LLVM  9.0.0svn
InstructionSimplify.cpp
Go to the documentation of this file.
1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements routines for folding instructions into simpler forms
10 // that do not require creating new instructions. This does constant folding
11 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
12 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
13 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
14 // simplified: This is usually true and assuming it simplifies the logic (if
15 // they have not been simplified then results are correct but maybe suboptimal).
16 //
17 //===----------------------------------------------------------------------===//
18 
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/Statistic.h"
31 #include "llvm/IR/ConstantRange.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GlobalAlias.h"
36 #include "llvm/IR/InstrTypes.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/ValueHandle.h"
41 #include "llvm/Support/KnownBits.h"
42 #include <algorithm>
43 using namespace llvm;
44 using namespace llvm::PatternMatch;
45 
46 #define DEBUG_TYPE "instsimplify"
47 
48 enum { RecursionLimit = 3 };
49 
50 STATISTIC(NumExpand, "Number of expansions");
51 STATISTIC(NumReassoc, "Number of reassociations");
52 
53 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
54 static Value *simplifyUnOp(unsigned, Value *, const SimplifyQuery &, unsigned);
55 static Value *simplifyFPUnOp(unsigned, Value *, const FastMathFlags &,
56  const SimplifyQuery &, unsigned);
57 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
58  unsigned);
59 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
60  const SimplifyQuery &, unsigned);
61 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
62  unsigned);
63 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
64  const SimplifyQuery &Q, unsigned MaxRecurse);
65 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
66 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
67 static Value *SimplifyCastInst(unsigned, Value *, Type *,
68  const SimplifyQuery &, unsigned);
70  unsigned);
71 
72 static Value *foldSelectWithBinaryOp(Value *Cond, Value *TrueVal,
73  Value *FalseVal) {
74  BinaryOperator::BinaryOps BinOpCode;
75  if (auto *BO = dyn_cast<BinaryOperator>(Cond))
76  BinOpCode = BO->getOpcode();
77  else
78  return nullptr;
79 
80  CmpInst::Predicate ExpectedPred, Pred1, Pred2;
81  if (BinOpCode == BinaryOperator::Or) {
82  ExpectedPred = ICmpInst::ICMP_NE;
83  } else if (BinOpCode == BinaryOperator::And) {
84  ExpectedPred = ICmpInst::ICMP_EQ;
85  } else
86  return nullptr;
87 
88  // %A = icmp eq %TV, %FV
89  // %B = icmp eq %X, %Y (and one of these is a select operand)
90  // %C = and %A, %B
91  // %D = select %C, %TV, %FV
92  // -->
93  // %FV
94 
95  // %A = icmp ne %TV, %FV
96  // %B = icmp ne %X, %Y (and one of these is a select operand)
97  // %C = or %A, %B
98  // %D = select %C, %TV, %FV
99  // -->
100  // %TV
101  Value *X, *Y;
102  if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
103  m_Specific(FalseVal)),
104  m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
105  Pred1 != Pred2 || Pred1 != ExpectedPred)
106  return nullptr;
107 
108  if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
109  return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
110 
111  return nullptr;
112 }
113 
114 /// For a boolean type or a vector of boolean type, return false or a vector
115 /// with every element false.
116 static Constant *getFalse(Type *Ty) {
117  return ConstantInt::getFalse(Ty);
118 }
119 
120 /// For a boolean type or a vector of boolean type, return true or a vector
121 /// with every element true.
122 static Constant *getTrue(Type *Ty) {
123  return ConstantInt::getTrue(Ty);
124 }
125 
126 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
127 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
128  Value *RHS) {
129  CmpInst *Cmp = dyn_cast<CmpInst>(V);
130  if (!Cmp)
131  return false;
132  CmpInst::Predicate CPred = Cmp->getPredicate();
133  Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
134  if (CPred == Pred && CLHS == LHS && CRHS == RHS)
135  return true;
136  return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
137  CRHS == LHS;
138 }
139 
140 /// Does the given value dominate the specified phi node?
141 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
143  if (!I)
144  // Arguments and constants dominate all instructions.
145  return true;
146 
147  // If we are processing instructions (and/or basic blocks) that have not been
148  // fully added to a function, the parent nodes may still be null. Simply
149  // return the conservative answer in these cases.
150  if (!I->getParent() || !P->getParent() || !I->getFunction())
151  return false;
152 
153  // If we have a DominatorTree then do a precise test.
154  if (DT)
155  return DT->dominates(I, P);
156 
157  // Otherwise, if the instruction is in the entry block and is not an invoke,
158  // then it obviously dominates all phi nodes.
159  if (I->getParent() == &I->getFunction()->getEntryBlock() &&
160  !isa<InvokeInst>(I))
161  return true;
162 
163  return false;
164 }
165 
166 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
167 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
168 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
169 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
170 /// Returns the simplified value, or null if no simplification was performed.
172  Instruction::BinaryOps OpcodeToExpand,
173  const SimplifyQuery &Q, unsigned MaxRecurse) {
174  // Recursion is always used, so bail out at once if we already hit the limit.
175  if (!MaxRecurse--)
176  return nullptr;
177 
178  // Check whether the expression has the form "(A op' B) op C".
179  if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
180  if (Op0->getOpcode() == OpcodeToExpand) {
181  // It does! Try turning it into "(A op C) op' (B op C)".
182  Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
183  // Do "A op C" and "B op C" both simplify?
184  if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
185  if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
186  // They do! Return "L op' R" if it simplifies or is already available.
187  // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
188  if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
189  && L == B && R == A)) {
190  ++NumExpand;
191  return LHS;
192  }
193  // Otherwise return "L op' R" if it simplifies.
194  if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
195  ++NumExpand;
196  return V;
197  }
198  }
199  }
200 
201  // Check whether the expression has the form "A op (B op' C)".
202  if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
203  if (Op1->getOpcode() == OpcodeToExpand) {
204  // It does! Try turning it into "(A op B) op' (A op C)".
205  Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
206  // Do "A op B" and "A op C" both simplify?
207  if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
208  if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
209  // They do! Return "L op' R" if it simplifies or is already available.
210  // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
211  if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
212  && L == C && R == B)) {
213  ++NumExpand;
214  return RHS;
215  }
216  // Otherwise return "L op' R" if it simplifies.
217  if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
218  ++NumExpand;
219  return V;
220  }
221  }
222  }
223 
224  return nullptr;
225 }
226 
227 /// Generic simplifications for associative binary operations.
228 /// Returns the simpler value, or null if none was found.
230  Value *LHS, Value *RHS,
231  const SimplifyQuery &Q,
232  unsigned MaxRecurse) {
233  assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
234 
235  // Recursion is always used, so bail out at once if we already hit the limit.
236  if (!MaxRecurse--)
237  return nullptr;
238 
241 
242  // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
243  if (Op0 && Op0->getOpcode() == Opcode) {
244  Value *A = Op0->getOperand(0);
245  Value *B = Op0->getOperand(1);
246  Value *C = RHS;
247 
248  // Does "B op C" simplify?
249  if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
250  // It does! Return "A op V" if it simplifies or is already available.
251  // If V equals B then "A op V" is just the LHS.
252  if (V == B) return LHS;
253  // Otherwise return "A op V" if it simplifies.
254  if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
255  ++NumReassoc;
256  return W;
257  }
258  }
259  }
260 
261  // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
262  if (Op1 && Op1->getOpcode() == Opcode) {
263  Value *A = LHS;
264  Value *B = Op1->getOperand(0);
265  Value *C = Op1->getOperand(1);
266 
267  // Does "A op B" simplify?
268  if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
269  // It does! Return "V op C" if it simplifies or is already available.
270  // If V equals B then "V op C" is just the RHS.
271  if (V == B) return RHS;
272  // Otherwise return "V op C" if it simplifies.
273  if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
274  ++NumReassoc;
275  return W;
276  }
277  }
278  }
279 
280  // The remaining transforms require commutativity as well as associativity.
281  if (!Instruction::isCommutative(Opcode))
282  return nullptr;
283 
284  // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
285  if (Op0 && Op0->getOpcode() == Opcode) {
286  Value *A = Op0->getOperand(0);
287  Value *B = Op0->getOperand(1);
288  Value *C = RHS;
289 
290  // Does "C op A" simplify?
291  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
292  // It does! Return "V op B" if it simplifies or is already available.
293  // If V equals A then "V op B" is just the LHS.
294  if (V == A) return LHS;
295  // Otherwise return "V op B" if it simplifies.
296  if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
297  ++NumReassoc;
298  return W;
299  }
300  }
301  }
302 
303  // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
304  if (Op1 && Op1->getOpcode() == Opcode) {
305  Value *A = LHS;
306  Value *B = Op1->getOperand(0);
307  Value *C = Op1->getOperand(1);
308 
309  // Does "C op A" simplify?
310  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
311  // It does! Return "B op V" if it simplifies or is already available.
312  // If V equals C then "B op V" is just the RHS.
313  if (V == C) return RHS;
314  // Otherwise return "B op V" if it simplifies.
315  if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
316  ++NumReassoc;
317  return W;
318  }
319  }
320  }
321 
322  return nullptr;
323 }
324 
325 /// In the case of a binary operation with a select instruction as an operand,
326 /// try to simplify the binop by seeing whether evaluating it on both branches
327 /// of the select results in the same value. Returns the common value if so,
328 /// otherwise returns null.
330  Value *RHS, const SimplifyQuery &Q,
331  unsigned MaxRecurse) {
332  // Recursion is always used, so bail out at once if we already hit the limit.
333  if (!MaxRecurse--)
334  return nullptr;
335 
336  SelectInst *SI;
337  if (isa<SelectInst>(LHS)) {
338  SI = cast<SelectInst>(LHS);
339  } else {
340  assert(isa<SelectInst>(RHS) && "No select instruction operand!");
341  SI = cast<SelectInst>(RHS);
342  }
343 
344  // Evaluate the BinOp on the true and false branches of the select.
345  Value *TV;
346  Value *FV;
347  if (SI == LHS) {
348  TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
349  FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
350  } else {
351  TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
352  FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
353  }
354 
355  // If they simplified to the same value, then return the common value.
356  // If they both failed to simplify then return null.
357  if (TV == FV)
358  return TV;
359 
360  // If one branch simplified to undef, return the other one.
361  if (TV && isa<UndefValue>(TV))
362  return FV;
363  if (FV && isa<UndefValue>(FV))
364  return TV;
365 
366  // If applying the operation did not change the true and false select values,
367  // then the result of the binop is the select itself.
368  if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
369  return SI;
370 
371  // If one branch simplified and the other did not, and the simplified
372  // value is equal to the unsimplified one, return the simplified value.
