1//===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
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 routines for folding instructions into simpler forms
10// that do not require creating new instructions. This does constant folding
11// ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
12// returning a constant ("and i32 %x, 0" -> "0") or an already existing value
13// ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
14// simplified: This is usually true and assuming it simplifies the logic (if
15// they have not been simplified then results are correct but maybe suboptimal).
16//
17//===----------------------------------------------------------------------===//
18
19#include "llvm/Analysis/InstructionSimplify.h"
20
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/ADT/SetVector.h"
23#include "llvm/ADT/Statistic.h"
24#include "llvm/Analysis/AliasAnalysis.h"
25#include "llvm/Analysis/AssumptionCache.h"
26#include "llvm/Analysis/CaptureTracking.h"
27#include "llvm/Analysis/CmpInstAnalysis.h"
28#include "llvm/Analysis/ConstantFolding.h"
29#include "llvm/Analysis/FloatingPointPredicateUtils.h"
30#include "llvm/Analysis/InstSimplifyFolder.h"
31#include "llvm/Analysis/Loads.h"
32#include "llvm/Analysis/LoopAnalysisManager.h"
33#include "llvm/Analysis/MemoryBuiltins.h"
34#include "llvm/Analysis/OverflowInstAnalysis.h"
35#include "llvm/Analysis/TargetLibraryInfo.h"
36#include "llvm/Analysis/ValueTracking.h"
37#include "llvm/Analysis/VectorUtils.h"
38#include "llvm/IR/ConstantFPRange.h"
39#include "llvm/IR/ConstantRange.h"
40#include "llvm/IR/DataLayout.h"
41#include "llvm/IR/Dominators.h"
42#include "llvm/IR/InstrTypes.h"
43#include "llvm/IR/Instructions.h"
44#include "llvm/IR/IntrinsicsAArch64.h"
45#include "llvm/IR/Operator.h"
46#include "llvm/IR/PatternMatch.h"
47#include "llvm/IR/Statepoint.h"
48#include "llvm/Support/KnownBits.h"
49#include "llvm/Support/KnownFPClass.h"
50#include <algorithm>
51#include <optional>
52using namespace llvm;
53using namespace llvm::PatternMatch;
54
55#define DEBUG_TYPE "instsimplify"
56
57enum { RecursionLimit = 3 };
58
59STATISTIC(NumExpand, "Number of expansions");
60STATISTIC(NumReassoc, "Number of reassociations");
61
62static Value *simplifyAndInst(Value *, Value *, const SimplifyQuery &,
63 unsigned);
64static Value *simplifyUnOp(unsigned, Value *, const SimplifyQuery &, unsigned);
65static Value *simplifyFPUnOp(unsigned, Value *, const FastMathFlags &,
66 const SimplifyQuery &, unsigned);
67static Value *simplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
68 unsigned);
69static Value *simplifyBinOp(unsigned, Value *, Value *, const FastMathFlags &,
70 const SimplifyQuery &, unsigned);
71static Value *simplifyCmpInst(CmpPredicate, Value *, Value *,
72 const SimplifyQuery &, unsigned);
73static Value *simplifyICmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
74 const SimplifyQuery &Q, unsigned MaxRecurse);
75static Value *simplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
76static Value *simplifyXorInst(Value *, Value *, const SimplifyQuery &,
77 unsigned);
78static Value *simplifyCastInst(unsigned, Value *, Type *, const SimplifyQuery &,
79 unsigned);
80static Value *simplifyGEPInst(Type *, Value *, ArrayRef<Value *>,
81 GEPNoWrapFlags, const SimplifyQuery &, unsigned);
82static Value *simplifySelectInst(Value *, Value *, Value *,
83 const SimplifyQuery &, unsigned);
84static Value *simplifyInstructionWithOperands(Instruction *I,
85 ArrayRef<Value *> NewOps,
86 const SimplifyQuery &SQ,
87 unsigned MaxRecurse);
88
89/// For a boolean type or a vector of boolean type, return false or a vector
90/// with every element false.
91static Constant *getFalse(Type *Ty) { return ConstantInt::getFalse(Ty); }
92
93/// For a boolean type or a vector of boolean type, return true or a vector
94/// with every element true.
95static Constant *getTrue(Type *Ty) { return ConstantInt::getTrue(Ty); }
96
97/// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
98static bool isSameCompare(Value *V, CmpPredicate Pred, Value *LHS, Value *RHS) {
99 CmpInst *Cmp = dyn_cast<CmpInst>(Val: V);
100 if (!Cmp)
101 return false;
102 CmpInst::Predicate CPred = Cmp->getPredicate();
103 Value *CLHS = Cmp->getOperand(i_nocapture: 0), *CRHS = Cmp->getOperand(i_nocapture: 1);
104 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
105 return true;
106 return CPred == CmpInst::getSwappedPredicate(pred: Pred) && CLHS == RHS &&
107 CRHS == LHS;
108}
109
110/// Simplify comparison with true or false branch of select:
111/// %sel = select i1 %cond, i32 %tv, i32 %fv
112/// %cmp = icmp sle i32 %sel, %rhs
113/// Compose new comparison by substituting %sel with either %tv or %fv
114/// and see if it simplifies.
115static Value *simplifyCmpSelCase(CmpPredicate Pred, Value *LHS, Value *RHS,
116 Value *Cond, const SimplifyQuery &Q,
117 unsigned MaxRecurse, Constant *TrueOrFalse) {
118 Value *SimplifiedCmp = simplifyCmpInst(Pred, LHS, RHS, Q, MaxRecurse);
119 if (SimplifiedCmp == Cond) {
120 // %cmp simplified to the select condition (%cond).
121 return TrueOrFalse;
122 } else if (!SimplifiedCmp && isSameCompare(V: Cond, Pred, LHS, RHS)) {
123 // It didn't simplify. However, if composed comparison is equivalent
124 // to the select condition (%cond) then we can replace it.
125 return TrueOrFalse;
126 }
127 return SimplifiedCmp;
128}
129
130/// Simplify comparison with true branch of select
131static Value *simplifyCmpSelTrueCase(CmpPredicate Pred, Value *LHS, Value *RHS,
132 Value *Cond, const SimplifyQuery &Q,
133 unsigned MaxRecurse) {
134 return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
135 TrueOrFalse: getTrue(Ty: Cond->getType()));
136}
137
138/// Simplify comparison with false branch of select
139static Value *simplifyCmpSelFalseCase(CmpPredicate Pred, Value *LHS, Value *RHS,
140 Value *Cond, const SimplifyQuery &Q,
141 unsigned MaxRecurse) {
142 return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
143 TrueOrFalse: getFalse(Ty: Cond->getType()));
144}
145
146/// We know comparison with both branches of select can be simplified, but they
147/// are not equal. This routine handles some logical simplifications.
148static Value *handleOtherCmpSelSimplifications(Value *TCmp, Value *FCmp,
149 Value *Cond,
150 const SimplifyQuery &Q,
151 unsigned MaxRecurse) {
152 // If the false value simplified to false, then the result of the compare
153 // is equal to "Cond && TCmp". This also catches the case when the false
154 // value simplified to false and the true value to true, returning "Cond".
155 // Folding select to and/or isn't poison-safe in general; impliesPoison
156 // checks whether folding it does not convert a well-defined value into
157 // poison.
158 if (match(V: FCmp, P: m_Zero()) && impliesPoison(ValAssumedPoison: TCmp, V: Cond))
159 if (Value *V = simplifyAndInst(Cond, TCmp, Q, MaxRecurse))
160 return V;
161 // If the true value simplified to true, then the result of the compare
162 // is equal to "Cond || FCmp".
163 if (match(V: TCmp, P: m_One()) && impliesPoison(ValAssumedPoison: FCmp, V: Cond))
164 if (Value *V = simplifyOrInst(Cond, FCmp, Q, MaxRecurse))
165 return V;
166 // Finally, if the false value simplified to true and the true value to
167 // false, then the result of the compare is equal to "!Cond".
168 if (match(V: FCmp, P: m_One()) && match(V: TCmp, P: m_Zero()))
169 if (Value *V = simplifyXorInst(
170 Cond, Constant::getAllOnesValue(Ty: Cond->getType()), Q, MaxRecurse))
171 return V;
172 return nullptr;
173}
174
175/// Does the given value dominate the specified phi node?
176static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
177 Instruction *I = dyn_cast<Instruction>(Val: V);
178 if (!I)
179 // Arguments and constants dominate all instructions.
180 return true;
181
182 // If we have a DominatorTree then do a precise test.
183 if (DT)
184 return DT->dominates(Def: I, User: P);
185
186 // Otherwise, if the instruction is in the entry block and is not an invoke,
187 // then it obviously dominates all phi nodes.
188 if (I->getParent()->isEntryBlock() && !isa<InvokeInst>(Val: I) &&
189 !isa<CallBrInst>(Val: I))
190 return true;
191
192 return false;
193}
194
195/// Try to simplify a binary operator of form "V op OtherOp" where V is
196/// "(B0 opex B1)" by distributing 'op' across 'opex' as
197/// "(B0 op OtherOp) opex (B1 op OtherOp)".
198static Value *expandBinOp(Instruction::BinaryOps Opcode, Value *V,
199 Value *OtherOp, Instruction::BinaryOps OpcodeToExpand,
200 const SimplifyQuery &Q, unsigned MaxRecurse) {
201 auto *B = dyn_cast<BinaryOperator>(Val: V);
202 if (!B || B->getOpcode() != OpcodeToExpand)
203 return nullptr;
204 Value *B0 = B->getOperand(i_nocapture: 0), *B1 = B->getOperand(i_nocapture: 1);
205 Value *L =
206 simplifyBinOp(Opcode, B0, OtherOp, Q.getWithoutUndef(), MaxRecurse);
207 if (!L)
208 return nullptr;
209 Value *R =
210 simplifyBinOp(Opcode, B1, OtherOp, Q.getWithoutUndef(), MaxRecurse);
211 if (!R)
212 return nullptr;
213
214 // Does the expanded pair of binops simplify to the existing binop?
215 if ((L == B0 && R == B1) ||
216 (Instruction::isCommutative(Opcode: OpcodeToExpand) && L == B1 && R == B0)) {
217 ++NumExpand;
218 return B;
219 }
220
221 // Otherwise, return "L op' R" if it simplifies.
222 Value *S = simplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse);
223 if (!S)
224 return nullptr;
225
226 ++NumExpand;
227 return S;
228}
229
230/// Try to simplify binops of form "A op (B op' C)" or the commuted variant by
231/// distributing op over op'.
232static Value *expandCommutativeBinOp(Instruction::BinaryOps Opcode, Value *L,
233 Value *R,
234 Instruction::BinaryOps OpcodeToExpand,
235 const SimplifyQuery &Q,
236 unsigned MaxRecurse) {
237 // Recursion is always used, so bail out at once if we already hit the limit.
238 if (!MaxRecurse--)
239 return nullptr;
240
241 if (Value *V = expandBinOp(Opcode, V: L, OtherOp: R, OpcodeToExpand, Q, MaxRecurse))
242 return V;
243 if (Value *V = expandBinOp(Opcode, V: R, OtherOp: L, OpcodeToExpand, Q, MaxRecurse))
244 return V;
245 return nullptr;
246}
247
248/// Generic simplifications for associative binary operations.
249/// Returns the simpler value, or null if none was found.
250static Value *simplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
251 Value *LHS, Value *RHS,
252 const SimplifyQuery &Q,
253 unsigned MaxRecurse) {
254 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
255
256 // Recursion is always used, so bail out at once if we already hit the limit.
257 if (!MaxRecurse--)
258 return nullptr;
259
260 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(Val: LHS);
261 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(Val: RHS);
262
263 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
264 if (Op0 && Op0->getOpcode() == Opcode) {
265 Value *A = Op0->getOperand(i_nocapture: 0);
266 Value *B = Op0->getOperand(i_nocapture: 1);
267 Value *C = RHS;
268
269 // Does "B op C" simplify?
270 if (Value *V = simplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
271 // It does! Return "A op V" if it simplifies or is already available.
272 // If V equals B then "A op V" is just the LHS.
273 if (V == B)
274 return LHS;
275 // Otherwise return "A op V" if it simplifies.
276 if (Value *W = simplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
277 ++NumReassoc;
278 return W;
279 }
280 }
281 }
282
283 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
284 if (Op1 && Op1->getOpcode() == Opcode) {
285 Value *A = LHS;
286 Value *B = Op1->getOperand(i_nocapture: 0);
287 Value *C = Op1->getOperand(i_nocapture: 1);
288
289 // Does "A op B" simplify?
290 if (Value *V = simplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
291 // It does! Return "V op C" if it simplifies or is already available.
292 // If V equals B then "V op C" is just the RHS.
293 if (V == B)
294 return RHS;
295 // Otherwise return "V op C" if it simplifies.
296 if (Value *W = simplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
297 ++NumReassoc;
298 return W;
299 }
300 }
301 }
302
303 // The remaining transforms require commutativity as well as associativity.
304 if (!Instruction::isCommutative(Opcode))
305 return nullptr;
306
307 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
308 if (Op0 && Op0->getOpcode() == Opcode) {
309 Value *A = Op0->getOperand(i_nocapture: 0);
310 Value *B = Op0->getOperand(i_nocapture: 1);
311 Value *C = RHS;
312
313 // Does "C op A" simplify?
314 if (Value *V = simplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
315 // It does! Return "V op B" if it simplifies or is already available.
316 // If V equals A then "V op B" is just the LHS.
317 if (V == A)
318 return LHS;
319 // Otherwise return "V op B" if it simplifies.
320 if (Value *W = simplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
321 ++NumReassoc;
322 return W;
323 }
324 }
325 }
326
327 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
328 if (Op1 && Op1->getOpcode() == Opcode) {
329 Value *A = LHS;
330 Value *B = Op1->getOperand(i_nocapture: 0);
331 Value *C = Op1->getOperand(i_nocapture: 1);
332
333 // Does "C op A" simplify?
334 if (Value *V = simplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
335 // It does! Return "B op V" if it simplifies or is already available.
336 // If V equals C then "B op V" is just the RHS.
337 if (V == C)
338 return RHS;
339 // Otherwise return "B op V" if it simplifies.
340 if (Value *W = simplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
341 ++NumReassoc;
342 return W;
343 }
344 }
345 }
346
347 return nullptr;
348}
349
350/// In the case of a binary operation with a select instruction as an operand,
351/// try to simplify the binop by seeing whether evaluating it on both branches
352/// of the select results in the same value. Returns the common value if so,
353/// otherwise returns null.
354static Value *threadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
355 Value *RHS, const SimplifyQuery &Q,
356 unsigned MaxRecurse) {
357 // Recursion is always used, so bail out at once if we already hit the limit.
358 if (!MaxRecurse--)
359 return nullptr;
360
361 SelectInst *SI;
362 if (isa<SelectInst>(Val: LHS)) {
363 SI = cast<SelectInst>(Val: LHS);
364 } else {
365 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
366 SI = cast<SelectInst>(Val: RHS);
367 }
368
369 // Evaluate the BinOp on the true and false branches of the select.
370 Value *TV;
371 Value *FV;
372 if (SI == LHS) {
373 TV = simplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
374 FV = simplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
375 } else {
376 TV = simplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
377 FV = simplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
378 }
379
380 // If they simplified to the same value, then return the common value.
381 // If they both failed to simplify then return null.
382 if (TV == FV)
383 return TV;
384
385 // If one branch simplified to undef, return the other one.
386 if (TV && Q.isUndefValue(V: TV))
387 return FV;
388 if (FV && Q.isUndefValue(V: FV))
389 return TV;
390
391 // If applying the operation did not change the true and false select values,
392 // then the result of the binop is the select itself.
393 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
394 return SI;
395
396 // If one branch simplified and the other did not, and the simplified
397 // value is equal to the unsimplified one, return the simplified value.
398 // For example, select (cond, X, X & Z) & Z -> X & Z.
399 if ((FV && !TV) || (TV && !FV)) {
400 // Check that the simplified value has the form "X op Y" where "op" is the
401 // same as the original operation.
402 Instruction *Simplified = dyn_cast<Instruction>(Val: FV ? FV : TV);
403 if (Simplified && Simplified->getOpcode() == unsigned(Opcode) &&
404 !Simplified->hasPoisonGeneratingFlags()) {
405 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
406 // We already know that "op" is the same as for the simplified value. See
407 // if the operands match too. If so, return the simplified value.
408 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
409 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
410 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
411 if (Simplified->getOperand(i: 0) == UnsimplifiedLHS &&
412 Simplified->getOperand(i: 1) == UnsimplifiedRHS)
413 return Simplified;
414 if (Simplified->isCommutative() &&
415 Simplified->getOperand(i: 1) == UnsimplifiedLHS &&
416 Simplified->getOperand(i: 0) == UnsimplifiedRHS)
417 return Simplified;
418 }
419 }
420
421 return nullptr;
422}
423
424/// In the case of a comparison with a select instruction, try to simplify the
425/// comparison by seeing whether both branches of the select result in the same
426/// value. Returns the common value if so, otherwise returns null.
427/// For example, if we have:
428/// %tmp = select i1 %cmp, i32 1, i32 2
429/// %cmp1 = icmp sle i32 %tmp, 3
430/// We can simplify %cmp1 to true, because both branches of select are
431/// less than 3. We compose new comparison by substituting %tmp with both
432/// branches of select and see if it can be simplified.
433static Value *threadCmpOverSelect(CmpPredicate Pred, Value *LHS, Value *RHS,
434 const SimplifyQuery &Q, unsigned MaxRecurse) {
435 // Recursion is always used, so bail out at once if we already hit the limit.
436 if (!MaxRecurse--)
437 return nullptr;
438
439 // Make sure the select is on the LHS.
440 if (!isa<SelectInst>(Val: LHS)) {
441 std::swap(a&: LHS, b&: RHS);
442 Pred = CmpInst::getSwappedPredicate(pred: Pred);
443 }
444 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
445 SelectInst *SI = cast<SelectInst>(Val: LHS);
446 Value *Cond = SI->getCondition();
447 Value *TV = SI->getTrueValue();
448 Value *FV = SI->getFalseValue();
449
450 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
451 // Does "cmp TV, RHS" simplify?
452 Value *TCmp = simplifyCmpSelTrueCase(Pred, LHS: TV, RHS, Cond, Q, MaxRecurse);
453 if (!TCmp)
454 return nullptr;
455
456 // Does "cmp FV, RHS" simplify?
457 Value *FCmp = simplifyCmpSelFalseCase(Pred, LHS: FV, RHS, Cond, Q, MaxRecurse);
458 if (!FCmp)
459 return nullptr;
460
461 // If both sides simplified to the same value, then use it as the result of
462 // the original comparison.
463 if (TCmp == FCmp)
464 return TCmp;
465
466 // The remaining cases only make sense if the select condition has the same
467 // type as the result of the comparison, so bail out if this is not so.
468 if (Cond->getType()->isVectorTy() == RHS->getType()->isVectorTy())
469 return handleOtherCmpSelSimplifications(TCmp, FCmp, Cond, Q, MaxRecurse);
470
471 return nullptr;
472}
473
474/// In the case of a binary operation with an operand that is a PHI instruction,
475/// try to simplify the binop by seeing whether evaluating it on the incoming
476/// phi values yields the same result for every value. If so returns the common
477/// value, otherwise returns null.
478static Value *threadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
479 Value *RHS, const SimplifyQuery &Q,
480 unsigned MaxRecurse) {
481 // Recursion is always used, so bail out at once if we already hit the limit.
482 if (!MaxRecurse--)
483 return nullptr;
484
485 PHINode *PI;
486 if (isa<PHINode>(Val: LHS)) {
487 PI = cast<PHINode>(Val: LHS);
488 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
489 if (!valueDominatesPHI(V: RHS, P: PI, DT: Q.DT))
490 return nullptr;
491 } else {
492 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
493 PI = cast<PHINode>(Val: RHS);
494 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
495 if (!valueDominatesPHI(V: LHS, P: PI, DT: Q.DT))
496 return nullptr;
497 }
498
499 // Evaluate the BinOp on the incoming phi values.
500 Value *CommonValue = nullptr;
501 for (Use &Incoming : PI->incoming_values()) {
502 // If the incoming value is the phi node itself, it can safely be skipped.
503 if (Incoming == PI)
504 continue;
505 Instruction *InTI = PI->getIncomingBlock(U: Incoming)->getTerminator();
506 Value *V = PI == LHS
507 ? simplifyBinOp(Opcode, Incoming, RHS,
508 Q.getWithInstruction(I: InTI), MaxRecurse)
509 : simplifyBinOp(Opcode, LHS, Incoming,
510 Q.getWithInstruction(I: InTI), MaxRecurse);
511 // If the operation failed to simplify, or simplified to a different value
512 // to previously, then give up.
513 if (!V || (CommonValue && V != CommonValue))
514 return nullptr;
515 CommonValue = V;
516 }
517
518 return CommonValue;
519}
520
521/// In the case of a comparison with a PHI instruction, try to simplify the
522/// comparison by seeing whether comparing with all of the incoming phi values
523/// yields the same result every time. If so returns the common result,
524/// otherwise returns null.
525static Value *threadCmpOverPHI(CmpPredicate Pred, Value *LHS, Value *RHS,
526 const SimplifyQuery &Q, unsigned MaxRecurse) {
527 // Recursion is always used, so bail out at once if we already hit the limit.
528 if (!MaxRecurse--)
529 return nullptr;
530
531 // Make sure the phi is on the LHS.
532 if (!isa<PHINode>(Val: LHS)) {
533 std::swap(a&: LHS, b&: RHS);
534 Pred = CmpInst::getSwappedPredicate(pred: Pred);
535 }
536 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
537 PHINode *PI = cast<PHINode>(Val: LHS);
538
539 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
540 if (!valueDominatesPHI(V: RHS, P: PI, DT: Q.DT))
541 return nullptr;
542
543 // Evaluate the BinOp on the incoming phi values.
544 Value *CommonValue = nullptr;
545 for (unsigned u = 0, e = PI->getNumIncomingValues(); u < e; ++u) {
546 Value *Incoming = PI->getIncomingValue(i: u);
547 Instruction *InTI = PI->getIncomingBlock(i: u)->getTerminator();
548 // If the incoming value is the phi node itself, it can safely be skipped.
549 if (Incoming == PI)
550 continue;
551 // Change the context instruction to the "edge" that flows into the phi.
552 // This is important because that is where incoming is actually "evaluated"
553 // even though it is used later somewhere else.
554 Value *V = simplifyCmpInst(Pred, Incoming, RHS, Q.getWithInstruction(I: InTI),
555 MaxRecurse);
556 // If the operation failed to simplify, or simplified to a different value
557 // to previously, then give up.
558 if (!V || (CommonValue && V != CommonValue))
559 return nullptr;
560 CommonValue = V;
561 }
562
563 return CommonValue;
564}
565
566static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
567 Value *&Op0, Value *&Op1,
568 const SimplifyQuery &Q) {
569 if (auto *CLHS = dyn_cast<Constant>(Val: Op0)) {
570 if (auto *CRHS = dyn_cast<Constant>(Val: Op1)) {
571 switch (Opcode) {
572 default:
573 break;
574 case Instruction::FAdd:
575 case Instruction::FSub:
576 case Instruction::FMul:
577 case Instruction::FDiv:
578 case Instruction::FRem:
579 if (Q.CxtI != nullptr)
580 return ConstantFoldFPInstOperands(Opcode, LHS: CLHS, RHS: CRHS, DL: Q.DL, I: Q.CxtI);
581 }
582 return ConstantFoldBinaryOpOperands(Opcode, LHS: CLHS, RHS: CRHS, DL: Q.DL);
583 }
584
585 // Canonicalize the constant to the RHS if this is a commutative operation.
586 if (Instruction::isCommutative(Opcode))
587 std::swap(a&: Op0, b&: Op1);
588 }
589 return nullptr;
590}
591
592/// Given operands for an Add, see if we can fold the result.
593/// If not, this returns null.
594static Value *simplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
595 const SimplifyQuery &Q, unsigned MaxRecurse) {
596 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Add, Op0, Op1, Q))
597 return C;
598
599 // X + poison -> poison
600 if (isa<PoisonValue>(Val: Op1))
601 return Op1;
602
603 // X + undef -> undef
604 if (Q.isUndefValue(V: Op1))
605 return Op1;
606
607 // X + 0 -> X
608 if (match(V: Op1, P: m_Zero()))
609 return Op0;
610
611 // If two operands are negative, return 0.
612 if (isKnownNegation(X: Op0, Y: Op1))
613 return Constant::getNullValue(Ty: Op0->getType());
614
615 // X + (Y - X) -> Y
616 // (Y - X) + X -> Y
617 // Eg: X + -X -> 0
618 Value *Y = nullptr;
619 if (match(V: Op1, P: m_Sub(L: m_Value(V&: Y), R: m_Specific(V: Op0))) ||
620 match(V: Op0, P: m_Sub(L: m_Value(V&: Y), R: m_Specific(V: Op1))))
621 return Y;
622
623 // X + ~X -> -1 since ~X = -X-1
624 Type *Ty = Op0->getType();
625 if (match(V: Op0, P: m_Not(V: m_Specific(V: Op1))) || match(V: Op1, P: m_Not(V: m_Specific(V: Op0))))
626 return Constant::getAllOnesValue(Ty);
627
628 // add nsw/nuw (xor Y, signmask), signmask --> Y
629 // The no-wrapping add guarantees that the top bit will be set by the add.
630 // Therefore, the xor must be clearing the already set sign bit of Y.
631 if ((IsNSW || IsNUW) && match(V: Op1, P: m_SignMask()) &&
632 match(V: Op0, P: m_Xor(L: m_Value(V&: Y), R: m_SignMask())))
633 return Y;
634
635 // add nuw %x, -1 -> -1, because %x can only be 0.
636 if (IsNUW && match(V: Op1, P: m_AllOnes()))
637 return Op1; // Which is -1.
638
639 /// i1 add -> xor.
640 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(BitWidth: 1))
641 if (Value *V = simplifyXorInst(Op0, Op1, Q, MaxRecurse - 1))
642 return V;
643
644 // Try some generic simplifications for associative operations.
645 if (Value *V =
646 simplifyAssociativeBinOp(Opcode: Instruction::Add, LHS: Op0, RHS: Op1, Q, MaxRecurse))
647 return V;
648
649 // Threading Add over selects and phi nodes is pointless, so don't bother.
650 // Threading over the select in "A + select(cond, B, C)" means evaluating
651 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
652 // only if B and C are equal. If B and C are equal then (since we assume
653 // that operands have already been simplified) "select(cond, B, C)" should
654 // have been simplified to the common value of B and C already. Analysing
655 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
656 // for threading over phi nodes.
657
658 return nullptr;
659}
660
661Value *llvm::simplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
662 const SimplifyQuery &Query) {
663 return ::simplifyAddInst(Op0, Op1, IsNSW, IsNUW, Q: Query, MaxRecurse: RecursionLimit);
664}
665
666/// Compute the base pointer and cumulative constant offsets for V.
667///
668/// This strips all constant offsets off of V, leaving it the base pointer, and
669/// accumulates the total constant offset applied in the returned constant.
670/// It returns zero if there are no constant offsets applied.
671///
672/// This is very similar to stripAndAccumulateConstantOffsets(), except it
673/// normalizes the offset bitwidth to the stripped pointer type, not the
674/// original pointer type.
675static APInt stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V) {
676 assert(V->getType()->isPtrOrPtrVectorTy());
677
678 APInt Offset = APInt::getZero(numBits: DL.getIndexTypeSizeInBits(Ty: V->getType()));
679 V = V->stripAndAccumulateConstantOffsets(DL, Offset,
680 /*AllowNonInbounds=*/true);
681 // As that strip may trace through `addrspacecast`, need to sext or trunc
682 // the offset calculated.
683 return Offset.sextOrTrunc(width: DL.getIndexTypeSizeInBits(Ty: V->getType()));
684}
685
686/// Compute the constant difference between two pointer values.
687/// If the difference is not a constant, returns zero.
688static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
689 Value *RHS) {
690 APInt LHSOffset = stripAndComputeConstantOffsets(DL, V&: LHS);
691 APInt RHSOffset = stripAndComputeConstantOffsets(DL, V&: RHS);
692
693 // If LHS and RHS are not related via constant offsets to the same base
694 // value, there is nothing we can do here.
695 if (LHS != RHS)
696 return nullptr;
697
698 // Otherwise, the difference of LHS - RHS can be computed as:
699 // LHS - RHS
700 // = (LHSOffset + Base) - (RHSOffset + Base)
701 // = LHSOffset - RHSOffset
702 Constant *Res = ConstantInt::get(Context&: LHS->getContext(), V: LHSOffset - RHSOffset);
703 if (auto *VecTy = dyn_cast<VectorType>(Val: LHS->getType()))
704 Res = ConstantVector::getSplat(EC: VecTy->getElementCount(), Elt: Res);
705 return Res;
706}
707
708/// Test if there is a dominating equivalence condition for the
709/// two operands. If there is, try to reduce the binary operation
710/// between the two operands.
711/// Example: Op0 - Op1 --> 0 when Op0 == Op1
712static Value *simplifyByDomEq(unsigned Opcode, Value *Op0, Value *Op1,
713 const SimplifyQuery &Q, unsigned MaxRecurse) {
714 // Recursive run it can not get any benefit
715 if (MaxRecurse != RecursionLimit)
716 return nullptr;
717
718 std::optional<bool> Imp =
719 isImpliedByDomCondition(Pred: CmpInst::ICMP_EQ, LHS: Op0, RHS: Op1, ContextI: Q.CxtI, DL: Q.DL);
720 if (Imp && *Imp) {
721 Type *Ty = Op0->getType();
722 switch (Opcode) {
723 case Instruction::Sub:
724 case Instruction::Xor:
725 case Instruction::URem:
726 case Instruction::SRem:
727 return Constant::getNullValue(Ty);
728
729 case Instruction::SDiv:
730 case Instruction::UDiv:
731 return ConstantInt::get(Ty, V: 1);
732
733 case Instruction::And:
734 case Instruction::Or:
735 // Could be either one - choose Op1 since that's more likely a constant.
736 return Op1;
737 default:
738 break;
739 }
740 }
741 return nullptr;
742}
743
744/// Given operands for a Sub, see if we can fold the result.
745/// If not, this returns null.
746static Value *simplifySubInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
747 const SimplifyQuery &Q, unsigned MaxRecurse) {
748 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Sub, Op0, Op1, Q))
749 return C;
750
751 // X - poison -> poison
752 // poison - X -> poison
753 if (isa<PoisonValue>(Val: Op0) || isa<PoisonValue>(Val: Op1))
754 return PoisonValue::get(T: Op0->getType());
755
756 // X - undef -> undef
757 // undef - X -> undef
758 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
759 return UndefValue::get(T: Op0->getType());
760
761 // X - 0 -> X
762 if (match(V: Op1, P: m_Zero()))
763 return Op0;
764
765 // X - X -> 0
766 if (Op0 == Op1)
767 return Constant::getNullValue(Ty: Op0->getType());
768
769 // Is this a negation?
770 if (match(V: Op0, P: m_Zero())) {
771 // 0 - X -> 0 if the sub is NUW.
772 if (IsNUW)
773 return Constant::getNullValue(Ty: Op0->getType());
774
775 KnownBits Known = computeKnownBits(V: Op1, Q);
776 if (Known.Zero.isMaxSignedValue()) {
777 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
778 // Op1 must be 0 because negating the minimum signed value is undefined.
779 if (IsNSW)
780 return Constant::getNullValue(Ty: Op0->getType());
781
782 // 0 - X -> X if X is 0 or the minimum signed value.
783 return Op1;
784 }
785 }
786
787 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
788 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
789 Value *X = nullptr, *Y = nullptr, *Z = Op1;
790 if (MaxRecurse && match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Value(V&: Y)))) { // (X + Y) - Z
791 // See if "V === Y - Z" simplifies.
792 if (Value *V = simplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse - 1))
793 // It does! Now see if "X + V" simplifies.
794 if (Value *W = simplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse - 1)) {
795 // It does, we successfully reassociated!
796 ++NumReassoc;
797 return W;
798 }
799 // See if "V === X - Z" simplifies.
800 if (Value *V = simplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse - 1))
801 // It does! Now see if "Y + V" simplifies.
802 if (Value *W = simplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse - 1)) {
803 // It does, we successfully reassociated!
804 ++NumReassoc;
805 return W;
806 }
807 }
808
809 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
810 // For example, X - (X + 1) -> -1
811 X = Op0;
812 if (MaxRecurse && match(V: Op1, P: m_Add(L: m_Value(V&: Y), R: m_Value(V&: Z)))) { // X - (Y + Z)
813 // See if "V === X - Y" simplifies.
814 if (Value *V = simplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse - 1))
815 // It does! Now see if "V - Z" simplifies.
816 if (Value *W = simplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse - 1)) {
817 // It does, we successfully reassociated!
818 ++NumReassoc;
819 return W;
820 }
821 // See if "V === X - Z" simplifies.
822 if (Value *V = simplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse - 1))
823 // It does! Now see if "V - Y" simplifies.
824 if (Value *W = simplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse - 1)) {
825 // It does, we successfully reassociated!
826 ++NumReassoc;
827 return W;
828 }
829 }
830
831 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
832 // For example, X - (X - Y) -> Y.
833 Z = Op0;
834 if (MaxRecurse && match(V: Op1, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y)))) // Z - (X - Y)
835 // See if "V === Z - X" simplifies.
836 if (Value *V = simplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse - 1))
837 // It does! Now see if "V + Y" simplifies.
838 if (Value *W = simplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse - 1)) {
839 // It does, we successfully reassociated!
840 ++NumReassoc;
841 return W;
842 }
843
844 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
845 if (MaxRecurse && match(V: Op0, P: m_Trunc(Op: m_Value(V&: X))) &&
846 match(V: Op1, P: m_Trunc(Op: m_Value(V&: Y))))
847 if (X->getType() == Y->getType())
848 // See if "V === X - Y" simplifies.
849 if (Value *V = simplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse - 1))
850 // It does! Now see if "trunc V" simplifies.
851 if (Value *W = simplifyCastInst(Instruction::Trunc, V, Op0->getType(),
852 Q, MaxRecurse - 1))
853 // It does, return the simplified "trunc V".
854 return W;
855
856 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
857 if (match(V: Op0, P: m_PtrToIntOrAddr(Op: m_Value(V&: X))) &&
858 match(V: Op1, P: m_PtrToIntOrAddr(Op: m_Value(V&: Y)))) {
859 if (Constant *Result = computePointerDifference(DL: Q.DL, LHS: X, RHS: Y))
860 return ConstantFoldIntegerCast(C: Result, DestTy: Op0->getType(), /*IsSigned*/ true,
861 DL: Q.DL);
862 }
863
864 // i1 sub -> xor.
865 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(BitWidth: 1))
866 if (Value *V = simplifyXorInst(Op0, Op1, Q, MaxRecurse - 1))
867 return V;
868
869 // Threading Sub over selects and phi nodes is pointless, so don't bother.
870 // Threading over the select in "A - select(cond, B, C)" means evaluating
871 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
872 // only if B and C are equal. If B and C are equal then (since we assume
873 // that operands have already been simplified) "select(cond, B, C)" should
874 // have been simplified to the common value of B and C already. Analysing
875 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
876 // for threading over phi nodes.
877
878 if (Value *V = simplifyByDomEq(Opcode: Instruction::Sub, Op0, Op1, Q, MaxRecurse))
879 return V;
880
881 // (sub nuw C_Mask, (xor X, C_Mask)) -> X
882 if (IsNUW) {
883 Value *X;
884 if (match(V: Op1, P: m_Xor(L: m_Value(V&: X), R: m_Specific(V: Op0))) &&
885 match(V: Op0, P: m_LowBitMask()))
886 return X;
887 }
888
889 return nullptr;
890}
891
892Value *llvm::simplifySubInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
893 const SimplifyQuery &Q) {
894 return ::simplifySubInst(Op0, Op1, IsNSW, IsNUW, Q, MaxRecurse: RecursionLimit);
895}
896
897/// Given operands for a Mul, see if we can fold the result.
898/// If not, this returns null.
899static Value *simplifyMulInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
900 const SimplifyQuery &Q, unsigned MaxRecurse) {
901 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Mul, Op0, Op1, Q))
902 return C;
903
904 // X * poison -> poison
905 if (isa<PoisonValue>(Val: Op1))
906 return Op1;
907
908 // X * undef -> 0
909 // X * 0 -> 0
910 if (Q.isUndefValue(V: Op1) || match(V: Op1, P: m_Zero()))
911 return Constant::getNullValue(Ty: Op0->getType());
912
913 // X * 1 -> X
914 if (match(V: Op1, P: m_One()))
915 return Op0;
916
917 // (X / Y) * Y -> X if the division is exact.
918 Value *X = nullptr;
919 if (Q.IIQ.UseInstrInfo &&
920 (match(V: Op0,
921 P: m_Exact(SubPattern: m_IDiv(L: m_Value(V&: X), R: m_Specific(V: Op1)))) || // (X / Y) * Y
922 match(V: Op1, P: m_Exact(SubPattern: m_IDiv(L: m_Value(V&: X), R: m_Specific(V: Op0)))))) // Y * (X / Y)
923 return X;
924
925 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
926 // mul i1 nsw is a special-case because -1 * -1 is poison (+1 is not
927 // representable). All other cases reduce to 0, so just return 0.
928 if (IsNSW)
929 return ConstantInt::getNullValue(Ty: Op0->getType());
930
931 // Treat "mul i1" as "and i1".
932 if (MaxRecurse)
933 if (Value *V = simplifyAndInst(Op0, Op1, Q, MaxRecurse - 1))
934 return V;
935 }
936
937 // Try some generic simplifications for associative operations.
938 if (Value *V =
939 simplifyAssociativeBinOp(Opcode: Instruction::Mul, LHS: Op0, RHS: Op1, Q, MaxRecurse))
940 return V;
941
942 // Mul distributes over Add. Try some generic simplifications based on this.
943 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::Mul, L: Op0, R: Op1,
944 OpcodeToExpand: Instruction::Add, Q, MaxRecurse))
945 return V;
946
947 // If the operation is with the result of a select instruction, check whether
948 // operating on either branch of the select always yields the same value.
949 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1))
950 if (Value *V =
951 threadBinOpOverSelect(Opcode: Instruction::Mul, LHS: Op0, RHS: Op1, Q, MaxRecurse))
952 return V;
953
954 // If the operation is with the result of a phi instruction, check whether
955 // operating on all incoming values of the phi always yields the same value.
956 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
957 if (Value *V =
958 threadBinOpOverPHI(Opcode: Instruction::Mul, LHS: Op0, RHS: Op1, Q, MaxRecurse))
959 return V;
960
961 return nullptr;
962}
963
964Value *llvm::simplifyMulInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
965 const SimplifyQuery &Q) {
966 return ::simplifyMulInst(Op0, Op1, IsNSW, IsNUW, Q, MaxRecurse: RecursionLimit);
967}
968
969/// Given a predicate and two operands, return true if the comparison is true.
970/// This is a helper for div/rem simplification where we return some other value
971/// when we can prove a relationship between the operands.
972static bool isICmpTrue(CmpPredicate Pred, Value *LHS, Value *RHS,
973 const SimplifyQuery &Q, unsigned MaxRecurse) {
974 Value *V = simplifyICmpInst(Predicate: Pred, LHS, RHS, Q, MaxRecurse);
975 Constant *C = dyn_cast_or_null<Constant>(Val: V);
976 return (C && C->isAllOnesValue());
977}
978
979/// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
980/// to simplify X % Y to X.
981static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
982 unsigned MaxRecurse, bool IsSigned) {
983 // Recursion is always used, so bail out at once if we already hit the limit.
984 if (!MaxRecurse--)
985 return false;
986
987 if (IsSigned) {
988 // (X srem Y) sdiv Y --> 0
989 if (match(V: X, P: m_SRem(L: m_Value(), R: m_Specific(V: Y))))
990 return true;
991
992 // |X| / |Y| --> 0
993 //
994 // We require that 1 operand is a simple constant. That could be extended to
995 // 2 variables if we computed the sign bit for each.
996 //
997 // Make sure that a constant is not the minimum signed value because taking
998 // the abs() of that is undefined.
999 Type *Ty = X->getType();
1000 const APInt *C;
1001 if (match(V: X, P: m_APInt(Res&: C)) && !C->isMinSignedValue()) {
1002 // Is the variable divisor magnitude always greater than the constant
1003 // dividend magnitude?
1004 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
1005 Constant *PosDividendC = ConstantInt::get(Ty, V: C->abs());
1006 Constant *NegDividendC = ConstantInt::get(Ty, V: -C->abs());
1007 if (isICmpTrue(Pred: CmpInst::ICMP_SLT, LHS: Y, RHS: NegDividendC, Q, MaxRecurse) ||
1008 isICmpTrue(Pred: CmpInst::ICMP_SGT, LHS: Y, RHS: PosDividendC, Q, MaxRecurse))
1009 return true;
1010 }
1011 if (match(V: Y, P: m_APInt(Res&: C))) {
1012 // Special-case: we can't take the abs() of a minimum signed value. If
1013 // that's the divisor, then all we have to do is prove that the dividend
1014 // is also not the minimum signed value.
1015 if (C->isMinSignedValue())
1016 return isICmpTrue(Pred: CmpInst::ICMP_NE, LHS: X, RHS: Y, Q, MaxRecurse);
1017
1018 // Is the variable dividend magnitude always less than the constant
1019 // divisor magnitude?
1020 // |X| < |C| --> X > -abs(C) and X < abs(C)
1021 Constant *PosDivisorC = ConstantInt::get(Ty, V: C->abs());
1022 Constant *NegDivisorC = ConstantInt::get(Ty, V: -C->abs());
1023 if (isICmpTrue(Pred: CmpInst::ICMP_SGT, LHS: X, RHS: NegDivisorC, Q, MaxRecurse) &&
1024 isICmpTrue(Pred: CmpInst::ICMP_SLT, LHS: X, RHS: PosDivisorC, Q, MaxRecurse))
1025 return true;
1026 }
1027 return false;
1028 }
1029
1030 // IsSigned == false.
1031
1032 // Is the unsigned dividend known to be less than a constant divisor?
1033 // TODO: Convert this (and above) to range analysis
1034 // ("computeConstantRangeIncludingKnownBits")?
1035 const APInt *C;
1036 if (match(V: Y, P: m_APInt(Res&: C)) && computeKnownBits(V: X, Q).getMaxValue().ult(RHS: *C))
1037 return true;
1038
1039 // Try again for any divisor:
1040 // Is the dividend unsigned less than the divisor?
1041 return isICmpTrue(Pred: ICmpInst::ICMP_ULT, LHS: X, RHS: Y, Q, MaxRecurse);
1042}
1043
1044/// Check for common or similar folds of integer division or integer remainder.
1045/// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
1046static Value *simplifyDivRem(Instruction::BinaryOps Opcode, Value *Op0,
1047 Value *Op1, const SimplifyQuery &Q,
1048 unsigned MaxRecurse) {
1049 bool IsDiv = (Opcode == Instruction::SDiv || Opcode == Instruction::UDiv);
1050 bool IsSigned = (Opcode == Instruction::SDiv || Opcode == Instruction::SRem);
1051
1052 Type *Ty = Op0->getType();
1053
1054 // X / undef -> poison
1055 // X % undef -> poison
1056 if (Q.isUndefValue(V: Op1) || isa<PoisonValue>(Val: Op1))
1057 return PoisonValue::get(T: Ty);
1058
1059 // X / 0 -> poison
1060 // X % 0 -> poison
1061 // We don't need to preserve faults!
1062 if (match(V: Op1, P: m_Zero()))
1063 return PoisonValue::get(T: Ty);
1064
1065 // poison / X -> poison
1066 // poison % X -> poison
1067 if (isa<PoisonValue>(Val: Op0))
1068 return Op0;
1069
1070 // undef / X -> 0
1071 // undef % X -> 0
1072 if (Q.isUndefValue(V: Op0))
1073 return Constant::getNullValue(Ty);
1074
1075 // 0 / X -> 0
1076 // 0 % X -> 0
1077 if (match(V: Op0, P: m_Zero()))
1078 return Constant::getNullValue(Ty: Op0->getType());
1079
1080 // X / X -> 1
1081 // X % X -> 0
1082 if (Op0 == Op1)
1083 return IsDiv ? ConstantInt::get(Ty, V: 1) : Constant::getNullValue(Ty);
1084
1085 KnownBits Known = computeKnownBits(V: Op1, Q);
1086 // X / 0 -> poison
1087 // X % 0 -> poison
1088 // If the divisor is known to be zero, just return poison. This can happen in
1089 // some cases where its provable indirectly the denominator is zero but it's
1090 // not trivially simplifiable (i.e known zero through a phi node).
1091 if (Known.isZero())
1092 return PoisonValue::get(T: Ty);
1093
1094 // X / 1 -> X
1095 // X % 1 -> 0
1096 // If the divisor can only be zero or one, we can't have division-by-zero
1097 // or remainder-by-zero, so assume the divisor is 1.
