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 *, FastMathFlags,
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_Ctpop(Op0: m_Value(V&: X)), R: m_APInt(Res&: C))) ||
1712 !match(V: Cmp1, P: m_ICmp(Pred&: Pred1, L: m_Specific(V: X), R: m_ZeroInt())) || C->isZero())
1713 return nullptr;
1714
1715 // (ctpop(X) == C) || (X != 0) --> X != 0 where C > 0
1716 if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_NE)
1717 return Cmp1;
1718 // (ctpop(X) != C) && (X == 0) --> X == 0 where C > 0
1719 if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_EQ)
1720 return Cmp1;
1721
1722 return nullptr;
1723}
1724
1725static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1726 const SimplifyQuery &Q) {
1727 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op0, UnsignedICmp: Op1, /*IsAnd=*/true, Q))
1728 return X;
1729 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op1, UnsignedICmp: Op0, /*IsAnd=*/true, Q))
1730 return X;
1731
1732 if (Value *X = simplifyAndOrOfICmpsWithConstants(Cmp0: Op0, Cmp1: Op1, IsAnd: true))
1733 return X;
1734
1735 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op0, Cmp1: Op1, IsAnd: true))
1736 return X;
1737 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op1, Cmp1: Op0, IsAnd: true))
1738 return X;
1739
1740 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ: Q.IIQ))
1741 return X;
1742 if (Value *X = simplifyAndOfICmpsWithAdd(Op0: Op1, Op1: Op0, IIQ: Q.IIQ))
1743 return X;
1744
1745 return nullptr;
1746}
1747
1748static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1749 const InstrInfoQuery &IIQ) {
1750 // (icmp (add V, C0), C1) | (icmp V, C0)
1751 CmpPredicate Pred0, Pred1;
1752 const APInt *C0, *C1;
1753 Value *V;
1754 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))))
1755 return nullptr;
1756
1757 if (!match(V: Op1, P: m_ICmp(Pred&: Pred1, L: m_Specific(V), R: m_Value())))
1758 return nullptr;
1759
1760 auto *AddInst = cast<BinaryOperator>(Val: Op0->getOperand(i_nocapture: 0));
1761 if (AddInst->getOperand(i_nocapture: 1) != Op1->getOperand(i_nocapture: 1))
1762 return nullptr;
1763
1764 Type *ITy = Op0->getType();
1765 bool IsNSW = IIQ.hasNoSignedWrap(Op: AddInst);
1766 bool IsNUW = IIQ.hasNoUnsignedWrap(Op: AddInst);
1767
1768 const APInt Delta = *C1 - *C0;
1769 if (C0->isStrictlyPositive()) {
1770 if (Delta == 2) {
1771 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1772 return getTrue(Ty: ITy);
1773 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && IsNSW)
1774 return getTrue(Ty: ITy);
1775 }
1776 if (Delta == 1) {
1777 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1778 return getTrue(Ty: ITy);
1779 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && IsNSW)
1780 return getTrue(Ty: ITy);
1781 }
1782 }
1783 if (C0->getBoolValue() && IsNUW) {
1784 if (Delta == 2)
1785 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1786 return getTrue(Ty: ITy);
1787 if (Delta == 1)
1788 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1789 return getTrue(Ty: ITy);
1790 }
1791
1792 return nullptr;
1793}
1794
1795static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1796 const SimplifyQuery &Q) {
1797 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op0, UnsignedICmp: Op1, /*IsAnd=*/false, Q))
1798 return X;
1799 if (Value *X = simplifyUnsignedRangeCheck(ZeroICmp: Op1, UnsignedICmp: Op0, /*IsAnd=*/false, Q))
1800 return X;
1801
1802 if (Value *X = simplifyAndOrOfICmpsWithConstants(Cmp0: Op0, Cmp1: Op1, IsAnd: false))
1803 return X;
1804
1805 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op0, Cmp1: Op1, IsAnd: false))
1806 return X;
1807 if (Value *X = simplifyAndOrOfICmpsWithCtpop(Cmp0: Op1, Cmp1: Op0, IsAnd: false))
1808 return X;
1809
1810 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ: Q.IIQ))
1811 return X;
1812 if (Value *X = simplifyOrOfICmpsWithAdd(Op0: Op1, Op1: Op0, IIQ: Q.IIQ))
1813 return X;
1814
1815 return nullptr;
1816}
1817
1818/// Test if a pair of compares with a shared operand and 2 constants has an
1819/// empty set intersection, full set union, or if one compare is a superset of
1820/// the other.
1821static Value *simplifyAndOrOfFCmpsWithConstants(FCmpInst *Cmp0, FCmpInst *Cmp1,
1822 bool IsAnd) {
1823 // Look for this pattern: {and/or} (fcmp X, C0), (fcmp X, C1)).
1824 if (Cmp0->getOperand(i_nocapture: 0) != Cmp1->getOperand(i_nocapture: 0))
1825 return nullptr;
1826
1827 const APFloat *C0, *C1;
1828 if (!match(V: Cmp0->getOperand(i_nocapture: 1), P: m_APFloat(Res&: C0)) ||
1829 !match(V: Cmp1->getOperand(i_nocapture: 1), P: m_APFloat(Res&: C1)))
1830 return nullptr;
1831
1832 auto Range0 = ConstantFPRange::makeExactFCmpRegion(
1833 Pred: IsAnd ? Cmp0->getPredicate() : Cmp0->getInversePredicate(), Other: *C0);
1834 auto Range1 = ConstantFPRange::makeExactFCmpRegion(
1835 Pred: IsAnd ? Cmp1->getPredicate() : Cmp1->getInversePredicate(), Other: *C1);
1836
1837 if (!Range0 || !Range1)
1838 return nullptr;
1839
1840 // For and-of-compares, check if the intersection is empty:
1841 // (fcmp X, C0) && (fcmp X, C1) --> empty set --> false
1842 if (Range0->intersectWith(CR: *Range1).isEmptySet())
1843 return ConstantInt::getBool(Ty: Cmp0->getType(), V: !IsAnd);
1844
1845 // Is one range a superset of the other?
1846 // If this is and-of-compares, take the smaller set:
1847 // (fcmp ogt X, 4) && (fcmp ogt X, 42) --> fcmp ogt X, 42
1848 // If this is or-of-compares, take the larger set:
1849 // (fcmp ogt X, 4) || (fcmp ogt X, 42) --> fcmp ogt X, 4
1850 if (Range0->contains(CR: *Range1))
1851 return Cmp1;
1852 if (Range1->contains(CR: *Range0))
1853 return Cmp0;
1854
1855 return nullptr;
1856}
1857
1858static Value *simplifyAndOrOfFCmps(const SimplifyQuery &Q, FCmpInst *LHS,
1859 FCmpInst *RHS, bool IsAnd) {
1860 Value *LHS0 = LHS->getOperand(i_nocapture: 0), *LHS1 = LHS->getOperand(i_nocapture: 1);
1861 Value *RHS0 = RHS->getOperand(i_nocapture: 0), *RHS1 = RHS->getOperand(i_nocapture: 1);
1862 if (LHS0->getType() != RHS0->getType())
1863 return nullptr;
1864
1865 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1866 auto AbsOrSelfLHS0 = m_CombineOr(Ps: m_Specific(V: LHS0), Ps: m_FAbs(Op0: m_Specific(V: LHS0)));
1867 if ((PredL == FCmpInst::FCMP_ORD || PredL == FCmpInst::FCMP_UNO) &&
1868 ((FCmpInst::isOrdered(predicate: PredR) && IsAnd) ||
1869 (FCmpInst::isUnordered(predicate: PredR) && !IsAnd))) {
1870 // (fcmp ord X, 0) & (fcmp o** X/abs(X), Y) --> fcmp o** X/abs(X), Y
1871 // (fcmp uno X, 0) & (fcmp o** X/abs(X), Y) --> false
1872 // (fcmp uno X, 0) | (fcmp u** X/abs(X), Y) --> fcmp u** X/abs(X), Y
1873 // (fcmp ord X, 0) | (fcmp u** X/abs(X), Y) --> true
1874 if ((match(V: RHS0, P: AbsOrSelfLHS0) || match(V: RHS1, P: AbsOrSelfLHS0)) &&
1875 match(V: LHS1, P: m_PosZeroFP()))
1876 return FCmpInst::isOrdered(predicate: PredL) == FCmpInst::isOrdered(predicate: PredR)
1877 ? static_cast<Value *>(RHS)
1878 : ConstantInt::getBool(Ty: LHS->getType(), V: !IsAnd);
1879 }
1880
1881 auto AbsOrSelfRHS0 = m_CombineOr(Ps: m_Specific(V: RHS0), Ps: m_FAbs(Op0: m_Specific(V: RHS0)));
1882 if ((PredR == FCmpInst::FCMP_ORD || PredR == FCmpInst::FCMP_UNO) &&
1883 ((FCmpInst::isOrdered(predicate: PredL) && IsAnd) ||
1884 (FCmpInst::isUnordered(predicate: PredL) && !IsAnd))) {
1885 // (fcmp o** X/abs(X), Y) & (fcmp ord X, 0) --> fcmp o** X/abs(X), Y
1886 // (fcmp o** X/abs(X), Y) & (fcmp uno X, 0) --> false
1887 // (fcmp u** X/abs(X), Y) | (fcmp uno X, 0) --> fcmp u** X/abs(X), Y
1888 // (fcmp u** X/abs(X), Y) | (fcmp ord X, 0) --> true
1889 if ((match(V: LHS0, P: AbsOrSelfRHS0) || match(V: LHS1, P: AbsOrSelfRHS0)) &&
1890 match(V: RHS1, P: m_PosZeroFP()))
1891 return FCmpInst::isOrdered(predicate: PredL) == FCmpInst::isOrdered(predicate: PredR)
1892 ? static_cast<Value *>(LHS)
1893 : ConstantInt::getBool(Ty: LHS->getType(), V: !IsAnd);
1894 }
1895
1896 if (auto *V = simplifyAndOrOfFCmpsWithConstants(Cmp0: LHS, Cmp1: RHS, IsAnd))
1897 return V;
1898
1899 return nullptr;
1900}
1901
1902static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q, Value *Op0,
1903 Value *Op1, bool IsAnd) {
1904 // Look through casts of the 'and' operands to find compares.
1905 auto *Cast0 = dyn_cast<CastInst>(Val: Op0);
1906 auto *Cast1 = dyn_cast<CastInst>(Val: Op1);
1907 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1908 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1909 Op0 = Cast0->getOperand(i_nocapture: 0);
1910 Op1 = Cast1->getOperand(i_nocapture: 0);
1911 }
1912
1913 Value *V = nullptr;
1914 auto *ICmp0 = dyn_cast<ICmpInst>(Val: Op0);
1915 auto *ICmp1 = dyn_cast<ICmpInst>(Val: Op1);
1916 if (ICmp0 && ICmp1)
1917 V = IsAnd ? simplifyAndOfICmps(Op0: ICmp0, Op1: ICmp1, Q)
1918 : simplifyOrOfICmps(Op0: ICmp0, Op1: ICmp1, Q);
1919
1920 auto *FCmp0 = dyn_cast<FCmpInst>(Val: Op0);
1921 auto *FCmp1 = dyn_cast<FCmpInst>(Val: Op1);
1922 if (FCmp0 && FCmp1)
1923 V = simplifyAndOrOfFCmps(Q, LHS: FCmp0, RHS: FCmp1, IsAnd);
1924
1925 if (!V)
1926 return nullptr;
1927 if (!Cast0)
1928 return V;
1929
1930 // If we looked through casts, we can only handle a constant simplification
1931 // because we are not allowed to create a cast instruction here.
1932 if (auto *C = dyn_cast<Constant>(Val: V))
1933 return ConstantFoldCastOperand(Opcode: Cast0->getOpcode(), C, DestTy: Cast0->getType(),
1934 DL: Q.DL);
1935
1936 return nullptr;
1937}
1938
1939static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
1940 const SimplifyQuery &Q,
1941 bool AllowRefinement,
1942 SmallVectorImpl<Instruction *> *DropFlags,
1943 unsigned MaxRecurse);
1944
1945static Value *simplifyAndOrWithICmpEq(unsigned Opcode, Value *Op0, Value *Op1,
1946 const SimplifyQuery &Q,
1947 unsigned MaxRecurse) {
1948 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1949 "Must be and/or");
1950 CmpPredicate Pred;
1951 Value *A, *B;
1952 if (!match(V: Op0, P: m_ICmpLike(Pred, L: m_Value(V&: A), R: m_Value(V&: B))) ||
1953 !ICmpInst::isEquality(P: Pred))
1954 return nullptr;
1955
1956 auto Simplify = [&](Value *Res) -> Value * {
1957 Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, Ty: Res->getType());
1958
1959 // and (icmp eq a, b), x implies (a==b) inside x.
1960 // or (icmp ne a, b), x implies (a==b) inside x.
1961 // If x simplifies to true/false, we can simplify the and/or.
1962 if (Pred ==
1963 (Opcode == Instruction::And ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
1964 if (Res == Absorber)
1965 return Absorber;
1966 if (Res == ConstantExpr::getBinOpIdentity(Opcode, Ty: Res->getType()))
1967 return Op0;
1968 return nullptr;
1969 }
1970
1971 // If we have and (icmp ne a, b), x and for a==b we can simplify x to false,
1972 // then we can drop the icmp, as x will already be false in the case where
1973 // the icmp is false. Similar for or and true.
1974 if (Res == Absorber)
1975 return Op1;
1976 return nullptr;
1977 };
1978
1979 // In the final case (Res == Absorber with inverted predicate), it is safe to
1980 // refine poison during simplification, but not undef. For simplicity always
1981 // disable undef-based folds here.
1982 if (Value *Res = simplifyWithOpReplaced(V: Op1, Op: A, RepOp: B, Q: Q.getWithoutUndef(),
1983 /* AllowRefinement */ true,
1984 /* DropFlags */ nullptr, MaxRecurse))
1985 return Simplify(Res);
1986 if (Value *Res = simplifyWithOpReplaced(V: Op1, Op: B, RepOp: A, Q: Q.getWithoutUndef(),
1987 /* AllowRefinement */ true,
1988 /* DropFlags */ nullptr, MaxRecurse))
1989 return Simplify(Res);
1990
1991 return nullptr;
1992}
1993
1994/// Given a bitwise logic op, check if the operands are add/sub with a common
1995/// source value and inverted constant (identity: C - X -> ~(X + ~C)).
1996static Value *simplifyLogicOfAddSub(Value *Op0, Value *Op1,
1997 Instruction::BinaryOps Opcode) {
1998 assert(Op0->getType() == Op1->getType() && "Mismatched binop types");
1999 assert(BinaryOperator::isBitwiseLogicOp(Opcode) && "Expected logic op");
2000 Value *X;
2001 Constant *C1, *C2;
2002 if ((match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Constant(C&: C1))) &&
2003 match(V: Op1, P: m_Sub(L: m_Constant(C&: C2), R: m_Specific(V: X)))) ||
2004 (match(V: Op1, P: m_Add(L: m_Value(V&: X), R: m_Constant(C&: C1))) &&
2005 match(V: Op0, P: m_Sub(L: m_Constant(C&: C2), R: m_Specific(V: X))))) {
2006 if (ConstantExpr::getNot(C: C1) == C2) {
2007 // (X + C) & (~C - X) --> (X + C) & ~(X + C) --> 0
2008 // (X + C) | (~C - X) --> (X + C) | ~(X + C) --> -1
2009 // (X + C) ^ (~C - X) --> (X + C) ^ ~(X + C) --> -1
2010 Type *Ty = Op0->getType();
2011 return Opcode == Instruction::And ? ConstantInt::getNullValue(Ty)
2012 : ConstantInt::getAllOnesValue(Ty);
2013 }
2014 }
2015 return nullptr;
2016}
2017
2018// Commutative patterns for and that will be tried with both operand orders.
2019static Value *simplifyAndCommutative(Value *Op0, Value *Op1,
2020 const SimplifyQuery &Q,
2021 unsigned MaxRecurse) {
2022 // ~A & A = 0
2023 if (match(V: Op0, P: m_Not(V: m_Specific(V: Op1))))
2024 return Constant::getNullValue(Ty: Op0->getType());
2025
2026 // (A | ?) & A = A
2027 if (match(V: Op0, P: m_c_Or(L: m_Specific(V: Op1), R: m_Value())))
2028 return Op1;
2029
2030 // (X | ~Y) & (X | Y) --> X
2031 Value *X, *Y;
2032 if (match(V: Op0, P: m_c_Or(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
2033 match(V: Op1, P: m_c_Or(L: m_Specific(V: X), R: m_Specific(V: Y))))
2034 return X;
2035
2036 // If we have a multiplication overflow check that is being 'and'ed with a
2037 // check that one of the multipliers is not zero, we can omit the 'and', and
2038 // only keep the overflow check.
2039 if (isCheckForZeroAndMulWithOverflow(Op0, Op1, IsAnd: true))
2040 return Op1;
2041
2042 // -A & A = A if A is a power of two or zero.
2043 if (match(V: Op0, P: m_Neg(V: m_Specific(V: Op1))) &&
2044 isKnownToBeAPowerOfTwo(V: Op1, DL: Q.DL, /*OrZero*/ true, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
2045 return Op1;
2046
2047 // This is a similar pattern used for checking if a value is a power-of-2:
2048 // (A - 1) & A --> 0 (if A is a power-of-2 or 0)
2049 if (match(V: Op0, P: m_Add(L: m_Specific(V: Op1), R: m_AllOnes())) &&
2050 isKnownToBeAPowerOfTwo(V: Op1, DL: Q.DL, /*OrZero*/ true, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
2051 return Constant::getNullValue(Ty: Op1->getType());
2052
2053 // (x << N) & ((x << M) - 1) --> 0, where x is known to be a power of 2 and
2054 // M <= N.
2055 const APInt *Shift1, *Shift2;
2056 if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: Shift1))) &&
2057 match(V: Op1, P: m_Add(L: m_Shl(L: m_Specific(V: X), R: m_APInt(Res&: Shift2)), R: m_AllOnes())) &&
2058 isKnownToBeAPowerOfTwo(V: X, DL: Q.DL, /*OrZero*/ true, AC: Q.AC, CxtI: Q.CxtI) &&
2059 Shift1->uge(RHS: *Shift2))
2060 return Constant::getNullValue(Ty: Op0->getType());
2061
2062 if (Value *V =
2063 simplifyAndOrWithICmpEq(Opcode: Instruction::And, Op0, Op1, Q, MaxRecurse))
2064 return V;
2065
2066 return nullptr;
2067}
2068
2069/// Given operands for an And, see if we can fold the result.
2070/// If not, this returns null.
2071static Value *simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2072 unsigned MaxRecurse) {
2073 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::And, Op0, Op1, Q))
2074 return C;
2075
2076 // X & poison -> poison
2077 if (isa<PoisonValue>(Val: Op1))
2078 return Op1;
2079
2080 // X & undef -> 0
2081 if (Q.isUndefValue(V: Op1))
2082 return Constant::getNullValue(Ty: Op0->getType());
2083
2084 // X & X = X
2085 if (Op0 == Op1)
2086 return Op0;
2087
2088 // X & 0 = 0
2089 if (match(V: Op1, P: m_Zero()))
2090 return Constant::getNullValue(Ty: Op0->getType());
2091
2092 // X & -1 = X
2093 if (match(V: Op1, P: m_AllOnes()))
2094 return Op0;
2095
2096 if (Value *Res = simplifyAndCommutative(Op0, Op1, Q, MaxRecurse))
2097 return Res;
2098 if (Value *Res = simplifyAndCommutative(Op0: Op1, Op1: Op0, Q, MaxRecurse))
2099 return Res;
2100
2101 if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Opcode: Instruction::And))
2102 return V;
2103
2104 // A mask that only clears known zeros of a shifted value is a no-op.
2105 const APInt *Mask;
2106 const APInt *ShAmt;
2107 Value *X, *Y;
2108 if (match(V: Op1, P: m_APInt(Res&: Mask))) {
2109 // If all bits in the inverted and shifted mask are clear:
2110 // and (shl X, ShAmt), Mask --> shl X, ShAmt
2111 if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: ShAmt))) &&
2112 (~(*Mask)).lshr(ShiftAmt: *ShAmt).isZero())
2113 return Op0;
2114
2115 // If all bits in the inverted and shifted mask are clear:
2116 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
2117 if (match(V: Op0, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: ShAmt))) &&
2118 (~(*Mask)).shl(ShiftAmt: *ShAmt).isZero())
2119 return Op0;
2120 }
2121
2122 // and 2^x-1, 2^C --> 0 where x <= C.
2123 const APInt *PowerC;
2124 Value *Shift;
2125 if (match(V: Op1, P: m_Power2(V&: PowerC)) &&
2126 match(V: Op0, P: m_Add(L: m_Value(V&: Shift), R: m_AllOnes())) &&
2127 isKnownToBeAPowerOfTwo(V: Shift, DL: Q.DL, /*OrZero*/ false, AC: Q.AC, CxtI: Q.CxtI,
2128 DT: Q.DT)) {
2129 KnownBits Known = computeKnownBits(V: Shift, Q);
2130 // Use getActiveBits() to make use of the additional power of two knowledge
2131 if (PowerC->getActiveBits() >= Known.getMaxValue().getActiveBits())
2132 return ConstantInt::getNullValue(Ty: Op1->getType());
2133 }
2134
2135 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, IsAnd: true))
2136 return V;
2137
2138 // zext(X) & sext(X) --> zext(X)
2139 // sext(X) & zext(X) --> zext(X)
2140 {
2141 Value *X = nullptr;
2142 if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_SExt(Op: m_Specific(V: X))))
2143 return Op0;
2144 if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op0, P: m_SExt(Op: m_Specific(V: X))))
2145 return Op1;
2146 }
2147
2148 // Try some generic simplifications for associative operations.
2149 if (Value *V =
2150 simplifyAssociativeBinOp(Opcode: Instruction::And, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2151 return V;
2152
2153 // And distributes over Or. Try some generic simplifications based on this.
2154 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::And, L: Op0, R: Op1,
2155 OpcodeToExpand: Instruction::Or, Q, MaxRecurse))
2156 return V;
2157
2158 // And distributes over Xor. Try some generic simplifications based on this.
2159 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::And, L: Op0, R: Op1,
2160 OpcodeToExpand: Instruction::Xor, Q, MaxRecurse))
2161 return V;
2162
2163 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1)) {
2164 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2165 // A & (A && B) -> A && B
2166 if (match(V: Op1, P: m_Select(C: m_Specific(V: Op0), L: m_Value(), R: m_Zero())))
2167 return Op1;
2168 else if (match(V: Op0, P: m_Select(C: m_Specific(V: Op1), L: m_Value(), R: m_Zero())))
2169 return Op0;
2170 }
2171 // If the operation is with the result of a select instruction, check
2172 // whether operating on either branch of the select always yields the same
2173 // value.
2174 if (Value *V =
2175 threadBinOpOverSelect(Opcode: Instruction::And, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2176 return V;
2177 }
2178
2179 // If the operation is with the result of a phi instruction, check whether
2180 // operating on all incoming values of the phi always yields the same value.
2181 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
2182 if (Value *V =
2183 threadBinOpOverPHI(Opcode: Instruction::And, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2184 return V;
2185
2186 // Assuming the effective width of Y is not larger than A, i.e. all bits
2187 // from X and Y are disjoint in (X << A) | Y,
2188 // if the mask of this AND op covers all bits of X or Y, while it covers
2189 // no bits from the other, we can bypass this AND op. E.g.,
2190 // ((X << A) | Y) & Mask -> Y,
2191 // if Mask = ((1 << effective_width_of(Y)) - 1)
2192 // ((X << A) | Y) & Mask -> X << A,
2193 // if Mask = ((1 << effective_width_of(X)) - 1) << A
2194 // SimplifyDemandedBits in InstCombine can optimize the general case.
2195 // This pattern aims to help other passes for a common case.
2196 Value *XShifted;
2197 if (Q.IIQ.UseInstrInfo && match(V: Op1, P: m_APInt(Res&: Mask)) &&
2198 match(V: Op0, P: m_c_Or(L: m_CombineAnd(Ps: m_NUWShl(L: m_Value(V&: X), R: m_APInt(Res&: ShAmt)),
2199 Ps: m_Value(V&: XShifted)),
2200 R: m_Value(V&: Y)))) {
2201 const unsigned Width = Op0->getType()->getScalarSizeInBits();
2202 const unsigned ShftCnt = ShAmt->getLimitedValue(Limit: Width);
2203 const KnownBits YKnown = computeKnownBits(V: Y, Q);
2204 const unsigned EffWidthY = YKnown.countMaxActiveBits();
2205 if (EffWidthY <= ShftCnt) {
2206 const KnownBits XKnown = computeKnownBits(V: X, Q);
2207 const unsigned EffWidthX = XKnown.countMaxActiveBits();
2208 const APInt EffBitsY = APInt::getLowBitsSet(numBits: Width, loBitsSet: EffWidthY);
2209 const APInt EffBitsX = APInt::getLowBitsSet(numBits: Width, loBitsSet: EffWidthX) << ShftCnt;
2210 // If the mask is extracting all bits from X or Y as is, we can skip
2211 // this AND op.
2212 if (EffBitsY.isSubsetOf(RHS: *Mask) && !EffBitsX.intersects(RHS: *Mask))
2213 return Y;
2214 if (EffBitsX.isSubsetOf(RHS: *Mask) && !EffBitsY.intersects(RHS: *Mask))
2215 return XShifted;
2216 }
2217 }
2218
2219 // ((X | Y) ^ X ) & ((X | Y) ^ Y) --> 0
2220 // ((X | Y) ^ Y ) & ((X | Y) ^ X) --> 0
2221 BinaryOperator *Or;
2222 if (match(V: Op0, P: m_c_Xor(L: m_Value(V&: X),
2223 R: m_CombineAnd(Ps: m_BinOp(I&: Or),
2224 Ps: m_c_Or(L: m_Deferred(V: X), R: m_Value(V&: Y))))) &&
2225 match(V: Op1, P: m_c_Xor(L: m_Specific(V: Or), R: m_Specific(V: Y))))
2226 return Constant::getNullValue(Ty: Op0->getType());
2227
2228 const APInt *C1;
2229 Value *A;
2230 // (A ^ C) & (A ^ ~C) -> 0
2231 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
2232 match(V: Op1, P: m_Xor(L: m_Specific(V: A), R: m_SpecificInt(V: ~*C1))))
2233 return Constant::getNullValue(Ty: Op0->getType());
2234
2235 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2236 if (std::optional<bool> Implied = isImpliedCondition(LHS: Op0, RHS: Op1, DL: Q.DL)) {
2237 // If Op0 is true implies Op1 is true, then Op0 is a subset of Op1.
2238 if (*Implied == true)
2239 return Op0;
2240 // If Op0 is true implies Op1 is false, then they are not true together.
2241 if (*Implied == false)
2242 return ConstantInt::getFalse(Ty: Op0->getType());
2243 }
2244 if (std::optional<bool> Implied = isImpliedCondition(LHS: Op1, RHS: Op0, DL: Q.DL)) {
2245 // If Op1 is true implies Op0 is true, then Op1 is a subset of Op0.
2246 if (*Implied)
2247 return Op1;
2248 // If Op1 is true implies Op0 is false, then they are not true together.
2249 if (!*Implied)
2250 return ConstantInt::getFalse(Ty: Op1->getType());
2251 }
2252 }
2253
2254 if (Value *V = simplifyByDomEq(Opcode: Instruction::And, Op0, Op1, Q, MaxRecurse))
2255 return V;
2256
2257 return nullptr;
2258}
2259
2260Value *llvm::simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2261 return ::simplifyAndInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
2262}
2263
2264// TODO: Many of these folds could use LogicalAnd/LogicalOr.
2265static Value *simplifyOrLogic(Value *X, Value *Y) {
2266 assert(X->getType() == Y->getType() && "Expected same type for 'or' ops");
2267 Type *Ty = X->getType();
2268
2269 // X | ~X --> -1
2270 if (match(V: Y, P: m_Not(V: m_Specific(V: X))))
2271 return ConstantInt::getAllOnesValue(Ty);
2272
2273 // X | ~(X & ?) = -1
2274 if (match(V: Y, P: m_Not(V: m_c_And(L: m_Specific(V: X), R: m_Value()))))
2275 return ConstantInt::getAllOnesValue(Ty);
2276
2277 // X | (X & ?) --> X
2278 if (match(V: Y, P: m_c_And(L: m_Specific(V: X), R: m_Value())))
2279 return X;
2280
2281 Value *A, *B;
2282
2283 // (A ^ B) | (A | B) --> A | B
2284 // (A ^ B) | (B | A) --> B | A
2285 if (match(V: X, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2286 match(V: Y, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2287 return Y;
2288
2289 // ~(A ^ B) | (A | B) --> -1
2290 // ~(A ^ B) | (B | A) --> -1
2291 if (match(V: X, P: m_Not(V: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)))) &&
2292 match(V: Y, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2293 return ConstantInt::getAllOnesValue(Ty);
2294
2295 // (A & ~B) | (A ^ B) --> A ^ B
2296 // (~B & A) | (A ^ B) --> A ^ B
2297 // (A & ~B) | (B ^ A) --> B ^ A
2298 // (~B & A) | (B ^ A) --> B ^ A
2299 if (match(V: X, P: m_c_And(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B)))) &&
2300 match(V: Y, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2301 return Y;
2302
2303 // (~A ^ B) | (A & B) --> ~A ^ B
2304 // (B ^ ~A) | (A & B) --> B ^ ~A
2305 // (~A ^ B) | (B & A) --> ~A ^ B
2306 // (B ^ ~A) | (B & A) --> B ^ ~A
2307 if (match(V: X, P: m_c_Xor(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2308 match(V: Y, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2309 return X;
2310
2311 // (~A | B) | (A ^ B) --> -1
2312 // (~A | B) | (B ^ A) --> -1
2313 // (B | ~A) | (A ^ B) --> -1
2314 // (B | ~A) | (B ^ A) --> -1
2315 if (match(V: X, P: m_c_Or(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2316 match(V: Y, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2317 return ConstantInt::getAllOnesValue(Ty);
2318
2319 // (~A & B) | ~(A | B) --> ~A
2320 // (~A & B) | ~(B | A) --> ~A
2321 // (B & ~A) | ~(A | B) --> ~A
2322 // (B & ~A) | ~(B | A) --> ~A
2323 Value *NotA;
2324 if (match(V: X, P: m_c_And(L: m_CombineAnd(Ps: m_Value(V&: NotA), Ps: m_Not(V: m_Value(V&: A))),
2325 R: m_Value(V&: B))) &&
2326 match(V: Y, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B)))))
2327 return NotA;
2328 // The same is true of Logical And
2329 // TODO: This could share the logic of the version above if there was a
2330 // version of LogicalAnd that allowed more than just i1 types.
2331 if (match(V: X, P: m_c_LogicalAnd(L: m_CombineAnd(Ps: m_Value(V&: NotA), Ps: m_Not(V: m_Value(V&: A))),
2332 R: m_Value(V&: B))) &&
2333 match(V: Y, P: m_Not(V: m_c_LogicalOr(L: m_Specific(V: A), R: m_Specific(V: B)))))
2334 return NotA;
2335
2336 // ~(A ^ B) | (A & B) --> ~(A ^ B)
2337 // ~(A ^ B) | (B & A) --> ~(A ^ B)
2338 Value *NotAB;
2339 if (match(V: X, P: m_CombineAnd(Ps: m_Not(V: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))),
2340 Ps: m_Value(V&: NotAB))) &&
2341 match(V: Y, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2342 return NotAB;
2343
2344 // ~(A & B) | (A ^ B) --> ~(A & B)
2345 // ~(A & B) | (B ^ A) --> ~(A & B)
2346 if (match(V: X, P: m_CombineAnd(Ps: m_Not(V: m_And(L: m_Value(V&: A), R: m_Value(V&: B))),
2347 Ps: m_Value(V&: NotAB))) &&
2348 match(V: Y, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2349 return NotAB;
2350
2351 return nullptr;
2352}
2353
2354/// Given operands for an Or, see if we can fold the result.
2355/// If not, this returns null.
2356static Value *simplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2357 unsigned MaxRecurse) {
2358 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Or, Op0, Op1, Q))
2359 return C;
2360
2361 // X | poison -> poison
2362 if (isa<PoisonValue>(Val: Op1))
2363 return Op1;
2364
2365 // X | undef -> -1
2366 // X | -1 = -1
2367 // Do not return Op1 because it may contain undef elements if it's a vector.
2368 if (Q.isUndefValue(V: Op1) || match(V: Op1, P: m_AllOnes()))
2369 return Constant::getAllOnesValue(Ty: Op0->getType());
2370
2371 // X | X = X
2372 // X | 0 = X
2373 if (Op0 == Op1 || match(V: Op1, P: m_Zero()))
2374 return Op0;
2375
2376 if (Value *R = simplifyOrLogic(X: Op0, Y: Op1))
2377 return R;
2378 if (Value *R = simplifyOrLogic(X: Op1, Y: Op0))
2379 return R;
2380
2381 if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Opcode: Instruction::Or))
2382 return V;
2383
2384 // Rotated -1 is still -1:
2385 // (-1 << X) | (-1 >> (C - X)) --> -1
2386 // (-1 >> X) | (-1 << (C - X)) --> -1
2387 // ...with C <= bitwidth (and commuted variants).
2388 Value *X, *Y;
2389 if ((match(V: Op0, P: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))) &&
2390 match(V: Op1, P: m_LShr(L: m_AllOnes(), R: m_Value(V&: Y)))) ||
2391 (match(V: Op1, P: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))) &&
2392 match(V: Op0, P: m_LShr(L: m_AllOnes(), R: m_Value(V&: Y))))) {
2393 const APInt *C;
2394 if ((match(V: X, P: m_Sub(L: m_APInt(Res&: C), R: m_Specific(V: Y))) ||
2395 match(V: Y, P: m_Sub(L: m_APInt(Res&: C), R: m_Specific(V: X)))) &&
2396 C->ule(RHS: X->getType()->getScalarSizeInBits())) {
2397 return ConstantInt::getAllOnesValue(Ty: X->getType());
2398 }
2399 }
2400
2401 // zext(X) | sext(X) --> sext(X)
2402 // sext(X) | zext(X) --> sext(X)
2403 {
2404 Value *X = nullptr;
2405 if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_SExt(Op: m_Specific(V: X))))
2406 return Op1;
2407 if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op0, P: m_SExt(Op: m_Specific(V: X))))
2408 return Op0;
2409 }
2410
2411 // A funnel shift (rotate) can be decomposed into simpler shifts. See if we
2412 // are mixing in another shift that is redundant with the funnel shift.
2413
2414 // (fshl X, ?, Y) | (shl X, Y) --> fshl X, ?, Y
2415 // (shl X, Y) | (fshl X, ?, Y) --> fshl X, ?, Y
2416 if (match(V: Op0,
2417 P: m_Intrinsic<Intrinsic::fshl>(Ops: m_Value(V&: X), Ops: m_Value(), Ops: m_Value(V&: Y))) &&
2418 match(V: Op1, P: m_Shl(L: m_Specific(V: X), R: m_Specific(V: Y))))
2419 return Op0;
2420 if (match(V: Op1,
2421 P: m_Intrinsic<Intrinsic::fshl>(Ops: m_Value(V&: X), Ops: m_Value(), Ops: m_Value(V&: Y))) &&
2422 match(V: Op0, P: m_Shl(L: m_Specific(V: X), R: m_Specific(V: Y))))
2423 return Op1;
2424
2425 // (fshr ?, X, Y) | (lshr X, Y) --> fshr ?, X, Y
2426 // (lshr X, Y) | (fshr ?, X, Y) --> fshr ?, X, Y
2427 if (match(V: Op0,
2428 P: m_Intrinsic<Intrinsic::fshr>(Ops: m_Value(), Ops: m_Value(V&: X), Ops: m_Value(V&: Y))) &&
2429 match(V: Op1, P: m_LShr(L: m_Specific(V: X), R: m_Specific(V: Y))))
2430 return Op0;
2431 if (match(V: Op1,
2432 P: m_Intrinsic<Intrinsic::fshr>(Ops: m_Value(), Ops: m_Value(V&: X), Ops: m_Value(V&: Y))) &&
2433 match(V: Op0, P: m_LShr(L: m_Specific(V: X), R: m_Specific(V: Y))))
2434 return Op1;
2435
2436 if (Value *V =
2437 simplifyAndOrWithICmpEq(Opcode: Instruction::Or, Op0, Op1, Q, MaxRecurse))
2438 return V;
2439 if (Value *V =
2440 simplifyAndOrWithICmpEq(Opcode: Instruction::Or, Op0: Op1, Op1: Op0, Q, MaxRecurse))
2441 return V;
2442
2443 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, IsAnd: false))
2444 return V;
2445
2446 // If we have a multiplication overflow check that is being 'and'ed with a
2447 // check that one of the multipliers is not zero, we can omit the 'and', and
2448 // only keep the overflow check.
