp pred X, 0
4321 if (RHSC->isNullValue() &&
4322 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(i: 0)->getType())
4323 return new ICmpInst(
4324 I.getPredicate(), LHSI->getOperand(i: 0),
4325 Constant::getNullValue(Ty: LHSI->getOperand(i: 0)->getType()));
4326 break;
4327
4328 case Instruction::Load:
4329 // Try to optimize things like "A[i] > 4" to index computations.
4330 if (GetElementPtrInst *GEP =
4331 dyn_cast<GetElementPtrInst>(Val: LHSI->getOperand(i: 0)))
4332 if (Instruction *Res =
4333 foldCmpLoadFromIndexedGlobal(LI: cast<LoadInst>(Val: LHSI), GEP, ICI&: I))
4334 return Res;
4335 break;
4336 }
4337
4338 return nullptr;
4339}
4340
4341Instruction *InstCombinerImpl::foldSelectICmp(CmpPredicate Pred, SelectInst *SI,
4342 Value *RHS, const ICmpInst &I) {
4343 // Try to fold the comparison into the select arms, which will cause the
4344 // select to be converted into a logical and/or.
4345 auto SimplifyOp = [&](Value *Op, bool SelectCondIsTrue) -> Value * {
4346 if (Value *Res = simplifyICmpInst(Pred, LHS: Op, RHS, Q: SQ))
4347 return Res;
4348 if (std::optional<bool> Impl = isImpliedCondition(
4349 LHS: SI->getCondition(), RHSPred: Pred, RHSOp0: Op, RHSOp1: RHS, DL, LHSIsTrue: SelectCondIsTrue))
4350 return ConstantInt::get(Ty: I.getType(), V: *Impl);
4351 return nullptr;
4352 };
4353
4354 ConstantInt *CI = nullptr;
4355 Value *Op1 = SimplifyOp(SI->getOperand(i_nocapture: 1), true);
4356 if (Op1)
4357 CI = dyn_cast<ConstantInt>(Val: Op1);
4358
4359 Value *Op2 = SimplifyOp(SI->getOperand(i_nocapture: 2), false);
4360 if (Op2)
4361 CI = dyn_cast<ConstantInt>(Val: Op2);
4362
4363 auto Simplifies = [&](Value *Op, unsigned Idx) {
4364 // A comparison of ucmp/scmp with a constant will fold into an icmp.
4365 const APInt *Dummy;
4366 return Op ||
4367 (isa<CmpIntrinsic>(Val: SI->getOperand(i_nocapture: Idx)) &&
4368 SI->getOperand(i_nocapture: Idx)->hasOneUse() && match(V: RHS, P: m_APInt(Res&: Dummy)));
4369 };
4370
4371 // We only want to perform this transformation if it will not lead to
4372 // additional code. This is true if either both sides of the select
4373 // fold to a constant (in which case the icmp is replaced with a select
4374 // which will usually simplify) or this is the only user of the
4375 // select (in which case we are trading a select+icmp for a simpler
4376 // select+icmp) or all uses of the select can be replaced based on
4377 // dominance information ("Global cases").
4378 bool Transform = false;
4379 if (Op1 && Op2)
4380 Transform = true;
4381 else if (Simplifies(Op1, 1) || Simplifies(Op2, 2)) {
4382 // Local case
4383 if (SI->hasOneUse())
4384 Transform = true;
4385 // Global cases
4386 else if (CI && !CI->isZero())
4387 // When Op1 is constant try replacing select with second operand.
4388 // Otherwise Op2 is constant and try replacing select with first
4389 // operand.
4390 Transform = replacedSelectWithOperand(SI, Icmp: &I, SIOpd: Op1 ? 2 : 1);
4391 }
4392 if (Transform) {
4393 if (!Op1)
4394 Op1 = Builder.CreateICmp(P: Pred, LHS: SI->getOperand(i_nocapture: 1), RHS, Name: I.getName());
4395 if (!Op2)
4396 Op2 = Builder.CreateICmp(P: Pred, LHS: SI->getOperand(i_nocapture: 2), RHS, Name: I.getName());
4397 return SelectInst::Create(C: SI->getOperand(i_nocapture: 0), S1: Op1, S2: Op2, NameStr: "", InsertBefore: nullptr,
4398 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : SI);
4399 }
4400
4401 return nullptr;
4402}
4403
4404// Returns whether V is a Mask ((X + 1) & X == 0) or ~Mask (-Pow2OrZero)
4405static bool isMaskOrZero(const Value *V, bool Not, const SimplifyQuery &Q,
4406 unsigned Depth = 0) {
4407 if (Not ? match(V, P: m_NegatedPower2OrZero()) : match(V, P: m_LowBitMaskOrZero()))
4408 return true;
4409 if (V->getType()->getScalarSizeInBits() == 1)
4410 return true;
4411 if (Depth++ >= MaxAnalysisRecursionDepth)
4412 return false;
4413 Value *X;
4414 const Instruction *I = dyn_cast<Instruction>(Val: V);
4415 if (!I)
4416 return false;
4417 switch (I->getOpcode()) {
4418 case Instruction::ZExt:
4419 // ZExt(Mask) is a Mask.
4420 return !Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4421 case Instruction::SExt:
4422 // SExt(Mask) is a Mask.
4423 // SExt(~Mask) is a ~Mask.
4424 return isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4425 case Instruction::And:
4426 case Instruction::Or:
4427 // Mask0 | Mask1 is a Mask.
4428 // Mask0 & Mask1 is a Mask.
4429 // ~Mask0 | ~Mask1 is a ~Mask.
4430 // ~Mask0 & ~Mask1 is a ~Mask.
4431 return isMaskOrZero(V: I->getOperand(i: 1), Not, Q, Depth) &&
4432 isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4433 case Instruction::Xor:
4434 if (match(V, P: m_Not(V: m_Value(V&: X))))
4435 return isMaskOrZero(V: X, Not: !Not, Q, Depth);
4436
4437 // (X ^ -X) is a ~Mask
4438 if (Not)
4439 return match(V, P: m_c_Xor(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))));
4440 // (X ^ (X - 1)) is a Mask
4441 else
4442 return match(V, P: m_c_Xor(L: m_Value(V&: X), R: m_Add(L: m_Deferred(V: X), R: m_AllOnes())));
4443 case Instruction::Select:
4444 // c ? Mask0 : Mask1 is a Mask.
4445 return isMaskOrZero(V: I->getOperand(i: 1), Not, Q, Depth) &&
4446 isMaskOrZero(V: I->getOperand(i: 2), Not, Q, Depth);
4447 case Instruction::Shl:
4448 // (~Mask) << X is a ~Mask.
4449 return Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4450 case Instruction::LShr:
4451 // Mask >> X is a Mask.
4452 return !Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4453 case Instruction::AShr:
4454 // Mask s>> X is a Mask.
4455 // ~Mask s>> X is a ~Mask.
4456 return isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4457 case Instruction::Add:
4458 // Pow2 - 1 is a Mask.
4459 if (!Not && match(V: I->getOperand(i: 1), P: m_AllOnes()))
4460 return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), DL: Q.DL, /*OrZero*/ true,
4461 AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT, UseInstrInfo: Depth);
4462 break;
4463 case Instruction::Sub:
4464 // -Pow2 is a ~Mask.
4465 if (Not && match(V: I->getOperand(i: 0), P: m_Zero()))
4466 return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), DL: Q.DL, /*OrZero*/ true,
4467 AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT, UseInstrInfo: Depth);
4468 break;
4469 case Instruction::Call: {
4470 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
4471 switch (II->getIntrinsicID()) {
4472 // min/max(Mask0, Mask1) is a Mask.
4473 // min/max(~Mask0, ~Mask1) is a ~Mask.
4474 case Intrinsic::umax:
4475 case Intrinsic::smax:
4476 case Intrinsic::umin:
4477 case Intrinsic::smin:
4478 return isMaskOrZero(V: II->getArgOperand(i: 1), Not, Q, Depth) &&
4479 isMaskOrZero(V: II->getArgOperand(i: 0), Not, Q, Depth);
4480
4481 // In the context of masks, bitreverse(Mask) == ~Mask
4482 case Intrinsic::bitreverse:
4483 return isMaskOrZero(V: II->getArgOperand(i: 0), Not: !Not, Q, Depth);
4484 default:
4485 break;
4486 }
4487 }
4488 break;
4489 }
4490 default:
4491 break;
4492 }
4493 return false;
4494}
4495
4496/// Some comparisons can be simplified.
4497/// In this case, we are looking for comparisons that look like
4498/// a check for a lossy truncation.
4499/// Folds:
4500/// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
4501/// icmp SrcPred (x & ~Mask), ~Mask to icmp DstPred x, ~Mask
4502/// icmp eq/ne (x & ~Mask), 0 to icmp DstPred x, Mask
4503/// icmp eq/ne (~x | Mask), -1 to icmp DstPred x, Mask
4504/// Where Mask is some pattern that produces all-ones in low bits:
4505/// (-1 >> y)
4506/// ((-1 << y) >> y) <- non-canonical, has extra uses
4507/// ~(-1 << y)
4508/// ((1 << y) + (-1)) <- non-canonical, has extra uses
4509/// The Mask can be a constant, too.
4510/// For some predicates, the operands are commutative.
4511/// For others, x can only be on a specific side.
4512static Value *foldICmpWithLowBitMaskedVal(CmpPredicate Pred, Value *Op0,
4513 Value *Op1, const SimplifyQuery &Q,
4514 InstCombiner &IC) {
4515
4516 ICmpInst::Predicate DstPred;
4517 switch (Pred) {
4518 case ICmpInst::Predicate::ICMP_EQ:
4519 // x & Mask == x
4520 // x & ~Mask == 0
4521 // ~x | Mask == -1
4522 // -> x u<= Mask
4523 // x & ~Mask == ~Mask
4524 // -> ~Mask u<= x
4525 DstPred = ICmpInst::Predicate::ICMP_ULE;
4526 break;
4527 case ICmpInst::Predicate::ICMP_NE:
4528 // x & Mask != x
4529 // x & ~Mask != 0
4530 // ~x | Mask != -1
4531 // -> x u> Mask
4532 // x & ~Mask != ~Mask
4533 // -> ~Mask u> x
4534 DstPred = ICmpInst::Predicate::ICMP_UGT;
4535 break;
4536 case ICmpInst::Predicate::ICMP_ULT:
4537 // x & Mask u< x
4538 // -> x u> Mask
4539 // x & ~Mask u< ~Mask
4540 // -> ~Mask u> x
4541 DstPred = ICmpInst::Predicate::ICMP_UGT;
4542 break;
4543 case ICmpInst::Predicate::ICMP_UGE:
4544 // x & Mask u>= x
4545 // -> x u<= Mask
4546 // x & ~Mask u>= ~Mask
4547 // -> ~Mask u<= x
4548 DstPred = ICmpInst::Predicate::ICMP_ULE;
4549 break;
4550 case ICmpInst::Predicate::ICMP_SLT:
4551 // x & Mask s< x [iff Mask s>= 0]
4552 // -> x s> Mask
4553 // x & ~Mask s< ~Mask [iff ~Mask != 0]
4554 // -> ~Mask s> x
4555 DstPred = ICmpInst::Predicate::ICMP_SGT;
4556 break;
4557 case ICmpInst::Predicate::ICMP_SGE:
4558 // x & Mask s>= x [iff Mask s>= 0]
4559 // -> x s<= Mask
4560 // x & ~Mask s>= ~Mask [iff ~Mask != 0]
4561 // -> ~Mask s<= x
4562 DstPred = ICmpInst::Predicate::ICMP_SLE;
4563 break;
4564 default:
4565 // We don't support sgt,sle
4566 // ult/ugt are simplified to true/false respectively.
4567 return nullptr;
4568 }
4569
4570 Value *X, *M;
4571 // Put search code in lambda for early positive returns.
4572 auto IsLowBitMask = [&]() {
4573 if (match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value(V&: M)))) {
4574 X = Op1;
4575 // Look for: x & Mask pred x
4576 if (isMaskOrZero(V: M, /*Not=*/false, Q)) {
4577 return !ICmpInst::isSigned(Pred) ||
4578 (match(V: M, P: m_NonNegative()) || isKnownNonNegative(V: M, SQ: Q));
4579 }
4580
4581 // Look for: x & ~Mask pred ~Mask
4582 if (isMaskOrZero(V: X, /*Not=*/true, Q)) {
4583 return !ICmpInst::isSigned(Pred) || isKnownNonZero(V: X, Q);
4584 }
4585 return false;
4586 }
4587 if (ICmpInst::isEquality(P: Pred) && match(V: Op1, P: m_AllOnes()) &&
4588 match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Value(V&: M))))) {
4589
4590 auto Check = [&]() {
4591 // Look for: ~x | Mask == -1
4592 if (isMaskOrZero(V: M, /*Not=*/false, Q)) {
4593 if (Value *NotX =
4594 IC.getFreelyInverted(V: X, WillInvertAllUses: X->hasOneUse(), Builder: &IC.Builder)) {
4595 X = NotX;
4596 return true;
4597 }
4598 }
4599 return false;
4600 };
4601 if (Check())
4602 return true;
4603 std::swap(a&: X, b&: M);
4604 return Check();
4605 }
4606 if (ICmpInst::isEquality(P: Pred) && match(V: Op1, P: m_Zero()) &&
4607 match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_Value(V&: M))))) {
4608 auto Check = [&]() {
4609 // Look for: x & ~Mask == 0
4610 if (isMaskOrZero(V: M, /*Not=*/true, Q)) {
4611 if (Value *NotM =
4612 IC.getFreelyInverted(V: M, WillInvertAllUses: M->hasOneUse(), Builder: &IC.Builder)) {
4613 M = NotM;
4614 return true;
4615 }
4616 }
4617 return false;
4618 };
4619 if (Check())
4620 return true;
4621 std::swap(a&: X, b&: M);
4622 return Check();
4623 }
4624 return false;
4625 };
4626
4627 if (!IsLowBitMask())
4628 return nullptr;
4629
4630 return IC.Builder.CreateICmp(P: DstPred, LHS: X, RHS: M);
4631}
4632
4633/// Some comparisons can be simplified.
4634/// In this case, we are looking for comparisons that look like
4635/// a check for a lossy signed truncation.
4636/// Folds: (MaskedBits is a constant.)
4637/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
4638/// Into:
4639/// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
4640/// Where KeptBits = bitwidth(%x) - MaskedBits
4641static Value *
4642foldICmpWithTruncSignExtendedVal(ICmpInst &I,
4643 InstCombiner::BuilderTy &Builder) {
4644 CmpPredicate SrcPred;
4645 Value *X;
4646 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
4647 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
4648 if (!match(V: &I, P: m_c_ICmp(Pred&: SrcPred,
4649 L: m_OneUse(SubPattern: m_AShr(L: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C0)),
4650 R: m_APInt(Res&: C1))),
4651 R: m_Deferred(V: X))))
4652 return nullptr;
4653
4654 // Potential handling of non-splats: for each element:
4655 // * if both are undef, replace with constant 0.
4656 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
4657 // * if both are not undef, and are different, bailout.
4658 // * else, only one is undef, then pick the non-undef one.
4659
4660 // The shift amount must be equal.
4661 if (*C0 != *C1)
4662 return nullptr;
4663 const APInt &MaskedBits = *C0;
4664 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
4665
4666 ICmpInst::Predicate DstPred;
4667 switch (SrcPred) {
4668 case ICmpInst::Predicate::ICMP_EQ:
4669 // ((%x << MaskedBits) a>> MaskedBits) == %x
4670 // =>
4671 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
4672 DstPred = ICmpInst::Predicate::ICMP_ULT;
4673 break;
4674 case ICmpInst::Predicate::ICMP_NE:
4675 // ((%x << MaskedBits) a>> MaskedBits) != %x
4676 // =>
4677 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
4678 DstPred = ICmpInst::Predicate::ICMP_UGE;
4679 break;
4680 // FIXME: are more folds possible?
4681 default:
4682 return nullptr;
4683 }
4684
4685 auto *XType = X->getType();
4686 const unsigned XBitWidth = XType->getScalarSizeInBits();
4687 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
4688 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
4689
4690 // KeptBits = bitwidth(%x) - MaskedBits
4691 const APInt KeptBits = BitWidth - MaskedBits;
4692 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
4693 // ICmpCst = (1 << KeptBits)
4694 const APInt ICmpCst = APInt(XBitWidth, 1).shl(ShiftAmt: KeptBits);
4695 assert(ICmpCst.isPowerOf2());
4696 // AddCst = (1 << (KeptBits-1))
4697 const APInt AddCst = ICmpCst.lshr(shiftAmt: 1);
4698 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
4699
4700 // T0 = add %x, AddCst
4701 Value *T0 = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: XType, V: AddCst));
4702 // T1 = T0 DstPred ICmpCst
4703 Value *T1 = Builder.CreateICmp(P: DstPred, LHS: T0, RHS: ConstantInt::get(Ty: XType, V: ICmpCst));
4704
4705 return T1;
4706}
4707
4708// Given pattern:
4709// icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
4710// we should move shifts to the same hand of 'and', i.e. rewrite as
4711// icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
4712// We are only interested in opposite logical shifts here.
4713// One of the shifts can be truncated.
4714// If we can, we want to end up creating 'lshr' shift.
4715static Value *
4716foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
4717 InstCombiner::BuilderTy &Builder) {
4718 if (!I.isEquality() || !match(V: I.getOperand(i_nocapture: 1), P: m_Zero()) ||
4719 !I.getOperand(i_nocapture: 0)->hasOneUse())
4720 return nullptr;
4721
4722 auto m_AnyLogicalShift = m_LogicalShift(L: m_Value(), R: m_Value());
4723
4724 // Look for an 'and' of two logical shifts, one of which may be truncated.
4725 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
4726 Instruction *XShift, *MaybeTruncation, *YShift;
4727 if (!match(
4728 V: I.getOperand(i_nocapture: 0),
4729 P: m_c_And(L: m_CombineAnd(L: m_AnyLogicalShift, R: m_Instruction(I&: XShift)),
4730 R: m_CombineAnd(L: m_TruncOrSelf(Op: m_CombineAnd(
4731 L: m_AnyLogicalShift, R: m_Instruction(I&: YShift))),
4732 R: m_Instruction(I&: MaybeTruncation)))))
4733 return nullptr;
4734
4735 // We potentially looked past 'trunc', but only when matching YShift,
4736 // therefore YShift must have the widest type.
4737 Instruction *WidestShift = YShift;
4738 // Therefore XShift must have the shallowest type.
4739 // Or they both have identical types if there was no truncation.
4740 Instruction *NarrowestShift = XShift;
4741
4742 Type *WidestTy = WidestShift->getType();
4743 Type *NarrowestTy = NarrowestShift->getType();
4744 assert(NarrowestTy == I.getOperand(0)->getType() &&
4745 "We did not look past any shifts while matching XShift though.");
4746 bool HadTrunc = WidestTy != I.getOperand(i_nocapture: 0)->getType();
4747
4748 // If YShift is a 'lshr', swap the shifts around.
4749 if (match(V: YShift, P: m_LShr(L: m_Value(), R: m_Value())))
4750 std::swap(a&: XShift, b&: YShift);
4751
4752 // The shifts must be in opposite directions.
4753 auto XShiftOpcode = XShift->getOpcode();
4754 if (XShiftOpcode == YShift->getOpcode())
4755 return nullptr; // Do not care about same-direction shifts here.
4756
4757 Value *X, *XShAmt, *Y, *YShAmt;
4758 match(V: XShift, P: m_BinOp(L: m_Value(V&: X), R: m_ZExtOrSelf(Op: m_Value(V&: XShAmt))));
4759 match(V: YShift, P: m_BinOp(L: m_Value(V&: Y), R: m_ZExtOrSelf(Op: m_Value(V&: YShAmt))));
4760
4761 // If one of the values being shifted is a constant, then we will end with
4762 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
4763 // however, we will need to ensure that we won't increase instruction count.
4764 if (!isa<Constant>(Val: X) && !isa<Constant>(Val: Y)) {
4765 // At least one of the hands of the 'and' should be one-use shift.
4766 if (!match(V: I.getOperand(i_nocapture: 0),
4767 P: m_c_And(L: m_OneUse(SubPattern: m_AnyLogicalShift), R: m_Value())))
4768 return nullptr;
4769 if (HadTrunc) {
4770 // Due to the 'trunc', we will need to widen X. For that either the old
4771 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
4772 if (!MaybeTruncation->hasOneUse() &&
4773 !NarrowestShift->getOperand(i: 1)->hasOneUse())
4774 return nullptr;
4775 }
4776 }
4777
4778 // We have two shift amounts from two different shifts. The types of those
4779 // shift amounts may not match. If that's the case let's bailout now.
4780 if (XShAmt->getType() != YShAmt->getType())
4781 return nullptr;
4782
4783 // As input, we have the following pattern:
4784 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
4785 // We want to rewrite that as:
4786 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
4787 // While we know that originally (Q+K) would not overflow
4788 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
4789 // shift amounts. so it may now overflow in smaller bitwidth.
4790 // To ensure that does not happen, we need to ensure that the total maximal
4791 // shift amount is still representable in that smaller bit width.
4792 unsigned MaximalPossibleTotalShiftAmount =
4793 (WidestTy->getScalarSizeInBits() - 1) +
4794 (NarrowestTy->getScalarSizeInBits() - 1);
4795 APInt MaximalRepresentableShiftAmount =
4796 APInt::getAllOnes(numBits: XShAmt->getType()->getScalarSizeInBits());
4797 if (MaximalRepresentableShiftAmount.ult(RHS: MaximalPossibleTotalShiftAmount))
4798 return nullptr;
4799
4800 // Can we fold (XShAmt+YShAmt) ?
4801 auto *NewShAmt = dyn_cast_or_null<Constant>(
4802 Val: simplifyAddInst(LHS: XShAmt, RHS: YShAmt, /*isNSW=*/IsNSW: false,
4803 /*isNUW=*/IsNUW: false, Q: SQ.getWithInstruction(I: &I)));
4804 if (!NewShAmt)
4805 return nullptr;
4806 if (NewShAmt->getType() != WidestTy) {
4807 NewShAmt =
4808 ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: NewShAmt, DestTy: WidestTy, DL: SQ.DL);
4809 if (!NewShAmt)
4810 return nullptr;
4811 }
4812 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
4813
4814 // Is the new shift amount smaller than the bit width?
4815 // FIXME: could also rely on ConstantRange.
4816 if (!match(V: NewShAmt,
4817 P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_ULT,
4818 Threshold: APInt(WidestBitWidth, WidestBitWidth))))
4819 return nullptr;
4820
4821 // An extra legality check is needed if we had trunc-of-lshr.
4822 if (HadTrunc && match(V: WidestShift, P: m_LShr(L: m_Value(), R: m_Value()))) {
4823 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
4824 WidestShift]() {
4825 // It isn't obvious whether it's worth it to analyze non-constants here.
4826 // Also, let's basically give up on non-splat cases, pessimizing vectors.
4827 // If *any* of these preconditions matches we can perform the fold.
4828 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
4829 ? NewShAmt->getSplatValue()
4830 : NewShAmt;
4831 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
4832 if (NewShAmtSplat &&
4833 (NewShAmtSplat->isNullValue() ||
4834 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
4835 return true;
4836 // We consider *min* leading zeros so a single outlier
4837 // blocks the transform as opposed to allowing it.
4838 if (auto *C = dyn_cast<Constant>(Val: NarrowestShift->getOperand(i: 0))) {
4839 KnownBits Known = computeKnownBits(V: C, DL: SQ.DL);
4840 unsigned MinLeadZero = Known.countMinLeadingZeros();
4841 // If the value being shifted has at most lowest bit set we can fold.
4842 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
4843 if (MaxActiveBits <= 1)
4844 return true;
4845 // Precondition: NewShAmt u<= countLeadingZeros(C)
4846 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(RHS: MinLeadZero))
4847 return true;
4848 }
4849 if (auto *C = dyn_cast<Constant>(Val: WidestShift->getOperand(i: 0))) {
4850 KnownBits Known = computeKnownBits(V: C, DL: SQ.DL);
4851 unsigned MinLeadZero = Known.countMinLeadingZeros();
4852 // If the value being shifted has at most lowest bit set we can fold.
4853 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
4854 if (MaxActiveBits <= 1)
4855 return true;
4856 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
4857 if (NewShAmtSplat) {
4858 APInt AdjNewShAmt =
4859 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
4860 if (AdjNewShAmt.ule(RHS: MinLeadZero))
4861 return true;
4862 }
4863 }
4864 return false; // Can't tell if it's ok.
4865 };
4866 if (!CanFold())
4867 return nullptr;
4868 }
4869
4870 // All good, we can do this fold.
4871 X = Builder.CreateZExt(V: X, DestTy: WidestTy);
4872 Y = Builder.CreateZExt(V: Y, DestTy: WidestTy);
4873 // The shift is the same that was for X.
4874 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
4875 ? Builder.CreateLShr(LHS: X, RHS: NewShAmt)
4876 : Builder.CreateShl(LHS: X, RHS: NewShAmt);
4877 Value *T1 = Builder.CreateAnd(LHS: T0, RHS: Y);
4878 return Builder.CreateICmp(P: I.getPredicate(), LHS: T1,
4879 RHS: Constant::getNullValue(Ty: WidestTy));
4880}
4881
4882/// Fold
4883/// (-1 u/ x) u< y
4884/// ((x * y) ?/ x) != y
4885/// to
4886/// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
4887/// Note that the comparison is commutative, while inverted (u>=, ==) predicate
4888/// will mean that we are looking for the opposite answer.
4889Value *InstCombinerImpl::foldMultiplicationOverflowCheck(ICmpInst &I) {
4890 CmpPredicate Pred;
4891 Value *X, *Y;
4892 Instruction *Mul;
4893 Instruction *Div;
4894 bool NeedNegation;
4895 // Look for: (-1 u/ x) u</u>= y
4896 if (!I.isEquality() &&
4897 match(V: &I, P: m_c_ICmp(Pred,
4898 L: m_CombineAnd(L: m_OneUse(SubPattern: m_UDiv(L: m_AllOnes(), R: m_Value(V&: X))),
4899 R: m_Instruction(I&: Div)),
4900 R: m_Value(V&: Y)))) {
4901 Mul = nullptr;
4902
4903 // Are we checking that overflow does not happen, or does happen?
4904 switch (Pred) {
4905 case ICmpInst::Predicate::ICMP_ULT:
4906 NeedNegation = false;
4907 break; // OK
4908 case ICmpInst::Predicate::ICMP_UGE:
4909 NeedNegation = true;
4910 break; // OK
4911 default:
4912 return nullptr; // Wrong predicate.
4913 }
4914 } else // Look for: ((x * y) / x) !=/== y
4915 if (I.isEquality() &&
4916 match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: Y),
4917 R: m_CombineAnd(L: m_OneUse(SubPattern: m_IDiv(
4918 L: m_CombineAnd(L: m_c_Mul(L: m_Deferred(V: Y),
4919 R: m_Value(V&: X)),
4920 R: m_Instruction(I&: Mul)),
4921 R: m_Deferred(V: X))),
4922 R: m_Instruction(I&: Div))))) {
4923 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
4924 } else
4925 return nullptr;
4926
4927 BuilderTy::InsertPointGuard Guard(Builder);
4928 // If the pattern included (x * y), we'll want to insert new instructions
4929 // right before that original multiplication so that we can replace it.
4930 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
4931 if (MulHadOtherUses)
4932 Builder.SetInsertPoint(Mul);
4933
4934 CallInst *Call = Builder.CreateIntrinsic(
4935 ID: Div->getOpcode() == Instruction::UDiv ? Intrinsic::umul_with_overflow
4936 : Intrinsic::smul_with_overflow,
4937 Types: X->getType(), Args: {X, Y}, /*FMFSource=*/nullptr, Name: "mul");
4938
4939 // If the multiplication was used elsewhere, to ensure that we don't leave
4940 // "duplicate" instructions, replace uses of that original multiplication
4941 // with the multiplication result from the with.overflow intrinsic.
