LLVM  13.0.0git
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 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/Statistic.h"
31 #include "llvm/IR/ConstantRange.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GlobalAlias.h"
36 #include "llvm/IR/InstrTypes.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/ValueHandle.h"
41 #include "llvm/Support/KnownBits.h"
42 #include <algorithm>
43 using namespace llvm;
44 using namespace llvm::PatternMatch;
45 
46 #define DEBUG_TYPE "instsimplify"
47 
48 enum { RecursionLimit = 3 };
49 
50 STATISTIC(NumExpand, "Number of expansions");
51 STATISTIC(NumReassoc, "Number of reassociations");
52 
53 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
54 static Value *simplifyUnOp(unsigned, Value *, const SimplifyQuery &, unsigned);
55 static Value *simplifyFPUnOp(unsigned, Value *, const FastMathFlags &,
56  const SimplifyQuery &, unsigned);
57 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
58  unsigned);
59 static Value *SimplifyBinOp(unsigned, Value *, Value *, const FastMathFlags &,
60  const SimplifyQuery &, unsigned);
61 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
62  unsigned);
63 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
64  const SimplifyQuery &Q, unsigned MaxRecurse);
65 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
66 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
67 static Value *SimplifyCastInst(unsigned, Value *, Type *,
68  const SimplifyQuery &, unsigned);
70  unsigned);
71 static Value *SimplifySelectInst(Value *, Value *, Value *,
72  const SimplifyQuery &, unsigned);
73 
75  Value *FalseVal) {
76  BinaryOperator::BinaryOps BinOpCode;
77  if (auto *BO = dyn_cast<BinaryOperator>(Cond))
78  BinOpCode = BO->getOpcode();
79  else
80  return nullptr;
81 
82  CmpInst::Predicate ExpectedPred, Pred1, Pred2;
83  if (BinOpCode == BinaryOperator::Or) {
84  ExpectedPred = ICmpInst::ICMP_NE;
85  } else if (BinOpCode == BinaryOperator::And) {
86  ExpectedPred = ICmpInst::ICMP_EQ;
87  } else
88  return nullptr;
89 
90  // %A = icmp eq %TV, %FV
91  // %B = icmp eq %X, %Y (and one of these is a select operand)
92  // %C = and %A, %B
93  // %D = select %C, %TV, %FV
94  // -->
95  // %FV
96 
97  // %A = icmp ne %TV, %FV
98  // %B = icmp ne %X, %Y (and one of these is a select operand)
99  // %C = or %A, %B
100  // %D = select %C, %TV, %FV
101  // -->
102  // %TV
103  Value *X, *Y;
106  m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
107  Pred1 != Pred2 || Pred1 != ExpectedPred)
108  return nullptr;
109 
110  if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
111  return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
112 
113  return nullptr;
114 }
115 
116 /// For a boolean type or a vector of boolean type, return false or a vector
117 /// with every element false.
118 static Constant *getFalse(Type *Ty) {
119  return ConstantInt::getFalse(Ty);
120 }
121 
122 /// For a boolean type or a vector of boolean type, return true or a vector
123 /// with every element true.
124 static Constant *getTrue(Type *Ty) {
125  return ConstantInt::getTrue(Ty);
126 }
127 
128 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
129 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
130  Value *RHS) {
131  CmpInst *Cmp = dyn_cast<CmpInst>(V);
132  if (!Cmp)
133  return false;
134  CmpInst::Predicate CPred = Cmp->getPredicate();
135  Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
136  if (CPred == Pred && CLHS == LHS && CRHS == RHS)
137  return true;
138  return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
139  CRHS == LHS;
140 }
141 
142 /// Simplify comparison with true or false branch of select:
143 /// %sel = select i1 %cond, i32 %tv, i32 %fv
144 /// %cmp = icmp sle i32 %sel, %rhs
145 /// Compose new comparison by substituting %sel with either %tv or %fv
146 /// and see if it simplifies.
148  Value *RHS, Value *Cond,
149  const SimplifyQuery &Q, unsigned MaxRecurse,
150  Constant *TrueOrFalse) {
151  Value *SimplifiedCmp = SimplifyCmpInst(Pred, LHS, RHS, Q, MaxRecurse);
152  if (SimplifiedCmp == Cond) {
153  // %cmp simplified to the select condition (%cond).
154  return TrueOrFalse;
155  } else if (!SimplifiedCmp && isSameCompare(Cond, Pred, LHS, RHS)) {
156  // It didn't simplify. However, if composed comparison is equivalent
157  // to the select condition (%cond) then we can replace it.
158  return TrueOrFalse;
159  }
160  return SimplifiedCmp;
161 }
162 
163 /// Simplify comparison with true branch of select
165  Value *RHS, Value *Cond,
166  const SimplifyQuery &Q,
167  unsigned MaxRecurse) {
168  return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
169  getTrue(Cond->getType()));
170 }
171 
172 /// Simplify comparison with false branch of select
174  Value *RHS, Value *Cond,
175  const SimplifyQuery &Q,
176  unsigned MaxRecurse) {
177  return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
178  getFalse(Cond->getType()));
179 }
180 
181 /// We know comparison with both branches of select can be simplified, but they
182 /// are not equal. This routine handles some logical simplifications.
184  Value *Cond,
185  const SimplifyQuery &Q,
186  unsigned MaxRecurse) {
187  // If the false value simplified to false, then the result of the compare
188  // is equal to "Cond && TCmp". This also catches the case when the false
189  // value simplified to false and the true value to true, returning "Cond".
190  if (match(FCmp, m_Zero()))
191  if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
192  return V;
193  // If the true value simplified to true, then the result of the compare
194  // is equal to "Cond || FCmp".
195  if (match(TCmp, m_One()))
196  if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
197  return V;
198  // Finally, if the false value simplified to true and the true value to
199  // false, then the result of the compare is equal to "!Cond".
200  if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
201  if (Value *V = SimplifyXorInst(
202  Cond, Constant::getAllOnesValue(Cond->getType()), Q, MaxRecurse))
203  return V;
204  return nullptr;
205 }
206 
207 /// Does the given value dominate the specified phi node?
208 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
209  Instruction *I = dyn_cast<Instruction>(V);
210  if (!I)
211  // Arguments and constants dominate all instructions.
212  return true;
213 
214  // If we are processing instructions (and/or basic blocks) that have not been
215  // fully added to a function, the parent nodes may still be null. Simply
216  // return the conservative answer in these cases.
217  if (!I->getParent() || !P->getParent() || !I->getFunction())
218  return false;
219 
220  // If we have a DominatorTree then do a precise test.
221  if (DT)
222  return DT->dominates(I, P);
223 
224  // Otherwise, if the instruction is in the entry block and is not an invoke,
225  // then it obviously dominates all phi nodes.
226  if (I->getParent() == &I->getFunction()->getEntryBlock() &&
227  !isa<InvokeInst>(I) && !isa<CallBrInst>(I))
228  return true;
229 
230  return false;
231 }
232 
233 /// Try to simplify a binary operator of form "V op OtherOp" where V is
234 /// "(B0 opex B1)" by distributing 'op' across 'opex' as
235 /// "(B0 op OtherOp) opex (B1 op OtherOp)".
237  Value *OtherOp, Instruction::BinaryOps OpcodeToExpand,
238  const SimplifyQuery &Q, unsigned MaxRecurse) {
239  auto *B = dyn_cast<BinaryOperator>(V);
240  if (!B || B->getOpcode() != OpcodeToExpand)
241  return nullptr;
242  Value *B0 = B->getOperand(0), *B1 = B->getOperand(1);
243  Value *L = SimplifyBinOp(Opcode, B0, OtherOp, Q.getWithoutUndef(),
244  MaxRecurse);
245  if (!L)
246  return nullptr;
247  Value *R = SimplifyBinOp(Opcode, B1, OtherOp, Q.getWithoutUndef(),
248  MaxRecurse);
249  if (!R)
250  return nullptr;
251 
252  // Does the expanded pair of binops simplify to the existing binop?
253  if ((L == B0 && R == B1) ||
254  (Instruction::isCommutative(OpcodeToExpand) && L == B1 && R == B0)) {
255  ++NumExpand;
256  return B;
257  }
258 
259  // Otherwise, return "L op' R" if it simplifies.
260  Value *S = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse);
261  if (!S)
262  return nullptr;
263 
264  ++NumExpand;
265  return S;
266 }
267 
268 /// Try to simplify binops of form "A op (B op' C)" or the commuted variant by
269 /// distributing op over op'.
271  Value *L, Value *R,
272  Instruction::BinaryOps OpcodeToExpand,
273  const SimplifyQuery &Q,
274  unsigned MaxRecurse) {
275  // Recursion is always used, so bail out at once if we already hit the limit.
276  if (!MaxRecurse--)
277  return nullptr;
278 
279  if (Value *V = expandBinOp(Opcode, L, R, OpcodeToExpand, Q, MaxRecurse))
280  return V;
281  if (Value *V = expandBinOp(Opcode, R, L, OpcodeToExpand, Q, MaxRecurse))
282  return V;
283  return nullptr;
284 }
285 
286 /// Generic simplifications for associative binary operations.
287 /// Returns the simpler value, or null if none was found.
289  Value *LHS, Value *RHS,
290  const SimplifyQuery &Q,
291  unsigned MaxRecurse) {
292  assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
293 
294  // Recursion is always used, so bail out at once if we already hit the limit.
295  if (!MaxRecurse--)
296  return nullptr;
297 
298  BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
299  BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
300 
301  // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
302  if (Op0 && Op0->getOpcode() == Opcode) {
303  Value *A = Op0->getOperand(0);
304  Value *B = Op0->getOperand(1);
305  Value *C = RHS;
306 
307  // Does "B op C" simplify?
308  if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
309  // It does! Return "A op V" if it simplifies or is already available.
310  // If V equals B then "A op V" is just the LHS.
311  if (V == B) return LHS;
312  // Otherwise return "A op V" if it simplifies.
313  if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
314  ++NumReassoc;
315  return W;
316  }
317  }
318  }
319 
320  // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
321  if (Op1 && Op1->getOpcode() == Opcode) {
322  Value *A = LHS;
323  Value *B = Op1->getOperand(0);
324  Value *C = Op1->getOperand(1);
325 
326  // Does "A op B" simplify?
327  if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
328  // It does! Return "V op C" if it simplifies or is already available.
329  // If V equals B then "V op C" is just the RHS.
330  if (V == B) return RHS;
331  // Otherwise return "V op C" if it simplifies.
332  if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
333  ++NumReassoc;
334  return W;
335  }
336  }
337  }
338 
339  // The remaining transforms require commutativity as well as associativity.
340  if (!Instruction::isCommutative(Opcode))
341  return nullptr;
342 
343  // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
344  if (Op0 && Op0->getOpcode() == Opcode) {
345  Value *A = Op0->getOperand(0);
346  Value *B = Op0->getOperand(1);
347  Value *C = RHS;
348 
349  // Does "C op A" simplify?
350  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
351  // It does! Return "V op B" if it simplifies or is already available.
352  // If V equals A then "V op B" is just the LHS.
353  if (V == A) return LHS;
354  // Otherwise return "V op B" if it simplifies.
355  if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
356  ++NumReassoc;
357  return W;
358  }
359  }
360  }
361 
362  // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
363  if (Op1 && Op1->getOpcode() == Opcode) {
364  Value *A = LHS;
365  Value *B = Op1->getOperand(0);
366  Value *C = Op1->getOperand(1);
367 
368  // Does "C op A" simplify?
369  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
370  // It does! Return "B op V" if it simplifies or is already available.
371  // If V equals C then "B op V" is just the RHS.
372  if (V == C) return RHS;
373  // Otherwise return "B op V" if it simplifies.
374  if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
375  ++NumReassoc;
376  return W;
377  }
378  }
379  }
380 
381  return nullptr;
382 }
383 
384 /// In the case of a binary operation with a select instruction as an operand,
385 /// try to simplify the binop by seeing whether evaluating it on both branches
386 /// of the select results in the same value. Returns the common value if so,
387 /// otherwise returns null.
389  Value *RHS, const SimplifyQuery &Q,
390  unsigned MaxRecurse) {
391  // Recursion is always used, so bail out at once if we already hit the limit.
392  if (!MaxRecurse--)
393  return nullptr;
394 
395  SelectInst *SI;
396  if (isa<SelectInst>(LHS)) {
397  SI = cast<SelectInst>(LHS);
398  } else {
399  assert(isa<SelectInst>(RHS) && "No select instruction operand!");
400  SI = cast<SelectInst>(RHS);
401  }
402 
403  // Evaluate the BinOp on the true and false branches of the select.
404  Value *TV;
405  Value *FV;
406  if (SI == LHS) {
407  TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
408  FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
409  } else {
410  TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
411  FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
412  }
413 
414  // If they simplified to the same value, then return the common value.
415  // If they both failed to simplify then return null.
416  if (TV == FV)
417  return TV;
418 
419  // If one branch simplified to undef, return the other one.
420  if (TV && Q.isUndefValue(TV))
421  return FV;
422  if (FV && Q.isUndefValue(FV))
423  return TV;
424 
425  // If applying the operation did not change the true and false select values,
426  // then the result of the binop is the select itself.
427  if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
428  return SI;
429 
430  // If one branch simplified and the other did not, and the simplified
431  // value is equal to the unsimplified one, return the simplified value.
432  // For example, select (cond, X, X & Z) & Z -> X & Z.
433  if ((FV && !TV) || (TV && !FV)) {
434  // Check that the simplified value has the form "X op Y" where "op" is the
435  // same as the original operation.
436  Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
437  if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
438  // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
439  // We already know that "op" is the same as for the simplified value. See
440  // if the operands match too. If so, return the simplified value.
441  Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
442  Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
443  Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
444  if (Simplified->getOperand(0) == UnsimplifiedLHS &&
445  Simplified->getOperand(1) == UnsimplifiedRHS)
446  return Simplified;
447  if (Simplified->isCommutative() &&
448  Simplified->getOperand(1) == UnsimplifiedLHS &&
449  Simplified->getOperand(0) == UnsimplifiedRHS)
450  return Simplified;
451  }
452  }
453 
454  return nullptr;
455 }
456 
457 /// In the case of a comparison with a select instruction, try to simplify the
458 /// comparison by seeing whether both branches of the select result in the same
459 /// value. Returns the common value if so, otherwise returns null.
460 /// For example, if we have:
461 /// %tmp = select i1 %cmp, i32 1, i32 2
462 /// %cmp1 = icmp sle i32 %tmp, 3
463 /// We can simplify %cmp1 to true, because both branches of select are
464 /// less than 3. We compose new comparison by substituting %tmp with both
465 /// branches of select and see if it can be simplified.
467  Value *RHS, const SimplifyQuery &Q,
468  unsigned MaxRecurse) {
469  // Recursion is always used, so bail out at once if we already hit the limit.
470  if (!MaxRecurse--)
471  return nullptr;
472 
473  // Make sure the select is on the LHS.
474  if (!isa<SelectInst>(LHS)) {
475  std::swap(LHS, RHS);
476  Pred = CmpInst::getSwappedPredicate(Pred);
477  }
478  assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
479  SelectInst *SI = cast<SelectInst>(LHS);
480  Value *Cond = SI->getCondition();
481  Value *TV = SI->getTrueValue();
482  Value *FV = SI->getFalseValue();
483 
484  // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
485  // Does "cmp TV, RHS" simplify?
486  Value *TCmp = simplifyCmpSelTrueCase(Pred, TV, RHS, Cond, Q, MaxRecurse);
487  if (!TCmp)
488  return nullptr;
489 
490  // Does "cmp FV, RHS" simplify?
491  Value *FCmp = simplifyCmpSelFalseCase(Pred, FV, RHS, Cond, Q, MaxRecurse);
492  if (!FCmp)
493  return nullptr;
494 
495  // If both sides simplified to the same value, then use it as the result of
496  // the original comparison.
497  if (TCmp == FCmp)
498  return TCmp;
499 
500  // The remaining cases only make sense if the select condition has the same
501  // type as the result of the comparison, so bail out if this is not so.
502  if (Cond->getType()->isVectorTy() == RHS->getType()->isVectorTy())
503  return handleOtherCmpSelSimplifications(TCmp, FCmp, Cond, Q, MaxRecurse);
504 
505  return nullptr;
506 }
507 
508 /// In the case of a binary operation with an operand that is a PHI instruction,
509 /// try to simplify the binop by seeing whether evaluating it on the incoming
510 /// phi values yields the same result for every value. If so returns the common
511 /// value, otherwise returns null.
513  Value *RHS, const SimplifyQuery &Q,
514  unsigned MaxRecurse) {
515  // Recursion is always used, so bail out at once if we already hit the limit.
516  if (!MaxRecurse--)
517  return nullptr;
518 
519  PHINode *PI;
520  if (isa<PHINode>(LHS)) {
521  PI = cast<PHINode>(LHS);
522  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
523  if (!valueDominatesPHI(RHS, PI, Q.DT))
524  return nullptr;
525  } else {
526  assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
527  PI = cast<PHINode>(RHS);
528  // Bail out if LHS and the phi may be mutually interdependent due to a loop.
529  if (!valueDominatesPHI(LHS, PI, Q.DT))
530  return nullptr;
531  }
532 
533  // Evaluate the BinOp on the incoming phi values.
534  Value *CommonValue = nullptr;
535  for (Value *Incoming : PI->incoming_values()) {
536  // If the incoming value is the phi node itself, it can safely be skipped.
537  if (Incoming == PI) continue;
538  Value *V = PI == LHS ?
539  SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
540  SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
541  // If the operation failed to simplify, or simplified to a different value
542  // to previously, then give up.
543  if (!V || (CommonValue && V != CommonValue))
544  return nullptr;
545  CommonValue = V;
546  }
547 
548  return CommonValue;
549 }
550 
551 /// In the case of a comparison with a PHI instruction, try to simplify the
552 /// comparison by seeing whether comparing with all of the incoming phi values
553 /// yields the same result every time. If so returns the common result,
554 /// otherwise returns null.
556  const SimplifyQuery &Q, unsigned MaxRecurse) {
557  // Recursion is always used, so bail out at once if we already hit the limit.
558  if (!MaxRecurse--)
559  return nullptr;
560 
561  // Make sure the phi is on the LHS.
562  if (!isa<PHINode>(LHS)) {
563  std::swap(LHS, RHS);
564  Pred = CmpInst::getSwappedPredicate(Pred);
565  }
566  assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
567  PHINode *PI = cast<PHINode>(LHS);
568 
569  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
570  if (!valueDominatesPHI(RHS, PI, Q.DT))
571  return nullptr;
572 
573  // Evaluate the BinOp on the incoming phi values.
574  Value *CommonValue = nullptr;
575  for (unsigned u = 0, e = PI->getNumIncomingValues(); u < e; ++u) {
576  Value *Incoming = PI->getIncomingValue(u);
577  Instruction *InTI = PI->getIncomingBlock(u)->getTerminator();
578  // If the incoming value is the phi node itself, it can safely be skipped.
579  if (Incoming == PI) continue;
580  // Change the context instruction to the "edge" that flows into the phi.
581  // This is important because that is where incoming is actually "evaluated"
582  // even though it is used later somewhere else.
583  Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q.getWithInstruction(InTI),
584  MaxRecurse);
585  // If the operation failed to simplify, or simplified to a different value
586  // to previously, then give up.