373  // For example, select (cond, X, X & Z) & Z -> X & Z.
374  if ((FV && !TV) || (TV && !FV)) {
375  // Check that the simplified value has the form "X op Y" where "op" is the
376  // same as the original operation.
377  Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
378  if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
379  // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
380  // We already know that "op" is the same as for the simplified value. See
381  // if the operands match too. If so, return the simplified value.
382  Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
383  Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
384  Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
385  if (Simplified->getOperand(0) == UnsimplifiedLHS &&
386  Simplified->getOperand(1) == UnsimplifiedRHS)
387  return Simplified;
388  if (Simplified->isCommutative() &&
389  Simplified->getOperand(1) == UnsimplifiedLHS &&
390  Simplified->getOperand(0) == UnsimplifiedRHS)
391  return Simplified;
392  }
393  }
394 
395  return nullptr;
396 }
397 
398 /// In the case of a comparison with a select instruction, try to simplify the
399 /// comparison by seeing whether both branches of the select result in the same
400 /// value. Returns the common value if so, otherwise returns null.
402  Value *RHS, const SimplifyQuery &Q,
403  unsigned MaxRecurse) {
404  // Recursion is always used, so bail out at once if we already hit the limit.
405  if (!MaxRecurse--)
406  return nullptr;
407 
408  // Make sure the select is on the LHS.
409  if (!isa<SelectInst>(LHS)) {
410  std::swap(LHS, RHS);
411  Pred = CmpInst::getSwappedPredicate(Pred);
412  }
413  assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
414  SelectInst *SI = cast<SelectInst>(LHS);
415  Value *Cond = SI->getCondition();
416  Value *TV = SI->getTrueValue();
417  Value *FV = SI->getFalseValue();
418 
419  // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
420  // Does "cmp TV, RHS" simplify?
421  Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
422  if (TCmp == Cond) {
423  // It not only simplified, it simplified to the select condition. Replace
424  // it with 'true'.
425  TCmp = getTrue(Cond->getType());
426  } else if (!TCmp) {
427  // It didn't simplify. However if "cmp TV, RHS" is equal to the select
428  // condition then we can replace it with 'true'. Otherwise give up.
429  if (!isSameCompare(Cond, Pred, TV, RHS))
430  return nullptr;
431  TCmp = getTrue(Cond->getType());
432  }
433 
434  // Does "cmp FV, RHS" simplify?
435  Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
436  if (FCmp == Cond) {
437  // It not only simplified, it simplified to the select condition. Replace
438  // it with 'false'.
439  FCmp = getFalse(Cond->getType());
440  } else if (!FCmp) {
441  // It didn't simplify. However if "cmp FV, RHS" is equal to the select
442  // condition then we can replace it with 'false'. Otherwise give up.
443  if (!isSameCompare(Cond, Pred, FV, RHS))
444  return nullptr;
445  FCmp = getFalse(Cond->getType());
446  }
447 
448  // If both sides simplified to the same value, then use it as the result of
449  // the original comparison.
450  if (TCmp == FCmp)
451  return TCmp;
452 
453  // The remaining cases only make sense if the select condition has the same
454  // type as the result of the comparison, so bail out if this is not so.
455  if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
456  return nullptr;
457  // If the false value simplified to false, then the result of the compare
458  // is equal to "Cond && TCmp". This also catches the case when the false
459  // value simplified to false and the true value to true, returning "Cond".
460  if (match(FCmp, m_Zero()))
461  if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
462  return V;
463  // If the true value simplified to true, then the result of the compare
464  // is equal to "Cond || FCmp".
465  if (match(TCmp, m_One()))
466  if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
467  return V;
468  // Finally, if the false value simplified to true and the true value to
469  // false, then the result of the compare is equal to "!Cond".
470  if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
471  if (Value *V =
472  SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
473  Q, MaxRecurse))
474  return V;
475 
476  return nullptr;
477 }
478 
479 /// In the case of a binary operation with an operand that is a PHI instruction,
480 /// try to simplify the binop by seeing whether evaluating it on the incoming
481 /// phi values yields the same result for every value. If so returns the common
482 /// value, otherwise returns null.
484  Value *RHS, const SimplifyQuery &Q,
485  unsigned MaxRecurse) {
486  // Recursion is always used, so bail out at once if we already hit the limit.
487  if (!MaxRecurse--)
488  return nullptr;
489 
490  PHINode *PI;
491  if (isa<PHINode>(LHS)) {
492  PI = cast<PHINode>(LHS);
493  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
494  if (!valueDominatesPHI(RHS, PI, Q.DT))
495  return nullptr;
496  } else {
497  assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
498  PI = cast<PHINode>(RHS);
499  // Bail out if LHS and the phi may be mutually interdependent due to a loop.
500  if (!valueDominatesPHI(LHS, PI, Q.DT))
501  return nullptr;
502  }
503 
504  // Evaluate the BinOp on the incoming phi values.
505  Value *CommonValue = nullptr;
506  for (Value *Incoming : PI->incoming_values()) {
507  // If the incoming value is the phi node itself, it can safely be skipped.
508  if (Incoming == PI) continue;
509  Value *V = PI == LHS ?
510  SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
511  SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
512  // If the operation failed to simplify, or simplified to a different value
513  // to previously, then give up.
514  if (!V || (CommonValue && V != CommonValue))
515  return nullptr;
516  CommonValue = V;
517  }
518 
519  return CommonValue;
520 }
521 
522 /// In the case of a comparison with a PHI instruction, try to simplify the
523 /// comparison by seeing whether comparing with all of the incoming phi values
524 /// yields the same result every time. If so returns the common result,
525 /// otherwise returns null.
527  const SimplifyQuery &Q, unsigned MaxRecurse) {
528  // Recursion is always used, so bail out at once if we already hit the limit.
529  if (!MaxRecurse--)
530  return nullptr;
531 
532  // Make sure the phi is on the LHS.
533  if (!isa<PHINode>(LHS)) {
534  std::swap(LHS, RHS);
535  Pred = CmpInst::getSwappedPredicate(Pred);
536  }
537  assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
538  PHINode *PI = cast<PHINode>(LHS);
539 
540  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
541  if (!valueDominatesPHI(RHS, PI, Q.DT))
542  return nullptr;
543 
544  // Evaluate the BinOp on the incoming phi values.
545  Value *CommonValue = nullptr;
546  for (Value *Incoming : PI->incoming_values()) {
547  // If the incoming value is the phi node itself, it can safely be skipped.
548  if (Incoming == PI) continue;
549  Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
550  // If the operation failed to simplify, or simplified to a different value
551  // to previously, then give up.
552  if (!V || (CommonValue && V != CommonValue))
553  return nullptr;
554  CommonValue = V;
555  }
556 
557  return CommonValue;
558 }
559 
561  Value *&Op0, Value *&Op1,
562  const SimplifyQuery &Q) {
563  if (auto *CLHS = dyn_cast<Constant>(Op0)) {
564  if (auto *CRHS = dyn_cast<Constant>(Op1))
565  return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
566 
567  // Canonicalize the constant to the RHS if this is a commutative operation.
568  if (Instruction::isCommutative(Opcode))
569  std::swap(Op0, Op1);
570  }
571  return nullptr;
572 }
573 
574 /// Given operands for an Add, see if we can fold the result.
575 /// If not, this returns null.
576 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
577  const SimplifyQuery &Q, unsigned MaxRecurse) {
578  if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
579  return C;
580 
581  // X + undef -> undef
582  if (match(Op1, m_Undef()))
583  return Op1;
584 
585  // X + 0 -> X
586  if (match(Op1, m_Zero()))
587  return Op0;
588 
589  // If two operands are negative, return 0.
590  if (isKnownNegation(Op0, Op1))
591  return Constant::getNullValue(Op0->getType());
592 
593  // X + (Y - X) -> Y
594  // (Y - X) + X -> Y
595  // Eg: X + -X -> 0
596  Value *Y = nullptr;
597  if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
598  match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
599  return Y;
600 
601  // X + ~X -> -1 since ~X = -X-1
602  Type *Ty = Op0->getType();
603  if (match(Op0, m_Not(m_Specific(Op1))) ||
604  match(Op1, m_Not(m_Specific(Op0))))
605  return Constant::getAllOnesValue(Ty);
606 
607  // add nsw/nuw (xor Y, signmask), signmask --> Y
608  // The no-wrapping add guarantees that the top bit will be set by the add.
609  // Therefore, the xor must be clearing the already set sign bit of Y.
610  if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
611  match(Op0, m_Xor(m_Value(Y), m_SignMask())))
612  return Y;
613 
614  // add nuw %x, -1 -> -1, because %x can only be 0.
615  if (IsNUW && match(Op1, m_AllOnes()))
616  return Op1; // Which is -1.
617 
618  /// i1 add -> xor.
619  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
620  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
621  return V;
622 
623  // Try some generic simplifications for associative operations.
624  if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
625  MaxRecurse))
626  return V;
627 
628  // Threading Add over selects and phi nodes is pointless, so don't bother.
629  // Threading over the select in "A + select(cond, B, C)" means evaluating
630  // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
631  // only if B and C are equal. If B and C are equal then (since we assume
632  // that operands have already been simplified) "select(cond, B, C)" should
633  // have been simplified to the common value of B and C already. Analysing
634  // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
635  // for threading over phi nodes.
636 
637  return nullptr;
638 }
639 
640 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
641  const SimplifyQuery &Query) {
642  return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
643 }
644 
645 /// Compute the base pointer and cumulative constant offsets for V.
646 ///
647 /// This strips all constant offsets off of V, leaving it the base pointer, and
648 /// accumulates the total constant offset applied in the returned constant. It
649 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
650 /// no constant offsets applied.
651 ///
652 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
653 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
654 /// folding.
656  bool AllowNonInbounds = false) {
658 
659  Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
660  APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
661 
662  // Even though we don't look through PHI nodes, we could be called on an
663  // instruction in an unreachable block, which may be on a cycle.
664  SmallPtrSet<Value *, 4> Visited;
665  Visited.insert(V);
666  do {
667  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
668  if ((!AllowNonInbounds && !GEP->isInBounds()) ||
669  !GEP->accumulateConstantOffset(DL, Offset))
670  break;
671  V = GEP->getPointerOperand();
672  } else if (Operator::getOpcode(V) == Instruction::BitCast) {
673  V = cast<Operator>(V)->getOperand(0);
674  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
675  if (GA->isInterposable())
676  break;
677  V = GA->getAliasee();
678  } else {
679  if (auto *Call = dyn_cast<CallBase>(V))
680  if (Value *RV = Call->getReturnedArgOperand()) {
681  V = RV;
682  continue;
683  }
684  break;
685  }
686  assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
687  } while (Visited.insert(V).second);
688 
689  Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
690  if (V->getType()->isVectorTy())
692  OffsetIntPtr);
693  return OffsetIntPtr;
694 }
695 
696 /// Compute the constant difference between two pointer values.
