LLVM  8.0.0svn
InstructionSimplify.cpp
Go to the documentation of this file.
1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
17 //
18 //===----------------------------------------------------------------------===//
19 
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/GlobalAlias.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Support/KnownBits.h"
41 #include <algorithm>
42 using namespace llvm;
43 using namespace llvm::PatternMatch;
44 
45 #define DEBUG_TYPE "instsimplify"
46 
47 enum { RecursionLimit = 3 };
48 
49 STATISTIC(NumExpand, "Number of expansions");
50 STATISTIC(NumReassoc, "Number of reassociations");
51 
52 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
54  unsigned);
55 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56  const SimplifyQuery &, unsigned);
57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
58  unsigned);
59 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60  const SimplifyQuery &Q, unsigned MaxRecurse);
61 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyCastInst(unsigned, Value *, Type *,
64  const SimplifyQuery &, unsigned);
66  unsigned);
67 
68 static Value *foldSelectWithBinaryOp(Value *Cond, Value *TrueVal,
69  Value *FalseVal) {
70  BinaryOperator::BinaryOps BinOpCode;
71  if (auto *BO = dyn_cast<BinaryOperator>(Cond))
72  BinOpCode = BO->getOpcode();
73  else
74  return nullptr;
75 
76  CmpInst::Predicate ExpectedPred, Pred1, Pred2;
77  if (BinOpCode == BinaryOperator::Or) {
78  ExpectedPred = ICmpInst::ICMP_NE;
79  } else if (BinOpCode == BinaryOperator::And) {
80  ExpectedPred = ICmpInst::ICMP_EQ;
81  } else
82  return nullptr;
83 
84  // %A = icmp eq %TV, %FV
85  // %B = icmp eq %X, %Y (and one of these is a select operand)
86  // %C = and %A, %B
87  // %D = select %C, %TV, %FV
88  // -->
89  // %FV
90 
91  // %A = icmp ne %TV, %FV
92  // %B = icmp ne %X, %Y (and one of these is a select operand)
93  // %C = or %A, %B
94  // %D = select %C, %TV, %FV
95  // -->
96  // %TV
97  Value *X, *Y;
98  if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
99  m_Specific(FalseVal)),
100  m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
101  Pred1 != Pred2 || Pred1 != ExpectedPred)
102  return nullptr;
103 
104  if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
105  return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
106 
107  return nullptr;
108 }
109 
110 /// For a boolean type or a vector of boolean type, return false or a vector
111 /// with every element false.
112 static Constant *getFalse(Type *Ty) {
113  return ConstantInt::getFalse(Ty);
114 }
115 
116 /// For a boolean type or a vector of boolean type, return true or a vector
117 /// with every element true.
118 static Constant *getTrue(Type *Ty) {
119  return ConstantInt::getTrue(Ty);
120 }
121 
122 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
123 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
124  Value *RHS) {
125  CmpInst *Cmp = dyn_cast<CmpInst>(V);
126  if (!Cmp)
127  return false;
128  CmpInst::Predicate CPred = Cmp->getPredicate();
129  Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
130  if (CPred == Pred && CLHS == LHS && CRHS == RHS)
131  return true;
132  return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
133  CRHS == LHS;
134 }
135 
136 /// Does the given value dominate the specified phi node?
137 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
139  if (!I)
140  // Arguments and constants dominate all instructions.
141  return true;
142 
143  // If we are processing instructions (and/or basic blocks) that have not been
144  // fully added to a function, the parent nodes may still be null. Simply
145  // return the conservative answer in these cases.
146  if (!I->getParent() || !P->getParent() || !I->getFunction())
147  return false;
148 
149  // If we have a DominatorTree then do a precise test.
150  if (DT)
151  return DT->dominates(I, P);
152 
153  // Otherwise, if the instruction is in the entry block and is not an invoke,
154  // then it obviously dominates all phi nodes.
155  if (I->getParent() == &I->getFunction()->getEntryBlock() &&
156  !isa<InvokeInst>(I))
157  return true;
158 
159  return false;
160 }
161 
162 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
163 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
164 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
165 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
166 /// Returns the simplified value, or null if no simplification was performed.
168  Instruction::BinaryOps OpcodeToExpand,
169  const SimplifyQuery &Q, unsigned MaxRecurse) {
170  // Recursion is always used, so bail out at once if we already hit the limit.
171  if (!MaxRecurse--)
172  return nullptr;
173 
174  // Check whether the expression has the form "(A op' B) op C".
175  if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
176  if (Op0->getOpcode() == OpcodeToExpand) {
177  // It does! Try turning it into "(A op C) op' (B op C)".
178  Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
179  // Do "A op C" and "B op C" both simplify?
180  if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
181  if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
182  // They do! Return "L op' R" if it simplifies or is already available.
183  // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
184  if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
185  && L == B && R == A)) {
186  ++NumExpand;
187  return LHS;
188  }
189  // Otherwise return "L op' R" if it simplifies.
190  if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
191  ++NumExpand;
192  return V;
193  }
194  }
195  }
196 
197  // Check whether the expression has the form "A op (B op' C)".
198  if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
199  if (Op1->getOpcode() == OpcodeToExpand) {
200  // It does! Try turning it into "(A op B) op' (A op C)".
201  Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
202  // Do "A op B" and "A op C" both simplify?
203  if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
204  if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
205  // They do! Return "L op' R" if it simplifies or is already available.
206  // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
207  if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
208  && L == C && R == B)) {
209  ++NumExpand;
210  return RHS;
211  }
212  // Otherwise return "L op' R" if it simplifies.
213  if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
214  ++NumExpand;
215  return V;
216  }
217  }
218  }
219 
220  return nullptr;
221 }
222 
223 /// Generic simplifications for associative binary operations.
224 /// Returns the simpler value, or null if none was found.
226  Value *LHS, Value *RHS,
227  const SimplifyQuery &Q,
228  unsigned MaxRecurse) {
229  assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
230 
231  // Recursion is always used, so bail out at once if we already hit the limit.
232  if (!MaxRecurse--)
233  return nullptr;
234 
237 
238  // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
239  if (Op0 && Op0->getOpcode() == Opcode) {
240  Value *A = Op0->getOperand(0);
241  Value *B = Op0->getOperand(1);
242  Value *C = RHS;
243 
244  // Does "B op C" simplify?
245  if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
246  // It does! Return "A op V" if it simplifies or is already available.
247  // If V equals B then "A op V" is just the LHS.
248  if (V == B) return LHS;
249  // Otherwise return "A op V" if it simplifies.
250  if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
251  ++NumReassoc;
252  return W;
253  }
254  }
255  }
256 
257  // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
258  if (Op1 && Op1->getOpcode() == Opcode) {
259  Value *A = LHS;
260  Value *B = Op1->getOperand(0);
261  Value *C = Op1->getOperand(1);
262 
263  // Does "A op B" simplify?
264  if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
265  // It does! Return "V op C" if it simplifies or is already available.
266  // If V equals B then "V op C" is just the RHS.
267  if (V == B) return RHS;
268  // Otherwise return "V op C" if it simplifies.
269  if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
270  ++NumReassoc;
271  return W;
272  }
273  }
274  }
275 
276  // The remaining transforms require commutativity as well as associativity.
277  if (!Instruction::isCommutative(Opcode))
278  return nullptr;
279 
280  // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
281  if (Op0 && Op0->getOpcode() == Opcode) {
282  Value *A = Op0->getOperand(0);
283  Value *B = Op0->getOperand(1);
284  Value *C = RHS;
285 
286  // Does "C op A" simplify?
287  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
288  // It does! Return "V op B" if it simplifies or is already available.
289  // If V equals A then "V op B" is just the LHS.
290  if (V == A) return LHS;
291  // Otherwise return "V op B" if it simplifies.
292  if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
293  ++NumReassoc;
294  return W;
295  }
296  }
297  }
298 
299  // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
300  if (Op1 && Op1->getOpcode() == Opcode) {
301  Value *A = LHS;
302  Value *B = Op1->getOperand(0);
303  Value *C = Op1->getOperand(1);
304 
305  // Does "C op A" simplify?
306  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
307  // It does! Return "B op V" if it simplifies or is already available.
308  // If V equals C then "B op V" is just the RHS.
309  if (V == C) return RHS;
310  // Otherwise return "B op V" if it simplifies.
311  if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
312  ++NumReassoc;
313  return W;
314  }
315  }
316  }
317 
318  return nullptr;
319 }
320 
321 /// In the case of a binary operation with a select instruction as an operand,
322 /// try to simplify the binop by seeing whether evaluating it on both branches
323 /// of the select results in the same value. Returns the common value if so,
324 /// otherwise returns null.
326  Value *RHS, const SimplifyQuery &Q,
327  unsigned MaxRecurse) {
328  // Recursion is always used, so bail out at once if we already hit the limit.
329  if (!MaxRecurse--)
330  return nullptr;
331 
332  SelectInst *SI;
333  if (isa<SelectInst>(LHS)) {
334  SI = cast<SelectInst>(LHS);
335  } else {
336  assert(isa<SelectInst>(RHS) && "No select instruction operand!");
337  SI = cast<SelectInst>(RHS);
338  }
339 
340  // Evaluate the BinOp on the true and false branches of the select.
341  Value *TV;
342  Value *FV;
343  if (SI == LHS) {
344  TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
345  FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
346  } else {
347  TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
348  FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
349  }
350 
351  // If they simplified to the same value, then return the common value.
352  // If they both failed to simplify then return null.
353  if (TV == FV)
354  return TV;
355 
356  // If one branch simplified to undef, return the other one.
357  if (TV && isa<UndefValue>(TV))
358  return FV;
359  if (FV && isa<UndefValue>(FV))
360  return TV;
361 
362  // If applying the operation did not change the true and false select values,
363  // then the result of the binop is the select itself.
364  if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
365  return SI;
366 
367  // If one branch simplified and the other did not, and the simplified
368  // value is equal to the unsimplified one, return the simplified value.
369  // For example, select (cond, X, X & Z) & Z -> X & Z.
370  if ((FV && !TV) || (TV && !FV)) {
371  // Check that the simplified value has the form "X op Y" where "op" is the
372  // same as the original operation.
373  Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
374  if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
375  // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
376  // We already know that "op" is the same as for the simplified value. See
377  // if the operands match too. If so, return the simplified value.
378  Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
379  Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
380  Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
381  if (Simplified->getOperand(0) == UnsimplifiedLHS &&
382  Simplified->getOperand(1) == UnsimplifiedRHS)
383  return Simplified;
384  if (Simplified->isCommutative() &&
385  Simplified->getOperand(1) == UnsimplifiedLHS &&
386  Simplified->getOperand(0) == UnsimplifiedRHS)
387  return Simplified;
388  }
389  }
390 
391  return nullptr;
392 }
393 
394 /// In the case of a comparison with a select instruction, try to simplify the
395 /// comparison by seeing whether both branches of the select result in the same
396 /// value. Returns the common value if so, otherwise returns null.
398  Value *RHS, const SimplifyQuery &Q,
399  unsigned MaxRecurse) {
400  // Recursion is always used, so bail out at once if we already hit the limit.
401  if (!MaxRecurse--)
402  return nullptr;
403 
404  // Make sure the select is on the LHS.
405  if (!isa<SelectInst>(LHS)) {
406  std::swap(LHS, RHS);
407  Pred = CmpInst::getSwappedPredicate(Pred);
408  }
409  assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
410  SelectInst *SI = cast<SelectInst>(LHS);
411  Value *Cond = SI->getCondition();
412  Value *TV = SI->getTrueValue();
413  Value *FV = SI->getFalseValue();
414 
415  // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
416  // Does "cmp TV, RHS" simplify?
417  Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
418  if (TCmp == Cond) {
419  // It not only simplified, it simplified to the select condition. Replace
420  // it with 'true'.
421  TCmp = getTrue(Cond->getType());
422  } else if (!TCmp) {
423  // It didn't simplify. However if "cmp TV, RHS" is equal to the select
424  // condition then we can replace it with 'true'. Otherwise give up.
425  if (!isSameCompare(Cond, Pred, TV, RHS))
426  return nullptr;
427  TCmp = getTrue(Cond->getType());
428  }
429 
430  // Does "cmp FV, RHS" simplify?
431  Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
432  if (FCmp == Cond) {
433  // It not only simplified, it simplified to the select condition. Replace
434  // it with 'false'.
435  FCmp = getFalse(Cond->getType());
436  } else if (!FCmp) {
437  // It didn't simplify. However if "cmp FV, RHS" is equal to the select
438  // condition then we can replace it with 'false'. Otherwise give up.
439  if (!isSameCompare(Cond, Pred, FV, RHS))
440  return nullptr;
441  FCmp = getFalse(Cond->getType());
442  }
443 
444  // If both sides simplified to the same value, then use it as the result of
445  // the original comparison.
446  if (TCmp == FCmp)
447  return TCmp;
448 
449  // The remaining cases only make sense if the select condition has the same
450  // type as the result of the comparison, so bail out if this is not so.
451  if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
452  return nullptr;
453  // If the false value simplified to false, then the result of the compare
454  // is equal to "Cond && TCmp". This also catches the case when the false
455  // value simplified to false and the true value to true, returning "Cond".
456  if (match(FCmp, m_Zero()))
457  if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
458  return V;
459  // If the true value simplified to true, then the result of the compare
460  // is equal to "Cond || FCmp".
461  if (match(TCmp, m_One()))
462  if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
463  return V;
464  // Finally, if the false value simplified to true and the true value to
465  // false, then the result of the compare is equal to "!Cond".
466  if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
467  if (Value *V =
468  SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
469  Q, MaxRecurse))
470  return V;
471 
472  return nullptr;
473 }
474 
475 /// In the case of a binary operation with an operand that is a PHI instruction,
476 /// try to simplify the binop by seeing whether evaluating it on the incoming
477 /// phi values yields the same result for every value. If so returns the common
478 /// value, otherwise returns null.
480  Value *RHS, const SimplifyQuery &Q,
481  unsigned MaxRecurse) {
482  // Recursion is always used, so bail out at once if we already hit the limit.
483  if (!MaxRecurse--)
484  return nullptr;
485 
486  PHINode *PI;
487  if (isa<PHINode>(LHS)) {
488  PI = cast<PHINode>(LHS);
489  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
490  if (!valueDominatesPHI(RHS, PI, Q.DT))
491  return nullptr;
492  } else {
493  assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
494  PI = cast<PHINode>(RHS);
495  // Bail out if LHS and the phi may be mutually interdependent due to a loop.
496  if (!valueDominatesPHI(LHS, PI, Q.DT))
497  return nullptr;
498  }
499 
500  // Evaluate the BinOp on the incoming phi values.
501  Value *CommonValue = nullptr;
502  for (Value *Incoming : PI->incoming_values()) {
503  // If the incoming value is the phi node itself, it can safely be skipped.
504  if (Incoming == PI) continue;
505  Value *V = PI == LHS ?
506  SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
507  SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
508  // If the operation failed to simplify, or simplified to a different value
509  // to previously, then give up.
510  if (!V || (CommonValue && V != CommonValue))
511  return nullptr;
512  CommonValue = V;
513  }
514 
515  return CommonValue;
516 }
517 
518 /// In the case of a comparison with a PHI instruction, try to simplify the
519 /// comparison by seeing whether comparing with all of the incoming phi values
520 /// yields the same result every time. If so returns the common result,
521 /// otherwise returns null.
523  const SimplifyQuery &Q, unsigned MaxRecurse) {
524  // Recursion is always used, so bail out at once if we already hit the limit.
525  if (!MaxRecurse--)
526  return nullptr;
527 
528  // Make sure the phi is on the LHS.
529  if (!isa<PHINode>(LHS)) {
530  std::swap(LHS, RHS);
531  Pred = CmpInst::getSwappedPredicate(Pred);
532  }
533  assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
534  PHINode *PI = cast<PHINode>(LHS);
535 
536  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
537  if (!valueDominatesPHI(RHS, PI, Q.DT))
538  return nullptr;
539 
540  // Evaluate the BinOp on the incoming phi values.
541  Value *CommonValue = nullptr;
542  for (Value *Incoming : PI->incoming_values()) {
543  // If the incoming value is the phi node itself, it can safely be skipped.
544  if (Incoming == PI) continue;
545  Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
546  // If the operation failed to simplify, or simplified to a different value
547  // to previously, then give up.
548  if (!V || (CommonValue && V != CommonValue))
549  return nullptr;
550  CommonValue = V;
551  }
552 
553  return CommonValue;
554 }
555 
557  Value *&Op0, Value *&Op1,
558  const SimplifyQuery &Q) {
559  if (auto *CLHS = dyn_cast<Constant>(Op0)) {
560  if (auto *CRHS = dyn_cast<Constant>(Op1))
561  return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
562 
563  // Canonicalize the constant to the RHS if this is a commutative operation.
564  if (Instruction::isCommutative(Opcode))
565  std::swap(Op0, Op1);
566  }
567  return nullptr;
568 }
569 
570 /// Given operands for an Add, see if we can fold the result.
571 /// If not, this returns null.
572 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
573  const SimplifyQuery &Q, unsigned MaxRecurse) {
574  if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
575  return C;
576 
577  // X + undef -> undef
578  if (match(Op1, m_Undef()))
579  return Op1;
580 
581  // X + 0 -> X
582  if (match(Op1, m_Zero()))
583  return Op0;
584 
585  // If two operands are negative, return 0.
586  if (isKnownNegation(Op0, Op1))
587  return Constant::getNullValue(Op0->getType());
588 
589  // X + (Y - X) -> Y
590  // (Y - X) + X -> Y
591  // Eg: X + -X -> 0
592  Value *Y = nullptr;
593  if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
594  match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
595  return Y;
596 
597  // X + ~X -> -1 since ~X = -X-1
598  Type *Ty = Op0->getType();
599  if (match(Op0, m_Not(m_Specific(Op1))) ||
600  match(Op1, m_Not(m_Specific(Op0))))
601  return Constant::getAllOnesValue(Ty);
602 
603  // add nsw/nuw (xor Y, signmask), signmask --> Y
604  // The no-wrapping add guarantees that the top bit will be set by the add.
605  // Therefore, the xor must be clearing the already set sign bit of Y.
606  if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
607  match(Op0, m_Xor(m_Value(Y), m_SignMask())))
608  return Y;
609 
610  // add nuw %x, -1 -> -1, because %x can only be 0.
611  if (IsNUW && match(Op1, m_AllOnes()))
612  return Op1; // Which is -1.