587  if (!V || (CommonValue && V != CommonValue))
588  return nullptr;
589  CommonValue = V;
590  }
591 
592  return CommonValue;
593 }
594 
596  Value *&Op0, Value *&Op1,
597  const SimplifyQuery &Q) {
598  if (auto *CLHS = dyn_cast<Constant>(Op0)) {
599  if (auto *CRHS = dyn_cast<Constant>(Op1))
600  return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
601 
602  // Canonicalize the constant to the RHS if this is a commutative operation.
603  if (Instruction::isCommutative(Opcode))
604  std::swap(Op0, Op1);
605  }
606  return nullptr;
607 }
608 
609 /// Given operands for an Add, see if we can fold the result.
610 /// If not, this returns null.
611 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
612  const SimplifyQuery &Q, unsigned MaxRecurse) {
613  if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
614  return C;
615 
616  // X + undef -> undef
617  if (Q.isUndefValue(Op1))
618  return Op1;
619 
620  // X + 0 -> X
621  if (match(Op1, m_Zero()))
622  return Op0;
623 
624  // If two operands are negative, return 0.
625  if (isKnownNegation(Op0, Op1))
626  return Constant::getNullValue(Op0->getType());
627 
628  // X + (Y - X) -> Y
629  // (Y - X) + X -> Y
630  // Eg: X + -X -> 0
631  Value *Y = nullptr;
632  if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
633  match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
634  return Y;
635 
636  // X + ~X -> -1 since ~X = -X-1
637  Type *Ty = Op0->getType();
638  if (match(Op0, m_Not(m_Specific(Op1))) ||
639  match(Op1, m_Not(m_Specific(Op0))))
640  return Constant::getAllOnesValue(Ty);
641 
642  // add nsw/nuw (xor Y, signmask), signmask --> Y
643  // The no-wrapping add guarantees that the top bit will be set by the add.
644  // Therefore, the xor must be clearing the already set sign bit of Y.
645  if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
646  match(Op0, m_Xor(m_Value(Y), m_SignMask())))
647  return Y;
648 
649  // add nuw %x, -1 -> -1, because %x can only be 0.
650  if (IsNUW && match(Op1, m_AllOnes()))
651  return Op1; // Which is -1.
652 
653  /// i1 add -> xor.
654  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
655  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
656  return V;
657 
658  // Try some generic simplifications for associative operations.
659  if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
660  MaxRecurse))
661  return V;
662 
663  // Threading Add over selects and phi nodes is pointless, so don't bother.
664  // Threading over the select in "A + select(cond, B, C)" means evaluating
665  // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
666  // only if B and C are equal. If B and C are equal then (since we assume
667  // that operands have already been simplified) "select(cond, B, C)" should
668  // have been simplified to the common value of B and C already. Analysing
669  // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
670  // for threading over phi nodes.
671 
672  return nullptr;
673 }
674 
675 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
676  const SimplifyQuery &Query) {
677  return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
678 }
679 
680 /// Compute the base pointer and cumulative constant offsets for V.
681 ///
682 /// This strips all constant offsets off of V, leaving it the base pointer, and
683 /// accumulates the total constant offset applied in the returned constant. It
684 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
685 /// no constant offsets applied.
686 ///
687 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
688 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
689 /// folding.
691  bool AllowNonInbounds = false) {
693 
694  Type *IntIdxTy = DL.getIndexType(V->getType())->getScalarType();
696 
697  V = V->stripAndAccumulateConstantOffsets(DL, Offset, AllowNonInbounds);
698  // As that strip may trace through `addrspacecast`, need to sext or trunc
699  // the offset calculated.
700  IntIdxTy = DL.getIndexType(V->getType())->getScalarType();
701  Offset = Offset.sextOrTrunc(IntIdxTy->getIntegerBitWidth());
702 
703  Constant *OffsetIntPtr = ConstantInt::get(IntIdxTy, Offset);
704  if (VectorType *VecTy = dyn_cast<VectorType>(V->getType()))
705  return ConstantVector::getSplat(VecTy->getElementCount(), OffsetIntPtr);
706  return OffsetIntPtr;
707 }
708 
709 /// Compute the constant difference between two pointer values.
710 /// If the difference is not a constant, returns zero.
712  Value *RHS) {
713  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
714  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
715 
716  // If LHS and RHS are not related via constant offsets to the same base
717  // value, there is nothing we can do here.
718  if (LHS != RHS)
719  return nullptr;
720 
721  // Otherwise, the difference of LHS - RHS can be computed as:
722  // LHS - RHS
723  // = (LHSOffset + Base) - (RHSOffset + Base)
724  // = LHSOffset - RHSOffset
725  return ConstantExpr::getSub(LHSOffset, RHSOffset);
726 }
727 
728 /// Given operands for a Sub, see if we can fold the result.
729 /// If not, this returns null.
730 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
731  const SimplifyQuery &Q, unsigned MaxRecurse) {
732  if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
733  return C;
734 
735  // X - undef -> undef
736  // undef - X -> undef
737  if (Q.isUndefValue(Op0) || Q.isUndefValue(Op1))
738  return UndefValue::get(Op0->getType());
739 
740  // X - 0 -> X
741  if (match(Op1, m_Zero()))
742  return Op0;
743 
744  // X - X -> 0
745  if (Op0 == Op1)
746  return Constant::getNullValue(Op0->getType());
747 
748  // Is this a negation?
749  if (match(Op0, m_Zero())) {
750  // 0 - X -> 0 if the sub is NUW.
751  if (isNUW)
752  return Constant::getNullValue(Op0->getType());
753 
754  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
755  if (Known.Zero.isMaxSignedValue()) {
756  // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
757  // Op1 must be 0 because negating the minimum signed value is undefined.
758  if (isNSW)
759  return Constant::getNullValue(Op0->getType());
760 
761  // 0 - X -> X if X is 0 or the minimum signed value.
762  return Op1;
763  }
764  }
765 
766  // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
767  // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
768  Value *X = nullptr, *Y = nullptr, *Z = Op1;
769  if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
770  // See if "V === Y - Z" simplifies.
771  if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
772  // It does! Now see if "X + V" simplifies.
773  if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
774  // It does, we successfully reassociated!
775  ++NumReassoc;
776  return W;
777  }
778  // See if "V === X - Z" simplifies.
779  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
780  // It does! Now see if "Y + V" simplifies.
781  if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
782  // It does, we successfully reassociated!
783  ++NumReassoc;
784  return W;
785  }
786  }
787 
788  // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
789  // For example, X - (X + 1) -> -1
790  X = Op0;
791  if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
792  // See if "V === X - Y" simplifies.
793  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
794  // It does! Now see if "V - Z" simplifies.
795  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
796  // It does, we successfully reassociated!
797  ++NumReassoc;
798  return W;
799  }
800  // See if "V === X - Z" simplifies.
801  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
802  // It does! Now see if "V - Y" simplifies.
803  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
804  // It does, we successfully reassociated!
805  ++NumReassoc;
806  return W;
807  }
808  }
809 
810  // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
811  // For example, X - (X - Y) -> Y.
812  Z = Op0;
813  if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
814  // See if "V === Z - X" simplifies.
815  if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
816  // It does! Now see if "V + Y" simplifies.
817  if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
818  // It does, we successfully reassociated!
819  ++NumReassoc;
820  return W;
821  }
822 
823  // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
824  if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
825  match(Op1, m_Trunc(m_Value(Y))))
826  if (X->getType() == Y->getType())
827  // See if "V === X - Y" simplifies.
828  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
829  // It does! Now see if "trunc V" simplifies.
830  if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
831  Q, MaxRecurse - 1))
832  // It does, return the simplified "trunc V".
833  return W;
834 
835  // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
836  if (match(Op0, m_PtrToInt(m_Value(X))) &&
837  match(Op1, m_PtrToInt(m_Value(Y))))
838  if (Constant *Result = computePointerDifference(Q.DL, X, Y))
839  return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
840 
841  // i1 sub -> xor.
842  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
843  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
844  return V;
845 
846  // Threading Sub over selects and phi nodes is pointless, so don't bother.
847  // Threading over the select in "A - select(cond, B, C)" means evaluating
848  // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
849  // only if B and C are equal. If B and C are equal then (since we assume
850  // that operands have already been simplified) "select(cond, B, C)" should
851  // have been simplified to the common value of B and C already. Analysing
852  // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
853  // for threading over phi nodes.
854 
855  return nullptr;
856 }
857 
858 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
859  const SimplifyQuery &Q) {
860  return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
861 }
862 
863 /// Given operands for a Mul, see if we can fold the result.
864 /// If not, this returns null.
865 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
866  unsigned MaxRecurse) {
867  if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
868  return C;
869 
870  // X * undef -> 0
871  // X * 0 -> 0
872  if (Q.isUndefValue(Op1) || match(Op1, m_Zero()))
873  return Constant::getNullValue(Op0->getType());
874 
875  // X * 1 -> X
876  if (match(Op1, m_One()))
877  return Op0;
878 
879  // (X / Y) * Y -> X if the division is exact.
880  Value *X = nullptr;
881  if (Q.IIQ.UseInstrInfo &&
882  (match(Op0,
883  m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
884  match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
885  return X;
886 
887  // i1 mul -> and.
888  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
889  if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
890  return V;
891 
892  // Try some generic simplifications for associative operations.
893  if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
894  MaxRecurse))
895  return V;
896 
897  // Mul distributes over Add. Try some generic simplifications based on this.
898  if (Value *V = expandCommutativeBinOp(Instruction::Mul, Op0, Op1,
899  Instruction::Add, Q, MaxRecurse))
900  return V;
901 
902  // If the operation is with the result of a select instruction, check whether
903  // operating on either branch of the select always yields the same value.
904  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
905  if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
906  MaxRecurse))
907  return V;
908 
909  // If the operation is with the result of a phi instruction, check whether
910  // operating on all incoming values of the phi always yields the same value.
911  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
912  if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
913  MaxRecurse))
914  return V;
915 
916  return nullptr;
917 }
918 
921 }
922 
923 /// Check for common or similar folds of integer division or integer remainder.
924 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
925 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv,
926  const SimplifyQuery &Q) {
927  Type *Ty = Op0->getType();
928 
929  // X / undef -> poison
930  // X % undef -> poison
931  if (Q.isUndefValue(Op1))
932  return PoisonValue::get(Ty);
933 
934  // X / 0 -> poison
935  // X % 0 -> poison
936  // We don't need to preserve faults!
937  if (match(Op1, m_Zero()))
938  return PoisonValue::get(Ty);
939 
940  // If any element of a constant divisor fixed width vector is zero or undef
941  // the behavior is undefined and we can fold the whole op to poison.
942  auto *Op1C = dyn_cast<Constant>(Op1);
943  auto *VTy = dyn_cast<FixedVectorType>(Ty);
944  if (Op1C && VTy) {
945  unsigned NumElts = VTy->getNumElements();
946  for (unsigned i = 0; i != NumElts; ++i) {
947  Constant *Elt = Op1C->getAggregateElement(i);
948  if (Elt && (Elt->isNullValue() || Q.isUndefValue(Elt)))
949  return PoisonValue::get(Ty);
950  }
951  }
952 
953  // undef / X -> 0
954  // undef % X -> 0
955  if (Q.isUndefValue(Op0))
956  return Constant::getNullValue(Ty);
957 
958  // 0 / X -> 0
959  // 0 % X -> 0
960  if (match(Op0, m_Zero()))
961  return Constant::getNullValue(Op0->getType());
962 
963  // X / X -> 1
964  // X % X -> 0
965  if (Op0 == Op1)
966  return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
967 
968  // X / 1 -> X
969  // X % 1 -> 0
970  // If this is a boolean op (single-bit element type), we can't have
971  // division-by-zero or remainder-by-zero, so assume the divisor is 1.
972  // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
973  Value *X;
974  if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
975  (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
976  return IsDiv ? Op0 : Constant::getNullValue(Ty);
977 
978  return nullptr;
979 }
980 
981 /// Given a predicate and two operands, return true if the comparison is true.
982 /// This is a helper for div/rem simplification where we return some other value
983 /// when we can prove a relationship between the operands.
984 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
985  const SimplifyQuery &Q, unsigned MaxRecurse) {
986  Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
987  Constant *C = dyn_cast_or_null<Constant>(V);
988  return (C && C->isAllOnesValue());
989 }
990 
991 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
992 /// to simplify X % Y to X.
993 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
994  unsigned MaxRecurse, bool IsSigned) {
995  // Recursion is always used, so bail out at once if we already hit the limit.
996  if (!MaxRecurse--)
997  return false;
998 
999  if (IsSigned) {
1000  // |X| / |Y| --> 0
1001  //
1002  // We require that 1 operand is a simple constant. That could be extended to
1003  // 2 variables if we computed the sign bit for each.
1004  //
1005  // Make sure that a constant is not the minimum signed value because taking
1006  // the abs() of that is undefined.
1007  Type *Ty = X->getType();
1008  const APInt *C;
1009  if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
1010  // Is the variable divisor magnitude always greater than the constant
1011  // dividend magnitude?
1012  // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
1013  Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
1014  Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
1015  if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
1016  isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
1017  return true;
1018  }
1019  if (match(Y, m_APInt(C))) {
1020  // Special-case: we can't take the abs() of a minimum signed value. If
1021  // that's the divisor, then all we have to do is prove that the dividend
1022  // is also not the minimum signed value.
1023  if (C->isMinSignedValue())
1024  return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1025 
1026  // Is the variable dividend magnitude always less than the constant
1027  // divisor magnitude?
1028  // |X| < |C| --> X > -abs(C) and X < abs(C)
1029  Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1030  Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1031  if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1032  isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1033  return true;
1034  }
1035  return false;
1036  }
1037 
1038  // IsSigned == false.
1039  // Is the dividend unsigned less than the divisor?
1040  return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1041 }
1042 
1043 /// These are simplifications common to SDiv and UDiv.
1045  const SimplifyQuery &Q, unsigned MaxRecurse) {
1046  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1047  return C;
1048 
1049  if (Value *V = simplifyDivRem(Op0, Op1, true, Q))
1050  return V;
1051 
1052  bool IsSigned = Opcode == Instruction::SDiv;
1053 
1054  // (X * Y) / Y -> X if the multiplication does not overflow.
1055  Value *X;
1056  if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1057  auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1058  // If the Mul does not overflow, then we are good to go.
1059  if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1060  (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1061  return X;
1062  // If X has the form X = A / Y, then X * Y cannot overflow.
1063  if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1064  (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1065  return X;
1066  }
1067 
1068  // (X rem Y) / Y -> 0
1069  if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1070  (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1071  return Constant::getNullValue(Op0->getType());
1072 
1073  // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1074  ConstantInt *C1, *C2;
1075  if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1076  match(Op1, m_ConstantInt(C2))) {
1077  bool Overflow;
1078  (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1079  if (Overflow)
1080  return Constant::getNullValue(Op0->getType());
1081  }
1082 
1083  // If the operation is with the result of a select instruction, check whether
1084  // operating on either branch of the select always yields the same value.
1085  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1086  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1087  return V;
1088 
1089  // If the operation is with the result of a phi instruction, check whether
1090  // operating on all incoming values of the phi always yields the same value.
1091  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1092  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1093  return V;
1094 
1095  if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1096  return Constant::getNullValue(Op0->getType());
1097 
1098  return nullptr;
1099 }
1100 
1101 /// These are simplifications common to SRem and URem.
1103  const SimplifyQuery &Q, unsigned MaxRecurse) {
1104  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1105  return C;
1106 
1107  if (Value *V = simplifyDivRem(Op0, Op1, false, Q))
1108  return V;
1109 
1110  // (X % Y) % Y -> X % Y
1111  if ((Opcode == Instruction::SRem &&
1112  match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1113  (Opcode == Instruction::URem &&
1114  match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1115  return Op0;
1116 
1117  // (X << Y) % X -> 0
1118  if (Q.IIQ.UseInstrInfo &&
1119  ((Opcode == Instruction::SRem &&
1120  match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1121  (Opcode == Instruction::URem &&
1122  match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1123  return Constant::getNullValue(Op0->getType());
1124 
1125  // If the operation is with the result of a select instruction, check whether
1126  // operating on either branch of the select always yields the same value.
1127  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1128  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1129  return V;
1130 
1131  // If the operation is with the result of a phi instruction, check whether
1132  // operating on all incoming values of the phi always yields the same value.
1133  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1134  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1135  return V;
1136 
1137  // If X / Y == 0, then X % Y == X.
1138  if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1139  return Op0;
1140 
1141  return nullptr;
1142 }
1143 
1144 /// Given operands for an SDiv, see if we can fold the result.
1145 /// If not, this returns null.
1146 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1147  unsigned MaxRecurse) {
1148  // If two operands are negated and no signed overflow, return -1.
1149  if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1150  return Constant::getAllOnesValue(Op0->getType());
1151 
1152  return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1153 }
1154 
1157 }
1158 
1159 /// Given operands for a UDiv, see if we can fold the result.
1160 /// If not, this returns null.
1161 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1162  unsigned MaxRecurse) {
1163  return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1164 }
1165 
1168 }
1169 
1170 /// Given operands for an SRem, see if we can fold the result.
1171 /// If not, this returns null.
1172 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1173  unsigned MaxRecurse) {
1174  // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1175  // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1176  Value *X;
1177  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1178  return ConstantInt::getNullValue(Op0->getType());
1179 
1180  // If the two operands are negated, return 0.
1181  if (isKnownNegation(Op0, Op1))
1182  return ConstantInt::getNullValue(Op0->getType());
1183 
1184  return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1185 }
1186 
1189 }
1190 
1191 /// Given operands for a URem, see if we can fold the result.
1192 /// If not, this returns null.
1193 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1194  unsigned MaxRecurse) {
1195  return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1196 }
1197 
1200 }
1201 
1202 /// Returns true if a shift by \c Amount always yields poison.
1203 static bool isPoisonShift(Value *Amount, const SimplifyQuery &Q) {
1204  Constant *C = dyn_cast<Constant>(Amount);
1205  if (!C)
1206  return false;
1207 
1208  // X shift by undef -> poison because it may shift by the bitwidth.
1209  if (Q.isUndefValue(C))
1210  return true;
1211 
1212  // Shifting by the bitwidth or more is undefined.
1213  if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1214  if (CI->getValue().uge(CI->getType()->getScalarSizeInBits()))
1215  return true;
1216 
1217  // If all lanes of a vector shift are undefined the whole shift is.
1218  if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1219  for (unsigned I = 0,
1220  E = cast<FixedVectorType>(C->getType())->getNumElements();
1221  I != E; ++I)
1222  if (!isPoisonShift(C->getAggregateElement(I), Q))
1223  return false;
1224  return true;
1225  }
1226 
1227  return false;
1228 }
1229 
1230 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1231 /// If not, this returns null.
1233  Value *Op1, bool IsNSW, const SimplifyQuery &Q,
1234  unsigned MaxRecurse) {
1235  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1236  return C;
1237 
1238  // 0 shift by X -> 0
1239  if (match(Op0, m_Zero()))
1240  return Constant::getNullValue(Op0->getType());
1241 
1242  // X shift by 0 -> X
1243  // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1244  // would be poison.
1245  Value *X;
1246  if (match(Op1, m_Zero()) ||
1247  (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1248  return Op0;
1249 
1250  // Fold undefined shifts.
1251  if (isPoisonShift(Op1, Q))
1252  return PoisonValue::get(Op0->getType());
1253 
1254  // If the operation is with the result of a select instruction, check whether
1255  // operating on either branch of the select always yields the same value.