697 /// If the difference is not a constant, returns zero.
699  Value *RHS) {
700  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
701  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
702 
703  // If LHS and RHS are not related via constant offsets to the same base
704  // value, there is nothing we can do here.
705  if (LHS != RHS)
706  return nullptr;
707 
708  // Otherwise, the difference of LHS - RHS can be computed as:
709  // LHS - RHS
710  // = (LHSOffset + Base) - (RHSOffset + Base)
711  // = LHSOffset - RHSOffset
712  return ConstantExpr::getSub(LHSOffset, RHSOffset);
713 }
714 
715 /// Given operands for a Sub, see if we can fold the result.
716 /// If not, this returns null.
717 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
718  const SimplifyQuery &Q, unsigned MaxRecurse) {
719  if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
720  return C;
721 
722  // X - undef -> undef
723  // undef - X -> undef
724  if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
725  return UndefValue::get(Op0->getType());
726 
727  // X - 0 -> X
728  if (match(Op1, m_Zero()))
729  return Op0;
730 
731  // X - X -> 0
732  if (Op0 == Op1)
733  return Constant::getNullValue(Op0->getType());
734 
735  // Is this a negation?
736  if (match(Op0, m_Zero())) {
737  // 0 - X -> 0 if the sub is NUW.
738  if (isNUW)
739  return Constant::getNullValue(Op0->getType());
740 
741  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
742  if (Known.Zero.isMaxSignedValue()) {
743  // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
744  // Op1 must be 0 because negating the minimum signed value is undefined.
745  if (isNSW)
746  return Constant::getNullValue(Op0->getType());
747 
748  // 0 - X -> X if X is 0 or the minimum signed value.
749  return Op1;
750  }
751  }
752 
753  // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
754  // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
755  Value *X = nullptr, *Y = nullptr, *Z = Op1;
756  if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
757  // See if "V === Y - Z" simplifies.
758  if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
759  // It does! Now see if "X + V" simplifies.
760  if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
761  // It does, we successfully reassociated!
762  ++NumReassoc;
763  return W;
764  }
765  // See if "V === X - Z" simplifies.
766  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
767  // It does! Now see if "Y + V" simplifies.
768  if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
769  // It does, we successfully reassociated!
770  ++NumReassoc;
771  return W;
772  }
773  }
774 
775  // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
776  // For example, X - (X + 1) -> -1
777  X = Op0;
778  if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
779  // See if "V === X - Y" simplifies.
780  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
781  // It does! Now see if "V - Z" simplifies.
782  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
783  // It does, we successfully reassociated!
784  ++NumReassoc;
785  return W;
786  }
787  // See if "V === X - Z" simplifies.
788  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
789  // It does! Now see if "V - Y" simplifies.
790  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
791  // It does, we successfully reassociated!
792  ++NumReassoc;
793  return W;
794  }
795  }
796 
797  // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
798  // For example, X - (X - Y) -> Y.
799  Z = Op0;
800  if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
801  // See if "V === Z - X" simplifies.
802  if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
803  // It does! Now see if "V + Y" simplifies.
804  if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
805  // It does, we successfully reassociated!
806  ++NumReassoc;
807  return W;
808  }
809 
810  // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
811  if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
812  match(Op1, m_Trunc(m_Value(Y))))
813  if (X->getType() == Y->getType())
814  // See if "V === X - Y" simplifies.
815  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
816  // It does! Now see if "trunc V" simplifies.
817  if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
818  Q, MaxRecurse - 1))
819  // It does, return the simplified "trunc V".
820  return W;
821 
822  // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
823  if (match(Op0, m_PtrToInt(m_Value(X))) &&
824  match(Op1, m_PtrToInt(m_Value(Y))))
825  if (Constant *Result = computePointerDifference(Q.DL, X, Y))
826  return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
827 
828  // i1 sub -> xor.
829  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
830  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
831  return V;
832 
833  // Threading Sub over selects and phi nodes is pointless, so don't bother.
834  // Threading over the select in "A - select(cond, B, C)" means evaluating
835  // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
836  // only if B and C are equal. If B and C are equal then (since we assume
837  // that operands have already been simplified) "select(cond, B, C)" should
838  // have been simplified to the common value of B and C already. Analysing
839  // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
840  // for threading over phi nodes.
841 
842  return nullptr;
843 }
844 
845 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
846  const SimplifyQuery &Q) {
847  return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
848 }
849 
850 /// Given operands for a Mul, see if we can fold the result.
851 /// If not, this returns null.
852 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
853  unsigned MaxRecurse) {
854  if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
855  return C;
856 
857  // X * undef -> 0
858  // X * 0 -> 0
859  if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
860  return Constant::getNullValue(Op0->getType());
861 
862  // X * 1 -> X
863  if (match(Op1, m_One()))
864  return Op0;
865 
866  // (X / Y) * Y -> X if the division is exact.
867  Value *X = nullptr;
868  if (Q.IIQ.UseInstrInfo &&
869  (match(Op0,
870  m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
871  match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
872  return X;
873 
874  // i1 mul -> and.
875  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
876  if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
877  return V;
878 
879  // Try some generic simplifications for associative operations.
880  if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
881  MaxRecurse))
882  return V;
883 
884  // Mul distributes over Add. Try some generic simplifications based on this.
885  if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
886  Q, MaxRecurse))
887  return V;
888 
889  // If the operation is with the result of a select instruction, check whether
890  // operating on either branch of the select always yields the same value.
891  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
892  if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
893  MaxRecurse))
894  return V;
895 
896  // If the operation is with the result of a phi instruction, check whether
897  // operating on all incoming values of the phi always yields the same value.
898  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
899  if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
900  MaxRecurse))
901  return V;
902 
903  return nullptr;
904 }
905 
908 }
909 
910 /// Check for common or similar folds of integer division or integer remainder.
911 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
912 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
913  Type *Ty = Op0->getType();
914 
915  // X / undef -> undef
916  // X % undef -> undef
917  if (match(Op1, m_Undef()))
918  return Op1;
919 
920  // X / 0 -> undef
921  // X % 0 -> undef
922  // We don't need to preserve faults!
923  if (match(Op1, m_Zero()))
924  return UndefValue::get(Ty);
925 
926  // If any element of a constant divisor vector is zero or undef, the whole op
927  // is undef.
928  auto *Op1C = dyn_cast<Constant>(Op1);
929  if (Op1C && Ty->isVectorTy()) {
930  unsigned NumElts = Ty->getVectorNumElements();
931  for (unsigned i = 0; i != NumElts; ++i) {
932  Constant *Elt = Op1C->getAggregateElement(i);
933  if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
934  return UndefValue::get(Ty);
935  }
936  }
937 
938  // undef / X -> 0
939  // undef % X -> 0
940  if (match(Op0, m_Undef()))
941  return Constant::getNullValue(Ty);
942 
943  // 0 / X -> 0
944  // 0 % X -> 0
945  if (match(Op0, m_Zero()))
946  return Constant::getNullValue(Op0->getType());
947 
948  // X / X -> 1
949  // X % X -> 0
950  if (Op0 == Op1)
951  return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
952 
953  // X / 1 -> X
954  // X % 1 -> 0
955  // If this is a boolean op (single-bit element type), we can't have
956  // division-by-zero or remainder-by-zero, so assume the divisor is 1.
957  // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
958  Value *X;
959  if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
960  (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
961  return IsDiv ? Op0 : Constant::getNullValue(Ty);
962 
963  return nullptr;
964 }
965 
966 /// Given a predicate and two operands, return true if the comparison is true.
967 /// This is a helper for div/rem simplification where we return some other value
968 /// when we can prove a relationship between the operands.
969 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
970  const SimplifyQuery &Q, unsigned MaxRecurse) {
971  Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
972  Constant *C = dyn_cast_or_null<Constant>(V);
973  return (C && C->isAllOnesValue());
974 }
975 
976 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
977 /// to simplify X % Y to X.
978 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
979  unsigned MaxRecurse, bool IsSigned) {
980  // Recursion is always used, so bail out at once if we already hit the limit.
981  if (!MaxRecurse--)
982  return false;
983 
984  if (IsSigned) {
985  // |X| / |Y| --> 0
986  //
987  // We require that 1 operand is a simple constant. That could be extended to
988  // 2 variables if we computed the sign bit for each.
989  //
990  // Make sure that a constant is not the minimum signed value because taking
991  // the abs() of that is undefined.
992  Type *Ty = X->getType();
993  const APInt *C;
994  if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
995  // Is the variable divisor magnitude always greater than the constant
996  // dividend magnitude?
997  // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
998  Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
999  Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
1000  if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
1001  isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
1002  return true;
1003  }
1004  if (match(Y, m_APInt(C))) {
1005  // Special-case: we can't take the abs() of a minimum signed value. If
1006  // that's the divisor, then all we have to do is prove that the dividend
1007  // is also not the minimum signed value.
1008  if (C->isMinSignedValue())
1009  return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1010 
1011  // Is the variable dividend magnitude always less than the constant
1012  // divisor magnitude?
1013  // |X| < |C| --> X > -abs(C) and X < abs(C)
1014  Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1015  Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1016  if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1017  isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1018  return true;
1019  }
1020  return false;
1021  }
1022 
1023  // IsSigned == false.
1024  // Is the dividend unsigned less than the divisor?
1025  return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1026 }
1027 
1028 /// These are simplifications common to SDiv and UDiv.
1030  const SimplifyQuery &Q, unsigned MaxRecurse) {
1031  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1032  return C;
1033 
1034  if (Value *V = simplifyDivRem(Op0, Op1, true))
1035  return V;
1036 
1037  bool IsSigned = Opcode == Instruction::SDiv;
1038 
1039  // (X * Y) / Y -> X if the multiplication does not overflow.
1040  Value *X;
1041  if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1042  auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1043  // If the Mul does not overflow, then we are good to go.
1044  if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1045  (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1046  return X;
1047  // If X has the form X = A / Y, then X * Y cannot overflow.
1048  if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1049  (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1050  return X;
1051  }
1052 
1053  // (X rem Y) / Y -> 0
1054  if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1055  (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1056  return Constant::getNullValue(Op0->getType());
1057 
1058  // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1059  ConstantInt *C1, *C2;
1060  if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1061  match(Op1, m_ConstantInt(C2))) {
1062  bool Overflow;
1063  (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1064  if (Overflow)
1065  return Constant::getNullValue(Op0->getType());
1066  }
1067 
1068  // If the operation is with the result of a select instruction, check whether
1069  // operating on either branch of the select always yields the same value.
