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<BranchInst>(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(predicate: 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(predicate: 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 (BranchInst *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 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1790 // TODO: We canonicalize to the longer form for scalars because we have
1791 // better analysis/folds for icmp, and codegen may be better with icmp.
1792 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isZero() &&
1793 match(V: And->getOperand(i_nocapture: 1), P: m_One()))
1794 return new TruncInst(And->getOperand(i_nocapture: 0), Cmp.getType());
1795
1796 const APInt *C2;
1797 Value *X;
1798 if (!match(V: And, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: C2))))
1799 return nullptr;
1800
1801 // (and X, highmask) s> [0, ~highmask] --> X s> ~highmask
1802 if (Cmp.getPredicate() == ICmpInst::ICMP_SGT && C1.ule(RHS: ~*C2) &&
1803 C2->isNegatedPowerOf2())
1804 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1805 ConstantInt::get(Ty: X->getType(), V: ~*C2));
1806 // (and X, highmask) s< [1, -highmask] --> X s< -highmask
1807 if (Cmp.getPredicate() == ICmpInst::ICMP_SLT && !C1.isSignMask() &&
1808 (C1 - 1).ule(RHS: ~*C2) && C2->isNegatedPowerOf2() && !C2->isSignMask())
1809 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1810 ConstantInt::get(Ty: X->getType(), V: -*C2));
1811
1812 // Don't perform the following transforms if the AND has multiple uses
1813 if (!And->hasOneUse())
1814 return nullptr;
1815
1816 if (Cmp.isEquality() && C1.isZero()) {
1817 // Restrict this fold to single-use 'and' (PR10267).
1818 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1819 if (C2->isSignMask()) {
1820 Constant *Zero = Constant::getNullValue(Ty: X->getType());
1821 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1822 return new ICmpInst(NewPred, X, Zero);
1823 }
1824
1825 APInt NewC2 = *C2;
1826 KnownBits Know = computeKnownBits(V: And->getOperand(i_nocapture: 0), CxtI: And);
1827 // Set high zeros of C2 to allow matching negated power-of-2.
1828 NewC2 = *C2 | APInt::getHighBitsSet(numBits: C2->getBitWidth(),
1829 hiBitsSet: Know.countMinLeadingZeros());
1830
1831 // Restrict this fold only for single-use 'and' (PR10267).
1832 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1833 if (NewC2.isNegatedPowerOf2()) {
1834 Constant *NegBOC = ConstantInt::get(Ty: And->getType(), V: -NewC2);
1835 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1836 return new ICmpInst(NewPred, X, NegBOC);
1837 }
1838 }
1839
1840 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1841 // the input width without changing the value produced, eliminate the cast:
1842 //
1843 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1844 //
1845 // We can do this transformation if the constants do not have their sign bits
1846 // set or if it is an equality comparison. Extending a relational comparison
1847 // when we're checking the sign bit would not work.
1848 Value *W;
1849 if (match(V: And->getOperand(i_nocapture: 0), P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: W)))) &&
1850 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1851 // TODO: Is this a good transform for vectors? Wider types may reduce
1852 // throughput. Should this transform be limited (even for scalars) by using
1853 // shouldChangeType()?
1854 if (!Cmp.getType()->isVectorTy()) {
1855 Type *WideType = W->getType();
1856 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1857 Constant *ZextC1 = ConstantInt::get(Ty: WideType, V: C1.zext(width: WideScalarBits));
1858 Constant *ZextC2 = ConstantInt::get(Ty: WideType, V: C2->zext(width: WideScalarBits));
1859 Value *NewAnd = Builder.CreateAnd(LHS: W, RHS: ZextC2, Name: And->getName());
1860 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1861 }
1862 }
1863
1864 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2: *C2))
1865 return I;
1866
1867 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1868 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1869 //
1870 // iff pred isn't signed
1871 if (!Cmp.isSigned() && C1.isZero() && And->getOperand(i_nocapture: 0)->hasOneUse() &&
1872 match(V: And->getOperand(i_nocapture: 1), P: m_One())) {
1873 Constant *One = cast<Constant>(Val: And->getOperand(i_nocapture: 1));
1874 Value *Or = And->getOperand(i_nocapture: 0);
1875 Value *A, *B, *LShr;
1876 if (match(V: Or, P: m_Or(L: m_Value(V&: LShr), R: m_Value(V&: A))) &&
1877 match(V: LShr, P: m_LShr(L: m_Specific(V: A), R: m_Value(V&: B)))) {
1878 unsigned UsesRemoved = 0;
1879 if (And->hasOneUse())
1880 ++UsesRemoved;
1881 if (Or->hasOneUse())
1882 ++UsesRemoved;
1883 if (LShr->hasOneUse())
1884 ++UsesRemoved;
1885
1886 // Compute A & ((1 << B) | 1)
1887 unsigned RequireUsesRemoved = match(V: B, P: m_ImmConstant()) ? 1 : 3;
1888 if (UsesRemoved >= RequireUsesRemoved) {
1889 Value *NewOr =
1890 Builder.CreateOr(LHS: Builder.CreateShl(LHS: One, RHS: B, Name: LShr->getName(),
1891 /*HasNUW=*/true),
1892 RHS: One, Name: Or->getName());
1893 Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: NewOr, Name: And->getName());
1894 return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(i_nocapture: 1));
1895 }
1896 }
1897 }
1898
1899 // (icmp eq (and (bitcast X to int), ExponentMask), ExponentMask) -->
1900 // llvm.is.fpclass(X, fcInf|fcNan)
1901 // (icmp ne (and (bitcast X to int), ExponentMask), ExponentMask) -->
1902 // llvm.is.fpclass(X, ~(fcInf|fcNan))
1903 // (icmp eq (and (bitcast X to int), ExponentMask), 0) -->
1904 // llvm.is.fpclass(X, fcSubnormal|fcZero)
1905 // (icmp ne (and (bitcast X to int), ExponentMask), 0) -->
1906 // llvm.is.fpclass(X, ~(fcSubnormal|fcZero))
1907 Value *V;
1908 if (!Cmp.getParent()->getParent()->hasFnAttribute(
1909 Kind: Attribute::NoImplicitFloat) &&
1910 Cmp.isEquality() &&
1911 match(V: X, P: m_OneUse(SubPattern: m_ElementWiseBitCast(Op: m_Value(V))))) {
1912 Type *FPType = V->getType()->getScalarType();
1913 if (FPType->isIEEELikeFPTy() && (C1.isZero() || C1 == *C2)) {
1914 APInt ExponentMask =
1915 APFloat::getInf(Sem: FPType->getFltSemantics()).bitcastToAPInt();
1916 if (*C2 == ExponentMask) {
1917 unsigned Mask = C1.isZero()
1918 ? FPClassTest::fcZero | FPClassTest::fcSubnormal
1919 : FPClassTest::fcNan | FPClassTest::fcInf;
1920 if (isICMP_NE)
1921 Mask = ~Mask & fcAllFlags;
1922 return replaceInstUsesWith(I&: Cmp, V: Builder.createIsFPClass(FPNum: V, Test: Mask));
1923 }
1924 }
1925 }
1926
1927 return nullptr;
1928}
1929
1930/// Fold icmp (and X, Y), C.
1931Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1932 BinaryOperator *And,
1933 const APInt &C) {
1934 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C1: C))
1935 return I;
1936
1937 const ICmpInst::Predicate Pred = Cmp.getPredicate();
1938 bool TrueIfNeg;
1939 if (isSignBitCheck(Pred, RHS: C, TrueIfSigned&: TrueIfNeg)) {
1940 // ((X - 1) & ~X) < 0 --> X == 0
1941 // ((X - 1) & ~X) >= 0 --> X != 0
1942 Value *X;
1943 if (match(V: And->getOperand(i_nocapture: 0), P: m_Add(L: m_Value(V&: X), R: m_AllOnes())) &&
1944 match(V: And->getOperand(i_nocapture: 1), P: m_Not(V: m_Specific(V: X)))) {
1945 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1946 return new ICmpInst(NewPred, X, ConstantInt::getNullValue(Ty: X->getType()));
1947 }
1948 // (X & -X) < 0 --> X == MinSignedC
1949 // (X & -X) > -1 --> X != MinSignedC
1950 if (match(V: And, P: m_c_And(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) {
1951 Constant *MinSignedC = ConstantInt::get(
1952 Ty: X->getType(),
1953 V: APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits()));
1954 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1955 return new ICmpInst(NewPred, X, MinSignedC);
1956 }
1957 }
1958
1959 // TODO: These all require that Y is constant too, so refactor with the above.
1960
1961 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1962 Value *X = And->getOperand(i_nocapture: 0);
1963 Value *Y = And->getOperand(i_nocapture: 1);
1964 if (auto *C2 = dyn_cast<ConstantInt>(Val: Y))
1965 if (auto *LI = dyn_cast<LoadInst>(Val: X))
1966 if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: LI->getOperand(i_nocapture: 0)))
1967 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(LI, GEP, ICI&: Cmp, AndCst: C2))
1968 return Res;
1969
1970 if (!Cmp.isEquality())
1971 return nullptr;
1972
1973 // X & -C == -C -> X > u ~C
1974 // X & -C != -C -> X <= u ~C
1975 // iff C is a power of 2
1976 if (Cmp.getOperand(i_nocapture: 1) == Y && C.isNegatedPowerOf2()) {
1977 auto NewPred =
1978 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1979 return new ICmpInst(NewPred, X, SubOne(C: cast<Constant>(Val: Cmp.getOperand(i_nocapture: 1))));
1980 }
1981
1982 // ((zext i1 X) & Y) == 0 --> !((trunc Y) & X)
1983 // ((zext i1 X) & Y) != 0 --> ((trunc Y) & X)
1984 // ((zext i1 X) & Y) == 1 --> ((trunc Y) & X)
1985 // ((zext i1 X) & Y) != 1 --> !((trunc Y) & X)
1986 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)))) &&
1987 X->getType()->isIntOrIntVectorTy(BitWidth: 1) && (C.isZero() || C.isOne())) {
1988 Value *TruncY = Builder.CreateTrunc(V: Y, DestTy: X->getType());
1989 if (C.isZero() ^ (Pred == CmpInst::ICMP_NE)) {
1990 Value *And = Builder.CreateAnd(LHS: TruncY, RHS: X);
1991 return BinaryOperator::CreateNot(Op: And);
1992 }
1993 return BinaryOperator::CreateAnd(V1: TruncY, V2: X);
1994 }
1995
1996 // (icmp eq/ne (and (shl -1, X), Y), 0)
1997 // -> (icmp eq/ne (lshr Y, X), 0)
1998 // We could technically handle any C == 0 or (C < 0 && isOdd(C)) but it seems
1999 // highly unlikely the non-zero case will ever show up in code.
2000 if (C.isZero() &&
2001 match(V: And, P: m_OneUse(SubPattern: m_c_And(L: m_OneUse(SubPattern: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))),
2002 R: m_Value(V&: Y))))) {
2003 Value *LShr = Builder.CreateLShr(LHS: Y, RHS: X);
2004 return new ICmpInst(Pred, LShr, Constant::getNullValue(Ty: LShr->getType()));
2005 }
2006
2007 // (icmp eq/ne (and (add A, Addend), Msk), C)
2008 // -> (icmp eq/ne (and A, Msk), (and (sub C, Addend), Msk))
2009 {
2010 Value *A;
2011 const APInt *Addend, *Msk;
2012 if (match(V: And, P: m_And(L: m_OneUse(SubPattern: m_Add(L: m_Value(V&: A), R: m_APInt(Res&: Addend))),
2013 R: m_LowBitMask(V&: Msk))) &&
2014 C.ule(RHS: *Msk)) {
2015 APInt NewComperand = (C - *Addend) & *Msk;
2016 Value *MaskA = Builder.CreateAnd(LHS: A, RHS: ConstantInt::get(Ty: A->getType(), V: *Msk));
2017 return new ICmpInst(Pred, MaskA,
2018 ConstantInt::get(Ty: MaskA->getType(), V: NewComperand));
2019 }
2020 }
2021
2022 return nullptr;
2023}
2024
2025/// Fold icmp eq/ne (or (xor/sub (X1, X2), xor/sub (X3, X4))), 0.
2026static Value *foldICmpOrXorSubChain(ICmpInst &Cmp, BinaryOperator *Or,
2027 InstCombiner::BuilderTy &Builder) {
2028 // Are we using xors or subs to bitwise check for a pair or pairs of
2029 // (in)equalities? Convert to a shorter form that has more potential to be
2030 // folded even further.
2031 // ((X1 ^/- X2) || (X3 ^/- X4)) == 0 --> (X1 == X2) && (X3 == X4)
2032 // ((X1 ^/- X2) || (X3 ^/- X4)) != 0 --> (X1 != X2) || (X3 != X4)
2033 // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) == 0 -->
2034 // (X1 == X2) && (X3 == X4) && (X5 == X6)
2035 // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) != 0 -->
2036 // (X1 != X2) || (X3 != X4) || (X5 != X6)
2037 SmallVector<std::pair<Value *, Value *>, 2> CmpValues;
2038 SmallVector<Value *, 16> WorkList(1, Or);
2039
2040 while (!WorkList.empty()) {
2041 auto MatchOrOperatorArgument = [&](Value *OrOperatorArgument) {
2042 Value *Lhs, *Rhs;
2043
2044 if (match(V: OrOperatorArgument,
2045 P: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: Lhs), R: m_Value(V&: Rhs))))) {
2046 CmpValues.emplace_back(Args&: Lhs, Args&: Rhs);
2047 return;
2048 }
2049
2050 if (match(V: OrOperatorArgument,
2051 P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: Lhs), R: m_Value(V&: Rhs))))) {
2052 CmpValues.emplace_back(Args&: Lhs, Args&: Rhs);
2053 return;
2054 }
2055
2056 WorkList.push_back(Elt: OrOperatorArgument);
2057 };
2058
2059 Value *CurrentValue = WorkList.pop_back_val();
2060 Value *OrOperatorLhs, *OrOperatorRhs;
2061
2062 if (!match(V: CurrentValue,
2063 P: m_Or(L: m_Value(V&: OrOperatorLhs), R: m_Value(V&: OrOperatorRhs)))) {
2064 return nullptr;
2065 }
2066
2067 MatchOrOperatorArgument(OrOperatorRhs);
2068 MatchOrOperatorArgument(OrOperatorLhs);
2069 }
2070
2071 ICmpInst::Predicate Pred = Cmp.getPredicate();
2072 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2073 Value *LhsCmp = Builder.CreateICmp(P: Pred, LHS: CmpValues.rbegin()->first,
2074 RHS: CmpValues.rbegin()->second);
2075
2076 for (auto It = CmpValues.rbegin() + 1; It != CmpValues.rend(); ++It) {
2077 Value *RhsCmp = Builder.CreateICmp(P: Pred, LHS: It->first, RHS: It->second);
2078 LhsCmp = Builder.CreateBinOp(Opc: BOpc, LHS: LhsCmp, RHS: RhsCmp);
2079 }
2080
2081 return LhsCmp;
2082}
2083
2084/// Fold icmp (or X, Y), C.
2085Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
2086 BinaryOperator *Or,
2087 const APInt &C) {
2088 ICmpInst::Predicate Pred = Cmp.getPredicate();
2089 if (C.isOne()) {
2090 // icmp slt signum(V) 1 --> icmp slt V, 1
2091 Value *V = nullptr;
2092 if (Pred == ICmpInst::ICMP_SLT && match(V: Or, P: m_Signum(V: m_Value(V))))
2093 return new ICmpInst(ICmpInst::ICMP_SLT, V,
2094 ConstantInt::get(Ty: V->getType(), V: 1));
2095 }
2096
2097 Value *OrOp0 = Or->getOperand(i_nocapture: 0), *OrOp1 = Or->getOperand(i_nocapture: 1);
2098
2099 // (icmp eq/ne (or disjoint x, C0), C1)
2100 // -> (icmp eq/ne x, C0^C1)
2101 if (Cmp.isEquality() && match(V: OrOp1, P: m_ImmConstant()) &&
2102 cast<PossiblyDisjointInst>(Val: Or)->isDisjoint()) {
2103 Value *NewC =
2104 Builder.CreateXor(LHS: OrOp1, RHS: ConstantInt::get(Ty: OrOp1->getType(), V: C));
2105 return new ICmpInst(Pred, OrOp0, NewC);
2106 }
2107
2108 const APInt *MaskC;
2109 if (match(V: OrOp1, P: m_APInt(Res&: MaskC)) && Cmp.isEquality()) {
2110 if (*MaskC == C && (C + 1).isPowerOf2()) {
2111 // X | C == C --> X <=u C
2112 // X | C != C --> X >u C
2113 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
2114 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
2115 return new ICmpInst(Pred, OrOp0, OrOp1);
2116 }
2117
2118 // More general: canonicalize 'equality with set bits mask' to
2119 // 'equality with clear bits mask'.
2120 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
2121 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
2122 if (Or->hasOneUse()) {
2123 Value *And = Builder.CreateAnd(LHS: OrOp0, RHS: ~(*MaskC));
2124 Constant *NewC = ConstantInt::get(Ty: Or->getType(), V: C ^ (*MaskC));
2125 return new ICmpInst(Pred, And, NewC);
2126 }
2127 }
2128
2129 // (X | (X-1)) s< 0 --> X s< 1
2130 // (X | (X-1)) s> -1 --> X s> 0
2131 Value *X;
2132 bool TrueIfSigned;
2133 if (isSignBitCheck(Pred, RHS: C, TrueIfSigned) &&
2134 match(V: Or, P: m_c_Or(L: m_Add(L: m_Value(V&: X), R: m_AllOnes()), R: m_Deferred(V: X)))) {
2135 auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT;
2136 Constant *NewC = ConstantInt::get(Ty: X->getType(), V: TrueIfSigned ? 1 : 0);
2137 return new ICmpInst(NewPred, X, NewC);
2138 }
2139
2140 const APInt *OrC;
2141 // icmp(X | OrC, C) --> icmp(X, 0)
2142 if (C.isNonNegative() && match(V: Or, P: m_Or(L: m_Value(V&: X), R: m_APInt(Res&: OrC)))) {
2143 switch (Pred) {
2144 // X | OrC s< C --> X s< 0 iff OrC s>= C s>= 0
2145 case ICmpInst::ICMP_SLT:
2146 // X | OrC s>= C --> X s>= 0 iff OrC s>= C s>= 0
2147 case ICmpInst::ICMP_SGE:
2148 if (OrC->sge(RHS: C))
2149 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType()));
2150 break;
2151 // X | OrC s<= C --> X s< 0 iff OrC s> C s>= 0
2152 case ICmpInst::ICMP_SLE:
2153 // X | OrC s> C --> X s>= 0 iff OrC s> C s>= 0
2154 case ICmpInst::ICMP_SGT:
2155 if (OrC->sgt(RHS: C))
2156 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), X,
2157 ConstantInt::getNullValue(Ty: X->getType()));
2158 break;
2159 default:
2160 break;
2161 }
2162 }
2163
2164 if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse())
2165 return nullptr;
2166
2167 Value *P, *Q;
2168 if (match(V: Or, P: m_Or(L: m_PtrToInt(Op: m_Value(V&: P)), R: m_PtrToInt(Op: m_Value(V&: Q))))) {
2169 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
2170 // -> and (icmp eq P, null), (icmp eq Q, null).
2171 Value *CmpP =
2172 Builder.CreateICmp(P: Pred, LHS: P, RHS: ConstantInt::getNullValue(Ty: P->getType()));
2173 Value *CmpQ =
2174 Builder.CreateICmp(P: Pred, LHS: Q, RHS: ConstantInt::getNullValue(Ty: Q->getType()));
2175 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2176 return BinaryOperator::Create(Op: BOpc, S1: CmpP, S2: CmpQ);
2177 }
2178
2179 if (Value *V = foldICmpOrXorSubChain(Cmp, Or, Builder))
2180 return replaceInstUsesWith(I&: Cmp, V);
2181
2182 return nullptr;
2183}
2184
2185/// Fold icmp (mul X, Y), C.
2186Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
2187 BinaryOperator *Mul,
2188 const APInt &C) {
2189 ICmpInst::Predicate Pred = Cmp.getPredicate();
2190 Type *MulTy = Mul->getType();
2191 Value *X = Mul->getOperand(i_nocapture: 0);
2192
2193 // If there's no overflow:
2194 // X * X == 0 --> X == 0
2195 // X * X != 0 --> X != 0
2196 if (Cmp.isEquality() && C.isZero() && X == Mul->getOperand(i_nocapture: 1) &&
2197 (Mul->hasNoUnsignedWrap() || Mul->hasNoSignedWrap()))
2198 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: MulTy));
2199
2200 const APInt *MulC;
2201 if (!match(V: Mul->getOperand(i_nocapture: 1), P: m_APInt(Res&: MulC)))
2202 return nullptr;
2203
2204 // If this is a test of the sign bit and the multiply is sign-preserving with
2205 // a constant operand, use the multiply LHS operand instead:
2206 // (X * +MulC) < 0 --> X < 0
2207 // (X * -MulC) < 0 --> X > 0
2208 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
2209 if (MulC->isNegative())
2210 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2211 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: MulTy));
2212 }
2213
2214 if (MulC->isZero())
2215 return nullptr;
2216
2217 // If the multiply does not wrap or the constant is odd, try to divide the
2218 // compare constant by the multiplication factor.
2219 if (Cmp.isEquality()) {
2220 // (mul nsw X, MulC) eq/ne C --> X eq/ne C /s MulC
2221 if (Mul->hasNoSignedWrap() && C.srem(RHS: *MulC).isZero()) {
2222 Constant *NewC = ConstantInt::get(Ty: MulTy, V: C.sdiv(RHS: *MulC));
2223 return new ICmpInst(Pred, X, NewC);
2224 }
2225
2226 // C % MulC == 0 is weaker than we could use if MulC is odd because it
2227 // correct to transform if MulC * N == C including overflow. I.e with i8
2228 // (icmp eq (mul X, 5), 101) -> (icmp eq X, 225) but since 101 % 5 != 0, we
2229 // miss that case.
2230 if (C.urem(RHS: *MulC).isZero()) {
2231 // (mul nuw X, MulC) eq/ne C --> X eq/ne C /u MulC
2232 // (mul X, OddC) eq/ne N * C --> X eq/ne N
2233 if ((*MulC & 1).isOne() || Mul->hasNoUnsignedWrap()) {
2234 Constant *NewC = ConstantInt::get(Ty: MulTy, V: C.udiv(RHS: *MulC));
2235 return new ICmpInst(Pred, X, NewC);
2236 }
2237 }
2238 }
2239
2240 // With a matching no-overflow guarantee, fold the constants:
2241 // (X * MulC) < C --> X < (C / MulC)
2242 // (X * MulC) > C --> X > (C / MulC)
2243 // TODO: Assert that Pred is not equal to SGE, SLE, UGE, ULE?
2244 Constant *NewC = nullptr;
2245 if (Mul->hasNoSignedWrap() && ICmpInst::isSigned(predicate: Pred)) {
2246 // MININT / -1 --> overflow.
2247 if (C.isMinSignedValue() && MulC->isAllOnes())
2248 return nullptr;
2249 if (MulC->isNegative())
2250 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2251
2252 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2253 NewC = ConstantInt::get(
2254 Ty: MulTy, V: APIntOps::RoundingSDiv(A: C, B: *MulC, RM: APInt::Rounding::UP));
2255 } else {
2256 assert((Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_SGT) &&
2257 "Unexpected predicate");
2258 NewC = ConstantInt::get(
2259 Ty: MulTy, V: APIntOps::RoundingSDiv(A: C, B: *MulC, RM: APInt::Rounding::DOWN));
2260 }
2261 } else if (Mul->hasNoUnsignedWrap() && ICmpInst::isUnsigned(predicate: Pred)) {
2262 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) {
2263 NewC = ConstantInt::get(
2264 Ty: MulTy, V: APIntOps::RoundingUDiv(A: C, B: *MulC, RM: APInt::Rounding::UP));
2265 } else {
2266 assert((Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
2267 "Unexpected predicate");
2268 NewC = ConstantInt::get(
2269 Ty: MulTy, V: APIntOps::RoundingUDiv(A: C, B: *MulC, RM: APInt::Rounding::DOWN));
2270 }
2271 }
2272
2273 return NewC ? new ICmpInst(Pred, X, NewC) : nullptr;
2274}
2275
2276/// Fold icmp (shl nuw C2, Y), C.
2277static Instruction *foldICmpShlLHSC(ICmpInst &Cmp, Instruction *Shl,
2278 const APInt &C) {
2279 Value *Y;
2280 const APInt *C2;
2281 if (!match(V: Shl, P: m_NUWShl(L: m_APInt(Res&: C2), R: m_Value(V&: Y))))
2282 return nullptr;
2283
2284 Type *ShiftType = Shl->getType();
2285 unsigned TypeBits = C.getBitWidth();
2286 ICmpInst::Predicate Pred = Cmp.getPredicate();
2287 if (Cmp.isUnsigned()) {
2288 if (C2->isZero() || C2->ugt(RHS: C))
2289 return nullptr;
2290 APInt Div, Rem;
2291 APInt::udivrem(LHS: C, RHS: *C2, Quotient&: Div, Remainder&: Rem);
2292 bool CIsPowerOf2 = Rem.isZero() && Div.isPowerOf2();
2293
2294 // (1 << Y) pred C -> Y pred Log2(C)
2295 if (!CIsPowerOf2) {
2296 // (1 << Y) < 30 -> Y <= 4
2297 // (1 << Y) <= 30 -> Y <= 4
2298 // (1 << Y) >= 30 -> Y > 4
2299 // (1 << Y) > 30 -> Y > 4
2300 if (Pred == ICmpInst::ICMP_ULT)
2301 Pred = ICmpInst::ICMP_ULE;
2302 else if (Pred == ICmpInst::ICMP_UGE)
2303 Pred = ICmpInst::ICMP_UGT;
2304 }
2305
2306 unsigned CLog2 = Div.logBase2();
2307 return new ICmpInst(Pred, Y, ConstantInt::get(Ty: ShiftType, V: CLog2));
2308 } else if (Cmp.isSigned() && C2->isOne()) {
2309 Constant *BitWidthMinusOne = ConstantInt::get(Ty: ShiftType, V: TypeBits - 1);
2310 // (1 << Y) > 0 -> Y != 31
2311 // (1 << Y) > C -> Y != 31 if C is negative.
2312 if (Pred == ICmpInst::ICMP_SGT && C.sle(RHS: 0))
2313 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2314
2315 // (1 << Y) < 0 -> Y == 31
2316 // (1 << Y) < 1 -> Y == 31
2317 // (1 << Y) < C -> Y == 31 if C is negative and not signed min.
2318 // Exclude signed min by subtracting 1 and lower the upper bound to 0.
2319 if (Pred == ICmpInst::ICMP_SLT && (C - 1).sle(RHS: 0))
2320 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2321 }
2322
2323 return nullptr;
2324}
2325
2326/// Fold icmp (shl X, Y), C.
2327Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2328 BinaryOperator *Shl,
2329 const APInt &C) {
2330 const APInt *ShiftVal;
2331 if (Cmp.isEquality() && match(V: Shl->getOperand(i_nocapture: 0), P: m_APInt(Res&: ShiftVal)))
2332 return foldICmpShlConstConst(I&: Cmp, A: Shl->getOperand(i_nocapture: 1), AP1: C, AP2: *ShiftVal);
2333
2334 ICmpInst::Predicate Pred = Cmp.getPredicate();
2335 // (icmp pred (shl nuw&nsw X, Y), Csle0)
2336 // -> (icmp pred X, Csle0)
2337 //
2338 // The idea is the nuw/nsw essentially freeze the sign bit for the shift op
2339 // so X's must be what is used.
2340 if (C.sle(RHS: 0) && Shl->hasNoUnsignedWrap() && Shl->hasNoSignedWrap())
2341 return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1));
2342
2343 // (icmp eq/ne (shl nuw|nsw X, Y), 0)
2344 // -> (icmp eq/ne X, 0)
2345 if (ICmpInst::isEquality(P: Pred) && C.isZero() &&
2346 (Shl->hasNoUnsignedWrap() || Shl->hasNoSignedWrap()))
2347 return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1));
2348
2349 // (icmp slt (shl nsw X, Y), 0/1)
2350 // -> (icmp slt X, 0/1)
2351 // (icmp sgt (shl nsw X, Y), 0/-1)
2352 // -> (icmp sgt X, 0/-1)
2353 //
2354 // NB: sge/sle with a constant will canonicalize to sgt/slt.
2355 if (Shl->hasNoSignedWrap() &&
2356 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT))
2357 if (C.isZero() || (Pred == ICmpInst::ICMP_SGT ? C.isAllOnes() : C.isOne()))
2358 return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1));
2359
2360 const APInt *ShiftAmt;
2361 if (!match(V: Shl->getOperand(i_nocapture: 1), P: m_APInt(Res&: ShiftAmt)))
2362 return foldICmpShlLHSC(Cmp, Shl, C);
2363
2364 // Check that the shift amount is in range. If not, don't perform undefined
2365 // shifts. When the shift is visited, it will be simplified.
2366 unsigned TypeBits = C.getBitWidth();
2367 if (ShiftAmt->uge(RHS: TypeBits))
2368 return nullptr;
2369
2370 Value *X = Shl->getOperand(i_nocapture: 0);
2371 Type *ShType = Shl->getType();
2372
2373 // NSW guarantees that we are only shifting out sign bits from the high bits,
2374 // so we can ASHR the compare constant without needing a mask and eliminate
2375 // the shift.
2376 if (Shl->hasNoSignedWrap()) {
2377 if (Pred == ICmpInst::ICMP_SGT) {
2378 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2379 APInt ShiftedC = C.ashr(ShiftAmt: *ShiftAmt);
2380 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2381 }
2382 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2383 C.ashr(ShiftAmt: *ShiftAmt).shl(ShiftAmt: *ShiftAmt) == C) {
2384 APInt ShiftedC = C.ashr(ShiftAmt: *ShiftAmt);
2385 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2386 }
2387 if (Pred == ICmpInst::ICMP_SLT) {
2388 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2389 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2390 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2391 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2392 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2393 APInt ShiftedC = (C - 1).ashr(ShiftAmt: *ShiftAmt) + 1;
2394 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2395 }
2396 }
2397
2398 // NUW guarantees that we are only shifting out zero bits from the high bits,
2399 // so we can LSHR the compare constant without needing a mask and eliminate
2400 // the shift.
2401 if (Shl->hasNoUnsignedWrap()) {
2402 if (Pred == ICmpInst::ICMP_UGT) {
2403 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2404 APInt ShiftedC = C.lshr(ShiftAmt: *ShiftAmt);
2405 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2406 }
2407 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2408 C.lshr(ShiftAmt: *ShiftAmt).shl(ShiftAmt: *ShiftAmt) == C) {
2409 APInt ShiftedC = C.lshr(ShiftAmt: *ShiftAmt);
2410 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2411 }
2412 if (Pred == ICmpInst::ICMP_ULT) {
2413 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2414 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2415 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2416 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2417 assert(C.ugt(0) && "ult 0 should have been eliminated");
2418 APInt ShiftedC = (C - 1).lshr(ShiftAmt: *ShiftAmt) + 1;
2419 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC));
2420 }
2421 }
2422
2423 if (Cmp.isEquality() && Shl->hasOneUse()) {
2424 // Strength-reduce the shift into an 'and'.
2425 Constant *Mask = ConstantInt::get(
2426 Ty: ShType,
2427 V: APInt::getLowBitsSet(numBits: TypeBits, loBitsSet: TypeBits - ShiftAmt->getZExtValue()));
2428 Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shl->getName() + ".mask");
2429 Constant *LShrC = ConstantInt::get(Ty: ShType, V: C.lshr(ShiftAmt: *ShiftAmt));
2430 return new ICmpInst(Pred, And, LShrC);
2431 }
2432
2433 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2434 bool TrueIfSigned = false;
2435 if (Shl->hasOneUse() && isSignBitCheck(Pred, RHS: C, TrueIfSigned)) {
2436 // (X << 31) <s 0 --> (X & 1) != 0
2437 Constant *Mask = ConstantInt::get(
2438 Ty: ShType,
2439 V: APInt::getOneBitSet(numBits: TypeBits, BitNo: TypeBits - ShiftAmt->getZExtValue() - 1));
2440 Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shl->getName() + ".mask");
2441 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2442 And, Constant::getNullValue(Ty: ShType));
2443 }
2444
2445 // Simplify 'shl' inequality test into 'and' equality test.
2446 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2447 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2448 if ((C + 1).isPowerOf2() &&
2449 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2450 Value *And = Builder.CreateAnd(LHS: X, RHS: (~C).lshr(shiftAmt: ShiftAmt->getZExtValue()));
2451 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2452 : ICmpInst::ICMP_NE,
2453 And, Constant::getNullValue(Ty: ShType));
2454 }
2455 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2456 if (C.isPowerOf2() &&
2457 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2458 Value *And =
2459 Builder.CreateAnd(LHS: X, RHS: (~(C - 1)).lshr(shiftAmt: ShiftAmt->getZExtValue()));
2460 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2461 : ICmpInst::ICMP_NE,
2462 And, Constant::getNullValue(Ty: ShType));
2463 }
2464 }
2465
2466 // Transform (icmp pred iM (shl iM %v, N), C)
2467 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2468 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2469 // This enables us to get rid of the shift in favor of a trunc that may be
2470 // free on the target. It has the additional benefit of comparing to a
2471 // smaller constant that may be more target-friendly.
2472 unsigned Amt = ShiftAmt->getLimitedValue(Limit: TypeBits - 1);
2473 if (Shl->hasOneUse() && Amt != 0 &&
2474 shouldChangeType(FromBitWidth: ShType->getScalarSizeInBits(), ToBitWidth: TypeBits - Amt)) {
2475 ICmpInst::Predicate CmpPred = Pred;
2476 APInt RHSC = C;
2477
2478 if (RHSC.countr_zero() < Amt && ICmpInst::isStrictPredicate(predicate: CmpPred)) {
2479 // Try the flipped strictness predicate.
2480 // e.g.:
2481 // icmp ult i64 (shl X, 32), 8589934593 ->
2482 // icmp ule i64 (shl X, 32), 8589934592 ->
2483 // icmp ule i32 (trunc X, i32), 2 ->
2484 // icmp ult i32 (trunc X, i32), 3
2485 if (auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(
2486 Pred, C: ConstantInt::get(Context&: ShType->getContext(), V: C))) {
2487 CmpPred = FlippedStrictness->first;
2488 RHSC = cast<ConstantInt>(Val: FlippedStrictness->second)->getValue();
2489 }
2490 }
2491
2492 if (RHSC.countr_zero() >= Amt) {
2493 Type *TruncTy = ShType->getWithNewBitWidth(NewBitWidth: TypeBits - Amt);
2494 Constant *NewC =
2495 ConstantInt::get(Ty: TruncTy, V: RHSC.ashr(ShiftAmt: *ShiftAmt).trunc(width: TypeBits - Amt));
2496 return new ICmpInst(CmpPred,
2497 Builder.CreateTrunc(V: X, DestTy: TruncTy, Name: "", /*IsNUW=*/false,
2498 IsNSW: Shl->hasNoSignedWrap()),
2499 NewC);
2500 }
2501 }
2502
2503 return nullptr;
2504}
2505
2506/// Fold icmp ({al}shr X, Y), C.
2507Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2508 BinaryOperator *Shr,
2509 const APInt &C) {
2510 // An exact shr only shifts out zero bits, so:
2511 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2512 Value *X = Shr->getOperand(i_nocapture: 0);
2513 CmpInst::Predicate Pred = Cmp.getPredicate();
2514 if (Cmp.isEquality() && Shr->isExact() && C.isZero())
2515 return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1));
2516
2517 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2518 const APInt *ShiftValC;
2519 if (match(V: X, P: m_APInt(Res&: ShiftValC))) {
2520 if (Cmp.isEquality())
2521 return foldICmpShrConstConst(I&: Cmp, A: Shr->getOperand(i_nocapture: 1), AP1: C, AP2: *ShiftValC);
2522
2523 // (ShiftValC >> Y) >s -1 --> Y != 0 with ShiftValC < 0
2524 // (ShiftValC >> Y) <s 0 --> Y == 0 with ShiftValC < 0
2525 bool TrueIfSigned;
2526 if (!IsAShr && ShiftValC->isNegative() &&
2527 isSignBitCheck(Pred, RHS: C, TrueIfSigned))
2528 return new ICmpInst(TrueIfSigned ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE,
2529 Shr->getOperand(i_nocapture: 1),
2530 ConstantInt::getNullValue(Ty: X->getType()));
2531
2532 // If the shifted constant is a power-of-2, test the shift amount directly:
2533 // (ShiftValC >> Y) >u C --> X <u (LZ(C) - LZ(ShiftValC))
2534 // (ShiftValC >> Y) <u C --> X >=u (LZ(C-1) - LZ(ShiftValC))
2535 if (!IsAShr && ShiftValC->isPowerOf2() &&
2536 (Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_ULT)) {
2537 bool IsUGT = Pred == CmpInst::ICMP_UGT;
2538 assert(ShiftValC->uge(C) && "Expected simplify of compare");
2539 assert((IsUGT || !C.isZero()) && "Expected X u< 0 to simplify");
2540
2541 unsigned CmpLZ = IsUGT ? C.countl_zero() : (C - 1).countl_zero();
2542 unsigned ShiftLZ = ShiftValC->countl_zero();
2543 Constant *NewC = ConstantInt::get(Ty: Shr->getType(), V: CmpLZ - ShiftLZ);
2544 auto NewPred = IsUGT ? CmpInst::ICMP_ULT : CmpInst::ICMP_UGE;
2545 return new ICmpInst(NewPred, Shr->getOperand(i_nocapture: 1), NewC);
2546 }
2547 }
2548
2549 const APInt *ShiftAmtC;
2550 if (!match(V: Shr->getOperand(i_nocapture: 1), P: m_APInt(Res&: ShiftAmtC)))
2551 return nullptr;
2552
2553 // Check that the shift amount is in range. If not, don't perform undefined
2554 // shifts. When the shift is visited it will be simplified.
2555 unsigned TypeBits = C.getBitWidth();
2556 unsigned ShAmtVal = ShiftAmtC->getLimitedValue(Limit: TypeBits);
2557 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2558 return nullptr;
2559
2560 bool IsExact = Shr->isExact();
2561 Type *ShrTy = Shr->getType();
2562 // TODO: If we could guarantee that InstSimplify would handle all of the
2563 // constant-value-based preconditions in the folds below, then we could assert
2564 // those conditions rather than checking them. This is difficult because of
2565 // undef/poison (PR34838).
2566 if (IsAShr && Shr->hasOneUse()) {
2567 if (IsExact && (Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) &&
2568 (C - 1).isPowerOf2() && C.countLeadingZeros() > ShAmtVal) {
2569 // When C - 1 is a power of two and the transform can be legally
2570 // performed, prefer this form so the produced constant is close to a
2571 // power of two.
2572 // icmp slt/ult (ashr exact X, ShAmtC), C
2573 // --> icmp slt/ult X, (C - 1) << ShAmtC) + 1
2574 APInt ShiftedC = (C - 1).shl(shiftAmt: ShAmtVal) + 1;
2575 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2576 }
2577 if (IsExact || Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) {
2578 // When ShAmtC can be shifted losslessly:
2579 // icmp PRED (ashr exact X, ShAmtC), C --> icmp PRED X, (C << ShAmtC)
2580 // icmp slt/ult (ashr X, ShAmtC), C --> icmp slt/ult X, (C << ShAmtC)
2581 APInt ShiftedC = C.shl(shiftAmt: ShAmtVal);
2582 if (ShiftedC.ashr(ShiftAmt: ShAmtVal) == C)
2583 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2584 }
2585 if (Pred == CmpInst::ICMP_SGT) {
2586 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2587 APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1;
2588 if (!C.isMaxSignedValue() && !(C + 1).shl(shiftAmt: ShAmtVal).isMinSignedValue() &&
2589 (ShiftedC + 1).ashr(ShiftAmt: ShAmtVal) == (C + 1))
2590 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2591 }
2592 if (Pred == CmpInst::ICMP_UGT) {
2593 // icmp ugt (ashr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2594 // 'C + 1 << ShAmtC' can overflow as a signed number, so the 2nd
2595 // clause accounts for that pattern.
2596 APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1;
2597 if ((ShiftedC + 1).ashr(ShiftAmt: ShAmtVal) == (C + 1) ||
2598 (C + 1).shl(shiftAmt: ShAmtVal).isMinSignedValue())
2599 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2600 }
2601
2602 // If the compare constant has significant bits above the lowest sign-bit,
2603 // then convert an unsigned cmp to a test of the sign-bit:
2604 // (ashr X, ShiftC) u> C --> X s< 0
2605 // (ashr X, ShiftC) u< C --> X s> -1
2606 if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2607 if (Pred == CmpInst::ICMP_UGT) {
2608 return new ICmpInst(CmpInst::ICMP_SLT, X,
2609 ConstantInt::getNullValue(Ty: ShrTy));
2610 }
2611 if (Pred == CmpInst::ICMP_ULT) {
2612 return new ICmpInst(CmpInst::ICMP_SGT, X,
2613 ConstantInt::getAllOnesValue(Ty: ShrTy));
2614 }
2615 }
2616 } else if (!IsAShr) {
2617 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2618 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2619 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2620 APInt ShiftedC = C.shl(shiftAmt: ShAmtVal);
2621 if (ShiftedC.lshr(shiftAmt: ShAmtVal) == C)
2622 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2623 }
2624 if (Pred == CmpInst::ICMP_UGT) {
2625 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2626 APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1;
2627 if ((ShiftedC + 1).lshr(shiftAmt: ShAmtVal) == (C + 1))
2628 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC));
2629 }
2630 }
2631
2632 if (!Cmp.isEquality())
2633 return nullptr;
2634
2635 // Handle equality comparisons of shift-by-constant.
2636
2637 // If the comparison constant changes with the shift, the comparison cannot
2638 // succeed (bits of the comparison constant cannot match the shifted value).
2639 // This should be known by InstSimplify and already be folded to true/false.