1256  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1257  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1258  return V;
1259 
1260  // If the operation is with the result of a phi instruction, check whether
1261  // operating on all incoming values of the phi always yields the same value.
1262  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1263  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1264  return V;
1265 
1266  // If any bits in the shift amount make that value greater than or equal to
1267  // the number of bits in the type, the shift is undefined.
1268  KnownBits KnownAmt = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1269  if (KnownAmt.getMinValue().uge(KnownAmt.getBitWidth()))
1270  return PoisonValue::get(Op0->getType());
1271 
1272  // If all valid bits in the shift amount are known zero, the first operand is
1273  // unchanged.
1274  unsigned NumValidShiftBits = Log2_32_Ceil(KnownAmt.getBitWidth());
1275  if (KnownAmt.countMinTrailingZeros() >= NumValidShiftBits)
1276  return Op0;
1277 
1278  // Check for nsw shl leading to a poison value.
1279  if (IsNSW) {
1280  assert(Opcode == Instruction::Shl && "Expected shl for nsw instruction");
1281  KnownBits KnownVal = computeKnownBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1282  KnownBits KnownShl = KnownBits::shl(KnownVal, KnownAmt);
1283 
1284  if (KnownVal.Zero.isSignBitSet())
1285  KnownShl.Zero.setSignBit();
1286  if (KnownVal.One.isSignBitSet())
1287  KnownShl.One.setSignBit();
1288 
1289  if (KnownShl.hasConflict())
1290  return PoisonValue::get(Op0->getType());
1291  }
1292 
1293  return nullptr;
1294 }
1295 
1296 /// Given operands for an Shl, LShr or AShr, see if we can
1297 /// fold the result. If not, this returns null.
1299  Value *Op1, bool isExact, const SimplifyQuery &Q,
1300  unsigned MaxRecurse) {
1301  if (Value *V =
1302  SimplifyShift(Opcode, Op0, Op1, /*IsNSW*/ false, Q, MaxRecurse))
1303  return V;
1304 
1305  // X >> X -> 0
1306  if (Op0 == Op1)
1307  return Constant::getNullValue(Op0->getType());
1308 
1309  // undef >> X -> 0
1310  // undef >> X -> undef (if it's exact)
1311  if (Q.isUndefValue(Op0))
1312  return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1313 
1314  // The low bit cannot be shifted out of an exact shift if it is set.
1315  if (isExact) {
1316  KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1317  if (Op0Known.One[0])
1318  return Op0;
1319  }
1320 
1321  return nullptr;
1322 }
1323 
1324 /// Given operands for an Shl, see if we can fold the result.
1325 /// If not, this returns null.
1326 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1327  const SimplifyQuery &Q, unsigned MaxRecurse) {
1328  if (Value *V =
1329  SimplifyShift(Instruction::Shl, Op0, Op1, isNSW, Q, MaxRecurse))
1330  return V;
1331 
1332  // undef << X -> 0
1333  // undef << X -> undef if (if it's NSW/NUW)
1334  if (Q.isUndefValue(Op0))
1335  return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1336 
1337  // (X >> A) << A -> X
1338  Value *X;
1339  if (Q.IIQ.UseInstrInfo &&
1340  match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1341  return X;
1342 
1343  // shl nuw i8 C, %x -> C iff C has sign bit set.
1344  if (isNUW && match(Op0, m_Negative()))
1345  return Op0;
1346  // NOTE: could use computeKnownBits() / LazyValueInfo,
1347  // but the cost-benefit analysis suggests it isn't worth it.
1348 
1349  return nullptr;
1350 }
1351 
1352 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1353  const SimplifyQuery &Q) {
1354  return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1355 }
1356 
1357 /// Given operands for an LShr, see if we can fold the result.
1358 /// If not, this returns null.
1359 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1360  const SimplifyQuery &Q, unsigned MaxRecurse) {
1361  if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1362  MaxRecurse))
1363  return V;
1364 
1365  // (X << A) >> A -> X
1366  Value *X;
1367  if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1368  return X;
1369 
1370  // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1371  // We can return X as we do in the above case since OR alters no bits in X.
1372  // SimplifyDemandedBits in InstCombine can do more general optimization for
1373  // bit manipulation. This pattern aims to provide opportunities for other
1374  // optimizers by supporting a simple but common case in InstSimplify.
1375  Value *Y;
1376  const APInt *ShRAmt, *ShLAmt;
1377  if (match(Op1, m_APInt(ShRAmt)) &&
1378  match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1379  *ShRAmt == *ShLAmt) {
1380  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1381  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1382  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1383  if (ShRAmt->uge(EffWidthY))
1384  return X;
1385  }
1386 
1387  return nullptr;
1388 }
1389 
1390 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1391  const SimplifyQuery &Q) {
1392  return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1393 }
1394 
1395 /// Given operands for an AShr, see if we can fold the result.
1396 /// If not, this returns null.
1397 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1398  const SimplifyQuery &Q, unsigned MaxRecurse) {
1399  if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1400  MaxRecurse))
1401  return V;
1402 
1403  // all ones >>a X -> -1
1404  // Do not return Op0 because it may contain undef elements if it's a vector.
1405  if (match(Op0, m_AllOnes()))
1406  return Constant::getAllOnesValue(Op0->getType());
1407 
1408  // (X << A) >> A -> X
1409  Value *X;
1410  if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1411  return X;
1412 
1413  // Arithmetic shifting an all-sign-bit value is a no-op.
1414  unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1415  if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1416  return Op0;
1417 
1418  return nullptr;
1419 }
1420 
1421 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1422  const SimplifyQuery &Q) {
1423  return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1424 }
1425 
1426 /// Commuted variants are assumed to be handled by calling this function again
1427 /// with the parameters swapped.
1429  ICmpInst *UnsignedICmp, bool IsAnd,
1430  const SimplifyQuery &Q) {
1431  Value *X, *Y;
1432 
1433  ICmpInst::Predicate EqPred;
1434  if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1435  !ICmpInst::isEquality(EqPred))
1436  return nullptr;
1437 
1438  ICmpInst::Predicate UnsignedPred;
1439 
1440  Value *A, *B;
1441  // Y = (A - B);
1442  if (match(Y, m_Sub(m_Value(A), m_Value(B)))) {
1443  if (match(UnsignedICmp,
1444  m_c_ICmp(UnsignedPred, m_Specific(A), m_Specific(B))) &&
1445  ICmpInst::isUnsigned(UnsignedPred)) {
1446  // A >=/<= B || (A - B) != 0 <--> true
1447  if ((UnsignedPred == ICmpInst::ICMP_UGE ||
1448  UnsignedPred == ICmpInst::ICMP_ULE) &&
1449  EqPred == ICmpInst::ICMP_NE && !IsAnd)
1450  return ConstantInt::getTrue(UnsignedICmp->getType());
1451  // A </> B && (A - B) == 0 <--> false
1452  if ((UnsignedPred == ICmpInst::ICMP_ULT ||
1453  UnsignedPred == ICmpInst::ICMP_UGT) &&
1454  EqPred == ICmpInst::ICMP_EQ && IsAnd)
1455  return ConstantInt::getFalse(UnsignedICmp->getType());
1456 
1457  // A </> B && (A - B) != 0 <--> A </> B
1458  // A </> B || (A - B) != 0 <--> (A - B) != 0
1459  if (EqPred == ICmpInst::ICMP_NE && (UnsignedPred == ICmpInst::ICMP_ULT ||
1460  UnsignedPred == ICmpInst::ICMP_UGT))
1461  return IsAnd ? UnsignedICmp : ZeroICmp;
1462 
1463  // A <=/>= B && (A - B) == 0 <--> (A - B) == 0
1464  // A <=/>= B || (A - B) == 0 <--> A <=/>= B
1465  if (EqPred == ICmpInst::ICMP_EQ && (UnsignedPred == ICmpInst::ICMP_ULE ||
1466  UnsignedPred == ICmpInst::ICMP_UGE))
1467  return IsAnd ? ZeroICmp : UnsignedICmp;
1468  }
1469 
1470  // Given Y = (A - B)
1471  // Y >= A && Y != 0 --> Y >= A iff B != 0
1472  // Y < A || Y == 0 --> Y < A iff B != 0
1473  if (match(UnsignedICmp,
1474  m_c_ICmp(UnsignedPred, m_Specific(Y), m_Specific(A)))) {
1475  if (UnsignedPred == ICmpInst::ICMP_UGE && IsAnd &&
1476  EqPred == ICmpInst::ICMP_NE &&
1477  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1478  return UnsignedICmp;
1479  if (UnsignedPred == ICmpInst::ICMP_ULT && !IsAnd &&
1480  EqPred == ICmpInst::ICMP_EQ &&
1481  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1482  return UnsignedICmp;
1483  }
1484  }
1485 
1486  if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1487  ICmpInst::isUnsigned(UnsignedPred))
1488  ;
1489  else if (match(UnsignedICmp,
1490  m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1491  ICmpInst::isUnsigned(UnsignedPred))
1492  UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1493  else
1494  return nullptr;
1495 
1496  // X > Y && Y == 0 --> Y == 0 iff X != 0
1497  // X > Y || Y == 0 --> X > Y iff X != 0
1498  if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ &&
1499  isKnownNonZero(X, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1500  return IsAnd ? ZeroICmp : UnsignedICmp;
1501 
1502  // X <= Y && Y != 0 --> X <= Y iff X != 0
1503  // X <= Y || Y != 0 --> Y != 0 iff X != 0
1504  if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE &&
1505  isKnownNonZero(X, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1506  return IsAnd ? UnsignedICmp : ZeroICmp;
1507 
1508  // The transforms below here are expected to be handled more generally with
1509  // simplifyAndOrOfICmpsWithLimitConst() or in InstCombine's
1510  // foldAndOrOfICmpsWithConstEq(). If we are looking to trim optimizer overlap,
1511  // these are candidates for removal.
1512 
1513  // X < Y && Y != 0 --> X < Y
1514  // X < Y || Y != 0 --> Y != 0
1515  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1516  return IsAnd ? UnsignedICmp : ZeroICmp;
1517 
1518  // X >= Y && Y == 0 --> Y == 0
1519  // X >= Y || Y == 0 --> X >= Y
1520  if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ)
1521  return IsAnd ? ZeroICmp : UnsignedICmp;
1522 
1523  // X < Y && Y == 0 --> false
1524  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1525  IsAnd)
1526  return getFalse(UnsignedICmp->getType());
1527 
1528  // X >= Y || Y != 0 --> true
1529  if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_NE &&
1530  !IsAnd)
1531  return getTrue(UnsignedICmp->getType());
1532 
1533  return nullptr;
1534 }
1535 
1536 /// Commuted variants are assumed to be handled by calling this function again
1537 /// with the parameters swapped.
1539  ICmpInst::Predicate Pred0, Pred1;
1540  Value *A ,*B;
1541  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1542  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1543  return nullptr;
1544 
1545  // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1546  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1547  // can eliminate Op1 from this 'and'.
1548  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1549  return Op0;
1550 
1551  // Check for any combination of predicates that are guaranteed to be disjoint.
1552  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1553  (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1554  (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1555  (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1556  return getFalse(Op0->getType());
1557 
1558  return nullptr;
1559 }
1560 
1561 /// Commuted variants are assumed to be handled by calling this function again
1562 /// with the parameters swapped.
1564  ICmpInst::Predicate Pred0, Pred1;
1565  Value *A ,*B;
1566  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1567  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1568  return nullptr;
1569 
1570  // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1571  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1572  // can eliminate Op0 from this 'or'.
1573  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1574  return Op1;
1575 
1576  // Check for any combination of predicates that cover the entire range of
1577  // possibilities.
1578  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1579  (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1580  (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1581  (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1582  return getTrue(Op0->getType());
1583 
1584  return nullptr;
1585 }
1586 
1587 /// Test if a pair of compares with a shared operand and 2 constants has an
1588 /// empty set intersection, full set union, or if one compare is a superset of
1589 /// the other.
1591  bool IsAnd) {
1592  // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1593  if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1594  return nullptr;
1595 
1596  const APInt *C0, *C1;
1597  if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1598  !match(Cmp1->getOperand(1), m_APInt(C1)))
1599  return nullptr;
1600 
1601  auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1602  auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1603 
1604  // For and-of-compares, check if the intersection is empty:
1605  // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1606  if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1607  return getFalse(Cmp0->getType());
1608 
1609  // For or-of-compares, check if the union is full:
1610  // (icmp X, C0) || (icmp X, C1) --> full set --> true
1611  if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1612  return getTrue(Cmp0->getType());
1613 
1614  // Is one range a superset of the other?
1615  // If this is and-of-compares, take the smaller set:
1616  // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1617  // If this is or-of-compares, take the larger set:
1618  // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1619  if (Range0.contains(Range1))
1620  return IsAnd ? Cmp1 : Cmp0;
1621  if (Range1.contains(Range0))
1622  return IsAnd ? Cmp0 : Cmp1;
1623 
1624  return nullptr;
1625 }
1626 
1628  bool IsAnd) {
1629  ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1630  if (!match(Cmp0->getOperand(1), m_Zero()) ||
1631  !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1632  return nullptr;
1633 
1634  if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1635  return nullptr;
1636 
1637  // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1638  Value *X = Cmp0->getOperand(0);
1639  Value *Y = Cmp1->getOperand(0);
1640 
1641  // If one of the compares is a masked version of a (not) null check, then
1642  // that compare implies the other, so we eliminate the other. Optionally, look
1643  // through a pointer-to-int cast to match a null check of a pointer type.
1644 
1645  // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1646  // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1647  // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1648  // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1649  if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1651  return Cmp1;
1652 
1653  // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1654  // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1655  // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1656  // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1657  if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1659  return Cmp0;
1660 
1661  return nullptr;
1662 }
1663 
1665  const InstrInfoQuery &IIQ) {
1666  // (icmp (add V, C0), C1) & (icmp V, C0)
1667  ICmpInst::Predicate Pred0, Pred1;
1668  const APInt *C0, *C1;
1669  Value *V;
1670  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1671  return nullptr;
1672 
1673  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1674  return nullptr;
1675 
1676  auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1677  if (AddInst->getOperand(1) != Op1->getOperand(1))
1678  return nullptr;
1679 
1680  Type *ITy = Op0->getType();
1681  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1682  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1683 
1684  const APInt Delta = *C1 - *C0;
1685  if (C0->isStrictlyPositive()) {
1686  if (Delta == 2) {
1687  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1688  return getFalse(ITy);
1689  if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1690  return getFalse(ITy);
1691  }
1692  if (Delta == 1) {
1693  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1694  return getFalse(ITy);
1695  if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1696  return getFalse(ITy);
1697  }
1698  }
1699  if (C0->getBoolValue() && isNUW) {
1700  if (Delta == 2)
1701  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1702  return getFalse(ITy);
1703  if (Delta == 1)
1704  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1705  return getFalse(ITy);
1706  }
1707 
1708  return nullptr;
1709 }
1710 
1711 /// Try to eliminate compares with signed or unsigned min/max constants.
1713  bool IsAnd) {
1714  // Canonicalize an equality compare as Cmp0.
1715  if (Cmp1->isEquality())
1716  std::swap(Cmp0, Cmp1);
1717  if (!Cmp0->isEquality())
1718  return nullptr;
1719 
1720  // The non-equality compare must include a common operand (X). Canonicalize
1721  // the common operand as operand 0 (the predicate is swapped if the common
1722  // operand was operand 1).
1723  ICmpInst::Predicate Pred0 = Cmp0->getPredicate();
1724  Value *X = Cmp0->getOperand(0);
1725  ICmpInst::Predicate Pred1;
1726  bool HasNotOp = match(Cmp1, m_c_ICmp(Pred1, m_Not(m_Specific(X)), m_Value()));
1727  if (!HasNotOp && !match(Cmp1, m_c_ICmp(Pred1, m_Specific(X), m_Value())))
1728  return nullptr;
1729  if (ICmpInst::isEquality(Pred1))
1730  return nullptr;
1731 
1732  // The equality compare must be against a constant. Flip bits if we matched
1733  // a bitwise not. Convert a null pointer constant to an integer zero value.
1734  APInt MinMaxC;
1735  const APInt *C;
1736  if (match(Cmp0->getOperand(1), m_APInt(C)))
1737  MinMaxC = HasNotOp ? ~*C : *C;
1738  else if (isa<ConstantPointerNull>(Cmp0->getOperand(1)))
1739  MinMaxC = APInt::getNullValue(8);
1740  else
1741  return nullptr;
1742 
1743  // DeMorganize if this is 'or': P0 || P1 --> !P0 && !P1.
1744  if (!IsAnd) {
1745  Pred0 = ICmpInst::getInversePredicate(Pred0);
1746  Pred1 = ICmpInst::getInversePredicate(Pred1);
1747  }
1748 
1749  // Normalize to unsigned compare and unsigned min/max value.
1750  // Example for 8-bit: -128 + 128 -> 0; 127 + 128 -> 255
1751  if (ICmpInst::isSigned(Pred1)) {
1752  Pred1 = ICmpInst::getUnsignedPredicate(Pred1);
1753  MinMaxC += APInt::getSignedMinValue(MinMaxC.getBitWidth());
1754  }
1755 
1756  // (X != MAX) && (X < Y) --> X < Y
1757  // (X == MAX) || (X >= Y) --> X >= Y
1758  if (MinMaxC.isMaxValue())
1759  if (Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT)
1760  return Cmp1;
1761 
1762  // (X != MIN) && (X > Y) --> X > Y
1763  // (X == MIN) || (X <= Y) --> X <= Y
1764  if (MinMaxC.isMinValue())
1765  if (Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_UGT)
1766  return Cmp1;
1767 
1768  return nullptr;
1769 }
1770 
1772  const SimplifyQuery &Q) {
1773  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true, Q))
1774  return X;
1775  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true, Q))
1776  return X;
1777 
1778  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1779  return X;
1780  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1781  return X;
1782 
1783  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1784  return X;
1785 
1786  if (Value *X = simplifyAndOrOfICmpsWithLimitConst(Op0, Op1, true))
1787  return X;
1788 
1789  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1790  return X;
1791 
1792  if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, Q.IIQ))
1793  return X;
1794  if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, Q.IIQ))
1795  return X;
1796 
1797  return nullptr;
1798 }
1799 
1801  const InstrInfoQuery &IIQ) {
1802  // (icmp (add V, C0), C1) | (icmp V, C0)
1803  ICmpInst::Predicate Pred0, Pred1;
1804  const APInt *C0, *C1;
1805  Value *V;
1806  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1807  return nullptr;
1808 
1809  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1810  return nullptr;
1811 
1812  auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1813  if (AddInst->getOperand(1) != Op1->getOperand(1))
1814  return nullptr;
1815 
1816  Type *ITy = Op0->getType();
1817  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1818  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1819 
1820  const APInt Delta = *C1 - *C0;
1821  if (C0->isStrictlyPositive()) {
1822  if (Delta == 2) {
1823  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1824  return getTrue(ITy);
1825  if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1826  return getTrue(ITy);
1827  }
1828  if (Delta == 1) {
1829  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1830  return getTrue(ITy);
1831  if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1832  return getTrue(ITy);
1833  }
1834  }
1835  if (C0->getBoolValue() && isNUW) {
1836  if (Delta == 2)
1837  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1838  return getTrue(ITy);
1839  if (Delta == 1)
1840  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1841  return getTrue(ITy);
1842  }
1843 
1844  return nullptr;
1845 }
1846 
1848  const SimplifyQuery &Q) {
1849  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false, Q))
1850  return X;
1851  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false, Q))
1852  return X;
1853 
1854  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1855  return X;
1856  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1857  return X;
1858 
1859  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1860  return X;
1861 
1862  if (Value *X = simplifyAndOrOfICmpsWithLimitConst(Op0, Op1, false))
1863  return X;
1864 
1865  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1866  return X;
1867 
1868  if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, Q.IIQ))
1869  return X;
1870  if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, Q.IIQ))
1871  return X;
1872 
1873  return nullptr;
1874 }
1875 
1877  FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1878  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1879  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1880  if (LHS0->getType() != RHS0->getType())
1881  return nullptr;
1882 
1883  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1884  if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1885  (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1886  // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1887  // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1888  // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1889  // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1890  // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1891  // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1892  // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1893  // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1894  if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1895  (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1896  return RHS;
1897 
1898  // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1899  // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1900  // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1901  // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1902  // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1903  // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1904  // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1905  // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1906  if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1907  (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1908  return LHS;
1909  }
1910 
1911  return nullptr;
1912 }
1913 
1915  Value *Op0, Value *Op1, bool IsAnd) {
1916  // Look through casts of the 'and' operands to find compares.