1070  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1071  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1072  return V;
1073 
1074  // If the operation is with the result of a phi instruction, check whether
1075  // operating on all incoming values of the phi always yields the same value.
1076  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1077  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1078  return V;
1079 
1080  if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1081  return Constant::getNullValue(Op0->getType());
1082 
1083  return nullptr;
1084 }
1085 
1086 /// These are simplifications common to SRem and URem.
1088  const SimplifyQuery &Q, unsigned MaxRecurse) {
1089  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1090  return C;
1091 
1092  if (Value *V = simplifyDivRem(Op0, Op1, false))
1093  return V;
1094 
1095  // (X % Y) % Y -> X % Y
1096  if ((Opcode == Instruction::SRem &&
1097  match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1098  (Opcode == Instruction::URem &&
1099  match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1100  return Op0;
1101 
1102  // (X << Y) % X -> 0
1103  if (Q.IIQ.UseInstrInfo &&
1104  ((Opcode == Instruction::SRem &&
1105  match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1106  (Opcode == Instruction::URem &&
1107  match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1108  return Constant::getNullValue(Op0->getType());
1109 
1110  // If the operation is with the result of a select instruction, check whether
1111  // operating on either branch of the select always yields the same value.
1112  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1113  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1114  return V;
1115 
1116  // If the operation is with the result of a phi instruction, check whether
1117  // operating on all incoming values of the phi always yields the same value.
1118  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1119  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1120  return V;
1121 
1122  // If X / Y == 0, then X % Y == X.
1123  if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1124  return Op0;
1125 
1126  return nullptr;
1127 }
1128 
1129 /// Given operands for an SDiv, see if we can fold the result.
1130 /// If not, this returns null.
1131 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1132  unsigned MaxRecurse) {
1133  // If two operands are negated and no signed overflow, return -1.
1134  if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1135  return Constant::getAllOnesValue(Op0->getType());
1136 
1137  return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1138 }
1139 
1142 }
1143 
1144 /// Given operands for a UDiv, see if we can fold the result.
1145 /// If not, this returns null.
1146 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1147  unsigned MaxRecurse) {
1148  return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1149 }
1150 
1153 }
1154 
1155 /// Given operands for an SRem, see if we can fold the result.
1156 /// If not, this returns null.
1157 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1158  unsigned MaxRecurse) {
1159  // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1160  // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1161  Value *X;
1162  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1163  return ConstantInt::getNullValue(Op0->getType());
1164 
1165  // If the two operands are negated, return 0.
1166  if (isKnownNegation(Op0, Op1))
1167  return ConstantInt::getNullValue(Op0->getType());
1168 
1169  return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1170 }
1171 
1174 }
1175 
1176 /// Given operands for a URem, see if we can fold the result.
1177 /// If not, this returns null.
1178 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1179  unsigned MaxRecurse) {
1180  return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1181 }
1182 
1185 }
1186 
1187 /// Returns true if a shift by \c Amount always yields undef.
1188 static bool isUndefShift(Value *Amount) {
1189  Constant *C = dyn_cast<Constant>(Amount);
1190  if (!C)
1191  return false;
1192 
1193  // X shift by undef -> undef because it may shift by the bitwidth.
1194  if (isa<UndefValue>(C))
1195  return true;
1196 
1197  // Shifting by the bitwidth or more is undefined.
1198  if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1199  if (CI->getValue().getLimitedValue() >=
1200  CI->getType()->getScalarSizeInBits())
1201  return true;
1202 
1203  // If all lanes of a vector shift are undefined the whole shift is.
1204  if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1205  for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1207  return false;
1208  return true;
1209  }
1210 
1211  return false;
1212 }
1213 
1214 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1215 /// If not, this returns null.
1217  Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1218  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1219  return C;
1220 
1221  // 0 shift by X -> 0
1222  if (match(Op0, m_Zero()))
1223  return Constant::getNullValue(Op0->getType());
1224 
1225  // X shift by 0 -> X
1226  // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1227  // would be poison.
1228  Value *X;
1229  if (match(Op1, m_Zero()) ||
1230  (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1231  return Op0;
1232 
1233  // Fold undefined shifts.
1234  if (isUndefShift(Op1))
1235  return UndefValue::get(Op0->getType());
1236 
1237  // If the operation is with the result of a select instruction, check whether
1238  // operating on either branch of the select always yields the same value.
1239  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1240  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1241  return V;
1242 
1243  // If the operation is with the result of a phi instruction, check whether
1244  // operating on all incoming values of the phi always yields the same value.
1245  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1246  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1247  return V;
1248 
1249  // If any bits in the shift amount make that value greater than or equal to
1250  // the number of bits in the type, the shift is undefined.
1251  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1252  if (Known.One.getLimitedValue() >= Known.getBitWidth())
1253  return UndefValue::get(Op0->getType());
1254 
1255  // If all valid bits in the shift amount are known zero, the first operand is
1256  // unchanged.
1257  unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1258  if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1259  return Op0;
1260 
1261  return nullptr;
1262 }
1263 
1264 /// Given operands for an Shl, LShr or AShr, see if we can
1265 /// fold the result. If not, this returns null.
1267  Value *Op1, bool isExact, const SimplifyQuery &Q,
1268  unsigned MaxRecurse) {
1269  if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1270  return V;
1271 
1272  // X >> X -> 0
1273  if (Op0 == Op1)
1274  return Constant::getNullValue(Op0->getType());
1275 
1276  // undef >> X -> 0
1277  // undef >> X -> undef (if it's exact)
1278  if (match(Op0, m_Undef()))
1279  return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1280 
1281  // The low bit cannot be shifted out of an exact shift if it is set.
1282  if (isExact) {
1283  KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1284  if (Op0Known.One[0])
1285  return Op0;
1286  }
1287 
1288  return nullptr;
1289 }
1290 
1291 /// Given operands for an Shl, see if we can fold the result.
1292 /// If not, this returns null.
1293 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1294  const SimplifyQuery &Q, unsigned MaxRecurse) {
1295  if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1296  return V;
1297 
1298  // undef << X -> 0
1299  // undef << X -> undef if (if it's NSW/NUW)
1300  if (match(Op0, m_Undef()))
1301  return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1302 
1303  // (X >> A) << A -> X
1304  Value *X;
1305  if (Q.IIQ.UseInstrInfo &&
1306  match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1307  return X;
1308 
1309  // shl nuw i8 C, %x -> C iff C has sign bit set.
1310  if (isNUW && match(Op0, m_Negative()))
1311  return Op0;
1312  // NOTE: could use computeKnownBits() / LazyValueInfo,
1313  // but the cost-benefit analysis suggests it isn't worth it.
1314 
1315  return nullptr;
1316 }
1317 
1318 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1319  const SimplifyQuery &Q) {
1320  return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1321 }
1322 
1323 /// Given operands for an LShr, see if we can fold the result.
1324 /// If not, this returns null.
1325 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1326  const SimplifyQuery &Q, unsigned MaxRecurse) {
1327  if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1328  MaxRecurse))
1329  return V;
1330 
1331  // (X << A) >> A -> X
1332  Value *X;
1333  if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1334  return X;
1335 
1336  // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1337  // We can return X as we do in the above case since OR alters no bits in X.
1338  // SimplifyDemandedBits in InstCombine can do more general optimization for
1339  // bit manipulation. This pattern aims to provide opportunities for other
1340  // optimizers by supporting a simple but common case in InstSimplify.
1341  Value *Y;
1342  const APInt *ShRAmt, *ShLAmt;
1343  if (match(Op1, m_APInt(ShRAmt)) &&
1344  match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1345  *ShRAmt == *ShLAmt) {
1346  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1347  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1348  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1349  if (ShRAmt->uge(EffWidthY))
1350  return X;
1351  }
1352 
1353  return nullptr;
1354 }
1355 
1356 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1357  const SimplifyQuery &Q) {
1358  return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1359 }
1360 
1361 /// Given operands for an AShr, see if we can fold the result.
1362 /// If not, this returns null.
1363 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1364  const SimplifyQuery &Q, unsigned MaxRecurse) {
1365  if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1366  MaxRecurse))
1367  return V;
1368 
1369  // all ones >>a X -> -1
1370  // Do not return Op0 because it may contain undef elements if it's a vector.
1371  if (match(Op0, m_AllOnes()))
1372  return Constant::getAllOnesValue(Op0->getType());
1373 
1374  // (X << A) >> A -> X
1375  Value *X;
1376  if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1377  return X;
1378 
1379  // Arithmetic shifting an all-sign-bit value is a no-op.
1380  unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1381  if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1382  return Op0;
1383 
1384  return nullptr;
1385 }
1386 
1387 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1388  const SimplifyQuery &Q) {
1389  return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1390 }
1391 
1392 /// Commuted variants are assumed to be handled by calling this function again
1393 /// with the parameters swapped.
1395  ICmpInst *UnsignedICmp, bool IsAnd) {
1396  Value *X, *Y;
1397 
1398  ICmpInst::Predicate EqPred;
1399  if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1400  !ICmpInst::isEquality(EqPred))
1401  return nullptr;
1402 
1403  ICmpInst::Predicate UnsignedPred;
1404  if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1405  ICmpInst::isUnsigned(UnsignedPred))
1406  ;
1407  else if (match(UnsignedICmp,
1408  m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1409  ICmpInst::isUnsigned(UnsignedPred))
1410  UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1411  else
1412  return nullptr;
1413 
1414  // X < Y && Y != 0 --> X < Y
1415  // X < Y || Y != 0 --> Y != 0
1416  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1417  return IsAnd ? UnsignedICmp : ZeroICmp;
1418 
1419  // X >= Y || Y != 0 --> true
1420  // X >= Y || Y == 0 --> X >= Y
1421  if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1422  if (EqPred == ICmpInst::ICMP_NE)
1423  return getTrue(UnsignedICmp->getType());
1424  return UnsignedICmp;
1425  }
1426 
1427  // X < Y && Y == 0 --> false
1428  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1429  IsAnd)
1430  return getFalse(UnsignedICmp->getType());
1431 
1432  return nullptr;
1433 }
1434 
1435 /// Commuted variants are assumed to be handled by calling this function again
1436 /// with the parameters swapped.
1438  ICmpInst::Predicate Pred0, Pred1;
1439  Value *A ,*B;
1440  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1441  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1442  return nullptr;
1443 
1444  // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1445  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1446  // can eliminate Op1 from this 'and'.
1447  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1448  return Op0;
1449 
1450  // Check for any combination of predicates that are guaranteed to be disjoint.