2640 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2641 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2642 "Expected icmp+shr simplify did not occur.");
2643
2644 // If the bits shifted out are known zero, compare the unshifted value:
2645 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2646 if (Shr->isExact())
2647 return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: C << ShAmtVal));
2648
2649 if (Shr->hasOneUse()) {
2650 // Canonicalize the shift into an 'and':
2651 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2652 APInt Val(APInt::getHighBitsSet(numBits: TypeBits, hiBitsSet: TypeBits - ShAmtVal));
2653 Constant *Mask = ConstantInt::get(Ty: ShrTy, V: Val);
2654 Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shr->getName() + ".mask");
2655 return new ICmpInst(Pred, And, ConstantInt::get(Ty: ShrTy, V: C << ShAmtVal));
2656 }
2657
2658 return nullptr;
2659}
2660
2661Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2662 BinaryOperator *SRem,
2663 const APInt &C) {
2664 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2665 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT) {
2666 // Canonicalize unsigned predicates to signed:
2667 // (X s% DivisorC) u> C -> (X s% DivisorC) s< 0
2668 // iff (C s< 0 ? ~C : C) u>= abs(DivisorC)-1
2669 // (X s% DivisorC) u< C+1 -> (X s% DivisorC) s> -1
2670 // iff (C+1 s< 0 ? ~C : C) u>= abs(DivisorC)-1
2671
2672 const APInt *DivisorC;
2673 if (!match(V: SRem->getOperand(i_nocapture: 1), P: m_APInt(Res&: DivisorC)))
2674 return nullptr;
2675 if (DivisorC->isZero())
2676 return nullptr;
2677
2678 APInt NormalizedC = C;
2679 if (Pred == ICmpInst::ICMP_ULT) {
2680 assert(!NormalizedC.isZero() &&
2681 "ult X, 0 should have been simplified already.");
2682 --NormalizedC;
2683 }
2684 if (C.isNegative())
2685 NormalizedC.flipAllBits();
2686 if (!NormalizedC.uge(RHS: DivisorC->abs() - 1))
2687 return nullptr;
2688
2689 Type *Ty = SRem->getType();
2690 if (Pred == ICmpInst::ICMP_UGT)
2691 return new ICmpInst(ICmpInst::ICMP_SLT, SRem,
2692 ConstantInt::getNullValue(Ty));
2693 return new ICmpInst(ICmpInst::ICMP_SGT, SRem,
2694 ConstantInt::getAllOnesValue(Ty));
2695 }
2696 // Match an 'is positive' or 'is negative' comparison of remainder by a
2697 // constant power-of-2 value:
2698 // (X % pow2C) sgt/slt 0
2699 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT &&
2700 Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2701 return nullptr;
2702
2703 // TODO: The one-use check is standard because we do not typically want to
2704 // create longer instruction sequences, but this might be a special-case
2705 // because srem is not good for analysis or codegen.
2706 if (!SRem->hasOneUse())
2707 return nullptr;
2708
2709 const APInt *DivisorC;
2710 if (!match(V: SRem->getOperand(i_nocapture: 1), P: m_Power2(V&: DivisorC)))
2711 return nullptr;
2712
2713 // For cmp_sgt/cmp_slt only zero valued C is handled.
2714 // For cmp_eq/cmp_ne only positive valued C is handled.
2715 if (((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT) &&
2716 !C.isZero()) ||
2717 ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2718 !C.isStrictlyPositive()))
2719 return nullptr;
2720
2721 // Mask off the sign bit and the modulo bits (low-bits).
2722 Type *Ty = SRem->getType();
2723 APInt SignMask = APInt::getSignMask(BitWidth: Ty->getScalarSizeInBits());
2724 Constant *MaskC = ConstantInt::get(Ty, V: SignMask | (*DivisorC - 1));
2725 Value *And = Builder.CreateAnd(LHS: SRem->getOperand(i_nocapture: 0), RHS: MaskC);
2726
2727 if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)
2728 return new ICmpInst(Pred, And, ConstantInt::get(Ty, V: C));
2729
2730 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2731 // bit is set. Example:
2732 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2733 if (Pred == ICmpInst::ICMP_SGT)
2734 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2735
2736 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2737 // bit is set. Example:
2738 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2739 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, V: SignMask));
2740}
2741
2742/// Fold icmp (udiv X, Y), C.
2743Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2744 BinaryOperator *UDiv,
2745 const APInt &C) {
2746 ICmpInst::Predicate Pred = Cmp.getPredicate();
2747 Value *X = UDiv->getOperand(i_nocapture: 0);
2748 Value *Y = UDiv->getOperand(i_nocapture: 1);
2749 Type *Ty = UDiv->getType();
2750
2751 const APInt *C2;
2752 if (!match(V: X, P: m_APInt(Res&: C2)))
2753 return nullptr;
2754
2755 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2756
2757 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2758 if (Pred == ICmpInst::ICMP_UGT) {
2759 assert(!C.isMaxValue() &&
2760 "icmp ugt X, UINT_MAX should have been simplified already.");
2761 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2762 ConstantInt::get(Ty, V: C2->udiv(RHS: C + 1)));
2763 }
2764
2765 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2766 if (Pred == ICmpInst::ICMP_ULT) {
2767 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2768 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2769 ConstantInt::get(Ty, V: C2->udiv(RHS: C)));
2770 }
2771
2772 return nullptr;
2773}
2774
2775/// Fold icmp ({su}div X, Y), C.
2776Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2777 BinaryOperator *Div,
2778 const APInt &C) {
2779 ICmpInst::Predicate Pred = Cmp.getPredicate();
2780 Value *X = Div->getOperand(i_nocapture: 0);
2781 Value *Y = Div->getOperand(i_nocapture: 1);
2782 Type *Ty = Div->getType();
2783 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2784
2785 // If unsigned division and the compare constant is bigger than
2786 // UMAX/2 (negative), there's only one pair of values that satisfies an
2787 // equality check, so eliminate the division:
2788 // (X u/ Y) == C --> (X == C) && (Y == 1)
2789 // (X u/ Y) != C --> (X != C) || (Y != 1)
2790 // Similarly, if signed division and the compare constant is exactly SMIN:
2791 // (X s/ Y) == SMIN --> (X == SMIN) && (Y == 1)
2792 // (X s/ Y) != SMIN --> (X != SMIN) || (Y != 1)
2793 if (Cmp.isEquality() && Div->hasOneUse() && C.isSignBitSet() &&
2794 (!DivIsSigned || C.isMinSignedValue())) {
2795 Value *XBig = Builder.CreateICmp(P: Pred, LHS: X, RHS: ConstantInt::get(Ty, V: C));
2796 Value *YOne = Builder.CreateICmp(P: Pred, LHS: Y, RHS: ConstantInt::get(Ty, V: 1));
2797 auto Logic = Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
2798 return BinaryOperator::Create(Op: Logic, S1: XBig, S2: YOne);
2799 }
2800
2801 // Fold: icmp pred ([us]div X, C2), C -> range test
2802 // Fold this div into the comparison, producing a range check.
2803 // Determine, based on the divide type, what the range is being
2804 // checked. If there is an overflow on the low or high side, remember
2805 // it, otherwise compute the range [low, hi) bounding the new value.
2806 // See: InsertRangeTest above for the kinds of replacements possible.
2807 const APInt *C2;
2808 if (!match(V: Y, P: m_APInt(Res&: C2)))
2809 return nullptr;
2810
2811 // FIXME: If the operand types don't match the type of the divide
2812 // then don't attempt this transform. The code below doesn't have the
2813 // logic to deal with a signed divide and an unsigned compare (and
2814 // vice versa). This is because (x /s C2) <s C produces different
2815 // results than (x /s C2) <u C or (x /u C2) <s C or even
2816 // (x /u C2) <u C. Simply casting the operands and result won't
2817 // work. :( The if statement below tests that condition and bails
2818 // if it finds it.
2819 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2820 return nullptr;
2821
2822 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2823 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2824 // division-by-constant cases should be present, we can not assert that they
2825 // have happened before we reach this icmp instruction.
2826 if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes()))
2827 return nullptr;
2828
2829 // Compute Prod = C * C2. We are essentially solving an equation of
2830 // form X / C2 = C. We solve for X by multiplying C2 and C.
2831 // By solving for X, we can turn this into a range check instead of computing
2832 // a divide.
2833 APInt Prod = C * *C2;
2834
2835 // Determine if the product overflows by seeing if the product is not equal to
2836 // the divide. Make sure we do the same kind of divide as in the LHS
2837 // instruction that we're folding.
2838 bool ProdOV = (DivIsSigned ? Prod.sdiv(RHS: *C2) : Prod.udiv(RHS: *C2)) != C;
2839
2840 // If the division is known to be exact, then there is no remainder from the
2841 // divide, so the covered range size is unit, otherwise it is the divisor.
2842 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2843
2844 // Figure out the interval that is being checked. For example, a comparison
2845 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2846 // Compute this interval based on the constants involved and the signedness of
2847 // the compare/divide. This computes a half-open interval, keeping track of
2848 // whether either value in the interval overflows. After analysis each
2849 // overflow variable is set to 0 if it's corresponding bound variable is valid
2850 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2851 int LoOverflow = 0, HiOverflow = 0;
2852 APInt LoBound, HiBound;
2853
2854 if (!DivIsSigned) { // udiv
2855 // e.g. X/5 op 3 --> [15, 20)
2856 LoBound = Prod;
2857 HiOverflow = LoOverflow = ProdOV;
2858 if (!HiOverflow) {
2859 // If this is not an exact divide, then many values in the range collapse
2860 // to the same result value.
2861 HiOverflow = addWithOverflow(Result&: HiBound, In1: LoBound, In2: RangeSize, IsSigned: false);
2862 }
2863 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2864 if (C.isZero()) { // (X / pos) op 0
2865 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2866 LoBound = -(RangeSize - 1);
2867 HiBound = RangeSize;
2868 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2869 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2870 HiOverflow = LoOverflow = ProdOV;
2871 if (!HiOverflow)
2872 HiOverflow = addWithOverflow(Result&: HiBound, In1: Prod, In2: RangeSize, IsSigned: true);
2873 } else { // (X / pos) op neg
2874 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2875 HiBound = Prod + 1;
2876 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2877 if (!LoOverflow) {
2878 APInt DivNeg = -RangeSize;
2879 LoOverflow = addWithOverflow(Result&: LoBound, In1: HiBound, In2: DivNeg, IsSigned: true) ? -1 : 0;
2880 }
2881 }
2882 } else if (C2->isNegative()) { // Divisor is < 0.
2883 if (Div->isExact())
2884 RangeSize.negate();
2885 if (C.isZero()) { // (X / neg) op 0
2886 // e.g. X/-5 op 0 --> [-4, 5)
2887 LoBound = RangeSize + 1;
2888 HiBound = -RangeSize;
2889 if (HiBound == *C2) { // -INTMIN = INTMIN
2890 HiOverflow = 1; // [INTMIN+1, overflow)
2891 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2892 }
2893 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2894 // e.g. X/-5 op 3 --> [-19, -14)
2895 HiBound = Prod + 1;
2896 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2897 if (!LoOverflow)
2898 LoOverflow =
2899 addWithOverflow(Result&: LoBound, In1: HiBound, In2: RangeSize, IsSigned: true) ? -1 : 0;
2900 } else { // (X / neg) op neg
2901 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2902 LoOverflow = HiOverflow = ProdOV;
2903 if (!HiOverflow)
2904 HiOverflow = subWithOverflow(Result&: HiBound, In1: Prod, In2: RangeSize, IsSigned: true);
2905 }
2906
2907 // Dividing by a negative swaps the condition. LT <-> GT
2908 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
2909 }
2910
2911 switch (Pred) {
2912 default:
2913 llvm_unreachable("Unhandled icmp predicate!");
2914 case ICmpInst::ICMP_EQ:
2915 if (LoOverflow && HiOverflow)
2916 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
2917 if (HiOverflow)
2918 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
2919 X, ConstantInt::get(Ty, V: LoBound));
2920 if (LoOverflow)
2921 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
2922 X, ConstantInt::get(Ty, V: HiBound));
2923 return replaceInstUsesWith(
2924 I&: Cmp, V: insertRangeTest(V: X, Lo: LoBound, Hi: HiBound, isSigned: DivIsSigned, Inside: true));
2925 case ICmpInst::ICMP_NE:
2926 if (LoOverflow && HiOverflow)
2927 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
2928 if (HiOverflow)
2929 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
2930 X, ConstantInt::get(Ty, V: LoBound));
2931 if (LoOverflow)
2932 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
2933 X, ConstantInt::get(Ty, V: HiBound));
2934 return replaceInstUsesWith(
2935 I&: Cmp, V: insertRangeTest(V: X, Lo: LoBound, Hi: HiBound, isSigned: DivIsSigned, Inside: false));
2936 case ICmpInst::ICMP_ULT:
2937 case ICmpInst::ICMP_SLT:
2938 if (LoOverflow == +1) // Low bound is greater than input range.
2939 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
2940 if (LoOverflow == -1) // Low bound is less than input range.
2941 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
2942 return new ICmpInst(Pred, X, ConstantInt::get(Ty, V: LoBound));
2943 case ICmpInst::ICMP_UGT:
2944 case ICmpInst::ICMP_SGT:
2945 if (HiOverflow == +1) // High bound greater than input range.
2946 return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse());
2947 if (HiOverflow == -1) // High bound less than input range.
2948 return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue());
2949 if (Pred == ICmpInst::ICMP_UGT)
2950 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, V: HiBound));
2951 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, V: HiBound));
2952 }
2953
2954 return nullptr;
2955}
2956
2957/// Fold icmp (sub X, Y), C.
2958Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2959 BinaryOperator *Sub,
2960 const APInt &C) {
2961 Value *X = Sub->getOperand(i_nocapture: 0), *Y = Sub->getOperand(i_nocapture: 1);
2962 ICmpInst::Predicate Pred = Cmp.getPredicate();
2963 Type *Ty = Sub->getType();
2964
2965 // (SubC - Y) == C) --> Y == (SubC - C)
2966 // (SubC - Y) != C) --> Y != (SubC - C)
2967 Constant *SubC;
2968 if (Cmp.isEquality() && match(V: X, P: m_ImmConstant(C&: SubC))) {
2969 return new ICmpInst(Pred, Y,
2970 ConstantExpr::getSub(C1: SubC, C2: ConstantInt::get(Ty, V: C)));
2971 }
2972
2973 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2974 const APInt *C2;
2975 APInt SubResult;
2976 ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate();
2977 bool HasNSW = Sub->hasNoSignedWrap();
2978 bool HasNUW = Sub->hasNoUnsignedWrap();
2979 if (match(V: X, P: m_APInt(Res&: C2)) &&
2980 ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) &&
2981 !subWithOverflow(Result&: SubResult, In1: *C2, In2: C, IsSigned: Cmp.isSigned()))
2982 return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, V: SubResult));
2983
2984 // X - Y == 0 --> X == Y.
2985 // X - Y != 0 --> X != Y.
2986 // TODO: We allow this with multiple uses as long as the other uses are not
2987 // in phis. The phi use check is guarding against a codegen regression
2988 // for a loop test. If the backend could undo this (and possibly
2989 // subsequent transforms), we would not need this hack.
2990 if (Cmp.isEquality() && C.isZero() &&
2991 none_of(Range: (Sub->users()), P: [](const User *U) { return isa<PHINode>(Val: U); }))
2992 return new ICmpInst(Pred, X, Y);
2993
2994 // The following transforms are only worth it if the only user of the subtract
2995 // is the icmp.
2996 // TODO: This is an artificial restriction for all of the transforms below
2997 // that only need a single replacement icmp. Can these use the phi test
2998 // like the transform above here?
2999 if (!Sub->hasOneUse())
3000 return nullptr;
3001
3002 if (Sub->hasNoSignedWrap()) {
3003 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
3004 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
3005 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3006
3007 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
3008 if (Pred == ICmpInst::ICMP_SGT && C.isZero())
3009 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3010
3011 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
3012 if (Pred == ICmpInst::ICMP_SLT && C.isZero())
3013 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3014
3015 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
3016 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
3017 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3018 }
3019
3020 if (!match(V: X, P: m_APInt(Res&: C2)))
3021 return nullptr;
3022
3023 // C2 - Y <u C -> (Y | (C - 1)) == C2
3024 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
3025 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
3026 (*C2 & (C - 1)) == (C - 1))
3027 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(LHS: Y, RHS: C - 1), X);
3028
3029 // C2 - Y >u C -> (Y | C) != C2
3030 // iff C2 & C == C and C + 1 is a power of 2
3031 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
3032 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(LHS: Y, RHS: C), X);
3033
3034 // We have handled special cases that reduce.
3035 // Canonicalize any remaining sub to add as:
3036 // (C2 - Y) > C --> (Y + ~C2) < ~C
3037 Value *Add = Builder.CreateAdd(LHS: Y, RHS: ConstantInt::get(Ty, V: ~(*C2)), Name: "notsub",
3038 HasNUW, HasNSW);
3039 return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, V: ~C));
3040}
3041
3042static Value *createLogicFromTable(const std::bitset<4> &Table, Value *Op0,
3043 Value *Op1, IRBuilderBase &Builder,
3044 bool HasOneUse) {
3045 auto FoldConstant = [&](bool Val) {
3046 Constant *Res = Val ? Builder.getTrue() : Builder.getFalse();
3047 if (Op0->getType()->isVectorTy())
3048 Res = ConstantVector::getSplat(
3049 EC: cast<VectorType>(Val: Op0->getType())->getElementCount(), Elt: Res);
3050 return Res;
3051 };
3052
3053 switch (Table.to_ulong()) {
3054 case 0: // 0 0 0 0
3055 return FoldConstant(false);
3056 case 1: // 0 0 0 1
3057 return HasOneUse ? Builder.CreateNot(V: Builder.CreateOr(LHS: Op0, RHS: Op1)) : nullptr;
3058 case 2: // 0 0 1 0
3059 return HasOneUse ? Builder.CreateAnd(LHS: Builder.CreateNot(V: Op0), RHS: Op1) : nullptr;
3060 case 3: // 0 0 1 1
3061 return Builder.CreateNot(V: Op0);
3062 case 4: // 0 1 0 0
3063 return HasOneUse ? Builder.CreateAnd(LHS: Op0, RHS: Builder.CreateNot(V: Op1)) : nullptr;
3064 case 5: // 0 1 0 1
3065 return Builder.CreateNot(V: Op1);
3066 case 6: // 0 1 1 0
3067 return Builder.CreateXor(LHS: Op0, RHS: Op1);
3068 case 7: // 0 1 1 1
3069 return HasOneUse ? Builder.CreateNot(V: Builder.CreateAnd(LHS: Op0, RHS: Op1)) : nullptr;
3070 case 8: // 1 0 0 0
3071 return Builder.CreateAnd(LHS: Op0, RHS: Op1);
3072 case 9: // 1 0 0 1
3073 return HasOneUse ? Builder.CreateNot(V: Builder.CreateXor(LHS: Op0, RHS: Op1)) : nullptr;
3074 case 10: // 1 0 1 0
3075 return Op1;
3076 case 11: // 1 0 1 1
3077 return HasOneUse ? Builder.CreateOr(LHS: Builder.CreateNot(V: Op0), RHS: Op1) : nullptr;
3078 case 12: // 1 1 0 0
3079 return Op0;
3080 case 13: // 1 1 0 1
3081 return HasOneUse ? Builder.CreateOr(LHS: Op0, RHS: Builder.CreateNot(V: Op1)) : nullptr;
3082 case 14: // 1 1 1 0
3083 return Builder.CreateOr(LHS: Op0, RHS: Op1);
3084 case 15: // 1 1 1 1
3085 return FoldConstant(true);
3086 default:
3087 llvm_unreachable("Invalid Operation");
3088 }
3089 return nullptr;
3090}
3091
3092Instruction *InstCombinerImpl::foldICmpBinOpWithConstantViaTruthTable(
3093 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3094 Value *A, *B;
3095 Constant *C1, *C2, *C3, *C4;
3096 if (!(match(V: BO->getOperand(i_nocapture: 0),
3097 P: m_Select(C: m_Value(V&: A), L: m_Constant(C&: C1), R: m_Constant(C&: C2)))) ||
3098 !match(V: BO->getOperand(i_nocapture: 1),
3099 P: m_Select(C: m_Value(V&: B), L: m_Constant(C&: C3), R: m_Constant(C&: C4))) ||
3100 Cmp.getType() != A->getType() || Cmp.getType() != B->getType())
3101 return nullptr;
3102
3103 std::bitset<4> Table;
3104 auto ComputeTable = [&](bool First, bool Second) -> std::optional<bool> {
3105 Constant *L = First ? C1 : C2;
3106 Constant *R = Second ? C3 : C4;
3107 if (auto *Res = ConstantFoldBinaryOpOperands(Opcode: BO->getOpcode(), LHS: L, RHS: R, DL)) {
3108 auto *Val = Res->getType()->isVectorTy() ? Res->getSplatValue() : Res;
3109 if (auto *CI = dyn_cast_or_null<ConstantInt>(Val))
3110 return ICmpInst::compare(LHS: CI->getValue(), RHS: C, Pred: Cmp.getPredicate());
3111 }
3112 return std::nullopt;
3113 };
3114
3115 for (unsigned I = 0; I < 4; ++I) {
3116 bool First = (I >> 1) & 1;
3117 bool Second = I & 1;
3118 if (auto Res = ComputeTable(First, Second))
3119 Table[I] = *Res;
3120 else
3121 return nullptr;
3122 }
3123
3124 // Synthesize optimal logic.
3125 if (auto *Cond = createLogicFromTable(Table, Op0: A, Op1: B, Builder, HasOneUse: BO->hasOneUse()))
3126 return replaceInstUsesWith(I&: Cmp, V: Cond);
3127 return nullptr;
3128}
3129
3130/// Fold icmp (add X, Y), C.
3131Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
3132 BinaryOperator *Add,
3133 const APInt &C) {
3134 Value *Y = Add->getOperand(i_nocapture: 1);
3135 Value *X = Add->getOperand(i_nocapture: 0);
3136
3137 Value *Op0, *Op1;
3138 Instruction *Ext0, *Ext1;
3139 const CmpPredicate Pred = Cmp.getCmpPredicate();
3140 if (match(V: Add,
3141 P: m_Add(L: m_CombineAnd(L: m_Instruction(I&: Ext0), R: m_ZExtOrSExt(Op: m_Value(V&: Op0))),
3142 R: m_CombineAnd(L: m_Instruction(I&: Ext1),
3143 R: m_ZExtOrSExt(Op: m_Value(V&: Op1))))) &&
3144 Op0->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
3145 Op1->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
3146 unsigned BW = C.getBitWidth();
3147 std::bitset<4> Table;
3148 auto ComputeTable = [&](bool Op0Val, bool Op1Val) {
3149 APInt Res(BW, 0);
3150 if (Op0Val)
3151 Res += APInt(BW, isa<ZExtInst>(Val: Ext0) ? 1 : -1, /*isSigned=*/true);
3152 if (Op1Val)
3153 Res += APInt(BW, isa<ZExtInst>(Val: Ext1) ? 1 : -1, /*isSigned=*/true);
3154 return ICmpInst::compare(LHS: Res, RHS: C, Pred);
3155 };
3156
3157 Table[0] = ComputeTable(false, false);
3158 Table[1] = ComputeTable(false, true);
3159 Table[2] = ComputeTable(true, false);
3160 Table[3] = ComputeTable(true, true);
3161 if (auto *Cond =
3162 createLogicFromTable(Table, Op0, Op1, Builder, HasOneUse: Add->hasOneUse()))
3163 return replaceInstUsesWith(I&: Cmp, V: Cond);
3164 }
3165
3166 // icmp ult (add nuw A, (lshr A, ShAmtC)), C --> icmp ult A, C
3167 // when C <= (1 << ShAmtC).
3168 const APInt *ShAmtC;
3169 Value *A;
3170 unsigned BitWidth = C.getBitWidth();
3171 if (Pred == ICmpInst::ICMP_ULT &&
3172 match(V: Add,
3173 P: m_c_NUWAdd(L: m_Value(V&: A), R: m_LShr(L: m_Deferred(V: A), R: m_APInt(Res&: ShAmtC)))) &&
3174 ShAmtC->ult(RHS: BitWidth) &&
3175 C.ule(RHS: APInt::getOneBitSet(numBits: BitWidth, BitNo: ShAmtC->getZExtValue())))
3176 return new ICmpInst(Pred, A, ConstantInt::get(Ty: A->getType(), V: C));
3177
3178 const APInt *C2;
3179 if (Cmp.isEquality() || !match(V: Y, P: m_APInt(Res&: C2)))
3180 return nullptr;
3181
3182 // Fold icmp pred (add X, C2), C.
3183 Type *Ty = Add->getType();
3184
3185 // If the add does not wrap, we can always adjust the compare by subtracting
3186 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
3187 // have been canonicalized to SGT/SLT/UGT/ULT.
3188 if (Add->hasNoUnsignedWrap() &&
3189 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT)) {
3190 bool Overflow;
3191 APInt NewC = C.usub_ov(RHS: *C2, Overflow);
3192 // If there is overflow, the result must be true or false.
3193 if (!Overflow)
3194 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
3195 return new ICmpInst(Pred, X, ConstantInt::get(Ty, V: NewC));
3196 }
3197
3198 CmpInst::Predicate ChosenPred = Pred.getPreferredSignedPredicate();
3199
3200 if (Add->hasNoSignedWrap() &&
3201 (ChosenPred == ICmpInst::ICMP_SGT || ChosenPred == ICmpInst::ICMP_SLT)) {
3202 bool Overflow;
3203 APInt NewC = C.ssub_ov(RHS: *C2, Overflow);
3204 if (!Overflow)
3205 // icmp samesign ugt/ult (add nsw X, C2), C
3206 // -> icmp sgt/slt X, (C - C2)
3207 return new ICmpInst(ChosenPred, X, ConstantInt::get(Ty, V: NewC));
3208 }
3209
3210 if (ICmpInst::isUnsigned(predicate: Pred) && Add->hasNoSignedWrap() &&
3211 C.isNonNegative() && (C - *C2).isNonNegative() &&
3212 computeConstantRange(V: X, /*ForSigned=*/true).add(Other: *C2).isAllNonNegative())
3213 return new ICmpInst(ICmpInst::getSignedPredicate(Pred), X,
3214 ConstantInt::get(Ty, V: C - *C2));
3215
3216 auto CR = ConstantRange::makeExactICmpRegion(Pred, Other: C).subtract(CI: *C2);
3217 const APInt &Upper = CR.getUpper();
3218 const APInt &Lower = CR.getLower();
3219 if (Cmp.isSigned()) {
3220 if (Lower.isSignMask())
3221 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, V: Upper));
3222 if (Upper.isSignMask())
3223 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, V: Lower));
3224 } else {
3225 if (Lower.isMinValue())
3226 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, V: Upper));
3227 if (Upper.isMinValue())
3228 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, V: Lower));
3229 }
3230
3231 // This set of folds is intentionally placed after folds that use no-wrapping
3232 // flags because those folds are likely better for later analysis/codegen.
3233 const APInt SMax = APInt::getSignedMaxValue(numBits: Ty->getScalarSizeInBits());
3234 const APInt SMin = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits());
3235
3236 // Fold compare with offset to opposite sign compare if it eliminates offset:
3237 // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
3238 if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
3239 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, V: -(*C2)));
3240
3241 // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
3242 if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
3243 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, V: ~(*C2)));
3244
3245 // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
3246 if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
3247 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, V: SMax - C));
3248
3249 // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
3250 if (Pred == CmpInst::ICMP_SLT && C == *C2)
3251 return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, V: C ^ SMax));
3252
3253 // (X + -1) <u C --> X <=u C (if X is never null)
3254 if (Pred == CmpInst::ICMP_ULT && C2->isAllOnes()) {
3255 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
3256 if (llvm::isKnownNonZero(V: X, Q))
3257 return new ICmpInst(ICmpInst::ICMP_ULE, X, ConstantInt::get(Ty, V: C));
3258 }
3259
3260 if (!Add->hasOneUse())
3261 return nullptr;
3262
3263 // X+C <u C2 -> (X & -C2) == C
3264 // iff C & (C2-1) == 0
3265 // C2 is a power of 2
3266 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
3267 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(LHS: X, RHS: -C),
3268 ConstantExpr::getNeg(C: cast<Constant>(Val: Y)));
3269
3270 // X+C2 <u C -> (X & C) == 2C
3271 // iff C == -(C2)
3272 // C2 is a power of 2
3273 if (Pred == ICmpInst::ICMP_ULT && C2->isPowerOf2() && C == -*C2)
3274 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(LHS: X, RHS: C),
3275 ConstantInt::get(Ty, V: C * 2));
3276
3277 // X+C >u C2 -> (X & ~C2) != C
3278 // iff C & C2 == 0
3279 // C2+1 is a power of 2
3280 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
3281 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(LHS: X, RHS: ~C),
3282 ConstantExpr::getNeg(C: cast<Constant>(Val: Y)));
3283
3284 // The range test idiom can use either ult or ugt. Arbitrarily canonicalize
3285 // to the ult form.
3286 // X+C2 >u C -> X+(C2-C-1) <u ~C
3287 if (Pred == ICmpInst::ICMP_UGT)
3288 return new ICmpInst(ICmpInst::ICMP_ULT,
3289 Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: *C2 - C - 1)),
3290 ConstantInt::get(Ty, V: ~C));
3291
3292 // zext(V) + C2 pred C -> V + C3 pred' C4
3293 Value *V;
3294 if (match(V: X, P: m_ZExt(Op: m_Value(V)))) {
3295 Type *NewCmpTy = V->getType();
3296 unsigned NewCmpBW = NewCmpTy->getScalarSizeInBits();
3297 if (shouldChangeType(From: Ty, To: NewCmpTy)) {
3298 ConstantRange SrcCR = CR.truncate(BitWidth: NewCmpBW, NoWrapKind: TruncInst::NoUnsignedWrap);
3299 CmpInst::Predicate EquivPred;
3300 APInt EquivInt;
3301 APInt EquivOffset;
3302
3303 SrcCR.getEquivalentICmp(Pred&: EquivPred, RHS&: EquivInt, Offset&: EquivOffset);
3304 return new ICmpInst(
3305 EquivPred,
3306 EquivOffset.isZero()
3307 ? V
3308 : Builder.CreateAdd(LHS: V, RHS: ConstantInt::get(Ty: NewCmpTy, V: EquivOffset)),
3309 ConstantInt::get(Ty: NewCmpTy, V: EquivInt));
3310 }
3311 }
3312
3313 return nullptr;
3314}
3315
3316bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
3317 Value *&RHS, ConstantInt *&Less,
3318 ConstantInt *&Equal,
3319 ConstantInt *&Greater) {
3320 // TODO: Generalize this to work with other comparison idioms or ensure
3321 // they get canonicalized into this form.
3322
3323 // select i1 (a == b),
3324 // i32 Equal,
3325 // i32 (select i1 (a < b), i32 Less, i32 Greater)
3326 // where Equal, Less and Greater are placeholders for any three constants.
3327 CmpPredicate PredA;
3328 if (!match(V: SI->getCondition(), P: m_ICmp(Pred&: PredA, L: m_Value(V&: LHS), R: m_Value(V&: RHS))) ||
3329 !ICmpInst::isEquality(P: PredA))
3330 return false;
3331 Value *EqualVal = SI->getTrueValue();
3332 Value *UnequalVal = SI->getFalseValue();
3333 // We still can get non-canonical predicate here, so canonicalize.
3334 if (PredA == ICmpInst::ICMP_NE)
3335 std::swap(a&: EqualVal, b&: UnequalVal);
3336 if (!match(V: EqualVal, P: m_ConstantInt(CI&: Equal)))
3337 return false;
3338 CmpPredicate PredB;
3339 Value *LHS2, *RHS2;
3340 if (!match(V: UnequalVal, P: m_Select(C: m_ICmp(Pred&: PredB, L: m_Value(V&: LHS2), R: m_Value(V&: RHS2)),
3341 L: m_ConstantInt(CI&: Less), R: m_ConstantInt(CI&: Greater))))
3342 return false;
3343 // We can get predicate mismatch here, so canonicalize if possible:
3344 // First, ensure that 'LHS' match.
3345 if (LHS2 != LHS) {
3346 // x sgt y <--> y slt x
3347 std::swap(a&: LHS2, b&: RHS2);
3348 PredB = ICmpInst::getSwappedPredicate(pred: PredB);
3349 }
3350 if (LHS2 != LHS)
3351 return false;
3352 // We also need to canonicalize 'RHS'.
3353 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(Val: RHS2)) {
3354 // x sgt C-1 <--> x sge C <--> not(x slt C)
3355 auto FlippedStrictness =
3356 getFlippedStrictnessPredicateAndConstant(Pred: PredB, C: cast<Constant>(Val: RHS2));
3357 if (!FlippedStrictness)
3358 return false;
3359 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE &&
3360 "basic correctness failure");
3361 RHS2 = FlippedStrictness->second;
3362 // And kind-of perform the result swap.
3363 std::swap(a&: Less, b&: Greater);
3364 PredB = ICmpInst::ICMP_SLT;
3365 }
3366 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
3367}
3368
3369Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
3370 SelectInst *Select,
3371 ConstantInt *C) {
3372
3373 assert(C && "Cmp RHS should be a constant int!");
3374 // If we're testing a constant value against the result of a three way
3375 // comparison, the result can be expressed directly in terms of the
3376 // original values being compared. Note: We could possibly be more
3377 // aggressive here and remove the hasOneUse test. The original select is
3378 // really likely to simplify or sink when we remove a test of the result.
3379 Value *OrigLHS, *OrigRHS;
3380 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
3381 if (Cmp.hasOneUse() &&
3382 matchThreeWayIntCompare(SI: Select, LHS&: OrigLHS, RHS&: OrigRHS, Less&: C1LessThan, Equal&: C2Equal,
3383 Greater&: C3GreaterThan)) {
3384 assert(C1LessThan && C2Equal && C3GreaterThan);
3385
3386 bool TrueWhenLessThan = ICmpInst::compare(
3387 LHS: C1LessThan->getValue(), RHS: C->getValue(), Pred: Cmp.getPredicate());
3388 bool TrueWhenEqual = ICmpInst::compare(LHS: C2Equal->getValue(), RHS: C->getValue(),
3389 Pred: Cmp.getPredicate());
3390 bool TrueWhenGreaterThan = ICmpInst::compare(
3391 LHS: C3GreaterThan->getValue(), RHS: C->getValue(), Pred: Cmp.getPredicate());
3392
3393 // This generates the new instruction that will replace the original Cmp
3394 // Instruction. Instead of enumerating the various combinations when
3395 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
3396 // false, we rely on chaining of ORs and future passes of InstCombine to
3397 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
3398
3399 // When none of the three constants satisfy the predicate for the RHS (C),
3400 // the entire original Cmp can be simplified to a false.
3401 Value *Cond = Builder.getFalse();
3402 if (TrueWhenLessThan)
3403 Cond = Builder.CreateOr(
3404 LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_SLT, LHS: OrigLHS, RHS: OrigRHS));
3405 if (TrueWhenEqual)
3406 Cond = Builder.CreateOr(
3407 LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_EQ, LHS: OrigLHS, RHS: OrigRHS));
3408 if (TrueWhenGreaterThan)
3409 Cond = Builder.CreateOr(
3410 LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_SGT, LHS: OrigLHS, RHS: OrigRHS));
3411
3412 return replaceInstUsesWith(I&: Cmp, V: Cond);
3413 }
3414 return nullptr;
3415}
3416
3417Instruction *InstCombinerImpl::foldICmpBitCast(ICmpInst &Cmp) {
3418 auto *Bitcast = dyn_cast<BitCastInst>(Val: Cmp.getOperand(i_nocapture: 0));
3419 if (!Bitcast)
3420 return nullptr;
3421
3422 ICmpInst::Predicate Pred = Cmp.getPredicate();
3423 Value *Op1 = Cmp.getOperand(i_nocapture: 1);
3424 Value *BCSrcOp = Bitcast->getOperand(i_nocapture: 0);
3425 Type *SrcType = Bitcast->getSrcTy();
3426 Type *DstType = Bitcast->getType();
3427
3428 // Make sure the bitcast doesn't change between scalar and vector and
3429 // doesn't change the number of vector elements.
3430 if (SrcType->isVectorTy() == DstType->isVectorTy() &&
3431 SrcType->getScalarSizeInBits() == DstType->getScalarSizeInBits()) {
3432 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
3433 Value *X;
3434 if (match(V: BCSrcOp, P: m_SIToFP(Op: m_Value(V&: X)))) {
3435 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
3436 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
3437 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
3438 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
3439 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
3440 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
3441 match(V: Op1, P: m_Zero()))
3442 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType()));
3443
3444 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
3445 if (Pred == ICmpInst::ICMP_SLT && match(V: Op1, P: m_One()))
3446 return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: 1));
3447
3448 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
3449 if (Pred == ICmpInst::ICMP_SGT && match(V: Op1, P: m_AllOnes()))
3450 return new ICmpInst(Pred, X,
3451 ConstantInt::getAllOnesValue(Ty: X->getType()));
3452 }
3453
3454 // Zero-equality checks are preserved through unsigned floating-point casts:
3455 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
3456 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
3457 if (match(V: BCSrcOp, P: m_UIToFP(Op: m_Value(V&: X))))
3458 if (Cmp.isEquality() && match(V: Op1, P: m_Zero()))
3459 return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType()));
3460
3461 const APInt *C;
3462 bool TrueIfSigned;
3463 if (match(V: Op1, P: m_APInt(Res&: C)) && Bitcast->hasOneUse()) {
3464 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
3465 // the FP extend/truncate because that cast does not change the sign-bit.
3466 // This is true for all standard IEEE-754 types and the X86 80-bit type.
3467 // The sign-bit is always the most significant bit in those types.
3468 if (isSignBitCheck(Pred, RHS: *C, TrueIfSigned) &&
3469 (match(V: BCSrcOp, P: m_FPExt(Op: m_Value(V&: X))) ||
3470 match(V: BCSrcOp, P: m_FPTrunc(Op: m_Value(V&: X))))) {
3471 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
3472 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
3473 Type *XType = X->getType();
3474
3475 // We can't currently handle Power style floating point operations here.
3476 if (!(XType->isPPC_FP128Ty() || SrcType->isPPC_FP128Ty())) {
3477 Type *NewType = Builder.getIntNTy(N: XType->getScalarSizeInBits());
3478 if (auto *XVTy = dyn_cast<VectorType>(Val: XType))
3479 NewType = VectorType::get(ElementType: NewType, EC: XVTy->getElementCount());
3480 Value *NewBitcast = Builder.CreateBitCast(V: X, DestTy: NewType);
3481 if (TrueIfSigned)
3482 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
3483 ConstantInt::getNullValue(Ty: NewType));
3484 else
3485 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
3486 ConstantInt::getAllOnesValue(Ty: NewType));
3487 }
3488 }
3489
3490 // icmp eq/ne (bitcast X to int), special fp -> llvm.is.fpclass(X, class)
3491 Type *FPType = SrcType->getScalarType();
3492 if (!Cmp.getParent()->getParent()->hasFnAttribute(
3493 Kind: Attribute::NoImplicitFloat) &&
3494 Cmp.isEquality() && FPType->isIEEELikeFPTy()) {
3495 FPClassTest Mask = APFloat(FPType->getFltSemantics(), *C).classify();
3496 if (Mask & (fcInf | fcZero)) {
3497 if (Pred == ICmpInst::ICMP_NE)
3498 Mask = ~Mask;
3499 return replaceInstUsesWith(I&: Cmp,
3500 V: Builder.createIsFPClass(FPNum: BCSrcOp, Test: Mask));
3501 }
3502 }
3503 }
3504 }
3505
3506 const APInt *C;
3507 if (!match(V: Cmp.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) || !DstType->isIntegerTy() ||
3508 !SrcType->isIntOrIntVectorTy())
3509 return nullptr;
3510
3511 // If this is checking if all elements of a vector compare are set or not,
3512 // invert the casted vector equality compare and test if all compare
3513 // elements are clear or not. Compare against zero is generally easier for
3514 // analysis and codegen.
3515 // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
3516 // Example: are all elements equal? --> are zero elements not equal?
3517 // TODO: Try harder to reduce compare of 2 freely invertible operands?
3518 if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse()) {
3519 if (Value *NotBCSrcOp =
3520 getFreelyInverted(V: BCSrcOp, WillInvertAllUses: BCSrcOp->hasOneUse(), Builder: &Builder)) {
3521 Value *Cast = Builder.CreateBitCast(V: NotBCSrcOp, DestTy: DstType);
3522 return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(Ty: DstType));
3523 }
3524 }
3525
3526 // If this is checking if all elements of an extended vector are clear or not,
3527 // compare in a narrow type to eliminate the extend:
3528 // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
3529 Value *X;
3530 if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() &&
3531 match(V: BCSrcOp, P: m_ZExtOrSExt(Op: m_Value(V&: X)))) {
3532 if (auto *VecTy = dyn_cast<FixedVectorType>(Val: X->getType())) {
3533 Type *NewType = Builder.getIntNTy(N: VecTy->getPrimitiveSizeInBits());
3534 Value *NewCast = Builder.CreateBitCast(V: X, DestTy: NewType);
3535 return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(Ty: NewType));
3536 }
3537 }
3538
3539 // Folding: icmp <pred> iN X, C
3540 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
3541 // and C is a splat of a K-bit pattern
3542 // and SC is a constant vector = <C', C', C', ..., C'>
3543 // Into:
3544 // %E = extractelement <M x iK> %vec, i32 C'
3545 // icmp <pred> iK %E, trunc(C)
3546 Value *Vec;
3547 ArrayRef<int> Mask;
3548 if (match(V: BCSrcOp, P: m_Shuffle(v1: m_Value(V&: Vec), v2: m_Undef(), mask: m_Mask(Mask)))) {
3549 // Check whether every element of Mask is the same constant
3550 if (all_equal(Range&: Mask)) {
3551 auto *VecTy = cast<VectorType>(Val: SrcType);
3552 auto *EltTy = cast<IntegerType>(Val: VecTy->getElementType());
3553 if (C->isSplat(SplatSizeInBits: EltTy->getBitWidth())) {
3554 // Fold the icmp based on the value of C
3555 // If C is M copies of an iK sized bit pattern,
3556 // then:
3557 // => %E = extractelement <N x iK> %vec, i32 Elem
3558 // icmp <pred> iK %SplatVal, <pattern>
3559 Value *Elem = Builder.getInt32(C: Mask[0]);
3560 Value *Extract = Builder.CreateExtractElement(Vec, Idx: Elem);
3561 Value *NewC = ConstantInt::get(Ty: EltTy, V: C->trunc(width: EltTy->getBitWidth()));
3562 return new ICmpInst(Pred, Extract, NewC);
3563 }
3564 }
3565 }
3566 return nullptr;
3567}
3568
3569/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3570/// where X is some kind of instruction.
