LLVM  14.0.0git
InstructionSimplify.cpp
Go to the documentation of this file.
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 
20 
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/Statistic.h"
35 #include "llvm/IR/ConstantRange.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/GlobalAlias.h"
40 #include "llvm/IR/InstrTypes.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Operator.h"
43 #include "llvm/IR/PatternMatch.h"
44 #include "llvm/IR/ValueHandle.h"
45 #include "llvm/Support/KnownBits.h"
46 #include <algorithm>
47 using namespace llvm;
48 using namespace llvm::PatternMatch;
49 
50 #define DEBUG_TYPE "instsimplify"
51 
52 enum { RecursionLimit = 3 };
53 
54 STATISTIC(NumExpand, "Number of expansions");
55 STATISTIC(NumReassoc, "Number of reassociations");
56 
57 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
58 static Value *simplifyUnOp(unsigned, Value *, const SimplifyQuery &, unsigned);
59 static Value *simplifyFPUnOp(unsigned, Value *, const FastMathFlags &,
60  const SimplifyQuery &, unsigned);
61 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
62  unsigned);
63 static Value *SimplifyBinOp(unsigned, Value *, Value *, const FastMathFlags &,
64  const SimplifyQuery &, unsigned);
65 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
66  unsigned);
67 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
68  const SimplifyQuery &Q, unsigned MaxRecurse);
69 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
70 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
71 static Value *SimplifyCastInst(unsigned, Value *, Type *,
72  const SimplifyQuery &, unsigned);
74  const SimplifyQuery &, unsigned);
75 static Value *SimplifySelectInst(Value *, Value *, Value *,
76  const SimplifyQuery &, unsigned);
77 
79  Value *FalseVal) {
80  BinaryOperator::BinaryOps BinOpCode;
81  if (auto *BO = dyn_cast<BinaryOperator>(Cond))
82  BinOpCode = BO->getOpcode();
83  else
84  return nullptr;
85 
86  CmpInst::Predicate ExpectedPred, Pred1, Pred2;
87  if (BinOpCode == BinaryOperator::Or) {
88  ExpectedPred = ICmpInst::ICMP_NE;
89  } else if (BinOpCode == BinaryOperator::And) {
90  ExpectedPred = ICmpInst::ICMP_EQ;
91  } else
92  return nullptr;
93 
94  // %A = icmp eq %TV, %FV
95  // %B = icmp eq %X, %Y (and one of these is a select operand)
96  // %C = and %A, %B
97  // %D = select %C, %TV, %FV
98  // -->
99  // %FV
100 
101  // %A = icmp ne %TV, %FV
102  // %B = icmp ne %X, %Y (and one of these is a select operand)
103  // %C = or %A, %B
104  // %D = select %C, %TV, %FV
105  // -->
106  // %TV
107  Value *X, *Y;
110  m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
111  Pred1 != Pred2 || Pred1 != ExpectedPred)
112  return nullptr;
113 
114  if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
115  return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
116 
117  return nullptr;
118 }
119 
120 /// For a boolean type or a vector of boolean type, return false or a vector
121 /// with every element false.
122 static Constant *getFalse(Type *Ty) {
123  return ConstantInt::getFalse(Ty);
124 }
125 
126 /// For a boolean type or a vector of boolean type, return true or a vector
127 /// with every element true.
128 static Constant *getTrue(Type *Ty) {
129  return ConstantInt::getTrue(Ty);
130 }
131 
132 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
133 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
134  Value *RHS) {
135  CmpInst *Cmp = dyn_cast<CmpInst>(V);
136  if (!Cmp)
137  return false;
138  CmpInst::Predicate CPred = Cmp->getPredicate();
139  Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
140  if (CPred == Pred && CLHS == LHS && CRHS == RHS)
141  return true;
142  return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
143  CRHS == LHS;
144 }
145 
146 /// Simplify comparison with true or false branch of select:
147 /// %sel = select i1 %cond, i32 %tv, i32 %fv
148 /// %cmp = icmp sle i32 %sel, %rhs
149 /// Compose new comparison by substituting %sel with either %tv or %fv
150 /// and see if it simplifies.
152  Value *RHS, Value *Cond,
153  const SimplifyQuery &Q, unsigned MaxRecurse,
154  Constant *TrueOrFalse) {
155  Value *SimplifiedCmp = SimplifyCmpInst(Pred, LHS, RHS, Q, MaxRecurse);
156  if (SimplifiedCmp == Cond) {
157  // %cmp simplified to the select condition (%cond).
158  return TrueOrFalse;
159  } else if (!SimplifiedCmp && isSameCompare(Cond, Pred, LHS, RHS)) {
160  // It didn't simplify. However, if composed comparison is equivalent
161  // to the select condition (%cond) then we can replace it.
162  return TrueOrFalse;
163  }
164  return SimplifiedCmp;
165 }
166 
167 /// Simplify comparison with true branch of select
169  Value *RHS, Value *Cond,
170  const SimplifyQuery &Q,
171  unsigned MaxRecurse) {
172  return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
173  getTrue(Cond->getType()));
174 }
175 
176 /// Simplify comparison with false branch of select
178  Value *RHS, Value *Cond,
179  const SimplifyQuery &Q,
180  unsigned MaxRecurse) {
181  return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
182  getFalse(Cond->getType()));
183 }
184 
185 /// We know comparison with both branches of select can be simplified, but they
186 /// are not equal. This routine handles some logical simplifications.
188  Value *Cond,
189  const SimplifyQuery &Q,
190  unsigned MaxRecurse) {
191  // If the false value simplified to false, then the result of the compare
192  // is equal to "Cond && TCmp". This also catches the case when the false
193  // value simplified to false and the true value to true, returning "Cond".
194  // Folding select to and/or isn't poison-safe in general; impliesPoison
195  // checks whether folding it does not convert a well-defined value into
196  // poison.
197  if (match(FCmp, m_Zero()) && impliesPoison(TCmp, Cond))
198  if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
199  return V;
200  // If the true value simplified to true, then the result of the compare
201  // is equal to "Cond || FCmp".
202  if (match(TCmp, m_One()) && impliesPoison(FCmp, Cond))
203  if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
204  return V;
205  // Finally, if the false value simplified to true and the true value to
206  // false, then the result of the compare is equal to "!Cond".
207  if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
208  if (Value *V = SimplifyXorInst(
209  Cond, Constant::getAllOnesValue(Cond->getType()), Q, MaxRecurse))
210  return V;
211  return nullptr;
212 }
213 
214 /// Does the given value dominate the specified phi node?
215 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
216  Instruction *I = dyn_cast<Instruction>(V);
217  if (!I)
218  // Arguments and constants dominate all instructions.
219  return true;
220 
221  // If we are processing instructions (and/or basic blocks) that have not been
222  // fully added to a function, the parent nodes may still be null. Simply
223  // return the conservative answer in these cases.
224  if (!I->getParent() || !P->getParent() || !I->getFunction())
225  return false;
226 
227  // If we have a DominatorTree then do a precise test.
228  if (DT)
229  return DT->dominates(I, P);
230 
231  // Otherwise, if the instruction is in the entry block and is not an invoke,
232  // then it obviously dominates all phi nodes.
233  if (I->getParent()->isEntryBlock() && !isa<InvokeInst>(I) &&
234  !isa<CallBrInst>(I))
235  return true;
236 
237  return false;
238 }
239 
240 /// Try to simplify a binary operator of form "V op OtherOp" where V is
241 /// "(B0 opex B1)" by distributing 'op' across 'opex' as
242 /// "(B0 op OtherOp) opex (B1 op OtherOp)".
244  Value *OtherOp, Instruction::BinaryOps OpcodeToExpand,
245  const SimplifyQuery &Q, unsigned MaxRecurse) {
246  auto *B = dyn_cast<BinaryOperator>(V);
247  if (!B || B->getOpcode() != OpcodeToExpand)
248  return nullptr;
249  Value *B0 = B->getOperand(0), *B1 = B->getOperand(1);
250  Value *L = SimplifyBinOp(Opcode, B0, OtherOp, Q.getWithoutUndef(),
251  MaxRecurse);
252  if (!L)
253  return nullptr;
254  Value *R = SimplifyBinOp(Opcode, B1, OtherOp, Q.getWithoutUndef(),
255  MaxRecurse);
256  if (!R)
257  return nullptr;
258 
259  // Does the expanded pair of binops simplify to the existing binop?
260  if ((L == B0 && R == B1) ||
261  (Instruction::isCommutative(OpcodeToExpand) && L == B1 && R == B0)) {
262  ++NumExpand;
263  return B;
264  }
265 
266  // Otherwise, return "L op' R" if it simplifies.
267  Value *S = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse);
268  if (!S)
269  return nullptr;
270 
271  ++NumExpand;
272  return S;
273 }
274 
275 /// Try to simplify binops of form "A op (B op' C)" or the commuted variant by
276 /// distributing op over op'.
278  Value *L, Value *R,
279  Instruction::BinaryOps OpcodeToExpand,
280  const SimplifyQuery &Q,
281  unsigned MaxRecurse) {
282  // Recursion is always used, so bail out at once if we already hit the limit.
283  if (!MaxRecurse--)
284  return nullptr;
285 
286  if (Value *V = expandBinOp(Opcode, L, R, OpcodeToExpand, Q, MaxRecurse))
287  return V;
288  if (Value *V = expandBinOp(Opcode, R, L, OpcodeToExpand, Q, MaxRecurse))
289  return V;
290  return nullptr;
291 }
292 
293 /// Generic simplifications for associative binary operations.
294 /// Returns the simpler value, or null if none was found.
296  Value *LHS, Value *RHS,
297  const SimplifyQuery &Q,
298  unsigned MaxRecurse) {
299  assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
300 
301  // Recursion is always used, so bail out at once if we already hit the limit.
302  if (!MaxRecurse--)
303  return nullptr;
304 
305  BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
306  BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
307 
308  // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
309  if (Op0 && Op0->getOpcode() == Opcode) {
310  Value *A = Op0->getOperand(0);
311  Value *B = Op0->getOperand(1);
312  Value *C = RHS;
313 
314  // Does "B op C" simplify?
315  if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
316  // It does! Return "A op V" if it simplifies or is already available.
317  // If V equals B then "A op V" is just the LHS.
318  if (V == B) return LHS;
319  // Otherwise return "A op V" if it simplifies.
320  if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
321  ++NumReassoc;
322  return W;
323  }
324  }
325  }
326 
327  // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
328  if (Op1 && Op1->getOpcode() == Opcode) {
329  Value *A = LHS;
330  Value *B = Op1->getOperand(0);
331  Value *C = Op1->getOperand(1);
332 
333  // Does "A op B" simplify?
334  if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
335  // It does! Return "V op C" if it simplifies or is already available.
336  // If V equals B then "V op C" is just the RHS.
337  if (V == B) return RHS;
338  // Otherwise return "V op C" if it simplifies.
339  if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
340  ++NumReassoc;
341  return W;
342  }
343  }
344  }
345 
346  // The remaining transforms require commutativity as well as associativity.
347  if (!Instruction::isCommutative(Opcode))
348  return nullptr;
349 
350  // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
351  if (Op0 && Op0->getOpcode() == Opcode) {
352  Value *A = Op0->getOperand(0);
353  Value *B = Op0->getOperand(1);
354  Value *C = RHS;
355 
356  // Does "C op A" simplify?
357  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
358  // It does! Return "V op B" if it simplifies or is already available.
359  // If V equals A then "V op B" is just the LHS.
360  if (V == A) return LHS;
361  // Otherwise return "V op B" if it simplifies.
362  if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
363  ++NumReassoc;
364  return W;
365  }
366  }
367  }
368 
369  // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
370  if (Op1 && Op1->getOpcode() == Opcode) {
371  Value *A = LHS;
372  Value *B = Op1->getOperand(0);
373  Value *C = Op1->getOperand(1);
374 
375  // Does "C op A" simplify?
376  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
377  // It does! Return "B op V" if it simplifies or is already available.
378  // If V equals C then "B op V" is just the RHS.
379  if (V == C) return RHS;
380  // Otherwise return "B op V" if it simplifies.
381  if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
382  ++NumReassoc;
383  return W;
384  }
385  }
386  }
387 
388  return nullptr;
389 }
390 
391 /// In the case of a binary operation with a select instruction as an operand,
392 /// try to simplify the binop by seeing whether evaluating it on both branches
393 /// of the select results in the same value. Returns the common value if so,
394 /// otherwise returns null.
396  Value *RHS, const SimplifyQuery &Q,
397  unsigned MaxRecurse) {
398  // Recursion is always used, so bail out at once if we already hit the limit.
399  if (!MaxRecurse--)
400  return nullptr;
401 
402  SelectInst *SI;
403  if (isa<SelectInst>(LHS)) {
404  SI = cast<SelectInst>(LHS);
405  } else {
406  assert(isa<SelectInst>(RHS) && "No select instruction operand!");
407  SI = cast<SelectInst>(RHS);
408  }
409 
410  // Evaluate the BinOp on the true and false branches of the select.
411  Value *TV;
412  Value *FV;
413  if (SI == LHS) {
414  TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
415  FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
416  } else {
417  TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
418  FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
419  }
420 
421  // If they simplified to the same value, then return the common value.
422  // If they both failed to simplify then return null.
423  if (TV == FV)
424  return TV;
425 
426  // If one branch simplified to undef, return the other one.
427  if (TV && Q.isUndefValue(TV))
428  return FV;
429  if (FV && Q.isUndefValue(FV))
430  return TV;
431 
432  // If applying the operation did not change the true and false select values,
433  // then the result of the binop is the select itself.
434  if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
435  return SI;
436 
437  // If one branch simplified and the other did not, and the simplified
438  // value is equal to the unsimplified one, return the simplified value.
439  // For example, select (cond, X, X & Z) & Z -> X & Z.
440  if ((FV && !TV) || (TV && !FV)) {
441  // Check that the simplified value has the form "X op Y" where "op" is the
442  // same as the original operation.
443  Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
444  if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
445  // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
446  // We already know that "op" is the same as for the simplified value. See
447  // if the operands match too. If so, return the simplified value.
448  Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
449  Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
450  Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
451  if (Simplified->getOperand(0) == UnsimplifiedLHS &&
452  Simplified->getOperand(1) == UnsimplifiedRHS)
453  return Simplified;
454  if (Simplified->isCommutative() &&
455  Simplified->getOperand(1) == UnsimplifiedLHS &&
456  Simplified->getOperand(0) == UnsimplifiedRHS)
457  return Simplified;
458  }
459  }
460 
461  return nullptr;
462 }
463 
464 /// In the case of a comparison with a select instruction, try to simplify the
465 /// comparison by seeing whether both branches of the select result in the same
466 /// value. Returns the common value if so, otherwise returns null.
467 /// For example, if we have:
468 /// %tmp = select i1 %cmp, i32 1, i32 2
469 /// %cmp1 = icmp sle i32 %tmp, 3
470 /// We can simplify %cmp1 to true, because both branches of select are
471 /// less than 3. We compose new comparison by substituting %tmp with both
472 /// branches of select and see if it can be simplified.
474  Value *RHS, const SimplifyQuery &Q,
475  unsigned MaxRecurse) {
476  // Recursion is always used, so bail out at once if we already hit the limit.
477  if (!MaxRecurse--)
478  return nullptr;
479 
480  // Make sure the select is on the LHS.
481  if (!isa<SelectInst>(LHS)) {
482  std::swap(LHS, RHS);
483  Pred = CmpInst::getSwappedPredicate(Pred);
484  }
485  assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
486  SelectInst *SI = cast<SelectInst>(LHS);
487  Value *Cond = SI->getCondition();
488  Value *TV = SI->getTrueValue();
489  Value *FV = SI->getFalseValue();
490 
491  // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
492  // Does "cmp TV, RHS" simplify?
493  Value *TCmp = simplifyCmpSelTrueCase(Pred, TV, RHS, Cond, Q, MaxRecurse);
494  if (!TCmp)
495  return nullptr;
496 
497  // Does "cmp FV, RHS" simplify?
498  Value *FCmp = simplifyCmpSelFalseCase(Pred, FV, RHS, Cond, Q, MaxRecurse);
499  if (!FCmp)
500  return nullptr;
501 
502  // If both sides simplified to the same value, then use it as the result of
503  // the original comparison.
504  if (TCmp == FCmp)
505  return TCmp;
506 
507  // The remaining cases only make sense if the select condition has the same
508  // type as the result of the comparison, so bail out if this is not so.
509  if (Cond->getType()->isVectorTy() == RHS->getType()->isVectorTy())
510  return handleOtherCmpSelSimplifications(TCmp, FCmp, Cond, Q, MaxRecurse);
511 
512  return nullptr;
513 }
514 
515 /// In the case of a binary operation with an operand that is a PHI instruction,
516 /// try to simplify the binop by seeing whether evaluating it on the incoming
517 /// phi values yields the same result for every value. If so returns the common
518 /// value, otherwise returns null.
520  Value *RHS, const SimplifyQuery &Q,
521  unsigned MaxRecurse) {
522  // Recursion is always used, so bail out at once if we already hit the limit.
523  if (!MaxRecurse--)
524  return nullptr;
525 
526  PHINode *PI;
527  if (isa<PHINode>(LHS)) {
528  PI = cast<PHINode>(LHS);
529  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
530  if (!valueDominatesPHI(RHS, PI, Q.DT))
531  return nullptr;
532  } else {
533  assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
534  PI = cast<PHINode>(RHS);
535  // Bail out if LHS and the phi may be mutually interdependent due to a loop.
536  if (!valueDominatesPHI(LHS, PI, Q.DT))
537  return nullptr;
538  }
539 
540  // Evaluate the BinOp on the incoming phi values.
541  Value *CommonValue = nullptr;
542  for (Value *Incoming : PI->incoming_values()) {
543  // If the incoming value is the phi node itself, it can safely be skipped.
544  if (Incoming == PI) continue;
545  Value *V = PI == LHS ?
546  SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
547  SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
548  // If the operation failed to simplify, or simplified to a different value
549  // to previously, then give up.
550  if (!V || (CommonValue && V != CommonValue))
551  return nullptr;
552  CommonValue = V;
553  }
554 
555  return CommonValue;
556 }
557 
558 /// In the case of a comparison with a PHI instruction, try to simplify the
559 /// comparison by seeing whether comparing with all of the incoming phi values
560 /// yields the same result every time. If so returns the common result,
561 /// otherwise returns null.
563  const SimplifyQuery &Q, unsigned MaxRecurse) {
564  // Recursion is always used, so bail out at once if we already hit the limit.
565  if (!MaxRecurse--)
566  return nullptr;
567 
568  // Make sure the phi is on the LHS.
569  if (!isa<PHINode>(LHS)) {
570  std::swap(LHS, RHS);
571  Pred = CmpInst::getSwappedPredicate(Pred);
572  }
573  assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
574  PHINode *PI = cast<PHINode>(LHS);
575 
576  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
577  if (!valueDominatesPHI(RHS, PI, Q.DT))
578  return nullptr;
579 
580  // Evaluate the BinOp on the incoming phi values.
581  Value *CommonValue = nullptr;
582  for (unsigned u = 0, e = PI->getNumIncomingValues(); u < e; ++u) {
583  Value *Incoming = PI->getIncomingValue(u);
584  Instruction *InTI = PI->getIncomingBlock(u)->getTerminator();
585  // If the incoming value is the phi node itself, it can safely be skipped.
586  if (Incoming == PI) continue;
587  // Change the context instruction to the "edge" that flows into the phi.