613 
614  /// i1 add -> xor.
615  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
616  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
617  return V;
618 
619  // Try some generic simplifications for associative operations.
620  if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
621  MaxRecurse))
622  return V;
623 
624  // Threading Add over selects and phi nodes is pointless, so don't bother.
625  // Threading over the select in "A + select(cond, B, C)" means evaluating
626  // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
627  // only if B and C are equal. If B and C are equal then (since we assume
628  // that operands have already been simplified) "select(cond, B, C)" should
629  // have been simplified to the common value of B and C already. Analysing
630  // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
631  // for threading over phi nodes.
632 
633  return nullptr;
634 }
635 
636 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
637  const SimplifyQuery &Query) {
638  return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
639 }
640 
641 /// Compute the base pointer and cumulative constant offsets for V.
642 ///
643 /// This strips all constant offsets off of V, leaving it the base pointer, and
644 /// accumulates the total constant offset applied in the returned constant. It
645 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
646 /// no constant offsets applied.
647 ///
648 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
649 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
650 /// folding.
652  bool AllowNonInbounds = false) {
654 
655  Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
656  APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
657 
658  // Even though we don't look through PHI nodes, we could be called on an
659  // instruction in an unreachable block, which may be on a cycle.
660  SmallPtrSet<Value *, 4> Visited;
661  Visited.insert(V);
662  do {
663  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
664  if ((!AllowNonInbounds && !GEP->isInBounds()) ||
665  !GEP->accumulateConstantOffset(DL, Offset))
666  break;
667  V = GEP->getPointerOperand();
668  } else if (Operator::getOpcode(V) == Instruction::BitCast) {
669  V = cast<Operator>(V)->getOperand(0);
670  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
671  if (GA->isInterposable())
672  break;
673  V = GA->getAliasee();
674  } else {
675  if (auto CS = CallSite(V))
676  if (Value *RV = CS.getReturnedArgOperand()) {
677  V = RV;
678  continue;
679  }
680  break;
681  }
682  assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
683  } while (Visited.insert(V).second);
684 
685  Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
686  if (V->getType()->isVectorTy())
688  OffsetIntPtr);
689  return OffsetIntPtr;
690 }
691 
692 /// Compute the constant difference between two pointer values.
693 /// If the difference is not a constant, returns zero.
695  Value *RHS) {
696  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
697  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
698 
699  // If LHS and RHS are not related via constant offsets to the same base
700  // value, there is nothing we can do here.
701  if (LHS != RHS)
702  return nullptr;
703 
704  // Otherwise, the difference of LHS - RHS can be computed as:
705  // LHS - RHS
706  // = (LHSOffset + Base) - (RHSOffset + Base)
707  // = LHSOffset - RHSOffset
708  return ConstantExpr::getSub(LHSOffset, RHSOffset);
709 }
710 
711 /// Given operands for a Sub, see if we can fold the result.
712 /// If not, this returns null.
713 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
714  const SimplifyQuery &Q, unsigned MaxRecurse) {
715  if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
716  return C;
717 
718  // X - undef -> undef
719  // undef - X -> undef
720  if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
721  return UndefValue::get(Op0->getType());
722 
723  // X - 0 -> X
724  if (match(Op1, m_Zero()))
725  return Op0;
726 
727  // X - X -> 0
728  if (Op0 == Op1)
729  return Constant::getNullValue(Op0->getType());
730 
731  // Is this a negation?
732  if (match(Op0, m_Zero())) {
733  // 0 - X -> 0 if the sub is NUW.
734  if (isNUW)
735  return Constant::getNullValue(Op0->getType());
736 
737  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
738  if (Known.Zero.isMaxSignedValue()) {
739  // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
740  // Op1 must be 0 because negating the minimum signed value is undefined.
741  if (isNSW)
742  return Constant::getNullValue(Op0->getType());
743 
744  // 0 - X -> X if X is 0 or the minimum signed value.
745  return Op1;
746  }
747  }
748 
749  // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
750  // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
751  Value *X = nullptr, *Y = nullptr, *Z = Op1;
752  if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
753  // See if "V === Y - Z" simplifies.
754  if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
755  // It does! Now see if "X + V" simplifies.
756  if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
757  // It does, we successfully reassociated!
758  ++NumReassoc;
759  return W;
760  }
761  // See if "V === X - Z" simplifies.
762  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
763  // It does! Now see if "Y + V" simplifies.
764  if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
765  // It does, we successfully reassociated!
766  ++NumReassoc;
767  return W;
768  }
769  }
770 
771  // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
772  // For example, X - (X + 1) -> -1
773  X = Op0;
774  if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
775  // See if "V === X - Y" simplifies.
776  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
777  // It does! Now see if "V - Z" simplifies.
778  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
779  // It does, we successfully reassociated!
780  ++NumReassoc;
781  return W;
782  }
783  // See if "V === X - Z" simplifies.
784  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
785  // It does! Now see if "V - Y" simplifies.
786  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
787  // It does, we successfully reassociated!
788  ++NumReassoc;
789  return W;
790  }
791  }
792 
793  // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
794  // For example, X - (X - Y) -> Y.
795  Z = Op0;
796  if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
797  // See if "V === Z - X" simplifies.
798  if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
799  // It does! Now see if "V + Y" simplifies.
800  if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
801  // It does, we successfully reassociated!
802  ++NumReassoc;
803  return W;
804  }
805 
806  // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
807  if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
808  match(Op1, m_Trunc(m_Value(Y))))
809  if (X->getType() == Y->getType())
810  // See if "V === X - Y" simplifies.
811  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
812  // It does! Now see if "trunc V" simplifies.
813  if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
814  Q, MaxRecurse - 1))
815  // It does, return the simplified "trunc V".
816  return W;
817 
818  // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
819  if (match(Op0, m_PtrToInt(m_Value(X))) &&
820  match(Op1, m_PtrToInt(m_Value(Y))))
821  if (Constant *Result = computePointerDifference(Q.DL, X, Y))
822  return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
823 
824  // i1 sub -> xor.
825  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
826  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
827  return V;
828 
829  // Threading Sub over selects and phi nodes is pointless, so don't bother.
830  // Threading over the select in "A - select(cond, B, C)" means evaluating
831  // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
832  // only if B and C are equal. If B and C are equal then (since we assume
833  // that operands have already been simplified) "select(cond, B, C)" should
834  // have been simplified to the common value of B and C already. Analysing
835  // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
836  // for threading over phi nodes.
837 
838  return nullptr;
839 }
840 
841 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
842  const SimplifyQuery &Q) {
843  return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
844 }
845 
846 /// Given operands for a Mul, see if we can fold the result.
847 /// If not, this returns null.
848 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
849  unsigned MaxRecurse) {
850  if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
851  return C;
852 
853  // X * undef -> 0
854  // X * 0 -> 0
855  if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
856  return Constant::getNullValue(Op0->getType());
857 
858  // X * 1 -> X
859  if (match(Op1, m_One()))
860  return Op0;
861 
862  // (X / Y) * Y -> X if the division is exact.
863  Value *X = nullptr;
864  if (Q.IIQ.UseInstrInfo &&
865  (match(Op0,
866  m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
867  match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
868  return X;
869 
870  // i1 mul -> and.
871  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
872  if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
873  return V;
874 
875  // Try some generic simplifications for associative operations.
876  if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
877  MaxRecurse))
878  return V;
879 
880  // Mul distributes over Add. Try some generic simplifications based on this.
881  if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
882  Q, MaxRecurse))
883  return V;
884 
885  // If the operation is with the result of a select instruction, check whether
886  // operating on either branch of the select always yields the same value.
887  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
888  if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
889  MaxRecurse))
890  return V;
891 
892  // If the operation is with the result of a phi instruction, check whether
893  // operating on all incoming values of the phi always yields the same value.
894  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
895  if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
896  MaxRecurse))
897  return V;
898 
899  return nullptr;
900 }
901 
904 }
905 
906 /// Check for common or similar folds of integer division or integer remainder.
907 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
908 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
909  Type *Ty = Op0->getType();
910 
911  // X / undef -> undef
912  // X % undef -> undef
913  if (match(Op1, m_Undef()))
914  return Op1;
915 
916  // X / 0 -> undef
917  // X % 0 -> undef
918  // We don't need to preserve faults!
919  if (match(Op1, m_Zero()))
920  return UndefValue::get(Ty);
921 
922  // If any element of a constant divisor vector is zero or undef, the whole op
923  // is undef.
924  auto *Op1C = dyn_cast<Constant>(Op1);
925  if (Op1C && Ty->isVectorTy()) {
926  unsigned NumElts = Ty->getVectorNumElements();
927  for (unsigned i = 0; i != NumElts; ++i) {
928  Constant *Elt = Op1C->getAggregateElement(i);
929  if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
930  return UndefValue::get(Ty);
931  }
932  }
933 
934  // undef / X -> 0
935  // undef % X -> 0
936  if (match(Op0, m_Undef()))
937  return Constant::getNullValue(Ty);
938 
939  // 0 / X -> 0
940  // 0 % X -> 0
941  if (match(Op0, m_Zero()))
942  return Constant::getNullValue(Op0->getType());
943 
944  // X / X -> 1
945  // X % X -> 0
946  if (Op0 == Op1)
947  return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
948 
949  // X / 1 -> X
950  // X % 1 -> 0
951  // If this is a boolean op (single-bit element type), we can't have
952  // division-by-zero or remainder-by-zero, so assume the divisor is 1.
953  // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
954  Value *X;
955  if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
956  (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
957  return IsDiv ? Op0 : Constant::getNullValue(Ty);
958 
959  return nullptr;
960 }
961 
962 /// Given a predicate and two operands, return true if the comparison is true.
963 /// This is a helper for div/rem simplification where we return some other value
964 /// when we can prove a relationship between the operands.
965 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
966  const SimplifyQuery &Q, unsigned MaxRecurse) {
967  Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
968  Constant *C = dyn_cast_or_null<Constant>(V);
969  return (C && C->isAllOnesValue());
970 }
971 
972 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
973 /// to simplify X % Y to X.
974 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
975  unsigned MaxRecurse, bool IsSigned) {
976  // Recursion is always used, so bail out at once if we already hit the limit.
977  if (!MaxRecurse--)
978  return false;
979 
980  if (IsSigned) {
981  // |X| / |Y| --> 0
982  //
983  // We require that 1 operand is a simple constant. That could be extended to
984  // 2 variables if we computed the sign bit for each.
985  //
986  // Make sure that a constant is not the minimum signed value because taking
987  // the abs() of that is undefined.
988  Type *Ty = X->getType();
989  const APInt *C;
990  if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
991  // Is the variable divisor magnitude always greater than the constant
992  // dividend magnitude?
993  // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
994  Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
995  Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
996  if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
997  isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
998  return true;
999  }
1000  if (match(Y, m_APInt(C))) {
1001  // Special-case: we can't take the abs() of a minimum signed value. If
1002  // that's the divisor, then all we have to do is prove that the dividend
1003  // is also not the minimum signed value.
1004  if (C->isMinSignedValue())
1005  return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1006 
1007  // Is the variable dividend magnitude always less than the constant
1008  // divisor magnitude?
1009  // |X| < |C| --> X > -abs(C) and X < abs(C)
1010  Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1011  Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1012  if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1013  isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1014  return true;
1015  }
1016  return false;
1017  }
1018 
1019  // IsSigned == false.
1020  // Is the dividend unsigned less than the divisor?
1021  return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1022 }
1023 
1024 /// These are simplifications common to SDiv and UDiv.
1026  const SimplifyQuery &Q, unsigned MaxRecurse) {
1027  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1028  return C;
1029 
1030  if (Value *V = simplifyDivRem(Op0, Op1, true))
1031  return V;
1032 
1033  bool IsSigned = Opcode == Instruction::SDiv;
1034 
1035  // (X * Y) / Y -> X if the multiplication does not overflow.
1036  Value *X;
1037  if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1038  auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1039  // If the Mul does not overflow, then we are good to go.
1040  if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1041  (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1042  return X;
1043  // If X has the form X = A / Y, then X * Y cannot overflow.
1044  if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1045  (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1046  return X;
1047  }
1048 
1049  // (X rem Y) / Y -> 0
1050  if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1051  (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1052  return Constant::getNullValue(Op0->getType());
1053 
1054  // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1055  ConstantInt *C1, *C2;
1056  if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1057  match(Op1, m_ConstantInt(C2))) {
1058  bool Overflow;
1059  (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1060  if (Overflow)
1061  return Constant::getNullValue(Op0->getType());
1062  }
1063 
1064  // If the operation is with the result of a select instruction, check whether
1065  // operating on either branch of the select always yields the same value.
1066  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1067  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1068  return V;
1069 
1070  // If the operation is with the result of a phi instruction, check whether
1071  // operating on all incoming values of the phi always yields the same value.
1072  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1073  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1074  return V;
1075 
1076  if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1077  return Constant::getNullValue(Op0->getType());
1078 
1079  return nullptr;
1080 }
1081 
1082 /// These are simplifications common to SRem and URem.
1084  const SimplifyQuery &Q, unsigned MaxRecurse) {
1085  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1086  return C;
1087 
1088  if (Value *V = simplifyDivRem(Op0, Op1, false))
1089  return V;
1090 
1091  // (X % Y) % Y -> X % Y
1092  if ((Opcode == Instruction::SRem &&
1093  match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1094  (Opcode == Instruction::URem &&
1095  match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1096  return Op0;
1097 
1098  // (X << Y) % X -> 0
1099  if (Q.IIQ.UseInstrInfo &&
1100  ((Opcode == Instruction::SRem &&
1101  match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1102  (Opcode == Instruction::URem &&
1103  match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1104  return Constant::getNullValue(Op0->getType());
1105 
1106  // If the operation is with the result of a select instruction, check whether
1107  // operating on either branch of the select always yields the same value.
1108  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1109  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1110  return V;
1111 
1112  // If the operation is with the result of a phi instruction, check whether
1113  // operating on all incoming values of the phi always yields the same value.
1114  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1115  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1116  return V;
1117 
1118  // If X / Y == 0, then X % Y == X.
1119  if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1120  return Op0;
1121 
1122  return nullptr;
1123 }
1124 
1125 /// Given operands for an SDiv, see if we can fold the result.
1126 /// If not, this returns null.
1127 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1128  unsigned MaxRecurse) {
1129  // If two operands are negated and no signed overflow, return -1.
1130  if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1131  return Constant::getAllOnesValue(Op0->getType());
1132 
1133  return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1134 }
1135 
1138 }
1139 
1140 /// Given operands for a UDiv, see if we can fold the result.
1141 /// If not, this returns null.
1142 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1143  unsigned MaxRecurse) {
1144  return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1145 }
1146 
1149 }
1150 
1151 /// Given operands for an SRem, see if we can fold the result.
1152 /// If not, this returns null.
1153 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1154  unsigned MaxRecurse) {
1155  // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1156  // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1157  Value *X;
1158  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1159  return ConstantInt::getNullValue(Op0->getType());
1160 
1161  // If the two operands are negated, return 0.
1162  if (isKnownNegation(Op0, Op1))
1163  return ConstantInt::getNullValue(Op0->getType());
1164 
1165  return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1166 }
1167 
1170 }
1171 
1172 /// Given operands for a URem, see if we can fold the result.
1173 /// If not, this returns null.
1174 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1175  unsigned MaxRecurse) {
1176  return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1177 }
1178 
1181 }
1182 
1183 /// Returns true if a shift by \c Amount always yields undef.
1184 static bool isUndefShift(Value *Amount) {
1185  Constant *C = dyn_cast<Constant>(Amount);
1186  if (!C)
1187  return false;
1188 
1189  // X shift by undef -> undef because it may shift by the bitwidth.
1190  if (isa<UndefValue>(C))
1191  return true;
1192 
1193  // Shifting by the bitwidth or more is undefined.
1194  if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1195  if (CI->getValue().getLimitedValue() >=
1196  CI->getType()->getScalarSizeInBits())
1197  return true;
1198 
1199  // If all lanes of a vector shift are undefined the whole shift is.
1200  if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1201  for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1203  return false;
1204  return true;
1205  }
1206 
1207  return false;
1208 }
1209 
1210 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1211 /// If not, this returns null.
1213  Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1214  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1215  return C;
1216 
1217  // 0 shift by X -> 0
1218  if (match(Op0, m_Zero()))
1219  return Constant::getNullValue(Op0->getType());
1220 
1221  // X shift by 0 -> X
1222  // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1223  // would be poison.
1224  Value *X;
1225  if (match(Op1, m_Zero()) ||
1226  (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1227  return Op0;
1228 
1229  // Fold undefined shifts.
1230  if (isUndefShift(Op1))
1231  return UndefValue::get(Op0->getType());
1232 
1233  // If the operation is with the result of a select instruction, check whether
1234  // operating on either branch of the select always yields the same value.
1235  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1236  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1237  return V;
1238 
1239  // If the operation is with the result of a phi instruction, check whether
1240  // operating on all incoming values of the phi always yields the same value.
1241  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1242  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1243  return V;
1244 
1245  // If any bits in the shift amount make that value greater than or equal to
1246  // the number of bits in the type, the shift is undefined.
1247  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1248  if (Known.One.getLimitedValue() >= Known.getBitWidth())
1249  return UndefValue::get(Op0->getType());
1250 
1251  // If all valid bits in the shift amount are known zero, the first operand is
1252  // unchanged.
1253  unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1254  if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1255  return Op0;
1256 
1257  return nullptr;
1258 }
1259 
1260 /// Given operands for an Shl, LShr or AShr, see if we can
1261 /// fold the result. If not, this returns null.
1263  Value *Op1, bool isExact, const SimplifyQuery &Q,
1264  unsigned MaxRecurse) {
1265  if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1266  return V;
1267 
1268  // X >> X -> 0
1269  if (Op0 == Op1)
1270  return Constant::getNullValue(Op0->getType());
1271 
1272  // undef >> X -> 0
1273  // undef >> X -> undef (if it's exact)
1274  if (match(Op0, m_Undef()))
1275  return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1276 
1277  // The low bit cannot be shifted out of an exact shift if it is set.
1278  if (isExact) {
1279  KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1280  if (Op0Known.One[0])
1281  return Op0;
1282  }
1283 
1284  return nullptr;
1285 }
1286 
1287 /// Given operands for an Shl, see if we can fold the result.
1288 /// If not, this returns null.