1098 // e.g. 1, zext (i8 X), sdiv X (Y and 1)
1099 if (Known.countMinLeadingZeros() == Known.getBitWidth() - 1)
1100 return IsDiv ? Op0 : Constant::getNullValue(Ty);
1101
1102 // If X * Y does not overflow, then:
1103 // X * Y / Y -> X
1104 // X * Y % Y -> 0
1105 Value *X;
1106 if (match(V: Op0, P: m_c_Mul(L: m_Value(V&: X), R: m_Specific(V: Op1)))) {
1107 auto *Mul = cast<OverflowingBinaryOperator>(Val: Op0);
1108 // The multiplication can't overflow if it is defined not to, or if
1109 // X == A / Y for some A.
1110 if ((IsSigned && Q.IIQ.hasNoSignedWrap(Op: Mul)) ||
1111 (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Op: Mul)) ||
1112 (IsSigned && match(V: X, P: m_SDiv(L: m_Value(), R: m_Specific(V: Op1)))) ||
1113 (!IsSigned && match(V: X, P: m_UDiv(L: m_Value(), R: m_Specific(V: Op1))))) {
1114 return IsDiv ? X : Constant::getNullValue(Ty: Op0->getType());
1115 }
1116 }
1117
1118 if (isDivZero(X: Op0, Y: Op1, Q, MaxRecurse, IsSigned))
1119 return IsDiv ? Constant::getNullValue(Ty: Op0->getType()) : Op0;
1120
1121 if (Value *V = simplifyByDomEq(Opcode, Op0, Op1, Q, MaxRecurse))
1122 return V;
1123
1124 // If the operation is with the result of a select instruction, check whether
1125 // operating on either branch of the select always yields the same value.
1126 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1))
1127 if (Value *V = threadBinOpOverSelect(Opcode, LHS: Op0, RHS: Op1, Q, MaxRecurse))
1128 return V;
1129
1130 // If the operation is with the result of a phi instruction, check whether
1131 // operating on all incoming values of the phi always yields the same value.
1132 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
1133 if (Value *V = threadBinOpOverPHI(Opcode, LHS: Op0, RHS: Op1, Q, MaxRecurse))
1134 return V;
1135
1136 return nullptr;
1137}
1138
1139/// These are simplifications common to SDiv and UDiv.
1140static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1141 bool IsExact, const SimplifyQuery &Q,
1142 unsigned MaxRecurse) {
1143 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1144 return C;
1145
1146 if (Value *V = simplifyDivRem(Opcode, Op0, Op1, Q, MaxRecurse))
1147 return V;
1148
1149 const APInt *DivC;
1150 if (IsExact && match(V: Op1, P: m_APInt(Res&: DivC))) {
1151 // If this is an exact divide by a constant, then the dividend (Op0) must
1152 // have at least as many trailing zeros as the divisor to divide evenly. If
1153 // it has less trailing zeros, then the result must be poison.
1154 if (DivC->countr_zero()) {
1155 KnownBits KnownOp0 = computeKnownBits(V: Op0, Q);
1156 if (KnownOp0.countMaxTrailingZeros() < DivC->countr_zero())
1157 return PoisonValue::get(T: Op0->getType());
1158 }
1159
1160 // udiv exact (mul nsw X, C), C --> X
1161 // sdiv exact (mul nuw X, C), C --> X
1162 // where C is not a power of 2.
1163 Value *X;
1164 if (!DivC->isPowerOf2() &&
1165 (Opcode == Instruction::UDiv
1166 ? match(V: Op0, P: m_NSWMul(L: m_Value(V&: X), R: m_Specific(V: Op1)))
1167 : match(V: Op0, P: m_NUWMul(L: m_Value(V&: X), R: m_Specific(V: Op1)))))
1168 return X;
1169 }
1170
1171 return nullptr;
1172}
1173
1174/// These are simplifications common to SRem and URem.
1175static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1176 const SimplifyQuery &Q, unsigned MaxRecurse) {
1177 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1178 return C;
1179
1180 if (Value *V = simplifyDivRem(Opcode, Op0, Op1, Q, MaxRecurse))
1181 return V;
1182
1183 // (X << Y) % X -> 0
1184 if (Q.IIQ.UseInstrInfo) {
1185 if ((Opcode == Instruction::SRem &&
1186 match(V: Op0, P: m_NSWShl(L: m_Specific(V: Op1), R: m_Value()))) ||
1187 (Opcode == Instruction::URem &&
1188 match(V: Op0, P: m_NUWShl(L: m_Specific(V: Op1), R: m_Value()))))
1189 return Constant::getNullValue(Ty: Op0->getType());
1190
1191 const APInt *C0;
1192 if (match(V: Op1, P: m_APInt(Res&: C0))) {
1193 // (srem (mul nsw X, C1), C0) -> 0 if C1 s% C0 == 0
1194 // (urem (mul nuw X, C1), C0) -> 0 if C1 u% C0 == 0
1195 if (Opcode == Instruction::SRem
1196 ? match(V: Op0,
1197 P: m_NSWMul(L: m_Value(), R: m_CheckedInt(CheckFn: [C0](const APInt &C) {
1198 return C.srem(RHS: *C0).isZero();
1199 })))
1200 : match(V: Op0,
1201 P: m_NUWMul(L: m_Value(), R: m_CheckedInt(CheckFn: [C0](const APInt &C) {
1202 return C.urem(RHS: *C0).isZero();
1203 }))))
1204 return Constant::getNullValue(Ty: Op0->getType());
1205 }
1206 }
1207 return nullptr;
1208}
1209
1210/// Given operands for an SDiv, see if we can fold the result.
1211/// If not, this returns null.
1212static Value *simplifySDivInst(Value *Op0, Value *Op1, bool IsExact,
1213 const SimplifyQuery &Q, unsigned MaxRecurse) {
1214 // If two operands are negated and no signed overflow, return -1.
1215 if (isKnownNegation(X: Op0, Y: Op1, /*NeedNSW=*/true))
1216 return Constant::getAllOnesValue(Ty: Op0->getType());
1217
1218 return simplifyDiv(Opcode: Instruction::SDiv, Op0, Op1, IsExact, Q, MaxRecurse);
1219}
1220
1221Value *llvm::simplifySDivInst(Value *Op0, Value *Op1, bool IsExact,
1222 const SimplifyQuery &Q) {
1223 return ::simplifySDivInst(Op0, Op1, IsExact, Q, MaxRecurse: RecursionLimit);
1224}
1225
1226/// Given operands for a UDiv, see if we can fold the result.
1227/// If not, this returns null.
1228static Value *simplifyUDivInst(Value *Op0, Value *Op1, bool IsExact,
1229 const SimplifyQuery &Q, unsigned MaxRecurse) {
1230 return simplifyDiv(Opcode: Instruction::UDiv, Op0, Op1, IsExact, Q, MaxRecurse);
1231}
1232
1233Value *llvm::simplifyUDivInst(Value *Op0, Value *Op1, bool IsExact,
1234 const SimplifyQuery &Q) {
1235 return ::simplifyUDivInst(Op0, Op1, IsExact, Q, MaxRecurse: RecursionLimit);
1236}
1237
1238/// Given operands for an SRem, see if we can fold the result.
1239/// If not, this returns null.
1240static Value *simplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1241 unsigned MaxRecurse) {
1242 // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1243 // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1244 Value *X;
1245 if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))
1246 return ConstantInt::getNullValue(Ty: Op0->getType());
1247
1248 // If the two operands are negated, return 0.
1249 if (isKnownNegation(X: Op0, Y: Op1))
1250 return ConstantInt::getNullValue(Ty: Op0->getType());
1251
1252 return simplifyRem(Opcode: Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1253}
1254
1255Value *llvm::simplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1256 return ::simplifySRemInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
1257}
1258
1259/// Given operands for a URem, see if we can fold the result.
1260/// If not, this returns null.
1261static Value *simplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1262 unsigned MaxRecurse) {
1263 return simplifyRem(Opcode: Instruction::URem, Op0, Op1, Q, MaxRecurse);
1264}
1265
1266Value *llvm::simplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1267 return ::simplifyURemInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
1268}
1269
1270/// Returns true if a shift by \c Amount always yields poison.
1271static bool isPoisonShift(Value *Amount, const SimplifyQuery &Q) {
1272 Constant *C = dyn_cast<Constant>(Val: Amount);
1273 if (!C)
1274 return false;
1275
1276 // X shift by undef -> poison because it may shift by the bitwidth.
1277 if (Q.isUndefValue(V: C))
1278 return true;
1279
1280 // Shifting by the bitwidth or more is poison. This covers scalars and
1281 // fixed/scalable vectors with splat constants.
1282 const APInt *AmountC;
1283 if (match(V: C, P: m_APInt(Res&: AmountC)) && AmountC->uge(RHS: AmountC->getBitWidth()))
1284 return true;
1285
1286 // Try harder for fixed-length vectors:
1287 // If all lanes of a vector shift are poison, the whole shift is poison.
1288 if (isa<ConstantVector>(Val: C) || isa<ConstantDataVector>(Val: C)) {
1289 for (unsigned I = 0,
1290 E = cast<FixedVectorType>(Val: C->getType())->getNumElements();
1291 I != E; ++I)
1292 if (!isPoisonShift(Amount: C->getAggregateElement(Elt: I), Q))
1293 return false;
1294 return true;
1295 }
1296
1297 return false;
1298}
1299
1300/// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1301/// If not, this returns null.
1302static Value *simplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1303 Value *Op1, bool IsNSW, const SimplifyQuery &Q,
1304 unsigned MaxRecurse) {
1305 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1306 return C;
1307
1308 // poison shift by X -> poison
1309 if (isa<PoisonValue>(Val: Op0))
1310 return Op0;
1311
1312 // 0 shift by X -> 0
1313 if (match(V: Op0, P: m_Zero()))
1314 return Constant::getNullValue(Ty: Op0->getType());
1315
1316 // X shift by 0 -> X
1317 // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1318 // would be poison.
1319 Value *X;
1320 if (match(V: Op1, P: m_Zero()) ||
1321 (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)))
1322 return Op0;
1323
1324 // Fold undefined shifts.
1325 if (isPoisonShift(Amount: Op1, Q))
1326 return PoisonValue::get(T: Op0->getType());
1327
1328 // If the operation is with the result of a select instruction, check whether
1329 // operating on either branch of the select always yields the same value.
1330 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1))
1331 if (Value *V = threadBinOpOverSelect(Opcode, LHS: Op0, RHS: Op1, Q, MaxRecurse))
1332 return V;
1333
1334 // If the operation is with the result of a phi instruction, check whether
1335 // operating on all incoming values of the phi always yields the same value.
1336 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
1337 if (Value *V = threadBinOpOverPHI(Opcode, LHS: Op0, RHS: Op1, Q, MaxRecurse))
1338 return V;
1339
1340 // If any bits in the shift amount make that value greater than or equal to
1341 // the number of bits in the type, the shift is undefined.
1342 KnownBits KnownAmt = computeKnownBits(V: Op1, Q);
1343 if (KnownAmt.getMinValue().uge(RHS: KnownAmt.getBitWidth()))
1344 return PoisonValue::get(T: Op0->getType());
1345
1346 // If all valid bits in the shift amount are known zero, the first operand is
1347 // unchanged.
1348 unsigned NumValidShiftBits = Log2_32_Ceil(Value: KnownAmt.getBitWidth());
1349 if (KnownAmt.countMinTrailingZeros() >= NumValidShiftBits)
1350 return Op0;
1351
1352 // Check for nsw shl leading to a poison value.
1353 if (IsNSW) {
1354 assert(Opcode == Instruction::Shl && "Expected shl for nsw instruction");
1355 KnownBits KnownVal = computeKnownBits(V: Op0, Q);
1356 KnownBits KnownShl = KnownBits::shl(LHS: KnownVal, RHS: KnownAmt);
1357
1358 if (KnownVal.Zero.isSignBitSet())
1359 KnownShl.Zero.setSignBit();
1360 if (KnownVal.One.isSignBitSet())
1361 KnownShl.One.setSignBit();
1362
1363 if (KnownShl.hasConflict())
1364 return PoisonValue::get(T: Op0->getType());
1365 }
1366
1367 return nullptr;
1368}
1369
1370/// Given operands for an LShr or AShr, see if we can fold the result. If not,
1371/// this returns null.
1372static Value *simplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1373 Value *Op1, bool IsExact,
1374 const SimplifyQuery &Q, unsigned MaxRecurse) {
1375 if (Value *V =
1376 simplifyShift(Opcode, Op0, Op1, /*IsNSW*/ false, Q, MaxRecurse))
1377 return V;
1378
1379 // X >> X -> 0
1380 if (Op0 == Op1)
1381 return Constant::getNullValue(Ty: Op0->getType());
1382
1383 // undef >> X -> 0
1384 // undef >> X -> undef (if it's exact)
1385 if (Q.isUndefValue(V: Op0))
1386 return IsExact ? Op0 : Constant::getNullValue(Ty: Op0->getType());
1387
1388 // The low bit cannot be shifted out of an exact shift if it is set.
1389 // TODO: Generalize by counting trailing zeros (see fold for exact division).
1390 if (IsExact) {
1391 KnownBits Op0Known = computeKnownBits(V: Op0, Q);
1392 if (Op0Known.One[0])
1393 return Op0;
1394 }
1395
1396 return nullptr;
1397}
1398
1399/// Given operands for an Shl, see if we can fold the result.
1400/// If not, this returns null.
1401static Value *simplifyShlInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
1402 const SimplifyQuery &Q, unsigned MaxRecurse) {
1403 if (Value *V =
1404 simplifyShift(Opcode: Instruction::Shl, Op0, Op1, IsNSW, Q, MaxRecurse))
1405 return V;
1406
1407 Type *Ty = Op0->getType();
1408 // undef << X -> 0
1409 // undef << X -> undef if (if it's NSW/NUW)
1410 if (Q.isUndefValue(V: Op0))
1411 return IsNSW || IsNUW ? Op0 : Constant::getNullValue(Ty);
1412
1413 // (X >> A) << A -> X
1414 Value *X;
1415 if (Q.IIQ.UseInstrInfo &&
1416 match(V: Op0, P: m_Exact(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1)))))
1417 return X;
1418
1419 // shl nuw i8 C, %x -> C iff C has sign bit set.
1420 if (IsNUW && match(V: Op0, P: m_Negative()))
1421 return Op0;
1422 // NOTE: could use computeKnownBits() / LazyValueInfo,
1423 // but the cost-benefit analysis suggests it isn't worth it.
1424
1425 // "nuw" guarantees that only zeros are shifted out, and "nsw" guarantees
1426 // that the sign-bit does not change, so the only input that does not
1427 // produce poison is 0, and "0 << (bitwidth-1) --> 0".
1428 if (IsNSW && IsNUW &&
1429 match(V: Op1, P: m_SpecificInt(V: Ty->getScalarSizeInBits() - 1)))
1430 return Constant::getNullValue(Ty);
1431
1432 return nullptr;
1433}
1434
1435Value *llvm::simplifyShlInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
1436 const SimplifyQuery &Q) {
1437 return ::simplifyShlInst(Op0, Op1, IsNSW, IsNUW, Q, MaxRecurse: RecursionLimit);
1438}
1439
1440/// Given operands for an LShr, see if we can fold the result.
1441/// If not, this returns null.
1442static Value *simplifyLShrInst(Value *Op0, Value *Op1, bool IsExact,
1443 const SimplifyQuery &Q, unsigned MaxRecurse) {
1444 if (Value *V = simplifyRightShift(Opcode: Instruction::LShr, Op0, Op1, IsExact, Q,
1445 MaxRecurse))
1446 return V;
1447
1448 // (X << A) >> A -> X
1449 Value *X;
1450 if (Q.IIQ.UseInstrInfo && match(V: Op0, P: m_NUWShl(L: m_Value(V&: X), R: m_Specific(V: Op1))))
1451 return X;
1452
1453 // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1454 // We can return X as we do in the above case since OR alters no bits in X.
1455 // SimplifyDemandedBits in InstCombine can do more general optimization for
1456 // bit manipulation. This pattern aims to provide opportunities for other
1457 // optimizers by supporting a simple but common case in InstSimplify.
1458 Value *Y;
1459 const APInt *ShRAmt, *ShLAmt;
1460 if (Q.IIQ.UseInstrInfo && match(V: Op1, P: m_APInt(Res&: ShRAmt)) &&
1461 match(V: Op0, P: m_c_Or(L: m_NUWShl(L: m_Value(V&: X), R: m_APInt(Res&: ShLAmt)), R: m_Value(V&: Y))) &&
1462 *ShRAmt == *ShLAmt) {
1463 const KnownBits YKnown = computeKnownBits(V: Y, Q);
1464 const unsigned EffWidthY = YKnown.countMaxActiveBits();
1465 if (ShRAmt->uge(RHS: EffWidthY))
1466 return X;
1467 }
1468
1469 return nullptr;
1470}
1471
1472Value *llvm::simplifyLShrInst(Value *Op0, Value *Op1, bool IsExact,
1473 const SimplifyQuery &Q) {
1474 return ::simplifyLShrInst(Op0, Op1, IsExact, Q, MaxRecurse: RecursionLimit);
1475}
1476
1477/// Given operands for an AShr, see if we can fold the result.
1478/// If not, this returns null.
1479static Value *simplifyAShrInst(Value *Op0, Value *Op1, bool IsExact,
1480 const SimplifyQuery &Q, unsigned MaxRecurse) {
1481 if (Value *V = simplifyRightShift(Opcode: Instruction::AShr, Op0, Op1, IsExact, Q,
1482 MaxRecurse))
1483 return V;
1484
1485 // -1 >>a X --> -1
1486 // (-1 << X) a>> X --> -1
1487 // We could return the original -1 constant to preserve poison elements.
1488 if (match(V: Op0, P: m_AllOnes()) ||
1489 match(V: Op0, P: m_Shl(L: m_AllOnes(), R: m_Specific(V: Op1))))
1490 return Constant::getAllOnesValue(Ty: Op0->getType());
1491
1492 // (X << A) >> A -> X
1493 Value *X;
1494 if (Q.IIQ.UseInstrInfo && match(V: Op0, P: m_NSWShl(L: m_Value(V&: X), R: m_Specific(V: Op1))))
1495 return X;
1496
1497 // Arithmetic shifting an all-sign-bit value is a no-op.
1498 unsigned NumSignBits = ComputeNumSignBits(Op: Op0, DL: Q.DL, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT);
1499 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1500 return Op0;
1501
1502 return nullptr;
1503}
1504
1505Value *llvm::simplifyAShrInst(Value *Op0, Value *Op1, bool IsExact,
1506 const SimplifyQuery &Q) {
1507 return ::simplifyAShrInst(Op0, Op1, IsExact, Q, MaxRecurse: RecursionLimit);
1508}
1509
1510/// Commuted variants are assumed to be handled by calling this function again
1511/// with the parameters swapped.
1512static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1513 ICmpInst *UnsignedICmp, bool IsAnd,
1514 const SimplifyQuery &Q) {
1515 Value *X, *Y;
1516
1517 CmpPredicate EqPred;
1518 if (!match(V: ZeroICmp, P: m_ICmp(Pred&: EqPred, L: m_Value(V&: Y), R: m_Zero())) ||
1519 !ICmpInst::isEquality(P: EqPred))
1520 return nullptr;
1521
1522 CmpPredicate UnsignedPred;
1523
1524 Value *A, *B;
1525 // Y = (A - B);
1526 if (match(V: Y, P: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B)))) {
1527 if (match(V: UnsignedICmp,
1528 P: m_c_ICmp(Pred&: UnsignedPred, L: m_Specific(V: A), R: m_Specific(V: B))) &&
1529 ICmpInst::isUnsigned(Pred: UnsignedPred)) {
1530 // A >=/<= B || (A - B) != 0 <--> true
1531 if ((UnsignedPred == ICmpInst::ICMP_UGE ||
1532 UnsignedPred == ICmpInst::ICMP_ULE) &&
1533 EqPred == ICmpInst::ICMP_NE && !IsAnd)
1534 return ConstantInt::getTrue(Ty: UnsignedICmp->getType());
1535 // A </> B && (A - B) == 0 <--> false
1536 if ((UnsignedPred == ICmpInst::ICMP_ULT ||
1537 UnsignedPred == ICmpInst::ICMP_UGT) &&
1538 EqPred == ICmpInst::ICMP_EQ && IsAnd)
1539 return ConstantInt::getFalse(Ty: UnsignedICmp->getType());
1540
1541 // A </> B && (A - B) != 0 <--> A </> B
1542 // A </> B || (A - B) != 0 <--> (A - B) != 0
1543 if (EqPred == ICmpInst::ICMP_NE && (UnsignedPred == ICmpInst::ICMP_ULT ||
1544 UnsignedPred == ICmpInst::ICMP_UGT))
1545 return IsAnd ? UnsignedICmp : ZeroICmp;
1546
1547 // A <=/>= B && (A - B) == 0 <--> (A - B) == 0
1548 // A <=/>= B || (A - B) == 0 <--> A <=/>= B
1549 if (EqPred == ICmpInst::ICMP_EQ && (UnsignedPred == ICmpInst::ICMP_ULE ||
1550 UnsignedPred == ICmpInst::ICMP_UGE))
1551 return IsAnd ? ZeroICmp : UnsignedICmp;
1552 }
1553
1554 // Given Y = (A - B)
1555 // Y >= A && Y != 0 --> Y >= A iff B != 0
1556 // Y < A || Y == 0 --> Y < A iff B != 0
1557 if (match(V: UnsignedICmp,
1558 P: m_c_ICmp(Pred&: UnsignedPred, L: m_Specific(V: Y), R: m_Specific(V: A)))) {
1559 if (UnsignedPred == ICmpInst::ICMP_UGE && IsAnd &&
1560 EqPred == ICmpInst::ICMP_NE && isKnownNonZero(V: B, Q))
1561 return UnsignedICmp;
1562 if (UnsignedPred == ICmpInst::ICMP_ULT && !IsAnd &&
1563 EqPred == ICmpInst::ICMP_EQ && isKnownNonZero(V: B, Q))
1564 return UnsignedICmp;
1565 }
1566 }
1567
1568 if (match(V: UnsignedICmp, P: m_ICmp(Pred&: UnsignedPred, L: m_Value(V&: X), R: m_Specific(V: Y))) &&
1569 ICmpInst::isUnsigned(Pred: UnsignedPred))
1570 ;
1571 else if (match(V: UnsignedICmp,
1572 P: m_ICmp(Pred&: UnsignedPred, L: m_Specific(V: Y), R: m_Value(V&: X))) &&
1573 ICmpInst::isUnsigned(Pred: UnsignedPred))
1574 UnsignedPred = ICmpInst::getSwappedPredicate(pred: UnsignedPred);
1575 else
1576 return nullptr;
1577
1578 // X > Y && Y == 0 --> Y == 0 iff X != 0
1579 // X > Y || Y == 0 --> X > Y iff X != 0
1580 if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ &&
1581 isKnownNonZero(V: X, Q))
1582 return IsAnd ? ZeroICmp : UnsignedICmp;
1583
1584 // X <= Y && Y != 0 --> X <= Y iff X != 0
1585 // X <= Y || Y != 0 --> Y != 0 iff X != 0
1586 if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE &&
1587 isKnownNonZero(V: X, Q))
1588 return IsAnd ? UnsignedICmp : ZeroICmp;
1589
1590 // The transforms below here are expected to be handled more generally with
1591 // simplifyAndOrOfICmpsWithLimitConst() or in InstCombine's
1592 // foldAndOrOfICmpsWithConstEq(). If we are looking to trim optimizer overlap,
1593 // these are candidates for removal.
1594
1595 // X < Y && Y != 0 --> X < Y
1596 // X < Y || Y != 0 --> Y != 0
1597 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1598 return IsAnd ? UnsignedICmp : ZeroICmp;
1599
1600 // X >= Y && Y == 0 --> Y == 0
1601 // X >= Y || Y == 0 --> X >= Y
1602 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ)
1603 return IsAnd ? ZeroICmp : UnsignedICmp;
1604
1605 // X < Y && Y == 0 --> false
1606 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1607 IsAnd)
1608 return getFalse(Ty: UnsignedICmp->getType());
1609
1610 // X >= Y || Y != 0 --> true
1611 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_NE &&
1612 !IsAnd)
1613 return getTrue(Ty: UnsignedICmp->getType());
1614
1615 return nullptr;
1616}
1617
1618/// Test if a pair of compares with a shared operand and 2 constants has an
1619/// empty set intersection, full set union, or if one compare is a superset of
1620/// the other.
1621static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1622 bool IsAnd) {
1623 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1624 if (Cmp0->getOperand(i_nocapture: 0) != Cmp1->getOperand(i_nocapture: 0))
1625 return nullptr;
1626
1627 const APInt *C0, *C1;
1628 if (!match(V: Cmp0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C0)) ||
1629 !match(V: Cmp1->getOperand(i_nocapture: 1), P: m_APInt(Res&: C1)))
1630 return nullptr;
1631
1632 auto Range0 = ConstantRange::makeExactICmpRegion(Pred: Cmp0->getPredicate(), Other: *C0);
1633 auto Range1 = ConstantRange::makeExactICmpRegion(Pred: Cmp1->getPredicate(), Other: *C1);
1634
1635 // For and-of-compares, check if the intersection is empty:
1636 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1637 if (IsAnd && Range0.intersectWith(CR: Range1).isEmptySet())
1638 return getFalse(Ty: Cmp0->getType());
1639
1640 // For or-of-compares, check if the union is full:
1641 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1642 if (!IsAnd && Range0.unionWith(CR: Range1).isFullSet())
1643 return getTrue(Ty: Cmp0->getType());
1644
1645 // Is one range a superset of the other?
1646 // If this is and-of-compares, take the smaller set:
1647 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1648 // If this is or-of-compares, take the larger set:
1649 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1650 if (Range0.contains(CR: Range1))
1651 return IsAnd ? Cmp1 : Cmp0;
1652 if (Range1.contains(CR: Range0))
1653 return IsAnd ? Cmp0 : Cmp1;
1654
1655 return nullptr;
1656}
1657
1658static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1659 const InstrInfoQuery &IIQ) {
1660 // (icmp (add V, C0), C1) & (icmp V, C0)
1661 CmpPredicate Pred0, Pred1;
1662 const APInt *C0, *C1;
1663 Value *V;
1664 if (!match(V: Op0, P: m_ICmp(Pred&: Pred0, L: m_Add(L: m_Value(V), R: m_APInt(Res&: C0)), R: m_APInt(Res&: C1))))
1665 return nullptr;
1666
1667 if (!match(V: Op1, P: m_ICmp(Pred&: Pred1, L: m_Specific(V), R: m_Value())))
1668 return nullptr;
1669
1670 auto *AddInst = cast<OverflowingBinaryOperator>(Val: Op0->getOperand(i_nocapture: 0));
1671 if (AddInst->getOperand(i_nocapture: 1) != Op1->getOperand(i_nocapture: 1))
1672 return nullptr;
1673
1674 Type *ITy = Op0->getType();
1675 bool IsNSW = IIQ.hasNoSignedWrap(Op: AddInst);
1676 bool IsNUW = IIQ.hasNoUnsignedWrap(Op: AddInst);
1677
1678 const APInt Delta = *C1 - *C0;
1679 if (C0->isStrictlyPositive()) {
1680 if (Delta == 2) {
1681 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1682 return getFalse(Ty: ITy);
1683 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && IsNSW)
1684 return getFalse(Ty: ITy);
1685 }
1686 if (Delta == 1) {
1687 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1688 return getFalse(Ty: ITy);
1689 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && IsNSW)
1690 return getFalse(Ty: ITy);
1691 }
1692 }
1693 if (C0->getBoolValue() && IsNUW) {
1694 if (Delta == 2)
1695 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1696 return getFalse(Ty: ITy);
1697 if (Delta == 1)
1698 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1699 return getFalse(Ty: ITy);
1700 }
1701
1702 return nullptr;
1703}
1704
1705/// Try to simplify and/or of icmp with ctpop intrinsic.
1706static Value *simplifyAndOrOfICmpsWithCtpop(ICmpInst *Cmp0, ICmpInst *Cmp1,
1707 bool IsAnd) {
1708 CmpPredicate Pred0, Pred1;
1709 Value *X;
1710 const APInt *C;
1711 if (!match(V: Cmp0, P: m_ICmp(Pred&: Pred0, L: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Value(V&: X)),
1712 R: m_APInt(Res&: C))) ||
1713 !match(V: Cmp1, P: m_ICmp(Pred&: Pred1, L: m_Specific(V: X), R: m_ZeroInt())) || C->isZero())
1714 return nullptr;
1715
1716 // (ctpop(X) == C) || (X != 0) --> X != 0 where C > 0
1717 if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_NE)
1718 return Cmp1;
1719 // (ctpop(X) != C) && (X == 0) --> X == 0 where C > 0
1720 if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_EQ)
1721 return Cmp1;
1722
1723 return nullptr;
1724}
1725
1726static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1727 const SimplifyQuery &Q) {
1728 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op0, UnsignedICmp: Op1, /*IsAnd=*/true, Q))
1729 return X;
1730 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op1, UnsignedICmp: Op0, /*IsAnd=*/true, Q))
1731 return X;
1732
1733 if (Value *X = simplifyAndOrOfICmpsWithConstants(Cmp0: Op0, Cmp1: Op1, IsAnd: true))
1734 return X;
1735
1736 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op0, Cmp1: Op1, IsAnd: true))
1737 return X;
1738 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op1, Cmp1: Op0, IsAnd: true))
1739 return X;
1740
1741 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ: Q.IIQ))
1742 return X;
1743 if (Value *X = simplifyAndOfICmpsWithAdd(Op0: Op1, Op1: Op0, IIQ: Q.IIQ))
1744 return X;
1745
1746 return nullptr;
1747}
1748
1749static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1750 const InstrInfoQuery &IIQ) {
1751 // (icmp (add V, C0), C1) | (icmp V, C0)
1752 CmpPredicate Pred0, Pred1;
1753 const APInt *C0, *C1;
1754 Value *V;
1755 if (!match(V: Op0, P: m_ICmp(Pred&: Pred0, L: m_Add(L: m_Value(V), R: m_APInt(Res&: C0)), R: m_APInt(Res&: C1))))
1756 return nullptr;
1757
1758 if (!match(V: Op1, P: m_ICmp(Pred&: Pred1, L: m_Specific(V), R: m_Value())))
1759 return nullptr;
1760
1761 auto *AddInst = cast<BinaryOperator>(Val: Op0->getOperand(i_nocapture: 0));
1762 if (AddInst->getOperand(i_nocapture: 1) != Op1->getOperand(i_nocapture: 1))
1763 return nullptr;
1764
1765 Type *ITy = Op0->getType();
1766 bool IsNSW = IIQ.hasNoSignedWrap(Op: AddInst);
1767 bool IsNUW = IIQ.hasNoUnsignedWrap(Op: AddInst);
1768
1769 const APInt Delta = *C1 - *C0;
1770 if (C0->isStrictlyPositive()) {
1771 if (Delta == 2) {
1772 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1773 return getTrue(Ty: ITy);
1774 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && IsNSW)
1775 return getTrue(Ty: ITy);
1776 }
1777 if (Delta == 1) {
1778 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1779 return getTrue(Ty: ITy);
1780 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && IsNSW)
1781 return getTrue(Ty: ITy);
1782 }
1783 }
1784 if (C0->getBoolValue() && IsNUW) {
1785 if (Delta == 2)
1786 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1787 return getTrue(Ty: ITy);
1788 if (Delta == 1)
1789 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1790 return getTrue(Ty: ITy);
1791 }
1792
1793 return nullptr;
1794}
1795
1796static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1797 const SimplifyQuery &Q) {
1798 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op0, UnsignedICmp: Op1, /*IsAnd=*/false, Q))
1799 return X;
1800 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op1, UnsignedICmp: Op0, /*IsAnd=*/false, Q))
1801 return X;
1802
1803 if (Value *X = simplifyAndOrOfICmpsWithConstants(Cmp0: Op0, Cmp1: Op1, IsAnd: false))
1804 return X;
1805
1806 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op0, Cmp1: Op1, IsAnd: false))
1807 return X;
1808 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op1, Cmp1: Op0, IsAnd: false))
1809 return X;
1810
1811 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ: Q.IIQ))
1812 return X;
1813 if (Value *X = simplifyOrOfICmpsWithAdd(Op0: Op1, Op1: Op0, IIQ: Q.IIQ))
1814 return X;
1815
1816 return nullptr;
1817}
1818
1819/// Test if a pair of compares with a shared operand and 2 constants has an
1820/// empty set intersection, full set union, or if one compare is a superset of
1821/// the other.
1822static Value *simplifyAndOrOfFCmpsWithConstants(FCmpInst *Cmp0, FCmpInst *Cmp1,
1823 bool IsAnd) {
1824 // Look for this pattern: {and/or} (fcmp X, C0), (fcmp X, C1)).
1825 if (Cmp0->getOperand(i_nocapture: 0) != Cmp1->getOperand(i_nocapture: 0))
1826 return nullptr;
1827
1828 const APFloat *C0, *C1;
1829 if (!match(V: Cmp0->getOperand(i_nocapture: 1), P: m_APFloat(Res&: C0)) ||
1830 !match(V: Cmp1->getOperand(i_nocapture: 1), P: m_APFloat(Res&: C1)))
1831 return nullptr;
1832
1833 auto Range0 = ConstantFPRange::makeExactFCmpRegion(
1834 Pred: IsAnd ? Cmp0->getPredicate() : Cmp0->getInversePredicate(), Other: *C0);
1835 auto Range1 = ConstantFPRange::makeExactFCmpRegion(
1836 Pred: IsAnd ? Cmp1->getPredicate() : Cmp1->getInversePredicate(), Other: *C1);
1837
1838 if (!Range0 || !Range1)
1839 return nullptr;
1840
1841 // For and-of-compares, check if the intersection is empty:
1842 // (fcmp X, C0) && (fcmp X, C1) --> empty set --> false
1843 if (Range0->intersectWith(CR: *Range1).isEmptySet())
1844 return ConstantInt::getBool(Ty: Cmp0->getType(), V: !IsAnd);
1845
1846 // Is one range a superset of the other?
1847 // If this is and-of-compares, take the smaller set:
1848 // (fcmp ogt X, 4) && (fcmp ogt X, 42) --> fcmp ogt X, 42
1849 // If this is or-of-compares, take the larger set:
1850 // (fcmp ogt X, 4) || (fcmp ogt X, 42) --> fcmp ogt X, 4
1851 if (Range0->contains(CR: *Range1))
1852 return Cmp1;
1853 if (Range1->contains(CR: *Range0))
1854 return Cmp0;
1855
1856 return nullptr;
1857}
1858
1859static Value *simplifyAndOrOfFCmps(const SimplifyQuery &Q, FCmpInst *LHS,
1860 FCmpInst *RHS, bool IsAnd) {
1861 Value *LHS0 = LHS->getOperand(i_nocapture: 0), *LHS1 = LHS->getOperand(i_nocapture: 1);
1862 Value *RHS0 = RHS->getOperand(i_nocapture: 0), *RHS1 = RHS->getOperand(i_nocapture: 1);
1863 if (LHS0->getType() != RHS0->getType())
1864 return nullptr;
1865
1866 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1867 auto AbsOrSelfLHS0 = m_CombineOr(L: m_Specific(V: LHS0), R: m_FAbs(Op0: m_Specific(V: LHS0)));
1868 if ((PredL == FCmpInst::FCMP_ORD || PredL == FCmpInst::FCMP_UNO) &&
1869 ((FCmpInst::isOrdered(predicate: PredR) && IsAnd) ||
1870 (FCmpInst::isUnordered(predicate: PredR) && !IsAnd))) {
1871 // (fcmp ord X, 0) & (fcmp o** X/abs(X), Y) --> fcmp o** X/abs(X), Y
1872 // (fcmp uno X, 0) & (fcmp o** X/abs(X), Y) --> false
1873 // (fcmp uno X, 0) | (fcmp u** X/abs(X), Y) --> fcmp u** X/abs(X), Y
1874 // (fcmp ord X, 0) | (fcmp u** X/abs(X), Y) --> true
1875 if ((match(V: RHS0, P: AbsOrSelfLHS0) || match(V: RHS1, P: AbsOrSelfLHS0)) &&
1876 match(V: LHS1, P: m_PosZeroFP()))
1877 return FCmpInst::isOrdered(predicate: PredL) == FCmpInst::isOrdered(predicate: PredR)
1878 ? static_cast<Value *>(RHS)
1879 : ConstantInt::getBool(Ty: LHS->getType(), V: !IsAnd);
1880 }
1881
1882 auto AbsOrSelfRHS0 = m_CombineOr(L: m_Specific(V: RHS0), R: m_FAbs(Op0: m_Specific(V: RHS0)));
1883 if ((PredR == FCmpInst::FCMP_ORD || PredR == FCmpInst::FCMP_UNO) &&
1884 ((FCmpInst::isOrdered(predicate: PredL) && IsAnd) ||
1885 (FCmpInst::isUnordered(predicate: PredL) && !IsAnd))) {
1886 // (fcmp o** X/abs(X), Y) & (fcmp ord X, 0) --> fcmp o** X/abs(X), Y
1887 // (fcmp o** X/abs(X), Y) & (fcmp uno X, 0) --> false
1888 // (fcmp u** X/abs(X), Y) | (fcmp uno X, 0) --> fcmp u** X/abs(X), Y
1889 // (fcmp u** X/abs(X), Y) | (fcmp ord X, 0) --> true
1890 if ((match(V: LHS0, P: AbsOrSelfRHS0) || match(V: LHS1, P: AbsOrSelfRHS0)) &&
1891 match(V: RHS1, P: m_PosZeroFP()))
1892 return FCmpInst::isOrdered(predicate: PredL) == FCmpInst::isOrdered(predicate: PredR)
1893 ? static_cast<Value *>(LHS)
1894 : ConstantInt::getBool(Ty: LHS->getType(), V: !IsAnd);
1895 }
1896
1897 if (auto *V = simplifyAndOrOfFCmpsWithConstants(Cmp0: LHS, Cmp1: RHS, IsAnd))
1898 return V;
1899
1900 return nullptr;
1901}
1902
1903static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q, Value *Op0,
1904 Value *Op1, bool IsAnd) {
1905 // Look through casts of the 'and' operands to find compares.
1906 auto *Cast0 = dyn_cast<CastInst>(Val: Op0);
1907 auto *Cast1 = dyn_cast<CastInst>(Val: Op1);
1908 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1909 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1910 Op0 = Cast0->getOperand(i_nocapture: 0);
1911 Op1 = Cast1->getOperand(i_nocapture: 0);
1912 }
1913
1914 Value *V = nullptr;
1915 auto *ICmp0 = dyn_cast<ICmpInst>(Val: Op0);
1916 auto *ICmp1 = dyn_cast<ICmpInst>(Val: Op1);
1917 if (ICmp0 && ICmp1)
1918 V = IsAnd ? simplifyAndOfICmps(Op0: ICmp0, Op1: ICmp1, Q)
1919 : simplifyOrOfICmps(Op0: ICmp0, Op1: ICmp1, Q);
1920
1921 auto *FCmp0 = dyn_cast<FCmpInst>(Val: Op0);
1922 auto *FCmp1 = dyn_cast<FCmpInst>(Val: Op1);
1923 if (FCmp0 && FCmp1)
1924 V = simplifyAndOrOfFCmps(Q, LHS: FCmp0, RHS: FCmp1, IsAnd);
1925
1926 if (!V)
1927 return nullptr;
1928 if (!Cast0)
1929 return V;
1930
1931 // If we looked through casts, we can only handle a constant simplification
1932 // because we are not allowed to create a cast instruction here.
1933 if (auto *C = dyn_cast<Constant>(Val: V))
1934 return ConstantFoldCastOperand(Opcode: Cast0->getOpcode(), C, DestTy: Cast0->getType(),
1935 DL: Q.DL);
1936
1937 return nullptr;
1938}
1939
1940static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
1941 const SimplifyQuery &Q,
1942 bool AllowRefinement,
1943 SmallVectorImpl<Instruction *> *DropFlags,
1944 unsigned MaxRecurse);
1945
1946static Value *simplifyAndOrWithICmpEq(unsigned Opcode, Value *Op0, Value *Op1,
1947 const SimplifyQuery &Q,
1948 unsigned MaxRecurse) {
1949 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1950 "Must be and/or");
1951 CmpPredicate Pred;
1952 Value *A, *B;
1953 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
1954 match(V: Op0, P: m_NUWTrunc(Op: m_Value(V&: A)))) {
1955 B = ConstantInt::getNullValue(Ty: A->getType());
1956 Pred = ICmpInst::ICMP_NE;
1957 } else if (!match(V: Op0, P: m_ICmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B))) ||
1958 !ICmpInst::isEquality(P: Pred))
1959 return nullptr;
1960
1961 auto Simplify = [&](Value *Res) -> Value * {
1962 Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, Ty: Res->getType());
1963
1964 // and (icmp eq a, b), x implies (a==b) inside x.
1965 // or (icmp ne a, b), x implies (a==b) inside x.
1966 // If x simplifies to true/false, we can simplify the and/or.
1967 if (Pred ==
1968 (Opcode == Instruction::And ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
1969 if (Res == Absorber)
1970 return Absorber;
1971 if (Res == ConstantExpr::getBinOpIdentity(Opcode, Ty: Res->getType()))
1972 return Op0;
1973 return nullptr;
1974 }
1975
1976 // If we have and (icmp ne a, b), x and for a==b we can simplify x to false,
1977 // then we can drop the icmp, as x will already be false in the case where
1978 // the icmp is false. Similar for or and true.
1979 if (Res == Absorber)
1980 return Op1;
1981 return nullptr;
1982 };
1983
1984 // In the final case (Res == Absorber with inverted predicate), it is safe to
1985 // refine poison during simplification, but not undef. For simplicity always
1986 // disable undef-based folds here.
1987 if (Value *Res = simplifyWithOpReplaced(V: Op1, Op: A, RepOp: B, Q: Q.getWithoutUndef(),
1988 /* AllowRefinement */ true,
1989 /* DropFlags */ nullptr, MaxRecurse))
1990 return Simplify(Res);
1991 if (Value *Res = simplifyWithOpReplaced(V: Op1, Op: B, RepOp: A, Q: Q.getWithoutUndef(),
1992 /* AllowRefinement */ true,
1993 /* DropFlags */ nullptr, MaxRecurse))
1994 return Simplify(Res);
1995
1996 return nullptr;
1997}
1998
1999/// Given a bitwise logic op, check if the operands are add/sub with a common
2000/// source value and inverted constant (identity: C - X -> ~(X + ~C)).
2001static Value *simplifyLogicOfAddSub(Value *Op0, Value *Op1,
2002 Instruction::BinaryOps Opcode) {
2003 assert(Op0->getType() == Op1->getType() && "Mismatched binop types");
2004 assert(BinaryOperator::isBitwiseLogicOp(Opcode) && "Expected logic op");
2005 Value *X;
2006 Constant *C1, *C2;
2007 if ((match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Constant(C&: C1))) &&
2008 match(V: Op1, P: m_Sub(L: m_Constant(C&: C2), R: m_Specific(V: X)))) ||
2009 (match(V: Op1, P: m_Add(L: m_Value(V&: X), R: m_Constant(C&: C1))) &&
2010 match(V: Op0, P: m_Sub(L: m_Constant(C&: C2), R: m_Specific(V: X))))) {
2011 if (ConstantExpr::getNot(C: C1) == C2) {
2012 // (X + C) & (~C - X) --> (X + C) & ~(X + C) --> 0
2013 // (X + C) | (~C - X) --> (X + C) | ~(X + C) --> -1
2014 // (X + C) ^ (~C - X) --> (X + C) ^ ~(X + C) --> -1
2015 Type *Ty = Op0->getType();
2016 return Opcode == Instruction::And ? ConstantInt::getNullValue(Ty)
2017 : ConstantInt::getAllOnesValue(Ty);
2018 }
2019 }
2020 return nullptr;
2021}
2022
2023// Commutative patterns for and that will be tried with both operand orders.
2024static Value *simplifyAndCommutative(Value *Op0, Value *Op1,
2025 const SimplifyQuery &Q,
2026 unsigned MaxRecurse) {
2027 // ~A & A = 0
2028 if (match(V: Op0, P: m_Not(V: m_Specific(V: Op1))))
2029 return Constant::getNullValue(Ty: Op0->getType());
2030
2031 // (A | ?) & A = A
2032 if (match(V: Op0, P: m_c_Or(L: m_Specific(V: Op1), R: m_Value())))
2033 return Op1;
2034
2035 // (X | ~Y) & (X | Y) --> X
2036 Value *X, *Y;
2037 if (match(V: Op0, P: m_c_Or(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
2038 match(V: Op1, P: m_c_Or(L: m_Specific(V: X), R: m_Specific(V: Y))))
2039 return X;
2040
2041 // If we have a multiplication overflow check that is being 'and'ed with a
2042 // check that one of the multipliers is not zero, we can omit the 'and', and
2043 // only keep the overflow check.