2449 if (isCheckForZeroAndMulWithOverflow(Op0, Op1, IsAnd: false))
2450 return Op1;
2451 if (isCheckForZeroAndMulWithOverflow(Op0: Op1, Op1: Op0, IsAnd: false))
2452 return Op0;
2453
2454 // Try some generic simplifications for associative operations.
2455 if (Value *V =
2456 simplifyAssociativeBinOp(Opcode: Instruction::Or, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2457 return V;
2458
2459 // Or distributes over And. Try some generic simplifications based on this.
2460 if (Value *V = expandCommutativeBinOp(Opcode: Instruction::Or, L: Op0, R: Op1,
2461 OpcodeToExpand: Instruction::And, Q, MaxRecurse))
2462 return V;
2463
2464 if (isa<SelectInst>(Val: Op0) || isa<SelectInst>(Val: Op1)) {
2465 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2466 // A | (A || B) -> A || B
2467 if (match(V: Op1, P: m_Select(C: m_Specific(V: Op0), L: m_One(), R: m_Value())))
2468 return Op1;
2469 else if (match(V: Op0, P: m_Select(C: m_Specific(V: Op1), L: m_One(), R: m_Value())))
2470 return Op0;
2471 }
2472 // If the operation is with the result of a select instruction, check
2473 // whether operating on either branch of the select always yields the same
2474 // value.
2475 if (Value *V =
2476 threadBinOpOverSelect(Opcode: Instruction::Or, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2477 return V;
2478 }
2479
2480 // (A & C1)|(B & C2)
2481 Value *A, *B;
2482 const APInt *C1, *C2;
2483 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
2484 match(V: Op1, P: m_And(L: m_Value(V&: B), R: m_APInt(Res&: C2)))) {
2485 if (*C1 == ~*C2) {
2486 // (A & C1)|(B & C2)
2487 // If we have: ((V + N) & C1) | (V & C2)
2488 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2489 // replace with V+N.
2490 Value *N;
2491 if (C2->isMask() && // C2 == 0+1+
2492 match(V: A, P: m_c_Add(L: m_Specific(V: B), R: m_Value(V&: N)))) {
2493 // Add commutes, try both ways.
2494 if (MaskedValueIsZero(V: N, Mask: *C2, SQ: Q))
2495 return A;
2496 }
2497 // Or commutes, try both ways.
2498 if (C1->isMask() && match(V: B, P: m_c_Add(L: m_Specific(V: A), R: m_Value(V&: N)))) {
2499 // Add commutes, try both ways.
2500 if (MaskedValueIsZero(V: N, Mask: *C1, SQ: Q))
2501 return B;
2502 }
2503 }
2504 }
2505
2506 // If the operation is with the result of a phi instruction, check whether
2507 // operating on all incoming values of the phi always yields the same value.
2508 if (isa<PHINode>(Val: Op0) || isa<PHINode>(Val: Op1))
2509 if (Value *V = threadBinOpOverPHI(Opcode: Instruction::Or, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2510 return V;
2511
2512 // (A ^ C) | (A ^ ~C) -> -1, i.e. all bits set to one.
2513 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_APInt(Res&: C1))) &&
2514 match(V: Op1, P: m_Xor(L: m_Specific(V: A), R: m_SpecificInt(V: ~*C1))))
2515 return Constant::getAllOnesValue(Ty: Op0->getType());
2516
2517 if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2518 if (std::optional<bool> Implied =
2519 isImpliedCondition(LHS: Op0, RHS: Op1, DL: Q.DL, LHSIsTrue: false)) {
2520 // If Op0 is false implies Op1 is false, then Op1 is a subset of Op0.
2521 if (*Implied == false)
2522 return Op0;
2523 // If Op0 is false implies Op1 is true, then at least one is always true.
2524 if (*Implied == true)
2525 return ConstantInt::getTrue(Ty: Op0->getType());
2526 }
2527 if (std::optional<bool> Implied =
2528 isImpliedCondition(LHS: Op1, RHS: Op0, DL: Q.DL, LHSIsTrue: false)) {
2529 // If Op1 is false implies Op0 is false, then Op0 is a subset of Op1.
2530 if (*Implied == false)
2531 return Op1;
2532 // If Op1 is false implies Op0 is true, then at least one is always true.
2533 if (*Implied == true)
2534 return ConstantInt::getTrue(Ty: Op1->getType());
2535 }
2536 }
2537
2538 if (Value *V = simplifyByDomEq(Opcode: Instruction::Or, Op0, Op1, Q, MaxRecurse))
2539 return V;
2540
2541 return nullptr;
2542}
2543
2544Value *llvm::simplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2545 return ::simplifyOrInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
2546}
2547
2548/// Given operands for a Xor, see if we can fold the result.
2549/// If not, this returns null.
2550static Value *simplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2551 unsigned MaxRecurse) {
2552 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::Xor, Op0, Op1, Q))
2553 return C;
2554
2555 // X ^ poison -> poison
2556 if (isa<PoisonValue>(Val: Op1))
2557 return Op1;
2558
2559 // A ^ undef -> undef
2560 if (Q.isUndefValue(V: Op1))
2561 return Op1;
2562
2563 // A ^ 0 = A
2564 if (match(V: Op1, P: m_Zero()))
2565 return Op0;
2566
2567 // A ^ A = 0
2568 if (Op0 == Op1)
2569 return Constant::getNullValue(Ty: Op0->getType());
2570
2571 // A ^ ~A = ~A ^ A = -1
2572 if (match(V: Op0, P: m_Not(V: m_Specific(V: Op1))) || match(V: Op1, P: m_Not(V: m_Specific(V: Op0))))
2573 return Constant::getAllOnesValue(Ty: Op0->getType());
2574
2575 auto foldAndOrNot = [](Value *X, Value *Y) -> Value * {
2576 Value *A, *B;
2577 // (~A & B) ^ (A | B) --> A -- There are 8 commuted variants.
2578 if (match(V: X, P: m_c_And(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2579 match(V: Y, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2580 return A;
2581
2582 // (~A | B) ^ (A & B) --> ~A -- There are 8 commuted variants.
2583 // The 'not' op must contain a complete -1 operand (no undef elements for
2584 // vector) for the transform to be safe.
2585 Value *NotA;
2586 if (match(V: X, P: m_c_Or(L: m_CombineAnd(Ps: m_Not(V: m_Value(V&: A)), Ps: m_Value(V&: NotA)),
2587 R: m_Value(V&: B))) &&
2588 match(V: Y, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
2589 return NotA;
2590
2591 return nullptr;
2592 };
2593 if (Value *R = foldAndOrNot(Op0, Op1))
2594 return R;
2595 if (Value *R = foldAndOrNot(Op1, Op0))
2596 return R;
2597
2598 if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Opcode: Instruction::Xor))
2599 return V;
2600
2601 // Try some generic simplifications for associative operations.
2602 if (Value *V =
2603 simplifyAssociativeBinOp(Opcode: Instruction::Xor, LHS: Op0, RHS: Op1, Q, MaxRecurse))
2604 return V;
2605
2606 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2607 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2608 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2609 // only if B and C are equal. If B and C are equal then (since we assume
2610 // that operands have already been simplified) "select(cond, B, C)" should
2611 // have been simplified to the common value of B and C already. Analysing
2612 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2613 // for threading over phi nodes.
2614
2615 if (Value *V = simplifyByDomEq(Opcode: Instruction::Xor, Op0, Op1, Q, MaxRecurse))
2616 return V;
2617
2618 // (xor (sub nuw C_Mask, X), C_Mask) -> X
2619 {
2620 Value *X;
2621 if (match(V: Op0, P: m_NUWSub(L: m_Specific(V: Op1), R: m_Value(V&: X))) &&
2622 match(V: Op1, P: m_LowBitMask()))
2623 return X;
2624 }
2625
2626 return nullptr;
2627}
2628
2629Value *llvm::simplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2630 return ::simplifyXorInst(Op0, Op1, Q, MaxRecurse: RecursionLimit);
2631}
2632
2633static Type *getCompareTy(Value *Op) {
2634 return CmpInst::makeCmpResultType(opnd_type: Op->getType());
2635}
2636
2637/// Rummage around inside V looking for something equivalent to the comparison
2638/// "LHS Pred RHS". Return such a value if found, otherwise return null.
2639/// Helper function for analyzing max/min idioms.
2640static Value *extractEquivalentCondition(Value *V, CmpPredicate Pred,
2641 Value *LHS, Value *RHS) {
2642 SelectInst *SI = dyn_cast<SelectInst>(Val: V);
2643 if (!SI)
2644 return nullptr;
2645 CmpInst *Cmp = dyn_cast<CmpInst>(Val: SI->getCondition());
2646 if (!Cmp)
2647 return nullptr;
2648 Value *CmpLHS = Cmp->getOperand(i_nocapture: 0), *CmpRHS = Cmp->getOperand(i_nocapture: 1);
2649 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2650 return Cmp;
2651 if (Pred == CmpInst::getSwappedPredicate(pred: Cmp->getPredicate()) &&
2652 LHS == CmpRHS && RHS == CmpLHS)
2653 return Cmp;
2654 return nullptr;
2655}
2656
2657static bool isByValArg(const Value *V) {
2658 const Argument *A = dyn_cast<Argument>(Val: V);
2659 return A && A->hasByValAttr();
2660}
2661
2662static bool isDereferenceableArg(const Value *V) {
2663 const Argument *A = dyn_cast<Argument>(Val: V);
2664 return A && A->getType()->isPointerTy() && A->getDereferenceableBytes() > 0;
2665}
2666
2667/// Return true if the underlying object (storage) must be disjoint from
2668/// storage returned by any noalias return call.
2669static bool isAllocDisjoint(const Value *V) {
2670 // For allocas, we consider only static ones (dynamic
2671 // allocas might be transformed into calls to malloc not simultaneously
2672 // live with the compared-to allocation). For globals, we exclude symbols
2673 // that might be resolve lazily to symbols in another dynamically-loaded
2674 // library (and, thus, could be malloc'ed by the implementation).
2675 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Val: V))
2676 return AI->isStaticAlloca();
2677 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V))
2678 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2679 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2680 !GV->isThreadLocal();
2681 // Byval arguments point to storage accessible to the caller, which is
2682 // disjoint from the allocated storage returned by a noalias pointer.
2683 // TODO: possibly extend this to `dereferenceable(N)` arguments once the LLVM
2684 // allocator model and its interaction with `noalias` on return values is
2685 // clarified.
2686 return isByValArg(V);
2687}
2688
2689/// Return true if V1 and V2 are each the base of some distict storage region
2690/// [V, object_size(V)] which do not overlap. Note that zero sized regions
2691/// *are* possible, and that zero sized regions do not overlap with any other.
2692static bool haveNonOverlappingStorage(const Value *V1, const Value *V2) {
2693 // Global variables always exist, so they always exist during the lifetime
2694 // of each other and all allocas. Global variables themselves usually have
2695 // non-overlapping storage, but since their addresses are constants, the
2696 // case involving two globals does not reach here and is instead handled in
2697 // constant folding.
2698 //
2699 // Two different allocas usually have different addresses...
2700 //
2701 // However, if there's an @llvm.stackrestore dynamically in between two
2702 // allocas, they may have the same address. It's tempting to reduce the
2703 // scope of the problem by only looking at *static* allocas here. That would
2704 // cover the majority of allocas while significantly reducing the likelihood
2705 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2706 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2707 // an entry block. Also, if we have a block that's not attached to a
2708 // function, we can't tell if it's "static" under the current definition.
2709 // Theoretically, this problem could be fixed by creating a new kind of
2710 // instruction kind specifically for static allocas. Such a new instruction
2711 // could be required to be at the top of the entry block, thus preventing it
2712 // from being subject to a @llvm.stackrestore. Instcombine could even
2713 // convert regular allocas into these special allocas. It'd be nifty.
2714 // However, until then, this problem remains open.
2715 //
2716 // So, we'll assume that two non-empty allocas have different addresses
2717 // for now.
2718 //
2719 // Furthermore, an argument marked with the `dereferenceable(N)` attribute is
2720 // guaranteed to point to N loadable bytes. Such a pointer cannot be a
2721 // one-past-the-end pointer whose address happens to coincide with the start
2722 // of another object (e.g., an alloca), as loading from a one-past-the-end
2723 // address would be UB (thus, in contrast with the premise).
2724
2725 // Byval args are backed by storage that does not overlap with allocas,
2726 // globals, other byval args, or any dereferenceable argument.
2727 if (isByValArg(V: V1))
2728 return isa<AllocaInst>(Val: V2) || isa<GlobalVariable>(Val: V2) || isByValArg(V: V2) ||
2729 isDereferenceableArg(V: V2);
2730 if (isByValArg(V: V2))
2731 return isa<AllocaInst>(Val: V1) || isa<GlobalVariable>(Val: V1) || isByValArg(V: V1) ||
2732 isDereferenceableArg(V: V1);
2733
2734 if ((isDereferenceableArg(V: V1) && isa<AllocaInst>(Val: V2)) ||
2735 (isDereferenceableArg(V: V2) && isa<AllocaInst>(Val: V1)))
2736 return true;
2737
2738 return isa<AllocaInst>(Val: V1) &&
2739 (isa<AllocaInst>(Val: V2) || isa<GlobalVariable>(Val: V2));
2740}
2741
2742// A significant optimization not implemented here is assuming that alloca
2743// addresses are not equal to incoming argument values. They don't *alias*,
2744// as we say, but that doesn't mean they aren't equal, so we take a
2745// conservative approach.
2746//
2747// This is inspired in part by C++11 5.10p1:
2748// "Two pointers of the same type compare equal if and only if they are both
2749// null, both point to the same function, or both represent the same
2750// address."
2751//
2752// This is pretty permissive.
2753//
2754// It's also partly due to C11 6.5.9p6:
2755// "Two pointers compare equal if and only if both are null pointers, both are
2756// pointers to the same object (including a pointer to an object and a
2757// subobject at its beginning) or function, both are pointers to one past the
2758// last element of the same array object, or one is a pointer to one past the
2759// end of one array object and the other is a pointer to the start of a
2760// different array object that happens to immediately follow the first array
2761// object in the address space.)
2762//
2763// C11's version is more restrictive, however there's no reason why an argument
2764// couldn't be a one-past-the-end value for a stack object in the caller and be
2765// equal to the beginning of a stack object in the callee.
2766//
2767// If the C and C++ standards are ever made sufficiently restrictive in this
2768// area, it may be possible to update LLVM's semantics accordingly and reinstate
2769// this optimization.
2770static Constant *computePointerICmp(CmpPredicate Pred, Value *LHS, Value *RHS,
2771 const SimplifyQuery &Q) {
2772 assert(LHS->getType() == RHS->getType() && "Must have same types");
2773 const DataLayout &DL = Q.DL;
2774 const TargetLibraryInfo *TLI = Q.TLI;
2775
2776 // We fold equality and unsigned predicates on pointer comparisons, but forbid
2777 // signed predicates since a GEP with inbounds could cross the sign boundary.
2778 if (CmpInst::isSigned(Pred))
2779 return nullptr;
2780
2781 // We have to switch to a signed predicate to handle negative indices from
2782 // the base pointer.
2783 Pred = ICmpInst::getSignedPredicate(Pred);
2784
2785 // Strip off any constant offsets so that we can reason about them.
2786 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2787 // here and compare base addresses like AliasAnalysis does, however there are
2788 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2789 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2790 // doesn't need to guarantee pointer inequality when it says NoAlias.
2791
2792 // Even if an non-inbounds GEP occurs along the path we can still optimize
2793 // equality comparisons concerning the result.
2794 bool AllowNonInbounds = ICmpInst::isEquality(P: Pred);
2795 unsigned IndexSize = DL.getIndexTypeSizeInBits(Ty: LHS->getType());
2796 APInt LHSOffset(IndexSize, 0), RHSOffset(IndexSize, 0);
2797 LHS = LHS->stripAndAccumulateConstantOffsets(DL, Offset&: LHSOffset, AllowNonInbounds);
2798 RHS = RHS->stripAndAccumulateConstantOffsets(DL, Offset&: RHSOffset, AllowNonInbounds);
2799
2800 // If LHS and RHS are related via constant offsets to the same base
2801 // value, we can replace it with an icmp which just compares the offsets.
2802 if (LHS == RHS)
2803 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2804 V: ICmpInst::compare(LHS: LHSOffset, RHS: RHSOffset, Pred));
2805
2806 // Various optimizations for (in)equality comparisons.
2807 if (ICmpInst::isEquality(P: Pred)) {
2808 // Different non-empty allocations that exist at the same time have
2809 // different addresses (if the program can tell). If the offsets are
2810 // within the bounds of their allocations (and not one-past-the-end,
2811 // so inbounds is not sufficient), and their allocations aren't the same,
2812 // the pointers are not equal.
2813 if (haveNonOverlappingStorage(V1: LHS, V2: RHS)) {
2814 // Size of object V, falling back to `dereferenceable(N)` attribute on an
2815 // argument when getObjectSize cannot determine a concrete size.
2816 auto GetKnownSize = [&](Value *V, uint64_t &Size) {
2817 bool CanBeNull;
2818 Size = V->getPointerDereferenceableBytes(DL, CanBeNull,
2819 /*CanBeFreed=*/nullptr);
2820 return Size != 0 && !CanBeNull;
2821 };
2822
2823 uint64_t LHSSize, RHSSize;
2824 if (GetKnownSize(LHS, LHSSize) && GetKnownSize(RHS, RHSSize)) {
2825 APInt Dist = LHSOffset - RHSOffset;
2826 if (Dist.isNonNegative() ? Dist.ult(RHS: LHSSize) : (-Dist).ult(RHS: RHSSize))
2827 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2828 V: !CmpInst::isTrueWhenEqual(predicate: Pred));
2829 }
2830 }
2831
2832 // If one side of the equality comparison must come from a noalias call
2833 // (meaning a system memory allocation function), and the other side must
2834 // come from a pointer that cannot overlap with dynamically-allocated
2835 // memory within the lifetime of the current function (allocas, byval
2836 // arguments, globals), then determine the comparison result here.
2837 SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2838 getUnderlyingObjects(V: LHS, Objects&: LHSUObjs);
2839 getUnderlyingObjects(V: RHS, Objects&: RHSUObjs);
2840
2841 // Is the set of underlying objects all noalias calls?
2842 auto IsNoAliasCall = [](ArrayRef<const Value *> Objects) {
2843 return all_of(Range&: Objects, P: isNoAliasCall);
2844 };
2845
2846 // Is the set of underlying objects all things which must be disjoint from
2847 // noalias calls. We assume that indexing from such disjoint storage
2848 // into the heap is undefined, and thus offsets can be safely ignored.
2849 auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2850 return all_of(Range&: Objects, P: ::isAllocDisjoint);
2851 };
2852
2853 if ((IsNoAliasCall(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2854 (IsNoAliasCall(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2855 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2856 V: !CmpInst::isTrueWhenEqual(predicate: Pred));
2857
2858 // Fold comparisons for non-escaping pointer even if the allocation call
2859 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2860 // dynamic allocation call could be either of the operands. Note that
2861 // the other operand can not be based on the alloc - if it were, then
2862 // the cmp itself would be a capture.
2863 Value *MI = nullptr;
2864 if (isAllocLikeFn(V: LHS, TLI) && llvm::isKnownNonZero(V: RHS, Q))
2865 MI = LHS;
2866 else if (isAllocLikeFn(V: RHS, TLI) && llvm::isKnownNonZero(V: LHS, Q))
2867 MI = RHS;
2868 if (MI) {
2869 // FIXME: This is incorrect, see PR54002. While we can assume that the
2870 // allocation is at an address that makes the comparison false, this
2871 // requires that *all* comparisons to that address be false, which
2872 // InstSimplify cannot guarantee.
2873 struct CustomCaptureTracker : public CaptureTracker {
2874 bool Captured = false;
2875 void tooManyUses() override { Captured = true; }
2876 Action captured(const Use *U, UseCaptureInfo CI) override {
2877 // TODO(captures): Use UseCaptureInfo.
2878 if (auto *ICmp = dyn_cast<ICmpInst>(Val: U->getUser())) {
2879 // Comparison against value stored in global variable. Given the
2880 // pointer does not escape, its value cannot be guessed and stored
2881 // separately in a global variable.
2882 unsigned OtherIdx = 1 - U->getOperandNo();
2883 auto *LI = dyn_cast<LoadInst>(Val: ICmp->getOperand(i_nocapture: OtherIdx));
2884 if (LI && isa<GlobalVariable>(Val: LI->getPointerOperand()))
2885 return Continue;
2886 }
2887
2888 Captured = true;
2889 return Stop;
2890 }
2891 };
2892 CustomCaptureTracker Tracker;
2893 PointerMayBeCaptured(V: MI, Tracker: &Tracker);
2894 if (!Tracker.Captured)
2895 return ConstantInt::get(Ty: getCompareTy(Op: LHS),
2896 V: CmpInst::isFalseWhenEqual(predicate: Pred));
2897 }
2898 }
2899
2900 // Otherwise, fail.
2901 return nullptr;
2902}
2903
2904/// Fold an icmp when its operands have i1 scalar type.
2905static Value *simplifyICmpOfBools(CmpPredicate Pred, Value *LHS, Value *RHS,
2906 const SimplifyQuery &Q) {
2907 Type *ITy = getCompareTy(Op: LHS); // The return type.
2908 Type *OpTy = LHS->getType(); // The operand type.
2909 if (!OpTy->isIntOrIntVectorTy(BitWidth: 1))
2910 return nullptr;
2911
2912 // A boolean compared to true/false can be reduced in 14 out of the 20
2913 // (10 predicates * 2 constants) possible combinations. The other
2914 // 6 cases require a 'not' of the LHS.
2915
2916 auto ExtractNotLHS = [](Value *V) -> Value * {
2917 Value *X;
2918 if (match(V, P: m_Not(V: m_Value(V&: X))))
2919 return X;
2920 return nullptr;
2921 };
2922
2923 if (match(V: RHS, P: m_Zero())) {
2924 switch (Pred) {
2925 case CmpInst::ICMP_NE: // X != 0 -> X
2926 case CmpInst::ICMP_UGT: // X >u 0 -> X
2927 case CmpInst::ICMP_SLT: // X <s 0 -> X
2928 return LHS;
2929
2930 case CmpInst::ICMP_EQ: // not(X) == 0 -> X != 0 -> X
2931 case CmpInst::ICMP_ULE: // not(X) <=u 0 -> X >u 0 -> X
2932 case CmpInst::ICMP_SGE: // not(X) >=s 0 -> X <s 0 -> X
2933 if (Value *X = ExtractNotLHS(LHS))
2934 return X;
2935 break;
2936
2937 case CmpInst::ICMP_ULT: // X <u 0 -> false
2938 case CmpInst::ICMP_SGT: // X >s 0 -> false
2939 return getFalse(Ty: ITy);
2940
2941 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2942 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2943 return getTrue(Ty: ITy);
2944
2945 default:
2946 break;
2947 }
2948 } else if (match(V: RHS, P: m_One())) {
2949 switch (Pred) {
2950 case CmpInst::ICMP_EQ: // X == 1 -> X
2951 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2952 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2953 return LHS;
2954
2955 case CmpInst::ICMP_NE: // not(X) != 1 -> X == 1 -> X
2956 case CmpInst::ICMP_ULT: // not(X) <=u 1 -> X >=u 1 -> X
2957 case CmpInst::ICMP_SGT: // not(X) >s 1 -> X <=s -1 -> X
2958 if (Value *X = ExtractNotLHS(LHS))
2959 return X;
2960 break;
2961
2962 case CmpInst::ICMP_UGT: // X >u 1 -> false
2963 case CmpInst::ICMP_SLT: // X <s -1 -> false
2964 return getFalse(Ty: ITy);
2965
2966 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2967 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2968 return getTrue(Ty: ITy);
2969
2970 default:
2971 break;
2972 }
2973 }
2974
2975 switch (Pred) {
2976 default:
2977 break;
2978 case ICmpInst::ICMP_UGE:
2979 if (isImpliedCondition(LHS: RHS, RHS: LHS, DL: Q.DL).value_or(u: false))
2980 return getTrue(Ty: ITy);
2981 break;
2982 case ICmpInst::ICMP_SGE:
2983 /// For signed comparison, the values for an i1 are 0 and -1
2984 /// respectively. This maps into a truth table of:
2985 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2986 /// 0 | 0 | 1 (0 >= 0) | 1
2987 /// 0 | 1 | 1 (0 >= -1) | 1
2988 /// 1 | 0 | 0 (-1 >= 0) | 0
2989 /// 1 | 1 | 1 (-1 >= -1) | 1
2990 if (isImpliedCondition(LHS, RHS, DL: Q.DL).value_or(u: false))
2991 return getTrue(Ty: ITy);
2992 break;
2993 case ICmpInst::ICMP_ULE:
2994 if (isImpliedCondition(LHS, RHS, DL: Q.DL).value_or(u: false))
2995 return getTrue(Ty: ITy);
2996 break;
2997 case ICmpInst::ICMP_SLE:
2998 /// SLE follows the same logic as SGE with the LHS and RHS swapped.
2999 if (isImpliedCondition(LHS: RHS, RHS: LHS, DL: Q.DL).value_or(u: false))
3000 return getTrue(Ty: ITy);
3001 break;
3002 }
3003
3004 return nullptr;
3005}
3006
3007/// Check if RHS is zero or can be transformed to an equivalent zero comparison.
3008/// E.g., icmp sgt X, -1 --> icmp sge X, 0
3009static bool matchEquivZeroRHS(CmpPredicate &Pred, const Value *RHS) {
3010 // icmp [pred] X, 0 --> as-is
3011 if (match(V: RHS, P: m_Zero()))
3012 return true;
3013
3014 // Handle comparisons with -1 (all ones)
3015 if (match(V: RHS, P: m_AllOnes())) {
3016 switch (Pred) {
3017 case ICmpInst::ICMP_SGT:
3018 // icmp sgt X, -1 --> icmp sge X, 0
3019 Pred = ICmpInst::ICMP_SGE;
3020 return true;
3021 case ICmpInst::ICMP_SLE:
3022 // icmp sle X, -1 --> icmp slt X, 0
3023 Pred = ICmpInst::ICMP_SLT;
3024 return true;
3025 // Note: unsigned comparisons with -1 (UINT_MAX) are not handled here:
3026 // - icmp ugt X, -1 is always false (nothing > UINT_MAX)
3027 // - icmp ule X, -1 is always true (everything <= UINT_MAX)
3028 default:
3029 return false;
3030 }
3031 }
3032
3033 // Handle comparisons with 1
3034 if (match(V: RHS, P: m_One())) {
3035 switch (Pred) {
3036 case ICmpInst::ICMP_SGE:
3037 // icmp sge X, 1 --> icmp sgt X, 0
3038 Pred = ICmpInst::ICMP_SGT;
3039 return true;
3040 case ICmpInst::ICMP_UGE:
3041 // icmp uge X, 1 --> icmp ugt X, 0
3042 Pred = ICmpInst::ICMP_UGT;
3043 return true;
3044 case ICmpInst::ICMP_SLT:
3045 // icmp slt X, 1 --> icmp sle X, 0
3046 Pred = ICmpInst::ICMP_SLE;
3047 return true;
3048 case ICmpInst::ICMP_ULT:
3049 // icmp ult X, 1 --> icmp ule X, 0
3050 Pred = ICmpInst::ICMP_ULE;
3051 return true;
3052 default:
3053 return false;
3054 }
3055 }
3056
3057 return false;
3058}
3059
3060/// Try hard to fold icmp with zero RHS because this is a common case.
3061/// Note that, this function also handles the equivalent zero RHS, e.g.,
3062/// icmp sgt X, -1 --> icmp sge X, 0
3063static Value *simplifyICmpWithZero(CmpPredicate Pred, Value *LHS, Value *RHS,
3064 const SimplifyQuery &Q) {
3065 // Check if RHS is zero or can be transformed to an equivalent zero comparison
3066 if (!matchEquivZeroRHS(Pred, RHS))
3067 return nullptr;
3068
3069 Type *ITy = getCompareTy(Op: LHS); // The return type.
3070 switch (Pred) {
3071 default:
3072 llvm_unreachable("Unknown ICmp predicate!");
3073 case ICmpInst::ICMP_ULT:
3074 return getFalse(Ty: ITy);
3075 case ICmpInst::ICMP_UGE:
3076 return getTrue(Ty: ITy);
3077 case ICmpInst::ICMP_EQ:
3078 case ICmpInst::ICMP_ULE:
3079 if (isKnownNonZero(V: LHS, Q))
3080 return getFalse(Ty: ITy);
3081 break;
3082 case ICmpInst::ICMP_NE:
3083 case ICmpInst::ICMP_UGT:
3084 if (isKnownNonZero(V: LHS, Q))
3085 return getTrue(Ty: ITy);
3086 break;
3087 case ICmpInst::ICMP_SLT: {
3088 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3089 if (LHSKnown.isNegative())
3090 return getTrue(Ty: ITy);
3091 if (LHSKnown.isNonNegative())
3092 return getFalse(Ty: ITy);
3093 break;
3094 }
3095 case ICmpInst::ICMP_SLE: {
3096 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3097 if (LHSKnown.isNegative())
3098 return getTrue(Ty: ITy);
3099 if (LHSKnown.isNonNegative() && isKnownNonZero(V: LHS, Q))
3100 return getFalse(Ty: ITy);
3101 break;
3102 }
3103 case ICmpInst::ICMP_SGE: {
3104 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3105 if (LHSKnown.isNegative())
3106 return getFalse(Ty: ITy);
3107 if (LHSKnown.isNonNegative())
3108 return getTrue(Ty: ITy);
3109 break;
3110 }
3111 case ICmpInst::ICMP_SGT: {
3112 KnownBits LHSKnown = computeKnownBits(V: LHS, Q);
3113 if (LHSKnown.isNegative())
3114 return getFalse(Ty: ITy);
3115 if (LHSKnown.isNonNegative() && isKnownNonZero(V: LHS, Q))
3116 return getTrue(Ty: ITy);
3117 break;
3118 }
3119 }
3120
3121 return nullptr;
3122}
3123
3124static Value *simplifyICmpWithConstant(CmpPredicate Pred, Value *LHS,
3125 Value *RHS, const SimplifyQuery &Q) {
3126 Type *ITy = getCompareTy(Op: RHS); // The return type.
3127
3128 Value *X;
3129 const APInt *C;
3130 if (!match(V: RHS, P: m_APIntAllowPoison(Res&: C)))
3131 return nullptr;
3132
3133 // Sign-bit checks can be optimized to true/false after unsigned
3134 // floating-point casts:
3135 // icmp slt (bitcast (uitofp X)), 0 --> false
3136 // icmp sgt (bitcast (uitofp X)), -1 --> true
3137 if (match(V: LHS, P: m_ElementWiseBitCast(Op: m_UIToFP(Op: m_Value(V&: X))))) {
3138 bool TrueIfSigned;
3139 if (isSignBitCheck(Pred, RHS: *C, TrueIfSigned))
3140 return ConstantInt::getBool(Ty: ITy, V: !TrueIfSigned);
3141 }
3142
3143 // Rule out tautological comparisons (eg., ult 0 or uge 0).
3144 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, Other: *C);
3145 if (RHS_CR.isEmptySet())
3146 return ConstantInt::getFalse(Ty: ITy);
3147 if (RHS_CR.isFullSet())
3148 return ConstantInt::getTrue(Ty: ITy);
3149
3150 ConstantRange LHS_CR = computeConstantRange(V: LHS, ForSigned: CmpInst::isSigned(Pred), SQ: Q);
3151 if (!LHS_CR.isFullSet()) {
3152 if (RHS_CR.contains(CR: LHS_CR))
3153 return ConstantInt::getTrue(Ty: ITy);
3154 if (RHS_CR.inverse().contains(CR: LHS_CR))
3155 return ConstantInt::getFalse(Ty: ITy);
3156 }
3157
3158 // (mul nuw/nsw X, MulC) != C --> true (if C is not a multiple of MulC)
3159 // (mul nuw/nsw X, MulC) == C --> false (if C is not a multiple of MulC)
3160 const APInt *MulC;
3161 if (Q.IIQ.UseInstrInfo && ICmpInst::isEquality(P: Pred) &&
3162 ((match(V: LHS, P: m_NUWMul(L: m_Value(), R: m_APIntAllowPoison(Res&: MulC))) &&
3163 *MulC != 0 && C->urem(RHS: *MulC) != 0) ||
3164 (match(V: LHS, P: m_NSWMul(L: m_Value(), R: m_APIntAllowPoison(Res&: MulC))) &&
3165 *MulC != 0 && C->srem(RHS: *MulC) != 0)))
3166 return ConstantInt::get(Ty: ITy, V: Pred == ICmpInst::ICMP_NE);
3167
3168 if (Pred == ICmpInst::ICMP_UGE && C->isOne() && isKnownNonZero(V: LHS, Q))
3169 return ConstantInt::getTrue(Ty: ITy);
3170
3171 return nullptr;
3172}
3173
3174enum class MonotonicType { GreaterEq, LowerEq };
3175
3176/// Get values V_i such that V uge V_i (GreaterEq) or V ule V_i (LowerEq).
3177static void getUnsignedMonotonicValues(SmallPtrSetImpl<Value *> &Res, Value *V,
3178 MonotonicType Type,
3179 const SimplifyQuery &Q,
3180 unsigned Depth = 0) {
3181 if (!Res.insert(Ptr: V).second)
3182 return;
3183
3184 // Can be increased if useful.
3185 if (++Depth > 1)
3186 return;
3187
3188 auto *I = dyn_cast<Instruction>(Val: V);
3189 if (!I)
3190 return;
3191
3192 Value *X, *Y;
3193 if (Type == MonotonicType::GreaterEq) {
3194 if (match(V: I, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) ||
3195 match(V: I, P: m_Intrinsic<Intrinsic::uadd_sat>(Ops: m_Value(V&: X), Ops: m_Value(V&: Y)))) {
3196 getUnsignedMonotonicValues(Res, V: X, Type, Q, Depth);
3197 getUnsignedMonotonicValues(Res, V: Y, Type, Q, Depth);
3198 }
3199 // X * Y >= X --> true
3200 if (match(V: I, P: m_NUWMul(L: m_Value(V&: X), R: m_Value(V&: Y)))) {
3201 if (isKnownNonZero(V: X, Q))
3202 getUnsignedMonotonicValues(Res, V: Y, Type, Q, Depth);
3203 if (isKnownNonZero(V: Y, Q))
3204 getUnsignedMonotonicValues(Res, V: X, Type, Q, Depth);
3205 }
3206 } else {
3207 assert(Type == MonotonicType::LowerEq);
3208 switch (I->getOpcode()) {
3209 case Instruction::And:
3210 getUnsignedMonotonicValues(Res, V: I->getOperand(i: 0), Type, Q, Depth);
3211 getUnsignedMonotonicValues(Res, V: I->getOperand(i: 1), Type, Q, Depth);
3212 break;
3213 case Instruction::URem:
3214 case Instruction::UDiv:
3215 case Instruction::LShr:
3216 getUnsignedMonotonicValues(Res, V: I->getOperand(i: 0), Type, Q, Depth);
3217 break;
3218 case Instruction::Call:
3219 if (match(V: I, P: m_Intrinsic<Intrinsic::usub_sat>(Ops: m_Value(V&: X))))
3220 getUnsignedMonotonicValues(Res, V: X, Type, Q, Depth);
3221 break;
3222 default:
3223 break;
3224 }
3225 }
3226}
3227
3228static Value *simplifyICmpUsingMonotonicValues(CmpPredicate Pred, Value *LHS,
3229 Value *RHS,
3230 const SimplifyQuery &Q) {
3231 if (Pred != ICmpInst::ICMP_UGE && Pred != ICmpInst::ICMP_ULT)
3232 return nullptr;
3233
3234 // We have LHS uge GreaterValues and LowerValues uge RHS. If any of the
3235 // GreaterValues and LowerValues are the same, it follows that LHS uge RHS.