588  // This is important because that is where incoming is actually "evaluated"
589  // even though it is used later somewhere else.
590  Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q.getWithInstruction(InTI),
591  MaxRecurse);
592  // If the operation failed to simplify, or simplified to a different value
593  // to previously, then give up.
594  if (!V || (CommonValue && V != CommonValue))
595  return nullptr;
596  CommonValue = V;
597  }
598 
599  return CommonValue;
600 }
601 
603  Value *&Op0, Value *&Op1,
604  const SimplifyQuery &Q) {
605  if (auto *CLHS = dyn_cast<Constant>(Op0)) {
606  if (auto *CRHS = dyn_cast<Constant>(Op1))
607  return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
608 
609  // Canonicalize the constant to the RHS if this is a commutative operation.
610  if (Instruction::isCommutative(Opcode))
611  std::swap(Op0, Op1);
612  }
613  return nullptr;
614 }
615 
616 /// Given operands for an Add, see if we can fold the result.
617 /// If not, this returns null.
618 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
619  const SimplifyQuery &Q, unsigned MaxRecurse) {
620  if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
621  return C;
622 
623  // X + undef -> undef
624  if (Q.isUndefValue(Op1))
625  return Op1;
626 
627  // X + 0 -> X
628  if (match(Op1, m_Zero()))
629  return Op0;
630 
631  // If two operands are negative, return 0.
632  if (isKnownNegation(Op0, Op1))
633  return Constant::getNullValue(Op0->getType());
634 
635  // X + (Y - X) -> Y
636  // (Y - X) + X -> Y
637  // Eg: X + -X -> 0
638  Value *Y = nullptr;
639  if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
640  match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
641  return Y;
642 
643  // X + ~X -> -1 since ~X = -X-1
644  Type *Ty = Op0->getType();
645  if (match(Op0, m_Not(m_Specific(Op1))) ||
646  match(Op1, m_Not(m_Specific(Op0))))
647  return Constant::getAllOnesValue(Ty);
648 
649  // add nsw/nuw (xor Y, signmask), signmask --> Y
650  // The no-wrapping add guarantees that the top bit will be set by the add.
651  // Therefore, the xor must be clearing the already set sign bit of Y.
652  if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
653  match(Op0, m_Xor(m_Value(Y), m_SignMask())))
654  return Y;
655 
656  // add nuw %x, -1 -> -1, because %x can only be 0.
657  if (IsNUW && match(Op1, m_AllOnes()))
658  return Op1; // Which is -1.
659 
660  /// i1 add -> xor.
661  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
662  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
663  return V;
664 
665  // Try some generic simplifications for associative operations.
666  if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
667  MaxRecurse))
668  return V;
669 
670  // Threading Add over selects and phi nodes is pointless, so don't bother.
671  // Threading over the select in "A + select(cond, B, C)" means evaluating
672  // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
673  // only if B and C are equal. If B and C are equal then (since we assume
674  // that operands have already been simplified) "select(cond, B, C)" should
675  // have been simplified to the common value of B and C already. Analysing
676  // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
677  // for threading over phi nodes.
678 
679  return nullptr;
680 }
681 
682 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
683  const SimplifyQuery &Query) {
684  return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
685 }
686 
687 /// Compute the base pointer and cumulative constant offsets for V.
688 ///
689 /// This strips all constant offsets off of V, leaving it the base pointer, and
690 /// accumulates the total constant offset applied in the returned constant. It
691 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
692 /// no constant offsets applied.
693 ///
694 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
695 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
696 /// folding.
698  bool AllowNonInbounds = false) {
700 
701  Type *IntIdxTy = DL.getIndexType(V->getType())->getScalarType();
703 
704  V = V->stripAndAccumulateConstantOffsets(DL, Offset, AllowNonInbounds);
705  // As that strip may trace through `addrspacecast`, need to sext or trunc
706  // the offset calculated.
707  IntIdxTy = DL.getIndexType(V->getType())->getScalarType();
708  Offset = Offset.sextOrTrunc(IntIdxTy->getIntegerBitWidth());
709 
710  Constant *OffsetIntPtr = ConstantInt::get(IntIdxTy, Offset);
711  if (VectorType *VecTy = dyn_cast<VectorType>(V->getType()))
712  return ConstantVector::getSplat(VecTy->getElementCount(), OffsetIntPtr);
713  return OffsetIntPtr;
714 }
715 
716 /// Compute the constant difference between two pointer values.
717 /// If the difference is not a constant, returns zero.
719  Value *RHS) {
720  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
721  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
722 
723  // If LHS and RHS are not related via constant offsets to the same base
724  // value, there is nothing we can do here.
725  if (LHS != RHS)
726  return nullptr;
727 
728  // Otherwise, the difference of LHS - RHS can be computed as:
729  // LHS - RHS
730  // = (LHSOffset + Base) - (RHSOffset + Base)
731  // = LHSOffset - RHSOffset
732  return ConstantExpr::getSub(LHSOffset, RHSOffset);
733 }
734 
735 /// Given operands for a Sub, see if we can fold the result.
736 /// If not, this returns null.
737 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
738  const SimplifyQuery &Q, unsigned MaxRecurse) {
739  if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
740  return C;
741 
742  // X - poison -> poison
743  // poison - X -> poison
744  if (isa<PoisonValue>(Op0) || isa<PoisonValue>(Op1))
745  return PoisonValue::get(Op0->getType());
746 
747  // X - undef -> undef
748  // undef - X -> undef
749  if (Q.isUndefValue(Op0) || Q.isUndefValue(Op1))
750  return UndefValue::get(Op0->getType());
751 
752  // X - 0 -> X
753  if (match(Op1, m_Zero()))
754  return Op0;
755 
756  // X - X -> 0
757  if (Op0 == Op1)
758  return Constant::getNullValue(Op0->getType());
759 
760  // Is this a negation?
761  if (match(Op0, m_Zero())) {
762  // 0 - X -> 0 if the sub is NUW.
763  if (isNUW)
764  return Constant::getNullValue(Op0->getType());
765 
766  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
767  if (Known.Zero.isMaxSignedValue()) {
768  // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
769  // Op1 must be 0 because negating the minimum signed value is undefined.
770  if (isNSW)
771  return Constant::getNullValue(Op0->getType());
772 
773  // 0 - X -> X if X is 0 or the minimum signed value.
774  return Op1;
775  }
776  }
777 
778  // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
779  // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
780  Value *X = nullptr, *Y = nullptr, *Z = Op1;
781  if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
782  // See if "V === Y - Z" simplifies.
783  if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
784  // It does! Now see if "X + V" simplifies.
785  if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
786  // It does, we successfully reassociated!
787  ++NumReassoc;
788  return W;
789  }
790  // See if "V === X - Z" simplifies.
791  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
792  // It does! Now see if "Y + V" simplifies.
793  if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
794  // It does, we successfully reassociated!
795  ++NumReassoc;
796  return W;
797  }
798  }
799 
800  // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
801  // For example, X - (X + 1) -> -1
802  X = Op0;
803  if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
804  // See if "V === X - Y" simplifies.
805  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
806  // It does! Now see if "V - Z" simplifies.
807  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
808  // It does, we successfully reassociated!
809  ++NumReassoc;
810  return W;
811  }
812  // See if "V === X - Z" simplifies.
813  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
814  // It does! Now see if "V - Y" simplifies.
815  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
816  // It does, we successfully reassociated!
817  ++NumReassoc;
818  return W;
819  }
820  }
821 
822  // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
823  // For example, X - (X - Y) -> Y.
824  Z = Op0;
825  if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
826  // See if "V === Z - X" simplifies.
827  if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
828  // It does! Now see if "V + Y" simplifies.
829  if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
830  // It does, we successfully reassociated!
831  ++NumReassoc;
832  return W;
833  }
834 
835  // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
836  if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
837  match(Op1, m_Trunc(m_Value(Y))))
838  if (X->getType() == Y->getType())
839  // See if "V === X - Y" simplifies.
840  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
841  // It does! Now see if "trunc V" simplifies.
842  if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
843  Q, MaxRecurse - 1))
844  // It does, return the simplified "trunc V".
845  return W;
846 
847  // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
848  if (match(Op0, m_PtrToInt(m_Value(X))) &&
849  match(Op1, m_PtrToInt(m_Value(Y))))
850  if (Constant *Result = computePointerDifference(Q.DL, X, Y))
851  return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
852 
853  // i1 sub -> xor.
854  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
855  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
856  return V;
857 
858  // Threading Sub over selects and phi nodes is pointless, so don't bother.
859  // Threading over the select in "A - select(cond, B, C)" means evaluating
860  // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
861  // only if B and C are equal. If B and C are equal then (since we assume
862  // that operands have already been simplified) "select(cond, B, C)" should
863  // have been simplified to the common value of B and C already. Analysing
864  // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
865  // for threading over phi nodes.
866 
867  return nullptr;
868 }
869 
870 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
871  const SimplifyQuery &Q) {
872  return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
873 }
874 
875 /// Given operands for a Mul, see if we can fold the result.
876 /// If not, this returns null.
877 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
878  unsigned MaxRecurse) {
879  if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
880  return C;
881 
882  // X * poison -> poison
883  if (isa<PoisonValue>(Op1))
884  return Op1;
885 
886  // X * undef -> 0
887  // X * 0 -> 0
888  if (Q.isUndefValue(Op1) || match(Op1, m_Zero()))
889  return Constant::getNullValue(Op0->getType());
890 
891  // X * 1 -> X
892  if (match(Op1, m_One()))
893  return Op0;
894 
895  // (X / Y) * Y -> X if the division is exact.
896  Value *X = nullptr;
897  if (Q.IIQ.UseInstrInfo &&
898  (match(Op0,
899  m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
900  match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
901  return X;
902 
903  // i1 mul -> and.
904  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
905  if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
906  return V;
907 
908  // Try some generic simplifications for associative operations.
909  if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
910  MaxRecurse))
911  return V;
912 
913  // Mul distributes over Add. Try some generic simplifications based on this.
914  if (Value *V = expandCommutativeBinOp(Instruction::Mul, Op0, Op1,
915  Instruction::Add, Q, MaxRecurse))
916  return V;
917 
918  // If the operation is with the result of a select instruction, check whether
919  // operating on either branch of the select always yields the same value.
920  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
921  if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
922  MaxRecurse))
923  return V;
924 
925  // If the operation is with the result of a phi instruction, check whether
926  // operating on all incoming values of the phi always yields the same value.
927  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
928  if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
929  MaxRecurse))
930  return V;
931 
932  return nullptr;
933 }
934 
937 }
938 
939 /// Check for common or similar folds of integer division or integer remainder.
940 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
942  Value *Op1, const SimplifyQuery &Q) {
943  bool IsDiv = (Opcode == Instruction::SDiv || Opcode == Instruction::UDiv);
944  bool IsSigned = (Opcode == Instruction::SDiv || Opcode == Instruction::SRem);
945 
946  Type *Ty = Op0->getType();
947 
948  // X / undef -> poison
949  // X % undef -> poison
950  if (Q.isUndefValue(Op1))
951  return PoisonValue::get(Ty);
952 
953  // X / 0 -> poison
954  // X % 0 -> poison
955  // We don't need to preserve faults!
956  if (match(Op1, m_Zero()))
957  return PoisonValue::get(Ty);
958 
959  // If any element of a constant divisor fixed width vector is zero or undef
960  // the behavior is undefined and we can fold the whole op to poison.
961  auto *Op1C = dyn_cast<Constant>(Op1);
962  auto *VTy = dyn_cast<FixedVectorType>(Ty);
963  if (Op1C && VTy) {
964  unsigned NumElts = VTy->getNumElements();
965  for (unsigned i = 0; i != NumElts; ++i) {
966  Constant *Elt = Op1C->getAggregateElement(i);
967  if (Elt && (Elt->isNullValue() || Q.isUndefValue(Elt)))
968  return PoisonValue::get(Ty);
969  }
970  }
971 
972  // poison / X -> poison
973  // poison % X -> poison
974  if (isa<PoisonValue>(Op0))
975  return Op0;
976 
977  // undef / X -> 0
978  // undef % X -> 0
979  if (Q.isUndefValue(Op0))
980  return Constant::getNullValue(Ty);
981 
982  // 0 / X -> 0
983  // 0 % X -> 0
984  if (match(Op0, m_Zero()))
985  return Constant::getNullValue(Op0->getType());
986 
987  // X / X -> 1
988  // X % X -> 0
989  if (Op0 == Op1)
990  return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
991 
992  // X / 1 -> X
993  // X % 1 -> 0
994  // If this is a boolean op (single-bit element type), we can't have
995  // division-by-zero or remainder-by-zero, so assume the divisor is 1.
996  // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
997  Value *X;
998  if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
999  (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1000  return IsDiv ? Op0 : Constant::getNullValue(Ty);
1001 
1002  // If X * Y does not overflow, then:
1003  // X * Y / Y -> X
1004  // X * Y % Y -> 0
1005  if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1006  auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1007  // The multiplication can't overflow if it is defined not to, or if
1008  // X == A / Y for some A.
1009  if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1010  (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)) ||
1011  (IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1012  (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1))))) {
1013  return IsDiv ? X : Constant::getNullValue(Op0->getType());
1014  }
1015  }
1016 
1017  return nullptr;
1018 }
1019 
1020 /// Given a predicate and two operands, return true if the comparison is true.
1021 /// This is a helper for div/rem simplification where we return some other value
1022 /// when we can prove a relationship between the operands.
1023 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
1024  const SimplifyQuery &Q, unsigned MaxRecurse) {
1025  Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
1026  Constant *C = dyn_cast_or_null<Constant>(V);
1027  return (C && C->isAllOnesValue());
1028 }
1029 
1030 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
1031 /// to simplify X % Y to X.
1032 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
1033  unsigned MaxRecurse, bool IsSigned) {
1034  // Recursion is always used, so bail out at once if we already hit the limit.
1035  if (!MaxRecurse--)
1036  return false;
1037 
1038  if (IsSigned) {
1039  // |X| / |Y| --> 0
1040  //
1041  // We require that 1 operand is a simple constant. That could be extended to
1042  // 2 variables if we computed the sign bit for each.
1043  //
1044  // Make sure that a constant is not the minimum signed value because taking
1045  // the abs() of that is undefined.
1046  Type *Ty = X->getType();
1047  const APInt *C;
1048  if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
1049  // Is the variable divisor magnitude always greater than the constant
1050  // dividend magnitude?
1051  // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
1052  Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
1053  Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
1054  if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
1055  isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
1056  return true;
1057  }
1058  if (match(Y, m_APInt(C))) {
1059  // Special-case: we can't take the abs() of a minimum signed value. If
1060  // that's the divisor, then all we have to do is prove that the dividend
1061  // is also not the minimum signed value.
1062  if (C->isMinSignedValue())
1063  return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1064 
1065  // Is the variable dividend magnitude always less than the constant
1066  // divisor magnitude?
1067  // |X| < |C| --> X > -abs(C) and X < abs(C)
1068  Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1069  Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1070  if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1071  isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1072  return true;
1073  }
1074  return false;
1075  }
1076 
1077  // IsSigned == false.
1078  // Is the dividend unsigned less than the divisor?
1079  return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1080 }
1081 
1082 /// These are simplifications common to SDiv and UDiv.
1084  const SimplifyQuery &Q, unsigned MaxRecurse) {
1085  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1086  return C;
1087 
1088  if (Value *V = simplifyDivRem(Opcode, Op0, Op1, Q))
1089  return V;
1090 
1091  bool IsSigned = Opcode == Instruction::SDiv;
1092 
1093  // (X rem Y) / Y -> 0
1094  if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1095  (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1096  return Constant::getNullValue(Op0->getType());
1097 
1098  // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1099  ConstantInt *C1, *C2;
1100  if (!IsSigned && match(Op0, m_UDiv(m_Value(), m_ConstantInt(C1))) &&
1101  match(Op1, m_ConstantInt(C2))) {
1102  bool Overflow;
1103  (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1104  if (Overflow)
1105  return Constant::getNullValue(Op0->getType());
1106  }
1107 
1108  // If the operation is with the result of a select instruction, check whether
1109  // operating on either branch of the select always yields the same value.
1110  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1111  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1112  return V;
1113 
1114  // If the operation is with the result of a phi instruction, check whether
1115  // operating on all incoming values of the phi always yields the same value.
1116  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1117  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1118  return V;
1119 
1120  if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1121  return Constant::getNullValue(Op0->getType());
1122 
1123  return nullptr;
1124 }
1125 
1126 /// These are simplifications common to SRem and URem.
1128  const SimplifyQuery &Q, unsigned MaxRecurse) {
1129  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1130  return C;
1131 
1132  if (Value *V = simplifyDivRem(Opcode, Op0, Op1, Q))
1133  return V;
1134 
1135  // (X % Y) % Y -> X % Y
1136  if ((Opcode == Instruction::SRem &&
1137  match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1138  (Opcode == Instruction::URem &&
1139  match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1140  return Op0;
1141 
1142  // (X << Y) % X -> 0
1143  if (Q.IIQ.UseInstrInfo &&
1144  ((Opcode == Instruction::SRem &&
1145  match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1146  (Opcode == Instruction::URem &&
1147  match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1148  return Constant::getNullValue(Op0->getType());
1149 
1150  // If the operation is with the result of a select instruction, check whether
1151  // operating on either branch of the select always yields the same value.
1152  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1153  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1154  return V;
1155 
1156  // If the operation is with the result of a phi instruction, check whether
1157  // operating on all incoming values of the phi always yields the same value.
1158  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1159  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1160  return V;
1161 
1162  // If X / Y == 0, then X % Y == X.
1163  if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1164  return Op0;
1165 
1166  return nullptr;
1167 }
1168 
1169 /// Given operands for an SDiv, see if we can fold the result.
1170 /// If not, this returns null.
1171 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1172  unsigned MaxRecurse) {
1173  // If two operands are negated and no signed overflow, return -1.
1174  if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1175  return Constant::getAllOnesValue(Op0->getType());
1176 
1177  return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1178 }
1179 
1182 }
1183 
1184 /// Given operands for a UDiv, see if we can fold the result.
1185 /// If not, this returns null.
1186 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1187  unsigned MaxRecurse) {
1188  return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1189 }
1190 
1193 }
1194 
1195 /// Given operands for an SRem, see if we can fold the result.
1196 /// If not, this returns null.
1197 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1198  unsigned MaxRecurse) {
1199  // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1200  // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1201  Value *X;
1202  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1203  return ConstantInt::getNullValue(Op0->getType());
1204 
1205  // If the two operands are negated, return 0.
1206  if (isKnownNegation(Op0, Op1))
1207  return ConstantInt::getNullValue(Op0->getType());
1208 
1209  return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1210 }
1211 
1214 }
1215 
1216 /// Given operands for a URem, see if we can fold the result.
1217 /// If not, this returns null.
1218 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1219  unsigned MaxRecurse) {
1220  return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1221 }
1222 
1225 }
1226 
1227 /// Returns true if a shift by \c Amount always yields poison.
1228 static bool isPoisonShift(Value *Amount, const SimplifyQuery &Q) {
1229  Constant *C = dyn_cast<Constant>(Amount);
1230  if (!C)
1231  return false;
1232 
1233  // X shift by undef -> poison because it may shift by the bitwidth.
1234  if (Q.isUndefValue(C))
1235  return true;
1236 
1237  // Shifting by the bitwidth or more is undefined.
1238  if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1239  if (CI->getValue().uge(CI->getType()->getScalarSizeInBits()))
1240  return true;
1241 
1242  // If all lanes of a vector shift are undefined the whole shift is.