3571Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
3572 const APInt *C;
3573
3574 if (match(V: Cmp.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) {
3575 if (auto *BO = dyn_cast<BinaryOperator>(Val: Cmp.getOperand(i_nocapture: 0)))
3576 if (Instruction *I = foldICmpBinOpWithConstant(Cmp, BO, C: *C))
3577 return I;
3578
3579 if (auto *SI = dyn_cast<SelectInst>(Val: Cmp.getOperand(i_nocapture: 0)))
3580 // For now, we only support constant integers while folding the
3581 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3582 // similar to the cases handled by binary ops above.
3583 if (auto *ConstRHS = dyn_cast<ConstantInt>(Val: Cmp.getOperand(i_nocapture: 1)))
3584 if (Instruction *I = foldICmpSelectConstant(Cmp, Select: SI, C: ConstRHS))
3585 return I;
3586
3587 if (auto *TI = dyn_cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 0)))
3588 if (Instruction *I = foldICmpTruncConstant(Cmp, Trunc: TI, C: *C))
3589 return I;
3590
3591 if (auto *II = dyn_cast<IntrinsicInst>(Val: Cmp.getOperand(i_nocapture: 0)))
3592 if (Instruction *I = foldICmpIntrinsicWithConstant(ICI&: Cmp, II, C: *C))
3593 return I;
3594
3595 // (extractval ([s/u]subo X, Y), 0) == 0 --> X == Y
3596 // (extractval ([s/u]subo X, Y), 0) != 0 --> X != Y
3597 // TODO: This checks one-use, but that is not strictly necessary.
3598 Value *Cmp0 = Cmp.getOperand(i_nocapture: 0);
3599 Value *X, *Y;
3600 if (C->isZero() && Cmp.isEquality() && Cmp0->hasOneUse() &&
3601 (match(V: Cmp0,
3602 P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::ssub_with_overflow>(
3603 Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) ||
3604 match(V: Cmp0,
3605 P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::usub_with_overflow>(
3606 Op0: m_Value(V&: X), Op1: m_Value(V&: Y))))))
3607 return new ICmpInst(Cmp.getPredicate(), X, Y);
3608 }
3609
3610 if (match(V: Cmp.getOperand(i_nocapture: 1), P: m_APIntAllowPoison(Res&: C)))
3611 return foldICmpInstWithConstantAllowPoison(Cmp, C: *C);
3612
3613 return nullptr;
3614}
3615
3616/// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3617/// icmp eq/ne BO, C.
3618Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
3619 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3620 // TODO: Some of these folds could work with arbitrary constants, but this
3621 // function is limited to scalar and vector splat constants.
3622 if (!Cmp.isEquality())
3623 return nullptr;
3624
3625 ICmpInst::Predicate Pred = Cmp.getPredicate();
3626 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3627 Constant *RHS = cast<Constant>(Val: Cmp.getOperand(i_nocapture: 1));
3628 Value *BOp0 = BO->getOperand(i_nocapture: 0), *BOp1 = BO->getOperand(i_nocapture: 1);
3629
3630 switch (BO->getOpcode()) {
3631 case Instruction::SRem:
3632 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3633 if (C.isZero() && BO->hasOneUse()) {
3634 const APInt *BOC;
3635 if (match(V: BOp1, P: m_APInt(Res&: BOC)) && BOC->sgt(RHS: 1) && BOC->isPowerOf2()) {
3636 Value *NewRem = Builder.CreateURem(LHS: BOp0, RHS: BOp1, Name: BO->getName());
3637 return new ICmpInst(Pred, NewRem,
3638 Constant::getNullValue(Ty: BO->getType()));
3639 }
3640 }
3641 break;
3642 case Instruction::Add: {
3643 // (A + C2) == C --> A == (C - C2)
3644 // (A + C2) != C --> A != (C - C2)
3645 // TODO: Remove the one-use limitation? See discussion in D58633.
3646 if (Constant *C2 = dyn_cast<Constant>(Val: BOp1)) {
3647 if (BO->hasOneUse())
3648 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(C1: RHS, C2));
3649 } else if (C.isZero()) {
3650 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3651 // efficiently invertible, or if the add has just this one use.
3652 if (Value *NegVal = dyn_castNegVal(V: BOp1))
3653 return new ICmpInst(Pred, BOp0, NegVal);
3654 if (Value *NegVal = dyn_castNegVal(V: BOp0))
3655 return new ICmpInst(Pred, NegVal, BOp1);
3656 if (BO->hasOneUse()) {
3657 // (add nuw A, B) != 0 -> (or A, B) != 0
3658 if (match(V: BO, P: m_NUWAdd(L: m_Value(), R: m_Value()))) {
3659 Value *Or = Builder.CreateOr(LHS: BOp0, RHS: BOp1);
3660 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty: BO->getType()));
3661 }
3662 Value *Neg = Builder.CreateNeg(V: BOp1);
3663 Neg->takeName(V: BO);
3664 return new ICmpInst(Pred, BOp0, Neg);
3665 }
3666 }
3667 break;
3668 }
3669 case Instruction::Xor:
3670 if (Constant *BOC = dyn_cast<Constant>(Val: BOp1)) {
3671 // For the xor case, we can xor two constants together, eliminating
3672 // the explicit xor.
3673 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(C1: RHS, C2: BOC));
3674 } else if (C.isZero()) {
3675 // Replace ((xor A, B) != 0) with (A != B)
3676 return new ICmpInst(Pred, BOp0, BOp1);
3677 }
3678 break;
3679 case Instruction::Or: {
3680 const APInt *BOC;
3681 if (match(V: BOp1, P: m_APInt(Res&: BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3682 // Comparing if all bits outside of a constant mask are set?
3683 // Replace (X | C) == -1 with (X & ~C) == ~C.
3684 // This removes the -1 constant.
3685 Constant *NotBOC = ConstantExpr::getNot(C: cast<Constant>(Val: BOp1));
3686 Value *And = Builder.CreateAnd(LHS: BOp0, RHS: NotBOC);
3687 return new ICmpInst(Pred, And, NotBOC);
3688 }
3689 // (icmp eq (or (select cond, 0, NonZero), Other), 0)
3690 // -> (and cond, (icmp eq Other, 0))
3691 // (icmp ne (or (select cond, NonZero, 0), Other), 0)
3692 // -> (or cond, (icmp ne Other, 0))
3693 Value *Cond, *TV, *FV, *Other, *Sel;
3694 if (C.isZero() &&
3695 match(V: BO,
3696 P: m_OneUse(SubPattern: m_c_Or(L: m_CombineAnd(L: m_Value(V&: Sel),
3697 R: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TV),
3698 R: m_Value(V&: FV))),
3699 R: m_Value(V&: Other)))) &&
3700 Cond->getType() == Cmp.getType()) {
3701 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
3702 // Easy case is if eq/ne matches whether 0 is trueval/falseval.
3703 if (Pred == ICmpInst::ICMP_EQ
3704 ? (match(V: TV, P: m_Zero()) && isKnownNonZero(V: FV, Q))
3705 : (match(V: FV, P: m_Zero()) && isKnownNonZero(V: TV, Q))) {
3706 Value *Cmp = Builder.CreateICmp(
3707 P: Pred, LHS: Other, RHS: Constant::getNullValue(Ty: Other->getType()));
3708 return BinaryOperator::Create(
3709 Op: Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, S1: Cmp,
3710 S2: Cond);
3711 }
3712 // Harder case is if eq/ne matches whether 0 is falseval/trueval. In this
3713 // case we need to invert the select condition so we need to be careful to
3714 // avoid creating extra instructions.
3715 // (icmp ne (or (select cond, 0, NonZero), Other), 0)
3716 // -> (or (not cond), (icmp ne Other, 0))
3717 // (icmp eq (or (select cond, NonZero, 0), Other), 0)
3718 // -> (and (not cond), (icmp eq Other, 0))
3719 //
3720 // Only do this if the inner select has one use, in which case we are
3721 // replacing `select` with `(not cond)`. Otherwise, we will create more
3722 // uses. NB: Trying to freely invert cond doesn't make sense here, as if
3723 // cond was freely invertable, the select arms would have been inverted.
3724 if (Sel->hasOneUse() &&
3725 (Pred == ICmpInst::ICMP_EQ
3726 ? (match(V: FV, P: m_Zero()) && isKnownNonZero(V: TV, Q))
3727 : (match(V: TV, P: m_Zero()) && isKnownNonZero(V: FV, Q)))) {
3728 Value *NotCond = Builder.CreateNot(V: Cond);
3729 Value *Cmp = Builder.CreateICmp(
3730 P: Pred, LHS: Other, RHS: Constant::getNullValue(Ty: Other->getType()));
3731 return BinaryOperator::Create(
3732 Op: Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, S1: Cmp,
3733 S2: NotCond);
3734 }
3735 }
3736 break;
3737 }
3738 case Instruction::UDiv:
3739 case Instruction::SDiv:
3740 if (BO->isExact()) {
3741 // div exact X, Y eq/ne 0 -> X eq/ne 0
3742 // div exact X, Y eq/ne 1 -> X eq/ne Y
3743 // div exact X, Y eq/ne C ->
3744 // if Y * C never-overflow && OneUse:
3745 // -> Y * C eq/ne X
3746 if (C.isZero())
3747 return new ICmpInst(Pred, BOp0, Constant::getNullValue(Ty: BO->getType()));
3748 else if (C.isOne())
3749 return new ICmpInst(Pred, BOp0, BOp1);
3750 else if (BO->hasOneUse()) {
3751 OverflowResult OR = computeOverflow(
3752 BinaryOp: Instruction::Mul, IsSigned: BO->getOpcode() == Instruction::SDiv, LHS: BOp1,
3753 RHS: Cmp.getOperand(i_nocapture: 1), CxtI: BO);
3754 if (OR == OverflowResult::NeverOverflows) {
3755 Value *YC =
3756 Builder.CreateMul(LHS: BOp1, RHS: ConstantInt::get(Ty: BO->getType(), V: C));
3757 return new ICmpInst(Pred, YC, BOp0);
3758 }
3759 }
3760 }
3761 if (BO->getOpcode() == Instruction::UDiv && C.isZero()) {
3762 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3763 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3764 return new ICmpInst(NewPred, BOp1, BOp0);
3765 }
3766 break;
3767 default:
3768 break;
3769 }
3770 return nullptr;
3771}
3772
3773static Instruction *foldCtpopPow2Test(ICmpInst &I, IntrinsicInst *CtpopLhs,
3774 const APInt &CRhs,
3775 InstCombiner::BuilderTy &Builder,
3776 const SimplifyQuery &Q) {
3777 assert(CtpopLhs->getIntrinsicID() == Intrinsic::ctpop &&
3778 "Non-ctpop intrin in ctpop fold");
3779 if (!CtpopLhs->hasOneUse())
3780 return nullptr;
3781
3782 // Power of 2 test:
3783 // isPow2OrZero : ctpop(X) u< 2
3784 // isPow2 : ctpop(X) == 1
3785 // NotPow2OrZero: ctpop(X) u> 1
3786 // NotPow2 : ctpop(X) != 1
3787 // If we know any bit of X can be folded to:
3788 // IsPow2 : X & (~Bit) == 0
3789 // NotPow2 : X & (~Bit) != 0
3790 const ICmpInst::Predicate Pred = I.getPredicate();
3791 if (((I.isEquality() || Pred == ICmpInst::ICMP_UGT) && CRhs == 1) ||
3792 (Pred == ICmpInst::ICMP_ULT && CRhs == 2)) {
3793 Value *Op = CtpopLhs->getArgOperand(i: 0);
3794 KnownBits OpKnown = computeKnownBits(V: Op, DL: Q.DL, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT);
3795 // No need to check for count > 1, that should be already constant folded.
3796 if (OpKnown.countMinPopulation() == 1) {
3797 Value *And = Builder.CreateAnd(
3798 LHS: Op, RHS: Constant::getIntegerValue(Ty: Op->getType(), V: ~(OpKnown.One)));
3799 return new ICmpInst(
3800 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_ULT)
3801 ? ICmpInst::ICMP_EQ
3802 : ICmpInst::ICMP_NE,
3803 And, Constant::getNullValue(Ty: Op->getType()));
3804 }
3805 }
3806
3807 return nullptr;
3808}
3809
3810/// Fold an equality icmp with LLVM intrinsic and constant operand.
3811Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3812 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3813 Type *Ty = II->getType();
3814 unsigned BitWidth = C.getBitWidth();
3815 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3816
3817 switch (II->getIntrinsicID()) {
3818 case Intrinsic::abs:
3819 // abs(A) == 0 -> A == 0
3820 // abs(A) == INT_MIN -> A == INT_MIN
3821 if (C.isZero() || C.isMinSignedValue())
3822 return new ICmpInst(Pred, II->getArgOperand(i: 0), ConstantInt::get(Ty, V: C));
3823 break;
3824
3825 case Intrinsic::bswap:
3826 // bswap(A) == C -> A == bswap(C)
3827 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3828 ConstantInt::get(Ty, V: C.byteSwap()));
3829
3830 case Intrinsic::bitreverse:
3831 // bitreverse(A) == C -> A == bitreverse(C)
3832 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3833 ConstantInt::get(Ty, V: C.reverseBits()));
3834
3835 case Intrinsic::ctlz:
3836 case Intrinsic::cttz: {
3837 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3838 if (C == BitWidth)
3839 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3840 ConstantInt::getNullValue(Ty));
3841
3842 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3843 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3844 // Limit to one use to ensure we don't increase instruction count.
3845 unsigned Num = C.getLimitedValue(Limit: BitWidth);
3846 if (Num != BitWidth && II->hasOneUse()) {
3847 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3848 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: Num + 1)
3849 : APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: Num + 1);
3850 APInt Mask2 = IsTrailing
3851 ? APInt::getOneBitSet(numBits: BitWidth, BitNo: Num)
3852 : APInt::getOneBitSet(numBits: BitWidth, BitNo: BitWidth - Num - 1);
3853 return new ICmpInst(Pred, Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask1),
3854 ConstantInt::get(Ty, V: Mask2));
3855 }
3856 break;
3857 }
3858
3859 case Intrinsic::ctpop: {
3860 // popcount(A) == 0 -> A == 0 and likewise for !=
3861 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3862 bool IsZero = C.isZero();
3863 if (IsZero || C == BitWidth)
3864 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3865 IsZero ? Constant::getNullValue(Ty)
3866 : Constant::getAllOnesValue(Ty));
3867
3868 break;
3869 }
3870
3871 case Intrinsic::fshl:
3872 case Intrinsic::fshr:
3873 if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) {
3874 const APInt *RotAmtC;
3875 // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3876 // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3877 if (match(V: II->getArgOperand(i: 2), P: m_APInt(Res&: RotAmtC)))
3878 return new ICmpInst(Pred, II->getArgOperand(i: 0),
3879 II->getIntrinsicID() == Intrinsic::fshl
3880 ? ConstantInt::get(Ty, V: C.rotr(rotateAmt: *RotAmtC))
3881 : ConstantInt::get(Ty, V: C.rotl(rotateAmt: *RotAmtC)));
3882 }
3883 break;
3884
3885 case Intrinsic::umax:
3886 case Intrinsic::uadd_sat: {
3887 // uadd.sat(a, b) == 0 -> (a | b) == 0
3888 // umax(a, b) == 0 -> (a | b) == 0
3889 if (C.isZero() && II->hasOneUse()) {
3890 Value *Or = Builder.CreateOr(LHS: II->getArgOperand(i: 0), RHS: II->getArgOperand(i: 1));
3891 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3892 }
3893 break;
3894 }
3895
3896 case Intrinsic::ssub_sat:
3897 // ssub.sat(a, b) == 0 -> a == b
3898 //
3899 // Note this doesn't work for ssub.sat.i1 because ssub.sat.i1 0, -1 = 0
3900 // (because 1 saturates to 0). Just skip the optimization for i1.
3901 if (C.isZero() && II->getType()->getScalarSizeInBits() > 1)
3902 return new ICmpInst(Pred, II->getArgOperand(i: 0), II->getArgOperand(i: 1));
3903 break;
3904 case Intrinsic::usub_sat: {
3905 // usub.sat(a, b) == 0 -> a <= b
3906 if (C.isZero()) {
3907 ICmpInst::Predicate NewPred =
3908 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3909 return new ICmpInst(NewPred, II->getArgOperand(i: 0), II->getArgOperand(i: 1));
3910 }
3911 break;
3912 }
3913 default:
3914 break;
3915 }
3916
3917 return nullptr;
3918}
3919
3920/// Fold an icmp with LLVM intrinsics
3921static Instruction *
3922foldICmpIntrinsicWithIntrinsic(ICmpInst &Cmp,
3923 InstCombiner::BuilderTy &Builder) {
3924 assert(Cmp.isEquality());
3925
3926 ICmpInst::Predicate Pred = Cmp.getPredicate();
3927 Value *Op0 = Cmp.getOperand(i_nocapture: 0);
3928 Value *Op1 = Cmp.getOperand(i_nocapture: 1);
3929 const auto *IIOp0 = dyn_cast<IntrinsicInst>(Val: Op0);
3930 const auto *IIOp1 = dyn_cast<IntrinsicInst>(Val: Op1);
3931 if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3932 return nullptr;
3933
3934 switch (IIOp0->getIntrinsicID()) {
3935 case Intrinsic::bswap:
3936 case Intrinsic::bitreverse:
3937 // If both operands are byte-swapped or bit-reversed, just compare the
3938 // original values.
3939 return new ICmpInst(Pred, IIOp0->getOperand(i_nocapture: 0), IIOp1->getOperand(i_nocapture: 0));
3940 case Intrinsic::fshl:
3941 case Intrinsic::fshr: {
3942 // If both operands are rotated by same amount, just compare the
3943 // original values.
3944 if (IIOp0->getOperand(i_nocapture: 0) != IIOp0->getOperand(i_nocapture: 1))
3945 break;
3946 if (IIOp1->getOperand(i_nocapture: 0) != IIOp1->getOperand(i_nocapture: 1))
3947 break;
3948 if (IIOp0->getOperand(i_nocapture: 2) == IIOp1->getOperand(i_nocapture: 2))
3949 return new ICmpInst(Pred, IIOp0->getOperand(i_nocapture: 0), IIOp1->getOperand(i_nocapture: 0));
3950
3951 // rotate(X, AmtX) == rotate(Y, AmtY)
3952 // -> rotate(X, AmtX - AmtY) == Y
3953 // Do this if either both rotates have one use or if only one has one use
3954 // and AmtX/AmtY are constants.
3955 unsigned OneUses = IIOp0->hasOneUse() + IIOp1->hasOneUse();
3956 if (OneUses == 2 ||
3957 (OneUses == 1 && match(V: IIOp0->getOperand(i_nocapture: 2), P: m_ImmConstant()) &&
3958 match(V: IIOp1->getOperand(i_nocapture: 2), P: m_ImmConstant()))) {
3959 Value *SubAmt =
3960 Builder.CreateSub(LHS: IIOp0->getOperand(i_nocapture: 2), RHS: IIOp1->getOperand(i_nocapture: 2));
3961 Value *CombinedRotate = Builder.CreateIntrinsic(
3962 RetTy: Op0->getType(), ID: IIOp0->getIntrinsicID(),
3963 Args: {IIOp0->getOperand(i_nocapture: 0), IIOp0->getOperand(i_nocapture: 0), SubAmt});
3964 return new ICmpInst(Pred, IIOp1->getOperand(i_nocapture: 0), CombinedRotate);
3965 }
3966 } break;
3967 default:
3968 break;
3969 }
3970
3971 return nullptr;
3972}
3973
3974/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3975/// where X is some kind of instruction and C is AllowPoison.
3976/// TODO: Move more folds which allow poison to this function.
3977Instruction *
3978InstCombinerImpl::foldICmpInstWithConstantAllowPoison(ICmpInst &Cmp,
3979 const APInt &C) {
3980 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3981 if (auto *II = dyn_cast<IntrinsicInst>(Val: Cmp.getOperand(i_nocapture: 0))) {
3982 switch (II->getIntrinsicID()) {
3983 default:
3984 break;
3985 case Intrinsic::fshl:
3986 case Intrinsic::fshr:
3987 if (Cmp.isEquality() && II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) {
3988 // (rot X, ?) == 0/-1 --> X == 0/-1
3989 if (C.isZero() || C.isAllOnes())
3990 return new ICmpInst(Pred, II->getArgOperand(i: 0), Cmp.getOperand(i_nocapture: 1));
3991 }
3992 break;
3993 }
3994 }
3995
3996 return nullptr;
3997}
3998
3999/// Fold an icmp with BinaryOp and constant operand: icmp Pred BO, C.
4000Instruction *InstCombinerImpl::foldICmpBinOpWithConstant(ICmpInst &Cmp,
4001 BinaryOperator *BO,
4002 const APInt &C) {
4003 switch (BO->getOpcode()) {
4004 case Instruction::Xor:
4005 if (Instruction *I = foldICmpXorConstant(Cmp, Xor: BO, C))
4006 return I;
4007 break;
4008 case Instruction::And:
4009 if (Instruction *I = foldICmpAndConstant(Cmp, And: BO, C))
4010 return I;
4011 break;
4012 case Instruction::Or:
4013 if (Instruction *I = foldICmpOrConstant(Cmp, Or: BO, C))
4014 return I;
4015 break;
4016 case Instruction::Mul:
4017 if (Instruction *I = foldICmpMulConstant(Cmp, Mul: BO, C))
4018 return I;
4019 break;
4020 case Instruction::Shl:
4021 if (Instruction *I = foldICmpShlConstant(Cmp, Shl: BO, C))
4022 return I;
4023 break;
4024 case Instruction::LShr:
4025 case Instruction::AShr:
4026 if (Instruction *I = foldICmpShrConstant(Cmp, Shr: BO, C))
4027 return I;
4028 break;
4029 case Instruction::SRem:
4030 if (Instruction *I = foldICmpSRemConstant(Cmp, SRem: BO, C))
4031 return I;
4032 break;
4033 case Instruction::UDiv:
4034 if (Instruction *I = foldICmpUDivConstant(Cmp, UDiv: BO, C))
4035 return I;
4036 [[fallthrough]];
4037 case Instruction::SDiv:
4038 if (Instruction *I = foldICmpDivConstant(Cmp, Div: BO, C))
4039 return I;
4040 break;
4041 case Instruction::Sub:
4042 if (Instruction *I = foldICmpSubConstant(Cmp, Sub: BO, C))
4043 return I;
4044 break;
4045 case Instruction::Add:
4046 if (Instruction *I = foldICmpAddConstant(Cmp, Add: BO, C))
4047 return I;
4048 break;
4049 default:
4050 break;
4051 }
4052
4053 // TODO: These folds could be refactored to be part of the above calls.
4054 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C))
4055 return I;
4056
4057 // Fall back to handling `icmp pred (select A ? C1 : C2) binop (select B ? C3
4058 // : C4), C5` pattern, by computing a truth table of the four constant
4059 // variants.
4060 return foldICmpBinOpWithConstantViaTruthTable(Cmp, BO, C);
4061}
4062
4063static Instruction *
4064foldICmpUSubSatOrUAddSatWithConstant(CmpPredicate Pred, SaturatingInst *II,
4065 const APInt &C,
4066 InstCombiner::BuilderTy &Builder) {
4067 // This transform may end up producing more than one instruction for the
4068 // intrinsic, so limit it to one user of the intrinsic.
4069 if (!II->hasOneUse())
4070 return nullptr;
4071
4072 // Let Y = [add/sub]_sat(X, C) pred C2
4073 // SatVal = The saturating value for the operation
4074 // WillWrap = Whether or not the operation will underflow / overflow
4075 // => Y = (WillWrap ? SatVal : (X binop C)) pred C2
4076 // => Y = WillWrap ? (SatVal pred C2) : ((X binop C) pred C2)
4077 //
4078 // When (SatVal pred C2) is true, then
4079 // Y = WillWrap ? true : ((X binop C) pred C2)
4080 // => Y = WillWrap || ((X binop C) pred C2)
4081 // else
4082 // Y = WillWrap ? false : ((X binop C) pred C2)
4083 // => Y = !WillWrap ? ((X binop C) pred C2) : false
4084 // => Y = !WillWrap && ((X binop C) pred C2)
4085 Value *Op0 = II->getOperand(i_nocapture: 0);
4086 Value *Op1 = II->getOperand(i_nocapture: 1);
4087
4088 const APInt *COp1;
4089 // This transform only works when the intrinsic has an integral constant or
4090 // splat vector as the second operand.
4091 if (!match(V: Op1, P: m_APInt(Res&: COp1)))
4092 return nullptr;
4093
4094 APInt SatVal;
4095 switch (II->getIntrinsicID()) {
4096 default:
4097 llvm_unreachable(
4098 "This function only works with usub_sat and uadd_sat for now!");
4099 case Intrinsic::uadd_sat:
4100 SatVal = APInt::getAllOnes(numBits: C.getBitWidth());
4101 break;
4102 case Intrinsic::usub_sat:
4103 SatVal = APInt::getZero(numBits: C.getBitWidth());
4104 break;
4105 }
4106
4107 // Check (SatVal pred C2)
4108 bool SatValCheck = ICmpInst::compare(LHS: SatVal, RHS: C, Pred);
4109
4110 // !WillWrap.
4111 ConstantRange C1 = ConstantRange::makeExactNoWrapRegion(
4112 BinOp: II->getBinaryOp(), Other: *COp1, NoWrapKind: II->getNoWrapKind());
4113
4114 // WillWrap.
4115 if (SatValCheck)
4116 C1 = C1.inverse();
4117
4118 ConstantRange C2 = ConstantRange::makeExactICmpRegion(Pred, Other: C);
4119 if (II->getBinaryOp() == Instruction::Add)
4120 C2 = C2.sub(Other: *COp1);
4121 else
4122 C2 = C2.add(Other: *COp1);
4123
4124 Instruction::BinaryOps CombiningOp =
4125 SatValCheck ? Instruction::BinaryOps::Or : Instruction::BinaryOps::And;
4126
4127 std::optional<ConstantRange> Combination;
4128 if (CombiningOp == Instruction::BinaryOps::Or)
4129 Combination = C1.exactUnionWith(CR: C2);
4130 else /* CombiningOp == Instruction::BinaryOps::And */
4131 Combination = C1.exactIntersectWith(CR: C2);
4132
4133 if (!Combination)
4134 return nullptr;
4135
4136 CmpInst::Predicate EquivPred;
4137 APInt EquivInt;
4138 APInt EquivOffset;
4139
4140 Combination->getEquivalentICmp(Pred&: EquivPred, RHS&: EquivInt, Offset&: EquivOffset);
4141
4142 return new ICmpInst(
4143 EquivPred,
4144 Builder.CreateAdd(LHS: Op0, RHS: ConstantInt::get(Ty: Op1->getType(), V: EquivOffset)),
4145 ConstantInt::get(Ty: Op1->getType(), V: EquivInt));
4146}
4147
4148static Instruction *
4149foldICmpOfCmpIntrinsicWithConstant(CmpPredicate Pred, IntrinsicInst *I,
4150 const APInt &C,
4151 InstCombiner::BuilderTy &Builder) {
4152 std::optional<ICmpInst::Predicate> NewPredicate = std::nullopt;
4153 switch (Pred) {
4154 case ICmpInst::ICMP_EQ:
4155 case ICmpInst::ICMP_NE:
4156 if (C.isZero())
4157 NewPredicate = Pred;
4158 else if (C.isOne())
4159 NewPredicate =
4160 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE;
4161 else if (C.isAllOnes())
4162 NewPredicate =
4163 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
4164 break;
4165
4166 case ICmpInst::ICMP_SGT:
4167 if (C.isAllOnes())
4168 NewPredicate = ICmpInst::ICMP_UGE;
4169 else if (C.isZero())
4170 NewPredicate = ICmpInst::ICMP_UGT;
4171 break;
4172
4173 case ICmpInst::ICMP_SLT:
4174 if (C.isZero())
4175 NewPredicate = ICmpInst::ICMP_ULT;
4176 else if (C.isOne())
4177 NewPredicate = ICmpInst::ICMP_ULE;
4178 break;
4179
4180 case ICmpInst::ICMP_ULT:
4181 if (C.ugt(RHS: 1))
4182 NewPredicate = ICmpInst::ICMP_UGE;
4183 break;
4184
4185 case ICmpInst::ICMP_UGT:
4186 if (!C.isZero() && !C.isAllOnes())
4187 NewPredicate = ICmpInst::ICMP_ULT;
4188 break;
4189
4190 default:
4191 break;
4192 }
4193
4194 if (!NewPredicate)
4195 return nullptr;
4196
4197 if (I->getIntrinsicID() == Intrinsic::scmp)
4198 NewPredicate = ICmpInst::getSignedPredicate(Pred: *NewPredicate);
4199 Value *LHS = I->getOperand(i_nocapture: 0);
4200 Value *RHS = I->getOperand(i_nocapture: 1);
4201 return new ICmpInst(*NewPredicate, LHS, RHS);
4202}
4203
4204/// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
4205Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
4206 IntrinsicInst *II,
4207 const APInt &C) {
4208 ICmpInst::Predicate Pred = Cmp.getPredicate();
4209
4210 // Handle folds that apply for any kind of icmp.
4211 switch (II->getIntrinsicID()) {
4212 default:
4213 break;
4214 case Intrinsic::uadd_sat:
4215 case Intrinsic::usub_sat:
4216 if (auto *Folded = foldICmpUSubSatOrUAddSatWithConstant(
4217 Pred, II: cast<SaturatingInst>(Val: II), C, Builder))
4218 return Folded;
4219 break;
4220 case Intrinsic::ctpop: {
4221 const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp);
4222 if (Instruction *R = foldCtpopPow2Test(I&: Cmp, CtpopLhs: II, CRhs: C, Builder, Q))
4223 return R;
4224 } break;
4225 case Intrinsic::scmp:
4226 case Intrinsic::ucmp:
4227 if (auto *Folded = foldICmpOfCmpIntrinsicWithConstant(Pred, I: II, C, Builder))
4228 return Folded;
4229 break;
4230 }
4231
4232 if (Cmp.isEquality())
4233 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
4234
4235 Type *Ty = II->getType();
4236 unsigned BitWidth = C.getBitWidth();
4237 switch (II->getIntrinsicID()) {
4238 case Intrinsic::ctpop: {
4239 // (ctpop X > BitWidth - 1) --> X == -1
4240 Value *X = II->getArgOperand(i: 0);
4241 if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
4242 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_EQ, S1: X,
4243 S2: ConstantInt::getAllOnesValue(Ty));
4244 // (ctpop X < BitWidth) --> X != -1
4245 if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
4246 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_NE, S1: X,
4247 S2: ConstantInt::getAllOnesValue(Ty));
4248 break;
4249 }
4250 case Intrinsic::ctlz: {
4251 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
4252 if (Pred == ICmpInst::ICMP_UGT && C.ult(RHS: BitWidth)) {
4253 unsigned Num = C.getLimitedValue();
4254 APInt Limit = APInt::getOneBitSet(numBits: BitWidth, BitNo: BitWidth - Num - 1);
4255 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_ULT,
4256 S1: II->getArgOperand(i: 0), S2: ConstantInt::get(Ty, V: Limit));
4257 }
4258
4259 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
4260 if (Pred == ICmpInst::ICMP_ULT && C.uge(RHS: 1) && C.ule(RHS: BitWidth)) {
4261 unsigned Num = C.getLimitedValue();
4262 APInt Limit = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - Num);
4263 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_UGT,
4264 S1: II->getArgOperand(i: 0), S2: ConstantInt::get(Ty, V: Limit));
4265 }
4266 break;
4267 }
4268 case Intrinsic::cttz: {
4269 // Limit to one use to ensure we don't increase instruction count.
4270 if (!II->hasOneUse())
4271 return nullptr;
4272
4273 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
4274 if (Pred == ICmpInst::ICMP_UGT && C.ult(RHS: BitWidth)) {
4275 APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: C.getLimitedValue() + 1);
4276 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_EQ,
4277 S1: Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask),
4278 S2: ConstantInt::getNullValue(Ty));
4279 }
4280
4281 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
4282 if (Pred == ICmpInst::ICMP_ULT && C.uge(RHS: 1) && C.ule(RHS: BitWidth)) {
4283 APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: C.getLimitedValue());
4284 return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_NE,
4285 S1: Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask),
4286 S2: ConstantInt::getNullValue(Ty));
4287 }
4288 break;
4289 }
4290 case Intrinsic::ssub_sat:
4291 // ssub.sat(a, b) spred 0 -> a spred b
4292 //
4293 // Note this doesn't work for ssub.sat.i1 because ssub.sat.i1 0, -1 = 0
4294 // (because 1 saturates to 0). Just skip the optimization for i1.
4295 if (ICmpInst::isSigned(predicate: Pred) && C.getBitWidth() > 1) {
4296 if (C.isZero())
4297 return new ICmpInst(Pred, II->getArgOperand(i: 0), II->getArgOperand(i: 1));
4298 // X s<= 0 is cannonicalized to X s< 1
4299 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
4300 return new ICmpInst(ICmpInst::ICMP_SLE, II->getArgOperand(i: 0),
4301 II->getArgOperand(i: 1));
4302 // X s>= 0 is cannonicalized to X s> -1
4303 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
4304 return new ICmpInst(ICmpInst::ICMP_SGE, II->getArgOperand(i: 0),
4305 II->getArgOperand(i: 1));
4306 }
4307 break;
4308 case Intrinsic::abs: {
4309 if (!II->hasOneUse())
4310 return nullptr;
4311
4312 Value *X = II->getArgOperand(i: 0);
4313 bool IsIntMinPoison =
4314 cast<ConstantInt>(Val: II->getArgOperand(i: 1))->getValue().isOne();
4315
4316 // If C >= 0:
4317 // abs(X) u> C --> X + C u> 2 * C
4318 if (Pred == CmpInst::ICMP_UGT && C.isNonNegative()) {
4319 return new ICmpInst(ICmpInst::ICMP_UGT,
4320 Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: C)),
4321 ConstantInt::get(Ty, V: 2 * C));
4322 }
4323
4324 // If abs(INT_MIN) is poison and C >= 1:
4325 // abs(X) u< C --> X + (C - 1) u<= 2 * (C - 1)
4326 if (IsIntMinPoison && Pred == CmpInst::ICMP_ULT && C.sge(RHS: 1)) {
4327 return new ICmpInst(ICmpInst::ICMP_ULE,
4328 Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: C - 1)),
4329 ConstantInt::get(Ty, V: 2 * (C - 1)));
4330 }
4331
4332 break;
4333 }
4334 default:
4335 break;
4336 }
4337
4338 return nullptr;
4339}
4340
4341/// Handle icmp with constant (but not simple integer constant) RHS.
4342Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
4343 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
4344 Constant *RHSC = dyn_cast<Constant>(Val: Op1);
4345 Instruction *LHSI = dyn_cast<Instruction>(Val: Op0);
4346 if (!RHSC || !LHSI)
4347 return nullptr;
4348
4349 switch (LHSI->getOpcode()) {
4350 case Instruction::IntToPtr:
4351 // icmp pred inttoptr(X), null -> icmp pred X, 0
4352 if (RHSC->isNullValue() &&
4353 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(i: 0)->getType())
4354 return new ICmpInst(
4355 I.getPredicate(), LHSI->getOperand(i: 0),
4356 Constant::getNullValue(Ty: LHSI->getOperand(i: 0)->getType()));
4357 break;
4358
4359 case Instruction::Load:
4360 // Try to optimize things like "A[i] > 4" to index computations.
4361 if (GetElementPtrInst *GEP =
4362 dyn_cast<GetElementPtrInst>(Val: LHSI->getOperand(i: 0)))
4363 if (Instruction *Res =
4364 foldCmpLoadFromIndexedGlobal(LI: cast<LoadInst>(Val: LHSI), GEP, ICI&: I))
4365 return Res;
4366 break;
4367 }
4368
4369 return nullptr;
4370}
4371
4372Instruction *InstCombinerImpl::foldSelectICmp(CmpPredicate Pred, SelectInst *SI,
4373 Value *RHS, const ICmpInst &I) {
4374 // Try to fold the comparison into the select arms, which will cause the
4375 // select to be converted into a logical and/or.
4376 auto SimplifyOp = [&](Value *Op, bool SelectCondIsTrue) -> Value * {
4377 if (Value *Res = simplifyICmpInst(Pred, LHS: Op, RHS, Q: SQ))
4378 return Res;
4379 if (std::optional<bool> Impl = isImpliedCondition(
4380 LHS: SI->getCondition(), RHSPred: Pred, RHSOp0: Op, RHSOp1: RHS, DL, LHSIsTrue: SelectCondIsTrue))
4381 return ConstantInt::get(Ty: I.getType(), V: *Impl);
4382 return nullptr;
4383 };
4384
4385 ConstantInt *CI = nullptr;
4386 Value *Op1 = SimplifyOp(SI->getOperand(i_nocapture: 1), true);
4387 if (Op1)
4388 CI = dyn_cast<ConstantInt>(Val: Op1);
4389
4390 Value *Op2 = SimplifyOp(SI->getOperand(i_nocapture: 2), false);
4391 if (Op2)
4392 CI = dyn_cast<ConstantInt>(Val: Op2);
4393
4394 auto Simplifies = [&](Value *Op, unsigned Idx) {
4395 // A comparison of ucmp/scmp with a constant will fold into an icmp.
4396 const APInt *Dummy;
4397 return Op ||
4398 (isa<CmpIntrinsic>(Val: SI->getOperand(i_nocapture: Idx)) &&
4399 SI->getOperand(i_nocapture: Idx)->hasOneUse() && match(V: RHS, P: m_APInt(Res&: Dummy)));
4400 };
4401
4402 // We only want to perform this transformation if it will not lead to
4403 // additional code. This is true if either both sides of the select
4404 // fold to a constant (in which case the icmp is replaced with a select
4405 // which will usually simplify) or this is the only user of the
4406 // select (in which case we are trading a select+icmp for a simpler
4407 // select+icmp) or all uses of the select can be replaced based on
4408 // dominance information ("Global cases").
4409 bool Transform = false;
4410 if (Op1 && Op2)
4411 Transform = true;
4412 else if (Simplifies(Op1, 1) || Simplifies(Op2, 2)) {
4413 // Local case
4414 if (SI->hasOneUse())
4415 Transform = true;
4416 // Global cases
4417 else if (CI && !CI->isZero())
4418 // When Op1 is constant try replacing select with second operand.
4419 // Otherwise Op2 is constant and try replacing select with first
4420 // operand.
4421 Transform = replacedSelectWithOperand(SI, Icmp: &I, SIOpd: Op1 ? 2 : 1);
4422 }
4423 if (Transform) {
4424 if (!Op1)
4425 Op1 = Builder.CreateICmp(P: Pred, LHS: SI->getOperand(i_nocapture: 1), RHS, Name: I.getName());
4426 if (!Op2)
4427 Op2 = Builder.CreateICmp(P: Pred, LHS: SI->getOperand(i_nocapture: 2), RHS, Name: I.getName());
4428 return SelectInst::Create(C: SI->getOperand(i_nocapture: 0), S1: Op1, S2: Op2, NameStr: "", InsertBefore: nullptr,
4429 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : SI);
4430 }
4431
4432 return nullptr;
4433}
4434
4435// Returns whether V is a Mask ((X + 1) & X == 0) or ~Mask (-Pow2OrZero)
4436static bool isMaskOrZero(const Value *V, bool Not, const SimplifyQuery &Q,
4437 unsigned Depth = 0) {
4438 if (Not ? match(V, P: m_NegatedPower2OrZero()) : match(V, P: m_LowBitMaskOrZero()))
4439 return true;
4440 if (V->getType()->getScalarSizeInBits() == 1)
4441 return true;
4442 if (Depth++ >= MaxAnalysisRecursionDepth)
4443 return false;
4444 Value *X;
4445 const Instruction *I = dyn_cast<Instruction>(Val: V);
4446 if (!I)
4447 return false;
4448 switch (I->getOpcode()) {
4449 case Instruction::ZExt:
4450 // ZExt(Mask) is a Mask.
4451 return !Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4452 case Instruction::SExt:
4453 // SExt(Mask) is a Mask.
4454 // SExt(~Mask) is a ~Mask.
4455 return isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4456 case Instruction::And:
4457 case Instruction::Or:
4458 // Mask0 | Mask1 is a Mask.
4459 // Mask0 & Mask1 is a Mask.
4460 // ~Mask0 | ~Mask1 is a ~Mask.
4461 // ~Mask0 & ~Mask1 is a ~Mask.
4462 return isMaskOrZero(V: I->getOperand(i: 1), Not, Q, Depth) &&
4463 isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4464 case Instruction::Xor:
4465 if (match(V, P: m_Not(V: m_Value(V&: X))))
4466 return isMaskOrZero(V: X, Not: !Not, Q, Depth);
4467
4468 // (X ^ -X) is a ~Mask
4469 if (Not)
4470 return match(V, P: m_c_Xor(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))));
4471 // (X ^ (X - 1)) is a Mask
4472 else
4473 return match(V, P: m_c_Xor(L: m_Value(V&: X), R: m_Add(L: m_Deferred(V: X), R: m_AllOnes())));
4474 case Instruction::Select:
4475 // c ? Mask0 : Mask1 is a Mask.
4476 return isMaskOrZero(V: I->getOperand(i: 1), Not, Q, Depth) &&
4477 isMaskOrZero(V: I->getOperand(i: 2), Not, Q, Depth);
4478 case Instruction::Shl:
4479 // (~Mask) << X is a ~Mask.
4480 return Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4481 case Instruction::LShr:
4482 // Mask >> X is a Mask.
4483 return !Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4484 case Instruction::AShr:
4485 // Mask s>> X is a Mask.
4486 // ~Mask s>> X is a ~Mask.
4487 return isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth);
4488 case Instruction::Add:
4489 // Pow2 - 1 is a Mask.
4490 if (!Not && match(V: I->getOperand(i: 1), P: m_AllOnes()))
4491 return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), DL: Q.DL, /*OrZero*/ true,
4492 AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT, UseInstrInfo: Depth);
4493 break;
4494 case Instruction::Sub:
4495 // -Pow2 is a ~Mask.
4496 if (Not && match(V: I->getOperand(i: 0), P: m_Zero()))
4497 return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), DL: Q.DL, /*OrZero*/ true,
4498 AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT, UseInstrInfo: Depth);
4499 break;
4500 case Instruction::Call: {
4501 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
4502 switch (II->getIntrinsicID()) {
4503 // min/max(Mask0, Mask1) is a Mask.