3236 SmallPtrSet<Value *, 4> GreaterValues;
3237 SmallPtrSet<Value *, 4> LowerValues;
3238 getUnsignedMonotonicValues(Res&: GreaterValues, V: LHS, Type: MonotonicType::GreaterEq, Q);
3239 getUnsignedMonotonicValues(Res&: LowerValues, V: RHS, Type: MonotonicType::LowerEq, Q);
3240 for (Value *GV : GreaterValues)
3241 if (LowerValues.contains(Ptr: GV))
3242 return ConstantInt::getBool(Ty: getCompareTy(Op: LHS),
3243 V: Pred == ICmpInst::ICMP_UGE);
3244 return nullptr;
3245}
3246
3247static Value *simplifyICmpWithBinOpOnLHS(CmpPredicate Pred, BinaryOperator *LBO,
3248 Value *RHS, const SimplifyQuery &Q,
3249 unsigned MaxRecurse) {
3250 Type *ITy = getCompareTy(Op: RHS); // The return type.
3251
3252 Value *Y = nullptr;
3253 // icmp pred (or X, Y), X
3254 if (match(V: LBO, P: m_c_Or(L: m_Value(V&: Y), R: m_Specific(V: RHS)))) {
3255 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
3256 KnownBits RHSKnown = computeKnownBits(V: RHS, Q);
3257 KnownBits YKnown = computeKnownBits(V: Y, Q);
3258 if (RHSKnown.isNonNegative() && YKnown.isNegative())
3259 return Pred == ICmpInst::ICMP_SLT ? getTrue(Ty: ITy) : getFalse(Ty: ITy);
3260 if (RHSKnown.isNegative() || YKnown.isNonNegative())
3261 return Pred == ICmpInst::ICMP_SLT ? getFalse(Ty: ITy) : getTrue(Ty: ITy);
3262 }
3263 }
3264
3265 // icmp pred (urem X, Y), Y
3266 if (match(V: LBO, P: m_URem(L: m_Value(), R: m_Specific(V: RHS)))) {
3267 switch (Pred) {
3268 default:
3269 break;
3270 case ICmpInst::ICMP_SGT:
3271 case ICmpInst::ICMP_SGE: {
3272 KnownBits Known = computeKnownBits(V: RHS, Q);
3273 if (!Known.isNonNegative())
3274 break;
3275 [[fallthrough]];
3276 }
3277 case ICmpInst::ICMP_EQ:
3278 case ICmpInst::ICMP_UGT:
3279 case ICmpInst::ICMP_UGE:
3280 return getFalse(Ty: ITy);
3281 case ICmpInst::ICMP_SLT:
3282 case ICmpInst::ICMP_SLE: {
3283 KnownBits Known = computeKnownBits(V: RHS, Q);
3284 if (!Known.isNonNegative())
3285 break;
3286 [[fallthrough]];
3287 }
3288 case ICmpInst::ICMP_NE:
3289 case ICmpInst::ICMP_ULT:
3290 case ICmpInst::ICMP_ULE:
3291 return getTrue(Ty: ITy);
3292 }
3293 }
3294
3295 // If x is nonzero:
3296 // x >>u C <u x --> true for C != 0.
3297 // x >>u C != x --> true for C != 0.
3298 // x >>u C >=u x --> false for C != 0.
3299 // x >>u C == x --> false for C != 0.
3300 // x udiv C <u x --> true for C != 1.
3301 // x udiv C != x --> true for C != 1.
3302 // x udiv C >=u x --> false for C != 1.
3303 // x udiv C == x --> false for C != 1.
3304 // TODO: allow non-constant shift amount/divisor
3305 const APInt *C;
3306 if ((match(V: LBO, P: m_LShr(L: m_Specific(V: RHS), R: m_APInt(Res&: C))) && *C != 0) ||
3307 (match(V: LBO, P: m_UDiv(L: m_Specific(V: RHS), R: m_APInt(Res&: C))) && *C != 1)) {
3308 if (isKnownNonZero(V: RHS, Q)) {
3309 switch (Pred) {
3310 default:
3311 break;
3312 case ICmpInst::ICMP_EQ:
3313 case ICmpInst::ICMP_UGE:
3314 case ICmpInst::ICMP_UGT:
3315 return getFalse(Ty: ITy);
3316 case ICmpInst::ICMP_NE:
3317 case ICmpInst::ICMP_ULT:
3318 case ICmpInst::ICMP_ULE:
3319 return getTrue(Ty: ITy);
3320 }
3321 }
3322 }
3323
3324 // (x*C1)/C2 <= x for C1 <= C2.
3325 // This holds even if the multiplication overflows: Assume that x != 0 and
3326 // arithmetic is modulo M. For overflow to occur we must have C1 >= M/x and
3327 // thus C2 >= M/x. It follows that (x*C1)/C2 <= (M-1)/C2 <= ((M-1)*x)/M < x.
3328 //
3329 // Additionally, either the multiplication and division might be represented
3330 // as shifts:
3331 // (x*C1)>>C2 <= x for C1 < 2**C2.
3332 // (x<<C1)/C2 <= x for 2**C1 < C2.
3333 const APInt *C1, *C2;
3334 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))) &&
3335 C1->ule(RHS: *C2)) ||
3336 (match(V: LBO, P: m_LShr(L: m_Mul(L: m_Specific(V: RHS), R: m_APInt(Res&: C1)), R: m_APInt(Res&: C2))) &&
3337 C1->ule(RHS: APInt(C2->getBitWidth(), 1) << *C2)) ||
3338 (match(V: LBO, P: m_UDiv(L: m_Shl(L: m_Specific(V: RHS), R: m_APInt(Res&: C1)), R: m_APInt(Res&: C2))) &&
3339 (APInt(C1->getBitWidth(), 1) << *C1).ule(RHS: *C2))) {
3340 if (Pred == ICmpInst::ICMP_UGT)
3341 return getFalse(Ty: ITy);
3342 if (Pred == ICmpInst::ICMP_ULE)
3343 return getTrue(Ty: ITy);
3344 }
3345
3346 // (sub C, X) == X, C is odd --> false
3347 // (sub C, X) != X, C is odd --> true
3348 if (match(V: LBO, P: m_Sub(L: m_APIntAllowPoison(Res&: C), R: m_Specific(V: RHS))) &&
3349 (*C & 1) == 1 && ICmpInst::isEquality(P: Pred))
3350 return (Pred == ICmpInst::ICMP_EQ) ? getFalse(Ty: ITy) : getTrue(Ty: ITy);
3351
3352 return nullptr;
3353}
3354
3355// If only one of the icmp's operands has NSW flags, try to prove that:
3356//
3357// icmp slt/sgt/sle/sge (x + C1), (x +nsw C2)
3358//
3359// is equivalent to:
3360//
3361// icmp slt/sgt/sle/sge C1, C2
3362//
3363// which is true if x + C2 has the NSW flags set and:
3364// *) C1 <= C2 && C1 >= 0, or
3365// *) C2 <= C1 && C1 <= 0.
3366//
3367static bool trySimplifyICmpWithAdds(CmpPredicate Pred, Value *LHS, Value *RHS,
3368 const InstrInfoQuery &IIQ) {
3369 // TODO: support other predicates.
3370 if (!ICmpInst::isSigned(Pred) || !IIQ.UseInstrInfo)
3371 return false;
3372
3373 // Canonicalize nsw add as RHS.
3374 if (!match(V: RHS, P: m_NSWAdd(L: m_Value(), R: m_Value())))
3375 std::swap(a&: LHS, b&: RHS);
3376 if (!match(V: RHS, P: m_NSWAdd(L: m_Value(), R: m_Value())))
3377 return false;
3378
3379 Value *X;
3380 const APInt *C1, *C2;
3381 if (!match(V: LHS, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C1))) ||
3382 !match(V: RHS, P: m_Add(L: m_Specific(V: X), R: m_APInt(Res&: C2))))
3383 return false;
3384
3385 return (C1->sle(RHS: *C2) && C1->isNonNegative()) ||
3386 (C2->sle(RHS: *C1) && C1->isNonPositive());
3387}
3388
3389/// TODO: A large part of this logic is duplicated in InstCombine's
3390/// foldICmpBinOp(). We should be able to share that and avoid the code
3391/// duplication.
3392static Value *simplifyICmpWithBinOp(CmpPredicate Pred, Value *LHS, Value *RHS,
3393 const SimplifyQuery &Q,
3394 unsigned MaxRecurse) {
3395 BinaryOperator *LBO = dyn_cast<BinaryOperator>(Val: LHS);
3396 BinaryOperator *RBO = dyn_cast<BinaryOperator>(Val: RHS);
3397 if (MaxRecurse && (LBO || RBO)) {
3398 // Analyze the case when either LHS or RHS is an add instruction.
3399 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3400 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
3401 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
3402 if (LBO && LBO->getOpcode() == Instruction::Add) {
3403 A = LBO->getOperand(i_nocapture: 0);
3404 B = LBO->getOperand(i_nocapture: 1);
3405 NoLHSWrapProblem =
3406 ICmpInst::isEquality(P: Pred) ||
3407 (CmpInst::isUnsigned(Pred) &&
3408 Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO))) ||
3409 (CmpInst::isSigned(Pred) &&
3410 Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO)));
3411 }
3412 if (RBO && RBO->getOpcode() == Instruction::Add) {
3413 C = RBO->getOperand(i_nocapture: 0);
3414 D = RBO->getOperand(i_nocapture: 1);
3415 NoRHSWrapProblem =
3416 ICmpInst::isEquality(P: Pred) ||
3417 (CmpInst::isUnsigned(Pred) &&
3418 Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: RBO))) ||
3419 (CmpInst::isSigned(Pred) &&
3420 Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: RBO)));
3421 }
3422
3423 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3424 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
3425 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: A == RHS ? B : A,
3426 RHS: Constant::getNullValue(Ty: RHS->getType()), Q,
3427 MaxRecurse: MaxRecurse - 1))
3428 return V;
3429
3430 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3431 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
3432 if (Value *V =
3433 simplifyICmpInst(Predicate: Pred, LHS: Constant::getNullValue(Ty: LHS->getType()),
3434 RHS: C == LHS ? D : C, Q, MaxRecurse: MaxRecurse - 1))
3435 return V;
3436
3437 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
3438 bool CanSimplify = (NoLHSWrapProblem && NoRHSWrapProblem) ||
3439 trySimplifyICmpWithAdds(Pred, LHS, RHS, IIQ: Q.IIQ);
3440 if (A && C && (A == C || A == D || B == C || B == D) && CanSimplify) {
3441 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3442 Value *Y, *Z;
3443 if (A == C) {
3444 // C + B == C + D -> B == D
3445 Y = B;
3446 Z = D;
3447 } else if (A == D) {
3448 // D + B == C + D -> B == C
3449 Y = B;
3450 Z = C;
3451 } else if (B == C) {
3452 // A + C == C + D -> A == D
3453 Y = A;
3454 Z = D;
3455 } else {
3456 assert(B == D);
3457 // A + D == C + D -> A == C
3458 Y = A;
3459 Z = C;
3460 }
3461 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: Y, RHS: Z, Q, MaxRecurse: MaxRecurse - 1))
3462 return V;
3463 }
3464 }
3465
3466 if (LBO)
3467 if (Value *V = simplifyICmpWithBinOpOnLHS(Pred, LBO, RHS, Q, MaxRecurse))
3468 return V;
3469
3470 if (RBO)
3471 if (Value *V = simplifyICmpWithBinOpOnLHS(
3472 Pred: ICmpInst::getSwappedPredicate(pred: Pred), LBO: RBO, RHS: LHS, Q, MaxRecurse))
3473 return V;
3474
3475 // 0 - (zext X) pred C
3476 if (!CmpInst::isUnsigned(Pred) && match(V: LHS, P: m_Neg(V: m_ZExt(Op: m_Value())))) {
3477 const APInt *C;
3478 if (match(V: RHS, P: m_APInt(Res&: C))) {
3479 if (C->isStrictlyPositive()) {
3480 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_NE)
3481 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3482 if (Pred == ICmpInst::ICMP_SGE || Pred == ICmpInst::ICMP_EQ)
3483 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3484 }
3485 if (C->isNonNegative()) {
3486 if (Pred == ICmpInst::ICMP_SLE)
3487 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3488 if (Pred == ICmpInst::ICMP_SGT)
3489 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3490 }
3491 }
3492 }
3493
3494 // If C2 is a power-of-2 and C is not:
3495 // (C2 << X) == C --> false
3496 // (C2 << X) != C --> true
3497 const APInt *C;
3498 if (match(V: LHS, P: m_Shl(L: m_Power2(), R: m_Value())) &&
3499 match(V: RHS, P: m_APIntAllowPoison(Res&: C)) && !C->isPowerOf2()) {
3500 // C2 << X can equal zero in some circumstances.
3501 // This simplification might be unsafe if C is zero.
3502 //
3503 // We know it is safe if:
3504 // - The shift is nsw. We can't shift out the one bit.
3505 // - The shift is nuw. We can't shift out the one bit.
3506 // - C2 is one.
3507 // - C isn't zero.
3508 if (Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO)) ||
3509 Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: LBO)) ||
3510 match(V: LHS, P: m_Shl(L: m_One(), R: m_Value())) || !C->isZero()) {
3511 if (Pred == ICmpInst::ICMP_EQ)
3512 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3513 if (Pred == ICmpInst::ICMP_NE)
3514 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3515 }
3516 }
3517
3518 // If C is a power-of-2:
3519 // (C << X) >u 0x8000 --> false
3520 // (C << X) <=u 0x8000 --> true
3521 if (match(V: LHS, P: m_Shl(L: m_Power2(), R: m_Value())) && match(V: RHS, P: m_SignMask())) {
3522 if (Pred == ICmpInst::ICMP_UGT)
3523 return ConstantInt::getFalse(Ty: getCompareTy(Op: RHS));
3524 if (Pred == ICmpInst::ICMP_ULE)
3525 return ConstantInt::getTrue(Ty: getCompareTy(Op: RHS));
3526 }
3527
3528 if (!MaxRecurse || !LBO || !RBO || LBO->getOpcode() != RBO->getOpcode())
3529 return nullptr;
3530
3531 if (LBO->getOperand(i_nocapture: 0) == RBO->getOperand(i_nocapture: 0)) {
3532 switch (LBO->getOpcode()) {
3533 default:
3534 break;
3535 case Instruction::Shl: {
3536 bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: LBO) && Q.IIQ.hasNoUnsignedWrap(Op: RBO);
3537 bool NSW = Q.IIQ.hasNoSignedWrap(Op: LBO) && Q.IIQ.hasNoSignedWrap(Op: RBO);
3538 if (!NUW || (ICmpInst::isSigned(Pred) && !NSW) ||
3539 !isKnownNonZero(V: LBO->getOperand(i_nocapture: 0), Q))
3540 break;
3541 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 1),
3542 RHS: RBO->getOperand(i_nocapture: 1), Q, MaxRecurse: MaxRecurse - 1))
3543 return V;
3544 break;
3545 }
3546 // If C1 & C2 == C1, A = X and/or C1, B = X and/or C2:
3547 // icmp ule A, B -> true
3548 // icmp ugt A, B -> false
3549 // icmp sle A, B -> true (C1 and C2 are the same sign)
3550 // icmp sgt A, B -> false (C1 and C2 are the same sign)
3551 case Instruction::And:
3552 case Instruction::Or: {
3553 const APInt *C1, *C2;
3554 if (ICmpInst::isRelational(P: Pred) &&
3555 match(V: LBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C1)) &&
3556 match(V: RBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C2))) {
3557 if (!C1->isSubsetOf(RHS: *C2)) {
3558 std::swap(a&: C1, b&: C2);
3559 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
3560 }
3561 if (C1->isSubsetOf(RHS: *C2)) {
3562 if (Pred == ICmpInst::ICMP_ULE)
3563 return ConstantInt::getTrue(Ty: getCompareTy(Op: LHS));
3564 if (Pred == ICmpInst::ICMP_UGT)
3565 return ConstantInt::getFalse(Ty: getCompareTy(Op: LHS));
3566 if (C1->isNonNegative() == C2->isNonNegative()) {
3567 if (Pred == ICmpInst::ICMP_SLE)
3568 return ConstantInt::getTrue(Ty: getCompareTy(Op: LHS));
3569 if (Pred == ICmpInst::ICMP_SGT)
3570 return ConstantInt::getFalse(Ty: getCompareTy(Op: LHS));
3571 }
3572 }
3573 }
3574 break;
3575 }
3576 }
3577 }
3578
3579 if (LBO->getOperand(i_nocapture: 1) == RBO->getOperand(i_nocapture: 1)) {
3580 switch (LBO->getOpcode()) {
3581 default:
3582 break;
3583 case Instruction::UDiv:
3584 case Instruction::LShr:
3585 if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(Op: LBO) ||
3586 !Q.IIQ.isExact(Op: RBO))
3587 break;
3588 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3589 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3590 return V;
3591 break;
3592 case Instruction::SDiv:
3593 if (!ICmpInst::isEquality(P: Pred) || !Q.IIQ.isExact(Op: LBO) ||
3594 !Q.IIQ.isExact(Op: RBO))
3595 break;
3596 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3597 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3598 return V;
3599 break;
3600 case Instruction::AShr:
3601 if (!Q.IIQ.isExact(Op: LBO) || !Q.IIQ.isExact(Op: RBO))
3602 break;
3603 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3604 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3605 return V;
3606 break;
3607 case Instruction::Shl: {
3608 bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: LBO) && Q.IIQ.hasNoUnsignedWrap(Op: RBO);
3609 bool NSW = Q.IIQ.hasNoSignedWrap(Op: LBO) && Q.IIQ.hasNoSignedWrap(Op: RBO);
3610 if (!NUW && !NSW)
3611 break;
3612 if (!NSW && ICmpInst::isSigned(Pred))
3613 break;
3614 if (Value *V = simplifyICmpInst(Predicate: Pred, LHS: LBO->getOperand(i_nocapture: 0),
3615 RHS: RBO->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3616 return V;
3617 break;
3618 }
3619 }
3620 }
3621 return nullptr;
3622}
3623
3624/// simplify integer comparisons where at least one operand of the compare
3625/// matches an integer min/max idiom.
3626static Value *simplifyICmpWithMinMax(CmpPredicate Pred, Value *LHS, Value *RHS,
3627 const SimplifyQuery &Q,
3628 unsigned MaxRecurse) {
3629 Type *ITy = getCompareTy(Op: LHS); // The return type.
3630 Value *A, *B;
3631 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
3632 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3633
3634 // Signed variants on "max(a,b)>=a -> true".
3635 if (match(V: LHS, P: m_SMax(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) && (A == RHS || B == RHS)) {
3636 if (A != RHS)
3637 std::swap(a&: A, b&: B); // smax(A, B) pred A.
3638 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3639 // We analyze this as smax(A, B) pred A.
3640 P = Pred;
3641 } else if (match(V: RHS, P: m_SMax(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3642 (A == LHS || B == LHS)) {
3643 if (A != LHS)
3644 std::swap(a&: A, b&: B); // A pred smax(A, B).
3645 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3646 // We analyze this as smax(A, B) swapped-pred A.
3647 P = CmpInst::getSwappedPredicate(pred: Pred);
3648 } else if (match(V: LHS, P: m_SMin(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3649 (A == RHS || B == RHS)) {
3650 if (A != RHS)
3651 std::swap(a&: A, b&: B); // smin(A, B) pred A.
3652 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3653 // We analyze this as smax(-A, -B) swapped-pred -A.
3654 // Note that we do not need to actually form -A or -B thanks to EqP.
3655 P = CmpInst::getSwappedPredicate(pred: Pred);
3656 } else if (match(V: RHS, P: m_SMin(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3657 (A == LHS || B == LHS)) {
3658 if (A != LHS)
3659 std::swap(a&: A, b&: B); // A pred smin(A, B).
3660 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3661 // We analyze this as smax(-A, -B) pred -A.
3662 // Note that we do not need to actually form -A or -B thanks to EqP.
3663 P = Pred;
3664 }
3665 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3666 // Cases correspond to "max(A, B) p A".
3667 switch (P) {
3668 default:
3669 break;
3670 case CmpInst::ICMP_EQ:
3671 case CmpInst::ICMP_SLE:
3672 // Equivalent to "A EqP B". This may be the same as the condition tested
3673 // in the max/min; if so, we can just return that.
3674 if (Value *V = extractEquivalentCondition(V: LHS, Pred: EqP, LHS: A, RHS: B))
3675 return V;
3676 if (Value *V = extractEquivalentCondition(V: RHS, Pred: EqP, LHS: A, RHS: B))
3677 return V;
3678 // Otherwise, see if "A EqP B" simplifies.
3679 if (MaxRecurse)
3680 if (Value *V = simplifyICmpInst(Predicate: EqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3681 return V;
3682 break;
3683 case CmpInst::ICMP_NE:
3684 case CmpInst::ICMP_SGT: {
3685 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(pred: EqP);
3686 // Equivalent to "A InvEqP B". This may be the same as the condition
3687 // tested in the max/min; if so, we can just return that.
3688 if (Value *V = extractEquivalentCondition(V: LHS, Pred: InvEqP, LHS: A, RHS: B))
3689 return V;
3690 if (Value *V = extractEquivalentCondition(V: RHS, Pred: InvEqP, LHS: A, RHS: B))
3691 return V;
3692 // Otherwise, see if "A InvEqP B" simplifies.
3693 if (MaxRecurse)
3694 if (Value *V = simplifyICmpInst(Predicate: InvEqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3695 return V;
3696 break;
3697 }
3698 case CmpInst::ICMP_SGE:
3699 // Always true.
3700 return getTrue(Ty: ITy);
3701 case CmpInst::ICMP_SLT:
3702 // Always false.
3703 return getFalse(Ty: ITy);
3704 }
3705 }
3706
3707 // Unsigned variants on "max(a,b)>=a -> true".
3708 P = CmpInst::BAD_ICMP_PREDICATE;
3709 if (match(V: LHS, P: m_UMax(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) && (A == RHS || B == RHS)) {
3710 if (A != RHS)
3711 std::swap(a&: A, b&: B); // umax(A, B) pred A.
3712 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3713 // We analyze this as umax(A, B) pred A.
3714 P = Pred;
3715 } else if (match(V: RHS, P: m_UMax(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3716 (A == LHS || B == LHS)) {
3717 if (A != LHS)
3718 std::swap(a&: A, b&: B); // A pred umax(A, B).
3719 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3720 // We analyze this as umax(A, B) swapped-pred A.
3721 P = CmpInst::getSwappedPredicate(pred: Pred);
3722 } else if (match(V: LHS, P: m_UMin(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3723 (A == RHS || B == RHS)) {
3724 if (A != RHS)
3725 std::swap(a&: A, b&: B); // umin(A, B) pred A.
3726 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3727 // We analyze this as umax(-A, -B) swapped-pred -A.
3728 // Note that we do not need to actually form -A or -B thanks to EqP.
3729 P = CmpInst::getSwappedPredicate(pred: Pred);
3730 } else if (match(V: RHS, P: m_UMin(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3731 (A == LHS || B == LHS)) {
3732 if (A != LHS)
3733 std::swap(a&: A, b&: B); // A pred umin(A, B).
3734 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3735 // We analyze this as umax(-A, -B) pred -A.
3736 // Note that we do not need to actually form -A or -B thanks to EqP.
3737 P = Pred;
3738 }
3739 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3740 // Cases correspond to "max(A, B) p A".
3741 switch (P) {
3742 default:
3743 break;
3744 case CmpInst::ICMP_EQ:
3745 case CmpInst::ICMP_ULE:
3746 // Equivalent to "A EqP B". This may be the same as the condition tested
3747 // in the max/min; if so, we can just return that.
3748 if (Value *V = extractEquivalentCondition(V: LHS, Pred: EqP, LHS: A, RHS: B))
3749 return V;
3750 if (Value *V = extractEquivalentCondition(V: RHS, Pred: EqP, LHS: A, RHS: B))
3751 return V;
3752 // Otherwise, see if "A EqP B" simplifies.
3753 if (MaxRecurse)
3754 if (Value *V = simplifyICmpInst(Predicate: EqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3755 return V;
3756 break;
3757 case CmpInst::ICMP_NE:
3758 case CmpInst::ICMP_UGT: {
3759 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(pred: EqP);
3760 // Equivalent to "A InvEqP B". This may be the same as the condition
3761 // tested in the max/min; if so, we can just return that.
3762 if (Value *V = extractEquivalentCondition(V: LHS, Pred: InvEqP, LHS: A, RHS: B))
3763 return V;
3764 if (Value *V = extractEquivalentCondition(V: RHS, Pred: InvEqP, LHS: A, RHS: B))
3765 return V;
3766 // Otherwise, see if "A InvEqP B" simplifies.
3767 if (MaxRecurse)
3768 if (Value *V = simplifyICmpInst(Predicate: InvEqP, LHS: A, RHS: B, Q, MaxRecurse: MaxRecurse - 1))
3769 return V;
3770 break;
3771 }
3772 case CmpInst::ICMP_UGE:
3773 return getTrue(Ty: ITy);
3774 case CmpInst::ICMP_ULT:
3775 return getFalse(Ty: ITy);
3776 }
3777 }
3778
3779 // Comparing 1 each of min/max with a common operand?
3780 // Canonicalize min operand to RHS.
3781 if (match(V: LHS, P: m_UMin(Op0: m_Value(), Op1: m_Value())) ||
3782 match(V: LHS, P: m_SMin(Op0: m_Value(), Op1: m_Value()))) {
3783 std::swap(a&: LHS, b&: RHS);
3784 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
3785 }
3786
3787 Value *C, *D;
3788 if (match(V: LHS, P: m_SMax(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3789 match(V: RHS, P: m_SMin(Op0: m_Value(V&: C), Op1: m_Value(V&: D))) &&
3790 (A == C || A == D || B == C || B == D)) {
3791 // smax(A, B) >=s smin(A, D) --> true
3792 if (Pred == CmpInst::ICMP_SGE)
3793 return getTrue(Ty: ITy);
3794 // smax(A, B) <s smin(A, D) --> false
3795 if (Pred == CmpInst::ICMP_SLT)
3796 return getFalse(Ty: ITy);
3797 } else if (match(V: LHS, P: m_UMax(Op0: m_Value(V&: A), Op1: m_Value(V&: B))) &&
3798 match(V: RHS, P: m_UMin(Op0: m_Value(V&: C), Op1: m_Value(V&: D))) &&
3799 (A == C || A == D || B == C || B == D)) {
3800 // umax(A, B) >=u umin(A, D) --> true
3801 if (Pred == CmpInst::ICMP_UGE)
3802 return getTrue(Ty: ITy);
3803 // umax(A, B) <u umin(A, D) --> false
3804 if (Pred == CmpInst::ICMP_ULT)
3805 return getFalse(Ty: ITy);
3806 }
3807
3808 return nullptr;
3809}
3810
3811static Value *simplifyICmpWithDominatingAssume(CmpPredicate Predicate,
3812 Value *LHS, Value *RHS,
3813 const SimplifyQuery &Q) {
3814 // Gracefully handle instructions that have not been inserted yet.
3815 if (!Q.AC || !Q.CxtI)
3816 return nullptr;
3817
3818 for (Value *AssumeBaseOp : {LHS, RHS}) {
3819 for (auto &AssumeVH : Q.AC->assumptionsFor(V: AssumeBaseOp)) {
3820 if (!AssumeVH)
3821 continue;
3822
3823 CallInst *Assume = cast<CallInst>(Val&: AssumeVH);
3824 if (std::optional<bool> Imp = isImpliedCondition(
3825 LHS: Assume->getArgOperand(i: 0), RHSPred: Predicate, RHSOp0: LHS, RHSOp1: RHS, DL: Q.DL))
3826 if (isValidAssumeForContext(I: Assume, Q))
3827 return ConstantInt::get(Ty: getCompareTy(Op: LHS), V: *Imp);
3828 }
3829 }
3830
3831 return nullptr;
3832}
3833
3834static Value *simplifyICmpWithIntrinsicOnLHS(CmpPredicate Pred, Value *LHS,
3835 Value *RHS) {
3836 auto *II = dyn_cast<IntrinsicInst>(Val: LHS);
3837 if (!II)
3838 return nullptr;
3839
3840 switch (II->getIntrinsicID()) {
3841 case Intrinsic::uadd_sat:
3842 // uadd.sat(X, Y) uge X + Y
3843 if (match(V: RHS, P: m_c_Add(L: m_Specific(V: II->getArgOperand(i: 0)),
3844 R: m_Specific(V: II->getArgOperand(i: 1))))) {
3845 if (Pred == ICmpInst::ICMP_UGE)
3846 return ConstantInt::getTrue(Ty: getCompareTy(Op: II));
3847 if (Pred == ICmpInst::ICMP_ULT)
3848 return ConstantInt::getFalse(Ty: getCompareTy(Op: II));
3849 }
3850 return nullptr;
3851 case Intrinsic::usub_sat:
3852 // usub.sat(X, Y) ule X - Y
3853 if (match(V: RHS, P: m_Sub(L: m_Specific(V: II->getArgOperand(i: 0)),
3854 R: m_Specific(V: II->getArgOperand(i: 1))))) {
3855 if (Pred == ICmpInst::ICMP_ULE)
3856 return ConstantInt::getTrue(Ty: getCompareTy(Op: II));
3857 if (Pred == ICmpInst::ICMP_UGT)
3858 return ConstantInt::getFalse(Ty: getCompareTy(Op: II));
3859 }
3860 return nullptr;
3861 default:
3862 return nullptr;
3863 }
3864}
3865
3866/// Helper method to get range from metadata or attribute.
3867static std::optional<ConstantRange> getRange(Value *V,
3868 const InstrInfoQuery &IIQ) {
3869 if (Instruction *I = dyn_cast<Instruction>(Val: V))
3870 if (MDNode *MD = IIQ.getMetadata(I, KindID: LLVMContext::MD_range))
3871 return getConstantRangeFromMetadata(RangeMD: *MD);
3872
3873 if (const Argument *A = dyn_cast<Argument>(Val: V))
3874 return A->getRange();
3875 else if (const CallBase *CB = dyn_cast<CallBase>(Val: V))
3876 return CB->getRange();
3877
3878 return std::nullopt;
3879}
3880
3881/// Given operands for an ICmpInst, see if we can fold the result.
3882/// If not, this returns null.
3883static Value *simplifyICmpInst(CmpPredicate Pred, Value *LHS, Value *RHS,
3884 const SimplifyQuery &Q, unsigned MaxRecurse) {
3885 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3886
3887 if (Constant *CLHS = dyn_cast<Constant>(Val: LHS)) {
3888 if (Constant *CRHS = dyn_cast<Constant>(Val: RHS))
3889 return ConstantFoldCompareInstOperands(Predicate: Pred, LHS: CLHS, RHS: CRHS, DL: Q.DL, TLI: Q.TLI);
3890
3891 // If we have a constant, make sure it is on the RHS.
3892 std::swap(a&: LHS, b&: RHS);
3893 Pred = CmpInst::getSwappedPredicate(pred: Pred);
3894 }
3895 assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3896
3897 Type *ITy = getCompareTy(Op: LHS); // The return type.
3898
3899 // icmp poison, X -> poison
3900 if (isa<PoisonValue>(Val: RHS))
3901 return PoisonValue::get(T: ITy);
3902
3903 // For EQ and NE, we can always pick a value for the undef to make the
3904 // predicate pass or fail, so we can return undef.
3905 // Matches behavior in llvm::ConstantFoldCompareInstruction.
3906 if (Q.isUndefValue(V: RHS) && ICmpInst::isEquality(P: Pred))
3907 return UndefValue::get(T: ITy);
3908
3909 // icmp X, X -> true/false
3910 // icmp X, undef -> true/false because undef could be X.
3911 if (LHS == RHS || Q.isUndefValue(V: RHS))
3912 return ConstantInt::get(Ty: ITy, V: CmpInst::isTrueWhenEqual(predicate: Pred));
3913
3914 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3915 return V;
3916
3917 // TODO: Sink/common this with other potentially expensive calls that use
3918 // ValueTracking? See comment below for isKnownNonEqual().
3919 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3920 return V;
3921
3922 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q))
3923 return V;
3924
3925 // If both operands have range metadata, use the metadata
3926 // to simplify the comparison.
3927 if (std::optional<ConstantRange> RhsCr = getRange(V: RHS, IIQ: Q.IIQ))
3928 if (std::optional<ConstantRange> LhsCr = getRange(V: LHS, IIQ: Q.IIQ)) {
3929 if (LhsCr->icmp(Pred, Other: *RhsCr))
3930 return ConstantInt::getTrue(Ty: ITy);
3931
3932 if (LhsCr->icmp(Pred: CmpInst::getInversePredicate(pred: Pred), Other: *RhsCr))
3933 return ConstantInt::getFalse(Ty: ITy);
3934 }
3935
3936 // Compare of cast, for example (zext X) != 0 -> X != 0
3937 if (isa<CastInst>(Val: LHS) && (isa<Constant>(Val: RHS) || isa<CastInst>(Val: RHS))) {
3938 Instruction *LI = cast<CastInst>(Val: LHS);
3939 Value *SrcOp = LI->getOperand(i: 0);
3940 Type *SrcTy = SrcOp->getType();
3941 Type *DstTy = LI->getType();
3942
3943 // Turn icmp (ptrtoint/ptrtoaddr x), (ptrtoint/ptrtoaddr/constant) into a
3944 // compare of the input if the integer type is the same size as the
3945 // pointer address type (icmp only compares the address of the pointer).
3946 if (MaxRecurse && (isa<PtrToIntInst, PtrToAddrInst>(Val: LI)) &&
3947 Q.DL.getAddressType(PtrTy: SrcTy) == DstTy) {
3948 if (Constant *RHSC = dyn_cast<Constant>(Val: RHS)) {
3949 // Transfer the cast to the constant.
3950 if (Value *V = simplifyICmpInst(Pred, LHS: SrcOp,
3951 RHS: ConstantExpr::getIntToPtr(C: RHSC, Ty: SrcTy),
3952 Q, MaxRecurse: MaxRecurse - 1))
3953 return V;
3954 } else if (isa<PtrToIntInst, PtrToAddrInst>(Val: RHS)) {
3955 auto *RI = cast<CastInst>(Val: RHS);
3956 if (RI->getOperand(i_nocapture: 0)->getType() == SrcTy)
3957 // Compare without the cast.
3958 if (Value *V = simplifyICmpInst(Pred, LHS: SrcOp, RHS: RI->getOperand(i_nocapture: 0), Q,
3959 MaxRecurse: MaxRecurse - 1))
3960 return V;
3961 }
3962 }
3963
3964 if (isa<ZExtInst>(Val: LHS)) {
3965 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3966 // same type.
3967 if (ZExtInst *RI = dyn_cast<ZExtInst>(Val: RHS)) {
3968 if (MaxRecurse && SrcTy == RI->getOperand(i_nocapture: 0)->getType())
3969 // Compare X and Y. Note that signed predicates become unsigned.
3970 if (Value *V =
3971 simplifyICmpInst(Pred: ICmpInst::getUnsignedPredicate(Pred), LHS: SrcOp,
3972 RHS: RI->getOperand(i_nocapture: 0), Q, MaxRecurse: MaxRecurse - 1))
3973 return V;
3974 }
3975 // Fold (zext X) ule (sext X), (zext X) sge (sext X) to true.
3976 else if (SExtInst *RI = dyn_cast<SExtInst>(Val: RHS)) {
3977 if (SrcOp == RI->getOperand(i_nocapture: 0)) {
3978 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_SGE)
3979 return ConstantInt::getTrue(Ty: ITy);
3980 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SLT)
3981 return ConstantInt::getFalse(Ty: ITy);
3982 }
3983 }
3984 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3985 // too. If not, then try to deduce the result of the comparison.
3986 else if (match(V: RHS, P: m_ImmConstant())) {
3987 Constant *C = dyn_cast<Constant>(Val: RHS);
3988 assert(C != nullptr);
3989
3990 // Compute the constant that would happen if we truncated to SrcTy then
3991 // reextended to DstTy.
3992 Constant *Trunc =
3993 ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy: SrcTy, DL: Q.DL);
3994 assert(Trunc && "Constant-fold of ImmConstant should not fail");
3995 Constant *RExt =
3996 ConstantFoldCastOperand(Opcode: CastInst::ZExt, C: Trunc, DestTy: DstTy, DL: Q.DL);
3997 assert(RExt && "Constant-fold of ImmConstant should not fail");
3998 Constant *AnyEq =
3999 ConstantFoldCompareInstOperands(Predicate: ICmpInst::ICMP_EQ, LHS: RExt, RHS: C, DL: Q.DL);
4000 assert(AnyEq && "Constant-fold of ImmConstant should not fail");
4001
4002 // If the re-extended constant didn't change any of the elements then
4003 // this is effectively also a case of comparing two zero-extended
4004 // values.
4005 if (AnyEq->isAllOnesValue() && MaxRecurse)
4006 if (Value *V = simplifyICmpInst(Pred: ICmpInst::getUnsignedPredicate(Pred),
4007 LHS: SrcOp, RHS: Trunc, Q, MaxRecurse: MaxRecurse - 1))
4008 return V;
4009
4010 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
4011 // there. Use this to work out the result of the comparison.
4012 if (AnyEq->isNullValue()) {
4013 switch (Pred) {
4014 default:
4015 llvm_unreachable("Unknown ICmp predicate!");
4016 // LHS <u RHS.
4017 case ICmpInst::ICMP_EQ:
4018 case ICmpInst::ICMP_UGT:
4019 case ICmpInst::ICMP_UGE:
4020 return Constant::getNullValue(Ty: ITy);
4021
4022 case ICmpInst::ICMP_NE:
4023 case ICmpInst::ICMP_ULT:
4024 case ICmpInst::ICMP_ULE:
4025 return Constant::getAllOnesValue(Ty: ITy);
4026
4027 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
4028 // is non-negative then LHS <s RHS.