1243  if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1244  for (unsigned I = 0,
1245  E = cast<FixedVectorType>(C->getType())->getNumElements();
1246  I != E; ++I)
1247  if (!isPoisonShift(C->getAggregateElement(I), Q))
1248  return false;
1249  return true;
1250  }
1251 
1252  return false;
1253 }
1254 
1255 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1256 /// If not, this returns null.
1258  Value *Op1, bool IsNSW, const SimplifyQuery &Q,
1259  unsigned MaxRecurse) {
1260  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1261  return C;
1262 
1263  // poison shift by X -> poison
1264  if (isa<PoisonValue>(Op0))
1265  return Op0;
1266 
1267  // 0 shift by X -> 0
1268  if (match(Op0, m_Zero()))
1269  return Constant::getNullValue(Op0->getType());
1270 
1271  // X shift by 0 -> X
1272  // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1273  // would be poison.
1274  Value *X;
1275  if (match(Op1, m_Zero()) ||
1276  (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1277  return Op0;
1278 
1279  // Fold undefined shifts.
1280  if (isPoisonShift(Op1, Q))
1281  return PoisonValue::get(Op0->getType());
1282 
1283  // If the operation is with the result of a select instruction, check whether
1284  // operating on either branch of the select always yields the same value.
1285  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1286  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1287  return V;
1288 
1289  // If the operation is with the result of a phi instruction, check whether
1290  // operating on all incoming values of the phi always yields the same value.
1291  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1292  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1293  return V;
1294 
1295  // If any bits in the shift amount make that value greater than or equal to
1296  // the number of bits in the type, the shift is undefined.
1297  KnownBits KnownAmt = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1298  if (KnownAmt.getMinValue().uge(KnownAmt.getBitWidth()))
1299  return PoisonValue::get(Op0->getType());
1300 
1301  // If all valid bits in the shift amount are known zero, the first operand is
1302  // unchanged.
1303  unsigned NumValidShiftBits = Log2_32_Ceil(KnownAmt.getBitWidth());
1304  if (KnownAmt.countMinTrailingZeros() >= NumValidShiftBits)
1305  return Op0;
1306 
1307  // Check for nsw shl leading to a poison value.
1308  if (IsNSW) {
1309  assert(Opcode == Instruction::Shl && "Expected shl for nsw instruction");
1310  KnownBits KnownVal = computeKnownBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1311  KnownBits KnownShl = KnownBits::shl(KnownVal, KnownAmt);
1312 
1313  if (KnownVal.Zero.isSignBitSet())
1314  KnownShl.Zero.setSignBit();
1315  if (KnownVal.One.isSignBitSet())
1316  KnownShl.One.setSignBit();
1317 
1318  if (KnownShl.hasConflict())
1319  return PoisonValue::get(Op0->getType());
1320  }
1321 
1322  return nullptr;
1323 }
1324 
1325 /// Given operands for an Shl, LShr or AShr, see if we can
1326 /// fold the result. If not, this returns null.
1328  Value *Op1, bool isExact, const SimplifyQuery &Q,
1329  unsigned MaxRecurse) {
1330  if (Value *V =
1331  SimplifyShift(Opcode, Op0, Op1, /*IsNSW*/ false, Q, MaxRecurse))
1332  return V;
1333 
1334  // X >> X -> 0
1335  if (Op0 == Op1)
1336  return Constant::getNullValue(Op0->getType());
1337 
1338  // undef >> X -> 0
1339  // undef >> X -> undef (if it's exact)
1340  if (Q.isUndefValue(Op0))
1341  return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1342 
1343  // The low bit cannot be shifted out of an exact shift if it is set.
1344  if (isExact) {
1345  KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1346  if (Op0Known.One[0])
1347  return Op0;
1348  }
1349 
1350  return nullptr;
1351 }
1352 
1353 /// Given operands for an Shl, see if we can fold the result.
1354 /// If not, this returns null.
1355 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1356  const SimplifyQuery &Q, unsigned MaxRecurse) {
1357  if (Value *V =
1358  SimplifyShift(Instruction::Shl, Op0, Op1, isNSW, Q, MaxRecurse))
1359  return V;
1360 
1361  // undef << X -> 0
1362  // undef << X -> undef if (if it's NSW/NUW)
1363  if (Q.isUndefValue(Op0))
1364  return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1365 
1366  // (X >> A) << A -> X
1367  Value *X;
1368  if (Q.IIQ.UseInstrInfo &&
1369  match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1370  return X;
1371 
1372  // shl nuw i8 C, %x -> C iff C has sign bit set.
1373  if (isNUW && match(Op0, m_Negative()))
1374  return Op0;
1375  // NOTE: could use computeKnownBits() / LazyValueInfo,
1376  // but the cost-benefit analysis suggests it isn't worth it.
1377 
1378  return nullptr;
1379 }
1380 
1381 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1382  const SimplifyQuery &Q) {
1383  return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1384 }
1385 
1386 /// Given operands for an LShr, see if we can fold the result.
1387 /// If not, this returns null.
1388 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1389  const SimplifyQuery &Q, unsigned MaxRecurse) {
1390  if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1391  MaxRecurse))
1392  return V;
1393 
1394  // (X << A) >> A -> X
1395  Value *X;
1396  if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1397  return X;
1398 
1399  // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1400  // We can return X as we do in the above case since OR alters no bits in X.
1401  // SimplifyDemandedBits in InstCombine can do more general optimization for
1402  // bit manipulation. This pattern aims to provide opportunities for other
1403  // optimizers by supporting a simple but common case in InstSimplify.
1404  Value *Y;
1405  const APInt *ShRAmt, *ShLAmt;
1406  if (match(Op1, m_APInt(ShRAmt)) &&
1407  match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1408  *ShRAmt == *ShLAmt) {
1409  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1410  const unsigned EffWidthY = YKnown.countMaxActiveBits();
1411  if (ShRAmt->uge(EffWidthY))
1412  return X;
1413  }
1414 
1415  return nullptr;
1416 }
1417 
1418 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1419  const SimplifyQuery &Q) {
1420  return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1421 }
1422 
1423 /// Given operands for an AShr, see if we can fold the result.
1424 /// If not, this returns null.
1425 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1426  const SimplifyQuery &Q, unsigned MaxRecurse) {
1427  if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1428  MaxRecurse))
1429  return V;
1430 
1431  // -1 >>a X --> -1
1432  // (-1 << X) a>> X --> -1
1433  // Do not return Op0 because it may contain undef elements if it's a vector.
1434  if (match(Op0, m_AllOnes()) ||
1435  match(Op0, m_Shl(m_AllOnes(), m_Specific(Op1))))
1436  return Constant::getAllOnesValue(Op0->getType());
1437 
1438  // (X << A) >> A -> X
1439  Value *X;
1440  if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1441  return X;
1442 
1443  // Arithmetic shifting an all-sign-bit value is a no-op.
1444  unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1445  if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1446  return Op0;
1447 
1448  return nullptr;
1449 }
1450 
1451 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1452  const SimplifyQuery &Q) {
1453  return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1454 }
1455 
1456 /// Commuted variants are assumed to be handled by calling this function again
1457 /// with the parameters swapped.
1459  ICmpInst *UnsignedICmp, bool IsAnd,
1460  const SimplifyQuery &Q) {
1461  Value *X, *Y;
1462 
1463  ICmpInst::Predicate EqPred;
1464  if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1465  !ICmpInst::isEquality(EqPred))
1466  return nullptr;
1467 
1468  ICmpInst::Predicate UnsignedPred;
1469 
1470  Value *A, *B;
1471  // Y = (A - B);
1472  if (match(Y, m_Sub(m_Value(A), m_Value(B)))) {
1473  if (match(UnsignedICmp,
1474  m_c_ICmp(UnsignedPred, m_Specific(A), m_Specific(B))) &&
1475  ICmpInst::isUnsigned(UnsignedPred)) {
1476  // A >=/<= B || (A - B) != 0 <--> true
1477  if ((UnsignedPred == ICmpInst::ICMP_UGE ||
1478  UnsignedPred == ICmpInst::ICMP_ULE) &&
1479  EqPred == ICmpInst::ICMP_NE && !IsAnd)
1480  return ConstantInt::getTrue(UnsignedICmp->getType());
1481  // A </> B && (A - B) == 0 <--> false
1482  if ((UnsignedPred == ICmpInst::ICMP_ULT ||
1483  UnsignedPred == ICmpInst::ICMP_UGT) &&
1484  EqPred == ICmpInst::ICMP_EQ && IsAnd)
1485  return ConstantInt::getFalse(UnsignedICmp->getType());
1486 
1487  // A </> B && (A - B) != 0 <--> A </> B
1488  // A </> B || (A - B) != 0 <--> (A - B) != 0
1489  if (EqPred == ICmpInst::ICMP_NE && (UnsignedPred == ICmpInst::ICMP_ULT ||
1490  UnsignedPred == ICmpInst::ICMP_UGT))
1491  return IsAnd ? UnsignedICmp : ZeroICmp;
1492 
1493  // A <=/>= B && (A - B) == 0 <--> (A - B) == 0
1494  // A <=/>= B || (A - B) == 0 <--> A <=/>= B
1495  if (EqPred == ICmpInst::ICMP_EQ && (UnsignedPred == ICmpInst::ICMP_ULE ||
1496  UnsignedPred == ICmpInst::ICMP_UGE))
1497  return IsAnd ? ZeroICmp : UnsignedICmp;
1498  }
1499 
1500  // Given Y = (A - B)
1501  // Y >= A && Y != 0 --> Y >= A iff B != 0
1502  // Y < A || Y == 0 --> Y < A iff B != 0
1503  if (match(UnsignedICmp,
1504  m_c_ICmp(UnsignedPred, m_Specific(Y), m_Specific(A)))) {
1505  if (UnsignedPred == ICmpInst::ICMP_UGE && IsAnd &&
1506  EqPred == ICmpInst::ICMP_NE &&
1507  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1508  return UnsignedICmp;
1509  if (UnsignedPred == ICmpInst::ICMP_ULT && !IsAnd &&
1510  EqPred == ICmpInst::ICMP_EQ &&
1511  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1512  return UnsignedICmp;
1513  }
1514  }
1515 
1516  if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1517  ICmpInst::isUnsigned(UnsignedPred))
1518  ;
1519  else if (match(UnsignedICmp,
1520  m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1521  ICmpInst::isUnsigned(UnsignedPred))
1522  UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1523  else
1524  return nullptr;
1525 
1526  // X > Y && Y == 0 --> Y == 0 iff X != 0
1527  // X > Y || Y == 0 --> X > Y iff X != 0
1528  if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ &&
1529  isKnownNonZero(X, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1530  return IsAnd ? ZeroICmp : UnsignedICmp;
1531 
1532  // X <= Y && Y != 0 --> X <= Y iff X != 0
1533  // X <= Y || Y != 0 --> Y != 0 iff X != 0
1534  if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE &&
1535  isKnownNonZero(X, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1536  return IsAnd ? UnsignedICmp : ZeroICmp;
1537 
1538  // The transforms below here are expected to be handled more generally with
1539  // simplifyAndOrOfICmpsWithLimitConst() or in InstCombine's
1540  // foldAndOrOfICmpsWithConstEq(). If we are looking to trim optimizer overlap,
1541  // these are candidates for removal.
1542 
1543  // X < Y && Y != 0 --> X < Y
1544  // X < Y || Y != 0 --> Y != 0
1545  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1546  return IsAnd ? UnsignedICmp : ZeroICmp;
1547 
1548  // X >= Y && Y == 0 --> Y == 0
1549  // X >= Y || Y == 0 --> X >= Y
1550  if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ)
1551  return IsAnd ? ZeroICmp : UnsignedICmp;
1552 
1553  // X < Y && Y == 0 --> false
1554  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1555  IsAnd)
1556  return getFalse(UnsignedICmp->getType());
1557 
1558  // X >= Y || Y != 0 --> true
1559  if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_NE &&
1560  !IsAnd)
1561  return getTrue(UnsignedICmp->getType());
1562 
1563  return nullptr;
1564 }
1565 
1566 /// Commuted variants are assumed to be handled by calling this function again
1567 /// with the parameters swapped.
1569  ICmpInst::Predicate Pred0, Pred1;
1570  Value *A ,*B;
1571  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1572  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1573  return nullptr;
1574 
1575  // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1576  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1577  // can eliminate Op1 from this 'and'.
1578  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1579  return Op0;
1580 
1581  // Check for any combination of predicates that are guaranteed to be disjoint.
1582  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1583  (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1584  (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1585  (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1586  return getFalse(Op0->getType());
1587 
1588  return nullptr;
1589 }
1590 
1591 /// Commuted variants are assumed to be handled by calling this function again
1592 /// with the parameters swapped.
1594  ICmpInst::Predicate Pred0, Pred1;
1595  Value *A ,*B;
1596  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1597  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1598  return nullptr;
1599 
1600  // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1601  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1602  // can eliminate Op0 from this 'or'.
1603  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1604  return Op1;
1605 
1606  // Check for any combination of predicates that cover the entire range of
1607  // possibilities.
1608  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1609  (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1610  (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1611  (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1612  return getTrue(Op0->getType());
1613 
1614  return nullptr;
1615 }
1616 
1617 /// Test if a pair of compares with a shared operand and 2 constants has an
1618 /// empty set intersection, full set union, or if one compare is a superset of
1619 /// the other.
1621  bool IsAnd) {
1622  // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1623  if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1624  return nullptr;
1625 
1626  const APInt *C0, *C1;
1627  if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1628  !match(Cmp1->getOperand(1), m_APInt(C1)))
1629  return nullptr;
1630 
1631  auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1632  auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1633 
1634  // For and-of-compares, check if the intersection is empty:
1635  // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1636  if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1637  return getFalse(Cmp0->getType());
1638 
1639  // For or-of-compares, check if the union is full:
1640  // (icmp X, C0) || (icmp X, C1) --> full set --> true
1641  if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1642  return getTrue(Cmp0->getType());
1643 
1644  // Is one range a superset of the other?
1645  // If this is and-of-compares, take the smaller set:
1646  // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1647  // If this is or-of-compares, take the larger set:
1648  // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1649  if (Range0.contains(Range1))
1650  return IsAnd ? Cmp1 : Cmp0;
1651  if (Range1.contains(Range0))
1652  return IsAnd ? Cmp0 : Cmp1;
1653 
1654  return nullptr;
1655 }
1656 
1658  bool IsAnd) {
1659  ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1660  if (!match(Cmp0->getOperand(1), m_Zero()) ||
1661  !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1662  return nullptr;
1663 
1664  if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1665  return nullptr;
1666 
1667  // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1668  Value *X = Cmp0->getOperand(0);
1669  Value *Y = Cmp1->getOperand(0);
1670 
1671  // If one of the compares is a masked version of a (not) null check, then
1672  // that compare implies the other, so we eliminate the other. Optionally, look
1673  // through a pointer-to-int cast to match a null check of a pointer type.
1674 
1675  // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1676  // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1677  // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1678  // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1679  if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1681  return Cmp1;
1682 
1683  // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1684  // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1685  // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1686  // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1687  if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1689  return Cmp0;
1690 
1691  return nullptr;
1692 }
1693 
1695  const InstrInfoQuery &IIQ) {
1696  // (icmp (add V, C0), C1) & (icmp V, C0)
1697  ICmpInst::Predicate Pred0, Pred1;
1698  const APInt *C0, *C1;
1699  Value *V;
1700  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1701  return nullptr;
1702 
1703  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1704  return nullptr;
1705 
1706  auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1707  if (AddInst->getOperand(1) != Op1->getOperand(1))
1708  return nullptr;
1709 
1710  Type *ITy = Op0->getType();
1711  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1712  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1713 
1714  const APInt Delta = *C1 - *C0;
1715  if (C0->isStrictlyPositive()) {
1716  if (Delta == 2) {
1717  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1718  return getFalse(ITy);
1719  if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1720  return getFalse(ITy);
1721  }
1722  if (Delta == 1) {
1723  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1724  return getFalse(ITy);
1725  if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1726  return getFalse(ITy);
1727  }
1728  }
1729  if (C0->getBoolValue() && isNUW) {
1730  if (Delta == 2)
1731  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1732  return getFalse(ITy);
1733  if (Delta == 1)
1734  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1735  return getFalse(ITy);
1736  }
1737 
1738  return nullptr;
1739 }
1740 
1741 /// Try to eliminate compares with signed or unsigned min/max constants.
1743  bool IsAnd) {
1744  // Canonicalize an equality compare as Cmp0.
1745  if (Cmp1->isEquality())
1746  std::swap(Cmp0, Cmp1);
1747  if (!Cmp0->isEquality())
1748  return nullptr;
1749 
1750  // The non-equality compare must include a common operand (X). Canonicalize
1751  // the common operand as operand 0 (the predicate is swapped if the common
1752  // operand was operand 1).
1753  ICmpInst::Predicate Pred0 = Cmp0->getPredicate();
1754  Value *X = Cmp0->getOperand(0);
1755  ICmpInst::Predicate Pred1;
1756  bool HasNotOp = match(Cmp1, m_c_ICmp(Pred1, m_Not(m_Specific(X)), m_Value()));
1757  if (!HasNotOp && !match(Cmp1, m_c_ICmp(Pred1, m_Specific(X), m_Value())))
1758  return nullptr;
1759  if (ICmpInst::isEquality(Pred1))
1760  return nullptr;
1761 
1762  // The equality compare must be against a constant. Flip bits if we matched
1763  // a bitwise not. Convert a null pointer constant to an integer zero value.
1764  APInt MinMaxC;
1765  const APInt *C;
1766  if (match(Cmp0->getOperand(1), m_APInt(C)))
1767  MinMaxC = HasNotOp ? ~*C : *C;
1768  else if (isa<ConstantPointerNull>(Cmp0->getOperand(1)))
1769  MinMaxC = APInt::getZero(8);
1770  else
1771  return nullptr;
1772 
1773  // DeMorganize if this is 'or': P0 || P1 --> !P0 && !P1.
1774  if (!IsAnd) {
1775  Pred0 = ICmpInst::getInversePredicate(Pred0);
1776  Pred1 = ICmpInst::getInversePredicate(Pred1);
1777  }
1778 
1779  // Normalize to unsigned compare and unsigned min/max value.