1917  auto *Cast0 = dyn_cast<CastInst>(Op0);
1918  auto *Cast1 = dyn_cast<CastInst>(Op1);
1919  if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1920  Cast0->getSrcTy() == Cast1->getSrcTy()) {
1921  Op0 = Cast0->getOperand(0);
1922  Op1 = Cast1->getOperand(0);
1923  }
1924 
1925  Value *V = nullptr;
1926  auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1927  auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1928  if (ICmp0 && ICmp1)
1929  V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q)
1930  : simplifyOrOfICmps(ICmp0, ICmp1, Q);
1931 
1932  auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1933  auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1934  if (FCmp0 && FCmp1)
1935  V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1936 
1937  if (!V)
1938  return nullptr;
1939  if (!Cast0)
1940  return V;
1941 
1942  // If we looked through casts, we can only handle a constant simplification
1943  // because we are not allowed to create a cast instruction here.
1944  if (auto *C = dyn_cast<Constant>(V))
1945  return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1946 
1947  return nullptr;
1948 }
1949 
1950 /// Check that the Op1 is in expected form, i.e.:
1951 /// %Agg = tail call { i4, i1 } @llvm.[us]mul.with.overflow.i4(i4 %X, i4 %???)
1952 /// %Op1 = extractvalue { i4, i1 } %Agg, 1
1954  Value *X) {
1955  auto *Extract = dyn_cast<ExtractValueInst>(Op1);
1956  // We should only be extracting the overflow bit.
1957  if (!Extract || !Extract->getIndices().equals(1))
1958  return false;
1959  Value *Agg = Extract->getAggregateOperand();
1960  // This should be a multiplication-with-overflow intrinsic.
1961  if (!match(Agg, m_CombineOr(m_Intrinsic<Intrinsic::umul_with_overflow>(),
1962  m_Intrinsic<Intrinsic::smul_with_overflow>())))
1963  return false;
1964  // One of its multipliers should be the value we checked for zero before.
1965  if (!match(Agg, m_CombineOr(m_Argument<0>(m_Specific(X)),
1966  m_Argument<1>(m_Specific(X)))))
1967  return false;
1968  return true;
1969 }
1970 
1971 /// The @llvm.[us]mul.with.overflow intrinsic could have been folded from some
1972 /// other form of check, e.g. one that was using division; it may have been
1973 /// guarded against division-by-zero. We can drop that check now.
1974 /// Look for:
1975 /// %Op0 = icmp ne i4 %X, 0
1976 /// %Agg = tail call { i4, i1 } @llvm.[us]mul.with.overflow.i4(i4 %X, i4 %???)
1977 /// %Op1 = extractvalue { i4, i1 } %Agg, 1
1978 /// %??? = and i1 %Op0, %Op1
1979 /// We can just return %Op1
1981  ICmpInst::Predicate Pred;
1982  Value *X;
1983  if (!match(Op0, m_ICmp(Pred, m_Value(X), m_Zero())) ||
1984  Pred != ICmpInst::Predicate::ICMP_NE)
1985  return nullptr;
1986  // Is Op1 in expected form?
1988  return nullptr;
1989  // Can omit 'and', and just return the overflow bit.
1990  return Op1;
1991 }
1992 
1993 /// The @llvm.[us]mul.with.overflow intrinsic could have been folded from some
1994 /// other form of check, e.g. one that was using division; it may have been
1995 /// guarded against division-by-zero. We can drop that check now.
1996 /// Look for:
1997 /// %Op0 = icmp eq i4 %X, 0
1998 /// %Agg = tail call { i4, i1 } @llvm.[us]mul.with.overflow.i4(i4 %X, i4 %???)
1999 /// %Op1 = extractvalue { i4, i1 } %Agg, 1
2000 /// %NotOp1 = xor i1 %Op1, true
2001 /// %or = or i1 %Op0, %NotOp1
2002 /// We can just return %NotOp1
2004  Value *NotOp1) {
2005  ICmpInst::Predicate Pred;
2006  Value *X;
2007  if (!match(Op0, m_ICmp(Pred, m_Value(X), m_Zero())) ||
2008  Pred != ICmpInst::Predicate::ICMP_EQ)
2009  return nullptr;
2010  // We expect the other hand of an 'or' to be a 'not'.
2011  Value *Op1;
2012  if (!match(NotOp1, m_Not(m_Value(Op1))))
2013  return nullptr;
2014  // Is Op1 in expected form?
2016  return nullptr;
2017  // Can omit 'and', and just return the inverted overflow bit.
2018  return NotOp1;
2019 }
2020 
2021 /// Given a bitwise logic op, check if the operands are add/sub with a common
2022 /// source value and inverted constant (identity: C - X -> ~(X + ~C)).
2024  Instruction::BinaryOps Opcode) {
2025  assert(Op0->getType() == Op1->getType() && "Mismatched binop types");
2026  assert(BinaryOperator::isBitwiseLogicOp(Opcode) && "Expected logic op");
2027  Value *X;
2028  Constant *C1, *C2;
2029  if ((match(Op0, m_Add(m_Value(X), m_Constant(C1))) &&
2030  match(Op1, m_Sub(m_Constant(C2), m_Specific(X)))) ||
2031  (match(Op1, m_Add(m_Value(X), m_Constant(C1))) &&
2032  match(Op0, m_Sub(m_Constant(C2), m_Specific(X))))) {
2033  if (ConstantExpr::getNot(C1) == C2) {
2034  // (X + C) & (~C - X) --> (X + C) & ~(X + C) --> 0
2035  // (X + C) | (~C - X) --> (X + C) | ~(X + C) --> -1
2036  // (X + C) ^ (~C - X) --> (X + C) ^ ~(X + C) --> -1
2037  Type *Ty = Op0->getType();
2038  return Opcode == Instruction::And ? ConstantInt::getNullValue(Ty)
2040  }
2041  }
2042  return nullptr;
2043 }
2044 
2045 /// Given operands for an And, see if we can fold the result.
2046 /// If not, this returns null.
2047 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2048  unsigned MaxRecurse) {
2049  if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
2050  return C;
2051 
2052  // X & undef -> 0
2053  if (Q.isUndefValue(Op1))
2054  return Constant::getNullValue(Op0->getType());
2055 
2056  // X & X = X
2057  if (Op0 == Op1)
2058  return Op0;
2059 
2060  // X & 0 = 0
2061  if (match(Op1, m_Zero()))
2062  return Constant::getNullValue(Op0->getType());
2063 
2064  // X & -1 = X
2065  if (match(Op1, m_AllOnes()))
2066  return Op0;
2067 
2068  // A & ~A = ~A & A = 0
2069  if (match(Op0, m_Not(m_Specific(Op1))) ||
2070  match(Op1, m_Not(m_Specific(Op0))))
2071  return Constant::getNullValue(Op0->getType());
2072 
2073  // (A | ?) & A = A
2074  if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
2075  return Op1;
2076 
2077  // A & (A | ?) = A
2078  if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
2079  return Op0;
2080 
2081  if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Instruction::And))
2082  return V;
2083 
2084  // A mask that only clears known zeros of a shifted value is a no-op.
2085  Value *X;
2086  const APInt *Mask;
2087  const APInt *ShAmt;
2088  if (match(Op1, m_APInt(Mask))) {
2089  // If all bits in the inverted and shifted mask are clear:
2090  // and (shl X, ShAmt), Mask --> shl X, ShAmt
2091  if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
2092  (~(*Mask)).lshr(*ShAmt).isNullValue())
2093  return Op0;
2094 
2095  // If all bits in the inverted and shifted mask are clear:
2096  // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
2097  if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
2098  (~(*Mask)).shl(*ShAmt).isNullValue())
2099  return Op0;
2100  }
2101 
2102  // If we have a multiplication overflow check that is being 'and'ed with a
2103  // check that one of the multipliers is not zero, we can omit the 'and', and
2104  // only keep the overflow check.
2105  if (Value *V = omitCheckForZeroBeforeMulWithOverflow(Op0, Op1))
2106  return V;
2107  if (Value *V = omitCheckForZeroBeforeMulWithOverflow(Op1, Op0))
2108  return V;
2109 
2110  // A & (-A) = A if A is a power of two or zero.
2111  if (match(Op0, m_Neg(m_Specific(Op1))) ||
2112  match(Op1, m_Neg(m_Specific(Op0)))) {
2113  if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
2114  Q.DT))
2115  return Op0;
2116  if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
2117  Q.DT))
2118  return Op1;
2119  }
2120 
2121  // This is a similar pattern used for checking if a value is a power-of-2:
2122  // (A - 1) & A --> 0 (if A is a power-of-2 or 0)
2123  // A & (A - 1) --> 0 (if A is a power-of-2 or 0)
2124  if (match(Op0, m_Add(m_Specific(Op1), m_AllOnes())) &&
2125  isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
2126  return Constant::getNullValue(Op1->getType());
2127  if (match(Op1, m_Add(m_Specific(Op0), m_AllOnes())) &&
2128  isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
2129  return Constant::getNullValue(Op0->getType());
2130 
2131  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
2132  return V;
2133 
2134  // Try some generic simplifications for associative operations.
2135  if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
2136  MaxRecurse))
2137  return V;
2138 
2139  // And distributes over Or. Try some generic simplifications based on this.
2140  if (Value *V = expandCommutativeBinOp(Instruction::And, Op0, Op1,
2141  Instruction::Or, Q, MaxRecurse))
2142  return V;
2143 
2144  // And distributes over Xor. Try some generic simplifications based on this.
2145  if (Value *V = expandCommutativeBinOp(Instruction::And, Op0, Op1,
2146  Instruction::Xor, Q, MaxRecurse))
2147  return V;
2148 
2149  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) {
2150  if (Op0->getType()->isIntOrIntVectorTy(1)) {
2151  // A & (A && B) -> A && B
2152  if (match(Op1, m_Select(m_Specific(Op0), m_Value(), m_Zero())))
2153  return Op1;
2154  else if (match(Op0, m_Select(m_Specific(Op1), m_Value(), m_Zero())))
2155  return Op0;
2156  }
2157  // If the operation is with the result of a select instruction, check
2158  // whether operating on either branch of the select always yields the same
2159  // value.
2160  if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
2161  MaxRecurse))
2162  return V;
2163  }
2164 
2165  // If the operation is with the result of a phi instruction, check whether
2166  // operating on all incoming values of the phi always yields the same value.
2167  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2168  if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
2169  MaxRecurse))
2170  return V;
2171 
2172  // Assuming the effective width of Y is not larger than A, i.e. all bits
2173  // from X and Y are disjoint in (X << A) | Y,
2174  // if the mask of this AND op covers all bits of X or Y, while it covers
2175  // no bits from the other, we can bypass this AND op. E.g.,
2176  // ((X << A) | Y) & Mask -> Y,
2177  // if Mask = ((1 << effective_width_of(Y)) - 1)
2178  // ((X << A) | Y) & Mask -> X << A,
2179  // if Mask = ((1 << effective_width_of(X)) - 1) << A
2180  // SimplifyDemandedBits in InstCombine can optimize the general case.
2181  // This pattern aims to help other passes for a common case.
2182  Value *Y, *XShifted;
2183  if (match(Op1, m_APInt(Mask)) &&
2185  m_Value(XShifted)),
2186  m_Value(Y)))) {
2187  const unsigned Width = Op0->getType()->getScalarSizeInBits();
2188  const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
2189  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2190  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
2191  if (EffWidthY <= ShftCnt) {
2192  const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
2193  Q.DT);
2194  const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
2195  const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
2196  const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
2197  // If the mask is extracting all bits from X or Y as is, we can skip
2198  // this AND op.
2199  if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
2200  return Y;
2201  if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
2202  return XShifted;
2203  }
2204  }
2205 
2206  return nullptr;
2207 }
2208 
2211 }
2212 
2213 /// Given operands for an Or, see if we can fold the result.
2214 /// If not, this returns null.
2215 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2216  unsigned MaxRecurse) {
2217  if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
2218  return C;
2219 
2220  // X | undef -> -1
2221  // X | -1 = -1
2222  // Do not return Op1 because it may contain undef elements if it's a vector.
2223  if (Q.isUndefValue(Op1) || match(Op1, m_AllOnes()))
2224  return Constant::getAllOnesValue(Op0->getType());
2225 
2226  // X | X = X
2227  // X | 0 = X
2228  if (Op0 == Op1 || match(Op1, m_Zero()))
2229  return Op0;
2230 
2231  // A | ~A = ~A | A = -1
2232  if (match(Op0, m_Not(m_Specific(Op1))) ||
2233  match(Op1, m_Not(m_Specific(Op0))))
2234  return Constant::getAllOnesValue(Op0->getType());
2235 
2236  // (A & ?) | A = A
2237  if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
2238  return Op1;
2239 
2240  // A | (A & ?) = A
2241  if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
2242  return Op0;
2243 
2244  // ~(A & ?) | A = -1
2245  if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
2246  return Constant::getAllOnesValue(Op1->getType());
2247 
2248  // A | ~(A & ?) = -1
2249  if (match(Op1, m_Not(m_c_And(m_Specific(Op0), m_Value()))))
2250  return Constant::getAllOnesValue(Op0->getType());
2251 
2252  if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Instruction::Or))
2253  return V;
2254 
2255  Value *A, *B, *NotA;
2256  // (A & ~B) | (A ^ B) -> (A ^ B)
2257  // (~B & A) | (A ^ B) -> (A ^ B)
2258  // (A & ~B) | (B ^ A) -> (B ^ A)
2259  // (~B & A) | (B ^ A) -> (B ^ A)
2260  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2261  (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
2262  match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
2263  return Op1;
2264 
2265  // Commute the 'or' operands.
2266  // (A ^ B) | (A & ~B) -> (A ^ B)
2267  // (A ^ B) | (~B & A) -> (A ^ B)
2268  // (B ^ A) | (A & ~B) -> (B ^ A)
2269  // (B ^ A) | (~B & A) -> (B ^ A)
2270  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2271  (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
2272  match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
2273  return Op0;
2274 
2275  // (A & B) | (~A ^ B) -> (~A ^ B)
2276  // (B & A) | (~A ^ B) -> (~A ^ B)
2277  // (A & B) | (B ^ ~A) -> (B ^ ~A)
2278  // (B & A) | (B ^ ~A) -> (B ^ ~A)
2279  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2280  (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2281  match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2282  return Op1;
2283 
2284  // Commute the 'or' operands.
2285  // (~A ^ B) | (A & B) -> (~A ^ B)
2286  // (~A ^ B) | (B & A) -> (~A ^ B)
2287  // (B ^ ~A) | (A & B) -> (B ^ ~A)
2288  // (B ^ ~A) | (B & A) -> (B ^ ~A)
2289  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2290  (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2291  match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2292  return Op0;
2293 
2294  // (~A & B) | ~(A | B) --> ~A
2295  // (~A & B) | ~(B | A) --> ~A
2296  // (B & ~A) | ~(A | B) --> ~A
2297  // (B & ~A) | ~(B | A) --> ~A
2298  if (match(Op0, m_c_And(m_CombineAnd(m_Value(NotA), m_Not(m_Value(A))),
2299  m_Value(B))) &&
2300  match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
2301  return NotA;
2302 
2303  // Commute the 'or' operands.
2304  // ~(A | B) | (~A & B) --> ~A
2305  // ~(B | A) | (~A & B) --> ~A
2306  // ~(A | B) | (B & ~A) --> ~A
2307  // ~(B | A) | (B & ~A) --> ~A
2308  if (match(Op1, m_c_And(m_CombineAnd(m_Value(NotA), m_Not(m_Value(A))),
2309  m_Value(B))) &&
2310  match(Op0, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
2311  return NotA;
2312 
2313  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
2314  return V;
2315 
2316  // If we have a multiplication overflow check that is being 'and'ed with a
2317  // check that one of the multipliers is not zero, we can omit the 'and', and
2318  // only keep the overflow check.
2320  return V;
2322  return V;
2323 
2324  // Try some generic simplifications for associative operations.
2325  if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
2326  MaxRecurse))
2327  return V;
2328 
2329  // Or distributes over And. Try some generic simplifications based on this.
2330  if (Value *V = expandCommutativeBinOp(Instruction::Or, Op0, Op1,
2331  Instruction::And, Q, MaxRecurse))
2332  return V;
2333 
2334  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1)) {
2335  if (Op0->getType()->isIntOrIntVectorTy(1)) {
2336  // A | (A || B) -> A || B
2337  if (match(Op1, m_Select(m_Specific(Op0), m_One(), m_Value())))
2338  return Op1;
2339  else if (match(Op0, m_Select(m_Specific(Op1), m_One(), m_Value())))
2340  return Op0;
2341  }
2342  // If the operation is with the result of a select instruction, check
2343  // whether operating on either branch of the select always yields the same
2344  // value.
2345  if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2346  MaxRecurse))
2347  return V;
2348  }
2349 
2350  // (A & C1)|(B & C2)
2351  const APInt *C1, *C2;
2352  if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2353  match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2354  if (*C1 == ~*C2) {
2355  // (A & C1)|(B & C2)
2356  // If we have: ((V + N) & C1) | (V & C2)
2357  // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2358  // replace with V+N.
2359  Value *N;
2360  if (C2->isMask() && // C2 == 0+1+
2361  match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2362  // Add commutes, try both ways.
2363  if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2364  return A;
2365  }
2366  // Or commutes, try both ways.
2367  if (C1->isMask() &&
2368  match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2369  // Add commutes, try both ways.
2370  if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2371  return B;
2372  }
2373  }
2374  }
2375 
2376  // If the operation is with the result of a phi instruction, check whether
2377  // operating on all incoming values of the phi always yields the same value.
2378  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2379  if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2380  return V;
2381 
2382  return nullptr;
2383 }
2384 
2387 }
2388 
2389 /// Given operands for a Xor, see if we can fold the result.
2390 /// If not, this returns null.