4029 case ICmpInst::ICMP_SGT:
4030 case ICmpInst::ICMP_SGE:
4031 return ConstantFoldCompareInstOperands(
4032 Predicate: ICmpInst::ICMP_SLT, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4033 DL: Q.DL);
4034 case ICmpInst::ICMP_SLT:
4035 case ICmpInst::ICMP_SLE:
4036 return ConstantFoldCompareInstOperands(
4037 Predicate: ICmpInst::ICMP_SGE, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4038 DL: Q.DL);
4039 }
4040 }
4041 }
4042 }
4043
4044 if (isa<SExtInst>(Val: LHS)) {
4045 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
4046 // same type.
4047 if (SExtInst *RI = dyn_cast<SExtInst>(Val: RHS)) {
4048 if (MaxRecurse && SrcTy == RI->getOperand(i_nocapture: 0)->getType())
4049 // Compare X and Y. Note that the predicate does not change.
4050 if (Value *V = simplifyICmpInst(Pred, LHS: SrcOp, RHS: RI->getOperand(i_nocapture: 0), Q,
4051 MaxRecurse: MaxRecurse - 1))
4052 return V;
4053 }
4054 // Fold (sext X) uge (zext X), (sext X) sle (zext X) to true.
4055 else if (ZExtInst *RI = dyn_cast<ZExtInst>(Val: RHS)) {
4056 if (SrcOp == RI->getOperand(i_nocapture: 0)) {
4057 if (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_SLE)
4058 return ConstantInt::getTrue(Ty: ITy);
4059 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SGT)
4060 return ConstantInt::getFalse(Ty: ITy);
4061 }
4062 }
4063 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
4064 // too. If not, then try to deduce the result of the comparison.
4065 else if (match(V: RHS, P: m_ImmConstant())) {
4066 Constant *C = cast<Constant>(Val: RHS);
4067
4068 // Compute the constant that would happen if we truncated to SrcTy then
4069 // reextended to DstTy.
4070 Constant *Trunc =
4071 ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy: SrcTy, DL: Q.DL);
4072 assert(Trunc && "Constant-fold of ImmConstant should not fail");
4073 Constant *RExt =
4074 ConstantFoldCastOperand(Opcode: CastInst::SExt, C: Trunc, DestTy: DstTy, DL: Q.DL);
4075 assert(RExt && "Constant-fold of ImmConstant should not fail");
4076 Constant *AnyEq =
4077 ConstantFoldCompareInstOperands(Predicate: ICmpInst::ICMP_EQ, LHS: RExt, RHS: C, DL: Q.DL);
4078 assert(AnyEq && "Constant-fold of ImmConstant should not fail");
4079
4080 // If the re-extended constant didn't change then this is effectively
4081 // also a case of comparing two sign-extended values.
4082 if (AnyEq->isAllOnesValue() && MaxRecurse)
4083 if (Value *V =
4084 simplifyICmpInst(Pred, LHS: SrcOp, RHS: Trunc, Q, MaxRecurse: MaxRecurse - 1))
4085 return V;
4086
4087 // Otherwise the upper bits of LHS are all equal, while RHS has varying
4088 // bits there. Use this to work out the result of the comparison.
4089 if (AnyEq->isNullValue()) {
4090 switch (Pred.getPreferredSignedPredicate()) {
4091 default:
4092 llvm_unreachable("Unknown ICmp predicate!");
4093 case ICmpInst::ICMP_EQ:
4094 return Constant::getNullValue(Ty: ITy);
4095 case ICmpInst::ICMP_NE:
4096 return Constant::getAllOnesValue(Ty: ITy);
4097
4098 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
4099 // LHS >s RHS.
4100 case ICmpInst::ICMP_SGT:
4101 case ICmpInst::ICMP_SGE:
4102 return ConstantFoldCompareInstOperands(
4103 Predicate: ICmpInst::ICMP_SLT, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4104 DL: Q.DL);
4105 case ICmpInst::ICMP_SLT:
4106 case ICmpInst::ICMP_SLE:
4107 return ConstantFoldCompareInstOperands(
4108 Predicate: ICmpInst::ICMP_SGE, LHS: C, RHS: Constant::getNullValue(Ty: C->getType()),
4109 DL: Q.DL);
4110
4111 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
4112 // LHS >u RHS.
4113 case ICmpInst::ICMP_UGT:
4114 case ICmpInst::ICMP_UGE:
4115 // Comparison is true iff the LHS <s 0.
4116 if (MaxRecurse)
4117 if (Value *V = simplifyICmpInst(Pred: ICmpInst::ICMP_SLT, LHS: SrcOp,
4118 RHS: Constant::getNullValue(Ty: SrcTy), Q,
4119 MaxRecurse: MaxRecurse - 1))
4120 return V;
4121 break;
4122 case ICmpInst::ICMP_ULT:
4123 case ICmpInst::ICMP_ULE:
4124 // Comparison is true iff the LHS >=s 0.
4125 if (MaxRecurse)
4126 if (Value *V = simplifyICmpInst(Pred: ICmpInst::ICMP_SGE, LHS: SrcOp,
4127 RHS: Constant::getNullValue(Ty: SrcTy), Q,
4128 MaxRecurse: MaxRecurse - 1))
4129 return V;
4130 break;
4131 }
4132 }
4133 }
4134 }
4135 }
4136
4137 // icmp eq|ne X, Y -> false|true if X != Y
4138 // This is potentially expensive, and we have already computedKnownBits for
4139 // compares with 0 above here, so only try this for a non-zero compare.
4140 if (ICmpInst::isEquality(P: Pred) && !match(V: RHS, P: m_Zero()) &&
4141 isKnownNonEqual(V1: LHS, V2: RHS, SQ: Q)) {
4142 return Pred == ICmpInst::ICMP_NE ? getTrue(Ty: ITy) : getFalse(Ty: ITy);
4143 }
4144
4145 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
4146 return V;
4147
4148 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
4149 return V;
4150
4151 if (Value *V = simplifyICmpWithIntrinsicOnLHS(Pred, LHS, RHS))
4152 return V;
4153 if (Value *V = simplifyICmpWithIntrinsicOnLHS(
4154 Pred: ICmpInst::getSwappedPredicate(pred: Pred), LHS: RHS, RHS: LHS))
4155 return V;
4156
4157 if (Value *V = simplifyICmpUsingMonotonicValues(Pred, LHS, RHS, Q))
4158 return V;
4159 if (Value *V = simplifyICmpUsingMonotonicValues(
4160 Pred: ICmpInst::getSwappedPredicate(pred: Pred), LHS: RHS, RHS: LHS, Q))
4161 return V;
4162
4163 if (Value *V = simplifyICmpWithDominatingAssume(Predicate: Pred, LHS, RHS, Q))
4164 return V;
4165
4166 if (std::optional<bool> Res =
4167 isImpliedByDomCondition(Pred, LHS, RHS, ContextI: Q.CxtI, DL: Q.DL))
4168 return ConstantInt::getBool(Ty: ITy, V: *Res);
4169
4170 // Simplify comparisons of related pointers using a powerful, recursive
4171 // GEP-walk when we have target data available..
4172 if (LHS->getType()->isPointerTy())
4173 if (auto *C = computePointerICmp(Pred, LHS, RHS, Q))
4174 return C;
4175
4176 // If the comparison is with the result of a select instruction, check whether
4177 // comparing with either branch of the select always yields the same value.
4178 if (isa<SelectInst>(Val: LHS) || isa<SelectInst>(Val: RHS))
4179 if (Value *V = threadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
4180 return V;
4181
4182 // If the comparison is with the result of a phi instruction, check whether
4183 // doing the compare with each incoming phi value yields a common result.
4184 if (isa<PHINode>(Val: LHS) || isa<PHINode>(Val: RHS))
4185 if (Value *V = threadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
4186 return V;
4187
4188 return nullptr;
4189}
4190
4191Value *llvm::simplifyICmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
4192 const SimplifyQuery &Q) {
4193 return ::simplifyICmpInst(Pred: Predicate, LHS, RHS, Q, MaxRecurse: RecursionLimit);
4194}
4195
4196/// Given operands for an FCmpInst, see if we can fold the result.
4197/// If not, this returns null.
4198static Value *simplifyFCmpInst(CmpPredicate Pred, Value *LHS, Value *RHS,
4199 FastMathFlags FMF, const SimplifyQuery &Q,
4200 unsigned MaxRecurse) {
4201 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
4202
4203 if (Constant *CLHS = dyn_cast<Constant>(Val: LHS)) {
4204 if (Constant *CRHS = dyn_cast<Constant>(Val: RHS)) {
4205 // if the folding isn't successfull, fall back to the rest of the logic
4206 if (auto *Result = ConstantFoldCompareInstOperands(Predicate: Pred, LHS: CLHS, RHS: CRHS, DL: Q.DL,
4207 TLI: Q.TLI, I: Q.CxtI))
4208 return Result;
4209 } else {
4210 // If we have a constant, make sure it is on the RHS.
4211 std::swap(a&: LHS, b&: RHS);
4212 Pred = CmpInst::getSwappedPredicate(pred: Pred);
4213 }
4214 }
4215
4216 // Fold trivial predicates.
4217 Type *RetTy = getCompareTy(Op: LHS);
4218 if (Pred == FCmpInst::FCMP_FALSE)
4219 return getFalse(Ty: RetTy);
4220 if (Pred == FCmpInst::FCMP_TRUE)
4221 return getTrue(Ty: RetTy);
4222
4223 // fcmp pred x, poison and fcmp pred poison, x
4224 // fold to poison
4225 if (isa<PoisonValue>(Val: LHS) || isa<PoisonValue>(Val: RHS))
4226 return PoisonValue::get(T: RetTy);
4227
4228 // fcmp pred x, undef and fcmp pred undef, x
4229 // fold to true if unordered, false if ordered
4230 if (Q.isUndefValue(V: LHS) || Q.isUndefValue(V: RHS)) {
4231 // Choosing NaN for the undef will always make unordered comparison succeed
4232 // and ordered comparison fail.
4233 return ConstantInt::get(Ty: RetTy, V: CmpInst::isUnordered(predicate: Pred));
4234 }
4235
4236 // fcmp x,x -> true/false. Not all compares are foldable.
4237 if (LHS == RHS) {
4238 if (CmpInst::isTrueWhenEqual(predicate: Pred))
4239 return getTrue(Ty: RetTy);
4240 if (CmpInst::isFalseWhenEqual(predicate: Pred))
4241 return getFalse(Ty: RetTy);
4242 }
4243
4244 // Fold (un)ordered comparison if we can determine there are no NaNs.
4245 //
4246 // This catches the 2 variable input case, constants are handled below as a
4247 // class-like compare.
4248 if (Pred == FCmpInst::FCMP_ORD || Pred == FCmpInst::FCMP_UNO) {
4249 KnownFPClass RHSClass = computeKnownFPClass(V: RHS, InterestedClasses: fcAllFlags, SQ: Q);
4250 KnownFPClass LHSClass = computeKnownFPClass(V: LHS, InterestedClasses: fcAllFlags, SQ: Q);
4251
4252 if (FMF.noNaNs() ||
4253 (RHSClass.isKnownNeverNaN() && LHSClass.isKnownNeverNaN()))
4254 return ConstantInt::get(Ty: RetTy, V: Pred == FCmpInst::FCMP_ORD);
4255
4256 if (RHSClass.isKnownAlwaysNaN() || LHSClass.isKnownAlwaysNaN())
4257 return ConstantInt::get(Ty: RetTy, V: Pred == CmpInst::FCMP_UNO);
4258 }
4259
4260 if (std::optional<bool> Res =
4261 isImpliedByDomCondition(Pred, LHS, RHS, ContextI: Q.CxtI, DL: Q.DL))
4262 return ConstantInt::getBool(Ty: RetTy, V: *Res);
4263
4264 const APFloat *C = nullptr;
4265 match(V: RHS, P: m_APFloatAllowPoison(Res&: C));
4266 std::optional<KnownFPClass> FullKnownClassLHS;
4267
4268 // Lazily compute the possible classes for LHS. Avoid computing it twice if
4269 // RHS is a 0.
4270 auto computeLHSClass = [=, &FullKnownClassLHS](FPClassTest InterestedFlags =
4271 fcAllFlags) {
4272 if (FullKnownClassLHS)
4273 return *FullKnownClassLHS;
4274 return computeKnownFPClass(V: LHS, FMF, InterestedClasses: InterestedFlags, SQ: Q);
4275 };
4276
4277 if (C && Q.CxtI) {
4278 // Fold out compares that express a class test.
4279 //
4280 // FIXME: Should be able to perform folds without context
4281 // instruction. Always pass in the context function?
4282
4283 const Function *ParentF = Q.CxtI->getFunction();
4284 auto [ClassVal, ClassTest] = fcmpToClassTest(Pred, F: *ParentF, LHS, ConstRHS: C);
4285 if (ClassVal) {
4286 FullKnownClassLHS = computeLHSClass();
4287 if ((FullKnownClassLHS->KnownFPClasses & ClassTest) == fcNone)
4288 return getFalse(Ty: RetTy);
4289 if ((FullKnownClassLHS->KnownFPClasses & ~ClassTest) == fcNone)
4290 return getTrue(Ty: RetTy);
4291 }
4292 }
4293
4294 // Handle fcmp with constant RHS.
4295 if (C) {
4296 // TODO: If we always required a context function, we wouldn't need to
4297 // special case nans.
4298 if (C->isNaN())
4299 return ConstantInt::get(Ty: RetTy, V: CmpInst::isUnordered(predicate: Pred));
4300
4301 // TODO: Need version fcmpToClassTest which returns implied class when the
4302 // compare isn't a complete class test. e.g. > 1.0 implies fcPositive, but
4303 // isn't implementable as a class call.
4304 if (C->isNegative() && !C->isNegZero()) {
4305 FPClassTest Interested = KnownFPClass::OrderedLessThanZeroMask;
4306
4307 // TODO: We can catch more cases by using a range check rather than
4308 // relying on CannotBeOrderedLessThanZero.
4309 switch (Pred) {
4310 case FCmpInst::FCMP_UGE:
4311 case FCmpInst::FCMP_UGT:
4312 case FCmpInst::FCMP_UNE: {
4313 KnownFPClass KnownClass = computeLHSClass(Interested);
4314
4315 // (X >= 0) implies (X > C) when (C < 0)
4316 if (KnownClass.cannotBeOrderedLessThanZero())
4317 return getTrue(Ty: RetTy);
4318 break;
4319 }
4320 case FCmpInst::FCMP_OEQ:
4321 case FCmpInst::FCMP_OLE:
4322 case FCmpInst::FCMP_OLT: {
4323 KnownFPClass KnownClass = computeLHSClass(Interested);
4324
4325 // (X >= 0) implies !(X < C) when (C < 0)
4326 if (KnownClass.cannotBeOrderedLessThanZero())
4327 return getFalse(Ty: RetTy);
4328 break;
4329 }
4330 default:
4331 break;
4332 }
4333 }
4334 // Check FCmp of [min/maxnum or min/maximumnum with const] with other const.
4335 const APFloat *C2;
4336 bool IsMax = match(V: LHS, P: m_FMaxNum_or_FMaximumNum(Op0: m_Value(), Op1: m_APFloat(Res&: C2)));
4337 bool IsMin = match(V: LHS, P: m_FMinNum_or_FMinimumNum(Op0: m_Value(), Op1: m_APFloat(Res&: C2)));
4338 if ((IsMax && *C2 > *C) || (IsMin && *C2 < *C)) {
4339 // The ordered relationship and min/maxnum or min/maximumnum guarantee
4340 // that we do not have NaN constants, so ordered/unordered preds are
4341 // handled the same.
4342 switch (Pred) {
4343 case FCmpInst::FCMP_OEQ:
4344 case FCmpInst::FCMP_UEQ:
4345 // minnum(X, LesserC) == C --> false
4346 // maxnum(X, GreaterC) == C --> false
4347 return getFalse(Ty: RetTy);
4348 case FCmpInst::FCMP_ONE:
4349 case FCmpInst::FCMP_UNE:
4350 // minnum(X, LesserC) != C --> true
4351 // maxnum(X, GreaterC) != C --> true
4352 return getTrue(Ty: RetTy);
4353 case FCmpInst::FCMP_OGE:
4354 case FCmpInst::FCMP_UGE:
4355 case FCmpInst::FCMP_OGT:
4356 case FCmpInst::FCMP_UGT:
4357 // minnum(X, LesserC) >= C --> false
4358 // minnum(X, LesserC) > C --> false
4359 // maxnum(X, GreaterC) >= C --> true
4360 // maxnum(X, GreaterC) > C --> true
4361 return ConstantInt::get(Ty: RetTy, V: IsMax);
4362 case FCmpInst::FCMP_OLE:
4363 case FCmpInst::FCMP_ULE:
4364 case FCmpInst::FCMP_OLT:
4365 case FCmpInst::FCMP_ULT:
4366 // minnum(X, LesserC) <= C --> true
4367 // minnum(X, LesserC) < C --> true
4368 // maxnum(X, GreaterC) <= C --> false
4369 // maxnum(X, GreaterC) < C --> false
4370 return ConstantInt::get(Ty: RetTy, V: !IsMax);
4371 default:
4372 // TRUE/FALSE/ORD/UNO should be handled before this.
4373 llvm_unreachable("Unexpected fcmp predicate");
4374 }
4375 }
4376 }
4377
4378 // TODO: Could fold this with above if there were a matcher which returned all
4379 // classes in a non-splat vector.
4380 if (match(V: RHS, P: m_AnyZeroFP())) {
4381 switch (Pred) {
4382 case FCmpInst::FCMP_OGE:
4383 case FCmpInst::FCMP_ULT: {
4384 FPClassTest Interested = KnownFPClass::OrderedLessThanZeroMask;
4385 if (!FMF.noNaNs())
4386 Interested |= fcNan;
4387
4388 KnownFPClass Known = computeLHSClass(Interested);
4389
4390 // Positive or zero X >= 0.0 --> true
4391 // Positive or zero X < 0.0 --> false
4392 if ((FMF.noNaNs() || Known.isKnownNeverNaN()) &&
4393 Known.cannotBeOrderedLessThanZero())
4394 return Pred == FCmpInst::FCMP_OGE ? getTrue(Ty: RetTy) : getFalse(Ty: RetTy);
4395 break;
4396 }
4397 case FCmpInst::FCMP_UGE:
4398 case FCmpInst::FCMP_OLT: {
4399 FPClassTest Interested = KnownFPClass::OrderedLessThanZeroMask;
4400 KnownFPClass Known = computeLHSClass(Interested);
4401
4402 // Positive or zero or nan X >= 0.0 --> true
4403 // Positive or zero or nan X < 0.0 --> false
4404 if (Known.cannotBeOrderedLessThanZero())
4405 return Pred == FCmpInst::FCMP_UGE ? getTrue(Ty: RetTy) : getFalse(Ty: RetTy);
4406 break;
4407 }
4408 default:
4409 break;
4410 }
4411 }
4412
4413 // If the comparison is with the result of a select instruction, check whether
4414 // comparing with either branch of the select always yields the same value.
4415 if (isa<SelectInst>(Val: LHS) || isa<SelectInst>(Val: RHS))
4416 if (Value *V = threadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
4417 return V;
4418
4419 // If the comparison is with the result of a phi instruction, check whether
4420 // doing the compare with each incoming phi value yields a common result.
4421 if (isa<PHINode>(Val: LHS) || isa<PHINode>(Val: RHS))
4422 if (Value *V = threadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
4423 return V;
4424
4425 return nullptr;
4426}
4427
4428Value *llvm::simplifyFCmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
4429 FastMathFlags FMF, const SimplifyQuery &Q) {
4430 return ::simplifyFCmpInst(Pred: Predicate, LHS, RHS, FMF, Q, MaxRecurse: RecursionLimit);
4431}
4432
4433static Value *simplifyWithOpsReplaced(Value *V,
4434 ArrayRef<std::pair<Value *, Value *>> Ops,
4435 const SimplifyQuery &Q,
4436 bool AllowRefinement,
4437 SmallVectorImpl<Instruction *> *DropFlags,
4438 unsigned MaxRecurse) {
4439 assert((AllowRefinement || !Q.CanUseUndef) &&
4440 "If AllowRefinement=false then CanUseUndef=false");
4441 for (const auto &OpAndRepOp : Ops) {
4442 // We cannot replace a constant, and shouldn't even try.
4443 if (isa<Constant>(Val: OpAndRepOp.first))
4444 return nullptr;
4445
4446 // Trivial replacement.
4447 if (V == OpAndRepOp.first)
4448 return OpAndRepOp.second;
4449 }
4450
4451 if (!MaxRecurse--)
4452 return nullptr;
4453
4454 auto *I = dyn_cast<Instruction>(Val: V);
4455 if (!I)
4456 return nullptr;
4457
4458 // The arguments of a phi node might refer to a value from a previous
4459 // cycle iteration.
4460 if (isa<PHINode>(Val: I))
4461 return nullptr;
4462
4463 // Don't fold away llvm.is.constant checks based on assumptions.
4464 if (match(V: I, P: m_Intrinsic<Intrinsic::is_constant>()))
4465 return nullptr;
4466
4467 // Don't simplify freeze.
4468 if (isa<FreezeInst>(Val: I))
4469 return nullptr;
4470
4471 for (const auto &OpAndRepOp : Ops) {
4472 // For vector types, the simplification must hold per-lane, so forbid
4473 // potentially cross-lane operations like shufflevector.
4474 if (OpAndRepOp.first->getType()->isVectorTy() &&
4475 !isNotCrossLaneOperation(I))
4476 return nullptr;
4477 }
4478
4479 // Replace Op with RepOp in instruction operands.
4480 SmallVector<Value *, 8> NewOps;
4481 bool AnyReplaced = false;
4482 for (Value *InstOp : I->operands()) {
4483 if (Value *NewInstOp = simplifyWithOpsReplaced(
4484 V: InstOp, Ops, Q, AllowRefinement, DropFlags, MaxRecurse)) {
4485 NewOps.push_back(Elt: NewInstOp);
4486 AnyReplaced = InstOp != NewInstOp;
4487 } else {
4488 NewOps.push_back(Elt: InstOp);
4489 }
4490
4491 // Bail out if any operand is undef and SimplifyQuery disables undef
4492 // simplification. Constant folding currently doesn't respect this option.
4493 if (isa<UndefValue>(Val: NewOps.back()) && !Q.CanUseUndef)
4494 return nullptr;
4495 }
4496
4497 if (!AnyReplaced)
4498 return nullptr;
4499
4500 if (!AllowRefinement) {
4501 // General InstSimplify functions may refine the result, e.g. by returning
4502 // a constant for a potentially poison value. To avoid this, implement only
4503 // a few non-refining but profitable transforms here.
4504
4505 if (auto *BO = dyn_cast<BinaryOperator>(Val: I)) {
4506 unsigned Opcode = BO->getOpcode();
4507 // id op x -> x, x op id -> x
4508 // Exclude floats, because x op id may produce a different NaN value.
4509 if (!BO->getType()->isFPOrFPVectorTy()) {
4510 if (NewOps[0] == ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType()))
4511 return NewOps[1];
4512 if (NewOps[1] == ConstantExpr::getBinOpIdentity(Opcode, Ty: I->getType(),
4513 /* RHS */ AllowRHSConstant: true))
4514 return NewOps[0];
4515 }
4516
4517 // x & x -> x, x | x -> x
4518 if ((Opcode == Instruction::And || Opcode == Instruction::Or) &&
4519 NewOps[0] == NewOps[1]) {
4520 // or disjoint x, x results in poison.
4521 if (auto *PDI = dyn_cast<PossiblyDisjointInst>(Val: BO)) {
4522 if (PDI->isDisjoint()) {
4523 if (!DropFlags)
4524 return nullptr;
4525 DropFlags->push_back(Elt: BO);
4526 }
4527 }
4528 return NewOps[0];
4529 }
4530
4531 // x - x -> 0, x ^ x -> 0. This is non-refining, because x is non-poison
4532 // by assumption and this case never wraps, so nowrap flags can be
4533 // ignored.
4534 if ((Opcode == Instruction::Sub || Opcode == Instruction::Xor) &&
4535 NewOps[0] == NewOps[1] &&
4536 any_of(Range&: Ops, P: [=](const auto &Rep) { return NewOps[0] == Rep.second; }))
4537 return Constant::getNullValue(Ty: I->getType());
4538
4539 // If we are substituting an absorber constant into a binop and extra
4540 // poison can't leak if we remove the select -- because both operands of
4541 // the binop are based on the same value -- then it may be safe to replace
4542 // the value with the absorber constant. Examples:
4543 // (Op == 0) ? 0 : (Op & -Op) --> Op & -Op
4544 // (Op == 0) ? 0 : (Op * (binop Op, C)) --> Op * (binop Op, C)
4545 // (Op == -1) ? -1 : (Op | (binop C, Op) --> Op | (binop C, Op)
4546 Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, Ty: I->getType());
4547 if ((NewOps[0] == Absorber || NewOps[1] == Absorber) &&
4548 any_of(Range&: Ops,
4549 P: [=](const auto &Rep) { return impliesPoison(BO, Rep.first); }))
4550 return Absorber;
4551 }
4552
4553 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
4554 // `x == y ? 0 : ucmp(x, y)` where under the replacement y -> x,
4555 // `ucmp(x, x)` becomes `0`.
4556 if ((II->getIntrinsicID() == Intrinsic::scmp ||
4557 II->getIntrinsicID() == Intrinsic::ucmp) &&
4558 NewOps[0] == NewOps[1]) {
4559 if (II->hasPoisonGeneratingAnnotations()) {
4560 if (!DropFlags)
4561 return nullptr;
4562
4563 DropFlags->push_back(Elt: II);
4564 }
4565
4566 return ConstantInt::get(Ty: I->getType(), V: 0);
4567 }
4568
4569 if (auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: II)) {
4570 const APInt Identity = MMI->getIdentity();
4571
4572 Value *Result = nullptr;
4573 if (match(V: NewOps[0], P: m_SpecificInt(V: Identity)))
4574 Result = NewOps[1];
4575 else if (match(V: NewOps[1], P: m_SpecificInt(V: Identity)))
4576 Result = NewOps[0];
4577
4578 if (Result) {
4579 if (II->hasPoisonGeneratingAnnotations()) {
4580 if (!DropFlags)
4581 return nullptr;
4582
4583 DropFlags->push_back(Elt: II);
4584 }
4585
4586 return Result;
4587 }
4588 }
4589 }
4590
4591 if (isa<GetElementPtrInst>(Val: I)) {
4592 // getelementptr x, 0 -> x.
4593 // This never returns poison, even if inbounds is set.
4594 if (NewOps.size() == 2 && match(V: NewOps[1], P: m_Zero()))
4595 return NewOps[0];
4596 }
4597 } else {
4598 // The simplification queries below may return the original value. Consider:
4599 // %div = udiv i32 %arg, %arg2
4600 // %mul = mul nsw i32 %div, %arg2
4601 // %cmp = icmp eq i32 %mul, %arg
4602 // %sel = select i1 %cmp, i32 %div, i32 undef
4603 // Replacing %arg by %mul, %div becomes "udiv i32 %mul, %arg2", which
4604 // simplifies back to %arg. This can only happen because %mul does not
4605 // dominate %div. To ensure a consistent return value contract, we make sure
4606 // that this case returns nullptr as well.
4607 auto PreventSelfSimplify = [V](Value *Simplified) {
4608 return Simplified != V ? Simplified : nullptr;
4609 };
4610
4611 return PreventSelfSimplify(
4612 ::simplifyInstructionWithOperands(I, NewOps, SQ: Q, MaxRecurse));
4613 }
4614
4615 // If all operands are constant after substituting Op for RepOp then we can
4616 // constant fold the instruction.
4617 SmallVector<Constant *, 8> ConstOps;
4618 for (Value *NewOp : NewOps) {
4619 if (Constant *ConstOp = dyn_cast<Constant>(Val: NewOp))
4620 ConstOps.push_back(Elt: ConstOp);
4621 else
4622 return nullptr;
4623 }
4624
4625 // Consider:
4626 // %cmp = icmp eq i32 %x, 2147483647
4627 // %add = add nsw i32 %x, 1
4628 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
4629 //
4630 // We can't replace %sel with %add unless we strip away the flags (which
4631 // will be done in InstCombine).
4632 // TODO: This may be unsound, because it only catches some forms of
4633 // refinement.
4634 if (!AllowRefinement) {
4635 auto *II = dyn_cast<IntrinsicInst>(Val: I);
4636 if (canCreatePoison(Op: cast<Operator>(Val: I), ConsiderFlagsAndMetadata: !DropFlags)) {
4637 // abs cannot create poison if the value is known to never be int_min.
4638 if (II && II->getIntrinsicID() == Intrinsic::abs) {
4639 if (!ConstOps[0]->isNotMinSignedValue())
4640 return nullptr;
4641 } else
4642 return nullptr;
4643 }
4644
4645 if (DropFlags && II) {
4646 // If we're going to change the poison flag of abs/ctz to false, also
4647 // perform constant folding that way, so we get an integer instead of a
4648 // poison value here.
4649 switch (II->getIntrinsicID()) {
4650 case Intrinsic::abs:
4651 case Intrinsic::ctlz:
4652 case Intrinsic::cttz:
4653 ConstOps[1] = ConstantInt::getFalse(Context&: I->getContext());
4654 break;
4655 default:
4656 break;
4657 }
4658 }
4659
4660 Constant *Res = ConstantFoldInstOperands(I, Ops: ConstOps, DL: Q.DL, TLI: Q.TLI,
4661 /*AllowNonDeterministic=*/false);
4662 if (DropFlags && Res && I->hasPoisonGeneratingAnnotations())
4663 DropFlags->push_back(Elt: I);
4664 return Res;
4665 }
4666
4667 return ConstantFoldInstOperands(I, Ops: ConstOps, DL: Q.DL, TLI: Q.TLI,
4668 /*AllowNonDeterministic=*/false);
4669}
4670
4671static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
4672 const SimplifyQuery &Q,
4673 bool AllowRefinement,
4674 SmallVectorImpl<Instruction *> *DropFlags,
4675 unsigned MaxRecurse) {
4676 return simplifyWithOpsReplaced(V, Ops: {{Op, RepOp}}, Q, AllowRefinement,
4677 DropFlags, MaxRecurse);
4678}
4679
4680Value *llvm::simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
4681 const SimplifyQuery &Q,
4682 bool AllowRefinement,
4683 SmallVectorImpl<Instruction *> *DropFlags) {
4684 // If refinement is disabled, also disable undef simplifications (which are
4685 // always refinements) in SimplifyQuery.
4686 if (!AllowRefinement)
4687 return ::simplifyWithOpReplaced(V, Op, RepOp, Q: Q.getWithoutUndef(),
4688 AllowRefinement, DropFlags, MaxRecurse: RecursionLimit);
4689 return ::simplifyWithOpReplaced(V, Op, RepOp, Q, AllowRefinement, DropFlags,
4690 MaxRecurse: RecursionLimit);
4691}
4692
4693/// Try to simplify a select instruction when its condition operand is an
4694/// integer comparison where one operand of the compare is a constant.
4695static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
4696 const APInt *Y, bool TrueWhenUnset) {
4697 const APInt *C;
4698
4699 // (X & Y) == 0 ? X & ~Y : X --> X
4700 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
4701 if (FalseVal == X && match(V: TrueVal, P: m_And(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4702 *Y == ~*C)
4703 return TrueWhenUnset ? FalseVal : TrueVal;
4704
4705 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
4706 // (X & Y) != 0 ? X : X & ~Y --> X
4707 if (TrueVal == X && match(V: FalseVal, P: m_And(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4708 *Y == ~*C)
4709 return TrueWhenUnset ? FalseVal : TrueVal;
4710
4711 if (Y->isPowerOf2()) {
4712 // (X & Y) == 0 ? X | Y : X --> X | Y
4713 // (X & Y) != 0 ? X | Y : X --> X
4714 if (FalseVal == X && match(V: TrueVal, P: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4715 *Y == *C) {
4716 // We can't return the or if it has the disjoint flag.
4717 if (TrueWhenUnset && cast<PossiblyDisjointInst>(Val: TrueVal)->isDisjoint())
4718 return nullptr;
4719 return TrueWhenUnset ? TrueVal : FalseVal;
4720 }
4721
4722 // (X & Y) == 0 ? X : X | Y --> X
4723 // (X & Y) != 0 ? X : X | Y --> X | Y
4724 if (TrueVal == X && match(V: FalseVal, P: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C))) &&
4725 *Y == *C) {
4726 // We can't return the or if it has the disjoint flag.
4727 if (!TrueWhenUnset && cast<PossiblyDisjointInst>(Val: FalseVal)->isDisjoint())
4728 return nullptr;
4729 return TrueWhenUnset ? TrueVal : FalseVal;
4730 }
4731 }
4732
4733 return nullptr;
4734}
4735
4736static Value *simplifyCmpSelOfMaxMin(Value *CmpLHS, Value *CmpRHS,
4737 CmpPredicate Pred, Value *TVal,
4738 Value *FVal) {
4739 // Canonicalize common cmp+sel operand as CmpLHS.
4740 if (CmpRHS == TVal || CmpRHS == FVal) {
4741 std::swap(a&: CmpLHS, b&: CmpRHS);
4742 Pred = ICmpInst::getSwappedPredicate(pred: Pred);
4743 }
4744
4745 // Canonicalize common cmp+sel operand as TVal.
4746 if (CmpLHS == FVal) {
4747 std::swap(a&: TVal, b&: FVal);
4748 Pred = ICmpInst::getInversePredicate(pred: Pred);
4749 }
4750
4751 // A vector select may be shuffling together elements that are equivalent
4752 // based on the max/min/select relationship.
4753 Value *X = CmpLHS, *Y = CmpRHS;
4754 bool PeekedThroughSelectShuffle = false;
4755 auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: FVal);
4756 if (Shuf && Shuf->isSelect()) {
4757 if (Shuf->getOperand(i_nocapture: 0) == Y)
4758 FVal = Shuf->getOperand(i_nocapture: 1);
4759 else if (Shuf->getOperand(i_nocapture: 1) == Y)
4760 FVal = Shuf->getOperand(i_nocapture: 0);
4761 else
4762 return nullptr;
4763 PeekedThroughSelectShuffle = true;
4764 }
4765
4766 // (X pred Y) ? X : max/min(X, Y)
4767 auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: FVal);
4768 if (!MMI || TVal != X ||
4769 !match(V: FVal, P: m_c_MaxOrMin(L: m_Specific(V: X), R: m_Specific(V: Y))))
4770 return nullptr;
4771
4772 // (X > Y) ? X : max(X, Y) --> max(X, Y)
4773 // (X >= Y) ? X : max(X, Y) --> max(X, Y)
4774 // (X < Y) ? X : min(X, Y) --> min(X, Y)
4775 // (X <= Y) ? X : min(X, Y) --> min(X, Y)
4776 //
4777 // The equivalence allows a vector select (shuffle) of max/min and Y. Ex:
4778 // (X > Y) ? X : (Z ? max(X, Y) : Y)
4779 // If Z is true, this reduces as above, and if Z is false:
4780 // (X > Y) ? X : Y --> max(X, Y)
4781 ICmpInst::Predicate MMPred = MMI->getPredicate();
4782 if (MMPred == CmpInst::getStrictPredicate(pred: Pred))
4783 return MMI;
4784
4785 // Other transforms are not valid with a shuffle.
4786 if (PeekedThroughSelectShuffle)
4787 return nullptr;
4788
4789 // (X == Y) ? X : max/min(X, Y) --> max/min(X, Y)
4790 if (Pred == CmpInst::ICMP_EQ)
4791 return MMI;
4792
4793 // (X != Y) ? X : max/min(X, Y) --> X
4794 if (Pred == CmpInst::ICMP_NE)
4795 return X;
4796
4797 // (X < Y) ? X : max(X, Y) --> X
4798 // (X <= Y) ? X : max(X, Y) --> X
4799 // (X > Y) ? X : min(X, Y) --> X
4800 // (X >= Y) ? X : min(X, Y) --> X
4801 ICmpInst::Predicate InvPred = CmpInst::getInversePredicate(pred: Pred);
4802 if (MMPred == CmpInst::getStrictPredicate(pred: InvPred))
4803 return X;
4804
4805 return nullptr;
4806}
4807
4808/// An alternative way to test if a bit is set or not.
4809/// uses e.g. sgt/slt or trunc instead of eq/ne.
4810static Value *simplifySelectWithBitTest(Value *CondVal, Value *TrueVal,
4811 Value *FalseVal) {
4812 if (auto Res = decomposeBitTest(Cond: CondVal))
4813 return simplifySelectBitTest(TrueVal, FalseVal, X: Res->X, Y: &Res->Mask,
4814 TrueWhenUnset: Res->Pred == ICmpInst::ICMP_EQ);
4815
4816 return nullptr;
4817}
4818
4819/// Try to simplify a select instruction when its condition operand is an
4820/// integer equality or floating-point equivalence comparison.