1780  // Example for 8-bit: -128 + 128 -> 0; 127 + 128 -> 255
1781  if (ICmpInst::isSigned(Pred1)) {
1782  Pred1 = ICmpInst::getUnsignedPredicate(Pred1);
1783  MinMaxC += APInt::getSignedMinValue(MinMaxC.getBitWidth());
1784  }
1785 
1786  // (X != MAX) && (X < Y) --> X < Y
1787  // (X == MAX) || (X >= Y) --> X >= Y
1788  if (MinMaxC.isMaxValue())
1789  if (Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT)
1790  return Cmp1;
1791 
1792  // (X != MIN) && (X > Y) --> X > Y
1793  // (X == MIN) || (X <= Y) --> X <= Y
1794  if (MinMaxC.isMinValue())
1795  if (Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_UGT)
1796  return Cmp1;
1797 
1798  return nullptr;
1799 }
1800 
1802  const SimplifyQuery &Q) {
1803  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true, Q))
1804  return X;
1805  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true, Q))
1806  return X;
1807 
1808  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1809  return X;
1810  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1811  return X;
1812 
1813  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1814  return X;
1815 
1816  if (Value *X = simplifyAndOrOfICmpsWithLimitConst(Op0, Op1, true))
1817  return X;
1818 
1819  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1820  return X;
1821 
1822  if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, Q.IIQ))
1823  return X;
1824  if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, Q.IIQ))
1825  return X;
1826 
1827  return nullptr;
1828 }
1829 
1831  const InstrInfoQuery &IIQ) {
1832  // (icmp (add V, C0), C1) | (icmp V, C0)
1833  ICmpInst::Predicate Pred0, Pred1;
1834  const APInt *C0, *C1;
1835  Value *V;
1836  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1837  return nullptr;
1838 
1839  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1840  return nullptr;
1841 
1842  auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1843  if (AddInst->getOperand(1) != Op1->getOperand(1))
1844  return nullptr;
1845 
1846  Type *ITy = Op0->getType();
1847  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1848  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1849 
1850  const APInt Delta = *C1 - *C0;
1851  if (C0->isStrictlyPositive()) {
1852  if (Delta == 2) {
1853  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1854  return getTrue(ITy);
1855  if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1856  return getTrue(ITy);
1857  }
1858  if (Delta == 1) {
1859  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1860  return getTrue(ITy);
1861  if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1862  return getTrue(ITy);
1863  }
1864  }
1865  if (C0->getBoolValue() && isNUW) {
1866  if (Delta == 2)
1867  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1868  return getTrue(ITy);
1869  if (Delta == 1)
1870  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1871  return getTrue(ITy);
1872  }
1873 
1874  return nullptr;
1875 }
1876 
1878  const SimplifyQuery &Q) {
1879  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false, Q))
1880  return X;
1881  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false, Q))
1882  return X;
1883 
1884  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1885  return X;
1886  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1887  return X;
1888 
1889  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1890  return X;
1891 
1892  if (Value *X = simplifyAndOrOfICmpsWithLimitConst(Op0, Op1, false))
1893  return X;
1894 
1895  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1896  return X;
1897 
1898  if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, Q.IIQ))
1899  return X;
1900  if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, Q.IIQ))
1901  return X;
1902 
1903  return nullptr;
1904 }
1905 
1907  FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1908  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1909  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1910  if (LHS0->getType() != RHS0->getType())
1911  return nullptr;
1912 
1913  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1914  if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1915  (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1916  // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1917  // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1918  // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1919  // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1920  // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1921  // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1922  // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1923  // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1924  if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1925  (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1926  return RHS;
1927 
1928  // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1929  // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1930  // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1931  // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1932  // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1933  // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1934  // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1935  // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1936  if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1937  (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1938  return LHS;
1939  }
1940 
1941  return nullptr;
1942 }
1943 
1945  Value *Op0, Value *Op1, bool IsAnd) {
1946  // Look through casts of the 'and' operands to find compares.
1947  auto *Cast0 = dyn_cast<CastInst>(Op0);
1948  auto *Cast1 = dyn_cast<CastInst>(Op1);
1949  if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1950  Cast0->getSrcTy() == Cast1->getSrcTy()) {
1951  Op0 = Cast0->getOperand(0);
1952  Op1 = Cast1->getOperand(0);
1953  }
1954 
1955  Value *V = nullptr;
1956  auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1957  auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1958  if (ICmp0 && ICmp1)
1959  V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q)
1960  : simplifyOrOfICmps(ICmp0, ICmp1, Q);
1961 
1962  auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1963  auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1964  if (FCmp0 && FCmp1)
1965  V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1966 
1967  if (!V)
1968  return nullptr;
1969  if (!Cast0)
1970  return V;
1971 
1972  // If we looked through casts, we can only handle a constant simplification
1973  // because we are not allowed to create a cast instruction here.
1974  if (auto *C = dyn_cast<Constant>(V))
1975  return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1976 
1977  return nullptr;
1978 }
1979 
1980 /// Given a bitwise logic op, check if the operands are add/sub with a common
1981 /// source value and inverted constant (identity: C - X -> ~(X + ~C)).
1983  Instruction::BinaryOps Opcode) {
1984  assert(Op0->getType() == Op1->getType() && "Mismatched binop types");
1985  assert(BinaryOperator::isBitwiseLogicOp(Opcode) && "Expected logic op");
1986  Value *X;
1987  Constant *C1, *C2;
1988  if ((match(Op0, m_Add(m_Value(X), m_Constant(C1))) &&
1989  match(Op1, m_Sub(m_Constant(C2), m_Specific(X)))) ||
1990  (match(Op1, m_Add(m_Value(X), m_Constant(C1))) &&
1991  match(Op0, m_Sub(m_Constant(C2), m_Specific(X))))) {
1992  if (ConstantExpr::getNot(C1) == C2) {
1993  // (X + C) & (~C - X) --> (X + C) & ~(X + C) --> 0
1994  // (X + C) | (~C - X) --> (X + C) | ~(X + C) --> -1
1995  // (X + C) ^ (~C - X) --> (X + C) ^ ~(X + C) --> -1
1996  Type *Ty = Op0->getType();
1997  return Opcode == Instruction::And ? ConstantInt::getNullValue(Ty)
1999  }
2000  }
2001  return nullptr;
2002 }
2003 
2004 /// Given operands for an And, see if we can fold the result.
2005 /// If not, this returns null.
2006 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2007  unsigned MaxRecurse) {
2008  if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
2009  return C;
2010 
2011  // X & poison -> poison
2012  if (isa<PoisonValue>(Op1))
2013  return Op1;
2014 
2015  // X & undef -> 0
2016  if (Q.isUndefValue(Op1))
2017  return Constant::getNullValue(Op0->getType());
2018 
2019  // X & X = X
2020  if (Op0 == Op1)
2021  return Op0;
2022 
2023  // X & 0 = 0
2024  if (match(Op1, m_Zero()))
2025  return Constant::getNullValue(Op0->getType());
2026 
2027  // X & -1 = X
2028  if (match(Op1, m_AllOnes()))
2029  return Op0;
2030 
2031  // A & ~A = ~A & A = 0
2032  if (match(Op0, m_Not(m_Specific(Op1))) ||
2033  match(Op1, m_Not(m_Specific(Op0))))
2034  return Constant::getNullValue(Op0->getType());
2035 
2036  // (A | ?) & A = A
2037  if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
2038  return Op1;
2039 
2040  // A & (A | ?) = A
2041  if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
2042  return Op0;
2043 
2044  // (X | Y) & (X | ~Y) --> X (commuted 8 ways)
2045  Value *X, *Y;
2046  if (match(Op0, m_c_Or(m_Value(X), m_Not(m_Value(Y)))) &&
2047  match(Op1, m_c_Or(m_Deferred(X), m_Deferred(Y))))
2048  return X;
2049  if (match(Op1, m_c_Or(m_Value(X), m_Not(m_Value(Y)))) &&
2050  match(Op0, m_c_Or(m_Deferred(X), m_Deferred(Y))))
2051  return X;
2052 
2053  if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Instruction::And))
2054  return V;
2055 
2056  // A mask that only clears known zeros of a shifted value is a no-op.
2057  const APInt *Mask;
2058  const APInt *ShAmt;
2059  if (match(Op1, m_APInt(Mask))) {
2060  // If all bits in the inverted and shifted mask are clear:
2061  // and (shl X, ShAmt), Mask --> shl X, ShAmt
2062  if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
2063  (~(*Mask)).lshr(*ShAmt).isZero())
2064  return Op0;
2065 
2066  // If all bits in the inverted and shifted mask are clear:
2067  // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
2068  if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
2069  (~(*Mask)).shl(*ShAmt).isZero())
2070  return Op0;
2071  }
2072 
2073  // If we have a multiplication overflow check that is being 'and'ed with a
2074  // check that one of the multipliers is not zero, we can omit the 'and', and
2075  // only keep the overflow check.
2076  if (isCheckForZeroAndMulWithOverflow(Op0, Op1, true))
2077  return Op1;
2078  if (isCheckForZeroAndMulWithOverflow(Op1, Op0, true))
2079  return Op0;
2080 
2081  // A & (-A) = A if A is a power of two or zero.
2082  if (match(Op0, m_Neg(m_Specific(Op1))) ||
2083  match(Op1, m_Neg(m_Specific(Op0)))) {
2084  if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
2085  Q.DT))
2086  return Op0;
2087  if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
2088  Q.DT))
2089  return Op1;
2090  }
2091 
2092  // This is a similar pattern used for checking if a value is a power-of-2:
2093  // (A - 1) & A --> 0 (if A is a power-of-2 or 0)
2094  // A & (A - 1) --> 0 (if A is a power-of-2 or 0)
2095  if (match(Op0, m_Add(m_Specific(Op1), m_AllOnes())) &&
2096  isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
2097  return Constant::getNullValue(Op1->getType());
2098  if (match(Op1, m_Add(m_Specific(Op0), m_AllOnes())) &&
2099  isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
2100  return Constant::getNullValue(Op0->getType());
2101 
2102  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
2103  return V;
2104 
2105  // Try some generic simplifications for associative operations.
2106  if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
2107  MaxRecurse))
2108  return V;
2109 
2110  // And distributes over Or. Try some generic simplifications based on this.
2111  if (Value *V = expandCommutativeBinOp(Instruction::And, Op0, Op1,
2112  Instruction::Or, Q, MaxRecurse))
2113  return V;
2114 
2115  // And distributes over Xor. Try some generic simplifications based on this.
2116  if (Value *V = expandCommutativeBinOp(Instruction::And, Op0, Op1,
2117  Instruction::Xor, Q, MaxRecurse))
2118  return V;
2119 
2120  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) {
2121  if (Op0->getType()->isIntOrIntVectorTy(1)) {
2122  // A & (A && B) -> A && B
2123  if (match(Op1, m_Select(m_Specific(Op0), m_Value(), m_Zero())))
2124  return Op1;
2125  else if (match(Op0, m_Select(m_Specific(Op1), m_Value(), m_Zero())))
2126  return Op0;
2127  }
2128  // If the operation is with the result of a select instruction, check
2129  // whether operating on either branch of the select always yields the same
2130  // value.
2131  if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
2132  MaxRecurse))
2133  return V;
2134  }
2135 
2136  // If the operation is with the result of a phi instruction, check whether
2137  // operating on all incoming values of the phi always yields the same value.
2138  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2139  if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
2140  MaxRecurse))
2141  return V;
2142 
2143  // Assuming the effective width of Y is not larger than A, i.e. all bits
2144  // from X and Y are disjoint in (X << A) | Y,
2145  // if the mask of this AND op covers all bits of X or Y, while it covers
2146  // no bits from the other, we can bypass this AND op. E.g.,
2147  // ((X << A) | Y) & Mask -> Y,
2148  // if Mask = ((1 << effective_width_of(Y)) - 1)
2149  // ((X << A) | Y) & Mask -> X << A,
2150  // if Mask = ((1 << effective_width_of(X)) - 1) << A
2151  // SimplifyDemandedBits in InstCombine can optimize the general case.
2152  // This pattern aims to help other passes for a common case.
2153  Value *XShifted;
2154  if (match(Op1, m_APInt(Mask)) &&
2156  m_Value(XShifted)),
2157  m_Value(Y)))) {
2158  const unsigned Width = Op0->getType()->getScalarSizeInBits();
2159  const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
2160  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2161  const unsigned EffWidthY = YKnown.countMaxActiveBits();
2162  if (EffWidthY <= ShftCnt) {
2163  const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
2164  Q.DT);
2165  const unsigned EffWidthX = XKnown.countMaxActiveBits();
2166  const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
2167  const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
2168  // If the mask is extracting all bits from X or Y as is, we can skip
2169  // this AND op.
2170  if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
2171  return Y;
2172  if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
2173  return XShifted;
2174  }
2175  }
2176 
2177  return nullptr;
2178 }
2179 
2182 }
2183 
2184 /// Given operands for an Or, see if we can fold the result.
2185 /// If not, this returns null.
2186 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2187  unsigned MaxRecurse) {
2188  if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
2189  return C;
2190 
2191  // X | poison -> poison
2192  if (isa<PoisonValue>(Op1))
2193  return Op1;
2194 
2195  // X | undef -> -1
2196  // X | -1 = -1
2197  // Do not return Op1 because it may contain undef elements if it's a vector.
2198  if (Q.isUndefValue(Op1) || match(Op1, m_AllOnes()))
2199  return Constant::getAllOnesValue(Op0->getType());
2200 
2201  // X | X = X
2202  // X | 0 = X
2203  if (Op0 == Op1 || match(Op1, m_Zero()))
2204  return Op0;
2205 
2206  // A | ~A = ~A | A = -1
2207  if (match(Op0, m_Not(m_Specific(Op1))) ||
2208  match(Op1, m_Not(m_Specific(Op0))))
2209  return Constant::getAllOnesValue(Op0->getType());
2210 
2211  // (A & ?) | A = A
2212  if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
2213  return Op1;
2214 
2215  // A | (A & ?) = A
2216  if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
2217  return Op0;
2218 
2219  // ~(A & ?) | A = -1
2220  if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
2221  return Constant::getAllOnesValue(Op1->getType());
2222 
2223  // A | ~(A & ?) = -1
2224  if (match(Op1, m_Not(m_c_And(m_Specific(Op0), m_Value()))))
2225  return Constant::getAllOnesValue(Op0->getType());
2226 
2227  if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Instruction::Or))
2228  return V;
2229 
2230  Value *A, *B, *NotA;
2231  // (A & ~B) | (A ^ B) -> (A ^ B)
2232  // (~B & A) | (A ^ B) -> (A ^ B)
2233  // (A & ~B) | (B ^ A) -> (B ^ A)
2234  // (~B & A) | (B ^ A) -> (B ^ A)
2235  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2236  (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
2237  match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
2238  return Op1;
2239 
2240  // Commute the 'or' operands.
2241  // (A ^ B) | (A & ~B) -> (A ^ B)
2242  // (A ^ B) | (~B & A) -> (A ^ B)
2243  // (B ^ A) | (A & ~B) -> (B ^ A)
2244  // (B ^ A) | (~B & A) -> (B ^ A)
2245  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2246  (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
2247  match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
2248  return Op0;
2249 
2250  // (A & B) | (~A ^ B) -> (~A ^ B)
2251  // (B & A) | (~A ^ B) -> (~A ^ B)
2252  // (A & B) | (B ^ ~A) -> (B ^ ~A)
2253  // (B & A) | (B ^ ~A) -> (B ^ ~A)
2254  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2255  (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2256  match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2257  return Op1;
2258 
2259  // Commute the 'or' operands.
2260  // (~A ^ B) | (A & B) -> (~A ^ B)
2261  // (~A ^ B) | (B & A) -> (~A ^ B)
2262  // (B ^ ~A) | (A & B) -> (B ^ ~A)
2263  // (B ^ ~A) | (B & A) -> (B ^ ~A)
2264  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2265  (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2266  match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2267  return Op0;
2268 
2269  // (~A & B) | ~(A | B) --> ~A
2270  // (~A & B) | ~(B | A) --> ~A
2271  // (B & ~A) | ~(A | B) --> ~A
2272  // (B & ~A) | ~(B | A) --> ~A
2273  if (match(Op0, m_c_And(m_CombineAnd(m_Value(NotA), m_Not(m_Value(A))),
2274  m_Value(B))) &&
2275  match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
2276  return NotA;
2277 
2278  // Commute the 'or' operands.
2279  // ~(A | B) | (~A & B) --> ~A
2280  // ~(B | A) | (~A & B) --> ~A
2281  // ~(A | B) | (B & ~A) --> ~A
2282  // ~(B | A) | (B & ~A) --> ~A
2283  if (match(Op1, m_c_And(m_CombineAnd(m_Value(NotA), m_Not(m_Value(A))),
2284  m_Value(B))) &&
2285  match(Op0, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
2286  return NotA;
2287 
2288  // Rotated -1 is still -1:
2289  // (-1 << X) | (-1 >> (C - X)) --> -1
2290  // (-1 >> X) | (-1 << (C - X)) --> -1
2291  // ...with C <= bitwidth (and commuted variants).
2292  Value *X, *Y;
2293  if ((match(Op0, m_Shl(m_AllOnes(), m_Value(X))) &&
2294  match(Op1, m_LShr(m_AllOnes(), m_Value(Y)))) ||
2295  (match(Op1, m_Shl(m_AllOnes(), m_Value(X))) &&
2296  match(Op0, m_LShr(m_AllOnes(), m_Value(Y))))) {
2297  const APInt *C;
2298  if ((match(X, m_Sub(m_APInt(C), m_Specific(Y))) ||
2299  match(Y, m_Sub(m_APInt(C), m_Specific(X)))) &&
2300  C->ule(X->getType()->getScalarSizeInBits())) {
2301  return ConstantInt::getAllOnesValue(X->getType());
2302  }
2303  }
2304 
2305  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
2306  return V;
2307 
2308  // If we have a multiplication overflow check that is being 'and'ed with a
2309  // check that one of the multipliers is not zero, we can omit the 'and', and
2310  // only keep the overflow check.
2311  if (isCheckForZeroAndMulWithOverflow(Op0, Op1, false))
2312  return Op1;
2313  if (isCheckForZeroAndMulWithOverflow(Op1, Op0, false))
2314  return Op0;
2315 
2316  // Try some generic simplifications for associative operations.
2317  if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
2318  MaxRecurse))
2319  return V;
2320 
2321  // Or distributes over And. Try some generic simplifications based on this.
2322  if (Value *V = expandCommutativeBinOp(Instruction::Or, Op0, Op1,
2323  Instruction::And, Q, MaxRecurse))
2324  return V;
2325 
2326  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) {
2327  if (Op0->getType()->isIntOrIntVectorTy(1)) {
2328  // A | (A || B) -> A || B
2329  if (match(Op1, m_Select(m_Specific(Op0), m_One(), m_Value())))
2330  return Op1;
2331  else if (match(Op0, m_Select(m_Specific(Op1), m_One(), m_Value())))
2332  return Op0;
2333  }
2334  // If the operation is with the result of a select instruction, check
2335  // whether operating on either branch of the select always yields the same
2336  // value.
2337  if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2338  MaxRecurse))
2339  return V;
2340  }
2341 
2342  // (A & C1)|(B & C2)
2343  const APInt *C1, *C2;
2344  if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2345  match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2346  if (*C1 == ~*C2) {
2347  // (A & C1)|(B & C2)
2348  // If we have: ((V + N) & C1) | (V & C2)
2349  // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2350  // replace with V+N.
2351  Value *N;
2352  if (C2->isMask() && // C2 == 0+1+
2353  match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2354  // Add commutes, try both ways.
2355  if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2356  return A;
2357  }
2358  // Or commutes, try both ways.
2359  if (C1->isMask() &&
2360  match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2361  // Add commutes, try both ways.
2362  if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2363  return B;
2364  }
2365  }
2366  }
2367 
2368  // If the operation is with the result of a phi instruction, check whether
2369  // operating on all incoming values of the phi always yields the same value.
2370  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2371  if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2372  return V;
2373 
2374  return nullptr;
2375 }
2376 
2379 }
2380 
2381 /// Given operands for a Xor, see if we can fold the result.
2382 /// If not, this returns null.
2383 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2384  unsigned MaxRecurse) {
2385  if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2386  return C;
2387 
2388  // A ^ undef -> undef
2389  if (Q.isUndefValue(Op1))
2390  return Op1;
2391 
2392  // A ^ 0 = A
2393  if (match(Op1, m_Zero()))
2394  return Op0;
2395 
2396  // A ^ A = 0
2397  if (Op0 == Op1)
2398  return Constant::getNullValue(Op0->getType());
2399 
2400  // A ^ ~A = ~A ^ A = -1
2401  if (match(Op0, m_Not(m_Specific(Op1))) ||
2402  match(Op1, m_Not(m_Specific(Op0))))
2403  return Constant::getAllOnesValue(Op0->getType());
2404 
2405  if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Instruction::Xor))
2406  return V;
2407 
2408  // Try some generic simplifications for associative operations.