1289 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1290  const SimplifyQuery &Q, unsigned MaxRecurse) {
1291  if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1292  return V;
1293 
1294  // undef << X -> 0
1295  // undef << X -> undef if (if it's NSW/NUW)
1296  if (match(Op0, m_Undef()))
1297  return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1298 
1299  // (X >> A) << A -> X
1300  Value *X;
1301  if (Q.IIQ.UseInstrInfo &&
1302  match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1303  return X;
1304 
1305  // shl nuw i8 C, %x -> C iff C has sign bit set.
1306  if (isNUW && match(Op0, m_Negative()))
1307  return Op0;
1308  // NOTE: could use computeKnownBits() / LazyValueInfo,
1309  // but the cost-benefit analysis suggests it isn't worth it.
1310 
1311  return nullptr;
1312 }
1313 
1314 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1315  const SimplifyQuery &Q) {
1316  return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1317 }
1318 
1319 /// Given operands for an LShr, see if we can fold the result.
1320 /// If not, this returns null.
1321 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1322  const SimplifyQuery &Q, unsigned MaxRecurse) {
1323  if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1324  MaxRecurse))
1325  return V;
1326 
1327  // (X << A) >> A -> X
1328  Value *X;
1329  if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1330  return X;
1331 
1332  // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1333  // We can return X as we do in the above case since OR alters no bits in X.
1334  // SimplifyDemandedBits in InstCombine can do more general optimization for
1335  // bit manipulation. This pattern aims to provide opportunities for other
1336  // optimizers by supporting a simple but common case in InstSimplify.
1337  Value *Y;
1338  const APInt *ShRAmt, *ShLAmt;
1339  if (match(Op1, m_APInt(ShRAmt)) &&
1340  match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1341  *ShRAmt == *ShLAmt) {
1342  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1343  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1344  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1345  if (ShRAmt->uge(EffWidthY))
1346  return X;
1347  }
1348 
1349  return nullptr;
1350 }
1351 
1352 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1353  const SimplifyQuery &Q) {
1354  return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1355 }
1356 
1357 /// Given operands for an AShr, see if we can fold the result.
1358 /// If not, this returns null.
1359 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1360  const SimplifyQuery &Q, unsigned MaxRecurse) {
1361  if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1362  MaxRecurse))
1363  return V;
1364 
1365  // all ones >>a X -> -1
1366  // Do not return Op0 because it may contain undef elements if it's a vector.
1367  if (match(Op0, m_AllOnes()))
1368  return Constant::getAllOnesValue(Op0->getType());
1369 
1370  // (X << A) >> A -> X
1371  Value *X;
1372  if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1373  return X;
1374 
1375  // Arithmetic shifting an all-sign-bit value is a no-op.
1376  unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1377  if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1378  return Op0;
1379 
1380  return nullptr;
1381 }
1382 
1383 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1384  const SimplifyQuery &Q) {
1385  return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1386 }
1387 
1388 /// Commuted variants are assumed to be handled by calling this function again
1389 /// with the parameters swapped.
1391  ICmpInst *UnsignedICmp, bool IsAnd) {
1392  Value *X, *Y;
1393 
1394  ICmpInst::Predicate EqPred;
1395  if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1396  !ICmpInst::isEquality(EqPred))
1397  return nullptr;
1398 
1399  ICmpInst::Predicate UnsignedPred;
1400  if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1401  ICmpInst::isUnsigned(UnsignedPred))
1402  ;
1403  else if (match(UnsignedICmp,
1404  m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1405  ICmpInst::isUnsigned(UnsignedPred))
1406  UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1407  else
1408  return nullptr;
1409 
1410  // X < Y && Y != 0 --> X < Y
1411  // X < Y || Y != 0 --> Y != 0
1412  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1413  return IsAnd ? UnsignedICmp : ZeroICmp;
1414 
1415  // X >= Y || Y != 0 --> true
1416  // X >= Y || Y == 0 --> X >= Y
1417  if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1418  if (EqPred == ICmpInst::ICMP_NE)
1419  return getTrue(UnsignedICmp->getType());
1420  return UnsignedICmp;
1421  }
1422 
1423  // X < Y && Y == 0 --> false
1424  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1425  IsAnd)
1426  return getFalse(UnsignedICmp->getType());
1427 
1428  return nullptr;
1429 }
1430 
1431 /// Commuted variants are assumed to be handled by calling this function again
1432 /// with the parameters swapped.
1434  ICmpInst::Predicate Pred0, Pred1;
1435  Value *A ,*B;
1436  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1437  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1438  return nullptr;
1439 
1440  // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1441  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1442  // can eliminate Op1 from this 'and'.
1443  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1444  return Op0;
1445 
1446  // Check for any combination of predicates that are guaranteed to be disjoint.
1447  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1448  (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1449  (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1450  (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1451  return getFalse(Op0->getType());
1452 
1453  return nullptr;
1454 }
1455 
1456 /// Commuted variants are assumed to be handled by calling this function again
1457 /// with the parameters swapped.
1459  ICmpInst::Predicate Pred0, Pred1;
1460  Value *A ,*B;
1461  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1462  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1463  return nullptr;
1464 
1465  // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1466  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1467  // can eliminate Op0 from this 'or'.
1468  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1469  return Op1;
1470 
1471  // Check for any combination of predicates that cover the entire range of
1472  // possibilities.
1473  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1474  (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1475  (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1476  (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1477  return getTrue(Op0->getType());
1478 
1479  return nullptr;
1480 }
1481 
1482 /// Test if a pair of compares with a shared operand and 2 constants has an
1483 /// empty set intersection, full set union, or if one compare is a superset of
1484 /// the other.
1486  bool IsAnd) {
1487  // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1488  if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1489  return nullptr;
1490 
1491  const APInt *C0, *C1;
1492  if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1493  !match(Cmp1->getOperand(1), m_APInt(C1)))
1494  return nullptr;
1495 
1496  auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1497  auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1498 
1499  // For and-of-compares, check if the intersection is empty:
1500  // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1501  if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1502  return getFalse(Cmp0->getType());
1503 
1504  // For or-of-compares, check if the union is full:
1505  // (icmp X, C0) || (icmp X, C1) --> full set --> true
1506  if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1507  return getTrue(Cmp0->getType());
1508 
1509  // Is one range a superset of the other?
1510  // If this is and-of-compares, take the smaller set:
1511  // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1512  // If this is or-of-compares, take the larger set:
1513  // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1514  if (Range0.contains(Range1))
1515  return IsAnd ? Cmp1 : Cmp0;
1516  if (Range1.contains(Range0))
1517  return IsAnd ? Cmp0 : Cmp1;
1518 
1519  return nullptr;
1520 }
1521 
1523  bool IsAnd) {
1524  ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1525  if (!match(Cmp0->getOperand(1), m_Zero()) ||
1526  !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1527  return nullptr;
1528 
1529  if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1530  return nullptr;
1531 
1532  // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1533  Value *X = Cmp0->getOperand(0);
1534  Value *Y = Cmp1->getOperand(0);
1535 
1536  // If one of the compares is a masked version of a (not) null check, then
1537  // that compare implies the other, so we eliminate the other. Optionally, look
1538  // through a pointer-to-int cast to match a null check of a pointer type.
1539 
1540  // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1541  // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1542  // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1543  // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1544  if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1546  return Cmp1;
1547 
1548  // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1549  // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1550  // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1551  // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1552  if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1554  return Cmp0;
1555 
1556  return nullptr;
1557 }
1558 
1560  const InstrInfoQuery &IIQ) {
1561  // (icmp (add V, C0), C1) & (icmp V, C0)
1562  ICmpInst::Predicate Pred0, Pred1;
1563  const APInt *C0, *C1;
1564  Value *V;
1565  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1566  return nullptr;
1567 
1568  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1569  return nullptr;
1570 
1571  auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1572  if (AddInst->getOperand(1) != Op1->getOperand(1))
1573  return nullptr;
1574 
1575  Type *ITy = Op0->getType();
1576  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1577  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1578 
1579  const APInt Delta = *C1 - *C0;
1580  if (C0->isStrictlyPositive()) {
1581  if (Delta == 2) {
1582  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1583  return getFalse(ITy);
1584  if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1585  return getFalse(ITy);
1586  }
1587  if (Delta == 1) {
1588  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1589  return getFalse(ITy);
1590  if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1591  return getFalse(ITy);
1592  }
1593  }
1594  if (C0->getBoolValue() && isNUW) {
1595  if (Delta == 2)
1596  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1597  return getFalse(ITy);
1598  if (Delta == 1)
1599  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1600  return getFalse(ITy);
1601  }
1602 
1603  return nullptr;
1604 }
1605 
1607  const InstrInfoQuery &IIQ) {
1608  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1609  return X;
1610  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1611  return X;
1612 
1613  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1614  return X;
1615  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1616  return X;
1617 
1618  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1619  return X;
1620 
1621  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1622  return X;
1623 
1624  if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ))
1625  return X;
1626  if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, IIQ))
1627  return X;
1628 
1629  return nullptr;
1630 }
1631 
1633  const InstrInfoQuery &IIQ) {
1634  // (icmp (add V, C0), C1) | (icmp V, C0)
1635  ICmpInst::Predicate Pred0, Pred1;
1636  const APInt *C0, *C1;
1637  Value *V;
1638  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1639  return nullptr;
1640 
1641  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1642  return nullptr;
1643 
1644  auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1645  if (AddInst->getOperand(1) != Op1->getOperand(1))
1646  return nullptr;
1647 
1648  Type *ITy = Op0->getType();
1649  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1650  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1651 
1652  const APInt Delta = *C1 - *C0;
1653  if (C0->isStrictlyPositive()) {
1654  if (Delta == 2) {
1655  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1656  return getTrue(ITy);
1657  if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1658  return getTrue(ITy);
1659  }
1660  if (Delta == 1) {
1661  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1662  return getTrue(ITy);
1663  if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1664  return getTrue(ITy);
1665  }
1666  }
1667  if (C0->getBoolValue() && isNUW) {
1668  if (Delta == 2)
1669  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1670  return getTrue(ITy);
1671  if (Delta == 1)
1672  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1673  return getTrue(ITy);
1674  }
1675 
1676  return nullptr;
1677 }
1678 
1680  const InstrInfoQuery &IIQ) {
1681  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1682  return X;
1683  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1684  return X;
1685 
1686  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1687  return X;
1688  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1689  return X;
1690 
1691  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1692  return X;
1693 
1694  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1695  return X;
1696 
1697  if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ))
1698  return X;
1699  if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, IIQ))
1700  return X;
1701 
1702  return nullptr;
1703 }
1704 
1706  FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1707  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1708  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1709  if (LHS0->getType() != RHS0->getType())
1710  return nullptr;
1711 
1712  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1713  if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1714  (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1715  // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1716  // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1717  // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1718  // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1719  // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1720  // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1721  // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1722  // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1723  if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1724  (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1725  return RHS;
1726 
1727  // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1728  // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1729  // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1730  // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1731  // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1732  // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1733  // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1734  // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1735  if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1736  (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1737  return LHS;
1738  }
1739 
1740  return nullptr;
1741 }
1742 
1744  Value *Op0, Value *Op1, bool IsAnd) {
1745  // Look through casts of the 'and' operands to find compares.
1746  auto *Cast0 = dyn_cast<CastInst>(Op0);
1747  auto *Cast1 = dyn_cast<CastInst>(Op1);
1748  if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1749  Cast0->getSrcTy() == Cast1->getSrcTy()) {
1750  Op0 = Cast0->getOperand(0);
1751  Op1 = Cast1->getOperand(0);
1752  }
1753 
1754  Value *V = nullptr;
1755  auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1756  auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1757  if (ICmp0 && ICmp1)
1758  V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q.IIQ)
1759  : simplifyOrOfICmps(ICmp0, ICmp1, Q.IIQ);
1760 
1761  auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1762  auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1763  if (FCmp0 && FCmp1)
1764  V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1765 
1766  if (!V)
1767  return nullptr;
1768  if (!Cast0)
1769  return V;
1770 
1771  // If we looked through casts, we can only handle a constant simplification
1772  // because we are not allowed to create a cast instruction here.
1773  if (auto *C = dyn_cast<Constant>(V))
1774  return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1775 
1776  return nullptr;
1777 }
1778 
1779 /// Given operands for an And, see if we can fold the result.
1780 /// If not, this returns null.
1781 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1782  unsigned MaxRecurse) {
1783  if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1784  return C;
1785 
1786  // X & undef -> 0
1787  if (match(Op1, m_Undef()))
1788  return Constant::getNullValue(Op0->getType());
1789 
1790  // X & X = X
1791  if (Op0 == Op1)
1792  return Op0;
1793 
1794  // X & 0 = 0
1795  if (match(Op1, m_Zero()))
1796  return Constant::getNullValue(Op0->getType());
1797 
1798  // X & -1 = X
1799  if (match(Op1, m_AllOnes()))
1800  return Op0;
1801 
1802  // A & ~A = ~A & A = 0
1803  if (match(Op0, m_Not(m_Specific(Op1))) ||
1804  match(Op1, m_Not(m_Specific(Op0))))
1805  return Constant::getNullValue(Op0->getType());
1806 
1807  // (A | ?) & A = A
1808  if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1809  return Op1;
1810 
1811  // A & (A | ?) = A
1812  if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1813  return Op0;
1814 
1815  // A mask that only clears known zeros of a shifted value is a no-op.
1816  Value *X;
1817  const APInt *Mask;
1818  const APInt *ShAmt;
1819  if (match(Op1, m_APInt(Mask))) {
1820  // If all bits in the inverted and shifted mask are clear:
1821  // and (shl X, ShAmt), Mask --> shl X, ShAmt
1822  if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1823  (~(*Mask)).lshr(*ShAmt).isNullValue())
1824  return Op0;
1825 
1826  // If all bits in the inverted and shifted mask are clear:
1827  // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1828  if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1829  (~(*Mask)).shl(*ShAmt).isNullValue())
1830  return Op0;
1831  }
1832 
1833  // A & (-A) = A if A is a power of two or zero.
1834  if (match(Op0, m_Neg(m_Specific(Op1))) ||
1835  match(Op1, m_Neg(m_Specific(Op0)))) {
1836  if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1837  Q.DT))
1838  return Op0;
1839  if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1840  Q.DT))
1841  return Op1;
1842  }
1843 
1844  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
1845  return V;
1846 
1847  // Try some generic simplifications for associative operations.
1848  if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1849  MaxRecurse))
1850  return V;
1851 
1852  // And distributes over Or. Try some generic simplifications based on this.
1853  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1854  Q, MaxRecurse))
1855  return V;
1856 
1857  // And distributes over Xor. Try some generic simplifications based on this.
1858  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1859  Q, MaxRecurse))
1860  return V;
1861 
1862  // If the operation is with the result of a select instruction, check whether
1863  // operating on either branch of the select always yields the same value.
1864  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1865  if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1866  MaxRecurse))
1867  return V;
1868 
1869  // If the operation is with the result of a phi instruction, check whether
1870  // operating on all incoming values of the phi always yields the same value.
1871  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1872  if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1873  MaxRecurse))
1874  return V;
1875 
1876  // Assuming the effective width of Y is not larger than A, i.e. all bits
1877  // from X and Y are disjoint in (X << A) | Y,
1878  // if the mask of this AND op covers all bits of X or Y, while it covers
1879  // no bits from the other, we can bypass this AND op. E.g.,
1880  // ((X << A) | Y) & Mask -> Y,
1881  // if Mask = ((1 << effective_width_of(Y)) - 1)
1882  // ((X << A) | Y) & Mask -> X << A,
1883  // if Mask = ((1 << effective_width_of(X)) - 1) << A
1884  // SimplifyDemandedBits in InstCombine can optimize the general case.
1885  // This pattern aims to help other passes for a common case.
1886  Value *Y, *XShifted;
1887  if (match(Op1, m_APInt(Mask)) &&
1888  match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
1889  m_Value(XShifted)),
1890  m_Value(Y)))) {
1891  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1892  const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
1893  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1894  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1895  if (EffWidthY <= ShftCnt) {
1896  const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
1897  Q.DT);
1898  const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
1899  const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
1900  const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
1901  // If the mask is extracting all bits from X or Y as is, we can skip
1902  // this AND op.
1903  if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
1904  return Y;
1905  if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
1906  return XShifted;
1907  }
1908  }
1909 
1910  return nullptr;
1911 }
1912 
1915 }
1916 
1917 /// Given operands for an Or, see if we can fold the result.
1918 /// If not, this returns null.
1919 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1920  unsigned MaxRecurse) {
1921  if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1922  return C;
1923 
1924  // X | undef -> -1
1925  // X | -1 = -1
1926  // Do not return Op1 because it may contain undef elements if it's a vector.
1927  if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1928  return Constant::getAllOnesValue(Op0->getType());
1929 
1930  // X | X = X
1931  // X | 0 = X
1932  if (Op0 == Op1 || match(Op1, m_Zero()))
1933  return Op0;
1934 
1935  // A | ~A = ~A | A = -1
1936  if (match(Op0, m_Not(m_Specific(Op1))) ||
1937  match(Op1, m_Not(m_Specific(Op0))))
1938  return Constant::getAllOnesValue(Op0->getType());
1939 
1940  // (A & ?) | A = A
1941  if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1942  return Op1;
1943 
1944  // A | (A & ?) = A
1945  if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1946  return Op0;
1947 
1948  // ~(A & ?) | A = -1
1949  if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1950  return Constant::getAllOnesValue(Op1->getType());
1951 
1952  // A | ~(A & ?) = -1
1953  if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1954  return Constant::getAllOnesValue(Op0->getType());
1955 
1956  Value *A, *B;
1957  // (A & ~B) | (A ^ B) -> (A ^ B)
1958  // (~B & A) | (A ^ B) -> (A ^ B)
1959  // (A & ~B) | (B ^ A) -> (B ^ A)
1960  // (~B & A) | (B ^ A) -> (B ^ A)
1961  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1962  (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1963  match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1964  return Op1;
1965 
1966  // Commute the 'or' operands.