4942 if (MulHadOtherUses)
4943 replaceInstUsesWith(I&: *Mul, V: Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "mul.val"));
4944
4945 Value *Res = Builder.CreateExtractValue(Agg: Call, Idxs: 1, Name: "mul.ov");
4946 if (NeedNegation) // This technically increases instruction count.
4947 Res = Builder.CreateNot(V: Res, Name: "mul.not.ov");
4948
4949 // If we replaced the mul, erase it. Do this after all uses of Builder,
4950 // as the mul is used as insertion point.
4951 if (MulHadOtherUses)
4952 eraseInstFromFunction(I&: *Mul);
4953
4954 return Res;
4955}
4956
4957static Instruction *foldICmpXNegX(ICmpInst &I,
4958 InstCombiner::BuilderTy &Builder) {
4959 CmpPredicate Pred;
4960 Value *X;
4961 if (match(V: &I, P: m_c_ICmp(Pred, L: m_NSWNeg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) {
4962
4963 if (ICmpInst::isSigned(Pred))
4964 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
4965 else if (ICmpInst::isUnsigned(Pred))
4966 Pred = ICmpInst::getSignedPredicate(Pred);
4967 // else for equality-comparisons just keep the predicate.
4968
4969 return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: X,
4970 S2: Constant::getNullValue(Ty: X->getType()), Name: I.getName());
4971 }
4972
4973 // A value is not equal to its negation unless that value is 0 or
4974 // MinSignedValue, ie: a != -a --> (a & MaxSignedVal) != 0
4975 if (match(V: &I, P: m_c_ICmp(Pred, L: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))), R: m_Deferred(V: X))) &&
4976 ICmpInst::isEquality(P: Pred)) {
4977 Type *Ty = X->getType();
4978 uint32_t BitWidth = Ty->getScalarSizeInBits();
4979 Constant *MaxSignedVal =
4980 ConstantInt::get(Ty, V: APInt::getSignedMaxValue(numBits: BitWidth));
4981 Value *And = Builder.CreateAnd(LHS: X, RHS: MaxSignedVal);
4982 Constant *Zero = Constant::getNullValue(Ty);
4983 return CmpInst::Create(Op: Instruction::ICmp, Pred, S1: And, S2: Zero);
4984 }
4985
4986 return nullptr;
4987}
4988
4989static Instruction *foldICmpAndXX(ICmpInst &I, const SimplifyQuery &Q,
4990 InstCombinerImpl &IC) {
4991 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A;
4992 // Normalize and operand as operand 0.
4993 CmpInst::Predicate Pred = I.getPredicate();
4994 if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value()))) {
4995 std::swap(a&: Op0, b&: Op1);
4996 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
4997 }
4998
4999 if (!match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value(V&: A))))
5000 return nullptr;
5001
5002 // (icmp (X & Y) u< X --> (X & Y) != X
5003 if (Pred == ICmpInst::ICMP_ULT)
5004 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5005
5006 // (icmp (X & Y) u>= X --> (X & Y) == X
5007 if (Pred == ICmpInst::ICMP_UGE)
5008 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5009
5010 if (ICmpInst::isEquality(P: Pred) && Op0->hasOneUse()) {
5011 // icmp (X & Y) eq/ne Y --> (X | ~Y) eq/ne -1 if Y is freely invertible and
5012 // Y is non-constant. If Y is constant the `X & C == C` form is preferable
5013 // so don't do this fold.
5014 if (!match(V: Op1, P: m_ImmConstant()))
5015 if (auto *NotOp1 =
5016 IC.getFreelyInverted(V: Op1, WillInvertAllUses: !Op1->hasNUsesOrMore(N: 3), Builder: &IC.Builder))
5017 return new ICmpInst(Pred, IC.Builder.CreateOr(LHS: A, RHS: NotOp1),
5018 Constant::getAllOnesValue(Ty: Op1->getType()));
5019 // icmp (X & Y) eq/ne Y --> (~X & Y) eq/ne 0 if X is freely invertible.
5020 if (auto *NotA = IC.getFreelyInverted(V: A, WillInvertAllUses: A->hasOneUse(), Builder: &IC.Builder))
5021 return new ICmpInst(Pred, IC.Builder.CreateAnd(LHS: Op1, RHS: NotA),
5022 Constant::getNullValue(Ty: Op1->getType()));
5023 }
5024
5025 if (!ICmpInst::isSigned(Pred))
5026 return nullptr;
5027
5028 KnownBits KnownY = IC.computeKnownBits(V: A, CxtI: &I);
5029 // (X & NegY) spred X --> (X & NegY) upred X
5030 if (KnownY.isNegative())
5031 return new ICmpInst(ICmpInst::getUnsignedPredicate(Pred), Op0, Op1);
5032
5033 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGT)
5034 return nullptr;
5035
5036 if (KnownY.isNonNegative())
5037 // (X & PosY) s<= X --> X s>= 0
5038 // (X & PosY) s> X --> X s< 0
5039 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
5040 Constant::getNullValue(Ty: Op1->getType()));
5041
5042 if (isKnownNegative(V: Op1, SQ: IC.getSimplifyQuery().getWithInstruction(I: &I)))
5043 // (NegX & Y) s<= NegX --> Y s< 0
5044 // (NegX & Y) s> NegX --> Y s>= 0
5045 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), A,
5046 Constant::getNullValue(Ty: A->getType()));
5047
5048 return nullptr;
5049}
5050
5051static Instruction *foldICmpOrXX(ICmpInst &I, const SimplifyQuery &Q,
5052 InstCombinerImpl &IC) {
5053 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A;
5054
5055 // Normalize or operand as operand 0.
5056 CmpInst::Predicate Pred = I.getPredicate();
5057 if (match(V: Op1, P: m_c_Or(L: m_Specific(V: Op0), R: m_Value(V&: A)))) {
5058 std::swap(a&: Op0, b&: Op1);
5059 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
5060 } else if (!match(V: Op0, P: m_c_Or(L: m_Specific(V: Op1), R: m_Value(V&: A)))) {
5061 return nullptr;
5062 }
5063
5064 // icmp (X | Y) u<= X --> (X | Y) == X
5065 if (Pred == ICmpInst::ICMP_ULE)
5066 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5067
5068 // icmp (X | Y) u> X --> (X | Y) != X
5069 if (Pred == ICmpInst::ICMP_UGT)
5070 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5071
5072 if (ICmpInst::isEquality(P: Pred) && Op0->hasOneUse()) {
5073 // icmp (X | Y) eq/ne Y --> (X & ~Y) eq/ne 0 if Y is freely invertible
5074 if (Value *NotOp1 = IC.getFreelyInverted(
5075 V: Op1, WillInvertAllUses: !isa<Constant>(Val: Op1) && !Op1->hasNUsesOrMore(N: 3), Builder: &IC.Builder))
5076 return new ICmpInst(Pred, IC.Builder.CreateAnd(LHS: A, RHS: NotOp1),
5077 Constant::getNullValue(Ty: Op1->getType()));
5078 // icmp (X | Y) eq/ne Y --> (~X | Y) eq/ne -1 if X is freely invertible.
5079 if (Value *NotA = IC.getFreelyInverted(V: A, WillInvertAllUses: A->hasOneUse(), Builder: &IC.Builder))
5080 return new ICmpInst(Pred, IC.Builder.CreateOr(LHS: Op1, RHS: NotA),
5081 Constant::getAllOnesValue(Ty: Op1->getType()));
5082 }
5083 return nullptr;
5084}
5085
5086static Instruction *foldICmpXorXX(ICmpInst &I, const SimplifyQuery &Q,
5087 InstCombinerImpl &IC) {
5088 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A;
5089 // Normalize xor operand as operand 0.
5090 CmpInst::Predicate Pred = I.getPredicate();
5091 if (match(V: Op1, P: m_c_Xor(L: m_Specific(V: Op0), R: m_Value()))) {
5092 std::swap(a&: Op0, b&: Op1);
5093 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
5094 }
5095 if (!match(V: Op0, P: m_c_Xor(L: m_Specific(V: Op1), R: m_Value(V&: A))))
5096 return nullptr;
5097
5098 // icmp (X ^ Y_NonZero) u>= X --> icmp (X ^ Y_NonZero) u> X
5099 // icmp (X ^ Y_NonZero) u<= X --> icmp (X ^ Y_NonZero) u< X
5100 // icmp (X ^ Y_NonZero) s>= X --> icmp (X ^ Y_NonZero) s> X
5101 // icmp (X ^ Y_NonZero) s<= X --> icmp (X ^ Y_NonZero) s< X
5102 CmpInst::Predicate PredOut = CmpInst::getStrictPredicate(pred: Pred);
5103 if (PredOut != Pred && isKnownNonZero(V: A, Q))
5104 return new ICmpInst(PredOut, Op0, Op1);
5105
5106 // These transform work when A is negative.
5107 // X s< X^A, X s<= X^A, X u> X^A, X u>= X^A --> X s< 0
5108 // X s> X^A, X s>= X^A, X u< X^A, X u<= X^A --> X s>= 0
5109 if (match(V: A, P: m_Negative())) {
5110 CmpInst::Predicate NewPred;
5111 switch (ICmpInst::getStrictPredicate(pred: Pred)) {
5112 default:
5113 return nullptr;
5114 case ICmpInst::ICMP_SLT:
5115 case ICmpInst::ICMP_UGT:
5116 NewPred = ICmpInst::ICMP_SLT;
5117 break;
5118 case ICmpInst::ICMP_SGT:
5119 case ICmpInst::ICMP_ULT:
5120 NewPred = ICmpInst::ICMP_SGE;
5121 break;
5122 }
5123 Constant *Const = Constant::getNullValue(Ty: Op0->getType());
5124 return new ICmpInst(NewPred, Op0, Const);
5125 }
5126
5127 return nullptr;
5128}
5129
5130/// Return true if X is a multiple of C.
5131/// TODO: Handle non-power-of-2 factors.
5132static bool isMultipleOf(Value *X, const APInt &C, const SimplifyQuery &Q) {
5133 if (C.isOne())
5134 return true;
5135
5136 if (!C.isPowerOf2())
5137 return false;
5138
5139 return MaskedValueIsZero(V: X, Mask: C - 1, SQ: Q);
5140}
5141
5142/// Try to fold icmp (binop), X or icmp X, (binop).
5143/// TODO: A large part of this logic is duplicated in InstSimplify's
5144/// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
5145/// duplication.
5146Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
5147 const SimplifyQuery &SQ) {
5148 const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
5149 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
5150
5151 // Special logic for binary operators.
5152 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Val: Op0);
5153 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Val: Op1);
5154 if (!BO0 && !BO1)
5155 return nullptr;
5156
5157 if (Instruction *NewICmp = foldICmpXNegX(I, Builder))
5158 return NewICmp;
5159
5160 const CmpInst::Predicate Pred = I.getPredicate();
5161 Value *X;
5162
5163 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
5164 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
5165 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Op1), R: m_Value(V&: X)))) &&
5166 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
5167 return new ICmpInst(Pred, Builder.CreateNot(V: Op1), X);
5168 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
5169 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: X)))) &&
5170 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
5171 return new ICmpInst(Pred, X, Builder.CreateNot(V: Op0));
5172
5173 {
5174 // (Op1 + X) + C u</u>= Op1 --> ~C - X u</u>= Op1
5175 Constant *C;
5176 if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_c_Add(L: m_Specific(V: Op1), R: m_Value(V&: X)),
5177 R: m_ImmConstant(C)))) &&
5178 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
5179 Constant *C2 = ConstantExpr::getNot(C);
5180 return new ICmpInst(Pred, Builder.CreateSub(LHS: C2, RHS: X), Op1);
5181 }
5182 // Op0 u>/u<= (Op0 + X) + C --> Op0 u>/u<= ~C - X
5183 if (match(V: Op1, P: m_OneUse(SubPattern: m_Add(L: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: X)),
5184 R: m_ImmConstant(C)))) &&
5185 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE)) {
5186 Constant *C2 = ConstantExpr::getNot(C);
5187 return new ICmpInst(Pred, Op0, Builder.CreateSub(LHS: C2, RHS: X));
5188 }
5189 }
5190
5191 // (icmp eq/ne (X, -P2), INT_MIN)
5192 // -> (icmp slt/sge X, INT_MIN + P2)
5193 if (ICmpInst::isEquality(P: Pred) && BO0 &&
5194 match(V: I.getOperand(i_nocapture: 1), P: m_SignMask()) &&
5195 match(V: BO0, P: m_And(L: m_Value(), R: m_NegatedPower2OrZero()))) {
5196 // Will Constant fold.
5197 Value *NewC = Builder.CreateSub(LHS: I.getOperand(i_nocapture: 1), RHS: BO0->getOperand(i_nocapture: 1));
5198 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SLT
5199 : ICmpInst::ICMP_SGE,
5200 BO0->getOperand(i_nocapture: 0), NewC);
5201 }
5202
5203 {
5204 // Similar to above: an unsigned overflow comparison may use offset + mask:
5205 // ((Op1 + C) & C) u< Op1 --> Op1 != 0
5206 // ((Op1 + C) & C) u>= Op1 --> Op1 == 0
5207 // Op0 u> ((Op0 + C) & C) --> Op0 != 0
5208 // Op0 u<= ((Op0 + C) & C) --> Op0 == 0
5209 BinaryOperator *BO;
5210 const APInt *C;
5211 if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) &&
5212 match(V: Op0, P: m_And(L: m_BinOp(I&: BO), R: m_LowBitMask(V&: C))) &&
5213 match(V: BO, P: m_Add(L: m_Specific(V: Op1), R: m_SpecificIntAllowPoison(V: *C)))) {
5214 CmpInst::Predicate NewPred =
5215 Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
5216 Constant *Zero = ConstantInt::getNullValue(Ty: Op1->getType());
5217 return new ICmpInst(NewPred, Op1, Zero);
5218 }
5219
5220 if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
5221 match(V: Op1, P: m_And(L: m_BinOp(I&: BO), R: m_LowBitMask(V&: C))) &&
5222 match(V: BO, P: m_Add(L: m_Specific(V: Op0), R: m_SpecificIntAllowPoison(V: *C)))) {
5223 CmpInst::Predicate NewPred =
5224 Pred == ICmpInst::ICMP_UGT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
5225 Constant *Zero = ConstantInt::getNullValue(Ty: Op1->getType());
5226 return new ICmpInst(NewPred, Op0, Zero);
5227 }
5228 }
5229
5230 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
5231 bool Op0HasNUW = false, Op1HasNUW = false;
5232 bool Op0HasNSW = false, Op1HasNSW = false;
5233 // Analyze the case when either Op0 or Op1 is an add instruction.
5234 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
5235 auto hasNoWrapProblem = [](const BinaryOperator &BO, CmpInst::Predicate Pred,
5236 bool &HasNSW, bool &HasNUW) -> bool {
5237 if (isa<OverflowingBinaryOperator>(Val: BO)) {
5238 HasNUW = BO.hasNoUnsignedWrap();
5239 HasNSW = BO.hasNoSignedWrap();
5240 return ICmpInst::isEquality(P: Pred) ||
5241 (CmpInst::isUnsigned(Pred) && HasNUW) ||
5242 (CmpInst::isSigned(Pred) && HasNSW);
5243 } else if (BO.getOpcode() == Instruction::Or) {
5244 HasNUW = true;
5245 HasNSW = true;
5246 return true;
5247 } else {
5248 return false;
5249 }
5250 };
5251 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
5252
5253 if (BO0) {
5254 match(V: BO0, P: m_AddLike(L: m_Value(V&: A), R: m_Value(V&: B)));
5255 NoOp0WrapProblem = hasNoWrapProblem(*BO0, Pred, Op0HasNSW, Op0HasNUW);
5256 }
5257 if (BO1) {
5258 match(V: BO1, P: m_AddLike(L: m_Value(V&: C), R: m_Value(V&: D)));
5259 NoOp1WrapProblem = hasNoWrapProblem(*BO1, Pred, Op1HasNSW, Op1HasNUW);
5260 }
5261
5262 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
5263 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
5264 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
5265 return new ICmpInst(Pred, A == Op1 ? B : A,
5266 Constant::getNullValue(Ty: Op1->getType()));
5267
5268 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
5269 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
5270 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
5271 return new ICmpInst(Pred, Constant::getNullValue(Ty: Op0->getType()),
5272 C == Op0 ? D : C);
5273
5274 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
5275 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
5276 NoOp1WrapProblem) {
5277 // Determine Y and Z in the form icmp (X+Y), (X+Z).
5278 Value *Y, *Z;
5279 if (A == C) {
5280 // C + B == C + D -> B == D
5281 Y = B;
5282 Z = D;
5283 } else if (A == D) {
5284 // D + B == C + D -> B == C
5285 Y = B;
5286 Z = C;
5287 } else if (B == C) {
5288 // A + C == C + D -> A == D
5289 Y = A;
5290 Z = D;
5291 } else {
5292 assert(B == D);
5293 // A + D == C + D -> A == C
5294 Y = A;
5295 Z = C;
5296 }
5297 return new ICmpInst(Pred, Y, Z);
5298 }
5299
5300 if (ICmpInst::isRelational(P: Pred)) {
5301 // Return if both X and Y is divisible by Z/-Z.
5302 // TODO: Generalize to check if (X - Y) is divisible by Z/-Z.
5303 auto ShareCommonDivisor = [&Q](Value *X, Value *Y, Value *Z,
5304 bool IsNegative) -> bool {
5305 const APInt *OffsetC;
5306 if (!match(V: Z, P: m_APInt(Res&: OffsetC)))
5307 return false;
5308
5309 // Fast path for Z == 1/-1.
5310 if (IsNegative ? OffsetC->isAllOnes() : OffsetC->isOne())
5311 return true;
5312
5313 APInt C = *OffsetC;
5314 if (IsNegative)
5315 C.negate();
5316 // Note: -INT_MIN is also negative.
5317 if (!C.isStrictlyPositive())
5318 return false;
5319
5320 return isMultipleOf(X, C, Q) && isMultipleOf(X: Y, C, Q);
5321 };
5322
5323 // TODO: The subtraction-related identities shown below also hold, but
5324 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
5325 // wouldn't happen even if they were implemented.
5326 //
5327 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
5328 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
5329 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
5330 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
5331
5332 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
5333 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
5334 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
5335 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
5336 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
5337 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
5338 if (A && NoOp0WrapProblem &&
5339 ShareCommonDivisor(A, Op1, B,
5340 ICmpInst::isLT(P: Pred) || ICmpInst::isGE(P: Pred)))
5341 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), A,
5342 Op1);
5343
5344 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
5345 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
5346 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
5347 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
5348 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
5349 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
5350 if (C && NoOp1WrapProblem &&
5351 ShareCommonDivisor(Op0, C, D,
5352 ICmpInst::isGT(P: Pred) || ICmpInst::isLE(P: Pred)))
5353 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), Op0,
5354 C);
5355 }
5356
5357 // if C1 has greater magnitude than C2:
5358 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
5359 // s.t. C3 = C1 - C2
5360 //
5361 // if C2 has greater magnitude than C1:
5362 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
5363 // s.t. C3 = C2 - C1
5364 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
5365 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) {
5366 const APInt *AP1, *AP2;
5367 // TODO: Support non-uniform vectors.
5368 // TODO: Allow poison passthrough if B or D's element is poison.
5369 if (match(V: B, P: m_APIntAllowPoison(Res&: AP1)) &&
5370 match(V: D, P: m_APIntAllowPoison(Res&: AP2)) &&
5371 AP1->isNegative() == AP2->isNegative()) {
5372 APInt AP1Abs = AP1->abs();
5373 APInt AP2Abs = AP2->abs();
5374 if (AP1Abs.uge(RHS: AP2Abs)) {
5375 APInt Diff = *AP1 - *AP2;
5376 Constant *C3 = Constant::getIntegerValue(Ty: BO0->getType(), V: Diff);
5377 Value *NewAdd = Builder.CreateAdd(
5378 LHS: A, RHS: C3, Name: "", HasNUW: Op0HasNUW && Diff.ule(RHS: *AP1), HasNSW: Op0HasNSW);
5379 return new ICmpInst(Pred, NewAdd, C);
5380 } else {
5381 APInt Diff = *AP2 - *AP1;
5382 Constant *C3 = Constant::getIntegerValue(Ty: BO0->getType(), V: Diff);
5383 Value *NewAdd = Builder.CreateAdd(
5384 LHS: C, RHS: C3, Name: "", HasNUW: Op1HasNUW && Diff.ule(RHS: *AP2), HasNSW: Op1HasNSW);
5385 return new ICmpInst(Pred, A, NewAdd);
5386 }
5387 }
5388 Constant *Cst1, *Cst2;
5389 if (match(V: B, P: m_ImmConstant(C&: Cst1)) && match(V: D, P: m_ImmConstant(C&: Cst2)) &&
5390 ICmpInst::isEquality(P: Pred)) {
5391 Constant *Diff = ConstantExpr::getSub(C1: Cst2, C2: Cst1);
5392 Value *NewAdd = Builder.CreateAdd(LHS: C, RHS: Diff);
5393 return new ICmpInst(Pred, A, NewAdd);
5394 }
5395 }
5396
5397 // Analyze the case when either Op0 or Op1 is a sub instruction.
5398 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
5399 A = nullptr;
5400 B = nullptr;
5401 C = nullptr;
5402 D = nullptr;
5403 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
5404 A = BO0->getOperand(i_nocapture: 0);
5405 B = BO0->getOperand(i_nocapture: 1);
5406 }
5407 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
5408 C = BO1->getOperand(i_nocapture: 0);
5409 D = BO1->getOperand(i_nocapture: 1);
5410 }
5411
5412 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
5413 if (A == Op1 && NoOp0WrapProblem)
5414 return new ICmpInst(Pred, Constant::getNullValue(Ty: Op1->getType()), B);
5415 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
5416 if (C == Op0 && NoOp1WrapProblem)
5417 return new ICmpInst(Pred, D, Constant::getNullValue(Ty: Op0->getType()));
5418
5419 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
5420 // (A - B) u>/u<= A --> B u>/u<= A
5421 if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
5422 return new ICmpInst(Pred, B, A);
5423 // C u</u>= (C - D) --> C u</u>= D
5424 if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
5425 return new ICmpInst(Pred, C, D);
5426 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
5427 if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
5428 isKnownNonZero(V: B, Q))
5429 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(pred: Pred), B, A);
5430 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
5431 if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
5432 isKnownNonZero(V: D, Q))
5433 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(pred: Pred), C, D);
5434
5435 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
5436 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
5437 return new ICmpInst(Pred, A, C);
5438
5439 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
5440 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
5441 return new ICmpInst(Pred, D, B);
5442
5443 // icmp (0-X) < cst --> x > -cst
5444 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
5445 Value *X;
5446 if (match(V: BO0, P: m_Neg(V: m_Value(V&: X))))
5447 if (Constant *RHSC = dyn_cast<Constant>(Val: Op1))
5448 if (RHSC->isNotMinSignedValue())
5449 return new ICmpInst(I.getSwappedPredicate(), X,
5450 ConstantExpr::getNeg(C: RHSC));
5451 }
5452
5453 if (Instruction *R = foldICmpXorXX(I, Q, IC&: *this))
5454 return R;
5455 if (Instruction *R = foldICmpOrXX(I, Q, IC&: *this))
5456 return R;
5457
5458 {
5459 // Try to remove shared multiplier from comparison:
5460 // X * Z pred Y * Z
5461 Value *X, *Y, *Z;
5462 if ((match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
5463 match(V: Op1, P: m_c_Mul(L: m_Specific(V: Z), R: m_Value(V&: Y)))) ||
5464 (match(V: Op0, P: m_Mul(L: m_Value(V&: Z), R: m_Value(V&: X))) &&
5465 match(V: Op1, P: m_c_Mul(L: m_Specific(V: Z), R: m_Value(V&: Y))))) {
5466 if (ICmpInst::isSigned(Pred)) {
5467 if (Op0HasNSW && Op1HasNSW) {
5468 KnownBits ZKnown = computeKnownBits(V: Z, CxtI: &I);
5469 if (ZKnown.isStrictlyPositive())
5470 return new ICmpInst(Pred, X, Y);
5471 if (ZKnown.isNegative())
5472 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), X, Y);
5473 Value *LessThan = simplifyICmpInst(Pred: ICmpInst::ICMP_SLT, LHS: X, RHS: Y,
5474 Q: SQ.getWithInstruction(I: &I));
5475 if (LessThan && match(V: LessThan, P: m_One()))
5476 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Z,
5477 Constant::getNullValue(Ty: Z->getType()));
5478 Value *GreaterThan = simplifyICmpInst(Pred: ICmpInst::ICMP_SGT, LHS: X, RHS: Y,
5479 Q: SQ.getWithInstruction(I: &I));
5480 if (GreaterThan && match(V: GreaterThan, P: m_One()))
5481 return new ICmpInst(Pred, Z, Constant::getNullValue(Ty: Z->getType()));
5482 }
5483 } else {
5484 bool NonZero;
5485 if (ICmpInst::isEquality(P: Pred)) {
5486 // If X != Y, fold (X *nw Z) eq/ne (Y *nw Z) -> Z eq/ne 0
5487 if (((Op0HasNSW && Op1HasNSW) || (Op0HasNUW && Op1HasNUW)) &&
5488 isKnownNonEqual(V1: X, V2: Y, SQ))
5489 return new ICmpInst(Pred, Z, Constant::getNullValue(Ty: Z->getType()));
5490
5491 KnownBits ZKnown = computeKnownBits(V: Z, CxtI: &I);
5492 // if Z % 2 != 0
5493 // X * Z eq/ne Y * Z -> X eq/ne Y
5494 if (ZKnown.countMaxTrailingZeros() == 0)
5495 return new ICmpInst(Pred, X, Y);
5496 NonZero = !ZKnown.One.isZero() || isKnownNonZero(V: Z, Q);
5497 // if Z != 0 and nsw(X * Z) and nsw(Y * Z)
5498 // X * Z eq/ne Y * Z -> X eq/ne Y
5499 if (NonZero && BO0 && BO1 && Op0HasNSW && Op1HasNSW)
5500 return new ICmpInst(Pred, X, Y);
5501 } else
5502 NonZero = isKnownNonZero(V: Z, Q);
5503
5504 // If Z != 0 and nuw(X * Z) and nuw(Y * Z)
5505 // X * Z u{lt/le/gt/ge}/eq/ne Y * Z -> X u{lt/le/gt/ge}/eq/ne Y
5506 if (NonZero && BO0 && BO1 && Op0HasNUW && Op1HasNUW)
5507 return new ICmpInst(Pred, X, Y);
5508 }
5509 }
5510 }
5511
5512 BinaryOperator *SRem = nullptr;
5513 // icmp (srem X, Y), Y
5514 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(i_nocapture: 1))
5515 SRem = BO0;
5516 // icmp Y, (srem X, Y)
5517 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
5518 Op0 == BO1->getOperand(i_nocapture: 1))
5519 SRem = BO1;
5520 if (SRem) {
5521 // We don't check hasOneUse to avoid increasing register pressure because
5522 // the value we use is the same value this instruction was already using.