2409  if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2410  MaxRecurse))
2411  return V;
2412 
2413  // Threading Xor over selects and phi nodes is pointless, so don't bother.
2414  // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2415  // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2416  // only if B and C are equal. If B and C are equal then (since we assume
2417  // that operands have already been simplified) "select(cond, B, C)" should
2418  // have been simplified to the common value of B and C already. Analysing
2419  // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2420  // for threading over phi nodes.
2421 
2422  return nullptr;
2423 }
2424 
2427 }
2428 
2429 
2431  return CmpInst::makeCmpResultType(Op->getType());
2432 }
2433 
2434 /// Rummage around inside V looking for something equivalent to the comparison
2435 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2436 /// Helper function for analyzing max/min idioms.
2438  Value *LHS, Value *RHS) {
2439  SelectInst *SI = dyn_cast<SelectInst>(V);
2440  if (!SI)
2441  return nullptr;
2442  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2443  if (!Cmp)
2444  return nullptr;
2445  Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2446  if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2447  return Cmp;
2448  if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2449  LHS == CmpRHS && RHS == CmpLHS)
2450  return Cmp;
2451  return nullptr;
2452 }
2453 
2454 // A significant optimization not implemented here is assuming that alloca
2455 // addresses are not equal to incoming argument values. They don't *alias*,
2456 // as we say, but that doesn't mean they aren't equal, so we take a
2457 // conservative approach.
2458 //
2459 // This is inspired in part by C++11 5.10p1:
2460 // "Two pointers of the same type compare equal if and only if they are both
2461 // null, both point to the same function, or both represent the same
2462 // address."
2463 //
2464 // This is pretty permissive.
2465 //
2466 // It's also partly due to C11 6.5.9p6:
2467 // "Two pointers compare equal if and only if both are null pointers, both are
2468 // pointers to the same object (including a pointer to an object and a
2469 // subobject at its beginning) or function, both are pointers to one past the
2470 // last element of the same array object, or one is a pointer to one past the
2471 // end of one array object and the other is a pointer to the start of a
2472 // different array object that happens to immediately follow the first array
2473 // object in the address space.)
2474 //
2475 // C11's version is more restrictive, however there's no reason why an argument
2476 // couldn't be a one-past-the-end value for a stack object in the caller and be
2477 // equal to the beginning of a stack object in the callee.
2478 //
2479 // If the C and C++ standards are ever made sufficiently restrictive in this
2480 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2481 // this optimization.
2482 static Constant *
2484  const SimplifyQuery &Q) {
2485  const DataLayout &DL = Q.DL;
2486  const TargetLibraryInfo *TLI = Q.TLI;
2487  const DominatorTree *DT = Q.DT;
2488  const Instruction *CxtI = Q.CxtI;
2489  const InstrInfoQuery &IIQ = Q.IIQ;
2490 
2491  // First, skip past any trivial no-ops.
2492  LHS = LHS->stripPointerCasts();
2493  RHS = RHS->stripPointerCasts();
2494 
2495  // A non-null pointer is not equal to a null pointer.
2496  if (isa<ConstantPointerNull>(RHS) && ICmpInst::isEquality(Pred) &&
2497  llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2498  IIQ.UseInstrInfo))
2499  return ConstantInt::get(GetCompareTy(LHS),
2500  !CmpInst::isTrueWhenEqual(Pred));
2501 
2502  // We can only fold certain predicates on pointer comparisons.
2503  switch (Pred) {
2504  default:
2505  return nullptr;
2506 
2507  // Equality comaprisons are easy to fold.
2508  case CmpInst::ICMP_EQ:
2509  case CmpInst::ICMP_NE:
2510  break;
2511 
2512  // We can only handle unsigned relational comparisons because 'inbounds' on
2513  // a GEP only protects against unsigned wrapping.
2514  case CmpInst::ICMP_UGT:
2515  case CmpInst::ICMP_UGE:
2516  case CmpInst::ICMP_ULT:
2517  case CmpInst::ICMP_ULE:
2518  // However, we have to switch them to their signed variants to handle
2519  // negative indices from the base pointer.
2520  Pred = ICmpInst::getSignedPredicate(Pred);
2521  break;
2522  }
2523 
2524  // Strip off any constant offsets so that we can reason about them.
2525  // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2526  // here and compare base addresses like AliasAnalysis does, however there are
2527  // numerous hazards. AliasAnalysis and its utilities rely on special rules
2528  // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2529  // doesn't need to guarantee pointer inequality when it says NoAlias.
2530  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2531  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2532 
2533  // If LHS and RHS are related via constant offsets to the same base
2534  // value, we can replace it with an icmp which just compares the offsets.
2535  if (LHS == RHS)
2536  return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2537 
2538  // Various optimizations for (in)equality comparisons.
2539  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2540  // Different non-empty allocations that exist at the same time have
2541  // different addresses (if the program can tell). Global variables always
2542  // exist, so they always exist during the lifetime of each other and all
2543  // allocas. Two different allocas usually have different addresses...
2544  //
2545  // However, if there's an @llvm.stackrestore dynamically in between two
2546  // allocas, they may have the same address. It's tempting to reduce the
2547  // scope of the problem by only looking at *static* allocas here. That would
2548  // cover the majority of allocas while significantly reducing the likelihood
2549  // of having an @llvm.stackrestore pop up in the middle. However, it's not
2550  // actually impossible for an @llvm.stackrestore to pop up in the middle of
2551  // an entry block. Also, if we have a block that's not attached to a
2552  // function, we can't tell if it's "static" under the current definition.
2553  // Theoretically, this problem could be fixed by creating a new kind of
2554  // instruction kind specifically for static allocas. Such a new instruction
2555  // could be required to be at the top of the entry block, thus preventing it
2556  // from being subject to a @llvm.stackrestore. Instcombine could even
2557  // convert regular allocas into these special allocas. It'd be nifty.
2558  // However, until then, this problem remains open.
2559  //
2560  // So, we'll assume that two non-empty allocas have different addresses
2561  // for now.
2562  //
2563  // With all that, if the offsets are within the bounds of their allocations
2564  // (and not one-past-the-end! so we can't use inbounds!), and their
2565  // allocations aren't the same, the pointers are not equal.
2566  //
2567  // Note that it's not necessary to check for LHS being a global variable
2568  // address, due to canonicalization and constant folding.
2569  if (isa<AllocaInst>(LHS) &&
2570  (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2571  ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2572  ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2573  uint64_t LHSSize, RHSSize;
2574  ObjectSizeOpts Opts;
2575  Opts.NullIsUnknownSize =
2576  NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2577  if (LHSOffsetCI && RHSOffsetCI &&
2578  getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2579  getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2580  const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2581  const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2582  if (!LHSOffsetValue.isNegative() &&
2583  !RHSOffsetValue.isNegative() &&
2584  LHSOffsetValue.ult(LHSSize) &&
2585  RHSOffsetValue.ult(RHSSize)) {
2586  return ConstantInt::get(GetCompareTy(LHS),
2587  !CmpInst::isTrueWhenEqual(Pred));
2588  }
2589  }
2590 
2591  // Repeat the above check but this time without depending on DataLayout
2592  // or being able to compute a precise size.
2593  if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2594  !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2595  LHSOffset->isNullValue() &&
2596  RHSOffset->isNullValue())
2597  return ConstantInt::get(GetCompareTy(LHS),
2598  !CmpInst::isTrueWhenEqual(Pred));
2599  }
2600 
2601  // Even if an non-inbounds GEP occurs along the path we can still optimize
2602  // equality comparisons concerning the result. We avoid walking the whole
2603  // chain again by starting where the last calls to
2604  // stripAndComputeConstantOffsets left off and accumulate the offsets.
2605  Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2606  Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2607  if (LHS == RHS)
2608  return ConstantExpr::getICmp(Pred,
2609  ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2610  ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2611 
2612  // If one side of the equality comparison must come from a noalias call
2613  // (meaning a system memory allocation function), and the other side must
2614  // come from a pointer that cannot overlap with dynamically-allocated
2615  // memory within the lifetime of the current function (allocas, byval
2616  // arguments, globals), then determine the comparison result here.
2617  SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2618  getUnderlyingObjects(LHS, LHSUObjs);
2619  getUnderlyingObjects(RHS, RHSUObjs);
2620 
2621  // Is the set of underlying objects all noalias calls?
2622  auto IsNAC = [](ArrayRef<const Value *> Objects) {
2623  return all_of(Objects, isNoAliasCall);
2624  };
2625 
2626  // Is the set of underlying objects all things which must be disjoint from
2627  // noalias calls. For allocas, we consider only static ones (dynamic
2628  // allocas might be transformed into calls to malloc not simultaneously
2629  // live with the compared-to allocation). For globals, we exclude symbols
2630  // that might be resolve lazily to symbols in another dynamically-loaded
2631  // library (and, thus, could be malloc'ed by the implementation).
2632  auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2633  return all_of(Objects, [](const Value *V) {
2634  if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2635  return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2636  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2637  return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2638  GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2639  !GV->isThreadLocal();
2640  if (const Argument *A = dyn_cast<Argument>(V))
2641  return A->hasByValAttr();
2642  return false;
2643  });
2644  };
2645 
2646  if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2647  (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2648  return ConstantInt::get(GetCompareTy(LHS),
2649  !CmpInst::isTrueWhenEqual(Pred));
2650 
2651  // Fold comparisons for non-escaping pointer even if the allocation call
2652  // cannot be elided. We cannot fold malloc comparison to null. Also, the
2653  // dynamic allocation call could be either of the operands.
2654  Value *MI = nullptr;
2655  if (isAllocLikeFn(LHS, TLI) &&
2656  llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2657  MI = LHS;
2658  else if (isAllocLikeFn(RHS, TLI) &&
2659  llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2660  MI = RHS;
2661  // FIXME: We should also fold the compare when the pointer escapes, but the
2662  // compare dominates the pointer escape
2663  if (MI && !PointerMayBeCaptured(MI, true, true))
2664  return ConstantInt::get(GetCompareTy(LHS),
2666  }
2667 
2668  // Otherwise, fail.
2669  return nullptr;
2670 }
2671 
2672 /// Fold an icmp when its operands have i1 scalar type.
2674  Value *RHS, const SimplifyQuery &Q) {
2675  Type *ITy = GetCompareTy(LHS); // The return type.
2676  Type *OpTy = LHS->getType(); // The operand type.
2677  if (!OpTy->isIntOrIntVectorTy(1))
2678  return nullptr;
2679 
2680  // A boolean compared to true/false can be simplified in 14 out of the 20
2681  // (10 predicates * 2 constants) possible combinations. Cases not handled here
2682  // require a 'not' of the LHS, so those must be transformed in InstCombine.
2683  if (match(RHS, m_Zero())) {
2684  switch (Pred) {
2685  case CmpInst::ICMP_NE: // X != 0 -> X
2686  case CmpInst::ICMP_UGT: // X >u 0 -> X
2687  case CmpInst::ICMP_SLT: // X <s 0 -> X
2688  return LHS;
2689 
2690  case CmpInst::ICMP_ULT: // X <u 0 -> false
2691  case CmpInst::ICMP_SGT: // X >s 0 -> false
2692  return getFalse(ITy);
2693 
2694  case CmpInst::ICMP_UGE: // X >=u 0 -> true
2695  case CmpInst::ICMP_SLE: // X <=s 0 -> true
2696  return getTrue(ITy);
2697 
2698  default: break;
2699  }
2700  } else if (match(RHS, m_One())) {
2701  switch (Pred) {
2702  case CmpInst::ICMP_EQ: // X == 1 -> X
2703  case CmpInst::ICMP_UGE: // X >=u 1 -> X
2704  case CmpInst::ICMP_SLE: // X <=s -1 -> X
2705  return LHS;
2706 
2707  case CmpInst::ICMP_UGT: // X >u 1 -> false
2708  case CmpInst::ICMP_SLT: // X <s -1 -> false
2709  return getFalse(ITy);
2710 
2711  case CmpInst::ICMP_ULE: // X <=u 1 -> true
2712  case CmpInst::ICMP_SGE: // X >=s -1 -> true
2713  return getTrue(ITy);
2714 
2715  default: break;
2716  }
2717  }
2718 
2719  switch (Pred) {
2720  default:
2721  break;
2722  case ICmpInst::ICMP_UGE:
2723  if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2724  return getTrue(ITy);
2725  break;
2726  case ICmpInst::ICMP_SGE:
2727  /// For signed comparison, the values for an i1 are 0 and -1
2728  /// respectively. This maps into a truth table of:
2729  /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2730  /// 0 | 0 | 1 (0 >= 0) | 1
2731  /// 0 | 1 | 1 (0 >= -1) | 1
2732  /// 1 | 0 | 0 (-1 >= 0) | 0
2733  /// 1 | 1 | 1 (-1 >= -1) | 1
2734  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2735  return getTrue(ITy);
2736  break;
2737  case ICmpInst::ICMP_ULE:
2738  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2739  return getTrue(ITy);
2740  break;
2741  }
2742 
2743  return nullptr;
2744 }
2745 
2746 /// Try hard to fold icmp with zero RHS because this is a common case.
2748  Value *RHS, const SimplifyQuery &Q) {
2749  if (!match(RHS, m_Zero()))
2750  return nullptr;
2751 
2752  Type *ITy = GetCompareTy(LHS); // The return type.
2753  switch (Pred) {
2754  default:
2755  llvm_unreachable("Unknown ICmp predicate!");
2756  case ICmpInst::ICMP_ULT:
2757  return getFalse(ITy);
2758  case ICmpInst::ICMP_UGE:
2759  return getTrue(ITy);
2760  case ICmpInst::ICMP_EQ:
2761  case ICmpInst::ICMP_ULE:
2762  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2763  return getFalse(ITy);
2764  break;
2765  case ICmpInst::ICMP_NE:
2766  case ICmpInst::ICMP_UGT:
2767  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2768  return getTrue(ITy);
2769  break;
2770  case ICmpInst::ICMP_SLT: {
2771  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2772  if (LHSKnown.isNegative())
2773  return getTrue(ITy);
2774  if (LHSKnown.isNonNegative())
2775  return getFalse(ITy);
2776  break;
2777  }
2778  case ICmpInst::ICMP_SLE: {
2779  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2780  if (LHSKnown.isNegative())
2781  return getTrue(ITy);
2782  if (LHSKnown.isNonNegative() &&
2783  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2784  return getFalse(ITy);
2785  break;
2786  }
2787  case ICmpInst::ICMP_SGE: {
2788  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2789  if (LHSKnown.isNegative())
2790  return getFalse(ITy);
2791  if (LHSKnown.isNonNegative())
2792  return getTrue(ITy);
2793  break;
2794  }
2795  case ICmpInst::ICMP_SGT: {
2796  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2797  if (LHSKnown.isNegative())
2798  return getFalse(ITy);
2799  if (LHSKnown.isNonNegative() &&
2800  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2801  return getTrue(ITy);
2802  break;
2803  }
2804  }
2805 
2806  return nullptr;
2807 }
2808 
2810  Value *RHS, const InstrInfoQuery &IIQ) {
2811  Type *ITy = GetCompareTy(RHS); // The return type.
2812 
2813  Value *X;
2814  // Sign-bit checks can be optimized to true/false after unsigned
2815  // floating-point casts:
2816  // icmp slt (bitcast (uitofp X)), 0 --> false
2817  // icmp sgt (bitcast (uitofp X)), -1 --> true
2818  if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2819  if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2820  return ConstantInt::getFalse(ITy);
2821  if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2822  return ConstantInt::getTrue(ITy);
2823  }
2824 
2825  const APInt *C;
2826  if (!match(RHS, m_APIntAllowUndef(C)))
2827  return nullptr;
2828 
2829  // Rule out tautological comparisons (eg., ult 0 or uge 0).
2831  if (RHS_CR.isEmptySet())
2832  return ConstantInt::getFalse(ITy);
2833  if (RHS_CR.isFullSet())
2834  return ConstantInt::getTrue(ITy);
2835 
2836  ConstantRange LHS_CR = computeConstantRange(LHS, IIQ.UseInstrInfo);
2837  if (!LHS_CR.isFullSet()) {
2838  if (RHS_CR.contains(LHS_CR))
2839  return ConstantInt::getTrue(ITy);
2840  if (RHS_CR.inverse().contains(LHS_CR))
2841  return ConstantInt::getFalse(ITy);
2842  }
2843 
2844  // (mul nuw/nsw X, MulC) != C --> true (if C is not a multiple of MulC)
2845  // (mul nuw/nsw X, MulC) == C --> false (if C is not a multiple of MulC)
2846  const APInt *MulC;
2847  if (ICmpInst::isEquality(Pred) &&
2848  ((match(LHS, m_NUWMul(m_Value(), m_APIntAllowUndef(MulC))) &&
2849  *MulC != 0 && C->urem(*MulC) != 0) ||
2850  (match(LHS, m_NSWMul(m_Value(), m_APIntAllowUndef(MulC))) &&
2851  *MulC != 0 && C->srem(*MulC) != 0)))
2852  return ConstantInt::get(ITy, Pred == ICmpInst::ICMP_NE);
2853 
2854  return nullptr;
2855 }
2856 
2858  CmpInst::Predicate Pred, BinaryOperator *LBO, Value *RHS,
2859  const SimplifyQuery &Q, unsigned MaxRecurse) {
2860  Type *ITy = GetCompareTy(RHS); // The return type.
2861 
2862  Value *Y = nullptr;
2863  // icmp pred (or X, Y), X
2864  if (match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2865  if (Pred == ICmpInst::ICMP_ULT)
2866  return getFalse(ITy);
2867  if (Pred == ICmpInst::ICMP_UGE)
2868  return getTrue(ITy);
2869 
2870  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2871  KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2872  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2873  if (RHSKnown.isNonNegative() && YKnown.isNegative())
2874  return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2875  if (RHSKnown.isNegative() || YKnown.isNonNegative())
2876  return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2877  }
2878  }
2879 
2880  // icmp pred (and X, Y), X
2881  if (match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2882  if (Pred == ICmpInst::ICMP_UGT)
2883  return getFalse(ITy);
2884  if (Pred == ICmpInst::ICMP_ULE)
2885  return getTrue(ITy);
2886  }
2887 
2888  // icmp pred (urem X, Y), Y
2889  if (match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2890  switch (Pred) {
2891  default:
2892  break;
2893  case ICmpInst::ICMP_SGT:
2894  case ICmpInst::ICMP_SGE: {
2895  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2896  if (!Known.isNonNegative())
2897  break;
2899  }
2900  case ICmpInst::ICMP_EQ:
2901  case ICmpInst::ICMP_UGT:
2902  case ICmpInst::ICMP_UGE:
2903  return getFalse(ITy);
2904  case ICmpInst::ICMP_SLT:
2905  case ICmpInst::ICMP_SLE: {
2906  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2907  if (!Known.isNonNegative())
2908  break;
2910  }
2911  case ICmpInst::ICMP_NE:
2912  case ICmpInst::ICMP_ULT:
2913  case ICmpInst::ICMP_ULE:
2914  return getTrue(ITy);
2915  }
2916  }
2917 
2918  // icmp pred (urem X, Y), X
2919  if (match(LBO, m_URem(m_Specific(RHS), m_Value()))) {
2920  if (Pred == ICmpInst::ICMP_ULE)
2921  return getTrue(ITy);
2922  if (Pred == ICmpInst::ICMP_UGT)
2923  return getFalse(ITy);
2924  }
2925 
2926  // x >> y <=u x
2927  // x udiv y <=u x.