4821static Value *simplifySelectWithEquivalence(
4822 ArrayRef<std::pair<Value *, Value *>> Replacements, Value *TrueVal,
4823 Value *FalseVal, const SimplifyQuery &Q, unsigned MaxRecurse) {
4824 Value *SimplifiedFalseVal =
4825 simplifyWithOpsReplaced(V: FalseVal, Ops: Replacements, Q: Q.getWithoutUndef(),
4826 /* AllowRefinement */ false,
4827 /* DropFlags */ nullptr, MaxRecurse);
4828 if (!SimplifiedFalseVal)
4829 SimplifiedFalseVal = FalseVal;
4830
4831 Value *SimplifiedTrueVal =
4832 simplifyWithOpsReplaced(V: TrueVal, Ops: Replacements, Q,
4833 /* AllowRefinement */ true,
4834 /* DropFlags */ nullptr, MaxRecurse);
4835 if (!SimplifiedTrueVal)
4836 SimplifiedTrueVal = TrueVal;
4837
4838 if (SimplifiedFalseVal == SimplifiedTrueVal)
4839 return FalseVal;
4840
4841 return nullptr;
4842}
4843
4844/// Try to simplify a select instruction when its condition operand is an
4845/// integer comparison.
4846static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
4847 Value *FalseVal,
4848 const SimplifyQuery &Q,
4849 unsigned MaxRecurse) {
4850 CmpPredicate Pred;
4851 Value *CmpLHS, *CmpRHS;
4852 if (!match(V: CondVal, P: m_ICmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS))))
4853 return nullptr;
4854
4855 if (Value *V = simplifyCmpSelOfMaxMin(CmpLHS, CmpRHS, Pred, TVal: TrueVal, FVal: FalseVal))
4856 return V;
4857
4858 // Canonicalize ne to eq predicate.
4859 if (Pred == ICmpInst::ICMP_NE) {
4860 Pred = ICmpInst::ICMP_EQ;
4861 std::swap(a&: TrueVal, b&: FalseVal);
4862 }
4863
4864 // Check for integer min/max with a limit constant:
4865 // X > MIN_INT ? X : MIN_INT --> X
4866 // X < MAX_INT ? X : MAX_INT --> X
4867 if (TrueVal->getType()->isIntOrIntVectorTy()) {
4868 Value *X, *Y;
4869 SelectPatternFlavor SPF =
4870 matchDecomposedSelectPattern(CmpI: cast<ICmpInst>(Val: CondVal), TrueVal, FalseVal,
4871 LHS&: X, RHS&: Y)
4872 .Flavor;
4873 if (SelectPatternResult::isMinOrMax(SPF) && Pred == getMinMaxPred(SPF)) {
4874 APInt LimitC = getMinMaxLimit(SPF: getInverseMinMaxFlavor(SPF),
4875 BitWidth: X->getType()->getScalarSizeInBits());
4876 if (match(V: Y, P: m_SpecificInt(V: LimitC)))
4877 return X;
4878 }
4879 }
4880
4881 if (Pred == ICmpInst::ICMP_EQ && match(V: CmpRHS, P: m_Zero())) {
4882 Value *X;
4883 const APInt *Y;
4884 if (match(V: CmpLHS, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: Y))))
4885 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
4886 /*TrueWhenUnset=*/true))
4887 return V;
4888
4889 // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
4890 Value *ShAmt;
4891 auto isFsh = m_CombineOr(Ps: m_FShl(Op0: m_Value(V&: X), Op1: m_Value(), Op2: m_Value(V&: ShAmt)),
4892 Ps: m_FShr(Op0: m_Value(), Op1: m_Value(V&: X), Op2: m_Value(V&: ShAmt)));
4893 // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
4894 // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
4895 if (match(V: TrueVal, P: isFsh) && FalseVal == X && CmpLHS == ShAmt)
4896 return X;
4897
4898 // Test for a zero-shift-guard-op around rotates. These are used to
4899 // avoid UB from oversized shifts in raw IR rotate patterns, but the
4900 // intrinsics do not have that problem.
4901 // We do not allow this transform for the general funnel shift case because
4902 // that would not preserve the poison safety of the original code.
4903 auto isRotate =
4904 m_CombineOr(Ps: m_FShl(Op0: m_Value(V&: X), Op1: m_Deferred(V: X), Op2: m_Value(V&: ShAmt)),
4905 Ps: m_FShr(Op0: m_Value(V&: X), Op1: m_Deferred(V: X), Op2: m_Value(V&: ShAmt)));
4906 // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
4907 // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
4908 if (match(V: FalseVal, P: isRotate) && TrueVal == X && CmpLHS == ShAmt &&
4909 Pred == ICmpInst::ICMP_EQ)
4910 return FalseVal;
4911
4912 // X == 0 ? abs(X) : -abs(X) --> -abs(X)
4913 // X == 0 ? -abs(X) : abs(X) --> abs(X)
4914 if (match(V: TrueVal, P: m_Intrinsic<Intrinsic::abs>(Ops: m_Specific(V: CmpLHS))) &&
4915 match(V: FalseVal, P: m_Neg(V: m_Intrinsic<Intrinsic::abs>(Ops: m_Specific(V: CmpLHS)))))
4916 return FalseVal;
4917 if (match(V: TrueVal,
4918 P: m_Neg(V: m_Intrinsic<Intrinsic::abs>(Ops: m_Specific(V: CmpLHS)))) &&
4919 match(V: FalseVal, P: m_Intrinsic<Intrinsic::abs>(Ops: m_Specific(V: CmpLHS))))
4920 return FalseVal;
4921 }
4922
4923 // If we have a scalar equality comparison, then we know the value in one of
4924 // the arms of the select. See if substituting this value into the arm and
4925 // simplifying the result yields the same value as the other arm.
4926 if (Pred == ICmpInst::ICMP_EQ) {
4927 if (CmpLHS->getType()->isIntOrIntVectorTy() ||
4928 canReplacePointersIfEqual(From: CmpLHS, To: CmpRHS, DL: Q.DL))
4929 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpLHS, CmpRHS}}, TrueVal,
4930 FalseVal, Q, MaxRecurse))
4931 return V;
4932 if (CmpLHS->getType()->isIntOrIntVectorTy() ||
4933 canReplacePointersIfEqual(From: CmpRHS, To: CmpLHS, DL: Q.DL))
4934 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpRHS, CmpLHS}}, TrueVal,
4935 FalseVal, Q, MaxRecurse))
4936 return V;
4937
4938 Value *X;
4939 Value *Y;
4940 // select((X | Y) == 0 ? X : 0) --> 0 (commuted 2 ways)
4941 if (match(V: CmpLHS, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
4942 match(V: CmpRHS, P: m_Zero())) {
4943 // (X | Y) == 0 implies X == 0 and Y == 0.
4944 if (Value *V = simplifySelectWithEquivalence(
4945 Replacements: {{X, CmpRHS}, {Y, CmpRHS}}, TrueVal, FalseVal, Q, MaxRecurse))
4946 return V;
4947 }
4948
4949 // select((X & Y) == -1 ? X : -1) --> -1 (commuted 2 ways)
4950 if (match(V: CmpLHS, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
4951 match(V: CmpRHS, P: m_AllOnes())) {
4952 // (X & Y) == -1 implies X == -1 and Y == -1.
4953 if (Value *V = simplifySelectWithEquivalence(
4954 Replacements: {{X, CmpRHS}, {Y, CmpRHS}}, TrueVal, FalseVal, Q, MaxRecurse))
4955 return V;
4956 }
4957 }
4958
4959 return nullptr;
4960}
4961
4962/// Try to simplify a select instruction when its condition operand is a
4963/// floating-point comparison.
4964static Value *simplifySelectWithFCmp(Value *Cond, Value *T, Value *F,
4965 FastMathFlags FMF, const SimplifyQuery &Q,
4966 unsigned MaxRecurse) {
4967 CmpPredicate Pred;
4968 Value *CmpLHS, *CmpRHS;
4969 if (!match(V: Cond, P: m_FCmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS))))
4970 return nullptr;
4971 FCmpInst *I = cast<FCmpInst>(Val: Cond);
4972
4973 bool IsEquiv = I->isEquivalence();
4974 if (I->isEquivalence(/*Invert=*/true)) {
4975 std::swap(a&: T, b&: F);
4976 Pred = FCmpInst::getInversePredicate(pred: Pred);
4977 IsEquiv = true;
4978 }
4979
4980 // This transforms is safe if at least one operand is known to not be zero.
4981 // Otherwise, the select can change the sign of a zero operand.
4982 if (IsEquiv) {
4983 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpLHS, CmpRHS}}, TrueVal: T, FalseVal: F, Q,
4984 MaxRecurse))
4985 return V;
4986 if (Value *V = simplifySelectWithEquivalence(Replacements: {{CmpRHS, CmpLHS}}, TrueVal: T, FalseVal: F, Q,
4987 MaxRecurse))
4988 return V;
4989 }
4990
4991 // Canonicalize CmpLHS to be T, and CmpRHS to be F, if they're swapped.
4992 if (CmpLHS == F && CmpRHS == T)
4993 std::swap(a&: CmpLHS, b&: CmpRHS);
4994
4995 if (CmpLHS != T || CmpRHS != F)
4996 return nullptr;
4997
4998 // This transform is also safe if we do not have (do not care about) -0.0.
4999 if (FMF.noSignedZeros()) {
5000 // (T == F) ? T : F --> F
5001 if (Pred == FCmpInst::FCMP_OEQ)
5002 return F;
5003
5004 // (T != F) ? T : F --> T
5005 if (Pred == FCmpInst::FCMP_UNE)
5006 return T;
5007 }
5008
5009 return nullptr;
5010}
5011
5012/// Look for the following pattern and simplify %to_fold to %identicalPhi.
5013/// Here %phi, %to_fold and %phi.next perform the same functionality as
5014/// %identicalPhi and hence the select instruction %to_fold can be folded
5015/// into %identicalPhi.
5016///
5017/// BB1:
5018/// %identicalPhi = phi [ X, %BB0 ], [ %identicalPhi.next, %BB1 ]
5019/// %phi = phi [ X, %BB0 ], [ %phi.next, %BB1 ]
5020/// ...
5021/// %identicalPhi.next = select %cmp, %val, %identicalPhi
5022/// (or select %cmp, %identicalPhi, %val)
5023/// %to_fold = select %cmp2, %identicalPhi, %phi
5024/// %phi.next = select %cmp, %val, %to_fold
5025/// (or select %cmp, %to_fold, %val)
5026///
5027/// Prove that %phi and %identicalPhi are the same by induction:
5028///
5029/// Base case: Both %phi and %identicalPhi are equal on entry to the loop.
5030/// Inductive case:
5031/// Suppose %phi and %identicalPhi are equal at iteration i.
5032/// We look at their values at iteration i+1 which are %phi.next and
5033/// %identicalPhi.next. They would have become different only when %cmp is
5034/// false and the corresponding values %to_fold and %identicalPhi differ
5035/// (similar reason for the other "or" case in the bracket).
5036///
5037/// The only condition when %to_fold and %identicalPh could differ is when %cmp2
5038/// is false and %to_fold is %phi, which contradicts our inductive hypothesis
5039/// that %phi and %identicalPhi are equal. Thus %phi and %identicalPhi are
5040/// always equal at iteration i+1.
5041bool isSelectWithIdenticalPHI(PHINode &PN, PHINode &IdenticalPN) {
5042 if (PN.getParent() != IdenticalPN.getParent())
5043 return false;
5044 if (PN.getNumIncomingValues() != 2)
5045 return false;
5046
5047 // Check that only the backedge incoming value is different.
5048 unsigned DiffVals = 0;
5049 BasicBlock *DiffValBB = nullptr;
5050 for (unsigned i = 0; i < 2; i++) {
5051 BasicBlock *PredBB = PN.getIncomingBlock(i);
5052 if (PN.getIncomingValue(i) !=
5053 IdenticalPN.getIncomingValueForBlock(BB: PredBB)) {
5054 DiffVals++;
5055 DiffValBB = PredBB;
5056 }
5057 }
5058 if (DiffVals != 1)
5059 return false;
5060 // Now check that the backedge incoming values are two select
5061 // instructions with the same condition. Either their true
5062 // values are the same, or their false values are the same.
5063 auto *SI = dyn_cast<SelectInst>(Val: PN.getIncomingValueForBlock(BB: DiffValBB));
5064 auto *IdenticalSI =
5065 dyn_cast<SelectInst>(Val: IdenticalPN.getIncomingValueForBlock(BB: DiffValBB));
5066 if (!SI || !IdenticalSI)
5067 return false;
5068 if (SI->getCondition() != IdenticalSI->getCondition())
5069 return false;
5070
5071 SelectInst *SIOtherVal = nullptr;
5072 Value *IdenticalSIOtherVal = nullptr;
5073 if (SI->getTrueValue() == IdenticalSI->getTrueValue()) {
5074 SIOtherVal = dyn_cast<SelectInst>(Val: SI->getFalseValue());
5075 IdenticalSIOtherVal = IdenticalSI->getFalseValue();
5076 } else if (SI->getFalseValue() == IdenticalSI->getFalseValue()) {
5077 SIOtherVal = dyn_cast<SelectInst>(Val: SI->getTrueValue());
5078 IdenticalSIOtherVal = IdenticalSI->getTrueValue();
5079 } else {
5080 return false;
5081 }
5082
5083 // Now check that the other values in select, i.e., %to_fold and
5084 // %identicalPhi, are essentially the same value.
5085 if (!SIOtherVal || IdenticalSIOtherVal != &IdenticalPN)
5086 return false;
5087 if (!(SIOtherVal->getTrueValue() == &IdenticalPN &&
5088 SIOtherVal->getFalseValue() == &PN) &&
5089 !(SIOtherVal->getTrueValue() == &PN &&
5090 SIOtherVal->getFalseValue() == &IdenticalPN))
5091 return false;
5092 return true;
5093}
5094
5095/// Given operands for a SelectInst, see if we can fold the result.
5096/// If not, this returns null.
5097static Value *simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
5098 FastMathFlags FMF, const SimplifyQuery &Q,
5099 unsigned MaxRecurse) {
5100 if (auto *CondC = dyn_cast<Constant>(Val: Cond)) {
5101 if (auto *TrueC = dyn_cast<Constant>(Val: TrueVal))
5102 if (auto *FalseC = dyn_cast<Constant>(Val: FalseVal))
5103 if (Constant *C = ConstantFoldSelectInstruction(Cond: CondC, V1: TrueC, V2: FalseC))
5104 return C;
5105
5106 // select poison, X, Y -> poison
5107 if (isa<PoisonValue>(Val: CondC))
5108 return PoisonValue::get(T: TrueVal->getType());
5109
5110 // select undef, X, Y -> X or Y
5111 if (Q.isUndefValue(V: CondC))
5112 return isa<Constant>(Val: FalseVal) ? FalseVal : TrueVal;
5113
5114 // select true, X, Y --> X
5115 // select false, X, Y --> Y
5116 // For vectors, allow undef/poison elements in the condition to match the
5117 // defined elements, so we can eliminate the select.
5118 if (match(V: CondC, P: m_One()))
5119 return TrueVal;
5120 if (match(V: CondC, P: m_Zero()))
5121 return FalseVal;
5122 }
5123
5124 assert(Cond->getType()->isIntOrIntVectorTy(1) &&
5125 "Select must have bool or bool vector condition");
5126 assert(TrueVal->getType() == FalseVal->getType() &&
5127 "Select must have same types for true/false ops");
5128
5129 if (Cond->getType() == TrueVal->getType()) {
5130 // select i1 Cond, i1 true, i1 false --> i1 Cond
5131 if (match(V: TrueVal, P: m_One()) && match(V: FalseVal, P: m_ZeroInt()))
5132 return Cond;
5133
5134 // (X && Y) ? X : Y --> Y (commuted 2 ways)
5135 if (match(V: Cond, P: m_c_LogicalAnd(L: m_Specific(V: TrueVal), R: m_Specific(V: FalseVal))))
5136 return FalseVal;
5137
5138 // (X || Y) ? X : Y --> X (commuted 2 ways)
5139 if (match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: TrueVal), R: m_Specific(V: FalseVal))))
5140 return TrueVal;
5141
5142 // (X || Y) ? false : X --> false (commuted 2 ways)
5143 if (match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: FalseVal), R: m_Value())) &&
5144 match(V: TrueVal, P: m_ZeroInt()))
5145 return ConstantInt::getFalse(Ty: Cond->getType());
5146
5147 // Match patterns that end in logical-and.
5148 if (match(V: FalseVal, P: m_ZeroInt())) {
5149 // !(X || Y) && X --> false (commuted 2 ways)
5150 if (match(V: Cond, P: m_Not(V: m_c_LogicalOr(L: m_Specific(V: TrueVal), R: m_Value()))))
5151 return ConstantInt::getFalse(Ty: Cond->getType());
5152 // X && !(X || Y) --> false (commuted 2 ways)
5153 if (match(V: TrueVal, P: m_Not(V: m_c_LogicalOr(L: m_Specific(V: Cond), R: m_Value()))))
5154 return ConstantInt::getFalse(Ty: Cond->getType());
5155
5156 // (X || Y) && Y --> Y (commuted 2 ways)
5157 if (match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: TrueVal), R: m_Value())))
5158 return TrueVal;
5159 // Y && (X || Y) --> Y (commuted 2 ways)
5160 if (match(V: TrueVal, P: m_c_LogicalOr(L: m_Specific(V: Cond), R: m_Value())))
5161 return Cond;
5162
5163 // (X || Y) && (X || !Y) --> X (commuted 8 ways)
5164 Value *X, *Y;
5165 if (match(V: Cond, P: m_c_LogicalOr(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
5166 match(V: TrueVal, P: m_c_LogicalOr(L: m_Specific(V: X), R: m_Specific(V: Y))))
5167 return X;
5168 if (match(V: TrueVal, P: m_c_LogicalOr(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) &&
5169 match(V: Cond, P: m_c_LogicalOr(L: m_Specific(V: X), R: m_Specific(V: Y))))
5170 return X;
5171 }
5172
5173 // Match patterns that end in logical-or.
5174 if (match(V: TrueVal, P: m_One())) {
5175 // !(X && Y) || X --> true (commuted 2 ways)
5176 if (match(V: Cond, P: m_Not(V: m_c_LogicalAnd(L: m_Specific(V: FalseVal), R: m_Value()))))
5177 return ConstantInt::getTrue(Ty: Cond->getType());
5178 // X || !(X && Y) --> true (commuted 2 ways)
5179 if (match(V: FalseVal, P: m_Not(V: m_c_LogicalAnd(L: m_Specific(V: Cond), R: m_Value()))))
5180 return ConstantInt::getTrue(Ty: Cond->getType());
5181
5182 // (X && Y) || Y --> Y (commuted 2 ways)
5183 if (match(V: Cond, P: m_c_LogicalAnd(L: m_Specific(V: FalseVal), R: m_Value())))
5184 return FalseVal;
5185 // Y || (X && Y) --> Y (commuted 2 ways)
5186 if (match(V: FalseVal, P: m_c_LogicalAnd(L: m_Specific(V: Cond), R: m_Value())))
5187 return Cond;
5188 }
5189 }
5190
5191 // select ?, X, X -> X
5192 if (TrueVal == FalseVal)
5193 return TrueVal;
5194
5195 if (Cond == TrueVal) {
5196 // select i1 X, i1 X, i1 false --> X (logical-and)
5197 if (match(V: FalseVal, P: m_ZeroInt()))
5198 return Cond;
5199 // select i1 X, i1 X, i1 true --> true
5200 if (match(V: FalseVal, P: m_One()))
5201 return ConstantInt::getTrue(Ty: Cond->getType());
5202 }
5203 if (Cond == FalseVal) {
5204 // select i1 X, i1 true, i1 X --> X (logical-or)
5205 if (match(V: TrueVal, P: m_One()))
5206 return Cond;
5207 // select i1 X, i1 false, i1 X --> false
5208 if (match(V: TrueVal, P: m_ZeroInt()))
5209 return ConstantInt::getFalse(Ty: Cond->getType());
5210 }
5211
5212 // If the true or false value is poison, we can fold to the other value.
5213 // If the true or false value is undef, we can fold to the other value as
5214 // long as the other value isn't poison.
5215 // select ?, poison, X -> X
5216 // select ?, undef, X -> X
5217 if (isa<PoisonValue>(Val: TrueVal) ||
5218 (Q.isUndefValue(V: TrueVal) && impliesPoison(ValAssumedPoison: FalseVal, V: Cond)))
5219 return FalseVal;
5220 // select ?, X, poison -> X
5221 // select ?, X, undef -> X
5222 if (isa<PoisonValue>(Val: FalseVal) ||
5223 (Q.isUndefValue(V: FalseVal) && impliesPoison(ValAssumedPoison: TrueVal, V: Cond)))
5224 return TrueVal;
5225
5226 // Deal with partial undef vector constants: select ?, VecC, VecC' --> VecC''
5227 Constant *TrueC, *FalseC;
5228 if (isa<FixedVectorType>(Val: TrueVal->getType()) &&
5229 match(V: TrueVal, P: m_Constant(C&: TrueC)) &&
5230 match(V: FalseVal, P: m_Constant(C&: FalseC))) {
5231 unsigned NumElts =
5232 cast<FixedVectorType>(Val: TrueC->getType())->getNumElements();
5233 SmallVector<Constant *, 16> NewC;
5234 for (unsigned i = 0; i != NumElts; ++i) {
5235 // Bail out on incomplete vector constants.
5236 Constant *TEltC = TrueC->getAggregateElement(Elt: i);
5237 Constant *FEltC = FalseC->getAggregateElement(Elt: i);
5238 if (!TEltC || !FEltC)
5239 break;
5240
5241 // If the elements match (undef or not), that value is the result. If only
5242 // one element is undef, choose the defined element as the safe result.
5243 if (TEltC == FEltC)
5244 NewC.push_back(Elt: TEltC);
5245 else if (isa<PoisonValue>(Val: TEltC) ||
5246 (Q.isUndefValue(V: TEltC) && isGuaranteedNotToBePoison(V: FEltC)))
5247 NewC.push_back(Elt: FEltC);
5248 else if (isa<PoisonValue>(Val: FEltC) ||
5249 (Q.isUndefValue(V: FEltC) && isGuaranteedNotToBePoison(V: TEltC)))
5250 NewC.push_back(Elt: TEltC);
5251 else
5252 break;
5253 }
5254 if (NewC.size() == NumElts)
5255 return ConstantVector::get(V: NewC);
5256 }
5257
5258 if (Value *V =
5259 simplifySelectWithICmpCond(CondVal: Cond, TrueVal, FalseVal, Q, MaxRecurse))
5260 return V;
5261
5262 if (Value *V = simplifySelectWithBitTest(CondVal: Cond, TrueVal, FalseVal))
5263 return V;
5264
5265 if (Value *V =
5266 simplifySelectWithFCmp(Cond, T: TrueVal, F: FalseVal, FMF, Q, MaxRecurse))
5267 return V;
5268
5269 std::optional<bool> Imp = isImpliedByDomCondition(Cond, ContextI: Q.CxtI, DL: Q.DL);
5270 if (Imp)
5271 return *Imp ? TrueVal : FalseVal;
5272 // Look for same PHIs in the true and false values.
5273 if (auto *TruePHI = dyn_cast<PHINode>(Val: TrueVal))
5274 if (auto *FalsePHI = dyn_cast<PHINode>(Val: FalseVal)) {
5275 if (isSelectWithIdenticalPHI(PN&: *TruePHI, IdenticalPN&: *FalsePHI))
5276 return FalseVal;
5277 if (isSelectWithIdenticalPHI(PN&: *FalsePHI, IdenticalPN&: *TruePHI))
5278 return TrueVal;
5279 }
5280 return nullptr;
5281}
5282
5283Value *llvm::simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
5284 FastMathFlags FMF, const SimplifyQuery &Q) {
5285 return ::simplifySelectInst(Cond, TrueVal, FalseVal, FMF, Q, MaxRecurse: RecursionLimit);
5286}
5287
5288/// Given operands for an GetElementPtrInst, see if we can fold the result.
5289/// If not, this returns null.
5290static Value *simplifyGEPInst(Type *SrcTy, Value *Ptr,
5291 ArrayRef<Value *> Indices, GEPNoWrapFlags NW,
5292 const SimplifyQuery &Q, unsigned) {
5293 // The type of the GEP pointer operand.
5294 unsigned AS =
5295 cast<PointerType>(Val: Ptr->getType()->getScalarType())->getAddressSpace();
5296
5297 // getelementptr P -> P.
5298 if (Indices.empty())
5299 return Ptr;
5300
5301 // Compute the (pointer) type returned by the GEP instruction.
5302 Type *LastType = GetElementPtrInst::getIndexedType(Ty: SrcTy, IdxList: Indices);
5303 Type *GEPTy = Ptr->getType();
5304 if (!GEPTy->isVectorTy()) {
5305 for (Value *Op : Indices) {
5306 // If one of the operands is a vector, the result type is a vector of
5307 // pointers. All vector operands must have the same number of elements.
5308 if (VectorType *VT = dyn_cast<VectorType>(Val: Op->getType())) {
5309 GEPTy = VectorType::get(ElementType: GEPTy, EC: VT->getElementCount());
5310 break;
5311 }
5312 }
5313 }
5314
5315 // All-zero GEP is a no-op, unless it performs a vector splat.
5316 if (Ptr->getType() == GEPTy && all_of(Range&: Indices, P: match_fn(P: m_Zero())))
5317 return Ptr;
5318
5319 // getelementptr poison, idx -> poison
5320 // getelementptr baseptr, poison -> poison
5321 if (isa<PoisonValue>(Val: Ptr) || any_of(Range&: Indices, P: IsaPred<PoisonValue>))
5322 return PoisonValue::get(T: GEPTy);
5323
5324 // getelementptr undef, idx -> undef
5325 if (Q.isUndefValue(V: Ptr))
5326 return UndefValue::get(T: GEPTy);
5327
5328 bool IsScalableVec =
5329 SrcTy->isScalableTy() || any_of(Range&: Indices, P: [](const Value *V) {
5330 return isa<ScalableVectorType>(Val: V->getType());
5331 });
5332
5333 if (Indices.size() == 1) {
5334 Type *Ty = SrcTy;
5335 if (!IsScalableVec && Ty->isSized()) {
5336 Value *P;
5337 uint64_t C;
5338 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
5339 // getelementptr P, N -> P if P points to a type of zero size.
5340 if (TyAllocSize == 0 && Ptr->getType() == GEPTy)
5341 return Ptr;
5342
5343 // The following transforms are only safe if the ptrtoint cast
5344 // doesn't truncate the address of the pointers. The non-address bits
5345 // must be the same, as the underlying objects are the same.
5346 if (Indices[0]->getType()->getScalarSizeInBits() >=
5347 Q.DL.getAddressSizeInBits(AS)) {
5348 auto CanSimplify = [GEPTy, &P, Ptr]() -> bool {
5349 return P->getType() == GEPTy &&
5350 getUnderlyingObject(V: P) == getUnderlyingObject(V: Ptr);
5351 };
5352 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
5353 if (TyAllocSize == 1 &&
5354 match(V: Indices[0], P: m_Sub(L: m_PtrToIntOrAddr(Op: m_Value(V&: P)),
5355 R: m_PtrToIntOrAddr(Op: m_Specific(V: Ptr)))) &&
5356 CanSimplify())
5357 return P;
5358
5359 // getelementptr V, (ashr (sub P, V), C) -> P if P points to a type of
5360 // size 1 << C.
5361 if (match(V: Indices[0], P: m_AShr(L: m_Sub(L: m_PtrToIntOrAddr(Op: m_Value(V&: P)),
5362 R: m_PtrToIntOrAddr(Op: m_Specific(V: Ptr))),
5363 R: m_ConstantInt(V&: C))) &&
5364 TyAllocSize == 1ULL << C && CanSimplify())
5365 return P;
5366
5367 // getelementptr V, (sdiv (sub P, V), C) -> P if P points to a type of
5368 // size C.
5369 if (match(V: Indices[0], P: m_SDiv(L: m_Sub(L: m_PtrToIntOrAddr(Op: m_Value(V&: P)),
5370 R: m_PtrToIntOrAddr(Op: m_Specific(V: Ptr))),
5371 R: m_SpecificInt(V: TyAllocSize))) &&
5372 CanSimplify())
5373 return P;
5374 }
5375 }
5376 }
5377
5378 if (!IsScalableVec && Q.DL.getTypeAllocSize(Ty: LastType) == 1 &&
5379 all_of(Range: Indices.drop_back(N: 1), P: match_fn(P: m_Zero()))) {
5380 unsigned IdxWidth =
5381 Q.DL.getIndexSizeInBits(AS: Ptr->getType()->getPointerAddressSpace());
5382 if (Q.DL.getTypeSizeInBits(Ty: Indices.back()->getType()) == IdxWidth) {
5383 APInt BasePtrOffset(IdxWidth, 0);
5384 Value *StrippedBasePtr =
5385 Ptr->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: BasePtrOffset);
5386
5387 // Avoid creating inttoptr of zero here: While LLVMs treatment of
5388 // inttoptr is generally conservative, this particular case is folded to
5389 // a null pointer, which will have incorrect provenance.
5390
5391 // gep (gep V, C), (sub 0, V) -> C
5392 if (match(V: Indices.back(),
5393 P: m_Neg(V: m_PtrToInt(Op: m_Specific(V: StrippedBasePtr)))) &&
5394 !BasePtrOffset.isZero()) {
5395 auto *CI = ConstantInt::get(Context&: GEPTy->getContext(), V: BasePtrOffset);
5396 return ConstantExpr::getIntToPtr(C: CI, Ty: GEPTy);
5397 }
5398 // gep (gep V, C), (xor V, -1) -> C-1
5399 if (match(V: Indices.back(),
5400 P: m_Xor(L: m_PtrToInt(Op: m_Specific(V: StrippedBasePtr)), R: m_AllOnes())) &&
5401 !BasePtrOffset.isOne()) {
5402 auto *CI = ConstantInt::get(Context&: GEPTy->getContext(), V: BasePtrOffset - 1);
5403 return ConstantExpr::getIntToPtr(C: CI, Ty: GEPTy);
5404 }
5405 }
5406 }
5407
5408 // Check to see if this is constant foldable.
5409 if (!isa<Constant>(Val: Ptr) || !all_of(Range&: Indices, P: IsaPred<Constant>))
5410 return nullptr;
5411
5412 if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy: SrcTy))
5413 return ConstantFoldGetElementPtr(Ty: SrcTy, C: cast<Constant>(Val: Ptr), InRange: std::nullopt,
5414 Idxs: Indices);
5415
5416 auto *CE =
5417 ConstantExpr::getGetElementPtr(Ty: SrcTy, C: cast<Constant>(Val: Ptr), IdxList: Indices, NW);
5418 return ConstantFoldConstant(C: CE, DL: Q.DL);
5419}
5420
5421Value *llvm::simplifyGEPInst(Type *SrcTy, Value *Ptr, ArrayRef<Value *> Indices,
5422 GEPNoWrapFlags NW, const SimplifyQuery &Q) {
5423 return ::simplifyGEPInst(SrcTy, Ptr, Indices, NW, Q, RecursionLimit);
5424}
5425
5426/// Given operands for an InsertValueInst, see if we can fold the result.
5427/// If not, this returns null.
5428static Value *simplifyInsertValueInst(Value *Agg, Value *Val,
5429 ArrayRef<unsigned> Idxs,
5430 const SimplifyQuery &Q, unsigned) {
5431 if (Constant *CAgg = dyn_cast<Constant>(Val: Agg))
5432 if (Constant *CVal = dyn_cast<Constant>(Val))
5433 return ConstantFoldInsertValueInstruction(Agg: CAgg, Val: CVal, Idxs);
5434
5435 // insertvalue x, poison, n -> x
5436 // insertvalue x, undef, n -> x if x cannot be poison
5437 if (isa<PoisonValue>(Val) ||
5438 (Q.isUndefValue(V: Val) && isGuaranteedNotToBePoison(V: Agg)))
5439 return Agg;
5440
5441 // insertvalue x, (extractvalue y, n), n
5442 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
5443 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
5444 EV->getIndices() == Idxs) {
5445 // insertvalue poison, (extractvalue y, n), n -> y
5446 // insertvalue undef, (extractvalue y, n), n -> y if y cannot be poison
5447 if (isa<PoisonValue>(Val: Agg) ||
5448 (Q.isUndefValue(V: Agg) &&
5449 isGuaranteedNotToBePoison(V: EV->getAggregateOperand())))
5450 return EV->getAggregateOperand();
5451
5452 // insertvalue y, (extractvalue y, n), n -> y
5453 if (Agg == EV->getAggregateOperand())
5454 return Agg;
5455 }
5456
5457 return nullptr;
5458}
5459
5460Value *llvm::simplifyInsertValueInst(Value *Agg, Value *Val,
5461 ArrayRef<unsigned> Idxs,
5462 const SimplifyQuery &Q) {
5463 return ::simplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
5464}
5465
5466Value *llvm::simplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
5467 const SimplifyQuery &Q) {
5468 // Try to constant fold.
5469 auto *VecC = dyn_cast<Constant>(Val: Vec);
5470 auto *ValC = dyn_cast<Constant>(Val);
5471 auto *IdxC = dyn_cast<Constant>(Val: Idx);
5472 if (VecC && ValC && IdxC)
5473 return ConstantExpr::getInsertElement(Vec: VecC, Elt: ValC, Idx: IdxC);
5474
5475 // For fixed-length vector, fold into poison if index is out of bounds.
5476 if (auto *CI = dyn_cast<ConstantInt>(Val: Idx)) {
5477 if (isa<FixedVectorType>(Val: Vec->getType()) &&
5478 CI->uge(Num: cast<FixedVectorType>(Val: Vec->getType())->getNumElements()))
5479 return PoisonValue::get(T: Vec->getType());
5480 }
5481
5482 // If index is undef, it might be out of bounds (see above case)
5483 if (Q.isUndefValue(V: Idx))
5484 return PoisonValue::get(T: Vec->getType());
5485
5486 // If the scalar is poison, or it is undef and there is no risk of
5487 // propagating poison from the vector value, simplify to the vector value.
5488 if (isa<PoisonValue>(Val) ||
5489 (Q.isUndefValue(V: Val) && isGuaranteedNotToBePoison(V: Vec)))
5490 return Vec;
5491
5492 // Inserting the splatted value into a constant splat does nothing.
5493 if (VecC && ValC && VecC->getSplatValue() == ValC)
5494 return Vec;
5495
5496 // If we are extracting a value from a vector, then inserting it into the same
5497 // place, that's the input vector:
5498 // insertelt Vec, (extractelt Vec, Idx), Idx --> Vec
5499 if (match(V: Val, P: m_ExtractElt(Val: m_Specific(V: Vec), Idx: m_Specific(V: Idx))))
5500 return Vec;
5501
5502 return nullptr;
5503}
5504
5505/// Given operands for an ExtractValueInst, see if we can fold the result.
5506/// If not, this returns null.
5507static Value *simplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
5508 const SimplifyQuery &, unsigned) {
5509 if (auto *CAgg = dyn_cast<Constant>(Val: Agg))
5510 return ConstantFoldExtractValueInstruction(Agg: CAgg, Idxs);
5511
5512 // extractvalue x, (insertvalue y, elt, n), n -> elt
5513 unsigned NumIdxs = Idxs.size();
5514 SmallPtrSet<InsertValueInst *, 8> VisitedSet;
5515 for (auto *IVI = dyn_cast<InsertValueInst>(Val: Agg); IVI != nullptr;
5516 IVI = dyn_cast<InsertValueInst>(Val: IVI->getAggregateOperand())) {
5517 // Protect against insertvalue cycles in unreachable code.
5518 if (!VisitedSet.insert(Ptr: IVI).second)
5519 break;
5520
5521 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
5522 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
5523 unsigned NumCommonIdxs = std::min(a: NumInsertValueIdxs, b: NumIdxs);
5524 if (InsertValueIdxs.slice(N: 0, M: NumCommonIdxs) ==
5525 Idxs.slice(N: 0, M: NumCommonIdxs)) {
5526 if (NumIdxs == NumInsertValueIdxs)
5527 return IVI->getInsertedValueOperand();
5528 break;
5529 }
5530 }
5531
5532 // Simplify umul_with_overflow where one operand is 1.
5533 Value *V;
5534 if (Idxs.size() == 1 &&
5535 (match(V: Agg,
5536 P: m_Intrinsic<Intrinsic::umul_with_overflow>(Ops: m_Value(V), Ops: m_One())) ||
5537 match(V: Agg, P: m_Intrinsic<Intrinsic::umul_with_overflow>(Ops: m_One(),
5538 Ops: m_Value(V))))) {
5539 if (Idxs[0] == 0)
5540 return V;
5541 assert(Idxs[0] == 1 && "invalid index");
5542 return getFalse(Ty: CmpInst::makeCmpResultType(opnd_type: V->getType()));
5543 }
5544
5545 return nullptr;
5546}
5547
5548Value *llvm::simplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
5549 const SimplifyQuery &Q) {
5550 return ::simplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
5551}
5552
5553/// Given operands for an ExtractElementInst, see if we can fold the result.
5554/// If not, this returns null.