5523 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(pred: Pred) : Pred) {
5524 default:
5525 break;
5526 case ICmpInst::ICMP_EQ:
5527 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
5528 case ICmpInst::ICMP_NE:
5529 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
5530 case ICmpInst::ICMP_SGT:
5531 case ICmpInst::ICMP_SGE:
5532 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(i_nocapture: 1),
5533 Constant::getAllOnesValue(Ty: SRem->getType()));
5534 case ICmpInst::ICMP_SLT:
5535 case ICmpInst::ICMP_SLE:
5536 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(i_nocapture: 1),
5537 Constant::getNullValue(Ty: SRem->getType()));
5538 }
5539 }
5540
5541 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
5542 (BO0->hasOneUse() || BO1->hasOneUse()) &&
5543 BO0->getOperand(i_nocapture: 1) == BO1->getOperand(i_nocapture: 1)) {
5544 switch (BO0->getOpcode()) {
5545 default:
5546 break;
5547 case Instruction::Add:
5548 case Instruction::Sub:
5549 case Instruction::Xor: {
5550 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
5551 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5552
5553 const APInt *C;
5554 if (match(V: BO0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) {
5555 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
5556 if (C->isSignMask()) {
5557 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
5558 return new ICmpInst(NewPred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5559 }
5560
5561 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
5562 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
5563 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
5564 NewPred = I.getSwappedPredicate(pred: NewPred);
5565 return new ICmpInst(NewPred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5566 }
5567 }
5568 break;
5569 }
5570 case Instruction::Mul: {
5571 if (!I.isEquality())
5572 break;
5573
5574 const APInt *C;
5575 if (match(V: BO0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero() &&
5576 !C->isOne()) {
5577 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
5578 // Mask = -1 >> count-trailing-zeros(C).
5579 if (unsigned TZs = C->countr_zero()) {
5580 Constant *Mask = ConstantInt::get(
5581 Ty: BO0->getType(),
5582 V: APInt::getLowBitsSet(numBits: C->getBitWidth(), loBitsSet: C->getBitWidth() - TZs));
5583 Value *And1 = Builder.CreateAnd(LHS: BO0->getOperand(i_nocapture: 0), RHS: Mask);
5584 Value *And2 = Builder.CreateAnd(LHS: BO1->getOperand(i_nocapture: 0), RHS: Mask);
5585 return new ICmpInst(Pred, And1, And2);
5586 }
5587 }
5588 break;
5589 }
5590 case Instruction::UDiv:
5591 case Instruction::LShr:
5592 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
5593 break;
5594 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5595
5596 case Instruction::SDiv:
5597 if (!(I.isEquality() || match(V: BO0->getOperand(i_nocapture: 1), P: m_NonNegative())) ||
5598 !BO0->isExact() || !BO1->isExact())
5599 break;
5600 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5601
5602 case Instruction::AShr:
5603 if (!BO0->isExact() || !BO1->isExact())
5604 break;
5605 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5606
5607 case Instruction::Shl: {
5608 bool NUW = Op0HasNUW && Op1HasNUW;
5609 bool NSW = Op0HasNSW && Op1HasNSW;
5610 if (!NUW && !NSW)
5611 break;
5612 if (!NSW && I.isSigned())
5613 break;
5614 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5615 }
5616 }
5617 }
5618
5619 if (BO0) {
5620 // Transform A & (L - 1) `ult` L --> L != 0
5621 auto LSubOne = m_Add(L: m_Specific(V: Op1), R: m_AllOnes());
5622 auto BitwiseAnd = m_c_And(L: m_Value(), R: LSubOne);
5623
5624 if (match(V: BO0, P: BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
5625 auto *Zero = Constant::getNullValue(Ty: BO0->getType());
5626 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
5627 }
5628 }
5629
5630 // For unsigned predicates / eq / ne:
5631 // icmp pred (x << 1), x --> icmp getSignedPredicate(pred) x, 0
5632 // icmp pred x, (x << 1) --> icmp getSignedPredicate(pred) 0, x
5633 if (!ICmpInst::isSigned(Pred)) {
5634 if (match(V: Op0, P: m_Shl(L: m_Specific(V: Op1), R: m_One())))
5635 return new ICmpInst(ICmpInst::getSignedPredicate(Pred), Op1,
5636 Constant::getNullValue(Ty: Op1->getType()));
5637 else if (match(V: Op1, P: m_Shl(L: m_Specific(V: Op0), R: m_One())))
5638 return new ICmpInst(ICmpInst::getSignedPredicate(Pred),
5639 Constant::getNullValue(Ty: Op0->getType()), Op0);
5640 }
5641
5642 if (Value *V = foldMultiplicationOverflowCheck(I))
5643 return replaceInstUsesWith(I, V);
5644
5645 if (Instruction *R = foldICmpAndXX(I, Q, IC&: *this))
5646 return R;
5647
5648 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
5649 return replaceInstUsesWith(I, V);
5650
5651 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
5652 return replaceInstUsesWith(I, V);
5653
5654 return nullptr;
5655}
5656
5657/// Fold icmp Pred min|max(X, Y), Z.
5658Instruction *InstCombinerImpl::foldICmpWithMinMax(Instruction &I,
5659 MinMaxIntrinsic *MinMax,
5660 Value *Z, CmpPredicate Pred) {
5661 Value *X = MinMax->getLHS();
5662 Value *Y = MinMax->getRHS();
5663 if (ICmpInst::isSigned(Pred) && !MinMax->isSigned())
5664 return nullptr;
5665 if (ICmpInst::isUnsigned(Pred) && MinMax->isSigned()) {
5666 // Revert the transform signed pred -> unsigned pred
5667 // TODO: We can flip the signedness of predicate if both operands of icmp
5668 // are negative.
5669 if (isKnownNonNegative(V: Z, SQ: SQ.getWithInstruction(I: &I)) &&
5670 isKnownNonNegative(V: MinMax, SQ: SQ.getWithInstruction(I: &I))) {
5671 Pred = ICmpInst::getFlippedSignednessPredicate(Pred);
5672 } else
5673 return nullptr;
5674 }
5675 SimplifyQuery Q = SQ.getWithInstruction(I: &I);
5676 auto IsCondKnownTrue = [](Value *Val) -> std::optional<bool> {
5677 if (!Val)
5678 return std::nullopt;
5679 if (match(V: Val, P: m_One()))
5680 return true;
5681 if (match(V: Val, P: m_Zero()))
5682 return false;
5683 return std::nullopt;
5684 };
5685 // Remove samesign here since it is illegal to keep it when we speculatively
5686 // execute comparisons. For example, `icmp samesign ult umax(X, -46), -32`
5687 // cannot be decomposed into `(icmp samesign ult X, -46) or (icmp samesign ult
5688 // -46, -32)`. `X` is allowed to be non-negative here.
5689 Pred = Pred.dropSameSign();
5690 auto CmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred, LHS: X, RHS: Z, Q));
5691 auto CmpYZ = IsCondKnownTrue(simplifyICmpInst(Pred, LHS: Y, RHS: Z, Q));
5692 if (!CmpXZ.has_value() && !CmpYZ.has_value())
5693 return nullptr;
5694 if (!CmpXZ.has_value()) {
5695 std::swap(a&: X, b&: Y);
5696 std::swap(lhs&: CmpXZ, rhs&: CmpYZ);
5697 }
5698
5699 auto FoldIntoCmpYZ = [&]() -> Instruction * {
5700 if (CmpYZ.has_value())
5701 return replaceInstUsesWith(I, V: ConstantInt::getBool(Ty: I.getType(), V: *CmpYZ));
5702 return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: Y, S2: Z);
5703 };
5704
5705 switch (Pred) {
5706 case ICmpInst::ICMP_EQ:
5707 case ICmpInst::ICMP_NE: {
5708 // If X == Z:
5709 // Expr Result
5710 // min(X, Y) == Z X <= Y
5711 // max(X, Y) == Z X >= Y
5712 // min(X, Y) != Z X > Y
5713 // max(X, Y) != Z X < Y
5714 if ((Pred == ICmpInst::ICMP_EQ) == *CmpXZ) {
5715 ICmpInst::Predicate NewPred =
5716 ICmpInst::getNonStrictPredicate(pred: MinMax->getPredicate());
5717 if (Pred == ICmpInst::ICMP_NE)
5718 NewPred = ICmpInst::getInversePredicate(pred: NewPred);
5719 return ICmpInst::Create(Op: Instruction::ICmp, Pred: NewPred, S1: X, S2: Y);
5720 }
5721 // Otherwise (X != Z):
5722 ICmpInst::Predicate NewPred = MinMax->getPredicate();
5723 auto MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred: NewPred, LHS: X, RHS: Z, Q));
5724 if (!MinMaxCmpXZ.has_value()) {
5725 std::swap(a&: X, b&: Y);
5726 std::swap(lhs&: CmpXZ, rhs&: CmpYZ);
5727 // Re-check pre-condition X != Z
5728 if (!CmpXZ.has_value() || (Pred == ICmpInst::ICMP_EQ) == *CmpXZ)
5729 break;
5730 MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred: NewPred, LHS: X, RHS: Z, Q));
5731 }
5732 if (!MinMaxCmpXZ.has_value())
5733 break;
5734 if (*MinMaxCmpXZ) {
5735 // Expr Fact Result
5736 // min(X, Y) == Z X < Z false
5737 // max(X, Y) == Z X > Z false
5738 // min(X, Y) != Z X < Z true
5739 // max(X, Y) != Z X > Z true
5740 return replaceInstUsesWith(
5741 I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred == ICmpInst::ICMP_NE));
5742 } else {
5743 // Expr Fact Result
5744 // min(X, Y) == Z X > Z Y == Z
5745 // max(X, Y) == Z X < Z Y == Z
5746 // min(X, Y) != Z X > Z Y != Z
5747 // max(X, Y) != Z X < Z Y != Z
5748 return FoldIntoCmpYZ();
5749 }
5750 break;
5751 }
5752 case ICmpInst::ICMP_SLT:
5753 case ICmpInst::ICMP_ULT:
5754 case ICmpInst::ICMP_SLE:
5755 case ICmpInst::ICMP_ULE:
5756 case ICmpInst::ICMP_SGT:
5757 case ICmpInst::ICMP_UGT:
5758 case ICmpInst::ICMP_SGE:
5759 case ICmpInst::ICMP_UGE: {
5760 bool IsSame = MinMax->getPredicate() == ICmpInst::getStrictPredicate(pred: Pred);
5761 if (*CmpXZ) {
5762 if (IsSame) {
5763 // Expr Fact Result
5764 // min(X, Y) < Z X < Z true
5765 // min(X, Y) <= Z X <= Z true
5766 // max(X, Y) > Z X > Z true
5767 // max(X, Y) >= Z X >= Z true
5768 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
5769 } else {
5770 // Expr Fact Result
5771 // max(X, Y) < Z X < Z Y < Z
5772 // max(X, Y) <= Z X <= Z Y <= Z
5773 // min(X, Y) > Z X > Z Y > Z
5774 // min(X, Y) >= Z X >= Z Y >= Z
5775 return FoldIntoCmpYZ();
5776 }
5777 } else {
5778 if (IsSame) {
5779 // Expr Fact Result
5780 // min(X, Y) < Z X >= Z Y < Z
5781 // min(X, Y) <= Z X > Z Y <= Z
5782 // max(X, Y) > Z X <= Z Y > Z
5783 // max(X, Y) >= Z X < Z Y >= Z
5784 return FoldIntoCmpYZ();
5785 } else {
5786 // Expr Fact Result
5787 // max(X, Y) < Z X >= Z false
5788 // max(X, Y) <= Z X > Z false
5789 // min(X, Y) > Z X <= Z false
5790 // min(X, Y) >= Z X < Z false
5791 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
5792 }
5793 }
5794 break;
5795 }
5796 default:
5797 break;
5798 }
5799
5800 return nullptr;
5801}
5802
5803/// Match and fold patterns like:
5804/// icmp eq/ne X, min(max(X, Lo), Hi)
5805/// which represents a range check and can be repsented as a ConstantRange.
5806///
5807/// For icmp eq, build ConstantRange [Lo, Hi + 1) and convert to:
5808/// (X - Lo) u< (Hi + 1 - Lo)
5809/// For icmp ne, build ConstantRange [Hi + 1, Lo) and convert to:
5810/// (X - (Hi + 1)) u< (Lo - (Hi + 1))
5811Instruction *InstCombinerImpl::foldICmpWithClamp(ICmpInst &I, Value *X,
5812 MinMaxIntrinsic *Min) {
5813 if (!I.isEquality() || !Min->hasOneUse() || !Min->isMin())
5814 return nullptr;
5815
5816 const APInt *Lo = nullptr, *Hi = nullptr;
5817 if (Min->isSigned()) {
5818 if (!match(V: Min->getLHS(), P: m_OneUse(SubPattern: m_SMax(L: m_Specific(V: X), R: m_APInt(Res&: Lo)))) ||
5819 !match(V: Min->getRHS(), P: m_APInt(Res&: Hi)) || !Lo->slt(RHS: *Hi))
5820 return nullptr;
5821 } else {
5822 if (!match(V: Min->getLHS(), P: m_OneUse(SubPattern: m_UMax(L: m_Specific(V: X), R: m_APInt(Res&: Lo)))) ||
5823 !match(V: Min->getRHS(), P: m_APInt(Res&: Hi)) || !Lo->ult(RHS: *Hi))
5824 return nullptr;
5825 }
5826
5827 ConstantRange CR = ConstantRange::getNonEmpty(Lower: *Lo, Upper: *Hi + 1);
5828 ICmpInst::Predicate Pred;
5829 APInt C, Offset;
5830 if (I.getPredicate() == ICmpInst::ICMP_EQ)
5831 CR.getEquivalentICmp(Pred, RHS&: C, Offset);
5832 else
5833 CR.inverse().getEquivalentICmp(Pred, RHS&: C, Offset);
5834
5835 if (!Offset.isZero())
5836 X = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: Offset));
5837
5838 return replaceInstUsesWith(
5839 I, V: Builder.CreateICmp(P: Pred, LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: C)));
5840}
5841
5842// Canonicalize checking for a power-of-2-or-zero value:
5843static Instruction *foldICmpPow2Test(ICmpInst &I,
5844 InstCombiner::BuilderTy &Builder) {
5845 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
5846 const CmpInst::Predicate Pred = I.getPredicate();
5847 Value *A = nullptr;
5848 bool CheckIs;
5849 if (I.isEquality()) {
5850 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
5851 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
5852 if (!match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Add(L: m_Value(V&: A), R: m_AllOnes()),
5853 R: m_Deferred(V: A)))) ||
5854 !match(V: Op1, P: m_ZeroInt()))
5855 A = nullptr;
5856
5857 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
5858 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
5859 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Neg(V: m_Specific(V: Op1)), R: m_Specific(V: Op1)))))
5860 A = Op1;
5861 else if (match(V: Op1,
5862 P: m_OneUse(SubPattern: m_c_And(L: m_Neg(V: m_Specific(V: Op0)), R: m_Specific(V: Op0)))))
5863 A = Op0;
5864
5865 CheckIs = Pred == ICmpInst::ICMP_EQ;
5866 } else if (ICmpInst::isUnsigned(Pred)) {
5867 // (A ^ (A-1)) u>= A --> ctpop(A) < 2 (two commuted variants)
5868 // ((A-1) ^ A) u< A --> ctpop(A) > 1 (two commuted variants)
5869
5870 if ((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
5871 match(V: Op0, P: m_OneUse(SubPattern: m_c_Xor(L: m_Add(L: m_Specific(V: Op1), R: m_AllOnes()),
5872 R: m_Specific(V: Op1))))) {
5873 A = Op1;
5874 CheckIs = Pred == ICmpInst::ICMP_UGE;
5875 } else if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
5876 match(V: Op1, P: m_OneUse(SubPattern: m_c_Xor(L: m_Add(L: m_Specific(V: Op0), R: m_AllOnes()),
5877 R: m_Specific(V: Op0))))) {
5878 A = Op0;
5879 CheckIs = Pred == ICmpInst::ICMP_ULE;
5880 }
5881 }
5882
5883 if (A) {
5884 Type *Ty = A->getType();
5885 CallInst *CtPop = Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V: A);
5886 return CheckIs ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop,
5887 ConstantInt::get(Ty, V: 2))
5888 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop,
5889 ConstantInt::get(Ty, V: 1));
5890 }
5891
5892 return nullptr;
5893}
5894
5895/// Find all possible pairs (BinOp, RHS) that BinOp V, RHS can be simplified.
5896using OffsetOp = std::pair<Instruction::BinaryOps, Value *>;
5897static void collectOffsetOp(Value *V, SmallVectorImpl<OffsetOp> &Offsets,
5898 bool AllowRecursion) {
5899 Instruction *Inst = dyn_cast<Instruction>(Val: V);
5900 if (!Inst || !Inst->hasOneUse())
5901 return;
5902
5903 switch (Inst->getOpcode()) {
5904 case Instruction::Add:
5905 Offsets.emplace_back(Args: Instruction::Sub, Args: Inst->getOperand(i: 1));
5906 Offsets.emplace_back(Args: Instruction::Sub, Args: Inst->getOperand(i: 0));
5907 break;
5908 case Instruction::Sub:
5909 Offsets.emplace_back(Args: Instruction::Add, Args: Inst->getOperand(i: 1));
5910 break;
5911 case Instruction::Xor:
5912 Offsets.emplace_back(Args: Instruction::Xor, Args: Inst->getOperand(i: 1));
5913 Offsets.emplace_back(Args: Instruction::Xor, Args: Inst->getOperand(i: 0));
5914 break;
5915 case Instruction::Shl:
5916 if (Inst->hasNoSignedWrap())
5917 Offsets.emplace_back(Args: Instruction::AShr, Args: Inst->getOperand(i: 1));
5918 if (Inst->hasNoUnsignedWrap())
5919 Offsets.emplace_back(Args: Instruction::LShr, Args: Inst->getOperand(i: 1));
5920 break;
5921 case Instruction::Select:
5922 if (AllowRecursion) {
5923 collectOffsetOp(V: Inst->getOperand(i: 1), Offsets, /*AllowRecursion=*/false);
5924 collectOffsetOp(V: Inst->getOperand(i: 2), Offsets, /*AllowRecursion=*/false);
5925 }
5926 break;
5927 default:
5928 break;
5929 }
5930}
5931
5932enum class OffsetKind { Invalid, Value, Select };
5933
5934struct OffsetResult {
5935 OffsetKind Kind;
5936 Value *V0, *V1, *V2;
5937 Instruction *MDFrom;
5938
5939 static OffsetResult invalid() {
5940 return {.Kind: OffsetKind::Invalid, .V0: nullptr, .V1: nullptr, .V2: nullptr, .MDFrom: nullptr};
5941 }
5942 static OffsetResult value(Value *V) {
5943 return {.Kind: OffsetKind::Value, .V0: V, .V1: nullptr, .V2: nullptr, .MDFrom: nullptr};
5944 }
5945 static OffsetResult select(Value *Cond, Value *TrueV, Value *FalseV,
5946 Instruction *MDFrom) {
5947 return {.Kind: OffsetKind::Select, .V0: Cond, .V1: TrueV, .V2: FalseV, .MDFrom: MDFrom};
5948 }
5949 bool isValid() const { return Kind != OffsetKind::Invalid; }
5950 Value *materialize(InstCombiner::BuilderTy &Builder) const {
5951 switch (Kind) {
5952 case OffsetKind::Invalid:
5953 llvm_unreachable("Invalid offset result");
5954 case OffsetKind::Value:
5955 return V0;
5956 case OffsetKind::Select:
5957 return Builder.CreateSelect(
5958 C: V0, True: V1, False: V2, Name: "", MDFrom: ProfcheckDisableMetadataFixes ? nullptr : MDFrom);
5959 }
5960 llvm_unreachable("Unknown OffsetKind enum");
5961 }
5962};
5963
5964/// Offset both sides of an equality icmp to see if we can save some
5965/// instructions: icmp eq/ne X, Y -> icmp eq/ne X op Z, Y op Z.
5966/// Note: This operation should not introduce poison.
5967static Instruction *foldICmpEqualityWithOffset(ICmpInst &I,
5968 InstCombiner::BuilderTy &Builder,
5969 const SimplifyQuery &SQ) {
5970 assert(I.isEquality() && "Expected an equality icmp");
5971 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
5972 if (!Op0->getType()->isIntOrIntVectorTy())
5973 return nullptr;
5974
5975 SmallVector<OffsetOp, 4> OffsetOps;
5976 collectOffsetOp(V: Op0, Offsets&: OffsetOps, /*AllowRecursion=*/true);
5977 collectOffsetOp(V: Op1, Offsets&: OffsetOps, /*AllowRecursion=*/true);
5978
5979 auto ApplyOffsetImpl = [&](Value *V, unsigned BinOpc, Value *RHS) -> Value * {
5980 switch (BinOpc) {
5981 // V = shl nsw X, RHS => X = ashr V, RHS
5982 case Instruction::AShr: {
5983 const APInt *CV, *CRHS;
5984 if (!(match(V, P: m_APInt(Res&: CV)) && match(V: RHS, P: m_APInt(Res&: CRHS)) &&
5985 CV->ashr(ShiftAmt: *CRHS).shl(ShiftAmt: *CRHS) == *CV) &&
5986 !match(V, P: m_NSWShl(L: m_Value(), R: m_Specific(V: RHS))))
5987 return nullptr;
5988 break;
5989 }
5990 // V = shl nuw X, RHS => X = lshr V, RHS
5991 case Instruction::LShr: {
5992 const APInt *CV, *CRHS;
5993 if (!(match(V, P: m_APInt(Res&: CV)) && match(V: RHS, P: m_APInt(Res&: CRHS)) &&
5994 CV->lshr(ShiftAmt: *CRHS).shl(ShiftAmt: *CRHS) == *CV) &&
5995 !match(V, P: m_NUWShl(L: m_Value(), R: m_Specific(V: RHS))))
5996 return nullptr;
5997 break;
5998 }
5999 default:
6000 break;
6001 }
6002
6003 Value *Simplified = simplifyBinOp(Opcode: BinOpc, LHS: V, RHS, Q: SQ);
6004 if (!Simplified)
6005 return nullptr;
6006 // Reject constant expressions as they don't simplify things.
6007 if (isa<Constant>(Val: Simplified) && !match(V: Simplified, P: m_ImmConstant()))
6008 return nullptr;
6009 // Check if the transformation introduces poison.
6010 return impliesPoison(ValAssumedPoison: RHS, V) ? Simplified : nullptr;
6011 };
6012
6013 auto ApplyOffset = [&](Value *V, unsigned BinOpc,
6014 Value *RHS) -> OffsetResult {
6015 if (auto *Sel = dyn_cast<SelectInst>(Val: V)) {
6016 if (!Sel->hasOneUse())
6017 return OffsetResult::invalid();
6018 Value *TrueVal = ApplyOffsetImpl(Sel->getTrueValue(), BinOpc, RHS);
6019 if (!TrueVal)
6020 return OffsetResult::invalid();
6021 Value *FalseVal = ApplyOffsetImpl(Sel->getFalseValue(), BinOpc, RHS);
6022 if (!FalseVal)
6023 return OffsetResult::invalid();
6024 return OffsetResult::select(Cond: Sel->getCondition(), TrueV: TrueVal, FalseV: FalseVal, MDFrom: Sel);
6025 }
6026 if (Value *Simplified = ApplyOffsetImpl(V, BinOpc, RHS))
6027 return OffsetResult::value(V: Simplified);
6028 return OffsetResult::invalid();
6029 };
6030
6031 for (auto [BinOp, RHS] : OffsetOps) {
6032 auto BinOpc = static_cast<unsigned>(BinOp);
6033
6034 auto Op0Result = ApplyOffset(Op0, BinOpc, RHS);
6035 if (!Op0Result.isValid())
6036 continue;
6037 auto Op1Result = ApplyOffset(Op1, BinOpc, RHS);
6038 if (!Op1Result.isValid())
6039 continue;
6040
6041 Value *NewLHS = Op0Result.materialize(Builder);
6042 Value *NewRHS = Op1Result.materialize(Builder);
6043 return new ICmpInst(I.getPredicate(), NewLHS, NewRHS);
6044 }
6045
6046 return nullptr;
6047}
6048
6049Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
6050 if (!I.isEquality())
6051 return nullptr;
6052
6053 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
6054 const CmpInst::Predicate Pred = I.getPredicate();
6055 Value *A, *B, *C, *D;
6056 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)))) {
6057 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6058 Value *OtherVal = A == Op1 ? B : A;
6059 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(Ty: A->getType()));
6060 }
6061
6062 if (match(V: Op1, P: m_Xor(L: m_Value(V&: C), R: m_Value(V&: D)))) {
6063 // A^c1 == C^c2 --> A == C^(c1^c2)
6064 ConstantInt *C1, *C2;
6065 if (match(V: B, P: m_ConstantInt(CI&: C1)) && match(V: D, P: m_ConstantInt(CI&: C2)) &&
6066 Op1->hasOneUse()) {
6067 Constant *NC = Builder.getInt(AI: C1->getValue() ^ C2->getValue());
6068 Value *Xor = Builder.CreateXor(LHS: C, RHS: NC);
6069 return new ICmpInst(Pred, A, Xor);
6070 }
6071
6072 // A^B == A^D -> B == D
6073 if (A == C)
6074 return new ICmpInst(Pred, B, D);
6075 if (A == D)
6076 return new ICmpInst(Pred, B, C);
6077 if (B == C)
6078 return new ICmpInst(Pred, A, D);
6079 if (B == D)
6080 return new ICmpInst(Pred, A, C);
6081 }
6082 }
6083
6084 if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) && (A == Op0 || B == Op0)) {
6085 // A == (A^B) -> B == 0
6086 Value *OtherVal = A == Op0 ? B : A;
6087 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(Ty: A->getType()));
6088 }
6089
6090 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6091 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
6092 match(V: Op1, P: m_And(L: m_Value(V&: C), R: m_Value(V&: D)))) {
6093 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
6094
6095 if (A == C) {
6096 X = B;
6097 Y = D;
6098 Z = A;
6099 } else if (A == D) {
6100 X = B;
6101 Y = C;
6102 Z = A;
6103 } else if (B == C) {
6104 X = A;
6105 Y = D;
6106 Z = B;
6107 } else if (B == D) {
6108 X = A;
6109 Y = C;
6110 Z = B;
6111 }
6112
6113 if (X) {
6114 // If X^Y is a negative power of two, then `icmp eq/ne (Z & NegP2), 0`
6115 // will fold to `icmp ult/uge Z, -NegP2` incurringb no additional
6116 // instructions.
6117 const APInt *C0, *C1;
6118 bool XorIsNegP2 = match(V: X, P: m_APInt(Res&: C0)) && match(V: Y, P: m_APInt(Res&: C1)) &&
6119 (*C0 ^ *C1).isNegatedPowerOf2();
6120
6121 // If either Op0/Op1 are both one use or X^Y will constant fold and one of
6122 // Op0/Op1 are one use, proceed. In those cases we are instruction neutral
6123 // but `icmp eq/ne A, 0` is easier to analyze than `icmp eq/ne A, B`.
6124 int UseCnt =
6125 int(Op0->hasOneUse()) + int(Op1->hasOneUse()) +
6126 (int(match(V: X, P: m_ImmConstant()) && match(V: Y, P: m_ImmConstant())));
6127 if (XorIsNegP2 || UseCnt >= 2) {
6128 // Build (X^Y) & Z
6129 Op1 = Builder.CreateXor(LHS: X, RHS: Y);
6130 Op1 = Builder.CreateAnd(LHS: Op1, RHS: Z);
6131 return new ICmpInst(Pred, Op1, Constant::getNullValue(Ty: Op1->getType()));
6132 }
6133 }
6134 }
6135
6136 {
6137 // Similar to above, but specialized for constant because invert is needed:
6138 // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
6139 Value *X, *Y;
6140 Constant *C;
6141 if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Constant(C)))) &&
6142 match(V: Op1, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: Y), R: m_Specific(V: C))))) {
6143 Value *Xor = Builder.CreateXor(LHS: X, RHS: Y);
6144 Value *And = Builder.CreateAnd(LHS: Xor, RHS: ConstantExpr::getNot(C));
6145 return new ICmpInst(Pred, And, Constant::getNullValue(Ty: And->getType()));
6146 }
6147 }
6148
6149 if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: A))) &&
6150 (Op0->hasOneUse() || Op1->hasOneUse())) {
6151 // (B & (Pow2C-1)) == zext A --> A == trunc B
6152 // (B & (Pow2C-1)) != zext A --> A != trunc B
6153 const APInt *MaskC;
6154 if (match(V: Op0, P: m_And(L: m_Value(V&: B), R: m_LowBitMask(V&: MaskC))) &&
6155 MaskC->countr_one() == A->getType()->getScalarSizeInBits())
6156 return new ICmpInst(Pred, A, Builder.CreateTrunc(V: B, DestTy: A->getType()));
6157 }
6158
6159 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
6160 // For lshr and ashr pairs.