2928  if (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2929  match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2930  // icmp pred (X op Y), X
2931  if (Pred == ICmpInst::ICMP_UGT)
2932  return getFalse(ITy);
2933  if (Pred == ICmpInst::ICMP_ULE)
2934  return getTrue(ITy);
2935  }
2936 
2937  // (x*C1)/C2 <= x for C1 <= C2.
2938  // This holds even if the multiplication overflows: Assume that x != 0 and
2939  // arithmetic is modulo M. For overflow to occur we must have C1 >= M/x and
2940  // thus C2 >= M/x. It follows that (x*C1)/C2 <= (M-1)/C2 <= ((M-1)*x)/M < x.
2941  //
2942  // Additionally, either the multiplication and division might be represented
2943  // as shifts:
2944  // (x*C1)>>C2 <= x for C1 < 2**C2.
2945  // (x<<C1)/C2 <= x for 2**C1 < C2.
2946  const APInt *C1, *C2;
2947  if ((match(LBO, m_UDiv(m_Mul(m_Specific(RHS), m_APInt(C1)), m_APInt(C2))) &&
2948  C1->ule(*C2)) ||
2949  (match(LBO, m_LShr(m_Mul(m_Specific(RHS), m_APInt(C1)), m_APInt(C2))) &&
2950  C1->ule(APInt(C2->getBitWidth(), 1) << *C2)) ||
2951  (match(LBO, m_UDiv(m_Shl(m_Specific(RHS), m_APInt(C1)), m_APInt(C2))) &&
2952  (APInt(C1->getBitWidth(), 1) << *C1).ule(*C2))) {
2953  if (Pred == ICmpInst::ICMP_UGT)
2954  return getFalse(ITy);
2955  if (Pred == ICmpInst::ICMP_ULE)
2956  return getTrue(ITy);
2957  }
2958 
2959  return nullptr;
2960 }
2961 
2962 
2963 // If only one of the icmp's operands has NSW flags, try to prove that:
2964 //
2965 // icmp slt (x + C1), (x +nsw C2)
2966 //
2967 // is equivalent to:
2968 //
2969 // icmp slt C1, C2
2970 //
2971 // which is true if x + C2 has the NSW flags set and:
2972 // *) C1 < C2 && C1 >= 0, or
2973 // *) C2 < C1 && C1 <= 0.
2974 //
2976  Value *RHS) {
2977  // TODO: only support icmp slt for now.
2978  if (Pred != CmpInst::ICMP_SLT)
2979  return false;
2980 
2981  // Canonicalize nsw add as RHS.
2982  if (!match(RHS, m_NSWAdd(m_Value(), m_Value())))
2983  std::swap(LHS, RHS);
2984  if (!match(RHS, m_NSWAdd(m_Value(), m_Value())))
2985  return false;
2986 
2987  Value *X;
2988  const APInt *C1, *C2;
2989  if (!match(LHS, m_c_Add(m_Value(X), m_APInt(C1))) ||
2990  !match(RHS, m_c_Add(m_Specific(X), m_APInt(C2))))
2991  return false;
2992 
2993  return (C1->slt(*C2) && C1->isNonNegative()) ||
2994  (C2->slt(*C1) && C1->isNonPositive());
2995 }
2996 
2997 
2998 /// TODO: A large part of this logic is duplicated in InstCombine's
2999 /// foldICmpBinOp(). We should be able to share that and avoid the code
3000 /// duplication.
3002  Value *RHS, const SimplifyQuery &Q,
3003  unsigned MaxRecurse) {
3004  BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
3005  BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
3006  if (MaxRecurse && (LBO || RBO)) {
3007  // Analyze the case when either LHS or RHS is an add instruction.
3008  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3009  // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
3010  bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
3011  if (LBO && LBO->getOpcode() == Instruction::Add) {
3012  A = LBO->getOperand(0);
3013  B = LBO->getOperand(1);
3014  NoLHSWrapProblem =
3015  ICmpInst::isEquality(Pred) ||
3016  (CmpInst::isUnsigned(Pred) &&
3017  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
3018  (CmpInst::isSigned(Pred) &&
3019  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
3020  }
3021  if (RBO && RBO->getOpcode() == Instruction::Add) {
3022  C = RBO->getOperand(0);
3023  D = RBO->getOperand(1);
3024  NoRHSWrapProblem =
3025  ICmpInst::isEquality(Pred) ||
3026  (CmpInst::isUnsigned(Pred) &&
3027  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
3028  (CmpInst::isSigned(Pred) &&
3029  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
3030  }
3031 
3032  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3033  if ((A == RHS || B == RHS) && NoLHSWrapProblem)
3034  if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
3035  Constant::getNullValue(RHS->getType()), Q,
3036  MaxRecurse - 1))
3037  return V;
3038 
3039  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3040  if ((C == LHS || D == LHS) && NoRHSWrapProblem)
3041  if (Value *V =
3043  C == LHS ? D : C, Q, MaxRecurse - 1))
3044  return V;
3045 
3046  // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
3047  bool CanSimplify = (NoLHSWrapProblem && NoRHSWrapProblem) ||
3048  trySimplifyICmpWithAdds(Pred, LHS, RHS);
3049  if (A && C && (A == C || A == D || B == C || B == D) && CanSimplify) {
3050  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3051  Value *Y, *Z;
3052  if (A == C) {
3053  // C + B == C + D -> B == D
3054  Y = B;
3055  Z = D;
3056  } else if (A == D) {
3057  // D + B == C + D -> B == C
3058  Y = B;
3059  Z = C;
3060  } else if (B == C) {
3061  // A + C == C + D -> A == D
3062  Y = A;
3063  Z = D;
3064  } else {
3065  assert(B == D);
3066  // A + D == C + D -> A == C
3067  Y = A;
3068  Z = C;
3069  }
3070  if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
3071  return V;
3072  }
3073  }
3074 
3075  if (LBO)
3076  if (Value *V = simplifyICmpWithBinOpOnLHS(Pred, LBO, RHS, Q, MaxRecurse))
3077  return V;
3078 
3079  if (RBO)
3081  ICmpInst::getSwappedPredicate(Pred), RBO, LHS, Q, MaxRecurse))
3082  return V;
3083 
3084  // 0 - (zext X) pred C
3085  if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
3086  const APInt *C;
3087  if (match(RHS, m_APInt(C))) {
3088  if (C->isStrictlyPositive()) {
3089  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_NE)
3090  return ConstantInt::getTrue(GetCompareTy(RHS));
3091  if (Pred == ICmpInst::ICMP_SGE || Pred == ICmpInst::ICMP_EQ)
3092  return ConstantInt::getFalse(GetCompareTy(RHS));
3093  }
3094  if (C->isNonNegative()) {
3095  if (Pred == ICmpInst::ICMP_SLE)
3096  return ConstantInt::getTrue(GetCompareTy(RHS));
3097  if (Pred == ICmpInst::ICMP_SGT)
3098  return ConstantInt::getFalse(GetCompareTy(RHS));
3099  }
3100  }
3101  }
3102 
3103  // If C2 is a power-of-2 and C is not:
3104  // (C2 << X) == C --> false
3105  // (C2 << X) != C --> true
3106  const APInt *C;
3107  if (match(LHS, m_Shl(m_Power2(), m_Value())) &&
3108  match(RHS, m_APIntAllowUndef(C)) && !C->isPowerOf2()) {
3109  // C2 << X can equal zero in some circumstances.
3110  // This simplification might be unsafe if C is zero.
3111  //
3112  // We know it is safe if:
3113  // - The shift is nsw. We can't shift out the one bit.
3114  // - The shift is nuw. We can't shift out the one bit.
3115  // - C2 is one.
3116  // - C isn't zero.
3117  if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
3118  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
3119  match(LHS, m_Shl(m_One(), m_Value())) || !C->isZero()) {
3120  if (Pred == ICmpInst::ICMP_EQ)
3121  return ConstantInt::getFalse(GetCompareTy(RHS));
3122  if (Pred == ICmpInst::ICMP_NE)
3123  return ConstantInt::getTrue(GetCompareTy(RHS));
3124  }
3125  }
3126 
3127  // TODO: This is overly constrained. LHS can be any power-of-2.
3128  // (1 << X) >u 0x8000 --> false
3129  // (1 << X) <=u 0x8000 --> true
3130  if (match(LHS, m_Shl(m_One(), m_Value())) && match(RHS, m_SignMask())) {
3131  if (Pred == ICmpInst::ICMP_UGT)
3132  return ConstantInt::getFalse(GetCompareTy(RHS));
3133  if (Pred == ICmpInst::ICMP_ULE)
3134  return ConstantInt::getTrue(GetCompareTy(RHS));
3135  }
3136 
3137  if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
3138  LBO->getOperand(1) == RBO->getOperand(1)) {
3139  switch (LBO->getOpcode()) {
3140  default:
3141  break;
3142  case Instruction::UDiv:
3143  case Instruction::LShr:
3144  if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
3145  !Q.IIQ.isExact(RBO))
3146  break;
3147  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3148  RBO->getOperand(0), Q, MaxRecurse - 1))
3149  return V;
3150  break;
3151  case Instruction::SDiv:
3152  if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
3153  !Q.IIQ.isExact(RBO))
3154  break;
3155  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3156  RBO->getOperand(0), Q, MaxRecurse - 1))
3157  return V;
3158  break;
3159  case Instruction::AShr:
3160  if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
3161  break;
3162  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3163  RBO->getOperand(0), Q, MaxRecurse - 1))
3164  return V;
3165  break;
3166  case Instruction::Shl: {
3167  bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
3168  bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
3169  if (!NUW && !NSW)
3170  break;
3171  if (!NSW && ICmpInst::isSigned(Pred))
3172  break;
3173  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3174  RBO->getOperand(0), Q, MaxRecurse - 1))
3175  return V;
3176  break;
3177  }
3178  }
3179  }
3180  return nullptr;
3181 }
3182 
3183 /// Simplify integer comparisons where at least one operand of the compare
3184 /// matches an integer min/max idiom.
3186  Value *RHS, const SimplifyQuery &Q,
3187  unsigned MaxRecurse) {
3188  Type *ITy = GetCompareTy(LHS); // The return type.
3189  Value *A, *B;
3191  CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3192 
3193  // Signed variants on "max(a,b)>=a -> true".
3194  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3195  if (A != RHS)
3196  std::swap(A, B); // smax(A, B) pred A.
3197  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3198  // We analyze this as smax(A, B) pred A.
3199  P = Pred;
3200  } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
3201  (A == LHS || B == LHS)) {
3202  if (A != LHS)
3203  std::swap(A, B); // A pred smax(A, B).
3204  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3205  // We analyze this as smax(A, B) swapped-pred A.
3207  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3208  (A == RHS || B == RHS)) {
3209  if (A != RHS)
3210  std::swap(A, B); // smin(A, B) pred A.
3211  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3212  // We analyze this as smax(-A, -B) swapped-pred -A.
3213  // Note that we do not need to actually form -A or -B thanks to EqP.
3215  } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
3216  (A == LHS || B == LHS)) {
3217  if (A != LHS)
3218  std::swap(A, B); // A pred smin(A, B).
3219  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3220  // We analyze this as smax(-A, -B) pred -A.
3221  // Note that we do not need to actually form -A or -B thanks to EqP.
3222  P = Pred;
3223  }
3224  if (P != CmpInst::BAD_ICMP_PREDICATE) {
3225  // Cases correspond to "max(A, B) p A".
3226  switch (P) {
3227  default:
3228  break;
3229  case CmpInst::ICMP_EQ:
3230  case CmpInst::ICMP_SLE:
3231  // Equivalent to "A EqP B". This may be the same as the condition tested
3232  // in the max/min; if so, we can just return that.
3233  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3234  return V;
3235  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3236  return V;
3237  // Otherwise, see if "A EqP B" simplifies.
3238  if (MaxRecurse)
3239  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3240  return V;
3241  break;
3242  case CmpInst::ICMP_NE:
3243  case CmpInst::ICMP_SGT: {
3245  // Equivalent to "A InvEqP B". This may be the same as the condition
3246  // tested in the max/min; if so, we can just return that.
3247  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3248  return V;
3249  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3250  return V;
3251  // Otherwise, see if "A InvEqP B" simplifies.
3252  if (MaxRecurse)
3253  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3254  return V;
3255  break;
3256  }
3257  case CmpInst::ICMP_SGE:
3258  // Always true.
3259  return getTrue(ITy);
3260  case CmpInst::ICMP_SLT:
3261  // Always false.
3262  return getFalse(ITy);
3263  }
3264  }
3265 
3266  // Unsigned variants on "max(a,b)>=a -> true".
3268  if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3269  if (A != RHS)
3270  std::swap(A, B); // umax(A, B) pred A.
3271  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3272  // We analyze this as umax(A, B) pred A.
3273  P = Pred;
3274  } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3275  (A == LHS || B == LHS)) {
3276  if (A != LHS)
3277  std::swap(A, B); // A pred umax(A, B).
3278  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3279  // We analyze this as umax(A, B) swapped-pred A.
3281  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3282  (A == RHS || B == RHS)) {
3283  if (A != RHS)
3284  std::swap(A, B); // umin(A, B) pred A.
3285  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3286  // We analyze this as umax(-A, -B) swapped-pred -A.
3287  // Note that we do not need to actually form -A or -B thanks to EqP.
3289  } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3290  (A == LHS || B == LHS)) {
3291  if (A != LHS)
3292  std::swap(A, B); // A pred umin(A, B).
3293  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3294  // We analyze this as umax(-A, -B) pred -A.
3295  // Note that we do not need to actually form -A or -B thanks to EqP.
3296  P = Pred;
3297  }
3298  if (P != CmpInst::BAD_ICMP_PREDICATE) {
3299  // Cases correspond to "max(A, B) p A".
3300  switch (P) {
3301  default:
3302  break;
3303  case CmpInst::ICMP_EQ:
3304  case CmpInst::ICMP_ULE:
3305  // Equivalent to "A EqP B". This may be the same as the condition tested
3306  // in the max/min; if so, we can just return that.
3307  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3308  return V;
3309  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3310  return V;
3311  // Otherwise, see if "A EqP B" simplifies.
3312  if (MaxRecurse)
3313  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3314  return V;
3315  break;
3316  case CmpInst::ICMP_NE:
3317  case CmpInst::ICMP_UGT: {
3319  // Equivalent to "A InvEqP B". This may be the same as the condition
3320  // tested in the max/min; if so, we can just return that.
3321  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3322  return V;
3323  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3324  return V;
3325  // Otherwise, see if "A InvEqP B" simplifies.
3326  if (MaxRecurse)
3327  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3328  return V;
3329  break;
3330  }
3331  case CmpInst::ICMP_UGE:
3332  return getTrue(ITy);
3333  case CmpInst::ICMP_ULT:
3334  return getFalse(ITy);
3335  }
3336  }
3337 
3338  // Comparing 1 each of min/max with a common operand?
3339  // Canonicalize min operand to RHS.
3340  if (match(LHS, m_UMin(m_Value(), m_Value())) ||
3341  match(LHS, m_SMin(m_Value(), m_Value()))) {
3342  std::swap(LHS, RHS);
3343  Pred = ICmpInst::getSwappedPredicate(Pred);
3344  }
3345 
3346  Value *C, *D;
3347  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3348  match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3349  (A == C || A == D || B == C || B == D)) {
3350  // smax(A, B) >=s smin(A, D) --> true
3351  if (Pred == CmpInst::ICMP_SGE)
3352  return getTrue(ITy);
3353  // smax(A, B) <s smin(A, D) --> false
3354  if (Pred == CmpInst::ICMP_SLT)
3355  return getFalse(ITy);
3356  } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3357  match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3358  (A == C || A == D || B == C || B == D)) {
3359  // umax(A, B) >=u umin(A, D) --> true
3360  if (Pred == CmpInst::ICMP_UGE)
3361  return getTrue(ITy);
3362  // umax(A, B) <u umin(A, D) --> false
3363  if (Pred == CmpInst::ICMP_ULT)
3364  return getFalse(ITy);
3365  }
3366 
3367  return nullptr;
3368 }
3369 
3371  Value *LHS, Value *RHS,
3372  const SimplifyQuery &Q) {
3373  // Gracefully handle instructions that have not been inserted yet.
3374  if (!Q.AC || !Q.CxtI || !Q.CxtI->getParent())
3375  return nullptr;
3376 
3377  for (Value *AssumeBaseOp : {LHS, RHS}) {
3378  for (auto &AssumeVH : Q.AC->assumptionsFor(AssumeBaseOp)) {
3379  if (!AssumeVH)
3380  continue;
3381 
3382  CallInst *Assume = cast<CallInst>(AssumeVH);
3383  if (Optional<bool> Imp =
3384  isImpliedCondition(Assume->getArgOperand(0), Predicate, LHS, RHS,
3385  Q.DL))
3386  if (isValidAssumeForContext(Assume, Q.CxtI, Q.DT))
3387  return ConstantInt::get(GetCompareTy(LHS), *Imp);
3388  }
3389  }
3390 
3391  return nullptr;
3392 }
3393 
3394 /// Given operands for an ICmpInst, see if we can fold the result.
3395 /// If not, this returns null.
3396 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3397  const SimplifyQuery &Q, unsigned MaxRecurse) {
3399  assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3400 
3401  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3402  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3403  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3404 
3405  // If we have a constant, make sure it is on the RHS.
3406  std::swap(LHS, RHS);
3407  Pred = CmpInst::getSwappedPredicate(Pred);
3408  }
3409  assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3410 
3411  Type *ITy = GetCompareTy(LHS); // The return type.
3412 
3413  // icmp poison, X -> poison
3414  if (isa<PoisonValue>(RHS))
3415  return PoisonValue::get(ITy);
3416 
3417  // For EQ and NE, we can always pick a value for the undef to make the
3418  // predicate pass or fail, so we can return undef.
3419  // Matches behavior in llvm::ConstantFoldCompareInstruction.
3420  if (Q.isUndefValue(RHS) && ICmpInst::isEquality(Pred))
3421  return UndefValue::get(ITy);
3422 
3423  // icmp X, X -> true/false
3424  // icmp X, undef -> true/false because undef could be X.
3425  if (LHS == RHS || Q.isUndefValue(RHS))
3426  return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3427 
3428  if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3429  return V;
3430 
3431  // TODO: Sink/common this with other potentially expensive calls that use
3432  // ValueTracking? See comment below for isKnownNonEqual().
3433  if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3434  return V;
3435 
3436  if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3437  return V;
3438 
3439  // If both operands have range metadata, use the metadata
3440  // to simplify the comparison.
3441  if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3442  auto RHS_Instr = cast<Instruction>(RHS);
3443  auto LHS_Instr = cast<Instruction>(LHS);
3444 
3445  if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3446  Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3447  auto RHS_CR = getConstantRangeFromMetadata(
3448  *RHS_Instr->getMetadata(LLVMContext::MD_range));
3449  auto LHS_CR = getConstantRangeFromMetadata(
3450  *LHS_Instr->getMetadata(LLVMContext::MD_range));
3451 
3452  if (LHS_CR.icmp(Pred, RHS_CR))
3453  return ConstantInt::getTrue(RHS->getContext());
3454 
3455  if (LHS_CR.icmp(CmpInst::getInversePredicate(Pred), RHS_CR))
3456  return ConstantInt::getFalse(RHS->getContext());
3457  }
3458  }
3459 
3460  // Compare of cast, for example (zext X) != 0 -> X != 0
3461  if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3462  Instruction *LI = cast<CastInst>(LHS);
3463  Value *SrcOp = LI->getOperand(0);
3464  Type *SrcTy = SrcOp->getType();
3465  Type *DstTy = LI->getType();
3466 
3467  // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3468  // if the integer type is the same size as the pointer type.