4504 // min/max(~Mask0, ~Mask1) is a ~Mask.
4505 case Intrinsic::umax:
4506 case Intrinsic::smax:
4507 case Intrinsic::umin:
4508 case Intrinsic::smin:
4509 return isMaskOrZero(V: II->getArgOperand(i: 1), Not, Q, Depth) &&
4510 isMaskOrZero(V: II->getArgOperand(i: 0), Not, Q, Depth);
4511
4512 // In the context of masks, bitreverse(Mask) == ~Mask
4513 case Intrinsic::bitreverse:
4514 return isMaskOrZero(V: II->getArgOperand(i: 0), Not: !Not, Q, Depth);
4515 default:
4516 break;
4517 }
4518 }
4519 break;
4520 }
4521 default:
4522 break;
4523 }
4524 return false;
4525}
4526
4527/// Some comparisons can be simplified.
4528/// In this case, we are looking for comparisons that look like
4529/// a check for a lossy truncation.
4530/// Folds:
4531/// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
4532/// icmp SrcPred (x & ~Mask), ~Mask to icmp DstPred x, ~Mask
4533/// icmp eq/ne (x & ~Mask), 0 to icmp DstPred x, Mask
4534/// icmp eq/ne (~x | Mask), -1 to icmp DstPred x, Mask
4535/// Where Mask is some pattern that produces all-ones in low bits:
4536/// (-1 >> y)
4537/// ((-1 << y) >> y) <- non-canonical, has extra uses
4538/// ~(-1 << y)
4539/// ((1 << y) + (-1)) <- non-canonical, has extra uses
4540/// The Mask can be a constant, too.
4541/// For some predicates, the operands are commutative.
4542/// For others, x can only be on a specific side.
4543static Value *foldICmpWithLowBitMaskedVal(CmpPredicate Pred, Value *Op0,
4544 Value *Op1, const SimplifyQuery &Q,
4545 InstCombiner &IC) {
4546
4547 ICmpInst::Predicate DstPred;
4548 switch (Pred) {
4549 case ICmpInst::Predicate::ICMP_EQ:
4550 // x & Mask == x
4551 // x & ~Mask == 0
4552 // ~x | Mask == -1
4553 // -> x u<= Mask
4554 // x & ~Mask == ~Mask
4555 // -> ~Mask u<= x
4556 DstPred = ICmpInst::Predicate::ICMP_ULE;
4557 break;
4558 case ICmpInst::Predicate::ICMP_NE:
4559 // x & Mask != x
4560 // x & ~Mask != 0
4561 // ~x | Mask != -1
4562 // -> x u> Mask
4563 // x & ~Mask != ~Mask
4564 // -> ~Mask u> x
4565 DstPred = ICmpInst::Predicate::ICMP_UGT;
4566 break;
4567 case ICmpInst::Predicate::ICMP_ULT:
4568 // x & Mask u< x
4569 // -> x u> Mask
4570 // x & ~Mask u< ~Mask
4571 // -> ~Mask u> x
4572 DstPred = ICmpInst::Predicate::ICMP_UGT;
4573 break;
4574 case ICmpInst::Predicate::ICMP_UGE:
4575 // x & Mask u>= x
4576 // -> x u<= Mask
4577 // x & ~Mask u>= ~Mask
4578 // -> ~Mask u<= x
4579 DstPred = ICmpInst::Predicate::ICMP_ULE;
4580 break;
4581 case ICmpInst::Predicate::ICMP_SLT:
4582 // x & Mask s< x [iff Mask s>= 0]
4583 // -> x s> Mask
4584 // x & ~Mask s< ~Mask [iff ~Mask != 0]
4585 // -> ~Mask s> x
4586 DstPred = ICmpInst::Predicate::ICMP_SGT;
4587 break;
4588 case ICmpInst::Predicate::ICMP_SGE:
4589 // x & Mask s>= x [iff Mask s>= 0]
4590 // -> x s<= Mask
4591 // x & ~Mask s>= ~Mask [iff ~Mask != 0]
4592 // -> ~Mask s<= x
4593 DstPred = ICmpInst::Predicate::ICMP_SLE;
4594 break;
4595 default:
4596 // We don't support sgt,sle
4597 // ult/ugt are simplified to true/false respectively.
4598 return nullptr;
4599 }
4600
4601 Value *X, *M;
4602 // Put search code in lambda for early positive returns.
4603 auto IsLowBitMask = [&]() {
4604 if (match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value(V&: M)))) {
4605 X = Op1;
4606 // Look for: x & Mask pred x
4607 if (isMaskOrZero(V: M, /*Not=*/false, Q)) {
4608 return !ICmpInst::isSigned(predicate: Pred) ||
4609 (match(V: M, P: m_NonNegative()) || isKnownNonNegative(V: M, SQ: Q));
4610 }
4611
4612 // Look for: x & ~Mask pred ~Mask
4613 if (isMaskOrZero(V: X, /*Not=*/true, Q)) {
4614 return !ICmpInst::isSigned(predicate: Pred) || isKnownNonZero(V: X, Q);
4615 }
4616 return false;
4617 }
4618 if (ICmpInst::isEquality(P: Pred) && match(V: Op1, P: m_AllOnes()) &&
4619 match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Value(V&: M))))) {
4620
4621 auto Check = [&]() {
4622 // Look for: ~x | Mask == -1
4623 if (isMaskOrZero(V: M, /*Not=*/false, Q)) {
4624 if (Value *NotX =
4625 IC.getFreelyInverted(V: X, WillInvertAllUses: X->hasOneUse(), Builder: &IC.Builder)) {
4626 X = NotX;
4627 return true;
4628 }
4629 }
4630 return false;
4631 };
4632 if (Check())
4633 return true;
4634 std::swap(a&: X, b&: M);
4635 return Check();
4636 }
4637 if (ICmpInst::isEquality(P: Pred) && match(V: Op1, P: m_Zero()) &&
4638 match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_Value(V&: M))))) {
4639 auto Check = [&]() {
4640 // Look for: x & ~Mask == 0
4641 if (isMaskOrZero(V: M, /*Not=*/true, Q)) {
4642 if (Value *NotM =
4643 IC.getFreelyInverted(V: M, WillInvertAllUses: M->hasOneUse(), Builder: &IC.Builder)) {
4644 M = NotM;
4645 return true;
4646 }
4647 }
4648 return false;
4649 };
4650 if (Check())
4651 return true;
4652 std::swap(a&: X, b&: M);
4653 return Check();
4654 }
4655 return false;
4656 };
4657
4658 if (!IsLowBitMask())
4659 return nullptr;
4660
4661 return IC.Builder.CreateICmp(P: DstPred, LHS: X, RHS: M);
4662}
4663
4664/// Some comparisons can be simplified.
4665/// In this case, we are looking for comparisons that look like
4666/// a check for a lossy signed truncation.
4667/// Folds: (MaskedBits is a constant.)
4668/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
4669/// Into:
4670/// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
4671/// Where KeptBits = bitwidth(%x) - MaskedBits
4672static Value *
4673foldICmpWithTruncSignExtendedVal(ICmpInst &I,
4674 InstCombiner::BuilderTy &Builder) {
4675 CmpPredicate SrcPred;
4676 Value *X;
4677 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
4678 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
4679 if (!match(V: &I, P: m_c_ICmp(Pred&: SrcPred,
4680 L: m_OneUse(SubPattern: m_AShr(L: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C0)),
4681 R: m_APInt(Res&: C1))),
4682 R: m_Deferred(V: X))))
4683 return nullptr;
4684
4685 // Potential handling of non-splats: for each element:
4686 // * if both are undef, replace with constant 0.
4687 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
4688 // * if both are not undef, and are different, bailout.
4689 // * else, only one is undef, then pick the non-undef one.
4690
4691 // The shift amount must be equal.
4692 if (*C0 != *C1)
4693 return nullptr;
4694 const APInt &MaskedBits = *C0;
4695 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
4696
4697 ICmpInst::Predicate DstPred;
4698 switch (SrcPred) {
4699 case ICmpInst::Predicate::ICMP_EQ:
4700 // ((%x << MaskedBits) a>> MaskedBits) == %x
4701 // =>
4702 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
4703 DstPred = ICmpInst::Predicate::ICMP_ULT;
4704 break;
4705 case ICmpInst::Predicate::ICMP_NE:
4706 // ((%x << MaskedBits) a>> MaskedBits) != %x
4707 // =>
4708 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
4709 DstPred = ICmpInst::Predicate::ICMP_UGE;
4710 break;
4711 // FIXME: are more folds possible?
4712 default:
4713 return nullptr;
4714 }
4715
4716 auto *XType = X->getType();
4717 const unsigned XBitWidth = XType->getScalarSizeInBits();
4718 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
4719 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
4720
4721 // KeptBits = bitwidth(%x) - MaskedBits
4722 const APInt KeptBits = BitWidth - MaskedBits;
4723 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
4724 // ICmpCst = (1 << KeptBits)
4725 const APInt ICmpCst = APInt(XBitWidth, 1).shl(ShiftAmt: KeptBits);
4726 assert(ICmpCst.isPowerOf2());
4727 // AddCst = (1 << (KeptBits-1))
4728 const APInt AddCst = ICmpCst.lshr(shiftAmt: 1);
4729 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
4730
4731 // T0 = add %x, AddCst
4732 Value *T0 = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: XType, V: AddCst));
4733 // T1 = T0 DstPred ICmpCst
4734 Value *T1 = Builder.CreateICmp(P: DstPred, LHS: T0, RHS: ConstantInt::get(Ty: XType, V: ICmpCst));
4735
4736 return T1;
4737}
4738
4739// Given pattern:
4740// icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
4741// we should move shifts to the same hand of 'and', i.e. rewrite as
4742// icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
4743// We are only interested in opposite logical shifts here.
4744// One of the shifts can be truncated.
4745// If we can, we want to end up creating 'lshr' shift.
4746static Value *
4747foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
4748 InstCombiner::BuilderTy &Builder) {
4749 if (!I.isEquality() || !match(V: I.getOperand(i_nocapture: 1), P: m_Zero()) ||
4750 !I.getOperand(i_nocapture: 0)->hasOneUse())
4751 return nullptr;
4752
4753 auto m_AnyLogicalShift = m_LogicalShift(L: m_Value(), R: m_Value());
4754
4755 // Look for an 'and' of two logical shifts, one of which may be truncated.
4756 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
4757 Instruction *XShift, *MaybeTruncation, *YShift;
4758 if (!match(
4759 V: I.getOperand(i_nocapture: 0),
4760 P: m_c_And(L: m_CombineAnd(L: m_AnyLogicalShift, R: m_Instruction(I&: XShift)),
4761 R: m_CombineAnd(L: m_TruncOrSelf(Op: m_CombineAnd(
4762 L: m_AnyLogicalShift, R: m_Instruction(I&: YShift))),
4763 R: m_Instruction(I&: MaybeTruncation)))))
4764 return nullptr;
4765
4766 // We potentially looked past 'trunc', but only when matching YShift,
4767 // therefore YShift must have the widest type.
4768 Instruction *WidestShift = YShift;
4769 // Therefore XShift must have the shallowest type.
4770 // Or they both have identical types if there was no truncation.
4771 Instruction *NarrowestShift = XShift;
4772
4773 Type *WidestTy = WidestShift->getType();
4774 Type *NarrowestTy = NarrowestShift->getType();
4775 assert(NarrowestTy == I.getOperand(0)->getType() &&
4776 "We did not look past any shifts while matching XShift though.");
4777 bool HadTrunc = WidestTy != I.getOperand(i_nocapture: 0)->getType();
4778
4779 // If YShift is a 'lshr', swap the shifts around.
4780 if (match(V: YShift, P: m_LShr(L: m_Value(), R: m_Value())))
4781 std::swap(a&: XShift, b&: YShift);
4782
4783 // The shifts must be in opposite directions.
4784 auto XShiftOpcode = XShift->getOpcode();
4785 if (XShiftOpcode == YShift->getOpcode())
4786 return nullptr; // Do not care about same-direction shifts here.
4787
4788 Value *X, *XShAmt, *Y, *YShAmt;
4789 match(V: XShift, P: m_BinOp(L: m_Value(V&: X), R: m_ZExtOrSelf(Op: m_Value(V&: XShAmt))));
4790 match(V: YShift, P: m_BinOp(L: m_Value(V&: Y), R: m_ZExtOrSelf(Op: m_Value(V&: YShAmt))));
4791
4792 // If one of the values being shifted is a constant, then we will end with
4793 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
4794 // however, we will need to ensure that we won't increase instruction count.
4795 if (!isa<Constant>(Val: X) && !isa<Constant>(Val: Y)) {
4796 // At least one of the hands of the 'and' should be one-use shift.
4797 if (!match(V: I.getOperand(i_nocapture: 0),
4798 P: m_c_And(L: m_OneUse(SubPattern: m_AnyLogicalShift), R: m_Value())))
4799 return nullptr;
4800 if (HadTrunc) {
4801 // Due to the 'trunc', we will need to widen X. For that either the old
4802 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
4803 if (!MaybeTruncation->hasOneUse() &&
4804 !NarrowestShift->getOperand(i: 1)->hasOneUse())
4805 return nullptr;
4806 }
4807 }
4808
4809 // We have two shift amounts from two different shifts. The types of those
4810 // shift amounts may not match. If that's the case let's bailout now.
4811 if (XShAmt->getType() != YShAmt->getType())
4812 return nullptr;
4813
4814 // As input, we have the following pattern:
4815 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
4816 // We want to rewrite that as:
4817 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
4818 // While we know that originally (Q+K) would not overflow
4819 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
4820 // shift amounts. so it may now overflow in smaller bitwidth.
4821 // To ensure that does not happen, we need to ensure that the total maximal
4822 // shift amount is still representable in that smaller bit width.
4823 unsigned MaximalPossibleTotalShiftAmount =
4824 (WidestTy->getScalarSizeInBits() - 1) +
4825 (NarrowestTy->getScalarSizeInBits() - 1);
4826 APInt MaximalRepresentableShiftAmount =
4827 APInt::getAllOnes(numBits: XShAmt->getType()->getScalarSizeInBits());
4828 if (MaximalRepresentableShiftAmount.ult(RHS: MaximalPossibleTotalShiftAmount))
4829 return nullptr;
4830
4831 // Can we fold (XShAmt+YShAmt) ?
4832 auto *NewShAmt = dyn_cast_or_null<Constant>(
4833 Val: simplifyAddInst(LHS: XShAmt, RHS: YShAmt, /*isNSW=*/IsNSW: false,
4834 /*isNUW=*/IsNUW: false, Q: SQ.getWithInstruction(I: &I)));
4835 if (!NewShAmt)
4836 return nullptr;
4837 if (NewShAmt->getType() != WidestTy) {
4838 NewShAmt =
4839 ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: NewShAmt, DestTy: WidestTy, DL: SQ.DL);
4840 if (!NewShAmt)
4841 return nullptr;
4842 }
4843 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
4844
4845 // Is the new shift amount smaller than the bit width?
4846 // FIXME: could also rely on ConstantRange.
4847 if (!match(V: NewShAmt,
4848 P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_ULT,
4849 Threshold: APInt(WidestBitWidth, WidestBitWidth))))
4850 return nullptr;
4851
4852 // An extra legality check is needed if we had trunc-of-lshr.
4853 if (HadTrunc && match(V: WidestShift, P: m_LShr(L: m_Value(), R: m_Value()))) {
4854 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
4855 WidestShift]() {
4856 // It isn't obvious whether it's worth it to analyze non-constants here.
4857 // Also, let's basically give up on non-splat cases, pessimizing vectors.
4858 // If *any* of these preconditions matches we can perform the fold.
4859 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
4860 ? NewShAmt->getSplatValue()
4861 : NewShAmt;
4862 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
4863 if (NewShAmtSplat &&
4864 (NewShAmtSplat->isNullValue() ||
4865 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
4866 return true;
4867 // We consider *min* leading zeros so a single outlier
4868 // blocks the transform as opposed to allowing it.
4869 if (auto *C = dyn_cast<Constant>(Val: NarrowestShift->getOperand(i: 0))) {
4870 KnownBits Known = computeKnownBits(V: C, DL: SQ.DL);
4871 unsigned MinLeadZero = Known.countMinLeadingZeros();
4872 // If the value being shifted has at most lowest bit set we can fold.
4873 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
4874 if (MaxActiveBits <= 1)
4875 return true;
4876 // Precondition: NewShAmt u<= countLeadingZeros(C)
4877 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(RHS: MinLeadZero))
4878 return true;
4879 }
4880 if (auto *C = dyn_cast<Constant>(Val: WidestShift->getOperand(i: 0))) {
4881 KnownBits Known = computeKnownBits(V: C, DL: SQ.DL);
4882 unsigned MinLeadZero = Known.countMinLeadingZeros();
4883 // If the value being shifted has at most lowest bit set we can fold.
4884 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
4885 if (MaxActiveBits <= 1)
4886 return true;
4887 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
4888 if (NewShAmtSplat) {
4889 APInt AdjNewShAmt =
4890 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
4891 if (AdjNewShAmt.ule(RHS: MinLeadZero))
4892 return true;
4893 }
4894 }
4895 return false; // Can't tell if it's ok.
4896 };
4897 if (!CanFold())
4898 return nullptr;
4899 }
4900
4901 // All good, we can do this fold.
4902 X = Builder.CreateZExt(V: X, DestTy: WidestTy);
4903 Y = Builder.CreateZExt(V: Y, DestTy: WidestTy);
4904 // The shift is the same that was for X.
4905 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
4906 ? Builder.CreateLShr(LHS: X, RHS: NewShAmt)
4907 : Builder.CreateShl(LHS: X, RHS: NewShAmt);
4908 Value *T1 = Builder.CreateAnd(LHS: T0, RHS: Y);
4909 return Builder.CreateICmp(P: I.getPredicate(), LHS: T1,
4910 RHS: Constant::getNullValue(Ty: WidestTy));
4911}
4912
4913/// Fold
4914/// (-1 u/ x) u< y
4915/// ((x * y) ?/ x) != y
4916/// to
4917/// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
4918/// Note that the comparison is commutative, while inverted (u>=, ==) predicate
4919/// will mean that we are looking for the opposite answer.
4920Value *InstCombinerImpl::foldMultiplicationOverflowCheck(ICmpInst &I) {
4921 CmpPredicate Pred;
4922 Value *X, *Y;
4923 Instruction *Mul;
4924 Instruction *Div;
4925 bool NeedNegation;
4926 // Look for: (-1 u/ x) u</u>= y
4927 if (!I.isEquality() &&
4928 match(V: &I, P: m_c_ICmp(Pred,
4929 L: m_CombineAnd(L: m_OneUse(SubPattern: m_UDiv(L: m_AllOnes(), R: m_Value(V&: X))),
4930 R: m_Instruction(I&: Div)),
4931 R: m_Value(V&: Y)))) {
4932 Mul = nullptr;
4933
4934 // Are we checking that overflow does not happen, or does happen?
4935 switch (Pred) {
4936 case ICmpInst::Predicate::ICMP_ULT:
4937 NeedNegation = false;
4938 break; // OK
4939 case ICmpInst::Predicate::ICMP_UGE:
4940 NeedNegation = true;
4941 break; // OK
4942 default:
4943 return nullptr; // Wrong predicate.
4944 }
4945 } else // Look for: ((x * y) / x) !=/== y
4946 if (I.isEquality() &&
4947 match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: Y),
4948 R: m_CombineAnd(L: m_OneUse(SubPattern: m_IDiv(
4949 L: m_CombineAnd(L: m_c_Mul(L: m_Deferred(V: Y),
4950 R: m_Value(V&: X)),
4951 R: m_Instruction(I&: Mul)),
4952 R: m_Deferred(V: X))),
4953 R: m_Instruction(I&: Div))))) {
4954 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
4955 } else
4956 return nullptr;
4957
4958 BuilderTy::InsertPointGuard Guard(Builder);
4959 // If the pattern included (x * y), we'll want to insert new instructions
4960 // right before that original multiplication so that we can replace it.
4961 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
4962 if (MulHadOtherUses)
4963 Builder.SetInsertPoint(Mul);
4964
4965 CallInst *Call = Builder.CreateIntrinsic(
4966 ID: Div->getOpcode() == Instruction::UDiv ? Intrinsic::umul_with_overflow
4967 : Intrinsic::smul_with_overflow,
4968 Types: X->getType(), Args: {X, Y}, /*FMFSource=*/nullptr, Name: "mul");
4969
4970 // If the multiplication was used elsewhere, to ensure that we don't leave
4971 // "duplicate" instructions, replace uses of that original multiplication
4972 // with the multiplication result from the with.overflow intrinsic.
4973 if (MulHadOtherUses)
4974 replaceInstUsesWith(I&: *Mul, V: Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "mul.val"));
4975
4976 Value *Res = Builder.CreateExtractValue(Agg: Call, Idxs: 1, Name: "mul.ov");
4977 if (NeedNegation) // This technically increases instruction count.
4978 Res = Builder.CreateNot(V: Res, Name: "mul.not.ov");
4979
4980 // If we replaced the mul, erase it. Do this after all uses of Builder,
4981 // as the mul is used as insertion point.
4982 if (MulHadOtherUses)
4983 eraseInstFromFunction(I&: *Mul);
4984
4985 return Res;
4986}
4987
4988static Instruction *foldICmpXNegX(ICmpInst &I,
4989 InstCombiner::BuilderTy &Builder) {
4990 CmpPredicate Pred;
4991 Value *X;
4992 if (match(V: &I, P: m_c_ICmp(Pred, L: m_NSWNeg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) {
4993
4994 if (ICmpInst::isSigned(predicate: Pred))
4995 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
4996 else if (ICmpInst::isUnsigned(predicate: Pred))
4997 Pred = ICmpInst::getSignedPredicate(Pred);
4998 // else for equality-comparisons just keep the predicate.
4999
5000 return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: X,
5001 S2: Constant::getNullValue(Ty: X->getType()), Name: I.getName());
5002 }
5003
5004 // A value is not equal to its negation unless that value is 0 or
5005 // MinSignedValue, ie: a != -a --> (a & MaxSignedVal) != 0
5006 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))) &&
5007 ICmpInst::isEquality(P: Pred)) {
5008 Type *Ty = X->getType();
5009 uint32_t BitWidth = Ty->getScalarSizeInBits();
5010 Constant *MaxSignedVal =
5011 ConstantInt::get(Ty, V: APInt::getSignedMaxValue(numBits: BitWidth));
5012 Value *And = Builder.CreateAnd(LHS: X, RHS: MaxSignedVal);
5013 Constant *Zero = Constant::getNullValue(Ty);
5014 return CmpInst::Create(Op: Instruction::ICmp, Pred, S1: And, S2: Zero);
5015 }
5016
5017 return nullptr;
5018}
5019
5020static Instruction *foldICmpAndXX(ICmpInst &I, const SimplifyQuery &Q,
5021 InstCombinerImpl &IC) {
5022 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A;
5023 // Normalize and operand as operand 0.
5024 CmpInst::Predicate Pred = I.getPredicate();
5025 if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value()))) {
5026 std::swap(a&: Op0, b&: Op1);
5027 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
5028 }
5029
5030 if (!match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value(V&: A))))
5031 return nullptr;
5032
5033 // (icmp (X & Y) u< X --> (X & Y) != X
5034 if (Pred == ICmpInst::ICMP_ULT)
5035 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5036
5037 // (icmp (X & Y) u>= X --> (X & Y) == X
5038 if (Pred == ICmpInst::ICMP_UGE)
5039 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5040
5041 if (ICmpInst::isEquality(P: Pred) && Op0->hasOneUse()) {
5042 // icmp (X & Y) eq/ne Y --> (X | ~Y) eq/ne -1 if Y is freely invertible and
5043 // Y is non-constant. If Y is constant the `X & C == C` form is preferable
5044 // so don't do this fold.
5045 if (!match(V: Op1, P: m_ImmConstant()))
5046 if (auto *NotOp1 =
5047 IC.getFreelyInverted(V: Op1, WillInvertAllUses: !Op1->hasNUsesOrMore(N: 3), Builder: &IC.Builder))
5048 return new ICmpInst(Pred, IC.Builder.CreateOr(LHS: A, RHS: NotOp1),
5049 Constant::getAllOnesValue(Ty: Op1->getType()));
5050 // icmp (X & Y) eq/ne Y --> (~X & Y) eq/ne 0 if X is freely invertible.
5051 if (auto *NotA = IC.getFreelyInverted(V: A, WillInvertAllUses: A->hasOneUse(), Builder: &IC.Builder))
5052 return new ICmpInst(Pred, IC.Builder.CreateAnd(LHS: Op1, RHS: NotA),
5053 Constant::getNullValue(Ty: Op1->getType()));
5054 }
5055
5056 if (!ICmpInst::isSigned(predicate: Pred))
5057 return nullptr;
5058
5059 KnownBits KnownY = IC.computeKnownBits(V: A, CxtI: &I);
5060 // (X & NegY) spred X --> (X & NegY) upred X
5061 if (KnownY.isNegative())
5062 return new ICmpInst(ICmpInst::getUnsignedPredicate(Pred), Op0, Op1);
5063
5064 if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGT)
5065 return nullptr;
5066
5067 if (KnownY.isNonNegative())
5068 // (X & PosY) s<= X --> X s>= 0
5069 // (X & PosY) s> X --> X s< 0
5070 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
5071 Constant::getNullValue(Ty: Op1->getType()));
5072
5073 if (isKnownNegative(V: Op1, SQ: IC.getSimplifyQuery().getWithInstruction(I: &I)))
5074 // (NegX & Y) s<= NegX --> Y s< 0
5075 // (NegX & Y) s> NegX --> Y s>= 0
5076 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), A,
5077 Constant::getNullValue(Ty: A->getType()));
5078
5079 return nullptr;
5080}
5081
5082static Instruction *foldICmpOrXX(ICmpInst &I, const SimplifyQuery &Q,
5083 InstCombinerImpl &IC) {
5084 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A;
5085
5086 // Normalize or operand as operand 0.
5087 CmpInst::Predicate Pred = I.getPredicate();
5088 if (match(V: Op1, P: m_c_Or(L: m_Specific(V: Op0), R: m_Value(V&: A)))) {
5089 std::swap(a&: Op0, b&: Op1);
5090 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
5091 } else if (!match(V: Op0, P: m_c_Or(L: m_Specific(V: Op1), R: m_Value(V&: A)))) {
5092 return nullptr;
5093 }
5094
5095 // icmp (X | Y) u<= X --> (X | Y) == X
5096 if (Pred == ICmpInst::ICMP_ULE)
5097 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5098
5099 // icmp (X | Y) u> X --> (X | Y) != X
5100 if (Pred == ICmpInst::ICMP_UGT)
5101 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5102
5103 if (ICmpInst::isEquality(P: Pred) && Op0->hasOneUse()) {
5104 // icmp (X | Y) eq/ne Y --> (X & ~Y) eq/ne 0 if Y is freely invertible
5105 if (Value *NotOp1 = IC.getFreelyInverted(
5106 V: Op1, WillInvertAllUses: !isa<Constant>(Val: Op1) && !Op1->hasNUsesOrMore(N: 3), Builder: &IC.Builder))
5107 return new ICmpInst(Pred, IC.Builder.CreateAnd(LHS: A, RHS: NotOp1),
5108 Constant::getNullValue(Ty: Op1->getType()));
5109 // icmp (X | Y) eq/ne Y --> (~X | Y) eq/ne -1 if X is freely invertible.
5110 if (Value *NotA = IC.getFreelyInverted(V: A, WillInvertAllUses: A->hasOneUse(), Builder: &IC.Builder))
5111 return new ICmpInst(Pred, IC.Builder.CreateOr(LHS: Op1, RHS: NotA),
5112 Constant::getAllOnesValue(Ty: Op1->getType()));
5113 }
5114 return nullptr;
5115}
5116
5117static Instruction *foldICmpXorXX(ICmpInst &I, const SimplifyQuery &Q,
5118 InstCombinerImpl &IC) {
5119 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A;
5120 // Normalize xor operand as operand 0.
5121 CmpInst::Predicate Pred = I.getPredicate();
5122 if (match(V: Op1, P: m_c_Xor(L: m_Specific(V: Op0), R: m_Value()))) {
5123 std::swap(a&: Op0, b&: Op1);
5124 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
5125 }
5126 if (!match(V: Op0, P: m_c_Xor(L: m_Specific(V: Op1), R: m_Value(V&: A))))
5127 return nullptr;
5128
5129 // icmp (X ^ Y_NonZero) u>= X --> icmp (X ^ Y_NonZero) u> X
5130 // icmp (X ^ Y_NonZero) u<= X --> icmp (X ^ Y_NonZero) u< X
5131 // icmp (X ^ Y_NonZero) s>= X --> icmp (X ^ Y_NonZero) s> X
5132 // icmp (X ^ Y_NonZero) s<= X --> icmp (X ^ Y_NonZero) s< X
5133 CmpInst::Predicate PredOut = CmpInst::getStrictPredicate(pred: Pred);
5134 if (PredOut != Pred && isKnownNonZero(V: A, Q))
5135 return new ICmpInst(PredOut, Op0, Op1);
5136
5137 // These transform work when A is negative.
5138 // X s< X^A, X s<= X^A, X u> X^A, X u>= X^A --> X s< 0
5139 // X s> X^A, X s>= X^A, X u< X^A, X u<= X^A --> X s>= 0
5140 if (match(V: A, P: m_Negative())) {
5141 CmpInst::Predicate NewPred;
5142 switch (ICmpInst::getStrictPredicate(pred: Pred)) {
5143 default:
5144 return nullptr;
5145 case ICmpInst::ICMP_SLT:
5146 case ICmpInst::ICMP_UGT:
5147 NewPred = ICmpInst::ICMP_SLT;
5148 break;
5149 case ICmpInst::ICMP_SGT:
5150 case ICmpInst::ICMP_ULT:
5151 NewPred = ICmpInst::ICMP_SGE;
5152 break;
5153 }
5154 Constant *Const = Constant::getNullValue(Ty: Op0->getType());
5155 return new ICmpInst(NewPred, Op0, Const);
5156 }
5157
5158 return nullptr;
5159}
5160
5161/// Return true if X is a multiple of C.
5162/// TODO: Handle non-power-of-2 factors.
5163static bool isMultipleOf(Value *X, const APInt &C, const SimplifyQuery &Q) {
5164 if (C.isOne())
5165 return true;
5166
5167 if (!C.isPowerOf2())
5168 return false;
5169
5170 return MaskedValueIsZero(V: X, Mask: C - 1, SQ: Q);
5171}
5172
5173/// Try to fold icmp (binop), X or icmp X, (binop).
5174/// TODO: A large part of this logic is duplicated in InstSimplify's
5175/// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
5176/// duplication.
5177Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
5178 const SimplifyQuery &SQ) {
5179 const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
5180 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
5181
5182 // Special logic for binary operators.
5183 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Val: Op0);
5184 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Val: Op1);
5185 if (!BO0 && !BO1)
5186 return nullptr;
5187
5188 if (Instruction *NewICmp = foldICmpXNegX(I, Builder))
5189 return NewICmp;
5190
5191 const CmpInst::Predicate Pred = I.getPredicate();
5192 Value *X;
5193
5194 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
5195 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
5196 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Op1), R: m_Value(V&: X)))) &&
5197 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
5198 return new ICmpInst(Pred, Builder.CreateNot(V: Op1), X);
5199 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
5200 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: X)))) &&
5201 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
5202 return new ICmpInst(Pred, X, Builder.CreateNot(V: Op0));
5203
5204 {
5205 // (Op1 + X) + C u</u>= Op1 --> ~C - X u</u>= Op1
5206 Constant *C;
5207 if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_c_Add(L: m_Specific(V: Op1), R: m_Value(V&: X)),
5208 R: m_ImmConstant(C)))) &&
5209 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
5210 Constant *C2 = ConstantExpr::getNot(C);
5211 return new ICmpInst(Pred, Builder.CreateSub(LHS: C2, RHS: X), Op1);
5212 }
5213 // Op0 u>/u<= (Op0 + X) + C --> Op0 u>/u<= ~C - X
5214 if (match(V: Op1, P: m_OneUse(SubPattern: m_Add(L: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: X)),
5215 R: m_ImmConstant(C)))) &&
5216 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE)) {
5217 Constant *C2 = ConstantExpr::getNot(C);
5218 return new ICmpInst(Pred, Op0, Builder.CreateSub(LHS: C2, RHS: X));
5219 }
5220 }
5221
5222 // (icmp eq/ne (X, -P2), INT_MIN)
5223 // -> (icmp slt/sge X, INT_MIN + P2)
5224 if (ICmpInst::isEquality(P: Pred) && BO0 &&
5225 match(V: I.getOperand(i_nocapture: 1), P: m_SignMask()) &&
5226 match(V: BO0, P: m_And(L: m_Value(), R: m_NegatedPower2OrZero()))) {
5227 // Will Constant fold.
5228 Value *NewC = Builder.CreateSub(LHS: I.getOperand(i_nocapture: 1), RHS: BO0->getOperand(i_nocapture: 1));
5229 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SLT
5230 : ICmpInst::ICMP_SGE,
5231 BO0->getOperand(i_nocapture: 0), NewC);
5232 }
5233
5234 {
5235 // Similar to above: an unsigned overflow comparison may use offset + mask:
5236 // ((Op1 + C) & C) u< Op1 --> Op1 != 0
5237 // ((Op1 + C) & C) u>= Op1 --> Op1 == 0
5238 // Op0 u> ((Op0 + C) & C) --> Op0 != 0
5239 // Op0 u<= ((Op0 + C) & C) --> Op0 == 0
5240 BinaryOperator *BO;
5241 const APInt *C;
5242 if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) &&
5243 match(V: Op0, P: m_And(L: m_BinOp(I&: BO), R: m_LowBitMask(V&: C))) &&
5244 match(V: BO, P: m_Add(L: m_Specific(V: Op1), R: m_SpecificIntAllowPoison(V: *C)))) {
5245 CmpInst::Predicate NewPred =
5246 Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
5247 Constant *Zero = ConstantInt::getNullValue(Ty: Op1->getType());
5248 return new ICmpInst(NewPred, Op1, Zero);
5249 }
5250
5251 if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
5252 match(V: Op1, P: m_And(L: m_BinOp(I&: BO), R: m_LowBitMask(V&: C))) &&
5253 match(V: BO, P: m_Add(L: m_Specific(V: Op0), R: m_SpecificIntAllowPoison(V: *C)))) {
5254 CmpInst::Predicate NewPred =
5255 Pred == ICmpInst::ICMP_UGT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
5256 Constant *Zero = ConstantInt::getNullValue(Ty: Op1->getType());
5257 return new ICmpInst(NewPred, Op0, Zero);
5258 }
5259 }
5260
5261 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
5262 bool Op0HasNUW = false, Op1HasNUW = false;
5263 bool Op0HasNSW = false, Op1HasNSW = false;
5264 // Analyze the case when either Op0 or Op1 is an add instruction.
5265 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
5266 auto hasNoWrapProblem = [](const BinaryOperator &BO, CmpInst::Predicate Pred,
5267 bool &HasNSW, bool &HasNUW) -> bool {
5268 if (isa<OverflowingBinaryOperator>(Val: BO)) {
5269 HasNUW = BO.hasNoUnsignedWrap();
5270 HasNSW = BO.hasNoSignedWrap();
5271 return ICmpInst::isEquality(P: Pred) ||
5272 (CmpInst::isUnsigned(predicate: Pred) && HasNUW) ||
5273 (CmpInst::isSigned(predicate: Pred) && HasNSW);
5274 } else if (BO.getOpcode() == Instruction::Or) {
5275 HasNUW = true;
5276 HasNSW = true;
5277 return true;
5278 } else {
5279 return false;
5280 }
5281 };
5282 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
5283
5284 if (BO0) {
5285 match(V: BO0, P: m_AddLike(L: m_Value(V&: A), R: m_Value(V&: B)));
5286 NoOp0WrapProblem = hasNoWrapProblem(*BO0, Pred, Op0HasNSW, Op0HasNUW);
5287 }
5288 if (BO1) {
5289 match(V: BO1, P: m_AddLike(L: m_Value(V&: C), R: m_Value(V&: D)));
5290 NoOp1WrapProblem = hasNoWrapProblem(*BO1, Pred, Op1HasNSW, Op1HasNUW);
5291 }
5292
5293 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
5294 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
5295 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
5296 return new ICmpInst(Pred, A == Op1 ? B : A,
5297 Constant::getNullValue(Ty: Op1->getType()));
5298
5299 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
5300 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
5301 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
5302 return new ICmpInst(Pred, Constant::getNullValue(Ty: Op0->getType()),
5303 C == Op0 ? D : C);
5304
5305 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
5306 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
5307 NoOp1WrapProblem) {
5308 // Determine Y and Z in the form icmp (X+Y), (X+Z).
5309 Value *Y, *Z;
5310 if (A == C) {
5311 // C + B == C + D -> B == D
5312 Y = B;
5313 Z = D;
5314 } else if (A == D) {
5315 // D + B == C + D -> B == C
5316 Y = B;
5317 Z = C;
5318 } else if (B == C) {
5319 // A + C == C + D -> A == D
5320 Y = A;
5321 Z = D;
5322 } else {
5323 assert(B == D);
5324 // A + D == C + D -> A == C
5325 Y = A;
5326 Z = C;
5327 }
5328 return new ICmpInst(Pred, Y, Z);
5329 }
5330
5331 if (ICmpInst::isRelational(P: Pred)) {
5332 // Return if both X and Y is divisible by Z/-Z.
5333 // TODO: Generalize to check if (X - Y) is divisible by Z/-Z.
5334 auto ShareCommonDivisor = [&Q](Value *X, Value *Y, Value *Z,
5335 bool IsNegative) -> bool {
5336 const APInt *OffsetC;
5337 if (!match(V: Z, P: m_APInt(Res&: OffsetC)))
5338 return false;
5339
5340 // Fast path for Z == 1/-1.
5341 if (IsNegative ? OffsetC->isAllOnes() : OffsetC->isOne())
5342 return true;
5343
5344 APInt C = *OffsetC;
5345 if (IsNegative)
5346 C.negate();
5347 // Note: -INT_MIN is also negative.
5348 if (!C.isStrictlyPositive())
5349 return false;
5350
5351 return isMultipleOf(X, C, Q) && isMultipleOf(X: Y, C, Q);
5352 };
5353
5354 // TODO: The subtraction-related identities shown below also hold, but
5355 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
5356 // wouldn't happen even if they were implemented.
5357 //
5358 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
5359 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
5360 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
5361 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
5362
5363 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
5364 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
5365 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
5366 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
5367 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
5368 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
5369 if (A && NoOp0WrapProblem &&
5370 ShareCommonDivisor(A, Op1, B,
5371 ICmpInst::isLT(P: Pred) || ICmpInst::isGE(P: Pred)))
5372 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), A,
5373 Op1);
5374
5375 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
5376 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
5377 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
5378 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
5379 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
5380 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
5381 if (C && NoOp1WrapProblem &&
5382 ShareCommonDivisor(Op0, C, D,
5383 ICmpInst::isGT(P: Pred) || ICmpInst::isLE(P: Pred)))
5384 return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), Op0,
5385 C);
5386 }
5387
5388 // if C1 has greater magnitude than C2:
5389 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
5390 // s.t. C3 = C1 - C2
5391 //
5392 // if C2 has greater magnitude than C1:
5393 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
5394 // s.t. C3 = C2 - C1
5395 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
5396 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) {
5397 const APInt *AP1, *AP2;
5398 // TODO: Support non-uniform vectors.
5399 // TODO: Allow poison passthrough if B or D's element is poison.
5400 if (match(V: B, P: m_APIntAllowPoison(Res&: AP1)) &&
5401 match(V: D, P: m_APIntAllowPoison(Res&: AP2)) &&
5402 AP1->isNegative() == AP2->isNegative()) {
5403 APInt AP1Abs = AP1->abs();
5404 APInt AP2Abs = AP2->abs();
5405 if (AP1Abs.uge(RHS: AP2Abs)) {
5406 APInt Diff = *AP1 - *AP2;
5407 Constant *C3 = Constant::getIntegerValue(Ty: BO0->getType(), V: Diff);
5408 Value *NewAdd = Builder.CreateAdd(
5409 LHS: A, RHS: C3, Name: "", HasNUW: Op0HasNUW && Diff.ule(RHS: *AP1), HasNSW: Op0HasNSW);
5410 return new ICmpInst(Pred, NewAdd, C);
5411 } else {
5412 APInt Diff = *AP2 - *AP1;
5413 Constant *C3 = Constant::getIntegerValue(Ty: BO0->getType(), V: Diff);
5414 Value *NewAdd = Builder.CreateAdd(
5415 LHS: C, RHS: C3, Name: "", HasNUW: Op1HasNUW && Diff.ule(RHS: *AP2), HasNSW: Op1HasNSW);
5416 return new ICmpInst(Pred, A, NewAdd);
5417 }
5418 }
5419 Constant *Cst1, *Cst2;
5420 if (match(V: B, P: m_ImmConstant(C&: Cst1)) && match(V: D, P: m_ImmConstant(C&: Cst2)) &&
5421 ICmpInst::isEquality(P: Pred)) {
5422 Constant *Diff = ConstantExpr::getSub(C1: Cst2, C2: Cst1);
5423 Value *NewAdd = Builder.CreateAdd(LHS: C, RHS: Diff);
5424 return new ICmpInst(Pred, A, NewAdd);
5425 }
5426 }
5427
5428 // Analyze the case when either Op0 or Op1 is a sub instruction.