5555static Value *simplifyExtractElementInst(Value *Vec, Value *Idx,
5556 const SimplifyQuery &Q, unsigned) {
5557 auto *VecVTy = cast<VectorType>(Val: Vec->getType());
5558 if (auto *CVec = dyn_cast<Constant>(Val: Vec)) {
5559 if (auto *CIdx = dyn_cast<Constant>(Val: Idx))
5560 return ConstantExpr::getExtractElement(Vec: CVec, Idx: CIdx);
5561
5562 if (Q.isUndefValue(V: Vec))
5563 return UndefValue::get(T: VecVTy->getElementType());
5564 }
5565
5566 // An undef extract index can be arbitrarily chosen to be an out-of-range
5567 // index value, which would result in the instruction being poison.
5568 if (Q.isUndefValue(V: Idx))
5569 return PoisonValue::get(T: VecVTy->getElementType());
5570
5571 // If extracting a specified index from the vector, see if we can recursively
5572 // find a previously computed scalar that was inserted into the vector.
5573 if (auto *IdxC = dyn_cast<ConstantInt>(Val: Idx)) {
5574 // For fixed-length vector, fold into undef if index is out of bounds.
5575 unsigned MinNumElts = VecVTy->getElementCount().getKnownMinValue();
5576 if (isa<FixedVectorType>(Val: VecVTy) && IdxC->getValue().uge(RHS: MinNumElts))
5577 return PoisonValue::get(T: VecVTy->getElementType());
5578 // Handle case where an element is extracted from a splat.
5579 if (IdxC->getValue().ult(RHS: MinNumElts))
5580 if (auto *Splat = getSplatValue(V: Vec))
5581 return Splat;
5582 if (Value *Elt = findScalarElement(V: Vec, EltNo: IdxC->getZExtValue()))
5583 return Elt;
5584 } else {
5585 // extractelt x, (insertelt y, elt, n), n -> elt
5586 // If the possibly-variable indices are trivially known to be equal
5587 // (because they are the same operand) then use the value that was
5588 // inserted directly.
5589 auto *IE = dyn_cast<InsertElementInst>(Val: Vec);
5590 if (IE && IE->getOperand(i_nocapture: 2) == Idx)
5591 return IE->getOperand(i_nocapture: 1);
5592
5593 // The index is not relevant if our vector is a splat.
5594 if (Value *Splat = getSplatValue(V: Vec))
5595 return Splat;
5596 }
5597 return nullptr;
5598}
5599
5600Value *llvm::simplifyExtractElementInst(Value *Vec, Value *Idx,
5601 const SimplifyQuery &Q) {
5602 return ::simplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
5603}
5604
5605/// See if we can fold the given phi. If not, returns null.
5606static Value *simplifyPHINode(PHINode *PN, ArrayRef<Value *> IncomingValues,
5607 const SimplifyQuery &Q) {
5608 // WARNING: no matter how worthwhile it may seem, we can not perform PHI CSE
5609 // here, because the PHI we may succeed simplifying to was not
5610 // def-reachable from the original PHI!
5611
5612 // If all of the PHI's incoming values are the same then replace the PHI node
5613 // with the common value.
5614 Value *CommonValue = nullptr;
5615 bool HasPoisonInput = false;
5616 bool HasUndefInput = false;
5617 for (Value *Incoming : IncomingValues) {
5618 // If the incoming value is the phi node itself, it can safely be skipped.
5619 if (Incoming == PN)
5620 continue;
5621 if (isa<PoisonValue>(Val: Incoming)) {
5622 HasPoisonInput = true;
5623 continue;
5624 }
5625 if (Q.isUndefValue(V: Incoming)) {
5626 // Remember that we saw an undef value, but otherwise ignore them.
5627 HasUndefInput = true;
5628 continue;
5629 }
5630 if (CommonValue && Incoming != CommonValue)
5631 return nullptr; // Not the same, bail out.
5632 CommonValue = Incoming;
5633 }
5634
5635 // If CommonValue is null then all of the incoming values were either undef,
5636 // poison or equal to the phi node itself.
5637 if (!CommonValue)
5638 return HasUndefInput ? UndefValue::get(T: PN->getType())
5639 : PoisonValue::get(T: PN->getType());
5640
5641 if (HasPoisonInput || HasUndefInput) {
5642 // If we have a PHI node like phi(X, undef, X), where X is defined by some
5643 // instruction, we cannot return X as the result of the PHI node unless it
5644 // dominates the PHI block.
5645 if (!valueDominatesPHI(V: CommonValue, P: PN, DT: Q.DT))
5646 return nullptr;
5647
5648 // Make sure we do not replace an undef value with poison.
5649 if (HasUndefInput &&
5650 !isGuaranteedNotToBePoison(V: CommonValue, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT))
5651 return nullptr;
5652 return CommonValue;
5653 }
5654
5655 return CommonValue;
5656}
5657
5658static Value *simplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
5659 const SimplifyQuery &Q, unsigned MaxRecurse) {
5660 if (auto *C = dyn_cast<Constant>(Val: Op))
5661 return ConstantFoldCastOperand(Opcode: CastOpc, C, DestTy: Ty, DL: Q.DL);
5662
5663 if (auto *CI = dyn_cast<CastInst>(Val: Op)) {
5664 auto *Src = CI->getOperand(i_nocapture: 0);
5665 Type *SrcTy = Src->getType();
5666 Type *MidTy = CI->getType();
5667 Type *DstTy = Ty;
5668 if (Src->getType() == Ty) {
5669 auto FirstOp = CI->getOpcode();
5670 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
5671 if (CastInst::isEliminableCastPair(firstOpcode: FirstOp, secondOpcode: SecondOp, SrcTy, MidTy, DstTy,
5672 DL: &Q.DL) == Instruction::BitCast)
5673 return Src;
5674 }
5675 }
5676
5677 // bitcast x -> x
5678 if (CastOpc == Instruction::BitCast)
5679 if (Op->getType() == Ty)
5680 return Op;
5681
5682 // ptrtoint (ptradd (Ptr, X - ptrtoint(Ptr))) -> X
5683 Value *Ptr, *X;
5684 if ((CastOpc == Instruction::PtrToInt || CastOpc == Instruction::PtrToAddr) &&
5685 match(V: Op,
5686 P: m_PtrAdd(PointerOp: m_Value(V&: Ptr),
5687 OffsetOp: m_Sub(L: m_Value(V&: X), R: m_PtrToIntOrAddr(Op: m_Deferred(V: Ptr))))) &&
5688 X->getType() == Ty && Ty == Q.DL.getIndexType(PtrTy: Ptr->getType()))
5689 return X;
5690
5691 // Fold a value-preserving zext/sext of a trunc back to the original value.
5692 if (CastOpc == Instruction::ZExt || CastOpc == Instruction::SExt) {
5693 if (auto *Trunc = dyn_cast<TruncInst>(Val: Op)) {
5694 Value *Src = Trunc->getOperand(i_nocapture: 0);
5695 bool NoWrap = CastOpc == Instruction::ZExt ? Trunc->hasNoUnsignedWrap()
5696 : Trunc->hasNoSignedWrap();
5697 if (Src->getType() == Ty && NoWrap)
5698 return Src;
5699 }
5700 }
5701
5702 return nullptr;
5703}
5704
5705Value *llvm::simplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
5706 const SimplifyQuery &Q) {
5707 return ::simplifyCastInst(CastOpc, Op, Ty, Q, MaxRecurse: RecursionLimit);
5708}
5709
5710/// For the given destination element of a shuffle, peek through shuffles to
5711/// match a root vector source operand that contains that element in the same
5712/// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
5713static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
5714 int MaskVal, Value *RootVec,
5715 unsigned MaxRecurse) {
5716 if (!MaxRecurse--)
5717 return nullptr;
5718
5719 // Bail out if any mask value is undefined. That kind of shuffle may be
5720 // simplified further based on demanded bits or other folds.
5721 if (MaskVal == -1)
5722 return nullptr;
5723
5724 // The mask value chooses which source operand we need to look at next.
5725 int InVecNumElts = cast<FixedVectorType>(Val: Op0->getType())->getNumElements();
5726 int RootElt = MaskVal;
5727 Value *SourceOp = Op0;
5728 if (MaskVal >= InVecNumElts) {
5729 RootElt = MaskVal - InVecNumElts;
5730 SourceOp = Op1;
5731 }
5732
5733 // If the source operand is a shuffle itself, look through it to find the
5734 // matching root vector.
5735 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(Val: SourceOp)) {
5736 return foldIdentityShuffles(
5737 DestElt, Op0: SourceShuf->getOperand(i_nocapture: 0), Op1: SourceShuf->getOperand(i_nocapture: 1),
5738 MaskVal: SourceShuf->getMaskValue(Elt: RootElt), RootVec, MaxRecurse);
5739 }
5740
5741 // The source operand is not a shuffle. Initialize the root vector value for
5742 // this shuffle if that has not been done yet.
5743 if (!RootVec)
5744 RootVec = SourceOp;
5745
5746 // Give up as soon as a source operand does not match the existing root value.
5747 if (RootVec != SourceOp)
5748 return nullptr;
5749
5750 // The element must be coming from the same lane in the source vector
5751 // (although it may have crossed lanes in intermediate shuffles).
5752 if (RootElt != DestElt)
5753 return nullptr;
5754
5755 return RootVec;
5756}
5757
5758static Value *simplifyShuffleVectorInst(Value *Op0, Value *Op1,
5759 ArrayRef<int> Mask, Type *RetTy,
5760 const SimplifyQuery &Q,
5761 unsigned MaxRecurse) {
5762 if (all_of(Range&: Mask, P: equal_to(Arg: PoisonMaskElem)))
5763 return PoisonValue::get(T: RetTy);
5764
5765 auto *InVecTy = cast<VectorType>(Val: Op0->getType());
5766 unsigned MaskNumElts = Mask.size();
5767 ElementCount InVecEltCount = InVecTy->getElementCount();
5768
5769 bool Scalable = InVecEltCount.isScalable();
5770
5771 SmallVector<int, 32> Indices;
5772 Indices.assign(in_start: Mask.begin(), in_end: Mask.end());
5773
5774 // Canonicalization: If mask does not select elements from an input vector,
5775 // replace that input vector with poison.
5776 if (!Scalable) {
5777 bool MaskSelects0 = false, MaskSelects1 = false;
5778 unsigned InVecNumElts = InVecEltCount.getKnownMinValue();
5779 for (unsigned i = 0; i != MaskNumElts; ++i) {
5780 if (Indices[i] == -1)
5781 continue;
5782 if ((unsigned)Indices[i] < InVecNumElts)
5783 MaskSelects0 = true;
5784 else
5785 MaskSelects1 = true;
5786 }
5787 if (!MaskSelects0)
5788 Op0 = PoisonValue::get(T: InVecTy);
5789 if (!MaskSelects1)
5790 Op1 = PoisonValue::get(T: InVecTy);
5791 }
5792
5793 auto *Op0Const = dyn_cast<Constant>(Val: Op0);
5794 auto *Op1Const = dyn_cast<Constant>(Val: Op1);
5795
5796 // If all operands are constant, constant fold the shuffle. This
5797 // transformation depends on the value of the mask which is not known at
5798 // compile time for scalable vectors
5799 if (Op0Const && Op1Const)
5800 return ConstantExpr::getShuffleVector(V1: Op0Const, V2: Op1Const, Mask);
5801
5802 // Canonicalization: if only one input vector is constant, it shall be the
5803 // second one. This transformation depends on the value of the mask which
5804 // is not known at compile time for scalable vectors
5805 if (!Scalable && Op0Const && !Op1Const) {
5806 std::swap(a&: Op0, b&: Op1);
5807 ShuffleVectorInst::commuteShuffleMask(Mask: Indices,
5808 InVecNumElts: InVecEltCount.getKnownMinValue());
5809 }
5810
5811 // A splat of an inserted scalar constant becomes a vector constant:
5812 // shuf (inselt ?, C, IndexC), undef, <IndexC, IndexC...> --> <C, C...>
5813 // NOTE: We may have commuted above, so analyze the updated Indices, not the
5814 // original mask constant.
5815 // NOTE: This transformation depends on the value of the mask which is not
5816 // known at compile time for scalable vectors
5817 Constant *C;
5818 ConstantInt *IndexC;
5819 if (!Scalable && match(V: Op0, P: m_InsertElt(Val: m_Value(), Elt: m_Constant(C),
5820 Idx: m_ConstantInt(CI&: IndexC)))) {
5821 // Match a splat shuffle mask of the insert index allowing undef elements.
5822 int InsertIndex = IndexC->getZExtValue();
5823 if (all_of(Range&: Indices, P: [InsertIndex](int MaskElt) {
5824 return MaskElt == InsertIndex || MaskElt == -1;
5825 })) {
5826 assert(isa<UndefValue>(Op1) && "Expected undef operand 1 for splat");
5827
5828 // Shuffle mask poisons become poison constant result elements.
5829 SmallVector<Constant *, 16> VecC(MaskNumElts, C);
5830 for (unsigned i = 0; i != MaskNumElts; ++i)
5831 if (Indices[i] == -1)
5832 VecC[i] = PoisonValue::get(T: C->getType());
5833 return ConstantVector::get(V: VecC);
5834 }
5835 }
5836
5837 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
5838 // value type is same as the input vectors' type.
5839 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Val: Op0))
5840 if (Q.isUndefValue(V: Op1) && RetTy == InVecTy &&
5841 all_equal(Range: OpShuf->getShuffleMask()))
5842 return Op0;
5843
5844 // All remaining transformation depend on the value of the mask, which is
5845 // not known at compile time for scalable vectors.
5846 if (Scalable)
5847 return nullptr;
5848
5849 // Don't fold a shuffle with undef mask elements. This may get folded in a
5850 // better way using demanded bits or other analysis.
5851 // TODO: Should we allow this?
5852 if (is_contained(Range&: Indices, Element: -1))
5853 return nullptr;
5854
5855 // Check if every element of this shuffle can be mapped back to the
5856 // corresponding element of a single root vector. If so, we don't need this
5857 // shuffle. This handles simple identity shuffles as well as chains of
5858 // shuffles that may widen/narrow and/or move elements across lanes and back.
5859 Value *RootVec = nullptr;
5860 for (unsigned i = 0; i != MaskNumElts; ++i) {
5861 // Note that recursion is limited for each vector element, so if any element
5862 // exceeds the limit, this will fail to simplify.
5863 RootVec =
5864 foldIdentityShuffles(DestElt: i, Op0, Op1, MaskVal: Indices[i], RootVec, MaxRecurse);
5865
5866 // We can't replace a widening/narrowing shuffle with one of its operands.
5867 if (!RootVec || RootVec->getType() != RetTy)
5868 return nullptr;
5869 }
5870 return RootVec;
5871}
5872
5873/// Given operands for a ShuffleVectorInst, fold the result or return null.
5874Value *llvm::simplifyShuffleVectorInst(Value *Op0, Value *Op1,
5875 ArrayRef<int> Mask, Type *RetTy,
5876 const SimplifyQuery &Q) {
5877 return ::simplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, MaxRecurse: RecursionLimit);
5878}
5879
5880static Constant *foldConstant(Instruction::UnaryOps Opcode, Value *&Op,
5881 const SimplifyQuery &Q) {
5882 if (auto *C = dyn_cast<Constant>(Val: Op))
5883 return ConstantFoldUnaryOpOperand(Opcode, Op: C, DL: Q.DL);
5884 return nullptr;
5885}
5886
5887/// Given the operand for an FNeg, see if we can fold the result. If not, this
5888/// returns null.
5889static Value *simplifyFNegInst(Value *Op, FastMathFlags FMF,
5890 const SimplifyQuery &Q, unsigned MaxRecurse) {
5891 if (Constant *C = foldConstant(Opcode: Instruction::FNeg, Op, Q))
5892 return C;
5893
5894 Value *X;
5895 // fneg (fneg X) ==> X
5896 if (match(V: Op, P: m_FNeg(X: m_Value(V&: X))))
5897 return X;
5898
5899 return nullptr;
5900}
5901
5902Value *llvm::simplifyFNegInst(Value *Op, FastMathFlags FMF,
5903 const SimplifyQuery &Q) {
5904 return ::simplifyFNegInst(Op, FMF, Q, MaxRecurse: RecursionLimit);
5905}
5906
5907/// Try to propagate existing NaN values when possible. If not, replace the
5908/// constant or elements in the constant with a canonical NaN.
5909static Constant *propagateNaN(Constant *In) {
5910 Type *Ty = In->getType();
5911 if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Ty)) {
5912 unsigned NumElts = VecTy->getNumElements();
5913 SmallVector<Constant *, 32> NewC(NumElts);
5914 for (unsigned i = 0; i != NumElts; ++i) {
5915 Constant *EltC = In->getAggregateElement(Elt: i);
5916 // Poison elements propagate. NaN propagates except signaling is quieted.
5917 // Replace unknown or undef elements with canonical NaN.
5918 if (EltC && isa<PoisonValue>(Val: EltC))
5919 NewC[i] = EltC;
5920 else if (EltC && EltC->isNaN())
5921 NewC[i] = ConstantFP::get(
5922 Ty: EltC->getType(), V: cast<ConstantFP>(Val: EltC)->getValue().makeQuiet());
5923 else
5924 NewC[i] = ConstantFP::getNaN(Ty: VecTy->getElementType());
5925 }
5926 return ConstantVector::get(V: NewC);
5927 }
5928
5929 // If it is not a fixed vector, but not a simple NaN either, return a
5930 // canonical NaN.
5931 if (!In->isNaN())
5932 return ConstantFP::getNaN(Ty);
5933
5934 // If we known this is a NaN, and it's scalable vector, we must have a splat
5935 // on our hands. Grab that before splatting a QNaN constant.
5936 if (isa<ScalableVectorType>(Val: Ty)) {
5937 auto *Splat = In->getSplatValue();
5938 assert(Splat && Splat->isNaN() &&
5939 "Found a scalable-vector NaN but not a splat");
5940 In = Splat;
5941 }
5942
5943 // Propagate an existing QNaN constant. If it is an SNaN, make it quiet, but
5944 // preserve the sign/payload.
5945 return ConstantFP::get(Ty, V: cast<ConstantFP>(Val: In)->getValue().makeQuiet());
5946}
5947
5948/// Perform folds that are common to any floating-point operation. This implies
5949/// transforms based on poison/undef/NaN because the operation itself makes no
5950/// difference to the result.
5951static Constant *simplifyFPOp(ArrayRef<Value *> Ops, FastMathFlags FMF,
5952 const SimplifyQuery &Q,
5953 fp::ExceptionBehavior ExBehavior,
5954 RoundingMode Rounding) {
5955 // Poison is independent of anything else. It always propagates from an
5956 // operand to a math result.
5957 if (any_of(Range&: Ops, P: IsaPred<PoisonValue>))
5958 return PoisonValue::get(T: Ops[0]->getType());
5959
5960 for (Value *V : Ops) {
5961 bool IsNan = match(V, P: m_NaN());
5962 bool IsInf = match(V, P: m_Inf());
5963 bool IsUndef = Q.isUndefValue(V);
5964
5965 // If this operation has 'nnan' or 'ninf' and at least 1 disallowed operand
5966 // (an undef operand can be chosen to be Nan/Inf), then the result of
5967 // this operation is poison.
5968 if (FMF.noNaNs() && (IsNan || IsUndef))
5969 return PoisonValue::get(T: V->getType());
5970 if (FMF.noInfs() && (IsInf || IsUndef))
5971 return PoisonValue::get(T: V->getType());
5972
5973 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding)) {
5974 // Undef does not propagate because undef means that all bits can take on
5975 // any value. If this is undef * NaN for example, then the result values
5976 // (at least the exponent bits) are limited. Assume the undef is a
5977 // canonical NaN and propagate that.
5978 if (IsUndef)
5979 return ConstantFP::getNaN(Ty: V->getType());
5980 if (IsNan)
5981 return propagateNaN(In: cast<Constant>(Val: V));
5982 } else if (ExBehavior != fp::ebStrict) {
5983 if (IsNan)
5984 return propagateNaN(In: cast<Constant>(Val: V));
5985 }
5986 }
5987 return nullptr;
5988}
5989
5990/// Given operands for an FAdd, see if we can fold the result. If not, this
5991/// returns null.
5992static Value *
5993simplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
5994 const SimplifyQuery &Q, unsigned MaxRecurse,
5995 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
5996 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
5997 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
5998 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FAdd, Op0, Op1, Q))
5999 return C;
6000
6001 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6002 return C;
6003
6004 // fadd X, -0 ==> X
6005 // With strict/constrained FP, we have these possible edge cases that do
6006 // not simplify to Op0:
6007 // fadd SNaN, -0.0 --> QNaN
6008 // fadd +0.0, -0.0 --> -0.0 (but only with round toward negative)
6009 if (canIgnoreSNaN(EB: ExBehavior, FMF) &&
6010 (!canRoundingModeBe(RM: Rounding, QRM: RoundingMode::TowardNegative) ||
6011 FMF.noSignedZeros()))
6012 if (match(V: Op1, P: m_NegZeroFP()))
6013 return Op0;
6014
6015 // fadd X, 0 ==> X, when we know X is not -0
6016 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6017 if (match(V: Op1, P: m_PosZeroFP()) &&
6018 (FMF.noSignedZeros() || cannotBeNegativeZero(V: Op0, SQ: Q)))
6019 return Op0;
6020
6021 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6022 return nullptr;
6023
6024 if (FMF.noNaNs()) {
6025 // With nnan: X + {+/-}Inf --> {+/-}Inf
6026 if (match(V: Op1, P: m_Inf()))
6027 return Op1;
6028
6029 // With nnan: -X + X --> 0.0 (and commuted variant)
6030 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
6031 // Negative zeros are allowed because we always end up with positive zero:
6032 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
6033 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
6034 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
6035 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
6036 if (match(V: Op0, P: m_FSub(L: m_AnyZeroFP(), R: m_Specific(V: Op1))) ||
6037 match(V: Op1, P: m_FSub(L: m_AnyZeroFP(), R: m_Specific(V: Op0))))
6038 return ConstantFP::getZero(Ty: Op0->getType());
6039
6040 if (match(V: Op0, P: m_FNeg(X: m_Specific(V: Op1))) ||
6041 match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0))))
6042 return ConstantFP::getZero(Ty: Op0->getType());
6043 }
6044
6045 // (X - Y) + Y --> X
6046 // Y + (X - Y) --> X
6047 Value *X;
6048 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
6049 (match(V: Op0, P: m_FSub(L: m_Value(V&: X), R: m_Specific(V: Op1))) ||
6050 match(V: Op1, P: m_FSub(L: m_Value(V&: X), R: m_Specific(V: Op0)))))
6051 return X;
6052
6053 return nullptr;
6054}
6055
6056/// Given operands for an FSub, see if we can fold the result. If not, this
6057/// returns null.
6058static Value *
6059simplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6060 const SimplifyQuery &Q, unsigned MaxRecurse,
6061 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6062 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6063 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6064 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FSub, Op0, Op1, Q))
6065 return C;
6066
6067 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6068 return C;
6069
6070 // fsub X, +0 ==> X
6071 if (canIgnoreSNaN(EB: ExBehavior, FMF) &&
6072 (!canRoundingModeBe(RM: Rounding, QRM: RoundingMode::TowardNegative) ||
6073 FMF.noSignedZeros()))
6074 if (match(V: Op1, P: m_PosZeroFP()))
6075 return Op0;
6076
6077 // fsub X, -0 ==> X, when we know X is not -0
6078 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6079 if (match(V: Op1, P: m_NegZeroFP()) &&
6080 (FMF.noSignedZeros() || cannotBeNegativeZero(V: Op0, SQ: Q)))
6081 return Op0;
6082
6083 // fsub -0.0, (fsub -0.0, X) ==> X
6084 // fsub -0.0, (fneg X) ==> X
6085 Value *X;
6086 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6087 if (match(V: Op0, P: m_NegZeroFP()) && match(V: Op1, P: m_FNeg(X: m_Value(V&: X))))
6088 return X;
6089
6090 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
6091 // fsub 0.0, (fneg X) ==> X if signed zeros are ignored.
6092 if (canIgnoreSNaN(EB: ExBehavior, FMF))
6093 if (FMF.noSignedZeros() && match(V: Op0, P: m_AnyZeroFP()) &&
6094 (match(V: Op1, P: m_FSub(L: m_AnyZeroFP(), R: m_Value(V&: X))) ||
6095 match(V: Op1, P: m_FNeg(X: m_Value(V&: X)))))
6096 return X;
6097
6098 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6099 return nullptr;
6100
6101 if (FMF.noNaNs()) {
6102 // fsub nnan x, x ==> 0.0
6103 if (Op0 == Op1)
6104 return Constant::getNullValue(Ty: Op0->getType());
6105
6106 // With nnan: {+/-}Inf - X --> {+/-}Inf
6107 if (match(V: Op0, P: m_Inf()))
6108 return Op0;
6109
6110 // With nnan: X - {+/-}Inf --> {-/+}Inf
6111 if (match(V: Op1, P: m_Inf()))
6112 return foldConstant(Opcode: Instruction::FNeg, Op&: Op1, Q);
6113 }
6114
6115 // Y - (Y - X) --> X
6116 // (X + Y) - Y --> X
6117 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
6118 (match(V: Op1, P: m_FSub(L: m_Specific(V: Op0), R: m_Value(V&: X))) ||
6119 match(V: Op0, P: m_c_FAdd(L: m_Specific(V: Op1), R: m_Value(V&: X)))))
6120 return X;
6121
6122 return nullptr;
6123}
6124
6125static Value *simplifyFMAFMul(Value *Op0, Value *Op1, FastMathFlags FMF,
6126 const SimplifyQuery &Q, unsigned MaxRecurse,
6127 fp::ExceptionBehavior ExBehavior,
6128 RoundingMode Rounding) {
6129 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6130 return C;
6131
6132 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6133 return nullptr;
6134
6135 // Canonicalize special constants as operand 1.
6136 if (match(V: Op0, P: m_FPOne()) || match(V: Op0, P: m_AnyZeroFP()))
6137 std::swap(a&: Op0, b&: Op1);
6138
6139 // X * 1.0 --> X
6140 if (match(V: Op1, P: m_FPOne()))
6141 return Op0;
6142
6143 if (match(V: Op1, P: m_AnyZeroFP())) {
6144 // X * 0.0 --> 0.0 (with nnan and nsz)
6145 if (FMF.noNaNs() && FMF.noSignedZeros())
6146 return ConstantFP::getZero(Ty: Op0->getType());
6147
6148 KnownFPClass Known = computeKnownFPClass(V: Op0, FMF, InterestedClasses: fcInf | fcNan, SQ: Q);
6149 if (Known.isKnownNever(Mask: fcInf | fcNan)) {
6150 // if nsz is set, return 0.0
6151 if (FMF.noSignedZeros())
6152 return ConstantFP::getZero(Ty: Op0->getType());
6153 // +normal number * (-)0.0 --> (-)0.0
6154 if (Known.SignBit == false)
6155 return Op1;
6156 // -normal number * (-)0.0 --> -(-)0.0
6157 if (Known.SignBit == true)
6158 return foldConstant(Opcode: Instruction::FNeg, Op&: Op1, Q);
6159 }
6160 }
6161
6162 // sqrt(X) * sqrt(X) --> X, if we can:
6163 // 1. Remove the intermediate rounding (reassociate).
6164 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
6165 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
6166 Value *X;
6167 if (Op0 == Op1 && match(V: Op0, P: m_Sqrt(Op0: m_Value(V&: X))) && FMF.allowReassoc() &&
6168 FMF.noNaNs() && FMF.noSignedZeros())
6169 return X;
6170
6171 return nullptr;
6172}
6173
6174/// Given the operands for an FMul, see if we can fold the result
6175static Value *
6176simplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6177 const SimplifyQuery &Q, unsigned MaxRecurse,
6178 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6179 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6180 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6181 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FMul, Op0, Op1, Q))
6182 return C;
6183
6184 // Now apply simplifications that do not require rounding.
6185 return simplifyFMAFMul(Op0, Op1, FMF, Q, MaxRecurse, ExBehavior, Rounding);
6186}
6187
6188Value *llvm::simplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6189 const SimplifyQuery &Q,
6190 fp::ExceptionBehavior ExBehavior,
6191 RoundingMode Rounding) {
6192 return ::simplifyFAddInst(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6193 Rounding);
6194}
6195
6196Value *llvm::simplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6197 const SimplifyQuery &Q,
6198 fp::ExceptionBehavior ExBehavior,
6199 RoundingMode Rounding) {
6200 return ::simplifyFSubInst(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6201 Rounding);
6202}
6203
6204Value *llvm::simplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6205 const SimplifyQuery &Q,
6206 fp::ExceptionBehavior ExBehavior,
6207 RoundingMode Rounding) {
6208 return ::simplifyFMulInst(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6209 Rounding);
6210}
6211
6212Value *llvm::simplifyFMAFMul(Value *Op0, Value *Op1, FastMathFlags FMF,
6213 const SimplifyQuery &Q,
6214 fp::ExceptionBehavior ExBehavior,
6215 RoundingMode Rounding) {
6216 return ::simplifyFMAFMul(Op0, Op1, FMF, Q, MaxRecurse: RecursionLimit, ExBehavior,
6217 Rounding);
6218}
6219
6220static Value *
6221simplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6222 const SimplifyQuery &Q, unsigned,
6223 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6224 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6225 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6226 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FDiv, Op0, Op1, Q))
6227 return C;
6228
6229 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6230 return C;
6231
6232 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6233 return nullptr;
6234
6235 // X / 1.0 -> X
6236 if (match(V: Op1, P: m_FPOne()))
6237 return Op0;
6238
6239 // 0 / X -> 0
6240 // Requires that NaNs are off (X could be zero) and signed zeroes are
6241 // ignored (X could be positive or negative, so the output sign is unknown).
6242 if (FMF.noNaNs() && FMF.noSignedZeros() && match(V: Op0, P: m_AnyZeroFP()))
6243 return ConstantFP::getZero(Ty: Op0->getType());
6244
6245 if (FMF.noNaNs()) {
6246 // X / X -> 1.0 is legal when NaNs are ignored.
6247 // We can ignore infinities because INF/INF is NaN.
6248 if (Op0 == Op1)
6249 return ConstantFP::get(Ty: Op0->getType(), V: 1.0);
6250
6251 // (X * Y) / Y --> X if we can reassociate to the above form.
6252 Value *X;
6253 if (FMF.allowReassoc() && match(V: Op0, P: m_c_FMul(L: m_Value(V&: X), R: m_Specific(V: Op1))))
6254 return X;
6255
6256 // -X / X -> -1.0 and
6257 // X / -X -> -1.0 are legal when NaNs are ignored.
6258 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
6259 if (match(V: Op0, P: m_FNegNSZ(X: m_Specific(V: Op1))) ||
6260 match(V: Op1, P: m_FNegNSZ(X: m_Specific(V: Op0))))
6261 return ConstantFP::get(Ty: Op0->getType(), V: -1.0);
6262
6263 // nnan ninf X / [-]0.0 -> poison
6264 if (FMF.noInfs() && match(V: Op1, P: m_AnyZeroFP()))
6265 return PoisonValue::get(T: Op1->getType());
6266 }
6267
6268 return nullptr;
6269}
6270
6271Value *llvm::simplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6272 const SimplifyQuery &Q,
6273 fp::ExceptionBehavior ExBehavior,
6274 RoundingMode Rounding) {
6275 return ::simplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit, ExBehavior,
6276 Rounding);
6277}
6278
6279static Value *
6280simplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6281 const SimplifyQuery &Q, unsigned,
6282 fp::ExceptionBehavior ExBehavior = fp::ebIgnore,
6283 RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
6284 if (isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6285 if (Constant *C = foldOrCommuteConstant(Opcode: Instruction::FRem, Op0, Op1, Q))
6286 return C;
6287
6288 if (Constant *C = simplifyFPOp(Ops: {Op0, Op1}, FMF, Q, ExBehavior, Rounding))
6289 return C;
6290
6291 if (!isDefaultFPEnvironment(EB: ExBehavior, RM: Rounding))
6292 return nullptr;
6293
6294 // Unlike fdiv, the result of frem always matches the sign of the dividend.
6295 // The constant match may include undef elements in a vector, so return a full
6296 // zero constant as the result.
6297 if (FMF.noNaNs()) {
6298 // +0 % X -> 0
6299 if (match(V: Op0, P: m_PosZeroFP()))
6300 return ConstantFP::getZero(Ty: Op0->getType());
6301 // -0 % X -> -0
6302 if (match(V: Op0, P: m_NegZeroFP()))
6303 return ConstantFP::getNegativeZero(Ty: Op0->getType());
6304 }
6305
6306 return nullptr;
6307}
6308
6309Value *llvm::simplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
6310 const SimplifyQuery &Q,
6311 fp::ExceptionBehavior ExBehavior,
6312 RoundingMode Rounding) {
6313 return ::simplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit, ExBehavior,
6314 Rounding);
6315}
6316
6317//=== Helper functions for higher up the class hierarchy.
6318
6319/// Given the operand for a UnaryOperator, see if we can fold the result.
6320/// If not, this returns null.
6321static Value *simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q,
6322 unsigned MaxRecurse) {
6323 switch (Opcode) {
6324 case Instruction::FNeg:
6325 return simplifyFNegInst(Op, FMF: FastMathFlags(), Q, MaxRecurse);
6326 default:
6327 llvm_unreachable("Unexpected opcode");
6328 }
6329}
6330
6331/// Given the operand for a UnaryOperator, see if we can fold the result.
6332/// If not, this returns null.
6333/// Try to use FastMathFlags when folding the result.
6334static Value *simplifyFPUnOp(unsigned Opcode, Value *Op,
6335 const FastMathFlags &FMF, const SimplifyQuery &Q,
6336 unsigned MaxRecurse) {
6337 switch (Opcode) {
6338 case Instruction::FNeg:
6339 return simplifyFNegInst(Op, FMF, Q, MaxRecurse);
6340 default:
6341 return simplifyUnOp(Opcode, Op, Q, MaxRecurse);
6342 }
6343}
6344
6345Value *llvm::simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q) {
6346 return ::simplifyUnOp(Opcode, Op, Q, MaxRecurse: RecursionLimit);
6347}
6348
6349Value *llvm::simplifyUnOp(unsigned Opcode, Value *Op, FastMathFlags FMF,
6350 const SimplifyQuery &Q) {
6351 return ::simplifyFPUnOp(Opcode, Op, FMF, Q, MaxRecurse: RecursionLimit);
6352}
6353
6354/// Given operands for a BinaryOperator, see if we can fold the result.
6355/// If not, this returns null.
6356static Value *simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6357 const SimplifyQuery &Q, unsigned MaxRecurse) {
6358 switch (Opcode) {
6359 case Instruction::Add:
6360 return simplifyAddInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6361 MaxRecurse);
6362 case Instruction::Sub:
6363 return simplifySubInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6364 MaxRecurse);
6365 case Instruction::Mul:
6366 return simplifyMulInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6367 MaxRecurse);
6368 case Instruction::SDiv:
6369 return simplifySDivInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6370 case Instruction::UDiv:
6371 return simplifyUDivInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6372 case Instruction::SRem:
6373 return simplifySRemInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6374 case Instruction::URem:
6375 return simplifyURemInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6376 case Instruction::Shl:
6377 return simplifyShlInst(Op0: LHS, Op1: RHS, /* IsNSW */ false, /* IsNUW */ false, Q,
6378 MaxRecurse);
6379 case Instruction::LShr:
6380 return simplifyLShrInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6381 case Instruction::AShr:
6382 return simplifyAShrInst(Op0: LHS, Op1: RHS, /* IsExact */ false, Q, MaxRecurse);
6383 case Instruction::And:
6384 return simplifyAndInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6385 case Instruction::Or:
6386 return simplifyOrInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6387 case Instruction::Xor:
6388 return simplifyXorInst(Op0: LHS, Op1: RHS, Q, MaxRecurse);
6389 case Instruction::FAdd:
6390 return simplifyFAddInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6391 case Instruction::FSub:
6392 return simplifyFSubInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6393 case Instruction::FMul:
6394 return simplifyFMulInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6395 case Instruction::FDiv:
6396 return simplifyFDivInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6397 case Instruction::FRem:
6398 return simplifyFRemInst(Op0: LHS, Op1: RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6399 default:
6400 llvm_unreachable("Unexpected opcode");
6401 }
6402}
6403
6404/// Given operands for a BinaryOperator, see if we can fold the result.
6405/// If not, this returns null.
6406/// Try to use FastMathFlags when folding the result.
6407static Value *simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6408 const FastMathFlags &FMF, const SimplifyQuery &Q,
6409 unsigned MaxRecurse) {
6410 switch (Opcode) {
6411 case Instruction::FAdd:
6412 return simplifyFAddInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6413 case Instruction::FSub:
6414 return simplifyFSubInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6415 case Instruction::FMul:
6416 return simplifyFMulInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6417 case Instruction::FDiv:
6418 return simplifyFDivInst(Op0: LHS, Op1: RHS, FMF, Q, MaxRecurse);
6419 default:
6420 return simplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
6421 }
6422}
6423
6424Value *llvm::simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6425 const SimplifyQuery &Q) {
6426 return ::simplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse: RecursionLimit);
6427}
6428
6429Value *llvm::simplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
6430 FastMathFlags FMF, const SimplifyQuery &Q) {
6431 return ::simplifyBinOp(Opcode, LHS, RHS, FMF, Q, MaxRecurse: RecursionLimit);
6432}
6433
6434/// Given operands for a CmpInst, see if we can fold the result.