2044 if (isCheckForZeroAndMulWithOverflow(Op0, Op1, IsAnd: true))
2045 return Op1;
2046
2047 // -A & A = A if A is a power of two or zero.
2048 if (match(V: Op0, P: m_Neg(V: m_Specific(V: Op1))) &&
2049 isKnownToBeAPowerOfTwo(V: Op1, DL: Q.DL, /*OrZero*/ true, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
2050 return Op1;
2051
2052 // This is a similar pattern used for checking if a value is a power-of-2:
2053 // (A - 1) & A --> 0 (if A is a power-of-2 or 0)
2054 if (match(V: Op0, P: m_Add(L: m_Specific(V: Op1), R: m_AllOnes())) &&
2055 isKnownToBeAPowerOfTwo(V: Op1, DL: Q.DL, /*OrZero*/ true, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
2056 return Constant::getNullValue(Ty: Op1->getType());
2057
2058 // (x << N) & ((x << M) - 1) --> 0, where x is known to be a power of 2 and
2059 // M <= N.
2060 const APInt *Shift1, *Shift2;
2061 if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: Shift1))) &&
2062 match(V: Op1, P: m_Add(L: m_Shl(L: m_Specific(V: X), R: m_APInt(Res&: Shift2)), R: m_AllOnes())) &&
2063 isKnownToBeAPowerOfTwo(V: X, DL: Q.DL, /*OrZero*/ true, AC: Q.AC, CxtI: Q.CxtI) &&
2064 Shift1->uge(RHS: *Shift2))
2065 return Constant::getNullValue(Ty: Op0->getType());
2066
2067 if (Value *V =
2068 simplifyAndOrWithICmpEq(Opcode: Instruction::And, Op0, Op1, Q, MaxRecurse))
2069 return V;
2070
2071 return nullptr;
2072}
2073
2074/// Given operands for an And, see if we can fold the result.
2075/// If not, this returns null.
2076static Value *simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2077 unsigned MaxRecurse) {
2078 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::And, Op0, Op1, Q))
2079 return C;
2080
2081 // X & poison -> poison
2082 if (isa<PoisonValue>(Val: Op1))
2083 return Op1;
2084
2085 // X & undef -> 0
2086 if (Q.isUndefValue(V: Op1))
2087 return Constant::getNullValue(Ty: Op0->getType());
2088
2089 // X & X = X
2090 if (Op0 == Op1)
2091 return Op0;
2092
2093 // X & 0 = 0
2094 if (match(V: Op1, P: m_Zero()))
2095 return Constant::getNullValue(Ty: Op0->getType());
2096
2097 // X & -1 = X
2098 if (match(V: Op1, P: m_AllOnes()))
2099 return Op0;
2100
2101 if (Value *Res = simplifyAndCommutative(Op0, Op1, Q, MaxRecurse))
2102 return Res;
2103 if (Value *Res = simplifyAndCommutative(Op0: Op1, Op1: Op0, Q, MaxRecurse))
2104 return Res;
2105
2106 if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Opcode: Instruction::And))
2107 return V;
2108
2109 // A mask that only clears known zeros of a shifted value is a no-op.
2110 const APInt *Mask;
2111 const APInt *ShAmt;
2112 Value *X, *Y;
2113 if (match(V: Op1, P: m_APInt(Res&: Mask))) {
2114 // If all bits in the inverted and shifted mask are clear:
2115 // and (shl X, ShAmt), Mask --> shl X, ShAmt
2116 if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: ShAmt))) &&
2117 (~(*Mask)).lshr(ShiftAmt: *ShAmt).isZero())
2118 return Op0;
2119
2120 // If all bits in the inverted and shifted mask are clear:
2121 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
2122 if (match(V: Op0, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: ShAmt))) &&
2123 (~(*Mask)).shl(ShiftAmt: *ShAmt).isZero())
2124 return Op0;
2125 }
2126
2127 // and 2^x-1, 2^C --> 0 where x <= C.
2128 const APInt *PowerC;
2129 Value *Shift;
2130 if (match(V: Op1, P: m_Power2(V&: PowerC)) &&
2131 match(V: Op0, P: m_Add(L: m_Value(V&: Shift), R: m_AllOnes())) &&
2132 isKnownToBeAPowerOfTwo(V: Shift, DL: Q.DL, /*OrZero*/ false, AC: Q.AC, CxtI: Q.CxtI,
2133 DT: Q.DT)) {
2134 KnownBits Known = computeKnownBits(V: Shift, Q);
2135 // Use getActiveBits() to make use of the additional power of two knowledge
2136 if (PowerC->getActiveBits() >= Known.getMaxValue().getActiveBits())
2137 return ConstantInt::getNullValue(Ty: Op1->getType());
2138 }
2139
2140 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, IsAnd: true))
2141 return V;
2142
2143 // Try some generic simplifications for associative operations.
2144 if (Value *V =
2145 simplifyAssociativeBinOp(Opcode: Instruction::And, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2146 return V;
2147
2148 // And distributes over Or. Try some generic simplifications based on this.
2149 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::And, L: Op0, R: Op1,
2150 OpcodeToExpand: Instruction::Or, Q, MaxRecurse))
2151 return V;
2152
2153 // And distributes over Xor. Try some generic simplifications based on this.
2154 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::And, L: Op0, R: Op1,
2155 OpcodeToExpand: Instruction::Xor, Q, MaxRecurse))
2156 return V;
2157
2158 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1)) {
2159 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2160 // A & (A && B) -> A && B
2161 if (match(V: Op1, P: m_Select(C: m_Specific(V: Op0), L: m_Value(), R: m_Zero())))
2162 return Op1;
2163 else if (match(V: Op0, P: m_Select(C: m_Specific(V: Op1), L: m_Value(), R: m_Zero())))
2164 return Op0;
2165 }
2166 // If the operation is with the result of a select instruction, check
2167 // whether operating on either branch of the select always yields the same
2168 // value.
2169 if (Value *V =
2170 threadBinOpOverSelect(Opcode: Instruction::And, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2171 return V;
2172 }
2173
2174 // If the operation is with the result of a phi instruction, check whether
2175 // operating on all incoming values of the phi always yields the same value.
2176 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
2177 if (Value *V =
2178 threadBinOpOverPHI(Opcode: Instruction::And, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2179 return V;
2180
2181 // Assuming the effective width of Y is not larger than A, i.e. all bits
2182 // from X and Y are disjoint in (X << A) | Y,
2183 // if the mask of this AND op covers all bits of X or Y, while it covers
2184 // no bits from the other, we can bypass this AND op. E.g.,
2185 // ((X << A) | Y) & Mask -> Y,
2186 // if Mask = ((1 << effective_width_of(Y)) - 1)
2187 // ((X << A) | Y) & Mask -> X << A,
2188 // if Mask = ((1 << effective_width_of(X)) - 1) << A
2189 // SimplifyDemandedBits in InstCombine can optimize the general case.
2190 // This pattern aims to help other passes for a common case.
2191 Value *XShifted;
2192 if (Q.IIQ.UseInstrInfo && match(V: Op1, P: m_APInt(Res&: Mask)) &&
2193 match(V: Op0, P: m_c_Or(L: m_CombineAnd(L: m_NUWShl(L: m_Value(V&: X), R: m_APInt(Res&: ShAmt)),
2194 R: m_Value(V&: XShifted)),
2195 R: m_Value(V&: Y)))) {
2196 const unsigned Width = Op0->getType()->getScalarSizeInBits();
2197 const unsigned ShftCnt = ShAmt->getLimitedValue(Limit: Width);
2198 const KnownBits YKnown = computeKnownBits(V: Y, Q);
2199 const unsigned EffWidthY = YKnown.countMaxActiveBits();
2200 if (EffWidthY <= ShftCnt) {
2201 const KnownBits XKnown = computeKnownBits(V: X, Q);
2202 const unsigned EffWidthX = XKnown.countMaxActiveBits();
2203 const APInt EffBitsY = APInt::getLowBitsSet(numBits: Width, loBitsSet: EffWidthY);
2204 const APInt EffBitsX = APInt::getLowBitsSet(numBits: Width, loBitsSet: EffWidthX) << ShftCnt;
2205 // If the mask is extracting all bits from X or Y as is, we can skip
2206 // this AND op.
2207 if (EffBitsY.isSubsetOf(RHS: *Mask) && !EffBitsX.intersects(RHS: *Mask))
2208 return Y;
2209 if (EffBitsX.isSubsetOf(RHS: *Mask) && !EffBitsY.intersects(RHS: *Mask))
2210 return XShifted;
2211 }
2212 }
2213
2214 // ((X | Y) ^ X ) & ((X | Y) ^ Y) --> 0
2215 // ((X | Y) ^ Y ) & ((X | Y) ^ X) --> 0
2216 BinaryOperator *Or;
2217 if (match(V: Op0, P: m_c_Xor(L: m_Value(V&: X),
2218 R: m_CombineAnd(L: m_BinOp(I&: Or),
2219 R: m_c_Or(L: m_Deferred(V: X), R: m_Value(V&: Y))))) &&
2220 match(V: Op1, P: m_c_Xor(L: m_Specific(V: Or), R: m_Specific(V: Y))))
2221 return Constant::getNullValue(Ty: Op0->getType());
2222
2223 const APInt *C1;
2224 Value *A;
2225 // (A ^ C) & (A ^ ~C) -> 0
2226 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
2227 match(V: Op1, P: m_Xor(L: m_Specific(V: A), R: m_SpecificInt(V: ~*C1))))
2228 return Constant::getNullValue(Ty: Op0->getType());
2229
2230 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2231 if (std::optional<bool> Implied = isImpliedCondition(LHS: Op0, RHS: Op1, DL: Q.DL)) {
2232 // If Op0 is true implies Op1 is true, then Op0 is a subset of Op1.
2233 if (*Implied == true)
2234 return Op0;
2235 // If Op0 is true implies Op1 is false, then they are not true together.
2236 if (*Implied == false)
2237 return ConstantInt::getFalse(Ty: Op0->getType());
2238 }
2239 if (std::optional<bool> Implied = isImpliedCondition(LHS: Op1, RHS: Op0, DL: Q.DL)) {
2240 // If Op1 is true implies Op0 is true, then Op1 is a subset of Op0.
2241 if (*Implied)
2242 return Op1;
2243 // If Op1 is true implies Op0 is false, then they are not true together.
2244 if (!*Implied)
2245 return ConstantInt::getFalse(Ty: Op1->getType());
2246 }
2247 }
2248
2249 if (Value *V = simplifyByDomEq(Opcode: Instruction::And, Op0, Op1, Q, MaxRecurse))
2250 return V;
2251
2252 return nullptr;
2253}
2254
2255Value *llvm::simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2256 return ::simplifyAndInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
2257}
2258
2259// TODO: Many of these folds could use LogicalAnd/LogicalOr.
2260static Value *simplifyOrLogic(Value *X, Value *Y) {
2261 assert(X->getType() == Y->getType() && "Expected same type for 'or' ops");
2262 Type *Ty = X->getType();
2263
2264 // X | ~X --> -1
2265 if (match(V: Y, P: m_Not(V: m_Specific(V: X))))
2266 return ConstantInt::getAllOnesValue(Ty);
2267
2268 // X | ~(X & ?) = -1
2269 if (match(V: Y, P: m_Not(V: m_c_And(L: m_Specific(V: X), R: m_Value()))))
2270 return ConstantInt::getAllOnesValue(Ty);
2271
2272 // X | (X & ?) --> X
2273 if (match(V: Y, P: m_c_And(L: m_Specific(V: X), R: m_Value())))
2274 return X;
2275
2276 Value *A, *B;
2277
2278 // (A ^ B) | (A | B) --> A | B
2279 // (A ^ B) | (B | A) --> B | A
2280 if (match(V: X, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2281 match(V: Y, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2282 return Y;
2283
2284 // ~(A ^ B) | (A | B) --> -1
2285 // ~(A ^ B) | (B | A) --> -1
2286 if (match(V: X, P: m_Not(V: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)))) &&
2287 match(V: Y, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2288 return ConstantInt::getAllOnesValue(Ty);
2289
2290 // (A & ~B) | (A ^ B) --> A ^ B
2291 // (~B & A) | (A ^ B) --> A ^ B
2292 // (A & ~B) | (B ^ A) --> B ^ A
2293 // (~B & A) | (B ^ A) --> B ^ A
2294 if (match(V: X, P: m_c_And(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B)))) &&
2295 match(V: Y, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2296 return Y;
2297
2298 // (~A ^ B) | (A & B) --> ~A ^ B
2299 // (B ^ ~A) | (A & B) --> B ^ ~A
2300 // (~A ^ B) | (B & A) --> ~A ^ B
2301 // (B ^ ~A) | (B & A) --> B ^ ~A
2302 if (match(V: X, P: m_c_Xor(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2303 match(V: Y, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2304 return X;
2305
2306 // (~A | B) | (A ^ B) --> -1
2307 // (~A | B) | (B ^ A) --> -1
2308 // (B | ~A) | (A ^ B) --> -1
2309 // (B | ~A) | (B ^ A) --> -1
2310 if (match(V: X, P: m_c_Or(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2311 match(V: Y, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2312 return ConstantInt::getAllOnesValue(Ty);
2313
2314 // (~A & B) | ~(A | B) --> ~A
2315 // (~A & B) | ~(B | A) --> ~A
2316 // (B & ~A) | ~(A | B) --> ~A
2317 // (B & ~A) | ~(B | A) --> ~A
2318 Value *NotA;
2319 if (match(V: X, P: m_c_And(L: m_CombineAnd(L: m_Value(V&: NotA), R: m_Not(V: m_Value(V&: A))),
2320 R: m_Value(V&: B))) &&
2321 match(V: Y, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B)))))
2322 return NotA;
2323 // The same is true of Logical And
2324 // TODO: This could share the logic of the version above if there was a
2325 // version of LogicalAnd that allowed more than just i1 types.
2326 if (match(V: X, P: m_c_LogicalAnd(L: m_CombineAnd(L: m_Value(V&: NotA), R: m_Not(V: m_Value(V&: A))),
2327 R: m_Value(V&: B))) &&
2328 match(V: Y, P: m_Not(V: m_c_LogicalOr(L: m_Specific(V: A), R: m_Specific(V: B)))))
2329 return NotA;
2330
2331 // ~(A ^ B) | (A & B) --> ~(A ^ B)
2332 // ~(A ^ B) | (B & A) --> ~(A ^ B)
2333 Value *NotAB;
2334 if (match(V: X, P: m_CombineAnd(L: m_Not(V: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))),
2335 R: m_Value(V&: NotAB))) &&
2336 match(V: Y, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2337 return NotAB;
2338
2339 // ~(A & B) | (A ^ B) --> ~(A & B)
2340 // ~(A & B) | (B ^ A) --> ~(A & B)
2341 if (match(V: X, P: m_CombineAnd(L: m_Not(V: m_And(L: m_Value(V&: A), R: m_Value(V&: B))),
2342 R: m_Value(V&: NotAB))) &&
2343 match(V: Y, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2344 return NotAB;
2345
2346 return nullptr;
2347}
2348
2349/// Given operands for an Or, see if we can fold the result.
2350/// If not, this returns null.
2351static Value *simplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2352 unsigned MaxRecurse) {
2353 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Or, Op0, Op1, Q))
2354 return C;
2355
2356 // X | poison -> poison
2357 if (isa<PoisonValue>(Val: Op1))
2358 return Op1;
2359
2360 // X | undef -> -1
2361 // X | -1 = -1
2362 // Do not return Op1 because it may contain undef elements if it's a vector.
2363 if (Q.isUndefValue(V: Op1) || match(V: Op1, P: m_AllOnes()))
2364 return Constant::getAllOnesValue(Ty: Op0->getType());
2365
2366 // X | X = X
2367 // X | 0 = X
2368 if (Op0 == Op1 || match(V: Op1, P: m_Zero()))
2369 return Op0;
2370
2371 if (Value *R = simplifyOrLogic(X: Op0, Y: Op1))
2372 return R;
2373 if (Value *R = simplifyOrLogic(X: Op1, Y: Op0))
2374 return R;
2375
2376 if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Opcode: Instruction::Or))
2377 return V;
2378
2379 // Rotated -1 is still -1:
2380 // (-1 << X) | (-1 >> (C - X)) --> -1
2381 // (-1 >> X) | (-1 << (C - X)) --> -1
2382 // ...with C <= bitwidth (and commuted variants).
2383 Value *X, *Y;
2384 if ((match(V: Op0, P: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))) &&
2385 match(V: Op1, P: m_LShr(L: m_AllOnes(), R: m_Value(V&: Y)))) ||
2386 (match(V: Op1, P: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))) &&
2387 match(V: Op0, P: m_LShr(L: m_AllOnes(), R: m_Value(V&: Y))))) {
2388 const APInt *C;
2389 if ((match(V: X, P: m_Sub(L: m_APInt(Res&: C), R: m_Specific(V: Y))) ||
2390 match(V: Y, P: m_Sub(L: m_APInt(Res&: C), R: m_Specific(V: X)))) &&
2391 C->ule(RHS: X->getType()->getScalarSizeInBits())) {
2392 return ConstantInt::getAllOnesValue(Ty: X->getType());
2393 }
2394 }
2395
2396 // A funnel shift (rotate) can be decomposed into simpler shifts. See if we
2397 // are mixing in another shift that is redundant with the funnel shift.
2398
2399 // (fshl X, ?, Y) | (shl X, Y) --> fshl X, ?, Y
2400 // (shl X, Y) | (fshl X, ?, Y) --> fshl X, ?, Y
2401 if (match(V: Op0,
2402 P: m_Intrinsic<Intrinsic::fshl>(Op0: m_Value(V&: X), Op1: m_Value(), Op2: m_Value(V&: Y))) &&
2403 match(V: Op1, P: m_Shl(L: m_Specific(V: X), R: m_Specific(V: Y))))
2404 return Op0;
2405 if (match(V: Op1,
2406 P: m_Intrinsic<Intrinsic::fshl>(Op0: m_Value(V&: X), Op1: m_Value(), Op2: m_Value(V&: Y))) &&
2407 match(V: Op0, P: m_Shl(L: m_Specific(V: X), R: m_Specific(V: Y))))
2408 return Op1;
2409
2410 // (fshr ?, X, Y) | (lshr X, Y) --> fshr ?, X, Y
2411 // (lshr X, Y) | (fshr ?, X, Y) --> fshr ?, X, Y
2412 if (match(V: Op0,
2413 P: m_Intrinsic<Intrinsic::fshr>(Op0: m_Value(), Op1: m_Value(V&: X), Op2: m_Value(V&: Y))) &&
2414 match(V: Op1, P: m_LShr(L: m_Specific(V: X), R: m_Specific(V: Y))))
2415 return Op0;
2416 if (match(V: Op1,
2417 P: m_Intrinsic<Intrinsic::fshr>(Op0: m_Value(), Op1: m_Value(V&: X), Op2: m_Value(V&: Y))) &&
2418 match(V: Op0, P: m_LShr(L: m_Specific(V: X), R: m_Specific(V: Y))))
2419 return Op1;
2420
2421 if (Value *V =
2422 simplifyAndOrWithICmpEq(Opcode: Instruction::Or, Op0, Op1, Q, MaxRecurse))
2423 return V;
2424 if (Value *V =
2425 simplifyAndOrWithICmpEq(Opcode: Instruction::Or, Op0: Op1, Op1: Op0, Q, MaxRecurse))
2426 return V;
2427
2428 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, IsAnd: false))
2429 return V;
2430
2431 // If we have a multiplication overflow check that is being 'and'ed with a
2432 // check that one of the multipliers is not zero, we can omit the 'and', and
2433 // only keep the overflow check.
2434 if (isCheckForZeroAndMulWithOverflow(Op0, Op1, IsAnd: false))
2435 return Op1;
2436 if (isCheckForZeroAndMulWithOverflow(Op0: Op1, Op1: Op0, IsAnd: false))
2437 return Op0;
2438
2439 // Try some generic simplifications for associative operations.
2440 if (Value *V =
2441 simplifyAssociativeBinOp(Opcode: Instruction::Or, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2442 return V;
2443
2444 // Or distributes over And. Try some generic simplifications based on this.
2445 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::Or, L: Op0, R: Op1,
2446 OpcodeToExpand: Instruction::And, Q, MaxRecurse))
2447 return V;
2448
2449 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1)) {
2450 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2451 // A | (A || B) -> A || B
2452 if (match(V: Op1, P: m_Select(C: m_Specific(V: Op0), L: m_One(), R: m_Value())))
2453 return Op1;
2454 else if (match(V: Op0, P: m_Select(C: m_Specific(V: Op1), L: m_One(), R: m_Value())))
2455 return Op0;
2456 }
2457 // If the operation is with the result of a select instruction, check
2458 // whether operating on either branch of the select always yields the same
2459 // value.
2460 if (Value *V =
2461 threadBinOpOverSelect(Opcode: Instruction::Or, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2462 return V;
2463 }
2464
2465 // (A & C1)|(B & C2)
2466 Value *A, *B;
2467 const APInt *C1, *C2;
2468 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
2469 match(V: Op1, P: m_And(L: m_Value(V&: B), R: m_APInt(Res&: C2)))) {
2470 if (*C1 == ~*C2) {
2471 // (A & C1)|(B & C2)
2472 // If we have: ((V + N) & C1) | (V & C2)
2473 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2474 // replace with V+N.
2475 Value *N;
2476 if (C2->isMask() && // C2 == 0+1+
2477 match(V: A, P: m_c_Add(L: m_Specific(V: B), R: m_Value(V&: N)))) {
2478 // Add commutes, try both ways.
2479 if (MaskedValueIsZero(V: N, Mask: *C2, SQ: Q))
2480 return A;
2481 }
2482 // Or commutes, try both ways.
2483 if (C1->isMask() && match(V: B, P: m_c_Add(L: m_Specific(V: A), R: m_Value(V&: N)))) {
2484 // Add commutes, try both ways.
2485 if (MaskedValueIsZero(V: N, Mask: *C1, SQ: Q))
2486 return B;
2487 }
2488 }
2489 }
2490
2491 // If the operation is with the result of a phi instruction, check whether
2492 // operating on all incoming values of the phi always yields the same value.
2493 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
2494 if (Value *V = threadBinOpOverPHI(Opcode: Instruction::Or, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2495 return V;
2496
2497 // (A ^ C) | (A ^ ~C) -> -1, i.e. all bits set to one.
2498 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
2499 match(V: Op1, P: m_Xor(L: m_Specific(V: A), R: m_SpecificInt(V: ~*C1))))
2500 return Constant::getAllOnesValue(Ty: Op0->getType());
2501
2502 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2503 if (std::optional<bool> Implied =
2504 isImpliedCondition(LHS: Op0, RHS: Op1, DL: Q.DL, LHSIsTrue: false)) {
2505 // If Op0 is false implies Op1 is false, then Op1 is a subset of Op0.
2506 if (*Implied == false)
2507 return Op0;
2508 // If Op0 is false implies Op1 is true, then at least one is always true.
2509 if (*Implied == true)
2510 return ConstantInt::getTrue(Ty: Op0->getType());
2511 }
2512 if (std::optional<bool> Implied =
2513 isImpliedCondition(LHS: Op1, RHS: Op0, DL: Q.DL, LHSIsTrue: false)) {
2514 // If Op1 is false implies Op0 is false, then Op0 is a subset of Op1.
2515 if (*Implied == false)
2516 return Op1;
2517 // If Op1 is false implies Op0 is true, then at least one is always true.
2518 if (*Implied == true)
2519 return ConstantInt::getTrue(Ty: Op1->getType());
2520 }
2521 }
2522
2523 if (Value *V = simplifyByDomEq(Opcode: Instruction::Or, Op0, Op1, Q, MaxRecurse))
2524 return V;
2525
2526 return nullptr;
2527}
2528
2529Value *llvm::simplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2530 return ::simplifyOrInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
2531}
2532
2533/// Given operands for a Xor, see if we can fold the result.
2534/// If not, this returns null.
2535static Value *simplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2536 unsigned MaxRecurse) {
2537 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Xor, Op0, Op1, Q))
2538 return C;
2539
2540 // X ^ poison -> poison
2541 if (isa<PoisonValue>(Val: Op1))
2542 return Op1;
2543
2544 // A ^ undef -> undef
2545 if (Q.isUndefValue(V: Op1))
2546 return Op1;
2547
2548 // A ^ 0 = A
2549 if (match(V: Op1, P: m_Zero()))
2550 return Op0;
2551
2552 // A ^ A = 0
2553 if (Op0 == Op1)
2554 return Constant::getNullValue(Ty: Op0->getType());
2555
2556 // A ^ ~A = ~A ^ A = -1
2557 if (match(V: Op0, P: m_Not(V: m_Specific(V: Op1))) || match(V: Op1, P: m_Not(V: m_Specific(V: Op0))))
2558 return Constant::getAllOnesValue(Ty: Op0->getType());
2559
2560 auto foldAndOrNot = [](Value *X, Value *Y) -> Value * {
2561 Value *A, *B;
2562 // (~A & B) ^ (A | B) --> A -- There are 8 commuted variants.
2563 if (match(V: X, P: m_c_And(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2564 match(V: Y, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2565 return A;
2566
2567 // (~A | B) ^ (A & B) --> ~A -- There are 8 commuted variants.
2568 // The 'not' op must contain a complete -1 operand (no undef elements for
2569 // vector) for the transform to be safe.
2570 Value *NotA;
2571 if (match(V: X, P: m_c_Or(L: m_CombineAnd(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: NotA)),
2572 R: m_Value(V&: B))) &&
2573 match(V: Y, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2574 return NotA;
2575
2576 return nullptr;
2577 };
2578 if (Value *R = foldAndOrNot(Op0, Op1))
2579 return R;
2580 if (Value *R = foldAndOrNot(Op1, Op0))
2581 return R;
2582
2583 if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Opcode: Instruction::Xor))
2584 return V;
2585
2586 // Try some generic simplifications for associative operations.
2587 if (Value *V =
2588 simplifyAssociativeBinOp(Opcode: Instruction::Xor, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2589 return V;
2590
2591 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2592 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2593 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2594 // only if B and C are equal. If B and C are equal then (since we assume
2595 // that operands have already been simplified) "select(cond, B, C)" should
2596 // have been simplified to the common value of B and C already. Analysing
2597 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2598 // for threading over phi nodes.
2599
2600 if (Value *V = simplifyByDomEq(Opcode: Instruction::Xor, Op0, Op1, Q, MaxRecurse))
2601 return V;
2602
2603 // (xor (sub nuw C_Mask, X), C_Mask) -> X
2604 {
2605 Value *X;
2606 if (match(V: Op0, P: m_NUWSub(L: m_Specific(V: Op1), R: m_Value(V&: X))) &&
2607 match(V: Op1, P: m_LowBitMask()))
2608 return X;
2609 }
2610
2611 return nullptr;
2612}
2613
2614Value *llvm::simplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2615 return ::simplifyXorInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
2616}
2617
2618static Type *getCompareTy(Value *Op) {
2619 return CmpInst::makeCmpResultType(opnd_type: Op->getType());
2620}
2621
2622/// Rummage around inside V looking for something equivalent to the comparison
2623/// "LHS Pred RHS". Return such a value if found, otherwise return null.
2624/// Helper function for analyzing max/min idioms.
2625static Value *extractEquivalentCondition(Value *V, CmpPredicate Pred,
2626 Value *LHS, Value *RHS) {
2627 SelectInst *SI = dyn_cast<SelectInst>(Val: V);
2628 if (!SI)
2629 return nullptr;
2630 CmpInst *Cmp = dyn_cast<CmpInst>(Val: SI->getCondition());
2631 if (!Cmp)
2632 return nullptr;
2633 Value *CmpLHS = Cmp->getOperand(i_nocapture: 0), *CmpRHS = Cmp->getOperand(i_nocapture: 1);
2634 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2635 return Cmp;
2636 if (Pred == CmpInst::getSwappedPredicate(pred: Cmp->getPredicate()) &&
2637 LHS == CmpRHS && RHS == CmpLHS)
2638 return Cmp;
2639 return nullptr;
2640}
2641
2642/// Return true if the underlying object (storage) must be disjoint from
2643/// storage returned by any noalias return call.
2644static bool isAllocDisjoint(const Value *V) {
2645 // For allocas, we consider only static ones (dynamic
2646 // allocas might be transformed into calls to malloc not simultaneously
2647 // live with the compared-to allocation). For globals, we exclude symbols
2648 // that might be resolve lazily to symbols in another dynamically-loaded
2649 // library (and, thus, could be malloc'ed by the implementation).
2650 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Val: V))
2651 return AI->isStaticAlloca();
2652 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V))
2653 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2654 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2655 !GV->isThreadLocal();
2656 if (const Argument *A = dyn_cast<Argument>(Val: V))
2657 return A->hasByValAttr();
2658 return false;
2659}
2660
2661/// Return true if V1 and V2 are each the base of some distict storage region
2662/// [V, object_size(V)] which do not overlap. Note that zero sized regions
2663/// *are* possible, and that zero sized regions do not overlap with any other.
2664static bool haveNonOverlappingStorage(const Value *V1, const Value *V2) {
2665 // Global variables always exist, so they always exist during the lifetime
2666 // of each other and all allocas. Global variables themselves usually have
2667 // non-overlapping storage, but since their addresses are constants, the
2668 // case involving two globals does not reach here and is instead handled in
2669 // constant folding.
2670 //
2671 // Two different allocas usually have different addresses...
2672 //
2673 // However, if there's an @llvm.stackrestore dynamically in between two
2674 // allocas, they may have the same address. It's tempting to reduce the
2675 // scope of the problem by only looking at *static* allocas here. That would
2676 // cover the majority of allocas while significantly reducing the likelihood
2677 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2678 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2679 // an entry block. Also, if we have a block that's not attached to a
2680 // function, we can't tell if it's "static" under the current definition.
2681 // Theoretically, this problem could be fixed by creating a new kind of
2682 // instruction kind specifically for static allocas. Such a new instruction
2683 // could be required to be at the top of the entry block, thus preventing it
2684 // from being subject to a @llvm.stackrestore. Instcombine could even
2685 // convert regular allocas into these special allocas. It'd be nifty.
2686 // However, until then, this problem remains open.
2687 //
2688 // So, we'll assume that two non-empty allocas have different addresses
2689 // for now.
2690 auto isByValArg = [](const Value *V) {
2691 const Argument *A = dyn_cast<Argument>(Val: V);
2692 return A && A->hasByValAttr();
2693 };
2694
2695 // Byval args are backed by store which does not overlap with each other,
2696 // allocas, or globals.
2697 if (isByValArg(V1))
2698 return isa<AllocaInst>(Val: V2) || isa<GlobalVariable>(Val: V2) || isByValArg(V2);
2699 if (isByValArg(V2))
2700 return isa<AllocaInst>(Val: V1) || isa<GlobalVariable>(Val: V1) || isByValArg(V1);
2701
2702 return isa<AllocaInst>(Val: V1) &&
2703 (isa<AllocaInst>(Val: V2) || isa<GlobalVariable>(Val: V2));
2704}
2705
2706// A significant optimization not implemented here is assuming that alloca
2707// addresses are not equal to incoming argument values. They don't *alias*,
2708// as we say, but that doesn't mean they aren't equal, so we take a
2709// conservative approach.
2710//
2711// This is inspired in part by C++11 5.10p1:
2712// "Two pointers of the same type compare equal if and only if they are both
2713// null, both point to the same function, or both represent the same
2714// address."
2715//
2716// This is pretty permissive.
2717//
2718// It's also partly due to C11 6.5.9p6:
2719// "Two pointers compare equal if and only if both are null pointers, both are
2720// pointers to the same object (including a pointer to an object and a
2721// subobject at its beginning) or function, both are pointers to one past the
2722// last element of the same array object, or one is a pointer to one past the
2723// end of one array object and the other is a pointer to the start of a
2724// different array object that happens to immediately follow the first array
2725// object in the address space.)
2726//
2727// C11's version is more restrictive, however there's no reason why an argument
2728// couldn't be a one-past-the-end value for a stack object in the caller and be
2729// equal to the beginning of a stack object in the callee.
2730//
2731// If the C and C++ standards are ever made sufficiently restrictive in this
2732// area, it may be possible to update LLVM's semantics accordingly and reinstate
2733// this optimization.
2734static Constant *computePointerICmp(CmpPredicate Pred, Value *LHS, Value *RHS,
2735 const SimplifyQuery &Q) {
2736 assert(LHS->getType() == RHS->getType() && "Must have same types");
2737 const DataLayout &DL = Q.DL;
2738 const TargetLibraryInfo *TLI = Q.TLI;
2739
2740 // We fold equality and unsigned predicates on pointer comparisons, but forbid
2741 // signed predicates since a GEP with inbounds could cross the sign boundary.
2742 if (CmpInst::isSigned(Pred))
2743 return nullptr;
2744
2745 // We have to switch to a signed predicate to handle negative indices from
2746 // the base pointer.
2747 Pred = ICmpInst::getSignedPredicate(Pred);
2748
2749 // Strip off any constant offsets so that we can reason about them.
2750 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2751 // here and compare base addresses like AliasAnalysis does, however there are
2752 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2753 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2754 // doesn't need to guarantee pointer inequality when it says NoAlias.
2755
2756 // Even if an non-inbounds GEP occurs along the path we can still optimize
2757 // equality comparisons concerning the result.
2758 bool AllowNonInbounds = ICmpInst::isEquality(P: Pred);
2759 unsigned IndexSize = DL.getIndexTypeSizeInBits(Ty: LHS->getType());
2760 APInt LHSOffset(IndexSize, 0), RHSOffset(IndexSize, 0);
2761 LHS = LHS->stripAndAccumulateConstantOffsets(DL, Offset&: LHSOffset, AllowNonInbounds);
2762 RHS = RHS->stripAndAccumulateConstantOffsets(DL, Offset&: RHSOffset, AllowNonInbounds);
2763
2764 // If LHS and RHS are related via constant offsets to the same base
2765 // value, we can replace it with an icmp which just compares the offsets.
2766 if (LHS == RHS)
2767 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2768 V: ICmpInst::compare(LHS: LHSOffset, RHS: RHSOffset, Pred));
2769
2770 // Various optimizations for (in)equality comparisons.
2771 if (ICmpInst::isEquality(P: Pred)) {
2772 // Different non-empty allocations that exist at the same time have
2773 // different addresses (if the program can tell). If the offsets are
2774 // within the bounds of their allocations (and not one-past-the-end!
2775 // so we can't use inbounds!), and their allocations aren't the same,
2776 // the pointers are not equal.
2777 if (haveNonOverlappingStorage(V1: LHS, V2: RHS)) {
2778 uint64_t LHSSize, RHSSize;
2779 ObjectSizeOpts Opts;
2780 Opts.EvalMode = ObjectSizeOpts::Mode::Min;
2781 auto *F = [](Value *V) -> Function * {
2782 if (auto *I = dyn_cast<Instruction>(Val: V))
2783 return I->getFunction();
2784 if (auto *A = dyn_cast<Argument>(Val: V))
2785 return A->getParent();
2786 return nullptr;
2787 }(LHS);
2788 Opts.NullIsUnknownSize = F ? NullPointerIsDefined(F) : true;
2789 if (getObjectSize(Ptr: LHS, Size&: LHSSize, DL, TLI, Opts) && LHSSize != 0 &&
2790 getObjectSize(Ptr: RHS, Size&: RHSSize, DL, TLI, Opts) && RHSSize != 0) {
2791 APInt Dist = LHSOffset - RHSOffset;
2792 if (Dist.isNonNegative() ? Dist.ult(RHS: LHSSize) : (-Dist).ult(RHS: RHSSize))
2793 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2794 V: !CmpInst::isTrueWhenEqual(predicate: Pred));
2795 }
2796 }
2797
2798 // If one side of the equality comparison must come from a noalias call
2799 // (meaning a system memory allocation function), and the other side must
2800 // come from a pointer that cannot overlap with dynamically-allocated
2801 // memory within the lifetime of the current function (allocas, byval
2802 // arguments, globals), then determine the comparison result here.
2803 SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2804 getUnderlyingObjects(V: LHS, Objects&: LHSUObjs);
2805 getUnderlyingObjects(V: RHS, Objects&: RHSUObjs);
2806
2807 // Is the set of underlying objects all noalias calls?
2808 auto IsNAC = [](ArrayRef<const Value *> Objects) {
2809 return all_of(Range&: Objects, P: isNoAliasCall);
2810 };
2811
2812 // Is the set of underlying objects all things which must be disjoint from
2813 // noalias calls. We assume that indexing from such disjoint storage
2814 // into the heap is undefined, and thus offsets can be safely ignored.
2815 auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2816 return all_of(Range&: Objects, P: ::isAllocDisjoint);
2817 };
2818
2819 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2820 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2821 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2822 V: !CmpInst::isTrueWhenEqual(predicate: Pred));
2823
2824 // Fold comparisons for non-escaping pointer even if the allocation call
2825 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2826 // dynamic allocation call could be either of the operands. Note that
2827 // the other operand can not be based on the alloc - if it were, then
2828 // the cmp itself would be a capture.
2829 Value *MI = nullptr;
2830 if (isAllocLikeFn(V: LHS, TLI) && llvm::isKnownNonZero(V: RHS, Q))
2831 MI = LHS;
2832 else if (isAllocLikeFn(V: RHS, TLI) && llvm::isKnownNonZero(V: LHS, Q))
2833 MI = RHS;
2834 if (MI) {
2835 // FIXME: This is incorrect, see PR54002. While we can assume that the
2836 // allocation is at an address that makes the comparison false, this
2837 // requires that *all* comparisons to that address be false, which
2838 // InstSimplify cannot guarantee.
2839 struct CustomCaptureTracker : public CaptureTracker {
2840 bool Captured = false;
2841 void tooManyUses() override { Captured = true; }
2842 Action captured(const Use *U, UseCaptureInfo CI) override {
2843 // TODO(captures): Use UseCaptureInfo.
2844 if (auto *ICmp = dyn_cast<ICmpInst>(Val: U->getUser())) {
2845 // Comparison against value stored in global variable. Given the
2846 // pointer does not escape, its value cannot be guessed and stored
2847 // separately in a global variable.
2848 unsigned OtherIdx = 1 - U->getOperandNo();
2849 auto *LI = dyn_cast<LoadInst>(Val: ICmp->getOperand(i_nocapture: OtherIdx));
2850 if (LI && isa<GlobalVariable>(Val: LI->getPointerOperand()))
2851 return Continue;
2852 }
2853
2854 Captured = true;
2855 return Stop;
2856 }
2857 };
2858 CustomCaptureTracker Tracker;
2859 PointerMayBeCaptured(V: MI, Tracker: &Tracker);
2860 if (!Tracker.Captured)
2861 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2862 V: CmpInst::isFalseWhenEqual(predicate: Pred));
2863 }
2864 }
2865
2866 // Otherwise, fail.
2867 return nullptr;
2868}
2869
2870/// Fold an icmp when its operands have i1 scalar type.
2871static Value *simplifyICmpOfBools(CmpPredicate Pred, Value *LHS, Value *RHS,
2872 const SimplifyQuery &Q) {
2873 Type *ITy = getCompareTy(Op: LHS); // The return type.
2874 Type *OpTy = LHS->getType(); // The operand type.
2875 if (!OpTy->isIntOrIntVectorTy(BitWidth: 1))
2876 return nullptr;
2877
2878 // A boolean compared to true/false can be reduced in 14 out of the 20
2879 // (10 predicates * 2 constants) possible combinations. The other
2880 // 6 cases require a 'not' of the LHS.
2881
2882 auto ExtractNotLHS = [](Value *V) -> Value * {
2883 Value *X;
2884 if (match(V, P: m_Not(V: m_Value(V&: X))))
2885 return X;
2886 return nullptr;
2887 };
2888
2889 if (match(V: RHS, P: m_Zero())) {
2890 switch (Pred) {
2891 case CmpInst::ICMP_NE: // X != 0 -> X
2892 case CmpInst::ICMP_UGT: // X >u 0 -> X
2893 case CmpInst::ICMP_SLT: // X <s 0 -> X
2894 return LHS;
2895
2896 case CmpInst::ICMP_EQ: // not(X) == 0 -> X != 0 -> X
2897 case CmpInst::ICMP_ULE: // not(X) <=u 0 -> X >u 0 -> X
2898 case CmpInst::ICMP_SGE: // not(X) >=s 0 -> X <s 0 -> X
2899 if (Value *X = ExtractNotLHS(LHS))
2900 return X;
2901 break;
2902
2903 case CmpInst::ICMP_ULT: // X <u 0 -> false
2904 case CmpInst::ICMP_SGT: // X >s 0 -> false
2905 return getFalse(Ty: ITy);
2906
2907 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2908 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2909 return getTrue(Ty: ITy);
2910
2911 default:
2912 break;
2913 }
2914 } else if (match(V: RHS, P: m_One())) {
2915 switch (Pred) {
2916 case CmpInst::ICMP_EQ: // X == 1 -> X
2917 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2918 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2919 return LHS;
2920
2921 case CmpInst::ICMP_NE: // not(X) != 1 -> X == 1 -> X
2922 case CmpInst::ICMP_ULT: // not(X) <=u 1 -> X >=u 1 -> X
2923 case CmpInst::ICMP_SGT: // not(X) >s 1 -> X <=s -1 -> X
2924 if (Value *X = ExtractNotLHS(LHS))
2925 return X;
2926 break;
2927
2928 case CmpInst::ICMP_UGT: // X >u 1 -> false
2929 case CmpInst::ICMP_SLT: // X <s -1 -> false
2930 return getFalse(Ty: ITy);
2931
2932 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2933 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2934 return getTrue(Ty: ITy);
2935
2936 default:
2937 break;
2938 }
2939 }
2940
2941 switch (Pred) {
2942 default:
2943 break;
2944 case ICmpInst::ICMP_UGE:
2945 if (isImpliedCondition(LHS: RHS, RHS: LHS, DL: Q.DL).value_or(u: false))
2946 return getTrue(Ty: ITy);
2947 break;
2948 case ICmpInst::ICMP_SGE:
2949 /// For signed comparison, the values for an i1 are 0 and -1
2950 /// respectively. This maps into a truth table of:
2951 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2952 /// 0 | 0 | 1 (0 >= 0) | 1
2953 /// 0 | 1 | 1 (0 >= -1) | 1
2954 /// 1 | 0 | 0 (-1 >= 0) | 0
2955 /// 1 | 1 | 1 (-1 >= -1) | 1
2956 if (isImpliedCondition(LHS, RHS, DL: Q.DL).value_or(u: false))
2957 return getTrue(Ty: ITy);
2958 break;
2959 case ICmpInst::ICMP_ULE:
2960 if (isImpliedCondition(LHS, RHS, DL: Q.DL).value_or(u: false))
2961 return getTrue(Ty: ITy);
2962 break;
2963 case ICmpInst::ICMP_SLE:
2964 /// SLE follows the same logic as SGE with the LHS and RHS swapped.
2965 if (isImpliedCondition(LHS: RHS, RHS: LHS, DL: Q.DL).value_or(u: false))
2966 return getTrue(Ty: ITy);
2967 break;
2968 }
2969
2970 return nullptr;
2971}
2972
2973/// Check if RHS is zero or can be transformed to an equivalent zero comparison.
2974/// E.g., icmp sgt X, -1 --> icmp sge X, 0
2975static bool matchEquivZeroRHS(CmpPredicate &Pred, const Value *RHS) {
2976 // icmp [pred] X, 0 --> as-is
2977 if (match(V: RHS, P: m_Zero()))
2978 return true;
2979
2980 // Handle comparisons with -1 (all ones)
2981 if (match(V: RHS, P: m_AllOnes())) {
2982 switch (Pred) {
2983 case ICmpInst::ICMP_SGT:
2984 // icmp sgt X, -1 --> icmp sge X, 0
2985 Pred = ICmpInst::ICMP_SGE;
2986 return true;
2987 case ICmpInst::ICMP_SLE:
2988 // icmp sle X, -1 --> icmp slt X, 0
2989 Pred = ICmpInst::ICMP_SLT;
2990 return true;
2991 // Note: unsigned comparisons with -1 (UINT_MAX) are not handled here:
2992 // - icmp ugt X, -1 is always false (nothing > UINT_MAX)
2993 // - icmp ule X, -1 is always true (everything <= UINT_MAX)
2994 default:
2995 return false;
2996 }
2997 }
2998
2999 // Handle comparisons with 1
3000 if (match(V: RHS, P: m_One())) {
3001 switch (Pred) {
3002 case ICmpInst::ICMP_SGE:
3003 // icmp sge X, 1 --> icmp sgt X, 0
3004 Pred = ICmpInst::ICMP_SGT;
3005 return true;
3006 case ICmpInst::ICMP_UGE:
3007 // icmp uge X, 1 --> icmp ugt X, 0
3008 Pred = ICmpInst::ICMP_UGT;
3009 return true;
3010 case ICmpInst::ICMP_SLT:
3011 // icmp slt X, 1 --> icmp sle X, 0
3012 Pred = ICmpInst::ICMP_SLE;
3013 return true;
3014 case ICmpInst::ICMP_ULT:
3015 // icmp ult X, 1 --> icmp ule X, 0
3016 Pred = ICmpInst::ICMP_ULE;
3017 return true;
3018 default:
3019 return false;
3020 }
3021 }
3022
3023 return false;
3024}
3025
3026/// Try hard to fold icmp with zero RHS because this is a common case.