2391 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2392  unsigned MaxRecurse) {
2393  if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2394  return C;
2395 
2396  // A ^ undef -> undef
2397  if (Q.isUndefValue(Op1))
2398  return Op1;
2399 
2400  // A ^ 0 = A
2401  if (match(Op1, m_Zero()))
2402  return Op0;
2403 
2404  // A ^ A = 0
2405  if (Op0 == Op1)
2406  return Constant::getNullValue(Op0->getType());
2407 
2408  // A ^ ~A = ~A ^ A = -1
2409  if (match(Op0, m_Not(m_Specific(Op1))) ||
2410  match(Op1, m_Not(m_Specific(Op0))))
2411  return Constant::getAllOnesValue(Op0->getType());
2412 
2413  if (Value *V = simplifyLogicOfAddSub(Op0, Op1, Instruction::Xor))
2414  return V;
2415 
2416  // Try some generic simplifications for associative operations.
2417  if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2418  MaxRecurse))
2419  return V;
2420 
2421  // Threading Xor over selects and phi nodes is pointless, so don't bother.
2422  // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2423  // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2424  // only if B and C are equal. If B and C are equal then (since we assume
2425  // that operands have already been simplified) "select(cond, B, C)" should
2426  // have been simplified to the common value of B and C already. Analysing
2427  // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2428  // for threading over phi nodes.
2429 
2430  return nullptr;
2431 }
2432 
2435 }
2436 
2437 
2439  return CmpInst::makeCmpResultType(Op->getType());
2440 }
2441 
2442 /// Rummage around inside V looking for something equivalent to the comparison
2443 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2444 /// Helper function for analyzing max/min idioms.
2446  Value *LHS, Value *RHS) {
2447  SelectInst *SI = dyn_cast<SelectInst>(V);
2448  if (!SI)
2449  return nullptr;
2450  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2451  if (!Cmp)
2452  return nullptr;
2453  Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2454  if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2455  return Cmp;
2456  if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2457  LHS == CmpRHS && RHS == CmpLHS)
2458  return Cmp;
2459  return nullptr;
2460 }
2461 
2462 // A significant optimization not implemented here is assuming that alloca
2463 // addresses are not equal to incoming argument values. They don't *alias*,
2464 // as we say, but that doesn't mean they aren't equal, so we take a
2465 // conservative approach.
2466 //
2467 // This is inspired in part by C++11 5.10p1:
2468 // "Two pointers of the same type compare equal if and only if they are both
2469 // null, both point to the same function, or both represent the same
2470 // address."
2471 //
2472 // This is pretty permissive.
2473 //
2474 // It's also partly due to C11 6.5.9p6:
2475 // "Two pointers compare equal if and only if both are null pointers, both are
2476 // pointers to the same object (including a pointer to an object and a
2477 // subobject at its beginning) or function, both are pointers to one past the
2478 // last element of the same array object, or one is a pointer to one past the
2479 // end of one array object and the other is a pointer to the start of a
2480 // different array object that happens to immediately follow the first array
2481 // object in the address space.)
2482 //
2483 // C11's version is more restrictive, however there's no reason why an argument
2484 // couldn't be a one-past-the-end value for a stack object in the caller and be
2485 // equal to the beginning of a stack object in the callee.
2486 //
2487 // If the C and C++ standards are ever made sufficiently restrictive in this
2488 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2489 // this optimization.
2490 static Constant *
2492  const SimplifyQuery &Q) {
2493  const DataLayout &DL = Q.DL;
2494  const TargetLibraryInfo *TLI = Q.TLI;
2495  const DominatorTree *DT = Q.DT;
2496  const Instruction *CxtI = Q.CxtI;
2497  const InstrInfoQuery &IIQ = Q.IIQ;
2498 
2499  // First, skip past any trivial no-ops.
2500  LHS = LHS->stripPointerCasts();
2501  RHS = RHS->stripPointerCasts();
2502 
2503  // A non-null pointer is not equal to a null pointer.
2504  if (isa<ConstantPointerNull>(RHS) && ICmpInst::isEquality(Pred) &&
2505  llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2506  IIQ.UseInstrInfo))
2507  return ConstantInt::get(GetCompareTy(LHS),
2508  !CmpInst::isTrueWhenEqual(Pred));
2509 
2510  // We can only fold certain predicates on pointer comparisons.
2511  switch (Pred) {
2512  default:
2513  return nullptr;
2514 
2515  // Equality comaprisons are easy to fold.
2516  case CmpInst::ICMP_EQ:
2517  case CmpInst::ICMP_NE:
2518  break;
2519 
2520  // We can only handle unsigned relational comparisons because 'inbounds' on
2521  // a GEP only protects against unsigned wrapping.
2522  case CmpInst::ICMP_UGT:
2523  case CmpInst::ICMP_UGE:
2524  case CmpInst::ICMP_ULT:
2525  case CmpInst::ICMP_ULE:
2526  // However, we have to switch them to their signed variants to handle
2527  // negative indices from the base pointer.
2528  Pred = ICmpInst::getSignedPredicate(Pred);
2529  break;
2530  }
2531 
2532  // Strip off any constant offsets so that we can reason about them.
2533  // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2534  // here and compare base addresses like AliasAnalysis does, however there are
2535  // numerous hazards. AliasAnalysis and its utilities rely on special rules
2536  // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2537  // doesn't need to guarantee pointer inequality when it says NoAlias.
2538  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2539  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2540 
2541  // If LHS and RHS are related via constant offsets to the same base
2542  // value, we can replace it with an icmp which just compares the offsets.
2543  if (LHS == RHS)
2544  return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2545 
2546  // Various optimizations for (in)equality comparisons.
2547  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2548  // Different non-empty allocations that exist at the same time have
2549  // different addresses (if the program can tell). Global variables always
2550  // exist, so they always exist during the lifetime of each other and all
2551  // allocas. Two different allocas usually have different addresses...
2552  //
2553  // However, if there's an @llvm.stackrestore dynamically in between two
2554  // allocas, they may have the same address. It's tempting to reduce the
2555  // scope of the problem by only looking at *static* allocas here. That would
2556  // cover the majority of allocas while significantly reducing the likelihood
2557  // of having an @llvm.stackrestore pop up in the middle. However, it's not
2558  // actually impossible for an @llvm.stackrestore to pop up in the middle of
2559  // an entry block. Also, if we have a block that's not attached to a
2560  // function, we can't tell if it's "static" under the current definition.
2561  // Theoretically, this problem could be fixed by creating a new kind of
2562  // instruction kind specifically for static allocas. Such a new instruction
2563  // could be required to be at the top of the entry block, thus preventing it
2564  // from being subject to a @llvm.stackrestore. Instcombine could even
2565  // convert regular allocas into these special allocas. It'd be nifty.
2566  // However, until then, this problem remains open.
2567  //
2568  // So, we'll assume that two non-empty allocas have different addresses
2569  // for now.
2570  //
2571  // With all that, if the offsets are within the bounds of their allocations
2572  // (and not one-past-the-end! so we can't use inbounds!), and their
2573  // allocations aren't the same, the pointers are not equal.
2574  //
2575  // Note that it's not necessary to check for LHS being a global variable
2576  // address, due to canonicalization and constant folding.
2577  if (isa<AllocaInst>(LHS) &&
2578  (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2579  ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2580  ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2581  uint64_t LHSSize, RHSSize;
2582  ObjectSizeOpts Opts;
2583  Opts.NullIsUnknownSize =
2584  NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2585  if (LHSOffsetCI && RHSOffsetCI &&
2586  getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2587  getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2588  const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2589  const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2590  if (!LHSOffsetValue.isNegative() &&
2591  !RHSOffsetValue.isNegative() &&
2592  LHSOffsetValue.ult(LHSSize) &&
2593  RHSOffsetValue.ult(RHSSize)) {
2594  return ConstantInt::get(GetCompareTy(LHS),
2595  !CmpInst::isTrueWhenEqual(Pred));
2596  }
2597  }
2598 
2599  // Repeat the above check but this time without depending on DataLayout
2600  // or being able to compute a precise size.
2601  if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2602  !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2603  LHSOffset->isNullValue() &&
2604  RHSOffset->isNullValue())
2605  return ConstantInt::get(GetCompareTy(LHS),
2606  !CmpInst::isTrueWhenEqual(Pred));
2607  }
2608 
2609  // Even if an non-inbounds GEP occurs along the path we can still optimize
2610  // equality comparisons concerning the result. We avoid walking the whole
2611  // chain again by starting where the last calls to
2612  // stripAndComputeConstantOffsets left off and accumulate the offsets.
2613  Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2614  Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2615  if (LHS == RHS)
2616  return ConstantExpr::getICmp(Pred,
2617  ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2618  ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2619 
2620  // If one side of the equality comparison must come from a noalias call
2621  // (meaning a system memory allocation function), and the other side must
2622  // come from a pointer that cannot overlap with dynamically-allocated
2623  // memory within the lifetime of the current function (allocas, byval
2624  // arguments, globals), then determine the comparison result here.
2625  SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2626  getUnderlyingObjects(LHS, LHSUObjs);
2627  getUnderlyingObjects(RHS, RHSUObjs);
2628 
2629  // Is the set of underlying objects all noalias calls?
2630  auto IsNAC = [](ArrayRef<const Value *> Objects) {
2631  return all_of(Objects, isNoAliasCall);
2632  };
2633 
2634  // Is the set of underlying objects all things which must be disjoint from
2635  // noalias calls. For allocas, we consider only static ones (dynamic
2636  // allocas might be transformed into calls to malloc not simultaneously
2637  // live with the compared-to allocation). For globals, we exclude symbols
2638  // that might be resolve lazily to symbols in another dynamically-loaded
2639  // library (and, thus, could be malloc'ed by the implementation).
2640  auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2641  return all_of(Objects, [](const Value *V) {
2642  if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2643  return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2644  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2645  return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2646  GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2647  !GV->isThreadLocal();
2648  if (const Argument *A = dyn_cast<Argument>(V))
2649  return A->hasByValAttr();
2650  return false;
2651  });
2652  };
2653 
2654  if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2655  (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2656  return ConstantInt::get(GetCompareTy(LHS),
2657  !CmpInst::isTrueWhenEqual(Pred));
2658 
2659  // Fold comparisons for non-escaping pointer even if the allocation call
2660  // cannot be elided. We cannot fold malloc comparison to null. Also, the
2661  // dynamic allocation call could be either of the operands.
2662  Value *MI = nullptr;
2663  if (isAllocLikeFn(LHS, TLI) &&
2664  llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2665  MI = LHS;
2666  else if (isAllocLikeFn(RHS, TLI) &&
2667  llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2668  MI = RHS;
2669  // FIXME: We should also fold the compare when the pointer escapes, but the
2670  // compare dominates the pointer escape
2671  if (MI && !PointerMayBeCaptured(MI, true, true))
2672  return ConstantInt::get(GetCompareTy(LHS),
2674  }
2675 
2676  // Otherwise, fail.
2677  return nullptr;
2678 }
2679 
2680 /// Fold an icmp when its operands have i1 scalar type.
2682  Value *RHS, const SimplifyQuery &Q) {
2683  Type *ITy = GetCompareTy(LHS); // The return type.
2684  Type *OpTy = LHS->getType(); // The operand type.
2685  if (!OpTy->isIntOrIntVectorTy(1))
2686  return nullptr;
2687 
2688  // A boolean compared to true/false can be simplified in 14 out of the 20
2689  // (10 predicates * 2 constants) possible combinations. Cases not handled here
2690  // require a 'not' of the LHS, so those must be transformed in InstCombine.
2691  if (match(RHS, m_Zero())) {
2692  switch (Pred) {
2693  case CmpInst::ICMP_NE: // X != 0 -> X
2694  case CmpInst::ICMP_UGT: // X >u 0 -> X
2695  case CmpInst::ICMP_SLT: // X <s 0 -> X
2696  return LHS;
2697 
2698  case CmpInst::ICMP_ULT: // X <u 0 -> false
2699  case CmpInst::ICMP_SGT: // X >s 0 -> false
2700  return getFalse(ITy);
2701 
2702  case CmpInst::ICMP_UGE: // X >=u 0 -> true
2703  case CmpInst::ICMP_SLE: // X <=s 0 -> true
2704  return getTrue(ITy);
2705 
2706  default: break;
2707  }
2708  } else if (match(RHS, m_One())) {
2709  switch (Pred) {
2710  case CmpInst::ICMP_EQ: // X == 1 -> X
2711  case CmpInst::ICMP_UGE: // X >=u 1 -> X
2712  case CmpInst::ICMP_SLE: // X <=s -1 -> X
2713  return LHS;
2714 
2715  case CmpInst::ICMP_UGT: // X >u 1 -> false
2716  case CmpInst::ICMP_SLT: // X <s -1 -> false
2717  return getFalse(ITy);
2718 
2719  case CmpInst::ICMP_ULE: // X <=u 1 -> true
2720  case CmpInst::ICMP_SGE: // X >=s -1 -> true
2721  return getTrue(ITy);
2722 
2723  default: break;
2724  }
2725  }
2726 
2727  switch (Pred) {
2728  default:
2729  break;
2730  case ICmpInst::ICMP_UGE:
2731  if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2732  return getTrue(ITy);
2733  break;
2734  case ICmpInst::ICMP_SGE:
2735  /// For signed comparison, the values for an i1 are 0 and -1
2736  /// respectively. This maps into a truth table of:
2737  /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2738  /// 0 | 0 | 1 (0 >= 0) | 1
2739  /// 0 | 1 | 1 (0 >= -1) | 1
2740  /// 1 | 0 | 0 (-1 >= 0) | 0
2741  /// 1 | 1 | 1 (-1 >= -1) | 1
2742  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2743  return getTrue(ITy);
2744  break;
2745  case ICmpInst::ICMP_ULE:
2746  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2747  return getTrue(ITy);
2748  break;
2749  }
2750 
2751  return nullptr;
2752 }
2753 
2754 /// Try hard to fold icmp with zero RHS because this is a common case.
2756  Value *RHS, const SimplifyQuery &Q) {
2757  if (!match(RHS, m_Zero()))
2758  return nullptr;
2759 
2760  Type *ITy = GetCompareTy(LHS); // The return type.
2761  switch (Pred) {
2762  default:
2763  llvm_unreachable("Unknown ICmp predicate!");
2764  case ICmpInst::ICMP_ULT:
2765  return getFalse(ITy);
2766  case ICmpInst::ICMP_UGE:
2767  return getTrue(ITy);
2768  case ICmpInst::ICMP_EQ:
2769  case ICmpInst::ICMP_ULE:
2770  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2771  return getFalse(ITy);
2772  break;
2773  case ICmpInst::ICMP_NE:
2774  case ICmpInst::ICMP_UGT:
2775  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2776  return getTrue(ITy);
2777  break;
2778  case ICmpInst::ICMP_SLT: {
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  return getFalse(ITy);
2784  break;
2785  }
2786  case ICmpInst::ICMP_SLE: {
2787  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2788  if (LHSKnown.isNegative())
2789  return getTrue(ITy);
2790  if (LHSKnown.isNonNegative() &&
2791  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2792  return getFalse(ITy);
2793  break;
2794  }
2795  case ICmpInst::ICMP_SGE: {
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  return getTrue(ITy);
2801  break;
2802  }
2803  case ICmpInst::ICMP_SGT: {
2804  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2805  if (LHSKnown.isNegative())
2806  return getFalse(ITy);
2807  if (LHSKnown.isNonNegative() &&
2808  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2809  return getTrue(ITy);
2810  break;
2811  }
2812  }
2813 
2814  return nullptr;
2815 }
2816 
2818  Value *RHS, const InstrInfoQuery &IIQ) {
2819  Type *ITy = GetCompareTy(RHS); // The return type.
2820 
2821  Value *X;
2822  // Sign-bit checks can be optimized to true/false after unsigned
2823  // floating-point casts:
2824  // icmp slt (bitcast (uitofp X)), 0 --> false
2825  // icmp sgt (bitcast (uitofp X)), -1 --> true
2826  if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2827  if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2828  return ConstantInt::getFalse(ITy);
2829  if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2830  return ConstantInt::getTrue(ITy);
2831  }
2832 
2833  const APInt *C;
2834  if (!match(RHS, m_APIntAllowUndef(C)))
2835  return nullptr;
2836 
2837  // Rule out tautological comparisons (eg., ult 0 or uge 0).
2839  if (RHS_CR.isEmptySet())
2840  return ConstantInt::getFalse(ITy);
2841  if (RHS_CR.isFullSet())
2842  return ConstantInt::getTrue(ITy);
2843 
2844  ConstantRange LHS_CR = computeConstantRange(LHS, IIQ.UseInstrInfo);
2845  if (!LHS_CR.isFullSet()) {
2846  if (RHS_CR.contains(LHS_CR))
2847  return ConstantInt::getTrue(ITy);
2848  if (RHS_CR.inverse().contains(LHS_CR))
2849  return ConstantInt::getFalse(ITy);
2850  }
2851 
2852  // (mul nuw/nsw X, MulC) != C --> true (if C is not a multiple of MulC)
2853  // (mul nuw/nsw X, MulC) == C --> false (if C is not a multiple of MulC)
2854  const APInt *MulC;
2855  if (ICmpInst::isEquality(Pred) &&
2856  ((match(LHS, m_NUWMul(m_Value(), m_APIntAllowUndef(MulC))) &&
2857  *MulC != 0 && C->urem(*MulC) != 0) ||
2858  (match(LHS, m_NSWMul(m_Value(), m_APIntAllowUndef(MulC))) &&
2859  *MulC != 0 && C->srem(*MulC) != 0)))
2860  return ConstantInt::get(ITy, Pred == ICmpInst::ICMP_NE);
2861 
2862  return nullptr;
2863 }
2864 
2866  CmpInst::Predicate Pred, BinaryOperator *LBO, Value *RHS,
2867  const SimplifyQuery &Q, unsigned MaxRecurse) {
2868  Type *ITy = GetCompareTy(RHS); // The return type.
2869 
2870  Value *Y = nullptr;
2871  // icmp pred (or X, Y), X
2872  if (match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2873  if (Pred == ICmpInst::ICMP_ULT)
2874  return getFalse(ITy);
2875  if (Pred == ICmpInst::ICMP_UGE)
2876  return getTrue(ITy);
2877 
2878  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2879  KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2880  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2881  if (RHSKnown.isNonNegative() && YKnown.isNegative())
2882  return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2883  if (RHSKnown.isNegative() || YKnown.isNonNegative())
2884  return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2885  }
2886  }
2887 
2888  // icmp pred (and X, Y), X
2889  if (match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2890  if (Pred == ICmpInst::ICMP_UGT)
2891  return getFalse(ITy);
2892  if (Pred == ICmpInst::ICMP_ULE)
2893  return getTrue(ITy);
2894  }
2895 
2896  // icmp pred (urem X, Y), Y
2897  if (match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2898  switch (Pred) {
2899  default:
2900  break;
2901  case ICmpInst::ICMP_SGT:
2902  case ICmpInst::ICMP_SGE: {
2903  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2904  if (!Known.isNonNegative())
2905  break;
2907  }
2908  case ICmpInst::ICMP_EQ:
2909  case ICmpInst::ICMP_UGT:
2910  case ICmpInst::ICMP_UGE:
2911  return getFalse(ITy);
2912  case ICmpInst::ICMP_SLT:
2913  case ICmpInst::ICMP_SLE: {
2914  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2915  if (!Known.isNonNegative())
2916  break;
2918  }
2919  case ICmpInst::ICMP_NE:
2920  case ICmpInst::ICMP_ULT:
2921  case ICmpInst::ICMP_ULE:
2922  return getTrue(ITy);
2923  }
2924  }
2925 
2926  // icmp pred (urem X, Y), X
2927  if (match(LBO, m_URem(m_Specific(RHS), m_Value()))) {
2928  if (Pred == ICmpInst::ICMP_ULE)
2929  return getTrue(ITy);
2930  if (Pred == ICmpInst::ICMP_UGT)
2931  return getFalse(ITy);
2932  }
2933 
2934  // x >> y <=u x
2935  // x udiv y <=u x.