6435static Value *simplifyCmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
6436 const SimplifyQuery &Q, unsigned MaxRecurse) {
6437 if (CmpInst::isIntPredicate(P: Predicate))
6438 return simplifyICmpInst(Pred: Predicate, LHS, RHS, Q, MaxRecurse);
6439 return simplifyFCmpInst(Pred: Predicate, LHS, RHS, FMF: FastMathFlags(), Q, MaxRecurse);
6440}
6441
6442Value *llvm::simplifyCmpInst(CmpPredicate Predicate, Value *LHS, Value *RHS,
6443 const SimplifyQuery &Q) {
6444 return ::simplifyCmpInst(Predicate, LHS, RHS, Q, MaxRecurse: RecursionLimit);
6445}
6446
6447static bool isIdempotent(Intrinsic::ID ID) {
6448 switch (ID) {
6449 default:
6450 return false;
6451
6452 // Unary idempotent: f(f(x)) = f(x)
6453 case Intrinsic::fabs:
6454 case Intrinsic::floor:
6455 case Intrinsic::ceil:
6456 case Intrinsic::trunc:
6457 case Intrinsic::rint:
6458 case Intrinsic::nearbyint:
6459 case Intrinsic::round:
6460 case Intrinsic::roundeven:
6461 case Intrinsic::canonicalize:
6462 case Intrinsic::arithmetic_fence:
6463 return true;
6464 }
6465}
6466
6467/// Return true if the intrinsic rounds a floating-point value to an integral
6468/// floating-point value (not an integer type).
6469static bool removesFPFraction(Intrinsic::ID ID) {
6470 switch (ID) {
6471 default:
6472 return false;
6473
6474 case Intrinsic::floor:
6475 case Intrinsic::ceil:
6476 case Intrinsic::trunc:
6477 case Intrinsic::rint:
6478 case Intrinsic::nearbyint:
6479 case Intrinsic::round:
6480 case Intrinsic::roundeven:
6481 return true;
6482 }
6483}
6484
6485static Value *simplifyRelativeLoad(Constant *Ptr, Constant *Offset,
6486 const DataLayout &DL) {
6487 GlobalValue *PtrSym;
6488 APInt PtrOffset;
6489 if (!IsConstantOffsetFromGlobal(C: Ptr, GV&: PtrSym, Offset&: PtrOffset, DL))
6490 return nullptr;
6491
6492 Type *Int32Ty = Type::getInt32Ty(C&: Ptr->getContext());
6493
6494 auto *OffsetConstInt = dyn_cast<ConstantInt>(Val: Offset);
6495 if (!OffsetConstInt || OffsetConstInt->getBitWidth() > 64)
6496 return nullptr;
6497
6498 APInt OffsetInt = OffsetConstInt->getValue().sextOrTrunc(
6499 width: DL.getIndexTypeSizeInBits(Ty: Ptr->getType()));
6500 if (OffsetInt.srem(RHS: 4) != 0)
6501 return nullptr;
6502
6503 Constant *Loaded =
6504 ConstantFoldLoadFromConstPtr(C: Ptr, Ty: Int32Ty, Offset: std::move(OffsetInt), DL);
6505 if (!Loaded)
6506 return nullptr;
6507
6508 auto *LoadedCE = dyn_cast<ConstantExpr>(Val: Loaded);
6509 if (!LoadedCE)
6510 return nullptr;
6511
6512 if (LoadedCE->getOpcode() == Instruction::Trunc) {
6513 LoadedCE = dyn_cast<ConstantExpr>(Val: LoadedCE->getOperand(i_nocapture: 0));
6514 if (!LoadedCE)
6515 return nullptr;
6516 }
6517
6518 if (LoadedCE->getOpcode() != Instruction::Sub)
6519 return nullptr;
6520
6521 auto *LoadedLHS = dyn_cast<ConstantExpr>(Val: LoadedCE->getOperand(i_nocapture: 0));
6522 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
6523 return nullptr;
6524 auto *LoadedLHSPtr = LoadedLHS->getOperand(i_nocapture: 0);
6525
6526 Constant *LoadedRHS = LoadedCE->getOperand(i_nocapture: 1);
6527 GlobalValue *LoadedRHSSym;
6528 APInt LoadedRHSOffset;
6529 if (!IsConstantOffsetFromGlobal(C: LoadedRHS, GV&: LoadedRHSSym, Offset&: LoadedRHSOffset,
6530 DL) ||
6531 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
6532 return nullptr;
6533
6534 return LoadedLHSPtr;
6535}
6536
6537// TODO: Need to pass in FastMathFlags
6538static Value *simplifyLdexp(Value *Op0, Value *Op1, const SimplifyQuery &Q,
6539 bool IsStrict) {
6540 // ldexp(poison, x) -> poison
6541 // ldexp(x, poison) -> poison
6542 if (isa<PoisonValue>(Val: Op0) || isa<PoisonValue>(Val: Op1))
6543 return Op0;
6544
6545 // ldexp(undef, x) -> nan
6546 if (Q.isUndefValue(V: Op0))
6547 return ConstantFP::getNaN(Ty: Op0->getType());
6548
6549 if (!IsStrict) {
6550 // TODO: Could insert a canonicalize for strict
6551
6552 // ldexp(x, undef) -> x
6553 if (Q.isUndefValue(V: Op1))
6554 return Op0;
6555 }
6556
6557 const APFloat *C = nullptr;
6558 match(V: Op0, P: PatternMatch::m_APFloat(Res&: C));
6559
6560 // These cases should be safe, even with strictfp.
6561 // ldexp(0.0, x) -> 0.0
6562 // ldexp(-0.0, x) -> -0.0
6563 // ldexp(inf, x) -> inf
6564 // ldexp(-inf, x) -> -inf
6565 if (C && (C->isZero() || C->isInfinity()))
6566 return Op0;
6567
6568 // These are canonicalization dropping, could do it if we knew how we could
6569 // ignore denormal flushes and target handling of nan payload bits.
6570 if (IsStrict)
6571 return nullptr;
6572
6573 // TODO: Could quiet this with strictfp if the exception mode isn't strict.
6574 if (C && C->isNaN())
6575 return ConstantFP::get(Ty: Op0->getType(), V: C->makeQuiet());
6576
6577 // ldexp(x, 0) -> x
6578
6579 // TODO: Could fold this if we know the exception mode isn't
6580 // strict, we know the denormal mode and other target modes.
6581 if (match(V: Op1, P: PatternMatch::m_ZeroInt()))
6582 return Op0;
6583
6584 return nullptr;
6585}
6586
6587static Value *simplifyUnaryIntrinsic(Intrinsic::ID IID, Value *Op0,
6588 FastMathFlags FMF,
6589 const SimplifyQuery &Q) {
6590 // Idempotent functions return the same result when called repeatedly.
6591 if (isIdempotent(ID: IID))
6592 if (auto *II = dyn_cast<IntrinsicInst>(Val: Op0))
6593 if (II->getIntrinsicID() == IID)
6594 return II;
6595
6596 if (removesFPFraction(ID: IID)) {
6597 // Converting from int or calling a rounding function always results in a
6598 // finite integral number or infinity. For those inputs, rounding functions
6599 // always return the same value, so the (2nd) rounding is eliminated. Ex:
6600 // floor (sitofp x) -> sitofp x
6601 // round (ceil x) -> ceil x
6602 auto *II = dyn_cast<IntrinsicInst>(Val: Op0);
6603 if ((II && removesFPFraction(ID: II->getIntrinsicID())) ||
6604 match(V: Op0, P: m_IToFP(Op: m_Value())))
6605 return Op0;
6606 }
6607
6608 Value *X;
6609 switch (IID) {
6610 case Intrinsic::fabs: {
6611 KnownFPClass KnownClass = computeKnownFPClass(V: Op0, InterestedClasses: fcAllFlags, SQ: Q);
6612 if (KnownClass.SignBit == false)
6613 return Op0;
6614
6615 if (KnownClass.cannotBeOrderedLessThanZero() &&
6616 KnownClass.isKnownNeverNaN() && FMF.noSignedZeros())
6617 return Op0;
6618
6619 break;
6620 }
6621 case Intrinsic::bswap:
6622 // bswap(bswap(x)) -> x
6623 if (match(V: Op0, P: m_BSwap(Op0: m_Value(V&: X))))
6624 return X;
6625 break;
6626 case Intrinsic::bitreverse:
6627 // bitreverse(bitreverse(x)) -> x
6628 if (match(V: Op0, P: m_BitReverse(Op0: m_Value(V&: X))))
6629 return X;
6630 break;
6631 case Intrinsic::ctpop: {
6632 // ctpop(X) -> 1 iff X is non-zero power of 2.
6633 if (isKnownToBeAPowerOfTwo(V: Op0, DL: Q.DL, /*OrZero*/ false, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
6634 return ConstantInt::get(Ty: Op0->getType(), V: 1);
6635 // If everything but the lowest bit is zero, that bit is the pop-count. Ex:
6636 // ctpop(and X, 1) --> and X, 1
6637 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
6638 if (MaskedValueIsZero(V: Op0, Mask: APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 1),
6639 SQ: Q))
6640 return Op0;
6641 break;
6642 }
6643 case Intrinsic::exp:
6644 // exp(log(x)) -> x
6645 if (FMF.allowReassoc() &&
6646 match(V: Op0, P: m_Intrinsic<Intrinsic::log>(Ops: m_Value(V&: X))))
6647 return X;
6648 break;
6649 case Intrinsic::exp2:
6650 // exp2(log2(x)) -> x
6651 if (FMF.allowReassoc() &&
6652 match(V: Op0, P: m_Intrinsic<Intrinsic::log2>(Ops: m_Value(V&: X))))
6653 return X;
6654 break;
6655 case Intrinsic::exp10:
6656 // exp10(log10(x)) -> x
6657 if (FMF.allowReassoc() &&
6658 match(V: Op0, P: m_Intrinsic<Intrinsic::log10>(Ops: m_Value(V&: X))))
6659 return X;
6660 break;
6661 case Intrinsic::log:
6662 // log(exp(x)) -> x
6663 if (FMF.allowReassoc() &&
6664 match(V: Op0, P: m_Intrinsic<Intrinsic::exp>(Ops: m_Value(V&: X))))
6665 return X;
6666 break;
6667 case Intrinsic::log2:
6668 // log2(exp2(x)) -> x
6669 if (FMF.allowReassoc() &&
6670 (match(V: Op0, P: m_Intrinsic<Intrinsic::exp2>(Ops: m_Value(V&: X))) ||
6671 match(V: Op0,
6672 P: m_Intrinsic<Intrinsic::pow>(Ops: m_SpecificFP(V: 2.0), Ops: m_Value(V&: X)))))
6673 return X;
6674 break;
6675 case Intrinsic::log10:
6676 // log10(pow(10.0, x)) -> x
6677 // log10(exp10(x)) -> x
6678 if (FMF.allowReassoc() &&
6679 (match(V: Op0, P: m_Intrinsic<Intrinsic::exp10>(Ops: m_Value(V&: X))) ||
6680 match(V: Op0,
6681 P: m_Intrinsic<Intrinsic::pow>(Ops: m_SpecificFP(V: 10.0), Ops: m_Value(V&: X)))))
6682 return X;
6683 break;
6684 case Intrinsic::vector_reverse:
6685 // vector.reverse(vector.reverse(x)) -> x
6686 if (match(V: Op0, P: m_VecReverse(Op0: m_Value(V&: X))))
6687 return X;
6688 // vector.reverse(splat(X)) -> splat(X)
6689 if (isSplatValue(V: Op0))
6690 return Op0;
6691 break;
6692 case Intrinsic::structured_gep:
6693 return Op0;
6694 default:
6695 break;
6696 }
6697
6698 return nullptr;
6699}
6700
6701/// Given a min/max intrinsic, see if it can be removed based on having an
6702/// operand that is another min/max intrinsic with shared operand(s). The caller
6703/// is expected to swap the operand arguments to handle commutation.
6704static Value *foldMinMaxSharedOp(Intrinsic::ID IID, Value *Op0, Value *Op1) {
6705 Value *X, *Y;
6706 if (!match(V: Op0, P: m_MaxOrMin(Op0: m_Value(V&: X), Op1: m_Value(V&: Y))))
6707 return nullptr;
6708
6709 auto *MM0 = dyn_cast<IntrinsicInst>(Val: Op0);
6710 if (!MM0)
6711 return nullptr;
6712 Intrinsic::ID IID0 = MM0->getIntrinsicID();
6713
6714 if (Op1 == X || Op1 == Y ||
6715 match(V: Op1, P: m_c_MaxOrMin(L: m_Specific(V: X), R: m_Specific(V: Y)))) {
6716 // max (max X, Y), X --> max X, Y
6717 if (IID0 == IID)
6718 return MM0;
6719 // max (min X, Y), X --> X
6720 if (IID0 == getInverseMinMaxIntrinsic(MinMaxID: IID))
6721 return Op1;
6722 }
6723 return nullptr;
6724}
6725
6726/// Given a min/max intrinsic, see if it can be removed based on having an
6727/// operand that is another min/max intrinsic with shared operand(s). The caller
6728/// is expected to swap the operand arguments to handle commutation.
6729static Value *foldMinimumMaximumSharedOp(Intrinsic::ID IID, Value *Op0,
6730 Value *Op1) {
6731 auto IsMinimumMaximumIntrinsic = [](Intrinsic::ID ID) {
6732 switch (ID) {
6733 case Intrinsic::maxnum:
6734 case Intrinsic::minnum:
6735 case Intrinsic::maximum:
6736 case Intrinsic::minimum:
6737 case Intrinsic::maximumnum:
6738 case Intrinsic::minimumnum:
6739 return true;
6740 default:
6741 return false;
6742 }
6743 };
6744
6745 assert(IsMinimumMaximumIntrinsic(IID) && "Unsupported intrinsic");
6746
6747 auto *M0 = dyn_cast<IntrinsicInst>(Val: Op0);
6748 // If Op0 is not the same intrinsic as IID, do not process.
6749 // This is a difference with integer min/max handling. We do not process the
6750 // case like max(min(X,Y),min(X,Y)) => min(X,Y). But it can be handled by GVN.
6751 if (!M0 || M0->getIntrinsicID() != IID)
6752 return nullptr;
6753 Value *X0 = M0->getOperand(i_nocapture: 0);
6754 Value *Y0 = M0->getOperand(i_nocapture: 1);
6755 // Simple case, m(m(X,Y), X) => m(X, Y)
6756 // m(m(X,Y), Y) => m(X, Y)
6757 // For minimum/maximum, X is NaN => m(NaN, Y) == NaN and m(NaN, NaN) == NaN.
6758 // For minimum/maximum, Y is NaN => m(X, NaN) == NaN and m(NaN, NaN) == NaN.
6759 // For minnum/maxnum, X is NaN => m(NaN, Y) == Y and m(Y, Y) == Y.
6760 // For minnum/maxnum, Y is NaN => m(X, NaN) == X and m(X, NaN) == X.
6761 if (X0 == Op1 || Y0 == Op1)
6762 return M0;
6763
6764 auto *M1 = dyn_cast<IntrinsicInst>(Val: Op1);
6765 if (!M1 || !IsMinimumMaximumIntrinsic(M1->getIntrinsicID()))
6766 return nullptr;
6767 Value *X1 = M1->getOperand(i_nocapture: 0);
6768 Value *Y1 = M1->getOperand(i_nocapture: 1);
6769 Intrinsic::ID IID1 = M1->getIntrinsicID();
6770 // we have a case m(m(X,Y),m'(X,Y)) taking into account m' is commutative.
6771 // if m' is m or inversion of m => m(m(X,Y),m'(X,Y)) == m(X,Y).
6772 // For minimum/maximum, X is NaN => m(NaN,Y) == m'(NaN, Y) == NaN.
6773 // For minimum/maximum, Y is NaN => m(X,NaN) == m'(X, NaN) == NaN.
6774 // For minnum/maxnum, X is NaN => m(NaN,Y) == m'(NaN, Y) == Y.
6775 // For minnum/maxnum, Y is NaN => m(X,NaN) == m'(X, NaN) == X.
6776 if ((X0 == X1 && Y0 == Y1) || (X0 == Y1 && Y0 == X1))
6777 if (IID1 == IID || getInverseMinMaxIntrinsic(MinMaxID: IID1) == IID)
6778 return M0;
6779
6780 return nullptr;
6781}
6782
6783enum class MinMaxOptResult {
6784 CannotOptimize = 0,
6785 UseNewConstVal = 1,
6786 UseOtherVal = 2,
6787 // For undef/poison, we can choose to either propgate undef/poison or
6788 // use the LHS value depending on what will allow more optimization.
6789 UseEither = 3
6790};
6791// Get the optimized value for a min/max instruction with a single constant
6792// input (either undef or scalar constantFP). The result may indicate to
6793// use the non-const LHS value, use a new constant value instead (with NaNs
6794// quieted), or to choose either option in the case of undef/poison.
6795static MinMaxOptResult OptimizeConstMinMax(const Constant *RHSConst,
6796 const Intrinsic::ID IID,
6797 FastMathFlags FMF,
6798 Constant **OutNewConstVal) {
6799 assert(OutNewConstVal != nullptr);
6800
6801 bool PropagateNaN = IID == Intrinsic::minimum || IID == Intrinsic::maximum;
6802 bool PropagateSNaN = IID == Intrinsic::minnum || IID == Intrinsic::maxnum;
6803 bool IsMin = IID == Intrinsic::minimum || IID == Intrinsic::minnum ||
6804 IID == Intrinsic::minimumnum;
6805
6806 // min/max(x, poison) -> either x or poison
6807 if (isa<UndefValue>(Val: RHSConst)) {
6808 *OutNewConstVal = const_cast<Constant *>(RHSConst);
6809 return MinMaxOptResult::UseEither;
6810 }
6811
6812 const ConstantFP *CFP = dyn_cast<ConstantFP>(Val: RHSConst);
6813 if (!CFP)
6814 return MinMaxOptResult::CannotOptimize;
6815 APFloat CAPF = CFP->getValueAPF();
6816
6817 // minnum(x, qnan) -> x
6818 // maxnum(x, qnan) -> x
6819 // minnum(x, snan) -> qnan
6820 // maxnum(x, snan) -> qnan
6821 // minimum(X, nan) -> qnan
6822 // maximum(X, nan) -> qnan
6823 // minimumnum(X, nan) -> x
6824 // maximumnum(X, nan) -> x
6825 if (CAPF.isNaN()) {
6826 if (PropagateNaN || (PropagateSNaN && CAPF.isSignaling())) {
6827 *OutNewConstVal = ConstantFP::get(Ty: CFP->getType(), V: CAPF.makeQuiet());
6828 return MinMaxOptResult::UseNewConstVal;
6829 }
6830 return MinMaxOptResult::UseOtherVal;
6831 }
6832
6833 if (CAPF.isInfinity() || (FMF.noInfs() && CAPF.isLargest())) {
6834 // minnum(X, -inf) -> -inf (ignoring sNaN -> qNaN propagation)
6835 // maxnum(X, +inf) -> +inf (ignoring sNaN -> qNaN propagation)
6836 // minimum(X, -inf) -> -inf if nnan
6837 // maximum(X, +inf) -> +inf if nnan
6838 // minimumnum(X, -inf) -> -inf
6839 // maximumnum(X, +inf) -> +inf
6840 if (CAPF.isNegative() == IsMin && (!PropagateNaN || FMF.noNaNs())) {
6841 *OutNewConstVal = const_cast<Constant *>(RHSConst);
6842 return MinMaxOptResult::UseNewConstVal;
6843 }
6844
6845 // minnum(X, +inf) -> X if nnan
6846 // maxnum(X, -inf) -> X if nnan
6847 // minimum(X, +inf) -> X (ignoring quieting of sNaNs)
6848 // maximum(X, -inf) -> X (ignoring quieting of sNaNs)
6849 // minimumnum(X, +inf) -> X if nnan
6850 // maximumnum(X, -inf) -> X if nnan
6851 if (CAPF.isNegative() != IsMin && (PropagateNaN || FMF.noNaNs()))
6852 return MinMaxOptResult::UseOtherVal;
6853 }
6854 return MinMaxOptResult::CannotOptimize;
6855}
6856
6857static Value *simplifySVEIntReduction(Intrinsic::ID IID, Type *ReturnType,
6858 Value *Op0, Value *Op1) {
6859 Constant *C0 = dyn_cast<Constant>(Val: Op0);
6860 Constant *C1 = dyn_cast<Constant>(Val: Op1);
6861 unsigned Width = ReturnType->getPrimitiveSizeInBits();
6862
6863 // All false predicate or reduction of neutral values ==> neutral result.
6864 switch (IID) {
6865 case Intrinsic::aarch64_sve_eorv:
6866 case Intrinsic::aarch64_sve_orv:
6867 case Intrinsic::aarch64_sve_saddv:
6868 case Intrinsic::aarch64_sve_uaddv:
6869 case Intrinsic::aarch64_sve_umaxv:
6870 if ((C0 && C0->isNullValue()) || (C1 && C1->isNullValue()))
6871 return ConstantInt::get(Ty: ReturnType, V: 0);
6872 break;
6873 case Intrinsic::aarch64_sve_andv:
6874 case Intrinsic::aarch64_sve_uminv:
6875 if ((C0 && C0->isNullValue()) || (C1 && C1->isAllOnesValue()))
6876 return ConstantInt::get(Ty: ReturnType, V: APInt::getMaxValue(numBits: Width));
6877 break;
6878 case Intrinsic::aarch64_sve_smaxv:
6879 if ((C0 && C0->isNullValue()) || (C1 && C1->isMinSignedValue()))
6880 return ConstantInt::get(Ty: ReturnType, V: APInt::getSignedMinValue(numBits: Width));
6881 break;
6882 case Intrinsic::aarch64_sve_sminv:
6883 if ((C0 && C0->isNullValue()) || (C1 && C1->isMaxSignedValue()))
6884 return ConstantInt::get(Ty: ReturnType, V: APInt::getSignedMaxValue(numBits: Width));
6885 break;
6886 }
6887
6888 switch (IID) {
6889 case Intrinsic::aarch64_sve_andv:
6890 case Intrinsic::aarch64_sve_orv:
6891 case Intrinsic::aarch64_sve_smaxv:
6892 case Intrinsic::aarch64_sve_sminv:
6893 case Intrinsic::aarch64_sve_umaxv:
6894 case Intrinsic::aarch64_sve_uminv:
6895 // sve_reduce_##(all, splat(X)) ==> X
6896 if (C0 && C0->isAllOnesValue()) {
6897 if (Value *SplatVal = getSplatValue(V: Op1)) {
6898 assert(SplatVal->getType() == ReturnType && "Unexpected result type!");
6899 return SplatVal;
6900 }
6901 }
6902 break;
6903 case Intrinsic::aarch64_sve_eorv:
6904 // sve_reduce_xor(all, splat(X)) ==> 0
6905 if (C0 && C0->isAllOnesValue())
6906 if (getSplatValue(V: Op1))
6907 return ConstantInt::get(Ty: ReturnType, V: 0);
6908 break;
6909 }
6910
6911 return nullptr;
6912}
6913
6914static Value *simplifyBinaryIntrinsic(Intrinsic::ID IID, Type *ReturnType,
6915 Value *Op0, Value *Op1, FastMathFlags FMF,
6916 const SimplifyQuery &Q) {
6917 unsigned BitWidth = ReturnType->getScalarSizeInBits();
6918 switch (IID) {
6919 case Intrinsic::get_active_lane_mask: {
6920 if (match(V: Op1, P: m_Zero()))
6921 return ConstantInt::getFalse(Ty: ReturnType);
6922
6923 if (!Q.CxtI)
6924 break;
6925
6926 const Function *F = Q.CxtI->getFunction();
6927 auto *ScalableTy = dyn_cast<ScalableVectorType>(Val: ReturnType);
6928 Attribute Attr = F->getFnAttribute(Kind: Attribute::VScaleRange);
6929 if (ScalableTy && Attr.isValid()) {
6930 std::optional<unsigned> VScaleMax = Attr.getVScaleRangeMax();
6931 if (!VScaleMax)
6932 break;
6933 uint64_t MaxPossibleMaskElements =
6934 (uint64_t)ScalableTy->getMinNumElements() * (*VScaleMax);
6935
6936 const APInt *Op1Val;
6937 if (match(V: Op0, P: m_Zero()) && match(V: Op1, P: m_APInt(Res&: Op1Val)) &&
6938 Op1Val->uge(RHS: MaxPossibleMaskElements))
6939 return ConstantInt::getAllOnesValue(Ty: ReturnType);
6940 }
6941 break;
6942 }
6943 case Intrinsic::abs:
6944 // abs(abs(x)) -> abs(x). We don't need to worry about the nsw arg here.
6945 // It is always ok to pick the earlier abs. We'll just lose nsw if its only
6946 // on the outer abs.
6947 if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Ops: m_Value(), Ops: m_Value())))
6948 return Op0;
6949 break;
6950
6951 case Intrinsic::cttz: {
6952 Value *X;
6953 if (match(V: Op0, P: m_Shl(L: m_One(), R: m_Value(V&: X))))
6954 return X;
6955 break;
6956 }
6957 case Intrinsic::ctlz: {
6958 Value *X;
6959 if (match(V: Op0, P: m_LShr(L: m_Negative(), R: m_Value(V&: X))))
6960 return X;
6961 if (match(V: Op0, P: m_AShr(L: m_Negative(), R: m_Value())))
6962 return Constant::getNullValue(Ty: ReturnType);
6963 break;
6964 }
6965 case Intrinsic::pdep: {
6966 if (match(V: Op0, P: m_Zero()))
6967 return Constant::getNullValue(Ty: ReturnType);
6968 if (match(V: Op1, P: m_Zero()))
6969 return Constant::getNullValue(Ty: ReturnType);
6970 if (match(V: Op1, P: m_AllOnes()))
6971 return Op0;
6972 break;
6973 }
6974 case Intrinsic::pext: {
6975 if (match(V: Op0, P: m_Zero()))
6976 return Constant::getNullValue(Ty: ReturnType);
6977 if (match(V: Op1, P: m_Zero()))
6978 return Constant::getNullValue(Ty: ReturnType);
6979 if (match(V: Op1, P: m_AllOnes()))
6980 return Op0;
6981 break;
6982 }
6983 case Intrinsic::ptrmask: {
6984 // NOTE: We can't apply this simplifications based on the value of Op1
6985 // because we need to preserve provenance.
6986 if (Q.isUndefValue(V: Op0) || match(V: Op0, P: m_Zero()))
6987 return Constant::getNullValue(Ty: Op0->getType());
6988
6989 assert(Op1->getType()->getScalarSizeInBits() ==
6990 Q.DL.getIndexTypeSizeInBits(Op0->getType()) &&
6991 "Invalid mask width");
6992 // If index-width (mask size) is less than pointer-size then mask is
6993 // 1-extended.
6994 if (match(V: Op1, P: m_PtrToIntOrAddr(Op: m_Specific(V: Op0))))
6995 return Op0;
6996
6997 // NOTE: We may have attributes associated with the return value of the
6998 // llvm.ptrmask intrinsic that will be lost when we just return the
6999 // operand. We should try to preserve them.
7000 if (match(V: Op1, P: m_AllOnes()) || Q.isUndefValue(V: Op1))
7001 return Op0;
7002
7003 Constant *C;
7004 if (match(V: Op1, P: m_ImmConstant(C))) {
7005 KnownBits PtrKnown = computeKnownBits(V: Op0, Q);
7006 // See if we only masking off bits we know are already zero due to
7007 // alignment.
7008 APInt IrrelevantPtrBits =
7009 PtrKnown.Zero.zextOrTrunc(width: C->getType()->getScalarSizeInBits());
7010 C = ConstantFoldBinaryOpOperands(
7011 Opcode: Instruction::Or, LHS: C, RHS: ConstantInt::get(Ty: C->getType(), V: IrrelevantPtrBits),
7012 DL: Q.DL);
7013 if (C != nullptr && C->isAllOnesValue())
7014 return Op0;
7015 }
7016 break;
7017 }
7018 case Intrinsic::smax:
7019 case Intrinsic::smin:
7020 case Intrinsic::umax:
7021 case Intrinsic::umin: {
7022 // If the arguments are the same, this is a no-op.
7023 if (Op0 == Op1)
7024 return Op0;
7025
7026 // Canonicalize immediate constant operand as Op1.
7027 if (match(V: Op0, P: m_ImmConstant()))
7028 std::swap(a&: Op0, b&: Op1);
7029
7030 // Assume undef is the limit value.
7031 if (Q.isUndefValue(V: Op1))
7032 return ConstantInt::get(
7033 Ty: ReturnType, V: MinMaxIntrinsic::getSaturationPoint(ID: IID, numBits: BitWidth));
7034
7035 const APInt *C;
7036 if (match(V: Op1, P: m_APIntAllowPoison(Res&: C))) {
7037 // Clamp to limit value. For example:
7038 // umax(i8 %x, i8 255) --> 255
7039 if (*C == MinMaxIntrinsic::getSaturationPoint(ID: IID, numBits: BitWidth))
7040 return ConstantInt::get(Ty: ReturnType, V: *C);
7041
7042 // If the constant op is the opposite of the limit value, the other must
7043 // be larger/smaller or equal. For example:
7044 // umin(i8 %x, i8 255) --> %x
7045 if (*C == MinMaxIntrinsic::getSaturationPoint(
7046 ID: getInverseMinMaxIntrinsic(MinMaxID: IID), numBits: BitWidth))
7047 return Op0;
7048
7049 // Remove nested call if constant operands allow it. Example:
7050 // max (max X, 7), 5 -> max X, 7
7051 auto *MinMax0 = dyn_cast<IntrinsicInst>(Val: Op0);
7052 if (MinMax0 && MinMax0->getIntrinsicID() == IID) {
7053 // TODO: loosen undef/splat restrictions for vector constants.
7054 Value *M00 = MinMax0->getOperand(i_nocapture: 0), *M01 = MinMax0->getOperand(i_nocapture: 1);
7055 const APInt *InnerC;
7056 if ((match(V: M00, P: m_APInt(Res&: InnerC)) || match(V: M01, P: m_APInt(Res&: InnerC))) &&
7057 ICmpInst::compare(LHS: *InnerC, RHS: *C,
7058 Pred: ICmpInst::getNonStrictPredicate(
7059 pred: MinMaxIntrinsic::getPredicate(ID: IID))))
7060 return Op0;
7061 }
7062 }
7063
7064 if (Value *V = foldMinMaxSharedOp(IID, Op0, Op1))
7065 return V;
7066 if (Value *V = foldMinMaxSharedOp(IID, Op0: Op1, Op1: Op0))
7067 return V;
7068
7069 ICmpInst::Predicate Pred =
7070 ICmpInst::getNonStrictPredicate(pred: MinMaxIntrinsic::getPredicate(ID: IID));
7071 if (isICmpTrue(Pred, LHS: Op0, RHS: Op1, Q: Q.getWithoutUndef(), MaxRecurse: RecursionLimit))
7072 return Op0;
7073 if (isICmpTrue(Pred, LHS: Op1, RHS: Op0, Q: Q.getWithoutUndef(), MaxRecurse: RecursionLimit))
7074 return Op1;
7075
7076 break;
7077 }
7078 case Intrinsic::scmp:
7079 case Intrinsic::ucmp: {
7080 // Fold to a constant if the relationship between operands can be
7081 // established with certainty
7082 if (isICmpTrue(Pred: CmpInst::ICMP_EQ, LHS: Op0, RHS: Op1, Q, MaxRecurse: RecursionLimit))
7083 return Constant::getNullValue(Ty: ReturnType);
7084
7085 ICmpInst::Predicate PredGT =
7086 IID == Intrinsic::scmp ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
7087 if (isICmpTrue(Pred: PredGT, LHS: Op0, RHS: Op1, Q, MaxRecurse: RecursionLimit))
7088 return ConstantInt::get(Ty: ReturnType, V: 1);
7089
7090 ICmpInst::Predicate PredLT =
7091 IID == Intrinsic::scmp ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
7092 if (isICmpTrue(Pred: PredLT, LHS: Op0, RHS: Op1, Q, MaxRecurse: RecursionLimit))
7093 return ConstantInt::getSigned(Ty: ReturnType, V: -1);
7094
7095 break;
7096 }
7097 case Intrinsic::usub_with_overflow:
7098 case Intrinsic::ssub_with_overflow:
7099 // X - X -> { 0, false }
7100 // X - undef -> { 0, false }
7101 // undef - X -> { 0, false }
7102 if (Op0 == Op1 || Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7103 return Constant::getNullValue(Ty: ReturnType);
7104 break;
7105 case Intrinsic::uadd_with_overflow:
7106 case Intrinsic::sadd_with_overflow:
7107 // X + undef -> { -1, false }
7108 // undef + x -> { -1, false }
7109 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1)) {
7110 return ConstantStruct::get(
7111 T: cast<StructType>(Val: ReturnType),
7112 V: {Constant::getAllOnesValue(Ty: ReturnType->getStructElementType(N: 0)),
7113 Constant::getNullValue(Ty: ReturnType->getStructElementType(N: 1))});
7114 }
7115 break;
7116 case Intrinsic::umul_with_overflow:
7117 case Intrinsic::smul_with_overflow:
7118 // 0 * X -> { 0, false }
7119 // X * 0 -> { 0, false }
7120 if (match(V: Op0, P: m_Zero()) || match(V: Op1, P: m_Zero()))
7121 return Constant::getNullValue(Ty: ReturnType);
7122 // undef * X -> { 0, false }
7123 // X * undef -> { 0, false }
7124 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7125 return Constant::getNullValue(Ty: ReturnType);
7126 break;
7127 case Intrinsic::uadd_sat:
7128 // sat(MAX + X) -> MAX
7129 // sat(X + MAX) -> MAX
7130 if (match(V: Op0, P: m_AllOnes()) || match(V: Op1, P: m_AllOnes()))
7131 return Constant::getAllOnesValue(Ty: ReturnType);
7132 [[fallthrough]];
7133 case Intrinsic::sadd_sat:
7134 // sat(X + undef) -> -1
7135 // sat(undef + X) -> -1
7136 // For unsigned: Assume undef is MAX, thus we saturate to MAX (-1).
7137 // For signed: Assume undef is ~X, in which case X + ~X = -1.
7138 if (Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7139 return Constant::getAllOnesValue(Ty: ReturnType);
7140
7141 // X + 0 -> X
7142 if (match(V: Op1, P: m_Zero()))
7143 return Op0;
7144 // 0 + X -> X
7145 if (match(V: Op0, P: m_Zero()))
7146 return Op1;
7147 break;
7148 case Intrinsic::usub_sat:
7149 // sat(0 - X) -> 0, sat(X - MAX) -> 0
7150 if (match(V: Op0, P: m_Zero()) || match(V: Op1, P: m_AllOnes()))
7151 return Constant::getNullValue(Ty: ReturnType);
7152 [[fallthrough]];
7153 case Intrinsic::ssub_sat:
7154 // X - X -> 0, X - undef -> 0, undef - X -> 0
7155 if (Op0 == Op1 || Q.isUndefValue(V: Op0) || Q.isUndefValue(V: Op1))
7156 return Constant::getNullValue(Ty: ReturnType);
7157 // X - 0 -> X
7158 if (match(V: Op1, P: m_Zero()))
7159 return Op0;
7160 break;
7161 case Intrinsic::load_relative:
7162 if (auto *C0 = dyn_cast<Constant>(Val: Op0))
7163 if (auto *C1 = dyn_cast<Constant>(Val: Op1))
7164 return simplifyRelativeLoad(Ptr: C0, Offset: C1, DL: Q.DL);
7165 break;
7166 case Intrinsic::powi:
7167 if (auto *Power = dyn_cast<ConstantInt>(Val: Op1)) {
7168 // powi(x, 0) -> 1.0
7169 if (Power->isZero())
7170 return ConstantFP::get(Ty: Op0->getType(), V: 1.0);
7171 // powi(x, 1) -> x
7172 if (Power->isOne())
7173 return Op0;
7174 }
7175 break;
7176 case Intrinsic::ldexp:
7177 return simplifyLdexp(Op0, Op1, Q, IsStrict: false);
7178 case Intrinsic::copysign:
7179 // copysign X, X --> X
7180 if (Op0 == Op1)
7181 return Op0;
7182 // copysign -X, X --> X
7183 // copysign X, -X --> -X
7184 if (match(V: Op0, P: m_FNeg(X: m_Specific(V: Op1))) ||
7185 match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0))))
7186 return Op1;
7187 break;
7188 case Intrinsic::is_fpclass: {
7189 uint64_t Mask = cast<ConstantInt>(Val: Op1)->getZExtValue();
7190 // If all tests are made, it doesn't matter what the value is.