3027/// Note that, this function also handles the equivalent zero RHS, e.g.,
3028/// icmp sgt X, -1 --> icmp sge X, 0
3029static Value *simplifyICmpWithZero(CmpPredicate Pred, Value *LHS, Value *RHS,
3030 const SimplifyQuery &Q) {
3031 // Check if RHS is zero or can be transformed to an equivalent zero comparison
3032 if (!matchEquivZeroRHS(Pred, RHS))
3033 return nullptr;
3034
3035 Type *ITy = getCompareTy(Op: LHS); // The return type.
3036 switch (Pred) {
3037 default:
3038 llvm_unreachable("Unknown ICmp predicate!");
3039 case ICmpInst::ICMP_ULT:
3040 return getFalse(Ty: ITy);
3041 case ICmpInst::ICMP_UGE:
3042 return getTrue(Ty: ITy);
3043 case ICmpInst::ICMP_EQ:
3044 case ICmpInst::ICMP_ULE:
3045 if (isKnownNonZero(V: LHS, Q))
3046 return getFalse(Ty: ITy);
3047 break;
3048 case ICmpInst::ICMP_NE:
3049 case ICmpInst::ICMP_UGT:
3050 if (isKnownNonZero(V: LHS, Q))
3051 return getTrue(Ty: ITy);
3052 break;
3053 case ICmpInst::ICMP_SLT: {
3054 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3055 if (LHSKnown.isNegative())
3056 return getTrue(Ty: ITy);
3057 if (LHSKnown.isNonNegative())
3058 return getFalse(Ty: ITy);
3059 break;
3060 }
3061 case ICmpInst::ICMP_SLE: {
3062 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3063 if (LHSKnown.isNegative())
3064 return getTrue(Ty: ITy);
3065 if (LHSKnown.isNonNegative() && isKnownNonZero(V: LHS, Q))
3066 return getFalse(Ty: ITy);
3067 break;
3068 }
3069 case ICmpInst::ICMP_SGE: {
3070 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3071 if (LHSKnown.isNegative())
3072 return getFalse(Ty: ITy);
3073 if (LHSKnown.isNonNegative())
3074 return getTrue(Ty: ITy);
3075 break;
3076 }
3077 case ICmpInst::ICMP_SGT: {
3078 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3079 if (LHSKnown.isNegative())
3080 return getFalse(Ty: ITy);
3081 if (LHSKnown.isNonNegative() && isKnownNonZero(V: LHS, Q))
3082 return getTrue(Ty: ITy);
3083 break;
3084 }
3085 }
3086
3087 return nullptr;
3088}
3089
3090static Value *simplifyICmpWithConstant(CmpPredicate Pred, Value *LHS,
3091 Value *RHS, const SimplifyQuery &Q) {
3092 Type *ITy = getCompareTy(Op: RHS); // The return type.
3093
3094 Value *X;
3095 const APInt *C;
3096 if (!match(V: RHS, P: m_APIntAllowPoison(Res&: C)))
3097 return nullptr;
3098
3099 // Sign-bit checks can be optimized to true/false after unsigned
3100 // floating-point casts:
3101 // icmp slt (bitcast (uitofp X)), 0 --> false
3102 // icmp sgt (bitcast (uitofp X)), -1 --> true
3103 if (match(V: LHS, P: m_ElementWiseBitCast(Op: m_UIToFP(Op: m_Value(V&: X))))) {
3104 bool TrueIfSigned;
3105 if (isSignBitCheck(Pred, RHS: *C, TrueIfSigned))
3106 return ConstantInt::getBool(Ty: ITy, V: !TrueIfSigned);
3107 }
3108
3109 // Rule out tautological comparisons (eg., ult 0 or uge 0).
3110 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, Other: *C);
3111 if (RHS_CR.isEmptySet())
3112 return ConstantInt::getFalse(Ty: ITy);
3113 if (RHS_CR.isFullSet())
3114 return ConstantInt::getTrue(Ty: ITy);
3115
3116 ConstantRange LHS_CR =
3117 computeConstantRange(V: LHS, ForSigned: CmpInst::isSigned(Pred), UseInstrInfo: Q.IIQ.UseInstrInfo);
3118 if (!LHS_CR.isFullSet()) {
3119 if (RHS_CR.contains(CR: LHS_CR))
3120 return ConstantInt::getTrue(Ty: ITy);
3121 if (RHS_CR.inverse().contains(CR: LHS_CR))
3122 return ConstantInt::getFalse(Ty: ITy);
3123 }
3124
3125 // (mul nuw/nsw X, MulC) != C --> true (if C is not a multiple of MulC)
3126 // (mul nuw/nsw X, MulC) == C --> false (if C is not a multiple of MulC)
3127 const APInt *MulC;
3128 if (Q.IIQ.UseInstrInfo && ICmpInst::isEquality(P: Pred) &&
3129 ((match(V: LHS, P: m_NUWMul(L: m_Value(), R: m_APIntAllowPoison(Res&: MulC))) &&
3130 *MulC != 0 && C->urem(RHS: *MulC) != 0) ||
3131 (match(V: LHS, P: m_NSWMul(L: m_Value(), R: m_APIntAllowPoison(Res&: MulC))) &&
3132 *MulC != 0 && C->srem(RHS: *MulC) != 0)))
3133 return ConstantInt::get(Ty: ITy, V: Pred == ICmpInst::ICMP_NE);
3134
3135 if (Pred == ICmpInst::ICMP_UGE && C->isOne() && isKnownNonZero(V: LHS, Q))
3136 return ConstantInt::getTrue(Ty: ITy);
3137
3138 return nullptr;
3139}
3140
3141enum class MonotonicType { GreaterEq, LowerEq };
3142
3143/// Get values V_i such that V uge V_i (GreaterEq) or V ule V_i (LowerEq).
3144static void getUnsignedMonotonicValues(SmallPtrSetImpl<Value *> &Res, Value *V,
3145 MonotonicType Type,
3146 const SimplifyQuery &Q,
3147 unsigned Depth = 0) {
3148 if (!Res.insert(Ptr: V).second)
3149 return;
3150
3151 // Can be increased if useful.
3152 if (++Depth > 1)
3153 return;
3154
3155 auto *I = dyn_cast<Instruction>(Val: V);
3156 if (!I)
3157 return;
3158
3159 Value *X, *Y;
3160 if (Type == MonotonicType::GreaterEq) {
3161 if (match(V: I, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) ||
3162 match(V: I, P: m_Intrinsic<Intrinsic::uadd_sat>(Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) {
3163 getUnsignedMonotonicValues(Res, V: X, Type, Q, Depth);
3164 getUnsignedMonotonicValues(Res, V: Y, Type, Q, Depth);
3165 }
3166 // X * Y >= X --> true
3167 if (match(V: I, P: m_NUWMul(L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3168 if (isKnownNonZero(V: X, Q))
3169 getUnsignedMonotonicValues(Res, V: Y, Type, Q, Depth);
3170 if (isKnownNonZero(V: Y, Q))
3171 getUnsignedMonotonicValues(Res, V: X, Type, Q, Depth);
3172 }
3173 } else {
3174 assert(Type == MonotonicType::LowerEq);
3175 switch (I->getOpcode()) {
3176 case Instruction::And:
3177 getUnsignedMonotonicValues(Res, V: I->getOperand(i: 0), Type, Q, Depth);
3178 getUnsignedMonotonicValues(Res, V: I->getOperand(i: 1), Type, Q, Depth);
3179 break;
3180 case Instruction::URem:
3181 case Instruction::UDiv:
3182 case Instruction::LShr:
3183 getUnsignedMonotonicValues(Res, V: I->getOperand(i: 0), Type, Q, Depth);
3184 break;
3185 case Instruction::Call:
3186 if (match(V: I, P: m_Intrinsic<Intrinsic::usub_sat>(Op0: m_Value(V&: X))))
3187 getUnsignedMonotonicValues(Res, V: X, Type, Q, Depth);
3188 break;
3189 default:
3190 break;
3191 }
3192 }
3193}
3194
3195static Value *simplifyICmpUsingMonotonicValues(CmpPredicate Pred, Value *LHS,
3196 Value *RHS,
3197 const SimplifyQuery &Q) {
3198 if (Pred != ICmpInst::ICMP_UGE && Pred != ICmpInst::ICMP_ULT)
3199 return nullptr;
3200
3201 // We have LHS uge GreaterValues and LowerValues uge RHS. If any of the
3202 // GreaterValues and LowerValues are the same, it follows that LHS uge RHS.
3203 SmallPtrSet<Value *, 4> GreaterValues;
3204 SmallPtrSet<Value *, 4> LowerValues;
3205 getUnsignedMonotonicValues(Res&: GreaterValues, V: LHS, Type: MonotonicType::GreaterEq, Q);
3206 getUnsignedMonotonicValues(Res&: LowerValues, V: RHS, Type: MonotonicType::LowerEq, Q);
3207 for (Value *GV : GreaterValues)
3208 if (LowerValues.contains(Ptr: GV))
3209 return ConstantInt::getBool(Ty: getCompareTy(Op: LHS),
3210 V: Pred == ICmpInst::ICMP_UGE);
3211 return nullptr;
3212}
3213
3214static Value *simplifyICmpWithBinOpOnLHS(CmpPredicate Pred, BinaryOperator *LBO,
3215 Value *RHS, const SimplifyQuery &Q,
3216 unsigned MaxRecurse) {
3217 Type *ITy = getCompareTy(Op: RHS); // The return type.
3218
3219 Value *Y = nullptr;
3220 // icmp pred (or X, Y), X
3221 if (match(V: LBO, P: m_c_Or(L: m_Value(V&: Y), R: m_Specific(V: RHS)))) {
3222 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
3223 KnownBits RHSKnown = computeKnownBits(V: RHS, Q);
3224 KnownBits YKnown = computeKnownBits(V: Y, Q);
3225 if (RHSKnown.isNonNegative() && YKnown.isNegative())
3226 return Pred == ICmpInst::ICMP_SLT ? getTrue(Ty: ITy) : getFalse(Ty: ITy);
3227 if (RHSKnown.isNegative() || YKnown.isNonNegative())
3228 return Pred == ICmpInst::ICMP_SLT ? getFalse(Ty: ITy) : getTrue(Ty: ITy);
3229 }
3230 }
3231
3232 // icmp pred (urem X, Y), Y
3233 if (match(V: LBO, P: m_URem(L: m_Value(), R: m_Specific(V: RHS)))) {
3234 switch (Pred) {
3235 default:
3236 break;
3237 case ICmpInst::ICMP_SGT:
3238 case ICmpInst::ICMP_SGE: {
3239 KnownBits Known = computeKnownBits(V: RHS, Q);
3240 if (!Known.isNonNegative())
3241 break;
3242 [[fallthrough]];
3243 }
3244 case ICmpInst::ICMP_EQ:
3245 case ICmpInst::ICMP_UGT:
3246 case ICmpInst::ICMP_UGE:
3247 return getFalse(Ty: ITy);
3248 case ICmpInst::ICMP_SLT:
3249 case ICmpInst::ICMP_SLE: {
3250 KnownBits Known = computeKnownBits(V: RHS, Q);
3251 if (!Known.isNonNegative())
3252 break;
3253 [[fallthrough]];
3254 }
3255 case ICmpInst::ICMP_NE:
3256 case ICmpInst::ICMP_ULT:
3257 case ICmpInst::ICMP_ULE:
3258 return getTrue(Ty: ITy);
3259 }
3260 }
3261
3262 // If x is nonzero:
3263 // x >>u C <u x --> true for C != 0.
3264 // x >>u C != x --> true for C != 0.
3265 // x >>u C >=u x --> false for C != 0.
3266 // x >>u C == x --> false for C != 0.
3267 // x udiv C <u x --> true for C != 1.
3268 // x udiv C != x --> true for C != 1.
3269 // x udiv C >=u x --> false for C != 1.
3270 // x udiv C == x --> false for C != 1.
3271 // TODO: allow non-constant shift amount/divisor
3272 const APInt *C;
3273 if ((match(V: LBO, P: m_LShr(L: m_Specific(V: RHS), R: m_APInt(Res&: C))) && *C != 0) ||
3274 (match(V: LBO, P: m_UDiv(L: m_Specific(V: RHS), R: m_APInt(Res&: C))) && *C != 1)) {
3275 if (isKnownNonZero(V: RHS, Q)) {
3276 switch (Pred) {
3277 default:
3278 break;
3279 case ICmpInst::ICMP_EQ:
3280 case ICmpInst::ICMP_UGE:
3281 case ICmpInst::ICMP_UGT:
3282 return getFalse(Ty: ITy);
3283 case ICmpInst::ICMP_NE:
3284 case ICmpInst::ICMP_ULT:
3285 case ICmpInst::ICMP_ULE:
3286 return getTrue(Ty: ITy);
3287 }
3288 }
3289 }
3290
3291 // (x*C1)/C2 <= x for C1 <= C2.
3292 // This holds even if the multiplication overflows: Assume that x != 0 and
3293 // arithmetic is modulo M. For overflow to occur we must have C1 >= M/x and
3294 // thus C2 >= M/x. It follows that (x*C1)/C2 <= (M-1)/C2 <= ((M-1)*x)/M < x.
3295 //
3296 // Additionally, either the multiplication and division might be represented
3297 // as shifts:
3298 // (x*C1)>>C2 <= x for C1 < 2**C2.
3299 // (x<<C1)/C2 <= x for 2**C1 < C2.
3300 const APInt *C1, *C2;
3301 if ((match(V: LBO, P: m_UDiv(L: m_Mul(L: m_Specific(V: RHS), R: m_APInt(Res&: C1)), R: m_APInt(Res&: C2))) &&
3302 C1->ule(RHS: *C2)) ||
3303 (match(V: LBO, P: m_LShr(L: m_Mul(L: m_Specific(V: RHS), R: m_APInt(Res&: C1)), R: m_APInt(Res&: C2))) &&
3304 C1->ule(RHS: APInt(C2->getBitWidth(), 1) << *C2)) ||
3305 (match(V: LBO, P: m_UDiv(L: m_Shl(L: m_Specific(V: RHS), R: m_APInt(Res&: C1)), R: m_APInt(Res&: C2))) &&
3306 (APInt(C1->getBitWidth(), 1) << *C1).ule(RHS: *C2))) {
3307 if (Pred == ICmpInst::ICMP_UGT)
3308 return getFalse(Ty: ITy);
3309 if (Pred == ICmpInst::ICMP_ULE)
3310 return getTrue(Ty: ITy);
3311 }
3312
3313 // (sub C, X) == X, C is odd --> false
3314 // (sub C, X) != X, C is odd --> true
3315 if (match(V: LBO, P: m_Sub(L: m_APIntAllowPoison(Res&: C), R: m_Specific(V: RHS))) &&
3316 (*C & 1) == 1 && ICmpInst::isEquality(P: Pred))
3317 return (Pred == ICmpInst::ICMP_EQ) ? getFalse(Ty: ITy) : getTrue(Ty: ITy);
3318
3319 return nullptr;
3320}
3321
3322// If only one of the icmp's operands has NSW flags, try to prove that:
3323//
3324// icmp slt/sgt/sle/sge (x + C1), (x +nsw C2)
3325//
3326// is equivalent to:
3327//
3328// icmp slt/sgt/sle/sge C1, C2
3329//
3330// which is true if x + C2 has the NSW flags set and:
3331// *) C1 <= C2 && C1 >= 0, or
3332// *) C2 <= C1 && C1 <= 0.
3333//
3334static bool trySimplifyICmpWithAdds(CmpPredicate Pred, Value *LHS, Value *RHS,
3335 const InstrInfoQuery &IIQ) {
3336 // TODO: support other predicates.
3337 if (!ICmpInst::isSigned(Pred) || !IIQ.UseInstrInfo)
3338 return false;
3339
3340 // Canonicalize nsw add as RHS.
3341 if (!match(V: RHS, P: m_NSWAdd(L: m_Value(), R: m_Value())))
3342 std::swap(a&: LHS, b&: RHS);
3343 if (!match(V: RHS, P: m_NSWAdd(L: m_Value(), R: m_Value())))
3344 return false;
3345
3346 Value *X;
3347 const APInt *C1, *C2;
3348 if (!match(V: LHS, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C1))) ||
3349 !match(V: RHS, P: m_Add(L: m_Specific(V: X), R: m_APInt(Res&: C2))))
3350 return false;
3351
3352 return (C1->sle(RHS: *C2) && C1->isNonNegative()) ||
3353 (C2->sle(RHS: *C1) && C1->isNonPositive());
3354}
3355
3356/// TODO: A large part of this logic is duplicated in InstCombine's
3357/// foldICmpBinOp(). We should be able to share that and avoid the code
3358/// duplication.
3359static Value *simplifyICmpWithBinOp(CmpPredicate Pred, Value *LHS, Value *RHS,
3360 const SimplifyQuery &Q,
3361 unsigned MaxRecurse) {
3362 BinaryOperator *LBO = dyn_cast<BinaryOperator>(Val: LHS);
3363 BinaryOperator *RBO = dyn_cast<BinaryOperator>(Val: RHS);
3364 if (MaxRecurse && (LBO || RBO)) {
3365 // Analyze the case when either LHS or RHS is an add instruction.
3366 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3367 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
3368 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
3369 if (LBO && LBO->getOpcode() == Instruction::Add) {
3370 A = LBO->getOperand(i_nocapture: 0);
3371 B = LBO->getOperand(i_nocapture: 1);
3372 NoLHSWrapProblem =
3373 ICmpInst::isEquality(P: Pred) ||
3374 (CmpInst::isUnsigned(Pred) &&
3375 Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO))) ||
3376 (CmpInst::isSigned(Pred) &&
3377 Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO)));
3378 }
3379 if (RBO && RBO->getOpcode() == Instruction::Add) {
3380 C = RBO->getOperand(i_nocapture: 0);
3381 D = RBO->getOperand(i_nocapture: 1);
3382 NoRHSWrapProblem =
3383 ICmpInst::isEquality(P: Pred) ||
3384 (CmpInst::isUnsigned(Pred) &&
3385 Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: RBO))) ||
3386 (CmpInst::isSigned(Pred) &&
3387 Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: RBO)));
3388 }
3389
3390 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3391 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
3392 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: A == RHS ? B : A,
3393 RHS: Constant::getNullValue(Ty: RHS->getType()), Q,
3394 MaxRecurse: MaxRecurse - 1))
3395 return V;
3396
3397 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3398 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
3399 if (Value *V =
3400 simplifyICmpInst(Predicate: Pred, LHS: Constant::getNullValue(Ty: LHS->getType()),
3401 RHS: C == LHS ? D : C, Q, MaxRecurse: MaxRecurse - 1))
3402 return V;
3403
3404 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
3405 bool CanSimplify = (NoLHSWrapProblem && NoRHSWrapProblem) ||
3406 trySimplifyICmpWithAdds(Pred, LHS, RHS, IIQ: Q.IIQ);
3407 if (A && C && (A == C || A == D || B == C || B == D) && CanSimplify) {
3408 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3409 Value *Y, *Z;
3410 if (A == C) {
3411 // C + B == C + D -> B == D
3412 Y = B;
3413 Z = D;
3414 } else if (A == D) {
3415 // D + B == C + D -> B == C
3416 Y = B;
3417 Z = C;
3418 } else if (B == C) {
3419 // A + C == C + D -> A == D
3420 Y = A;
3421 Z = D;
3422 } else {
3423 assert(B == D);
3424 // A + D == C + D -> A == C
3425 Y = A;
3426 Z = C;
3427 }
3428 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: Y, RHS: Z, Q, MaxRecurse: MaxRecurse - 1))
3429 return V;
3430 }
3431 }
3432
3433 if (LBO)
3434 if (Value *V = simplifyICmpWithBinOpOnLHS(Pred, LBO, RHS, Q, MaxRecurse))
3435 return V;
3436
3437 if (RBO)
3438 if (Value *V = simplifyICmpWithBinOpOnLHS(
3439 Pred: ICmpInst::getSwappedPredicate(pred: Pred), LBO: RBO, RHS: LHS, Q, MaxRecurse))
3440 return V;
3441
3442 // 0 - (zext X) pred C
3443 if (!CmpInst::isUnsigned(Pred) && match(V: LHS, P: m_Neg(V: m_ZExt(Op: m_Value())))) {
3444 const APInt *C;
3445 if (match(V: RHS, P: m_APInt(Res&: C))) {
3446 if (C->isStrictlyPositive()) {
3447 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_NE)
3448 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3449 if (Pred == ICmpInst::ICMP_SGE || Pred == ICmpInst::ICMP_EQ)
3450 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3451 }
3452 if (C->isNonNegative()) {
3453 if (Pred == ICmpInst::ICMP_SLE)
3454 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3455 if (Pred == ICmpInst::ICMP_SGT)
3456 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3457 }
3458 }
3459 }
3460
3461 // If C2 is a power-of-2 and C is not:
3462 // (C2 << X) == C --> false
3463 // (C2 << X) != C --> true
3464 const APInt *C;
3465 if (match(V: LHS, P: m_Shl(L: m_Power2(), R: m_Value())) &&
3466 match(V: RHS, P: m_APIntAllowPoison(Res&: C)) && !C->isPowerOf2()) {
3467 // C2 << X can equal zero in some circumstances.
3468 // This simplification might be unsafe if C is zero.
3469 //
3470 // We know it is safe if:
3471 // - The shift is nsw. We can't shift out the one bit.
3472 // - The shift is nuw. We can't shift out the one bit.
3473 // - C2 is one.
3474 // - C isn't zero.
3475 if (Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO)) ||
3476 Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO)) ||
3477 match(V: LHS, P: m_Shl(L: m_One(), R: m_Value())) || !C->isZero()) {
3478 if (Pred == ICmpInst::ICMP_EQ)
3479 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3480 if (Pred == ICmpInst::ICMP_NE)
3481 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3482 }
3483 }
3484
3485 // If C is a power-of-2:
3486 // (C << X) >u 0x8000 --> false
3487 // (C << X) <=u 0x8000 --> true
3488 if (match(V: LHS, P: m_Shl(L: m_Power2(), R: m_Value())) && match(V: RHS, P: m_SignMask())) {
3489 if (Pred == ICmpInst::ICMP_UGT)
3490 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3491 if (Pred == ICmpInst::ICMP_ULE)
3492 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3493 }
3494
3495 if (!MaxRecurse || !LBO || !RBO || LBO->getOpcode() != RBO->getOpcode())
3496 return nullptr;
3497
3498 if (LBO->getOperand(i_nocapture: 0) == RBO->getOperand(i_nocapture: 0)) {
3499 switch (LBO->getOpcode()) {
3500 default:
3501 break;
3502 case Instruction::Shl: {
3503 bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: LBO) && Q.IIQ.hasNoUnsignedWrap(Op: RBO);
3504 bool NSW = Q.IIQ.hasNoSignedWrap(Op: LBO) && Q.IIQ.hasNoSignedWrap(Op: RBO);
3505 if (!NUW || (ICmpInst::isSigned(Pred) && !NSW) ||
3506 !isKnownNonZero(V: LBO->getOperand(i_nocapture: 0), Q))
3507 break;
3508 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 1),
3509 RHS: RBO->getOperand(i_nocapture: 1), Q, MaxRecurse: MaxRecurse - 1))
3510 return V;
3511 break;
3512 }
3513 // If C1 & C2 == C1, A = X and/or C1, B = X and/or C2:
3514 // icmp ule A, B -> true
3515 // icmp ugt A, B -> false
3516 // icmp sle A, B -> true (C1 and C2 are the same sign)
3517 // icmp sgt A, B -> false (C1 and C2 are the same sign)
3518 case Instruction::And:
3519 case Instruction::Or: {
3520 const APInt *C1, *C2;
3521 if (ICmpInst::isRelational(P: Pred) &&
3522 match(V: LBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C1)) &&
3523 match(V: RBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C2))) {
3524 if (!C1->isSubsetOf(RHS: *C2)) {
3525 std::swap(a&: C1, b&: C2);
3526 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
3527 }
3528 if (C1->isSubsetOf(RHS: *C2)) {
3529 if (Pred == ICmpInst::ICMP_ULE)
3530 return ConstantInt::getTrue(Ty: getCompareTy(Op: LHS));
3531 if (Pred == ICmpInst::ICMP_UGT)
3532 return ConstantInt::getFalse(Ty: getCompareTy(Op: LHS));
3533 if (C1->isNonNegative() == C2->isNonNegative()) {
3534 if (Pred == ICmpInst::ICMP_SLE)
3535 return ConstantInt::getTrue(Ty: getCompareTy(Op: LHS));
3536 if (Pred == ICmpInst::ICMP_SGT)
3537 return ConstantInt::getFalse(Ty: getCompareTy(Op: LHS));
3538 }
3539 }
3540 }
3541 break;
3542 }
3543 }
3544 }
3545
3546 if (LBO->getOperand(i_nocapture: 1) == RBO->getOperand(i_nocapture: 1)) {
3547 switch (LBO->getOpcode()) {
3548 default:
3549 break;
3550 case Instruction::UDiv:
3551 case Instruction::LShr:
3552 if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(Op: LBO) ||
3553 !Q.IIQ.isExact(Op: RBO))
3554 break;
3555 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3556 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3557 return V;
3558 break;
3559 case Instruction::SDiv:
3560 if (!ICmpInst::isEquality(P: Pred) || !Q.IIQ.isExact(Op: LBO) ||
3561 !Q.IIQ.isExact(Op: RBO))
3562 break;
3563 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3564 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3565 return V;
3566 break;
3567 case Instruction::AShr:
3568 if (!Q.IIQ.isExact(Op: LBO) || !Q.IIQ.isExact(Op: RBO))
3569 break;
3570 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3571 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3572 return V;
3573 break;
3574 case Instruction::Shl: {
3575 bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: LBO) && Q.IIQ.hasNoUnsignedWrap(Op: RBO);
3576 bool NSW = Q.IIQ.hasNoSignedWrap(Op: LBO) && Q.IIQ.hasNoSignedWrap(Op: RBO);
3577 if (!NUW && !NSW)
3578 break;
3579 if (!NSW && ICmpInst::isSigned(Pred))
3580 break;
3581 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3582 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3583 return V;
3584 break;
3585 }
3586 }
3587 }
3588 return nullptr;
3589}
3590
3591/// simplify integer comparisons where at least one operand of the compare
3592/// matches an integer min/max idiom.
3593static Value *simplifyICmpWithMinMax(CmpPredicate Pred, Value *LHS, Value *RHS,
3594 const SimplifyQuery &Q,
3595 unsigned MaxRecurse) {
3596 Type *ITy = getCompareTy(Op: LHS); // The return type.
3597 Value *A, *B;
3598 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
3599 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3600
3601 // Signed variants on "max(a,b)>=a -> true".
3602 if (match(V: LHS, P: m_SMax(L: m_Value(V&: A), R: m_Value(V&: B))) && (A == RHS || B == RHS)) {
3603 if (A != RHS)
3604 std::swap(a&: A, b&: B); // smax(A, B) pred A.
3605 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3606 // We analyze this as smax(A, B) pred A.
3607 P = Pred;
3608 } else if (match(V: RHS, P: m_SMax(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3609 (A == LHS || B == LHS)) {
3610 if (A != LHS)
3611 std::swap(a&: A, b&: B); // A pred smax(A, B).
3612 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3613 // We analyze this as smax(A, B) swapped-pred A.
3614 P = CmpInst::getSwappedPredicate(pred: Pred);
3615 } else if (match(V: LHS, P: m_SMin(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3616 (A == RHS || B == RHS)) {
3617 if (A != RHS)
3618 std::swap(a&: A, b&: B); // smin(A, B) pred A.
3619 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3620 // We analyze this as smax(-A, -B) swapped-pred -A.
3621 // Note that we do not need to actually form -A or -B thanks to EqP.
3622 P = CmpInst::getSwappedPredicate(pred: Pred);
3623 } else if (match(V: RHS, P: m_SMin(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3624 (A == LHS || B == LHS)) {
3625 if (A != LHS)
3626 std::swap(a&: A, b&: B); // A pred smin(A, B).
3627 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3628 // We analyze this as smax(-A, -B) pred -A.
3629 // Note that we do not need to actually form -A or -B thanks to EqP.
3630 P = Pred;
3631 }
3632 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3633 // Cases correspond to "max(A, B) p A".
3634 switch (P) {
3635 default:
3636 break;
3637 case CmpInst::ICMP_EQ:
3638 case CmpInst::ICMP_SLE:
3639 // Equivalent to "A EqP B". This may be the same as the condition tested
3640 // in the max/min; if so, we can just return that.
3641 if (Value *V = extractEquivalentCondition(V: LHS, Pred: EqP, LHS: A, RHS: B))
3642 return V;
3643 if (Value *V = extractEquivalentCondition(V: RHS, Pred: EqP, LHS: A, RHS: B))
3644 return V;
3645 // Otherwise, see if "A EqP B" simplifies.
3646 if (MaxRecurse)
3647 if (Value *V = simplifyICmpInst(Predicate: EqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3648 return V;
3649 break;
3650 case CmpInst::ICMP_NE:
3651 case CmpInst::ICMP_SGT: {
3652 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(pred: EqP);
3653 // Equivalent to "A InvEqP B". This may be the same as the condition
3654 // tested in the max/min; if so, we can just return that.
3655 if (Value *V = extractEquivalentCondition(V: LHS, Pred: InvEqP, LHS: A, RHS: B))
3656 return V;
3657 if (Value *V = extractEquivalentCondition(V: RHS, Pred: InvEqP, LHS: A, RHS: B))
3658 return V;
3659 // Otherwise, see if "A InvEqP B" simplifies.
3660 if (MaxRecurse)
3661 if (Value *V = simplifyICmpInst(Predicate: InvEqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3662 return V;
3663 break;
3664 }
3665 case CmpInst::ICMP_SGE:
3666 // Always true.
3667 return getTrue(Ty: ITy);
3668 case CmpInst::ICMP_SLT:
3669 // Always false.
3670 return getFalse(Ty: ITy);
3671 }
3672 }
3673
3674 // Unsigned variants on "max(a,b)>=a -> true".
3675 P = CmpInst::BAD_ICMP_PREDICATE;
3676 if (match(V: LHS, P: m_UMax(L: m_Value(V&: A), R: m_Value(V&: B))) && (A == RHS || B == RHS)) {
3677 if (A != RHS)
3678 std::swap(a&: A, b&: B); // umax(A, B) pred A.
3679 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3680 // We analyze this as umax(A, B) pred A.
3681 P = Pred;
3682 } else if (match(V: RHS, P: m_UMax(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3683 (A == LHS || B == LHS)) {
3684 if (A != LHS)
3685 std::swap(a&: A, b&: B); // A pred umax(A, B).
3686 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3687 // We analyze this as umax(A, B) swapped-pred A.
3688 P = CmpInst::getSwappedPredicate(pred: Pred);
3689 } else if (match(V: LHS, P: m_UMin(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3690 (A == RHS || B == RHS)) {
3691 if (A != RHS)
3692 std::swap(a&: A, b&: B); // umin(A, B) pred A.
3693 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3694 // We analyze this as umax(-A, -B) swapped-pred -A.
3695 // Note that we do not need to actually form -A or -B thanks to EqP.
3696 P = CmpInst::getSwappedPredicate(pred: Pred);
3697 } else if (match(V: RHS, P: m_UMin(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3698 (A == LHS || B == LHS)) {
3699 if (A != LHS)
3700 std::swap(a&: A, b&: B); // A pred umin(A, B).
3701 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3702 // We analyze this as umax(-A, -B) pred -A.
3703 // Note that we do not need to actually form -A or -B thanks to EqP.
3704 P = Pred;
3705 }
3706 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3707 // Cases correspond to "max(A, B) p A".
3708 switch (P) {
3709 default:
3710 break;
3711 case CmpInst::ICMP_EQ:
3712 case CmpInst::ICMP_ULE:
3713 // Equivalent to "A EqP B". This may be the same as the condition tested
3714 // in the max/min; if so, we can just return that.
3715 if (Value *V = extractEquivalentCondition(V: LHS, Pred: EqP, LHS: A, RHS: B))
3716 return V;
3717 if (Value *V = extractEquivalentCondition(V: RHS, Pred: EqP, LHS: A, RHS: B))
3718 return V;
3719 // Otherwise, see if "A EqP B" simplifies.
3720 if (MaxRecurse)
3721 if (Value *V = simplifyICmpInst(Predicate: EqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3722 return V;
3723 break;
3724 case CmpInst::ICMP_NE:
3725 case CmpInst::ICMP_UGT: {
3726 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(pred: EqP);
3727 // Equivalent to "A InvEqP B". This may be the same as the condition
3728 // tested in the max/min; if so, we can just return that.
3729 if (Value *V = extractEquivalentCondition(V: LHS, Pred: InvEqP, LHS: A, RHS: B))
3730 return V;
3731 if (Value *V = extractEquivalentCondition(V: RHS, Pred: InvEqP, LHS: A, RHS: B))
3732 return V;
3733 // Otherwise, see if "A InvEqP B" simplifies.
3734 if (MaxRecurse)
3735 if (Value *V = simplifyICmpInst(Predicate: InvEqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3736 return V;
3737 break;
3738 }
3739 case CmpInst::ICMP_UGE:
3740 return getTrue(Ty: ITy);
3741 case CmpInst::ICMP_ULT:
3742 return getFalse(Ty: ITy);
3743 }
3744 }
3745
3746 // Comparing 1 each of min/max with a common operand?
3747 // Canonicalize min operand to RHS.
3748 if (match(V: LHS, P: m_UMin(L: m_Value(), R: m_Value())) ||
3749 match(V: LHS, P: m_SMin(L: m_Value(), R: m_Value()))) {
3750 std::swap(a&: LHS, b&: RHS);
3751 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
3752 }
3753
3754 Value *C, *D;
3755 if (match(V: LHS, P: m_SMax(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3756 match(V: RHS, P: m_SMin(L: m_Value(V&: C), R: m_Value(V&: D))) &&
3757 (A == C || A == D || B == C || B == D)) {
3758 // smax(A, B) >=s smin(A, D) --> true
3759 if (Pred == CmpInst::ICMP_SGE)
3760 return getTrue(Ty: ITy);
3761 // smax(A, B) <s smin(A, D) --> false
3762 if (Pred == CmpInst::ICMP_SLT)
3763 return getFalse(Ty: ITy);
3764 } else if (match(V: LHS, P: m_UMax(L: m_Value(V&: A), R: m_Value(V&: B))) &&
3765 match(V: RHS, P: m_UMin(L: m_Value(V&: C), R: m_Value(V&: D))) &&
3766 (A == C || A == D || B == C || B == D)) {
3767 // umax(A, B) >=u umin(A, D) --> true
3768 if (Pred == CmpInst::ICMP_UGE)
3769 return getTrue(Ty: ITy);
3770 // umax(A, B) <u umin(A, D) --> false
3771 if (Pred == CmpInst::ICMP_ULT)
3772 return getFalse(Ty: ITy);
3773 }
3774
3775 return nullptr;
3776}
3777
3778static Value *simplifyICmpWithDominatingAssume(CmpPredicate Predicate,
3779 Value *LHS, Value *RHS,
3780 const SimplifyQuery &Q) {
3781 // Gracefully handle instructions that have not been inserted yet.
3782 if (!Q.AC || !Q.CxtI)
3783 return nullptr;
3784
3785 for (Value *AssumeBaseOp : {LHS, RHS}) {
3786 for (auto &AssumeVH : Q.AC->assumptionsFor(V: AssumeBaseOp)) {
3787 if (!AssumeVH)
3788 continue;
3789
3790 CallInst *Assume = cast<CallInst>(Val&: AssumeVH);
3791 if (std::optional<bool> Imp = isImpliedCondition(
3792 LHS: Assume->getArgOperand(i: 0), RHSPred: Predicate, RHSOp0: LHS, RHSOp1: RHS, DL: Q.DL))
3793 if (isValidAssumeForContext(I: Assume, CxtI: Q.CxtI, DT: Q.DT))
3794 return ConstantInt::get(Ty: getCompareTy(Op: LHS), V: *Imp);
3795 }
3796 }
3797
3798 return nullptr;
3799}
3800
3801static Value *simplifyICmpWithIntrinsicOnLHS(CmpPredicate Pred, Value *LHS,
3802 Value *RHS) {
3803 auto *II = dyn_cast<IntrinsicInst>(Val: LHS);
3804 if (!II)
3805 return nullptr;
3806
3807 switch (II->getIntrinsicID()) {
3808 case Intrinsic::uadd_sat:
3809 // uadd.sat(X, Y) uge X + Y
3810 if (match(V: RHS, P: m_c_Add(L: m_Specific(V: II->getArgOperand(i: 0)),
3811 R: m_Specific(V: II->getArgOperand(i: 1))))) {
3812 if (Pred == ICmpInst::ICMP_UGE)
3813 return ConstantInt::getTrue(Ty: getCompareTy(Op: II));
3814 if (Pred == ICmpInst::ICMP_ULT)
3815 return ConstantInt::getFalse(Ty: getCompareTy(Op: II));
3816 }
3817 return nullptr;
3818 case Intrinsic::usub_sat:
3819 // usub.sat(X, Y) ule X - Y
3820 if (match(V: RHS, P: m_Sub(L: m_Specific(V: II->getArgOperand(i: 0)),
3821 R: m_Specific(V: II->getArgOperand(i: 1))))) {
3822 if (Pred == ICmpInst::ICMP_ULE)
3823 return ConstantInt::getTrue(Ty: getCompareTy(Op: II));
3824 if (Pred == ICmpInst::ICMP_UGT)
3825 return ConstantInt::getFalse(Ty: getCompareTy(Op: II));
3826 }
3827 return nullptr;
3828 default:
3829 return nullptr;
3830 }
3831}
3832
3833/// Helper method to get range from metadata or attribute.
3834static std::optional<ConstantRange> getRange(Value *V,
3835 const InstrInfoQuery &IIQ) {
3836 if (Instruction *I = dyn_cast<Instruction>(Val: V))
3837 if (MDNode *MD = IIQ.getMetadata(I, KindID: LLVMContext::MD_range))
3838 return getConstantRangeFromMetadata(RangeMD: *MD);
3839
3840 if (const Argument *A = dyn_cast<Argument>(Val: V))
3841 return A->getRange();
3842 else if (const CallBase *CB = dyn_cast<CallBase>(Val: V))
3843 return CB->getRange();
3844
3845 return std::nullopt;
3846}
3847
3848/// Given operands for an ICmpInst, see if we can fold the result.
3849/// If not, this returns null.
3850static Value *simplifyICmpInst(CmpPredicate Pred, Value *LHS, Value *RHS,
3851 const SimplifyQuery &Q, unsigned MaxRecurse) {
3852 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3853
3854 if (Constant *CLHS = dyn_cast<Constant>(Val: LHS)) {
3855 if (Constant *CRHS = dyn_cast<Constant>(Val: RHS))
3856 return ConstantFoldCompareInstOperands(Predicate: Pred, LHS: CLHS, RHS: CRHS, DL: Q.DL, TLI: Q.TLI);
3857
3858 // If we have a constant, make sure it is on the RHS.
3859 std::swap(a&: LHS, b&: RHS);
3860 Pred = CmpInst::getSwappedPredicate(pred: Pred);
3861 }
3862 assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3863
3864 Type *ITy = getCompareTy(Op: LHS); // The return type.
3865
3866 // icmp poison, X -> poison
3867 if (isa<PoisonValue>(Val: RHS))
3868 return PoisonValue::get(T: ITy);
3869
3870 // For EQ and NE, we can always pick a value for the undef to make the
3871 // predicate pass or fail, so we can return undef.
3872 // Matches behavior in llvm::ConstantFoldCompareInstruction.
3873 if (Q.isUndefValue(V: RHS) && ICmpInst::isEquality(P: Pred))
3874 return UndefValue::get(T: ITy);
3875
3876 // icmp X, X -> true/false
3877 // icmp X, undef -> true/false because undef could be X.
3878 if (LHS == RHS || Q.isUndefValue(V: RHS))
3879 return ConstantInt::get(Ty: ITy, V: CmpInst::isTrueWhenEqual(predicate: Pred));
3880
3881 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3882 return V;
3883
3884 // TODO: Sink/common this with other potentially expensive calls that use
3885 // ValueTracking? See comment below for isKnownNonEqual().
3886 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3887 return V;
3888
3889 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q))
3890 return V;
3891
3892 // If both operands have range metadata, use the metadata
3893 // to simplify the comparison.
3894 if (std::optional<ConstantRange> RhsCr = getRange(V: RHS, IIQ: Q.IIQ))
3895 if (std::optional<ConstantRange> LhsCr = getRange(V: LHS, IIQ: Q.IIQ)) {
3896 if (LhsCr->icmp(Pred, Other: *RhsCr))
3897 return ConstantInt::getTrue(Ty: ITy);
3898
3899 if (LhsCr->icmp(Pred: CmpInst::getInversePredicate(pred: Pred), Other: *RhsCr))
3900 return ConstantInt::getFalse(Ty: ITy);
3901 }
3902
3903 // Compare of cast, for example (zext X) != 0 -> X != 0
3904 if (isa<CastInst>(Val: LHS) && (isa<Constant>(Val: RHS) || isa<CastInst>(Val: RHS))) {
3905 Instruction *LI = cast<CastInst>(Val: LHS);
3906 Value *SrcOp = LI->getOperand(i: 0);
3907 Type *SrcTy = SrcOp->getType();
3908 Type *DstTy = LI->getType();
3909
3910 // Turn icmp (ptrtoint/ptrtoaddr x), (ptrtoint/ptrtoaddr/constant) into a
3911 // compare of the input if the integer type is the same size as the
3912 // pointer address type (icmp only compares the address of the pointer).
3913 if (MaxRecurse && (isa<PtrToIntInst, PtrToAddrInst>(Val: LI)) &&
3914 Q.DL.getAddressType(PtrTy: SrcTy) == DstTy) {
3915 if (Constant *RHSC = dyn_cast<Constant>(Val: RHS)) {
3916 // Transfer the cast to the constant.
3917 if (Value *V = simplifyICmpInst(Pred, LHS: SrcOp,
3918 RHS: ConstantExpr::getIntToPtr(C: RHSC, Ty: SrcTy),
3919 Q, MaxRecurse: MaxRecurse - 1))
3920 return V;
3921 } else if (isa<PtrToIntInst, PtrToAddrInst>(Val: RHS)) {
3922 auto *RI = cast<CastInst>(Val: RHS);
3923 if (RI->getOperand(i_nocapture: 0)->getType() == SrcTy)
3924 // Compare without the cast.
3925 if (Value *V = simplifyICmpInst(Pred, LHS: SrcOp, RHS: RI->getOperand(i_nocapture: 0), Q,
3926 MaxRecurse: MaxRecurse - 1))
3927 return V;
3928 }
3929 }
3930
3931 if (isa<ZExtInst>(Val: LHS)) {
3932 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3933 // same type.
3934 if (ZExtInst *RI = dyn_cast<ZExtInst>(Val: RHS)) {
3935 if (MaxRecurse && SrcTy == RI->getOperand(i_nocapture: 0)->getType())
3936 // Compare X and Y. Note that signed predicates become unsigned.
3937 if (Value *V =
3938 simplifyICmpInst(Pred: ICmpInst::getUnsignedPredicate(Pred), LHS: SrcOp,
3939 RHS: RI->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3940 return V;
3941 }
3942 // Fold (zext X) ule (sext X), (zext X) sge (sext X) to true.
3943 else if (SExtInst *RI = dyn_cast<SExtInst>(Val: RHS)) {
3944 if (SrcOp == RI->getOperand(i_nocapture: 0)) {
3945 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_SGE)
3946 return ConstantInt::getTrue(Ty: ITy);
3947 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SLT)
3948 return ConstantInt::getFalse(Ty: ITy);
3949 }
3950 }
3951 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3952 // too. If not, then try to deduce the result of the comparison.
3953 else if (match(V: RHS, P: m_ImmConstant())) {
3954 Constant *C = dyn_cast<Constant>(Val: RHS);
3955 assert(C != nullptr);
3956
3957 // Compute the constant that would happen if we truncated to SrcTy then
3958 // reextended to DstTy.