6161 const APInt *AP1, *AP2;
6162 if ((match(V: Op0, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: AP1)))) &&
6163 match(V: Op1, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: B), R: m_APIntAllowPoison(Res&: AP2))))) ||
6164 (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: AP1)))) &&
6165 match(V: Op1, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: B), R: m_APIntAllowPoison(Res&: AP2)))))) {
6166 if (*AP1 != *AP2)
6167 return nullptr;
6168 unsigned TypeBits = AP1->getBitWidth();
6169 unsigned ShAmt = AP1->getLimitedValue(Limit: TypeBits);
6170 if (ShAmt < TypeBits && ShAmt != 0) {
6171 ICmpInst::Predicate NewPred =
6172 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6173 Value *Xor = Builder.CreateXor(LHS: A, RHS: B, Name: I.getName() + ".unshifted");
6174 APInt CmpVal = APInt::getOneBitSet(numBits: TypeBits, BitNo: ShAmt);
6175 return new ICmpInst(NewPred, Xor, ConstantInt::get(Ty: A->getType(), V: CmpVal));
6176 }
6177 }
6178
6179 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
6180 ConstantInt *Cst1;
6181 if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: A), R: m_ConstantInt(CI&: Cst1)))) &&
6182 match(V: Op1, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: B), R: m_Specific(V: Cst1))))) {
6183 unsigned TypeBits = Cst1->getBitWidth();
6184 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(Limit: TypeBits);
6185 if (ShAmt < TypeBits && ShAmt != 0) {
6186 Value *Xor = Builder.CreateXor(LHS: A, RHS: B, Name: I.getName() + ".unshifted");
6187 APInt AndVal = APInt::getLowBitsSet(numBits: TypeBits, loBitsSet: TypeBits - ShAmt);
6188 Value *And =
6189 Builder.CreateAnd(LHS: Xor, RHS: Builder.getInt(AI: AndVal), Name: I.getName() + ".mask");
6190 return new ICmpInst(Pred, And, Constant::getNullValue(Ty: Cst1->getType()));
6191 }
6192 }
6193
6194 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
6195 // "icmp (and X, mask), cst"
6196 uint64_t ShAmt = 0;
6197 if (Op0->hasOneUse() &&
6198 match(V: Op0, P: m_Trunc(Op: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: A), R: m_ConstantInt(V&: ShAmt))))) &&
6199 match(V: Op1, P: m_ConstantInt(CI&: Cst1)) &&
6200 // Only do this when A has multiple uses. This is most important to do
6201 // when it exposes other optimizations.
6202 !A->hasOneUse()) {
6203 unsigned ASize = cast<IntegerType>(Val: A->getType())->getPrimitiveSizeInBits();
6204
6205 if (ShAmt < ASize) {
6206 APInt MaskV =
6207 APInt::getLowBitsSet(numBits: ASize, loBitsSet: Op0->getType()->getPrimitiveSizeInBits());
6208 MaskV <<= ShAmt;
6209
6210 APInt CmpV = Cst1->getValue().zext(width: ASize);
6211 CmpV <<= ShAmt;
6212
6213 Value *Mask = Builder.CreateAnd(LHS: A, RHS: Builder.getInt(AI: MaskV));
6214 return new ICmpInst(Pred, Mask, Builder.getInt(AI: CmpV));
6215 }
6216 }
6217
6218 if (Instruction *ICmp = foldICmpIntrinsicWithIntrinsic(Cmp&: I, Builder))
6219 return ICmp;
6220
6221 // Match icmp eq (trunc (lshr A, BW), (ashr (trunc A), BW-1)), which checks
6222 // the top BW/2 + 1 bits are all the same. Create "A >=s INT_MIN && A <=s
6223 // INT_MAX", which we generate as "icmp ult (add A, 2^(BW-1)), 2^BW" to skip a
6224 // few steps of instcombine.
6225 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
6226 if (match(V: Op0, P: m_AShr(L: m_Trunc(Op: m_Value(V&: A)), R: m_SpecificInt(V: BitWidth - 1))) &&
6227 match(V: Op1, P: m_Trunc(Op: m_LShr(L: m_Specific(V: A), R: m_SpecificInt(V: BitWidth)))) &&
6228 A->getType()->getScalarSizeInBits() == BitWidth * 2 &&
6229 (I.getOperand(i_nocapture: 0)->hasOneUse() || I.getOperand(i_nocapture: 1)->hasOneUse())) {
6230 APInt C = APInt::getOneBitSet(numBits: BitWidth * 2, BitNo: BitWidth - 1);
6231 Value *Add = Builder.CreateAdd(LHS: A, RHS: ConstantInt::get(Ty: A->getType(), V: C));
6232 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT
6233 : ICmpInst::ICMP_UGE,
6234 Add, ConstantInt::get(Ty: A->getType(), V: C.shl(shiftAmt: 1)));
6235 }
6236
6237 // Canonicalize:
6238 // Assume B_Pow2 != 0
6239 // 1. A & B_Pow2 != B_Pow2 -> A & B_Pow2 == 0
6240 // 2. A & B_Pow2 == B_Pow2 -> A & B_Pow2 != 0
6241 if (match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value())) &&
6242 isKnownToBeAPowerOfTwo(V: Op1, /* OrZero */ false, CxtI: &I))
6243 return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op0,
6244 ConstantInt::getNullValue(Ty: Op0->getType()));
6245
6246 if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value())) &&
6247 isKnownToBeAPowerOfTwo(V: Op0, /* OrZero */ false, CxtI: &I))
6248 return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op1,
6249 ConstantInt::getNullValue(Ty: Op1->getType()));
6250
6251 // Canonicalize:
6252 // icmp eq/ne X, OneUse(rotate-right(X))
6253 // -> icmp eq/ne X, rotate-left(X)
6254 // We generally try to convert rotate-right -> rotate-left, this just
6255 // canonicalizes another case.
6256 if (match(V: &I, P: m_c_ICmp(L: m_Value(V&: A),
6257 R: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::fshr>(
6258 Op0: m_Deferred(V: A), Op1: m_Deferred(V: A), Op2: m_Value(V&: B))))))
6259 return new ICmpInst(
6260 Pred, A,
6261 Builder.CreateIntrinsic(RetTy: Op0->getType(), ID: Intrinsic::fshl, Args: {A, A, B}));
6262
6263 // Canonicalize:
6264 // icmp eq/ne OneUse(A ^ Cst), B --> icmp eq/ne (A ^ B), Cst
6265 Constant *Cst;
6266 if (match(V: &I, P: m_c_ICmp(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: A), R: m_ImmConstant(C&: Cst))),
6267 R: m_CombineAnd(L: m_Value(V&: B), R: m_Unless(M: m_ImmConstant())))))
6268 return new ICmpInst(Pred, Builder.CreateXor(LHS: A, RHS: B), Cst);
6269
6270 {
6271 // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
6272 auto m_Matcher =
6273 m_CombineOr(L: m_CombineOr(L: m_c_Add(L: m_Value(V&: B), R: m_Deferred(V: A)),
6274 R: m_c_Xor(L: m_Value(V&: B), R: m_Deferred(V: A))),
6275 R: m_Sub(L: m_Value(V&: B), R: m_Deferred(V: A)));
6276 std::optional<bool> IsZero = std::nullopt;
6277 if (match(V: &I, P: m_c_ICmp(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_Matcher)),
6278 R: m_Deferred(V: A))))
6279 IsZero = false;
6280 // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
6281 else if (match(V: &I,
6282 P: m_ICmp(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_Matcher)), R: m_Zero())))
6283 IsZero = true;
6284
6285 if (IsZero && isKnownToBeAPowerOfTwo(V: A, /* OrZero */ true, CxtI: &I))
6286 // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
6287 // -> (icmp eq/ne (and X, P2), 0)
6288 // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
6289 // -> (icmp eq/ne (and X, P2), P2)
6290 return new ICmpInst(Pred, Builder.CreateAnd(LHS: B, RHS: A),
6291 *IsZero ? A
6292 : ConstantInt::getNullValue(Ty: A->getType()));
6293 }
6294
6295 if (auto *Res = foldICmpEqualityWithOffset(
6296 I, Builder, SQ: getSimplifyQuery().getWithInstruction(I: &I)))
6297 return Res;
6298
6299 return nullptr;
6300}
6301
6302Instruction *InstCombinerImpl::foldICmpWithTrunc(ICmpInst &ICmp) {
6303 ICmpInst::Predicate Pred = ICmp.getPredicate();
6304 Value *Op0 = ICmp.getOperand(i_nocapture: 0), *Op1 = ICmp.getOperand(i_nocapture: 1);
6305
6306 // Try to canonicalize trunc + compare-to-constant into a mask + cmp.
6307 // The trunc masks high bits while the compare may effectively mask low bits.
6308 Value *X;
6309 const APInt *C;
6310 if (!match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: X)))) || !match(V: Op1, P: m_APInt(Res&: C)))
6311 return nullptr;
6312
6313 // This matches patterns corresponding to tests of the signbit as well as:
6314 // (trunc X) pred C2 --> (X & Mask) == C
6315 if (auto Res = decomposeBitTestICmp(LHS: Op0, RHS: Op1, Pred, /*LookThroughTrunc=*/true,
6316 /*AllowNonZeroC=*/true)) {
6317 Value *And = Builder.CreateAnd(LHS: Res->X, RHS: Res->Mask);
6318 Constant *C = ConstantInt::get(Ty: Res->X->getType(), V: Res->C);
6319 return new ICmpInst(Res->Pred, And, C);
6320 }
6321
6322 unsigned SrcBits = X->getType()->getScalarSizeInBits();
6323 if (auto *II = dyn_cast<IntrinsicInst>(Val: X)) {
6324 if (II->getIntrinsicID() == Intrinsic::cttz ||
6325 II->getIntrinsicID() == Intrinsic::ctlz) {
6326 unsigned MaxRet = SrcBits;
6327 // If the "is_zero_poison" argument is set, then we know at least
6328 // one bit is set in the input, so the result is always at least one
6329 // less than the full bitwidth of that input.
6330 if (match(V: II->getArgOperand(i: 1), P: m_One()))
6331 MaxRet--;
6332
6333 // Make sure the destination is wide enough to hold the largest output of
6334 // the intrinsic.
6335 if (llvm::Log2_32(Value: MaxRet) + 1 <= Op0->getType()->getScalarSizeInBits())
6336 if (Instruction *I =
6337 foldICmpIntrinsicWithConstant(Cmp&: ICmp, II, C: C->zext(width: SrcBits)))
6338 return I;
6339 }
6340 }
6341
6342 return nullptr;
6343}
6344
6345Instruction *InstCombinerImpl::foldICmpWithZextOrSext(ICmpInst &ICmp) {
6346 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
6347 auto *CastOp0 = cast<CastInst>(Val: ICmp.getOperand(i_nocapture: 0));
6348 Value *X;
6349 if (!match(V: CastOp0, P: m_ZExtOrSExt(Op: m_Value(V&: X))))
6350 return nullptr;
6351
6352 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
6353 bool IsSignedCmp = ICmp.isSigned();
6354
6355 // icmp Pred (ext X), (ext Y)
6356 Value *Y;
6357 if (match(V: ICmp.getOperand(i_nocapture: 1), P: m_ZExtOrSExt(Op: m_Value(V&: Y)))) {
6358 bool IsZext0 = isa<ZExtInst>(Val: ICmp.getOperand(i_nocapture: 0));
6359 bool IsZext1 = isa<ZExtInst>(Val: ICmp.getOperand(i_nocapture: 1));
6360
6361 if (IsZext0 != IsZext1) {
6362 // If X and Y and both i1
6363 // (icmp eq/ne (zext X) (sext Y))
6364 // eq -> (icmp eq (or X, Y), 0)
6365 // ne -> (icmp ne (or X, Y), 0)
6366 if (ICmp.isEquality() && X->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
6367 Y->getType()->isIntOrIntVectorTy(BitWidth: 1))
6368 return new ICmpInst(ICmp.getPredicate(), Builder.CreateOr(LHS: X, RHS: Y),
6369 Constant::getNullValue(Ty: X->getType()));
6370
6371 // If we have mismatched casts and zext has the nneg flag, we can
6372 // treat the "zext nneg" as "sext". Otherwise, we cannot fold and quit.
6373
6374 auto *NonNegInst0 = dyn_cast<PossiblyNonNegInst>(Val: ICmp.getOperand(i_nocapture: 0));
6375 auto *NonNegInst1 = dyn_cast<PossiblyNonNegInst>(Val: ICmp.getOperand(i_nocapture: 1));
6376
6377 bool IsNonNeg0 = NonNegInst0 && NonNegInst0->hasNonNeg();
6378 bool IsNonNeg1 = NonNegInst1 && NonNegInst1->hasNonNeg();
6379
6380 if ((IsZext0 && IsNonNeg0) || (IsZext1 && IsNonNeg1))
6381 IsSignedExt = true;
6382 else
6383 return nullptr;
6384 }
6385
6386 // Not an extension from the same type?
6387 Type *XTy = X->getType(), *YTy = Y->getType();
6388 if (XTy != YTy) {
6389 // One of the casts must have one use because we are creating a new cast.
6390 if (!ICmp.getOperand(i_nocapture: 0)->hasOneUse() && !ICmp.getOperand(i_nocapture: 1)->hasOneUse())
6391 return nullptr;
6392 // Extend the narrower operand to the type of the wider operand.
6393 CastInst::CastOps CastOpcode =
6394 IsSignedExt ? Instruction::SExt : Instruction::ZExt;
6395 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
6396 X = Builder.CreateCast(Op: CastOpcode, V: X, DestTy: YTy);
6397 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
6398 Y = Builder.CreateCast(Op: CastOpcode, V: Y, DestTy: XTy);
6399 else
6400 return nullptr;
6401 }
6402
6403 // (zext X) == (zext Y) --> X == Y
6404 // (sext X) == (sext Y) --> X == Y
6405 if (ICmp.isEquality())
6406 return new ICmpInst(ICmp.getPredicate(), X, Y);
6407
6408 // A signed comparison of sign extended values simplifies into a
6409 // signed comparison.
6410 if (IsSignedCmp && IsSignedExt)
6411 return new ICmpInst(ICmp.getPredicate(), X, Y);
6412
6413 // The other three cases all fold into an unsigned comparison.
6414 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
6415 }
6416
6417 // Below here, we are only folding a compare with constant.
6418 auto *C = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1));
6419 if (!C)
6420 return nullptr;
6421
6422 // If a lossless truncate is possible...
6423 Type *SrcTy = CastOp0->getSrcTy();
6424 Constant *Res = getLosslessInvCast(C, InvCastTo: SrcTy, CastOp: CastOp0->getOpcode(), DL);
6425 if (Res) {
6426 if (ICmp.isEquality())
6427 return new ICmpInst(ICmp.getPredicate(), X, Res);
6428
6429 // A signed comparison of sign extended values simplifies into a
6430 // signed comparison.
6431 if (IsSignedExt && IsSignedCmp)
6432 return new ICmpInst(ICmp.getPredicate(), X, Res);
6433
6434 // The other three cases all fold into an unsigned comparison.
6435 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res);
6436 }
6437
6438 // The re-extended constant changed, partly changed (in the case of a vector),
6439 // or could not be determined to be equal (in the case of a constant
6440 // expression), so the constant cannot be represented in the shorter type.
6441 // All the cases that fold to true or false will have already been handled
6442 // by simplifyICmpInst, so only deal with the tricky case.
6443 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(Val: C))
6444 return nullptr;
6445
6446 // Is source op positive?
6447 // icmp ult (sext X), C --> icmp sgt X, -1
6448 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
6449 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(Ty: SrcTy));
6450
6451 // Is source op negative?
6452 // icmp ugt (sext X), C --> icmp slt X, 0
6453 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
6454 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(Ty: SrcTy));
6455}
6456
6457/// Handle icmp (cast x), (cast or constant).
6458Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
6459 // If any operand of ICmp is a inttoptr roundtrip cast then remove it as
6460 // icmp compares only pointer's value.
6461 // icmp (inttoptr (ptrtoint p1)), p2 --> icmp p1, p2.
6462 Value *SimplifiedOp0 = simplifyIntToPtrRoundTripCast(Val: ICmp.getOperand(i_nocapture: 0));
6463 Value *SimplifiedOp1 = simplifyIntToPtrRoundTripCast(Val: ICmp.getOperand(i_nocapture: 1));
6464 if (SimplifiedOp0 || SimplifiedOp1)
6465 return new ICmpInst(ICmp.getPredicate(),
6466 SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(i_nocapture: 0),
6467 SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(i_nocapture: 1));
6468
6469 auto *CastOp0 = dyn_cast<CastInst>(Val: ICmp.getOperand(i_nocapture: 0));
6470 if (!CastOp0)
6471 return nullptr;
6472 if (!isa<Constant>(Val: ICmp.getOperand(i_nocapture: 1)) && !isa<CastInst>(Val: ICmp.getOperand(i_nocapture: 1)))
6473 return nullptr;
6474
6475 Value *Op0Src = CastOp0->getOperand(i_nocapture: 0);
6476 Type *SrcTy = CastOp0->getSrcTy();
6477 Type *DestTy = CastOp0->getDestTy();
6478
6479 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6480 // integer type is the same size as the pointer type.
6481 auto CompatibleSizes = [&](Type *PtrTy, Type *IntTy) {
6482 if (isa<VectorType>(Val: PtrTy)) {
6483 PtrTy = cast<VectorType>(Val: PtrTy)->getElementType();
6484 IntTy = cast<VectorType>(Val: IntTy)->getElementType();
6485 }
6486 return DL.getPointerTypeSizeInBits(PtrTy) == IntTy->getIntegerBitWidth();
6487 };
6488 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
6489 CompatibleSizes(SrcTy, DestTy)) {
6490 Value *NewOp1 = nullptr;
6491 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(Val: ICmp.getOperand(i_nocapture: 1))) {
6492 Value *PtrSrc = PtrToIntOp1->getOperand(i_nocapture: 0);
6493 if (PtrSrc->getType() == Op0Src->getType())
6494 NewOp1 = PtrToIntOp1->getOperand(i_nocapture: 0);
6495 } else if (auto *RHSC = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1))) {
6496 NewOp1 = ConstantExpr::getIntToPtr(C: RHSC, Ty: SrcTy);
6497 }
6498
6499 if (NewOp1)
6500 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
6501 }
6502
6503 // Do the same in the other direction for icmp (inttoptr x), (inttoptr/c).
6504 if (CastOp0->getOpcode() == Instruction::IntToPtr &&
6505 CompatibleSizes(DestTy, SrcTy)) {
6506 Value *NewOp1 = nullptr;
6507 if (auto *IntToPtrOp1 = dyn_cast<IntToPtrInst>(Val: ICmp.getOperand(i_nocapture: 1))) {
6508 Value *IntSrc = IntToPtrOp1->getOperand(i_nocapture: 0);
6509 if (IntSrc->getType() == Op0Src->getType())
6510 NewOp1 = IntToPtrOp1->getOperand(i_nocapture: 0);
6511 } else if (auto *RHSC = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1))) {
6512 NewOp1 = ConstantFoldConstant(C: ConstantExpr::getPtrToInt(C: RHSC, Ty: SrcTy), DL);
6513 }
6514
6515 if (NewOp1)
6516 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
6517 }
6518
6519 if (Instruction *R = foldICmpWithTrunc(ICmp))
6520 return R;
6521
6522 return foldICmpWithZextOrSext(ICmp);
6523}
6524
6525static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS,
6526 bool IsSigned) {
6527 switch (BinaryOp) {
6528 default:
6529 llvm_unreachable("Unsupported binary op");
6530 case Instruction::Add:
6531 case Instruction::Sub:
6532 return match(V: RHS, P: m_Zero());
6533 case Instruction::Mul:
6534 return !(RHS->getType()->isIntOrIntVectorTy(BitWidth: 1) && IsSigned) &&
6535 match(V: RHS, P: m_One());
6536 }
6537}
6538
6539OverflowResult
6540InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
6541 bool IsSigned, Value *LHS, Value *RHS,
6542 Instruction *CxtI) const {
6543 switch (BinaryOp) {
6544 default:
6545 llvm_unreachable("Unsupported binary op");
6546 case Instruction::Add:
6547 if (IsSigned)
6548 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
6549 else
6550 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
6551 case Instruction::Sub:
6552 if (IsSigned)
6553 return computeOverflowForSignedSub(LHS, RHS, CxtI);
6554 else
6555 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
6556 case Instruction::Mul:
6557 if (IsSigned)
6558 return computeOverflowForSignedMul(LHS, RHS, CxtI);
6559 else
6560 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
6561 }
6562}
6563
6564bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
6565 bool IsSigned, Value *LHS,
6566 Value *RHS, Instruction &OrigI,
6567 Value *&Result,
6568 Constant *&Overflow) {
6569 if (OrigI.isCommutative() && isa<Constant>(Val: LHS) && !isa<Constant>(Val: RHS))
6570 std::swap(a&: LHS, b&: RHS);
6571
6572 // If the overflow check was an add followed by a compare, the insertion point
6573 // may be pointing to the compare. We want to insert the new instructions
6574 // before the add in case there are uses of the add between the add and the
6575 // compare.
6576 Builder.SetInsertPoint(&OrigI);
6577
6578 Type *OverflowTy = Type::getInt1Ty(C&: LHS->getContext());
6579 if (auto *LHSTy = dyn_cast<VectorType>(Val: LHS->getType()))
6580 OverflowTy = VectorType::get(ElementType: OverflowTy, EC: LHSTy->getElementCount());
6581
6582 if (isNeutralValue(BinaryOp, RHS, IsSigned)) {
6583 Result = LHS;
6584 Overflow = ConstantInt::getFalse(Ty: OverflowTy);
6585 return true;
6586 }
6587
6588 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, CxtI: &OrigI)) {
6589 case OverflowResult::MayOverflow:
6590 return false;
6591 case OverflowResult::AlwaysOverflowsLow:
6592 case OverflowResult::AlwaysOverflowsHigh:
6593 Result = Builder.CreateBinOp(Opc: BinaryOp, LHS, RHS);
6594 Result->takeName(V: &OrigI);
6595 Overflow = ConstantInt::getTrue(Ty: OverflowTy);
6596 return true;
6597 case OverflowResult::NeverOverflows:
6598 Result = Builder.CreateBinOp(Opc: BinaryOp, LHS, RHS);
6599 Result->takeName(V: &OrigI);
6600 Overflow = ConstantInt::getFalse(Ty: OverflowTy);
6601 if (auto *Inst = dyn_cast<Instruction>(Val: Result)) {
6602 if (IsSigned)
6603 Inst->setHasNoSignedWrap();
6604 else
6605 Inst->setHasNoUnsignedWrap();
6606 }
6607 return true;
6608 }
6609
6610 llvm_unreachable("Unexpected overflow result");
6611}
6612
6613/// Recognize and process idiom involving test for multiplication
6614/// overflow.
6615///
6616/// The caller has matched a pattern of the form:
6617/// I = cmp u (mul(zext A, zext B), V
6618/// The function checks if this is a test for overflow and if so replaces
6619/// multiplication with call to 'mul.with.overflow' intrinsic.
6620///
6621/// \param I Compare instruction.
6622/// \param MulVal Result of 'mult' instruction. It is one of the arguments of
6623/// the compare instruction. Must be of integer type.
6624/// \param OtherVal The other argument of compare instruction.
6625/// \returns Instruction which must replace the compare instruction, NULL if no
6626/// replacement required.
6627static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
6628 const APInt *OtherVal,
6629 InstCombinerImpl &IC) {
6630 // Don't bother doing this transformation for pointers, don't do it for
6631 // vectors.
6632 if (!isa<IntegerType>(Val: MulVal->getType()))
6633 return nullptr;
6634
6635 auto *MulInstr = dyn_cast<Instruction>(Val: MulVal);
6636 if (!MulInstr)
6637 return nullptr;
6638 assert(MulInstr->getOpcode() == Instruction::Mul);
6639
6640 auto *LHS = cast<ZExtInst>(Val: MulInstr->getOperand(i: 0)),
6641 *RHS = cast<ZExtInst>(Val: MulInstr->getOperand(i: 1));
6642 assert(LHS->getOpcode() == Instruction::ZExt);
6643 assert(RHS->getOpcode() == Instruction::ZExt);
6644 Value *A = LHS->getOperand(i_nocapture: 0), *B = RHS->getOperand(i_nocapture: 0);
6645
6646 // Calculate type and width of the result produced by mul.with.overflow.
6647 Type *TyA = A->getType(), *TyB = B->getType();
6648 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
6649 WidthB = TyB->getPrimitiveSizeInBits();
6650 unsigned MulWidth;
6651 Type *MulType;
6652 if (WidthB > WidthA) {
6653 MulWidth = WidthB;
6654 MulType = TyB;
6655 } else {
6656 MulWidth = WidthA;
6657 MulType = TyA;
6658 }
6659
6660 // In order to replace the original mul with a narrower mul.with.overflow,
6661 // all uses must ignore upper bits of the product. The number of used low
6662 // bits must be not greater than the width of mul.with.overflow.
6663 if (MulVal->hasNUsesOrMore(N: 2))
6664 for (User *U : MulVal->users()) {
6665 if (U == &I)
6666 continue;
6667 if (TruncInst *TI = dyn_cast<TruncInst>(Val: U)) {
6668 // Check if truncation ignores bits above MulWidth.
6669 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
6670 if (TruncWidth > MulWidth)
6671 return nullptr;
6672 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: U)) {
6673 // Check if AND ignores bits above MulWidth.
6674 if (BO->getOpcode() != Instruction::And)
6675 return nullptr;
6676 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) {
6677 const APInt &CVal = CI->getValue();
6678 if (CVal.getBitWidth() - CVal.countl_zero() > MulWidth)
6679 return nullptr;
6680 } else {
6681 // In this case we could have the operand of the binary operation
6682 // being defined in another block, and performing the replacement
6683 // could break the dominance relation.
6684 return nullptr;
6685 }
6686 } else {
6687 // Other uses prohibit this transformation.
6688 return nullptr;
6689 }
6690 }
6691
6692 // Recognize patterns
6693 switch (I.getPredicate()) {
6694 case ICmpInst::ICMP_UGT: {
6695 // Recognize pattern:
6696 // mulval = mul(zext A, zext B)
6697 // cmp ugt mulval, max
6698 APInt MaxVal = APInt::getMaxValue(numBits: MulWidth);
6699 MaxVal = MaxVal.zext(width: OtherVal->getBitWidth());
6700 if (MaxVal.eq(RHS: *OtherVal))
6701 break; // Recognized
6702 return nullptr;
6703 }
6704
6705 case ICmpInst::ICMP_ULT: {
6706 // Recognize pattern:
6707 // mulval = mul(zext A, zext B)
6708 // cmp ule mulval, max + 1
6709 APInt MaxVal = APInt::getOneBitSet(numBits: OtherVal->getBitWidth(), BitNo: MulWidth);
6710 if (MaxVal.eq(RHS: *OtherVal))
6711 break; // Recognized
6712 return nullptr;
6713 }
6714
6715 default:
6716 return nullptr;
6717 }
6718
6719 InstCombiner::BuilderTy &Builder = IC.Builder;
6720 Builder.SetInsertPoint(MulInstr);
6721
6722 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
6723 Value *MulA = A, *MulB = B;
6724 if (WidthA < MulWidth)
6725 MulA = Builder.CreateZExt(V: A, DestTy: MulType);
6726 if (WidthB < MulWidth)
6727 MulB = Builder.CreateZExt(V: B, DestTy: MulType);
6728 CallInst *Call =
6729 Builder.CreateIntrinsic(ID: Intrinsic::umul_with_overflow, Types: MulType,
6730 Args: {MulA, MulB}, /*FMFSource=*/nullptr, Name: "umul");
6731 IC.addToWorklist(I: MulInstr);
6732
6733 // If there are uses of mul result other than the comparison, we know that
6734 // they are truncation or binary AND. Change them to use result of
6735 // mul.with.overflow and adjust properly mask/size.