3469  if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3470  Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3471  if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3472  // Transfer the cast to the constant.
3473  if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3474  ConstantExpr::getIntToPtr(RHSC, SrcTy),
3475  Q, MaxRecurse-1))
3476  return V;
3477  } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3478  if (RI->getOperand(0)->getType() == SrcTy)
3479  // Compare without the cast.
3480  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3481  Q, MaxRecurse-1))
3482  return V;
3483  }
3484  }
3485 
3486  if (isa<ZExtInst>(LHS)) {
3487  // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3488  // same type.
3489  if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3490  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3491  // Compare X and Y. Note that signed predicates become unsigned.
3493  SrcOp, RI->getOperand(0), Q,
3494  MaxRecurse-1))
3495  return V;
3496  }
3497  // Fold (zext X) ule (sext X), (zext X) sge (sext X) to true.
3498  else if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3499  if (SrcOp == RI->getOperand(0)) {
3500  if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_SGE)
3501  return ConstantInt::getTrue(ITy);
3502  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SLT)
3503  return ConstantInt::getFalse(ITy);
3504  }
3505  }
3506  // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3507  // too. If not, then try to deduce the result of the comparison.
3508  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3509  // Compute the constant that would happen if we truncated to SrcTy then
3510  // reextended to DstTy.
3511  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3512  Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3513 
3514  // If the re-extended constant didn't change then this is effectively
3515  // also a case of comparing two zero-extended values.
3516  if (RExt == CI && MaxRecurse)
3518  SrcOp, Trunc, Q, MaxRecurse-1))
3519  return V;
3520 
3521  // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3522  // there. Use this to work out the result of the comparison.
3523  if (RExt != CI) {
3524  switch (Pred) {
3525  default: llvm_unreachable("Unknown ICmp predicate!");
3526  // LHS <u RHS.
3527  case ICmpInst::ICMP_EQ:
3528  case ICmpInst::ICMP_UGT:
3529  case ICmpInst::ICMP_UGE:
3530  return ConstantInt::getFalse(CI->getContext());
3531 
3532  case ICmpInst::ICMP_NE:
3533  case ICmpInst::ICMP_ULT:
3534  case ICmpInst::ICMP_ULE:
3535  return ConstantInt::getTrue(CI->getContext());
3536 
3537  // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3538  // is non-negative then LHS <s RHS.
3539  case ICmpInst::ICMP_SGT:
3540  case ICmpInst::ICMP_SGE:
3541  return CI->getValue().isNegative() ?
3542  ConstantInt::getTrue(CI->getContext()) :
3543  ConstantInt::getFalse(CI->getContext());
3544 
3545  case ICmpInst::ICMP_SLT:
3546  case ICmpInst::ICMP_SLE:
3547  return CI->getValue().isNegative() ?
3548  ConstantInt::getFalse(CI->getContext()) :
3549  ConstantInt::getTrue(CI->getContext());
3550  }
3551  }
3552  }
3553  }
3554 
3555  if (isa<SExtInst>(LHS)) {
3556  // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3557  // same type.
3558  if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3559  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3560  // Compare X and Y. Note that the predicate does not change.
3561  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3562  Q, MaxRecurse-1))
3563  return V;
3564  }
3565  // Fold (sext X) uge (zext X), (sext X) sle (zext X) to true.
3566  else if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3567  if (SrcOp == RI->getOperand(0)) {
3568  if (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_SLE)
3569  return ConstantInt::getTrue(ITy);
3570  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SGT)
3571  return ConstantInt::getFalse(ITy);
3572  }
3573  }
3574  // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3575  // too. If not, then try to deduce the result of the comparison.
3576  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3577  // Compute the constant that would happen if we truncated to SrcTy then
3578  // reextended to DstTy.
3579  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3580  Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3581 
3582  // If the re-extended constant didn't change then this is effectively
3583  // also a case of comparing two sign-extended values.
3584  if (RExt == CI && MaxRecurse)
3585  if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3586  return V;
3587 
3588  // Otherwise the upper bits of LHS are all equal, while RHS has varying
3589  // bits there. Use this to work out the result of the comparison.
3590  if (RExt != CI) {
3591  switch (Pred) {
3592  default: llvm_unreachable("Unknown ICmp predicate!");
3593  case ICmpInst::ICMP_EQ:
3594  return ConstantInt::getFalse(CI->getContext());
3595  case ICmpInst::ICMP_NE:
3596  return ConstantInt::getTrue(CI->getContext());
3597 
3598  // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3599  // LHS >s RHS.
3600  case ICmpInst::ICMP_SGT:
3601  case ICmpInst::ICMP_SGE:
3602  return CI->getValue().isNegative() ?
3603  ConstantInt::getTrue(CI->getContext()) :
3604  ConstantInt::getFalse(CI->getContext());
3605  case ICmpInst::ICMP_SLT:
3606  case ICmpInst::ICMP_SLE:
3607  return CI->getValue().isNegative() ?
3608  ConstantInt::getFalse(CI->getContext()) :
3609  ConstantInt::getTrue(CI->getContext());
3610 
3611  // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3612  // LHS >u RHS.
3613  case ICmpInst::ICMP_UGT:
3614  case ICmpInst::ICMP_UGE:
3615  // Comparison is true iff the LHS <s 0.
3616  if (MaxRecurse)
3618  Constant::getNullValue(SrcTy),
3619  Q, MaxRecurse-1))
3620  return V;
3621  break;
3622  case ICmpInst::ICMP_ULT:
3623  case ICmpInst::ICMP_ULE:
3624  // Comparison is true iff the LHS >=s 0.
3625  if (MaxRecurse)
3627  Constant::getNullValue(SrcTy),
3628  Q, MaxRecurse-1))
3629  return V;
3630  break;
3631  }
3632  }
3633  }
3634  }
3635  }
3636 
3637  // icmp eq|ne X, Y -> false|true if X != Y
3638  // This is potentially expensive, and we have already computedKnownBits for
3639  // compares with 0 above here, so only try this for a non-zero compare.
3640  if (ICmpInst::isEquality(Pred) && !match(RHS, m_Zero()) &&
3641  isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3642  return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3643  }
3644 
3645  if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3646  return V;
3647 
3648  if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3649  return V;
3650 
3651  if (Value *V = simplifyICmpWithDominatingAssume(Pred, LHS, RHS, Q))
3652  return V;
3653 
3654  // Simplify comparisons of related pointers using a powerful, recursive
3655  // GEP-walk when we have target data available..
3656  if (LHS->getType()->isPointerTy())
3657  if (auto *C = computePointerICmp(Pred, LHS, RHS, Q))
3658  return C;
3659  if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3660  if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3661  if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3662  Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3663  Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3664  Q.DL.getTypeSizeInBits(CRHS->getType()))
3665  if (auto *C = computePointerICmp(Pred, CLHS->getPointerOperand(),
3666  CRHS->getPointerOperand(), Q))
3667  return C;
3668 
3669  if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3670  if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3671  if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3672  GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3673  (ICmpInst::isEquality(Pred) ||
3674  (GLHS->isInBounds() && GRHS->isInBounds() &&
3675  Pred == ICmpInst::getSignedPredicate(Pred)))) {
3676  // The bases are equal and the indices are constant. Build a constant
3677  // expression GEP with the same indices and a null base pointer to see
3678  // what constant folding can make out of it.
3679  Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3680  SmallVector<Value *, 4> IndicesLHS(GLHS->indices());
3682  GLHS->getSourceElementType(), Null, IndicesLHS);
3683 
3684  SmallVector<Value *, 4> IndicesRHS(GRHS->indices());
3686  GLHS->getSourceElementType(), Null, IndicesRHS);
3687  Constant *NewICmp = ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3688  return ConstantFoldConstant(NewICmp, Q.DL);
3689  }
3690  }
3691  }
3692 
3693  // If the comparison is with the result of a select instruction, check whether
3694  // comparing with either branch of the select always yields the same value.
3695  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3696  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3697  return V;
3698 
3699  // If the comparison is with the result of a phi instruction, check whether
3700  // doing the compare with each incoming phi value yields a common result.
3701  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3702  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3703  return V;
3704 
3705  return nullptr;
3706 }
3707 
3709  const SimplifyQuery &Q) {
3711 }
3712 
3713 /// Given operands for an FCmpInst, see if we can fold the result.
3714 /// If not, this returns null.
3715 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3716  FastMathFlags FMF, const SimplifyQuery &Q,
3717  unsigned MaxRecurse) {
3719  assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3720 
3721  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3722  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3723  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3724 
3725  // If we have a constant, make sure it is on the RHS.
3726  std::swap(LHS, RHS);
3727  Pred = CmpInst::getSwappedPredicate(Pred);
3728  }
3729 
3730  // Fold trivial predicates.
3731  Type *RetTy = GetCompareTy(LHS);
3732  if (Pred == FCmpInst::FCMP_FALSE)
3733  return getFalse(RetTy);
3734  if (Pred == FCmpInst::FCMP_TRUE)
3735  return getTrue(RetTy);
3736 
3737  // Fold (un)ordered comparison if we can determine there are no NaNs.
3738  if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3739  if (FMF.noNaNs() ||
3740  (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3741  return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3742 
3743  // NaN is unordered; NaN is not ordered.
3745  "Comparison must be either ordered or unordered");
3746  if (match(RHS, m_NaN()))
3747  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3748 
3749  // fcmp pred x, poison and fcmp pred poison, x
3750  // fold to poison
3751  if (isa<PoisonValue>(LHS) || isa<PoisonValue>(RHS))
3752  return PoisonValue::get(RetTy);
3753 
3754  // fcmp pred x, undef and fcmp pred undef, x
3755  // fold to true if unordered, false if ordered
3756  if (Q.isUndefValue(LHS) || Q.isUndefValue(RHS)) {
3757  // Choosing NaN for the undef will always make unordered comparison succeed
3758  // and ordered comparison fail.
3759  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3760  }
3761 
3762  // fcmp x,x -> true/false. Not all compares are foldable.
3763  if (LHS == RHS) {
3764  if (CmpInst::isTrueWhenEqual(Pred))
3765  return getTrue(RetTy);
3766  if (CmpInst::isFalseWhenEqual(Pred))
3767  return getFalse(RetTy);
3768  }
3769 
3770  // Handle fcmp with constant RHS.
3771  // TODO: Use match with a specific FP value, so these work with vectors with
3772  // undef lanes.
3773  const APFloat *C;
3774  if (match(RHS, m_APFloat(C))) {
3775  // Check whether the constant is an infinity.
3776  if (C->isInfinity()) {
3777  if (C->isNegative()) {
3778  switch (Pred) {
3779  case FCmpInst::FCMP_OLT:
3780  // No value is ordered and less than negative infinity.
3781  return getFalse(RetTy);
3782  case FCmpInst::FCMP_UGE:
3783  // All values are unordered with or at least negative infinity.
3784  return getTrue(RetTy);
3785  default:
3786  break;
3787  }
3788  } else {
3789  switch (Pred) {
3790  case FCmpInst::FCMP_OGT:
3791  // No value is ordered and greater than infinity.
3792  return getFalse(RetTy);
3793  case FCmpInst::FCMP_ULE:
3794  // All values are unordered with and at most infinity.
3795  return getTrue(RetTy);
3796  default:
3797  break;
3798  }
3799  }
3800 
3801  // LHS == Inf
3802  if (Pred == FCmpInst::FCMP_OEQ && isKnownNeverInfinity(LHS, Q.TLI))
3803  return getFalse(RetTy);
3804  // LHS != Inf
3805  if (Pred == FCmpInst::FCMP_UNE && isKnownNeverInfinity(LHS, Q.TLI))
3806  return getTrue(RetTy);
3807  // LHS == Inf || LHS == NaN
3808  if (Pred == FCmpInst::FCMP_UEQ && isKnownNeverInfinity(LHS, Q.TLI) &&
3809  isKnownNeverNaN(LHS, Q.TLI))
3810  return getFalse(RetTy);
3811  // LHS != Inf && LHS != NaN
3812  if (Pred == FCmpInst::FCMP_ONE && isKnownNeverInfinity(LHS, Q.TLI) &&
3813  isKnownNeverNaN(LHS, Q.TLI))
3814  return getTrue(RetTy);
3815  }
3816  if (C->isNegative() && !C->isNegZero()) {
3817  assert(!C->isNaN() && "Unexpected NaN constant!");
3818  // TODO: We can catch more cases by using a range check rather than
3819  // relying on CannotBeOrderedLessThanZero.
3820  switch (Pred) {
3821  case FCmpInst::FCMP_UGE:
3822  case FCmpInst::FCMP_UGT:
3823  case FCmpInst::FCMP_UNE:
3824  // (X >= 0) implies (X > C) when (C < 0)
3825  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3826  return getTrue(RetTy);
3827  break;
3828  case FCmpInst::FCMP_OEQ:
3829  case FCmpInst::FCMP_OLE:
3830  case FCmpInst::FCMP_OLT:
3831  // (X >= 0) implies !(X < C) when (C < 0)
3832  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3833  return getFalse(RetTy);
3834  break;
3835  default:
3836  break;
3837  }
3838  }
3839 
3840  // Check comparison of [minnum/maxnum with constant] with other constant.
3841  const APFloat *C2;
3842  if ((match(LHS, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_APFloat(C2))) &&
3843  *C2 < *C) ||
3844  (match(LHS, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_APFloat(C2))) &&
3845  *C2 > *C)) {
3846  bool IsMaxNum =
3847  cast<IntrinsicInst>(LHS)->getIntrinsicID() == Intrinsic::maxnum;
3848  // The ordered relationship and minnum/maxnum guarantee that we do not
3849  // have NaN constants, so ordered/unordered preds are handled the same.
3850  switch (Pred) {
3852  // minnum(X, LesserC) == C --> false
3853  // maxnum(X, GreaterC) == C --> false
3854  return getFalse(RetTy);
3856  // minnum(X, LesserC) != C --> true
3857  // maxnum(X, GreaterC) != C --> true
3858  return getTrue(RetTy);
3861  // minnum(X, LesserC) >= C --> false
3862  // minnum(X, LesserC) > C --> false
3863  // maxnum(X, GreaterC) >= C --> true
3864  // maxnum(X, GreaterC) > C --> true
3865  return ConstantInt::get(RetTy, IsMaxNum);
3868  // minnum(X, LesserC) <= C --> true
3869  // minnum(X, LesserC) < C --> true
3870  // maxnum(X, GreaterC) <= C --> false
3871  // maxnum(X, GreaterC) < C --> false
3872  return ConstantInt::get(RetTy, !IsMaxNum);
3873  default:
3874  // TRUE/FALSE/ORD/UNO should be handled before this.
3875  llvm_unreachable("Unexpected fcmp predicate");
3876  }
3877  }
3878  }
3879 
3880  if (match(RHS, m_AnyZeroFP())) {
3881  switch (Pred) {
3882  case FCmpInst::FCMP_OGE:
3883  case FCmpInst::FCMP_ULT:
3884  // Positive or zero X >= 0.0 --> true
3885  // Positive or zero X < 0.0 --> false
3886  if ((FMF.noNaNs() || isKnownNeverNaN(LHS, Q.TLI)) &&
3888  return Pred == FCmpInst::FCMP_OGE ? getTrue(RetTy) : getFalse(RetTy);
3889  break;
3890  case FCmpInst::FCMP_UGE:
3891  case FCmpInst::FCMP_OLT:
3892  // Positive or zero or nan X >= 0.0 --> true
3893  // Positive or zero or nan X < 0.0 --> false
3894  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3895  return Pred == FCmpInst::FCMP_UGE ? getTrue(RetTy) : getFalse(RetTy);
3896  break;
3897  default:
3898  break;
3899  }
3900  }
3901 
3902  // If the comparison is with the result of a select instruction, check whether
3903  // comparing with either branch of the select always yields the same value.
3904  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3905  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3906  return V;
3907 
3908  // If the comparison is with the result of a phi instruction, check whether
3909  // doing the compare with each incoming phi value yields a common result.
3910  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3911  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3912  return V;
3913 
3914  return nullptr;
3915 }
3916 
3918  FastMathFlags FMF, const SimplifyQuery &Q) {
3920 }
3921 
3923  const SimplifyQuery &Q,
3924  bool AllowRefinement,
3925  unsigned MaxRecurse) {
3926  assert(!Op->getType()->isVectorTy() && "This is not safe for vectors");
3927 
3928  // Trivial replacement.
3929  if (V == Op)
3930  return RepOp;
3931 
3932  // We cannot replace a constant, and shouldn't even try.
3933  if (isa<Constant>(Op))
3934  return nullptr;
3935 
3936  auto *I = dyn_cast<Instruction>(V);
3937  if (!I || !is_contained(I->operands(), Op))
3938  return nullptr;
3939 
3940  // Replace Op with RepOp in instruction operands.
3941  SmallVector<Value *, 8> NewOps(I->getNumOperands());
3942  transform(I->operands(), NewOps.begin(),
3943  [&](Value *V) { return V == Op ? RepOp : V; });
3944 
3945  if (!AllowRefinement) {
3946  // General InstSimplify functions may refine the result, e.g. by returning
3947  // a constant for a potentially poison value. To avoid this, implement only
3948  // a few non-refining but profitable transforms here.
3949 
3950  if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3951  unsigned Opcode = BO->getOpcode();
3952  // id op x -> x, x op id -> x
3953  if (NewOps[0] == ConstantExpr::getBinOpIdentity(Opcode, I->getType()))
3954  return NewOps[1];
3955  if (NewOps[1] == ConstantExpr::getBinOpIdentity(Opcode, I->getType(),
3956  /* RHS */ true))
3957  return NewOps[0];
3958 
3959  // x & x -> x, x | x -> x
3960  if ((Opcode == Instruction::And || Opcode == Instruction::Or) &&
3961  NewOps[0] == NewOps[1])
3962  return NewOps[0];
3963  }
3964 
3965  if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3966  // getelementptr x, 0 -> x
3967  if (NewOps.size() == 2 && match(NewOps[1], m_Zero()) &&
3968  !GEP->isInBounds())
3969  return NewOps[0];
3970  }
3971  } else if (MaxRecurse) {
3972  // The simplification queries below may return the original value. Consider:
3973  // %div = udiv i32 %arg, %arg2
3974  // %mul = mul nsw i32 %div, %arg2
3975  // %cmp = icmp eq i32 %mul, %arg
3976  // %sel = select i1 %cmp, i32 %div, i32 undef
3977  // Replacing %arg by %mul, %div becomes "udiv i32 %mul, %arg2", which
3978  // simplifies back to %arg. This can only happen because %mul does not
3979  // dominate %div. To ensure a consistent return value contract, we make sure
3980  // that this case returns nullptr as well.