1451  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1452  (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1453  (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1454  (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1455  return getFalse(Op0->getType());
1456 
1457  return nullptr;
1458 }
1459 
1460 /// Commuted variants are assumed to be handled by calling this function again
1461 /// with the parameters swapped.
1463  ICmpInst::Predicate Pred0, Pred1;
1464  Value *A ,*B;
1465  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1466  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1467  return nullptr;
1468 
1469  // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1470  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1471  // can eliminate Op0 from this 'or'.
1472  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1473  return Op1;
1474 
1475  // Check for any combination of predicates that cover the entire range of
1476  // possibilities.
1477  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1478  (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1479  (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1480  (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1481  return getTrue(Op0->getType());
1482 
1483  return nullptr;
1484 }
1485 
1486 /// Test if a pair of compares with a shared operand and 2 constants has an
1487 /// empty set intersection, full set union, or if one compare is a superset of
1488 /// the other.
1490  bool IsAnd) {
1491  // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1492  if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1493  return nullptr;
1494 
1495  const APInt *C0, *C1;
1496  if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1497  !match(Cmp1->getOperand(1), m_APInt(C1)))
1498  return nullptr;
1499 
1500  auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1501  auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1502 
1503  // For and-of-compares, check if the intersection is empty:
1504  // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1505  if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1506  return getFalse(Cmp0->getType());
1507 
1508  // For or-of-compares, check if the union is full:
1509  // (icmp X, C0) || (icmp X, C1) --> full set --> true
1510  if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1511  return getTrue(Cmp0->getType());
1512 
1513  // Is one range a superset of the other?
1514  // If this is and-of-compares, take the smaller set:
1515  // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1516  // If this is or-of-compares, take the larger set:
1517  // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1518  if (Range0.contains(Range1))
1519  return IsAnd ? Cmp1 : Cmp0;
1520  if (Range1.contains(Range0))
1521  return IsAnd ? Cmp0 : Cmp1;
1522 
1523  return nullptr;
1524 }
1525 
1527  bool IsAnd) {
1528  ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1529  if (!match(Cmp0->getOperand(1), m_Zero()) ||
1530  !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1531  return nullptr;
1532 
1533  if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1534  return nullptr;
1535 
1536  // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1537  Value *X = Cmp0->getOperand(0);
1538  Value *Y = Cmp1->getOperand(0);
1539 
1540  // If one of the compares is a masked version of a (not) null check, then
1541  // that compare implies the other, so we eliminate the other. Optionally, look
1542  // through a pointer-to-int cast to match a null check of a pointer type.
1543 
1544  // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1545  // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1546  // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1547  // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1548  if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1550  return Cmp1;
1551 
1552  // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1553  // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1554  // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1555  // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1556  if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1558  return Cmp0;
1559 
1560  return nullptr;
1561 }
1562 
1564  const InstrInfoQuery &IIQ) {
1565  // (icmp (add V, C0), C1) & (icmp V, C0)
1566  ICmpInst::Predicate Pred0, Pred1;
1567  const APInt *C0, *C1;
1568  Value *V;
1569  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1570  return nullptr;
1571 
1572  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1573  return nullptr;
1574 
1575  auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1576  if (AddInst->getOperand(1) != Op1->getOperand(1))
1577  return nullptr;
1578 
1579  Type *ITy = Op0->getType();
1580  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1581  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1582 
1583  const APInt Delta = *C1 - *C0;
1584  if (C0->isStrictlyPositive()) {
1585  if (Delta == 2) {
1586  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1587  return getFalse(ITy);
1588  if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1589  return getFalse(ITy);
1590  }
1591  if (Delta == 1) {
1592  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1593  return getFalse(ITy);
1594  if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1595  return getFalse(ITy);
1596  }
1597  }
1598  if (C0->getBoolValue() && isNUW) {
1599  if (Delta == 2)
1600  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1601  return getFalse(ITy);
1602  if (Delta == 1)
1603  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1604  return getFalse(ITy);
1605  }
1606 
1607  return nullptr;
1608 }
1609 
1611  const InstrInfoQuery &IIQ) {
1612  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1613  return X;
1614  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1615  return X;
1616 
1617  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1618  return X;
1619  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1620  return X;
1621 
1622  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1623  return X;
1624 
1625  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1626  return X;
1627 
1628  if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ))
1629  return X;
1630  if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, IIQ))
1631  return X;
1632 
1633  return nullptr;
1634 }
1635 
1637  const InstrInfoQuery &IIQ) {
1638  // (icmp (add V, C0), C1) | (icmp V, C0)
1639  ICmpInst::Predicate Pred0, Pred1;
1640  const APInt *C0, *C1;
1641  Value *V;
1642  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1643  return nullptr;
1644 
1645  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1646  return nullptr;
1647 
1648  auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1649  if (AddInst->getOperand(1) != Op1->getOperand(1))
1650  return nullptr;
1651 
1652  Type *ITy = Op0->getType();
1653  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1654  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1655 
1656  const APInt Delta = *C1 - *C0;
1657  if (C0->isStrictlyPositive()) {
1658  if (Delta == 2) {
1659  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1660  return getTrue(ITy);
1661  if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1662  return getTrue(ITy);
1663  }
1664  if (Delta == 1) {
1665  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1666  return getTrue(ITy);
1667  if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1668  return getTrue(ITy);
1669  }
1670  }
1671  if (C0->getBoolValue() && isNUW) {
1672  if (Delta == 2)
1673  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1674  return getTrue(ITy);
1675  if (Delta == 1)
1676  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1677  return getTrue(ITy);
1678  }
1679 
1680  return nullptr;
1681 }
1682 
1684  const InstrInfoQuery &IIQ) {
1685  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1686  return X;
1687  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1688  return X;
1689 
1690  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1691  return X;
1692  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1693  return X;
1694 
1695  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1696  return X;
1697 
1698  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1699  return X;
1700 
1701  if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ))
1702  return X;
1703  if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, IIQ))
1704  return X;
1705 
1706  return nullptr;
1707 }
1708 
1710  FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1711  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1712  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1713  if (LHS0->getType() != RHS0->getType())
1714  return nullptr;
1715 
1716  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1717  if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1718  (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1719  // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1720  // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1721  // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1722  // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1723  // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1724  // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1725  // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1726  // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1727  if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1728  (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1729  return RHS;
1730 
1731  // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1732  // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1733  // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1734  // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1735  // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1736  // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1737  // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1738  // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1739  if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1740  (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1741  return LHS;
1742  }
1743 
1744  return nullptr;
1745 }
1746 
1748  Value *Op0, Value *Op1, bool IsAnd) {
1749  // Look through casts of the 'and' operands to find compares.
1750  auto *Cast0 = dyn_cast<CastInst>(Op0);
1751  auto *Cast1 = dyn_cast<CastInst>(Op1);
1752  if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1753  Cast0->getSrcTy() == Cast1->getSrcTy()) {
1754  Op0 = Cast0->getOperand(0);
1755  Op1 = Cast1->getOperand(0);
1756  }
1757 
1758  Value *V = nullptr;
1759  auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1760  auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1761  if (ICmp0 && ICmp1)
1762  V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q.IIQ)
1763  : simplifyOrOfICmps(ICmp0, ICmp1, Q.IIQ);
1764 
1765  auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1766  auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1767  if (FCmp0 && FCmp1)
1768  V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1769 
1770  if (!V)
1771  return nullptr;
1772  if (!Cast0)
1773  return V;
1774 
1775  // If we looked through casts, we can only handle a constant simplification
1776  // because we are not allowed to create a cast instruction here.
1777  if (auto *C = dyn_cast<Constant>(V))
1778  return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1779 
1780  return nullptr;
1781 }
1782 
1783 /// Given operands for an And, see if we can fold the result.
1784 /// If not, this returns null.
1785 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1786  unsigned MaxRecurse) {
1787  if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1788  return C;
1789 
1790  // X & undef -> 0
1791  if (match(Op1, m_Undef()))
1792  return Constant::getNullValue(Op0->getType());
1793 
1794  // X & X = X
1795  if (Op0 == Op1)
1796  return Op0;
1797 
1798  // X & 0 = 0
1799  if (match(Op1, m_Zero()))
1800  return Constant::getNullValue(Op0->getType());
1801 
1802  // X & -1 = X
1803  if (match(Op1, m_AllOnes()))
1804  return Op0;
1805 
1806  // A & ~A = ~A & A = 0
1807  if (match(Op0, m_Not(m_Specific(Op1))) ||
1808  match(Op1, m_Not(m_Specific(Op0))))
1809  return Constant::getNullValue(Op0->getType());
1810 
1811  // (A | ?) & A = A
1812  if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1813  return Op1;
1814 
1815  // A & (A | ?) = A
1816  if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1817  return Op0;
1818 
1819  // A mask that only clears known zeros of a shifted value is a no-op.
1820  Value *X;
1821  const APInt *Mask;
1822  const APInt *ShAmt;
1823  if (match(Op1, m_APInt(Mask))) {
1824  // If all bits in the inverted and shifted mask are clear:
1825  // and (shl X, ShAmt), Mask --> shl X, ShAmt
1826  if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1827  (~(*Mask)).lshr(*ShAmt).isNullValue())
1828  return Op0;
1829 
1830  // If all bits in the inverted and shifted mask are clear:
1831  // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1832  if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1833  (~(*Mask)).shl(*ShAmt).isNullValue())
1834  return Op0;
1835  }
1836 
1837  // A & (-A) = A if A is a power of two or zero.
1838  if (match(Op0, m_Neg(m_Specific(Op1))) ||
1839  match(Op1, m_Neg(m_Specific(Op0)))) {
1840  if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1841  Q.DT))
1842  return Op0;
1843  if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1844  Q.DT))
1845  return Op1;
1846  }
1847 
1848  // This is a similar pattern used for checking if a value is a power-of-2:
1849  // (A - 1) & A --> 0 (if A is a power-of-2 or 0)
1850  // A & (A - 1) --> 0 (if A is a power-of-2 or 0)
1851  if (match(Op0, m_Add(m_Specific(Op1), m_AllOnes())) &&
1852  isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
1853  return Constant::getNullValue(Op1->getType());
1854  if (match(Op1, m_Add(m_Specific(Op0), m_AllOnes())) &&
1855  isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
1856  return Constant::getNullValue(Op0->getType());
1857 
1858  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
1859  return V;
1860 
1861  // Try some generic simplifications for associative operations.
1862  if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1863  MaxRecurse))
1864  return V;
1865 
1866  // And distributes over Or. Try some generic simplifications based on this.
1867  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1868  Q, MaxRecurse))
1869  return V;
1870 
1871  // And distributes over Xor. Try some generic simplifications based on this.