6736 if (MulVal->hasNUsesOrMore(N: 2)) {
6737 Value *Mul = Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "umul.value");
6738 for (User *U : make_early_inc_range(Range: MulVal->users())) {
6739 if (U == &I)
6740 continue;
6741 if (TruncInst *TI = dyn_cast<TruncInst>(Val: U)) {
6742 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
6743 IC.replaceInstUsesWith(I&: *TI, V: Mul);
6744 else
6745 TI->setOperand(i_nocapture: 0, Val_nocapture: Mul);
6746 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: U)) {
6747 assert(BO->getOpcode() == Instruction::And);
6748 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
6749 ConstantInt *CI = cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1));
6750 APInt ShortMask = CI->getValue().trunc(width: MulWidth);
6751 Value *ShortAnd = Builder.CreateAnd(LHS: Mul, RHS: ShortMask);
6752 Value *Zext = Builder.CreateZExt(V: ShortAnd, DestTy: BO->getType());
6753 IC.replaceInstUsesWith(I&: *BO, V: Zext);
6754 } else {
6755 llvm_unreachable("Unexpected Binary operation");
6756 }
6757 IC.addToWorklist(I: cast<Instruction>(Val: U));
6758 }
6759 }
6760
6761 // The original icmp gets replaced with the overflow value, maybe inverted
6762 // depending on predicate.
6763 if (I.getPredicate() == ICmpInst::ICMP_ULT) {
6764 Value *Res = Builder.CreateExtractValue(Agg: Call, Idxs: 1);
6765 return BinaryOperator::CreateNot(Op: Res);
6766 }
6767
6768 return ExtractValueInst::Create(Agg: Call, Idxs: 1);
6769}
6770
6771/// When performing a comparison against a constant, it is possible that not all
6772/// the bits in the LHS are demanded. This helper method computes the mask that
6773/// IS demanded.
6774static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
6775 const APInt *RHS;
6776 if (!match(V: I.getOperand(i_nocapture: 1), P: m_APInt(Res&: RHS)))
6777 return APInt::getAllOnes(numBits: BitWidth);
6778
6779 // If this is a normal comparison, it demands all bits. If it is a sign bit
6780 // comparison, it only demands the sign bit.
6781 bool UnusedBit;
6782 if (isSignBitCheck(Pred: I.getPredicate(), RHS: *RHS, TrueIfSigned&: UnusedBit))
6783 return APInt::getSignMask(BitWidth);
6784
6785 switch (I.getPredicate()) {
6786 // For a UGT comparison, we don't care about any bits that
6787 // correspond to the trailing ones of the comparand. The value of these
6788 // bits doesn't impact the outcome of the comparison, because any value
6789 // greater than the RHS must differ in a bit higher than these due to carry.
6790 case ICmpInst::ICMP_UGT:
6791 return APInt::getBitsSetFrom(numBits: BitWidth, loBit: RHS->countr_one());
6792
6793 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
6794 // Any value less than the RHS must differ in a higher bit because of carries.
6795 case ICmpInst::ICMP_ULT:
6796 return APInt::getBitsSetFrom(numBits: BitWidth, loBit: RHS->countr_zero());
6797
6798 default:
6799 return APInt::getAllOnes(numBits: BitWidth);
6800 }
6801}
6802
6803/// Check that one use is in the same block as the definition and all
6804/// other uses are in blocks dominated by a given block.
6805///
6806/// \param DI Definition
6807/// \param UI Use
6808/// \param DB Block that must dominate all uses of \p DI outside
6809/// the parent block
6810/// \return true when \p UI is the only use of \p DI in the parent block
6811/// and all other uses of \p DI are in blocks dominated by \p DB.
6812///
6813bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
6814 const Instruction *UI,
6815 const BasicBlock *DB) const {
6816 assert(DI && UI && "Instruction not defined\n");
6817 // Ignore incomplete definitions.
6818 if (!DI->getParent())
6819 return false;
6820 // DI and UI must be in the same block.
6821 if (DI->getParent() != UI->getParent())
6822 return false;
6823 // Protect from self-referencing blocks.
6824 if (DI->getParent() == DB)
6825 return false;
6826 for (const User *U : DI->users()) {
6827 auto *Usr = cast<Instruction>(Val: U);
6828 if (Usr != UI && !DT.dominates(A: DB, B: Usr->getParent()))
6829 return false;
6830 }
6831 return true;
6832}
6833
6834/// Return true when the instruction sequence within a block is select-cmp-br.
6835static bool isChainSelectCmpBranch(const SelectInst *SI) {
6836 const BasicBlock *BB = SI->getParent();
6837 if (!BB)
6838 return false;
6839 auto *BI = dyn_cast_or_null<CondBrInst>(Val: BB->getTerminator());
6840 if (!BI)
6841 return false;
6842 auto *IC = dyn_cast<ICmpInst>(Val: BI->getCondition());
6843 if (!IC || (IC->getOperand(i_nocapture: 0) != SI && IC->getOperand(i_nocapture: 1) != SI))
6844 return false;
6845 return true;
6846}
6847
6848/// True when a select result is replaced by one of its operands
6849/// in select-icmp sequence. This will eventually result in the elimination
6850/// of the select.
6851///
6852/// \param SI Select instruction
6853/// \param Icmp Compare instruction
6854/// \param SIOpd Operand that replaces the select
6855///
6856/// Notes:
6857/// - The replacement is global and requires dominator information
6858/// - The caller is responsible for the actual replacement
6859///
6860/// Example:
6861///
6862/// entry:
6863/// %4 = select i1 %3, %C* %0, %C* null
6864/// %5 = icmp eq %C* %4, null
6865/// br i1 %5, label %9, label %7
6866/// ...
6867/// ; <label>:7 ; preds = %entry
6868/// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
6869/// ...
6870///
6871/// can be transformed to
6872///
6873/// %5 = icmp eq %C* %0, null
6874/// %6 = select i1 %3, i1 %5, i1 true
6875/// br i1 %6, label %9, label %7
6876/// ...
6877/// ; <label>:7 ; preds = %entry
6878/// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
6879///
6880/// Similar when the first operand of the select is a constant or/and
6881/// the compare is for not equal rather than equal.
6882///
6883/// NOTE: The function is only called when the select and compare constants
6884/// are equal, the optimization can work only for EQ predicates. This is not a
6885/// major restriction since a NE compare should be 'normalized' to an equal
6886/// compare, which usually happens in the combiner and test case
6887/// select-cmp-br.ll checks for it.
6888bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
6889 const ICmpInst *Icmp,
6890 const unsigned SIOpd) {
6891 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
6892 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
6893 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(Idx: 1);
6894 // The check for the single predecessor is not the best that can be
6895 // done. But it protects efficiently against cases like when SI's
6896 // home block has two successors, Succ and Succ1, and Succ1 predecessor
6897 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
6898 // replaced can be reached on either path. So the uniqueness check
6899 // guarantees that the path all uses of SI (outside SI's parent) are on
6900 // is disjoint from all other paths out of SI. But that information
6901 // is more expensive to compute, and the trade-off here is in favor
6902 // of compile-time. It should also be noticed that we check for a single
6903 // predecessor and not only uniqueness. This to handle the situation when
6904 // Succ and Succ1 points to the same basic block.
6905 if (Succ->getSinglePredecessor() && dominatesAllUses(DI: SI, UI: Icmp, DB: Succ)) {
6906 NumSel++;
6907 SI->replaceUsesOutsideBlock(V: SI->getOperand(i_nocapture: SIOpd), BB: SI->getParent());
6908 return true;
6909 }
6910 }
6911 return false;
6912}
6913
6914/// Try to fold the comparison based on range information we can get by checking
6915/// whether bits are known to be zero or one in the inputs.
6916Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
6917 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
6918 Type *Ty = Op0->getType();
6919 ICmpInst::Predicate Pred = I.getPredicate();
6920
6921 // Get scalar or pointer size.
6922 unsigned BitWidth = Ty->isIntOrIntVectorTy()
6923 ? Ty->getScalarSizeInBits()
6924 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
6925
6926 if (!BitWidth)
6927 return nullptr;
6928
6929 KnownBits Op0Known(BitWidth);
6930 KnownBits Op1Known(BitWidth);
6931
6932 {
6933 // Don't use dominating conditions when folding icmp using known bits. This
6934 // may convert signed into unsigned predicates in ways that other passes
6935 // (especially IndVarSimplify) may not be able to reliably undo.
6936 SimplifyQuery Q = SQ.getWithoutDomCondCache().getWithInstruction(I: &I);
6937 if (SimplifyDemandedBits(I: &I, Op: 0, DemandedMask: getDemandedBitsLHSMask(I, BitWidth),
6938 Known&: Op0Known, Q))
6939 return &I;
6940
6941 if (SimplifyDemandedBits(I: &I, Op: 1, DemandedMask: APInt::getAllOnes(numBits: BitWidth), Known&: Op1Known, Q))
6942 return &I;
6943 }
6944
6945 if (!isa<Constant>(Val: Op0) && Op0Known.isConstant())
6946 return new ICmpInst(
6947 Pred, ConstantExpr::getIntegerValue(Ty, V: Op0Known.getConstant()), Op1);
6948 if (!isa<Constant>(Val: Op1) && Op1Known.isConstant())
6949 return new ICmpInst(
6950 Pred, Op0, ConstantExpr::getIntegerValue(Ty, V: Op1Known.getConstant()));
6951
6952 if (std::optional<bool> Res = ICmpInst::compare(LHS: Op0Known, RHS: Op1Known, Pred))
6953 return replaceInstUsesWith(I, V: ConstantInt::getBool(Ty: I.getType(), V: *Res));
6954
6955 // Given the known and unknown bits, compute a range that the LHS could be
6956 // in. Compute the Min, Max and RHS values based on the known bits. For the
6957 // EQ and NE we use unsigned values.
6958 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6959 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6960 if (I.isSigned()) {
6961 Op0Min = Op0Known.getSignedMinValue();
6962 Op0Max = Op0Known.getSignedMaxValue();
6963 Op1Min = Op1Known.getSignedMinValue();
6964 Op1Max = Op1Known.getSignedMaxValue();
6965 } else {
6966 Op0Min = Op0Known.getMinValue();
6967 Op0Max = Op0Known.getMaxValue();
6968 Op1Min = Op1Known.getMinValue();
6969 Op1Max = Op1Known.getMaxValue();
6970 }
6971
6972 // Don't break up a clamp pattern -- (min(max X, Y), Z) -- by replacing a
6973 // min/max canonical compare with some other compare. That could lead to
6974 // conflict with select canonicalization and infinite looping.
6975 // FIXME: This constraint may go away if min/max intrinsics are canonical.
6976 auto isMinMaxCmp = [&](Instruction &Cmp) {
6977 if (!Cmp.hasOneUse())
6978 return false;
6979 Value *A, *B;
6980 SelectPatternFlavor SPF = matchSelectPattern(V: Cmp.user_back(), LHS&: A, RHS&: B).Flavor;
6981 if (!SelectPatternResult::isMinOrMax(SPF))
6982 return false;
6983 return match(V: Op0, P: m_MaxOrMin(L: m_Value(), R: m_Value())) ||
6984 match(V: Op1, P: m_MaxOrMin(L: m_Value(), R: m_Value()));
6985 };
6986 if (!isMinMaxCmp(I)) {
6987 switch (Pred) {
6988 default:
6989 break;
6990 case ICmpInst::ICMP_ULT: {
6991 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6992 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6993 const APInt *CmpC;
6994 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
6995 // A <u C -> A == C-1 if min(A)+1 == C
6996 if (*CmpC == Op0Min + 1)
6997 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6998 ConstantInt::get(Ty: Op1->getType(), V: *CmpC - 1));
6999 // X <u C --> X == 0, if the number of zero bits in the bottom of X
7000 // exceeds the log2 of C.
7001 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
7002 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7003 Constant::getNullValue(Ty: Op1->getType()));
7004 }
7005 break;
7006 }
7007 case ICmpInst::ICMP_UGT: {
7008 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
7009 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7010 const APInt *CmpC;
7011 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7012 // A >u C -> A == C+1 if max(a)-1 == C
7013 if (*CmpC == Op0Max - 1)
7014 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7015 ConstantInt::get(Ty: Op1->getType(), V: *CmpC + 1));
7016 // X >u C --> X != 0, if the number of zero bits in the bottom of X
7017 // exceeds the log2 of C.
7018 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
7019 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
7020 Constant::getNullValue(Ty: Op1->getType()));
7021 }
7022 break;
7023 }
7024 case ICmpInst::ICMP_SLT: {
7025 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
7026 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7027 const APInt *CmpC;
7028 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7029 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
7030 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7031 ConstantInt::get(Ty: Op1->getType(), V: *CmpC - 1));
7032 }
7033 break;
7034 }
7035 case ICmpInst::ICMP_SGT: {
7036 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
7037 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7038 const APInt *CmpC;
7039 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7040 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
7041 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7042 ConstantInt::get(Ty: Op1->getType(), V: *CmpC + 1));
7043 }
7044 break;
7045 }
7046 }
7047 }
7048
7049 // Based on the range information we know about the LHS, see if we can
7050 // simplify this comparison. For example, (x&4) < 8 is always true.
7051 switch (Pred) {
7052 default:
7053 break;
7054 case ICmpInst::ICMP_EQ:
7055 case ICmpInst::ICMP_NE: {
7056 // If all bits are known zero except for one, then we know at most one bit
7057 // is set. If the comparison is against zero, then this is a check to see if
7058 // *that* bit is set.
7059 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
7060 if (Op1Known.isZero()) {
7061 // If the LHS is an AND with the same constant, look through it.
7062 Value *LHS = nullptr;
7063 const APInt *LHSC;
7064 if (!match(V: Op0, P: m_And(L: m_Value(V&: LHS), R: m_APInt(Res&: LHSC))) ||
7065 *LHSC != Op0KnownZeroInverted)
7066 LHS = Op0;
7067
7068 Value *X;
7069 const APInt *C1;
7070 if (match(V: LHS, P: m_Shl(L: m_Power2(V&: C1), R: m_Value(V&: X)))) {
7071 Type *XTy = X->getType();
7072 unsigned Log2C1 = C1->countr_zero();
7073 APInt C2 = Op0KnownZeroInverted;
7074 APInt C2Pow2 = (C2 & ~(*C1 - 1)) + *C1;
7075 if (C2Pow2.isPowerOf2()) {
7076 // iff (C1 is pow2) & ((C2 & ~(C1-1)) + C1) is pow2):
7077 // ((C1 << X) & C2) == 0 -> X >= (Log2(C2+C1) - Log2(C1))
7078 // ((C1 << X) & C2) != 0 -> X < (Log2(C2+C1) - Log2(C1))
7079 unsigned Log2C2 = C2Pow2.countr_zero();
7080 auto *CmpC = ConstantInt::get(Ty: XTy, V: Log2C2 - Log2C1);
7081 auto NewPred =
7082 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
7083 return new ICmpInst(NewPred, X, CmpC);
7084 }
7085 }
7086 }
7087
7088 // Op0 eq C_Pow2 -> Op0 ne 0 if Op0 is known to be C_Pow2 or zero.
7089 if (Op1Known.isConstant() && Op1Known.getConstant().isPowerOf2() &&
7090 (Op0Known & Op1Known) == Op0Known)
7091 return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op0,
7092 ConstantInt::getNullValue(Ty: Op1->getType()));
7093 break;
7094 }
7095 case ICmpInst::ICMP_SGE:
7096 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
7097 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7098 break;
7099 case ICmpInst::ICMP_SLE:
7100 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
7101 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7102 break;
7103 case ICmpInst::ICMP_UGE:
7104 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
7105 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7106 break;
7107 case ICmpInst::ICMP_ULE:
7108 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
7109 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7110 break;
7111 }
7112
7113 // Turn a signed comparison into an unsigned one if both operands are known to
7114 // have the same sign. Set samesign if possible (except for equality
7115 // predicates).
7116 if ((I.isSigned() || (I.isUnsigned() && !I.hasSameSign())) &&
7117 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
7118 (Op0Known.One.isNegative() && Op1Known.One.isNegative()))) {
7119 I.setPredicate(I.getUnsignedPredicate());
7120 I.setSameSign();
7121 return &I;
7122 }
7123
7124 return nullptr;
7125}
7126
7127/// If one operand of an icmp is effectively a bool (value range of {0,1}),
7128/// then try to reduce patterns based on that limit.
7129Instruction *InstCombinerImpl::foldICmpUsingBoolRange(ICmpInst &I) {
7130 Value *X, *Y;
7131 CmpPredicate Pred;
7132
7133 // X must be 0 and bool must be true for "ULT":
7134 // X <u (zext i1 Y) --> (X == 0) & Y
7135 if (match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: X), R: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: Y))))) &&
7136 Y->getType()->isIntOrIntVectorTy(BitWidth: 1) && Pred == ICmpInst::ICMP_ULT)
7137 return BinaryOperator::CreateAnd(V1: Builder.CreateIsNull(Arg: X), V2: Y);
7138
7139 // X must be 0 or bool must be true for "ULE":
7140 // X <=u (sext i1 Y) --> (X == 0) | Y
7141 if (match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: X), R: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: Y))))) &&
7142 Y->getType()->isIntOrIntVectorTy(BitWidth: 1) && Pred == ICmpInst::ICMP_ULE)
7143 return BinaryOperator::CreateOr(V1: Builder.CreateIsNull(Arg: X), V2: Y);
7144
7145 // icmp eq/ne X, (zext/sext (icmp eq/ne X, C))
7146 CmpPredicate Pred1, Pred2;
7147 const APInt *C;
7148 Instruction *ExtI;
7149 if (match(V: &I, P: m_c_ICmp(Pred&: Pred1, L: m_Value(V&: X),
7150 R: m_CombineAnd(L: m_Instruction(I&: ExtI),
7151 R: m_ZExtOrSExt(Op: m_ICmp(Pred&: Pred2, L: m_Deferred(V: X),
7152 R: m_APInt(Res&: C)))))) &&
7153 ICmpInst::isEquality(P: Pred1) && ICmpInst::isEquality(P: Pred2)) {
7154 bool IsSExt = ExtI->getOpcode() == Instruction::SExt;
7155 bool HasOneUse = ExtI->hasOneUse() && ExtI->getOperand(i: 0)->hasOneUse();
7156 auto CreateRangeCheck = [&] {
7157 Value *CmpV1 =
7158 Builder.CreateICmp(P: Pred1, LHS: X, RHS: Constant::getNullValue(Ty: X->getType()));
7159 Value *CmpV2 = Builder.CreateICmp(
7160 P: Pred1, LHS: X, RHS: ConstantInt::getSigned(Ty: X->getType(), V: IsSExt ? -1 : 1));
7161 return BinaryOperator::Create(
7162 Op: Pred1 == ICmpInst::ICMP_EQ ? Instruction::Or : Instruction::And,
7163 S1: CmpV1, S2: CmpV2);
7164 };
7165 if (C->isZero()) {
7166 if (Pred2 == ICmpInst::ICMP_EQ) {
7167 // icmp eq X, (zext/sext (icmp eq X, 0)) --> false
7168 // icmp ne X, (zext/sext (icmp eq X, 0)) --> true
7169 return replaceInstUsesWith(
7170 I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred1 == ICmpInst::ICMP_NE));
7171 } else if (!IsSExt || HasOneUse) {
7172 // icmp eq X, (zext (icmp ne X, 0)) --> X == 0 || X == 1
7173 // icmp ne X, (zext (icmp ne X, 0)) --> X != 0 && X != 1
7174 // icmp eq X, (sext (icmp ne X, 0)) --> X == 0 || X == -1
7175 // icmp ne X, (sext (icmp ne X, 0)) --> X != 0 && X != -1
7176 return CreateRangeCheck();
7177 }
7178 } else if (IsSExt ? C->isAllOnes() : C->isOne()) {
7179 if (Pred2 == ICmpInst::ICMP_NE) {
7180 // icmp eq X, (zext (icmp ne X, 1)) --> false
7181 // icmp ne X, (zext (icmp ne X, 1)) --> true
7182 // icmp eq X, (sext (icmp ne X, -1)) --> false
7183 // icmp ne X, (sext (icmp ne X, -1)) --> true
7184 return replaceInstUsesWith(
7185 I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred1 == ICmpInst::ICMP_NE));
7186 } else if (!IsSExt || HasOneUse) {
7187 // icmp eq X, (zext (icmp eq X, 1)) --> X == 0 || X == 1
7188 // icmp ne X, (zext (icmp eq X, 1)) --> X != 0 && X != 1
7189 // icmp eq X, (sext (icmp eq X, -1)) --> X == 0 || X == -1
7190 // icmp ne X, (sext (icmp eq X, -1)) --> X != 0 && X == -1
7191 return CreateRangeCheck();
7192 }
7193 } else {
7194 // when C != 0 && C != 1:
7195 // icmp eq X, (zext (icmp eq X, C)) --> icmp eq X, 0
7196 // icmp eq X, (zext (icmp ne X, C)) --> icmp eq X, 1
7197 // icmp ne X, (zext (icmp eq X, C)) --> icmp ne X, 0
7198 // icmp ne X, (zext (icmp ne X, C)) --> icmp ne X, 1
7199 // when C != 0 && C != -1:
7200 // icmp eq X, (sext (icmp eq X, C)) --> icmp eq X, 0
7201 // icmp eq X, (sext (icmp ne X, C)) --> icmp eq X, -1
7202 // icmp ne X, (sext (icmp eq X, C)) --> icmp ne X, 0
7203 // icmp ne X, (sext (icmp ne X, C)) --> icmp ne X, -1
7204 return ICmpInst::Create(
7205 Op: Instruction::ICmp, Pred: Pred1, S1: X,
7206 S2: ConstantInt::getSigned(Ty: X->getType(), V: Pred2 == ICmpInst::ICMP_NE
7207 ? (IsSExt ? -1 : 1)
7208 : 0));
7209 }
7210 }
7211
7212 return nullptr;
7213}
7214
7215/// If we have an icmp le or icmp ge instruction with a constant operand, turn
7216/// it into the appropriate icmp lt or icmp gt instruction. This transform
7217/// allows them to be folded in visitICmpInst.
7218static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
7219 ICmpInst::Predicate Pred = I.getPredicate();
7220 if (ICmpInst::isEquality(P: Pred) || !ICmpInst::isIntPredicate(P: Pred) ||
7221 InstCombiner::isCanonicalPredicate(Pred))
7222 return nullptr;
7223
7224 Value *Op0 = I.getOperand(i_nocapture: 0);
7225 Value *Op1 = I.getOperand(i_nocapture: 1);
7226 auto *Op1C = dyn_cast<Constant>(Val: Op1);
7227 if (!Op1C)
7228 return nullptr;
7229
7230 auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, C: Op1C);
7231 if (!FlippedStrictness)
7232 return nullptr;
7233
7234 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
7235}
7236
7237/// If we have a comparison with a non-canonical predicate, if we can update
7238/// all the users, invert the predicate and adjust all the users.
7239CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
7240 // Is the predicate already canonical?
7241 CmpInst::Predicate Pred = I.getPredicate();
7242 if (InstCombiner::isCanonicalPredicate(Pred))
7243 return nullptr;
7244
7245 // Can all users be adjusted to predicate inversion?
7246 if (!InstCombiner::canFreelyInvertAllUsersOf(V: &I, /*IgnoredUser=*/nullptr))
7247 return nullptr;
7248
7249 // Ok, we can canonicalize comparison!
7250 // Let's first invert the comparison's predicate.
7251 I.setPredicate(CmpInst::getInversePredicate(pred: Pred));
7252 I.setName(I.getName() + ".not");
7253
7254 // And, adapt users.
7255 freelyInvertAllUsersOf(V: &I);
7256
7257 return &I;
7258}
7259
7260/// Integer compare with boolean values can always be turned into bitwise ops.
7261static Instruction *canonicalizeICmpBool(ICmpInst &I,
7262 InstCombiner::BuilderTy &Builder) {
7263 Value *A = I.getOperand(i_nocapture: 0), *B = I.getOperand(i_nocapture: 1);
7264 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
7265
7266 // A boolean compared to true/false can be simplified to Op0/true/false in
7267 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
7268 // Cases not handled by InstSimplify are always 'not' of Op0.