5429 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
5430 A = nullptr;
5431 B = nullptr;
5432 C = nullptr;
5433 D = nullptr;
5434 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
5435 A = BO0->getOperand(i_nocapture: 0);
5436 B = BO0->getOperand(i_nocapture: 1);
5437 }
5438 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
5439 C = BO1->getOperand(i_nocapture: 0);
5440 D = BO1->getOperand(i_nocapture: 1);
5441 }
5442
5443 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
5444 if (A == Op1 && NoOp0WrapProblem)
5445 return new ICmpInst(Pred, Constant::getNullValue(Ty: Op1->getType()), B);
5446 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
5447 if (C == Op0 && NoOp1WrapProblem)
5448 return new ICmpInst(Pred, D, Constant::getNullValue(Ty: Op0->getType()));
5449
5450 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
5451 // (A - B) u>/u<= A --> B u>/u<= A
5452 if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
5453 return new ICmpInst(Pred, B, A);
5454 // C u</u>= (C - D) --> C u</u>= D
5455 if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
5456 return new ICmpInst(Pred, C, D);
5457 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
5458 if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
5459 isKnownNonZero(V: B, Q))
5460 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(pred: Pred), B, A);
5461 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
5462 if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
5463 isKnownNonZero(V: D, Q))
5464 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(pred: Pred), C, D);
5465
5466 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
5467 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
5468 return new ICmpInst(Pred, A, C);
5469
5470 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
5471 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
5472 return new ICmpInst(Pred, D, B);
5473
5474 // icmp (0-X) < cst --> x > -cst
5475 if (NoOp0WrapProblem && ICmpInst::isSigned(predicate: Pred)) {
5476 Value *X;
5477 if (match(V: BO0, P: m_Neg(V: m_Value(V&: X))))
5478 if (Constant *RHSC = dyn_cast<Constant>(Val: Op1))
5479 if (RHSC->isNotMinSignedValue())
5480 return new ICmpInst(I.getSwappedPredicate(), X,
5481 ConstantExpr::getNeg(C: RHSC));
5482 }
5483
5484 if (Instruction *R = foldICmpXorXX(I, Q, IC&: *this))
5485 return R;
5486 if (Instruction *R = foldICmpOrXX(I, Q, IC&: *this))
5487 return R;
5488
5489 {
5490 // Try to remove shared multiplier from comparison:
5491 // X * Z pred Y * Z
5492 Value *X, *Y, *Z;
5493 if ((match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Z))) &&
5494 match(V: Op1, P: m_c_Mul(L: m_Specific(V: Z), R: m_Value(V&: Y)))) ||
5495 (match(V: Op0, P: m_Mul(L: m_Value(V&: Z), R: m_Value(V&: X))) &&
5496 match(V: Op1, P: m_c_Mul(L: m_Specific(V: Z), R: m_Value(V&: Y))))) {
5497 if (ICmpInst::isSigned(predicate: Pred)) {
5498 if (Op0HasNSW && Op1HasNSW) {
5499 KnownBits ZKnown = computeKnownBits(V: Z, CxtI: &I);
5500 if (ZKnown.isStrictlyPositive())
5501 return new ICmpInst(Pred, X, Y);
5502 if (ZKnown.isNegative())
5503 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), X, Y);
5504 Value *LessThan = simplifyICmpInst(Pred: ICmpInst::ICMP_SLT, LHS: X, RHS: Y,
5505 Q: SQ.getWithInstruction(I: &I));
5506 if (LessThan && match(V: LessThan, P: m_One()))
5507 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Z,
5508 Constant::getNullValue(Ty: Z->getType()));
5509 Value *GreaterThan = simplifyICmpInst(Pred: ICmpInst::ICMP_SGT, LHS: X, RHS: Y,
5510 Q: SQ.getWithInstruction(I: &I));
5511 if (GreaterThan && match(V: GreaterThan, P: m_One()))
5512 return new ICmpInst(Pred, Z, Constant::getNullValue(Ty: Z->getType()));
5513 }
5514 } else {
5515 bool NonZero;
5516 if (ICmpInst::isEquality(P: Pred)) {
5517 // If X != Y, fold (X *nw Z) eq/ne (Y *nw Z) -> Z eq/ne 0
5518 if (((Op0HasNSW && Op1HasNSW) || (Op0HasNUW && Op1HasNUW)) &&
5519 isKnownNonEqual(V1: X, V2: Y, SQ))
5520 return new ICmpInst(Pred, Z, Constant::getNullValue(Ty: Z->getType()));
5521
5522 KnownBits ZKnown = computeKnownBits(V: Z, CxtI: &I);
5523 // if Z % 2 != 0
5524 // X * Z eq/ne Y * Z -> X eq/ne Y
5525 if (ZKnown.countMaxTrailingZeros() == 0)
5526 return new ICmpInst(Pred, X, Y);
5527 NonZero = !ZKnown.One.isZero() || isKnownNonZero(V: Z, Q);
5528 // if Z != 0 and nsw(X * Z) and nsw(Y * Z)
5529 // X * Z eq/ne Y * Z -> X eq/ne Y
5530 if (NonZero && BO0 && BO1 && Op0HasNSW && Op1HasNSW)
5531 return new ICmpInst(Pred, X, Y);
5532 } else
5533 NonZero = isKnownNonZero(V: Z, Q);
5534
5535 // If Z != 0 and nuw(X * Z) and nuw(Y * Z)
5536 // X * Z u{lt/le/gt/ge}/eq/ne Y * Z -> X u{lt/le/gt/ge}/eq/ne Y
5537 if (NonZero && BO0 && BO1 && Op0HasNUW && Op1HasNUW)
5538 return new ICmpInst(Pred, X, Y);
5539 }
5540 }
5541 }
5542
5543 BinaryOperator *SRem = nullptr;
5544 // icmp (srem X, Y), Y
5545 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(i_nocapture: 1))
5546 SRem = BO0;
5547 // icmp Y, (srem X, Y)
5548 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
5549 Op0 == BO1->getOperand(i_nocapture: 1))
5550 SRem = BO1;
5551 if (SRem) {
5552 // We don't check hasOneUse to avoid increasing register pressure because
5553 // the value we use is the same value this instruction was already using.
5554 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(pred: Pred) : Pred) {
5555 default:
5556 break;
5557 case ICmpInst::ICMP_EQ:
5558 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
5559 case ICmpInst::ICMP_NE:
5560 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
5561 case ICmpInst::ICMP_SGT:
5562 case ICmpInst::ICMP_SGE:
5563 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(i_nocapture: 1),
5564 Constant::getAllOnesValue(Ty: SRem->getType()));
5565 case ICmpInst::ICMP_SLT:
5566 case ICmpInst::ICMP_SLE:
5567 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(i_nocapture: 1),
5568 Constant::getNullValue(Ty: SRem->getType()));
5569 }
5570 }
5571
5572 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
5573 (BO0->hasOneUse() || BO1->hasOneUse()) &&
5574 BO0->getOperand(i_nocapture: 1) == BO1->getOperand(i_nocapture: 1)) {
5575 switch (BO0->getOpcode()) {
5576 default:
5577 break;
5578 case Instruction::Add:
5579 case Instruction::Sub:
5580 case Instruction::Xor: {
5581 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
5582 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5583
5584 const APInt *C;
5585 if (match(V: BO0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) {
5586 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
5587 if (C->isSignMask()) {
5588 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
5589 return new ICmpInst(NewPred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5590 }
5591
5592 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
5593 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
5594 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
5595 NewPred = I.getSwappedPredicate(pred: NewPred);
5596 return new ICmpInst(NewPred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5597 }
5598 }
5599 break;
5600 }
5601 case Instruction::Mul: {
5602 if (!I.isEquality())
5603 break;
5604
5605 const APInt *C;
5606 if (match(V: BO0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero() &&
5607 !C->isOne()) {
5608 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
5609 // Mask = -1 >> count-trailing-zeros(C).
5610 if (unsigned TZs = C->countr_zero()) {
5611 Constant *Mask = ConstantInt::get(
5612 Ty: BO0->getType(),
5613 V: APInt::getLowBitsSet(numBits: C->getBitWidth(), loBitsSet: C->getBitWidth() - TZs));
5614 Value *And1 = Builder.CreateAnd(LHS: BO0->getOperand(i_nocapture: 0), RHS: Mask);
5615 Value *And2 = Builder.CreateAnd(LHS: BO1->getOperand(i_nocapture: 0), RHS: Mask);
5616 return new ICmpInst(Pred, And1, And2);
5617 }
5618 }
5619 break;
5620 }
5621 case Instruction::UDiv:
5622 case Instruction::LShr:
5623 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
5624 break;
5625 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5626
5627 case Instruction::SDiv:
5628 if (!(I.isEquality() || match(V: BO0->getOperand(i_nocapture: 1), P: m_NonNegative())) ||
5629 !BO0->isExact() || !BO1->isExact())
5630 break;
5631 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5632
5633 case Instruction::AShr:
5634 if (!BO0->isExact() || !BO1->isExact())
5635 break;
5636 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5637
5638 case Instruction::Shl: {
5639 bool NUW = Op0HasNUW && Op1HasNUW;
5640 bool NSW = Op0HasNSW && Op1HasNSW;
5641 if (!NUW && !NSW)
5642 break;
5643 if (!NSW && I.isSigned())
5644 break;
5645 return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0));
5646 }
5647 }
5648 }
5649
5650 if (BO0) {
5651 // Transform A & (L - 1) `ult` L --> L != 0
5652 auto LSubOne = m_Add(L: m_Specific(V: Op1), R: m_AllOnes());
5653 auto BitwiseAnd = m_c_And(L: m_Value(), R: LSubOne);
5654
5655 if (match(V: BO0, P: BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
5656 auto *Zero = Constant::getNullValue(Ty: BO0->getType());
5657 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
5658 }
5659 }
5660
5661 // For unsigned predicates / eq / ne:
5662 // icmp pred (x << 1), x --> icmp getSignedPredicate(pred) x, 0
5663 // icmp pred x, (x << 1) --> icmp getSignedPredicate(pred) 0, x
5664 if (!ICmpInst::isSigned(predicate: Pred)) {
5665 if (match(V: Op0, P: m_Shl(L: m_Specific(V: Op1), R: m_One())))
5666 return new ICmpInst(ICmpInst::getSignedPredicate(Pred), Op1,
5667 Constant::getNullValue(Ty: Op1->getType()));
5668 else if (match(V: Op1, P: m_Shl(L: m_Specific(V: Op0), R: m_One())))
5669 return new ICmpInst(ICmpInst::getSignedPredicate(Pred),
5670 Constant::getNullValue(Ty: Op0->getType()), Op0);
5671 }
5672
5673 if (Value *V = foldMultiplicationOverflowCheck(I))
5674 return replaceInstUsesWith(I, V);
5675
5676 if (Instruction *R = foldICmpAndXX(I, Q, IC&: *this))
5677 return R;
5678
5679 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
5680 return replaceInstUsesWith(I, V);
5681
5682 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
5683 return replaceInstUsesWith(I, V);
5684
5685 return nullptr;
5686}
5687
5688/// Fold icmp Pred min|max(X, Y), Z.
5689Instruction *InstCombinerImpl::foldICmpWithMinMax(Instruction &I,
5690 MinMaxIntrinsic *MinMax,
5691 Value *Z, CmpPredicate Pred) {
5692 Value *X = MinMax->getLHS();
5693 Value *Y = MinMax->getRHS();
5694 if (ICmpInst::isSigned(predicate: Pred) && !MinMax->isSigned())
5695 return nullptr;
5696 if (ICmpInst::isUnsigned(predicate: Pred) && MinMax->isSigned()) {
5697 // Revert the transform signed pred -> unsigned pred
5698 // TODO: We can flip the signedness of predicate if both operands of icmp
5699 // are negative.
5700 if (isKnownNonNegative(V: Z, SQ: SQ.getWithInstruction(I: &I)) &&
5701 isKnownNonNegative(V: MinMax, SQ: SQ.getWithInstruction(I: &I))) {
5702 Pred = ICmpInst::getFlippedSignednessPredicate(Pred);
5703 } else
5704 return nullptr;
5705 }
5706 SimplifyQuery Q = SQ.getWithInstruction(I: &I);
5707 auto IsCondKnownTrue = [](Value *Val) -> std::optional<bool> {
5708 if (!Val)
5709 return std::nullopt;
5710 if (match(V: Val, P: m_One()))
5711 return true;
5712 if (match(V: Val, P: m_Zero()))
5713 return false;
5714 return std::nullopt;
5715 };
5716 // Remove samesign here since it is illegal to keep it when we speculatively
5717 // execute comparisons. For example, `icmp samesign ult umax(X, -46), -32`
5718 // cannot be decomposed into `(icmp samesign ult X, -46) or (icmp samesign ult
5719 // -46, -32)`. `X` is allowed to be non-negative here.
5720 Pred = Pred.dropSameSign();
5721 auto CmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred, LHS: X, RHS: Z, Q));
5722 auto CmpYZ = IsCondKnownTrue(simplifyICmpInst(Pred, LHS: Y, RHS: Z, Q));
5723 if (!CmpXZ.has_value() && !CmpYZ.has_value())
5724 return nullptr;
5725 if (!CmpXZ.has_value()) {
5726 std::swap(a&: X, b&: Y);
5727 std::swap(lhs&: CmpXZ, rhs&: CmpYZ);
5728 }
5729
5730 auto FoldIntoCmpYZ = [&]() -> Instruction * {
5731 if (CmpYZ.has_value())
5732 return replaceInstUsesWith(I, V: ConstantInt::getBool(Ty: I.getType(), V: *CmpYZ));
5733 return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: Y, S2: Z);
5734 };
5735
5736 switch (Pred) {
5737 case ICmpInst::ICMP_EQ:
5738 case ICmpInst::ICMP_NE: {
5739 // If X == Z:
5740 // Expr Result
5741 // min(X, Y) == Z X <= Y
5742 // max(X, Y) == Z X >= Y
5743 // min(X, Y) != Z X > Y
5744 // max(X, Y) != Z X < Y
5745 if ((Pred == ICmpInst::ICMP_EQ) == *CmpXZ) {
5746 ICmpInst::Predicate NewPred =
5747 ICmpInst::getNonStrictPredicate(pred: MinMax->getPredicate());
5748 if (Pred == ICmpInst::ICMP_NE)
5749 NewPred = ICmpInst::getInversePredicate(pred: NewPred);
5750 return ICmpInst::Create(Op: Instruction::ICmp, Pred: NewPred, S1: X, S2: Y);
5751 }
5752 // Otherwise (X != Z):
5753 ICmpInst::Predicate NewPred = MinMax->getPredicate();
5754 auto MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred: NewPred, LHS: X, RHS: Z, Q));
5755 if (!MinMaxCmpXZ.has_value()) {
5756 std::swap(a&: X, b&: Y);
5757 std::swap(lhs&: CmpXZ, rhs&: CmpYZ);
5758 // Re-check pre-condition X != Z
5759 if (!CmpXZ.has_value() || (Pred == ICmpInst::ICMP_EQ) == *CmpXZ)
5760 break;
5761 MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred: NewPred, LHS: X, RHS: Z, Q));
5762 }
5763 if (!MinMaxCmpXZ.has_value())
5764 break;
5765 if (*MinMaxCmpXZ) {
5766 // Expr Fact Result
5767 // min(X, Y) == Z X < Z false
5768 // max(X, Y) == Z X > Z false
5769 // min(X, Y) != Z X < Z true
5770 // max(X, Y) != Z X > Z true
5771 return replaceInstUsesWith(
5772 I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred == ICmpInst::ICMP_NE));
5773 } else {
5774 // Expr Fact Result
5775 // min(X, Y) == Z X > Z Y == Z
5776 // max(X, Y) == Z X < Z Y == Z
5777 // min(X, Y) != Z X > Z Y != Z
5778 // max(X, Y) != Z X < Z Y != Z
5779 return FoldIntoCmpYZ();
5780 }
5781 break;
5782 }
5783 case ICmpInst::ICMP_SLT:
5784 case ICmpInst::ICMP_ULT:
5785 case ICmpInst::ICMP_SLE:
5786 case ICmpInst::ICMP_ULE:
5787 case ICmpInst::ICMP_SGT:
5788 case ICmpInst::ICMP_UGT:
5789 case ICmpInst::ICMP_SGE:
5790 case ICmpInst::ICMP_UGE: {
5791 bool IsSame = MinMax->getPredicate() == ICmpInst::getStrictPredicate(pred: Pred);
5792 if (*CmpXZ) {
5793 if (IsSame) {
5794 // Expr Fact Result
5795 // min(X, Y) < Z X < Z true
5796 // min(X, Y) <= Z X <= Z true
5797 // max(X, Y) > Z X > Z true
5798 // max(X, Y) >= Z X >= Z true
5799 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
5800 } else {
5801 // Expr Fact Result
5802 // max(X, Y) < Z X < Z Y < Z
5803 // max(X, Y) <= Z X <= Z Y <= Z
5804 // min(X, Y) > Z X > Z Y > Z
5805 // min(X, Y) >= Z X >= Z Y >= Z
5806 return FoldIntoCmpYZ();
5807 }
5808 } else {
5809 if (IsSame) {
5810 // Expr Fact Result
5811 // min(X, Y) < Z X >= Z Y < Z
5812 // min(X, Y) <= Z X > Z Y <= Z
5813 // max(X, Y) > Z X <= Z Y > Z
5814 // max(X, Y) >= Z X < Z Y >= Z
5815 return FoldIntoCmpYZ();
5816 } else {
5817 // Expr Fact Result
5818 // max(X, Y) < Z X >= Z false
5819 // max(X, Y) <= Z X > Z false
5820 // min(X, Y) > Z X <= Z false
5821 // min(X, Y) >= Z X < Z false
5822 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
5823 }
5824 }
5825 break;
5826 }
5827 default:
5828 break;
5829 }
5830
5831 return nullptr;
5832}
5833
5834/// Match and fold patterns like:
5835/// icmp eq/ne X, min(max(X, Lo), Hi)
5836/// which represents a range check and can be repsented as a ConstantRange.
5837///
5838/// For icmp eq, build ConstantRange [Lo, Hi + 1) and convert to:
5839/// (X - Lo) u< (Hi + 1 - Lo)
5840/// For icmp ne, build ConstantRange [Hi + 1, Lo) and convert to:
5841/// (X - (Hi + 1)) u< (Lo - (Hi + 1))
5842Instruction *InstCombinerImpl::foldICmpWithClamp(ICmpInst &I, Value *X,
5843 MinMaxIntrinsic *Min) {
5844 if (!I.isEquality() || !Min->hasOneUse() || !Min->isMin())
5845 return nullptr;
5846
5847 const APInt *Lo = nullptr, *Hi = nullptr;
5848 if (Min->isSigned()) {
5849 if (!match(V: Min->getLHS(), P: m_OneUse(SubPattern: m_SMax(L: m_Specific(V: X), R: m_APInt(Res&: Lo)))) ||
5850 !match(V: Min->getRHS(), P: m_APInt(Res&: Hi)) || !Lo->slt(RHS: *Hi))
5851 return nullptr;
5852 } else {
5853 if (!match(V: Min->getLHS(), P: m_OneUse(SubPattern: m_UMax(L: m_Specific(V: X), R: m_APInt(Res&: Lo)))) ||
5854 !match(V: Min->getRHS(), P: m_APInt(Res&: Hi)) || !Lo->ult(RHS: *Hi))
5855 return nullptr;
5856 }
5857
5858 ConstantRange CR = ConstantRange::getNonEmpty(Lower: *Lo, Upper: *Hi + 1);
5859 ICmpInst::Predicate Pred;
5860 APInt C, Offset;
5861 if (I.getPredicate() == ICmpInst::ICMP_EQ)
5862 CR.getEquivalentICmp(Pred, RHS&: C, Offset);
5863 else
5864 CR.inverse().getEquivalentICmp(Pred, RHS&: C, Offset);
5865
5866 if (!Offset.isZero())
5867 X = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: Offset));
5868
5869 return replaceInstUsesWith(
5870 I, V: Builder.CreateICmp(P: Pred, LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: C)));
5871}
5872
5873// Canonicalize checking for a power-of-2-or-zero value:
5874static Instruction *foldICmpPow2Test(ICmpInst &I,
5875 InstCombiner::BuilderTy &Builder) {
5876 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
5877 const CmpInst::Predicate Pred = I.getPredicate();
5878 Value *A = nullptr;
5879 bool CheckIs;
5880 if (I.isEquality()) {
5881 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
5882 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
5883 if (!match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Add(L: m_Value(V&: A), R: m_AllOnes()),
5884 R: m_Deferred(V: A)))) ||
5885 !match(V: Op1, P: m_ZeroInt()))
5886 A = nullptr;
5887
5888 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
5889 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
5890 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Neg(V: m_Specific(V: Op1)), R: m_Specific(V: Op1)))))
5891 A = Op1;
5892 else if (match(V: Op1,
5893 P: m_OneUse(SubPattern: m_c_And(L: m_Neg(V: m_Specific(V: Op0)), R: m_Specific(V: Op0)))))
5894 A = Op0;
5895
5896 CheckIs = Pred == ICmpInst::ICMP_EQ;
5897 } else if (ICmpInst::isUnsigned(predicate: Pred)) {
5898 // (A ^ (A-1)) u>= A --> ctpop(A) < 2 (two commuted variants)
5899 // ((A-1) ^ A) u< A --> ctpop(A) > 1 (two commuted variants)
5900
5901 if ((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
5902 match(V: Op0, P: m_OneUse(SubPattern: m_c_Xor(L: m_Add(L: m_Specific(V: Op1), R: m_AllOnes()),
5903 R: m_Specific(V: Op1))))) {
5904 A = Op1;
5905 CheckIs = Pred == ICmpInst::ICMP_UGE;
5906 } else if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
5907 match(V: Op1, P: m_OneUse(SubPattern: m_c_Xor(L: m_Add(L: m_Specific(V: Op0), R: m_AllOnes()),
5908 R: m_Specific(V: Op0))))) {
5909 A = Op0;
5910 CheckIs = Pred == ICmpInst::ICMP_ULE;
5911 }
5912 }
5913
5914 if (A) {
5915 Type *Ty = A->getType();
5916 CallInst *CtPop = Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V: A);
5917 return CheckIs ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop,
5918 ConstantInt::get(Ty, V: 2))
5919 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop,
5920 ConstantInt::get(Ty, V: 1));
5921 }
5922
5923 return nullptr;
5924}
5925
5926/// Find all possible pairs (BinOp, RHS) that BinOp V, RHS can be simplified.
5927using OffsetOp = std::pair<Instruction::BinaryOps, Value *>;
5928static void collectOffsetOp(Value *V, SmallVectorImpl<OffsetOp> &Offsets,
5929 bool AllowRecursion) {
5930 Instruction *Inst = dyn_cast<Instruction>(Val: V);
5931 if (!Inst || !Inst->hasOneUse())
5932 return;
5933
5934 switch (Inst->getOpcode()) {
5935 case Instruction::Add:
5936 Offsets.emplace_back(Args: Instruction::Sub, Args: Inst->getOperand(i: 1));
5937 Offsets.emplace_back(Args: Instruction::Sub, Args: Inst->getOperand(i: 0));
5938 break;
5939 case Instruction::Sub:
5940 Offsets.emplace_back(Args: Instruction::Add, Args: Inst->getOperand(i: 1));
5941 break;
5942 case Instruction::Xor:
5943 Offsets.emplace_back(Args: Instruction::Xor, Args: Inst->getOperand(i: 1));
5944 Offsets.emplace_back(Args: Instruction::Xor, Args: Inst->getOperand(i: 0));
5945 break;
5946 case Instruction::Shl:
5947 if (Inst->hasNoSignedWrap())
5948 Offsets.emplace_back(Args: Instruction::AShr, Args: Inst->getOperand(i: 1));
5949 if (Inst->hasNoUnsignedWrap())
5950 Offsets.emplace_back(Args: Instruction::LShr, Args: Inst->getOperand(i: 1));
5951 break;
5952 case Instruction::Select:
5953 if (AllowRecursion) {
5954 collectOffsetOp(V: Inst->getOperand(i: 1), Offsets, /*AllowRecursion=*/false);
5955 collectOffsetOp(V: Inst->getOperand(i: 2), Offsets, /*AllowRecursion=*/false);
5956 }
5957 break;
5958 default:
5959 break;
5960 }
5961}
5962
5963enum class OffsetKind { Invalid, Value, Select };
5964
5965struct OffsetResult {
5966 OffsetKind Kind;
5967 Value *V0, *V1, *V2;
5968 Instruction *MDFrom;
5969
5970 static OffsetResult invalid() {
5971 return {.Kind: OffsetKind::Invalid, .V0: nullptr, .V1: nullptr, .V2: nullptr, .MDFrom: nullptr};
5972 }
5973 static OffsetResult value(Value *V) {
5974 return {.Kind: OffsetKind::Value, .V0: V, .V1: nullptr, .V2: nullptr, .MDFrom: nullptr};
5975 }
5976 static OffsetResult select(Value *Cond, Value *TrueV, Value *FalseV,
5977 Instruction *MDFrom) {
5978 return {.Kind: OffsetKind::Select, .V0: Cond, .V1: TrueV, .V2: FalseV, .MDFrom: MDFrom};
5979 }
5980 bool isValid() const { return Kind != OffsetKind::Invalid; }
5981 Value *materialize(InstCombiner::BuilderTy &Builder) const {
5982 switch (Kind) {
5983 case OffsetKind::Invalid:
5984 llvm_unreachable("Invalid offset result");
5985 case OffsetKind::Value:
5986 return V0;
5987 case OffsetKind::Select:
5988 return Builder.CreateSelect(
5989 C: V0, True: V1, False: V2, Name: "", MDFrom: ProfcheckDisableMetadataFixes ? nullptr : MDFrom);
5990 }
5991 llvm_unreachable("Unknown OffsetKind enum");
5992 }
5993};
5994
5995/// Offset both sides of an equality icmp to see if we can save some
5996/// instructions: icmp eq/ne X, Y -> icmp eq/ne X op Z, Y op Z.
5997/// Note: This operation should not introduce poison.
5998static Instruction *foldICmpEqualityWithOffset(ICmpInst &I,
5999 InstCombiner::BuilderTy &Builder,
6000 const SimplifyQuery &SQ) {
6001 assert(I.isEquality() && "Expected an equality icmp");
6002 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
6003 if (!Op0->getType()->isIntOrIntVectorTy())
6004 return nullptr;
6005
6006 SmallVector<OffsetOp, 4> OffsetOps;
6007 collectOffsetOp(V: Op0, Offsets&: OffsetOps, /*AllowRecursion=*/true);
6008 collectOffsetOp(V: Op1, Offsets&: OffsetOps, /*AllowRecursion=*/true);
6009
6010 auto ApplyOffsetImpl = [&](Value *V, unsigned BinOpc, Value *RHS) -> Value * {
6011 switch (BinOpc) {
6012 // V = shl nsw X, RHS => X = ashr V, RHS
6013 case Instruction::AShr: {
6014 const APInt *CV, *CRHS;
6015 if (!(match(V, P: m_APInt(Res&: CV)) && match(V: RHS, P: m_APInt(Res&: CRHS)) &&
6016 CV->ashr(ShiftAmt: *CRHS).shl(ShiftAmt: *CRHS) == *CV) &&
6017 !match(V, P: m_NSWShl(L: m_Value(), R: m_Specific(V: RHS))))
6018 return nullptr;
6019 break;
6020 }
6021 // V = shl nuw X, RHS => X = lshr V, RHS
6022 case Instruction::LShr: {
6023 const APInt *CV, *CRHS;
6024 if (!(match(V, P: m_APInt(Res&: CV)) && match(V: RHS, P: m_APInt(Res&: CRHS)) &&
6025 CV->lshr(ShiftAmt: *CRHS).shl(ShiftAmt: *CRHS) == *CV) &&
6026 !match(V, P: m_NUWShl(L: m_Value(), R: m_Specific(V: RHS))))
6027 return nullptr;
6028 break;
6029 }
6030 default:
6031 break;
6032 }
6033
6034 Value *Simplified = simplifyBinOp(Opcode: BinOpc, LHS: V, RHS, Q: SQ);
6035 if (!Simplified)
6036 return nullptr;
6037 // Reject constant expressions as they don't simplify things.
6038 if (isa<Constant>(Val: Simplified) && !match(V: Simplified, P: m_ImmConstant()))
6039 return nullptr;
6040 // Check if the transformation introduces poison.
6041 return impliesPoison(ValAssumedPoison: RHS, V) ? Simplified : nullptr;
6042 };
6043
6044 auto ApplyOffset = [&](Value *V, unsigned BinOpc,
6045 Value *RHS) -> OffsetResult {
6046 if (auto *Sel = dyn_cast<SelectInst>(Val: V)) {
6047 if (!Sel->hasOneUse())
6048 return OffsetResult::invalid();
6049 Value *TrueVal = ApplyOffsetImpl(Sel->getTrueValue(), BinOpc, RHS);
6050 if (!TrueVal)
6051 return OffsetResult::invalid();
6052 Value *FalseVal = ApplyOffsetImpl(Sel->getFalseValue(), BinOpc, RHS);
6053 if (!FalseVal)
6054 return OffsetResult::invalid();
6055 return OffsetResult::select(Cond: Sel->getCondition(), TrueV: TrueVal, FalseV: FalseVal, MDFrom: Sel);
6056 }
6057 if (Value *Simplified = ApplyOffsetImpl(V, BinOpc, RHS))
6058 return OffsetResult::value(V: Simplified);
6059 return OffsetResult::invalid();
6060 };
6061
6062 for (auto [BinOp, RHS] : OffsetOps) {
6063 auto BinOpc = static_cast<unsigned>(BinOp);
6064
6065 auto Op0Result = ApplyOffset(Op0, BinOpc, RHS);
6066 if (!Op0Result.isValid())
6067 continue;
6068 auto Op1Result = ApplyOffset(Op1, BinOpc, RHS);
6069 if (!Op1Result.isValid())
6070 continue;
6071
6072 Value *NewLHS = Op0Result.materialize(Builder);
6073 Value *NewRHS = Op1Result.materialize(Builder);
6074 return new ICmpInst(I.getPredicate(), NewLHS, NewRHS);
6075 }
6076
6077 return nullptr;
6078}
6079
6080Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
6081 if (!I.isEquality())
6082 return nullptr;
6083
6084 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
6085 const CmpInst::Predicate Pred = I.getPredicate();
6086 Value *A, *B, *C, *D;
6087 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)))) {
6088 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6089 Value *OtherVal = A == Op1 ? B : A;
6090 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(Ty: A->getType()));
6091 }
6092
6093 if (match(V: Op1, P: m_Xor(L: m_Value(V&: C), R: m_Value(V&: D)))) {
6094 // A^c1 == C^c2 --> A == C^(c1^c2)
6095 ConstantInt *C1, *C2;
6096 if (match(V: B, P: m_ConstantInt(CI&: C1)) && match(V: D, P: m_ConstantInt(CI&: C2)) &&
6097 Op1->hasOneUse()) {
6098 Constant *NC = Builder.getInt(AI: C1->getValue() ^ C2->getValue());
6099 Value *Xor = Builder.CreateXor(LHS: C, RHS: NC);
6100 return new ICmpInst(Pred, A, Xor);
6101 }
6102
6103 // A^B == A^D -> B == D
6104 if (A == C)
6105 return new ICmpInst(Pred, B, D);
6106 if (A == D)
6107 return new ICmpInst(Pred, B, C);
6108 if (B == C)
6109 return new ICmpInst(Pred, A, D);
6110 if (B == D)
6111 return new ICmpInst(Pred, A, C);
6112 }
6113 }
6114
6115 if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) && (A == Op0 || B == Op0)) {
6116 // A == (A^B) -> B == 0
6117 Value *OtherVal = A == Op0 ? B : A;
6118 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(Ty: A->getType()));
6119 }
6120
6121 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6122 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
6123 match(V: Op1, P: m_And(L: m_Value(V&: C), R: m_Value(V&: D)))) {
6124 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
6125
6126 if (A == C) {
6127 X = B;
6128 Y = D;
6129 Z = A;
6130 } else if (A == D) {
6131 X = B;
6132 Y = C;
6133 Z = A;
6134 } else if (B == C) {
6135 X = A;
6136 Y = D;
6137 Z = B;
6138 } else if (B == D) {
6139 X = A;
6140 Y = C;
6141 Z = B;
6142 }
6143
6144 if (X) {
6145 // If X^Y is a negative power of two, then `icmp eq/ne (Z & NegP2), 0`
6146 // will fold to `icmp ult/uge Z, -NegP2` incurringb no additional
6147 // instructions.
6148 const APInt *C0, *C1;
6149 bool XorIsNegP2 = match(V: X, P: m_APInt(Res&: C0)) && match(V: Y, P: m_APInt(Res&: C1)) &&
6150 (*C0 ^ *C1).isNegatedPowerOf2();
6151
6152 // If either Op0/Op1 are both one use or X^Y will constant fold and one of
6153 // Op0/Op1 are one use, proceed. In those cases we are instruction neutral
6154 // but `icmp eq/ne A, 0` is easier to analyze than `icmp eq/ne A, B`.
6155 int UseCnt =
6156 int(Op0->hasOneUse()) + int(Op1->hasOneUse()) +
6157 (int(match(V: X, P: m_ImmConstant()) && match(V: Y, P: m_ImmConstant())));
6158 if (XorIsNegP2 || UseCnt >= 2) {
6159 // Build (X^Y) & Z
6160 Op1 = Builder.CreateXor(LHS: X, RHS: Y);
6161 Op1 = Builder.CreateAnd(LHS: Op1, RHS: Z);
6162 return new ICmpInst(Pred, Op1, Constant::getNullValue(Ty: Op1->getType()));
6163 }
6164 }
6165 }
6166
6167 {
6168 // Similar to above, but specialized for constant because invert is needed:
6169 // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
6170 Value *X, *Y;
6171 Constant *C;
6172 if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Constant(C)))) &&
6173 match(V: Op1, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: Y), R: m_Specific(V: C))))) {
6174 Value *Xor = Builder.CreateXor(LHS: X, RHS: Y);
6175 Value *And = Builder.CreateAnd(LHS: Xor, RHS: ConstantExpr::getNot(C));
6176 return new ICmpInst(Pred, And, Constant::getNullValue(Ty: And->getType()));
6177 }
6178 }
6179
6180 if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: A))) &&
6181 (Op0->hasOneUse() || Op1->hasOneUse())) {
6182 // (B & (Pow2C-1)) == zext A --> A == trunc B
6183 // (B & (Pow2C-1)) != zext A --> A != trunc B
6184 const APInt *MaskC;
6185 if (match(V: Op0, P: m_And(L: m_Value(V&: B), R: m_LowBitMask(V&: MaskC))) &&
6186 MaskC->countr_one() == A->getType()->getScalarSizeInBits())
6187 return new ICmpInst(Pred, A, Builder.CreateTrunc(V: B, DestTy: A->getType()));
6188 }
6189
6190 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
6191 // For lshr and ashr pairs.
6192 const APInt *AP1, *AP2;
6193 if ((match(V: Op0, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: AP1)))) &&
6194 match(V: Op1, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: B), R: m_APIntAllowPoison(Res&: AP2))))) ||
6195 (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: AP1)))) &&
6196 match(V: Op1, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: B), R: m_APIntAllowPoison(Res&: AP2)))))) {
6197 if (*AP1 != *AP2)
6198 return nullptr;
6199 unsigned TypeBits = AP1->getBitWidth();
6200 unsigned ShAmt = AP1->getLimitedValue(Limit: TypeBits);
6201 if (ShAmt < TypeBits && ShAmt != 0) {
6202 ICmpInst::Predicate NewPred =
6203 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6204 Value *Xor = Builder.CreateXor(LHS: A, RHS: B, Name: I.getName() + ".unshifted");
6205 APInt CmpVal = APInt::getOneBitSet(numBits: TypeBits, BitNo: ShAmt);
6206 return new ICmpInst(NewPred, Xor, ConstantInt::get(Ty: A->getType(), V: CmpVal));
6207 }
6208 }
6209
6210 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
6211 ConstantInt *Cst1;
6212 if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: A), R: m_ConstantInt(CI&: Cst1)))) &&
6213 match(V: Op1, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: B), R: m_Specific(V: Cst1))))) {
6214 unsigned TypeBits = Cst1->getBitWidth();
6215 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(Limit: TypeBits);
6216 if (ShAmt < TypeBits && ShAmt != 0) {
6217 Value *Xor = Builder.CreateXor(LHS: A, RHS: B, Name: I.getName() + ".unshifted");
6218 APInt AndVal = APInt::getLowBitsSet(numBits: TypeBits, loBitsSet: TypeBits - ShAmt);
6219 Value *And =
6220 Builder.CreateAnd(LHS: Xor, RHS: Builder.getInt(AI: AndVal), Name: I.getName() + ".mask");
6221 return new ICmpInst(Pred, And, Constant::getNullValue(Ty: Cst1->getType()));
6222 }
6223 }
6224
6225 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
6226 // "icmp (and X, mask), cst"
6227 uint64_t ShAmt = 0;
6228 if (Op0->hasOneUse() &&
6229 match(V: Op0, P: m_Trunc(Op: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: A), R: m_ConstantInt(V&: ShAmt))))) &&
6230 match(V: Op1, P: m_ConstantInt(CI&: Cst1)) &&
6231 // Only do this when A has multiple uses. This is most important to do
6232 // when it exposes other optimizations.
6233 !A->hasOneUse()) {
6234 unsigned ASize = cast<IntegerType>(Val: A->getType())->getPrimitiveSizeInBits();
6235
6236 if (ShAmt < ASize) {
6237 APInt MaskV =
6238 APInt::getLowBitsSet(numBits: ASize, loBitsSet: Op0->getType()->getPrimitiveSizeInBits());
6239 MaskV <<= ShAmt;
6240
6241 APInt CmpV = Cst1->getValue().zext(width: ASize);
6242 CmpV <<= ShAmt;
6243
6244 Value *Mask = Builder.CreateAnd(LHS: A, RHS: Builder.getInt(AI: MaskV));
6245 return new ICmpInst(Pred, Mask, Builder.getInt(AI: CmpV));
6246 }
6247 }
6248
6249 if (Instruction *ICmp = foldICmpIntrinsicWithIntrinsic(Cmp&: I, Builder))
6250 return ICmp;
6251
6252 // Match icmp eq (trunc (lshr A, BW), (ashr (trunc A), BW-1)), which checks
6253 // the top BW/2 + 1 bits are all the same. Create "A >=s INT_MIN && A <=s
6254 // INT_MAX", which we generate as "icmp ult (add A, 2^(BW-1)), 2^BW" to skip a
6255 // few steps of instcombine.
6256 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
6257 if (match(V: Op0, P: m_AShr(L: m_Trunc(Op: m_Value(V&: A)), R: m_SpecificInt(V: BitWidth - 1))) &&
6258 match(V: Op1, P: m_Trunc(Op: m_LShr(L: m_Specific(V: A), R: m_SpecificInt(V: BitWidth)))) &&
6259 A->getType()->getScalarSizeInBits() == BitWidth * 2 &&
6260 (I.getOperand(i_nocapture: 0)->hasOneUse() || I.getOperand(i_nocapture: 1)->hasOneUse())) {
6261 APInt C = APInt::getOneBitSet(numBits: BitWidth * 2, BitNo: BitWidth - 1);
6262 Value *Add = Builder.CreateAdd(LHS: A, RHS: ConstantInt::get(Ty: A->getType(), V: C));
6263 return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT
6264 : ICmpInst::ICMP_UGE,
6265 Add, ConstantInt::get(Ty: A->getType(), V: C.shl(shiftAmt: 1)));
6266 }
6267
6268 // Canonicalize:
6269 // Assume B_Pow2 != 0
6270 // 1. A & B_Pow2 != B_Pow2 -> A & B_Pow2 == 0
6271 // 2. A & B_Pow2 == B_Pow2 -> A & B_Pow2 != 0
6272 if (match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value())) &&
6273 isKnownToBeAPowerOfTwo(V: Op1, /* OrZero */ false, CxtI: &I))
6274 return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op0,
6275 ConstantInt::getNullValue(Ty: Op0->getType()));
6276
6277 if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value())) &&
6278 isKnownToBeAPowerOfTwo(V: Op0, /* OrZero */ false, CxtI: &I))
6279 return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op1,
6280 ConstantInt::getNullValue(Ty: Op1->getType()));
6281
6282 // Canonicalize:
6283 // icmp eq/ne X, OneUse(rotate-right(X))
6284 // -> icmp eq/ne X, rotate-left(X)
6285 // We generally try to convert rotate-right -> rotate-left, this just
6286 // canonicalizes another case.
6287 if (match(V: &I, P: m_c_ICmp(L: m_Value(V&: A),
6288 R: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::fshr>(
6289 Op0: m_Deferred(V: A), Op1: m_Deferred(V: A), Op2: m_Value(V&: B))))))
6290 return new ICmpInst(
6291 Pred, A,
6292 Builder.CreateIntrinsic(RetTy: Op0->getType(), ID: Intrinsic::fshl, Args: {A, A, B}));
6293
6294 // Canonicalize:
6295 // icmp eq/ne OneUse(A ^ Cst), B --> icmp eq/ne (A ^ B), Cst
6296 Constant *Cst;
6297 if (match(V: &I, P: m_c_ICmp(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: A), R: m_ImmConstant(C&: Cst))),
6298 R: m_CombineAnd(L: m_Value(V&: B), R: m_Unless(M: m_ImmConstant())))))
6299 return new ICmpInst(Pred, Builder.CreateXor(LHS: A, RHS: B), Cst);
6300
6301 {
6302 // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
6303 auto m_Matcher =
6304 m_CombineOr(L: m_CombineOr(L: m_c_Add(L: m_Value(V&: B), R: m_Deferred(V: A)),
6305 R: m_c_Xor(L: m_Value(V&: B), R: m_Deferred(V: A))),
6306 R: m_Sub(L: m_Value(V&: B), R: m_Deferred(V: A)));
6307 std::optional<bool> IsZero = std::nullopt;
6308 if (match(V: &I, P: m_c_ICmp(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_Matcher)),
6309 R: m_Deferred(V: A))))
6310 IsZero = false;
6311 // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
6312 else if (match(V: &I,
6313 P: m_ICmp(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_Matcher)), R: m_Zero())))
6314 IsZero = true;
6315
6316 if (IsZero && isKnownToBeAPowerOfTwo(V: A, /* OrZero */ true, CxtI: &I))
6317 // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
6318 // -> (icmp eq/ne (and X, P2), 0)
6319 // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
6320 // -> (icmp eq/ne (and X, P2), P2)
6321 return new ICmpInst(Pred, Builder.CreateAnd(LHS: B, RHS: A),
6322 *IsZero ? A
6323 : ConstantInt::getNullValue(Ty: A->getType()));
6324 }
6325
6326 if (auto *Res = foldICmpEqualityWithOffset(
6327 I, Builder, SQ: getSimplifyQuery().getWithInstruction(I: &I)))
6328 return Res;
6329
6330 return nullptr;
6331}
6332
6333Instruction *InstCombinerImpl::foldICmpWithTrunc(ICmpInst &ICmp) {
6334 ICmpInst::Predicate Pred = ICmp.getPredicate();
6335 Value *Op0 = ICmp.getOperand(i_nocapture: 0), *Op1 = ICmp.getOperand(i_nocapture: 1);
6336
6337 // Try to canonicalize trunc + compare-to-constant into a mask + cmp.