7191 if ((Mask & fcAllFlags) == fcAllFlags)
7192 return ConstantInt::get(Ty: ReturnType, V: true);
7193 if ((Mask & fcAllFlags) == 0)
7194 return ConstantInt::get(Ty: ReturnType, V: false);
7195 if (Q.isUndefValue(V: Op0))
7196 return UndefValue::get(T: ReturnType);
7197 break;
7198 }
7199 case Intrinsic::maxnum:
7200 case Intrinsic::minnum:
7201 case Intrinsic::maximum:
7202 case Intrinsic::minimum:
7203 case Intrinsic::maximumnum:
7204 case Intrinsic::minimumnum: {
7205 // In several cases here, we deviate from exact IEEE 754 semantics
7206 // to enable optimizations (as allowed by the LLVM IR spec).
7207 //
7208 // For instance, we may return one of the arguments unmodified instead of
7209 // inserting an llvm.canonicalize to transform input sNaNs into qNaNs,
7210 // or may assume all NaN inputs are qNaNs.
7211
7212 // If the arguments are the same, this is a no-op (ignoring NaN quieting)
7213 if (Op0 == Op1)
7214 return Op0;
7215
7216 // Canonicalize constant operand as Op1.
7217 if (isa<Constant>(Val: Op0))
7218 std::swap(a&: Op0, b&: Op1);
7219
7220 if (Constant *C = dyn_cast<Constant>(Val: Op1)) {
7221 MinMaxOptResult OptResult = MinMaxOptResult::CannotOptimize;
7222 Constant *NewConst = nullptr;
7223
7224 if (VectorType *VTy = dyn_cast<VectorType>(Val: C->getType())) {
7225 ElementCount ElemCount = VTy->getElementCount();
7226
7227 if (Constant *SplatVal = C->getSplatValue()) {
7228 // Handle splat vectors (including scalable vectors)
7229 OptResult = OptimizeConstMinMax(RHSConst: SplatVal, IID, FMF, OutNewConstVal: &NewConst);
7230 if (OptResult == MinMaxOptResult::UseNewConstVal)
7231 NewConst = ConstantVector::getSplat(EC: ElemCount, Elt: NewConst);
7232
7233 } else if (ElemCount.isFixed()) {
7234 // Storage to build up new const return value (with NaNs quieted)
7235 SmallVector<Constant *, 16> NewC(ElemCount.getFixedValue());
7236
7237 // Check elementwise whether we can optimize to either a constant
7238 // value or return the LHS value. We cannot mix and match LHS +
7239 // constant elements, as this would require inserting a new
7240 // VectorShuffle instruction, which is not allowed in simplifyBinOp.
7241 OptResult = MinMaxOptResult::UseEither;
7242 for (unsigned i = 0; i != ElemCount.getFixedValue(); ++i) {
7243 auto *Elt = C->getAggregateElement(Elt: i);
7244 if (!Elt) {
7245 OptResult = MinMaxOptResult::CannotOptimize;
7246 break;
7247 }
7248 auto ElemResult = OptimizeConstMinMax(RHSConst: Elt, IID, FMF, OutNewConstVal: &NewConst);
7249 if (ElemResult == MinMaxOptResult::CannotOptimize ||
7250 (ElemResult != OptResult &&
7251 OptResult != MinMaxOptResult::UseEither &&
7252 ElemResult != MinMaxOptResult::UseEither)) {
7253 OptResult = MinMaxOptResult::CannotOptimize;
7254 break;
7255 }
7256 NewC[i] = NewConst;
7257 if (ElemResult != MinMaxOptResult::UseEither)
7258 OptResult = ElemResult;
7259 }
7260 if (OptResult == MinMaxOptResult::UseNewConstVal)
7261 NewConst = ConstantVector::get(V: NewC);
7262 }
7263 } else {
7264 // Handle scalar inputs
7265 OptResult = OptimizeConstMinMax(RHSConst: C, IID, FMF, OutNewConstVal: &NewConst);
7266 }
7267
7268 if (OptResult == MinMaxOptResult::UseOtherVal ||
7269 OptResult == MinMaxOptResult::UseEither)
7270 return Op0; // Return the other arg (ignoring NaN quieting)
7271 else if (OptResult == MinMaxOptResult::UseNewConstVal)
7272 return NewConst;
7273 }
7274
7275 // Min/max of the same operation with common operand:
7276 // m(m(X, Y)), X --> m(X, Y) (4 commuted variants)
7277 if (Value *V = foldMinimumMaximumSharedOp(IID, Op0, Op1))
7278 return V;
7279 if (Value *V = foldMinimumMaximumSharedOp(IID, Op0: Op1, Op1: Op0))
7280 return V;
7281
7282 break;
7283 }
7284 case Intrinsic::vector_extract: {
7285 // (extract_vector (insert_vector _, X, 0), 0) -> X
7286 unsigned IdxN = cast<ConstantInt>(Val: Op1)->getZExtValue();
7287 Value *X = nullptr;
7288 if (match(V: Op0, P: m_Intrinsic<Intrinsic::vector_insert>(Ops: m_Value(), Ops: m_Value(V&: X),
7289 Ops: m_Zero())) &&
7290 IdxN == 0 && X->getType() == ReturnType)
7291 return X;
7292
7293 break;
7294 }
7295
7296 case Intrinsic::aarch64_sve_andv:
7297 case Intrinsic::aarch64_sve_eorv:
7298 case Intrinsic::aarch64_sve_orv:
7299 case Intrinsic::aarch64_sve_saddv:
7300 case Intrinsic::aarch64_sve_smaxv:
7301 case Intrinsic::aarch64_sve_sminv:
7302 case Intrinsic::aarch64_sve_uaddv:
7303 case Intrinsic::aarch64_sve_umaxv:
7304 case Intrinsic::aarch64_sve_uminv:
7305 return simplifySVEIntReduction(IID, ReturnType, Op0, Op1);
7306 default:
7307 break;
7308 }
7309
7310 return nullptr;
7311}
7312
7313/// interleaveN(extractvalue(deinterleaveN(x), 0), ...,
7314/// extractvalue(deinterleaveN(x), N-1)) --> x
7315static Value *simplifyIdentityInterleave(Intrinsic::ID IID,
7316 ArrayRef<Value *> Args) {
7317 unsigned Factor = getInterleaveIntrinsicFactor(ID: IID);
7318 if (!Factor || Factor != Args.size())
7319 return nullptr;
7320
7321 Intrinsic::ID DeinterleaveID = Intrinsic::getDeinterleaveIntrinsicID(Factor);
7322 IntrinsicInst *DI = nullptr;
7323 for (unsigned Idx = 0; Idx != Factor; ++Idx) {
7324 auto *EV = dyn_cast<ExtractValueInst>(Val: Args[Idx]);
7325 if (!EV || EV->getNumIndices() != 1 || *EV->idx_begin() != Idx)
7326 return nullptr;
7327
7328 auto *CurDI = dyn_cast<IntrinsicInst>(Val: EV->getAggregateOperand());
7329 if (!CurDI || CurDI->getIntrinsicID() != DeinterleaveID)
7330 return nullptr;
7331
7332 if (!DI)
7333 DI = CurDI;
7334 else if (DI != CurDI)
7335 return nullptr;
7336 }
7337
7338 return DI->getArgOperand(i: 0);
7339}
7340
7341Value *llvm::simplifyIntrinsic(Intrinsic::ID IID, Type *ReturnType,
7342 ArrayRef<Value *> Args, FastMathFlags FMF,
7343 const SimplifyQuery &Q, Function *CxtF,
7344 fp::ExceptionBehavior ExBehavior,
7345 RoundingMode Rounding) {
7346 unsigned NumOperands = Args.size();
7347 if (IID != Intrinsic::not_intrinsic && intrinsicPropagatesPoison(IID) &&
7348 any_of(Range&: Args, P: IsaPred<PoisonValue>))
7349 return PoisonValue::get(T: ReturnType);
7350
7351 // Defer to ConstantFolding if all args are constants.
7352 if (all_of(Range&: Args, P: IsaPred<Constant>))
7353 if (Constant *C = ConstantFoldIntrinsic(
7354 ID: IID, Ops: ArrayRef((Constant *const *)Args.data(), Args.size()),
7355 Ty: ReturnType, DL: Q.DL, CxtF))
7356 return C;
7357
7358 // Most of the intrinsics with no operands have some kind of side effect.
7359 // Don't simplify.
7360 if (!NumOperands) {
7361 switch (IID) {
7362 case Intrinsic::vscale: {
7363 if (!CxtF)
7364 return nullptr;
7365 ConstantRange CR = getVScaleRange(F: CxtF, BitWidth: 64);
7366 if (const APInt *C = CR.getSingleElement())
7367 return ConstantInt::get(Ty: ReturnType, V: C->getZExtValue());
7368 return nullptr;
7369 }
7370 default:
7371 return nullptr;
7372 }
7373 }
7374
7375 if (Value *V = simplifyIdentityInterleave(IID, Args))
7376 return V;
7377
7378 if (NumOperands == 1)
7379 return simplifyUnaryIntrinsic(IID, Op0: Args[0], FMF, Q);
7380
7381 if (NumOperands == 2)
7382 return simplifyBinaryIntrinsic(IID, ReturnType, Op0: Args[0], Op1: Args[1], FMF, Q);
7383
7384 // Handle intrinsics with 3 or more arguments.
7385 switch (IID) {
7386 case Intrinsic::masked_load:
7387 case Intrinsic::masked_gather: {
7388 Value *MaskArg = Args[1];
7389 Value *PassthruArg = Args[2];
7390 // If the mask is all zeros or poison, the "passthru" argument is the
7391 // result.
7392 if (match(V: MaskArg, P: m_ZeroOrPoison()))
7393 return PassthruArg;
7394 return nullptr;
7395 }
7396 case Intrinsic::fshl:
7397 case Intrinsic::fshr: {
7398 Value *Op0 = Args[0], *Op1 = Args[1], *ShAmtArg = Args[2];
7399
7400 // If both operands are undef, the result is undef.
7401 if (Q.isUndefValue(V: Op0) && Q.isUndefValue(V: Op1))
7402 return UndefValue::get(T: ReturnType);
7403
7404 // If shift amount is undef, assume it is zero.
7405 if (Q.isUndefValue(V: ShAmtArg))
7406 return Args[IID == Intrinsic::fshl ? 0 : 1];
7407
7408 const APInt *ShAmtC;
7409 if (match(V: ShAmtArg, P: m_APInt(Res&: ShAmtC))) {
7410 // If there's effectively no shift, return the 1st arg or 2nd arg.
7411 APInt BitWidth = APInt(ShAmtC->getBitWidth(), ShAmtC->getBitWidth());
7412 const APInt ShAmt = ShAmtC->urem(RHS: BitWidth);
7413 if (ShAmt.isZero())
7414 return Args[IID == Intrinsic::fshl ? 0 : 1];
7415
7416 // fshl (lshr X, C1), (shl X, C2), C1 -> X when C1 + C2 == BW
7417 // fshr (lshr X, C1), (shl X, C2), C2 -> X when C1 + C2 == BW
7418 const APInt *C1, *C2;
7419 Value *X;
7420 if (match(V: Op0, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1))) &&
7421 match(V: Op1, P: m_Shl(L: m_Specific(V: X), R: m_APInt(Res&: C2))) &&
7422 *C1 + *C2 == BitWidth && ShAmt == *(IID == Intrinsic::fshl ? C1 : C2))
7423 return X;
7424 }
7425
7426 // Rotating zero by anything is zero.
7427 if (match(V: Op0, P: m_Zero()) && match(V: Op1, P: m_Zero()))
7428 return ConstantInt::getNullValue(Ty: ReturnType);
7429
7430 // Rotating -1 by anything is -1.
7431 if (match(V: Op0, P: m_AllOnes()) && match(V: Op1, P: m_AllOnes()))
7432 return ConstantInt::getAllOnesValue(Ty: ReturnType);
7433
7434 return nullptr;
7435 }
7436 case Intrinsic::experimental_constrained_fma:
7437 return simplifyFPOp(Ops: Args, FMF: {}, Q, ExBehavior, Rounding);
7438 case Intrinsic::fma:
7439 case Intrinsic::fmuladd:
7440 return simplifyFPOp(Ops: Args, FMF: {}, Q, ExBehavior: fp::ebIgnore,
7441 Rounding: RoundingMode::NearestTiesToEven);
7442 case Intrinsic::smul_fix:
7443 case Intrinsic::smul_fix_sat: {
7444 Value *Op0 = Args[0];
7445 Value *Op1 = Args[1];
7446 Value *Op2 = Args[2];
7447
7448 // Canonicalize constant operand as Op1 (ConstantFolding handles the case
7449 // when both Op0 and Op1 are constant so we do not care about that special
7450 // case here).
7451 if (isa<Constant>(Val: Op0))
7452 std::swap(a&: Op0, b&: Op1);
7453
7454 // X * 0 -> 0
7455 if (match(V: Op1, P: m_Zero()))
7456 return Constant::getNullValue(Ty: ReturnType);
7457
7458 // X * undef -> 0
7459 if (Q.isUndefValue(V: Op1))
7460 return Constant::getNullValue(Ty: ReturnType);
7461
7462 // X * (1 << Scale) -> X
7463 APInt ScaledOne =
7464 APInt::getOneBitSet(numBits: ReturnType->getScalarSizeInBits(),
7465 BitNo: cast<ConstantInt>(Val: Op2)->getZExtValue());
7466 if (ScaledOne.isNonNegative() && match(V: Op1, P: m_SpecificInt(V: ScaledOne)))
7467 return Op0;
7468
7469 return nullptr;
7470 }
7471 case Intrinsic::vector_insert: {
7472 Value *Vec = Args[0];
7473 Value *SubVec = Args[1];
7474 Value *Idx = Args[2];
7475
7476 // (insert_vector Y, (extract_vector X, 0), 0) -> X
7477 // where: Y is X, or Y is undef
7478 unsigned IdxN = cast<ConstantInt>(Val: Idx)->getZExtValue();
7479 Value *X = nullptr;
7480 if (match(V: SubVec,
7481 P: m_Intrinsic<Intrinsic::vector_extract>(Ops: m_Value(V&: X), Ops: m_Zero())) &&
7482 (Q.isUndefValue(V: Vec) || Vec == X) && IdxN == 0 &&
7483 X->getType() == ReturnType)
7484 return X;
7485
7486 return nullptr;
7487 }
7488 case Intrinsic::vector_splice_right: {
7489 // splice.right(splice.left(poison, x, offset), poison, offset) -> x
7490 Value *X, *Offset = Args[2];
7491 if (match(V: Args[0], P: m_Intrinsic<Intrinsic::vector_splice_left>(
7492 Ops: m_Poison(), Ops: m_Value(V&: X), Ops: m_Specific(V: Offset))) &&
7493 isa<PoisonValue>(Val: Args[1]))
7494 return X;
7495 [[fallthrough]];
7496 }
7497 case Intrinsic::vector_splice_left: {
7498 Value *Offset = Args[2];
7499 auto *Ty = cast<VectorType>(Val: ReturnType);
7500 if (Q.isUndefValue(V: Offset))
7501 return PoisonValue::get(T: Ty);
7502
7503 unsigned BitWidth = Offset->getType()->getScalarSizeInBits();
7504 ConstantRange NumElts(
7505 APInt(BitWidth, Ty->getElementCount().getKnownMinValue()));
7506 if (Ty->isScalableTy())
7507 NumElts = NumElts.multiply(Other: CxtF ? getVScaleRange(F: CxtF, BitWidth)
7508 : ConstantRange::getFull(BitWidth));
7509
7510 // If we know Offset > NumElts, simplify to poison.
7511 ConstantRange CR = computeConstantRangeIncludingKnownBits(V: Offset, ForSigned: false, SQ: Q);
7512 if (CR.getUnsignedMin().ugt(RHS: NumElts.getUnsignedMax()))
7513 return PoisonValue::get(T: Ty);
7514
7515 // splice.left(a, b, 0) --> a, splice.right(a, b, 0) --> b
7516 if (CR.isSingleElement() && CR.getSingleElement()->isZero())
7517 return IID == Intrinsic::vector_splice_left ? Args[0] : Args[1];
7518
7519 return nullptr;
7520 }
7521 case Intrinsic::experimental_constrained_fadd:
7522 return simplifyFAddInst(Op0: Args[0], Op1: Args[1], FMF, Q, ExBehavior, Rounding);
7523 case Intrinsic::experimental_constrained_fsub:
7524 return simplifyFSubInst(Op0: Args[0], Op1: Args[1], FMF, Q, ExBehavior, Rounding);
7525 case Intrinsic::experimental_constrained_fmul:
7526 return simplifyFMulInst(Op0: Args[0], Op1: Args[1], FMF, Q, ExBehavior, Rounding);
7527 case Intrinsic::experimental_constrained_fdiv:
7528 return simplifyFDivInst(Op0: Args[0], Op1: Args[1], FMF, Q, ExBehavior, Rounding);
7529 case Intrinsic::experimental_constrained_frem:
7530 return simplifyFRemInst(Op0: Args[0], Op1: Args[1], FMF, Q, ExBehavior, Rounding);
7531 case Intrinsic::experimental_constrained_ldexp:
7532 return simplifyLdexp(Op0: Args[0], Op1: Args[1], Q, IsStrict: true);
7533 case Intrinsic::experimental_vp_reverse: {
7534 Value *Vec = Args[0];
7535 Value *EVL = Args[2];
7536
7537 Value *X;
7538 // vp.reverse(vp.reverse(X)) == X (mask doesn't matter)
7539 if (match(V: Vec, P: m_Intrinsic<Intrinsic::experimental_vp_reverse>(
7540 Ops: m_Value(V&: X), Ops: m_Value(), Ops: m_Specific(V: EVL))))
7541 return X;
7542
7543 // vp.reverse(splat(X)) -> splat(X) (regardless of mask and EVL)
7544 if (isSplatValue(V: Vec))
7545 return Vec;
7546 return nullptr;
7547 }
7548 default:
7549 return nullptr;
7550 }
7551}
7552
7553static Value *simplifyIntrinsic(CallBase *Call, ArrayRef<Value *> Args,
7554 const SimplifyQuery &Q) {
7555 // Operand bundles should not be in Args.
7556 assert(Call->arg_size() == Args.size());
7557 Intrinsic::ID IID = Call->getCalledFunction()->getIntrinsicID();
7558 Type *ReturnType = Call->getCalledFunction()->getReturnType();
7559
7560 switch (IID) {
7561 case Intrinsic::experimental_gc_relocate: {
7562 GCRelocateInst &GCR = *cast<GCRelocateInst>(Val: Call);
7563 Value *DerivedPtr = GCR.getDerivedPtr();
7564 Value *BasePtr = GCR.getBasePtr();
7565
7566 // Undef is undef, even after relocation.
7567 if (isa<UndefValue>(Val: DerivedPtr) || isa<UndefValue>(Val: BasePtr)) {
7568 return UndefValue::get(T: GCR.getType());
7569 }
7570
7571 if (auto *PT = dyn_cast<PointerType>(Val: GCR.getType())) {
7572 // For now, the assumption is that the relocation of null will be null
7573 // for most any collector. If this ever changes, a corresponding hook
7574 // should be added to GCStrategy and this code should check it first.
7575 if (isa<ConstantPointerNull>(Val: DerivedPtr)) {
7576 // Use null-pointer of gc_relocate's type to replace it.
7577 return ConstantPointerNull::get(T: PT);
7578 }
7579 }
7580 return nullptr;
7581 }
7582 default: {
7583 // Use the default FP environment if none is found.
7584 fp::ExceptionBehavior ExBehavior = fp::ebIgnore;
7585 RoundingMode Rounding = RoundingMode::NearestTiesToEven;
7586 if (auto *Constrained = dyn_cast<ConstrainedFPIntrinsic>(Val: Call)) {
7587 ExBehavior = Constrained->getExceptionBehavior().value_or(u&: ExBehavior);
7588 Rounding = Constrained->getRoundingMode().value_or(u&: Rounding);
7589 }
7590 return simplifyIntrinsic(IID, ReturnType, Args,
7591 FMF: Call->getFastMathFlagsOrNone(), Q,
7592 CxtF: Call->getFunction(), ExBehavior, Rounding);
7593 }
7594 }
7595}
7596
7597static Value *tryConstantFoldCall(CallBase *Call, ArrayRef<Value *> Args,
7598 const SimplifyQuery &Q) {
7599 auto *F = Call->getCalledFunction();
7600 if (!F || !canConstantFoldCallTo(Call, F))
7601 return nullptr;
7602
7603 SmallVector<Constant *, 4> ConstantArgs;
7604 ConstantArgs.reserve(N: Args.size());
7605 for (Value *Arg : Args) {
7606 Constant *C = dyn_cast<Constant>(Val: Arg);
7607 if (!C) {
7608 if (isa<MetadataAsValue>(Val: Arg))
7609 continue;
7610 return nullptr;
7611 }
7612 ConstantArgs.push_back(Elt: C);
7613 }
7614
7615 return ConstantFoldCall(Call, F, Operands: ConstantArgs, TLI: Q.TLI);
7616}
7617
7618Value *llvm::simplifyCall(CallBase *Call, Value *Callee, ArrayRef<Value *> Args,
7619 const SimplifyQuery &Q) {
7620 // Args should not contain operand bundle operands.
7621 assert(Call->arg_size() == Args.size());
7622
7623 // musttail calls can only be simplified if they are also DCEd.
7624 // As we can't guarantee this here, don't simplify them.
7625 if (Call->isMustTailCall())
7626 return nullptr;
7627
7628 // call undef -> poison
7629 // call null -> poison
7630 if (isa<UndefValue>(Val: Callee) || isa<ConstantPointerNull>(Val: Callee))
7631 return PoisonValue::get(T: Call->getType());
7632
7633 if (Value *V = tryConstantFoldCall(Call, Args, Q))
7634 return V;
7635
7636 auto *F = dyn_cast<Function>(Val: Callee);
7637 if (F && F->isIntrinsic())
7638 if (Value *Ret = ::simplifyIntrinsic(Call, Args, Q))
7639 return Ret;
7640
7641 return nullptr;
7642}
7643
7644Value *llvm::simplifyConstrainedFPCall(CallBase *Call, const SimplifyQuery &Q) {
7645 assert(isa<ConstrainedFPIntrinsic>(Call));
7646 SmallVector<Value *, 4> Args(Call->args());
7647 if (Value *V = tryConstantFoldCall(Call, Args, Q))
7648 return V;
7649 if (Value *Ret = ::simplifyIntrinsic(Call, Args, Q))
7650 return Ret;
7651 return nullptr;
7652}
7653
7654/// Given operands for a Freeze, see if we can fold the result.
7655static Value *simplifyFreezeInst(Value *Op0, const SimplifyQuery &Q) {
7656 // Use a utility function defined in ValueTracking.
7657 if (llvm::isGuaranteedNotToBeUndefOrPoison(V: Op0, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT))
7658 return Op0;
7659 // We have room for improvement.
7660 return nullptr;
7661}
7662
7663Value *llvm::simplifyFreezeInst(Value *Op0, const SimplifyQuery &Q) {
7664 return ::simplifyFreezeInst(Op0, Q);
7665}
7666
7667Value *llvm::simplifyLoadInst(LoadInst *LI, Value *PtrOp,
7668 const SimplifyQuery &Q) {
7669 if (LI->isVolatile())
7670 return nullptr;
7671
7672 if (auto *PtrOpC = dyn_cast<Constant>(Val: PtrOp))
7673 return ConstantFoldLoadFromConstPtr(C: PtrOpC, Ty: LI->getType(), DL: Q.DL);
7674
7675 // We can only fold the load if it is from a constant global with definitive
7676 // initializer. Skip expensive logic if this is not the case.
7677 auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: PtrOp));
7678 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
7679 return nullptr;
7680
7681 // If GlobalVariable's initializer is uniform, then return the constant
7682 // regardless of its offset.
7683 if (Constant *C = ConstantFoldLoadFromUniformValue(C: GV->getInitializer(),
7684 Ty: LI->getType(), DL: Q.DL))
7685 return C;
7686
7687 // Try to convert operand into a constant by stripping offsets while looking
7688 // through invariant.group intrinsics.
7689 APInt Offset(Q.DL.getIndexTypeSizeInBits(Ty: PtrOp->getType()), 0);
7690 PtrOp = PtrOp->stripAndAccumulateConstantOffsets(
7691 DL: Q.DL, Offset, /* AllowNonInbounts */ AllowNonInbounds: true,
7692 /* AllowInvariantGroup */ true);
7693 if (PtrOp == GV) {
7694 // Index size may have changed due to address space casts.
7695 Offset = Offset.sextOrTrunc(width: Q.DL.getIndexTypeSizeInBits(Ty: PtrOp->getType()));
7696 return ConstantFoldLoadFromConstPtr(C: GV, Ty: LI->getType(), Offset: std::move(Offset),
7697 DL: Q.DL);
7698 }
7699
7700 return nullptr;
7701}
7702
7703/// See if we can compute a simplified version of this instruction.
7704/// If not, this returns null.
7705
7706static Value *simplifyInstructionWithOperands(Instruction *I,
7707 ArrayRef<Value *> NewOps,
7708 const SimplifyQuery &SQ,
7709 unsigned MaxRecurse) {
7710 assert(I->getFunction() && "instruction should be inserted in a function");
7711 assert((!SQ.CxtI || SQ.CxtI->getFunction() == I->getFunction()) &&
7712 "context instruction should be in the same function");
7713
7714 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
7715
7716 switch (I->getOpcode()) {
7717 default:
7718 if (all_of(Range&: NewOps, P: IsaPred<Constant>)) {
7719 SmallVector<Constant *, 8> NewConstOps(NewOps.size());
7720 transform(Range&: NewOps, d_first: NewConstOps.begin(),
7721 F: [](Value *V) { return cast<Constant>(Val: V); });
7722 return ConstantFoldInstOperands(I, Ops: NewConstOps, DL: Q.DL, TLI: Q.TLI);
7723 }
7724 return nullptr;
7725 case Instruction::FNeg:
7726 return simplifyFNegInst(Op: NewOps[0], FMF: I->getFastMathFlags(), Q, MaxRecurse);
7727 case Instruction::FAdd:
7728 return simplifyFAddInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7729 MaxRecurse);
7730 case Instruction::Add:
7731 return simplifyAddInst(
7732 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7733 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7734 case Instruction::FSub:
7735 return simplifyFSubInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7736 MaxRecurse);
7737 case Instruction::Sub:
7738 return simplifySubInst(
7739 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7740 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7741 case Instruction::FMul:
7742 return simplifyFMulInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7743 MaxRecurse);
7744 case Instruction::Mul:
7745 return simplifyMulInst(
7746 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7747 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7748 case Instruction::SDiv:
7749 return simplifySDivInst(Op0: NewOps[0], Op1: NewOps[1],
7750 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7751 MaxRecurse);
7752 case Instruction::UDiv:
7753 return simplifyUDivInst(Op0: NewOps[0], Op1: NewOps[1],
7754 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7755 MaxRecurse);
7756 case Instruction::FDiv:
7757 return simplifyFDivInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7758 MaxRecurse);
7759 case Instruction::SRem:
7760 return simplifySRemInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7761 case Instruction::URem:
7762 return simplifyURemInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7763 case Instruction::FRem:
7764 return simplifyFRemInst(Op0: NewOps[0], Op1: NewOps[1], FMF: I->getFastMathFlags(), Q,
7765 MaxRecurse);
7766 case Instruction::Shl:
7767 return simplifyShlInst(
7768 Op0: NewOps[0], Op1: NewOps[1], IsNSW: Q.IIQ.hasNoSignedWrap(Op: cast<BinaryOperator>(Val: I)),
7769 IsNUW: Q.IIQ.hasNoUnsignedWrap(Op: cast<BinaryOperator>(Val: I)), Q, MaxRecurse);
7770 case Instruction::LShr:
7771 return simplifyLShrInst(Op0: NewOps[0], Op1: NewOps[1],
7772 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7773 MaxRecurse);
7774 case Instruction::AShr:
7775 return simplifyAShrInst(Op0: NewOps[0], Op1: NewOps[1],
7776 IsExact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)), Q,
7777 MaxRecurse);
7778 case Instruction::And:
7779 return simplifyAndInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7780 case Instruction::Or:
7781 return simplifyOrInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7782 case Instruction::Xor:
7783 return simplifyXorInst(Op0: NewOps[0], Op1: NewOps[1], Q, MaxRecurse);
7784 case Instruction::ICmp:
7785 return simplifyICmpInst(Pred: cast<ICmpInst>(Val: I)->getCmpPredicate(), LHS: NewOps[0],
7786 RHS: NewOps[1], Q, MaxRecurse);
7787 case Instruction::FCmp:
7788 return simplifyFCmpInst(Pred: cast<FCmpInst>(Val: I)->getPredicate(), LHS: NewOps[0],
7789 RHS: NewOps[1], FMF: I->getFastMathFlags(), Q, MaxRecurse);
7790 case Instruction::Select: {
7791 FastMathFlags FMF;
7792 if (auto *FPMO = dyn_cast<FPMathOperator>(Val: I))
7793 FMF = FPMO->getFastMathFlags();
7794 return simplifySelectInst(Cond: NewOps[0], TrueVal: NewOps[1], FalseVal: NewOps[2], FMF, Q,
7795 MaxRecurse);
7796 }
7797 case Instruction::GetElementPtr: {
7798 auto *GEPI = cast<GetElementPtrInst>(Val: I);
7799 return simplifyGEPInst(SrcTy: GEPI->getSourceElementType(), Ptr: NewOps[0],
7800 Indices: ArrayRef(NewOps).slice(N: 1), NW: GEPI->getNoWrapFlags(), Q,
7801 MaxRecurse);
7802 }
7803 case Instruction::InsertValue: {
7804 InsertValueInst *IV = cast<InsertValueInst>(Val: I);
7805 return simplifyInsertValueInst(Agg: NewOps[0], Val: NewOps[1], Idxs: IV->getIndices(), Q,
7806 MaxRecurse);
7807 }
7808 case Instruction::InsertElement:
7809 return simplifyInsertElementInst(Vec: NewOps[0], Val: NewOps[1], Idx: NewOps[2], Q);
7810 case Instruction::ExtractValue: {
7811 auto *EVI = cast<ExtractValueInst>(Val: I);
7812 return simplifyExtractValueInst(Agg: NewOps[0], Idxs: EVI->getIndices(), Q,
7813 MaxRecurse);
7814 }
7815 case Instruction::ExtractElement:
7816 return simplifyExtractElementInst(Vec: NewOps[0], Idx: NewOps[1], Q, MaxRecurse);
7817 case Instruction::ShuffleVector: {
7818 auto *SVI = cast<ShuffleVectorInst>(Val: I);
7819 return simplifyShuffleVectorInst(Op0: NewOps[0], Op1: NewOps[1],
7820 Mask: SVI->getShuffleMask(), RetTy: SVI->getType(), Q,
7821 MaxRecurse);
7822 }
7823 case Instruction::PHI:
7824 return simplifyPHINode(PN: cast<PHINode>(Val: I), IncomingValues: NewOps, Q);
7825 case Instruction::Call:
7826 return simplifyCall(
7827 Call: cast<CallInst>(Val: I), Callee: NewOps.back(),
7828 Args: NewOps.drop_back(N: 1 + cast<CallInst>(Val: I)->getNumTotalBundleOperands()), Q);
7829 case Instruction::Freeze:
7830 return llvm::simplifyFreezeInst(Op0: NewOps[0], Q);
7831#define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
7832#include "llvm/IR/Instruction.def"
7833#undef HANDLE_CAST_INST
7834 return simplifyCastInst(CastOpc: I->getOpcode(), Op: NewOps[0], Ty: I->getType(), Q,
7835 MaxRecurse);
7836 case Instruction::Alloca:
7837 // No simplifications for Alloca and it can't be constant folded.
7838 return nullptr;
7839 case Instruction::Load:
7840 return simplifyLoadInst(LI: cast<LoadInst>(Val: I), PtrOp: NewOps[0], Q);
7841 }
7842}
7843
7844Value *llvm::simplifyInstructionWithOperands(Instruction *I,
7845 ArrayRef<Value *> NewOps,
7846 const SimplifyQuery &SQ) {
7847 assert(NewOps.size() == I->getNumOperands() &&
7848 "Number of operands should match the instruction!");
7849 return ::simplifyInstructionWithOperands(I, NewOps, SQ, MaxRecurse: RecursionLimit);
7850}
7851
7852Value *llvm::simplifyInstruction(Instruction *I, const SimplifyQuery &SQ) {
7853 SmallVector<Value *, 8> Ops(I->operands());
7854 Value *Result = ::simplifyInstructionWithOperands(I, NewOps: Ops, SQ, MaxRecurse: RecursionLimit);
7855
7856 /// If called on unreachable code, the instruction may simplify to itself.
7857 /// Make life easier for users by detecting that case here, and returning a
7858 /// safe value instead.
7859 return Result == I ? PoisonValue::get(T: I->getType()) : Result;
7860}
7861
7862/// Implementation of recursive simplification through an instruction's
7863/// uses.
7864///
7865/// This is the common implementation of the recursive simplification routines.
7866/// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
7867/// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
7868/// instructions to process and attempt to simplify it using
7869/// InstructionSimplify. Recursively visited users which could not be
7870/// simplified themselves are to the optional UnsimplifiedUsers set for
7871/// further processing by the caller.
7872///
7873/// This routine returns 'true' only when *it* simplifies something. The passed
7874/// in simplified value does not count toward this.
7875static bool replaceAndRecursivelySimplifyImpl(
7876 Instruction *I, Value *SimpleV, const TargetLibraryInfo *TLI,
7877 const DominatorTree *DT, AssumptionCache *AC,
7878 SmallSetVector<Instruction *, 8> *UnsimplifiedUsers = nullptr) {
7879 bool Simplified = false;
7880 SmallSetVector<Instruction *, 8> Worklist;
7881 const DataLayout &DL = I->getDataLayout();
7882
7883 // If we have an explicit value to collapse to, do that round of the
7884 // simplification loop by hand initially.
7885 if (SimpleV) {
7886 for (User *U : I->users())
7887 if (U != I)
7888 Worklist.insert(X: cast<Instruction>(Val: U));
7889
7890 // Replace the instruction with its simplified value.
7891 I->replaceAllUsesWith(V: SimpleV);
7892
7893 if (!I->isEHPad() && !I->isTerminator() && !I->mayHaveSideEffects())
7894 I->eraseFromParent();
7895 } else {
7896 Worklist.insert(X: I);
7897 }
7898
7899 // Note that we must test the size on each iteration, the worklist can grow.
7900 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
7901 I = Worklist[Idx];
7902
7903 // See if this instruction simplifies.
7904 SimpleV = simplifyInstruction(I, SQ: {DL, TLI, DT, AC});
7905 if (!SimpleV) {
7906 if (UnsimplifiedUsers)
7907 UnsimplifiedUsers->insert(X: I);
7908 continue;
7909 }
7910
7911 Simplified = true;
7912
7913 // Stash away all the uses of the old instruction so we can check them for
7914 // recursive simplifications after a RAUW. This is cheaper than checking all
7915 // uses of To on the recursive step in most cases.
7916 for (User *U : I->users())
7917 Worklist.insert(X: cast<Instruction>(Val: U));
7918
7919 // Replace the instruction with its simplified value.
7920 I->replaceAllUsesWith(V: SimpleV);
7921
7922 if (!I->isEHPad() && !I->isTerminator() && !I->mayHaveSideEffects())
7923 I->eraseFromParent();
7924 }
7925 return Simplified;
7926}
7927
7928bool llvm::replaceAndRecursivelySimplify(
7929 Instruction *I, Value *SimpleV, const TargetLibraryInfo *TLI,
7930 const DominatorTree *DT, AssumptionCache *AC,
7931 SmallSetVector<Instruction *, 8> *UnsimplifiedUsers) {
7932 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
7933 assert(SimpleV && "Must provide a simplified value.");
7934 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC,
7935 UnsimplifiedUsers);
7936}
7937
7938namespace llvm {
7939const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
7940 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
7941 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
7942 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
7943 auto *TLI = TLIWP ? &TLIWP->getTLI(F) : nullptr;
7944 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
7945 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
7946 return {F.getDataLayout(), TLI, DT, AC};
7947}
7948
7949const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
7950 const DataLayout &DL) {
7951 return {DL, &AR.TLI, &AR.DT, &AR.AC};
7952}
7953
7954template <class T, class... TArgs>
7955const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
7956 Function &F) {
7957 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
7958 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
7959 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
7960 return {F.getDataLayout(), TLI, DT, AC};
7961}
7962template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,
7963 Function &);
7964
7965bool SimplifyQuery::isUndefValue(Value *V) const {
7966 if (!CanUseUndef)
7967 return false;
7968
7969 return match(V, P: m_Undef());
7970}
7971
7972} // namespace llvm
7973
7974void InstSimplifyFolder::anchor() {}
7975