1967  // (A ^ B) | (A & ~B) -> (A ^ B)
1968  // (A ^ B) | (~B & A) -> (A ^ B)
1969  // (B ^ A) | (A & ~B) -> (B ^ A)
1970  // (B ^ A) | (~B & A) -> (B ^ A)
1971  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1972  (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1973  match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1974  return Op0;
1975 
1976  // (A & B) | (~A ^ B) -> (~A ^ B)
1977  // (B & A) | (~A ^ B) -> (~A ^ B)
1978  // (A & B) | (B ^ ~A) -> (B ^ ~A)
1979  // (B & A) | (B ^ ~A) -> (B ^ ~A)
1980  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1981  (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1982  match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1983  return Op1;
1984 
1985  // (~A ^ B) | (A & B) -> (~A ^ B)
1986  // (~A ^ B) | (B & A) -> (~A ^ B)
1987  // (B ^ ~A) | (A & B) -> (B ^ ~A)
1988  // (B ^ ~A) | (B & A) -> (B ^ ~A)
1989  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1990  (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1991  match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1992  return Op0;
1993 
1994  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
1995  return V;
1996 
1997  // Try some generic simplifications for associative operations.
1998  if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1999  MaxRecurse))
2000  return V;
2001 
2002  // Or distributes over And. Try some generic simplifications based on this.
2003  if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2004  MaxRecurse))
2005  return V;
2006 
2007  // If the operation is with the result of a select instruction, check whether
2008  // operating on either branch of the select always yields the same value.
2009  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2010  if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2011  MaxRecurse))
2012  return V;
2013 
2014  // (A & C1)|(B & C2)
2015  const APInt *C1, *C2;
2016  if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2017  match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2018  if (*C1 == ~*C2) {
2019  // (A & C1)|(B & C2)
2020  // If we have: ((V + N) & C1) | (V & C2)
2021  // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2022  // replace with V+N.
2023  Value *N;
2024  if (C2->isMask() && // C2 == 0+1+
2025  match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2026  // Add commutes, try both ways.
2027  if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2028  return A;
2029  }
2030  // Or commutes, try both ways.
2031  if (C1->isMask() &&
2032  match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2033  // Add commutes, try both ways.
2034  if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2035  return B;
2036  }
2037  }
2038  }
2039 
2040  // If the operation is with the result of a phi instruction, check whether
2041  // operating on all incoming values of the phi always yields the same value.
2042  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2043  if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2044  return V;
2045 
2046  return nullptr;
2047 }
2048 
2051 }
2052 
2053 /// Given operands for a Xor, see if we can fold the result.
2054 /// If not, this returns null.
2055 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2056  unsigned MaxRecurse) {
2057  if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2058  return C;
2059 
2060  // A ^ undef -> undef
2061  if (match(Op1, m_Undef()))
2062  return Op1;
2063 
2064  // A ^ 0 = A
2065  if (match(Op1, m_Zero()))
2066  return Op0;
2067 
2068  // A ^ A = 0
2069  if (Op0 == Op1)
2070  return Constant::getNullValue(Op0->getType());
2071 
2072  // A ^ ~A = ~A ^ A = -1
2073  if (match(Op0, m_Not(m_Specific(Op1))) ||
2074  match(Op1, m_Not(m_Specific(Op0))))
2075  return Constant::getAllOnesValue(Op0->getType());
2076 
2077  // Try some generic simplifications for associative operations.
2078  if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2079  MaxRecurse))
2080  return V;
2081 
2082  // Threading Xor over selects and phi nodes is pointless, so don't bother.
2083  // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2084  // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2085  // only if B and C are equal. If B and C are equal then (since we assume
2086  // that operands have already been simplified) "select(cond, B, C)" should
2087  // have been simplified to the common value of B and C already. Analysing
2088  // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2089  // for threading over phi nodes.
2090 
2091  return nullptr;
2092 }
2093 
2096 }
2097 
2098 
2100  return CmpInst::makeCmpResultType(Op->getType());
2101 }
2102 
2103 /// Rummage around inside V looking for something equivalent to the comparison
2104 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2105 /// Helper function for analyzing max/min idioms.
2107  Value *LHS, Value *RHS) {
2109  if (!SI)
2110  return nullptr;
2111  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2112  if (!Cmp)
2113  return nullptr;
2114  Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2115  if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2116  return Cmp;
2117  if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2118  LHS == CmpRHS && RHS == CmpLHS)
2119  return Cmp;
2120  return nullptr;
2121 }
2122 
2123 // A significant optimization not implemented here is assuming that alloca
2124 // addresses are not equal to incoming argument values. They don't *alias*,
2125 // as we say, but that doesn't mean they aren't equal, so we take a
2126 // conservative approach.
2127 //
2128 // This is inspired in part by C++11 5.10p1:
2129 // "Two pointers of the same type compare equal if and only if they are both
2130 // null, both point to the same function, or both represent the same
2131 // address."
2132 //
2133 // This is pretty permissive.
2134 //
2135 // It's also partly due to C11 6.5.9p6:
2136 // "Two pointers compare equal if and only if both are null pointers, both are
2137 // pointers to the same object (including a pointer to an object and a
2138 // subobject at its beginning) or function, both are pointers to one past the
2139 // last element of the same array object, or one is a pointer to one past the
2140 // end of one array object and the other is a pointer to the start of a
2141 // different array object that happens to immediately follow the first array
2142 // object in the address space.)
2143 //
2144 // C11's version is more restrictive, however there's no reason why an argument
2145 // couldn't be a one-past-the-end value for a stack object in the caller and be
2146 // equal to the beginning of a stack object in the callee.
2147 //
2148 // If the C and C++ standards are ever made sufficiently restrictive in this
2149 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2150 // this optimization.
2151 static Constant *
2153  const DominatorTree *DT, CmpInst::Predicate Pred,
2154  AssumptionCache *AC, const Instruction *CxtI,
2155  const InstrInfoQuery &IIQ, Value *LHS, Value *RHS) {
2156  // First, skip past any trivial no-ops.
2157  LHS = LHS->stripPointerCasts();
2158  RHS = RHS->stripPointerCasts();
2159 
2160  // A non-null pointer is not equal to a null pointer.
2161  if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2162  IIQ.UseInstrInfo) &&
2163  isa<ConstantPointerNull>(RHS) &&
2164  (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2165  return ConstantInt::get(GetCompareTy(LHS),
2166  !CmpInst::isTrueWhenEqual(Pred));
2167 
2168  // We can only fold certain predicates on pointer comparisons.
2169  switch (Pred) {
2170  default:
2171  return nullptr;
2172 
2173  // Equality comaprisons are easy to fold.
2174  case CmpInst::ICMP_EQ:
2175  case CmpInst::ICMP_NE:
2176  break;
2177 
2178  // We can only handle unsigned relational comparisons because 'inbounds' on
2179  // a GEP only protects against unsigned wrapping.
2180  case CmpInst::ICMP_UGT:
2181  case CmpInst::ICMP_UGE:
2182  case CmpInst::ICMP_ULT:
2183  case CmpInst::ICMP_ULE:
2184  // However, we have to switch them to their signed variants to handle
2185  // negative indices from the base pointer.
2186  Pred = ICmpInst::getSignedPredicate(Pred);
2187  break;
2188  }
2189 
2190  // Strip off any constant offsets so that we can reason about them.
2191  // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2192  // here and compare base addresses like AliasAnalysis does, however there are
2193  // numerous hazards. AliasAnalysis and its utilities rely on special rules
2194  // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2195  // doesn't need to guarantee pointer inequality when it says NoAlias.
2196  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2197  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2198 
2199  // If LHS and RHS are related via constant offsets to the same base
2200  // value, we can replace it with an icmp which just compares the offsets.
2201  if (LHS == RHS)
2202  return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2203 
2204  // Various optimizations for (in)equality comparisons.
2205  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2206  // Different non-empty allocations that exist at the same time have
2207  // different addresses (if the program can tell). Global variables always
2208  // exist, so they always exist during the lifetime of each other and all
2209  // allocas. Two different allocas usually have different addresses...
2210  //
2211  // However, if there's an @llvm.stackrestore dynamically in between two
2212  // allocas, they may have the same address. It's tempting to reduce the
2213  // scope of the problem by only looking at *static* allocas here. That would
2214  // cover the majority of allocas while significantly reducing the likelihood
2215  // of having an @llvm.stackrestore pop up in the middle. However, it's not
2216  // actually impossible for an @llvm.stackrestore to pop up in the middle of
2217  // an entry block. Also, if we have a block that's not attached to a
2218  // function, we can't tell if it's "static" under the current definition.
2219  // Theoretically, this problem could be fixed by creating a new kind of
2220  // instruction kind specifically for static allocas. Such a new instruction
2221  // could be required to be at the top of the entry block, thus preventing it
2222  // from being subject to a @llvm.stackrestore. Instcombine could even
2223  // convert regular allocas into these special allocas. It'd be nifty.
2224  // However, until then, this problem remains open.
2225  //
2226  // So, we'll assume that two non-empty allocas have different addresses
2227  // for now.
2228  //
2229  // With all that, if the offsets are within the bounds of their allocations
2230  // (and not one-past-the-end! so we can't use inbounds!), and their
2231  // allocations aren't the same, the pointers are not equal.
2232  //
2233  // Note that it's not necessary to check for LHS being a global variable
2234  // address, due to canonicalization and constant folding.
2235  if (isa<AllocaInst>(LHS) &&
2236  (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2237  ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2238  ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2239  uint64_t LHSSize, RHSSize;
2240  ObjectSizeOpts Opts;
2241  Opts.NullIsUnknownSize =
2242  NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2243  if (LHSOffsetCI && RHSOffsetCI &&
2244  getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2245  getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2246  const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2247  const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2248  if (!LHSOffsetValue.isNegative() &&
2249  !RHSOffsetValue.isNegative() &&
2250  LHSOffsetValue.ult(LHSSize) &&
2251  RHSOffsetValue.ult(RHSSize)) {
2252  return ConstantInt::get(GetCompareTy(LHS),
2253  !CmpInst::isTrueWhenEqual(Pred));
2254  }
2255  }
2256 
2257  // Repeat the above check but this time without depending on DataLayout
2258  // or being able to compute a precise size.
2259  if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2260  !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2261  LHSOffset->isNullValue() &&
2262  RHSOffset->isNullValue())
2263  return ConstantInt::get(GetCompareTy(LHS),
2264  !CmpInst::isTrueWhenEqual(Pred));
2265  }
2266 
2267  // Even if an non-inbounds GEP occurs along the path we can still optimize
2268  // equality comparisons concerning the result. We avoid walking the whole
2269  // chain again by starting where the last calls to
2270  // stripAndComputeConstantOffsets left off and accumulate the offsets.
2271  Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2272  Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2273  if (LHS == RHS)
2274  return ConstantExpr::getICmp(Pred,
2275  ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2276  ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2277 
2278  // If one side of the equality comparison must come from a noalias call
2279  // (meaning a system memory allocation function), and the other side must
2280  // come from a pointer that cannot overlap with dynamically-allocated
2281  // memory within the lifetime of the current function (allocas, byval
2282  // arguments, globals), then determine the comparison result here.
2283  SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2284  GetUnderlyingObjects(LHS, LHSUObjs, DL);
2285  GetUnderlyingObjects(RHS, RHSUObjs, DL);
2286 
2287  // Is the set of underlying objects all noalias calls?
2288  auto IsNAC = [](ArrayRef<Value *> Objects) {
2289  return all_of(Objects, isNoAliasCall);
2290  };
2291 
2292  // Is the set of underlying objects all things which must be disjoint from
2293  // noalias calls. For allocas, we consider only static ones (dynamic
2294  // allocas might be transformed into calls to malloc not simultaneously
2295  // live with the compared-to allocation). For globals, we exclude symbols
2296  // that might be resolve lazily to symbols in another dynamically-loaded
2297  // library (and, thus, could be malloc'ed by the implementation).
2298  auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2299  return all_of(Objects, [](Value *V) {
2300  if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2301  return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2302  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2303  return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2304  GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2305  !GV->isThreadLocal();
2306  if (const Argument *A = dyn_cast<Argument>(V))
2307  return A->hasByValAttr();
2308  return false;
2309  });
2310  };
2311 
2312  if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2313  (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2314  return ConstantInt::get(GetCompareTy(LHS),
2315  !CmpInst::isTrueWhenEqual(Pred));
2316 
2317  // Fold comparisons for non-escaping pointer even if the allocation call
2318  // cannot be elided. We cannot fold malloc comparison to null. Also, the
2319  // dynamic allocation call could be either of the operands.
2320  Value *MI = nullptr;
2321  if (isAllocLikeFn(LHS, TLI) &&
2322  llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2323  MI = LHS;
2324  else if (isAllocLikeFn(RHS, TLI) &&
2325  llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2326  MI = RHS;
2327  // FIXME: We should also fold the compare when the pointer escapes, but the
2328  // compare dominates the pointer escape
2329  if (MI && !PointerMayBeCaptured(MI, true, true))
2330  return ConstantInt::get(GetCompareTy(LHS),
2332  }
2333 
2334  // Otherwise, fail.
2335  return nullptr;
2336 }
2337 
2338 /// Fold an icmp when its operands have i1 scalar type.
2340  Value *RHS, const SimplifyQuery &Q) {
2341  Type *ITy = GetCompareTy(LHS); // The return type.
2342  Type *OpTy = LHS->getType(); // The operand type.
2343  if (!OpTy->isIntOrIntVectorTy(1))
2344  return nullptr;
2345 
2346  // A boolean compared to true/false can be simplified in 14 out of the 20
2347  // (10 predicates * 2 constants) possible combinations. Cases not handled here
2348  // require a 'not' of the LHS, so those must be transformed in InstCombine.
2349  if (match(RHS, m_Zero())) {
2350  switch (Pred) {
2351  case CmpInst::ICMP_NE: // X != 0 -> X
2352  case CmpInst::ICMP_UGT: // X >u 0 -> X
2353  case CmpInst::ICMP_SLT: // X <s 0 -> X
2354  return LHS;
2355 
2356  case CmpInst::ICMP_ULT: // X <u 0 -> false
2357  case CmpInst::ICMP_SGT: // X >s 0 -> false
2358  return getFalse(ITy);
2359 
2360  case CmpInst::ICMP_UGE: // X >=u 0 -> true
2361  case CmpInst::ICMP_SLE: // X <=s 0 -> true
2362  return getTrue(ITy);
2363 
2364  default: break;
2365  }
2366  } else if (match(RHS, m_One())) {
2367  switch (Pred) {
2368  case CmpInst::ICMP_EQ: // X == 1 -> X
2369  case CmpInst::ICMP_UGE: // X >=u 1 -> X
2370  case CmpInst::ICMP_SLE: // X <=s -1 -> X
2371  return LHS;
2372 
2373  case CmpInst::ICMP_UGT: // X >u 1 -> false
2374  case CmpInst::ICMP_SLT: // X <s -1 -> false
2375  return getFalse(ITy);
2376 
2377  case CmpInst::ICMP_ULE: // X <=u 1 -> true
2378  case CmpInst::ICMP_SGE: // X >=s -1 -> true
2379  return getTrue(ITy);
2380 
2381  default: break;
2382  }
2383  }
2384 
2385  switch (Pred) {
2386  default:
2387  break;
2388  case ICmpInst::ICMP_UGE:
2389  if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2390  return getTrue(ITy);
2391  break;
2392  case ICmpInst::ICMP_SGE:
2393  /// For signed comparison, the values for an i1 are 0 and -1
2394  /// respectively. This maps into a truth table of:
2395  /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2396  /// 0 | 0 | 1 (0 >= 0) | 1
2397  /// 0 | 1 | 1 (0 >= -1) | 1
2398  /// 1 | 0 | 0 (-1 >= 0) | 0
2399  /// 1 | 1 | 1 (-1 >= -1) | 1
2400  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2401  return getTrue(ITy);
2402  break;
2403  case ICmpInst::ICMP_ULE:
2404  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2405  return getTrue(ITy);
2406  break;
2407  }
2408 
2409  return nullptr;
2410 }
2411 
2412 /// Try hard to fold icmp with zero RHS because this is a common case.
2414  Value *RHS, const SimplifyQuery &Q) {
2415  if (!match(RHS, m_Zero()))
2416  return nullptr;
2417 
2418  Type *ITy = GetCompareTy(LHS); // The return type.
2419  switch (Pred) {
2420  default:
2421  llvm_unreachable("Unknown ICmp predicate!");
2422  case ICmpInst::ICMP_ULT:
2423  return getFalse(ITy);
2424  case ICmpInst::ICMP_UGE:
2425  return getTrue(ITy);
2426  case ICmpInst::ICMP_EQ:
2427  case ICmpInst::ICMP_ULE:
2428  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2429  return getFalse(ITy);
2430  break;
2431  case ICmpInst::ICMP_NE:
2432  case ICmpInst::ICMP_UGT:
2433  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2434  return getTrue(ITy);
2435  break;
2436  case ICmpInst::ICMP_SLT: {
2437  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2438  if (LHSKnown.isNegative())
2439  return getTrue(ITy);
2440  if (LHSKnown.isNonNegative())
2441  return getFalse(ITy);
2442  break;
2443  }
2444  case ICmpInst::ICMP_SLE: {
2445  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2446  if (LHSKnown.isNegative())
2447  return getTrue(ITy);
2448  if (LHSKnown.isNonNegative() &&
2449  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2450  return getFalse(ITy);
2451  break;
2452  }
2453  case ICmpInst::ICMP_SGE: {
2454  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2455  if (LHSKnown.isNegative())
2456  return getFalse(ITy);
2457  if (LHSKnown.isNonNegative())
2458  return getTrue(ITy);
2459  break;
2460  }
2461  case ICmpInst::ICMP_SGT: {
2462  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2463  if (LHSKnown.isNegative())
2464  return getFalse(ITy);
2465  if (LHSKnown.isNonNegative() &&
2466  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2467  return getTrue(ITy);
2468  break;
2469  }
2470  }
2471 
2472  return nullptr;
2473 }
2474 
2475 /// Many binary operators with a constant operand have an easy-to-compute
2476 /// range of outputs. This can be used to fold a comparison to always true or
2477 /// always false.
2479  const InstrInfoQuery &IIQ) {
2480  unsigned Width = Lower.getBitWidth();
2481  const APInt *C;
2482  switch (BO.getOpcode()) {
2483  case Instruction::Add:
2484  if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2485  // FIXME: If we have both nuw and nsw, we should reduce the range further.
2486  if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
2487  // 'add nuw x, C' produces [C, UINT_MAX].
2488  Lower = *C;
2489  } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
2490  if (C->isNegative()) {
2491  // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2492  Lower = APInt::getSignedMinValue(Width);
2493  Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2494  } else {
2495  // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2496  Lower = APInt::getSignedMinValue(Width) + *C;
2497  Upper = APInt::getSignedMaxValue(Width) + 1;
2498  }
2499  }
2500  }
2501  break;
2502 
2503  case Instruction::And:
2504  if (match(BO.getOperand(1), m_APInt(C)))
2505  // 'and x, C' produces [0, C].