6338 // The trunc masks high bits while the compare may effectively mask low bits.
6339 Value *X;
6340 const APInt *C;
6341 if (!match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: X)))) || !match(V: Op1, P: m_APInt(Res&: C)))
6342 return nullptr;
6343
6344 // This matches patterns corresponding to tests of the signbit as well as:
6345 // (trunc X) pred C2 --> (X & Mask) == C
6346 if (auto Res = decomposeBitTestICmp(LHS: Op0, RHS: Op1, Pred, /*LookThroughTrunc=*/true,
6347 /*AllowNonZeroC=*/true)) {
6348 Value *And = Builder.CreateAnd(LHS: Res->X, RHS: Res->Mask);
6349 Constant *C = ConstantInt::get(Ty: Res->X->getType(), V: Res->C);
6350 return new ICmpInst(Res->Pred, And, C);
6351 }
6352
6353 unsigned SrcBits = X->getType()->getScalarSizeInBits();
6354 if (auto *II = dyn_cast<IntrinsicInst>(Val: X)) {
6355 if (II->getIntrinsicID() == Intrinsic::cttz ||
6356 II->getIntrinsicID() == Intrinsic::ctlz) {
6357 unsigned MaxRet = SrcBits;
6358 // If the "is_zero_poison" argument is set, then we know at least
6359 // one bit is set in the input, so the result is always at least one
6360 // less than the full bitwidth of that input.
6361 if (match(V: II->getArgOperand(i: 1), P: m_One()))
6362 MaxRet--;
6363
6364 // Make sure the destination is wide enough to hold the largest output of
6365 // the intrinsic.
6366 if (llvm::Log2_32(Value: MaxRet) + 1 <= Op0->getType()->getScalarSizeInBits())
6367 if (Instruction *I =
6368 foldICmpIntrinsicWithConstant(Cmp&: ICmp, II, C: C->zext(width: SrcBits)))
6369 return I;
6370 }
6371 }
6372
6373 return nullptr;
6374}
6375
6376Instruction *InstCombinerImpl::foldICmpWithZextOrSext(ICmpInst &ICmp) {
6377 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
6378 auto *CastOp0 = cast<CastInst>(Val: ICmp.getOperand(i_nocapture: 0));
6379 Value *X;
6380 if (!match(V: CastOp0, P: m_ZExtOrSExt(Op: m_Value(V&: X))))
6381 return nullptr;
6382
6383 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
6384 bool IsSignedCmp = ICmp.isSigned();
6385
6386 // icmp Pred (ext X), (ext Y)
6387 Value *Y;
6388 if (match(V: ICmp.getOperand(i_nocapture: 1), P: m_ZExtOrSExt(Op: m_Value(V&: Y)))) {
6389 bool IsZext0 = isa<ZExtInst>(Val: ICmp.getOperand(i_nocapture: 0));
6390 bool IsZext1 = isa<ZExtInst>(Val: ICmp.getOperand(i_nocapture: 1));
6391
6392 if (IsZext0 != IsZext1) {
6393 // If X and Y and both i1
6394 // (icmp eq/ne (zext X) (sext Y))
6395 // eq -> (icmp eq (or X, Y), 0)
6396 // ne -> (icmp ne (or X, Y), 0)
6397 if (ICmp.isEquality() && X->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
6398 Y->getType()->isIntOrIntVectorTy(BitWidth: 1))
6399 return new ICmpInst(ICmp.getPredicate(), Builder.CreateOr(LHS: X, RHS: Y),
6400 Constant::getNullValue(Ty: X->getType()));
6401
6402 // If we have mismatched casts and zext has the nneg flag, we can
6403 // treat the "zext nneg" as "sext". Otherwise, we cannot fold and quit.
6404
6405 auto *NonNegInst0 = dyn_cast<PossiblyNonNegInst>(Val: ICmp.getOperand(i_nocapture: 0));
6406 auto *NonNegInst1 = dyn_cast<PossiblyNonNegInst>(Val: ICmp.getOperand(i_nocapture: 1));
6407
6408 bool IsNonNeg0 = NonNegInst0 && NonNegInst0->hasNonNeg();
6409 bool IsNonNeg1 = NonNegInst1 && NonNegInst1->hasNonNeg();
6410
6411 if ((IsZext0 && IsNonNeg0) || (IsZext1 && IsNonNeg1))
6412 IsSignedExt = true;
6413 else
6414 return nullptr;
6415 }
6416
6417 // Not an extension from the same type?
6418 Type *XTy = X->getType(), *YTy = Y->getType();
6419 if (XTy != YTy) {
6420 // One of the casts must have one use because we are creating a new cast.
6421 if (!ICmp.getOperand(i_nocapture: 0)->hasOneUse() && !ICmp.getOperand(i_nocapture: 1)->hasOneUse())
6422 return nullptr;
6423 // Extend the narrower operand to the type of the wider operand.
6424 CastInst::CastOps CastOpcode =
6425 IsSignedExt ? Instruction::SExt : Instruction::ZExt;
6426 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
6427 X = Builder.CreateCast(Op: CastOpcode, V: X, DestTy: YTy);
6428 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
6429 Y = Builder.CreateCast(Op: CastOpcode, V: Y, DestTy: XTy);
6430 else
6431 return nullptr;
6432 }
6433
6434 // (zext X) == (zext Y) --> X == Y
6435 // (sext X) == (sext Y) --> X == Y
6436 if (ICmp.isEquality())
6437 return new ICmpInst(ICmp.getPredicate(), X, Y);
6438
6439 // A signed comparison of sign extended values simplifies into a
6440 // signed comparison.
6441 if (IsSignedCmp && IsSignedExt)
6442 return new ICmpInst(ICmp.getPredicate(), X, Y);
6443
6444 // The other three cases all fold into an unsigned comparison.
6445 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
6446 }
6447
6448 // Below here, we are only folding a compare with constant.
6449 auto *C = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1));
6450 if (!C)
6451 return nullptr;
6452
6453 // If a lossless truncate is possible...
6454 Type *SrcTy = CastOp0->getSrcTy();
6455 Constant *Res = getLosslessInvCast(C, InvCastTo: SrcTy, CastOp: CastOp0->getOpcode(), DL);
6456 if (Res) {
6457 if (ICmp.isEquality())
6458 return new ICmpInst(ICmp.getPredicate(), X, Res);
6459
6460 // A signed comparison of sign extended values simplifies into a
6461 // signed comparison.
6462 if (IsSignedExt && IsSignedCmp)
6463 return new ICmpInst(ICmp.getPredicate(), X, Res);
6464
6465 // The other three cases all fold into an unsigned comparison.
6466 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res);
6467 }
6468
6469 // The re-extended constant changed, partly changed (in the case of a vector),
6470 // or could not be determined to be equal (in the case of a constant
6471 // expression), so the constant cannot be represented in the shorter type.
6472 // All the cases that fold to true or false will have already been handled
6473 // by simplifyICmpInst, so only deal with the tricky case.
6474 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(Val: C))
6475 return nullptr;
6476
6477 // Is source op positive?
6478 // icmp ult (sext X), C --> icmp sgt X, -1
6479 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
6480 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(Ty: SrcTy));
6481
6482 // Is source op negative?
6483 // icmp ugt (sext X), C --> icmp slt X, 0
6484 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
6485 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(Ty: SrcTy));
6486}
6487
6488/// Handle icmp (cast x), (cast or constant).
6489Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
6490 // If any operand of ICmp is a inttoptr roundtrip cast then remove it as
6491 // icmp compares only pointer's value.
6492 // icmp (inttoptr (ptrtoint p1)), p2 --> icmp p1, p2.
6493 Value *SimplifiedOp0 = simplifyIntToPtrRoundTripCast(Val: ICmp.getOperand(i_nocapture: 0));
6494 Value *SimplifiedOp1 = simplifyIntToPtrRoundTripCast(Val: ICmp.getOperand(i_nocapture: 1));
6495 if (SimplifiedOp0 || SimplifiedOp1)
6496 return new ICmpInst(ICmp.getPredicate(),
6497 SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(i_nocapture: 0),
6498 SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(i_nocapture: 1));
6499
6500 auto *CastOp0 = dyn_cast<CastInst>(Val: ICmp.getOperand(i_nocapture: 0));
6501 if (!CastOp0)
6502 return nullptr;
6503 if (!isa<Constant>(Val: ICmp.getOperand(i_nocapture: 1)) && !isa<CastInst>(Val: ICmp.getOperand(i_nocapture: 1)))
6504 return nullptr;
6505
6506 Value *Op0Src = CastOp0->getOperand(i_nocapture: 0);
6507 Type *SrcTy = CastOp0->getSrcTy();
6508 Type *DestTy = CastOp0->getDestTy();
6509
6510 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6511 // integer type is the same size as the pointer type.
6512 auto CompatibleSizes = [&](Type *PtrTy, Type *IntTy) {
6513 if (isa<VectorType>(Val: PtrTy)) {
6514 PtrTy = cast<VectorType>(Val: PtrTy)->getElementType();
6515 IntTy = cast<VectorType>(Val: IntTy)->getElementType();
6516 }
6517 return DL.getPointerTypeSizeInBits(PtrTy) == IntTy->getIntegerBitWidth();
6518 };
6519 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
6520 CompatibleSizes(SrcTy, DestTy)) {
6521 Value *NewOp1 = nullptr;
6522 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(Val: ICmp.getOperand(i_nocapture: 1))) {
6523 Value *PtrSrc = PtrToIntOp1->getOperand(i_nocapture: 0);
6524 if (PtrSrc->getType() == Op0Src->getType())
6525 NewOp1 = PtrToIntOp1->getOperand(i_nocapture: 0);
6526 } else if (auto *RHSC = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1))) {
6527 NewOp1 = ConstantExpr::getIntToPtr(C: RHSC, Ty: SrcTy);
6528 }
6529
6530 if (NewOp1)
6531 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
6532 }
6533
6534 // Do the same in the other direction for icmp (inttoptr x), (inttoptr/c).
6535 if (CastOp0->getOpcode() == Instruction::IntToPtr &&
6536 CompatibleSizes(DestTy, SrcTy)) {
6537 Value *NewOp1 = nullptr;
6538 if (auto *IntToPtrOp1 = dyn_cast<IntToPtrInst>(Val: ICmp.getOperand(i_nocapture: 1))) {
6539 Value *IntSrc = IntToPtrOp1->getOperand(i_nocapture: 0);
6540 if (IntSrc->getType() == Op0Src->getType())
6541 NewOp1 = IntToPtrOp1->getOperand(i_nocapture: 0);
6542 } else if (auto *RHSC = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1))) {
6543 NewOp1 = ConstantFoldConstant(C: ConstantExpr::getPtrToInt(C: RHSC, Ty: SrcTy), DL);
6544 }
6545
6546 if (NewOp1)
6547 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
6548 }
6549
6550 if (Instruction *R = foldICmpWithTrunc(ICmp))
6551 return R;
6552
6553 return foldICmpWithZextOrSext(ICmp);
6554}
6555
6556static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS,
6557 bool IsSigned) {
6558 switch (BinaryOp) {
6559 default:
6560 llvm_unreachable("Unsupported binary op");
6561 case Instruction::Add:
6562 case Instruction::Sub:
6563 return match(V: RHS, P: m_Zero());
6564 case Instruction::Mul:
6565 return !(RHS->getType()->isIntOrIntVectorTy(BitWidth: 1) && IsSigned) &&
6566 match(V: RHS, P: m_One());
6567 }
6568}
6569
6570OverflowResult
6571InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
6572 bool IsSigned, Value *LHS, Value *RHS,
6573 Instruction *CxtI) const {
6574 switch (BinaryOp) {
6575 default:
6576 llvm_unreachable("Unsupported binary op");
6577 case Instruction::Add:
6578 if (IsSigned)
6579 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
6580 else
6581 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
6582 case Instruction::Sub:
6583 if (IsSigned)
6584 return computeOverflowForSignedSub(LHS, RHS, CxtI);
6585 else
6586 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
6587 case Instruction::Mul:
6588 if (IsSigned)
6589 return computeOverflowForSignedMul(LHS, RHS, CxtI);
6590 else
6591 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
6592 }
6593}
6594
6595bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
6596 bool IsSigned, Value *LHS,
6597 Value *RHS, Instruction &OrigI,
6598 Value *&Result,
6599 Constant *&Overflow) {
6600 if (OrigI.isCommutative() && isa<Constant>(Val: LHS) && !isa<Constant>(Val: RHS))
6601 std::swap(a&: LHS, b&: RHS);
6602
6603 // If the overflow check was an add followed by a compare, the insertion point
6604 // may be pointing to the compare. We want to insert the new instructions
6605 // before the add in case there are uses of the add between the add and the
6606 // compare.
6607 Builder.SetInsertPoint(&OrigI);
6608
6609 Type *OverflowTy = Type::getInt1Ty(C&: LHS->getContext());
6610 if (auto *LHSTy = dyn_cast<VectorType>(Val: LHS->getType()))
6611 OverflowTy = VectorType::get(ElementType: OverflowTy, EC: LHSTy->getElementCount());
6612
6613 if (isNeutralValue(BinaryOp, RHS, IsSigned)) {
6614 Result = LHS;
6615 Overflow = ConstantInt::getFalse(Ty: OverflowTy);
6616 return true;
6617 }
6618
6619 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, CxtI: &OrigI)) {
6620 case OverflowResult::MayOverflow:
6621 return false;
6622 case OverflowResult::AlwaysOverflowsLow:
6623 case OverflowResult::AlwaysOverflowsHigh:
6624 Result = Builder.CreateBinOp(Opc: BinaryOp, LHS, RHS);
6625 Result->takeName(V: &OrigI);
6626 Overflow = ConstantInt::getTrue(Ty: OverflowTy);
6627 return true;
6628 case OverflowResult::NeverOverflows:
6629 Result = Builder.CreateBinOp(Opc: BinaryOp, LHS, RHS);
6630 Result->takeName(V: &OrigI);
6631 Overflow = ConstantInt::getFalse(Ty: OverflowTy);
6632 if (auto *Inst = dyn_cast<Instruction>(Val: Result)) {
6633 if (IsSigned)
6634 Inst->setHasNoSignedWrap();
6635 else
6636 Inst->setHasNoUnsignedWrap();
6637 }
6638 return true;
6639 }
6640
6641 llvm_unreachable("Unexpected overflow result");
6642}
6643
6644/// Recognize and process idiom involving test for multiplication
6645/// overflow.
6646///
6647/// The caller has matched a pattern of the form:
6648/// I = cmp u (mul(zext A, zext B), V
6649/// The function checks if this is a test for overflow and if so replaces
6650/// multiplication with call to 'mul.with.overflow' intrinsic.
6651///
6652/// \param I Compare instruction.
6653/// \param MulVal Result of 'mult' instruction. It is one of the arguments of
6654/// the compare instruction. Must be of integer type.
6655/// \param OtherVal The other argument of compare instruction.
6656/// \returns Instruction which must replace the compare instruction, NULL if no
6657/// replacement required.
6658static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
6659 const APInt *OtherVal,
6660 InstCombinerImpl &IC) {
6661 // Don't bother doing this transformation for pointers, don't do it for
6662 // vectors.
6663 if (!isa<IntegerType>(Val: MulVal->getType()))
6664 return nullptr;
6665
6666 auto *MulInstr = dyn_cast<Instruction>(Val: MulVal);
6667 if (!MulInstr)
6668 return nullptr;
6669 assert(MulInstr->getOpcode() == Instruction::Mul);
6670
6671 auto *LHS = cast<ZExtInst>(Val: MulInstr->getOperand(i: 0)),
6672 *RHS = cast<ZExtInst>(Val: MulInstr->getOperand(i: 1));
6673 assert(LHS->getOpcode() == Instruction::ZExt);
6674 assert(RHS->getOpcode() == Instruction::ZExt);
6675 Value *A = LHS->getOperand(i_nocapture: 0), *B = RHS->getOperand(i_nocapture: 0);
6676
6677 // Calculate type and width of the result produced by mul.with.overflow.
6678 Type *TyA = A->getType(), *TyB = B->getType();
6679 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
6680 WidthB = TyB->getPrimitiveSizeInBits();
6681 unsigned MulWidth;
6682 Type *MulType;
6683 if (WidthB > WidthA) {
6684 MulWidth = WidthB;
6685 MulType = TyB;
6686 } else {
6687 MulWidth = WidthA;
6688 MulType = TyA;
6689 }
6690
6691 // In order to replace the original mul with a narrower mul.with.overflow,
6692 // all uses must ignore upper bits of the product. The number of used low
6693 // bits must be not greater than the width of mul.with.overflow.
6694 if (MulVal->hasNUsesOrMore(N: 2))
6695 for (User *U : MulVal->users()) {
6696 if (U == &I)
6697 continue;
6698 if (TruncInst *TI = dyn_cast<TruncInst>(Val: U)) {
6699 // Check if truncation ignores bits above MulWidth.
6700 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
6701 if (TruncWidth > MulWidth)
6702 return nullptr;
6703 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: U)) {
6704 // Check if AND ignores bits above MulWidth.
6705 if (BO->getOpcode() != Instruction::And)
6706 return nullptr;
6707 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) {
6708 const APInt &CVal = CI->getValue();
6709 if (CVal.getBitWidth() - CVal.countl_zero() > MulWidth)
6710 return nullptr;
6711 } else {
6712 // In this case we could have the operand of the binary operation
6713 // being defined in another block, and performing the replacement
6714 // could break the dominance relation.
6715 return nullptr;
6716 }
6717 } else {
6718 // Other uses prohibit this transformation.
6719 return nullptr;
6720 }
6721 }
6722
6723 // Recognize patterns
6724 switch (I.getPredicate()) {
6725 case ICmpInst::ICMP_UGT: {
6726 // Recognize pattern:
6727 // mulval = mul(zext A, zext B)
6728 // cmp ugt mulval, max
6729 APInt MaxVal = APInt::getMaxValue(numBits: MulWidth);
6730 MaxVal = MaxVal.zext(width: OtherVal->getBitWidth());
6731 if (MaxVal.eq(RHS: *OtherVal))
6732 break; // Recognized
6733 return nullptr;
6734 }
6735
6736 case ICmpInst::ICMP_ULT: {
6737 // Recognize pattern:
6738 // mulval = mul(zext A, zext B)
6739 // cmp ule mulval, max + 1
6740 APInt MaxVal = APInt::getOneBitSet(numBits: OtherVal->getBitWidth(), BitNo: MulWidth);
6741 if (MaxVal.eq(RHS: *OtherVal))
6742 break; // Recognized
6743 return nullptr;
6744 }
6745
6746 default:
6747 return nullptr;
6748 }
6749
6750 InstCombiner::BuilderTy &Builder = IC.Builder;
6751 Builder.SetInsertPoint(MulInstr);
6752
6753 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
6754 Value *MulA = A, *MulB = B;
6755 if (WidthA < MulWidth)
6756 MulA = Builder.CreateZExt(V: A, DestTy: MulType);
6757 if (WidthB < MulWidth)
6758 MulB = Builder.CreateZExt(V: B, DestTy: MulType);
6759 CallInst *Call =
6760 Builder.CreateIntrinsic(ID: Intrinsic::umul_with_overflow, Types: MulType,
6761 Args: {MulA, MulB}, /*FMFSource=*/nullptr, Name: "umul");
6762 IC.addToWorklist(I: MulInstr);
6763
6764 // If there are uses of mul result other than the comparison, we know that
6765 // they are truncation or binary AND. Change them to use result of
6766 // mul.with.overflow and adjust properly mask/size.
6767 if (MulVal->hasNUsesOrMore(N: 2)) {
6768 Value *Mul = Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "umul.value");
6769 for (User *U : make_early_inc_range(Range: MulVal->users())) {
6770 if (U == &I)
6771 continue;
6772 if (TruncInst *TI = dyn_cast<TruncInst>(Val: U)) {
6773 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
6774 IC.replaceInstUsesWith(I&: *TI, V: Mul);
6775 else
6776 TI->setOperand(i_nocapture: 0, Val_nocapture: Mul);
6777 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: U)) {
6778 assert(BO->getOpcode() == Instruction::And);
6779 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
6780 ConstantInt *CI = cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1));
6781 APInt ShortMask = CI->getValue().trunc(width: MulWidth);
6782 Value *ShortAnd = Builder.CreateAnd(LHS: Mul, RHS: ShortMask);
6783 Value *Zext = Builder.CreateZExt(V: ShortAnd, DestTy: BO->getType());
6784 IC.replaceInstUsesWith(I&: *BO, V: Zext);
6785 } else {
6786 llvm_unreachable("Unexpected Binary operation");
6787 }
6788 IC.addToWorklist(I: cast<Instruction>(Val: U));
6789 }
6790 }
6791
6792 // The original icmp gets replaced with the overflow value, maybe inverted
6793 // depending on predicate.
6794 if (I.getPredicate() == ICmpInst::ICMP_ULT) {
6795 Value *Res = Builder.CreateExtractValue(Agg: Call, Idxs: 1);
6796 return BinaryOperator::CreateNot(Op: Res);
6797 }
6798
6799 return ExtractValueInst::Create(Agg: Call, Idxs: 1);
6800}
6801
6802/// When performing a comparison against a constant, it is possible that not all
6803/// the bits in the LHS are demanded. This helper method computes the mask that
6804/// IS demanded.
6805static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
6806 const APInt *RHS;
6807 if (!match(V: I.getOperand(i_nocapture: 1), P: m_APInt(Res&: RHS)))
6808 return APInt::getAllOnes(numBits: BitWidth);
6809
6810 // If this is a normal comparison, it demands all bits. If it is a sign bit
6811 // comparison, it only demands the sign bit.
6812 bool UnusedBit;
6813 if (isSignBitCheck(Pred: I.getPredicate(), RHS: *RHS, TrueIfSigned&: UnusedBit))
6814 return APInt::getSignMask(BitWidth);
6815
6816 switch (I.getPredicate()) {
6817 // For a UGT comparison, we don't care about any bits that
6818 // correspond to the trailing ones of the comparand. The value of these
6819 // bits doesn't impact the outcome of the comparison, because any value
6820 // greater than the RHS must differ in a bit higher than these due to carry.
6821 case ICmpInst::ICMP_UGT:
6822 return APInt::getBitsSetFrom(numBits: BitWidth, loBit: RHS->countr_one());
6823
6824 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
6825 // Any value less than the RHS must differ in a higher bit because of carries.
6826 case ICmpInst::ICMP_ULT:
6827 return APInt::getBitsSetFrom(numBits: BitWidth, loBit: RHS->countr_zero());
6828
6829 default:
6830 return APInt::getAllOnes(numBits: BitWidth);
6831 }
6832}
6833
6834/// Check that one use is in the same block as the definition and all
6835/// other uses are in blocks dominated by a given block.
6836///
6837/// \param DI Definition
6838/// \param UI Use
6839/// \param DB Block that must dominate all uses of \p DI outside
6840/// the parent block
6841/// \return true when \p UI is the only use of \p DI in the parent block
6842/// and all other uses of \p DI are in blocks dominated by \p DB.
6843///
6844bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
6845 const Instruction *UI,
6846 const BasicBlock *DB) const {
6847 assert(DI && UI && "Instruction not defined\n");
6848 // Ignore incomplete definitions.
6849 if (!DI->getParent())
6850 return false;
6851 // DI and UI must be in the same block.
6852 if (DI->getParent() != UI->getParent())
6853 return false;
6854 // Protect from self-referencing blocks.
6855 if (DI->getParent() == DB)
6856 return false;
6857 for (const User *U : DI->users()) {
6858 auto *Usr = cast<Instruction>(Val: U);
6859 if (Usr != UI && !DT.dominates(A: DB, B: Usr->getParent()))
6860 return false;
6861 }
6862 return true;
6863}
6864
6865/// Return true when the instruction sequence within a block is select-cmp-br.
6866static bool isChainSelectCmpBranch(const SelectInst *SI) {
6867 const BasicBlock *BB = SI->getParent();
6868 if (!BB)
6869 return false;
6870 auto *BI = dyn_cast_or_null<BranchInst>(Val: BB->getTerminator());
6871 if (!BI || BI->getNumSuccessors() != 2)
6872 return false;
6873 auto *IC = dyn_cast<ICmpInst>(Val: BI->getCondition());
6874 if (!IC || (IC->getOperand(i_nocapture: 0) != SI && IC->getOperand(i_nocapture: 1) != SI))
6875 return false;
6876 return true;
6877}
6878
6879/// True when a select result is replaced by one of its operands
6880/// in select-icmp sequence. This will eventually result in the elimination
6881/// of the select.
6882///
6883/// \param SI Select instruction
6884/// \param Icmp Compare instruction
6885/// \param SIOpd Operand that replaces the select
6886///
6887/// Notes:
6888/// - The replacement is global and requires dominator information
6889/// - The caller is responsible for the actual replacement
6890///
6891/// Example:
6892///
6893/// entry:
6894/// %4 = select i1 %3, %C* %0, %C* null
6895/// %5 = icmp eq %C* %4, null
6896/// br i1 %5, label %9, label %7
6897/// ...
6898/// ; <label>:7 ; preds = %entry
6899/// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
6900/// ...
6901///
6902/// can be transformed to
6903///
6904/// %5 = icmp eq %C* %0, null
6905/// %6 = select i1 %3, i1 %5, i1 true
6906/// br i1 %6, label %9, label %7
6907/// ...
6908/// ; <label>:7 ; preds = %entry
6909/// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
6910///
6911/// Similar when the first operand of the select is a constant or/and
6912/// the compare is for not equal rather than equal.
6913///
6914/// NOTE: The function is only called when the select and compare constants
6915/// are equal, the optimization can work only for EQ predicates. This is not a
6916/// major restriction since a NE compare should be 'normalized' to an equal
6917/// compare, which usually happens in the combiner and test case
6918/// select-cmp-br.ll checks for it.
6919bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
6920 const ICmpInst *Icmp,
6921 const unsigned SIOpd) {
6922 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
6923 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
6924 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(Idx: 1);
6925 // The check for the single predecessor is not the best that can be
6926 // done. But it protects efficiently against cases like when SI's
6927 // home block has two successors, Succ and Succ1, and Succ1 predecessor
6928 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
6929 // replaced can be reached on either path. So the uniqueness check
6930 // guarantees that the path all uses of SI (outside SI's parent) are on
6931 // is disjoint from all other paths out of SI. But that information
6932 // is more expensive to compute, and the trade-off here is in favor
6933 // of compile-time. It should also be noticed that we check for a single
6934 // predecessor and not only uniqueness. This to handle the situation when
6935 // Succ and Succ1 points to the same basic block.
6936 if (Succ->getSinglePredecessor() && dominatesAllUses(DI: SI, UI: Icmp, DB: Succ)) {
6937 NumSel++;
6938 SI->replaceUsesOutsideBlock(V: SI->getOperand(i_nocapture: SIOpd), BB: SI->getParent());
6939 return true;
6940 }
6941 }
6942 return false;
6943}
6944
6945/// Try to fold the comparison based on range information we can get by checking
6946/// whether bits are known to be zero or one in the inputs.
6947Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
6948 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
6949 Type *Ty = Op0->getType();
6950 ICmpInst::Predicate Pred = I.getPredicate();
6951
6952 // Get scalar or pointer size.
6953 unsigned BitWidth = Ty->isIntOrIntVectorTy()
6954 ? Ty->getScalarSizeInBits()
6955 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
6956
6957 if (!BitWidth)
6958 return nullptr;
6959
6960 KnownBits Op0Known(BitWidth);
6961 KnownBits Op1Known(BitWidth);
6962
6963 {
6964 // Don't use dominating conditions when folding icmp using known bits. This
6965 // may convert signed into unsigned predicates in ways that other passes
6966 // (especially IndVarSimplify) may not be able to reliably undo.
6967 SimplifyQuery Q = SQ.getWithoutDomCondCache().getWithInstruction(I: &I);
6968 if (SimplifyDemandedBits(I: &I, Op: 0, DemandedMask: getDemandedBitsLHSMask(I, BitWidth),
6969 Known&: Op0Known, Q))
6970 return &I;
6971
6972 if (SimplifyDemandedBits(I: &I, Op: 1, DemandedMask: APInt::getAllOnes(numBits: BitWidth), Known&: Op1Known, Q))
6973 return &I;
6974 }
6975
6976 if (!isa<Constant>(Val: Op0) && Op0Known.isConstant())
6977 return new ICmpInst(
6978 Pred, ConstantExpr::getIntegerValue(Ty, V: Op0Known.getConstant()), Op1);
6979 if (!isa<Constant>(Val: Op1) && Op1Known.isConstant())
6980 return new ICmpInst(
6981 Pred, Op0, ConstantExpr::getIntegerValue(Ty, V: Op1Known.getConstant()));
6982
6983 if (std::optional<bool> Res = ICmpInst::compare(LHS: Op0Known, RHS: Op1Known, Pred))
6984 return replaceInstUsesWith(I, V: ConstantInt::getBool(Ty: I.getType(), V: *Res));
6985
6986 // Given the known and unknown bits, compute a range that the LHS could be
6987 // in. Compute the Min, Max and RHS values based on the known bits. For the
6988 // EQ and NE we use unsigned values.
6989 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6990 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6991 if (I.isSigned()) {
6992 Op0Min = Op0Known.getSignedMinValue();
6993 Op0Max = Op0Known.getSignedMaxValue();
6994 Op1Min = Op1Known.getSignedMinValue();
6995 Op1Max = Op1Known.getSignedMaxValue();
6996 } else {
6997 Op0Min = Op0Known.getMinValue();
6998 Op0Max = Op0Known.getMaxValue();
6999 Op1Min = Op1Known.getMinValue();
7000 Op1Max = Op1Known.getMaxValue();
7001 }
7002
7003 // Don't break up a clamp pattern -- (min(max X, Y), Z) -- by replacing a
7004 // min/max canonical compare with some other compare. That could lead to
7005 // conflict with select canonicalization and infinite looping.
7006 // FIXME: This constraint may go away if min/max intrinsics are canonical.
7007 auto isMinMaxCmp = [&](Instruction &Cmp) {
7008 if (!Cmp.hasOneUse())
7009 return false;
7010 Value *A, *B;
7011 SelectPatternFlavor SPF = matchSelectPattern(V: Cmp.user_back(), LHS&: A, RHS&: B).Flavor;
7012 if (!SelectPatternResult::isMinOrMax(SPF))
7013 return false;
7014 return match(V: Op0, P: m_MaxOrMin(L: m_Value(), R: m_Value())) ||
7015 match(V: Op1, P: m_MaxOrMin(L: m_Value(), R: m_Value()));
7016 };
7017 if (!isMinMaxCmp(I)) {
7018 switch (Pred) {
7019 default:
7020 break;
7021 case ICmpInst::ICMP_ULT: {
7022 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
7023 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7024 const APInt *CmpC;
7025 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7026 // A <u C -> A == C-1 if min(A)+1 == C
7027 if (*CmpC == Op0Min + 1)
7028 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7029 ConstantInt::get(Ty: Op1->getType(), V: *CmpC - 1));
7030 // X <u C --> X == 0, if the number of zero bits in the bottom of X
7031 // exceeds the log2 of C.
7032 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
7033 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7034 Constant::getNullValue(Ty: Op1->getType()));
7035 }
7036 break;
7037 }
7038 case ICmpInst::ICMP_UGT: {
7039 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
7040 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7041 const APInt *CmpC;
7042 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7043 // A >u C -> A == C+1 if max(a)-1 == C
7044 if (*CmpC == Op0Max - 1)
7045 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7046 ConstantInt::get(Ty: Op1->getType(), V: *CmpC + 1));
7047 // X >u C --> X != 0, if the number of zero bits in the bottom of X
7048 // exceeds the log2 of C.
7049 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
7050 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
7051 Constant::getNullValue(Ty: Op1->getType()));
7052 }
7053 break;
7054 }
7055 case ICmpInst::ICMP_SLT: {
7056 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
7057 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7058 const APInt *CmpC;
7059 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7060 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
7061 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7062 ConstantInt::get(Ty: Op1->getType(), V: *CmpC - 1));
7063 }
7064 break;
7065 }
7066 case ICmpInst::ICMP_SGT: {
7067 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
7068 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
7069 const APInt *CmpC;
7070 if (match(V: Op1, P: m_APInt(Res&: CmpC))) {
7071 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
7072 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
7073 ConstantInt::get(Ty: Op1->getType(), V: *CmpC + 1));
7074 }
7075 break;
7076 }
7077 }
7078 }
7079
7080 // Based on the range information we know about the LHS, see if we can
7081 // simplify this comparison. For example, (x&4) < 8 is always true.
7082 switch (Pred) {
7083 default:
7084 break;
7085 case ICmpInst::ICMP_EQ:
7086 case ICmpInst::ICMP_NE: {
7087 // If all bits are known zero except for one, then we know at most one bit
7088 // is set. If the comparison is against zero, then this is a check to see if
7089 // *that* bit is set.
7090 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
7091 if (Op1Known.isZero()) {
7092 // If the LHS is an AND with the same constant, look through it.
7093 Value *LHS = nullptr;
7094 const APInt *LHSC;
7095 if (!match(V: Op0, P: m_And(L: m_Value(V&: LHS), R: m_APInt(Res&: LHSC))) ||
7096 *LHSC != Op0KnownZeroInverted)
7097 LHS = Op0;
7098
7099 Value *X;
7100 const APInt *C1;
7101 if (match(V: LHS, P: m_Shl(L: m_Power2(V&: C1), R: m_Value(V&: X)))) {
7102 Type *XTy = X->getType();
7103 unsigned Log2C1 = C1->countr_zero();
7104 APInt C2 = Op0KnownZeroInverted;
7105 APInt C2Pow2 = (C2 & ~(*C1 - 1)) + *C1;
7106 if (C2Pow2.isPowerOf2()) {
7107 // iff (C1 is pow2) & ((C2 & ~(C1-1)) + C1) is pow2):
7108 // ((C1 << X) & C2) == 0 -> X >= (Log2(C2+C1) - Log2(C1))
7109 // ((C1 << X) & C2) != 0 -> X < (Log2(C2+C1) - Log2(C1))
7110 unsigned Log2C2 = C2Pow2.countr_zero();
7111 auto *CmpC = ConstantInt::get(Ty: XTy, V: Log2C2 - Log2C1);
7112 auto NewPred =
7113 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
7114 return new ICmpInst(NewPred, X, CmpC);
7115 }
7116 }
7117 }
7118
7119 // Op0 eq C_Pow2 -> Op0 ne 0 if Op0 is known to be C_Pow2 or zero.
7120 if (Op1Known.isConstant() && Op1Known.getConstant().isPowerOf2() &&
7121 (Op0Known & Op1Known) == Op0Known)
7122 return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op0,
7123 ConstantInt::getNullValue(Ty: Op1->getType()));
7124 break;
7125 }
7126 case ICmpInst::ICMP_SGE:
7127 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
7128 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7129 break;
7130 case ICmpInst::ICMP_SLE:
7131 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
7132 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7133 break;
7134 case ICmpInst::ICMP_UGE:
7135 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
7136 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7137 break;
7138 case ICmpInst::ICMP_ULE:
7139 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
7140 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
7141 break;
7142 }
7143
7144 // Turn a signed comparison into an unsigned one if both operands are known to
7145 // have the same sign. Set samesign if possible (except for equality
7146 // predicates).
7147 if ((I.isSigned() || (I.isUnsigned() && !I.hasSameSign())) &&
7148 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
7149 (Op0Known.One.isNegative() && Op1Known.One.isNegative()))) {
7150 I.setPredicate(I.getUnsignedPredicate());
7151 I.setSameSign();
7152 return &I;
7153 }
7154
7155 return nullptr;
7156}
7157
7158/// If one operand of an icmp is effectively a bool (value range of {0,1}),
7159/// then try to reduce patterns based on that limit.
7160Instruction *InstCombinerImpl::foldICmpUsingBoolRange(ICmpInst &I) {
7161 Value *X, *Y;
7162 CmpPredicate Pred;
7163
7164 // X must be 0 and bool must be true for "ULT":
7165 // X <u (zext i1 Y) --> (X == 0) & Y
7166 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))))) &&
7167 Y->getType()->isIntOrIntVectorTy(BitWidth: 1) && Pred == ICmpInst::ICMP_ULT)
7168 return BinaryOperator::CreateAnd(V1: Builder.CreateIsNull(Arg: X), V2: Y);
7169
7170 // X must be 0 or bool must be true for "ULE":
7171 // X <=u (sext i1 Y) --> (X == 0) | Y
7172 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))))) &&
7173 Y->getType()->isIntOrIntVectorTy(BitWidth: 1) && Pred == ICmpInst::ICMP_ULE)
7174 return BinaryOperator::CreateOr(V1: Builder.CreateIsNull(Arg: X), V2: Y);
7175
7176 // icmp eq/ne X, (zext/sext (icmp eq/ne X, C))
7177 CmpPredicate Pred1, Pred2;
7178 const APInt *C;
7179 Instruction *ExtI;
7180 if (match(V: &I, P: m_c_ICmp(Pred&: Pred1, L: m_Value(V&: X),
7181 R: m_CombineAnd(L: m_Instruction(I&: ExtI),
7182 R: m_ZExtOrSExt(Op: m_ICmp(Pred&: Pred2, L: m_Deferred(V: X),
7183 R: m_APInt(Res&: C)))))) &&
7184 ICmpInst::isEquality(P: Pred1) && ICmpInst::isEquality(P: Pred2)) {
7185 bool IsSExt = ExtI->getOpcode() == Instruction::SExt;
7186 bool HasOneUse = ExtI->hasOneUse() && ExtI->getOperand(i: 0)->hasOneUse();
7187 auto CreateRangeCheck = [&] {
7188 Value *CmpV1 =
7189 Builder.CreateICmp(P: Pred1, LHS: X, RHS: Constant::getNullValue(Ty: X->getType()));
7190 Value *CmpV2 = Builder.CreateICmp(
7191 P: Pred1, LHS: X, RHS: ConstantInt::getSigned(Ty: X->getType(), V: IsSExt ? -1 : 1));
7192 return BinaryOperator::Create(
7193 Op: Pred1 == ICmpInst::ICMP_EQ ? Instruction::Or : Instruction::And,
7194 S1: CmpV1, S2: CmpV2);
7195 };
7196 if (C->isZero()) {
7197 if (Pred2 == ICmpInst::ICMP_EQ) {
7198 // icmp eq X, (zext/sext (icmp eq X, 0)) --> false
7199 // icmp ne X, (zext/sext (icmp eq X, 0)) --> true
7200 return replaceInstUsesWith(
7201 I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred1 == ICmpInst::ICMP_NE));
7202 } else if (!IsSExt || HasOneUse) {
7203 // icmp eq X, (zext (icmp ne X, 0)) --> X == 0 || X == 1
7204 // icmp ne X, (zext (icmp ne X, 0)) --> X != 0 && X != 1
7205 // icmp eq X, (sext (icmp ne X, 0)) --> X == 0 || X == -1
7206 // icmp ne X, (sext (icmp ne X, 0)) --> X != 0 && X != -1
7207 return CreateRangeCheck();
7208 }
7209 } else if (IsSExt ? C->isAllOnes() : C->isOne()) {
7210 if (Pred2 == ICmpInst::ICMP_NE) {
7211 // icmp eq X, (zext (icmp ne X, 1)) --> false
7212 // icmp ne X, (zext (icmp ne X, 1)) --> true
7213 // icmp eq X, (sext (icmp ne X, -1)) --> false
7214 // icmp ne X, (sext (icmp ne X, -1)) --> true
7215 return replaceInstUsesWith(
7216 I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred1 == ICmpInst::ICMP_NE));
7217 } else if (!IsSExt || HasOneUse) {
7218 // icmp eq X, (zext (icmp eq X, 1)) --> X == 0 || X == 1
7219 // icmp ne X, (zext (icmp eq X, 1)) --> X != 0 && X != 1
7220 // icmp eq X, (sext (icmp eq X, -1)) --> X == 0 || X == -1
7221 // icmp ne X, (sext (icmp eq X, -1)) --> X != 0 && X == -1
7222 return CreateRangeCheck();
7223 }
7224 } else {
7225 // when C != 0 && C != 1:
7226 // icmp eq X, (zext (icmp eq X, C)) --> icmp eq X, 0
7227 // icmp eq X, (zext (icmp ne X, C)) --> icmp eq X, 1
7228 // icmp ne X, (zext (icmp eq X, C)) --> icmp ne X, 0
7229 // icmp ne X, (zext (icmp ne X, C)) --> icmp ne X, 1
7230 // when C != 0 && C != -1:
7231 // icmp eq X, (sext (icmp eq X, C)) --> icmp eq X, 0
7232 // icmp eq X, (sext (icmp ne X, C)) --> icmp eq X, -1
7233 // icmp ne X, (sext (icmp eq X, C)) --> icmp ne X, 0
7234 // icmp ne X, (sext (icmp ne X, C)) --> icmp ne X, -1
7235 return ICmpInst::Create(
7236 Op: Instruction::ICmp, Pred: Pred1, S1: X,
7237 S2: ConstantInt::getSigned(Ty: X->getType(), V: Pred2 == ICmpInst::ICMP_NE
7238 ? (IsSExt ? -1 : 1)
7239 : 0));
7240 }
7241 }
7242
7243 return nullptr;
7244}
7245
7246/// If we have an icmp le or icmp ge instruction with a constant operand, turn
7247/// it into the appropriate icmp lt or icmp gt instruction. This transform
7248/// allows them to be folded in visitICmpInst.
7249static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
7250 ICmpInst::Predicate Pred = I.getPredicate();
7251 if (ICmpInst::isEquality(P: Pred) || !ICmpInst::isIntPredicate(P: Pred) ||
7252 InstCombiner::isCanonicalPredicate(Pred))
7253 return nullptr;
7254
7255 Value *Op0 = I.getOperand(i_nocapture: 0);
7256 Value *Op1 = I.getOperand(i_nocapture: 1);
7257 auto *Op1C = dyn_cast<Constant>(Val: Op1);
7258 if (!Op1C)
7259 return nullptr;
7260
7261 auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, C: Op1C);
7262 if (!FlippedStrictness)
7263 return nullptr;
7264
7265 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
7266}
7267
7268/// If we have a comparison with a non-canonical predicate, if we can update
7269/// all the users, invert the predicate and adjust all the users.
7270CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
7271 // Is the predicate already canonical?
7272 CmpInst::Predicate Pred = I.getPredicate();
7273 if (InstCombiner::isCanonicalPredicate(Pred))
7274 return nullptr;
7275
7276 // Can all users be adjusted to predicate inversion?