1872  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1873  Q, MaxRecurse))
1874  return V;
1875 
1876  // If the operation is with the result of a select instruction, check whether
1877  // operating on either branch of the select always yields the same value.
1878  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1879  if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1880  MaxRecurse))
1881  return V;
1882 
1883  // If the operation is with the result of a phi instruction, check whether
1884  // operating on all incoming values of the phi always yields the same value.
1885  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1886  if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1887  MaxRecurse))
1888  return V;
1889 
1890  // Assuming the effective width of Y is not larger than A, i.e. all bits
1891  // from X and Y are disjoint in (X << A) | Y,
1892  // if the mask of this AND op covers all bits of X or Y, while it covers
1893  // no bits from the other, we can bypass this AND op. E.g.,
1894  // ((X << A) | Y) & Mask -> Y,
1895  // if Mask = ((1 << effective_width_of(Y)) - 1)
1896  // ((X << A) | Y) & Mask -> X << A,
1897  // if Mask = ((1 << effective_width_of(X)) - 1) << A
1898  // SimplifyDemandedBits in InstCombine can optimize the general case.
1899  // This pattern aims to help other passes for a common case.
1900  Value *Y, *XShifted;
1901  if (match(Op1, m_APInt(Mask)) &&
1902  match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
1903  m_Value(XShifted)),
1904  m_Value(Y)))) {
1905  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1906  const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
1907  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1908  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1909  if (EffWidthY <= ShftCnt) {
1910  const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
1911  Q.DT);
1912  const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
1913  const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
1914  const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
1915  // If the mask is extracting all bits from X or Y as is, we can skip
1916  // this AND op.
1917  if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
1918  return Y;
1919  if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
1920  return XShifted;
1921  }
1922  }
1923 
1924  return nullptr;
1925 }
1926 
1929 }
1930 
1931 /// Given operands for an Or, see if we can fold the result.
1932 /// If not, this returns null.
1933 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1934  unsigned MaxRecurse) {
1935  if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1936  return C;
1937 
1938  // X | undef -> -1
1939  // X | -1 = -1
1940  // Do not return Op1 because it may contain undef elements if it's a vector.
1941  if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1942  return Constant::getAllOnesValue(Op0->getType());
1943 
1944  // X | X = X
1945  // X | 0 = X
1946  if (Op0 == Op1 || match(Op1, m_Zero()))
1947  return Op0;
1948 
1949  // A | ~A = ~A | A = -1
1950  if (match(Op0, m_Not(m_Specific(Op1))) ||
1951  match(Op1, m_Not(m_Specific(Op0))))
1952  return Constant::getAllOnesValue(Op0->getType());
1953 
1954  // (A & ?) | A = A
1955  if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1956  return Op1;
1957 
1958  // A | (A & ?) = A
1959  if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1960  return Op0;
1961 
1962  // ~(A & ?) | A = -1
1963  if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1964  return Constant::getAllOnesValue(Op1->getType());
1965 
1966  // A | ~(A & ?) = -1
1967  if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1968  return Constant::getAllOnesValue(Op0->getType());
1969 
1970  Value *A, *B;
1971  // (A & ~B) | (A ^ B) -> (A ^ B)
1972  // (~B & A) | (A ^ B) -> (A ^ B)
1973  // (A & ~B) | (B ^ A) -> (B ^ A)
1974  // (~B & A) | (B ^ A) -> (B ^ A)
1975  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1976  (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1977  match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1978  return Op1;
1979 
1980  // Commute the 'or' operands.
1981  // (A ^ B) | (A & ~B) -> (A ^ B)
1982  // (A ^ B) | (~B & A) -> (A ^ B)
1983  // (B ^ A) | (A & ~B) -> (B ^ A)
1984  // (B ^ A) | (~B & A) -> (B ^ A)
1985  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1986  (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1987  match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1988  return Op0;
1989 
1990  // (A & B) | (~A ^ B) -> (~A ^ B)
1991  // (B & A) | (~A ^ B) -> (~A ^ B)
1992  // (A & B) | (B ^ ~A) -> (B ^ ~A)
1993  // (B & A) | (B ^ ~A) -> (B ^ ~A)
1994  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1995  (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1996  match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1997  return Op1;
1998 
1999  // (~A ^ B) | (A & B) -> (~A ^ B)
2000  // (~A ^ B) | (B & A) -> (~A ^ B)
2001  // (B ^ ~A) | (A & B) -> (B ^ ~A)
2002  // (B ^ ~A) | (B & A) -> (B ^ ~A)
2003  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2004  (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2005  match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2006  return Op0;
2007 
2008  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
2009  return V;
2010 
2011  // Try some generic simplifications for associative operations.
2012  if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
2013  MaxRecurse))
2014  return V;
2015 
2016  // Or distributes over And. Try some generic simplifications based on this.
2017  if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2018  MaxRecurse))
2019  return V;
2020 
2021  // If the operation is with the result of a select instruction, check whether
2022  // operating on either branch of the select always yields the same value.
2023  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2024  if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2025  MaxRecurse))
2026  return V;
2027 
2028  // (A & C1)|(B & C2)
2029  const APInt *C1, *C2;
2030  if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2031  match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2032  if (*C1 == ~*C2) {
2033  // (A & C1)|(B & C2)
2034  // If we have: ((V + N) & C1) | (V & C2)
2035  // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2036  // replace with V+N.
2037  Value *N;
2038  if (C2->isMask() && // C2 == 0+1+
2039  match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2040  // Add commutes, try both ways.
2041  if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2042  return A;
2043  }
2044  // Or commutes, try both ways.
2045  if (C1->isMask() &&
2046  match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2047  // Add commutes, try both ways.
2048  if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2049  return B;
2050  }
2051  }
2052  }
2053 
2054  // If the operation is with the result of a phi instruction, check whether
2055  // operating on all incoming values of the phi always yields the same value.
2056  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2057  if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2058  return V;
2059 
2060  return nullptr;
2061 }
2062 
2065 }
2066 
2067 /// Given operands for a Xor, see if we can fold the result.
2068 /// If not, this returns null.
2069 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2070  unsigned MaxRecurse) {
2071  if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2072  return C;
2073 
2074  // A ^ undef -> undef
2075  if (match(Op1, m_Undef()))
2076  return Op1;
2077 
2078  // A ^ 0 = A
2079  if (match(Op1, m_Zero()))
2080  return Op0;
2081 
2082  // A ^ A = 0
2083  if (Op0 == Op1)
2084  return Constant::getNullValue(Op0->getType());
2085 
2086  // A ^ ~A = ~A ^ A = -1
2087  if (match(Op0, m_Not(m_Specific(Op1))) ||
2088  match(Op1, m_Not(m_Specific(Op0))))
2089  return Constant::getAllOnesValue(Op0->getType());
2090 
2091  // Try some generic simplifications for associative operations.
2092  if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2093  MaxRecurse))
2094  return V;
2095 
2096  // Threading Xor over selects and phi nodes is pointless, so don't bother.
2097  // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2098  // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2099  // only if B and C are equal. If B and C are equal then (since we assume
2100  // that operands have already been simplified) "select(cond, B, C)" should
2101  // have been simplified to the common value of B and C already. Analysing
2102  // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2103  // for threading over phi nodes.
2104 
2105  return nullptr;
2106 }
2107 
2110 }
2111 
2112 
2114  return CmpInst::makeCmpResultType(Op->getType());
2115 }
2116 
2117 /// Rummage around inside V looking for something equivalent to the comparison
2118 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2119 /// Helper function for analyzing max/min idioms.
2121  Value *LHS, Value *RHS) {
2123  if (!SI)
2124  return nullptr;
2125  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2126  if (!Cmp)
2127  return nullptr;
2128  Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2129  if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2130  return Cmp;
2131  if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2132  LHS == CmpRHS && RHS == CmpLHS)
2133  return Cmp;
2134  return nullptr;
2135 }
2136 
2137 // A significant optimization not implemented here is assuming that alloca
2138 // addresses are not equal to incoming argument values. They don't *alias*,
2139 // as we say, but that doesn't mean they aren't equal, so we take a
2140 // conservative approach.
2141 //
2142 // This is inspired in part by C++11 5.10p1:
2143 // "Two pointers of the same type compare equal if and only if they are both
2144 // null, both point to the same function, or both represent the same
2145 // address."
2146 //
2147 // This is pretty permissive.
2148 //
2149 // It's also partly due to C11 6.5.9p6:
2150 // "Two pointers compare equal if and only if both are null pointers, both are
2151 // pointers to the same object (including a pointer to an object and a
2152 // subobject at its beginning) or function, both are pointers to one past the
2153 // last element of the same array object, or one is a pointer to one past the
2154 // end of one array object and the other is a pointer to the start of a
2155 // different array object that happens to immediately follow the first array
2156 // object in the address space.)
2157 //
2158 // C11's version is more restrictive, however there's no reason why an argument
2159 // couldn't be a one-past-the-end value for a stack object in the caller and be
2160 // equal to the beginning of a stack object in the callee.
2161 //
2162 // If the C and C++ standards are ever made sufficiently restrictive in this
2163 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2164 // this optimization.
2165 static Constant *
2167  const DominatorTree *DT, CmpInst::Predicate Pred,
2168  AssumptionCache *AC, const Instruction *CxtI,
2169  const InstrInfoQuery &IIQ, Value *LHS, Value *RHS) {
2170  // First, skip past any trivial no-ops.
2171  LHS = LHS->stripPointerCasts();
2172  RHS = RHS->stripPointerCasts();
2173 
2174  // A non-null pointer is not equal to a null pointer.
2175  if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2176  IIQ.UseInstrInfo) &&
2177  isa<ConstantPointerNull>(RHS) &&
2178  (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2179  return ConstantInt::get(GetCompareTy(LHS),
2180  !CmpInst::isTrueWhenEqual(Pred));
2181 
2182  // We can only fold certain predicates on pointer comparisons.
2183  switch (Pred) {
2184  default:
2185  return nullptr;
2186 
2187  // Equality comaprisons are easy to fold.
2188  case CmpInst::ICMP_EQ:
2189  case CmpInst::ICMP_NE:
2190  break;
2191 
2192  // We can only handle unsigned relational comparisons because 'inbounds' on
2193  // a GEP only protects against unsigned wrapping.
2194  case CmpInst::ICMP_UGT:
2195  case CmpInst::ICMP_UGE:
2196  case CmpInst::ICMP_ULT:
2197  case CmpInst::ICMP_ULE:
2198  // However, we have to switch them to their signed variants to handle
2199  // negative indices from the base pointer.