2506  Upper = *C + 1;
2507  break;
2508 
2509  case Instruction::Or:
2510  if (match(BO.getOperand(1), m_APInt(C)))
2511  // 'or x, C' produces [C, UINT_MAX].
2512  Lower = *C;
2513  break;
2514 
2515  case Instruction::AShr:
2516  if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2517  // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2518  Lower = APInt::getSignedMinValue(Width).ashr(*C);
2519  Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2520  } else if (match(BO.getOperand(0), m_APInt(C))) {
2521  unsigned ShiftAmount = Width - 1;
2522  if (!C->isNullValue() && IIQ.isExact(&BO))
2523  ShiftAmount = C->countTrailingZeros();
2524  if (C->isNegative()) {
2525  // 'ashr C, x' produces [C, C >> (Width-1)]
2526  Lower = *C;
2527  Upper = C->ashr(ShiftAmount) + 1;
2528  } else {
2529  // 'ashr C, x' produces [C >> (Width-1), C]
2530  Lower = C->ashr(ShiftAmount);
2531  Upper = *C + 1;
2532  }
2533  }
2534  break;
2535 
2536  case Instruction::LShr:
2537  if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2538  // 'lshr x, C' produces [0, UINT_MAX >> C].
2539  Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2540  } else if (match(BO.getOperand(0), m_APInt(C))) {
2541  // 'lshr C, x' produces [C >> (Width-1), C].
2542  unsigned ShiftAmount = Width - 1;
2543  if (!C->isNullValue() && IIQ.isExact(&BO))
2544  ShiftAmount = C->countTrailingZeros();
2545  Lower = C->lshr(ShiftAmount);
2546  Upper = *C + 1;
2547  }
2548  break;
2549 
2550  case Instruction::Shl:
2551  if (match(BO.getOperand(0), m_APInt(C))) {
2552  if (IIQ.hasNoUnsignedWrap(&BO)) {
2553  // 'shl nuw C, x' produces [C, C << CLZ(C)]
2554  Lower = *C;
2555  Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2556  } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2557  if (C->isNegative()) {
2558  // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2559  unsigned ShiftAmount = C->countLeadingOnes() - 1;
2560  Lower = C->shl(ShiftAmount);
2561  Upper = *C + 1;
2562  } else {
2563  // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2564  unsigned ShiftAmount = C->countLeadingZeros() - 1;
2565  Lower = *C;
2566  Upper = C->shl(ShiftAmount) + 1;
2567  }
2568  }
2569  }
2570  break;
2571 
2572  case Instruction::SDiv:
2573  if (match(BO.getOperand(1), m_APInt(C))) {
2574  APInt IntMin = APInt::getSignedMinValue(Width);
2575  APInt IntMax = APInt::getSignedMaxValue(Width);
2576  if (C->isAllOnesValue()) {
2577  // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2578  // where C != -1 and C != 0 and C != 1
2579  Lower = IntMin + 1;
2580  Upper = IntMax + 1;
2581  } else if (C->countLeadingZeros() < Width - 1) {
2582  // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2583  // where C != -1 and C != 0 and C != 1
2584  Lower = IntMin.sdiv(*C);
2585  Upper = IntMax.sdiv(*C);
2586  if (Lower.sgt(Upper))
2587  std::swap(Lower, Upper);
2588  Upper = Upper + 1;
2589  assert(Upper != Lower && "Upper part of range has wrapped!");
2590  }
2591  } else if (match(BO.getOperand(0), m_APInt(C))) {
2592  if (C->isMinSignedValue()) {
2593  // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2594  Lower = *C;
2595  Upper = Lower.lshr(1) + 1;
2596  } else {
2597  // 'sdiv C, x' produces [-|C|, |C|].
2598  Upper = C->abs() + 1;
2599  Lower = (-Upper) + 1;
2600  }
2601  }
2602  break;
2603 
2604  case Instruction::UDiv:
2605  if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2606  // 'udiv x, C' produces [0, UINT_MAX / C].
2607  Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2608  } else if (match(BO.getOperand(0), m_APInt(C))) {
2609  // 'udiv C, x' produces [0, C].
2610  Upper = *C + 1;
2611  }
2612  break;
2613 
2614  case Instruction::SRem:
2615  if (match(BO.getOperand(1), m_APInt(C))) {
2616  // 'srem x, C' produces (-|C|, |C|).
2617  Upper = C->abs();
2618  Lower = (-Upper) + 1;
2619  }
2620  break;
2621 
2622  case Instruction::URem:
2623  if (match(BO.getOperand(1), m_APInt(C)))
2624  // 'urem x, C' produces [0, C).
2625  Upper = *C;
2626  break;
2627 
2628  default:
2629  break;
2630  }
2631 }
2632 
2634  Value *RHS, const InstrInfoQuery &IIQ) {
2635  Type *ITy = GetCompareTy(RHS); // The return type.
2636 
2637  Value *X;
2638  // Sign-bit checks can be optimized to true/false after unsigned
2639  // floating-point casts:
2640  // icmp slt (bitcast (uitofp X)), 0 --> false
2641  // icmp sgt (bitcast (uitofp X)), -1 --> true
2642  if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2643  if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2644  return ConstantInt::getFalse(ITy);
2645  if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2646  return ConstantInt::getTrue(ITy);
2647  }
2648 
2649  const APInt *C;
2650  if (!match(RHS, m_APInt(C)))
2651  return nullptr;
2652 
2653  // Rule out tautological comparisons (eg., ult 0 or uge 0).
2655  if (RHS_CR.isEmptySet())
2656  return ConstantInt::getFalse(ITy);
2657  if (RHS_CR.isFullSet())
2658  return ConstantInt::getTrue(ITy);
2659 
2660  // Find the range of possible values for binary operators.
2661  unsigned Width = C->getBitWidth();
2662  APInt Lower = APInt(Width, 0);
2663  APInt Upper = APInt(Width, 0);
2664  if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2665  setLimitsForBinOp(*BO, Lower, Upper, IIQ);
2666 
2667  ConstantRange LHS_CR =
2668  Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2669 
2670  if (auto *I = dyn_cast<Instruction>(LHS))
2671  if (auto *Ranges = IIQ.getMetadata(I, LLVMContext::MD_range))
2672  LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2673 
2674  if (!LHS_CR.isFullSet()) {
2675  if (RHS_CR.contains(LHS_CR))
2676  return ConstantInt::getTrue(ITy);
2677  if (RHS_CR.inverse().contains(LHS_CR))
2678  return ConstantInt::getFalse(ITy);
2679  }
2680 
2681  return nullptr;
2682 }
2683 
2684 /// TODO: A large part of this logic is duplicated in InstCombine's
2685 /// foldICmpBinOp(). We should be able to share that and avoid the code
2686 /// duplication.
2688  Value *RHS, const SimplifyQuery &Q,
2689  unsigned MaxRecurse) {
2690  Type *ITy = GetCompareTy(LHS); // The return type.
2691 
2692  BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2693  BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2694  if (MaxRecurse && (LBO || RBO)) {
2695  // Analyze the case when either LHS or RHS is an add instruction.
2696  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2697  // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2698  bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2699  if (LBO && LBO->getOpcode() == Instruction::Add) {
2700  A = LBO->getOperand(0);
2701  B = LBO->getOperand(1);
2702  NoLHSWrapProblem =
2703  ICmpInst::isEquality(Pred) ||
2704  (CmpInst::isUnsigned(Pred) &&
2705  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2706  (CmpInst::isSigned(Pred) &&
2707  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2708  }
2709  if (RBO && RBO->getOpcode() == Instruction::Add) {
2710  C = RBO->getOperand(0);
2711  D = RBO->getOperand(1);
2712  NoRHSWrapProblem =
2713  ICmpInst::isEquality(Pred) ||
2714  (CmpInst::isUnsigned(Pred) &&
2715  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2716  (CmpInst::isSigned(Pred) &&
2717  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2718  }
2719 
2720  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2721  if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2722  if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2723  Constant::getNullValue(RHS->getType()), Q,
2724  MaxRecurse - 1))
2725  return V;
2726 
2727  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2728  if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2729  if (Value *V =
2731  C == LHS ? D : C, Q, MaxRecurse - 1))
2732  return V;
2733 
2734  // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2735  if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2736  NoRHSWrapProblem) {
2737  // Determine Y and Z in the form icmp (X+Y), (X+Z).
2738  Value *Y, *Z;
2739  if (A == C) {
2740  // C + B == C + D -> B == D
2741  Y = B;
2742  Z = D;
2743  } else if (A == D) {
2744  // D + B == C + D -> B == C
2745  Y = B;
2746  Z = C;
2747  } else if (B == C) {
2748  // A + C == C + D -> A == D
2749  Y = A;
2750  Z = D;
2751  } else {
2752  assert(B == D);
2753  // A + D == C + D -> A == C
2754  Y = A;
2755  Z = C;
2756  }
2757  if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2758  return V;
2759  }
2760  }
2761 
2762  {
2763  Value *Y = nullptr;
2764  // icmp pred (or X, Y), X
2765  if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2766  if (Pred == ICmpInst::ICMP_ULT)
2767  return getFalse(ITy);
2768  if (Pred == ICmpInst::ICMP_UGE)
2769  return getTrue(ITy);
2770 
2771  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2772  KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2773  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2774  if (RHSKnown.isNonNegative() && YKnown.isNegative())
2775  return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2776  if (RHSKnown.isNegative() || YKnown.isNonNegative())
2777  return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2778  }
2779  }
2780  // icmp pred X, (or X, Y)
2781  if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2782  if (Pred == ICmpInst::ICMP_ULE)
2783  return getTrue(ITy);
2784  if (Pred == ICmpInst::ICMP_UGT)
2785  return getFalse(ITy);
2786 
2787  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2788  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2789  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2790  if (LHSKnown.isNonNegative() && YKnown.isNegative())
2791  return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2792  if (LHSKnown.isNegative() || YKnown.isNonNegative())
2793  return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2794  }
2795  }
2796  }
2797 
2798  // icmp pred (and X, Y), X
2799  if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2800  if (Pred == ICmpInst::ICMP_UGT)
2801  return getFalse(ITy);
2802  if (Pred == ICmpInst::ICMP_ULE)
2803  return getTrue(ITy);
2804  }
2805  // icmp pred X, (and X, Y)
2806  if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2807  if (Pred == ICmpInst::ICMP_UGE)
2808  return getTrue(ITy);
2809  if (Pred == ICmpInst::ICMP_ULT)
2810  return getFalse(ITy);
2811  }
2812 
2813  // 0 - (zext X) pred C
2814  if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2815  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2816  if (RHSC->getValue().isStrictlyPositive()) {
2817  if (Pred == ICmpInst::ICMP_SLT)
2818  return ConstantInt::getTrue(RHSC->getContext());
2819  if (Pred == ICmpInst::ICMP_SGE)
2820  return ConstantInt::getFalse(RHSC->getContext());
2821  if (Pred == ICmpInst::ICMP_EQ)
2822  return ConstantInt::getFalse(RHSC->getContext());
2823  if (Pred == ICmpInst::ICMP_NE)
2824  return ConstantInt::getTrue(RHSC->getContext());
2825  }
2826  if (RHSC->getValue().isNonNegative()) {
2827  if (Pred == ICmpInst::ICMP_SLE)
2828  return ConstantInt::getTrue(RHSC->getContext());
2829  if (Pred == ICmpInst::ICMP_SGT)
2830  return ConstantInt::getFalse(RHSC->getContext());
2831  }
2832  }
2833  }
2834 
2835  // icmp pred (urem X, Y), Y
2836  if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2837  switch (Pred) {
2838  default:
2839  break;
2840  case ICmpInst::ICMP_SGT:
2841  case ICmpInst::ICMP_SGE: {
2842  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2843  if (!Known.isNonNegative())
2844  break;
2846  }
2847  case ICmpInst::ICMP_EQ:
2848  case ICmpInst::ICMP_UGT:
2849  case ICmpInst::ICMP_UGE:
2850  return getFalse(ITy);
2851  case ICmpInst::ICMP_SLT:
2852  case ICmpInst::ICMP_SLE: {
2853  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2854  if (!Known.isNonNegative())
2855  break;
2857  }
2858  case ICmpInst::ICMP_NE:
2859  case ICmpInst::ICMP_ULT:
2860  case ICmpInst::ICMP_ULE:
2861  return getTrue(ITy);
2862  }
2863  }
2864 
2865  // icmp pred X, (urem Y, X)
2866  if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2867  switch (Pred) {
2868  default:
2869  break;
2870  case ICmpInst::ICMP_SGT:
2871  case ICmpInst::ICMP_SGE: {
2872  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2873  if (!Known.isNonNegative())
2874  break;
2876  }
2877  case ICmpInst::ICMP_NE:
2878  case ICmpInst::ICMP_UGT:
2879  case ICmpInst::ICMP_UGE:
2880  return getTrue(ITy);
2881  case ICmpInst::ICMP_SLT:
2882  case ICmpInst::ICMP_SLE: {
2883  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2884  if (!Known.isNonNegative())
2885  break;
2887  }
2888  case ICmpInst::ICMP_EQ:
2889  case ICmpInst::ICMP_ULT:
2890  case ICmpInst::ICMP_ULE:
2891  return getFalse(ITy);
2892  }
2893  }
2894 
2895  // x >> y <=u x
2896  // x udiv y <=u x.
2897  if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2898  match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2899  // icmp pred (X op Y), X
2900  if (Pred == ICmpInst::ICMP_UGT)
2901  return getFalse(ITy);
2902  if (Pred == ICmpInst::ICMP_ULE)
2903  return getTrue(ITy);
2904  }
2905 
2906  // x >=u x >> y
2907  // x >=u x udiv y.
2908  if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2909  match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2910  // icmp pred X, (X op Y)
2911  if (Pred == ICmpInst::ICMP_ULT)
2912  return getFalse(ITy);
2913  if (Pred == ICmpInst::ICMP_UGE)
2914  return getTrue(ITy);
2915  }
2916 
2917  // handle:
2918  // CI2 << X == CI
2919  // CI2 << X != CI
2920  //
2921  // where CI2 is a power of 2 and CI isn't
2922  if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2923  const APInt *CI2Val, *CIVal = &CI->getValue();
2924  if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2925  CI2Val->isPowerOf2()) {
2926  if (!CIVal->isPowerOf2()) {
2927  // CI2 << X can equal zero in some circumstances,
2928  // this simplification is unsafe if CI is zero.
2929  //
2930  // We know it is safe if:
2931  // - The shift is nsw, we can't shift out the one bit.
2932  // - The shift is nuw, we can't shift out the one bit.
2933  // - CI2 is one
2934  // - CI isn't zero
2935  if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2936  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2937  CI2Val->isOneValue() || !CI->isZero()) {
2938  if (Pred == ICmpInst::ICMP_EQ)
2939  return ConstantInt::getFalse(RHS->getContext());
2940  if (Pred == ICmpInst::ICMP_NE)
2941  return ConstantInt::getTrue(RHS->getContext());
2942  }
2943  }
2944  if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2945  if (Pred == ICmpInst::ICMP_UGT)
2946  return ConstantInt::getFalse(RHS->getContext());
2947  if (Pred == ICmpInst::ICMP_ULE)
2948  return ConstantInt::getTrue(RHS->getContext());
2949  }
2950  }
2951  }
2952 
2953  if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2954  LBO->getOperand(1) == RBO->getOperand(1)) {
2955  switch (LBO->getOpcode()) {
2956  default:
2957  break;
2958  case Instruction::UDiv:
2959  case Instruction::LShr:
2960  if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2961  !Q.IIQ.isExact(RBO))
2962  break;
2963  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2964  RBO->getOperand(0), Q, MaxRecurse - 1))
2965  return V;
2966  break;
2967  case Instruction::SDiv:
2968  if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2969  !Q.IIQ.isExact(RBO))
2970  break;
2971  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2972  RBO->getOperand(0), Q, MaxRecurse - 1))
2973  return V;
2974  break;
2975  case Instruction::AShr:
2976  if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2977  break;
2978  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2979  RBO->getOperand(0), Q, MaxRecurse - 1))
2980  return V;
2981  break;
2982  case Instruction::Shl: {
2983  bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
2984  bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
2985  if (!NUW && !NSW)
2986  break;
2987  if (!NSW && ICmpInst::isSigned(Pred))
2988  break;
2989  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2990  RBO->getOperand(0), Q, MaxRecurse - 1))
2991  return V;
2992  break;
2993  }
2994  }
2995  }
2996  return nullptr;
2997 }
2998 
3000  Value *Op1) {
3001  // We need a comparison with a constant.
3002  const APInt *C;
3003  if (!match(Op1, m_APInt(C)))
3004  return nullptr;
3005 
3006  // matchSelectPattern returns the negation part of an abs pattern in SP1.
3007  // If the negate has an NSW flag, abs(INT_MIN) is undefined. Without that
3008  // constraint, we can't make a contiguous range for the result of abs.
3010  Value *SP0, *SP1;
3011  SelectPatternFlavor SPF = matchSelectPattern(Op0, SP0, SP1).Flavor;
3012  if (SPF == SelectPatternFlavor::SPF_ABS &&
3013  cast<Instruction>(SP1)->hasNoSignedWrap())
3014  // The result of abs(X) is >= 0 (with nsw).
3015  AbsPred = ICmpInst::ICMP_SGE;
3016  if (SPF == SelectPatternFlavor::SPF_NABS)
3017  // The result of -abs(X) is <= 0.
3018  AbsPred = ICmpInst::ICMP_SLE;
3019 
3020  if (AbsPred == ICmpInst::BAD_ICMP_PREDICATE)
3021  return nullptr;
3022 
3023  // If there is no intersection between abs/nabs and the range of this icmp,
3024  // the icmp must be false. If the abs/nabs range is a subset of the icmp
3025  // range, the icmp must be true.
3026  APInt Zero = APInt::getNullValue(C->getBitWidth());
3027  ConstantRange AbsRange = ConstantRange::makeExactICmpRegion(AbsPred, Zero);
3029  if (AbsRange.intersectWith(CmpRange).isEmptySet())
3030  return getFalse(GetCompareTy(Op0));
3031  if (CmpRange.contains(AbsRange))
3032  return getTrue(GetCompareTy(Op0));
3033 
3034  return nullptr;
3035 }
3036 
3037 /// Simplify integer comparisons where at least one operand of the compare
3038 /// matches an integer min/max idiom.