3959 Constant *Trunc =
3960 ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy: SrcTy, DL: Q.DL);
3961 assert(Trunc && "Constant-fold of ImmConstant should not fail");
3962 Constant *RExt =
3963 ConstantFoldCastOperand(Opcode: CastInst::ZExt, C: Trunc, DestTy: DstTy, DL: Q.DL);
3964 assert(RExt && "Constant-fold of ImmConstant should not fail");
3965 Constant *AnyEq =
3966 ConstantFoldCompareInstOperands(Predicate: ICmpInst::ICMP_EQ, LHS: RExt, RHS: C, DL: Q.DL);
3967 assert(AnyEq && "Constant-fold of ImmConstant should not fail");
3968
3969 // If the re-extended constant didn't change any of the elements then
3970 // this is effectively also a case of comparing two zero-extended
3971 // values.
3972 if (AnyEq->isAllOnesValue() && MaxRecurse)
3973 if (Value *V = simplifyICmpInst(Pred: ICmpInst::getUnsignedPredicate(Pred),
3974 LHS: SrcOp, RHS: Trunc, Q, MaxRecurse: MaxRecurse - 1))
3975 return V;
3976
3977 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3978 // there. Use this to work out the result of the comparison.
3979 if (AnyEq->isNullValue()) {
3980 switch (Pred) {
3981 default:
3982 llvm_unreachable("Unknown ICmp predicate!");
3983 // LHS <u RHS.
3984 case ICmpInst::ICMP_EQ:
3985 case ICmpInst::ICMP_UGT:
3986 case ICmpInst::ICMP_UGE:
3987 return Constant::getNullValue(Ty: ITy);
3988
3989 case ICmpInst::ICMP_NE:
3990 case ICmpInst::ICMP_ULT:
3991 case ICmpInst::ICMP_ULE:
3992 return Constant::getAllOnesValue(Ty: ITy);
3993
3994 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3995 // is non-negative then LHS <s RHS.
3996 case ICmpInst::ICMP_SGT:
3997 case ICmpInst::ICMP_SGE:
3998 return ConstantFoldCompareInstOperands(
3999 Predicate: ICmpInst::ICMP_SLT, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4000 DL: Q.DL);
4001 case ICmpInst::ICMP_SLT:
4002 case ICmpInst::ICMP_SLE:
4003 return ConstantFoldCompareInstOperands(
4004 Predicate: ICmpInst::ICMP_SGE, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4005 DL: Q.DL);
4006 }
4007 }
4008 }
4009 }
4010
4011 if (isa<SExtInst>(Val: LHS)) {
4012 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
4013 // same type.
4014 if (SExtInst *RI = dyn_cast<SExtInst>(Val: RHS)) {
4015 if (MaxRecurse && SrcTy == RI->getOperand(i_nocapture: 0)->getType())
4016 // Compare X and Y. Note that the predicate does not change.
4017 if (Value *V = simplifyICmpInst(Pred, LHS: SrcOp, RHS: RI->getOperand(i_nocapture: 0), Q,
4018 MaxRecurse: MaxRecurse - 1))
4019 return V;
4020 }
4021 // Fold (sext X) uge (zext X), (sext X) sle (zext X) to true.
4022 else if (ZExtInst *RI = dyn_cast<ZExtInst>(Val: RHS)) {
4023 if (SrcOp == RI->getOperand(i_nocapture: 0)) {
4024 if (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_SLE)
4025 return ConstantInt::getTrue(Ty: ITy);
4026 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SGT)
4027 return ConstantInt::getFalse(Ty: ITy);
4028 }
4029 }
4030 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
4031 // too. If not, then try to deduce the result of the comparison.
4032 else if (match(V: RHS, P: m_ImmConstant())) {
4033 Constant *C = cast<Constant>(Val: RHS);
4034
4035 // Compute the constant that would happen if we truncated to SrcTy then
4036 // reextended to DstTy.
4037 Constant *Trunc =
4038 ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy: SrcTy, DL: Q.DL);
4039 assert(Trunc && "Constant-fold of ImmConstant should not fail");
4040 Constant *RExt =
4041 ConstantFoldCastOperand(Opcode: CastInst::SExt, C: Trunc, DestTy: DstTy, DL: Q.DL);
4042 assert(RExt && "Constant-fold of ImmConstant should not fail");
4043 Constant *AnyEq =
4044 ConstantFoldCompareInstOperands(Predicate: ICmpInst::ICMP_EQ, LHS: RExt, RHS: C, DL: Q.DL);
4045 assert(AnyEq && "Constant-fold of ImmConstant should not fail");
4046
4047 // If the re-extended constant didn't change then this is effectively
4048 // also a case of comparing two sign-extended values.
4049 if (AnyEq->isAllOnesValue() && MaxRecurse)
4050 if (Value *V =
4051 simplifyICmpInst(Pred, LHS: SrcOp, RHS: Trunc, Q, MaxRecurse: MaxRecurse - 1))
4052 return V;
4053
4054 // Otherwise the upper bits of LHS are all equal, while RHS has varying
4055 // bits there. Use this to work out the result of the comparison.
4056 if (AnyEq->isNullValue()) {
4057 switch (Pred) {
4058 default:
4059 llvm_unreachable("Unknown ICmp predicate!");
4060 case ICmpInst::ICMP_EQ:
4061 return Constant::getNullValue(Ty: ITy);
4062 case ICmpInst::ICMP_NE:
4063 return Constant::getAllOnesValue(Ty: ITy);
4064
4065 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
4066 // LHS >s RHS.
4067 case ICmpInst::ICMP_SGT:
4068 case ICmpInst::ICMP_SGE:
4069 return ConstantFoldCompareInstOperands(
4070 Predicate: ICmpInst::ICMP_SLT, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4071 DL: Q.DL);
4072 case ICmpInst::ICMP_SLT:
4073 case ICmpInst::ICMP_SLE:
4074 return ConstantFoldCompareInstOperands(
4075 Predicate: ICmpInst::ICMP_SGE, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4076 DL: Q.DL);
4077
4078 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
4079 // LHS >u RHS.
4080 case ICmpInst::ICMP_UGT:
4081 case ICmpInst::ICMP_UGE:
4082 // Comparison is true iff the LHS <s 0.
4083 if (MaxRecurse)
4084 if (Value *V = simplifyICmpInst(Pred: ICmpInst::ICMP_SLT, LHS: SrcOp,
4085 RHS: Constant::getNullValue(Ty: SrcTy), Q,
4086 MaxRecurse: MaxRecurse - 1))
4087 return V;
4088 break;
4089 case ICmpInst::ICMP_ULT:
4090 case ICmpInst::ICMP_ULE:
4091 // Comparison is true iff the LHS >=s 0.
4092 if (MaxRecurse)
4093 if (Value *V = simplifyICmpInst(Pred: ICmpInst::ICMP_SGE, LHS: SrcOp,
4094 RHS: Constant::getNullValue(Ty: SrcTy), Q,
4095 MaxRecurse: MaxRecurse - 1))
4096 return V;
4097 break;
4098 }
4099 }
4100 }
4101 }
4102 }
4103
4104 // icmp eq|ne X, Y -> false|true if X != Y
4105 // This is potentially expensive, and we have already computedKnownBits for
4106 // compares with 0 above here, so only try this for a non-zero compare.
4107 if (ICmpInst::isEquality(P: Pred) && !match(V: RHS, P: m_Zero()) &&
4108 isKnownNonEqual(V1: LHS, V2: RHS, SQ: Q)) {
4109 return Pred == ICmpInst::ICMP_NE ? getTrue(Ty: ITy) : getFalse(Ty: ITy);
4110 }
4111
4112 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
4113 return V;
4114
4115 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
4116 return V;
4117
4118 if (Value *V = simplifyICmpWithIntrinsicOnLHS(Pred, LHS, RHS))
4119 return V;
4120 if (Value *V = simplifyICmpWithIntrinsicOnLHS(
4121 Pred: ICmpInst::getSwappedPredicate(pred: Pred), LHS: RHS, RHS: LHS))
4122 return V;
4123
4124 if (Value *V = simplifyICmpUsingMonotonicValues(Pred, LHS, RHS, Q))
4125 return V;
4126 if (Value *V = simplifyICmpUsingMonotonicValues(
4127 Pred: ICmpInst::getSwappedPredicate(pred: Pred), LHS: RHS, RHS: LHS, Q))
4128 return V;
4129
4130 if (Value *V = simplifyICmpWithDominatingAssume(Predicate: Pred, LHS, RHS, Q))
4131 return V;
4132
4133 if (std::optional<bool> Res =
4134 isImpliedByDomCondition(Pred, LHS, RHS, ContextI: Q.CxtI, DL: Q.DL))
4135 return ConstantInt::getBool(Ty: ITy, V: *Res);
4136
4137 // Simplify comparisons of related pointers using a powerful, recursive
4138 // GEP-walk when we have target data available..
4139 if (LHS->getType()->isPointerTy())
4140 if (auto *C = computePointerICmp(Pred, LHS, RHS, Q))
4141 return C;
4142
4143 // If the comparison is with the result of a select instruction, check whether
4144 // comparing with either branch of the select always yields the same value.
4145 if (isa<SelectInst>(Val: LHS) || isa<SelectInst>(Val: RHS))
4146 if (Value *V = threadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
4147 return V;
4148
4149 // If the comparison is with the result of a phi instruction, check whether
4150 // doing the compare with each incoming phi value yields a common result.
4151 if (isa<PHINode>(Val: LHS) || isa<PHINode>(Val: RHS))
4152 if (Value *V = threadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
4153 return V;
4154
4155 return nullptr;
4156}
4157
4158Value *llvm::simplifyICmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
4159 const SimplifyQuery &Q) {
4160 return ::simplifyICmpInst(Pred: Predicate, LHS, RHS, Q, MaxRecurse: RecursionLimit);
4161}
4162
4163/// Given operands for an FCmpInst, see if we can fold the result.
4164/// If not, this returns null.
4165static Value *simplifyFCmpInst(CmpPredicate Pred, Value *LHS, Value *RHS,
4166 FastMathFlags FMF, const SimplifyQuery &Q,
4167 unsigned MaxRecurse) {
4168 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
4169
4170 if (Constant *CLHS = dyn_cast<Constant>(Val: LHS)) {
4171 if (Constant *CRHS = dyn_cast<Constant>(Val: RHS)) {
4172 // if the folding isn't successfull, fall back to the rest of the logic
4173 if (auto *Result = ConstantFoldCompareInstOperands(Predicate: Pred, LHS: CLHS, RHS: CRHS, DL: Q.DL,
4174 TLI: Q.TLI, I: Q.CxtI))
4175 return Result;
4176 } else {
4177 // If we have a constant, make sure it is on the RHS.
4178 std::swap(a&: LHS, b&: RHS);
4179 Pred = CmpInst::getSwappedPredicate(pred: Pred);
4180 }
4181 }
4182
4183 // Fold trivial predicates.
4184 Type *RetTy = getCompareTy(Op: LHS);
4185 if (Pred == FCmpInst::FCMP_FALSE)
4186 return getFalse(Ty: RetTy);
4187 if (Pred == FCmpInst::FCMP_TRUE)
4188 return getTrue(Ty: RetTy);
4189
4190 // fcmp pred x, poison and fcmp pred poison, x
4191 // fold to poison
4192 if (isa<PoisonValue>(Val: LHS) || isa<PoisonValue>(Val: RHS))
4193 return PoisonValue::get(T: RetTy);
4194
4195 // fcmp pred x, undef and fcmp pred undef, x
4196 // fold to true if unordered, false if ordered
4197 if (Q.isUndefValue(V: LHS) || Q.isUndefValue(V: RHS)) {
4198 // Choosing NaN for the undef will always make unordered comparison succeed
4199 // and ordered comparison fail.
4200 return ConstantInt::get(Ty: RetTy, V: CmpInst::isUnordered(predicate: Pred));
4201 }
4202
4203 // fcmp x,x -> true/false. Not all compares are foldable.
4204 if (LHS == RHS) {
4205 if (CmpInst::isTrueWhenEqual(predicate: Pred))
4206 return getTrue(Ty: RetTy);
4207 if (CmpInst::isFalseWhenEqual(predicate: Pred))
4208 return getFalse(Ty: RetTy);
4209 }
4210
4211 // Fold (un)ordered comparison if we can determine there are no NaNs.
4212 //
4213 // This catches the 2 variable input case, constants are handled below as a
4214 // class-like compare.
4215 if (Pred == FCmpInst::FCMP_ORD || Pred == FCmpInst::FCMP_UNO) {
4216 KnownFPClass RHSClass = computeKnownFPClass(V: RHS, InterestedClasses: fcAllFlags, SQ: Q);
4217 KnownFPClass LHSClass = computeKnownFPClass(V: LHS, InterestedClasses: fcAllFlags, SQ: Q);
4218
4219 if (FMF.noNaNs() ||
4220 (RHSClass.isKnownNeverNaN() && LHSClass.isKnownNeverNaN()))
4221 return ConstantInt::get(Ty: RetTy, V: Pred == FCmpInst::FCMP_ORD);
4222
4223 if (RHSClass.isKnownAlwaysNaN() || LHSClass.isKnownAlwaysNaN())
4224 return ConstantInt::get(Ty: RetTy, V: Pred == CmpInst::FCMP_UNO);
4225 }
4226
4227 if (std::optional<bool> Res =
4228 isImpliedByDomCondition(Pred, LHS, RHS, ContextI: Q.CxtI, DL: Q.DL))
4229 return ConstantInt::getBool(Ty: RetTy, V: *Res);
4230
4231 const APFloat *C = nullptr;
4232 match(V: RHS, P: m_APFloatAllowPoison(Res&: C));
4233 std::optional<KnownFPClass> FullKnownClassLHS;
4234
4235 // Lazily compute the possible classes for LHS. Avoid computing it twice if
4236 // RHS is a 0.
4237 auto computeLHSClass = [=, &FullKnownClassLHS](FPClassTest InterestedFlags =
4238 fcAllFlags) {
4239 if (FullKnownClassLHS)
4240 return *FullKnownClassLHS;
4241 return computeKnownFPClass(V: LHS, FMF, InterestedClasses: InterestedFlags, SQ: Q);
4242 };
4243
4244 if (C && Q.CxtI) {
4245 // Fold out compares that express a class test.
4246 //
4247 // FIXME: Should be able to perform folds without context
4248 // instruction. Always pass in the context function?
4249
4250 const Function *ParentF = Q.CxtI->getFunction();
4251 auto [ClassVal, ClassTest] = fcmpToClassTest(Pred, F: *ParentF, LHS, ConstRHS: C);
4252 if (ClassVal) {
4253 FullKnownClassLHS = computeLHSClass();
4254 if ((FullKnownClassLHS->KnownFPClasses & ClassTest) == fcNone)
4255 return getFalse(Ty: RetTy);
4256 if ((FullKnownClassLHS->KnownFPClasses & ~ClassTest) == fcNone)
4257 return getTrue(Ty: RetTy);
4258 }
4259 }
4260
4261 // Handle fcmp with constant RHS.
4262 if (C) {
4263 // TODO: If we always required a context function, we wouldn't need to
4264 // special case nans.
4265 if (C->isNaN())
4266 return ConstantInt::get(Ty: RetTy, V: CmpInst::isUnordered(predicate: Pred));
4267
4268 // TODO: Need version fcmpToClassTest which returns implied class when the
4269 // compare isn't a complete class test. e.g. > 1.0 implies fcPositive, but
4270 // isn't implementable as a class call.
4271 if (C->isNegative() && !C->isNegZero()) {
4272 FPClassTest Interested = KnownFPClass::OrderedLessThanZeroMask;
4273
4274 // TODO: We can catch more cases by using a range check rather than
4275 // relying on CannotBeOrderedLessThanZero.
4276 switch (Pred) {
4277 case FCmpInst::FCMP_UGE:
4278 case FCmpInst::FCMP_UGT:
4279 case FCmpInst::FCMP_UNE: {
4280 KnownFPClass KnownClass = computeLHSClass(Interested);
4281
4282 // (X >= 0) implies (X > C) when (C < 0)
4283 if (KnownClass.cannotBeOrderedLessThanZero())
4284 return getTrue(Ty: RetTy);
4285 break;
4286 }
4287 case FCmpInst::FCMP_OEQ:
4288 case FCmpInst::FCMP_OLE:
4289 case FCmpInst::FCMP_OLT: {
4290 KnownFPClass KnownClass = computeLHSClass(Interested);
4291
4292 // (X >= 0) implies !(X < C) when (C < 0)
4293 if (KnownClass.cannotBeOrderedLessThanZero())
4294 return getFalse(Ty: RetTy);
4295 break;
4296 }
4297 default:
4298 break;
4299 }
4300 }
4301 // Check FCmp of [min/maxnum or min/maximumnum with const] with other const.
4302 const APFloat *C2;
4303 bool IsMax = match(V: LHS, P: m_FMaxNum_or_FMaximumNum(Op0: m_Value(), Op1: m_APFloat(Res&: C2)));
4304 bool IsMin = match(V: LHS, P: m_FMinNum_or_FMinimumNum(Op0: m_Value(), Op1: m_APFloat(Res&: C2)));
4305 if ((IsMax && *C2 > *C) || (IsMin && *C2 < *C)) {
4306 // The ordered relationship and min/maxnum or min/maximumnum guarantee
4307 // that we do not have NaN constants, so ordered/unordered preds are
4308 // handled the same.
4309 switch (Pred) {
4310 case FCmpInst::FCMP_OEQ:
4311 case FCmpInst::FCMP_UEQ:
4312 // minnum(X, LesserC) == C --> false
4313 // maxnum(X, GreaterC) == C --> false
4314 return getFalse(Ty: RetTy);
4315 case FCmpInst::FCMP_ONE:
4316 case FCmpInst::FCMP_UNE:
4317 // minnum(X, LesserC) != C --> true
4318 // maxnum(X, GreaterC) != C --> true
4319 return getTrue(Ty: RetTy);
4320 case FCmpInst::FCMP_OGE:
4321 case FCmpInst::FCMP_UGE:
4322 case FCmpInst::FCMP_OGT:
4323 case FCmpInst::FCMP_UGT:
4324 // minnum(X, LesserC) >= C --> false
4325 // minnum(X, LesserC) > C --> false
4326 // maxnum(X, GreaterC) >= C --> true
4327 // maxnum(X, GreaterC) > C --> true
4328 return ConstantInt::get(Ty: RetTy, V: IsMax);
4329 case FCmpInst::FCMP_OLE:
4330 case FCmpInst::FCMP_ULE:
4331 case FCmpInst::FCMP_OLT:
4332 case FCmpInst::FCMP_ULT:
4333 // minnum(X, LesserC) <= C --> true
4334 // minnum(X, LesserC) < C --> true
4335 // maxnum(X, GreaterC) <= C --> false
4336 // maxnum(X, GreaterC) < C --> false
4337 return ConstantInt::get(Ty: RetTy, V: !IsMax);
4338 default:
4339 // TRUE/FALSE/ORD/UNO should be handled before this.
4340 llvm_unreachable("Unexpected fcmp predicate");
4341 }
4342 }
4343 }
4344
4345 // TODO: Could fold this with above if there were a matcher which returned all
4346 // classes in a non-splat vector.
4347 if (match(V: RHS, P: m_AnyZeroFP())) {
4348 switch (Pred) {
4349 case FCmpInst::FCMP_OGE:
4350 case FCmpInst::FCMP_ULT: {
4351 FPClassTest Interested = KnownFPClass::OrderedLessThanZeroMask;
4352 if (!FMF.noNaNs())
4353 Interested |= fcNan;
4354
4355 KnownFPClass Known = computeLHSClass(Interested);
4356
4357 // Positive or zero X >= 0.0 --> true
4358 // Positive or zero X < 0.0 --> false
4359 if ((FMF.noNaNs() || Known.isKnownNeverNaN()) &&
4360 Known.cannotBeOrderedLessThanZero())
4361 return Pred == FCmpInst::FCMP_OGE ? getTrue(Ty: RetTy) : getFalse(Ty: RetTy);
4362 break;
4363 }
4364 case FCmpInst::FCMP_UGE:
4365 case FCmpInst::FCMP_OLT: {
4366 FPClassTest Interested = KnownFPClass::OrderedLessThanZeroMask;
4367 KnownFPClass Known = computeLHSClass(Interested);
4368
4369 // Positive or zero or nan X >= 0.0 --> true
4370 // Positive or zero or nan X < 0.0 --> false
4371 if (Known.cannotBeOrderedLessThanZero())
4372 return Pred == FCmpInst::FCMP_UGE ? getTrue(Ty: RetTy) : getFalse(Ty: RetTy);
4373 break;
4374 }
4375 default:
4376 break;
4377 }
4378 }
4379
4380 // If the comparison is with the result of a select instruction, check whether
4381 // comparing with either branch of the select always yields the same value.
4382 if (isa<SelectInst>(Val: LHS) || isa<SelectInst>(Val: RHS))
4383 if (Value *V = threadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
4384 return V;
4385
4386 // If the comparison is with the result of a phi instruction, check whether
4387 // doing the compare with each incoming phi value yields a common result.
4388 if (isa<PHINode>(Val: LHS) || isa<PHINode>(Val: RHS))
4389 if (Value *V = threadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
4390 return V;
4391
4392 return nullptr;
4393}
4394
4395Value *llvm::simplifyFCmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
4396 FastMathFlags FMF, const SimplifyQuery &Q) {
4397 return ::simplifyFCmpInst(Pred: Predicate, LHS, RHS, FMF, Q, MaxRecurse: RecursionLimit);
4398}
4399
4400static Value *simplifyWithOpsReplaced(Value *V,
4401 ArrayRef<std::pair<Value *, Value *>> Ops,
4402 const SimplifyQuery &Q,
4403 bool AllowRefinement,
4404 SmallVectorImpl<Instruction *> *DropFlags,
4405 unsigned MaxRecurse) {
4406 assert((AllowRefinement || !Q.CanUseUndef) &&
4407 "If AllowRefinement=false then CanUseUndef=false");
4408 for (const auto &OpAndRepOp : Ops) {
4409 // We cannot replace a constant, and shouldn't even try.
4410 if (isa<Constant>(Val: OpAndRepOp.first))
4411 return nullptr;
4412
4413 // Trivial replacement.
4414 if (V == OpAndRepOp.first)
4415 return OpAndRepOp.second;
4416 }
4417
4418 if (!MaxRecurse--)
4419 return nullptr;
4420
4421 auto *I = dyn_cast<Instruction>(Val: V);
4422 if (!I)
4423 return nullptr;
4424
4425 // The arguments of a phi node might refer to a value from a previous
4426 // cycle iteration.
4427 if (isa<PHINode>(Val: I))
4428 return nullptr;
4429
4430 // Don't fold away llvm.is.constant checks based on assumptions.
4431 if (match(V: I, P: m_Intrinsic<Intrinsic::is_constant>()))
4432 return nullptr;
4433
4434 // Don't simplify freeze.
4435 if (isa<FreezeInst>(Val: I))
4436 return nullptr;
4437
4438 for (const auto &OpAndRepOp : Ops) {
4439 // For vector types, the simplification must hold per-lane, so forbid
4440 // potentially cross-lane operations like shufflevector.
4441 if (OpAndRepOp.first->getType()->isVectorTy() &&
4442 !isNotCrossLaneOperation(I))
4443 return nullptr;
4444 }
4445
4446 // Replace Op with RepOp in instruction operands.
4447 SmallVector<Value *, 8> NewOps;
4448 bool AnyReplaced = false;
4449 for (Value *InstOp : I->operands()) {
4450 if (Value *NewInstOp = simplifyWithOpsReplaced(
4451 V: InstOp, Ops, Q, AllowRefinement, DropFlags, MaxRecurse)) {
4452 NewOps.push_back(Elt: NewInstOp);
4453 AnyReplaced = InstOp != NewInstOp;
4454 } else {
4455 NewOps.push_back(Elt: InstOp);
4456 }
4457
4458 // Bail out if any operand is undef and SimplifyQuery disables undef
4459 // simplification. Constant folding currently doesn't respect this option.
4460 if (isa<UndefValue>(Val: NewOps.back()) && !Q.CanUseUndef)
4461 return nullptr;
4462 }
4463
4464 if (!AnyReplaced)
4465 return nullptr;
4466
4467 if (!AllowRefinement) {
4468 // General InstSimplify functions may refine the result, e.g. by returning
4469 // a constant for a potentially poison value. To avoid this, implement only
4470 // a few non-refining but profitable transforms here.
4471
4472 if (auto *BO = dyn_cast<BinaryOperator>(Val: I)) {
4473 unsigned Opcode = BO->getOpcode();
4474 // id op x -> x, x op id -> x
4475 // Exclude floats, because x op id may produce a different NaN value.
4476 if (!BO->getType()->isFPOrFPVectorTy()) {
4477 if (NewOps[0] == ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType()))
4478 return NewOps[1];
4479 if (NewOps[1] == ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType(),
4480 /* RHS */ AllowRHSConstant: true))
4481 return NewOps[0];
4482 }
4483
4484 // x & x -> x, x | x -> x
4485 if ((Opcode == Instruction::And || Opcode == Instruction::Or) &&
4486 NewOps[0] == NewOps[1]) {
4487 // or disjoint x, x results in poison.
4488 if (auto *PDI = dyn_cast<PossiblyDisjointInst>(Val: BO)) {
4489 if (PDI->isDisjoint()) {
4490 if (!DropFlags)
4491 return nullptr;
4492 DropFlags->push_back(Elt: BO);
4493 }
4494 }
4495 return NewOps[0];
4496 }
4497
4498 // x - x -> 0, x ^ x -> 0. This is non-refining, because x is non-poison
4499 // by assumption and this case never wraps, so nowrap flags can be
4500 // ignored.
4501 if ((Opcode == Instruction::Sub || Opcode == Instruction::Xor) &&
4502 NewOps[0] == NewOps[1] &&
4503 any_of(Range&: Ops, P: [=](const auto &Rep) { return NewOps[0] == Rep.second; }))
4504 return Constant::getNullValue(Ty: I->getType());
4505
4506 // If we are substituting an absorber constant into a binop and extra
4507 // poison can't leak if we remove the select -- because both operands of
4508 // the binop are based on the same value -- then it may be safe to replace
4509 // the value with the absorber constant. Examples:
4510 // (Op == 0) ? 0 : (Op & -Op) --> Op & -Op
4511 // (Op == 0) ? 0 : (Op * (binop Op, C)) --> Op * (binop Op, C)
4512 // (Op == -1) ? -1 : (Op | (binop C, Op) --> Op | (binop C, Op)
4513 Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, Ty: I->getType());
4514 if ((NewOps[0] == Absorber || NewOps[1] == Absorber) &&
4515 any_of(Range&: Ops,
4516 P: [=](const auto &Rep) { return impliesPoison(BO, Rep.first); }))
4517 return Absorber;
4518 }
4519
4520 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
4521 // `x == y ? 0 : ucmp(x, y)` where under the replacement y -> x,
4522 // `ucmp(x, x)` becomes `0`.
4523 if ((II->getIntrinsicID() == Intrinsic::scmp ||
4524 II->getIntrinsicID() == Intrinsic::ucmp) &&
4525 NewOps[0] == NewOps[1]) {
4526 if (II->hasPoisonGeneratingAnnotations()) {
4527 if (!DropFlags)
4528 return nullptr;
4529
4530 DropFlags->push_back(Elt: II);
4531 }
4532
4533 return ConstantInt::get(Ty: I->getType(), V: 0);
4534 }
4535 }
4536
4537 if (isa<GetElementPtrInst>(Val: I)) {
4538 // getelementptr x, 0 -> x.
4539 // This never returns poison, even if inbounds is set.
4540 if (NewOps.size() == 2 && match(V: NewOps[1], P: m_Zero()))
4541 return NewOps[0];
4542 }
4543 } else {
4544 // The simplification queries below may return the original value. Consider:
4545 // %div = udiv i32 %arg, %arg2
4546 // %mul = mul nsw i32 %div, %arg2
4547 // %cmp = icmp eq i32 %mul, %arg
4548 // %sel = select i1 %cmp, i32 %div, i32 undef
4549 // Replacing %arg by %mul, %div becomes "udiv i32 %mul, %arg2", which
4550 // simplifies back to %arg. This can only happen because %mul does not
4551 // dominate %div. To ensure a consistent return value contract, we make sure
4552 // that this case returns nullptr as well.
4553 auto PreventSelfSimplify = [V](Value *Simplified) {
4554 return Simplified != V ? Simplified : nullptr;
4555 };
4556
4557 return PreventSelfSimplify(
4558 ::simplifyInstructionWithOperands(I, NewOps, SQ: Q, MaxRecurse));
4559 }
4560
4561 // If all operands are constant after substituting Op for RepOp then we can
4562 // constant fold the instruction.
4563 SmallVector<Constant *, 8> ConstOps;
4564 for (Value *NewOp : NewOps) {
4565 if (Constant *ConstOp = dyn_cast<Constant>(Val: NewOp))
4566 ConstOps.push_back(Elt: ConstOp);
4567 else
4568 return nullptr;
4569 }
4570
4571 // Consider:
4572 // %cmp = icmp eq i32 %x, 2147483647
4573 // %add = add nsw i32 %x, 1
4574 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
4575 //
4576 // We can't replace %sel with %add unless we strip away the flags (which
4577 // will be done in InstCombine).
4578 // TODO: This may be unsound, because it only catches some forms of
4579 // refinement.
4580 if (!AllowRefinement) {
4581 auto *II = dyn_cast<IntrinsicInst>(Val: I);
4582 if (canCreatePoison(Op: cast<Operator>(Val: I), ConsiderFlagsAndMetadata: !DropFlags)) {
4583 // abs cannot create poison if the value is known to never be int_min.
4584 if (II && II->getIntrinsicID() == Intrinsic::abs) {
4585 if (!ConstOps[0]->isNotMinSignedValue())
4586 return nullptr;
4587 } else
4588 return nullptr;
4589 }
4590
4591 if (DropFlags && II) {
4592 // If we're going to change the poison flag of abs/ctz to false, also
4593 // perform constant folding that way, so we get an integer instead of a
4594 // poison value here.
4595 switch (II->getIntrinsicID()) {
4596 case Intrinsic::abs:
4597 case Intrinsic::ctlz:
4598 case Intrinsic::cttz:
4599 ConstOps[1] = ConstantInt::getFalse(Context&: I->getContext());
4600 break;
4601 default:
4602 break;
4603 }
4604 }
4605
4606 Constant *Res = ConstantFoldInstOperands(I, Ops: ConstOps, DL: Q.DL, TLI: Q.TLI,
4607 /*AllowNonDeterministic=*/false);
4608 if (DropFlags && Res && I->hasPoisonGeneratingAnnotations())
4609 DropFlags->push_back(Elt: I);
4610 return Res;
4611 }
4612
4613 return ConstantFoldInstOperands(I, Ops: ConstOps, DL: Q.DL, TLI: Q.TLI,
4614 /*AllowNonDeterministic=*/false);
4615}
4616
4617static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
4618 const SimplifyQuery &Q,
4619 bool AllowRefinement,
4620 SmallVectorImpl<Instruction *> *DropFlags,
4621 unsigned MaxRecurse) {
4622 return simplifyWithOpsReplaced(V, Ops: {{Op, RepOp}}, Q, AllowRefinement,
4623 DropFlags, MaxRecurse);
4624}
4625
4626Value *llvm::simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
4627 const SimplifyQuery &Q,
4628 bool AllowRefinement,
4629 SmallVectorImpl<Instruction *> *DropFlags) {
4630 // If refinement is disabled, also disable undef simplifications (which are
4631 // always refinements) in SimplifyQuery.
4632 if (!AllowRefinement)
4633 return ::simplifyWithOpReplaced(V, Op, RepOp, Q: Q.getWithoutUndef(),
4634 AllowRefinement, DropFlags, MaxRecurse: RecursionLimit);
4635 return ::simplifyWithOpReplaced(V, Op, RepOp, Q, AllowRefinement, DropFlags,
4636 MaxRecurse: RecursionLimit);
4637}
4638
4639/// Try to simplify a select instruction when its condition operand is an
4640/// integer comparison where one operand of the compare is a constant.
4641static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
4642 const APInt *Y, bool TrueWhenUnset) {
4643 const APInt *C;
4644
4645 // (X & Y) == 0 ? X & ~Y : X --> X
4646 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
4647 if (FalseVal == X && match(V: TrueVal, P: m_And(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4648 *Y == ~*C)
4649 return TrueWhenUnset ? FalseVal : TrueVal;
4650
4651 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
4652 // (X & Y) != 0 ? X : X & ~Y --> X
4653 if (TrueVal == X && match(V: FalseVal, P: m_And(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4654 *Y == ~*C)
4655 return TrueWhenUnset ? FalseVal : TrueVal;
4656
4657 if (Y->isPowerOf2()) {
4658 // (X & Y) == 0 ? X | Y : X --> X | Y
4659 // (X & Y) != 0 ? X | Y : X --> X
4660 if (FalseVal == X && match(V: TrueVal, P: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4661 *Y == *C) {
4662 // We can't return the or if it has the disjoint flag.
4663 if (TrueWhenUnset && cast<PossiblyDisjointInst>(Val: TrueVal)->isDisjoint())
4664 return nullptr;
4665 return TrueWhenUnset ? TrueVal : FalseVal;
4666 }
4667
4668 // (X & Y) == 0 ? X : X | Y --> X
4669 // (X & Y) != 0 ? X : X | Y --> X | Y
4670 if (TrueVal == X && match(V: FalseVal, P: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4671 *Y == *C) {
4672 // We can't return the or if it has the disjoint flag.
4673 if (!TrueWhenUnset && cast<PossiblyDisjointInst>(Val: FalseVal)->isDisjoint())
4674 return nullptr;
4675 return TrueWhenUnset ? TrueVal : FalseVal;
4676 }
4677 }
4678
4679 return nullptr;
4680}
4681
4682static Value *simplifyCmpSelOfMaxMin(Value *CmpLHS, Value *CmpRHS,
4683 CmpPredicate Pred, Value *TVal,
4684 Value *FVal) {
4685 // Canonicalize common cmp+sel operand as CmpLHS.
4686 if (CmpRHS == TVal || CmpRHS == FVal) {
4687 std::swap(a&: CmpLHS, b&: CmpRHS);
4688 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
4689 }
4690
4691 // Canonicalize common cmp+sel operand as TVal.
4692 if (CmpLHS == FVal) {
4693 std::swap(a&: TVal, b&: FVal);
4694 Pred = ICmpInst::getInversePredicate(pred: Pred);
4695 }
4696
4697 // A vector select may be shuffling together elements that are equivalent
4698 // based on the max/min/select relationship.
4699 Value *X = CmpLHS, *Y = CmpRHS;
4700 bool PeekedThroughSelectShuffle = false;
4701 auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: FVal);
4702 if (Shuf && Shuf->isSelect()) {
4703 if (Shuf->getOperand(i_nocapture: 0) == Y)
4704 FVal = Shuf->getOperand(i_nocapture: 1);
4705 else if (Shuf->getOperand(i_nocapture: 1) == Y)
4706 FVal = Shuf->getOperand(i_nocapture: 0);
4707 else
4708 return nullptr;
4709 PeekedThroughSelectShuffle = true;
4710 }
4711
4712 // (X pred Y) ? X : max/min(X, Y)
4713 auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: FVal);
4714 if (!MMI || TVal != X ||
4715 !match(V: FVal, P: m_c_MaxOrMin(L: m_Specific(V: X), R: m_Specific(V: Y))))
4716 return nullptr;
4717
4718 // (X > Y) ? X : max(X, Y) --> max(X, Y)
4719 // (X >= Y) ? X : max(X, Y) --> max(X, Y)
4720 // (X < Y) ? X : min(X, Y) --> min(X, Y)
4721 // (X <= Y) ? X : min(X, Y) --> min(X, Y)
4722 //
4723 // The equivalence allows a vector select (shuffle) of max/min and Y. Ex:
4724 // (X > Y) ? X : (Z ? max(X, Y) : Y)
4725 // If Z is true, this reduces as above, and if Z is false:
4726 // (X > Y) ? X : Y --> max(X, Y)
4727 ICmpInst::Predicate MMPred = MMI->getPredicate();
4728 if (MMPred == CmpInst::getStrictPredicate(pred: Pred))
4729 return MMI;
4730
4731 // Other transforms are not valid with a shuffle.
4732 if (PeekedThroughSelectShuffle)
4733 return nullptr;
4734
4735 // (X == Y) ? X : max/min(X, Y) --> max/min(X, Y)
4736 if (Pred == CmpInst::ICMP_EQ)
4737 return MMI;
4738
4739 // (X != Y) ? X : max/min(X, Y) --> X
4740 if (Pred == CmpInst::ICMP_NE)
4741 return X;
4742
4743 // (X < Y) ? X : max(X, Y) --> X
4744 // (X <= Y) ? X : max(X, Y) --> X
4745 // (X > Y) ? X : min(X, Y) --> X
4746 // (X >= Y) ? X : min(X, Y) --> X
4747 ICmpInst::Predicate InvPred = CmpInst::getInversePredicate(pred: Pred);
4748 if (MMPred == CmpInst::getStrictPredicate(pred: InvPred))
4749 return X;
4750
4751 return nullptr;
4752}
4753
4754/// An alternative way to test if a bit is set or not.
4755/// uses e.g. sgt/slt or trunc instead of eq/ne.
4756static Value *simplifySelectWithBitTest(Value *CondVal, Value *TrueVal,
4757 Value *FalseVal) {
4758 if (auto Res = decomposeBitTest(Cond: CondVal))
4759 return simplifySelectBitTest(TrueVal, FalseVal, X: Res->X, Y: &Res->Mask,
4760 TrueWhenUnset: Res->Pred == ICmpInst::ICMP_EQ);
4761
4762 return nullptr;
4763}
4764
4765/// Try to simplify a select instruction when its condition operand is an
4766/// integer equality or floating-point equivalence comparison.
4767static Value *simplifySelectWithEquivalence(
4768 ArrayRef<std::pair<Value *, Value *>> Replacements, Value *TrueVal,
4769 Value *FalseVal, const SimplifyQuery &Q, unsigned MaxRecurse) {
4770 Value *SimplifiedFalseVal =
4771 simplifyWithOpsReplaced(V: FalseVal, Ops: Replacements, Q: Q.getWithoutUndef(),
4772 /* AllowRefinement */ false,
4773 /* DropFlags */ nullptr, MaxRecurse);
4774 if (!SimplifiedFalseVal)
4775 SimplifiedFalseVal = FalseVal;
4776
4777 Value *SimplifiedTrueVal =
4778 simplifyWithOpsReplaced(V: TrueVal, Ops: Replacements, Q,
4779 /* AllowRefinement */ true,
4780 /* DropFlags */ nullptr, MaxRecurse);
4781 if (!SimplifiedTrueVal)
4782 SimplifiedTrueVal = TrueVal;
4783
4784 if (SimplifiedFalseVal == SimplifiedTrueVal)
4785 return FalseVal;
4786
4787 return nullptr;
4788}
4789
4790/// Try to simplify a select instruction when its condition operand is an
4791/// integer comparison.
4792static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
4793 Value *FalseVal,
4794 const SimplifyQuery &Q,
4795 unsigned MaxRecurse) {
4796 CmpPredicate Pred;
4797 Value *CmpLHS, *CmpRHS;
4798 if (!match(V: CondVal, P: m_ICmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS))))
4799 return nullptr;
4800
4801 if (Value *V = simplifyCmpSelOfMaxMin(CmpLHS, CmpRHS, Pred, TVal: TrueVal, FVal: FalseVal))
4802 return V;
4803
4804 // Canonicalize ne to eq predicate.
4805 if (Pred == ICmpInst::ICMP_NE) {
4806 Pred = ICmpInst::ICMP_EQ;
4807 std::swap(a&: TrueVal, b&: FalseVal);
4808 }
4809
4810 // Check for integer min/max with a limit constant:
4811 // X > MIN_INT ? X : MIN_INT --> X
4812 // X < MAX_INT ? X : MAX_INT --> X
4813 if (TrueVal->getType()->isIntOrIntVectorTy()) {
4814 Value *X, *Y;
4815 SelectPatternFlavor SPF =
4816 matchDecomposedSelectPattern(CmpI: cast<ICmpInst>(Val: CondVal), TrueVal, FalseVal,
4817 LHS&: X, RHS&: Y)
4818 .Flavor;
4819 if (SelectPatternResult::isMinOrMax(SPF) && Pred == getMinMaxPred(SPF)) {
4820 APInt LimitC = getMinMaxLimit(SPF: getInverseMinMaxFlavor(SPF),
4821 BitWidth: X->getType()->getScalarSizeInBits());
4822 if (match(V: Y, P: m_SpecificInt(V: LimitC)))
4823 return X;
4824 }
4825 }
4826
4827 if (Pred == ICmpInst::ICMP_EQ && match(V: CmpRHS, P: m_Zero())) {
4828 Value *X;
4829 const APInt *Y;
4830 if (match(V: CmpLHS, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: Y))))
4831 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
4832 /*TrueWhenUnset=*/true))
4833 return V;
4834
4835 // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
4836 Value *ShAmt;
4837 auto isFsh = m_CombineOr(L: m_FShl(Op0: m_Value(V&: X), Op1: m_Value(), Op2: m_Value(V&: ShAmt)),
4838 R: m_FShr(Op0: m_Value(), Op1: m_Value(V&: X), Op2: m_Value(V&: ShAmt)));
4839 // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
4840 // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
4841 if (match(V: TrueVal, P: isFsh) && FalseVal == X && CmpLHS == ShAmt)
4842 return X;
4843
4844 // Test for a zero-shift-guard-op around rotates. These are used to
4845 // avoid UB from oversized shifts in raw IR rotate patterns, but the
4846 // intrinsics do not have that problem.
4847 // We do not allow this transform for the general funnel shift case because
4848 // that would not preserve the poison safety of the original code.
4849 auto isRotate =
4850 m_CombineOr(L: m_FShl(Op0: m_Value(V&: X), Op1: m_Deferred(V: X), Op2: m_Value(V&: ShAmt)),
4851 R: m_FShr(Op0: m_Value(V&: X), Op1: m_Deferred(V: X), Op2: m_Value(V&: ShAmt)));
4852 // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
4853 // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
4854 if (match(V: FalseVal, P: isRotate) && TrueVal == X && CmpLHS == ShAmt &&
4855 Pred == ICmpInst::ICMP_EQ)
4856 return FalseVal;
4857
4858 // X == 0 ? abs(X) : -abs(X) --> -abs(X)
4859 // X == 0 ? -abs(X) : abs(X) --> abs(X)
4860 if (match(V: TrueVal, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Specific(V: CmpLHS))) &&
4861 match(V: FalseVal, P: m_Neg(V: m_Intrinsic<Intrinsic::abs>(Op0: m_Specific(V: CmpLHS)))))
4862 return FalseVal;
4863 if (match(V: TrueVal,
4864 P: m_Neg(V: m_Intrinsic<Intrinsic::abs>(Op0: m_Specific(V: CmpLHS)))) &&
4865 match(V: FalseVal, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Specific(V: CmpLHS))))
4866 return FalseVal;
4867 }
4868
4869 // If we have a scalar equality comparison, then we know the value in one of
4870 // the arms of the select. See if substituting this value into the arm and
4871 // simplifying the result yields the same value as the other arm.
4872 if (Pred == ICmpInst::ICMP_EQ) {
4873 if (CmpLHS->getType()->isIntOrIntVectorTy() ||
4874 canReplacePointersIfEqual(From: CmpLHS, To: CmpRHS, DL: Q.DL))
4875 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpLHS, CmpRHS}}, TrueVal,
4876 FalseVal, Q, MaxRecurse))
4877 return V;
4878 if (CmpLHS->getType()->isIntOrIntVectorTy() ||
4879 canReplacePointersIfEqual(From: CmpRHS, To: CmpLHS, DL: Q.DL))
4880 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpRHS, CmpLHS}}, TrueVal,
4881 FalseVal, Q, MaxRecurse))
4882 return V;
4883
4884 Value *X;
4885 Value *Y;
4886 // select((X | Y) == 0 ? X : 0) --> 0 (commuted 2 ways)
4887 if (match(V: CmpLHS, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
4888 match(V: CmpRHS, P: m_Zero())) {
4889 // (X | Y) == 0 implies X == 0 and Y == 0.
4890 if (Value *V = simplifySelectWithEquivalence(
4891 Replacements: {{X, CmpRHS}, {Y, CmpRHS}}, TrueVal, FalseVal, Q, MaxRecurse))
4892 return V;
4893 }
4894
4895 // select((X & Y) == -1 ? X : -1) --> -1 (commuted 2 ways)
4896 if (match(V: CmpLHS, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
4897 match(V: CmpRHS, P: m_AllOnes())) {
4898 // (X & Y) == -1 implies X == -1 and Y == -1.
4899 if (Value *V = simplifySelectWithEquivalence(
4900 Replacements: {{X, CmpRHS}, {Y, CmpRHS}}, TrueVal, FalseVal, Q, MaxRecurse))
4901 return V;
4902 }
4903 }
4904
4905 return nullptr;
4906}
4907
4908/// Try to simplify a select instruction when its condition operand is a
4909/// floating-point comparison.