2200  Pred = ICmpInst::getSignedPredicate(Pred);
2201  break;
2202  }
2203 
2204  // Strip off any constant offsets so that we can reason about them.
2205  // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2206  // here and compare base addresses like AliasAnalysis does, however there are
2207  // numerous hazards. AliasAnalysis and its utilities rely on special rules
2208  // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2209  // doesn't need to guarantee pointer inequality when it says NoAlias.
2210  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2211  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2212 
2213  // If LHS and RHS are related via constant offsets to the same base
2214  // value, we can replace it with an icmp which just compares the offsets.
2215  if (LHS == RHS)
2216  return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2217 
2218  // Various optimizations for (in)equality comparisons.
2219  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2220  // Different non-empty allocations that exist at the same time have
2221  // different addresses (if the program can tell). Global variables always
2222  // exist, so they always exist during the lifetime of each other and all
2223  // allocas. Two different allocas usually have different addresses...
2224  //
2225  // However, if there's an @llvm.stackrestore dynamically in between two
2226  // allocas, they may have the same address. It's tempting to reduce the
2227  // scope of the problem by only looking at *static* allocas here. That would
2228  // cover the majority of allocas while significantly reducing the likelihood
2229  // of having an @llvm.stackrestore pop up in the middle. However, it's not
2230  // actually impossible for an @llvm.stackrestore to pop up in the middle of
2231  // an entry block. Also, if we have a block that's not attached to a
2232  // function, we can't tell if it's "static" under the current definition.
2233  // Theoretically, this problem could be fixed by creating a new kind of
2234  // instruction kind specifically for static allocas. Such a new instruction
2235  // could be required to be at the top of the entry block, thus preventing it
2236  // from being subject to a @llvm.stackrestore. Instcombine could even
2237  // convert regular allocas into these special allocas. It'd be nifty.
2238  // However, until then, this problem remains open.
2239  //
2240  // So, we'll assume that two non-empty allocas have different addresses
2241  // for now.
2242  //
2243  // With all that, if the offsets are within the bounds of their allocations
2244  // (and not one-past-the-end! so we can't use inbounds!), and their
2245  // allocations aren't the same, the pointers are not equal.
2246  //
2247  // Note that it's not necessary to check for LHS being a global variable
2248  // address, due to canonicalization and constant folding.
2249  if (isa<AllocaInst>(LHS) &&
2250  (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2251  ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2252  ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2253  uint64_t LHSSize, RHSSize;
2254  ObjectSizeOpts Opts;
2255  Opts.NullIsUnknownSize =
2256  NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2257  if (LHSOffsetCI && RHSOffsetCI &&
2258  getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2259  getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2260  const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2261  const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2262  if (!LHSOffsetValue.isNegative() &&
2263  !RHSOffsetValue.isNegative() &&
2264  LHSOffsetValue.ult(LHSSize) &&
2265  RHSOffsetValue.ult(RHSSize)) {
2266  return ConstantInt::get(GetCompareTy(LHS),
2267  !CmpInst::isTrueWhenEqual(Pred));
2268  }
2269  }
2270 
2271  // Repeat the above check but this time without depending on DataLayout
2272  // or being able to compute a precise size.
2273  if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2274  !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2275  LHSOffset->isNullValue() &&
2276  RHSOffset->isNullValue())
2277  return ConstantInt::get(GetCompareTy(LHS),
2278  !CmpInst::isTrueWhenEqual(Pred));
2279  }
2280 
2281  // Even if an non-inbounds GEP occurs along the path we can still optimize
2282  // equality comparisons concerning the result. We avoid walking the whole
2283  // chain again by starting where the last calls to
2284  // stripAndComputeConstantOffsets left off and accumulate the offsets.
2285  Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2286  Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2287  if (LHS == RHS)
2288  return ConstantExpr::getICmp(Pred,
2289  ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2290  ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2291 
2292  // If one side of the equality comparison must come from a noalias call
2293  // (meaning a system memory allocation function), and the other side must
2294  // come from a pointer that cannot overlap with dynamically-allocated
2295  // memory within the lifetime of the current function (allocas, byval
2296  // arguments, globals), then determine the comparison result here.
2297  SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2298  GetUnderlyingObjects(LHS, LHSUObjs, DL);
2299  GetUnderlyingObjects(RHS, RHSUObjs, DL);
2300 
2301  // Is the set of underlying objects all noalias calls?
2302  auto IsNAC = [](ArrayRef<const Value *> Objects) {
2303  return all_of(Objects, isNoAliasCall);
2304  };
2305 
2306  // Is the set of underlying objects all things which must be disjoint from
2307  // noalias calls. For allocas, we consider only static ones (dynamic
2308  // allocas might be transformed into calls to malloc not simultaneously
2309  // live with the compared-to allocation). For globals, we exclude symbols
2310  // that might be resolve lazily to symbols in another dynamically-loaded
2311  // library (and, thus, could be malloc'ed by the implementation).
2312  auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2313  return all_of(Objects, [](const Value *V) {
2314  if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2315  return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2316  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2317  return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2318  GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2319  !GV->isThreadLocal();
2320  if (const Argument *A = dyn_cast<Argument>(V))
2321  return A->hasByValAttr();
2322  return false;
2323  });
2324  };
2325 
2326  if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2327  (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2328  return ConstantInt::get(GetCompareTy(LHS),
2329  !CmpInst::isTrueWhenEqual(Pred));
2330 
2331  // Fold comparisons for non-escaping pointer even if the allocation call
2332  // cannot be elided. We cannot fold malloc comparison to null. Also, the
2333  // dynamic allocation call could be either of the operands.
2334  Value *MI = nullptr;
2335  if (isAllocLikeFn(LHS, TLI) &&
2336  llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2337  MI = LHS;
2338  else if (isAllocLikeFn(RHS, TLI) &&
2339  llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2340  MI = RHS;
2341  // FIXME: We should also fold the compare when the pointer escapes, but the
2342  // compare dominates the pointer escape
2343  if (MI && !PointerMayBeCaptured(MI, true, true))
2344  return ConstantInt::get(GetCompareTy(LHS),
2346  }
2347 
2348  // Otherwise, fail.
2349  return nullptr;
2350 }
2351 
2352 /// Fold an icmp when its operands have i1 scalar type.
2354  Value *RHS, const SimplifyQuery &Q) {
2355  Type *ITy = GetCompareTy(LHS); // The return type.
2356  Type *OpTy = LHS->getType(); // The operand type.
2357  if (!OpTy->isIntOrIntVectorTy(1))
2358  return nullptr;
2359 
2360  // A boolean compared to true/false can be simplified in 14 out of the 20
2361  // (10 predicates * 2 constants) possible combinations. Cases not handled here
2362  // require a 'not' of the LHS, so those must be transformed in InstCombine.
2363  if (match(RHS, m_Zero())) {
2364  switch (Pred) {
2365  case CmpInst::ICMP_NE: // X != 0 -> X
2366  case CmpInst::ICMP_UGT: // X >u 0 -> X
2367  case CmpInst::ICMP_SLT: // X <s 0 -> X
2368  return LHS;
2369 
2370  case CmpInst::ICMP_ULT: // X <u 0 -> false
2371  case CmpInst::ICMP_SGT: // X >s 0 -> false
2372  return getFalse(ITy);
2373 
2374  case CmpInst::ICMP_UGE: // X >=u 0 -> true
2375  case CmpInst::ICMP_SLE: // X <=s 0 -> true
2376  return getTrue(ITy);
2377 
2378  default: break;
2379  }
2380  } else if (match(RHS, m_One())) {
2381  switch (Pred) {
2382  case CmpInst::ICMP_EQ: // X == 1 -> X
2383  case CmpInst::ICMP_UGE: // X >=u 1 -> X
2384  case CmpInst::ICMP_SLE: // X <=s -1 -> X
2385  return LHS;
2386 
2387  case CmpInst::ICMP_UGT: // X >u 1 -> false
2388  case CmpInst::ICMP_SLT: // X <s -1 -> false
2389  return getFalse(ITy);
2390 
2391  case CmpInst::ICMP_ULE: // X <=u 1 -> true
2392  case CmpInst::ICMP_SGE: // X >=s -1 -> true
2393  return getTrue(ITy);
2394 
2395  default: break;
2396  }
2397  }
2398 
2399  switch (Pred) {
2400  default:
2401  break;
2402  case ICmpInst::ICMP_UGE:
2403  if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2404  return getTrue(ITy);
2405  break;
2406  case ICmpInst::ICMP_SGE:
2407  /// For signed comparison, the values for an i1 are 0 and -1
2408  /// respectively. This maps into a truth table of:
2409  /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2410  /// 0 | 0 | 1 (0 >= 0) | 1
2411  /// 0 | 1 | 1 (0 >= -1) | 1
2412  /// 1 | 0 | 0 (-1 >= 0) | 0
2413  /// 1 | 1 | 1 (-1 >= -1) | 1
2414  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2415  return getTrue(ITy);
2416  break;
2417  case ICmpInst::ICMP_ULE:
2418  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2419  return getTrue(ITy);
2420  break;
2421  }
2422 
2423  return nullptr;
2424 }
2425 
2426 /// Try hard to fold icmp with zero RHS because this is a common case.
2428  Value *RHS, const SimplifyQuery &Q) {
2429  if (!match(RHS, m_Zero()))
2430  return nullptr;
2431 
2432  Type *ITy = GetCompareTy(LHS); // The return type.
2433  switch (Pred) {
2434  default:
2435  llvm_unreachable("Unknown ICmp predicate!");
2436  case ICmpInst::ICMP_ULT:
2437  return getFalse(ITy);
2438  case ICmpInst::ICMP_UGE:
2439  return getTrue(ITy);
2440  case ICmpInst::ICMP_EQ:
2441  case ICmpInst::ICMP_ULE:
2442  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2443  return getFalse(ITy);
2444  break;
2445  case ICmpInst::ICMP_NE:
2446  case ICmpInst::ICMP_UGT:
2447  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2448  return getTrue(ITy);
2449  break;
2450  case ICmpInst::ICMP_SLT: {
2451  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2452  if (LHSKnown.isNegative())
2453  return getTrue(ITy);
2454  if (LHSKnown.isNonNegative())
2455  return getFalse(ITy);
2456  break;
2457  }
2458  case ICmpInst::ICMP_SLE: {
2459  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2460  if (LHSKnown.isNegative())
2461  return getTrue(ITy);
2462  if (LHSKnown.isNonNegative() &&
2463  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2464  return getFalse(ITy);
2465  break;
2466  }
2467  case ICmpInst::ICMP_SGE: {
2468  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2469  if (LHSKnown.isNegative())
2470  return getFalse(ITy);
2471  if (LHSKnown.isNonNegative())
2472  return getTrue(ITy);
2473  break;
2474  }
2475  case ICmpInst::ICMP_SGT: {
2476  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2477  if (LHSKnown.isNegative())
2478  return getFalse(ITy);
2479  if (LHSKnown.isNonNegative() &&
2480  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2481  return getTrue(ITy);
2482  break;
2483  }
2484  }
2485 
2486  return nullptr;
2487 }
2488 
2490  Value *RHS, const InstrInfoQuery &IIQ) {
2491  Type *ITy = GetCompareTy(RHS); // The return type.
2492 
2493  Value *X;
2494  // Sign-bit checks can be optimized to true/false after unsigned
2495  // floating-point casts:
2496  // icmp slt (bitcast (uitofp X)), 0 --> false
2497  // icmp sgt (bitcast (uitofp X)), -1 --> true
2498  if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2499  if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2500  return ConstantInt::getFalse(ITy);
2501  if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2502  return ConstantInt::getTrue(ITy);
2503  }
2504 
2505  const APInt *C;
2506  if (!match(RHS, m_APInt(C)))
2507  return nullptr;
2508 
2509  // Rule out tautological comparisons (eg., ult 0 or uge 0).