2936  if (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2937  match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2938  // icmp pred (X op Y), X
2939  if (Pred == ICmpInst::ICMP_UGT)
2940  return getFalse(ITy);
2941  if (Pred == ICmpInst::ICMP_ULE)
2942  return getTrue(ITy);
2943  }
2944 
2945  // (x*C1)/C2 <= x for C1 <= C2.
2946  // This holds even if the multiplication overflows: Assume that x != 0 and
2947  // arithmetic is modulo M. For overflow to occur we must have C1 >= M/x and
2948  // thus C2 >= M/x. It follows that (x*C1)/C2 <= (M-1)/C2 <= ((M-1)*x)/M < x.
2949  //
2950  // Additionally, either the multiplication and division might be represented
2951  // as shifts:
2952  // (x*C1)>>C2 <= x for C1 < 2**C2.
2953  // (x<<C1)/C2 <= x for 2**C1 < C2.
2954  const APInt *C1, *C2;
2955  if ((match(LBO, m_UDiv(m_Mul(m_Specific(RHS), m_APInt(C1)), m_APInt(C2))) &&
2956  C1->ule(*C2)) ||
2957  (match(LBO, m_LShr(m_Mul(m_Specific(RHS), m_APInt(C1)), m_APInt(C2))) &&
2958  C1->ule(APInt(C2->getBitWidth(), 1) << *C2)) ||
2959  (match(LBO, m_UDiv(m_Shl(m_Specific(RHS), m_APInt(C1)), m_APInt(C2))) &&
2960  (APInt(C1->getBitWidth(), 1) << *C1).ule(*C2))) {
2961  if (Pred == ICmpInst::ICMP_UGT)
2962  return getFalse(ITy);
2963  if (Pred == ICmpInst::ICMP_ULE)
2964  return getTrue(ITy);
2965  }
2966 
2967  return nullptr;
2968 }
2969 
2970 
2971 // If only one of the icmp's operands has NSW flags, try to prove that:
2972 //
2973 // icmp slt (x + C1), (x +nsw C2)
2974 //
2975 // is equivalent to:
2976 //
2977 // icmp slt C1, C2
2978 //
2979 // which is true if x + C2 has the NSW flags set and:
2980 // *) C1 < C2 && C1 >= 0, or
2981 // *) C2 < C1 && C1 <= 0.
2982 //
2984  Value *RHS) {
2985  // TODO: only support icmp slt for now.
2986  if (Pred != CmpInst::ICMP_SLT)
2987  return false;
2988 
2989  // Canonicalize nsw add as RHS.
2990  if (!match(RHS, m_NSWAdd(m_Value(), m_Value())))
2991  std::swap(LHS, RHS);
2992  if (!match(RHS, m_NSWAdd(m_Value(), m_Value())))
2993  return false;
2994 
2995  Value *X;
2996  const APInt *C1, *C2;
2997  if (!match(LHS, m_c_Add(m_Value(X), m_APInt(C1))) ||
2998  !match(RHS, m_c_Add(m_Specific(X), m_APInt(C2))))
2999  return false;
3000 
3001  return (C1->slt(*C2) && C1->isNonNegative()) ||
3002  (C2->slt(*C1) && C1->isNonPositive());
3003 }
3004 
3005 
3006 /// TODO: A large part of this logic is duplicated in InstCombine's
3007 /// foldICmpBinOp(). We should be able to share that and avoid the code
3008 /// duplication.
3010  Value *RHS, const SimplifyQuery &Q,
3011  unsigned MaxRecurse) {
3012  BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
3013  BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
3014  if (MaxRecurse && (LBO || RBO)) {
3015  // Analyze the case when either LHS or RHS is an add instruction.
3016  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3017  // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
3018  bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
3019  if (LBO && LBO->getOpcode() == Instruction::Add) {
3020  A = LBO->getOperand(0);
3021  B = LBO->getOperand(1);
3022  NoLHSWrapProblem =
3023  ICmpInst::isEquality(Pred) ||
3024  (CmpInst::isUnsigned(Pred) &&
3025  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
3026  (CmpInst::isSigned(Pred) &&
3027  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
3028  }
3029  if (RBO && RBO->getOpcode() == Instruction::Add) {
3030  C = RBO->getOperand(0);
3031  D = RBO->getOperand(1);
3032  NoRHSWrapProblem =
3033  ICmpInst::isEquality(Pred) ||
3034  (CmpInst::isUnsigned(Pred) &&
3035  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
3036  (CmpInst::isSigned(Pred) &&
3037  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
3038  }
3039 
3040  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3041  if ((A == RHS || B == RHS) && NoLHSWrapProblem)
3042  if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
3043  Constant::getNullValue(RHS->getType()), Q,
3044  MaxRecurse - 1))
3045  return V;
3046 
3047  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3048  if ((C == LHS || D == LHS) && NoRHSWrapProblem)
3049  if (Value *V =
3051  C == LHS ? D : C, Q, MaxRecurse - 1))
3052  return V;
3053 
3054  // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
3055  bool CanSimplify = (NoLHSWrapProblem && NoRHSWrapProblem) ||
3056  trySimplifyICmpWithAdds(Pred, LHS, RHS);
3057  if (A && C && (A == C || A == D || B == C || B == D) && CanSimplify) {
3058  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3059  Value *Y, *Z;
3060  if (A == C) {
3061  // C + B == C + D -> B == D
3062  Y = B;
3063  Z = D;
3064  } else if (A == D) {
3065  // D + B == C + D -> B == C
3066  Y = B;
3067  Z = C;
3068  } else if (B == C) {
3069  // A + C == C + D -> A == D
3070  Y = A;
3071  Z = D;
3072  } else {
3073  assert(B == D);
3074  // A + D == C + D -> A == C
3075  Y = A;
3076  Z = C;
3077  }
3078  if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
3079  return V;
3080  }
3081  }
3082 
3083  if (LBO)
3084  if (Value *V = simplifyICmpWithBinOpOnLHS(Pred, LBO, RHS, Q, MaxRecurse))
3085  return V;
3086 
3087  if (RBO)
3089  ICmpInst::getSwappedPredicate(Pred), RBO, LHS, Q, MaxRecurse))
3090  return V;
3091 
3092  // 0 - (zext X) pred C
3093  if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
3094  const APInt *C;
3095  if (match(RHS, m_APInt(C))) {
3096  if (C->isStrictlyPositive()) {
3097  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_NE)
3098  return ConstantInt::getTrue(GetCompareTy(RHS));
3099  if (Pred == ICmpInst::ICMP_SGE || Pred == ICmpInst::ICMP_EQ)
3100  return ConstantInt::getFalse(GetCompareTy(RHS));
3101  }
3102  if (C->isNonNegative()) {
3103  if (Pred == ICmpInst::ICMP_SLE)
3104  return ConstantInt::getTrue(GetCompareTy(RHS));
3105  if (Pred == ICmpInst::ICMP_SGT)
3106  return ConstantInt::getFalse(GetCompareTy(RHS));
3107  }
3108  }
3109  }
3110 
3111  // If C2 is a power-of-2 and C is not:
3112  // (C2 << X) == C --> false
3113  // (C2 << X) != C --> true
3114  const APInt *C;
3115  if (match(LHS, m_Shl(m_Power2(), m_Value())) &&
3116  match(RHS, m_APIntAllowUndef(C)) && !C->isPowerOf2()) {
3117  // C2 << X can equal zero in some circumstances.
3118  // This simplification might be unsafe if C is zero.
3119  //
3120  // We know it is safe if:
3121  // - The shift is nsw. We can't shift out the one bit.
3122  // - The shift is nuw. We can't shift out the one bit.
3123  // - C2 is one.
3124  // - C isn't zero.
3125  if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
3126  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
3127  match(LHS, m_Shl(m_One(), m_Value())) || !C->isNullValue()) {
3128  if (Pred == ICmpInst::ICMP_EQ)
3129  return ConstantInt::getFalse(GetCompareTy(RHS));
3130  if (Pred == ICmpInst::ICMP_NE)
3131  return ConstantInt::getTrue(GetCompareTy(RHS));
3132  }
3133  }
3134 
3135  // TODO: This is overly constrained. LHS can be any power-of-2.
3136  // (1 << X) >u 0x8000 --> false
3137  // (1 << X) <=u 0x8000 --> true
3138  if (match(LHS, m_Shl(m_One(), m_Value())) && match(RHS, m_SignMask())) {
3139  if (Pred == ICmpInst::ICMP_UGT)
3140  return ConstantInt::getFalse(GetCompareTy(RHS));
3141  if (Pred == ICmpInst::ICMP_ULE)
3142  return ConstantInt::getTrue(GetCompareTy(RHS));
3143  }
3144 
3145  if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
3146  LBO->getOperand(1) == RBO->getOperand(1)) {
3147  switch (LBO->getOpcode()) {
3148  default:
3149  break;
3150  case Instruction::UDiv:
3151  case Instruction::LShr:
3152  if (ICmpInst::isSigned(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::SDiv:
3160  if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
3161  !Q.IIQ.isExact(RBO))
3162  break;
3163  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3164  RBO->getOperand(0), Q, MaxRecurse - 1))
3165  return V;
3166  break;
3167  case Instruction::AShr:
3168  if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
3169  break;
3170  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3171  RBO->getOperand(0), Q, MaxRecurse - 1))
3172  return V;
3173  break;
3174  case Instruction::Shl: {
3175  bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
3176  bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
3177  if (!NUW && !NSW)
3178  break;
3179  if (!NSW && ICmpInst::isSigned(Pred))
3180  break;
3181  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3182  RBO->getOperand(0), Q, MaxRecurse - 1))
3183  return V;
3184  break;
3185  }
3186  }
3187  }
3188  return nullptr;
3189 }
3190 
3191 /// Simplify integer comparisons where at least one operand of the compare
3192 /// matches an integer min/max idiom.
3194  Value *RHS, const SimplifyQuery &Q,
3195  unsigned MaxRecurse) {
3196  Type *ITy = GetCompareTy(LHS); // The return type.
3197  Value *A, *B;
3199  CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3200 
3201  // Signed variants on "max(a,b)>=a -> true".
3202  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3203  if (A != RHS)
3204  std::swap(A, B); // smax(A, B) pred A.
3205  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3206  // We analyze this as smax(A, B) pred A.
3207  P = Pred;
3208  } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
3209  (A == LHS || B == LHS)) {
3210  if (A != LHS)
3211  std::swap(A, B); // A pred smax(A, B).
3212  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3213  // We analyze this as smax(A, B) swapped-pred A.
3215  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3216  (A == RHS || B == RHS)) {
3217  if (A != RHS)
3218  std::swap(A, B); // smin(A, B) pred A.
3219  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3220  // We analyze this as smax(-A, -B) swapped-pred -A.
3221  // Note that we do not need to actually form -A or -B thanks to EqP.
3223  } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
3224  (A == LHS || B == LHS)) {
3225  if (A != LHS)
3226  std::swap(A, B); // A pred smin(A, B).
3227  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3228  // We analyze this as smax(-A, -B) pred -A.
3229  // Note that we do not need to actually form -A or -B thanks to EqP.
3230  P = Pred;
3231  }
3232  if (P != CmpInst::BAD_ICMP_PREDICATE) {
3233  // Cases correspond to "max(A, B) p A".
3234  switch (P) {
3235  default:
3236  break;
3237  case CmpInst::ICMP_EQ:
3238  case CmpInst::ICMP_SLE:
3239  // Equivalent to "A EqP B". This may be the same as the condition tested
3240  // in the max/min; if so, we can just return that.
3241  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3242  return V;
3243  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3244  return V;
3245  // Otherwise, see if "A EqP B" simplifies.
3246  if (MaxRecurse)
3247  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3248  return V;
3249  break;
3250  case CmpInst::ICMP_NE:
3251  case CmpInst::ICMP_SGT: {
3253  // Equivalent to "A InvEqP B". This may be the same as the condition
3254  // tested in the max/min; if so, we can just return that.
3255  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3256  return V;
3257  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3258  return V;
3259  // Otherwise, see if "A InvEqP B" simplifies.
3260  if (MaxRecurse)
3261  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3262  return V;
3263  break;
3264  }
3265  case CmpInst::ICMP_SGE:
3266  // Always true.
3267  return getTrue(ITy);
3268  case CmpInst::ICMP_SLT:
3269  // Always false.
3270  return getFalse(ITy);
3271  }
3272  }
3273 
3274  // Unsigned variants on "max(a,b)>=a -> true".
3276  if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3277  if (A != RHS)
3278  std::swap(A, B); // umax(A, B) pred A.
3279  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3280  // We analyze this as umax(A, B) pred A.
3281  P = Pred;
3282  } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3283  (A == LHS || B == LHS)) {
3284  if (A != LHS)
3285  std::swap(A, B); // A pred umax(A, B).
3286  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3287  // We analyze this as umax(A, B) swapped-pred A.
3289  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3290  (A == RHS || B == RHS)) {
3291  if (A != RHS)
3292  std::swap(A, B); // umin(A, B) pred A.
3293  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3294  // We analyze this as umax(-A, -B) swapped-pred -A.
3295  // Note that we do not need to actually form -A or -B thanks to EqP.
3297  } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3298  (A == LHS || B == LHS)) {
3299  if (A != LHS)
3300  std::swap(A, B); // A pred umin(A, B).
3301  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3302  // We analyze this as umax(-A, -B) pred -A.
3303  // Note that we do not need to actually form -A or -B thanks to EqP.
3304  P = Pred;
3305  }
3306  if (P != CmpInst::BAD_ICMP_PREDICATE) {
3307  // Cases correspond to "max(A, B) p A".
3308  switch (P) {
3309  default:
3310  break;
3311  case CmpInst::ICMP_EQ:
3312  case CmpInst::ICMP_ULE:
3313  // Equivalent to "A EqP B". This may be the same as the condition tested
3314  // in the max/min; if so, we can just return that.
3315  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3316  return V;
3317  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3318  return V;
3319  // Otherwise, see if "A EqP B" simplifies.
3320  if (MaxRecurse)
3321  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3322  return V;
3323  break;
3324  case CmpInst::ICMP_NE:
3325  case CmpInst::ICMP_UGT: {
3327  // Equivalent to "A InvEqP B". This may be the same as the condition
3328  // tested in the max/min; if so, we can just return that.
3329  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3330  return V;
3331  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3332  return V;
3333  // Otherwise, see if "A InvEqP B" simplifies.
3334  if (MaxRecurse)
3335  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3336  return V;
3337  break;
3338  }
3339  case CmpInst::ICMP_UGE:
3340  return getTrue(ITy);
3341  case CmpInst::ICMP_ULT:
3342  return getFalse(ITy);
3343  }
3344  }
3345 
3346  // Comparing 1 each of min/max with a common operand?
3347  // Canonicalize min operand to RHS.
3348  if (match(LHS, m_UMin(m_Value(), m_Value())) ||
3349  match(LHS, m_SMin(m_Value(), m_Value()))) {
3350  std::swap(LHS, RHS);
3351  Pred = ICmpInst::getSwappedPredicate(Pred);
3352  }
3353 
3354  Value *C, *D;
3355  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3356  match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3357  (A == C || A == D || B == C || B == D)) {
3358  // smax(A, B) >=s smin(A, D) --> true
3359  if (Pred == CmpInst::ICMP_SGE)
3360  return getTrue(ITy);
3361  // smax(A, B) <s smin(A, D) --> false
3362  if (Pred == CmpInst::ICMP_SLT)
3363  return getFalse(ITy);
3364  } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3365  match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3366  (A == C || A == D || B == C || B == D)) {
3367  // umax(A, B) >=u umin(A, D) --> true
3368  if (Pred == CmpInst::ICMP_UGE)
3369  return getTrue(ITy);
3370  // umax(A, B) <u umin(A, D) --> false
3371  if (Pred == CmpInst::ICMP_ULT)
3372  return getFalse(ITy);
3373  }
3374 
3375  return nullptr;
3376 }
3377 
3379  Value *LHS, Value *RHS,
3380  const SimplifyQuery &Q) {
3381  // Gracefully handle instructions that have not been inserted yet.
3382  if (!Q.AC || !Q.CxtI || !Q.CxtI->getParent())
3383  return nullptr;
3384 
3385  for (Value *AssumeBaseOp : {LHS, RHS}) {
3386  for (auto &AssumeVH : Q.AC->assumptionsFor(AssumeBaseOp)) {
3387  if (!AssumeVH)
3388  continue;
3389 
3390  CallInst *Assume = cast<CallInst>(AssumeVH);
3391  if (Optional<bool> Imp =
3392  isImpliedCondition(Assume->getArgOperand(0), Predicate, LHS, RHS,
3393  Q.DL))
3394  if (isValidAssumeForContext(Assume, Q.CxtI, Q.DT))
3395  return ConstantInt::get(GetCompareTy(LHS), *Imp);
3396  }
3397  }
3398 
3399  return nullptr;
3400 }
3401 
3402 /// Given operands for an ICmpInst, see if we can fold the result.
3403 /// If not, this returns null.
3404 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3405  const SimplifyQuery &Q, unsigned MaxRecurse) {
3407  assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3408 
3409  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3410  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3411  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3412 
3413  // If we have a constant, make sure it is on the RHS.
3414  std::swap(LHS, RHS);
3415  Pred = CmpInst::getSwappedPredicate(Pred);
3416  }
3417  assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3418 
3419  Type *ITy = GetCompareTy(LHS); // The return type.
3420 
3421  // For EQ and NE, we can always pick a value for the undef to make the
3422  // predicate pass or fail, so we can return undef.
3423  // Matches behavior in llvm::ConstantFoldCompareInstruction.
3424  if (Q.isUndefValue(RHS) && ICmpInst::isEquality(Pred))
3425  return UndefValue::get(ITy);
3426 
3427  // icmp X, X -> true/false
3428  // icmp X, undef -> true/false because undef could be X.
3429  if (LHS == RHS || Q.isUndefValue(RHS))
3430  return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3431 
3432  if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3433  return V;
3434 
3435  // TODO: Sink/common this with other potentially expensive calls that use
3436  // ValueTracking? See comment below for isKnownNonEqual().
3437  if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3438  return V;
3439 
3440  if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3441  return V;
3442 
3443  // If both operands have range metadata, use the metadata
3444  // to simplify the comparison.
3445  if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3446  auto RHS_Instr = cast<Instruction>(RHS);
3447  auto LHS_Instr = cast<Instruction>(LHS);
3448 
3449  if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3450  Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3451  auto RHS_CR = getConstantRangeFromMetadata(
3452  *RHS_Instr->getMetadata(LLVMContext::MD_range));
3453  auto LHS_CR = getConstantRangeFromMetadata(
3454  *LHS_Instr->getMetadata(LLVMContext::MD_range));
3455 
3456  if (LHS_CR.icmp(Pred, RHS_CR))
3457  return ConstantInt::getTrue(RHS->getContext());
3458 
3459  if (LHS_CR.icmp(CmpInst::getInversePredicate(Pred), RHS_CR))
3460  return ConstantInt::getFalse(RHS->getContext());
3461  }
3462  }
3463 
3464  // Compare of cast, for example (zext X) != 0 -> X != 0
3465  if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3466  Instruction *LI = cast<CastInst>(LHS);
3467  Value *SrcOp = LI->getOperand(0);
3468  Type *SrcTy = SrcOp->getType();
3469  Type *DstTy = LI->getType();
3470 
3471  // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3472  // if the integer type is the same size as the pointer type.