7277 if (!InstCombiner::canFreelyInvertAllUsersOf(V: &I, /*IgnoredUser=*/nullptr))
7278 return nullptr;
7279
7280 // Ok, we can canonicalize comparison!
7281 // Let's first invert the comparison's predicate.
7282 I.setPredicate(CmpInst::getInversePredicate(pred: Pred));
7283 I.setName(I.getName() + ".not");
7284
7285 // And, adapt users.
7286 freelyInvertAllUsersOf(V: &I);
7287
7288 return &I;
7289}
7290
7291/// Integer compare with boolean values can always be turned into bitwise ops.
7292static Instruction *canonicalizeICmpBool(ICmpInst &I,
7293 InstCombiner::BuilderTy &Builder) {
7294 Value *A = I.getOperand(i_nocapture: 0), *B = I.getOperand(i_nocapture: 1);
7295 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
7296
7297 // A boolean compared to true/false can be simplified to Op0/true/false in
7298 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
7299 // Cases not handled by InstSimplify are always 'not' of Op0.
7300 if (match(V: B, P: m_Zero())) {
7301 switch (I.getPredicate()) {
7302 case CmpInst::ICMP_EQ: // A == 0 -> !A
7303 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
7304 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
7305 return BinaryOperator::CreateNot(Op: A);
7306 default:
7307 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
7308 }
7309 } else if (match(V: B, P: m_One())) {
7310 switch (I.getPredicate()) {
7311 case CmpInst::ICMP_NE: // A != 1 -> !A
7312 case CmpInst::ICMP_ULT: // A <u 1 -> !A
7313 case CmpInst::ICMP_SGT: // A >s -1 -> !A
7314 return BinaryOperator::CreateNot(Op: A);
7315 default:
7316 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
7317 }
7318 }
7319
7320 switch (I.getPredicate()) {
7321 default:
7322 llvm_unreachable("Invalid icmp instruction!");
7323 case ICmpInst::ICMP_EQ:
7324 // icmp eq i1 A, B -> ~(A ^ B)
7325 return BinaryOperator::CreateNot(Op: Builder.CreateXor(LHS: A, RHS: B));
7326
7327 case ICmpInst::ICMP_NE:
7328 // icmp ne i1 A, B -> A ^ B
7329 return BinaryOperator::CreateXor(V1: A, V2: B);
7330
7331 case ICmpInst::ICMP_UGT:
7332 // icmp ugt -> icmp ult
7333 std::swap(a&: A, b&: B);
7334 [[fallthrough]];
7335 case ICmpInst::ICMP_ULT:
7336 // icmp ult i1 A, B -> ~A & B
7337 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: A), V2: B);
7338
7339 case ICmpInst::ICMP_SGT:
7340 // icmp sgt -> icmp slt
7341 std::swap(a&: A, b&: B);
7342 [[fallthrough]];
7343 case ICmpInst::ICMP_SLT:
7344 // icmp slt i1 A, B -> A & ~B
7345 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: B), V2: A);
7346
7347 case ICmpInst::ICMP_UGE:
7348 // icmp uge -> icmp ule
7349 std::swap(a&: A, b&: B);
7350 [[fallthrough]];
7351 case ICmpInst::ICMP_ULE:
7352 // icmp ule i1 A, B -> ~A | B
7353 return BinaryOperator::CreateOr(V1: Builder.CreateNot(V: A), V2: B);
7354
7355 case ICmpInst::ICMP_SGE:
7356 // icmp sge -> icmp sle
7357 std::swap(a&: A, b&: B);
7358 [[fallthrough]];
7359 case ICmpInst::ICMP_SLE:
7360 // icmp sle i1 A, B -> A | ~B
7361 return BinaryOperator::CreateOr(V1: Builder.CreateNot(V: B), V2: A);
7362 }
7363}
7364
7365// Transform pattern like:
7366// (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
7367// (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
7368// Into:
7369// (X l>> Y) != 0
7370// (X l>> Y) == 0
7371static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
7372 InstCombiner::BuilderTy &Builder) {
7373 CmpPredicate Pred, NewPred;
7374 Value *X, *Y;
7375 if (match(V: &Cmp,
7376 P: m_c_ICmp(Pred, L: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: Y))), R: m_Value(V&: X)))) {
7377 switch (Pred) {
7378 case ICmpInst::ICMP_ULE:
7379 NewPred = ICmpInst::ICMP_NE;
7380 break;
7381 case ICmpInst::ICMP_UGT:
7382 NewPred = ICmpInst::ICMP_EQ;
7383 break;
7384 default:
7385 return nullptr;
7386 }
7387 } else if (match(V: &Cmp, P: m_c_ICmp(Pred,
7388 L: m_OneUse(SubPattern: m_CombineOr(
7389 L: m_Not(V: m_Shl(L: m_AllOnes(), R: m_Value(V&: Y))),
7390 R: m_Add(L: m_Shl(L: m_One(), R: m_Value(V&: Y)),
7391 R: m_AllOnes()))),
7392 R: m_Value(V&: X)))) {
7393 // The variant with 'add' is not canonical, (the variant with 'not' is)
7394 // we only get it because it has extra uses, and can't be canonicalized,
7395
7396 switch (Pred) {
7397 case ICmpInst::ICMP_ULT:
7398 NewPred = ICmpInst::ICMP_NE;
7399 break;
7400 case ICmpInst::ICMP_UGE:
7401 NewPred = ICmpInst::ICMP_EQ;
7402 break;
7403 default:
7404 return nullptr;
7405 }
7406 } else
7407 return nullptr;
7408
7409 Value *NewX = Builder.CreateLShr(LHS: X, RHS: Y, Name: X->getName() + ".highbits");
7410 Constant *Zero = Constant::getNullValue(Ty: NewX->getType());
7411 return CmpInst::Create(Op: Instruction::ICmp, Pred: NewPred, S1: NewX, S2: Zero);
7412}
7413
7414static Instruction *foldVectorCmp(CmpInst &Cmp,
7415 InstCombiner::BuilderTy &Builder) {
7416 const CmpInst::Predicate Pred = Cmp.getPredicate();
7417 Value *LHS = Cmp.getOperand(i_nocapture: 0), *RHS = Cmp.getOperand(i_nocapture: 1);
7418 Value *V1, *V2;
7419
7420 auto createCmpReverse = [&](CmpInst::Predicate Pred, Value *X, Value *Y) {
7421 Value *V = Builder.CreateCmp(Pred, LHS: X, RHS: Y, Name: Cmp.getName());
7422 if (auto *I = dyn_cast<Instruction>(Val: V))
7423 I->copyIRFlags(V: &Cmp);
7424 Module *M = Cmp.getModule();
7425 Function *F = Intrinsic::getOrInsertDeclaration(
7426 M, id: Intrinsic::vector_reverse, Tys: V->getType());
7427 return CallInst::Create(Func: F, Args: V);
7428 };
7429
7430 if (match(V: LHS, P: m_VecReverse(Op0: m_Value(V&: V1)))) {
7431 // cmp Pred, rev(V1), rev(V2) --> rev(cmp Pred, V1, V2)
7432 if (match(V: RHS, P: m_VecReverse(Op0: m_Value(V&: V2))) &&
7433 (LHS->hasOneUse() || RHS->hasOneUse()))
7434 return createCmpReverse(Pred, V1, V2);
7435
7436 // cmp Pred, rev(V1), RHSSplat --> rev(cmp Pred, V1, RHSSplat)
7437 if (LHS->hasOneUse() && isSplatValue(V: RHS))
7438 return createCmpReverse(Pred, V1, RHS);
7439 }
7440 // cmp Pred, LHSSplat, rev(V2) --> rev(cmp Pred, LHSSplat, V2)
7441 else if (isSplatValue(V: LHS) && match(V: RHS, P: m_OneUse(SubPattern: m_VecReverse(Op0: m_Value(V&: V2)))))
7442 return createCmpReverse(Pred, LHS, V2);
7443
7444 ArrayRef<int> M;
7445 if (!match(V: LHS, P: m_Shuffle(v1: m_Value(V&: V1), v2: m_Undef(), mask: m_Mask(M))))
7446 return nullptr;
7447
7448 // If both arguments of the cmp are shuffles that use the same mask and
7449 // shuffle within a single vector, move the shuffle after the cmp:
7450 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
7451 Type *V1Ty = V1->getType();
7452 if (match(V: RHS, P: m_Shuffle(v1: m_Value(V&: V2), v2: m_Undef(), mask: m_SpecificMask(M))) &&
7453 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
7454 Value *NewCmp = Builder.CreateCmp(Pred, LHS: V1, RHS: V2);
7455 return new ShuffleVectorInst(NewCmp, M);
7456 }
7457
7458 // Try to canonicalize compare with splatted operand and splat constant.
7459 // TODO: We could generalize this for more than splats. See/use the code in
7460 // InstCombiner::foldVectorBinop().
7461 Constant *C;
7462 if (!LHS->hasOneUse() || !match(V: RHS, P: m_Constant(C)))
7463 return nullptr;
7464
7465 // Length-changing splats are ok, so adjust the constants as needed:
7466 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
7467 Constant *ScalarC = C->getSplatValue(/* AllowPoison */ true);
7468 int MaskSplatIndex;
7469 if (ScalarC && match(Mask: M, P: m_SplatOrPoisonMask(MaskSplatIndex))) {
7470 // We allow poison in matching, but this transform removes it for safety.
7471 // Demanded elements analysis should be able to recover some/all of that.
7472 C = ConstantVector::getSplat(EC: cast<VectorType>(Val: V1Ty)->getElementCount(),
7473 Elt: ScalarC);
7474 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
7475 Value *NewCmp = Builder.CreateCmp(Pred, LHS: V1, RHS: C);
7476 return new ShuffleVectorInst(NewCmp, NewM);
7477 }
7478
7479 return nullptr;
7480}
7481
7482// extract(uadd.with.overflow(A, B), 0) ult A
7483// -> extract(uadd.with.overflow(A, B), 1)
7484static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
7485 CmpInst::Predicate Pred = I.getPredicate();
7486 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
7487
7488 Value *UAddOv;
7489 Value *A, *B;
7490 auto UAddOvResultPat = m_ExtractValue<0>(
7491 V: m_Intrinsic<Intrinsic::uadd_with_overflow>(Op0: m_Value(V&: A), Op1: m_Value(V&: B)));
7492 if (match(V: Op0, P: UAddOvResultPat) &&
7493 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
7494 (Pred == ICmpInst::ICMP_EQ && match(V: Op1, P: m_ZeroInt()) &&
7495 (match(V: A, P: m_One()) || match(V: B, P: m_One()))) ||
7496 (Pred == ICmpInst::ICMP_NE && match(V: Op1, P: m_AllOnes()) &&
7497 (match(V: A, P: m_AllOnes()) || match(V: B, P: m_AllOnes())))))
7498 // extract(uadd.with.overflow(A, B), 0) < A
7499 // extract(uadd.with.overflow(A, 1), 0) == 0
7500 // extract(uadd.with.overflow(A, -1), 0) != -1
7501 UAddOv = cast<ExtractValueInst>(Val: Op0)->getAggregateOperand();
7502 else if (match(V: Op1, P: UAddOvResultPat) && Pred == ICmpInst::ICMP_UGT &&
7503 (Op0 == A || Op0 == B))
7504 // A > extract(uadd.with.overflow(A, B), 0)
7505 UAddOv = cast<ExtractValueInst>(Val: Op1)->getAggregateOperand();
7506 else
7507 return nullptr;
7508
7509 return ExtractValueInst::Create(Agg: UAddOv, Idxs: 1);
7510}
7511
7512static Instruction *foldICmpInvariantGroup(ICmpInst &I) {
7513 if (!I.getOperand(i_nocapture: 0)->getType()->isPointerTy() ||
7514 NullPointerIsDefined(
7515 F: I.getParent()->getParent(),
7516 AS: I.getOperand(i_nocapture: 0)->getType()->getPointerAddressSpace())) {
7517 return nullptr;
7518 }
7519 Instruction *Op;
7520 if (match(V: I.getOperand(i_nocapture: 0), P: m_Instruction(I&: Op)) &&
7521 match(V: I.getOperand(i_nocapture: 1), P: m_Zero()) &&
7522 Op->isLaunderOrStripInvariantGroup()) {
7523 return ICmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(),
7524 S1: Op->getOperand(i: 0), S2: I.getOperand(i_nocapture: 1));
7525 }
7526 return nullptr;
7527}
7528
7529/// This function folds patterns produced by lowering of reduce idioms, such as
7530/// llvm.vector.reduce.and which are lowered into instruction chains. This code
7531/// attempts to generate fewer number of scalar comparisons instead of vector
7532/// comparisons when possible.
7533static Instruction *foldReductionIdiom(ICmpInst &I,
7534 InstCombiner::BuilderTy &Builder,
7535 const DataLayout &DL) {
7536 if (I.getType()->isVectorTy())
7537 return nullptr;
7538 CmpPredicate OuterPred, InnerPred;
7539 Value *LHS, *RHS;
7540
7541 // Match lowering of @llvm.vector.reduce.and. Turn
7542 /// %vec_ne = icmp ne <8 x i8> %lhs, %rhs
7543 /// %scalar_ne = bitcast <8 x i1> %vec_ne to i8
7544 /// %res = icmp <pred> i8 %scalar_ne, 0
7545 ///
7546 /// into
7547 ///
7548 /// %lhs.scalar = bitcast <8 x i8> %lhs to i64
7549 /// %rhs.scalar = bitcast <8 x i8> %rhs to i64
7550 /// %res = icmp <pred> i64 %lhs.scalar, %rhs.scalar
7551 ///
7552 /// for <pred> in {ne, eq}.
7553 if (!match(V: &I, P: m_ICmp(Pred&: OuterPred,
7554 L: m_OneUse(SubPattern: m_BitCast(Op: m_OneUse(
7555 SubPattern: m_ICmp(Pred&: InnerPred, L: m_Value(V&: LHS), R: m_Value(V&: RHS))))),
7556 R: m_Zero())))
7557 return nullptr;
7558 auto *LHSTy = dyn_cast<FixedVectorType>(Val: LHS->getType());
7559 if (!LHSTy || !LHSTy->getElementType()->isIntegerTy())
7560 return nullptr;
7561 unsigned NumBits =
7562 LHSTy->getNumElements() * LHSTy->getElementType()->getIntegerBitWidth();
7563 // TODO: Relax this to "not wider than max legal integer type"?
7564 if (!DL.isLegalInteger(Width: NumBits))
7565 return nullptr;
7566
7567 if (ICmpInst::isEquality(P: OuterPred) && InnerPred == ICmpInst::ICMP_NE) {
7568 auto *ScalarTy = Builder.getIntNTy(N: NumBits);
7569 LHS = Builder.CreateBitCast(V: LHS, DestTy: ScalarTy, Name: LHS->getName() + ".scalar");
7570 RHS = Builder.CreateBitCast(V: RHS, DestTy: ScalarTy, Name: RHS->getName() + ".scalar");
7571 return ICmpInst::Create(Op: Instruction::ICmp, Pred: OuterPred, S1: LHS, S2: RHS,
7572 Name: I.getName());
7573 }
7574
7575 return nullptr;
7576}
7577
7578// This helper will be called with icmp operands in both orders.
7579Instruction *InstCombinerImpl::foldICmpCommutative(CmpPredicate Pred,
7580 Value *Op0, Value *Op1,
7581 ICmpInst &CxtI) {
7582 // Try to optimize 'icmp GEP, P' or 'icmp P, GEP'.
7583 if (auto *GEP = dyn_cast<GEPOperator>(Val: Op0))
7584 if (Instruction *NI = foldGEPICmp(GEPLHS: GEP, RHS: Op1, Cond: Pred, I&: CxtI))
7585 return NI;
7586
7587 if (auto *SI = dyn_cast<SelectInst>(Val: Op0))
7588 if (Instruction *NI = foldSelectICmp(Pred, SI, RHS: Op1, I: CxtI))
7589 return NI;
7590
7591 if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op0)) {
7592 if (Instruction *Res = foldICmpWithMinMax(I&: CxtI, MinMax, Z: Op1, Pred))
7593 return Res;
7594
7595 if (Instruction *Res = foldICmpWithClamp(I&: CxtI, X: Op1, Min: MinMax))
7596 return Res;
7597 }
7598
7599 {
7600 Value *X;
7601 const APInt *C;
7602 // icmp X+Cst, X
7603 if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C))) && Op1 == X)
7604 return foldICmpAddOpConst(X, C: *C, Pred);
7605 }
7606
7607 // abs(X) >= X --> true
7608 // abs(X) u<= X --> true
7609 // abs(X) < X --> false
7610 // abs(X) u> X --> false
7611 // abs(X) u>= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7612 // abs(X) <= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7613 // abs(X) == X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN`
7614 // abs(X) u< X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7615 // abs(X) > X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7616 // abs(X) != X --> IsIntMinPosion ? `X < 0` : `X > INTMIN`
7617 {
7618 Value *X;
7619 Constant *C;
7620 if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X), Op1: m_Constant(C))) &&
7621 match(V: Op1, P: m_Specific(V: X))) {
7622 Value *NullValue = Constant::getNullValue(Ty: X->getType());
7623 Value *AllOnesValue = Constant::getAllOnesValue(Ty: X->getType());
7624 const APInt SMin =
7625 APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits());
7626 bool IsIntMinPosion = C->isAllOnesValue();
7627 switch (Pred) {
7628 case CmpInst::ICMP_ULE:
7629 case CmpInst::ICMP_SGE:
7630 return replaceInstUsesWith(I&: CxtI, V: ConstantInt::getTrue(Ty: CxtI.getType()));
7631 case CmpInst::ICMP_UGT:
7632 case CmpInst::ICMP_SLT:
7633 return replaceInstUsesWith(I&: CxtI, V: ConstantInt::getFalse(Ty: CxtI.getType()));
7634 case CmpInst::ICMP_UGE:
7635 case CmpInst::ICMP_SLE:
7636 case CmpInst::ICMP_EQ: {
7637 return replaceInstUsesWith(
7638 I&: CxtI, V: IsIntMinPosion
7639 ? Builder.CreateICmpSGT(LHS: X, RHS: AllOnesValue)
7640 : Builder.CreateICmpULT(
7641 LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: SMin + 1)));
7642 }
7643 case CmpInst::ICMP_ULT:
7644 case CmpInst::ICMP_SGT:
7645 case CmpInst::ICMP_NE: {
7646 return replaceInstUsesWith(
7647 I&: CxtI, V: IsIntMinPosion
7648 ? Builder.CreateICmpSLT(LHS: X, RHS: NullValue)
7649 : Builder.CreateICmpUGT(
7650 LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: SMin)));
7651 }
7652 default:
7653 llvm_unreachable("Invalid predicate!");
7654 }
7655 }
7656 }
7657
7658 const SimplifyQuery Q = SQ.getWithInstruction(I: &CxtI);
7659 if (Value *V = foldICmpWithLowBitMaskedVal(Pred, Op0, Op1, Q, IC&: *this))
7660 return replaceInstUsesWith(I&: CxtI, V);
7661
7662 // Folding (X / Y) pred X => X swap(pred) 0 for constant Y other than 0 or 1
7663 auto CheckUGT1 = [](const APInt &Divisor) { return Divisor.ugt(RHS: 1); };
7664 {
7665 if (match(V: Op0, P: m_UDiv(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckUGT1)))) {
7666 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7667 Constant::getNullValue(Ty: Op1->getType()));
7668 }
7669
7670 if (!ICmpInst::isUnsigned(predicate: Pred) &&
7671 match(V: Op0, P: m_SDiv(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckUGT1)))) {
7672 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7673 Constant::getNullValue(Ty: Op1->getType()));
7674 }
7675 }
7676
7677 // Another case of this fold is (X >> Y) pred X => X swap(pred) 0 if Y != 0
7678 auto CheckNE0 = [](const APInt &Shift) { return !Shift.isZero(); };
7679 {
7680 if (match(V: Op0, P: m_LShr(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckNE0)))) {
7681 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7682 Constant::getNullValue(Ty: Op1->getType()));
7683 }
7684
7685 if ((Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SGE) &&
7686 match(V: Op0, P: m_AShr(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckNE0)))) {
7687 return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1,
7688 Constant::getNullValue(Ty: Op1->getType()));
7689 }
7690 }
7691
7692 return nullptr;
7693}
7694
7695Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
7696 bool Changed = false;
7697 const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
7698 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
7699 unsigned Op0Cplxity = getComplexity(V: Op0);
7700 unsigned Op1Cplxity = getComplexity(V: Op1);
7701
7702 /// Orders the operands of the compare so that they are listed from most
7703 /// complex to least complex. This puts constants before unary operators,
7704 /// before binary operators.
7705 if (Op0Cplxity < Op1Cplxity) {
7706 I.swapOperands();
7707 std::swap(a&: Op0, b&: Op1);
7708 Changed = true;
7709 }
7710
7711 if (Value *V = simplifyICmpInst(Pred: I.getCmpPredicate(), LHS: Op0, RHS: Op1, Q))
7712 return replaceInstUsesWith(I, V);
7713
7714 // Comparing -val or val with non-zero is the same as just comparing val
7715 // ie, abs(val) != 0 -> val != 0
7716 if (I.getPredicate() == ICmpInst::ICMP_NE && match(V: Op1, P: m_Zero())) {
7717 Value *Cond, *SelectTrue, *SelectFalse;
7718 if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: SelectTrue),
7719 R: m_Value(V&: SelectFalse)))) {
7720 if (Value *V = dyn_castNegVal(V: SelectTrue)) {
7721 if (V == SelectFalse)
7722 return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), S1: V, S2: Op1);
7723 } else if (Value *V = dyn_castNegVal(V: SelectFalse)) {
7724 if (V == SelectTrue)
7725 return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), S1: V, S2: Op1);
7726 }
7727 }
7728 }
7729
7730 if (Instruction *Res = foldICmpTruncWithTruncOrExt(Cmp&: I, Q))
7731 return Res;
7732
7733 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1))
7734 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
7735 return Res;
7736
7737 if (Instruction *Res = canonicalizeCmpWithConstant(I))
7738 return Res;
7739
7740 if (Instruction *Res = canonicalizeICmpPredicate(I))
7741 return Res;
7742
7743 if (Instruction *Res = foldICmpWithConstant(Cmp&: I))
7744 return Res;
7745
7746 if (Instruction *Res = foldICmpWithDominatingICmp(Cmp&: I))
7747 return Res;
7748
7749 if (Instruction *Res = foldICmpUsingBoolRange(I))
7750 return Res;
7751
7752 if (Instruction *Res = foldICmpUsingKnownBits(I))
7753 return Res;
7754
7755 if (Instruction *Res = foldIsMultipleOfAPowerOfTwo(Cmp&: I))
7756 return Res;
7757
7758 // Test if the ICmpInst instruction is used exclusively by a select as
7759 // part of a minimum or maximum operation. If so, refrain from doing
7760 // any other folding. This helps out other analyses which understand
7761 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
7762 // and CodeGen. And in this case, at least one of the comparison
7763 // operands has at least one user besides the compare (the select),
7764 // which would often largely negate the benefit of folding anyway.
7765 //
7766 // Do the same for the other patterns recognized by matchSelectPattern.
7767 if (I.hasOneUse())
7768 if (SelectInst *SI = dyn_cast<SelectInst>(Val: I.user_back())) {
7769 Value *A, *B;
7770 SelectPatternResult SPR = matchSelectPattern(V: SI, LHS&: A, RHS&: B);
7771 if (SPR.Flavor != SPF_UNKNOWN)
7772 return nullptr;
7773 }
7774
7775 // Do this after checking for min/max to prevent infinite looping.
7776 if (Instruction *Res = foldICmpWithZero(Cmp&: I))
7777 return Res;
7778
7779 // FIXME: We only do this after checking for min/max to prevent infinite
7780 // looping caused by a reverse canonicalization of these patterns for min/max.
7781 // FIXME: The organization of folds is a mess. These would naturally go into
7782 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
7783 // down here after the min/max restriction.
7784 ICmpInst::Predicate Pred = I.getPredicate();
7785 const APInt *C;
7786 if (match(V: Op1, P: m_APInt(Res&: C))) {
7787 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
7788 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
7789 Constant *Zero = Constant::getNullValue(Ty: Op0->getType());
7790 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
7791 }
7792
7793 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
7794 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
7795 Constant *AllOnes = Constant::getAllOnesValue(Ty: Op0->getType());
7796 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
7797 }
7798 }
7799
7800 // The folds in here may rely on wrapping flags and special constants, so
7801 // they can break up min/max idioms in some cases but not seemingly similar
7802 // patterns.
7803 // FIXME: It may be possible to enhance select folding to make this
7804 // unnecessary. It may also be moot if we canonicalize to min/max
7805 // intrinsics.
7806 if (Instruction *Res = foldICmpBinOp(I, SQ: Q))
7807 return Res;
7808
7809 if (Instruction *Res = foldICmpInstWithConstant(Cmp&: I))
7810 return Res;
7811
7812 // Try to match comparison as a sign bit test. Intentionally do this after
7813 // foldICmpInstWithConstant() to potentially let other folds to happen first.
7814 if (Instruction *New = foldSignBitTest(I))
7815 return New;
7816
7817 if (auto *PN = dyn_cast<PHINode>(Val: Op0))
7818 if (Instruction *NV = foldOpIntoPhi(I, PN))
7819 return NV;
7820 if (auto *PN = dyn_cast<PHINode>(Val: Op1))
7821 if (Instruction *NV = foldOpIntoPhi(I, PN))
7822 return NV;
7823
7824 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
7825 return Res;
7826
7827 if (Instruction *Res = foldICmpCommutative(Pred: I.getCmpPredicate(), Op0, Op1, CxtI&: I))
7828 return Res;
7829 if (Instruction *Res =
7830 foldICmpCommutative(Pred: I.getSwappedCmpPredicate(), Op0: Op1, Op1: Op0, CxtI&: I))
7831 return Res;
7832
7833 if (I.isCommutative()) {
7834 if (auto Pair = matchSymmetricPair(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1))) {
7835 replaceOperand(I, OpNum: 0, V: Pair->first);
7836 replaceOperand(I, OpNum: 1, V: Pair->second);
7837 return &I;
7838 }
7839 }
7840
7841 // In case of a comparison with two select instructions having the same
7842 // condition, check whether one of the resulting branches can be simplified.
7843 // If so, just compare the other branch and select the appropriate result.
7844 // For example:
7845 // %tmp1 = select i1 %cmp, i32 %y, i32 %x
7846 // %tmp2 = select i1 %cmp, i32 %z, i32 %x
7847 // %cmp2 = icmp slt i32 %tmp2, %tmp1
7848 // The icmp will result false for the false value of selects and the result
7849 // will depend upon the comparison of true values of selects if %cmp is
7850 // true. Thus, transform this into:
7851 // %cmp = icmp slt i32 %y, %z
7852 // %sel = select i1 %cond, i1 %cmp, i1 false
7853 // This handles similar cases to transform.
7854 {
7855 Value *Cond, *A, *B, *C, *D;
7856 if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: A), R: m_Value(V&: B))) &&
7857 match(V: Op1, P: m_Select(C: m_Specific(V: Cond), L: m_Value(V&: C), R: m_Value(V&: D))) &&
7858 (Op0->hasOneUse() || Op1->hasOneUse())) {
7859 // Check whether comparison of TrueValues can be simplified
7860 if (Value *Res = simplifyICmpInst(Pred, LHS: A, RHS: C, Q: SQ)) {
7861 Value *NewICMP = Builder.CreateICmp(P: Pred, LHS: B, RHS: D);
7862 return SelectInst::Create(
7863 C: Cond, S1: Res, S2: NewICMP, /*NameStr=*/"", /*InsertBefore=*/nullptr,
7864 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : cast<Instruction>(Val: Op0));
7865 }
7866 // Check whether comparison of FalseValues can be simplified
7867 if (Value *Res = simplifyICmpInst(Pred, LHS: B, RHS: D, Q: SQ)) {
7868 Value *NewICMP = Builder.CreateICmp(P: Pred, LHS: A, RHS: C);
7869 return SelectInst::Create(
7870 C: Cond, S1: NewICMP, S2: Res, /*NameStr=*/"", /*InsertBefore=*/nullptr,
7871 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : cast<Instruction>(Val: Op0));
7872 }
7873 }
7874 }
7875
7876 // icmp slt (sub nsw x, y), (add nsw x, y) --> icmp sgt y, 0
7877 // icmp ult (sub nuw x, y), (add nuw x, y) --> icmp ugt y, 0
7878 // icmp eq (sub nsw/nuw x, y), (add nsw/nuw x, y) --> icmp eq y, 0
7879 {
7880 Value *A, *B;
7881 CmpPredicate CmpPred;
7882 if (match(V: &I, P: m_c_ICmp(Pred&: CmpPred, L: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B)),
7883 R: m_c_Add(L: m_Deferred(V: A), R: m_Deferred(V: B))))) {
7884 auto *I0 = cast<OverflowingBinaryOperator>(Val: Op0);
7885 auto *I1 = cast<OverflowingBinaryOperator>(Val: Op1);
7886 bool I0NUW = I0->hasNoUnsignedWrap();
7887 bool I1NUW = I1->hasNoUnsignedWrap();
7888 bool I0NSW = I0->hasNoSignedWrap();
7889 bool I1NSW = I1->hasNoSignedWrap();
7890 if ((ICmpInst::isUnsigned(predicate: Pred) && I0NUW && I1NUW) ||
7891 (ICmpInst::isSigned(predicate: Pred) && I0NSW && I1NSW) ||
7892 (ICmpInst::isEquality(P: Pred) &&
7893 ((I0NUW || I0NSW) && (I1NUW || I1NSW)))) {
7894 return new ICmpInst(CmpPredicate::getSwapped(P: CmpPred), B,
7895 ConstantInt::get(Ty: Op0->getType(), V: 0));
7896 }
7897 }
7898 }
7899
7900 // Try to optimize equality comparisons against alloca-based pointers.
7901 if (Op0->getType()->isPointerTy() && I.isEquality()) {
7902 assert(Op1->getType()->isPointerTy() &&
7903 "Comparing pointer with non-pointer?");
7904 if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Op0)))
7905 if (foldAllocaCmp(Alloca))
7906 return nullptr;
7907 if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Op1)))
7908 if (foldAllocaCmp(Alloca))
7909 return nullptr;
7910 }
7911
7912 if (Instruction *Res = foldICmpBitCast(Cmp&: I))
7913 return Res;
7914
7915 // TODO: Hoist this above the min/max bailout.
7916 if (Instruction *R = foldICmpWithCastOp(ICmp&: I))
7917 return R;
7918
7919 {
7920 Value *X, *Y;
7921 // Transform (X & ~Y) == 0 --> (X & Y) != 0
7922 // and (X & ~Y) != 0 --> (X & Y) == 0
7923 // if A is a power of 2.
7924 if (match(V: Op0, P: m_And(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
7925 match(V: Op1, P: m_Zero()) && isKnownToBeAPowerOfTwo(V: X, OrZero: false, CxtI: &I) &&
7926 I.isEquality())
7927 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(LHS: X, RHS: Y),
7928 Op1);
7929
7930 // Op0 pred Op1 -> ~Op1 pred ~Op0, if this allows us to drop an instruction.
7931 if (Op0->getType()->isIntOrIntVectorTy()) {
7932 bool ConsumesOp0, ConsumesOp1;
7933 if (isFreeToInvert(V: Op0, WillInvertAllUses: Op0->hasOneUse(), DoesConsume&: ConsumesOp0) &&
7934 isFreeToInvert(V: Op1, WillInvertAllUses: Op1->hasOneUse(), DoesConsume&: ConsumesOp1) &&
7935 (ConsumesOp0 || ConsumesOp1)) {
7936 Value *InvOp0 = getFreelyInverted(V: Op0, WillInvertAllUses: Op0->hasOneUse(), Builder: &Builder);
7937 Value *InvOp1 = getFreelyInverted(V: Op1, WillInvertAllUses: Op1->hasOneUse(), Builder: &Builder);
7938 assert(InvOp0 && InvOp1 &&
7939 "Mismatch between isFreeToInvert and getFreelyInverted");
7940 return new ICmpInst(I.getSwappedPredicate(), InvOp0, InvOp1);
7941 }
7942 }
7943
7944 Instruction *AddI = nullptr;
7945 if (match(V: &I, P: m_UAddWithOverflow(L: m_Value(V&: X), R: m_Value(V&: Y),
7946 S: m_Instruction(I&: AddI))) &&
7947 isa<IntegerType>(Val: X->getType())) {
7948 Value *Result;
7949 Constant *Overflow;
7950 // m_UAddWithOverflow can match patterns that do not include an explicit
7951 // "add" instruction, so check the opcode of the matched op.
7952 if (AddI->getOpcode() == Instruction::Add &&
7953 OptimizeOverflowCheck(BinaryOp: Instruction::Add, /*Signed*/ IsSigned: false, LHS: X, RHS: Y, OrigI&: *AddI,
7954 Result, Overflow)) {
7955 replaceInstUsesWith(I&: *AddI, V: Result);
7956 eraseInstFromFunction(I&: *AddI);
7957 return replaceInstUsesWith(I, V: Overflow);
7958 }
7959 }
7960
7961 // (zext X) * (zext Y) --> llvm.umul.with.overflow.
7962 if (match(V: Op0, P: m_NUWMul(L: m_ZExt(Op: m_Value(V&: X)), R: m_ZExt(Op: m_Value(V&: Y)))) &&
7963 match(V: Op1, P: m_APInt(Res&: C))) {
7964 if (Instruction *R = processUMulZExtIdiom(I, MulVal: Op0, OtherVal: C, IC&: *this))
7965 return R;
7966 }
7967
7968 // Signbit test folds
7969 // Fold (X u>> BitWidth - 1 Pred ZExt(i1)) --> X s< 0 Pred i1
7970 // Fold (X s>> BitWidth - 1 Pred SExt(i1)) --> X s< 0 Pred i1
7971 Instruction *ExtI;
7972 if ((I.isUnsigned() || I.isEquality()) &&
7973 match(V: Op1,
7974 P: m_CombineAnd(L: m_Instruction(I&: ExtI), R: m_ZExtOrSExt(Op: m_Value(V&: Y)))) &&
7975 Y->getType()->getScalarSizeInBits() == 1 &&
7976 (Op0->hasOneUse() || Op1->hasOneUse())) {
7977 unsigned OpWidth = Op0->getType()->getScalarSizeInBits();
7978 Instruction *ShiftI;
7979 if (match(V: Op0, P: m_CombineAnd(L: m_Instruction(I&: ShiftI),
7980 R: m_Shr(L: m_Value(V&: X), R: m_SpecificIntAllowPoison(
7981 V: OpWidth - 1))))) {
7982 unsigned ExtOpc = ExtI->getOpcode();
7983 unsigned ShiftOpc = ShiftI->getOpcode();
7984 if ((ExtOpc == Instruction::ZExt && ShiftOpc == Instruction::LShr) ||
7985 (ExtOpc == Instruction::SExt && ShiftOpc == Instruction::AShr)) {
7986 Value *SLTZero =
7987 Builder.CreateICmpSLT(LHS: X, RHS: Constant::getNullValue(Ty: X->getType()));
7988 Value *Cmp = Builder.CreateICmp(P: Pred, LHS: SLTZero, RHS: Y, Name: I.getName());
7989 return replaceInstUsesWith(I, V: Cmp);
7990 }
7991 }
7992 }
7993 }
7994
7995 if (Instruction *Res = foldICmpEquality(I))
7996 return Res;
7997
7998 if (Instruction *Res = foldICmpPow2Test(I, Builder))
7999 return Res;
8000
8001 if (Instruction *Res = foldICmpOfUAddOv(I))
8002 return Res;
8003
8004 // The 'cmpxchg' instruction returns an aggregate containing the old value and
8005 // an i1 which indicates whether or not we successfully did the swap.
8006 //
8007 // Replace comparisons between the old value and the expected value with the
8008 // indicator that 'cmpxchg' returns.
8009 //
8010 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
8011 // spuriously fail. In those cases, the old value may equal the expected
8012 // value but it is possible for the swap to not occur.
8013 if (I.getPredicate() == ICmpInst::ICMP_EQ)
8014 if (auto *EVI = dyn_cast<ExtractValueInst>(Val: Op0))
8015 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(Val: EVI->getAggregateOperand()))
8016 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
8017 !ACXI->isWeak())
8018 return ExtractValueInst::Create(Agg: ACXI, Idxs: 1);
8019
8020 if (Instruction *Res = foldICmpWithHighBitMask(Cmp&: I, Builder))
8021 return Res;
8022
8023 if (I.getType()->isVectorTy())
8024 if (Instruction *Res = foldVectorCmp(Cmp&: I, Builder))
8025 return Res;
8026
8027 if (Instruction *Res = foldICmpInvariantGroup(I))
8028 return Res;
8029
8030 if (Instruction *Res = foldReductionIdiom(I, Builder, DL))
8031 return Res;
8032
8033 {
8034 Value *A;
8035 const APInt *C1, *C2;
8036 ICmpInst::Predicate Pred = I.getPredicate();
8037 if (ICmpInst::isEquality(P: Pred)) {
8038 // sext(a) & c1 == c2 --> a & c3 == trunc(c2)
8039 // sext(a) & c1 != c2 --> a & c3 != trunc(c2)
8040 if (match(V: Op0, P: m_And(L: m_SExt(Op: m_Value(V&: A)), R: m_APInt(Res&: C1))) &&
8041 match(V: Op1, P: m_APInt(Res&: C2))) {
8042 Type *InputTy = A->getType();
8043 unsigned InputBitWidth = InputTy->getScalarSizeInBits();
8044 // c2 must be non-negative at the bitwidth of a.
8045 if (C2->getActiveBits() < InputBitWidth) {
8046 APInt TruncC1 = C1->trunc(width: InputBitWidth);
8047 // Check if there are 1s in C1 high bits of size InputBitWidth.
8048 if (C1->uge(RHS: APInt::getOneBitSet(numBits: C1->getBitWidth(), BitNo: InputBitWidth)))
8049 TruncC1.setBit(InputBitWidth - 1);
8050 Value *AndInst = Builder.CreateAnd(LHS: A, RHS: TruncC1);
8051 return new ICmpInst(
8052 Pred, AndInst,
8053 ConstantInt::get(Ty: InputTy, V: C2->trunc(width: InputBitWidth)));
8054 }
8055 }
8056 }
8057 }
8058
8059 return Changed ? &I : nullptr;
8060}
8061
8062/// Fold fcmp ([us]itofp x, cst) if possible.
8063Instruction *InstCombinerImpl::foldFCmpIntToFPConst(FCmpInst &I,
8064 Instruction *LHSI,
8065 Constant *RHSC) {
8066 const APFloat *RHS;
8067 if (!match(V: RHSC, P: m_APFloat(Res&: RHS)))
8068 return nullptr;
8069
8070 // Get the width of the mantissa. We don't want to hack on conversions that
8071 // might lose information from the integer, e.g. "i64 -> float"
8072 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
8073 if (MantissaWidth == -1)
8074 return nullptr; // Unknown.
8075
8076 Type *IntTy = LHSI->getOperand(i: 0)->getType();
8077 unsigned IntWidth = IntTy->getScalarSizeInBits();
8078 bool LHSUnsigned = isa<UIToFPInst>(Val: LHSI);
8079
8080 if (I.isEquality()) {
8081 FCmpInst::Predicate P = I.getPredicate();
8082 bool IsExact = false;
8083 APSInt RHSCvt(IntWidth, LHSUnsigned);
8084 RHS->convertToInteger(Result&: RHSCvt, RM: APFloat::rmNearestTiesToEven, IsExact: &IsExact);
8085
8086 // If the floating point constant isn't an integer value, we know if we will
8087 // ever compare equal / not equal to it.
8088 if (!IsExact) {
8089 // TODO: Can never be -0.0 and other non-representable values
8090 APFloat RHSRoundInt(*RHS);
8091 RHSRoundInt.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
8092 if (*RHS != RHSRoundInt) {
8093 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
8094 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8095
8096 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
8097 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8098 }
8099 }
8100
8101 // TODO: If the constant is exactly representable, is it always OK to do
8102 // equality compares as integer?
8103 }
8104
8105 // Check to see that the input is converted from an integer type that is small
8106 // enough that preserves all bits. TODO: check here for "known" sign bits.
8107 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
8108
8109 // Following test does NOT adjust IntWidth downwards for signed inputs,
8110 // because the most negative value still requires all the mantissa bits
8111 // to distinguish it from one less than that value.
8112 if ((int)IntWidth > MantissaWidth) {
8113 // Conversion would lose accuracy. Check if loss can impact comparison.
8114 int Exp = ilogb(Arg: *RHS);
8115 if (Exp == APFloat::IEK_Inf) {
8116 int MaxExponent = ilogb(Arg: APFloat::getLargest(Sem: RHS->getSemantics()));
8117 if (MaxExponent < (int)IntWidth - !LHSUnsigned)
8118 // Conversion could create infinity.
8119 return nullptr;
8120 } else {
8121 // Note that if RHS is zero or NaN, then Exp is negative
8122 // and first condition is trivially false.
8123 if (MantissaWidth <= Exp && Exp <= (int)IntWidth - !LHSUnsigned)
8124 // Conversion could affect comparison.
8125 return nullptr;
8126 }
8127 }
8128
8129 // Otherwise, we can potentially simplify the comparison. We know that it
8130 // will always come through as an integer value and we know the constant is
8131 // not a NAN (it would have been previously simplified).
8132 assert(!RHS->isNaN() && "NaN comparison not already folded!");
8133
8134 ICmpInst::Predicate Pred;
8135 switch (I.getPredicate()) {
8136 default:
8137 llvm_unreachable("Unexpected predicate!");
8138 case FCmpInst::FCMP_UEQ:
8139 case FCmpInst::FCMP_OEQ:
8140 Pred = ICmpInst::ICMP_EQ;
8141 break;
8142 case FCmpInst::FCMP_UGT:
8143 case FCmpInst::FCMP_OGT:
8144 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
8145 break;
8146 case FCmpInst::FCMP_UGE:
8147 case FCmpInst::FCMP_OGE:
8148 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
8149 break;
8150 case FCmpInst::FCMP_ULT:
8151 case FCmpInst::FCMP_OLT:
8152 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
8153 break;
8154 case FCmpInst::FCMP_ULE:
8155 case FCmpInst::FCMP_OLE:
8156 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
8157 break;
8158 case FCmpInst::FCMP_UNE:
8159 case FCmpInst::FCMP_ONE:
8160 Pred = ICmpInst::ICMP_NE;
8161 break;
8162 case FCmpInst::FCMP_ORD:
8163 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8164 case FCmpInst::FCMP_UNO:
8165 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8166 }
8167
8168 // Now we know that the APFloat is a normal number, zero or inf.