4910static Value *simplifySelectWithFCmp(Value *Cond, Value *T, Value *F,
4911 const SimplifyQuery &Q,
4912 unsigned MaxRecurse) {
4913 CmpPredicate Pred;
4914 Value *CmpLHS, *CmpRHS;
4915 if (!match(V: Cond, P: m_FCmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS))))
4916 return nullptr;
4917 FCmpInst *I = cast<FCmpInst>(Val: Cond);
4918
4919 bool IsEquiv = I->isEquivalence();
4920 if (I->isEquivalence(/*Invert=*/true)) {
4921 std::swap(a&: T, b&: F);
4922 Pred = FCmpInst::getInversePredicate(pred: Pred);
4923 IsEquiv = true;
4924 }
4925
4926 // This transforms is safe if at least one operand is known to not be zero.
4927 // Otherwise, the select can change the sign of a zero operand.
4928 if (IsEquiv) {
4929 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpLHS, CmpRHS}}, TrueVal: T, FalseVal: F, Q,
4930 MaxRecurse))
4931 return V;
4932 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpRHS, CmpLHS}}, TrueVal: T, FalseVal: F, Q,
4933 MaxRecurse))
4934 return V;
4935 }
4936
4937 // Canonicalize CmpLHS to be T, and CmpRHS to be F, if they're swapped.
4938 if (CmpLHS == F && CmpRHS == T)
4939 std::swap(a&: CmpLHS, b&: CmpRHS);
4940
4941 if (CmpLHS != T || CmpRHS != F)
4942 return nullptr;
4943
4944 // This transform is also safe if we do not have (do not care about) -0.0.
4945 if (Q.CxtI && isa<FPMathOperator>(Val: Q.CxtI) && Q.CxtI->hasNoSignedZeros()) {
4946 // (T == F) ? T : F --> F
4947 if (Pred == FCmpInst::FCMP_OEQ)
4948 return F;
4949
4950 // (T != F) ? T : F --> T
4951 if (Pred == FCmpInst::FCMP_UNE)
4952 return T;
4953 }
4954
4955 return nullptr;
4956}
4957
4958/// Look for the following pattern and simplify %to_fold to %identicalPhi.
4959/// Here %phi, %to_fold and %phi.next perform the same functionality as
4960/// %identicalPhi and hence the select instruction %to_fold can be folded
4961/// into %identicalPhi.
4962///
4963/// BB1:
4964/// %identicalPhi = phi [ X, %BB0 ], [ %identicalPhi.next, %BB1 ]
4965/// %phi = phi [ X, %BB0 ], [ %phi.next, %BB1 ]
4966/// ...
4967/// %identicalPhi.next = select %cmp, %val, %identicalPhi
4968/// (or select %cmp, %identicalPhi, %val)
4969/// %to_fold = select %cmp2, %identicalPhi, %phi
4970/// %phi.next = select %cmp, %val, %to_fold
4971/// (or select %cmp, %to_fold, %val)
4972///
4973/// Prove that %phi and %identicalPhi are the same by induction:
4974///
4975/// Base case: Both %phi and %identicalPhi are equal on entry to the loop.
4976/// Inductive case:
4977/// Suppose %phi and %identicalPhi are equal at iteration i.
4978/// We look at their values at iteration i+1 which are %phi.next and
4979/// %identicalPhi.next. They would have become different only when %cmp is
4980/// false and the corresponding values %to_fold and %identicalPhi differ
4981/// (similar reason for the other "or" case in the bracket).
4982///
4983/// The only condition when %to_fold and %identicalPh could differ is when %cmp2
4984/// is false and %to_fold is %phi, which contradicts our inductive hypothesis
4985/// that %phi and %identicalPhi are equal. Thus %phi and %identicalPhi are
4986/// always equal at iteration i+1.
4987bool isSelectWithIdenticalPHI(PHINode &PN, PHINode &IdenticalPN) {
4988 if (PN.getParent() != IdenticalPN.getParent())
4989 return false;
4990 if (PN.getNumIncomingValues() != 2)
4991 return false;
4992
4993 // Check that only the backedge incoming value is different.
4994 unsigned DiffVals = 0;
4995 BasicBlock *DiffValBB = nullptr;
4996 for (unsigned i = 0; i < 2; i++) {
4997 BasicBlock *PredBB = PN.getIncomingBlock(i);
4998 if (PN.getIncomingValue(i) !=
4999 IdenticalPN.getIncomingValueForBlock(BB: PredBB)) {
5000 DiffVals++;
5001 DiffValBB = PredBB;
5002 }
5003 }
5004 if (DiffVals != 1)
5005 return false;
5006 // Now check that the backedge incoming values are two select
5007 // instructions with the same condition. Either their true
5008 // values are the same, or their false values are the same.
5009 auto *SI = dyn_cast<SelectInst>(Val: PN.getIncomingValueForBlock(BB: DiffValBB));
5010 auto *IdenticalSI =
5011 dyn_cast<SelectInst>(Val: IdenticalPN.getIncomingValueForBlock(BB: DiffValBB));
5012 if (!SI || !IdenticalSI)
5013 return false;
5014 if (SI->getCondition() != IdenticalSI->getCondition())
5015 return false;
5016
5017 SelectInst *SIOtherVal = nullptr;
5018 Value *IdenticalSIOtherVal = nullptr;
5019 if (SI->getTrueValue() == IdenticalSI->getTrueValue()) {
5020 SIOtherVal = dyn_cast<SelectInst>(Val: SI->getFalseValue());
5021 IdenticalSIOtherVal = IdenticalSI->getFalseValue();
5022 } else if (SI->getFalseValue() == IdenticalSI->getFalseValue()) {
5023 SIOtherVal = dyn_cast<SelectInst>(Val: SI->getTrueValue());
5024 IdenticalSIOtherVal = IdenticalSI->getTrueValue();
5025 } else {
5026 return false;
5027 }
5028
5029 // Now check that the other values in select, i.e., %to_fold and
5030 // %identicalPhi, are essentially the same value.
5031 if (!SIOtherVal || IdenticalSIOtherVal != &IdenticalPN)
5032 return false;
5033 if (!(SIOtherVal->getTrueValue() == &IdenticalPN &&
5034 SIOtherVal->getFalseValue() == &PN) &&
5035 !(SIOtherVal->getTrueValue() == &PN &&
5036 SIOtherVal->getFalseValue() == &IdenticalPN))
5037 return false;
5038 return true;
5039}
5040
5041/// Given operands for a SelectInst, see if we can fold the result.
5042/// If not, this returns null.
5043static Value *simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
5044 const SimplifyQuery &Q, unsigned MaxRecurse) {
5045 if (auto *CondC = dyn_cast<Constant>(Val: Cond)) {
5046 if (auto *TrueC = dyn_cast<Constant>(Val: TrueVal))
5047 if (auto *FalseC = dyn_cast<Constant>(Val: FalseVal))
5048 if (Constant *C = ConstantFoldSelectInstruction(Cond: CondC, V1: TrueC, V2: FalseC))
5049 return C;
5050
5051 // select poison, X, Y -> poison
5052 if (isa<PoisonValue>(Val: CondC))
5053 return PoisonValue::get(T: TrueVal->getType());
5054
5055 // select undef, X, Y -> X or Y
5056 if (Q.isUndefValue(V: CondC))
5057 return isa<Constant>(Val: FalseVal) ? FalseVal : TrueVal;
5058
5059 // select true, X, Y --> X
5060 // select false, X, Y --> Y
5061 // For vectors, allow undef/poison elements in the condition to match the
5062 // defined elements, so we can eliminate the select.
5063 if (match(V: CondC, P: m_One()))
5064 return TrueVal;
5065 if (match(V: CondC, P: m_Zero()))
5066 return FalseVal;
5067 }
5068
5069 assert(Cond->getType()->isIntOrIntVectorTy(1) &&
5070 "Select must have bool or bool vector condition");
5071 assert(TrueVal->getType() == FalseVal->getType() &&
5072 "Select must have same types for true/false ops");
5073
5074 if (Cond->getType() == TrueVal->getType()) {
5075 // select i1 Cond, i1 true, i1 false --> i1 Cond
5076 if (match(V: TrueVal, P: m_One()) && match(V: FalseVal, P: m_ZeroInt()))
5077 return Cond;
5078
5079 // (X && Y) ? X : Y --> Y (commuted 2 ways)
5080 if (match(V: Cond, P: m_c_LogicalAnd(L: m_Specific(V: TrueVal), R: m_Specific(V: FalseVal))))
5081 return FalseVal;
5082
5083 // (X || Y) ? X : Y --> X (commuted 2 ways)
5084 if (match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: TrueVal), R: m_Specific(V: FalseVal))))
5085 return TrueVal;
5086
5087 // (X || Y) ? false : X --> false (commuted 2 ways)
5088 if (match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: FalseVal), R: m_Value())) &&
5089 match(V: TrueVal, P: m_ZeroInt()))
5090 return ConstantInt::getFalse(Ty: Cond->getType());
5091
5092 // Match patterns that end in logical-and.
5093 if (match(V: FalseVal, P: m_ZeroInt())) {
5094 // !(X || Y) && X --> false (commuted 2 ways)
5095 if (match(V: Cond, P: m_Not(V: m_c_LogicalOr(L: m_Specific(V: TrueVal), R: m_Value()))))
5096 return ConstantInt::getFalse(Ty: Cond->getType());
5097 // X && !(X || Y) --> false (commuted 2 ways)
5098 if (match(V: TrueVal, P: m_Not(V: m_c_LogicalOr(L: m_Specific(V: Cond), R: m_Value()))))
5099 return ConstantInt::getFalse(Ty: Cond->getType());
5100
5101 // (X || Y) && Y --> Y (commuted 2 ways)
5102 if (match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: TrueVal), R: m_Value())))
5103 return TrueVal;
5104 // Y && (X || Y) --> Y (commuted 2 ways)
5105 if (match(V: TrueVal, P: m_c_LogicalOr(L: m_Specific(V: Cond), R: m_Value())))
5106 return Cond;
5107
5108 // (X || Y) && (X || !Y) --> X (commuted 8 ways)
5109 Value *X, *Y;
5110 if (match(V: Cond, P: m_c_LogicalOr(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
5111 match(V: TrueVal, P: m_c_LogicalOr(L: m_Specific(V: X), R: m_Specific(V: Y))))
5112 return X;
5113 if (match(V: TrueVal, P: m_c_LogicalOr(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
5114 match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: X), R: m_Specific(V: Y))))
5115 return X;
5116 }
5117
5118 // Match patterns that end in logical-or.
5119 if (match(V: TrueVal, P: m_One())) {
5120 // !(X && Y) || X --> true (commuted 2 ways)
5121 if (match(V: Cond, P: m_Not(V: m_c_LogicalAnd(L: m_Specific(V: FalseVal), R: m_Value()))))
5122 return ConstantInt::getTrue(Ty: Cond->getType());
5123 // X || !(X && Y) --> true (commuted 2 ways)
5124 if (match(V: FalseVal, P: m_Not(V: m_c_LogicalAnd(L: m_Specific(V: Cond), R: m_Value()))))
5125 return ConstantInt::getTrue(Ty: Cond->getType());
5126
5127 // (X && Y) || Y --> Y (commuted 2 ways)
5128 if (match(V: Cond, P: m_c_LogicalAnd(L: m_Specific(V: FalseVal), R: m_Value())))
5129 return FalseVal;
5130 // Y || (X && Y) --> Y (commuted 2 ways)
5131 if (match(V: FalseVal, P: m_c_LogicalAnd(L: m_Specific(V: Cond), R: m_Value())))
5132 return Cond;
5133 }
5134 }
5135
5136 // select ?, X, X -> X
5137 if (TrueVal == FalseVal)
5138 return TrueVal;
5139
5140 if (Cond == TrueVal) {
5141 // select i1 X, i1 X, i1 false --> X (logical-and)
5142 if (match(V: FalseVal, P: m_ZeroInt()))
5143 return Cond;
5144 // select i1 X, i1 X, i1 true --> true
5145 if (match(V: FalseVal, P: m_One()))
5146 return ConstantInt::getTrue(Ty: Cond->getType());
5147 }
5148 if (Cond == FalseVal) {
5149 // select i1 X, i1 true, i1 X --> X (logical-or)
5150 if (match(V: TrueVal, P: m_One()))
5151 return Cond;
5152 // select i1 X, i1 false, i1 X --> false
5153 if (match(V: TrueVal, P: m_ZeroInt()))
5154 return ConstantInt::getFalse(Ty: Cond->getType());
5155 }
5156
5157 // If the true or false value is poison, we can fold to the other value.
5158 // If the true or false value is undef, we can fold to the other value as
5159 // long as the other value isn't poison.
5160 // select ?, poison, X -> X
5161 // select ?, undef, X -> X
5162 if (isa<PoisonValue>(Val: TrueVal) ||
5163 (Q.isUndefValue(V: TrueVal) && impliesPoison(ValAssumedPoison: FalseVal, V: Cond)))
5164 return FalseVal;
5165 // select ?, X, poison -> X
5166 // select ?, X, undef -> X
5167 if (isa<PoisonValue>(Val: FalseVal) ||
5168 (Q.isUndefValue(V: FalseVal) && impliesPoison(ValAssumedPoison: TrueVal, V: Cond)))
5169 return TrueVal;
5170
5171 // Deal with partial undef vector constants: select ?, VecC, VecC' --> VecC''
5172 Constant *TrueC, *FalseC;
5173 if (isa<FixedVectorType>(Val: TrueVal->getType()) &&
5174 match(V: TrueVal, P: m_Constant(C&: TrueC)) &&
5175 match(V: FalseVal, P: m_Constant(C&: FalseC))) {
5176 unsigned NumElts =
5177 cast<FixedVectorType>(Val: TrueC->getType())->getNumElements();
5178 SmallVector<Constant *, 16> NewC;
5179 for (unsigned i = 0; i != NumElts; ++i) {
5180 // Bail out on incomplete vector constants.
5181 Constant *TEltC = TrueC->getAggregateElement(Elt: i);
5182 Constant *FEltC = FalseC->getAggregateElement(Elt: i);
5183 if (!TEltC || !FEltC)
5184 break;
5185
5186 // If the elements match (undef or not), that value is the result. If only
5187 // one element is undef, choose the defined element as the safe result.
5188 if (TEltC == FEltC)
5189 NewC.push_back(Elt: TEltC);
5190 else if (isa<PoisonValue>(Val: TEltC) ||
5191 (Q.isUndefValue(V: TEltC) && isGuaranteedNotToBePoison(V: FEltC)))
5192 NewC.push_back(Elt: FEltC);
5193 else if (isa<PoisonValue>(Val: FEltC) ||
5194 (Q.isUndefValue(V: FEltC) && isGuaranteedNotToBePoison(V: TEltC)))
5195 NewC.push_back(Elt: TEltC);
5196 else
5197 break;
5198 }
5199 if (NewC.size() == NumElts)
5200 return ConstantVector::get(V: NewC);
5201 }
5202
5203 if (Value *V =
5204 simplifySelectWithICmpCond(CondVal: Cond, TrueVal, FalseVal, Q, MaxRecurse))
5205 return V;
5206
5207 if (Value *V = simplifySelectWithBitTest(CondVal: Cond, TrueVal, FalseVal))
5208 return V;
5209
5210 if (Value *V = simplifySelectWithFCmp(Cond, T: TrueVal, F: FalseVal, Q, MaxRecurse))
5211 return V;
5212
5213 std::optional<bool> Imp = isImpliedByDomCondition(Cond, ContextI: Q.CxtI, DL: Q.DL);
5214 if (Imp)
5215 return *Imp ? TrueVal : FalseVal;
5216 // Look for same PHIs in the true and false values.
5217 if (auto *TruePHI = dyn_cast<PHINode>(Val: TrueVal))
5218 if (auto *FalsePHI = dyn_cast<PHINode>(Val: FalseVal)) {
5219 if (isSelectWithIdenticalPHI(PN&: *TruePHI, IdenticalPN&: *FalsePHI))
5220 return FalseVal;
5221 if (isSelectWithIdenticalPHI(PN&: *FalsePHI, IdenticalPN&: *TruePHI))
5222 return TrueVal;
5223 }
5224 return nullptr;
5225}
5226
5227Value *llvm::simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
5228 const SimplifyQuery &Q) {
5229 return ::simplifySelectInst(Cond, TrueVal, FalseVal, Q, MaxRecurse: RecursionLimit);
5230}
5231
5232/// Given operands for an GetElementPtrInst, see if we can fold the result.
5233/// If not, this returns null.
5234static Value *simplifyGEPInst(Type *SrcTy, Value *Ptr,
5235 ArrayRef<Value *> Indices, GEPNoWrapFlags NW,
5236 const SimplifyQuery &Q, unsigned) {
5237 // The type of the GEP pointer operand.
5238 unsigned AS =
5239 cast<PointerType>(Val: Ptr->getType()->getScalarType())->getAddressSpace();
5240
5241 // getelementptr P -> P.
5242 if (Indices.empty())
5243 return Ptr;
5244
5245 // Compute the (pointer) type returned by the GEP instruction.
5246 Type *LastType = GetElementPtrInst::getIndexedType(Ty: SrcTy, IdxList: Indices);
5247 Type *GEPTy = Ptr->getType();
5248 if (!GEPTy->isVectorTy()) {
5249 for (Value *Op : Indices) {
5250 // If one of the operands is a vector, the result type is a vector of
5251 // pointers. All vector operands must have the same number of elements.
5252 if (VectorType *VT = dyn_cast<VectorType>(Val: Op->getType())) {
5253 GEPTy = VectorType::get(ElementType: GEPTy, EC: VT->getElementCount());
5254 break;
5255 }
5256 }
5257 }
5258
5259 // All-zero GEP is a no-op, unless it performs a vector splat.
5260 if (Ptr->getType() == GEPTy && all_of(Range&: Indices, P: match_fn(P: m_Zero())))
5261 return Ptr;
5262
5263 // getelementptr poison, idx -> poison
5264 // getelementptr baseptr, poison -> poison
5265 if (isa<PoisonValue>(Val: Ptr) || any_of(Range&: Indices, P: IsaPred<PoisonValue>))
5266 return PoisonValue::get(T: GEPTy);
5267
5268 // getelementptr undef, idx -> undef
5269 if (Q.isUndefValue(V: Ptr))
5270 return UndefValue::get(T: GEPTy);
5271
5272 bool IsScalableVec =
5273 SrcTy->isScalableTy() || any_of(Range&: Indices, P: [](const Value *V) {
5274 return isa<ScalableVectorType>(Val: V->getType());
5275 });
5276
5277 if (Indices.size() == 1) {
5278 Type *Ty = SrcTy;
5279 if (!IsScalableVec && Ty->isSized()) {
5280 Value *P;
5281 uint64_t C;
5282 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
5283 // getelementptr P, N -> P if P points to a type of zero size.
5284 if (TyAllocSize == 0 && Ptr->getType() == GEPTy)
5285 return Ptr;
5286
5287 // The following transforms are only safe if the ptrtoint cast
5288 // doesn't truncate the address of the pointers. The non-address bits
5289 // must be the same, as the underlying objects are the same.
5290 if (Indices[0]->getType()->getScalarSizeInBits() >=
5291 Q.DL.getAddressSizeInBits(AS)) {
5292 auto CanSimplify = [GEPTy, &P, Ptr]() -> bool {
5293 return P->getType() == GEPTy &&
5294 getUnderlyingObject(V: P) == getUnderlyingObject(V: Ptr);
5295 };
5296 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
5297 if (TyAllocSize == 1 &&
5298 match(V: Indices[0], P: m_Sub(L: m_PtrToIntOrAddr(Op: m_Value(V&: P)),
5299 R: m_PtrToIntOrAddr(Op: m_Specific(V: Ptr)))) &&
5300 CanSimplify())
5301 return P;
5302
5303 // getelementptr V, (ashr (sub P, V), C) -> P if P points to a type of
5304 // size 1 << C.
5305 if (match(V: Indices[0], P: m_AShr(L: m_Sub(L: m_PtrToIntOrAddr(Op: m_Value(V&: P)),
5306 R: m_PtrToIntOrAddr(Op: m_Specific(V: Ptr))),
5307 R: m_ConstantInt(V&: C))) &&
5308 TyAllocSize == 1ULL << C && CanSimplify())
5309 return P;
5310
5311 // getelementptr V, (sdiv (sub P, V), C) -> P if P points to a type of
5312 // size C.
5313 if (match(V: Indices[0], P: m_SDiv(L: m_Sub(L: m_PtrToIntOrAddr(Op: m_Value(V&: P)),
5314 R: m_PtrToIntOrAddr(Op: m_Specific(V: Ptr))),
5315 R: m_SpecificInt(V: TyAllocSize))) &&
5316 CanSimplify())
5317 return P;
5318 }
5319 }
5320 }
5321
5322 if (!IsScalableVec && Q.DL.getTypeAllocSize(Ty: LastType) == 1 &&
5323 all_of(Range: Indices.drop_back(N: 1), P: match_fn(P: m_Zero()))) {
5324 unsigned IdxWidth =
5325 Q.DL.getIndexSizeInBits(AS: Ptr->getType()->getPointerAddressSpace());
5326 if (Q.DL.getTypeSizeInBits(Ty: Indices.back()->getType()) == IdxWidth) {
5327 APInt BasePtrOffset(IdxWidth, 0);
5328 Value *StrippedBasePtr =
5329 Ptr->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: BasePtrOffset);
5330
5331 // Avoid creating inttoptr of zero here: While LLVMs treatment of
5332 // inttoptr is generally conservative, this particular case is folded to
5333 // a null pointer, which will have incorrect provenance.
5334
5335 // gep (gep V, C), (sub 0, V) -> C
5336 if (match(V: Indices.back(),
5337 P: m_Neg(V: m_PtrToInt(Op: m_Specific(V: StrippedBasePtr)))) &&
5338 !BasePtrOffset.isZero()) {
5339 auto *CI = ConstantInt::get(Context&: GEPTy->getContext(), V: BasePtrOffset);
5340 return ConstantExpr::getIntToPtr(C: CI, Ty: GEPTy);
5341 }
5342 // gep (gep V, C), (xor V, -1) -> C-1
5343 if (match(V: Indices.back(),
5344 P: m_Xor(L: m_PtrToInt(Op: m_Specific(V: StrippedBasePtr)), R: m_AllOnes())) &&
5345 !BasePtrOffset.isOne()) {
5346 auto *CI = ConstantInt::get(Context&: GEPTy->getContext(), V: BasePtrOffset - 1);
5347 return ConstantExpr::getIntToPtr(C: CI, Ty: GEPTy);
5348 }
5349 }
5350 }
5351
5352 // Check to see if this is constant foldable.
5353 if (!isa<Constant>(Val: Ptr) || !all_of(Range&: Indices, P: IsaPred<Constant>))
5354 return nullptr;
5355
5356 if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy: SrcTy))
5357 return ConstantFoldGetElementPtr(Ty: SrcTy, C: cast<Constant>(Val: Ptr), InRange: std::nullopt,
5358 Idxs: Indices);
5359
5360 auto *CE =
5361 ConstantExpr::getGetElementPtr(Ty: SrcTy, C: cast<Constant>(Val: Ptr), IdxList: Indices, NW);
5362 return ConstantFoldConstant(C: CE, DL: Q.DL);
5363}
5364
5365Value *llvm::simplifyGEPInst(Type *SrcTy, Value *Ptr, ArrayRef<Value *> Indices,
5366 GEPNoWrapFlags NW, const SimplifyQuery &Q) {
5367 return ::simplifyGEPInst(SrcTy, Ptr, Indices, NW, Q, RecursionLimit);
5368}
5369
5370/// Given operands for an InsertValueInst, see if we can fold the result.
5371/// If not, this returns null.
5372static Value *simplifyInsertValueInst(Value *Agg, Value *Val,
5373 ArrayRef<unsigned> Idxs,
5374 const SimplifyQuery &Q, unsigned) {
5375 if (Constant *CAgg = dyn_cast<Constant>(Val: Agg))
5376 if (Constant *CVal = dyn_cast<Constant>(Val))
5377 return ConstantFoldInsertValueInstruction(Agg: CAgg, Val: CVal, Idxs);
5378
5379 // insertvalue x, poison, n -> x
5380 // insertvalue x, undef, n -> x if x cannot be poison
5381 if (isa<PoisonValue>(Val) ||
5382 (Q.isUndefValue(V: Val) && isGuaranteedNotToBePoison(V: Agg)))
5383 return Agg;
5384
5385 // insertvalue x, (extractvalue y, n), n
5386 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
5387 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
5388 EV->getIndices() == Idxs) {
5389 // insertvalue poison, (extractvalue y, n), n -> y
5390 // insertvalue undef, (extractvalue y, n), n -> y if y cannot be poison
5391 if (isa<PoisonValue>(Val: Agg) ||
5392 (Q.isUndefValue(V: Agg) &&
5393 isGuaranteedNotToBePoison(V: EV->getAggregateOperand())))
5394 return EV->getAggregateOperand();
5395
5396 // insertvalue y, (extractvalue y, n), n -> y
5397 if (Agg == EV->getAggregateOperand())
5398 return Agg;
5399 }
5400
5401 return nullptr;
5402}
5403
5404Value *llvm::simplifyInsertValueInst(Value *Agg, Value *Val,
5405 ArrayRef<unsigned> Idxs,
5406 const SimplifyQuery &Q) {
5407 return ::simplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
5408}
5409
5410Value *llvm::simplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
5411 const SimplifyQuery &Q) {
5412 // Try to constant fold.
5413 auto *VecC = dyn_cast<Constant>(Val: Vec);
5414 auto *ValC = dyn_cast<Constant>(Val);
5415 auto *IdxC = dyn_cast<Constant>(Val: Idx);
5416 if (VecC && ValC && IdxC)
5417 return ConstantExpr::getInsertElement(Vec: VecC, Elt: ValC, Idx: IdxC);
5418
5419 // For fixed-length vector, fold into poison if index is out of bounds.
5420 if (auto *CI = dyn_cast<ConstantInt>(Val: Idx)) {
5421 if (isa<FixedVectorType>(Val: Vec->getType()) &&
5422 CI->uge(Num: cast<FixedVectorType>(Val: Vec->getType())->getNumElements()))
5423 return PoisonValue::get(T: Vec->getType());
5424 }
5425
5426 // If index is undef, it might be out of bounds (see above case)
5427 if (Q.isUndefValue(V: Idx))
5428 return PoisonValue::get(T: Vec->getType());
5429
5430 // If the scalar is poison, or it is undef and there is no risk of
5431 // propagating poison from the vector value, simplify to the vector value.
5432 if (isa<PoisonValue>(Val) ||
5433 (Q.isUndefValue(V: Val) && isGuaranteedNotToBePoison(V: Vec)))
5434 return Vec;
5435
5436 // Inserting the splatted value into a constant splat does nothing.
5437 if (VecC && ValC && VecC->getSplatValue() == ValC)
5438 return Vec;
5439
5440 // If we are extracting a value from a vector, then inserting it into the same
5441 // place, that's the input vector:
5442 // insertelt Vec, (extractelt Vec, Idx), Idx --> Vec
5443 if (match(V: Val, P: m_ExtractElt(Val: m_Specific(V: Vec), Idx: m_Specific(V: Idx))))
5444 return Vec;
5445
5446 return nullptr;
5447}
5448
5449/// Given operands for an ExtractValueInst, see if we can fold the result.
5450/// If not, this returns null.
5451static Value *simplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
5452 const SimplifyQuery &, unsigned) {
5453 if (auto *CAgg = dyn_cast<Constant>(Val: Agg))
5454 return ConstantFoldExtractValueInstruction(Agg: CAgg, Idxs);
5455
5456 // extractvalue x, (insertvalue y, elt, n), n -> elt
5457 unsigned NumIdxs = Idxs.size();
5458 for (auto *IVI = dyn_cast<InsertValueInst>(Val: Agg); IVI != nullptr;
5459 IVI = dyn_cast<InsertValueInst>(Val: IVI->getAggregateOperand())) {
5460 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
5461 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
5462 unsigned NumCommonIdxs = std::min(a: NumInsertValueIdxs, b: NumIdxs);
5463 if (InsertValueIdxs.slice(N: 0, M: NumCommonIdxs) ==
5464 Idxs.slice(N: 0, M: NumCommonIdxs)) {
5465 if (NumIdxs == NumInsertValueIdxs)
5466 return IVI->getInsertedValueOperand();
5467 break;
5468 }
5469 }
5470
5471 // Simplify umul_with_overflow where one operand is 1.
5472 Value *V;
5473 if (Idxs.size() == 1 &&
5474 (match(V: Agg,
5475 P: m_Intrinsic<Intrinsic::umul_with_overflow>(Op0: m_Value(V), Op1: m_One())) ||
5476 match(V: Agg, P: m_Intrinsic<Intrinsic::umul_with_overflow>(Op0: m_One(),
5477 Op1: m_Value(V))))) {
5478 if (Idxs[0] == 0)
5479 return V;
5480 assert(Idxs[0] == 1 && "invalid index");
5481 return getFalse(Ty: CmpInst::makeCmpResultType(opnd_type: V->getType()));
5482 }
5483
5484 return nullptr;
5485}
5486
5487Value *llvm::simplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
5488 const SimplifyQuery &Q) {
5489 return ::simplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
5490}
5491
5492/// Given operands for an ExtractElementInst, see if we can fold the result.
5493/// If not, this returns null.
5494static Value *simplifyExtractElementInst(Value *Vec, Value *Idx,
5495 const SimplifyQuery &Q, unsigned) {
5496 auto *VecVTy = cast<VectorType>(Val: Vec->getType());
5497 if (auto *CVec = dyn_cast<Constant>(Val: Vec)) {
5498 if (auto *CIdx = dyn_cast<Constant>(Val: Idx))
5499 return ConstantExpr::getExtractElement(Vec: CVec, Idx: CIdx);
5500
5501 if (Q.isUndefValue(V: Vec))
5502 return UndefValue::get(T: VecVTy->getElementType());
5503 }
5504
5505 // An undef extract index can be arbitrarily chosen to be an out-of-range
5506 // index value, which would result in the instruction being poison.
5507 if (Q.isUndefValue(V: Idx))
5508 return PoisonValue::get(T: VecVTy->getElementType());
5509
5510 // If extracting a specified index from the vector, see if we can recursively
5511 // find a previously computed scalar that was inserted into the vector.
5512 if (auto *IdxC = dyn_cast<ConstantInt>(Val: Idx)) {
5513 // For fixed-length vector, fold into undef if index is out of bounds.
5514 unsigned MinNumElts = VecVTy->getElementCount().getKnownMinValue();
5515 if (isa<FixedVectorType>(Val: VecVTy) && IdxC->getValue().uge(RHS: MinNumElts))
5516 return PoisonValue::get(T: VecVTy->getElementType());
5517 // Handle case where an element is extracted from a splat.
5518 if (IdxC->getValue().ult(RHS: MinNumElts))
5519 if (auto *Splat = getSplatValue(V: Vec))
5520 return Splat;
5521 if (Value *Elt = findScalarElement(V: Vec, EltNo: IdxC->getZExtValue()))
5522 return Elt;
5523 } else {
5524 // extractelt x, (insertelt y, elt, n), n -> elt
5525 // If the possibly-variable indices are trivially known to be equal
5526 // (because they are the same operand) then use the value that was
5527 // inserted directly.
5528 auto *IE = dyn_cast<InsertElementInst>(Val: Vec);
5529 if (IE && IE->getOperand(i_nocapture: 2) == Idx)
5530 return IE->getOperand(i_nocapture: 1);
5531
5532 // The index is not relevant if our vector is a splat.
5533 if (Value *Splat = getSplatValue(V: Vec))
5534 return Splat;
5535 }
5536 return nullptr;
5537}
5538
5539Value *llvm::simplifyExtractElementInst(Value *Vec, Value *Idx,
5540 const SimplifyQuery &Q) {
5541 return ::simplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
5542}
5543
5544/// See if we can fold the given phi. If not, returns null.
5545static Value *simplifyPHINode(PHINode *PN, ArrayRef<Value *> IncomingValues,
5546 const SimplifyQuery &Q) {
5547 // WARNING: no matter how worthwhile it may seem, we can not perform PHI CSE
5548 // here, because the PHI we may succeed simplifying to was not
5549 // def-reachable from the original PHI!
5550
5551 // If all of the PHI's incoming values are the same then replace the PHI node
5552 // with the common value.
5553 Value *CommonValue = nullptr;
5554 bool HasPoisonInput = false;
5555 bool HasUndefInput = false;
5556 for (Value *Incoming : IncomingValues) {
5557 // If the incoming value is the phi node itself, it can safely be skipped.
5558 if (Incoming == PN)
5559 continue;
5560 if (isa<PoisonValue>(Val: Incoming)) {
5561 HasPoisonInput = true;
5562 continue;
5563 }
5564 if (Q.isUndefValue(V: Incoming)) {
5565 // Remember that we saw an undef value, but otherwise ignore them.
5566 HasUndefInput = true;
5567 continue;
5568 }
5569 if (CommonValue && Incoming != CommonValue)
5570 return nullptr; // Not the same, bail out.
5571 CommonValue = Incoming;
5572 }
5573
5574 // If CommonValue is null then all of the incoming values were either undef,
5575 // poison or equal to the phi node itself.
5576 if (!CommonValue)
5577 return HasUndefInput ? UndefValue::get(T: PN->getType())
5578 : PoisonValue::get(T: PN->getType());
5579
5580 if (HasPoisonInput || HasUndefInput) {
5581 // If we have a PHI node like phi(X, undef, X), where X is defined by some
5582 // instruction, we cannot return X as the result of the PHI node unless it
5583 // dominates the PHI block.
5584 if (!valueDominatesPHI(V: CommonValue, P: PN, DT: Q.DT))
5585 return nullptr;
5586
5587 // Make sure we do not replace an undef value with poison.
5588 if (HasUndefInput &&
5589 !isGuaranteedNotToBePoison(V: CommonValue, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT))
5590 return nullptr;
5591 return CommonValue;
5592 }
5593
5594 return CommonValue;
5595}
5596
5597static Value *simplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
5598 const SimplifyQuery &Q, unsigned MaxRecurse) {
5599 if (auto *C = dyn_cast<Constant>(Val: Op))
5600 return ConstantFoldCastOperand(Opcode: CastOpc, C, DestTy: Ty, DL: Q.DL);
5601
5602 if (auto *CI = dyn_cast<CastInst>(Val: Op)) {
5603 auto *Src = CI->getOperand(i_nocapture: 0);
5604 Type *SrcTy = Src->getType();
5605 Type *MidTy = CI->getType();
5606 Type *DstTy = Ty;
5607 if (Src->getType() == Ty) {
5608 auto FirstOp = CI->getOpcode();
5609 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
5610 if (CastInst::isEliminableCastPair(firstOpcode: FirstOp, secondOpcode: SecondOp, SrcTy, MidTy, DstTy,
5611 DL: &Q.DL) == Instruction::BitCast)
5612 return Src;
5613 }
5614 }
5615
5616 // bitcast x -> x
5617 if (CastOpc == Instruction::BitCast)
5618 if (Op->getType() == Ty)
5619 return Op;
5620
5621 // ptrtoint (ptradd (Ptr, X - ptrtoint(Ptr))) -> X
5622 Value *Ptr, *X;
5623 if ((CastOpc == Instruction::PtrToInt || CastOpc == Instruction::PtrToAddr) &&
5624 match(V: Op,
5625 P: m_PtrAdd(PointerOp: m_Value(V&: Ptr),
5626 OffsetOp: m_Sub(L: m_Value(V&: X), R: m_PtrToIntOrAddr(Op: m_Deferred(V: Ptr))))) &&
5627 X->getType() == Ty && Ty == Q.DL.getIndexType(PtrTy: Ptr->getType()))
5628 return X;
5629
5630 return nullptr;
5631}
5632
5633Value *llvm::simplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
5634 const SimplifyQuery &Q) {
5635 return ::simplifyCastInst(CastOpc, Op, Ty, Q, MaxRecurse: RecursionLimit);
5636}
5637
5638/// For the given destination element of a shuffle, peek through shuffles to
5639/// match a root vector source operand that contains that element in the same
5640/// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
5641static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
5642 int MaskVal, Value *RootVec,
5643 unsigned MaxRecurse) {
5644 if (!MaxRecurse--)
5645 return nullptr;
5646
5647 // Bail out if any mask value is undefined. That kind of shuffle may be
5648 // simplified further based on demanded bits or other folds.
5649 if (MaskVal == -1)
5650 return nullptr;
5651
5652 // The mask value chooses which source operand we need to look at next.
5653 int InVecNumElts = cast<FixedVectorType>(Val: Op0->getType())->getNumElements();
5654 int RootElt = MaskVal;
5655 Value *SourceOp = Op0;
5656 if (MaskVal >= InVecNumElts) {
5657 RootElt = MaskVal - InVecNumElts;
5658 SourceOp = Op1;
5659 }
5660
5661 // If the source operand is a shuffle itself, look through it to find the
5662 // matching root vector.
5663 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(Val: SourceOp)) {
5664 return foldIdentityShuffles(
5665 DestElt, Op0: SourceShuf->getOperand(i_nocapture: 0), Op1: SourceShuf->getOperand(i_nocapture: 1),
5666 MaskVal: SourceShuf->getMaskValue(Elt: RootElt), RootVec, MaxRecurse);
5667 }
5668
5669 // The source operand is not a shuffle. Initialize the root vector value for
5670 // this shuffle if that has not been done yet.
5671 if (!RootVec)
5672 RootVec = SourceOp;
5673
5674 // Give up as soon as a source operand does not match the existing root value.
5675 if (RootVec != SourceOp)
5676 return nullptr;
5677
5678 // The element must be coming from the same lane in the source vector
5679 // (although it may have crossed lanes in intermediate shuffles).
5680 if (RootElt != DestElt)
5681 return nullptr;
5682
5683 return RootVec;
5684}
5685
5686static Value *simplifyShuffleVectorInst(Value *Op0, Value *Op1,
5687 ArrayRef<int> Mask, Type *RetTy,
5688 const SimplifyQuery &Q,
5689 unsigned MaxRecurse) {
5690 if (all_of(Range&: Mask, P: equal_to(Arg: PoisonMaskElem)))
5691 return PoisonValue::get(T: RetTy);
5692
5693 auto *InVecTy = cast<VectorType>(Val: Op0->getType());
5694 unsigned MaskNumElts = Mask.size();
5695 ElementCount InVecEltCount = InVecTy->getElementCount();
5696
5697 bool Scalable = InVecEltCount.isScalable();
5698
5699 SmallVector<int, 32> Indices;
5700 Indices.assign(in_start: Mask.begin(), in_end: Mask.end());
5701
5702 // Canonicalization: If mask does not select elements from an input vector,
5703 // replace that input vector with poison.
5704 if (!Scalable) {
5705 bool MaskSelects0 = false, MaskSelects1 = false;
5706 unsigned InVecNumElts = InVecEltCount.getKnownMinValue();
5707 for (unsigned i = 0; i != MaskNumElts; ++i) {
5708 if (Indices[i] == -1)
5709 continue;
5710 if ((unsigned)Indices[i] < InVecNumElts)
5711 MaskSelects0 = true;
5712 else
5713 MaskSelects1 = true;
5714 }
5715 if (!MaskSelects0)
5716 Op0 = PoisonValue::get(T: InVecTy);
5717 if (!MaskSelects1)
5718 Op1 = PoisonValue::get(T: InVecTy);
5719 }
5720
5721 auto *Op0Const = dyn_cast<Constant>(Val: Op0);
5722 auto *Op1Const = dyn_cast<Constant>(Val: Op1);
5723
5724 // If all operands are constant, constant fold the shuffle. This
5725 // transformation depends on the value of the mask which is not known at
5726 // compile time for scalable vectors
5727 if (Op0Const && Op1Const)
5728 return ConstantExpr::getShuffleVector(V1: Op0Const, V2: Op1Const, Mask);
5729
5730 // Canonicalization: if only one input vector is constant, it shall be the
5731 // second one. This transformation depends on the value of the mask which
5732 // is not known at compile time for scalable vectors
5733 if (!Scalable && Op0Const && !Op1Const) {
5734 std::swap(a&: Op0, b&: Op1);
5735 ShuffleVectorInst::commuteShuffleMask(Mask: Indices,
5736 InVecNumElts: InVecEltCount.getKnownMinValue());
5737 }
5738
5739 // A splat of an inserted scalar constant becomes a vector constant:
5740 // shuf (inselt ?, C, IndexC), undef, <IndexC, IndexC...> --> <C, C...>
5741 // NOTE: We may have commuted above, so analyze the updated Indices, not the
5742 // original mask constant.
5743 // NOTE: This transformation depends on the value of the mask which is not
5744 // known at compile time for scalable vectors
5745 Constant *C;
5746 ConstantInt *IndexC;
5747 if (!Scalable && match(V: Op0, P: m_InsertElt(Val: m_Value(), Elt: m_Constant(C),
5748 Idx: m_ConstantInt(CI&: IndexC)))) {
5749 // Match a splat shuffle mask of the insert index allowing undef elements.
5750 int InsertIndex = IndexC->getZExtValue();
5751 if (all_of(Range&: Indices, P: [InsertIndex](int MaskElt) {
5752 return MaskElt == InsertIndex || MaskElt == -1;
5753 })) {
5754 assert(isa<UndefValue>(Op1) && "Expected undef operand 1 for splat");
5755
5756 // Shuffle mask poisons become poison constant result elements.
5757 SmallVector<Constant *, 16> VecC(MaskNumElts, C);
5758 for (unsigned i = 0; i != MaskNumElts; ++i)
5759 if (Indices[i] == -1)
5760 VecC[i] = PoisonValue::get(T: C->getType());
5761 return ConstantVector::get(V: VecC);
5762 }
5763 }
5764
5765 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
5766 // value type is same as the input vectors' type.
5767 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Val: Op0))
5768 if (Q.isUndefValue(V: Op1) && RetTy == InVecTy &&
5769 all_equal(Range: OpShuf->getShuffleMask()))
5770 return Op0;
5771
5772 // All remaining transformation depend on the value of the mask, which is
5773 // not known at compile time for scalable vectors.
5774 if (Scalable)
5775 return nullptr;
5776
5777 // Don't fold a shuffle with undef mask elements. This may get folded in a
5778 // better way using demanded bits or other analysis.
5779 // TODO: Should we allow this?
5780 if (is_contained(Range&: Indices, Element: -1))
5781 return nullptr;
5782
5783 // Check if every element of this shuffle can be mapped back to the
5784 // corresponding element of a single root vector. If so, we don't need this
5785 // shuffle. This handles simple identity shuffles as well as chains of
5786 // shuffles that may widen/narrow and/or move elements across lanes and back.
5787 Value *RootVec = nullptr;
5788 for (unsigned i = 0; i != MaskNumElts; ++i) {
5789 // Note that recursion is limited for each vector element, so if any element
5790 // exceeds the limit, this will fail to simplify.
5791 RootVec =
5792 foldIdentityShuffles(DestElt: i, Op0, Op1, MaskVal: Indices[i], RootVec, MaxRecurse);
5793
5794 // We can't replace a widening/narrowing shuffle with one of its operands.
5795 if (!RootVec || RootVec->getType() != RetTy)
5796 return nullptr;
5797 }
5798 return RootVec;
5799}
5800
5801/// Given operands for a ShuffleVectorInst, fold the result or return null.
5802Value *llvm::simplifyShuffleVectorInst(Value *Op0, Value *Op1,
5803 ArrayRef<int> Mask, Type *RetTy,
5804 const SimplifyQuery &Q) {
5805 return ::simplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, MaxRecurse: RecursionLimit);
5806}
5807
5808static Constant *foldConstant(Instruction::UnaryOps Opcode, Value *&Op,
5809 const SimplifyQuery &Q) {
5810 if (auto *C = dyn_cast<Constant>(Val: Op))
5811 return ConstantFoldUnaryOpOperand(Opcode, Op: C, DL: Q.DL);
5812 return nullptr;
5813}
5814
5815/// Given the operand for an FNeg, see if we can fold the result. If not, this
5816/// returns null.
5817static Value *simplifyFNegInst(Value *Op, FastMathFlags FMF,
5818 const SimplifyQuery &Q, unsigned MaxRecurse) {
5819 if (Constant *C = foldConstant(Opcode: Instruction::FNeg, Op, Q))
5820 return C;
5821
5822 Value *X;
5823 // fneg (fneg X) ==> X
5824 if (match(V: Op, P: m_FNeg(X: m_Value(V&: X))))
5825 return X;
5826
5827 return nullptr;
5828}
5829
5830Value *llvm::simplifyFNegInst(Value *Op, FastMathFlags FMF,
5831 const SimplifyQuery &Q) {
5832 return ::simplifyFNegInst(Op, FMF, Q, MaxRecurse: RecursionLimit);
5833}
5834
5835/// Try to propagate existing NaN values when possible. If not, replace the
5836/// constant or elements in the constant with a canonical NaN.