3473  if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3474  Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3475  if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3476  // Transfer the cast to the constant.
3477  if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3478  ConstantExpr::getIntToPtr(RHSC, SrcTy),
3479  Q, MaxRecurse-1))
3480  return V;
3481  } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3482  if (RI->getOperand(0)->getType() == SrcTy)
3483  // Compare without the cast.
3484  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3485  Q, MaxRecurse-1))
3486  return V;
3487  }
3488  }
3489 
3490  if (isa<ZExtInst>(LHS)) {
3491  // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3492  // same type.
3493  if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3494  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3495  // Compare X and Y. Note that signed predicates become unsigned.
3497  SrcOp, RI->getOperand(0), Q,
3498  MaxRecurse-1))
3499  return V;
3500  }
3501  // Fold (zext X) ule (sext X), (zext X) sge (sext X) to true.
3502  else if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3503  if (SrcOp == RI->getOperand(0)) {
3504  if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_SGE)
3505  return ConstantInt::getTrue(ITy);
3506  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SLT)
3507  return ConstantInt::getFalse(ITy);
3508  }
3509  }
3510  // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3511  // too. If not, then try to deduce the result of the comparison.
3512  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3513  // Compute the constant that would happen if we truncated to SrcTy then
3514  // reextended to DstTy.
3515  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3516  Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3517 
3518  // If the re-extended constant didn't change then this is effectively
3519  // also a case of comparing two zero-extended values.
3520  if (RExt == CI && MaxRecurse)
3522  SrcOp, Trunc, Q, MaxRecurse-1))
3523  return V;
3524 
3525  // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3526  // there. Use this to work out the result of the comparison.
3527  if (RExt != CI) {
3528  switch (Pred) {
3529  default: llvm_unreachable("Unknown ICmp predicate!");
3530  // LHS <u RHS.
3531  case ICmpInst::ICMP_EQ:
3532  case ICmpInst::ICMP_UGT:
3533  case ICmpInst::ICMP_UGE:
3534  return ConstantInt::getFalse(CI->getContext());
3535 
3536  case ICmpInst::ICMP_NE:
3537  case ICmpInst::ICMP_ULT:
3538  case ICmpInst::ICMP_ULE:
3539  return ConstantInt::getTrue(CI->getContext());
3540 
3541  // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3542  // is non-negative then LHS <s RHS.
3543  case ICmpInst::ICMP_SGT:
3544  case ICmpInst::ICMP_SGE:
3545  return CI->getValue().isNegative() ?
3546  ConstantInt::getTrue(CI->getContext()) :
3547  ConstantInt::getFalse(CI->getContext());
3548 
3549  case ICmpInst::ICMP_SLT:
3550  case ICmpInst::ICMP_SLE:
3551  return CI->getValue().isNegative() ?
3552  ConstantInt::getFalse(CI->getContext()) :
3553  ConstantInt::getTrue(CI->getContext());
3554  }
3555  }
3556  }
3557  }
3558 
3559  if (isa<SExtInst>(LHS)) {
3560  // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3561  // same type.
3562  if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3563  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3564  // Compare X and Y. Note that the predicate does not change.
3565  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3566  Q, MaxRecurse-1))
3567  return V;
3568  }
3569  // Fold (sext X) uge (zext X), (sext X) sle (zext X) to true.
3570  else if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3571  if (SrcOp == RI->getOperand(0)) {
3572  if (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_SLE)
3573  return ConstantInt::getTrue(ITy);
3574  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SGT)
3575  return ConstantInt::getFalse(ITy);
3576  }
3577  }
3578  // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3579  // too. If not, then try to deduce the result of the comparison.
3580  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3581  // Compute the constant that would happen if we truncated to SrcTy then
3582  // reextended to DstTy.
3583  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3584  Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3585 
3586  // If the re-extended constant didn't change then this is effectively
3587  // also a case of comparing two sign-extended values.
3588  if (RExt == CI && MaxRecurse)
3589  if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3590  return V;
3591 
3592  // Otherwise the upper bits of LHS are all equal, while RHS has varying
3593  // bits there. Use this to work out the result of the comparison.
3594  if (RExt != CI) {
3595  switch (Pred) {
3596  default: llvm_unreachable("Unknown ICmp predicate!");
3597  case ICmpInst::ICMP_EQ:
3598  return ConstantInt::getFalse(CI->getContext());
3599  case ICmpInst::ICMP_NE:
3600  return ConstantInt::getTrue(CI->getContext());
3601 
3602  // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3603  // LHS >s RHS.
3604  case ICmpInst::ICMP_SGT:
3605  case ICmpInst::ICMP_SGE:
3606  return CI->getValue().isNegative() ?
3607  ConstantInt::getTrue(CI->getContext()) :
3608  ConstantInt::getFalse(CI->getContext());
3609  case ICmpInst::ICMP_SLT:
3610  case ICmpInst::ICMP_SLE:
3611  return CI->getValue().isNegative() ?
3612  ConstantInt::getFalse(CI->getContext()) :
3613  ConstantInt::getTrue(CI->getContext());
3614 
3615  // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3616  // LHS >u RHS.
3617  case ICmpInst::ICMP_UGT:
3618  case ICmpInst::ICMP_UGE:
3619  // Comparison is true iff the LHS <s 0.
3620  if (MaxRecurse)
3622  Constant::getNullValue(SrcTy),
3623  Q, MaxRecurse-1))
3624  return V;
3625  break;
3626  case ICmpInst::ICMP_ULT:
3627  case ICmpInst::ICMP_ULE:
3628  // Comparison is true iff the LHS >=s 0.
3629  if (MaxRecurse)
3631  Constant::getNullValue(SrcTy),
3632  Q, MaxRecurse-1))
3633  return V;
3634  break;
3635  }
3636  }
3637  }
3638  }
3639  }
3640 
3641  // icmp eq|ne X, Y -> false|true if X != Y
3642  // This is potentially expensive, and we have already computedKnownBits for
3643  // compares with 0 above here, so only try this for a non-zero compare.
3644  if (ICmpInst::isEquality(Pred) && !match(RHS, m_Zero()) &&
3645  isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3646  return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3647  }
3648 
3649  if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3650  return V;
3651 
3652  if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3653  return V;
3654 
3655  if (Value *V = simplifyICmpWithDominatingAssume(Pred, LHS, RHS, Q))
3656  return V;
3657 
3658  // Simplify comparisons of related pointers using a powerful, recursive
3659  // GEP-walk when we have target data available..
3660  if (LHS->getType()->isPointerTy())
3661  if (auto *C = computePointerICmp(Pred, LHS, RHS, Q))
3662  return C;
3663  if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3664  if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3665  if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3666  Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3667  Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3668  Q.DL.getTypeSizeInBits(CRHS->getType()))
3669  if (auto *C = computePointerICmp(Pred, CLHS->getPointerOperand(),
3670  CRHS->getPointerOperand(), Q))
3671  return C;
3672 
3673  if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3674  if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3675  if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3676  GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3677  (ICmpInst::isEquality(Pred) ||
3678  (GLHS->isInBounds() && GRHS->isInBounds() &&
3679  Pred == ICmpInst::getSignedPredicate(Pred)))) {
3680  // The bases are equal and the indices are constant. Build a constant
3681  // expression GEP with the same indices and a null base pointer to see
3682  // what constant folding can make out of it.
3683  Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3684  SmallVector<Value *, 4> IndicesLHS(GLHS->indices());
3686  GLHS->getSourceElementType(), Null, IndicesLHS);
3687 
3688  SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3690  GLHS->getSourceElementType(), Null, IndicesRHS);
3691  Constant *NewICmp = ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3692  return ConstantFoldConstant(NewICmp, Q.DL);
3693  }
3694  }
3695  }
3696 
3697  // If the comparison is with the result of a select instruction, check whether
3698  // comparing with either branch of the select always yields the same value.
3699  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3700  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3701  return V;
3702 
3703  // If the comparison is with the result of a phi instruction, check whether
3704  // doing the compare with each incoming phi value yields a common result.
3705  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3706  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3707  return V;
3708 
3709  return nullptr;
3710 }
3711 
3713  const SimplifyQuery &Q) {
3715 }
3716 
3717 /// Given operands for an FCmpInst, see if we can fold the result.
3718 /// If not, this returns null.
3719 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3720  FastMathFlags FMF, const SimplifyQuery &Q,
3721  unsigned MaxRecurse) {
3723  assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3724 
3725  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3726  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3727  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3728 
3729  // If we have a constant, make sure it is on the RHS.
3730  std::swap(LHS, RHS);
3731  Pred = CmpInst::getSwappedPredicate(Pred);
3732  }
3733 
3734  // Fold trivial predicates.
3735  Type *RetTy = GetCompareTy(LHS);
3736  if (Pred == FCmpInst::FCMP_FALSE)
3737  return getFalse(RetTy);
3738  if (Pred == FCmpInst::FCMP_TRUE)
3739  return getTrue(RetTy);
3740 
3741  // Fold (un)ordered comparison if we can determine there are no NaNs.
3742  if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3743  if (FMF.noNaNs() ||
3744  (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3745  return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3746 
3747  // NaN is unordered; NaN is not ordered.
3749  "Comparison must be either ordered or unordered");
3750  if (match(RHS, m_NaN()))
3751  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3752 
3753  // fcmp pred x, undef and fcmp pred undef, x
3754  // fold to true if unordered, false if ordered
3755  if (Q.isUndefValue(LHS) || Q.isUndefValue(RHS)) {
3756  // Choosing NaN for the undef will always make unordered comparison succeed
3757  // and ordered comparison fail.
3758  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3759  }
3760 
3761  // fcmp x,x -> true/false. Not all compares are foldable.
3762  if (LHS == RHS) {
3763  if (CmpInst::isTrueWhenEqual(Pred))
3764  return getTrue(RetTy);
3765  if (CmpInst::isFalseWhenEqual(Pred))
3766  return getFalse(RetTy);
3767  }
3768 
3769  // Handle fcmp with constant RHS.
3770  // TODO: Use match with a specific FP value, so these work with vectors with
3771  // undef lanes.
3772  const APFloat *C;
3773  if (match(RHS, m_APFloat(C))) {
3774  // Check whether the constant is an infinity.
3775  if (C->isInfinity()) {
3776  if (C->isNegative()) {
3777  switch (Pred) {
3778  case FCmpInst::FCMP_OLT:
3779  // No value is ordered and less than negative infinity.
3780  return getFalse(RetTy);
3781  case FCmpInst::FCMP_UGE:
3782  // All values are unordered with or at least negative infinity.
3783  return getTrue(RetTy);
3784  default:
3785  break;
3786  }
3787  } else {
3788  switch (Pred) {
3789  case FCmpInst::FCMP_OGT:
3790  // No value is ordered and greater than infinity.
3791  return getFalse(RetTy);
3792  case FCmpInst::FCMP_ULE:
3793  // All values are unordered with and at most infinity.
3794  return getTrue(RetTy);
3795  default:
3796  break;
3797  }
3798  }
3799 
3800  // LHS == Inf
3801  if (Pred == FCmpInst::FCMP_OEQ && isKnownNeverInfinity(LHS, Q.TLI))
3802  return getFalse(RetTy);
3803  // LHS != Inf
3804  if (Pred == FCmpInst::FCMP_UNE && isKnownNeverInfinity(LHS, Q.TLI))
3805  return getTrue(RetTy);
3806  // LHS == Inf || LHS == NaN
3807  if (Pred == FCmpInst::FCMP_UEQ && isKnownNeverInfinity(LHS, Q.TLI) &&
3808  isKnownNeverNaN(LHS, Q.TLI))
3809  return getFalse(RetTy);
3810  // LHS != Inf && LHS != NaN
3811  if (Pred == FCmpInst::FCMP_ONE && isKnownNeverInfinity(LHS, Q.TLI) &&
3812  isKnownNeverNaN(LHS, Q.TLI))
3813  return getTrue(RetTy);
3814  }
3815  if (C->isNegative() && !C->isNegZero()) {
3816  assert(!C->isNaN() && "Unexpected NaN constant!");
3817  // TODO: We can catch more cases by using a range check rather than
3818  // relying on CannotBeOrderedLessThanZero.
3819  switch (Pred) {
3820  case FCmpInst::FCMP_UGE:
3821  case FCmpInst::FCMP_UGT:
3822  case FCmpInst::FCMP_UNE:
3823  // (X >= 0) implies (X > C) when (C < 0)
3824  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3825  return getTrue(RetTy);
3826  break;
3827  case FCmpInst::FCMP_OEQ:
3828  case FCmpInst::FCMP_OLE:
3829  case FCmpInst::FCMP_OLT:
3830  // (X >= 0) implies !(X < C) when (C < 0)
3831  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3832  return getFalse(RetTy);
3833  break;
3834  default:
3835  break;
3836  }
3837  }
3838 
3839  // Check comparison of [minnum/maxnum with constant] with other constant.
3840  const APFloat *C2;
3841  if ((match(LHS, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_APFloat(C2))) &&
3842  *C2 < *C) ||
3843  (match(LHS, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_APFloat(C2))) &&
3844  *C2 > *C)) {
3845  bool IsMaxNum =
3846  cast<IntrinsicInst>(LHS)->getIntrinsicID() == Intrinsic::maxnum;
3847  // The ordered relationship and minnum/maxnum guarantee that we do not
3848  // have NaN constants, so ordered/unordered preds are handled the same.
3849  switch (Pred) {
3851  // minnum(X, LesserC) == C --> false
3852  // maxnum(X, GreaterC) == C --> false
3853  return getFalse(RetTy);
3855  // minnum(X, LesserC) != C --> true
3856  // maxnum(X, GreaterC) != C --> true
3857  return getTrue(RetTy);
3860  // minnum(X, LesserC) >= C --> false
3861  // minnum(X, LesserC) > C --> false
3862  // maxnum(X, GreaterC) >= C --> true
3863  // maxnum(X, GreaterC) > C --> true
3864  return ConstantInt::get(RetTy, IsMaxNum);
3867  // minnum(X, LesserC) <= C --> true
3868  // minnum(X, LesserC) < C --> true
3869  // maxnum(X, GreaterC) <= C --> false
3870  // maxnum(X, GreaterC) < C --> false
3871  return ConstantInt::get(RetTy, !IsMaxNum);
3872  default:
3873  // TRUE/FALSE/ORD/UNO should be handled before this.
3874  llvm_unreachable("Unexpected fcmp predicate");
3875  }
3876  }
3877  }
3878 
3879  if (match(RHS, m_AnyZeroFP())) {
3880  switch (Pred) {
3881  case FCmpInst::FCMP_OGE:
3882  case FCmpInst::FCMP_ULT:
3883  // Positive or zero X >= 0.0 --> true
3884  // Positive or zero X < 0.0 --> false
3885  if ((FMF.noNaNs() || isKnownNeverNaN(LHS, Q.TLI)) &&
3887  return Pred == FCmpInst::FCMP_OGE ? getTrue(RetTy) : getFalse(RetTy);
3888  break;
3889  case FCmpInst::FCMP_UGE:
3890  case FCmpInst::FCMP_OLT:
3891  // Positive or zero or nan X >= 0.0 --> true
3892  // Positive or zero or nan X < 0.0 --> false
3893  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3894  return Pred == FCmpInst::FCMP_UGE ? getTrue(RetTy) : getFalse(RetTy);
3895  break;
3896  default:
3897  break;
3898  }
3899  }
3900 
3901  // If the comparison is with the result of a select instruction, check whether
3902  // comparing with either branch of the select always yields the same value.
3903  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3904  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3905  return V;
3906 
3907  // If the comparison is with the result of a phi instruction, check whether
3908  // doing the compare with each incoming phi value yields a common result.
3909  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3910  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3911  return V;
3912 
3913  return nullptr;
3914 }
3915 
3917  FastMathFlags FMF, const SimplifyQuery &Q) {
3919 }
3920 
3922  const SimplifyQuery &Q,
3923  bool AllowRefinement,
3924  unsigned MaxRecurse) {
3925  // Trivial replacement.
3926  if (V == Op)
3927  return RepOp;
3928 
3929  // We cannot replace a constant, and shouldn't even try.
3930  if (isa<Constant>(Op))
3931  return nullptr;
3932 
3933  auto *I = dyn_cast<Instruction>(V);
3934  if (!I || !is_contained(I->operands(), Op))
3935  return nullptr;
3936 
3937  // Replace Op with RepOp in instruction operands.
3938  SmallVector<Value *, 8> NewOps(I->getNumOperands());
3939  transform(I->operands(), NewOps.begin(),
3940  [&](Value *V) { return V == Op ? RepOp : V; });
3941 
3942  if (!AllowRefinement) {
3943  // General InstSimplify functions may refine the result, e.g. by returning
3944  // a constant for a potentially poison value. To avoid this, implement only
3945  // a few non-refining but profitable transforms here.
3946 
3947  if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3948  unsigned Opcode = BO->getOpcode();
3949  // id op x -> x, x op id -> x
3950  if (NewOps[0] == ConstantExpr::getBinOpIdentity(Opcode, I->getType()))
3951  return NewOps[1];
3952  if (NewOps[1] == ConstantExpr::getBinOpIdentity(Opcode, I->getType(),
3953  /* RHS */ true))
3954  return NewOps[0];
3955 
3956  // x & x -> x, x | x -> x
3957  if ((Opcode == Instruction::And || Opcode == Instruction::Or) &&
3958  NewOps[0] == NewOps[1])
3959  return NewOps[0];
3960  }
3961 
3962  if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3963  // getelementptr x, 0 -> x
3964  if (NewOps.size() == 2 && match(NewOps[1], m_Zero()) &&
3965  !GEP->isInBounds())
3966  return NewOps[0];
3967  }
3968  } else if (MaxRecurse) {
3969  // The simplification queries below may return the original value. Consider:
3970  // %div = udiv i32 %arg, %arg2
3971  // %mul = mul nsw i32 %div, %arg2
3972  // %cmp = icmp eq i32 %mul, %arg
3973  // %sel = select i1 %cmp, i32 %div, i32 undef
3974  // Replacing %arg by %mul, %div becomes "udiv i32 %mul, %arg2", which
3975  // simplifies back to %arg. This can only happen because %mul does not
3976  // dominate %div. To ensure a consistent return value contract, we make sure
3977  // that this case returns nullptr as well.
3978  auto PreventSelfSimplify = [V](Value *Simplified) {
3979  return Simplified != V ? Simplified : nullptr;
3980  };
3981 
3982  if (auto *B = dyn_cast<BinaryOperator>(I))
3983  return PreventSelfSimplify(SimplifyBinOp(B->getOpcode(), NewOps[0],
3984  NewOps[1], Q, MaxRecurse - 1));
3985 
3986  if (CmpInst *C = dyn_cast<CmpInst>(I))
3987  return PreventSelfSimplify(SimplifyCmpInst(C->getPredicate(), NewOps[0],
3988  NewOps[1], Q, MaxRecurse - 1));
3989 
3990  if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
3991  return PreventSelfSimplify(SimplifyGEPInst(GEP->getSourceElementType(),
3992  NewOps, Q, MaxRecurse - 1));
3993 
3994  if (isa<SelectInst>(I))
3995  return PreventSelfSimplify(
3996  SimplifySelectInst(NewOps[0], NewOps[1], NewOps[2], Q,
3997  MaxRecurse - 1));
3998  // TODO: We could hand off more cases to instsimplify here.