3040  Value *RHS, const SimplifyQuery &Q,
3041  unsigned MaxRecurse) {
3042  Type *ITy = GetCompareTy(LHS); // The return type.
3043  Value *A, *B;
3045  CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3046 
3047  // Signed variants on "max(a,b)>=a -> true".
3048  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3049  if (A != RHS)
3050  std::swap(A, B); // smax(A, B) pred A.
3051  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3052  // We analyze this as smax(A, B) pred A.
3053  P = Pred;
3054  } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
3055  (A == LHS || B == LHS)) {
3056  if (A != LHS)
3057  std::swap(A, B); // A pred smax(A, B).
3058  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3059  // We analyze this as smax(A, B) swapped-pred A.
3060  P = CmpInst::getSwappedPredicate(Pred);
3061  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3062  (A == RHS || B == RHS)) {
3063  if (A != RHS)
3064  std::swap(A, B); // smin(A, B) pred A.
3065  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3066  // We analyze this as smax(-A, -B) swapped-pred -A.
3067  // Note that we do not need to actually form -A or -B thanks to EqP.
3068  P = CmpInst::getSwappedPredicate(Pred);
3069  } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
3070  (A == LHS || B == LHS)) {
3071  if (A != LHS)
3072  std::swap(A, B); // A pred smin(A, B).
3073  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3074  // We analyze this as smax(-A, -B) pred -A.
3075  // Note that we do not need to actually form -A or -B thanks to EqP.
3076  P = Pred;
3077  }
3078  if (P != CmpInst::BAD_ICMP_PREDICATE) {
3079  // Cases correspond to "max(A, B) p A".
3080  switch (P) {
3081  default:
3082  break;
3083  case CmpInst::ICMP_EQ:
3084  case CmpInst::ICMP_SLE:
3085  // Equivalent to "A EqP B". This may be the same as the condition tested
3086  // in the max/min; if so, we can just return that.
3087  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3088  return V;
3089  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3090  return V;
3091  // Otherwise, see if "A EqP B" simplifies.
3092  if (MaxRecurse)
3093  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3094  return V;
3095  break;
3096  case CmpInst::ICMP_NE:
3097  case CmpInst::ICMP_SGT: {
3099  // Equivalent to "A InvEqP B". This may be the same as the condition
3100  // tested in the max/min; if so, we can just return that.
3101  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3102  return V;
3103  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3104  return V;
3105  // Otherwise, see if "A InvEqP B" simplifies.
3106  if (MaxRecurse)
3107  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3108  return V;
3109  break;
3110  }
3111  case CmpInst::ICMP_SGE:
3112  // Always true.
3113  return getTrue(ITy);
3114  case CmpInst::ICMP_SLT:
3115  // Always false.
3116  return getFalse(ITy);
3117  }
3118  }
3119 
3120  // Unsigned variants on "max(a,b)>=a -> true".
3122  if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3123  if (A != RHS)
3124  std::swap(A, B); // umax(A, B) pred A.
3125  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3126  // We analyze this as umax(A, B) pred A.
3127  P = Pred;
3128  } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3129  (A == LHS || B == LHS)) {
3130  if (A != LHS)
3131  std::swap(A, B); // A pred umax(A, B).
3132  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3133  // We analyze this as umax(A, B) swapped-pred A.
3134  P = CmpInst::getSwappedPredicate(Pred);
3135  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3136  (A == RHS || B == RHS)) {
3137  if (A != RHS)
3138  std::swap(A, B); // umin(A, B) pred A.
3139  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3140  // We analyze this as umax(-A, -B) swapped-pred -A.
3141  // Note that we do not need to actually form -A or -B thanks to EqP.
3142  P = CmpInst::getSwappedPredicate(Pred);
3143  } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3144  (A == LHS || B == LHS)) {
3145  if (A != LHS)
3146  std::swap(A, B); // A pred umin(A, B).
3147  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3148  // We analyze this as umax(-A, -B) pred -A.
3149  // Note that we do not need to actually form -A or -B thanks to EqP.
3150  P = Pred;
3151  }
3152  if (P != CmpInst::BAD_ICMP_PREDICATE) {
3153  // Cases correspond to "max(A, B) p A".
3154  switch (P) {
3155  default:
3156  break;
3157  case CmpInst::ICMP_EQ:
3158  case CmpInst::ICMP_ULE:
3159  // Equivalent to "A EqP B". This may be the same as the condition tested
3160  // in the max/min; if so, we can just return that.
3161  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3162  return V;
3163  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3164  return V;
3165  // Otherwise, see if "A EqP B" simplifies.
3166  if (MaxRecurse)
3167  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3168  return V;
3169  break;
3170  case CmpInst::ICMP_NE:
3171  case CmpInst::ICMP_UGT: {
3173  // Equivalent to "A InvEqP B". This may be the same as the condition
3174  // tested in the max/min; if so, we can just return that.
3175  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3176  return V;
3177  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3178  return V;
3179  // Otherwise, see if "A InvEqP B" simplifies.
3180  if (MaxRecurse)
3181  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3182  return V;
3183  break;
3184  }
3185  case CmpInst::ICMP_UGE:
3186  // Always true.
3187  return getTrue(ITy);
3188  case CmpInst::ICMP_ULT:
3189  // Always false.
3190  return getFalse(ITy);
3191  }
3192  }
3193 
3194  // Variants on "max(x,y) >= min(x,z)".
3195  Value *C, *D;
3196  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3197  match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3198  (A == C || A == D || B == C || B == D)) {
3199  // max(x, ?) pred min(x, ?).
3200  if (Pred == CmpInst::ICMP_SGE)
3201  // Always true.
3202  return getTrue(ITy);
3203  if (Pred == CmpInst::ICMP_SLT)
3204  // Always false.
3205  return getFalse(ITy);
3206  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3207  match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3208  (A == C || A == D || B == C || B == D)) {
3209  // min(x, ?) pred max(x, ?).
3210  if (Pred == CmpInst::ICMP_SLE)
3211  // Always true.
3212  return getTrue(ITy);
3213  if (Pred == CmpInst::ICMP_SGT)
3214  // Always false.
3215  return getFalse(ITy);
3216  } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3217  match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3218  (A == C || A == D || B == C || B == D)) {
3219  // max(x, ?) pred min(x, ?).
3220  if (Pred == CmpInst::ICMP_UGE)
3221  // Always true.
3222  return getTrue(ITy);
3223  if (Pred == CmpInst::ICMP_ULT)
3224  // Always false.
3225  return getFalse(ITy);
3226  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3227  match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3228  (A == C || A == D || B == C || B == D)) {
3229  // min(x, ?) pred max(x, ?).
3230  if (Pred == CmpInst::ICMP_ULE)
3231  // Always true.
3232  return getTrue(ITy);
3233  if (Pred == CmpInst::ICMP_UGT)
3234  // Always false.
3235  return getFalse(ITy);
3236  }
3237 
3238  return nullptr;
3239 }
3240 
3241 /// Given operands for an ICmpInst, see if we can fold the result.
3242 /// If not, this returns null.
3243 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3244  const SimplifyQuery &Q, unsigned MaxRecurse) {
3245  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3246  assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3247 
3248  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3249  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3250  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3251 
3252  // If we have a constant, make sure it is on the RHS.
3253  std::swap(LHS, RHS);
3254  Pred = CmpInst::getSwappedPredicate(Pred);
3255  }
3256 
3257  Type *ITy = GetCompareTy(LHS); // The return type.
3258 
3259  // icmp X, X -> true/false
3260  // icmp X, undef -> true/false because undef could be X.
3261  if (LHS == RHS || isa<UndefValue>(RHS))
3262  return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3263 
3264  if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3265  return V;
3266 
3267  if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3268  return V;
3269 
3270  if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3271  return V;
3272 
3273  // If both operands have range metadata, use the metadata
3274  // to simplify the comparison.
3275  if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3276  auto RHS_Instr = cast<Instruction>(RHS);
3277  auto LHS_Instr = cast<Instruction>(LHS);
3278 
3279  if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3280  Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3281  auto RHS_CR = getConstantRangeFromMetadata(
3282  *RHS_Instr->getMetadata(LLVMContext::MD_range));
3283  auto LHS_CR = getConstantRangeFromMetadata(
3284  *LHS_Instr->getMetadata(LLVMContext::MD_range));
3285 
3286  auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3287  if (Satisfied_CR.contains(LHS_CR))
3288  return ConstantInt::getTrue(RHS->getContext());
3289 
3290  auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3291  CmpInst::getInversePredicate(Pred), RHS_CR);
3292  if (InversedSatisfied_CR.contains(LHS_CR))
3293  return ConstantInt::getFalse(RHS->getContext());
3294  }
3295  }
3296 
3297  // Compare of cast, for example (zext X) != 0 -> X != 0
3298  if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3299  Instruction *LI = cast<CastInst>(LHS);
3300  Value *SrcOp = LI->getOperand(0);
3301  Type *SrcTy = SrcOp->getType();
3302  Type *DstTy = LI->getType();
3303 
3304  // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3305  // if the integer type is the same size as the pointer type.
3306  if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3307  Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3308  if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3309  // Transfer the cast to the constant.
3310  if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3311  ConstantExpr::getIntToPtr(RHSC, SrcTy),
3312  Q, MaxRecurse-1))
3313  return V;
3314  } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3315  if (RI->getOperand(0)->getType() == SrcTy)
3316  // Compare without the cast.
3317  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3318  Q, MaxRecurse-1))
3319  return V;
3320  }
3321  }
3322 
3323  if (isa<ZExtInst>(LHS)) {
3324  // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3325  // same type.
3326  if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3327  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3328  // Compare X and Y. Note that signed predicates become unsigned.
3330  SrcOp, RI->getOperand(0), Q,
3331  MaxRecurse-1))
3332  return V;
3333  }
3334  // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3335  // too. If not, then try to deduce the result of the comparison.
3336  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3337  // Compute the constant that would happen if we truncated to SrcTy then
3338  // reextended to DstTy.
3339  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3340  Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3341 
3342  // If the re-extended constant didn't change then this is effectively
3343  // also a case of comparing two zero-extended values.
3344  if (RExt == CI && MaxRecurse)
3346  SrcOp, Trunc, Q, MaxRecurse-1))
3347  return V;
3348 
3349  // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3350  // there. Use this to work out the result of the comparison.
3351  if (RExt != CI) {
3352  switch (Pred) {
3353  default: llvm_unreachable("Unknown ICmp predicate!");
3354  // LHS <u RHS.
3355  case ICmpInst::ICMP_EQ:
3356  case ICmpInst::ICMP_UGT:
3357  case ICmpInst::ICMP_UGE:
3358  return ConstantInt::getFalse(CI->getContext());
3359 
3360  case ICmpInst::ICMP_NE:
3361  case ICmpInst::ICMP_ULT:
3362  case ICmpInst::ICMP_ULE:
3363  return ConstantInt::getTrue(CI->getContext());
3364 
3365  // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3366  // is non-negative then LHS <s RHS.
3367  case ICmpInst::ICMP_SGT:
3368  case ICmpInst::ICMP_SGE:
3369  return CI->getValue().isNegative() ?
3370  ConstantInt::getTrue(CI->getContext()) :
3371  ConstantInt::getFalse(CI->getContext());
3372 
3373  case ICmpInst::ICMP_SLT:
3374  case ICmpInst::ICMP_SLE:
3375  return CI->getValue().isNegative() ?
3376  ConstantInt::getFalse(CI->getContext()) :
3377  ConstantInt::getTrue(CI->getContext());
3378  }
3379  }
3380  }
3381  }
3382 
3383  if (isa<SExtInst>(LHS)) {
3384  // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3385  // same type.
3386  if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3387  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3388  // Compare X and Y. Note that the predicate does not change.
3389  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3390  Q, MaxRecurse-1))
3391  return V;
3392  }
3393  // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3394  // too. If not, then try to deduce the result of the comparison.
3395  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3396  // Compute the constant that would happen if we truncated to SrcTy then
3397  // reextended to DstTy.
3398  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3399  Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3400 
3401  // If the re-extended constant didn't change then this is effectively
3402  // also a case of comparing two sign-extended values.
3403  if (RExt == CI && MaxRecurse)
3404  if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3405  return V;
3406 
3407  // Otherwise the upper bits of LHS are all equal, while RHS has varying
3408  // bits there. Use this to work out the result of the comparison.
3409  if (RExt != CI) {
3410  switch (Pred) {
3411  default: llvm_unreachable("Unknown ICmp predicate!");
3412  case ICmpInst::ICMP_EQ:
3413  return ConstantInt::getFalse(CI->getContext());
3414  case ICmpInst::ICMP_NE:
3415  return ConstantInt::getTrue(CI->getContext());
3416 
3417  // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3418  // LHS >s RHS.
3419  case ICmpInst::ICMP_SGT:
3420  case ICmpInst::ICMP_SGE:
3421  return CI->getValue().isNegative() ?
3422  ConstantInt::getTrue(CI->getContext()) :
3423  ConstantInt::getFalse(CI->getContext());
3424  case ICmpInst::ICMP_SLT:
3425  case ICmpInst::ICMP_SLE:
3426  return CI->getValue().isNegative() ?
3427  ConstantInt::getFalse(CI->getContext()) :
3428  ConstantInt::getTrue(CI->getContext());
3429 
3430  // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3431  // LHS >u RHS.
3432  case ICmpInst::ICMP_UGT:
3433  case ICmpInst::ICMP_UGE:
3434  // Comparison is true iff the LHS <s 0.
3435  if (MaxRecurse)
3436  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3437  Constant::getNullValue(SrcTy),
3438  Q, MaxRecurse-1))
3439  return V;
3440  break;
3441  case ICmpInst::ICMP_ULT:
3442  case ICmpInst::ICMP_ULE:
3443  // Comparison is true iff the LHS >=s 0.
3444  if (MaxRecurse)
3445  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3446  Constant::getNullValue(SrcTy),
3447  Q, MaxRecurse-1))
3448  return V;
3449  break;
3450  }
3451  }
3452  }
3453  }
3454  }
3455 
3456  // icmp eq|ne X, Y -> false|true if X != Y
3457  if (ICmpInst::isEquality(Pred) &&
3458  isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3459  return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3460  }
3461 
3462  if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3463  return V;
3464 
3465  if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3466  return V;
3467 
3468  if (Value *V = simplifyICmpWithAbsNabs(Pred, LHS, RHS))
3469  return V;
3470 
3471  // Simplify comparisons of related pointers using a powerful, recursive
3472  // GEP-walk when we have target data available..
3473  if (LHS->getType()->isPointerTy())
3474  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3475  Q.IIQ, LHS, RHS))
3476  return C;
3477  if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3478  if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3479  if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3480  Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3481  Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3482  Q.DL.getTypeSizeInBits(CRHS->getType()))
3483  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3484  Q.IIQ, CLHS->getPointerOperand(),
3485  CRHS->getPointerOperand()))
3486  return C;
3487 
3488  if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3489  if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3490  if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3491  GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3492  (ICmpInst::isEquality(Pred) ||
3493  (GLHS->isInBounds() && GRHS->isInBounds() &&
3494  Pred == ICmpInst::getSignedPredicate(Pred)))) {
3495  // The bases are equal and the indices are constant. Build a constant
3496  // expression GEP with the same indices and a null base pointer to see
3497  // what constant folding can make out of it.
3498  Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3499  SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3501  GLHS->getSourceElementType(), Null, IndicesLHS);
3502 
3503  SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3505  GLHS->getSourceElementType(), Null, IndicesRHS);
3506  return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3507  }
3508  }
3509  }
3510 
3511  // If the comparison is with the result of a select instruction, check whether
3512  // comparing with either branch of the select always yields the same value.
3513  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3514  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3515  return V;
3516 
3517  // If the comparison is with the result of a phi instruction, check whether
3518  // doing the compare with each incoming phi value yields a common result.
3519  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3520  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3521  return V;
3522 
3523  return nullptr;
3524 }
3525 
3526 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3527  const SimplifyQuery &Q) {
3528  return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3529 }
3530 
3531 /// Given operands for an FCmpInst, see if we can fold the result.
3532 /// If not, this returns null.
3533 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3534  FastMathFlags FMF, const SimplifyQuery &Q,
3535  unsigned MaxRecurse) {
3536  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3537  assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3538 
3539  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3540  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3541  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3542 
3543  // If we have a constant, make sure it is on the RHS.
3544  std::swap(LHS, RHS);
3545  Pred = CmpInst::getSwappedPredicate(Pred);
3546  }
3547 
3548  // Fold trivial predicates.
3549  Type *RetTy = GetCompareTy(LHS);
3550  if (Pred == FCmpInst::FCMP_FALSE)
3551  return getFalse(RetTy);
3552  if (Pred == FCmpInst::FCMP_TRUE)
3553  return getTrue(RetTy);
3554 
3555  // Fold (un)ordered comparison if we can determine there are no NaNs.
3556  if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3557  if (FMF.noNaNs() ||
3558  (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3559  return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3560 
3561  // NaN is unordered; NaN is not ordered.
3563  "Comparison must be either ordered or unordered");
3564  if (match(RHS, m_NaN()))
3565  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3566 
3567  // fcmp pred x, undef and fcmp pred undef, x
3568  // fold to true if unordered, false if ordered
3569  if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3570  // Choosing NaN for the undef will always make unordered comparison succeed
3571  // and ordered comparison fail.
3572  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3573  }
3574 
3575  // fcmp x,x -> true/false. Not all compares are foldable.
3576  if (LHS == RHS) {
3577  if (CmpInst::isTrueWhenEqual(Pred))
3578  return getTrue(RetTy);
3579  if (CmpInst::isFalseWhenEqual(Pred))
3580  return getFalse(RetTy);
3581  }
3582 
3583  // Handle fcmp with constant RHS.
3584  const APFloat *C;
3585  if (match(RHS, m_APFloat(C))) {
3586  // Check whether the constant is an infinity.
3587  if (C->isInfinity()) {
3588  if (C->isNegative()) {
3589  switch (Pred) {
3590  case FCmpInst::FCMP_OLT:
3591  // No value is ordered and less than negative infinity.
3592  return getFalse(RetTy);
3593  case FCmpInst::FCMP_UGE:
3594  // All values are unordered with or at least negative infinity.