5837static Constant *propagateNaN(Constant *In) {
5838 Type *Ty = In->getType();
5839 if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Ty)) {
5840 unsigned NumElts = VecTy->getNumElements();
5841 SmallVector<Constant *, 32> NewC(NumElts);
5842 for (unsigned i = 0; i != NumElts; ++i) {
5843 Constant *EltC = In->getAggregateElement(Elt: i);
5844 // Poison elements propagate. NaN propagates except signaling is quieted.
5845 // Replace unknown or undef elements with canonical NaN.
5846 if (EltC && isa<PoisonValue>(Val: EltC))
5847 NewC[i] = EltC;
5848 else if (EltC && EltC->isNaN())
5849 NewC[i] = ConstantFP::get(
5850 Ty: EltC->getType(), V: cast<ConstantFP>(Val: EltC)->getValue().makeQuiet());
5851 else
5852 NewC[i] = ConstantFP::getNaN(Ty: VecTy->getElementType());
5853 }
5854 return ConstantVector::get(V: NewC);
5855 }
5856
5857 // If it is not a fixed vector, but not a simple NaN either, return a
5858 // canonical NaN.
5859 if (!In->isNaN())
5860 return ConstantFP::getNaN(Ty);
5861
5862 // If we known this is a NaN, and it's scalable vector, we must have a splat
5863 // on our hands. Grab that before splatting a QNaN constant.
5864 if (isa<ScalableVectorType>(Val: Ty)) {
5865 auto *Splat = In->getSplatValue();
5866 assert(Splat && Splat->isNaN() &&
5867 "Found a scalable-vector NaN but not a splat");
5868 In = Splat;
5869 }
5870
5871 // Propagate an existing QNaN constant. If it is an SNaN, make it quiet, but
5872 // preserve the sign/payload.
5873 return ConstantFP::get(Ty, V: cast<ConstantFP>(Val: In)->getValue().makeQuiet());
5874}
5875
5876/// Perform folds that are common to any floating-point operation. This implies
5877/// transforms based on poison/undef/NaN because the operation itself makes no
5878/// difference to the result.
5879static Constant *simplifyFPOp(ArrayRef<Value *> Ops, FastMathFlags FMF,
5880 const SimplifyQuery &Q,
5881 fp::ExceptionBehavior ExBehavior,
5882 RoundingMode Rounding) {
5883 // Poison is independent of anything else. It always propagates from an
5884 // operand to a math result.
5885 if (any_of(Range&: Ops, P: IsaPred<PoisonValue>))
5886 return PoisonValue::get(T: Ops[0]->getType());
5887
5888 for (Value *V : Ops) {
5889 bool IsNan = match(V, P: m_NaN());
5890 bool IsInf = match(V, P: m_Inf());
5891 bool IsUndef = Q.isUndefValue(V);
5892
5893 // If this operation has 'nnan' or 'ninf' and at least 1 disallowed operand
5894 // (an undef operand can be chosen to be Nan/Inf), then the result of
5895 // this operation is poison.
5896 if (FMF.noNaNs() && (IsNan || IsUndef))
5897 return PoisonValue::get(T: V->getType());
5898 if (FMF.noInfs() && (IsInf || IsUndef))
5899 return PoisonValue::get(T: V->getType());
5900
5901 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding)) {
5902 // Undef does not propagate because undef means that all bits can take on
5903 // any value. If this is undef * NaN for example, then the result values
5904 // (at least the exponent bits) are limited. Assume the undef is a
5905 // canonical NaN and propagate that.
5906 if (IsUndef)
5907 return ConstantFP::getNaN(Ty: V->getType());
5908 if (IsNan)
5909 return propagateNaN(In: cast<Constant>(Val: V));
5910 } else if (ExBehavior != fp::ebStrict) {
5911 if (IsNan)
5912 return propagateNaN(In: cast<Constant>(Val: V));
5913 }
5914 }
5915 return nullptr;
5916}
5917
5918/// Given operands for an FAdd, see if we can fold the result. If not, this
5919/// returns null.
5920static Value *
5921simplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
5922 const SimplifyQuery &Q, unsigned MaxRecurse,
5923 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
5924 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
5925 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
5926 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FAdd, Op0, Op1, Q))
5927 return C;
5928
5929 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
5930 return C;
5931
5932 // fadd X, -0 ==> X
5933 // With strict/constrained FP, we have these possible edge cases that do
5934 // not simplify to Op0:
5935 // fadd SNaN, -0.0 --> QNaN
5936 // fadd +0.0, -0.0 --> -0.0 (but only with round toward negative)
5937 if (canIgnoreSNaN(EB: ExBehavior, FMF) &&
5938 (!canRoundingModeBe(RM: Rounding, QRM: RoundingMode::TowardNegative) ||
5939 FMF.noSignedZeros()))
5940 if (match(V: Op1, P: m_NegZeroFP()))
5941 return Op0;
5942
5943 // fadd X, 0 ==> X, when we know X is not -0
5944 if (canIgnoreSNaN(EB: ExBehavior, FMF))
5945 if (match(V: Op1, P: m_PosZeroFP()) &&
5946 (FMF.noSignedZeros() || cannotBeNegativeZero(V: Op0, SQ: Q)))
5947 return Op0;
5948
5949 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
5950 return nullptr;
5951
5952 if (FMF.noNaNs()) {
5953 // With nnan: X + {+/-}Inf --> {+/-}Inf
5954 if (match(V: Op1, P: m_Inf()))
5955 return Op1;
5956
5957 // With nnan: -X + X --> 0.0 (and commuted variant)
5958 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
5959 // Negative zeros are allowed because we always end up with positive zero:
5960 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
5961 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
5962 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
5963 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
5964 if (match(V: Op0, P: m_FSub(L: m_AnyZeroFP(), R: m_Specific(V: Op1))) ||
5965 match(V: Op1, P: m_FSub(L: m_AnyZeroFP(), R: m_Specific(V: Op0))))
5966 return ConstantFP::getZero(Ty: Op0->getType());
5967
5968 if (match(V: Op0, P: m_FNeg(X: m_Specific(V: Op1))) ||
5969 match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0))))
5970 return ConstantFP::getZero(Ty: Op0->getType());
5971 }
5972
5973 // (X - Y) + Y --> X
5974 // Y + (X - Y) --> X
5975 Value *X;
5976 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
5977 (match(V: Op0, P: m_FSub(L: m_Value(V&: X), R: m_Specific(V: Op1))) ||
5978 match(V: Op1, P: m_FSub(L: m_Value(V&: X), R: m_Specific(V: Op0)))))
5979 return X;
5980
5981 return nullptr;
5982}
5983
5984/// Given operands for an FSub, see if we can fold the result. If not, this
5985/// returns null.
5986static Value *
5987simplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
5988 const SimplifyQuery &Q, unsigned MaxRecurse,
5989 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
5990 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
5991 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
5992 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FSub, Op0, Op1, Q))
5993 return C;
5994
5995 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
5996 return C;
5997
5998 // fsub X, +0 ==> X
5999 if (canIgnoreSNaN(EB: ExBehavior, FMF) &&
6000 (!canRoundingModeBe(RM: Rounding, QRM: RoundingMode::TowardNegative) ||
6001 FMF.noSignedZeros()))
6002 if (match(V: Op1, P: m_PosZeroFP()))
6003 return Op0;
6004
6005 // fsub X, -0 ==> X, when we know X is not -0
6006 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6007 if (match(V: Op1, P: m_NegZeroFP()) &&
6008 (FMF.noSignedZeros() || cannotBeNegativeZero(V: Op0, SQ: Q)))
6009 return Op0;
6010
6011 // fsub -0.0, (fsub -0.0, X) ==> X
6012 // fsub -0.0, (fneg X) ==> X
6013 Value *X;
6014 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6015 if (match(V: Op0, P: m_NegZeroFP()) && match(V: Op1, P: m_FNeg(X: m_Value(V&: X))))
6016 return X;
6017
6018 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
6019 // fsub 0.0, (fneg X) ==> X if signed zeros are ignored.
6020 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6021 if (FMF.noSignedZeros() && match(V: Op0, P: m_AnyZeroFP()) &&
6022 (match(V: Op1, P: m_FSub(L: m_AnyZeroFP(), R: m_Value(V&: X))) ||
6023 match(V: Op1, P: m_FNeg(X: m_Value(V&: X)))))
6024 return X;
6025
6026 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6027 return nullptr;
6028
6029 if (FMF.noNaNs()) {
6030 // fsub nnan x, x ==> 0.0
6031 if (Op0 == Op1)
6032 return Constant::getNullValue(Ty: Op0->getType());
6033
6034 // With nnan: {+/-}Inf - X --> {+/-}Inf
6035 if (match(V: Op0, P: m_Inf()))
6036 return Op0;
6037
6038 // With nnan: X - {+/-}Inf --> {-/+}Inf
6039 if (match(V: Op1, P: m_Inf()))
6040 return foldConstant(Opcode: Instruction::FNeg, Op&: Op1, Q);
6041 }
6042
6043 // Y - (Y - X) --> X
6044 // (X + Y) - Y --> X
6045 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
6046 (match(V: Op1, P: m_FSub(L: m_Specific(V: Op0), R: m_Value(V&: X))) ||
6047 match(V: Op0, P: m_c_FAdd(L: m_Specific(V: Op1), R: m_Value(V&: X)))))
6048 return X;
6049
6050 return nullptr;
6051}
6052
6053static Value *simplifyFMAFMul(Value *Op0, Value *Op1, FastMathFlags FMF,
6054 const SimplifyQuery &Q, unsigned MaxRecurse,
6055 fp::ExceptionBehavior ExBehavior,
6056 RoundingMode Rounding) {
6057 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6058 return C;
6059
6060 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6061 return nullptr;
6062
6063 // Canonicalize special constants as operand 1.
6064 if (match(V: Op0, P: m_FPOne()) || match(V: Op0, P: m_AnyZeroFP()))
6065 std::swap(a&: Op0, b&: Op1);
6066
6067 // X * 1.0 --> X
6068 if (match(V: Op1, P: m_FPOne()))
6069 return Op0;
6070
6071 if (match(V: Op1, P: m_AnyZeroFP())) {
6072 // X * 0.0 --> 0.0 (with nnan and nsz)
6073 if (FMF.noNaNs() && FMF.noSignedZeros())
6074 return ConstantFP::getZero(Ty: Op0->getType());
6075
6076 KnownFPClass Known = computeKnownFPClass(V: Op0, FMF, InterestedClasses: fcInf | fcNan, SQ: Q);
6077 if (Known.isKnownNever(Mask: fcInf | fcNan)) {
6078 // if nsz is set, return 0.0
6079 if (FMF.noSignedZeros())
6080 return ConstantFP::getZero(Ty: Op0->getType());
6081 // +normal number * (-)0.0 --> (-)0.0
6082 if (Known.SignBit == false)
6083 return Op1;
6084 // -normal number * (-)0.0 --> -(-)0.0
6085 if (Known.SignBit == true)
6086 return foldConstant(Opcode: Instruction::FNeg, Op&: Op1, Q);
6087 }
6088 }
6089
6090 // sqrt(X) * sqrt(X) --> X, if we can:
6091 // 1. Remove the intermediate rounding (reassociate).
6092 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
6093 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
6094 Value *X;
6095 if (Op0 == Op1 && match(V: Op0, P: m_Sqrt(Op0: m_Value(V&: X))) && FMF.allowReassoc() &&
6096 FMF.noNaNs() && FMF.noSignedZeros())
6097 return X;
6098
6099 return nullptr;
6100}
6101
6102/// Given the operands for an FMul, see if we can fold the result
6103static Value *
6104simplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6105 const SimplifyQuery &Q, unsigned MaxRecurse,
6106 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6107 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6108 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6109 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FMul, Op0, Op1, Q))
6110 return C;
6111
6112 // Now apply simplifications that do not require rounding.
6113 return simplifyFMAFMul(Op0, Op1, FMF, Q, MaxRecurse, ExBehavior, Rounding);
6114}
6115
6116Value *llvm::simplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6117 const SimplifyQuery &Q,
6118 fp::ExceptionBehavior ExBehavior,
6119 RoundingMode Rounding) {
6120 return ::simplifyFAddInst(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6121 Rounding);
6122}
6123
6124Value *llvm::simplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6125 const SimplifyQuery &Q,
6126 fp::ExceptionBehavior ExBehavior,
6127 RoundingMode Rounding) {
6128 return ::simplifyFSubInst(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6129 Rounding);
6130}
6131
6132Value *llvm::simplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6133 const SimplifyQuery &Q,
6134 fp::ExceptionBehavior ExBehavior,
6135 RoundingMode Rounding) {
6136 return ::simplifyFMulInst(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6137 Rounding);
6138}
6139
6140Value *llvm::simplifyFMAFMul(Value *Op0, Value *Op1, FastMathFlags FMF,
6141 const SimplifyQuery &Q,
6142 fp::ExceptionBehavior ExBehavior,
6143 RoundingMode Rounding) {
6144 return ::simplifyFMAFMul(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6145 Rounding);
6146}
6147
6148static Value *
6149simplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6150 const SimplifyQuery &Q, unsigned,
6151 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6152 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6153 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6154 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FDiv, Op0, Op1, Q))
6155 return C;
6156
6157 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6158 return C;
6159
6160 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6161 return nullptr;
6162
6163 // X / 1.0 -> X
6164 if (match(V: Op1, P: m_FPOne()))
6165 return Op0;
6166
6167 // 0 / X -> 0
6168 // Requires that NaNs are off (X could be zero) and signed zeroes are
6169 // ignored (X could be positive or negative, so the output sign is unknown).
6170 if (FMF.noNaNs() && FMF.noSignedZeros() && match(V: Op0, P: m_AnyZeroFP()))
6171 return ConstantFP::getZero(Ty: Op0->getType());
6172
6173 if (FMF.noNaNs()) {
6174 // X / X -> 1.0 is legal when NaNs are ignored.
6175 // We can ignore infinities because INF/INF is NaN.
6176 if (Op0 == Op1)
6177 return ConstantFP::get(Ty: Op0->getType(), V: 1.0);
6178
6179 // (X * Y) / Y --> X if we can reassociate to the above form.
6180 Value *X;
6181 if (FMF.allowReassoc() && match(V: Op0, P: m_c_FMul(L: m_Value(V&: X), R: m_Specific(V: Op1))))
6182 return X;
6183
6184 // -X / X -> -1.0 and
6185 // X / -X -> -1.0 are legal when NaNs are ignored.
6186 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
6187 if (match(V: Op0, P: m_FNegNSZ(X: m_Specific(V: Op1))) ||
6188 match(V: Op1, P: m_FNegNSZ(X: m_Specific(V: Op0))))
6189 return ConstantFP::get(Ty: Op0->getType(), V: -1.0);
6190
6191 // nnan ninf X / [-]0.0 -> poison
6192 if (FMF.noInfs() && match(V: Op1, P: m_AnyZeroFP()))
6193 return PoisonValue::get(T: Op1->getType());
6194 }
6195
6196 return nullptr;
6197}
6198
6199Value *llvm::simplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6200 const SimplifyQuery &Q,
6201 fp::ExceptionBehavior ExBehavior,
6202 RoundingMode Rounding) {
6203 return ::simplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit, ExBehavior,
6204 Rounding);
6205}
6206
6207static Value *
6208simplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6209 const SimplifyQuery &Q, unsigned,
6210 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6211 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6212 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6213 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FRem, Op0, Op1, Q))
6214 return C;
6215
6216 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6217 return C;
6218
6219 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6220 return nullptr;
6221
6222 // Unlike fdiv, the result of frem always matches the sign of the dividend.
6223 // The constant match may include undef elements in a vector, so return a full
6224 // zero constant as the result.
6225 if (FMF.noNaNs()) {
6226 // +0 % X -> 0
6227 if (match(V: Op0, P: m_PosZeroFP()))
6228 return ConstantFP::getZero(Ty: Op0->getType());
6229 // -0 % X -> -0
6230 if (match(V: Op0, P: m_NegZeroFP()))
6231 return ConstantFP::getNegativeZero(Ty: Op0->getType());
6232 }
6233
6234 return nullptr;
6235}
6236
6237Value *llvm::simplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6238 const SimplifyQuery &Q,
6239 fp::ExceptionBehavior ExBehavior,
6240 RoundingMode Rounding) {
6241 return ::simplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit, ExBehavior,
6242 Rounding);
6243}
6244
6245//=== Helper functions for higher up the class hierarchy.
6246
6247/// Given the operand for a UnaryOperator, see if we can fold the result.
6248/// If not, this returns null.
6249static Value *simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q,
6250 unsigned MaxRecurse) {
6251 switch (Opcode) {
6252 case Instruction::FNeg:
6253 return simplifyFNegInst(Op, FMF: FastMathFlags(), Q, MaxRecurse);
6254 default:
6255 llvm_unreachable("Unexpected opcode");
6256 }
6257}
6258
6259/// Given the operand for a UnaryOperator, see if we can fold the result.
6260/// If not, this returns null.
6261/// Try to use FastMathFlags when folding the result.
6262static Value *simplifyFPUnOp(unsigned Opcode, Value *Op,
6263 const FastMathFlags &FMF, const SimplifyQuery &Q,
6264 unsigned MaxRecurse) {
6265 switch (Opcode) {
6266 case Instruction::FNeg:
6267 return simplifyFNegInst(Op, FMF, Q, MaxRecurse);
6268 default:
6269 return simplifyUnOp(Opcode, Op, Q, MaxRecurse);
6270 }
6271}
6272
6273Value *llvm::simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q) {
6274 return ::simplifyUnOp(Opcode, Op, Q, MaxRecurse: RecursionLimit);
6275}
6276
6277Value *llvm::simplifyUnOp(unsigned Opcode, Value *Op, FastMathFlags FMF,
6278 const SimplifyQuery &Q) {
6279 return ::simplifyFPUnOp(Opcode, Op, FMF, Q, MaxRecurse: RecursionLimit);
6280}
6281
6282/// Given operands for a BinaryOperator, see if we can fold the result.
6283/// If not, this returns null.
6284static Value *simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6285 const SimplifyQuery &Q, unsigned MaxRecurse) {
6286 switch (Opcode) {
6287 case Instruction::Add:
6288 return simplifyAddInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6289 MaxRecurse);
6290 case Instruction::Sub:
6291 return simplifySubInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6292 MaxRecurse);
6293 case Instruction::Mul:
6294 return simplifyMulInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6295 MaxRecurse);
6296 case Instruction::SDiv:
6297 return simplifySDivInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6298 case Instruction::UDiv:
6299 return simplifyUDivInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6300 case Instruction::SRem:
6301 return simplifySRemInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6302 case Instruction::URem:
6303 return simplifyURemInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6304 case Instruction::Shl:
6305 return simplifyShlInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6306 MaxRecurse);
6307 case Instruction::LShr:
6308 return simplifyLShrInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6309 case Instruction::AShr:
6310 return simplifyAShrInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6311 case Instruction::And:
6312 return simplifyAndInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6313 case Instruction::Or:
6314 return simplifyOrInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6315 case Instruction::Xor:
6316 return simplifyXorInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6317 case Instruction::FAdd:
6318 return simplifyFAddInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6319 case Instruction::FSub:
6320 return simplifyFSubInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6321 case Instruction::FMul:
6322 return simplifyFMulInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6323 case Instruction::FDiv:
6324 return simplifyFDivInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6325 case Instruction::FRem:
6326 return simplifyFRemInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6327 default:
6328 llvm_unreachable("Unexpected opcode");
6329 }
6330}
6331
6332/// Given operands for a BinaryOperator, see if we can fold the result.
6333/// If not, this returns null.
6334/// Try to use FastMathFlags when folding the result.
6335static Value *simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6336 const FastMathFlags &FMF, const SimplifyQuery &Q,
6337 unsigned MaxRecurse) {
6338 switch (Opcode) {
6339 case Instruction::FAdd:
6340 return simplifyFAddInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6341 case Instruction::FSub:
6342 return simplifyFSubInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6343 case Instruction::FMul:
6344 return simplifyFMulInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6345 case Instruction::FDiv:
6346 return simplifyFDivInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6347 default:
6348 return simplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
6349 }
6350}
6351
6352Value *llvm::simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6353 const SimplifyQuery &Q) {
6354 return ::simplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse: RecursionLimit);
6355}
6356
6357Value *llvm::simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6358 FastMathFlags FMF, const SimplifyQuery &Q) {
6359 return ::simplifyBinOp(Opcode, LHS, RHS, FMF, Q, MaxRecurse: RecursionLimit);
6360}
6361
6362/// Given operands for a CmpInst, see if we can fold the result.
6363static Value *simplifyCmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
6364 const SimplifyQuery &Q, unsigned MaxRecurse) {
6365 if (CmpInst::isIntPredicate(P: Predicate))
6366 return simplifyICmpInst(Pred: Predicate, LHS, RHS, Q, MaxRecurse);
6367 return simplifyFCmpInst(Pred: Predicate, LHS, RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6368}
6369
6370Value *llvm::simplifyCmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
6371 const SimplifyQuery &Q) {
6372 return ::simplifyCmpInst(Predicate, LHS, RHS, Q, MaxRecurse: RecursionLimit);
6373}
6374
6375static bool isIdempotent(Intrinsic::ID ID) {
6376 switch (ID) {
6377 default:
6378 return false;
6379
6380 // Unary idempotent: f(f(x)) = f(x)
6381 case Intrinsic::fabs:
6382 case Intrinsic::floor:
6383 case Intrinsic::ceil:
6384 case Intrinsic::trunc:
6385 case Intrinsic::rint:
6386 case Intrinsic::nearbyint:
6387 case Intrinsic::round:
6388 case Intrinsic::roundeven:
6389 case Intrinsic::canonicalize:
6390 case Intrinsic::arithmetic_fence:
6391 return true;
6392 }
6393}
6394
6395/// Return true if the intrinsic rounds a floating-point value to an integral
6396/// floating-point value (not an integer type).
6397static bool removesFPFraction(Intrinsic::ID ID) {
6398 switch (ID) {
6399 default:
6400 return false;
6401
6402 case Intrinsic::floor:
6403 case Intrinsic::ceil:
6404 case Intrinsic::trunc:
6405 case Intrinsic::rint:
6406 case Intrinsic::nearbyint:
6407 case Intrinsic::round:
6408 case Intrinsic::roundeven:
6409 return true;
6410 }
6411}
6412
6413static Value *simplifyRelativeLoad(Constant *Ptr, Constant *Offset,
6414 const DataLayout &DL) {
6415 GlobalValue *PtrSym;
6416 APInt PtrOffset;
6417 if (!IsConstantOffsetFromGlobal(C: Ptr, GV&: PtrSym, Offset&: PtrOffset, DL))
6418 return nullptr;
6419
6420 Type *Int32Ty = Type::getInt32Ty(C&: Ptr->getContext());
6421
6422 auto *OffsetConstInt = dyn_cast<ConstantInt>(Val: Offset);
6423 if (!OffsetConstInt || OffsetConstInt->getBitWidth() > 64)
6424 return nullptr;
6425
6426 APInt OffsetInt = OffsetConstInt->getValue().sextOrTrunc(
6427 width: DL.getIndexTypeSizeInBits(Ty: Ptr->getType()));
6428 if (OffsetInt.srem(RHS: 4) != 0)
6429 return nullptr;
6430
6431 Constant *Loaded =
6432 ConstantFoldLoadFromConstPtr(C: Ptr, Ty: Int32Ty, Offset: std::move(OffsetInt), DL);
6433 if (!Loaded)
6434 return nullptr;
6435
6436 auto *LoadedCE = dyn_cast<ConstantExpr>(Val: Loaded);
6437 if (!LoadedCE)
6438 return nullptr;
6439
6440 if (LoadedCE->getOpcode() == Instruction::Trunc) {
6441 LoadedCE = dyn_cast<ConstantExpr>(Val: LoadedCE->getOperand(i_nocapture: 0));
6442 if (!LoadedCE)
6443 return nullptr;
6444 }
6445
6446 if (LoadedCE->getOpcode() != Instruction::Sub)
6447 return nullptr;
6448
6449 auto *LoadedLHS = dyn_cast<ConstantExpr>(Val: LoadedCE->getOperand(i_nocapture: 0));
6450 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
6451 return nullptr;
6452 auto *LoadedLHSPtr = LoadedLHS->getOperand(i_nocapture: 0);
6453
6454 Constant *LoadedRHS = LoadedCE->getOperand(i_nocapture: 1);
6455 GlobalValue *LoadedRHSSym;
6456 APInt LoadedRHSOffset;
6457 if (!IsConstantOffsetFromGlobal(C: LoadedRHS, GV&: LoadedRHSSym, Offset&: LoadedRHSOffset,
6458 DL) ||
6459 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
6460 return nullptr;
6461
6462 return LoadedLHSPtr;
6463}
6464
6465// TODO: Need to pass in FastMathFlags
6466static Value *simplifyLdexp(Value *Op0, Value *Op1, const SimplifyQuery &Q,
6467 bool IsStrict) {
6468 // ldexp(poison, x) -> poison
6469 // ldexp(x, poison) -> poison
6470 if (isa<PoisonValue>(Val: Op0) || isa<PoisonValue>(Val: Op1))
6471 return Op0;
6472
6473 // ldexp(undef, x) -> nan
6474 if (Q.isUndefValue(V: Op0))
6475 return ConstantFP::getNaN(Ty: Op0->getType());
6476
6477 if (!IsStrict) {
6478 // TODO: Could insert a canonicalize for strict
6479
6480 // ldexp(x, undef) -> x
6481 if (Q.isUndefValue(V: Op1))
6482 return Op0;
6483 }
6484
6485 const APFloat *C = nullptr;
6486 match(V: Op0, P: PatternMatch::m_APFloat(Res&: C));
6487
6488 // These cases should be safe, even with strictfp.
6489 // ldexp(0.0, x) -> 0.0
6490 // ldexp(-0.0, x) -> -0.0
6491 // ldexp(inf, x) -> inf
6492 // ldexp(-inf, x) -> -inf
6493 if (C && (C->isZero() || C->isInfinity()))
6494 return Op0;
6495
6496 // These are canonicalization dropping, could do it if we knew how we could
6497 // ignore denormal flushes and target handling of nan payload bits.
6498 if (IsStrict)
6499 return nullptr;
6500
6501 // TODO: Could quiet this with strictfp if the exception mode isn't strict.
6502 if (C && C->isNaN())
6503 return ConstantFP::get(Ty: Op0->getType(), V: C->makeQuiet());
6504
6505 // ldexp(x, 0) -> x
6506
6507 // TODO: Could fold this if we know the exception mode isn't
6508 // strict, we know the denormal mode and other target modes.
6509 if (match(V: Op1, P: PatternMatch::m_ZeroInt()))
6510 return Op0;
6511
6512 return nullptr;
6513}
6514
6515static Value *simplifyUnaryIntrinsic(Function *F, Value *Op0,
6516 const SimplifyQuery &Q,
6517 const CallBase *Call) {
6518 // Idempotent functions return the same result when called repeatedly.
6519 Intrinsic::ID IID = F->getIntrinsicID();
6520 if (isIdempotent(ID: IID))
6521 if (auto *II = dyn_cast<IntrinsicInst>(Val: Op0))
6522 if (II->getIntrinsicID() == IID)
6523 return II;
6524
6525 if (removesFPFraction(ID: IID)) {
6526 // Converting from int or calling a rounding function always results in a
6527 // finite integral number or infinity. For those inputs, rounding functions
6528 // always return the same value, so the (2nd) rounding is eliminated. Ex:
6529 // floor (sitofp x) -> sitofp x
6530 // round (ceil x) -> ceil x
6531 auto *II = dyn_cast<IntrinsicInst>(Val: Op0);
6532 if ((II && removesFPFraction(ID: II->getIntrinsicID())) ||
6533 match(V: Op0, P: m_SIToFP(Op: m_Value())) || match(V: Op0, P: m_UIToFP(Op: m_Value())))
6534 return Op0;
6535 }
6536
6537 Value *X;
6538 switch (IID) {
6539 case Intrinsic::fabs: {
6540 KnownFPClass KnownClass = computeKnownFPClass(V: Op0, InterestedClasses: fcAllFlags, SQ: Q);
6541 if (KnownClass.SignBit == false)
6542 return Op0;
6543
6544 if (KnownClass.cannotBeOrderedLessThanZero() &&
6545 KnownClass.isKnownNeverNaN() && Call->hasNoSignedZeros())
6546 return Op0;
6547
6548 break;
6549 }
6550 case Intrinsic::bswap:
6551 // bswap(bswap(x)) -> x
6552 if (match(V: Op0, P: m_BSwap(Op0: m_Value(V&: X))))
6553 return X;
6554 break;
6555 case Intrinsic::bitreverse:
6556 // bitreverse(bitreverse(x)) -> x
6557 if (match(V: Op0, P: m_BitReverse(Op0: m_Value(V&: X))))
6558 return X;
6559 break;
6560 case Intrinsic::ctpop: {
6561 // ctpop(X) -> 1 iff X is non-zero power of 2.
6562 if (isKnownToBeAPowerOfTwo(V: Op0, DL: Q.DL, /*OrZero*/ false, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
6563 return ConstantInt::get(Ty: Op0->getType(), V: 1);
6564 // If everything but the lowest bit is zero, that bit is the pop-count. Ex:
6565 // ctpop(and X, 1) --> and X, 1
6566 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
6567 if (MaskedValueIsZero(V: Op0, Mask: APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 1),
6568 SQ: Q))
6569 return Op0;
6570 break;
6571 }
6572 case Intrinsic::exp:
6573 // exp(log(x)) -> x
6574 if (Call->hasAllowReassoc() &&
6575 match(V: Op0, P: m_Intrinsic<Intrinsic::log>(Op0: m_Value(V&: X))))
6576 return X;
6577 break;
6578 case Intrinsic::exp2:
6579 // exp2(log2(x)) -> x
6580 if (Call->hasAllowReassoc() &&
6581 match(V: Op0, P: m_Intrinsic<Intrinsic::log2>(Op0: m_Value(V&: X))))
6582 return X;
6583 break;
6584 case Intrinsic::exp10:
6585 // exp10(log10(x)) -> x
6586 if (Call->hasAllowReassoc() &&
6587 match(V: Op0, P: m_Intrinsic<Intrinsic::log10>(Op0: m_Value(V&: X))))
6588 return X;
6589 break;
6590 case Intrinsic::log:
6591 // log(exp(x)) -> x
6592 if (Call->hasAllowReassoc() &&
6593 match(V: Op0, P: m_Intrinsic<Intrinsic::exp>(Op0: m_Value(V&: X))))
6594 return X;
6595 break;
6596 case Intrinsic::log2:
6597 // log2(exp2(x)) -> x
6598 if (Call->hasAllowReassoc() &&
6599 (match(V: Op0, P: m_Intrinsic<Intrinsic::exp2>(Op0: m_Value(V&: X))) ||
6600 match(V: Op0,
6601 P: m_Intrinsic<Intrinsic::pow>(Op0: m_SpecificFP(V: 2.0), Op1: m_Value(V&: X)))))
6602 return X;
6603 break;
6604 case Intrinsic::log10:
6605 // log10(pow(10.0, x)) -> x
6606 // log10(exp10(x)) -> x
6607 if (Call->hasAllowReassoc() &&
6608 (match(V: Op0, P: m_Intrinsic<Intrinsic::exp10>(Op0: m_Value(V&: X))) ||
6609 match(V: Op0,
6610 P: m_Intrinsic<Intrinsic::pow>(Op0: m_SpecificFP(V: 10.0), Op1: m_Value(V&: X)))))
6611 return X;
6612 break;
6613 case Intrinsic::vector_reverse:
6614 // vector.reverse(vector.reverse(x)) -> x
6615 if (match(V: Op0, P: m_VecReverse(Op0: m_Value(V&: X))))
6616 return X;
6617 // vector.reverse(splat(X)) -> splat(X)
6618 if (isSplatValue(V: Op0))
6619 return Op0;
6620 break;
6621 case Intrinsic::structured_gep:
6622 return cast<StructuredGEPInst>(Val: Call)->getPointerOperand();
6623 default:
6624 break;
6625 }
6626
6627 return nullptr;
6628}
6629
6630/// Given a min/max intrinsic, see if it can be removed based on having an
6631/// operand that is another min/max intrinsic with shared operand(s). The caller
6632/// is expected to swap the operand arguments to handle commutation.
6633static Value *foldMinMaxSharedOp(Intrinsic::ID IID, Value *Op0, Value *Op1) {
6634 Value *X, *Y;
6635 if (!match(V: Op0, P: m_MaxOrMin(L: m_Value(V&: X), R: m_Value(V&: Y))))
6636 return nullptr;
6637
6638 auto *MM0 = dyn_cast<IntrinsicInst>(Val: Op0);
6639 if (!MM0)
6640 return nullptr;
6641 Intrinsic::ID IID0 = MM0->getIntrinsicID();
6642
6643 if (Op1 == X || Op1 == Y ||
6644 match(V: Op1, P: m_c_MaxOrMin(L: m_Specific(V: X), R: m_Specific(V: Y)))) {
6645 // max (max X, Y), X --> max X, Y
6646 if (IID0 == IID)
6647 return MM0;
6648 // max (min X, Y), X --> X
6649 if (IID0 == getInverseMinMaxIntrinsic(MinMaxID: IID))
6650 return Op1;
6651 }
6652 return nullptr;
6653}
6654
6655/// Given a min/max intrinsic, see if it can be removed based on having an
6656/// operand that is another min/max intrinsic with shared operand(s). The caller
6657/// is expected to swap the operand arguments to handle commutation.
6658static Value *foldMinimumMaximumSharedOp(Intrinsic::ID IID, Value *Op0,
6659 Value *Op1) {
6660 auto IsMinimumMaximumIntrinsic = [](Intrinsic::ID ID) {
6661 switch (ID) {
6662 case Intrinsic::maxnum:
6663 case Intrinsic::minnum:
6664 case Intrinsic::maximum:
6665 case Intrinsic::minimum:
6666 case Intrinsic::maximumnum:
6667 case Intrinsic::minimumnum:
6668 return true;
6669 default:
6670 return false;
6671 }
6672 };
6673
6674 assert(IsMinimumMaximumIntrinsic(IID) && "Unsupported intrinsic");
6675
6676 auto *M0 = dyn_cast<IntrinsicInst>(Val: Op0);
6677 // If Op0 is not the same intrinsic as IID, do not process.
6678 // This is a difference with integer min/max handling. We do not process the
6679 // case like max(min(X,Y),min(X,Y)) => min(X,Y). But it can be handled by GVN.
6680 if (!M0 || M0->getIntrinsicID() != IID)
6681 return nullptr;
6682 Value *X0 = M0->getOperand(i_nocapture: 0);
6683 Value *Y0 = M0->getOperand(i_nocapture: 1);
6684 // Simple case, m(m(X,Y), X) => m(X, Y)
6685 // m(m(X,Y), Y) => m(X, Y)
6686 // For minimum/maximum, X is NaN => m(NaN, Y) == NaN and m(NaN, NaN) == NaN.
6687 // For minimum/maximum, Y is NaN => m(X, NaN) == NaN and m(NaN, NaN) == NaN.
6688 // For minnum/maxnum, X is NaN => m(NaN, Y) == Y and m(Y, Y) == Y.
6689 // For minnum/maxnum, Y is NaN => m(X, NaN) == X and m(X, NaN) == X.
6690 if (X0 == Op1 || Y0 == Op1)
6691 return M0;
6692
6693 auto *M1 = dyn_cast<IntrinsicInst>(Val: Op1);
6694 if (!M1 || !IsMinimumMaximumIntrinsic(M1->getIntrinsicID()))
6695 return nullptr;
6696 Value *X1 = M1->getOperand(i_nocapture: 0);
6697 Value *Y1 = M1->getOperand(i_nocapture: 1);
6698 Intrinsic::ID IID1 = M1->getIntrinsicID();
6699 // we have a case m(m(X,Y),m'(X,Y)) taking into account m' is commutative.
6700 // if m' is m or inversion of m => m(m(X,Y),m'(X,Y)) == m(X,Y).
6701 // For minimum/maximum, X is NaN => m(NaN,Y) == m'(NaN, Y) == NaN.
6702 // For minimum/maximum, Y is NaN => m(X,NaN) == m'(X, NaN) == NaN.
6703 // For minnum/maxnum, X is NaN => m(NaN,Y) == m'(NaN, Y) == Y.
6704 // For minnum/maxnum, Y is NaN => m(X,NaN) == m'(X, NaN) == X.
6705 if ((X0 == X1 && Y0 == Y1) || (X0 == Y1 && Y0 == X1))
6706 if (IID1 == IID || getInverseMinMaxIntrinsic(MinMaxID: IID1) == IID)
6707 return M0;
6708
6709 return nullptr;
6710}
6711
6712enum class MinMaxOptResult {
6713 CannotOptimize = 0,
6714 UseNewConstVal = 1,
6715 UseOtherVal = 2,
6716 // For undef/poison, we can choose to either propgate undef/poison or
6717 // use the LHS value depending on what will allow more optimization.
6718 UseEither = 3
6719};
6720// Get the optimized value for a min/max instruction with a single constant
6721// input (either undef or scalar constantFP). The result may indicate to
6722// use the non-const LHS value, use a new constant value instead (with NaNs
6723// quieted), or to choose either option in the case of undef/poison.
6724static MinMaxOptResult OptimizeConstMinMax(const Constant *RHSConst,
6725 const Intrinsic::ID IID,
6726 const CallBase *Call,
6727 Constant **OutNewConstVal) {
6728 assert(OutNewConstVal != nullptr);
6729
6730 bool PropagateNaN = IID == Intrinsic::minimum || IID == Intrinsic::maximum;
6731 bool PropagateSNaN = IID == Intrinsic::minnum || IID == Intrinsic::maxnum;
6732 bool IsMin = IID == Intrinsic::minimum || IID == Intrinsic::minnum ||
6733 IID == Intrinsic::minimumnum;
6734
6735 // min/max(x, poison) -> either x or poison
6736 if (isa<UndefValue>(Val: RHSConst)) {
6737 *OutNewConstVal = const_cast<Constant *>(RHSConst);
6738 return MinMaxOptResult::UseEither;
6739 }
6740
6741 const ConstantFP *CFP = dyn_cast<ConstantFP>(Val: RHSConst);
6742 if (!CFP)
6743 return MinMaxOptResult::CannotOptimize;
6744 APFloat CAPF = CFP->getValueAPF();
6745
6746 // minnum(x, qnan) -> x
6747 // maxnum(x, qnan) -> x
6748 // minnum(x, snan) -> qnan
6749 // maxnum(x, snan) -> qnan
6750 // minimum(X, nan) -> qnan
6751 // maximum(X, nan) -> qnan
6752 // minimumnum(X, nan) -> x
6753 // maximumnum(X, nan) -> x
6754 if (CAPF.isNaN()) {
6755 if (PropagateNaN || (PropagateSNaN && CAPF.isSignaling())) {
6756 *OutNewConstVal = ConstantFP::get(Ty: CFP->getType(), V: CAPF.makeQuiet());
6757 return MinMaxOptResult::UseNewConstVal;
6758 }
6759 return MinMaxOptResult::UseOtherVal;
6760 }
6761
6762 if (CAPF.isInfinity() || (Call && Call->hasNoInfs() && CAPF.isLargest())) {
6763 // minnum(X, -inf) -> -inf (ignoring sNaN -> qNaN propagation)
6764 // maxnum(X, +inf) -> +inf (ignoring sNaN -> qNaN propagation)
6765 // minimum(X, -inf) -> -inf if nnan
6766 // maximum(X, +inf) -> +inf if nnan
6767 // minimumnum(X, -inf) -> -inf
6768 // maximumnum(X, +inf) -> +inf
6769 if (CAPF.isNegative() == IsMin &&
6770 (!PropagateNaN || (Call && Call->hasNoNaNs()))) {
6771 *OutNewConstVal = const_cast<Constant *>(RHSConst);
6772 return MinMaxOptResult::UseNewConstVal;
6773 }
6774
6775 // minnum(X, +inf) -> X if nnan
6776 // maxnum(X, -inf) -> X if nnan
6777 // minimum(X, +inf) -> X (ignoring quieting of sNaNs)
6778 // maximum(X, -inf) -> X (ignoring quieting of sNaNs)
6779 // minimumnum(X, +inf) -> X if nnan
6780 // maximumnum(X, -inf) -> X if nnan
6781 if (CAPF.isNegative() != IsMin &&
6782 (PropagateNaN || (Call && Call->hasNoNaNs())))
6783 return MinMaxOptResult::UseOtherVal;
6784 }
6785 return MinMaxOptResult::CannotOptimize;
6786}
6787
6788static Value *simplifySVEIntReduction(Intrinsic::ID IID, Type *ReturnType,
6789 Value *Op0, Value *Op1) {
6790 Constant *C0 = dyn_cast<Constant>(Val: Op0);
6791 Constant *C1 = dyn_cast<Constant>(Val: Op1);
6792 unsigned Width = ReturnType->getPrimitiveSizeInBits();
6793
6794 // All false predicate or reduction of neutral values ==> neutral result.
6795 switch (IID) {
6796 case Intrinsic::aarch64_sve_eorv:
6797 case Intrinsic::aarch64_sve_orv:
6798 case Intrinsic::aarch64_sve_saddv:
6799 case Intrinsic::aarch64_sve_uaddv:
6800 case Intrinsic::aarch64_sve_umaxv:
6801 if ((C0 && C0->isNullValue()) || (C1 && C1->isNullValue()))
6802 return ConstantInt::get(Ty: ReturnType, V: 0);
6803 break;
6804 case Intrinsic::aarch64_sve_andv:
6805 case Intrinsic::aarch64_sve_uminv:
6806 if ((C0 && C0->isNullValue()) || (C1 && C1->isAllOnesValue()))
6807 return ConstantInt::get(Ty: ReturnType, V: APInt::getMaxValue(numBits: Width));
6808 break;
6809 case Intrinsic::aarch64_sve_smaxv:
6810 if ((C0 && C0->isNullValue()) || (C1 && C1->isMinSignedValue()))
6811 return ConstantInt::get(Ty: ReturnType, V: APInt::getSignedMinValue(numBits: Width));
6812 break;
6813 case Intrinsic::aarch64_sve_sminv:
6814 if ((C0 && C0->isNullValue()) || (C1 && C1->isMaxSignedValue()))
6815 return ConstantInt::get(Ty: ReturnType, V: APInt::getSignedMaxValue(numBits: Width));
6816 break;
6817 }
6818
6819 switch (IID) {
6820 case Intrinsic::aarch64_sve_andv:
6821 case Intrinsic::aarch64_sve_orv:
6822 case Intrinsic::aarch64_sve_smaxv:
6823 case Intrinsic::aarch64_sve_sminv:
6824 case Intrinsic::aarch64_sve_umaxv:
6825 case Intrinsic::aarch64_sve_uminv:
6826 // sve_reduce_##(all, splat(X)) ==> X
6827 if (C0 && C0->isAllOnesValue()) {
6828 if (Value *SplatVal = getSplatValue(V: Op1)) {
6829 assert(SplatVal->getType() == ReturnType && "Unexpected result type!");
6830 return SplatVal;
6831 }
6832 }
6833 break;
6834 case Intrinsic::aarch64_sve_eorv:
6835 // sve_reduce_xor(all, splat(X)) ==> 0
6836 if (C0 && C0->isAllOnesValue())
6837 return ConstantInt::get(Ty: ReturnType, V: 0);
6838 break;
6839 }
6840
6841 return nullptr;
6842}
6843
6844Value *llvm::simplifyBinaryIntrinsic(Intrinsic::ID IID, Type *ReturnType,
6845 Value *Op0, Value *Op1,
6846 const SimplifyQuery &Q,
6847 const CallBase *Call) {
6848 unsigned BitWidth = ReturnType->getScalarSizeInBits();
6849 switch (IID) {
6850 case Intrinsic::get_active_lane_mask: {
6851 if (match(V: Op1, P: m_Zero()))
6852 return ConstantInt::getFalse(Ty: ReturnType);
6853
6854 const Function *F = Call->getFunction();
6855 auto *ScalableTy = dyn_cast<ScalableVectorType>(Val: ReturnType);
6856 Attribute Attr = F->getFnAttribute(Kind: Attribute::VScaleRange);
6857 if (ScalableTy && Attr.isValid()) {
6858 std::optional<unsigned> VScaleMax = Attr.getVScaleRangeMax();
6859 if (!VScaleMax)
6860 break;
6861 uint64_t MaxPossibleMaskElements =
6862 (uint64_t)ScalableTy->getMinNumElements() * (*VScaleMax);
6863
6864 const APInt *Op1Val;
6865 if (match(V: Op0, P: m_Zero()) && match(V: Op1, P: m_APInt(Res&: Op1Val)) &&
6866 Op1Val->uge(RHS: MaxPossibleMaskElements))
6867 return ConstantInt::getAllOnesValue(Ty: ReturnType);
6868 }
6869 break;
6870 }
6871 case Intrinsic::abs:
6872 // abs(abs(x)) -> abs(x). We don't need to worry about the nsw arg here.