3999  }
4000 
4001  // If all operands are constant after substituting Op for RepOp then we can
4002  // constant fold the instruction.
4003  SmallVector<Constant *, 8> ConstOps;
4004  for (Value *NewOp : NewOps) {
4005  if (Constant *ConstOp = dyn_cast<Constant>(NewOp))
4006  ConstOps.push_back(ConstOp);
4007  else
4008  return nullptr;
4009  }
4010 
4011  // Consider:
4012  // %cmp = icmp eq i32 %x, 2147483647
4013  // %add = add nsw i32 %x, 1
4014  // %sel = select i1 %cmp, i32 -2147483648, i32 %add
4015  //
4016  // We can't replace %sel with %add unless we strip away the flags (which
4017  // will be done in InstCombine).
4018  // TODO: This may be unsound, because it only catches some forms of
4019  // refinement.
4020  if (!AllowRefinement && canCreatePoison(cast<Operator>(I)))
4021  return nullptr;
4022 
4023  if (CmpInst *C = dyn_cast<CmpInst>(I))
4024  return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
4025  ConstOps[1], Q.DL, Q.TLI);
4026 
4027  if (LoadInst *LI = dyn_cast<LoadInst>(I))
4028  if (!LI->isVolatile())
4029  return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
4030 
4031  return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
4032 }
4033 
4035  const SimplifyQuery &Q,
4036  bool AllowRefinement) {
4037  return ::SimplifyWithOpReplaced(V, Op, RepOp, Q, AllowRefinement,
4038  RecursionLimit);
4039 }
4040 
4041 /// Try to simplify a select instruction when its condition operand is an
4042 /// integer comparison where one operand of the compare is a constant.
4044  const APInt *Y, bool TrueWhenUnset) {
4045  const APInt *C;
4046 
4047  // (X & Y) == 0 ? X & ~Y : X --> X
4048  // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
4049  if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
4050  *Y == ~*C)
4051  return TrueWhenUnset ? FalseVal : TrueVal;
4052 
4053  // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
4054  // (X & Y) != 0 ? X : X & ~Y --> X
4055  if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
4056  *Y == ~*C)
4057  return TrueWhenUnset ? FalseVal : TrueVal;
4058 
4059  if (Y->isPowerOf2()) {
4060  // (X & Y) == 0 ? X | Y : X --> X | Y
4061  // (X & Y) != 0 ? X | Y : X --> X
4062  if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
4063  *Y == *C)
4064  return TrueWhenUnset ? TrueVal : FalseVal;
4065 
4066  // (X & Y) == 0 ? X : X | Y --> X
4067  // (X & Y) != 0 ? X : X | Y --> X | Y
4068  if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
4069  *Y == *C)
4070  return TrueWhenUnset ? TrueVal : FalseVal;
4071  }
4072 
4073  return nullptr;
4074 }
4075 
4076 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
4077 /// eq/ne.
4079  ICmpInst::Predicate Pred,
4080  Value *TrueVal, Value *FalseVal) {
4081  Value *X;
4082  APInt Mask;
4083  if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
4084  return nullptr;
4085 
4087  Pred == ICmpInst::ICMP_EQ);
4088 }
4089 
4090 /// Try to simplify a select instruction when its condition operand is an
4091 /// integer comparison.
4093  Value *FalseVal, const SimplifyQuery &Q,
4094  unsigned MaxRecurse) {
4095  ICmpInst::Predicate Pred;
4096  Value *CmpLHS, *CmpRHS;
4097  if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
4098  return nullptr;
4099 
4100  // Canonicalize ne to eq predicate.
4101  if (Pred == ICmpInst::ICMP_NE) {
4102  Pred = ICmpInst::ICMP_EQ;
4104  }
4105 
4106  if (Pred == ICmpInst::ICMP_EQ && match(CmpRHS, m_Zero())) {
4107  Value *X;
4108  const APInt *Y;
4109  if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
4111  /*TrueWhenUnset=*/true))
4112  return V;
4113 
4114  // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
4115  Value *ShAmt;
4116  auto isFsh = m_CombineOr(m_FShl(m_Value(X), m_Value(), m_Value(ShAmt)),
4117  m_FShr(m_Value(), m_Value(X), m_Value(ShAmt)));
4118  // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
4119  // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
4120  if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt)
4121  return X;
4122 
4123  // Test for a zero-shift-guard-op around rotates. These are used to
4124  // avoid UB from oversized shifts in raw IR rotate patterns, but the
4125  // intrinsics do not have that problem.
4126  // We do not allow this transform for the general funnel shift case because
4127  // that would not preserve the poison safety of the original code.
4128  auto isRotate =
4130  m_FShr(m_Value(X), m_Deferred(X), m_Value(ShAmt)));
4131  // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
4132  // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
4133  if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
4134  Pred == ICmpInst::ICMP_EQ)
4135  return FalseVal;
4136 
4137  // X == 0 ? abs(X) : -abs(X) --> -abs(X)
4138  // X == 0 ? -abs(X) : abs(X) --> abs(X)
4139  if (match(TrueVal, m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS))) &&
4140  match(FalseVal, m_Neg(m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS)))))
4141  return FalseVal;
4142  if (match(TrueVal,
4143  m_Neg(m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS)))) &&
4144  match(FalseVal, m_Intrinsic<Intrinsic::abs>(m_Specific(CmpLHS))))
4145  return FalseVal;
4146  }
4147 
4148  // Check for other compares that behave like bit test.
4149  if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
4150  TrueVal, FalseVal))
4151  return V;
4152 
4153  // If we have a scalar equality comparison, then we know the value in one of
4154  // the arms of the select. See if substituting this value into the arm and
4155  // simplifying the result yields the same value as the other arm.
4156  // Note that the equivalence/replacement opportunity does not hold for vectors
4157  // because each element of a vector select is chosen independently.
4158  if (Pred == ICmpInst::ICMP_EQ && !CondVal->getType()->isVectorTy()) {
4159  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q,
4160  /* AllowRefinement */ false, MaxRecurse) ==
4161  TrueVal ||
4162  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q,
4163  /* AllowRefinement */ false, MaxRecurse) ==
4164  TrueVal)
4165  return FalseVal;
4166  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q,
4167  /* AllowRefinement */ true, MaxRecurse) ==
4168  FalseVal ||
4169  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q,
4170  /* AllowRefinement */ true, MaxRecurse) ==
4171  FalseVal)
4172  return FalseVal;
4173  }
4174 
4175  return nullptr;
4176 }
4177 
4178 /// Try to simplify a select instruction when its condition operand is a
4179 /// floating-point comparison.
4181  const SimplifyQuery &Q) {
4182  FCmpInst::Predicate Pred;
4183  if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
4184  !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
4185  return nullptr;
4186 
4187  // This transform is safe if we do not have (do not care about) -0.0 or if
4188  // at least one operand is known to not be -0.0. Otherwise, the select can
4189  // change the sign of a zero operand.
4190  bool HasNoSignedZeros = Q.CxtI && isa<FPMathOperator>(Q.CxtI) &&
4191  Q.CxtI->hasNoSignedZeros();
4192  const APFloat *C;
4193  if (HasNoSignedZeros || (match(T, m_APFloat(C)) && C->isNonZero()) ||
4194  (match(F, m_APFloat(C)) && C->isNonZero())) {
4195  // (T == F) ? T : F --> F
4196  // (F == T) ? T : F --> F
4197  if (Pred == FCmpInst::FCMP_OEQ)
4198  return F;
4199 
4200  // (T != F) ? T : F --> T
4201  // (F != T) ? T : F --> T
4202  if (Pred == FCmpInst::FCMP_UNE)
4203  return T;
4204  }
4205 
4206  return nullptr;
4207 }
4208 
4209 /// Given operands for a SelectInst, see if we can fold the result.
4210 /// If not, this returns null.
4212  const SimplifyQuery &Q, unsigned MaxRecurse) {
4213  if (auto *CondC = dyn_cast<Constant>(Cond)) {
4214  if (auto *TrueC = dyn_cast<Constant>(TrueVal))
4215  if (auto *FalseC = dyn_cast<Constant>(FalseVal))
4216  return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
4217 
4218  // select undef, X, Y -> X or Y
4219  if (Q.isUndefValue(CondC))
4220  return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
4221 
4222  // TODO: Vector constants with undef elements don't simplify.
4223 
4224  // select true, X, Y -> X
4225  if (CondC->isAllOnesValue())
4226  return TrueVal;
4227  // select false, X, Y -> Y
4228  if (CondC->isNullValue())
4229  return FalseVal;
4230  }
4231 
4232  // select i1 Cond, i1 true, i1 false --> i1 Cond
4233  assert(Cond->getType()->isIntOrIntVectorTy(1) &&
4234  "Select must have bool or bool vector condition");
4235  assert(TrueVal->getType() == FalseVal->getType() &&
4236  "Select must have same types for true/false ops");
4237  if (Cond->getType() == TrueVal->getType() &&
4239  return Cond;
4240 
4241  // select ?, X, X -> X
4242  if (TrueVal == FalseVal)
4243  return TrueVal;
4244 
4245  // If the true or false value is undef, we can fold to the other value as
4246  // long as the other value isn't poison.
4247  // select ?, undef, X -> X
4248  if (Q.isUndefValue(TrueVal) &&
4250  return FalseVal;
4251  // select ?, X, undef -> X
4252  if (Q.isUndefValue(FalseVal) &&
4254  return TrueVal;
4255 
4256  // Deal with partial undef vector constants: select ?, VecC, VecC' --> VecC''
4257  Constant *TrueC, *FalseC;
4258  if (isa<FixedVectorType>(TrueVal->getType()) &&
4259  match(TrueVal, m_Constant(TrueC)) &&
4260  match(FalseVal, m_Constant(FalseC))) {
4261  unsigned NumElts =
4262  cast<FixedVectorType>(TrueC->getType())->getNumElements();
4264  for (unsigned i = 0; i != NumElts; ++i) {
4265  // Bail out on incomplete vector constants.
4266  Constant *TEltC = TrueC->getAggregateElement(i);
4267  Constant *FEltC = FalseC->getAggregateElement(i);
4268  if (!TEltC || !FEltC)
4269  break;
4270 
4271  // If the elements match (undef or not), that value is the result. If only
4272  // one element is undef, choose the defined element as the safe result.
4273  if (TEltC == FEltC)
4274  NewC.push_back(TEltC);
4275  else if (Q.isUndefValue(TEltC) &&
4277  NewC.push_back(FEltC);
4278  else if (Q.isUndefValue(FEltC) &&
4280  NewC.push_back(TEltC);
4281  else
4282  break;
4283  }
4284  if (NewC.size() == NumElts)
4285  return ConstantVector::get(NewC);
4286  }
4287 
4288  if (Value *V =
4289  simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
4290  return V;
4291 
4293  return V;
4294 
4296  return V;
4297 
4299  if (Imp)
4300  return *Imp ? TrueVal : FalseVal;
4301 
4302  return nullptr;
4303 }
4304 
4306  const SimplifyQuery &Q) {
4308 }
4309 
4310 /// Given operands for an GetElementPtrInst, see if we can fold the result.
4311 /// If not, this returns null.
4313  const SimplifyQuery &Q, unsigned) {
4314  // The type of the GEP pointer operand.
4315  unsigned AS =
4316  cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
4317 
4318  // getelementptr P -> P.
4319  if (Ops.size() == 1)
4320  return Ops[0];
4321 
4322  // Compute the (pointer) type returned by the GEP instruction.
4323  Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
4324  Type *GEPTy = PointerType::get(LastType, AS);
4325  for (Value *Op : Ops) {
4326  // If one of the operands is a vector, the result type is a vector of
4327  // pointers. All vector operands must have the same number of elements.
4328  if (VectorType *VT = dyn_cast<VectorType>(Op->getType())) {
4329  GEPTy = VectorType::get(GEPTy, VT->getElementCount());
4330  break;
4331  }
4332  }
4333 
4334  // getelementptr poison, idx -> poison
4335  // getelementptr baseptr, poison -> poison
4336  if (any_of(Ops, [](const auto *V) { return isa<PoisonValue>(V); }))
4337  return PoisonValue::get(GEPTy);
4338 
4339  if (Q.isUndefValue(Ops[0]))
4340  return UndefValue::get(GEPTy);
4341 
4342  bool IsScalableVec =
4343  isa<ScalableVectorType>(SrcTy) || any_of(Ops, [](const Value *V) {
4344  return isa<ScalableVectorType>(V->getType());
4345  });
4346 
4347  if (Ops.size() == 2) {
4348  // getelementptr P, 0 -> P.
4349  if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
4350  return Ops[0];
4351 
4352  Type *Ty = SrcTy;
4353  if (!IsScalableVec && Ty->isSized()) {
4354  Value *P;
4355  uint64_t C;
4356  uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
4357  // getelementptr P, N -> P if P points to a type of zero size.
4358  if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
4359  return Ops[0];
4360 
4361  // The following transforms are only safe if the ptrtoint cast
4362  // doesn't truncate the pointers.
4363  if (Ops[1]->getType()->getScalarSizeInBits() ==
4364  Q.DL.getPointerSizeInBits(AS)) {
4365  auto CanSimplify = [GEPTy, &P, V = Ops[0]]() -> bool {
4366  return P->getType() == GEPTy &&
4368  };
4369  // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
4370  if (TyAllocSize == 1 &&
4371  match(Ops[1], m_Sub(m_PtrToInt(m_Value(P)),
4372  m_PtrToInt(m_Specific(Ops[0])))) &&
4373  CanSimplify())
4374  return P;
4375 
4376  // getelementptr V, (ashr (sub P, V), C) -> P if P points to a type of
4377  // size 1 << C.
4378  if (match(Ops[1], m_AShr(m_Sub(m_PtrToInt(m_Value(P)),
4379  m_PtrToInt(m_Specific(Ops[0]))),
4380  m_ConstantInt(C))) &&
4381  TyAllocSize == 1ULL << C && CanSimplify())
4382  return P;
4383 
4384  // getelementptr V, (sdiv (sub P, V), C) -> P if P points to a type of
4385  // size C.
4386  if (match(Ops[1], m_SDiv(m_Sub(m_PtrToInt(m_Value(P)),
4387  m_PtrToInt(m_Specific(Ops[0]))),
4388  m_SpecificInt(TyAllocSize))) &&
4389  CanSimplify())
4390  return P;
4391  }
4392  }
4393  }
4394 
4395  if (!IsScalableVec && Q.DL.getTypeAllocSize(LastType) == 1 &&
4396  all_of(Ops.slice(1).drop_back(1),
4397  [](Value *Idx) { return match(Idx, m_Zero()); })) {
4398  unsigned IdxWidth =
4399  Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
4400  if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
4401  APInt BasePtrOffset(IdxWidth, 0);
4402  Value *StrippedBasePtr =
4404  BasePtrOffset);
4405 
4406  // Avoid creating inttoptr of zero here: While LLVMs treatment of
4407  // inttoptr is generally conservative, this particular case is folded to
4408  // a null pointer, which will have incorrect provenance.
4409 
4410  // gep (gep V, C), (sub 0, V) -> C
4411  if (match(Ops.back(),
4412  m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr)))) &&
4413  !BasePtrOffset.isNullValue()) {
4414  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
4415  return ConstantExpr::getIntToPtr(CI, GEPTy);
4416  }
4417  // gep (gep V, C), (xor V, -1) -> C-1
4418  if (match(Ops.back(),
4419  m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes())) &&
4420  !BasePtrOffset.isOneValue()) {
4421  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
4422  return ConstantExpr::getIntToPtr(CI, GEPTy);
4423  }
4424  }
4425  }
4426 
4427  // Check to see if this is constant foldable.
4428  if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
4429  return nullptr;
4430 
4431  auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
4432  Ops.slice(1));
4433  return ConstantFoldConstant(CE, Q.DL);
4434 }
4435 
4437  const SimplifyQuery &Q) {
4438  return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
4439 }
4440 
4441 /// Given operands for an InsertValueInst, see if we can fold the result.
4442 /// If not, this returns null.
4444  ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
4445  unsigned) {
4446  if (Constant *CAgg = dyn_cast<Constant>(Agg))
4447  if (Constant *CVal = dyn_cast<Constant>(Val))
4448  return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
4449 
4450  // insertvalue x, undef, n -> x
4451  if (Q.isUndefValue(Val))
4452  return Agg;
4453 
4454  // insertvalue x, (extractvalue y, n), n
4455  if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
4456  if (EV->getAggregateOperand()->getType() == Agg->getType() &&
4457  EV->getIndices() == Idxs) {
4458  // insertvalue undef, (extractvalue y, n), n -> y
4459  if (Q.isUndefValue(Agg))
4460  return EV->getAggregateOperand();
4461 
4462  // insertvalue y, (extractvalue y, n), n -> y
4463  if (Agg == EV->getAggregateOperand())
4464  return Agg;
4465  }
4466 
4467  return nullptr;
4468 }
4469 
4471  ArrayRef<unsigned> Idxs,
4472  const SimplifyQuery &Q) {
4474 }
4475 
4477  const SimplifyQuery &Q) {
4478  // Try to constant fold.
4479  auto *VecC = dyn_cast<Constant>(Vec);
4480  auto *ValC = dyn_cast<Constant>(Val);
4481  auto *IdxC = dyn_cast<Constant>(Idx);
4482  if (VecC && ValC && IdxC)
4483  return ConstantExpr::getInsertElement(VecC, ValC, IdxC);
4484 
4485  // For fixed-length vector, fold into poison if index is out of bounds.
4486  if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4487  if (isa<FixedVectorType>(Vec->getType()) &&
4488  CI->uge(cast<FixedVectorType>(Vec->getType())->getNumElements()))
4489  return PoisonValue::get(Vec->getType());
4490  }
4491 
4492  // If index is undef, it might be out of bounds (see above case)
4493  if (Q.isUndefValue(Idx))
4494  return PoisonValue::get(Vec->getType());
4495 
4496  // If the scalar is poison, or it is undef and there is no risk of
4497  // propagating poison from the vector value, simplify to the vector value.
4498  if (isa<PoisonValue>(Val) ||
4499  (Q.isUndefValue(Val) && isGuaranteedNotToBePoison(Vec)))
4500  return Vec;
4501 
4502  // If we are extracting a value from a vector, then inserting it into the same
4503  // place, that's the input vector:
4504  // insertelt Vec, (extractelt Vec, Idx), Idx --> Vec
4505  if (match(Val, m_ExtractElt(m_Specific(Vec), m_Specific(Idx))))
4506  return Vec;
4507 
4508  return nullptr;
4509 }
4510 
4511 /// Given operands for an ExtractValueInst, see if we can fold the result.
4512 /// If not, this returns null.
4514  const SimplifyQuery &, unsigned) {
4515  if (auto *CAgg = dyn_cast<Constant>(Agg))
4516  return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4517 
4518  // extractvalue x, (insertvalue y, elt, n), n -> elt
4519  unsigned NumIdxs = Idxs.size();
4520  for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4521  IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4522  ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4523  unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4524  unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4525  if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4526  Idxs.slice(0, NumCommonIdxs)) {
4527  if (NumIdxs == NumInsertValueIdxs)
4528  return IVI->getInsertedValueOperand();
4529  break;
4530  }
4531  }
4532 
4533  return nullptr;
4534 }
4535