3595  return getTrue(RetTy);
3596  default:
3597  break;
3598  }
3599  } else {
3600  switch (Pred) {
3601  case FCmpInst::FCMP_OGT:
3602  // No value is ordered and greater than infinity.
3603  return getFalse(RetTy);
3604  case FCmpInst::FCMP_ULE:
3605  // All values are unordered with and at most infinity.
3606  return getTrue(RetTy);
3607  default:
3608  break;
3609  }
3610  }
3611  }
3612  if (C->isZero()) {
3613  switch (Pred) {
3614  case FCmpInst::FCMP_OGE:
3615  if (FMF.noNaNs() && CannotBeOrderedLessThanZero(LHS, Q.TLI))
3616  return getTrue(RetTy);
3617  break;
3618  case FCmpInst::FCMP_UGE:
3619  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3620  return getTrue(RetTy);
3621  break;
3622  case FCmpInst::FCMP_ULT:
3623  if (FMF.noNaNs() && CannotBeOrderedLessThanZero(LHS, Q.TLI))
3624  return getFalse(RetTy);
3625  break;
3626  case FCmpInst::FCMP_OLT:
3627  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3628  return getFalse(RetTy);
3629  break;
3630  default:
3631  break;
3632  }
3633  } else if (C->isNegative()) {
3634  assert(!C->isNaN() && "Unexpected NaN constant!");
3635  // TODO: We can catch more cases by using a range check rather than
3636  // relying on CannotBeOrderedLessThanZero.
3637  switch (Pred) {
3638  case FCmpInst::FCMP_UGE:
3639  case FCmpInst::FCMP_UGT:
3640  case FCmpInst::FCMP_UNE:
3641  // (X >= 0) implies (X > C) when (C < 0)
3642  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3643  return getTrue(RetTy);
3644  break;
3645  case FCmpInst::FCMP_OEQ:
3646  case FCmpInst::FCMP_OLE:
3647  case FCmpInst::FCMP_OLT:
3648  // (X >= 0) implies !(X < C) when (C < 0)
3649  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3650  return getFalse(RetTy);
3651  break;
3652  default:
3653  break;
3654  }
3655  }
3656  }
3657 
3658  // If the comparison is with the result of a select instruction, check whether
3659  // comparing with either branch of the select always yields the same value.
3660  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3661  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3662  return V;
3663 
3664  // If the comparison is with the result of a phi instruction, check whether
3665  // doing the compare with each incoming phi value yields a common result.
3666  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3667  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3668  return V;
3669 
3670  return nullptr;
3671 }
3672 
3673 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3674  FastMathFlags FMF, const SimplifyQuery &Q) {
3675  return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3676 }
3677 
3678 /// See if V simplifies when its operand Op is replaced with RepOp.
3679 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3680  const SimplifyQuery &Q,
3681  unsigned MaxRecurse) {
3682  // Trivial replacement.
3683  if (V == Op)
3684  return RepOp;
3685 
3686  // We cannot replace a constant, and shouldn't even try.
3687  if (isa<Constant>(Op))
3688  return nullptr;
3689 
3690  auto *I = dyn_cast<Instruction>(V);
3691  if (!I)
3692  return nullptr;
3693 
3694  // If this is a binary operator, try to simplify it with the replaced op.
3695  if (auto *B = dyn_cast<BinaryOperator>(I)) {
3696  // Consider:
3697  // %cmp = icmp eq i32 %x, 2147483647
3698  // %add = add nsw i32 %x, 1
3699  // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3700  //
3701  // We can't replace %sel with %add unless we strip away the flags.
3702  if (isa<OverflowingBinaryOperator>(B))
3703  if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3704  return nullptr;
3705  if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3706  return nullptr;
3707 
3708  if (MaxRecurse) {
3709  if (B->getOperand(0) == Op)
3710  return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3711  MaxRecurse - 1);
3712  if (B->getOperand(1) == Op)
3713  return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3714  MaxRecurse - 1);
3715  }
3716  }
3717 
3718  // Same for CmpInsts.
3719  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3720  if (MaxRecurse) {
3721  if (C->getOperand(0) == Op)
3722  return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3723  MaxRecurse - 1);
3724  if (C->getOperand(1) == Op)
3725  return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3726  MaxRecurse - 1);
3727  }
3728  }
3729 
3730  // Same for GEPs.
3731  if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3732  if (MaxRecurse) {
3733  SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3734  transform(GEP->operands(), NewOps.begin(),
3735  [&](Value *V) { return V == Op ? RepOp : V; });
3736  return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3737  MaxRecurse - 1);
3738  }
3739  }
3740 
3741  // TODO: We could hand off more cases to instsimplify here.
3742 
3743  // If all operands are constant after substituting Op for RepOp then we can
3744  // constant fold the instruction.
3745  if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3746  // Build a list of all constant operands.
3747  SmallVector<Constant *, 8> ConstOps;
3748  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3749  if (I->getOperand(i) == Op)
3750  ConstOps.push_back(CRepOp);
3751  else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3752  ConstOps.push_back(COp);
3753  else
3754  break;
3755  }
3756 
3757  // All operands were constants, fold it.
3758  if (ConstOps.size() == I->getNumOperands()) {
3759  if (CmpInst *C = dyn_cast<CmpInst>(I))
3760  return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3761  ConstOps[1], Q.DL, Q.TLI);
3762 
3763  if (LoadInst *LI = dyn_cast<LoadInst>(I))
3764  if (!LI->isVolatile())
3765  return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3766 
3767  return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3768  }
3769  }
3770 
3771  return nullptr;
3772 }
3773 
3774 /// Try to simplify a select instruction when its condition operand is an
3775 /// integer comparison where one operand of the compare is a constant.
3776 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3777  const APInt *Y, bool TrueWhenUnset) {
3778  const APInt *C;
3779 
3780  // (X & Y) == 0 ? X & ~Y : X --> X
3781  // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3782  if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3783  *Y == ~*C)
3784  return TrueWhenUnset ? FalseVal : TrueVal;
3785 
3786  // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3787  // (X & Y) != 0 ? X : X & ~Y --> X
3788  if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3789  *Y == ~*C)
3790  return TrueWhenUnset ? FalseVal : TrueVal;
3791 
3792  if (Y->isPowerOf2()) {
3793  // (X & Y) == 0 ? X | Y : X --> X | Y
3794  // (X & Y) != 0 ? X | Y : X --> X
3795  if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3796  *Y == *C)
3797  return TrueWhenUnset ? TrueVal : FalseVal;
3798 
3799  // (X & Y) == 0 ? X : X | Y --> X
3800  // (X & Y) != 0 ? X : X | Y --> X | Y
3801  if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3802  *Y == *C)
3803  return TrueWhenUnset ? TrueVal : FalseVal;
3804  }
3805 
3806  return nullptr;
3807 }
3808 
3809 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3810 /// eq/ne.
3812  ICmpInst::Predicate Pred,
3813  Value *TrueVal, Value *FalseVal) {
3814  Value *X;
3815  APInt Mask;
3816  if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3817  return nullptr;
3818 
3819  return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3820  Pred == ICmpInst::ICMP_EQ);
3821 }
3822 
3823 /// Try to simplify a select instruction when its condition operand is an
3824 /// integer comparison.
3825 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3826  Value *FalseVal, const SimplifyQuery &Q,
3827  unsigned MaxRecurse) {
3828  ICmpInst::Predicate Pred;
3829  Value *CmpLHS, *CmpRHS;
3830  if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3831  return nullptr;
3832 
3833  if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3834  Value *X;
3835  const APInt *Y;
3836  if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3837  if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3838  Pred == ICmpInst::ICMP_EQ))
3839  return V;
3840 
3841  // Test for zero-shift-guard-ops around funnel shifts. These are used to
3842  // avoid UB from oversized shifts in raw IR rotate patterns, but the
3843  // intrinsics do not have that problem.
3844  Value *ShAmt;
3845  auto isFsh = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(),
3846  m_Value(ShAmt)),
3847  m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X),
3848  m_Value(ShAmt)));
3849  // (ShAmt != 0) ? fshl(X, *, ShAmt) : X --> fshl(X, *, ShAmt)
3850  // (ShAmt != 0) ? fshr(*, X, ShAmt) : X --> fshr(*, X, ShAmt)
3851  // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
3852  // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
3853  if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt)
3854  return Pred == ICmpInst::ICMP_NE ? TrueVal : X;
3855 
3856  // (ShAmt == 0) ? X : fshl(X, *, ShAmt) --> fshl(X, *, ShAmt)
3857  // (ShAmt == 0) ? X : fshr(*, X, ShAmt) --> fshr(*, X, ShAmt)
3858  // (ShAmt != 0) ? X : fshl(X, *, ShAmt) --> X
3859  // (ShAmt != 0) ? X : fshr(*, X, ShAmt) --> X
3860  if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt)
3861  return Pred == ICmpInst::ICMP_EQ ? FalseVal : X;
3862  }
3863 
3864  // Check for other compares that behave like bit test.
3865  if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3866  TrueVal, FalseVal))
3867  return V;
3868 
3869  // If we have an equality comparison, then we know the value in one of the
3870  // arms of the select. See if substituting this value into the arm and
3871  // simplifying the result yields the same value as the other arm.
3872  if (Pred == ICmpInst::ICMP_EQ) {
3873  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3874  TrueVal ||
3875  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3876  TrueVal)
3877  return FalseVal;
3878  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3879  FalseVal ||
3880  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3881  FalseVal)
3882  return FalseVal;
3883  } else if (Pred == ICmpInst::ICMP_NE) {
3884  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3885  FalseVal ||
3886  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3887  FalseVal)
3888  return TrueVal;
3889  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3890  TrueVal ||
3891  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3892  TrueVal)
3893  return TrueVal;
3894  }
3895 
3896  return nullptr;
3897 }
3898 
3899 /// Try to simplify a select instruction when its condition operand is a
3900 /// floating-point comparison.
3902  FCmpInst::Predicate Pred;
3903  if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
3904  !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
3905  return nullptr;
3906 
3907  // TODO: The transform may not be valid with -0.0. An incomplete way of
3908  // testing for that possibility is to check if at least one operand is a
3909  // non-zero constant.
3910  const APFloat *C;
3911  if ((match(T, m_APFloat(C)) && C->isNonZero()) ||
3912  (match(F, m_APFloat(C)) && C->isNonZero())) {
3913  // (T == F) ? T : F --> F
3914  // (F == T) ? T : F --> F
3915  if (Pred == FCmpInst::FCMP_OEQ)
3916  return F;
3917 
3918  // (T != F) ? T : F --> T
3919  // (F != T) ? T : F --> T
3920  if (Pred == FCmpInst::FCMP_UNE)
3921  return T;
3922  }
3923 
3924  return nullptr;
3925 }
3926 
3927 /// Given operands for a SelectInst, see if we can fold the result.
3928 /// If not, this returns null.
3929 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3930  const SimplifyQuery &Q, unsigned MaxRecurse) {
3931  if (auto *CondC = dyn_cast<Constant>(Cond)) {
3932  if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3933  if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3934  return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3935 
3936  // select undef, X, Y -> X or Y
3937  if (isa<UndefValue>(CondC))
3938  return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3939 
3940  // TODO: Vector constants with undef elements don't simplify.
3941 
3942  // select true, X, Y -> X
3943  if (CondC->isAllOnesValue())
3944  return TrueVal;
3945  // select false, X, Y -> Y
3946  if (CondC->isNullValue())
3947  return FalseVal;
3948  }
3949 
3950  // select ?, X, X -> X
3951  if (TrueVal == FalseVal)
3952  return TrueVal;
3953 
3954  if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3955  return FalseVal;
3956  if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3957  return TrueVal;
3958 
3959  if (Value *V =
3960  simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3961  return V;
3962 
3963  if (Value *V = simplifySelectWithFCmp(Cond, TrueVal, FalseVal))
3964  return V;
3965 
3966  if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3967  return V;
3968 
3969  Optional<bool> Imp = isImpliedByDomCondition(Cond, Q.CxtI, Q.DL);
3970  if (Imp)
3971  return *Imp ? TrueVal : FalseVal;
3972 
3973  return nullptr;
3974 }
3975 
3976 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3977  const SimplifyQuery &Q) {
3978  return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3979 }
3980 
3981 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3982 /// If not, this returns null.
3984  const SimplifyQuery &Q, unsigned) {
3985  // The type of the GEP pointer operand.
3986  unsigned AS =
3987  cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3988 
3989  // getelementptr P -> P.
3990  if (Ops.size() == 1)
3991  return Ops[0];
3992 
3993  // Compute the (pointer) type returned by the GEP instruction.
3994  Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3995  Type *GEPTy = PointerType::get(LastType, AS);
3996  if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3997  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3998  else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3999  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
4000 
4001  if (isa<UndefValue>(Ops[0]))
4002  return UndefValue::get(GEPTy);
4003 
4004  if (Ops.size() == 2) {
4005  // getelementptr P, 0 -> P.
4006  if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
4007  return Ops[0];
4008 
4009  Type *Ty = SrcTy;
4010  if (Ty->isSized()) {
4011  Value *P;
4012  uint64_t C;
4013  uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
4014  // getelementptr P, N -> P if P points to a type of zero size.
4015  if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
4016  return Ops[0];
4017 
4018  // The following transforms are only safe if the ptrtoint cast
4019  // doesn't truncate the pointers.
4020  if (Ops[1]->getType()->getScalarSizeInBits() ==
4021  Q.DL.getIndexSizeInBits(AS)) {
4022  auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
4023  if (match(P, m_Zero()))
4024  return Constant::getNullValue(GEPTy);
4025  Value *Temp;
4026  if (match(P, m_PtrToInt(m_Value(Temp))))
4027  if (Temp->getType() == GEPTy)
4028  return Temp;
4029  return nullptr;
4030  };
4031 
4032  // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
4033  if (TyAllocSize == 1 &&
4034  match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
4035  if (Value *R = PtrToIntOrZero(P))
4036  return R;
4037 
4038  // getelementptr V, (ashr (sub P, V), C) -> Q
4039  // if P points to a type of size 1 << C.
4040  if (match(Ops[1],
4041  m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
4042  m_ConstantInt(C))) &&
4043  TyAllocSize == 1ULL << C)
4044  if (Value *R = PtrToIntOrZero(P))
4045  return R;
4046 
4047  // getelementptr V, (sdiv (sub P, V), C) -> Q
4048  // if P points to a type of size C.
4049  if (match(Ops[1],
4050  m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
4051  m_SpecificInt(TyAllocSize))))
4052  if (Value *R = PtrToIntOrZero(P))
4053  return R;
4054  }
4055  }
4056  }
4057 
4058  if (Q.DL.getTypeAllocSize(LastType) == 1 &&
4059  all_of(Ops.slice(1).drop_back(1),
4060  [](Value *Idx) { return match(Idx, m_Zero()); })) {
4061  unsigned IdxWidth =
4062  Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
4063  if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
4064  APInt BasePtrOffset(IdxWidth, 0);
4065  Value *StrippedBasePtr =
4066  Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
4067  BasePtrOffset);
4068 
4069  // gep (gep V, C), (sub 0, V) -> C
4070  if (match(Ops.back(),
4071  m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
4072  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
4073  return ConstantExpr::getIntToPtr(CI, GEPTy);
4074  }
4075  // gep (gep V, C), (xor V, -1) -> C-1
4076  if (match(Ops.back(),
4077  m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
4078  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
4079  return ConstantExpr::getIntToPtr(CI, GEPTy);
4080  }
4081  }
4082  }
4083 
4084  // Check to see if this is constant foldable.
4085  if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
4086  return nullptr;
4087 
4088  auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
4089  Ops.slice(1));
4090  if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
4091  return CEFolded;
4092  return CE;
4093 }
4094 
4096  const SimplifyQuery &Q) {
4097  return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
4098 }
4099 
4100 /// Given operands for an InsertValueInst, see if we can fold the result.
4101 /// If not, this returns null.
4103  ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
4104  unsigned) {
4105  if (Constant *CAgg = dyn_cast<Constant>(Agg))
4106  if (Constant *CVal = dyn_cast<Constant>(Val))
4107  return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
4108 
4109  // insertvalue x, undef, n -> x
4110  if (match(Val, m_Undef()))
4111  return Agg;
4112 
4113  // insertvalue x, (extractvalue y, n), n
4114  if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
4115  if (EV->getAggregateOperand()->getType() == Agg->getType() &&
4116  EV->getIndices() == Idxs) {
4117  // insertvalue undef, (extractvalue y, n), n -> y
4118  if (match(Agg, m_Undef()))
4119  return EV->getAggregateOperand();
4120 
4121  // insertvalue y, (extractvalue y, n), n -> y
4122  if (Agg == EV->getAggregateOperand())
4123  return Agg;
4124  }
4125 
4126  return nullptr;
4127 }
4128 
4130  ArrayRef<unsigned> Idxs,
4131  const SimplifyQuery &Q) {
4133 }
4134 
4136  const SimplifyQuery &Q) {
4137  // Try to constant fold.
4138  auto *VecC = dyn_cast<Constant>(Vec);
4139  auto *ValC = dyn_cast<Constant>(Val);
4140  auto *IdxC = dyn_cast<Constant>(Idx);
4141  if (VecC && ValC && IdxC)
4142  return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
4143 
4144  // Fold into undef if index is out of bounds.
4145  if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4146  uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
4147  if (CI->uge(NumElements))
4148  return UndefValue::get(Vec->getType());
4149  }
4150 
4151  // If index is undef, it might be out of bounds (see above case)
4152  if (isa<UndefValue>(Idx))
4153  return UndefValue::get(Vec->getType());
4154 
4155  return nullptr;
4156 }
4157 
4158 /// Given operands for an ExtractValueInst, see if we can fold the result.
4159 /// If not, this returns null.
4161  const SimplifyQuery &, unsigned) {
4162  if (auto *CAgg = dyn_cast<Constant>(Agg))
4163  return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4164 
4165  // extractvalue x, (insertvalue y, elt, n), n -> elt
4166  unsigned NumIdxs = Idxs.size();
4167  for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4168  IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4169  ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4170  unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4171  unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4172  if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4173  Idxs.slice(0, NumCommonIdxs)) {
4174  if (NumIdxs == NumInsertValueIdxs)
4175  return IVI->getInsertedValueOperand();
4176  break;
4177  }
4178  }
4179 
4180  return nullptr;
4181 }
4182 
4184  const SimplifyQuery &Q) {
4186 }
4187 
4188 /// Given operands for an ExtractElementInst, see if we can fold the result.