6873 // It is always ok to pick the earlier abs. We'll just lose nsw if its only
6874 // on the outer abs.
6875 if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(), Op1: m_Value())))
6876 return Op0;
6877 break;
6878
6879 case Intrinsic::cttz: {
6880 Value *X;
6881 if (match(V: Op0, P: m_Shl(L: m_One(), R: m_Value(V&: X))))
6882 return X;
6883 break;
6884 }
6885 case Intrinsic::ctlz: {
6886 Value *X;
6887 if (match(V: Op0, P: m_LShr(L: m_Negative(), R: m_Value(V&: X))))
6888 return X;
6889 if (match(V: Op0, P: m_AShr(L: m_Negative(), R: m_Value())))
6890 return Constant::getNullValue(Ty: ReturnType);
6891 break;
6892 }
6893 case Intrinsic::ptrmask: {
6894 // NOTE: We can't apply this simplifications based on the value of Op1
6895 // because we need to preserve provenance.
6896 if (Q.isUndefValue(V: Op0) || match(V: Op0, P: m_Zero()))
6897 return Constant::getNullValue(Ty: Op0->getType());
6898
6899 assert(Op1->getType()->getScalarSizeInBits() ==
6900 Q.DL.getIndexTypeSizeInBits(Op0->getType()) &&
6901 "Invalid mask width");
6902 // If index-width (mask size) is less than pointer-size then mask is
6903 // 1-extended.
6904 if (match(V: Op1, P: m_PtrToIntOrAddr(Op: m_Specific(V: Op0))))
6905 return Op0;
6906
6907 // NOTE: We may have attributes associated with the return value of the
6908 // llvm.ptrmask intrinsic that will be lost when we just return the
6909 // operand. We should try to preserve them.
6910 if (match(V: Op1, P: m_AllOnes()) || Q.isUndefValue(V: Op1))
6911 return Op0;
6912
6913 Constant *C;
6914 if (match(V: Op1, P: m_ImmConstant(C))) {
6915 KnownBits PtrKnown = computeKnownBits(V: Op0, Q);
6916 // See if we only masking off bits we know are already zero due to
6917 // alignment.
6918 APInt IrrelevantPtrBits =
6919 PtrKnown.Zero.zextOrTrunc(width: C->getType()->getScalarSizeInBits());
6920 C = ConstantFoldBinaryOpOperands(
6921 Opcode: Instruction::Or, LHS: C, RHS: ConstantInt::get(Ty: C->getType(), V: IrrelevantPtrBits),
6922 DL: Q.DL);
6923 if (C != nullptr && C->isAllOnesValue())
6924 return Op0;
6925 }
6926 break;
6927 }
6928 case Intrinsic::smax:
6929 case Intrinsic::smin:
6930 case Intrinsic::umax:
6931 case Intrinsic::umin: {
6932 // If the arguments are the same, this is a no-op.
6933 if (Op0 == Op1)
6934 return Op0;
6935
6936 // Canonicalize immediate constant operand as Op1.
6937 if (match(V: Op0, P: m_ImmConstant()))
6938 std::swap(a&: Op0, b&: Op1);
6939
6940 // Assume undef is the limit value.
6941 if (Q.isUndefValue(V: Op1))
6942 return ConstantInt::get(
6943 Ty: ReturnType, V: MinMaxIntrinsic::getSaturationPoint(ID: IID, numBits: BitWidth));
6944
6945 const APInt *C;
6946 if (match(V: Op1, P: m_APIntAllowPoison(Res&: C))) {
6947 // Clamp to limit value. For example:
6948 // umax(i8 %x, i8 255) --> 255
6949 if (*C == MinMaxIntrinsic::getSaturationPoint(ID: IID, numBits: BitWidth))
6950 return ConstantInt::get(Ty: ReturnType, V: *C);
6951
6952 // If the constant op is the opposite of the limit value, the other must
6953 // be larger/smaller or equal. For example:
6954 // umin(i8 %x, i8 255) --> %x
6955 if (*C == MinMaxIntrinsic::getSaturationPoint(
6956 ID: getInverseMinMaxIntrinsic(MinMaxID: IID), numBits: BitWidth))
6957 return Op0;
6958
6959 // Remove nested call if constant operands allow it. Example:
6960 // max (max X, 7), 5 -> max X, 7
6961 auto *MinMax0 = dyn_cast<IntrinsicInst>(Val: Op0);
6962 if (MinMax0 && MinMax0->getIntrinsicID() == IID) {
6963 // TODO: loosen undef/splat restrictions for vector constants.
6964 Value *M00 = MinMax0->getOperand(i_nocapture: 0), *M01 = MinMax0->getOperand(i_nocapture: 1);
6965 const APInt *InnerC;
6966 if ((match(V: M00, P: m_APInt(Res&: InnerC)) || match(V: M01, P: m_APInt(Res&: InnerC))) &&
6967 ICmpInst::compare(LHS: *InnerC, RHS: *C,
6968 Pred: ICmpInst::getNonStrictPredicate(
6969 pred: MinMaxIntrinsic::getPredicate(ID: IID))))
6970 return Op0;
6971 }
6972 }
6973
6974 if (Value *V = foldMinMaxSharedOp(IID, Op0, Op1))
6975 return V;
6976 if (Value *V = foldMinMaxSharedOp(IID, Op0: Op1, Op1: Op0))
6977 return V;
6978
6979 ICmpInst::Predicate Pred =
6980 ICmpInst::getNonStrictPredicate(pred: MinMaxIntrinsic::getPredicate(ID: IID));
6981 if (isICmpTrue(Pred, LHS: Op0, RHS: Op1, Q: Q.getWithoutUndef(), MaxRecurse: RecursionLimit))
6982 return Op0;
6983 if (isICmpTrue(Pred, LHS: Op1, RHS: Op0, Q: Q.getWithoutUndef(), MaxRecurse: RecursionLimit))
6984 return Op1;
6985
6986 break;
6987 }
6988 case Intrinsic::scmp:
6989 case Intrinsic::ucmp: {
6990 // Fold to a constant if the relationship between operands can be
6991 // established with certainty
6992 if (isICmpTrue(Pred: CmpInst::ICMP_EQ, LHS: Op0, RHS: Op1, Q, MaxRecurse: RecursionLimit))
6993 return Constant::getNullValue(Ty: ReturnType);
6994
6995 ICmpInst::Predicate PredGT =
6996 IID == Intrinsic::scmp ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
6997 if (isICmpTrue(Pred: PredGT, LHS: Op0, RHS: Op1, Q, MaxRecurse: RecursionLimit))
6998 return ConstantInt::get(Ty: ReturnType, V: 1);
6999
7000 ICmpInst::Predicate PredLT =
7001 IID == Intrinsic::scmp ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
7002 if (isICmpTrue(Pred: PredLT, LHS: Op0, RHS: Op1, Q, MaxRecurse: RecursionLimit))
7003 return ConstantInt::getSigned(Ty: ReturnType, V: -1);
7004
7005 break;
7006 }
7007 case Intrinsic::usub_with_overflow:
7008 case Intrinsic::ssub_with_overflow:
7009 // X - X -> { 0, false }
7010 // X - undef -> { 0, false }
7011 // undef - X -> { 0, false }
7012 if (Op0 == Op1 || Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7013 return Constant::getNullValue(Ty: ReturnType);
7014 break;
7015 case Intrinsic::uadd_with_overflow:
7016 case Intrinsic::sadd_with_overflow:
7017 // X + undef -> { -1, false }
7018 // undef + x -> { -1, false }
7019 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1)) {
7020 return ConstantStruct::get(
7021 T: cast<StructType>(Val: ReturnType),
7022 V: {Constant::getAllOnesValue(Ty: ReturnType->getStructElementType(N: 0)),
7023 Constant::getNullValue(Ty: ReturnType->getStructElementType(N: 1))});
7024 }
7025 break;
7026 case Intrinsic::umul_with_overflow:
7027 case Intrinsic::smul_with_overflow:
7028 // 0 * X -> { 0, false }
7029 // X * 0 -> { 0, false }
7030 if (match(V: Op0, P: m_Zero()) || match(V: Op1, P: m_Zero()))
7031 return Constant::getNullValue(Ty: ReturnType);
7032 // undef * X -> { 0, false }
7033 // X * undef -> { 0, false }
7034 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7035 return Constant::getNullValue(Ty: ReturnType);
7036 break;
7037 case Intrinsic::uadd_sat:
7038 // sat(MAX + X) -> MAX
7039 // sat(X + MAX) -> MAX
7040 if (match(V: Op0, P: m_AllOnes()) || match(V: Op1, P: m_AllOnes()))
7041 return Constant::getAllOnesValue(Ty: ReturnType);
7042 [[fallthrough]];
7043 case Intrinsic::sadd_sat:
7044 // sat(X + undef) -> -1
7045 // sat(undef + X) -> -1
7046 // For unsigned: Assume undef is MAX, thus we saturate to MAX (-1).
7047 // For signed: Assume undef is ~X, in which case X + ~X = -1.
7048 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7049 return Constant::getAllOnesValue(Ty: ReturnType);
7050
7051 // X + 0 -> X
7052 if (match(V: Op1, P: m_Zero()))
7053 return Op0;
7054 // 0 + X -> X
7055 if (match(V: Op0, P: m_Zero()))
7056 return Op1;
7057 break;
7058 case Intrinsic::usub_sat:
7059 // sat(0 - X) -> 0, sat(X - MAX) -> 0
7060 if (match(V: Op0, P: m_Zero()) || match(V: Op1, P: m_AllOnes()))
7061 return Constant::getNullValue(Ty: ReturnType);
7062 [[fallthrough]];
7063 case Intrinsic::ssub_sat:
7064 // X - X -> 0, X - undef -> 0, undef - X -> 0
7065 if (Op0 == Op1 || Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7066 return Constant::getNullValue(Ty: ReturnType);
7067 // X - 0 -> X
7068 if (match(V: Op1, P: m_Zero()))
7069 return Op0;
7070 break;
7071 case Intrinsic::load_relative:
7072 if (auto *C0 = dyn_cast<Constant>(Val: Op0))
7073 if (auto *C1 = dyn_cast<Constant>(Val: Op1))
7074 return simplifyRelativeLoad(Ptr: C0, Offset: C1, DL: Q.DL);
7075 break;
7076 case Intrinsic::powi:
7077 if (auto *Power = dyn_cast<ConstantInt>(Val: Op1)) {
7078 // powi(x, 0) -> 1.0
7079 if (Power->isZero())
7080 return ConstantFP::get(Ty: Op0->getType(), V: 1.0);
7081 // powi(x, 1) -> x
7082 if (Power->isOne())
7083 return Op0;
7084 }
7085 break;
7086 case Intrinsic::ldexp:
7087 return simplifyLdexp(Op0, Op1, Q, IsStrict: false);
7088 case Intrinsic::copysign:
7089 // copysign X, X --> X
7090 if (Op0 == Op1)
7091 return Op0;
7092 // copysign -X, X --> X
7093 // copysign X, -X --> -X
7094 if (match(V: Op0, P: m_FNeg(X: m_Specific(V: Op1))) ||
7095 match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0))))
7096 return Op1;
7097 break;
7098 case Intrinsic::is_fpclass: {
7099 uint64_t Mask = cast<ConstantInt>(Val: Op1)->getZExtValue();
7100 // If all tests are made, it doesn't matter what the value is.
7101 if ((Mask & fcAllFlags) == fcAllFlags)
7102 return ConstantInt::get(Ty: ReturnType, V: true);
7103 if ((Mask & fcAllFlags) == 0)
7104 return ConstantInt::get(Ty: ReturnType, V: false);
7105 if (Q.isUndefValue(V: Op0))
7106 return UndefValue::get(T: ReturnType);
7107 break;
7108 }
7109 case Intrinsic::maxnum:
7110 case Intrinsic::minnum:
7111 case Intrinsic::maximum:
7112 case Intrinsic::minimum:
7113 case Intrinsic::maximumnum:
7114 case Intrinsic::minimumnum: {
7115 // In several cases here, we deviate from exact IEEE 754 semantics
7116 // to enable optimizations (as allowed by the LLVM IR spec).
7117 //
7118 // For instance, we may return one of the arguments unmodified instead of
7119 // inserting an llvm.canonicalize to transform input sNaNs into qNaNs,
7120 // or may assume all NaN inputs are qNaNs.
7121
7122 // If the arguments are the same, this is a no-op (ignoring NaN quieting)
7123 if (Op0 == Op1)
7124 return Op0;
7125
7126 // Canonicalize constant operand as Op1.
7127 if (isa<Constant>(Val: Op0))
7128 std::swap(a&: Op0, b&: Op1);
7129
7130 if (Constant *C = dyn_cast<Constant>(Val: Op1)) {
7131 MinMaxOptResult OptResult = MinMaxOptResult::CannotOptimize;
7132 Constant *NewConst = nullptr;
7133
7134 if (VectorType *VTy = dyn_cast<VectorType>(Val: C->getType())) {
7135 ElementCount ElemCount = VTy->getElementCount();
7136
7137 if (Constant *SplatVal = C->getSplatValue()) {
7138 // Handle splat vectors (including scalable vectors)
7139 OptResult = OptimizeConstMinMax(RHSConst: SplatVal, IID, Call, OutNewConstVal: &NewConst);
7140 if (OptResult == MinMaxOptResult::UseNewConstVal)
7141 NewConst = ConstantVector::getSplat(EC: ElemCount, Elt: NewConst);
7142
7143 } else if (ElemCount.isFixed()) {
7144 // Storage to build up new const return value (with NaNs quieted)
7145 SmallVector<Constant *, 16> NewC(ElemCount.getFixedValue());
7146
7147 // Check elementwise whether we can optimize to either a constant
7148 // value or return the LHS value. We cannot mix and match LHS +
7149 // constant elements, as this would require inserting a new
7150 // VectorShuffle instruction, which is not allowed in simplifyBinOp.
7151 OptResult = MinMaxOptResult::UseEither;
7152 for (unsigned i = 0; i != ElemCount.getFixedValue(); ++i) {
7153 auto *Elt = C->getAggregateElement(Elt: i);
7154 if (!Elt) {
7155 OptResult = MinMaxOptResult::CannotOptimize;
7156 break;
7157 }
7158 auto ElemResult = OptimizeConstMinMax(RHSConst: Elt, IID, Call, OutNewConstVal: &NewConst);
7159 if (ElemResult == MinMaxOptResult::CannotOptimize ||
7160 (ElemResult != OptResult &&
7161 OptResult != MinMaxOptResult::UseEither &&
7162 ElemResult != MinMaxOptResult::UseEither)) {
7163 OptResult = MinMaxOptResult::CannotOptimize;
7164 break;
7165 }
7166 NewC[i] = NewConst;
7167 if (ElemResult != MinMaxOptResult::UseEither)
7168 OptResult = ElemResult;
7169 }
7170 if (OptResult == MinMaxOptResult::UseNewConstVal)
7171 NewConst = ConstantVector::get(V: NewC);
7172 }
7173 } else {
7174 // Handle scalar inputs
7175 OptResult = OptimizeConstMinMax(RHSConst: C, IID, Call, OutNewConstVal: &NewConst);
7176 }
7177
7178 if (OptResult == MinMaxOptResult::UseOtherVal ||
7179 OptResult == MinMaxOptResult::UseEither)
7180 return Op0; // Return the other arg (ignoring NaN quieting)
7181 else if (OptResult == MinMaxOptResult::UseNewConstVal)
7182 return NewConst;
7183 }
7184
7185 // Min/max of the same operation with common operand:
7186 // m(m(X, Y)), X --> m(X, Y) (4 commuted variants)
7187 if (Value *V = foldMinimumMaximumSharedOp(IID, Op0, Op1))
7188 return V;
7189 if (Value *V = foldMinimumMaximumSharedOp(IID, Op0: Op1, Op1: Op0))
7190 return V;
7191
7192 break;
7193 }
7194 case Intrinsic::vector_extract: {
7195 // (extract_vector (insert_vector _, X, 0), 0) -> X
7196 unsigned IdxN = cast<ConstantInt>(Val: Op1)->getZExtValue();
7197 Value *X = nullptr;
7198 if (match(V: Op0, P: m_Intrinsic<Intrinsic::vector_insert>(Op0: m_Value(), Op1: m_Value(V&: X),
7199 Op2: m_Zero())) &&
7200 IdxN == 0 && X->getType() == ReturnType)
7201 return X;
7202
7203 break;
7204 }
7205
7206 case Intrinsic::aarch64_sve_andv:
7207 case Intrinsic::aarch64_sve_eorv:
7208 case Intrinsic::aarch64_sve_orv:
7209 case Intrinsic::aarch64_sve_saddv:
7210 case Intrinsic::aarch64_sve_smaxv:
7211 case Intrinsic::aarch64_sve_sminv:
7212 case Intrinsic::aarch64_sve_uaddv:
7213 case Intrinsic::aarch64_sve_umaxv:
7214 case Intrinsic::aarch64_sve_uminv:
7215 return simplifySVEIntReduction(IID, ReturnType, Op0, Op1);
7216 default:
7217 break;
7218 }
7219
7220 return nullptr;
7221}
7222
7223static Value *simplifyIntrinsic(CallBase *Call, Value *Callee,
7224 ArrayRef<Value *> Args,
7225 const SimplifyQuery &Q) {
7226 // Operand bundles should not be in Args.
7227 assert(Call->arg_size() == Args.size());
7228 unsigned NumOperands = Args.size();
7229 Function *F = cast<Function>(Val: Callee);
7230 Intrinsic::ID IID = F->getIntrinsicID();
7231
7232 if (IID != Intrinsic::not_intrinsic && intrinsicPropagatesPoison(IID) &&
7233 any_of(Range&: Args, P: IsaPred<PoisonValue>))
7234 return PoisonValue::get(T: F->getReturnType());
7235 // Most of the intrinsics with no operands have some kind of side effect.
7236 // Don't simplify.
7237 if (!NumOperands) {
7238 switch (IID) {
7239 case Intrinsic::vscale: {
7240 Type *RetTy = F->getReturnType();
7241 ConstantRange CR = getVScaleRange(F: Call->getFunction(), BitWidth: 64);
7242 if (const APInt *C = CR.getSingleElement())
7243 return ConstantInt::get(Ty: RetTy, V: C->getZExtValue());
7244 return nullptr;
7245 }
7246 default:
7247 return nullptr;
7248 }
7249 }
7250
7251 if (NumOperands == 1)
7252 return simplifyUnaryIntrinsic(F, Op0: Args[0], Q, Call);
7253
7254 if (NumOperands == 2)
7255 return simplifyBinaryIntrinsic(IID, ReturnType: F->getReturnType(), Op0: Args[0], Op1: Args[1], Q,
7256 Call);
7257
7258 // Handle intrinsics with 3 or more arguments.
7259 switch (IID) {
7260 case Intrinsic::masked_load:
7261 case Intrinsic::masked_gather: {
7262 Value *MaskArg = Args[1];
7263 Value *PassthruArg = Args[2];
7264 // If the mask is all zeros or undef, the "passthru" argument is the result.
7265 if (maskIsAllZeroOrUndef(Mask: MaskArg))
7266 return PassthruArg;
7267 return nullptr;
7268 }
7269 case Intrinsic::fshl:
7270 case Intrinsic::fshr: {
7271 Value *Op0 = Args[0], *Op1 = Args[1], *ShAmtArg = Args[2];
7272
7273 // If both operands are undef, the result is undef.
7274 if (Q.isUndefValue(V: Op0) && Q.isUndefValue(V: Op1))
7275 return UndefValue::get(T: F->getReturnType());
7276
7277 // If shift amount is undef, assume it is zero.
7278 if (Q.isUndefValue(V: ShAmtArg))
7279 return Args[IID == Intrinsic::fshl ? 0 : 1];
7280
7281 const APInt *ShAmtC;
7282 if (match(V: ShAmtArg, P: m_APInt(Res&: ShAmtC))) {
7283 // If there's effectively no shift, return the 1st arg or 2nd arg.
7284 APInt BitWidth = APInt(ShAmtC->getBitWidth(), ShAmtC->getBitWidth());
7285 if (ShAmtC->urem(RHS: BitWidth).isZero())
7286 return Args[IID == Intrinsic::fshl ? 0 : 1];
7287 }
7288
7289 // Rotating zero by anything is zero.
7290 if (match(V: Op0, P: m_Zero()) && match(V: Op1, P: m_Zero()))
7291 return ConstantInt::getNullValue(Ty: F->getReturnType());
7292
7293 // Rotating -1 by anything is -1.
7294 if (match(V: Op0, P: m_AllOnes()) && match(V: Op1, P: m_AllOnes()))
7295 return ConstantInt::getAllOnesValue(Ty: F->getReturnType());
7296
7297 return nullptr;
7298 }
7299 case Intrinsic::experimental_constrained_fma: {
7300 auto *FPI = cast<ConstrainedFPIntrinsic>(Val: Call);
7301 if (Value *V = simplifyFPOp(Ops: Args, FMF: {}, Q, ExBehavior: *FPI->getExceptionBehavior(),
7302 Rounding: *FPI->getRoundingMode()))
7303 return V;
7304 return nullptr;
7305 }
7306 case Intrinsic::fma:
7307 case Intrinsic::fmuladd: {
7308 if (Value *V = simplifyFPOp(Ops: Args, FMF: {}, Q, ExBehavior: fp::ebIgnore,
7309 Rounding: RoundingMode::NearestTiesToEven))
7310 return V;
7311 return nullptr;
7312 }
7313 case Intrinsic::smul_fix:
7314 case Intrinsic::smul_fix_sat: {
7315 Value *Op0 = Args[0];
7316 Value *Op1 = Args[1];
7317 Value *Op2 = Args[2];
7318 Type *ReturnType = F->getReturnType();
7319
7320 // Canonicalize constant operand as Op1 (ConstantFolding handles the case
7321 // when both Op0 and Op1 are constant so we do not care about that special
7322 // case here).
7323 if (isa<Constant>(Val: Op0))
7324 std::swap(a&: Op0, b&: Op1);
7325
7326 // X * 0 -> 0
7327 if (match(V: Op1, P: m_Zero()))
7328 return Constant::getNullValue(Ty: ReturnType);
7329
7330 // X * undef -> 0
7331 if (Q.isUndefValue(V: Op1))
7332 return Constant::getNullValue(Ty: ReturnType);
7333
7334 // X * (1 << Scale) -> X
7335 APInt ScaledOne =
7336 APInt::getOneBitSet(numBits: ReturnType->getScalarSizeInBits(),
7337 BitNo: cast<ConstantInt>(Val: Op2)->getZExtValue());
7338 if (ScaledOne.isNonNegative() && match(V: Op1, P: m_SpecificInt(V: ScaledOne)))
7339 return Op0;
7340
7341 return nullptr;
7342 }
7343 case Intrinsic::vector_insert: {
7344 Value *Vec = Args[0];
7345 Value *SubVec = Args[1];
7346 Value *Idx = Args[2];
7347 Type *ReturnType = F->getReturnType();
7348
7349 // (insert_vector Y, (extract_vector X, 0), 0) -> X
7350 // where: Y is X, or Y is undef
7351 unsigned IdxN = cast<ConstantInt>(Val: Idx)->getZExtValue();
7352 Value *X = nullptr;
7353 if (match(V: SubVec,
7354 P: m_Intrinsic<Intrinsic::vector_extract>(Op0: m_Value(V&: X), Op1: m_Zero())) &&
7355 (Q.isUndefValue(V: Vec) || Vec == X) && IdxN == 0 &&
7356 X->getType() == ReturnType)
7357 return X;
7358
7359 return nullptr;
7360 }
7361 case Intrinsic::vector_splice_left:
7362 case Intrinsic::vector_splice_right: {
7363 Value *Offset = Args[2];
7364 auto *Ty = cast<VectorType>(Val: F->getReturnType());
7365 if (Q.isUndefValue(V: Offset))
7366 return PoisonValue::get(T: Ty);
7367
7368 unsigned BitWidth = Offset->getType()->getScalarSizeInBits();
7369 ConstantRange NumElts(
7370 APInt(BitWidth, Ty->getElementCount().getKnownMinValue()));
7371 if (Ty->isScalableTy())
7372 NumElts = NumElts.multiply(Other: getVScaleRange(F: Call->getFunction(), BitWidth));
7373
7374 // If we know Offset > NumElts, simplify to poison.
7375 ConstantRange CR = computeConstantRangeIncludingKnownBits(V: Offset, ForSigned: false, SQ: Q);
7376 if (CR.getUnsignedMin().ugt(RHS: NumElts.getUnsignedMax()))
7377 return PoisonValue::get(T: Ty);
7378
7379 // splice.left(a, b, 0) --> a, splice.right(a, b, 0) --> b
7380 if (CR.isSingleElement() && CR.getSingleElement()->isZero())
7381 return IID == Intrinsic::vector_splice_left ? Args[0] : Args[1];
7382
7383 return nullptr;
7384 }
7385 case Intrinsic::experimental_constrained_fadd: {
7386 auto *FPI = cast<ConstrainedFPIntrinsic>(Val: Call);
7387 return simplifyFAddInst(Op0: Args[0], Op1: Args[1], FMF: FPI->getFastMathFlags(), Q,
7388 ExBehavior: *FPI->getExceptionBehavior(),
7389 Rounding: *FPI->getRoundingMode());
7390 }
7391 case Intrinsic::experimental_constrained_fsub: {
7392 auto *FPI = cast<ConstrainedFPIntrinsic>(Val: Call);
7393 return simplifyFSubInst(Op0: Args[0], Op1: Args[1], FMF: FPI->getFastMathFlags(), Q,
7394 ExBehavior: *FPI->getExceptionBehavior(),
7395 Rounding: *FPI->getRoundingMode());
7396 }
7397 case Intrinsic::experimental_constrained_fmul: {
7398 auto *FPI = cast<ConstrainedFPIntrinsic>(Val: Call);
7399 return simplifyFMulInst(Op0: Args[0], Op1: Args[1], FMF: FPI->getFastMathFlags(), Q,
7400 ExBehavior: *FPI->getExceptionBehavior(),
7401 Rounding: *FPI->getRoundingMode());
7402 }
7403 case Intrinsic::experimental_constrained_fdiv: {
7404 auto *FPI = cast<ConstrainedFPIntrinsic>(Val: Call);
7405 return simplifyFDivInst(Op0: Args[0], Op1: Args[1], FMF: FPI->getFastMathFlags(), Q,
7406 ExBehavior: *FPI->getExceptionBehavior(),
7407 Rounding: *FPI->getRoundingMode());
7408 }
7409 case Intrinsic::experimental_constrained_frem: {
7410 auto *FPI = cast<ConstrainedFPIntrinsic>(Val: Call);
7411 return simplifyFRemInst(Op0: Args[0], Op1: Args[1], FMF: FPI->getFastMathFlags(), Q,
7412 ExBehavior: *FPI->getExceptionBehavior(),
7413 Rounding: *FPI->getRoundingMode());
7414 }
7415 case Intrinsic::experimental_constrained_ldexp:
7416 return simplifyLdexp(Op0: Args[0], Op1: Args[1], Q, IsStrict: true);
7417 case Intrinsic::experimental_gc_relocate: {
7418 GCRelocateInst &GCR = *cast<GCRelocateInst>(Val: Call);
7419 Value *DerivedPtr = GCR.getDerivedPtr();
7420 Value *BasePtr = GCR.getBasePtr();
7421
7422 // Undef is undef, even after relocation.
7423 if (isa<UndefValue>(Val: DerivedPtr) || isa<UndefValue>(Val: BasePtr)) {
7424 return UndefValue::get(T: GCR.getType());
7425 }
7426
7427 if (auto *PT = dyn_cast<PointerType>(Val: GCR.getType())) {
7428 // For now, the assumption is that the relocation of null will be null
7429 // for most any collector. If this ever changes, a corresponding hook
7430 // should be added to GCStrategy and this code should check it first.
7431 if (isa<ConstantPointerNull>(Val: DerivedPtr)) {
7432 // Use null-pointer of gc_relocate's type to replace it.
7433 return ConstantPointerNull::get(T: PT);
7434 }
7435 }
7436 return nullptr;
7437 }
7438 case Intrinsic::experimental_vp_reverse: {
7439 Value *Vec = Call->getArgOperand(i: 0);
7440 Value *EVL = Call->getArgOperand(i: 2);
7441
7442 Value *X;
7443 // vp.reverse(vp.reverse(X)) == X (mask doesn't matter)
7444 if (match(V: Vec, P: m_Intrinsic<Intrinsic::experimental_vp_reverse>(
7445 Op0: m_Value(V&: X), Op1: m_Value(), Op2: m_Specific(V: EVL))))
7446 return X;
7447
7448 // vp.reverse(splat(X)) -> splat(X) (regardless of mask and EVL)
7449 if (isSplatValue(V: Vec))
7450 return Vec;
7451 return nullptr;
7452 }
7453 default:
7454 return nullptr;
7455 }
7456}
7457
7458static Value *tryConstantFoldCall(CallBase *Call, Value *Callee,
7459 ArrayRef<Value *> Args,
7460 const SimplifyQuery &Q) {
7461 auto *F = dyn_cast<Function>(Val: Callee);
7462 if (!F || !canConstantFoldCallTo(Call, F))
7463 return nullptr;
7464
7465 SmallVector<Constant *, 4> ConstantArgs;
7466 ConstantArgs.reserve(N: Args.size());
7467 for (Value *Arg : Args) {
7468 Constant *C = dyn_cast<Constant>(Val: Arg);
7469 if (!C) {
7470 if (isa<MetadataAsValue>(Val: Arg))
7471 continue;
7472 return nullptr;
7473 }
7474 ConstantArgs.push_back(Elt: C);
7475 }
7476
7477 return ConstantFoldCall(Call, F, Operands: ConstantArgs, TLI: Q.TLI);
7478}
7479
7480Value *llvm::simplifyCall(CallBase *Call, Value *Callee, ArrayRef<Value *> Args,
7481 const SimplifyQuery &Q) {
7482 // Args should not contain operand bundle operands.
7483 assert(Call->arg_size() == Args.size());
7484
7485 // musttail calls can only be simplified if they are also DCEd.
7486 // As we can't guarantee this here, don't simplify them.
7487 if (Call->isMustTailCall())
7488 return nullptr;
7489
7490 // call undef -> poison
7491 // call null -> poison
7492 if (isa<UndefValue>(Val: Callee) || isa<ConstantPointerNull>(Val: Callee))
7493 return PoisonValue::get(T: Call->getType());
7494
7495 if (Value *V = tryConstantFoldCall(Call, Callee, Args, Q))
7496 return V;
7497
7498 auto *F = dyn_cast<Function>(Val: Callee);
7499 if (F && F->isIntrinsic())
7500 if (Value *Ret = simplifyIntrinsic(Call, Callee, Args, Q))
7501 return Ret;
7502
7503 return nullptr;
7504}
7505
7506Value *llvm::simplifyConstrainedFPCall(CallBase *Call, const SimplifyQuery &Q) {
7507 assert(isa<ConstrainedFPIntrinsic>(Call));
7508 SmallVector<Value *, 4> Args(Call->args());
7509 if (Value *V = tryConstantFoldCall(Call, Callee: Call->getCalledOperand(), Args, Q))
7510 return V;
7511 if (Value *Ret = simplifyIntrinsic(Call, Callee: Call->getCalledOperand(), Args, Q))
7512 return Ret;
7513 return nullptr;
7514}
7515
7516/// Given operands for a Freeze, see if we can fold the result.
7517static Value *simplifyFreezeInst(Value *Op0, const SimplifyQuery &Q) {
7518 // Use a utility function defined in ValueTracking.
7519 if (llvm::isGuaranteedNotToBeUndefOrPoison(V: Op0, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT))
7520 return Op0;
7521 // We have room for improvement.
7522 return nullptr;
7523}
7524
7525Value *llvm::simplifyFreezeInst(Value *Op0, const SimplifyQuery &Q) {
7526 return ::simplifyFreezeInst(Op0, Q);
7527}
7528
7529Value *llvm::simplifyLoadInst(LoadInst *LI, Value *PtrOp,
7530 const SimplifyQuery &Q) {
7531 if (LI->isVolatile())
7532 return nullptr;
7533
7534 if (auto *PtrOpC = dyn_cast<Constant>(Val: PtrOp))
7535 return ConstantFoldLoadFromConstPtr(C: PtrOpC, Ty: LI->getType(), DL: Q.DL);
7536
7537 // We can only fold the load if it is from a constant global with definitive
7538 // initializer. Skip expensive logic if this is not the case.
7539 auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: PtrOp));
7540 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
7541 return nullptr;
7542
7543 // If GlobalVariable's initializer is uniform, then return the constant
7544 // regardless of its offset.
7545 if (Constant *C = ConstantFoldLoadFromUniformValue(C: GV->getInitializer(),
7546 Ty: LI->getType(), DL: Q.DL))
7547 return C;
7548
7549 // Try to convert operand into a constant by stripping offsets while looking
7550 // through invariant.group intrinsics.
7551 APInt Offset(Q.DL.getIndexTypeSizeInBits(Ty: PtrOp->getType()), 0);
7552 PtrOp = PtrOp->stripAndAccumulateConstantOffsets(
7553 DL: Q.DL, Offset, /* AllowNonInbounts */ AllowNonInbounds: true,
7554 /* AllowInvariantGroup */ true);
7555 if (PtrOp == GV) {
7556 // Index size may have changed due to address space casts.
7557 Offset = Offset.sextOrTrunc(width: Q.DL.getIndexTypeSizeInBits(Ty: PtrOp->getType()));
7558 return ConstantFoldLoadFromConstPtr(C: GV, Ty: LI->getType(), Offset: std::move(Offset),
7559 DL: Q.DL);
7560 }
7561
7562 return nullptr;
7563}
7564
7565/// See if we can compute a simplified version of this instruction.
7566/// If not, this returns null.
7567
7568static Value *simplifyInstructionWithOperands(Instruction *I,
7569 ArrayRef<Value *> NewOps,
7570 const SimplifyQuery &SQ,
7571 unsigned MaxRecurse) {
7572 assert(I->getFunction() && "instruction should be inserted in a function");
7573 assert((!SQ.CxtI || SQ.CxtI->getFunction() == I->getFunction()) &&
7574 "context instruction should be in the same function");
7575
7576 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
7577
7578 switch (I->getOpcode()) {
7579 default:
7580 if (all_of(Range&: NewOps, P: IsaPred<Constant>)) {
7581 SmallVector<Constant *, 8> NewConstOps(NewOps.size());
7582 transform(Range&: NewOps, d_first: NewConstOps.begin(),
7583 F: [](Value *V) { return cast<Constant>(Val: V); });
7584 return ConstantFoldInstOperands(I, Ops: NewConstOps, DL: Q.DL, TLI: Q.TLI);
7585 }
7586 return nullptr;
7587 case Instruction::FNeg:
7588 return simplifyFNegInst(Op: NewOps[0], FMF: I->getFastMathFlags(), Q, MaxRecurse);
7589 case Instruction::FAdd:
7590 return simplifyFAddInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7591 MaxRecurse);
7592 case Instruction::Add:
7593 return simplifyAddInst(
7594 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7595 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7596 case Instruction::FSub:
7597 return simplifyFSubInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7598 MaxRecurse);
7599 case Instruction::Sub:
7600 return simplifySubInst(
7601 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7602 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7603 case Instruction::FMul:
7604 return simplifyFMulInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7605 MaxRecurse);
7606 case Instruction::Mul:
7607 return simplifyMulInst(
7608 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7609 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7610 case Instruction::SDiv:
7611 return simplifySDivInst(Op0: NewOps[0], Op1: NewOps[1],
7612 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7613 MaxRecurse);
7614 case Instruction::UDiv:
7615 return simplifyUDivInst(Op0: NewOps[0], Op1: NewOps[1],
7616 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7617 MaxRecurse);
7618 case Instruction::FDiv:
7619 return simplifyFDivInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7620 MaxRecurse);
7621 case Instruction::SRem:
7622 return simplifySRemInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7623 case Instruction::URem:
7624 return simplifyURemInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7625 case Instruction::FRem:
7626 return simplifyFRemInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7627 MaxRecurse);
7628 case Instruction::Shl:
7629 return simplifyShlInst(
7630 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7631 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7632 case Instruction::LShr:
7633 return simplifyLShrInst(Op0: NewOps[0], Op1: NewOps[1],
7634 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7635 MaxRecurse);
7636 case Instruction::AShr:
7637 return simplifyAShrInst(Op0: NewOps[0], Op1: NewOps[1],
7638 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7639 MaxRecurse);
7640 case Instruction::And:
7641 return simplifyAndInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7642 case Instruction::Or:
7643 return simplifyOrInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7644 case Instruction::Xor:
7645 return simplifyXorInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7646 case Instruction::ICmp:
7647 return simplifyICmpInst(Pred: cast<ICmpInst>(Val: I)->getCmpPredicate(), LHS: NewOps[0],
7648 RHS: NewOps[1], Q, MaxRecurse);
7649 case Instruction::FCmp:
7650 return simplifyFCmpInst(Pred: cast<FCmpInst>(Val: I)->getPredicate(), LHS: NewOps[0],
7651 RHS: NewOps[1], FMF: I->getFastMathFlags(), Q, MaxRecurse);
7652 case Instruction::Select:
7653 return simplifySelectInst(Cond: NewOps[0], TrueVal: NewOps[1], FalseVal: NewOps[2], Q, MaxRecurse);
7654 case Instruction::GetElementPtr: {
7655 auto *GEPI = cast<GetElementPtrInst>(Val: I);
7656 return simplifyGEPInst(SrcTy: GEPI->getSourceElementType(), Ptr: NewOps[0],
7657 Indices: ArrayRef(NewOps).slice(N: 1), NW: GEPI->getNoWrapFlags(), Q,
7658 MaxRecurse);
7659 }
7660 case Instruction::InsertValue: {
7661 InsertValueInst *IV = cast<InsertValueInst>(Val: I);
7662 return simplifyInsertValueInst(Agg: NewOps[0], Val: NewOps[1], Idxs: IV->getIndices(), Q,
7663 MaxRecurse);
7664 }
7665 case Instruction::InsertElement:
7666 return simplifyInsertElementInst(Vec: NewOps[0], Val: NewOps[1], Idx: NewOps[2], Q);
7667 case Instruction::ExtractValue: {
7668 auto *EVI = cast<ExtractValueInst>(Val: I);
7669 return simplifyExtractValueInst(Agg: NewOps[0], Idxs: EVI->getIndices(), Q,
7670 MaxRecurse);
7671 }
7672 case Instruction::ExtractElement:
7673 return simplifyExtractElementInst(Vec: NewOps[0], Idx: NewOps[1], Q, MaxRecurse);
7674 case Instruction::ShuffleVector: {
7675 auto *SVI = cast<ShuffleVectorInst>(Val: I);
7676 return simplifyShuffleVectorInst(Op0: NewOps[0], Op1: NewOps[1],
7677 Mask: SVI->getShuffleMask(), RetTy: SVI->getType(), Q,
7678 MaxRecurse);
7679 }
7680 case Instruction::PHI:
7681 return simplifyPHINode(PN: cast<PHINode>(Val: I), IncomingValues: NewOps, Q);
7682 case Instruction::Call:
7683 return simplifyCall(
7684 Call: cast<CallInst>(Val: I), Callee: NewOps.back(),
7685 Args: NewOps.drop_back(N: 1 + cast<CallInst>(Val: I)->getNumTotalBundleOperands()), Q);
7686 case Instruction::Freeze:
7687 return llvm::simplifyFreezeInst(Op0: NewOps[0], Q);
7688#define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
7689#include "llvm/IR/Instruction.def"
7690#undef HANDLE_CAST_INST
7691 return simplifyCastInst(CastOpc: I->getOpcode(), Op: NewOps[0], Ty: I->getType(), Q,
7692 MaxRecurse);
7693 case Instruction::Alloca:
7694 // No simplifications for Alloca and it can't be constant folded.
7695 return nullptr;
7696 case Instruction::Load:
7697 return simplifyLoadInst(LI: cast<LoadInst>(Val: I), PtrOp: NewOps[0], Q);
7698 }
7699}
7700
7701Value *llvm::simplifyInstructionWithOperands(Instruction *I,
7702 ArrayRef<Value *> NewOps,
7703 const SimplifyQuery &SQ) {
7704 assert(NewOps.size() == I->getNumOperands() &&
7705 "Number of operands should match the instruction!");
7706 return ::simplifyInstructionWithOperands(I, NewOps, SQ, MaxRecurse: RecursionLimit);
7707}
7708
7709Value *llvm::simplifyInstruction(Instruction *I, const SimplifyQuery &SQ) {
7710 SmallVector<Value *, 8> Ops(I->operands());
7711 Value *Result = ::simplifyInstructionWithOperands(I, NewOps: Ops, SQ, MaxRecurse: RecursionLimit);
7712
7713 /// If called on unreachable code, the instruction may simplify to itself.
7714 /// Make life easier for users by detecting that case here, and returning a
7715 /// safe value instead.
7716 return Result == I ? PoisonValue::get(T: I->getType()) : Result;
7717}
7718
7719/// Implementation of recursive simplification through an instruction's
7720/// uses.
7721///
7722/// This is the common implementation of the recursive simplification routines.
7723/// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
7724/// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
7725/// instructions to process and attempt to simplify it using
7726/// InstructionSimplify. Recursively visited users which could not be
7727/// simplified themselves are to the optional UnsimplifiedUsers set for
7728/// further processing by the caller.
7729///
7730/// This routine returns 'true' only when *it* simplifies something. The passed
7731/// in simplified value does not count toward this.
7732static bool replaceAndRecursivelySimplifyImpl(
7733 Instruction *I, Value *SimpleV, const TargetLibraryInfo *TLI,
7734 const DominatorTree *DT, AssumptionCache *AC,
7735 SmallSetVector<Instruction *, 8> *UnsimplifiedUsers = nullptr) {
7736 bool Simplified = false;
7737 SmallSetVector<Instruction *, 8> Worklist;
7738 const DataLayout &DL = I->getDataLayout();
7739
7740 // If we have an explicit value to collapse to, do that round of the
7741 // simplification loop by hand initially.
7742 if (SimpleV) {
7743 for (User *U : I->users())
7744 if (U != I)
7745 Worklist.insert(X: cast<Instruction>(Val: U));
7746
7747 // Replace the instruction with its simplified value.
7748 I->replaceAllUsesWith(V: SimpleV);
7749
7750 if (!I->isEHPad() && !I->isTerminator() && !I->mayHaveSideEffects())
7751 I->eraseFromParent();
7752 } else {
7753 Worklist.insert(X: I);
7754 }
7755
7756 // Note that we must test the size on each iteration, the worklist can grow.
7757 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
7758 I = Worklist[Idx];
7759
7760 // See if this instruction simplifies.
7761 SimpleV = simplifyInstruction(I, SQ: {DL, TLI, DT, AC});
7762 if (!SimpleV) {
7763 if (UnsimplifiedUsers)
7764 UnsimplifiedUsers->insert(X: I);
7765 continue;
7766 }
7767
7768 Simplified = true;
7769
7770 // Stash away all the uses of the old instruction so we can check them for
7771 // recursive simplifications after a RAUW. This is cheaper than checking all
7772 // uses of To on the recursive step in most cases.
7773 for (User *U : I->users())
7774 Worklist.insert(X: cast<Instruction>(Val: U));
7775
7776 // Replace the instruction with its simplified value.
7777 I->replaceAllUsesWith(V: SimpleV);
7778
7779 if (!I->isEHPad() && !I->isTerminator() && !I->mayHaveSideEffects())
7780 I->eraseFromParent();
7781 }
7782 return Simplified;
7783}
7784
7785bool llvm::replaceAndRecursivelySimplify(
7786 Instruction *I, Value *SimpleV, const TargetLibraryInfo *TLI,
7787 const DominatorTree *DT, AssumptionCache *AC,
7788 SmallSetVector<Instruction *, 8> *UnsimplifiedUsers) {
7789 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
7790 assert(SimpleV && "Must provide a simplified value.");
7791 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC,
7792 UnsimplifiedUsers);
7793}
7794
7795namespace llvm {
7796const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
7797 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
7798 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
7799 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
7800 auto *TLI = TLIWP ? &TLIWP->getTLI(F) : nullptr;
7801 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
7802 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
7803 return {F.getDataLayout(), TLI, DT, AC};
7804}
7805
7806const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
7807 const DataLayout &DL) {
7808 return {DL, &AR.TLI, &AR.DT, &AR.AC};
7809}
7810
7811template <class T, class... TArgs>
7812const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
7813 Function &F) {
7814 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
7815 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
7816 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
7817 return {F.getDataLayout(), TLI, DT, AC};
7818}
7819template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,
7820 Function &);
7821
7822bool SimplifyQuery::isUndefValue(Value *V) const {
7823 if (!CanUseUndef)
7824 return false;
7825
7826 return match(V, P: m_Undef());
7827}
7828
7829} // namespace llvm
7830
7831void InstSimplifyFolder::anchor() {}
7832