8169
8170 // See if the FP constant is too large for the integer. For example,
8171 // comparing an i8 to 300.0.
8172 if (!LHSUnsigned) {
8173 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
8174 // and large values.
8175 APFloat SMax(RHS->getSemantics());
8176 SMax.convertFromAPInt(Input: APInt::getSignedMaxValue(numBits: IntWidth), IsSigned: true,
8177 RM: APFloat::rmNearestTiesToEven);
8178 if (SMax < *RHS) { // smax < 13123.0
8179 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
8180 Pred == ICmpInst::ICMP_SLE)
8181 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8182 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8183 }
8184 } else {
8185 // If the RHS value is > UnsignedMax, fold the comparison. This handles
8186 // +INF and large values.
8187 APFloat UMax(RHS->getSemantics());
8188 UMax.convertFromAPInt(Input: APInt::getMaxValue(numBits: IntWidth), IsSigned: false,
8189 RM: APFloat::rmNearestTiesToEven);
8190 if (UMax < *RHS) { // umax < 13123.0
8191 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
8192 Pred == ICmpInst::ICMP_ULE)
8193 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8194 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8195 }
8196 }
8197
8198 if (!LHSUnsigned) {
8199 // See if the RHS value is < SignedMin.
8200 APFloat SMin(RHS->getSemantics());
8201 SMin.convertFromAPInt(Input: APInt::getSignedMinValue(numBits: IntWidth), IsSigned: true,
8202 RM: APFloat::rmNearestTiesToEven);
8203 if (SMin > *RHS) { // smin > 12312.0
8204 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
8205 Pred == ICmpInst::ICMP_SGE)
8206 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8207 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8208 }
8209 } else {
8210 // See if the RHS value is < UnsignedMin.
8211 APFloat UMin(RHS->getSemantics());
8212 UMin.convertFromAPInt(Input: APInt::getMinValue(numBits: IntWidth), IsSigned: false,
8213 RM: APFloat::rmNearestTiesToEven);
8214 if (UMin > *RHS) { // umin > 12312.0
8215 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
8216 Pred == ICmpInst::ICMP_UGE)
8217 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8218 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8219 }
8220 }
8221
8222 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
8223 // [0, UMAX], but it may still be fractional. Check whether this is the case
8224 // using the IsExact flag.
8225 // Don't do this for zero, because -0.0 is not fractional.
8226 APSInt RHSInt(IntWidth, LHSUnsigned);
8227 bool IsExact;
8228 RHS->convertToInteger(Result&: RHSInt, RM: APFloat::rmTowardZero, IsExact: &IsExact);
8229 if (!RHS->isZero()) {
8230 if (!IsExact) {
8231 // If we had a comparison against a fractional value, we have to adjust
8232 // the compare predicate and sometimes the value. RHSC is rounded towards
8233 // zero at this point.
8234 switch (Pred) {
8235 default:
8236 llvm_unreachable("Unexpected integer comparison!");
8237 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
8238 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8239 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
8240 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8241 case ICmpInst::ICMP_ULE:
8242 // (float)int <= 4.4 --> int <= 4
8243 // (float)int <= -4.4 --> false
8244 if (RHS->isNegative())
8245 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8246 break;
8247 case ICmpInst::ICMP_SLE:
8248 // (float)int <= 4.4 --> int <= 4
8249 // (float)int <= -4.4 --> int < -4
8250 if (RHS->isNegative())
8251 Pred = ICmpInst::ICMP_SLT;
8252 break;
8253 case ICmpInst::ICMP_ULT:
8254 // (float)int < -4.4 --> false
8255 // (float)int < 4.4 --> int <= 4
8256 if (RHS->isNegative())
8257 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8258 Pred = ICmpInst::ICMP_ULE;
8259 break;
8260 case ICmpInst::ICMP_SLT:
8261 // (float)int < -4.4 --> int < -4
8262 // (float)int < 4.4 --> int <= 4
8263 if (!RHS->isNegative())
8264 Pred = ICmpInst::ICMP_SLE;
8265 break;
8266 case ICmpInst::ICMP_UGT:
8267 // (float)int > 4.4 --> int > 4
8268 // (float)int > -4.4 --> true
8269 if (RHS->isNegative())
8270 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8271 break;
8272 case ICmpInst::ICMP_SGT:
8273 // (float)int > 4.4 --> int > 4
8274 // (float)int > -4.4 --> int >= -4
8275 if (RHS->isNegative())
8276 Pred = ICmpInst::ICMP_SGE;
8277 break;
8278 case ICmpInst::ICMP_UGE:
8279 // (float)int >= -4.4 --> true
8280 // (float)int >= 4.4 --> int > 4
8281 if (RHS->isNegative())
8282 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8283 Pred = ICmpInst::ICMP_UGT;
8284 break;
8285 case ICmpInst::ICMP_SGE:
8286 // (float)int >= -4.4 --> int >= -4
8287 // (float)int >= 4.4 --> int > 4
8288 if (!RHS->isNegative())
8289 Pred = ICmpInst::ICMP_SGT;
8290 break;
8291 }
8292 }
8293 }
8294
8295 // Lower this FP comparison into an appropriate integer version of the
8296 // comparison.
8297 return new ICmpInst(Pred, LHSI->getOperand(i: 0),
8298 ConstantInt::get(Ty: LHSI->getOperand(i: 0)->getType(), V: RHSInt));
8299}
8300
8301/// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
8302static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
8303 Constant *RHSC) {
8304 // When C is not 0.0 and infinities are not allowed:
8305 // (C / X) < 0.0 is a sign-bit test of X
8306 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
8307 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
8308 //
8309 // Proof:
8310 // Multiply (C / X) < 0.0 by X * X / C.
8311 // - X is non zero, if it is the flag 'ninf' is violated.
8312 // - C defines the sign of X * X * C. Thus it also defines whether to swap
8313 // the predicate. C is also non zero by definition.
8314 //
8315 // Thus X * X / C is non zero and the transformation is valid. [qed]
8316
8317 FCmpInst::Predicate Pred = I.getPredicate();
8318
8319 // Check that predicates are valid.
8320 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
8321 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
8322 return nullptr;
8323
8324 // Check that RHS operand is zero.
8325 if (!match(V: RHSC, P: m_AnyZeroFP()))
8326 return nullptr;
8327
8328 // Check fastmath flags ('ninf').
8329 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
8330 return nullptr;
8331
8332 // Check the properties of the dividend. It must not be zero to avoid a
8333 // division by zero (see Proof).
8334 const APFloat *C;
8335 if (!match(V: LHSI->getOperand(i: 0), P: m_APFloat(Res&: C)))
8336 return nullptr;
8337
8338 if (C->isZero())
8339 return nullptr;
8340
8341 // Get swapped predicate if necessary.
8342 if (C->isNegative())
8343 Pred = I.getSwappedPredicate();
8344
8345 return new FCmpInst(Pred, LHSI->getOperand(i: 1), RHSC, "", &I);
8346}
8347
8348// Transform 'fptrunc(x) cmp C' to 'x cmp ext(C)' if possible.
8349// Patterns include:
8350// fptrunc(x) < C --> x < ext(C)
8351// fptrunc(x) <= C --> x <= ext(C)
8352// fptrunc(x) > C --> x > ext(C)
8353// fptrunc(x) >= C --> x >= ext(C)
8354// where 'ext(C)' is the extension of 'C' to the type of 'x' with a small bias
8355// due to precision loss.
8356static Instruction *foldFCmpFpTrunc(FCmpInst &I, const Instruction &FPTrunc,
8357 const Constant &C) {
8358 FCmpInst::Predicate Pred = I.getPredicate();
8359 bool RoundDown = false;
8360
8361 if (Pred == FCmpInst::FCMP_OGE || Pred == FCmpInst::FCMP_UGE ||
8362 Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_ULT)
8363 RoundDown = true;
8364 else if (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT ||
8365 Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)
8366 RoundDown = false;
8367 else
8368 return nullptr;
8369
8370 const APFloat *CValue;
8371 if (!match(V: &C, P: m_APFloat(Res&: CValue)))
8372 return nullptr;
8373
8374 if (CValue->isNaN() || CValue->isInfinity())
8375 return nullptr;
8376
8377 auto ConvertFltSema = [](const APFloat &Src, const fltSemantics &Sema) {
8378 bool LosesInfo;
8379 APFloat Dest = Src;
8380 Dest.convert(ToSemantics: Sema, RM: APFloat::rmNearestTiesToEven, losesInfo: &LosesInfo);
8381 return Dest;
8382 };
8383
8384 auto NextValue = [](const APFloat &Value, bool RoundDown) {
8385 APFloat NextValue = Value;
8386 NextValue.next(nextDown: RoundDown);
8387 return NextValue;
8388 };
8389
8390 APFloat NextCValue = NextValue(*CValue, RoundDown);
8391
8392 Type *DestType = FPTrunc.getOperand(i: 0)->getType();
8393 const fltSemantics &DestFltSema =
8394 DestType->getScalarType()->getFltSemantics();
8395
8396 APFloat ExtCValue = ConvertFltSema(*CValue, DestFltSema);
8397 APFloat ExtNextCValue = ConvertFltSema(NextCValue, DestFltSema);
8398
8399 // When 'NextCValue' is infinity, use an imaged 'NextCValue' that equals
8400 // 'CValue + bias' to avoid the infinity after conversion. The bias is
8401 // estimated as 'CValue - PrevCValue', where 'PrevCValue' is the previous
8402 // value of 'CValue'.
8403 if (NextCValue.isInfinity()) {
8404 APFloat PrevCValue = NextValue(*CValue, !RoundDown);
8405 APFloat Bias = ConvertFltSema(*CValue - PrevCValue, DestFltSema);
8406
8407 ExtNextCValue = ExtCValue + Bias;
8408 }
8409
8410 APFloat ExtMidValue =
8411 scalbn(X: ExtCValue + ExtNextCValue, Exp: -1, RM: APFloat::rmNearestTiesToEven);
8412
8413 const fltSemantics &SrcFltSema =
8414 C.getType()->getScalarType()->getFltSemantics();
8415
8416 // 'MidValue' might be rounded to 'NextCValue'. Correct it here.
8417 APFloat MidValue = ConvertFltSema(ExtMidValue, SrcFltSema);
8418 if (MidValue != *CValue)
8419 ExtMidValue.next(nextDown: !RoundDown);
8420
8421 // Check whether 'ExtMidValue' is a valid result since the assumption on
8422 // imaged 'NextCValue' might not hold for new float types.
8423 // ppc_fp128 can't pass here when converting from max float because of
8424 // APFloat implementation.
8425 if (NextCValue.isInfinity()) {
8426 // ExtMidValue --- narrowed ---> Finite
8427 if (ConvertFltSema(ExtMidValue, SrcFltSema).isInfinity())
8428 return nullptr;
8429
8430 // NextExtMidValue --- narrowed ---> Infinity
8431 APFloat NextExtMidValue = NextValue(ExtMidValue, RoundDown);
8432 if (ConvertFltSema(NextExtMidValue, SrcFltSema).isFinite())
8433 return nullptr;
8434 }
8435
8436 return new FCmpInst(Pred, FPTrunc.getOperand(i: 0),
8437 ConstantFP::get(Ty: DestType, V: ExtMidValue), "", &I);
8438}
8439
8440/// Optimize fabs(X) compared with zero.
8441static Instruction *foldFabsWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
8442 Value *X;
8443 if (!match(V: I.getOperand(i_nocapture: 0), P: m_FAbs(Op0: m_Value(V&: X))))
8444 return nullptr;
8445
8446 const APFloat *C;
8447 if (!match(V: I.getOperand(i_nocapture: 1), P: m_APFloat(Res&: C)))
8448 return nullptr;
8449
8450 if (!C->isPosZero()) {
8451 if (!C->isSmallestNormalized())
8452 return nullptr;
8453
8454 const Function *F = I.getFunction();
8455 DenormalMode Mode = F->getDenormalMode(FPType: C->getSemantics());
8456 if (Mode.Input == DenormalMode::PreserveSign ||
8457 Mode.Input == DenormalMode::PositiveZero) {
8458
8459 auto replaceFCmp = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
8460 Constant *Zero = ConstantFP::getZero(Ty: X->getType());
8461 return new FCmpInst(P, X, Zero, "", I);
8462 };
8463
8464 switch (I.getPredicate()) {
8465 case FCmpInst::FCMP_OLT:
8466 // fcmp olt fabs(x), smallest_normalized_number -> fcmp oeq x, 0.0
8467 return replaceFCmp(&I, FCmpInst::FCMP_OEQ, X);
8468 case FCmpInst::FCMP_UGE:
8469 // fcmp uge fabs(x), smallest_normalized_number -> fcmp une x, 0.0
8470 return replaceFCmp(&I, FCmpInst::FCMP_UNE, X);
8471 case FCmpInst::FCMP_OGE:
8472 // fcmp oge fabs(x), smallest_normalized_number -> fcmp one x, 0.0
8473 return replaceFCmp(&I, FCmpInst::FCMP_ONE, X);
8474 case FCmpInst::FCMP_ULT:
8475 // fcmp ult fabs(x), smallest_normalized_number -> fcmp ueq x, 0.0
8476 return replaceFCmp(&I, FCmpInst::FCMP_UEQ, X);
8477 default:
8478 break;
8479 }
8480 }
8481
8482 return nullptr;
8483 }
8484
8485 auto replacePredAndOp0 = [&IC](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
8486 I->setPredicate(P);
8487 return IC.replaceOperand(I&: *I, OpNum: 0, V: X);
8488 };
8489
8490 switch (I.getPredicate()) {
8491 case FCmpInst::FCMP_UGE:
8492 case FCmpInst::FCMP_OLT:
8493 // fabs(X) >= 0.0 --> true
8494 // fabs(X) < 0.0 --> false
8495 llvm_unreachable("fcmp should have simplified");
8496
8497 case FCmpInst::FCMP_OGT:
8498 // fabs(X) > 0.0 --> X != 0.0
8499 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
8500
8501 case FCmpInst::FCMP_UGT:
8502 // fabs(X) u> 0.0 --> X u!= 0.0
8503 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
8504
8505 case FCmpInst::FCMP_OLE:
8506 // fabs(X) <= 0.0 --> X == 0.0
8507 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
8508
8509 case FCmpInst::FCMP_ULE:
8510 // fabs(X) u<= 0.0 --> X u== 0.0
8511 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
8512
8513 case FCmpInst::FCMP_OGE:
8514 // fabs(X) >= 0.0 --> !isnan(X)
8515 assert(!I.hasNoNaNs() && "fcmp should have simplified");
8516 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
8517
8518 case FCmpInst::FCMP_ULT:
8519 // fabs(X) u< 0.0 --> isnan(X)
8520 assert(!I.hasNoNaNs() && "fcmp should have simplified");
8521 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
8522
8523 case FCmpInst::FCMP_OEQ:
8524 case FCmpInst::FCMP_UEQ:
8525 case FCmpInst::FCMP_ONE:
8526 case FCmpInst::FCMP_UNE:
8527 case FCmpInst::FCMP_ORD:
8528 case FCmpInst::FCMP_UNO:
8529 // Look through the fabs() because it doesn't change anything but the sign.
8530 // fabs(X) == 0.0 --> X == 0.0,
8531 // fabs(X) != 0.0 --> X != 0.0
8532 // isnan(fabs(X)) --> isnan(X)
8533 // !isnan(fabs(X) --> !isnan(X)
8534 return replacePredAndOp0(&I, I.getPredicate(), X);
8535
8536 default:
8537 return nullptr;
8538 }
8539}
8540
8541/// Optimize sqrt(X) compared with zero.
8542static Instruction *foldSqrtWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
8543 Value *X;
8544 if (!match(V: I.getOperand(i_nocapture: 0), P: m_Sqrt(Op0: m_Value(V&: X))))
8545 return nullptr;
8546
8547 if (!match(V: I.getOperand(i_nocapture: 1), P: m_PosZeroFP()))
8548 return nullptr;
8549
8550 auto ReplacePredAndOp0 = [&](FCmpInst::Predicate P) {
8551 I.setPredicate(P);
8552 return IC.replaceOperand(I, OpNum: 0, V: X);
8553 };
8554
8555 // Clear ninf flag if sqrt doesn't have it.
8556 if (!cast<Instruction>(Val: I.getOperand(i_nocapture: 0))->hasNoInfs())
8557 I.setHasNoInfs(false);
8558
8559 switch (I.getPredicate()) {
8560 case FCmpInst::FCMP_OLT:
8561 case FCmpInst::FCMP_UGE:
8562 // sqrt(X) < 0.0 --> false
8563 // sqrt(X) u>= 0.0 --> true
8564 llvm_unreachable("fcmp should have simplified");
8565 case FCmpInst::FCMP_ULT:
8566 case FCmpInst::FCMP_ULE:
8567 case FCmpInst::FCMP_OGT:
8568 case FCmpInst::FCMP_OGE:
8569 case FCmpInst::FCMP_OEQ:
8570 case FCmpInst::FCMP_UNE:
8571 // sqrt(X) u< 0.0 --> X u< 0.0
8572 // sqrt(X) u<= 0.0 --> X u<= 0.0
8573 // sqrt(X) > 0.0 --> X > 0.0
8574 // sqrt(X) >= 0.0 --> X >= 0.0
8575 // sqrt(X) == 0.0 --> X == 0.0
8576 // sqrt(X) u!= 0.0 --> X u!= 0.0
8577 return IC.replaceOperand(I, OpNum: 0, V: X);
8578
8579 case FCmpInst::FCMP_OLE:
8580 // sqrt(X) <= 0.0 --> X == 0.0
8581 return ReplacePredAndOp0(FCmpInst::FCMP_OEQ);
8582 case FCmpInst::FCMP_UGT:
8583 // sqrt(X) u> 0.0 --> X u!= 0.0
8584 return ReplacePredAndOp0(FCmpInst::FCMP_UNE);
8585 case FCmpInst::FCMP_UEQ:
8586 // sqrt(X) u== 0.0 --> X u<= 0.0
8587 return ReplacePredAndOp0(FCmpInst::FCMP_ULE);
8588 case FCmpInst::FCMP_ONE:
8589 // sqrt(X) != 0.0 --> X > 0.0
8590 return ReplacePredAndOp0(FCmpInst::FCMP_OGT);
8591 case FCmpInst::FCMP_ORD:
8592 // !isnan(sqrt(X)) --> X >= 0.0
8593 return ReplacePredAndOp0(FCmpInst::FCMP_OGE);
8594 case FCmpInst::FCMP_UNO:
8595 // isnan(sqrt(X)) --> X u< 0.0
8596 return ReplacePredAndOp0(FCmpInst::FCMP_ULT);
8597 default:
8598 llvm_unreachable("Unexpected predicate!");
8599 }
8600}
8601
8602static Instruction *foldFCmpFNegCommonOp(FCmpInst &I) {
8603 CmpInst::Predicate Pred = I.getPredicate();
8604 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
8605
8606 // Canonicalize fneg as Op1.
8607 if (match(V: Op0, P: m_FNeg(X: m_Value())) && !match(V: Op1, P: m_FNeg(X: m_Value()))) {
8608 std::swap(a&: Op0, b&: Op1);
8609 Pred = I.getSwappedPredicate();
8610 }
8611
8612 if (!match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0))))
8613 return nullptr;
8614
8615 // Replace the negated operand with 0.0:
8616 // fcmp Pred Op0, -Op0 --> fcmp Pred Op0, 0.0
8617 Constant *Zero = ConstantFP::getZero(Ty: Op0->getType());
8618 return new FCmpInst(Pred, Op0, Zero, "", &I);
8619}
8620
8621static Instruction *foldFCmpFSubIntoFCmp(FCmpInst &I, Instruction *LHSI,
8622 Constant *RHSC, InstCombinerImpl &CI) {
8623 const CmpInst::Predicate Pred = I.getPredicate();
8624 Value *X = LHSI->getOperand(i: 0);
8625 Value *Y = LHSI->getOperand(i: 1);
8626 switch (Pred) {
8627 default:
8628 break;
8629 case FCmpInst::FCMP_UGT:
8630 case FCmpInst::FCMP_ULT:
8631 case FCmpInst::FCMP_UNE:
8632 case FCmpInst::FCMP_OEQ:
8633 case FCmpInst::FCMP_OGE:
8634 case FCmpInst::FCMP_OLE:
8635 // The optimization is not valid if X and Y are infinities of the same
8636 // sign, i.e. the inf - inf = nan case. If the fsub has the ninf or nnan
8637 // flag then we can assume we do not have that case. Otherwise we might be
8638 // able to prove that either X or Y is not infinity.
8639 if (!LHSI->hasNoNaNs() && !LHSI->hasNoInfs() &&
8640 !isKnownNeverInfinity(V: Y,
8641 SQ: CI.getSimplifyQuery().getWithInstruction(I: &I)) &&
8642 !isKnownNeverInfinity(V: X, SQ: CI.getSimplifyQuery().getWithInstruction(I: &I)))
8643 break;
8644
8645 [[fallthrough]];
8646 case FCmpInst::FCMP_OGT:
8647 case FCmpInst::FCMP_OLT:
8648 case FCmpInst::FCMP_ONE:
8649 case FCmpInst::FCMP_UEQ:
8650 case FCmpInst::FCMP_UGE:
8651 case FCmpInst::FCMP_ULE:
8652 // fcmp pred (x - y), 0 --> fcmp pred x, y
8653 if (match(V: RHSC, P: m_AnyZeroFP()) &&
8654 I.getFunction()->getDenormalMode(
8655 FPType: LHSI->getType()->getScalarType()->getFltSemantics()) ==
8656 DenormalMode::getIEEE()) {
8657 CI.replaceOperand(I, OpNum: 0, V: X);
8658 CI.replaceOperand(I, OpNum: 1, V: Y);
8659 I.setHasNoInfs(LHSI->hasNoInfs());
8660 if (LHSI->hasNoNaNs())
8661 I.setHasNoNaNs(true);
8662 return &I;
8663 }
8664 break;
8665 }
8666
8667 return nullptr;
8668}
8669
8670static Instruction *foldFCmpWithFloorAndCeil(FCmpInst &I,
8671 InstCombinerImpl &IC) {
8672 Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1);
8673 Type *OpType = LHS->getType();
8674 CmpInst::Predicate Pred = I.getPredicate();
8675
8676 bool FloorX = match(V: LHS, P: m_Intrinsic<Intrinsic::floor>(Op0: m_Specific(V: RHS)));
8677 bool CeilX = match(V: LHS, P: m_Intrinsic<Intrinsic::ceil>(Op0: m_Specific(V: RHS)));
8678
8679 if (!FloorX && !CeilX) {
8680 if ((FloorX = match(V: RHS, P: m_Intrinsic<Intrinsic::floor>(Op0: m_Specific(V: LHS)))) ||
8681 (CeilX = match(V: RHS, P: m_Intrinsic<Intrinsic::ceil>(Op0: m_Specific(V: LHS))))) {
8682 std::swap(a&: LHS, b&: RHS);
8683 Pred = I.getSwappedPredicate();
8684 }
8685 }
8686
8687 switch (Pred) {
8688 case FCmpInst::FCMP_OLE:
8689 // fcmp ole floor(x), x => fcmp ord x, 0
8690 if (FloorX)
8691 return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(Ty: OpType),
8692 "", &I);
8693 break;
8694 case FCmpInst::FCMP_OGT:
8695 // fcmp ogt floor(x), x => false
8696 if (FloorX)
8697 return IC.replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8698 break;
8699 case FCmpInst::FCMP_OGE:
8700 // fcmp oge ceil(x), x => fcmp ord x, 0
8701 if (CeilX)
8702 return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(Ty: OpType),
8703 "", &I);
8704 break;
8705 case FCmpInst::FCMP_OLT:
8706 // fcmp olt ceil(x), x => false
8707 if (CeilX)
8708 return IC.replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
8709 break;
8710 case FCmpInst::FCMP_ULE:
8711 // fcmp ule floor(x), x => true
8712 if (FloorX)
8713 return IC.replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8714 break;
8715 case FCmpInst::FCMP_UGT:
8716 // fcmp ugt floor(x), x => fcmp uno x, 0
8717 if (FloorX)
8718 return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(Ty: OpType),
8719 "", &I);
8720 break;
8721 case FCmpInst::FCMP_UGE:
8722 // fcmp uge ceil(x), x => true
8723 if (CeilX)
8724 return IC.replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
8725 break;
8726 case FCmpInst::FCMP_ULT:
8727 // fcmp ult ceil(x), x => fcmp uno x, 0
8728 if (CeilX)
8729 return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(Ty: OpType),
8730 "", &I);
8731 break;
8732 default:
8733 break;
8734 }
8735
8736 return nullptr;
8737}
8738
8739/// Returns true if a select that implements a min/max is redundant and
8740/// select result can be replaced with its non-constant operand, e.g.,
8741/// select ( (si/ui-to-fp A) <= C ), C, (si/ui-to-fp A)
8742/// where C is the FP constant equal to the minimum integer value
8743/// representable by A.
8744static bool isMinMaxCmpSelectEliminable(SelectPatternFlavor Flavor, Value *A,
8745 Value *B) {
8746 const APFloat *APF;
8747 if (!match(V: B, P: m_APFloat(Res&: APF)))
8748 return false;
8749
8750 auto *I = dyn_cast<Instruction>(Val: A);
8751 if (!I || !(I->getOpcode() == Instruction::SIToFP ||
8752 I->getOpcode() == Instruction::UIToFP))
8753 return false;
8754
8755 bool IsUnsigned = I->getOpcode() == Instruction::UIToFP;
8756 unsigned BitWidth = I->getOperand(i: 0)->getType()->getScalarSizeInBits();
8757 APSInt IntBoundary = (Flavor == SPF_FMAXNUM)
8758 ? APSInt::getMinValue(numBits: BitWidth, Unsigned: IsUnsigned)
8759 : APSInt::getMaxValue(numBits: BitWidth, Unsigned: IsUnsigned);
8760 APSInt ConvertedInt(BitWidth, IsUnsigned);
8761 bool IsExact;
8762 APFloat::opStatus Status =
8763 APF->convertToInteger(Result&: ConvertedInt, RM: APFloat::rmTowardZero, IsExact: &IsExact);
8764 return Status == APFloat::opOK && IsExact && ConvertedInt == IntBoundary;
8765}
8766
8767Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
8768 bool Changed = false;
8769
8770 /// Orders the operands of the compare so that they are listed from most
8771 /// complex to least complex. This puts constants before unary operators,
8772 /// before binary operators.
8773 if (getComplexity(V: I.getOperand(i_nocapture: 0)) < getComplexity(V: I.getOperand(i_nocapture: 1))) {
8774 I.swapOperands();
8775 Changed = true;
8776 }
8777
8778 const CmpInst::Predicate Pred = I.getPredicate();
8779 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
8780 if (Value *V = simplifyFCmpInst(Predicate: Pred, LHS: Op0, RHS: Op1, FMF: I.getFastMathFlags(),
8781 Q: SQ.getWithInstruction(I: &I)))
8782 return replaceInstUsesWith(I, V);
8783
8784 // Simplify 'fcmp pred X, X'
8785 Type *OpType = Op0->getType();
8786 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
8787 if (Op0 == Op1) {
8788 switch (Pred) {
8789 default:
8790 break;
8791 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
8792 case FCmpInst::FCMP_ULT: // True if unordered or less than
8793 case FCmpInst::FCMP_UGT: // True if unordered or greater than
8794 case FCmpInst::FCMP_UNE: // True if unordered or not equal
8795 // Canonicalize these to be 'fcmp uno %X, 0.0'.
8796 I.setPredicate(FCmpInst::FCMP_UNO);
8797 I.setOperand(i_nocapture: 1, Val_nocapture: Constant::getNullValue(Ty: OpType));
8798 return &I;
8799
8800 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
8801 case FCmpInst::FCMP_OEQ: // True if ordered and equal
8802 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
8803 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
8804 // Canonicalize these to be 'fcmp ord %X, 0.0'.
8805 I.setPredicate(FCmpInst::FCMP_ORD);
8806 I.setOperand(i_nocapture: 1, Val_nocapture: Constant::getNullValue(Ty: OpType));
8807 return &I;
8808 }
8809 }
8810
8811 if (I.isCommutative()) {
8812 if (auto Pair = matchSymmetricPair(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1))) {
8813 replaceOperand(I, OpNum: 0, V: Pair->first);
8814 replaceOperand(I, OpNum: 1, V: Pair->second);
8815 return &I;
8816 }
8817 }
8818
8819 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
8820 // then canonicalize the operand to 0.0.
8821 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
8822 if (!match(V: Op0, P: m_PosZeroFP()) &&
8823 isKnownNeverNaN(V: Op0, SQ: getSimplifyQuery().getWithInstruction(I: &I)))
8824 return replaceOperand(I, OpNum: 0, V: ConstantFP::getZero(Ty: OpType));
8825
8826 if (!match(V: Op1, P: m_PosZeroFP()) &&
8827 isKnownNeverNaN(V: Op1, SQ: getSimplifyQuery().getWithInstruction(I: &I)))
8828 return replaceOperand(I, OpNum: 1, V: ConstantFP::getZero(Ty: OpType));
8829 }
8830
8831 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
8832 Value *X, *Y;
8833 if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_FNeg(X: m_Value(V&: Y))))
8834 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
8835
8836 if (Instruction *R = foldFCmpFNegCommonOp(I))
8837 return R;
8838
8839 // Test if the FCmpInst instruction is used exclusively by a select as
8840 // part of a minimum or maximum operation. If so, refrain from doing
8841 // any other folding. This helps out other analyses which understand
8842 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
8843 // and CodeGen. And in this case, at least one of the comparison
8844 // operands has at least one user besides the compare (the select),
8845 // which would often largely negate the benefit of folding anyway.
8846 if (I.hasOneUse())
8847 if (SelectInst *SI = dyn_cast<SelectInst>(Val: I.user_back())) {
8848 Value *A, *B;
8849 SelectPatternResult SPR = matchSelectPattern(V: SI, LHS&: A, RHS&: B);
8850 bool IsRedundantMinMaxClamp =
8851 (SPR.Flavor == SPF_FMAXNUM || SPR.Flavor == SPF_FMINNUM) &&
8852 isMinMaxCmpSelectEliminable(Flavor: SPR.Flavor, A, B);
8853 if (SPR.Flavor != SPF_UNKNOWN && !IsRedundantMinMaxClamp)
8854 return nullptr;
8855 }
8856
8857 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
8858 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
8859 if (match(V: Op1, P: m_AnyZeroFP()) && !match(V: Op1, P: m_PosZeroFP()))
8860 return replaceOperand(I, OpNum: 1, V: ConstantFP::getZero(Ty: OpType));
8861
8862 // Canonicalize:
8863 // fcmp olt X, +inf -> fcmp one X, +inf
8864 // fcmp ole X, +inf -> fcmp ord X, 0
8865 // fcmp ogt X, +inf -> false
8866 // fcmp oge X, +inf -> fcmp oeq X, +inf
8867 // fcmp ult X, +inf -> fcmp une X, +inf
8868 // fcmp ule X, +inf -> true
8869 // fcmp ugt X, +inf -> fcmp uno X, 0
8870 // fcmp uge X, +inf -> fcmp ueq X, +inf
8871 // fcmp olt X, -inf -> false
8872 // fcmp ole X, -inf -> fcmp oeq X, -inf
8873 // fcmp ogt X, -inf -> fcmp one X, -inf
8874 // fcmp oge X, -inf -> fcmp ord X, 0
8875 // fcmp ult X, -inf -> fcmp uno X, 0
8876 // fcmp ule X, -inf -> fcmp ueq X, -inf
8877 // fcmp ugt X, -inf -> fcmp une X, -inf
8878 // fcmp uge X, -inf -> true
8879 const APFloat *C;
8880 if (match(V: Op1, P: m_APFloat(Res&: C)) && C->isInfinity()) {
8881 switch (C->isNegative() ? FCmpInst::getSwappedPredicate(pred: Pred) : Pred) {
8882 default:
8883 break;
8884 case FCmpInst::FCMP_ORD:
8885 case FCmpInst::FCMP_UNO:
8886 case FCmpInst::FCMP_TRUE:
8887 case FCmpInst::FCMP_FALSE:
8888 case FCmpInst::FCMP_OGT:
8889 case FCmpInst::FCMP_ULE:
8890 llvm_unreachable("Should be simplified by InstSimplify");
8891 case FCmpInst::FCMP_OLT:
8892 return new FCmpInst(FCmpInst::FCMP_ONE, Op0, Op1, "", &I);
8893 case FCmpInst::FCMP_OLE:
8894 return new FCmpInst(FCmpInst::FCMP_ORD, Op0, ConstantFP::getZero(Ty: OpType),
8895 "", &I);
8896 case FCmpInst::FCMP_OGE:
8897 return new FCmpInst(FCmpInst::FCMP_OEQ, Op0, Op1, "", &I);
8898 case FCmpInst::FCMP_ULT:
8899 return new FCmpInst(FCmpInst::FCMP_UNE, Op0, Op1, "", &I);
8900 case FCmpInst::FCMP_UGT:
8901 return new FCmpInst(FCmpInst::FCMP_UNO, Op0, ConstantFP::getZero(Ty: OpType),
8902 "", &I);
8903 case FCmpInst::FCMP_UGE:
8904 return new FCmpInst(FCmpInst::FCMP_UEQ, Op0, Op1, "", &I);
8905 }
8906 }
8907
8908 // Ignore signbit of bitcasted int when comparing equality to FP 0.0:
8909 // fcmp oeq/une (bitcast X), 0.0 --> (and X, SignMaskC) ==/!= 0
8910 if (match(V: Op1, P: m_PosZeroFP()) &&
8911 match(V: Op0, P: m_OneUse(SubPattern: m_ElementWiseBitCast(Op: m_Value(V&: X))))) {
8912 ICmpInst::Predicate IntPred = ICmpInst::BAD_ICMP_PREDICATE;
8913 if (Pred == FCmpInst::FCMP_OEQ)
8914 IntPred = ICmpInst::ICMP_EQ;
8915 else if (Pred == FCmpInst::FCMP_UNE)
8916 IntPred = ICmpInst::ICMP_NE;
8917
8918 if (IntPred != ICmpInst::BAD_ICMP_PREDICATE) {
8919 Type *IntTy = X->getType();
8920 const APInt &SignMask = ~APInt::getSignMask(BitWidth: IntTy->getScalarSizeInBits());
8921 Value *MaskX = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty: IntTy, V: SignMask));
8922 return new ICmpInst(IntPred, MaskX, ConstantInt::getNullValue(Ty: IntTy));
8923 }
8924 }
8925
8926 // Handle fcmp with instruction LHS and constant RHS.
8927 Instruction *LHSI;
8928 Constant *RHSC;
8929 if (match(V: Op0, P: m_Instruction(I&: LHSI)) && match(V: Op1, P: m_Constant(C&: RHSC))) {
8930 switch (LHSI->getOpcode()) {
8931 case Instruction::Select:
8932 // fcmp eq (cond ? x : -x), 0 --> fcmp eq x, 0
8933 if (FCmpInst::isEquality(Pred) && match(V: RHSC, P: m_AnyZeroFP()) &&
8934 match(V: LHSI, P: m_c_Select(L: m_FNeg(X: m_Value(V&: X)), R: m_Deferred(V: X))))
8935 return replaceOperand(I, OpNum: 0, V: X);
8936 if (Instruction *NV = FoldOpIntoSelect(Op&: I, SI: cast<SelectInst>(Val: LHSI)))
8937 return NV;
8938 break;
8939 case Instruction::FSub:
8940 if (LHSI->hasOneUse())
8941 if (Instruction *NV = foldFCmpFSubIntoFCmp(I, LHSI, RHSC, CI&: *this))
8942 return NV;
8943 break;
8944 case Instruction::PHI:
8945 if (Instruction *NV = foldOpIntoPhi(I, PN: cast<PHINode>(Val: LHSI)))
8946 return NV;
8947 break;
8948 case Instruction::SIToFP:
8949 case Instruction::UIToFP:
8950 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
8951 return NV;
8952 break;
8953 case Instruction::FDiv:
8954 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
8955 return NV;
8956 break;
8957 case Instruction::Load:
8958 if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: LHSI->getOperand(i: 0)))
8959 if (Instruction *Res =
8960 foldCmpLoadFromIndexedGlobal(LI: cast<LoadInst>(Val: LHSI), GEP, ICI&: I))
8961 return Res;
8962 break;
8963 case Instruction::FPTrunc:
8964 if (Instruction *NV = foldFCmpFpTrunc(I, FPTrunc: *LHSI, C: *RHSC))
8965 return NV;
8966 break;
8967 }
8968 }
8969
8970 if (Instruction *R = foldFabsWithFcmpZero(I, IC&: *this))
8971 return R;
8972
8973 if (Instruction *R = foldSqrtWithFcmpZero(I, IC&: *this))
8974 return R;
8975
8976 if (Instruction *R = foldFCmpWithFloorAndCeil(I, IC&: *this))
8977 return R;
8978
8979 if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X)))) {
8980 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
8981 Constant *C;
8982 if (match(V: Op1, P: m_Constant(C)))
8983 if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL))
8984 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
8985 }
8986
8987 // fcmp (fadd X, 0.0), Y --> fcmp X, Y
8988 if (match(V: Op0, P: m_FAdd(L: m_Value(V&: X), R: m_AnyZeroFP())))
8989 return new FCmpInst(Pred, X, Op1, "", &I);
8990
8991 // fcmp X, (fadd Y, 0.0) --> fcmp X, Y
8992 if (match(V: Op1, P: m_FAdd(L: m_Value(V&: Y), R: m_AnyZeroFP())))
8993 return new FCmpInst(Pred, Op0, Y, "", &I);
8994
8995 if (match(V: Op0, P: m_FPExt(Op: m_Value(V&: X)))) {
8996 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
8997 if (match(V: Op1, P: m_FPExt(Op: m_Value(V&: Y))) && X->getType() == Y->getType())
8998 return new FCmpInst(Pred, X, Y, "", &I);
8999
9000 const APFloat *C;
9001 if (match(V: Op1, P: m_APFloat(Res&: C))) {
9002 const fltSemantics &FPSem =
9003 X->getType()->getScalarType()->getFltSemantics();
9004 bool Lossy;
9005 APFloat TruncC = *C;
9006 TruncC.convert(ToSemantics: FPSem, RM: APFloat::rmNearestTiesToEven, losesInfo: &Lossy);
9007
9008 if (Lossy) {
9009 // X can't possibly equal the higher-precision constant, so reduce any
9010 // equality comparison.
9011 // TODO: Other predicates can be handled via getFCmpCode().
9012 switch (Pred) {
9013 case FCmpInst::FCMP_OEQ:
9014 // X is ordered and equal to an impossible constant --> false
9015 return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType()));
9016 case FCmpInst::FCMP_ONE:
9017 // X is ordered and not equal to an impossible constant --> ordered
9018 return new FCmpInst(FCmpInst::FCMP_ORD, X,
9019 ConstantFP::getZero(Ty: X->getType()));
9020 case FCmpInst::FCMP_UEQ:
9021 // X is unordered or equal to an impossible constant --> unordered
9022 return new FCmpInst(FCmpInst::FCMP_UNO, X,
9023 ConstantFP::getZero(Ty: X->getType()));
9024 case FCmpInst::FCMP_UNE:
9025 // X is unordered or not equal to an impossible constant --> true
9026 return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType()));
9027 default:
9028 break;
9029 }
9030 }
9031
9032 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
9033 // Avoid lossy conversions and denormals.
9034 // Zero is a special case that's OK to convert.
9035 APFloat Fabs = TruncC;
9036 Fabs.clearSign();
9037 if (!Lossy &&
9038 (Fabs.isZero() || !(Fabs < APFloat::getSmallestNormalized(Sem: FPSem)))) {
9039 Constant *NewC = ConstantFP::get(Ty: X->getType(), V: TruncC);
9040 return new FCmpInst(Pred, X, NewC, "", &I);
9041 }
9042 }
9043 }
9044
9045 // Convert a sign-bit test of an FP value into a cast and integer compare.
9046 // TODO: Simplify if the copysign constant is 0.0 or NaN.
9047 // TODO: Handle non-zero compare constants.
9048 // TODO: Handle other predicates.
9049 if (match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::copysign>(Op0: m_APFloat(Res&: C),
9050 Op1: m_Value(V&: X)))) &&
9051 match(V: Op1, P: m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) {
9052 Type *IntType = Builder.getIntNTy(N: X->getType()->getScalarSizeInBits());
9053 if (auto *VecTy = dyn_cast<VectorType>(Val: OpType))
9054 IntType = VectorType::get(ElementType: IntType, EC: VecTy->getElementCount());
9055
9056 // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0
9057 if (Pred == FCmpInst::FCMP_OLT) {
9058 Value *IntX = Builder.CreateBitCast(V: X, DestTy: IntType);
9059 return new ICmpInst(ICmpInst::ICMP_SLT, IntX,
9060 ConstantInt::getNullValue(Ty: IntType));
9061 }
9062 }
9063
9064 {
9065 Value *CanonLHS = nullptr;
9066 match(V: Op0, P: m_Intrinsic<Intrinsic::canonicalize>(Op0: m_Value(V&: CanonLHS)));
9067 // (canonicalize(x) == x) => (x == x)
9068 if (CanonLHS == Op1)
9069 return new FCmpInst(Pred, Op1, Op1, "", &I);
9070
9071 Value *CanonRHS = nullptr;
9072 match(V: Op1, P: m_Intrinsic<Intrinsic::canonicalize>(Op0: m_Value(V&: CanonRHS)));
9073 // (x == canonicalize(x)) => (x == x)
9074 if (CanonRHS == Op0)
9075 return new FCmpInst(Pred, Op0, Op0, "", &I);
9076
9077 // (canonicalize(x) == canonicalize(y)) => (x == y)
9078 if (CanonLHS && CanonRHS)
9079 return new FCmpInst(Pred, CanonLHS, CanonRHS, "", &I);
9080 }
9081
9082 if (I.getType()->isVectorTy())
9083 if (Instruction *Res = foldVectorCmp(Cmp&: I, Builder))
9084 return Res;
9085
9086 return Changed ? &I : nullptr;
9087}
9088