3981  auto PreventSelfSimplify = [V](Value *Simplified) {
3982  return Simplified != V ? Simplified : nullptr;
3983  };
3984 
3985  if (auto *B = dyn_cast<BinaryOperator>(I))
3986  return PreventSelfSimplify(SimplifyBinOp(B->getOpcode(), NewOps[0],
3987  NewOps[1], Q, MaxRecurse - 1));
3988 
3989  if (CmpInst *C = dyn_cast<CmpInst>(I))
3990  return PreventSelfSimplify(SimplifyCmpInst(C->getPredicate(), NewOps[0],
3991  NewOps[1], Q, MaxRecurse - 1));
3992 
3993  if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
3994  return PreventSelfSimplify(SimplifyGEPInst(GEP->getSourceElementType(),
3995  NewOps, GEP->isInBounds(), Q,
3996  MaxRecurse - 1));
3997 
3998  if (isa<SelectInst>(I))
3999  return PreventSelfSimplify(
4000  SimplifySelectInst(NewOps[0], NewOps[1], NewOps[2], Q,
4001  MaxRecurse - 1));
4002  // TODO: We could hand off more cases to instsimplify here.
4003  }
4004 
4005  // If all operands are constant after substituting Op for RepOp then we can
4006  // constant fold the instruction.
4007  SmallVector<Constant *, 8> ConstOps;
4008  for (Value *NewOp : NewOps) {
4009  if (Constant *ConstOp = dyn_cast<Constant>(NewOp))
4010  ConstOps.push_back(ConstOp);
4011  else
4012  return nullptr;
4013  }
4014 
4015  // Consider:
4016  // %cmp = icmp eq i32 %x, 2147483647
4017  // %add = add nsw i32 %x, 1
4018  // %sel = select i1 %cmp, i32 -2147483648, i32 %add
4019  //
4020  // We can't replace %sel with %add unless we strip away the flags (which
4021  // will be done in InstCombine).
4022  // TODO: This may be unsound, because it only catches some forms of
4023  // refinement.
4024  if (!AllowRefinement && canCreatePoison(cast<Operator>(I)))
4025  return nullptr;
4026 
4027  if (CmpInst *C = dyn_cast<CmpInst>(I))
4028  return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
4029  ConstOps[1], Q.DL, Q.TLI);
4030 
4031  if (LoadInst *LI = dyn_cast<LoadInst>(I))
4032  if (!LI->isVolatile())
4033  return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
4034 
4035  return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
4036 }
4037 
4039  const SimplifyQuery &Q,
4040  bool AllowRefinement) {
4041  return ::simplifyWithOpReplaced(V, Op, RepOp, Q, AllowRefinement,
4042  RecursionLimit);
4043 }
4044 
4045 /// Try to simplify a select instruction when its condition operand is an
4046 /// integer comparison where one operand of the compare is a constant.
4048  const APInt *Y, bool TrueWhenUnset) {
4049  const APInt *C;
4050 
4051  // (X & Y) == 0 ? X & ~Y : X --> X
4052  // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
4053  if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
4054  *Y == ~*C)
4055  return TrueWhenUnset ? FalseVal : TrueVal;
4056 
4057  // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
4058  // (X & Y) != 0 ? X : X & ~Y --> X
4059  if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
4060  *Y == ~*C)
4061  return TrueWhenUnset ? FalseVal : TrueVal;
4062 
4063  if (Y->isPowerOf2()) {
4064  // (X & Y) == 0 ? X | Y : X --> X | Y
4065  // (X & Y) != 0 ? X | Y : X --> X
4066  if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
4067  *Y == *C)
4068  return TrueWhenUnset ? TrueVal : FalseVal;
4069 
4070  // (X & Y) == 0 ? X : X | Y --> X
4071  // (X & Y) != 0 ? X : X | Y --> X | Y
4072  if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
4073  *Y == *C)
4074  return TrueWhenUnset ? TrueVal : FalseVal;
4075  }
4076 
4077  return nullptr;
4078 }
4079 
4080 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
4081 /// eq/ne.
4083  ICmpInst::Predicate Pred,
4084  Value *TrueVal, Value *FalseVal) {
4085  Value *X;
4086  APInt Mask;
4087  if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
4088  return nullptr;
4089 
4091  Pred == ICmpInst::ICMP_EQ);
4092 }
4093 
4094 /// Try to simplify a select instruction when its condition operand is an
4095 /// integer comparison.
4097  Value *FalseVal, const SimplifyQuery &Q,
4098  unsigned MaxRecurse) {
4099  ICmpInst::Predicate Pred;
4100  Value *CmpLHS, *CmpRHS;
4101  if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
4102  return nullptr;
4103 
4104  // Canonicalize ne to eq predicate.
4105  if (Pred == ICmpInst::ICMP_NE) {
4106  Pred = ICmpInst::ICMP_EQ;
4108  }
4109 
4110  // Check for integer min/max with a limit constant:
4111  // X > MIN_INT ? X : MIN_INT --> X
4112  // X < MAX_INT ? X : MAX_INT --> X
4113  if (TrueVal->getType()->isIntOrIntVectorTy()) {
4114  Value *X, *Y;
4115  SelectPatternFlavor SPF =
4116  matchDecomposedSelectPattern(cast<ICmpInst>(CondVal), TrueVal, FalseVal,
4117  X, Y).Flavor;
4118  if (SelectPatternResult::isMinOrMax(SPF) && Pred == getMinMaxPred(SPF)) {
4120  X->getType()->getScalarSizeInBits());
4121  if (match(Y, m_SpecificInt(LimitC)))
4122  return X;
4123  }
4124  }
4125 
4126  if (Pred == ICmpInst::ICMP_EQ && match(CmpRHS, m_Zero())) {
4127  Value *X;
4128  const APInt *Y;
4129  if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
4131  /*TrueWhenUnset=*/true))
4132  return V;
4133 
4134  // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
4135  Value *ShAmt;
4136  auto isFsh = m_CombineOr(m_FShl(m_Value(X), m_Value(), m_Value(ShAmt)),
4137  m_FShr(m_Value(), m_Value(X), m_Value(ShAmt)));
4138  // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
4139  // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
4140  if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt)
4141  return X;
4142 
4143  // Test for a zero-shift-guard-op around rotates. These are used to
4144  // avoid UB from oversized shifts in raw IR rotate patterns, but the
4145  // intrinsics do not have that problem.
4146  // We do not allow this transform for the general funnel shift case because
4147  // that would not preserve the poison safety of the original code.
4148  auto isRotate =
4150  m_FShr(m_Value(X), m_Deferred(X), m_Value(ShAmt)));
4151  // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
4152  // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
4153  if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
4154  Pred == ICmpInst::ICMP_EQ)
4155  return FalseVal;
4156 
4157  // X == 0 ? abs(X) : -abs(X) --> -abs(X)
4158  // X == 0 ? -abs(X) : abs(X) --> abs(X)
4159  if (match(TrueVal, m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS))) &&
4160  match(FalseVal, m_Neg(m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS)))))
4161  return FalseVal;
4162  if (match(TrueVal,
4163  m_Neg(m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS)))) &&
4164  match(FalseVal, m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS))))
4165  return FalseVal;
4166  }
4167 
4168  // Check for other compares that behave like bit test.
4169  if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
4170  TrueVal, FalseVal))
4171  return V;
4172 
4173  // If we have a scalar equality comparison, then we know the value in one of
4174  // the arms of the select. See if substituting this value into the arm and
4175  // simplifying the result yields the same value as the other arm.
4176  // Note that the equivalence/replacement opportunity does not hold for vectors
4177  // because each element of a vector select is chosen independently.
4178  if (Pred == ICmpInst::ICMP_EQ && !CondVal->getType()->isVectorTy()) {
4179  if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q,
4180  /* AllowRefinement */ false, MaxRecurse) ==
4181  TrueVal ||
4182  simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q,
4183  /* AllowRefinement */ false, MaxRecurse) ==
4184  TrueVal)
4185  return FalseVal;
4186  if (simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q,
4187  /* AllowRefinement */ true, MaxRecurse) ==
4188  FalseVal ||
4189  simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q,
4190  /* AllowRefinement */ true, MaxRecurse) ==
4191  FalseVal)
4192  return FalseVal;
4193  }
4194 
4195  return nullptr;
4196 }
4197 
4198 /// Try to simplify a select instruction when its condition operand is a
4199 /// floating-point comparison.
4201  const SimplifyQuery &Q) {
4202  FCmpInst::Predicate Pred;
4203  if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
4204  !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
4205  return nullptr;
4206 
4207  // This transform is safe if we do not have (do not care about) -0.0 or if
4208  // at least one operand is known to not be -0.0. Otherwise, the select can
4209  // change the sign of a zero operand.
4210  bool HasNoSignedZeros = Q.CxtI && isa<FPMathOperator>(Q.CxtI) &&
4211  Q.CxtI->hasNoSignedZeros();
4212  const APFloat *C;
4213  if (HasNoSignedZeros || (match(T, m_APFloat(C)) && C->isNonZero()) ||
4214  (match(F, m_APFloat(C)) && C->isNonZero())) {
4215  // (T == F) ? T : F --> F
4216  // (F == T) ? T : F --> F
4217  if (Pred == FCmpInst::FCMP_OEQ)
4218  return F;
4219 
4220  // (T != F) ? T : F --> T
4221  // (F != T) ? T : F --> T
4222  if (Pred == FCmpInst::FCMP_UNE)
4223  return T;
4224  }
4225 
4226  return nullptr;
4227 }
4228 
4229 /// Given operands for a SelectInst, see if we can fold the result.
4230 /// If not, this returns null.
4232  const SimplifyQuery &Q, unsigned MaxRecurse) {
4233  if (auto *CondC = dyn_cast<Constant>(Cond)) {
4234  if (auto *TrueC = dyn_cast<Constant>(TrueVal))
4235  if (auto *FalseC = dyn_cast<Constant>(FalseVal))
4236  return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
4237 
4238  // select poison, X, Y -> poison
4239  if (isa<PoisonValue>(CondC))
4240  return PoisonValue::get(TrueVal->getType());
4241 
4242  // select undef, X, Y -> X or Y
4243  if (Q.isUndefValue(CondC))
4244  return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
4245 
4246  // select true, X, Y --> X
4247  // select false, X, Y --> Y
4248  // For vectors, allow undef/poison elements in the condition to match the
4249  // defined elements, so we can eliminate the select.
4250  if (match(CondC, m_One()))
4251  return TrueVal;
4252  if (match(CondC, m_Zero()))
4253  return FalseVal;
4254  }
4255 
4256  assert(Cond->getType()->isIntOrIntVectorTy(1) &&
4257  "Select must have bool or bool vector condition");
4258  assert(TrueVal->getType() == FalseVal->getType() &&
4259  "Select must have same types for true/false ops");
4260 
4261  if (Cond->getType() == TrueVal->getType()) {
4262  // select i1 Cond, i1 true, i1 false --> i1 Cond
4263  if (match(TrueVal, m_One()) && match(FalseVal, m_ZeroInt()))
4264  return Cond;
4265 
4266  // (X || Y) && (X || !Y) --> X (commuted 8 ways)
4267  Value *X, *Y;
4268  if (match(FalseVal, m_ZeroInt())) {
4269  if (match(Cond, m_c_LogicalOr(m_Value(X), m_Not(m_Value(Y)))) &&
4271  return X;
4274  return X;
4275  }
4276  }
4277 
4278  // select ?, X, X -> X
4279  if (TrueVal == FalseVal)
4280  return TrueVal;
4281 
4282  // If the true or false value is poison, we can fold to the other value.
4283  // If the true or false value is undef, we can fold to the other value as
4284  // long as the other value isn't poison.
4285  // select ?, poison, X -> X
4286  // select ?, undef, X -> X
4287  if (isa<PoisonValue>(TrueVal) ||
4288  (Q.isUndefValue(TrueVal) &&
4290  return FalseVal;
4291  // select ?, X, poison -> X
4292  // select ?, X, undef -> X
4293  if (isa<PoisonValue>(FalseVal) ||
4294  (Q.isUndefValue(FalseVal) &&
4296  return TrueVal;
4297 
4298  // Deal with partial undef vector constants: select ?, VecC, VecC' --> VecC''
4299  Constant *TrueC, *FalseC;
4300  if (isa<FixedVectorType>(TrueVal->getType()) &&
4301  match(TrueVal, m_Constant(TrueC)) &&
4302  match(FalseVal, m_Constant(FalseC))) {
4303  unsigned NumElts =
4304  cast<FixedVectorType>(TrueC->getType())->getNumElements();
4306  for (unsigned i = 0; i != NumElts; ++i) {
4307  // Bail out on incomplete vector constants.
4308  Constant *TEltC = TrueC->getAggregateElement(i);
4309  Constant *FEltC = FalseC->getAggregateElement(i);
4310  if (!TEltC || !FEltC)
4311  break;
4312 
4313  // If the elements match (undef or not), that value is the result. If only
4314  // one element is undef, choose the defined element as the safe result.
4315  if (TEltC == FEltC)
4316  NewC.push_back(TEltC);
4317  else if (isa<PoisonValue>(TEltC) ||
4318  (Q.isUndefValue(TEltC) && isGuaranteedNotToBePoison(FEltC)))
4319  NewC.push_back(FEltC);
4320  else if (isa<PoisonValue>(FEltC) ||
4321  (Q.isUndefValue(FEltC) && isGuaranteedNotToBePoison(TEltC)))
4322  NewC.push_back(TEltC);
4323  else
4324  break;
4325  }
4326  if (NewC.size() == NumElts)
4327  return ConstantVector::get(NewC);
4328  }
4329 
4330  if (Value *V =
4331  simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
4332  return V;
4333 
4335  return V;
4336 
4338  return V;
4339 
4341  if (Imp)
4342  return *Imp ? TrueVal : FalseVal;
4343 
4344  return nullptr;
4345 }
4346 
4348  const SimplifyQuery &Q) {
4350 }
4351 
4352 /// Given operands for an GetElementPtrInst, see if we can fold the result.
4353 /// If not, this returns null.
4354 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, bool InBounds,
4355  const SimplifyQuery &Q, unsigned) {
4356  // The type of the GEP pointer operand.
4357  unsigned AS =
4358  cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
4359 
4360  // getelementptr P -> P.
4361  if (Ops.size() == 1)
4362  return Ops[0];
4363 
4364  // Compute the (pointer) type returned by the GEP instruction.
4365  Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
4366  Type *GEPTy = PointerType::get(LastType, AS);
4367  for (Value *Op : Ops) {
4368  // If one of the operands is a vector, the result type is a vector of
4369  // pointers. All vector operands must have the same number of elements.
4370  if (VectorType *VT = dyn_cast<VectorType>(Op->getType())) {
4371  GEPTy = VectorType::get(GEPTy, VT->getElementCount());
4372  break;
4373  }
4374  }
4375 
4376  // getelementptr poison, idx -> poison
4377  // getelementptr baseptr, poison -> poison
4378  if (any_of(Ops, [](const auto *V) { return isa<PoisonValue>(V); }))
4379  return PoisonValue::get(GEPTy);
4380 
4381  if (Q.isUndefValue(Ops[0]))
4382  return UndefValue::get(GEPTy);
4383 
4384  bool IsScalableVec =
4385  isa<ScalableVectorType>(SrcTy) || any_of(Ops, [](const Value *V) {
4386  return isa<ScalableVectorType>(V->getType());
4387  });
4388 
4389  if (Ops.size() == 2) {
4390  // getelementptr P, 0 -> P.
4391  if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
4392  return Ops[0];
4393 
4394  Type *Ty = SrcTy;
4395  if (!IsScalableVec && Ty->isSized()) {
4396  Value *P;
4397  uint64_t C;
4398  uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
4399  // getelementptr P, N -> P if P points to a type of zero size.
4400  if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
4401  return Ops[0];
4402 
4403  // The following transforms are only safe if the ptrtoint cast
4404  // doesn't truncate the pointers.
4405  if (Ops[1]->getType()->getScalarSizeInBits() ==
4406  Q.DL.getPointerSizeInBits(AS)) {
4407  auto CanSimplify = [GEPTy, &P, V = Ops[0]]() -> bool {
4408  return P->getType() == GEPTy &&
4410  };
4411  // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
4412  if (TyAllocSize == 1 &&
4413  match(Ops[1], m_Sub(m_PtrToInt(m_Value(P)),
4414  m_PtrToInt(m_Specific(Ops[0])))) &&
4415  CanSimplify())
4416  return P;
4417 
4418  // getelementptr V, (ashr (sub P, V), C) -> P if P points to a type of
4419  // size 1 << C.
4420  if (match(Ops[1], m_AShr(m_Sub(m_PtrToInt(m_Value(P)),
4421  m_PtrToInt(m_Specific(Ops[0]))),
4422  m_ConstantInt(C))) &&
4423  TyAllocSize == 1ULL << C && CanSimplify())
4424  return P;
4425 
4426  // getelementptr V, (sdiv (sub P, V), C) -> P if P points to a type of
4427  // size C.
4428  if (match(Ops[1], m_SDiv(m_Sub(m_PtrToInt(m_Value(P)),
4429  m_PtrToInt(m_Specific(Ops[0]))),
4430  m_SpecificInt(TyAllocSize))) &&
4431  CanSimplify())
4432  return P;
4433  }
4434  }
4435  }
4436 
4437  if (!IsScalableVec && Q.DL.getTypeAllocSize(LastType) == 1 &&
4438  all_of(Ops.slice(1).drop_back(1),
4439  [](Value *Idx) { return match(Idx, m_Zero()); })) {
4440  unsigned IdxWidth =
4441  Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
4442  if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
4443  APInt BasePtrOffset(IdxWidth, 0);
4444  Value *StrippedBasePtr =
4446  BasePtrOffset);
4447 
4448  // Avoid creating inttoptr of zero here: While LLVMs treatment of
4449  // inttoptr is generally conservative, this particular case is folded to
4450  // a null pointer, which will have incorrect provenance.
4451 
4452  // gep (gep V, C), (sub 0, V) -> C
4453  if (match(Ops.back(),
4454  m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr)))) &&
4455  !BasePtrOffset.isZero()) {
4456  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
4457  return ConstantExpr::getIntToPtr(CI, GEPTy);
4458  }
4459  // gep (gep V, C), (xor V, -1) -> C-1
4460  if (match(Ops.back(),
4461  m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes())) &&
4462  !BasePtrOffset.isOne()) {
4463  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
4464  return ConstantExpr::getIntToPtr(CI, GEPTy);
4465  }
4466  }
4467  }
4468 
4469  // Check to see if this is constant foldable.
4470  if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
4471  return nullptr;
4472 
4473  auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
4474  Ops.slice(1), InBounds);
4475  return ConstantFoldConstant(CE, Q.DL);
4476 }
4477 
4479  const SimplifyQuery &Q) {
4480  return ::SimplifyGEPInst(SrcTy, Ops, InBounds, Q, RecursionLimit);
4481 }
4482 
4483 /// Given operands for an InsertValueInst, see if we can fold the result.
4484 /// If not, this returns null.
4486  ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
4487  unsigned) {
4488  if (Constant *CAgg = dyn_cast<Constant>(Agg))
4489  if (Constant *CVal = dyn_cast<Constant>(Val))
4490  return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
4491 
4492  // insertvalue x, undef, n -> x
4493  if (Q.isUndefValue(Val))
4494  return Agg;
4495 
4496  // insertvalue x, (extractvalue y, n), n
4497  if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
4498  if (EV->getAggregateOperand()->getType() == Agg->getType() &&
4499  EV->getIndices() == Idxs) {
4500  // insertvalue undef, (extractvalue y, n), n -> y
4501  if (Q.isUndefValue(Agg))
4502  return EV->getAggregateOperand();
4503 
4504  // insertvalue y, (extractvalue y, n), n -> y
4505  if (Agg == EV->getAggregateOperand())