7269 if (match(V: B, P: m_Zero())) {
7270 switch (I.getPredicate()) {
7271 case CmpInst::ICMP_EQ: // A == 0 -> !A
7272 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
7273 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
7274 return BinaryOperator::CreateNot(Op: A);
7275 default:
7276 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
7277 }
7278 } else if (match(V: B, P: m_One())) {
7279 switch (I.getPredicate()) {
7280 case CmpInst::ICMP_NE: // A != 1 -> !A
7281 case CmpInst::ICMP_ULT: // A <u 1 -> !A
7282 case CmpInst::ICMP_SGT: // A >s -1 -> !A
7283 return BinaryOperator::CreateNot(Op: A);
7284 default:
7285 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
7286 }
7287 }
7288
7289 switch (I.getPredicate()) {
7290 default:
7291 llvm_unreachable("Invalid icmp instruction!");
7292 case ICmpInst::ICMP_EQ:
7293 // icmp eq i1 A, B -> ~(A ^ B)
7294 return BinaryOperator::CreateNot(Op: Builder.CreateXor(LHS: A, RHS: B));
7295
7296 case ICmpInst::ICMP_NE:
7297 // icmp ne i1 A, B -> A ^ B
7298 return BinaryOperator::CreateXor(V1: A, V2: B);
7299
7300 case ICmpInst::ICMP_UGT:
7301 // icmp ugt -> icmp ult
7302 std::swap(a&: A, b&: B);
7303 [[fallthrough]];
7304 case ICmpInst::ICMP_ULT:
7305 // icmp ult i1 A, B -> ~A & B
7306 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: A), V2: B);
7307
7308 case ICmpInst::ICMP_SGT:
7309 // icmp sgt -> icmp slt
7310 std::swap(a&: A, b&: B);
7311 [[fallthrough]];
7312 case ICmpInst::ICMP_SLT:
7313 // icmp slt i1 A, B -> A & ~B
7314 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: B), V2: A);
7315
7316 case ICmpInst::ICMP_UGE:
7317 // icmp uge -> icmp ule
7318 std::swap(a&: A, b&: B);
7319 [[fallthrough]];
7320 case ICmpInst::ICMP_ULE:
7321 // icmp ule i1 A, B -> ~A | B
7322 return BinaryOperator::CreateOr(V1: Builder.CreateNot(V: A), V2: B);
7323
7324 case ICmpInst::ICMP_SGE:
7325 // icmp sge -> icmp sle
7326 std::swap(a&: A, b&: B);
7327 [[fallthrough]];
7328 case ICmpInst::ICMP_SLE:
7329 // icmp sle i1 A, B -> A | ~B
7330 return BinaryOperator::CreateOr(V1: Builder.CreateNot(V: B), V2: A);
7331 }
7332}
7333
7334// Transform pattern like:
7335// (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
7336// (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
7337// Into:
7338// (X l>> Y) != 0
7339// (X l>> Y) == 0
7340static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
7341 InstCombiner::BuilderTy &Builder) {
7342 CmpPredicate Pred, NewPred;
7343 Value *X, *Y;
7344 if (match(V: &Cmp,
7345 P: m_c_ICmp(Pred, L: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: Y))), R: m_Value(V&: X)))) {
7346 switch (Pred) {
7347 case ICmpInst::ICMP_ULE:
7348 NewPred = ICmpInst::ICMP_NE;
7349 break;
7350 case ICmpInst::ICMP_UGT:
7351 NewPred = ICmpInst::ICMP_EQ;
7352 break;
7353 default:
7354 return nullptr;
7355 }
7356 } else if (match(V: &Cmp, P: m_c_ICmp(Pred,
7357 L: m_OneUse(SubPattern: m_CombineOr(
7358 L: m_Not(V: m_Shl(L: m_AllOnes(), R: m_Value(V&: Y))),
7359 R: m_Add(L: m_Shl(L: m_One(), R: m_Value(V&: Y)),
7360 R: m_AllOnes()))),
7361 R: m_Value(V&: X)))) {
7362 // The variant with 'add' is not canonical, (the variant with 'not' is)
7363 // we only get it because it has extra uses, and can't be canonicalized,
7364
7365 switch (Pred) {
7366 case ICmpInst::ICMP_ULT:
7367 NewPred = ICmpInst::ICMP_NE;
7368 break;
7369 case ICmpInst::ICMP_UGE:
7370 NewPred = ICmpInst::ICMP_EQ;
7371 break;
7372 default:
7373 return nullptr;
7374 }
7375 } else
7376 return nullptr;
7377
7378 Value *NewX = Builder.CreateLShr(LHS: X, RHS: Y, Name: X->getName() + ".highbits");
7379 Constant *Zero = Constant::getNullValue(Ty: NewX->getType());
7380 return CmpInst::Create(Op: Instruction::ICmp, Pred: NewPred, S1: NewX, S2: Zero);
7381}
7382
7383static Instruction *foldVectorCmp(CmpInst &Cmp,
7384 InstCombiner::BuilderTy &Builder) {
7385 const CmpInst::Predicate Pred = Cmp.getPredicate();
7386 Value *LHS = Cmp.getOperand(i_nocapture: 0), *RHS = Cmp.getOperand(i_nocapture: 1);
7387 Value *V1, *V2;
7388
7389 auto createCmpReverse = [&](CmpInst::Predicate Pred, Value *X, Value *Y) {
7390 Value *V = Builder.CreateCmp(Pred, LHS: X, RHS: Y, Name: Cmp.getName());
7391 if (auto *I = dyn_cast<Instruction>(Val: V))
7392 I->copyIRFlags(V: &Cmp);
7393 Module *M = Cmp.getModule();
7394 Function *F = Intrinsic::getOrInsertDeclaration(
7395 M, id: Intrinsic::vector_reverse, Tys: V->getType());
7396 return CallInst::Create(Func: F, Args: V);
7397 };
7398
7399 if (match(V: LHS, P: m_VecReverse(Op0: m_Value(V&: V1)))) {
7400 // cmp Pred, rev(V1), rev(V2) --> rev(cmp Pred, V1, V2)
7401 if (match(V: RHS, P: m_VecReverse(Op0: m_Value(V&: V2))) &&
7402 (LHS->hasOneUse() || RHS->hasOneUse()))
7403 return createCmpReverse(Pred, V1, V2);
7404
7405 // cmp Pred, rev(V1), RHSSplat --> rev(cmp Pred, V1, RHSSplat)
7406 if (LHS->hasOneUse() && isSplatValue(V: RHS))
7407 return createCmpReverse(Pred, V1, RHS);
7408 }
7409 // cmp Pred, LHSSplat, rev(V2) --> rev(cmp Pred, LHSSplat, V2)
7410 else if (isSplatValue(V: LHS) && match(V: RHS, P: m_OneUse(SubPattern: m_VecReverse(Op0: m_Value(V&: V2)))))
7411 return createCmpReverse(Pred, LHS, V2);
7412
7413 ArrayRef<int> M;
7414 if (!match(V: LHS, P: m_Shuffle(v1: m_Value(V&: V1), v2: m_Undef(), mask: m_Mask(M))))
7415 return nullptr;
7416
7417 // If both arguments of the cmp are shuffles that use the same mask and
7418 // shuffle within a single vector, move the shuffle after the cmp:
7419 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
7420 Type *V1Ty = V1->getType();
7421 if (match(V: RHS, P: m_Shuffle(v1: m_Value(V&: V2), v2: m_Undef(), mask: m_SpecificMask(M))) &&
7422 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
7423 Value *NewCmp = Builder.CreateCmp(Pred, LHS: V1, RHS: V2);
7424 return new ShuffleVectorInst(NewCmp, M);
7425 }
7426
7427 // Try to canonicalize compare with splatted operand and splat constant.
7428 // TODO: We could generalize this for more than splats. See/use the code in
7429 // InstCombiner::foldVectorBinop().
7430 Constant *C;
7431 if (!LHS->hasOneUse() || !match(V: RHS, P: m_Constant(C)))
7432 return nullptr;
7433
7434 // Length-changing splats are ok, so adjust the constants as needed:
7435 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
7436 Constant *ScalarC = C->getSplatValue(/* AllowPoison */ true);
7437 int MaskSplatIndex;
7438 if (ScalarC && match(Mask: M, P: m_SplatOrPoisonMask(MaskSplatIndex))) {
7439 // We allow poison in matching, but this transform removes it for safety.
7440 // Demanded elements analysis should be able to recover some/all of that.
7441 C = ConstantVector::getSplat(EC: cast<VectorType>(Val: V1Ty)->getElementCount(),
7442 Elt: ScalarC);
7443 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
7444 Value *NewCmp = Builder.CreateCmp(Pred, LHS: V1, RHS: C);
7445 return new ShuffleVectorInst(NewCmp, NewM);
7446 }
7447
7448 return nullptr;
7449}
7450
7451// extract(uadd.with.overflow(A, B), 0) ult A
7452// -> extract(uadd.with.overflow(A, B), 1)
7453static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
7454 CmpInst::Predicate Pred = I.getPredicate();
7455 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
7456
7457 Value *UAddOv;
7458 Value *A, *B;
7459 auto UAddOvResultPat = m_ExtractValue<0>(
7460 V: m_Intrinsic<Intrinsic::uadd_with_overflow>(Op0: m_Value(V&: A), Op1: m_Value(V&: B)));
7461 if (match(V: Op0, P: UAddOvResultPat) &&
7462 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
7463 (Pred == ICmpInst::ICMP_EQ && match(V: Op1, P: m_ZeroInt()) &&
7464 (match(V: A, P: m_One()) || match(V: B, P: m_One()))) ||
7465 (Pred == ICmpInst::ICMP_NE && match(V: Op1, P: m_AllOnes()) &&
7466 (match(V: A, P: m_AllOnes()) || match(V: B, P: m_AllOnes())))))
7467 // extract(uadd.with.overflow(A, B), 0) < A
7468 // extract(uadd.with.overflow(A, 1), 0) == 0
7469 // extract(uadd.with.overflow(A, -1), 0) != -1
7470 UAddOv = cast<ExtractValueInst>(Val: Op0)->getAggregateOperand();
7471 else if (match(V: Op1, P: UAddOvResultPat) && Pred == ICmpInst::ICMP_UGT &&
7472 (Op0 == A || Op0 == B))
7473 // A > extract(uadd.with.overflow(A, B), 0)
7474 UAddOv = cast<ExtractValueInst>(Val: Op1)->getAggregateOperand();
7475 else
7476 return nullptr;
7477
7478 return ExtractValueInst::Create(Agg: UAddOv, Idxs: 1);
7479}
7480
7481static Instruction *foldICmpInvariantGroup(ICmpInst &I) {
7482 if (!I.getOperand(i_nocapture: 0)->getType()->isPointerTy() ||
7483 NullPointerIsDefined(
7484 F: I.getParent()->getParent(),
7485 AS: I.getOperand(i_nocapture: 0)->getType()->getPointerAddressSpace())) {
7486 return nullptr;
7487 }
7488 Instruction *Op;
7489 if (match(V: I.getOperand(i_nocapture: 0), P: m_Instruction(I&: Op)) &&
7490 match(V: I.getOperand(i_nocapture: 1), P: m_Zero()) &&
7491 Op->isLaunderOrStripInvariantGroup()) {
7492 return ICmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(),
7493 S1: Op->getOperand(i: 0), S2: I.getOperand(i_nocapture: 1));
7494 }
7495 return nullptr;
7496}
7497
7498static Instruction *foldICmpOfVectorReduce(ICmpInst &I, const DataLayout &DL,
7499 IRBuilderBase &Builder) {
7500 if (!ICmpInst::isEquality(P: I.getPredicate()))
7501 return nullptr;
7502
7503 // The caller puts constants after non-constants.
7504 Value *Op = I.getOperand(i_nocapture: 0);
7505 Value *Const = I.getOperand(i_nocapture: 1);
7506
7507 // For Cond an equality condition, fold
7508 //
7509 // icmp (eq|ne) (vreduce_(or|and) Op), (Zero|AllOnes) ->
7510 // icmp (eq|ne) Op, (Zero|AllOnes)
7511 //
7512 // with a bitcast.
7513 Value *Vec;
7514 if ((match(V: Const, P: m_ZeroInt()) &&
7515 match(V: Op, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_or>(
7516 Op0: m_Value(V&: Vec))))) ||
7517 (match(V: Const, P: m_AllOnes()) &&
7518 match(V: Op, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_and>(
7519 Op0: m_Value(V&: Vec)))))) {
7520 auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType());
7521 if (!VecTy)
7522 return nullptr;
7523 Type *VecEltTy = VecTy->getElementType();
7524 unsigned ScalarBW =
7525 DL.getTypeSizeInBits(Ty: VecEltTy) * VecTy->getNumElements();
7526 if (!DL.fitsInLegalInteger(Width: ScalarBW))
7527 return nullptr;
7528 Type *ScalarTy = IntegerType::get(C&: I.getContext(), NumBits: ScalarBW);
7529 Value *NewConst = match(V: Const, P: m_ZeroInt())
7530 ? ConstantInt::get(Ty: ScalarTy, V: 0)
7531 : ConstantInt::getAllOnesValue(Ty: ScalarTy);
7532 return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(),
7533 S1: Builder.CreateBitCast(V: Vec, DestTy: ScalarTy), S2: NewConst);
7534 }
7535 return nullptr;
7536}
7537
7538/// This function folds patterns produced by lowering of reduce idioms, such as
7539/// llvm.vector.reduce.and which are lowered into instruction chains. This code
7540/// attempts to generate fewer number of scalar comparisons instead of vector
7541/// comparisons when possible.
7542static Instruction *foldReductionIdiom(ICmpInst &I,
7543 InstCombiner::BuilderTy &Builder,
7544 const DataLayout &DL) {
7545 if (I.getType()->isVectorTy())
7546 return nullptr;
7547 CmpPredicate OuterPred, InnerPred;
7548 Value *LHS, *RHS;
7549
7550 // Match lowering of @llvm.vector.reduce.and. Turn
7551 /// %vec_ne = icmp ne <8 x i8> %lhs, %rhs
7552 /// %scalar_ne = bitcast <8 x i1> %vec_ne to i8
7553 /// %res = icmp <pred> i8 %scalar_ne, 0
7554 ///
7555 /// into
7556 ///
7557 /// %lhs.scalar = bitcast <8 x i8> %lhs to i64
7558 /// %rhs.scalar = bitcast <8 x i8> %rhs to i64
7559 /// %res = icmp <pred> i64 %lhs.scalar, %rhs.scalar
7560 ///
7561 /// for <pred> in {ne, eq}.
7562 if (!match(V: &I, P: m_ICmp(Pred&: OuterPred,
7563 L: m_OneUse(SubPattern: m_BitCast(Op: m_OneUse(
7564 SubPattern: m_ICmp(Pred&: InnerPred, L: m_Value(V&: LHS), R: m_Value(V&: RHS))))),
7565 R: m_Zero())))
7566 return nullptr;
7567 auto *LHSTy = dyn_cast<FixedVectorType>(Val: LHS->getType());
7568 if (!LHSTy || !LHSTy->getElementType()->isIntegerTy())
7569 return nullptr;
7570 unsigned NumBits =
7571 LHSTy->getNumElements() * LHSTy->getElementType()->getIntegerBitWidth();
7572 // TODO: Relax this to "not wider than max legal integer type"?
7573 if (!DL.isLegalInteger(Width: NumBits))
7574 return nullptr;
7575
7576 if (ICmpInst::isEquality(P: OuterPred) && InnerPred == ICmpInst::ICMP_NE) {
7577 auto *ScalarTy = Builder.getIntNTy(N: NumBits);
7578 LHS = Builder.CreateBitCast(V: LHS, DestTy: ScalarTy, Name: LHS->getName() + ".scalar");
7579 RHS = Builder.CreateBitCast(V: RHS, DestTy: ScalarTy, Name: RHS->getName() + ".scalar");
7580 return ICmpInst::Create(Op: Instruction::ICmp, Pred: OuterPred, S1: LHS, S2: RHS,
7581 Name: I.getName());
7582 }
7583
7584 return nullptr;
7585}
7586
7587// This helper will be called with icmp operands in both orders.
7588Instruction *InstCombinerImpl::foldICmpCommutative(CmpPredicate Pred,
7589 Value *Op0, Value *Op1,
7590 ICmpInst &CxtI) {
7591 // Try to optimize 'icmp GEP, P' or 'icmp P, GEP'.
7592 if (auto *GEP = dyn_cast<GEPOperator>(Val: Op0))
7593 if (Instruction *NI = foldGEPICmp(GEPLHS: GEP, RHS: Op1, Cond: Pred, I&: CxtI))
7594 return NI;
7595
7596 if (auto *SI = dyn_cast<SelectInst>(Val: Op0))
7597 if (Instruction *NI = foldSelectICmp(Pred, SI, RHS: Op1, I: CxtI))
7598 return NI;
7599
7600 if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op0)) {
7601 if (Instruction *Res = foldICmpWithMinMax(I&: CxtI, MinMax, Z: Op1, Pred))
7602 return Res;
7603
7604 if (Instruction *Res = foldICmpWithClamp(I&: CxtI, X: Op1, Min: MinMax))
7605 return Res;
7606 }
7607
7608 {
7609 Value *X;
7610 const APInt *C;
7611 // icmp X+Cst, X
7612 if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C))) && Op1 == X)
7613 return foldICmpAddOpConst(X, C: *C, Pred);
7614 }
7615
7616 // abs(X) >= X --> true
7617 // abs(X) u<= X --> true
7618 // abs(X) < X --> false
7619 // abs(X) u> X --> false
7620 // abs(X) u>= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7621 // abs(X) <= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7622 // abs(X) == X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7623 // abs(X) u< X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7624 // abs(X) > X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7625 // abs(X) != X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7626 {
7627 Value *X;
7628 Constant *C;
7629 if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X), Op1: m_Constant(C))) &&
7630 match(V: Op1, P: m_Specific(V: X))) {
7631 Value *NullValue = Constant::getNullValue(Ty: X->getType());
7632 Value *AllOnesValue = Constant::getAllOnesValue(Ty: X->getType());
7633 const APInt SMin =
7634 APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits());
7635 bool IsIntMinPosion = C->isAllOnesValue();
7636 switch (Pred) {
7637 case CmpInst::ICMP_ULE:
7638 case CmpInst::ICMP_SGE:
7639 return replaceInstUsesWith(I&: CxtI, V: ConstantInt::getTrue(Ty: CxtI.getType()));
7640 case CmpInst::ICMP_UGT:
7641 case CmpInst::ICMP_SLT:
7642 return replaceInstUsesWith(I&: CxtI, V: ConstantInt::getFalse(Ty: CxtI.getType()));
7643 case CmpInst::ICMP_UGE:
7644 case CmpInst::ICMP_SLE:
7645 case CmpInst::ICMP_EQ: {
7646 return replaceInstUsesWith(
7647 I&: CxtI, V: IsIntMinPosion
7648 ? Builder.CreateICmpSGT(LHS: X, RHS: AllOnesValue)
7649 : Builder.CreateICmpULT(
7650 LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: SMin + 1)));
7651 }
7652 case CmpInst::ICMP_ULT:
7653 case CmpInst::ICMP_SGT:
7654 case CmpInst::ICMP_NE: {
7655 return replaceInstUsesWith(
7656 I&: CxtI, V: IsIntMinPosion
7657 ? Builder.CreateICmpSLT(LHS: X, RHS: NullValue)
7658 : Builder.CreateICmpUGT(
7659 LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: SMin)));
7660 }
7661 default:
7662 llvm_unreachable("Invalid predicate!");
7663 }
7664 }
7665 }
7666
7667 const SimplifyQuery Q = SQ.getWithInstruction(I: &CxtI);
7668 if (Value *V = foldICmpWithLowBitMaskedVal(Pred, Op0, Op1, Q, IC&: *this))
7669 return replaceInstUsesWith(I&: CxtI, V);
7670
7671 // Folding (X / Y) pred X => X swap(pred) 0 for constant Y other than 0 or 1
7672 auto CheckUGT1 = [](const APInt &Divisor) { return Divisor.ugt(RHS: 1); };
7673 {
7674 if (match(V: Op0, P: m_UDiv(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckUGT1)))) {
7675 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7676 Constant::getNullValue(Ty: Op1->getType()));
7677 }
7678
7679 if (!ICmpInst::isUnsigned(Pred) &&
7680 match(V: Op0, P: m_SDiv(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckUGT1)))) {
7681 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7682 Constant::getNullValue(Ty: Op1->getType()));
7683 }
7684 }
7685
7686 // Another case of this fold is (X >> Y) pred X => X swap(pred) 0 if Y != 0
7687 auto CheckNE0 = [](const APInt &Shift) { return !Shift.isZero(); };
7688 {
7689 if (match(V: Op0, P: m_LShr(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckNE0)))) {
7690 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7691 Constant::getNullValue(Ty: Op1->getType()));
7692 }
7693
7694 if ((Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SGE) &&
7695 match(V: Op0, P: m_AShr(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckNE0)))) {
7696 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7697 Constant::getNullValue(Ty: Op1->getType()));
7698 }
7699 }
7700
7701 return nullptr;
7702}
7703
7704Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
7705 bool Changed = false;
7706 const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
7707 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
7708 unsigned Op0Cplxity = getComplexity(V: Op0);
7709 unsigned Op1Cplxity = getComplexity(V: Op1);
7710
7711 /// Orders the operands of the compare so that they are listed from most
7712 /// complex to least complex. This puts constants before unary operators,
7713 /// before binary operators.
7714 if (Op0Cplxity < Op1Cplxity) {
7715 I.swapOperands();
7716 std::swap(a&: Op0, b&: Op1);
7717 Changed = true;
7718 }
7719
7720 if (Value *V = simplifyICmpInst(Pred: I.getCmpPredicate(), LHS: Op0, RHS: Op1, Q))
7721 return replaceInstUsesWith(I, V);
7722
7723 // Comparing -val or val with non-zero is the same as just comparing val
7724 // ie, abs(val) != 0 -> val != 0
7725 if (I.getPredicate() == ICmpInst::ICMP_NE && match(V: Op1, P: m_Zero())) {
7726 Value *Cond, *SelectTrue, *SelectFalse;
7727 if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: SelectTrue),
7728 R: m_Value(V&: SelectFalse)))) {
7729 if (Value *V = dyn_castNegVal(V: SelectTrue)) {
7730 if (V == SelectFalse)
7731 return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), S1: V, S2: Op1);
7732 } else if (Value *V = dyn_castNegVal(V: SelectFalse)) {
7733 if (V == SelectTrue)
7734 return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), S1: V, S2: Op1);
7735 }
7736 }
7737 }
7738
7739 if (Instruction *Res = foldICmpTruncWithTruncOrExt(Cmp&: I, Q))
7740 return Res;
7741
7742 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1))
7743 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
7744 return Res;
7745
7746 if (Instruction *Res = canonicalizeCmpWithConstant(I))
7747 return Res;
7748
7749 if (Instruction *Res = canonicalizeICmpPredicate(I))
7750 return Res;
7751
7752 if (Instruction *Res = foldICmpWithConstant(Cmp&: I))
7753 return Res;
7754
7755 if (Instruction *Res = foldICmpWithDominatingICmp(Cmp&: I))
7756 return Res;
7757
7758 if (Instruction *Res = foldICmpUsingBoolRange(I))
7759 return Res;
7760
7761 if (Instruction *Res = foldICmpUsingKnownBits(I))
7762 return Res;
7763
7764 if (Instruction *Res = foldIsMultipleOfAPowerOfTwo(Cmp&: I))
7765 return Res;
7766
7767 // Test if the ICmpInst instruction is used exclusively by a select as
7768 // part of a minimum or maximum operation. If so, refrain from doing
7769 // any other folding. This helps out other analyses which understand
7770 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
7771 // and CodeGen. And in this case, at least one of the comparison
7772 // operands has at least one user besides the compare (the select),
7773 // which would often largely negate the benefit of folding anyway.
7774 //
7775 // Do the same for the other patterns recognized by matchSelectPattern.
7776 if (I.hasOneUse())
7777 if (SelectInst *SI = dyn_cast<SelectInst>(Val: I.user_back())) {
7778 Value *A, *B;
7779 SelectPatternResult SPR = matchSelectPattern(V: SI, LHS&: A, RHS&: B);
7780 if (SPR.Flavor != SPF_UNKNOWN)
7781 return nullptr;
7782 }
7783
7784 // Do this after checking for min/max to prevent infinite looping.
7785 if (Instruction *Res = foldICmpWithZero(Cmp&: I))
7786 return Res;
7787
7788 // FIXME: We only do this after checking for min/max to prevent infinite
7789 // looping caused by a reverse canonicalization of these patterns for min/max.
7790 // FIXME: The organization of folds is a mess. These would naturally go into
7791 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
7792 // down here after the min/max restriction.
7793 ICmpInst::Predicate Pred = I.getPredicate();
7794 const APInt *C;
7795 if (match(V: Op1, P: m_APInt(Res&: C))) {
7796 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
7797 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
7798 Constant *Zero = Constant::getNullValue(Ty: Op0->getType());
7799 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
7800 }
7801
7802 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
7803 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
7804 Constant *AllOnes = Constant::getAllOnesValue(Ty: Op0->getType());
7805 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
7806 }
7807 }
7808
7809 // The folds in here may rely on wrapping flags and special constants, so
7810 // they can break up min/max idioms in some cases but not seemingly similar
7811 // patterns.
7812 // FIXME: It may be possible to enhance select folding to make this
7813 // unnecessary. It may also be moot if we canonicalize to min/max
7814 // intrinsics.
7815 if (Instruction *Res = foldICmpBinOp(I, SQ: Q))
7816 return Res;
7817
7818 if (Instruction *Res = foldICmpInstWithConstant(Cmp&: I))
7819 return Res;
7820
7821 // Try to match comparison as a sign bit test. Intentionally do this after
7822 // foldICmpInstWithConstant() to potentially let other folds to happen first.
7823 if (Instruction *New = foldSignBitTest(I))
7824 return New;
7825
7826 if (auto *PN = dyn_cast<PHINode>(Val: Op0))
7827 if (Instruction *NV = foldOpIntoPhi(I, PN))
7828 return NV;
7829 if (auto *PN = dyn_cast<PHINode>(Val: Op1))
7830 if (Instruction *NV = foldOpIntoPhi(I, PN))
7831 return NV;
7832
7833 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
7834 return Res;
7835
7836 if (Instruction *Res = foldICmpCommutative(Pred: I.getCmpPredicate(), Op0, Op1, CxtI&: I))
7837 return Res;
7838 if (Instruction *Res =
7839 foldICmpCommutative(Pred: I.getSwappedCmpPredicate(), Op0: Op1, Op1: Op0, CxtI&: I))
7840 return Res;
7841
7842 if (I.isCommutative()) {
7843 if (auto Pair = matchSymmetricPair(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1))) {
7844 replaceOperand(I, OpNum: 0, V: Pair->first);
7845 replaceOperand(I, OpNum: 1, V: Pair->second);
7846 return &I;
7847 }
7848 }
7849
7850 // In case of a comparison with two select instructions having the same
7851 // condition, check whether one of the resulting branches can be simplified.
7852 // If so, just compare the other branch and select the appropriate result.
7853 // For example:
7854 // %tmp1 = select i1 %cmp, i32 %y, i32 %x
7855 // %tmp2 = select i1 %cmp, i32 %z, i32 %x
7856 // %cmp2 = icmp slt i32 %tmp2, %tmp1
7857 // The icmp will result false for the false value of selects and the result
7858 // will depend upon the comparison of true values of selects if %cmp is
7859 // true. Thus, transform this into:
7860 // %cmp = icmp slt i32 %y, %z
7861 // %sel = select i1 %cond, i1 %cmp, i1 false
7862 // This handles similar cases to transform.
7863 {
7864 Value *Cond, *A, *B, *C, *D;
7865 if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: A), R: m_Value(V&: B))) &&
7866 match(V: Op1, P: m_Select(C: m_Specific(V: Cond), L: m_Value(V&: C), R: m_Value(V&: D))) &&
7867 (Op0->hasOneUse() || Op1->hasOneUse())) {
7868 // Check whether comparison of TrueValues can be simplified
7869 if (Value *Res = simplifyICmpInst(Pred, LHS: A, RHS: C, Q: SQ)) {
7870 Value *NewICMP = Builder.CreateICmp(P: Pred, LHS: B, RHS: D);
7871 return SelectInst::Create(
7872 C: Cond, S1: Res, S2: NewICMP, /*NameStr=*/"", /*InsertBefore=*/nullptr,
7873 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : cast<Instruction>(Val: Op0));
7874 }
7875 // Check whether comparison of FalseValues can be simplified
7876 if (Value *Res = simplifyICmpInst(Pred, LHS: B, RHS: D, Q: SQ)) {
7877 Value *NewICMP = Builder.CreateICmp(P: Pred, LHS: A, RHS: C);
7878 return SelectInst::Create(
7879 C: Cond, S1: NewICMP, S2: Res, /*NameStr=*/"", /*InsertBefore=*/nullptr,
7880 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : cast<Instruction>(Val: Op0));
7881 }
7882 }
7883 }
7884
7885 // icmp slt (sub nsw x, y), (add nsw x, y) --> icmp sgt y, 0
7886 // icmp ult (sub nuw x, y), (add nuw x, y) --> icmp ugt y, 0
7887 // icmp eq (sub nsw/nuw x, y), (add nsw/nuw x, y) --> icmp eq y, 0
7888 {
7889 Value *A, *B;
7890 CmpPredicate CmpPred;
7891 if (match(V: &I, P: m_c_ICmp(Pred&: CmpPred, L: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B)),
7892 R: m_c_Add(L: m_Deferred(V: A), R: m_Deferred(V: B))))) {
7893 auto *I0 = cast<OverflowingBinaryOperator>(Val: Op0);
7894 auto *I1 = cast<OverflowingBinaryOperator>(Val: Op1);
7895 bool I0NUW = I0->hasNoUnsignedWrap();
7896 bool I1NUW = I1->hasNoUnsignedWrap();
7897 bool I0NSW = I0->hasNoSignedWrap();
7898 bool I1NSW = I1->hasNoSignedWrap();
7899 if ((ICmpInst::isUnsigned(Pred) && I0NUW && I1NUW) ||
7900 (ICmpInst::isSigned(Pred) && I0NSW && I1NSW) ||
7901 (ICmpInst::isEquality(P: Pred) &&
7902 ((I0NUW || I0NSW) && (I1NUW || I1NSW)))) {
7903 return new ICmpInst(CmpPredicate::getSwapped(P: CmpPred), B,
7904 ConstantInt::get(Ty: Op0->getType(), V: 0));
7905 }
7906 }
7907 }
7908
7909 // Try to optimize equality comparisons against alloca-based pointers.