4189 /// If not, this returns null.
4191  unsigned) {
4192  if (auto *CVec = dyn_cast<Constant>(Vec)) {
4193  if (auto *CIdx = dyn_cast<Constant>(Idx))
4194  return ConstantFoldExtractElementInstruction(CVec, CIdx);
4195 
4196  // The index is not relevant if our vector is a splat.
4197  if (auto *Splat = CVec->getSplatValue())
4198  return Splat;
4199 
4200  if (isa<UndefValue>(Vec))
4201  return UndefValue::get(Vec->getType()->getVectorElementType());
4202  }
4203 
4204  // If extracting a specified index from the vector, see if we can recursively
4205  // find a previously computed scalar that was inserted into the vector.
4206  if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4207  if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4208  // definitely out of bounds, thus undefined result
4209  return UndefValue::get(Vec->getType()->getVectorElementType());
4210  if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4211  return Elt;
4212  }
4213 
4214  // An undef extract index can be arbitrarily chosen to be an out-of-range
4215  // index value, which would result in the instruction being undef.
4216  if (isa<UndefValue>(Idx))
4217  return UndefValue::get(Vec->getType()->getVectorElementType());
4218 
4219  return nullptr;
4220 }
4221 
4223  const SimplifyQuery &Q) {
4225 }
4226 
4227 /// See if we can fold the given phi. If not, returns null.
4228 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4229  // If all of the PHI's incoming values are the same then replace the PHI node
4230  // with the common value.
4231  Value *CommonValue = nullptr;
4232  bool HasUndefInput = false;
4233  for (Value *Incoming : PN->incoming_values()) {
4234  // If the incoming value is the phi node itself, it can safely be skipped.
4235  if (Incoming == PN) continue;
4236  if (isa<UndefValue>(Incoming)) {
4237  // Remember that we saw an undef value, but otherwise ignore them.
4238  HasUndefInput = true;
4239  continue;
4240  }
4241  if (CommonValue && Incoming != CommonValue)
4242  return nullptr; // Not the same, bail out.
4243  CommonValue = Incoming;
4244  }
4245 
4246  // If CommonValue is null then all of the incoming values were either undef or
4247  // equal to the phi node itself.
4248  if (!CommonValue)
4249  return UndefValue::get(PN->getType());
4250 
4251  // If we have a PHI node like phi(X, undef, X), where X is defined by some
4252  // instruction, we cannot return X as the result of the PHI node unless it
4253  // dominates the PHI block.
4254  if (HasUndefInput)
4255  return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4256 
4257  return CommonValue;
4258 }
4259 
4260 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4261  Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4262  if (auto *C = dyn_cast<Constant>(Op))
4263  return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4264 
4265  if (auto *CI = dyn_cast<CastInst>(Op)) {
4266  auto *Src = CI->getOperand(0);
4267  Type *SrcTy = Src->getType();
4268  Type *MidTy = CI->getType();
4269  Type *DstTy = Ty;
4270  if (Src->getType() == Ty) {
4271  auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4272  auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4273  Type *SrcIntPtrTy =
4274  SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4275  Type *MidIntPtrTy =
4276  MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4277  Type *DstIntPtrTy =
4278  DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4279  if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4280  SrcIntPtrTy, MidIntPtrTy,
4281  DstIntPtrTy) == Instruction::BitCast)
4282  return Src;
4283  }
4284  }
4285 
4286  // bitcast x -> x
4287  if (CastOpc == Instruction::BitCast)
4288  if (Op->getType() == Ty)
4289  return Op;
4290 
4291  return nullptr;
4292 }
4293 
4294 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4295  const SimplifyQuery &Q) {
4296  return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4297 }
4298 
4299 /// For the given destination element of a shuffle, peek through shuffles to
4300 /// match a root vector source operand that contains that element in the same
4301 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4302 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4303  int MaskVal, Value *RootVec,
4304  unsigned MaxRecurse) {
4305  if (!MaxRecurse--)
4306  return nullptr;
4307 
4308  // Bail out if any mask value is undefined. That kind of shuffle may be
4309  // simplified further based on demanded bits or other folds.
4310  if (MaskVal == -1)
4311  return nullptr;
4312 
4313  // The mask value chooses which source operand we need to look at next.
4314  int InVecNumElts = Op0->getType()->getVectorNumElements();
4315  int RootElt = MaskVal;
4316  Value *SourceOp = Op0;
4317  if (MaskVal >= InVecNumElts) {
4318  RootElt = MaskVal - InVecNumElts;
4319  SourceOp = Op1;
4320  }
4321 
4322  // If the source operand is a shuffle itself, look through it to find the
4323  // matching root vector.
4324  if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4325  return foldIdentityShuffles(
4326  DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4327  SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4328  }
4329 
4330  // TODO: Look through bitcasts? What if the bitcast changes the vector element
4331  // size?
4332 
4333  // The source operand is not a shuffle. Initialize the root vector value for
4334  // this shuffle if that has not been done yet.
4335  if (!RootVec)
4336  RootVec = SourceOp;
4337 
4338  // Give up as soon as a source operand does not match the existing root value.
4339  if (RootVec != SourceOp)
4340  return nullptr;
4341 
4342  // The element must be coming from the same lane in the source vector
4343  // (although it may have crossed lanes in intermediate shuffles).
4344  if (RootElt != DestElt)
4345  return nullptr;
4346 
4347  return RootVec;
4348 }
4349 
4351  Type *RetTy, const SimplifyQuery &Q,
4352  unsigned MaxRecurse) {
4353  if (isa<UndefValue>(Mask))
4354  return UndefValue::get(RetTy);
4355 
4356  Type *InVecTy = Op0->getType();
4357  unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4358  unsigned InVecNumElts = InVecTy->getVectorNumElements();
4359 
4360  SmallVector<int, 32> Indices;
4361  ShuffleVectorInst::getShuffleMask(Mask, Indices);
4362  assert(MaskNumElts == Indices.size() &&
4363  "Size of Indices not same as number of mask elements?");
4364 
4365  // Canonicalization: If mask does not select elements from an input vector,
4366  // replace that input vector with undef.
4367  bool MaskSelects0 = false, MaskSelects1 = false;
4368  for (unsigned i = 0; i != MaskNumElts; ++i) {
4369  if (Indices[i] == -1)
4370  continue;
4371  if ((unsigned)Indices[i] < InVecNumElts)
4372  MaskSelects0 = true;
4373  else
4374  MaskSelects1 = true;
4375  }
4376  if (!MaskSelects0)
4377  Op0 = UndefValue::get(InVecTy);
4378  if (!MaskSelects1)
4379  Op1 = UndefValue::get(InVecTy);
4380 
4381  auto *Op0Const = dyn_cast<Constant>(Op0);
4382  auto *Op1Const = dyn_cast<Constant>(Op1);
4383 
4384  // If all operands are constant, constant fold the shuffle.
4385  if (Op0Const && Op1Const)
4386  return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4387 
4388  // Canonicalization: if only one input vector is constant, it shall be the
4389  // second one.
4390  if (Op0Const && !Op1Const) {
4391  std::swap(Op0, Op1);
4392  ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4393  }
4394 
4395  // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4396  // value type is same as the input vectors' type.
4397  if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4398  if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4399  OpShuf->getMask()->getSplatValue())
4400  return Op0;
4401 
4402  // Don't fold a shuffle with undef mask elements. This may get folded in a
4403  // better way using demanded bits or other analysis.
4404  // TODO: Should we allow this?
4405  if (find(Indices, -1) != Indices.end())
4406  return nullptr;
4407 
4408  // Check if every element of this shuffle can be mapped back to the
4409  // corresponding element of a single root vector. If so, we don't need this
4410  // shuffle. This handles simple identity shuffles as well as chains of
4411  // shuffles that may widen/narrow and/or move elements across lanes and back.
4412  Value *RootVec = nullptr;
4413  for (unsigned i = 0; i != MaskNumElts; ++i) {
4414  // Note that recursion is limited for each vector element, so if any element
4415  // exceeds the limit, this will fail to simplify.
4416  RootVec =
4417  foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4418 
4419  // We can't replace a widening/narrowing shuffle with one of its operands.
4420  if (!RootVec || RootVec->getType() != RetTy)
4421  return nullptr;
4422  }
4423  return RootVec;
4424 }
4425 
4426 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4428  Type *RetTy, const SimplifyQuery &Q) {
4429  return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4430 }
4431 
4433  // If the input is a vector with undef elements, just return a default NaN.
4434  if (!In->isNaN())
4435  return ConstantFP::getNaN(In->getType());
4436 
4437  // Propagate the existing NaN constant when possible.
4438  // TODO: Should we quiet a signaling NaN?
4439  return In;
4440 }
4441 
4442 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4443  if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4444  return ConstantFP::getNaN(Op0->getType());
4445 
4446  if (match(Op0, m_NaN()))
4447  return propagateNaN(cast<Constant>(Op0));
4448  if (match(Op1, m_NaN()))
4449  return propagateNaN(cast<Constant>(Op1));
4450 
4451  return nullptr;
4452 }
4453 
4454 /// Given operands for an FAdd, see if we can fold the result. If not, this
4455 /// returns null.
4457  const SimplifyQuery &Q, unsigned MaxRecurse) {
4458  if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4459  return C;
4460 
4461  if (Constant *C = simplifyFPBinop(Op0, Op1))
4462  return C;
4463 
4464  // fadd X, -0 ==> X
4465  if (match(Op1, m_NegZeroFP()))
4466  return Op0;
4467 
4468  // fadd X, 0 ==> X, when we know X is not -0
4469  if (match(Op1, m_PosZeroFP()) &&
4470  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4471  return Op0;
4472 
4473  // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4474  // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4475  // Negative zeros are allowed because we always end up with positive zero:
4476  // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4477  // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4478  // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4479  // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4480  if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4481  match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4482  return ConstantFP::getNullValue(Op0->getType());
4483 
4484  // (X - Y) + Y --> X
4485  // Y + (X - Y) --> X
4486  Value *X;
4487  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4488  (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4489  match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4490  return X;
4491 
4492  return nullptr;
4493 }
4494 
4495 /// Given operands for an FSub, see if we can fold the result. If not, this
4496 /// returns null.
4498  const SimplifyQuery &Q, unsigned MaxRecurse) {
4499  if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4500  return C;
4501 
4502  if (Constant *C = simplifyFPBinop(Op0, Op1))
4503  return C;
4504 
4505  // fsub X, +0 ==> X
4506  if (match(Op1, m_PosZeroFP()))
4507  return Op0;
4508 
4509  // fsub X, -0 ==> X, when we know X is not -0
4510  if (match(Op1, m_NegZeroFP()) &&
4511  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4512  return Op0;
4513 
4514  // fsub -0.0, (fsub -0.0, X) ==> X
4515  Value *X;
4516  if (match(Op0, m_NegZeroFP()) &&
4517  match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4518  return X;
4519 
4520  // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4521  if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4522  match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4523  return X;
4524 
4525  // fsub nnan x, x ==> 0.0
4526  if (FMF.noNaNs() && Op0 == Op1)
4527  return Constant::getNullValue(Op0->getType());
4528 
4529  // Y - (Y - X) --> X
4530  // (X + Y) - Y --> X
4531  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4532  (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4533  match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4534  return X;
4535 
4536  return nullptr;
4537 }
4538 
4539 /// Given the operands for an FMul, see if we can fold the result
4541  const SimplifyQuery &Q, unsigned MaxRecurse) {
4542  if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4543  return C;
4544 
4545  if (Constant *C = simplifyFPBinop(Op0, Op1))
4546  return C;
4547 
4548  // fmul X, 1.0 ==> X
4549  if (match(Op1, m_FPOne()))
4550  return Op0;
4551 
4552  // fmul nnan nsz X, 0 ==> 0
4553  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4554  return ConstantFP::getNullValue(Op0->getType());
4555 
4556  // sqrt(X) * sqrt(X) --> X, if we can:
4557  // 1. Remove the intermediate rounding (reassociate).
4558  // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4559  // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4560  Value *X;
4561  if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4562  FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4563  return X;
4564 
4565  return nullptr;
4566 }
4567 
4569  const SimplifyQuery &Q) {
4570  return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4571 }
4572 
4573 
4575  const SimplifyQuery &Q) {
4576  return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4577 }
4578 
4580  const SimplifyQuery &Q) {
4581  return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4582 }
4583 
4585  const SimplifyQuery &Q, unsigned) {
4586  if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4587  return C;
4588 
4589  if (Constant *C = simplifyFPBinop(Op0, Op1))
4590  return C;
4591 
4592  // X / 1.0 -> X
4593  if (match(Op1, m_FPOne()))
4594  return Op0;
4595 
4596  // 0 / X -> 0
4597  // Requires that NaNs are off (X could be zero) and signed zeroes are
4598  // ignored (X could be positive or negative, so the output sign is unknown).
4599  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4600  return ConstantFP::getNullValue(Op0->getType());
4601 
4602  if (FMF.noNaNs()) {
4603  // X / X -> 1.0 is legal when NaNs are ignored.
4604  // We can ignore infinities because INF/INF is NaN.
4605  if (Op0 == Op1)
4606  return ConstantFP::get(Op0->getType(), 1.0);
4607 
4608  // (X * Y) / Y --> X if we can reassociate to the above form.
4609  Value *X;
4610  if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4611  return X;
4612 
4613  // -X / X -> -1.0 and
4614  // X / -X -> -1.0 are legal when NaNs are ignored.
4615  // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4616  if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4617  match(Op1, m_FNegNSZ(m_Specific(Op0))))
4618  return ConstantFP::get(Op0->getType(), -1.0);
4619  }
4620 
4621  return nullptr;
4622 }
4623 
4625  const SimplifyQuery &Q) {
4626  return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4627 }
4628 
4630  const SimplifyQuery &Q, unsigned) {
4631  if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4632  return C;
4633 
4634  if (Constant *C = simplifyFPBinop(Op0, Op1))
4635  return C;
4636 
4637  // Unlike fdiv, the result of frem always matches the sign of the dividend.
4638  // The constant match may include undef elements in a vector, so return a full
4639  // zero constant as the result.
4640  if (FMF.noNaNs()) {
4641  // +0 % X -> 0
4642  if (match(Op0, m_PosZeroFP()))
4643  return ConstantFP::getNullValue(Op0->getType());
4644  // -0 % X -> -0
4645  if (match(Op0, m_NegZeroFP()))
4646  return ConstantFP::getNegativeZero(Op0->getType());
4647  }
4648 
4649  return nullptr;
4650 }
4651 
4653  const SimplifyQuery &Q) {
4654  return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4655 }
4656 
4657 //=== Helper functions for higher up the class hierarchy.
4658 
4659 /// Given operands for a BinaryOperator, see if we can fold the result.
4660 /// If not, this returns null.
4661 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4662  const SimplifyQuery &Q, unsigned MaxRecurse) {
4663  switch (Opcode) {
4664  case Instruction::Add:
4665  return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4666  case Instruction::Sub:
4667  return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4668  case Instruction::Mul:
4669  return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4670  case Instruction::SDiv:
4671  return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4672  case Instruction::UDiv:
4673  return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4674  case Instruction::SRem:
4675  return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4676  case Instruction::URem:
4677  return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4678  case Instruction::Shl:
4679  return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4680  case Instruction::LShr:
4681  return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4682  case Instruction::AShr:
4683  return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4684  case Instruction::And:
4685  return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4686  case Instruction::Or:
4687  return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4688  case Instruction::Xor:
4689  return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4690  case Instruction::FAdd:
4691  return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4692  case Instruction::FSub:
4693  return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4694  case Instruction::FMul:
4695  return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4696  case Instruction::FDiv:
4697  return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4698  case Instruction::FRem:
4699  return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4700  default:
4701  llvm_unreachable("Unexpected opcode");
4702  }
4703 }
4704 
4705 /// Given operands for a BinaryOperator, see if we can fold the result.
4706 /// If not, this returns null.
4707 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4708 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4709 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4710  const FastMathFlags &FMF, const SimplifyQuery &Q,
4711  unsigned MaxRecurse) {
4712  switch (Opcode) {
4713  case Instruction::FAdd:
4714  return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4715  case Instruction::FSub:
4716  return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4717  case Instruction::FMul:
4718  return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4719  case Instruction::FDiv:
4720  return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4721  default:
4722  return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4723  }
4724 }
4725 
4726 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4727  const SimplifyQuery &Q) {
4728  return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4729 }
4730 
4731 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4732  FastMathFlags FMF, const SimplifyQuery &Q) {
4733  return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4734 }
4735 
4736 /// Given operands for a CmpInst, see if we can fold the result.
4737 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4738  const SimplifyQuery &Q, unsigned MaxRecurse) {
4740  return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4741  return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4742 }
4743 
4744 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4745  const SimplifyQuery &Q) {
4746  return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4747 }
4748 
4750  switch (ID) {
4751  default: return false;
4752 
4753  // Unary idempotent: f(f(x)) = f(x)
4754  case Intrinsic::fabs:
4755  case Intrinsic::floor:
4756  case Intrinsic::ceil:
4757  case Intrinsic::trunc:
4758  case Intrinsic::rint:
4759  case Intrinsic::nearbyint:
4760  case Intrinsic::round:
4761  case Intrinsic::canonicalize:
4762  return true;
4763  }
4764 }
4765 
4767  const DataLayout &DL) {
4768  GlobalValue *PtrSym;
4769  APInt PtrOffset;
4770  if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4771  return nullptr;
4772 
4773  Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4775  Type *Int32PtrTy = Int32Ty->getPointerTo();
4776  Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4777 
4778  auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4779  if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4780  return nullptr;
4781 
4782  uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4783  if (OffsetInt % 4 != 0)
4784  return nullptr;
4785 
4787  Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4788  ConstantInt::get(Int64Ty, OffsetInt / 4));
4789  Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4790  if (!Loaded)
4791  return nullptr;
4792 
4793  auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4794  if (!LoadedCE)
4795  return nullptr;
4796 
4797  if (LoadedCE->getOpcode() == Instruction::Trunc) {
4798  LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4799  if (!LoadedCE)
4800  return nullptr;
4801  }
4802 
4803  if (LoadedCE->getOpcode() != Instruction::Sub)
4804  return nullptr;
4805 
4806  auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4807  if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4808  return nullptr;
4809  auto *LoadedLHSPtr = LoadedLHS->