7910 if (Op0->getType()->isPointerTy() && I.isEquality()) {
7911 assert(Op1->getType()->isPointerTy() &&
7912 "Comparing pointer with non-pointer?");
7913 if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Op0)))
7914 if (foldAllocaCmp(Alloca))
7915 return nullptr;
7916 if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Op1)))
7917 if (foldAllocaCmp(Alloca))
7918 return nullptr;
7919 }
7920
7921 if (Instruction *Res = foldICmpBitCast(Cmp&: I))
7922 return Res;
7923
7924 // TODO: Hoist this above the min/max bailout.
7925 if (Instruction *R = foldICmpWithCastOp(ICmp&: I))
7926 return R;
7927
7928 {
7929 Value *X, *Y;
7930 // Transform (X & ~Y) == 0 --> (X & Y) != 0
7931 // and (X & ~Y) != 0 --> (X & Y) == 0
7932 // if A is a power of 2.
7933 if (match(V: Op0, P: m_And(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
7934 match(V: Op1, P: m_Zero()) && isKnownToBeAPowerOfTwo(V: X, OrZero: false, CxtI: &I) &&
7935 I.isEquality())
7936 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(LHS: X, RHS: Y),
7937 Op1);
7938
7939 // Op0 pred Op1 -> ~Op1 pred ~Op0, if this allows us to drop an instruction.
7940 if (Op0->getType()->isIntOrIntVectorTy()) {
7941 bool ConsumesOp0, ConsumesOp1;
7942 if (isFreeToInvert(V: Op0, WillInvertAllUses: Op0->hasOneUse(), DoesConsume&: ConsumesOp0) &&
7943 isFreeToInvert(V: Op1, WillInvertAllUses: Op1->hasOneUse(), DoesConsume&: ConsumesOp1) &&
7944 (ConsumesOp0 || ConsumesOp1)) {
7945 Value *InvOp0 = getFreelyInverted(V: Op0, WillInvertAllUses: Op0->hasOneUse(), Builder: &Builder);
7946 Value *InvOp1 = getFreelyInverted(V: Op1, WillInvertAllUses: Op1->hasOneUse(), Builder: &Builder);
7947 assert(InvOp0 && InvOp1 &&
7948 "Mismatch between isFreeToInvert and getFreelyInverted");
7949 return new ICmpInst(I.getSwappedPredicate(), InvOp0, InvOp1);
7950 }
7951 }
7952
7953 Instruction *AddI = nullptr;
7954 if (match(V: &I, P: m_UAddWithOverflow(L: m_Value(V&: X), R: m_Value(V&: Y),
7955 S: m_Instruction(I&: AddI))) &&
7956 isa<IntegerType>(Val: X->getType())) {
7957 Value *Result;
7958 Constant *Overflow;
7959 // m_UAddWithOverflow can match patterns that do not include an explicit
7960 // "add" instruction, so check the opcode of the matched op.
7961 if (AddI->getOpcode() == Instruction::Add &&
7962 OptimizeOverflowCheck(BinaryOp: Instruction::Add, /*Signed*/ IsSigned: false, LHS: X, RHS: Y, OrigI&: *AddI,
7963 Result, Overflow)) {
7964 replaceInstUsesWith(I&: *AddI, V: Result);
7965 eraseInstFromFunction(I&: *AddI);
7966 return replaceInstUsesWith(I, V: Overflow);
7967 }
7968 }
7969
7970 // (zext X) * (zext Y) --> llvm.umul.with.overflow.
7971 if (match(V: Op0, P: m_NUWMul(L: m_ZExt(Op: m_Value(V&: X)), R: m_ZExt(Op: m_Value(V&: Y)))) &&
7972 match(V: Op1, P: m_APInt(Res&: C))) {
7973 if (Instruction *R = processUMulZExtIdiom(I, MulVal: Op0, OtherVal: C, IC&: *this))
7974 return R;
7975 }
7976
7977 // Signbit test folds
7978 // Fold (X u>> BitWidth - 1 Pred ZExt(i1)) --> X s< 0 Pred i1
7979 // Fold (X s>> BitWidth - 1 Pred SExt(i1)) --> X s< 0 Pred i1
7980 Instruction *ExtI;
7981 if ((I.isUnsigned() || I.isEquality()) &&
7982 match(V: Op1,
7983 P: m_CombineAnd(L: m_Instruction(I&: ExtI), R: m_ZExtOrSExt(Op: m_Value(V&: Y)))) &&
7984 Y->getType()->getScalarSizeInBits() == 1 &&
7985 (Op0->hasOneUse() || Op1->hasOneUse())) {
7986 unsigned OpWidth = Op0->getType()->getScalarSizeInBits();
7987 Instruction *ShiftI;
7988 if (match(V: Op0, P: m_CombineAnd(L: m_Instruction(I&: ShiftI),
7989 R: m_Shr(L: m_Value(V&: X), R: m_SpecificIntAllowPoison(
7990 V: OpWidth - 1))))) {
7991 unsigned ExtOpc = ExtI->getOpcode();
7992 unsigned ShiftOpc = ShiftI->getOpcode();
7993 if ((ExtOpc == Instruction::ZExt && ShiftOpc == Instruction::LShr) ||
7994 (ExtOpc == Instruction::SExt && ShiftOpc == Instruction::AShr)) {
7995 Value *SLTZero =
7996 Builder.CreateICmpSLT(LHS: X, RHS: Constant::getNullValue(Ty: X->getType()));
7997 Value *Cmp = Builder.CreateICmp(P: Pred, LHS: SLTZero, RHS: Y, Name: I.getName());
7998 return replaceInstUsesWith(I, V: Cmp);
7999 }
8000 }
8001 }
8002 }
8003
8004 if (Instruction *Res = foldICmpEquality(I))
8005 return Res;
8006
8007 if (Instruction *Res = foldICmpPow2Test(I, Builder))
8008 return Res;
8009
8010 if (Instruction *Res = foldICmpOfUAddOv(I))
8011 return Res;
8012
8013 if (Instruction *Res = foldICmpOfVectorReduce(I, DL, Builder))
8014 return Res;
8015
8016 // The 'cmpxchg' instruction returns an aggregate containing the old value and
8017 // an i1 which indicates whether or not we successfully did the swap.
8018 //
8019 // Replace comparisons between the old value and the expected value with the
8020 // indicator that 'cmpxchg' returns.
8021 //
8022 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
8023 // spuriously fail. In those cases, the old value may equal the expected
8024 // value but it is possible for the swap to not occur.
8025 if (I.getPredicate() == ICmpInst::ICMP_EQ)
8026 if (auto *EVI = dyn_cast<ExtractValueInst>(Val: Op0))
8027 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(Val: EVI->getAggregateOperand()))
8028 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
8029 !ACXI->isWeak())
8030 return ExtractValueInst::Create(Agg: ACXI, Idxs: 1);
8031
8032 if (Instruction *Res = foldICmpWithHighBitMask(Cmp&: I, Builder))
8033 return Res;
8034
8035 if (I.getType()->isVectorTy())
8036 if (Instruction *Res = foldVectorCmp(Cmp&: I, Builder))
8037 return Res;
8038
8039 if (Instruction *Res = foldICmpInvariantGroup(I))
8040 return Res;
8041
8042 if (Instruction *Res = foldReductionIdiom(I, Builder, DL))
8043 return Res;
8044
8045 {
8046 Value *A;
8047 const APInt *C1, *C2;
8048 ICmpInst::Predicate Pred = I.getPredicate();
8049 if (ICmpInst::isEquality(P: Pred)) {
8050 // sext(a) & c1 == c2 --> a & c3 == trunc(c2)
8051 // sext(a) & c1 != c2 --> a & c3 != trunc(c2)
8052 if (match(V: Op0, P: m_And(L: m_SExt(Op: m_Value(V&: A)), R: m_APInt(Res&: C1))) &&
8053 match(V: Op1, P: m_APInt(Res&: C2))) {
8054 Type *InputTy = A->getType();
8055 unsigned InputBitWidth = InputTy->getScalarSizeInBits();
8056 // c2 must be non-negative at the bitwidth of a.
8057 if (C2->getActiveBits() < InputBitWidth) {
8058 APInt TruncC1 = C1->trunc(width: InputBitWidth);
8059 // Check if there are 1s in C1 high bits of size InputBitWidth.
8060 if (C1->uge(RHS: APInt::getOneBitSet(numBits: C1->getBitWidth(), BitNo: InputBitWidth)))
8061 TruncC1.setBit(InputBitWidth - 1);
8062 Value *AndInst = Builder.CreateAnd(LHS: A, RHS: TruncC1);
8063 return new ICmpInst(
8064 Pred, AndInst,
8065 ConstantInt::get(Ty: InputTy, V: C2->trunc(width: InputBitWidth)));
8066 }
8067 }
8068 }
8069 }
8070
8071 return Changed ? &I : nullptr;
8072}
8073
8074/// Fold fcmp ([us]itofp x, cst) if possible.
8075Instruction *InstCombinerImpl::foldFCmpIntToFPConst(FCmpInst &I,
8076 Instruction *LHSI,
8077 Constant *RHSC) {
8078 const APFloat *RHS;
8079 if (!match(V: RHSC, P: m_APFloat(Res&: RHS)))
8080 return nullptr;
8081
8082 // Get the width of the mantissa. We don't want to hack on conversions that
8083 // might lose information from the integer, e.g. "i64 -> float"
8084 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
8085 if (MantissaWidth == -1)
8086 return nullptr; // Unknown.
8087
8088 Type *IntTy = LHSI->getOperand(i: 0)->getType();
8089 unsigned IntWidth = IntTy->getScalarSizeInBits();
8090 bool LHSUnsigned = isa<UIToFPInst>(Val: LHSI);
8091
8092 if (I.isEquality()) {
8093 FCmpInst::Predicate P = I.getPredicate();
8094 bool IsExact = false;
8095 APSInt RHSCvt(IntWidth, LHSUnsigned);
8096 RHS->convertToInteger(Result&: RHSCvt, RM: APFloat::rmNearestTiesToEven, IsExact: &IsExact);
8097
8098 // If the floating point constant isn't an integer value, we know if we will
8099 // ever compare equal / not equal to it.
8100 if (!IsExact) {
8101 // TODO: Can never be -0.0 and other non-representable values
8102 APFloat RHSRoundInt(*RHS);
8103 RHSRoundInt.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
8104 if (*RHS != RHSRoundInt) {
8105 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
8106 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8107
8108 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
8109 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8110 }
8111 }
8112
8113 // TODO: If the constant is exactly representable, is it always OK to do
8114 // equality compares as integer?
8115 }
8116
8117 // Check to see that the input is converted from an integer type that is small
8118 // enough that preserves all bits. TODO: check here for "known" sign bits.
8119 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
8120
8121 // Following test does NOT adjust IntWidth downwards for signed inputs,
8122 // because the most negative value still requires all the mantissa bits
8123 // to distinguish it from one less than that value.
8124 if ((int)IntWidth > MantissaWidth) {
8125 // Conversion would lose accuracy. Check if loss can impact comparison.
8126 int Exp = ilogb(Arg: *RHS);
8127 if (Exp == APFloat::IEK_Inf) {
8128 int MaxExponent = ilogb(Arg: APFloat::getLargest(Sem: RHS->getSemantics()));
8129 if (MaxExponent < (int)IntWidth - !LHSUnsigned)
8130 // Conversion could create infinity.
8131 return nullptr;
8132 } else {
8133 // Note that if RHS is zero or NaN, then Exp is negative
8134 // and first condition is trivially false.
8135 if (MantissaWidth <= Exp && Exp <= (int)IntWidth - !LHSUnsigned)
8136 // Conversion could affect comparison.
8137 return nullptr;
8138 }
8139 }
8140
8141 // Otherwise, we can potentially simplify the comparison. We know that it
8142 // will always come through as an integer value and we know the constant is
8143 // not a NAN (it would have been previously simplified).
8144 assert(!RHS->isNaN() && "NaN comparison not already folded!");
8145
8146 ICmpInst::Predicate Pred;
8147 switch (I.getPredicate()) {
8148 default:
8149 llvm_unreachable("Unexpected predicate!");
8150 case FCmpInst::FCMP_UEQ:
8151 case FCmpInst::FCMP_OEQ:
8152 Pred = ICmpInst::ICMP_EQ;
8153 break;
8154 case FCmpInst::FCMP_UGT:
8155 case FCmpInst::FCMP_OGT:
8156 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
8157 break;
8158 case FCmpInst::FCMP_UGE:
8159 case FCmpInst::FCMP_OGE:
8160 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
8161 break;
8162 case FCmpInst::FCMP_ULT:
8163 case FCmpInst::FCMP_OLT:
8164 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
8165 break;
8166 case FCmpInst::FCMP_ULE:
8167 case FCmpInst::FCMP_OLE:
8168 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
8169 break;
8170 case FCmpInst::FCMP_UNE:
8171 case FCmpInst::FCMP_ONE:
8172 Pred = ICmpInst::ICMP_NE;
8173 break;
8174 case FCmpInst::FCMP_ORD:
8175 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8176 case FCmpInst::FCMP_UNO:
8177 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8178 }
8179
8180 // Now we know that the APFloat is a normal number, zero or inf.
8181
8182 // See if the FP constant is too large for the integer. For example,
8183 // comparing an i8 to 300.0.
8184 if (!LHSUnsigned) {
8185 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
8186 // and large values.
8187 APFloat SMax(RHS->getSemantics());
8188 SMax.convertFromAPInt(Input: APInt::getSignedMaxValue(numBits: IntWidth), IsSigned: true,
8189 RM: APFloat::rmNearestTiesToEven);
8190 if (SMax < *RHS) { // smax < 13123.0
8191 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
8192 Pred == ICmpInst::ICMP_SLE)
8193 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8194 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8195 }
8196 } else {
8197 // If the RHS value is > UnsignedMax, fold the comparison. This handles
8198 // +INF and large values.
8199 APFloat UMax(RHS->getSemantics());
8200 UMax.convertFromAPInt(Input: APInt::getMaxValue(numBits: IntWidth), IsSigned: false,
8201 RM: APFloat::rmNearestTiesToEven);
8202 if (UMax < *RHS) { // umax < 13123.0
8203 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
8204 Pred == ICmpInst::ICMP_ULE)
8205 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8206 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8207 }
8208 }
8209
8210 if (!LHSUnsigned) {
8211 // See if the RHS value is < SignedMin.
8212 APFloat SMin(RHS->getSemantics());
8213 SMin.convertFromAPInt(Input: APInt::getSignedMinValue(numBits: IntWidth), IsSigned: true,
8214 RM: APFloat::rmNearestTiesToEven);
8215 if (SMin > *RHS) { // smin > 12312.0
8216 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
8217 Pred == ICmpInst::ICMP_SGE)
8218 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8219 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8220 }
8221 } else {
8222 // See if the RHS value is < UnsignedMin.
8223 APFloat UMin(RHS->getSemantics());
8224 UMin.convertFromAPInt(Input: APInt::getMinValue(numBits: IntWidth), IsSigned: false,
8225 RM: APFloat::rmNearestTiesToEven);
8226 if (UMin > *RHS) { // umin > 12312.0
8227 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
8228 Pred == ICmpInst::ICMP_UGE)
8229 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8230 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8231 }
8232 }
8233
8234 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
8235 // [0, UMAX], but it may still be fractional. Check whether this is the case
8236 // using the IsExact flag.
8237 // Don't do this for zero, because -0.0 is not fractional.
8238 APSInt RHSInt(IntWidth, LHSUnsigned);
8239 bool IsExact;
8240 RHS->convertToInteger(Result&: RHSInt, RM: APFloat::rmTowardZero, IsExact: &IsExact);
8241 if (!RHS->isZero()) {
8242 if (!IsExact) {
8243 // If we had a comparison against a fractional value, we have to adjust
8244 // the compare predicate and sometimes the value. RHSC is rounded towards
8245 // zero at this point.
8246 switch (Pred) {
8247 default:
8248 llvm_unreachable("Unexpected integer comparison!");
8249 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
8250 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8251 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
8252 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8253 case ICmpInst::ICMP_ULE:
8254 // (float)int <= 4.4 --> int <= 4
8255 // (float)int <= -4.4 --> false
8256 if (RHS->isNegative())
8257 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8258 break;
8259 case ICmpInst::ICMP_SLE:
8260 // (float)int <= 4.4 --> int <= 4
8261 // (float)int <= -4.4 --> int < -4
8262 if (RHS->isNegative())
8263 Pred = ICmpInst::ICMP_SLT;
8264 break;
8265 case ICmpInst::ICMP_ULT:
8266 // (float)int < -4.4 --> false
8267 // (float)int < 4.4 --> int <= 4
8268 if (RHS->isNegative())
8269 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8270 Pred = ICmpInst::ICMP_ULE;
8271 break;
8272 case ICmpInst::ICMP_SLT:
8273 // (float)int < -4.4 --> int < -4
8274 // (float)int < 4.4 --> int <= 4
8275 if (!RHS->isNegative())
8276 Pred = ICmpInst::ICMP_SLE;
8277 break;
8278 case ICmpInst::ICMP_UGT:
8279 // (float)int > 4.4 --> int > 4
8280 // (float)int > -4.4 --> true
8281 if (RHS->isNegative())
8282 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8283 break;
8284 case ICmpInst::ICMP_SGT:
8285 // (float)int > 4.4 --> int > 4
8286 // (float)int > -4.4 --> int >= -4
8287 if (RHS->isNegative())
8288 Pred = ICmpInst::ICMP_SGE;
8289 break;
8290 case ICmpInst::ICMP_UGE:
8291 // (float)int >= -4.4 --> true
8292 // (float)int >= 4.4 --> int > 4
8293 if (RHS->isNegative())
8294 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8295 Pred = ICmpInst::ICMP_UGT;
8296 break;
8297 case ICmpInst::ICMP_SGE:
8298 // (float)int >= -4.4 --> int >= -4
8299 // (float)int >= 4.4 --> int > 4
8300 if (!RHS->isNegative())
8301 Pred = ICmpInst::ICMP_SGT;
8302 break;
8303 }
8304 }
8305 }
8306
8307 // Lower this FP comparison into an appropriate integer version of the
8308 // comparison.
8309 return new ICmpInst(Pred, LHSI->getOperand(i: 0),
8310 ConstantInt::get(Ty: LHSI->getOperand(i: 0)->getType(), V: RHSInt));
8311}
8312
8313/// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
8314static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
8315 Constant *RHSC) {
8316 // When C is not 0.0 and infinities are not allowed:
8317 // (C / X) < 0.0 is a sign-bit test of X
8318 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
8319 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
8320 //
8321 // Proof:
8322 // Multiply (C / X) < 0.0 by X * X / C.
8323 // - X is non zero, if it is the flag 'ninf' is violated.
8324 // - C defines the sign of X * X * C. Thus it also defines whether to swap
8325 // the predicate. C is also non zero by definition.
8326 //
8327 // Thus X * X / C is non zero and the transformation is valid. [qed]
8328
8329 FCmpInst::Predicate Pred = I.getPredicate();
8330
8331 // Check that predicates are valid.
8332 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
8333 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
8334 return nullptr;
8335
8336 // Check that RHS operand is zero.
8337 if (!match(V: RHSC, P: m_AnyZeroFP()))
8338 return nullptr;
8339
8340 // Check fastmath flags ('ninf').
8341 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
8342 return nullptr;
8343
8344 // Check the properties of the dividend. It must not be zero to avoid a
8345 // division by zero (see Proof).
8346 const APFloat *C;
8347 if (!match(V: LHSI->getOperand(i: 0), P: m_APFloat(Res&: C)))
8348 return nullptr;
8349
8350 if (C->isZero())
8351 return nullptr;
8352
8353 // Get swapped predicate if necessary.
8354 if (C->isNegative())
8355 Pred = I.getSwappedPredicate();
8356
8357 return new FCmpInst(Pred, LHSI->getOperand(i: 1), RHSC, "", &I);
8358}
8359
8360// Transform 'fptrunc(x) cmp C' to 'x cmp ext(C)' if possible.
8361// Patterns include:
8362// fptrunc(x) < C --> x < ext(C)
8363// fptrunc(x) <= C --> x <= ext(C)
8364// fptrunc(x) > C --> x > ext(C)
8365// fptrunc(x) >= C --> x >= ext(C)
8366// where 'ext(C)' is the extension of 'C' to the type of 'x' with a small bias
8367// due to precision loss.
8368static Instruction *foldFCmpFpTrunc(FCmpInst &I, const Instruction &FPTrunc,
8369 const Constant &C) {
8370 FCmpInst::Predicate Pred = I.getPredicate();
8371 bool RoundDown = false;
8372
8373 if (Pred == FCmpInst::FCMP_OGE || Pred == FCmpInst::FCMP_UGE ||
8374 Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_ULT)
8375 RoundDown = true;
8376 else if (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT ||
8377 Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)
8378 RoundDown = false;
8379 else
8380 return nullptr;
8381
8382 const APFloat *CValue;
8383 if (!match(V: &C, P: m_APFloat(Res&: CValue)))
8384 return nullptr;
8385
8386 if (CValue->isNaN() || CValue->isInfinity())
8387 return nullptr;
8388
8389 auto ConvertFltSema = [](const APFloat &Src, const fltSemantics &Sema) {
8390 bool LosesInfo;
8391 APFloat Dest = Src;
8392 Dest.convert(ToSemantics: Sema, RM: APFloat::rmNearestTiesToEven, losesInfo: &LosesInfo);
8393 return Dest;
8394 };
8395
8396 auto NextValue = [](const APFloat &Value, bool RoundDown) {
8397 APFloat NextValue = Value;
8398 NextValue.next(nextDown: RoundDown);
8399 return NextValue;
8400 };
8401
8402 APFloat NextCValue = NextValue(*CValue, RoundDown);
8403
8404 Type *DestType = FPTrunc.getOperand(i: 0)->getType();
8405 const fltSemantics &DestFltSema =
8406 DestType->getScalarType()->getFltSemantics();
8407
8408 APFloat ExtCValue = ConvertFltSema(*CValue, DestFltSema);
8409 APFloat ExtNextCValue = ConvertFltSema(NextCValue, DestFltSema);
8410
8411 // When 'NextCValue' is infinity, use an imaged 'NextCValue' that equals
8412 // 'CValue + bias' to avoid the infinity after conversion. The bias is
8413 // estimated as 'CValue - PrevCValue', where 'PrevCValue' is the previous
8414 // value of 'CValue'.
8415 if (NextCValue.isInfinity()) {
8416 APFloat PrevCValue = NextValue(*CValue, !RoundDown);
8417 APFloat Bias = ConvertFltSema(*CValue - PrevCValue, DestFltSema);
8418
8419 ExtNextCValue = ExtCValue + Bias;
8420 }
8421
8422 APFloat ExtMidValue =
8423 scalbn(X: ExtCValue + ExtNextCValue, Exp: -1, RM: APFloat::rmNearestTiesToEven);
8424
8425 const fltSemantics &SrcFltSema =
8426 C.getType()->getScalarType()->getFltSemantics();
8427
8428 // 'MidValue' might be rounded to 'NextCValue'. Correct it here.
8429 APFloat MidValue = ConvertFltSema(ExtMidValue, SrcFltSema);
8430 if (MidValue != *CValue)
8431 ExtMidValue.next(nextDown: !RoundDown);
8432
8433 // Check whether 'ExtMidValue' is a valid result since the assumption on
8434 // imaged 'NextCValue' might not hold for new float types.
8435 // ppc_fp128 can't pass here when converting from max float because of
8436 // APFloat implementation.
8437 if (NextCValue.isInfinity()) {
8438 // ExtMidValue --- narrowed ---> Finite
8439 if (ConvertFltSema(ExtMidValue, SrcFltSema).isInfinity())
8440 return nullptr;
8441
8442 // NextExtMidValue --- narrowed ---> Infinity
8443 APFloat NextExtMidValue = NextValue(ExtMidValue, RoundDown);
8444 if (ConvertFltSema(NextExtMidValue, SrcFltSema).isFinite())
8445 return nullptr;
8446 }
8447
8448 return new FCmpInst(Pred, FPTrunc.getOperand(i: 0),
8449 ConstantFP::get(Ty: DestType, V: ExtMidValue), "", &I);
8450}
8451
8452/// Optimize fabs(X) compared with zero.
8453static Instruction *foldFabsWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
8454 Value *X;
8455 if (!match(V: I.getOperand(i_nocapture: 0), P: m_FAbs(Op0: m_Value(V&: X))))
8456 return nullptr;
8457
8458 const APFloat *C;
8459 if (!match(V: I.getOperand(i_nocapture: 1), P: m_APFloat(Res&: C)))
8460 return nullptr;
8461
8462 if (!C->isPosZero()) {
8463 if (!C->isSmallestNormalized())
8464 return nullptr;
8465
8466 const Function *F = I.getFunction();
8467 DenormalMode Mode = F->getDenormalMode(FPType: C->getSemantics());
8468 if (Mode.Input == DenormalMode::PreserveSign ||
8469 Mode.Input == DenormalMode::PositiveZero) {
8470
8471 auto replaceFCmp = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
8472 Constant *Zero = ConstantFP::getZero(Ty: X->getType());
8473 return new FCmpInst(P, X, Zero, "", I);
8474 };
8475
8476 switch (I.getPredicate()) {
8477 case FCmpInst::FCMP_OLT:
8478 // fcmp olt fabs(x), smallest_normalized_number -> fcmp oeq x, 0.0
8479 return replaceFCmp(&I, FCmpInst::FCMP_OEQ, X);
8480 case FCmpInst::FCMP_UGE:
8481 // fcmp uge fabs(x), smallest_normalized_number -> fcmp une x, 0.0
8482 return replaceFCmp(&I, FCmpInst::FCMP_UNE, X);
8483 case FCmpInst::FCMP_OGE:
8484 // fcmp oge fabs(x), smallest_normalized_number -> fcmp one x, 0.0
8485 return replaceFCmp(&I, FCmpInst::FCMP_ONE, X);
8486 case FCmpInst::FCMP_ULT:
8487 // fcmp ult fabs(x), smallest_normalized_number -> fcmp ueq x, 0.0
8488 return replaceFCmp(&I, FCmpInst::FCMP_UEQ, X);
8489 default:
8490 break;
8491 }
8492 }
8493
8494 return nullptr;
8495 }
8496
8497 auto replacePredAndOp0 = [&IC](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
8498 I->setPredicate(P);
8499 return IC.replaceOperand(I&: *I, OpNum: 0, V: X);
8500 };
8501
8502 switch (I.getPredicate()) {
8503 case FCmpInst::FCMP_UGE:
8504 case FCmpInst::FCMP_OLT:
8505 // fabs(X) >= 0.0 --> true
8506 // fabs(X) < 0.0 --> false
8507 llvm_unreachable("fcmp should have simplified");
8508
8509 case FCmpInst::FCMP_OGT:
8510 // fabs(X) > 0.0 --> X != 0.0
8511 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
8512
8513 case FCmpInst::FCMP_UGT:
8514 // fabs(X) u> 0.0 --> X u!= 0.0
8515 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
8516
8517 case FCmpInst::FCMP_OLE:
8518 // fabs(X) <= 0.0 --> X == 0.0
8519 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
8520
8521 case FCmpInst::FCMP_ULE:
8522 // fabs(X) u<= 0.0 --> X u== 0.0
8523 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
8524
8525 case FCmpInst::FCMP_OGE:
8526 // fabs(X) >= 0.0 --> !isnan(X)
8527 assert(!I.hasNoNaNs() && "fcmp should have simplified");
8528 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
8529
8530 case FCmpInst::FCMP_ULT:
8531 // fabs(X) u< 0.0 --> isnan(X)
8532 assert(!I.hasNoNaNs() && "fcmp should have simplified");
8533 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
8534
8535 case FCmpInst::FCMP_OEQ:
8536 case FCmpInst::FCMP_UEQ:
8537 case FCmpInst::FCMP_ONE:
8538 case FCmpInst::FCMP_UNE:
8539 case FCmpInst::FCMP_ORD:
8540 case FCmpInst::FCMP_UNO:
8541 // Look through the fabs() because it doesn't change anything but the sign.
8542 // fabs(X) == 0.0 --> X == 0.0,
8543 // fabs(X) != 0.0 --> X != 0.0
8544 // isnan(fabs(X)) --> isnan(X)
8545 // !isnan(fabs(X) --> !isnan(X)
8546 return replacePredAndOp0(&I, I.getPredicate(), X);
8547
8548 default:
8549 return nullptr;
8550 }
8551}
8552
8553/// Optimize sqrt(X) compared with zero.
8554static Instruction *foldSqrtWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
8555 Value *X;
8556 if (!match(V: I.getOperand(i_nocapture: 0), P: m_Sqrt(Op0: m_Value(V&: X))))
8557 return nullptr;
8558
8559 if (!match(V: I.getOperand(i_nocapture: 1), P: m_PosZeroFP()))
8560 return nullptr;
8561
8562 auto ReplacePredAndOp0 = [&](FCmpInst::Predicate P) {
8563 I.setPredicate(P);
8564 return IC.replaceOperand(I, OpNum: 0, V: X);
8565 };
8566
8567 // Clear ninf flag if sqrt doesn't have it.
8568 if (!cast<Instruction>(Val: I.getOperand(i_nocapture: 0))->hasNoInfs())
8569 I.setHasNoInfs(false);
8570
8571 switch (I.getPredicate()) {
8572 case FCmpInst::FCMP_OLT:
8573 case FCmpInst::FCMP_UGE:
8574 // sqrt(X) < 0.0 --> false
8575 // sqrt(X) u>= 0.0 --> true
8576 llvm_unreachable("fcmp should have simplified");
8577 case FCmpInst::FCMP_ULT:
8578 case FCmpInst::FCMP_ULE:
8579 case FCmpInst::FCMP_OGT:
8580 case FCmpInst::FCMP_OGE:
8581 case FCmpInst::FCMP_OEQ:
8582 case FCmpInst::FCMP_UNE:
8583 // sqrt(X) u< 0.0 --> X u< 0.0
8584 // sqrt(X) u<= 0.0 --> X u<= 0.0
8585 // sqrt(X) > 0.0 --> X > 0.0
8586 // sqrt(X) >= 0.0 --> X >= 0.0
8587 // sqrt(X) == 0.0 --> X == 0.0
8588 // sqrt(X) u!= 0.0 --> X u!= 0.0
8589 return IC.replaceOperand(I, OpNum: 0, V: X);
8590
8591 case FCmpInst::FCMP_OLE:
8592 // sqrt(X) <= 0.0 --> X == 0.0
8593 return ReplacePredAndOp0(FCmpInst::FCMP_OEQ);
8594 case FCmpInst::FCMP_UGT:
8595 // sqrt(X) u> 0.0 --> X u!= 0.0
8596 return ReplacePredAndOp0(FCmpInst::FCMP_UNE);
8597 case FCmpInst::FCMP_UEQ:
8598 // sqrt(X) u== 0.0 --> X u<= 0.0
8599 return ReplacePredAndOp0(FCmpInst::FCMP_ULE);
8600 case FCmpInst::FCMP_ONE:
8601 // sqrt(X) != 0.0 --> X > 0.0
8602 return ReplacePredAndOp0(FCmpInst::FCMP_OGT);
8603 case FCmpInst::FCMP_ORD:
8604 // !isnan(sqrt(X)) --> X >= 0.0
8605 return ReplacePredAndOp0(FCmpInst::FCMP_OGE);
8606 case FCmpInst::FCMP_UNO:
8607 // isnan(sqrt(X)) --> X u< 0.0
8608 return ReplacePredAndOp0(FCmpInst::FCMP_ULT);
8609 default:
8610 llvm_unreachable("Unexpected predicate!");
8611 }
8612}
8613
8614static Instruction *foldFCmpFNegCommonOp(FCmpInst &I) {
8615 CmpInst::Predicate Pred = I.getPredicate();
8616 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
8617
8618 // Canonicalize fneg as Op1.
8619 if (match(V: Op0, P: m_FNeg(X: m_Value())) && !match(V: Op1, P: m_FNeg(X: m_Value()))) {
8620 std::swap(a&: Op0, b&: Op1);
8621 Pred = I.getSwappedPredicate();
8622 }
8623
8624 if (!match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0))))
8625 return nullptr;
8626
8627 // Replace the negated operand with 0.0:
8628 // fcmp Pred Op0, -Op0 --> fcmp Pred Op0, 0.0
8629 Constant *Zero = ConstantFP::getZero(Ty: Op0->getType());
8630 return new FCmpInst(Pred, Op0, Zero, "", &I);
8631}
8632
8633static Instruction *foldFCmpFSubIntoFCmp(FCmpInst &I, Instruction *LHSI,
8634 Constant *RHSC, InstCombinerImpl &CI) {
8635 const CmpInst::Predicate Pred = I.getPredicate();
8636 Value *X = LHSI->getOperand(i: 0);
8637 Value *Y = LHSI->getOperand(i: 1);
8638 switch (Pred) {
8639 default:
8640 break;
8641 case FCmpInst::FCMP_UGT:
8642 case FCmpInst::FCMP_ULT:
8643 case FCmpInst::FCMP_UNE:
8644 case FCmpInst::FCMP_OEQ:
8645 case FCmpInst::FCMP_OGE:
8646 case FCmpInst::FCMP_OLE:
8647 // The optimization is not valid if X and Y are infinities of the same
8648 // sign, i.e. the inf - inf = nan case. If the fsub has the ninf or nnan
8649 // flag then we can assume we do not have that case. Otherwise we might be
8650 // able to prove that either X or Y is not infinity.
8651 if (!LHSI->hasNoNaNs() && !LHSI->hasNoInfs() &&
8652 !isKnownNeverInfinity(V: Y,
8653 SQ: CI.getSimplifyQuery().getWithInstruction(I: &I)) &&
8654 !isKnownNeverInfinity(V: X, SQ: CI.getSimplifyQuery().getWithInstruction(I: &I)))
8655 break;
8656
8657 [[fallthrough]];
8658 case FCmpInst::FCMP_OGT:
8659 case FCmpInst::FCMP_OLT:
8660 case FCmpInst::FCMP_ONE:
8661 case FCmpInst::FCMP_UEQ:
8662 case FCmpInst::FCMP_UGE:
8663 case FCmpInst::FCMP_ULE:
8664 // fcmp pred (x - y), 0 --> fcmp pred x, y
8665 if (match(V: RHSC, P: m_AnyZeroFP()) &&
8666 I.getFunction()->getDenormalMode(
8667 FPType: LHSI->getType()->getScalarType()->getFltSemantics()) ==
8668 DenormalMode::getIEEE()) {
8669 CI.replaceOperand(I, OpNum: 0, V: X);
8670 CI.replaceOperand(I, OpNum: 1, V: Y);
8671 I.setHasNoInfs(LHSI->hasNoInfs());
8672 if (LHSI->hasNoNaNs())
8673 I.setHasNoNaNs(true);
8674 return &I;
8675 }
8676 break;
8677 }
8678
8679 return nullptr;
8680}
8681
8682static Instruction *foldFCmpWithFloorAndCeil(FCmpInst &I,
8683 InstCombinerImpl &IC) {
8684 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
8685 Type *OpType = LHS->getType();
8686 CmpInst::Predicate Pred = I.getPredicate();
8687
8688 bool FloorX = match(V: LHS, P: m_Intrinsic<Intrinsic::floor>(Op0: m_Specific(V: RHS)));
8689 bool CeilX = match(V: LHS, P: m_Intrinsic<Intrinsic::ceil>(Op0: m_Specific(V: RHS)));
8690
8691 if (!FloorX && !CeilX) {
8692 if ((FloorX = match(V: RHS, P: m_Intrinsic<Intrinsic::floor>(Op0: m_Specific(V: LHS)))) ||
8693 (CeilX = match(V: RHS, P: m_Intrinsic<Intrinsic::ceil>(Op0: m_Specific(V: LHS))))) {
8694 std::swap(a&: LHS, b&: RHS);
8695 Pred = I.getSwappedPredicate();
8696 }
8697 }
8698
8699 switch (Pred) {
8700 case FCmpInst::FCMP_OLE:
8701 // fcmp ole floor(x), x => fcmp ord x, 0
8702 if (FloorX)
8703 return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(Ty: OpType),
8704 "", &I);
8705 break;
8706 case FCmpInst::FCMP_OGT:
8707 // fcmp ogt floor(x), x => false
8708 if (FloorX)
8709 return IC.replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8710 break;
8711 case FCmpInst::FCMP_OGE:
8712 // fcmp oge ceil(x), x => fcmp ord x, 0
8713 if (CeilX)
8714 return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(Ty: OpType),
8715 "", &I);
8716 break;
8717 case FCmpInst::FCMP_OLT:
8718 // fcmp olt ceil(x), x => false
8719 if (CeilX)
8720 return IC.replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8721 break;
8722 case FCmpInst::FCMP_ULE:
8723 // fcmp ule floor(x), x => true
8724 if (FloorX)
8725 return IC.replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8726 break;
8727 case FCmpInst::FCMP_UGT:
8728 // fcmp ugt floor(x), x => fcmp uno x, 0
8729 if (FloorX)
8730 return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(Ty: OpType),
8731 "", &I);
8732 break;
8733 case FCmpInst::FCMP_UGE:
8734 // fcmp uge ceil(x), x => true
8735 if (CeilX)
8736 return IC.replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8737 break;
8738 case FCmpInst::FCMP_ULT:
8739 // fcmp ult ceil(x), x => fcmp uno x, 0
8740 if (CeilX)
8741 return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(Ty: OpType),
8742 "", &I);
8743 break;
8744 default:
8745 break;
8746 }
8747
8748 return nullptr;
8749}
8750
8751/// Returns true if a select that implements a min/max is redundant and
8752/// select result can be replaced with its non-constant operand, e.g.,
8753/// select ( (si/ui-to-fp A) <= C ), C, (si/ui-to-fp A)
8754/// where C is the FP constant equal to the minimum integer value
8755/// representable by A.
8756static bool isMinMaxCmpSelectEliminable(SelectPatternFlavor Flavor, Value *A,
8757 Value *B) {
8758 const APFloat *APF;
8759 if (!match(V: B, P: m_APFloat(Res&: APF)))
8760 return false;
8761
8762 auto *I = dyn_cast<Instruction>(Val: A);
8763 if (!I || !(I->getOpcode() == Instruction::SIToFP ||
8764 I->getOpcode() == Instruction::UIToFP))
8765 return false;
8766
8767 bool IsUnsigned = I->getOpcode() == Instruction::UIToFP;
8768 unsigned BitWidth = I->getOperand(i: 0)->getType()->getScalarSizeInBits();
8769 APSInt IntBoundary = (Flavor == SPF_FMAXNUM)
8770 ? APSInt::getMinValue(numBits: BitWidth, Unsigned: IsUnsigned)
8771 : APSInt::getMaxValue(numBits: BitWidth, Unsigned: IsUnsigned);
8772 APSInt ConvertedInt(BitWidth, IsUnsigned);
8773 bool IsExact;
8774 APFloat::opStatus Status =
8775 APF->convertToInteger(Result&: ConvertedInt, RM: APFloat::rmTowardZero, IsExact: &IsExact);
8776 return Status == APFloat::opOK && IsExact && ConvertedInt == IntBoundary;
8777}
8778
8779Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
8780 bool Changed = false;
8781
8782 /// Orders the operands of the compare so that they are listed from most
8783 /// complex to least complex. This puts constants before unary operators,
8784 /// before binary operators.
8785 if (getComplexity(V: I.getOperand(i_nocapture: 0)) < getComplexity(V: I.getOperand(i_nocapture: 1))) {
8786 I.swapOperands();
8787 Changed = true;
8788 }
8789
8790 const CmpInst::Predicate Pred = I.getPredicate();
8791 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
8792 if (Value *V = simplifyFCmpInst(Predicate: Pred, LHS: Op0, RHS: Op1, FMF: I.getFastMathFlags(),
8793 Q: SQ.getWithInstruction(I: &I)))
8794 return replaceInstUsesWith(I, V);
8795
8796 // Simplify 'fcmp pred X, X'
8797 Type *OpType = Op0->getType();
8798 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
8799 if (Op0 == Op1) {
8800 switch (Pred) {
8801 default:
8802 break;
8803 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
8804 case FCmpInst::FCMP_ULT: // True if unordered or less than
8805 case FCmpInst::FCMP_UGT: // True if unordered or greater than
8806 case FCmpInst::FCMP_UNE: // True if unordered or not equal
8807 // Canonicalize these to be 'fcmp uno %X, 0.0'.
8808 I.setPredicate(FCmpInst::FCMP_UNO);
8809 I.setOperand(i_nocapture: 1, Val_nocapture: Constant::getNullValue(Ty: OpType));
8810 return &I;
8811
8812 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
8813 case FCmpInst::FCMP_OEQ: // True if ordered and equal
8814 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
8815 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
8816 // Canonicalize these to be 'fcmp ord %X, 0.0'.
8817 I.setPredicate(FCmpInst::FCMP_ORD);
8818 I.setOperand(i_nocapture: 1, Val_nocapture: Constant::getNullValue(Ty: OpType));
8819 return &I;
8820 }
8821 }
8822
8823 if (I.isCommutative()) {
8824 if (auto Pair = matchSymmetricPair(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1))) {
8825 replaceOperand(I, OpNum: 0, V: Pair->first);
8826 replaceOperand(I, OpNum: 1, V: Pair->second);
8827 return &I;
8828 }
8829 }
8830
8831 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
8832 // then canonicalize the operand to 0.0.
8833 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
8834 if (!match(V: Op0, P: m_PosZeroFP()) &&
8835 isKnownNeverNaN(V: Op0, SQ: getSimplifyQuery().getWithInstruction(I: &I)))
8836 return replaceOperand(I, OpNum: 0, V: ConstantFP::getZero(Ty: OpType));
8837
8838 if (!match(V: Op1, P: m_PosZeroFP()) &&
8839 isKnownNeverNaN(V: Op1, SQ: getSimplifyQuery().getWithInstruction(I: &I)))
8840 return replaceOperand(I, OpNum: 1, V: ConstantFP::getZero(Ty: OpType));
8841 }
8842
8843 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
8844 Value *X, *Y;
8845 if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_FNeg(X: m_Value(V&: Y))))
8846 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
8847
8848 if (Instruction *R = foldFCmpFNegCommonOp(I))
8849 return R;
8850
8851 // Test if the FCmpInst instruction is used exclusively by a select as
8852 // part of a minimum or maximum operation. If so, refrain from doing
8853 // any other folding. This helps out other analyses which understand
8854 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
8855 // and CodeGen. And in this case, at least one of the comparison
8856 // operands has at least one user besides the compare (the select),
8857 // which would often largely negate the benefit of folding anyway.
8858 if (I.hasOneUse())
8859 if (SelectInst *SI = dyn_cast<SelectInst>(Val: I.user_back())) {
8860 Value *A, *B;
8861 SelectPatternResult SPR = matchSelectPattern(V: SI, LHS&: A, RHS&: B);
8862 bool IsRedundantMinMaxClamp =
8863 (SPR.Flavor == SPF_FMAXNUM || SPR.Flavor == SPF_FMINNUM) &&
8864 isMinMaxCmpSelectEliminable(Flavor: SPR.Flavor, A, B);
8865 if (SPR.Flavor != SPF_UNKNOWN && !IsRedundantMinMaxClamp)
8866 return nullptr;
8867 }
8868
8869 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
8870 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
8871 if (match(V: Op1, P: m_AnyZeroFP()) && !match(V: Op1, P: m_PosZeroFP()))
8872 return replaceOperand(I, OpNum: 1, V: ConstantFP::getZero(Ty: OpType));
8873
8874 // Canonicalize:
8875 // fcmp olt X, +inf -> fcmp one X, +inf
8876 // fcmp ole X, +inf -> fcmp ord X, 0
8877 // fcmp ogt X, +inf -> false
8878 // fcmp oge X, +inf -> fcmp oeq X, +inf
8879 // fcmp ult X, +inf -> fcmp une X, +inf
8880 // fcmp ule X, +inf -> true
8881 // fcmp ugt X, +inf -> fcmp uno X, 0
8882 // fcmp uge X, +inf -> fcmp ueq X, +inf
8883 // fcmp olt X, -inf -> false
8884 // fcmp ole X, -inf -> fcmp oeq X, -inf
8885 // fcmp ogt X, -inf -> fcmp one X, -inf
8886 // fcmp oge X, -inf -> fcmp ord X, 0
8887 // fcmp ult X, -inf -> fcmp uno X, 0
8888 // fcmp ule X, -inf -> fcmp ueq X, -inf
8889 // fcmp ugt X, -inf -> fcmp une X, -inf
8890 // fcmp uge X, -inf -> true
8891 const APFloat *C;
8892 if (match(V: Op1, P: m_APFloat(Res&: C)) && C->isInfinity()) {
8893 switch (C->isNegative() ? FCmpInst::getSwappedPredicate(pred: Pred) : Pred) {
8894 default:
8895 break;
8896 case FCmpInst::FCMP_ORD:
8897 case FCmpInst::FCMP_UNO:
8898 case FCmpInst::FCMP_TRUE:
8899 case FCmpInst::FCMP_FALSE:
8900 case FCmpInst::FCMP_OGT:
8901 case FCmpInst::FCMP_ULE:
8902 llvm_unreachable("Should be simplified by InstSimplify");
8903 case FCmpInst::FCMP_OLT:
8904 return new FCmpInst(FCmpInst::FCMP_ONE, Op0, Op1, "", &I);
8905 case FCmpInst::FCMP_OLE:
8906 return new FCmpInst(FCmpInst::FCMP_ORD, Op0, ConstantFP::getZero(Ty: OpType),
8907 "", &I);
8908 case FCmpInst::FCMP_OGE:
8909 return new FCmpInst(FCmpInst::FCMP_OEQ, Op0, Op1, "", &I);
8910 case FCmpInst::FCMP_ULT:
8911 return new FCmpInst(FCmpInst::FCMP_UNE, Op0, Op1, "", &I);
8912 case FCmpInst::FCMP_UGT:
8913 return new FCmpInst(FCmpInst::FCMP_UNO, Op0, ConstantFP::getZero(Ty: OpType),
8914 "", &I);
8915 case FCmpInst::FCMP_UGE:
8916 return new FCmpInst(FCmpInst::FCMP_UEQ, Op0, Op1, "", &I);
8917 }
8918 }
8919
8920 // Ignore signbit of bitcasted int when comparing equality to FP 0.0:
8921 // fcmp oeq/une (bitcast X), 0.0 --> (and X, SignMaskC) ==/!= 0
8922 if (match(V: Op1, P: m_PosZeroFP()) &&
8923 match(V: Op0, P: m_OneUse(SubPattern: m_ElementWiseBitCast(Op: m_Value(V&: X)))) &&
8924 !F.getDenormalMode(FPType: Op1->getType()->getScalarType()->getFltSemantics())
8925 .inputsMayBeZero()) {
8926 ICmpInst::Predicate IntPred = ICmpInst::BAD_ICMP_PREDICATE;
8927 if (Pred == FCmpInst::FCMP_OEQ)
8928 IntPred = ICmpInst::ICMP_EQ;
8929 else if (Pred == FCmpInst::FCMP_UNE)
8930 IntPred = ICmpInst::ICMP_NE;
8931
8932 if (IntPred != ICmpInst::BAD_ICMP_PREDICATE) {
8933 Type *IntTy = X->getType();
8934 const APInt &SignMask = ~APInt::getSignMask(BitWidth: IntTy->getScalarSizeInBits());
8935 Value *MaskX = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty: IntTy, V: SignMask));
8936 return new ICmpInst(IntPred, MaskX, ConstantInt::getNullValue(Ty: IntTy));
8937 }
8938 }
8939
8940 // Handle fcmp with instruction LHS and constant RHS.
8941 Instruction *LHSI;
8942 Constant *RHSC;
8943 if (match(V: Op0, P: m_Instruction(I&: LHSI)) && match(V: Op1, P: m_Constant(C&: RHSC))) {
8944 switch (LHSI->getOpcode()) {
8945 case Instruction::Select:
8946 // fcmp eq (cond ? x : -x), 0 --> fcmp eq x, 0
8947 if (FCmpInst::isEquality(Pred) && match(V: RHSC, P: m_AnyZeroFP()) &&
8948 match(V: LHSI, P: m_c_Select(L: m_FNeg(X: m_Value(V&: X)), R: m_Deferred(V: X))))
8949 return replaceOperand(I, OpNum: 0, V: X);
8950 if (Instruction *NV = FoldOpIntoSelect(Op&: I, SI: cast<SelectInst>(Val: LHSI)))
8951 return NV;
8952 break;
8953 case Instruction::FSub:
8954 if (LHSI->hasOneUse())
8955 if (Instruction *NV = foldFCmpFSubIntoFCmp(I, LHSI, RHSC, CI&: *this))
8956 return NV;
8957 break;
8958 case Instruction::PHI:
8959 if (Instruction *NV = foldOpIntoPhi(I, PN: cast<PHINode>(Val: LHSI)))
8960 return NV;
8961 break;
8962 case Instruction::SIToFP:
8963 case Instruction::UIToFP:
8964 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
8965 return NV;
8966 break;
8967 case Instruction::FDiv:
8968 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
8969 return NV;
8970 break;
8971 case Instruction::Load:
8972 if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: LHSI->getOperand(i: 0)))
8973 if (Instruction *Res =
8974 foldCmpLoadFromIndexedGlobal(LI: cast<LoadInst>(Val: LHSI), GEP, ICI&: I))
8975 return Res;
8976 break;
8977 case Instruction::FPTrunc:
8978 if (Instruction *NV = foldFCmpFpTrunc(I, FPTrunc: *LHSI, C: *RHSC))
8979 return NV;
8980 break;
8981 }
8982 }
8983
8984 if (Instruction *R = foldFabsWithFcmpZero(I, IC&: *this))
8985 return R;
8986
8987 if (Instruction *R = foldSqrtWithFcmpZero(I, IC&: *this))
8988 return R;
8989
8990 if (Instruction *R = foldFCmpWithFloorAndCeil(I, IC&: *this))
8991 return R;
8992
8993 if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X)))) {
8994 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
8995 Constant *C;
8996 if (match(V: Op1, P: m_Constant(C)))
8997 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL))
8998 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
8999 }
9000
9001 // fcmp (fadd X, 0.0), Y --> fcmp X, Y
9002 if (match(V: Op0, P: m_FAdd(L: m_Value(V&: X), R: m_AnyZeroFP())))
9003 return new FCmpInst(Pred, X, Op1, "", &I);
9004
9005 // fcmp X, (fadd Y, 0.0) --> fcmp X, Y
9006 if (match(V: Op1, P: m_FAdd(L: m_Value(V&: Y), R: m_AnyZeroFP())))
9007 return new FCmpInst(Pred, Op0, Y, "", &I);
9008
9009 if (match(V: Op0, P: m_FPExt(Op: m_Value(V&: X)))) {
9010 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
9011 if (match(V: Op1, P: m_FPExt(Op: m_Value(V&: Y))) && X->getType() == Y->getType())
9012 return new FCmpInst(Pred, X, Y, "", &I);
9013
9014 const APFloat *C;
9015 if (match(V: Op1, P: m_APFloat(Res&: C))) {
9016 const fltSemantics &FPSem =
9017 X->getType()->getScalarType()->getFltSemantics();
9018 bool Lossy;
9019 APFloat TruncC = *C;
9020 TruncC.convert(ToSemantics: FPSem, RM: APFloat::rmNearestTiesToEven, losesInfo: &Lossy);
9021
9022 if (Lossy) {
9023 // X can't possibly equal the higher-precision constant, so reduce any
9024 // equality comparison.
9025 // TODO: Other predicates can be handled via getFCmpCode().
9026 switch (Pred) {
9027 case FCmpInst::FCMP_OEQ:
9028 // X is ordered and equal to an impossible constant --> false
9029 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
9030 case FCmpInst::FCMP_ONE:
9031 // X is ordered and not equal to an impossible constant --> ordered
9032 return new FCmpInst(FCmpInst::FCMP_ORD, X,
9033 ConstantFP::getZero(Ty: X->getType()));
9034 case FCmpInst::FCMP_UEQ:
9035 // X is unordered or equal to an impossible constant --> unordered
9036 return new FCmpInst(FCmpInst::FCMP_UNO, X,
9037 ConstantFP::getZero(Ty: X->getType()));
9038 case FCmpInst::FCMP_UNE:
9039 // X is unordered or not equal to an impossible constant --> true
9040 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
9041 default:
9042 break;
9043 }
9044 }
9045
9046 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
9047 // Avoid lossy conversions and denormals.
9048 // Zero is a special case that's OK to convert.
9049 APFloat Fabs = TruncC;
9050 Fabs.clearSign();
9051 if (!Lossy &&
9052 (Fabs.isZero() || !(Fabs < APFloat::getSmallestNormalized(Sem: FPSem)))) {
9053 Constant *NewC = ConstantFP::get(Ty: X->getType(), V: TruncC);
9054 return new FCmpInst(Pred, X, NewC, "", &I);
9055 }
9056 }
9057 }
9058
9059 // Convert a sign-bit test of an FP value into a cast and integer compare.
9060 // TODO: Simplify if the copysign constant is 0.0 or NaN.
9061 // TODO: Handle non-zero compare constants.
9062 // TODO: Handle other predicates.
9063 if (match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::copysign>(Op0: m_APFloat(Res&: C),
9064 Op1: m_Value(V&: X)))) &&
9065 match(V: Op1, P: m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) {
9066 Type *IntType = Builder.getIntNTy(N: X->getType()->getScalarSizeInBits());
9067 if (auto *VecTy = dyn_cast<VectorType>(Val: OpType))
9068 IntType = VectorType::get(ElementType: IntType, EC: VecTy->getElementCount());
9069
9070 // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0
9071 if (Pred == FCmpInst::FCMP_OLT) {
9072 Value *IntX = Builder.CreateBitCast(V: X, DestTy: IntType);
9073 return new ICmpInst(ICmpInst::ICMP_SLT, IntX,
9074 ConstantInt::getNullValue(Ty: IntType));
9075 }
9076 }
9077
9078 {
9079 Value *CanonLHS = nullptr;
9080 match(V: Op0, P: m_Intrinsic<Intrinsic::canonicalize>(Op0: m_Value(V&: CanonLHS)));
9081 // (canonicalize(x) == x) => (x == x)
9082 if (CanonLHS == Op1)
9083 return new FCmpInst(Pred, Op1, Op1, "", &I);
9084
9085 Value *CanonRHS = nullptr;
9086 match(V: Op1, P: m_Intrinsic<Intrinsic::canonicalize>(Op0: m_Value(V&: CanonRHS)));
9087 // (x == canonicalize(x)) => (x == x)
9088 if (CanonRHS == Op0)
9089 return new FCmpInst(Pred, Op0, Op0, "", &I);
9090
9091 // (canonicalize(x) == canonicalize(y)) => (x == y)
9092 if (CanonLHS && CanonRHS)
9093 return new FCmpInst(Pred, CanonLHS, CanonRHS, "", &I);
9094 }
9095
9096 if (I.getType()->isVectorTy())
9097 if (Instruction *Res = foldVectorCmp(Cmp&: I, Builder))
9098 return Res;
9099
9100 return Changed ? &I : nullptr;
9101}
9102