2511  if (RHS_CR.isEmptySet())
2512  return ConstantInt::getFalse(ITy);
2513  if (RHS_CR.isFullSet())
2514  return ConstantInt::getTrue(ITy);
2515 
2516  ConstantRange LHS_CR = computeConstantRange(LHS, IIQ.UseInstrInfo);
2517  if (!LHS_CR.isFullSet()) {
2518  if (RHS_CR.contains(LHS_CR))
2519  return ConstantInt::getTrue(ITy);
2520  if (RHS_CR.inverse().contains(LHS_CR))
2521  return ConstantInt::getFalse(ITy);
2522  }
2523 
2524  return nullptr;
2525 }
2526 
2527 /// TODO: A large part of this logic is duplicated in InstCombine's
2528 /// foldICmpBinOp(). We should be able to share that and avoid the code
2529 /// duplication.
2531  Value *RHS, const SimplifyQuery &Q,
2532  unsigned MaxRecurse) {
2533  Type *ITy = GetCompareTy(LHS); // The return type.
2534 
2535  BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2536  BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2537  if (MaxRecurse && (LBO || RBO)) {
2538  // Analyze the case when either LHS or RHS is an add instruction.
2539  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2540  // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2541  bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2542  if (LBO && LBO->getOpcode() == Instruction::Add) {
2543  A = LBO->getOperand(0);
2544  B = LBO->getOperand(1);
2545  NoLHSWrapProblem =
2546  ICmpInst::isEquality(Pred) ||
2547  (CmpInst::isUnsigned(Pred) &&
2548  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2549  (CmpInst::isSigned(Pred) &&
2550  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2551  }
2552  if (RBO && RBO->getOpcode() == Instruction::Add) {
2553  C = RBO->getOperand(0);
2554  D = RBO->getOperand(1);
2555  NoRHSWrapProblem =
2556  ICmpInst::isEquality(Pred) ||
2557  (CmpInst::isUnsigned(Pred) &&
2558  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2559  (CmpInst::isSigned(Pred) &&
2560  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2561  }
2562 
2563  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2564  if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2565  if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2566  Constant::getNullValue(RHS->getType()), Q,
2567  MaxRecurse - 1))
2568  return V;
2569 
2570  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2571  if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2572  if (Value *V =
2574  C == LHS ? D : C, Q, MaxRecurse - 1))
2575  return V;
2576 
2577  // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2578  if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2579  NoRHSWrapProblem) {
2580  // Determine Y and Z in the form icmp (X+Y), (X+Z).
2581  Value *Y, *Z;
2582  if (A == C) {
2583  // C + B == C + D -> B == D
2584  Y = B;
2585  Z = D;
2586  } else if (A == D) {
2587  // D + B == C + D -> B == C
2588  Y = B;
2589  Z = C;
2590  } else if (B == C) {
2591  // A + C == C + D -> A == D
2592  Y = A;
2593  Z = D;
2594  } else {
2595  assert(B == D);
2596  // A + D == C + D -> A == C
2597  Y = A;
2598  Z = C;
2599  }
2600  if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2601  return V;
2602  }
2603  }
2604 
2605  {
2606  Value *Y = nullptr;
2607  // icmp pred (or X, Y), X
2608  if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2609  if (Pred == ICmpInst::ICMP_ULT)
2610  return getFalse(ITy);
2611  if (Pred == ICmpInst::ICMP_UGE)
2612  return getTrue(ITy);
2613 
2614  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2615  KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2616  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2617  if (RHSKnown.isNonNegative() && YKnown.isNegative())
2618  return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2619  if (RHSKnown.isNegative() || YKnown.isNonNegative())
2620  return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2621  }
2622  }
2623  // icmp pred X, (or X, Y)
2624  if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2625  if (Pred == ICmpInst::ICMP_ULE)
2626  return getTrue(ITy);
2627  if (Pred == ICmpInst::ICMP_UGT)
2628  return getFalse(ITy);
2629 
2630  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2631  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2632  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2633  if (LHSKnown.isNonNegative() && YKnown.isNegative())
2634  return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2635  if (LHSKnown.isNegative() || YKnown.isNonNegative())
2636  return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2637  }
2638  }
2639  }
2640 
2641  // icmp pred (and X, Y), X
2642  if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2643  if (Pred == ICmpInst::ICMP_UGT)
2644  return getFalse(ITy);
2645  if (Pred == ICmpInst::ICMP_ULE)
2646  return getTrue(ITy);
2647  }
2648  // icmp pred X, (and X, Y)
2649  if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2650  if (Pred == ICmpInst::ICMP_UGE)
2651  return getTrue(ITy);
2652  if (Pred == ICmpInst::ICMP_ULT)
2653  return getFalse(ITy);
2654  }
2655 
2656  // 0 - (zext X) pred C
2657  if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2658  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2659  if (RHSC->getValue().isStrictlyPositive()) {
2660  if (Pred == ICmpInst::ICMP_SLT)
2661  return ConstantInt::getTrue(RHSC->getContext());
2662  if (Pred == ICmpInst::ICMP_SGE)
2663  return ConstantInt::getFalse(RHSC->getContext());
2664  if (Pred == ICmpInst::ICMP_EQ)
2665  return ConstantInt::getFalse(RHSC->getContext());
2666  if (Pred == ICmpInst::ICMP_NE)
2667  return ConstantInt::getTrue(RHSC->getContext());
2668  }
2669  if (RHSC->getValue().isNonNegative()) {
2670  if (Pred == ICmpInst::ICMP_SLE)
2671  return ConstantInt::getTrue(RHSC->getContext());
2672  if (Pred == ICmpInst::ICMP_SGT)
2673  return ConstantInt::getFalse(RHSC->getContext());
2674  }
2675  }
2676  }
2677 
2678  // icmp pred (urem X, Y), Y
2679  if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2680  switch (Pred) {
2681  default:
2682  break;
2683  case ICmpInst::ICMP_SGT:
2684  case ICmpInst::ICMP_SGE: {
2685  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2686  if (!Known.isNonNegative())
2687  break;
2689  }
2690  case ICmpInst::ICMP_EQ:
2691  case ICmpInst::ICMP_UGT:
2692  case ICmpInst::ICMP_UGE:
2693  return getFalse(ITy);
2694  case ICmpInst::ICMP_SLT:
2695  case ICmpInst::ICMP_SLE: {
2696  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2697  if (!Known.isNonNegative())
2698  break;
2700  }
2701  case ICmpInst::ICMP_NE:
2702  case ICmpInst::ICMP_ULT:
2703  case ICmpInst::ICMP_ULE:
2704  return getTrue(ITy);
2705  }
2706  }
2707 
2708  // icmp pred X, (urem Y, X)
2709  if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2710  switch (Pred) {
2711  default:
2712  break;
2713  case ICmpInst::ICMP_SGT:
2714  case ICmpInst::ICMP_SGE: {
2715  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2716  if (!Known.isNonNegative())
2717  break;
2719  }
2720  case ICmpInst::ICMP_NE:
2721  case ICmpInst::ICMP_UGT:
2722  case ICmpInst::ICMP_UGE:
2723  return getTrue(ITy);
2724  case ICmpInst::ICMP_SLT:
2725  case ICmpInst::ICMP_SLE: {
2726  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2727  if (!Known.isNonNegative())
2728  break;
2730  }
2731  case ICmpInst::ICMP_EQ:
2732  case ICmpInst::ICMP_ULT:
2733  case ICmpInst::ICMP_ULE:
2734  return getFalse(ITy);
2735  }
2736  }
2737 
2738  // x >> y <=u x
2739  // x udiv y <=u x.
2740  if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2741  match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2742  // icmp pred (X op Y), X
2743  if (Pred == ICmpInst::ICMP_UGT)
2744  return getFalse(ITy);
2745  if (Pred == ICmpInst::ICMP_ULE)
2746  return getTrue(ITy);
2747  }
2748 
2749  // x >=u x >> y
2750  // x >=u x udiv y.
2751  if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2752  match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2753  // icmp pred X, (X op Y)
2754  if (Pred == ICmpInst::ICMP_ULT)
2755  return getFalse(ITy);
2756  if (Pred == ICmpInst::ICMP_UGE)
2757  return getTrue(ITy);
2758  }
2759 
2760  // handle:
2761  // CI2 << X == CI
2762  // CI2 << X != CI
2763  //
2764  // where CI2 is a power of 2 and CI isn't
2765  if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2766  const APInt *CI2Val, *CIVal = &CI->getValue();
2767  if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2768  CI2Val->isPowerOf2()) {
2769  if (!CIVal->isPowerOf2()) {
2770  // CI2 << X can equal zero in some circumstances,
2771  // this simplification is unsafe if CI is zero.
2772  //
2773  // We know it is safe if:
2774  // - The shift is nsw, we can't shift out the one bit.
2775  // - The shift is nuw, we can't shift out the one bit.
2776  // - CI2 is one
2777  // - CI isn't zero
2778  if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2779  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2780  CI2Val->isOneValue() || !CI->isZero()) {
2781  if (Pred == ICmpInst::ICMP_EQ)
2782  return ConstantInt::getFalse(RHS->getContext());
2783  if (Pred == ICmpInst::ICMP_NE)
2784  return ConstantInt::getTrue(RHS->getContext());
2785  }
2786  }
2787  if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2788  if (Pred == ICmpInst::ICMP_UGT)
2789  return ConstantInt::getFalse(RHS->getContext());
2790  if (Pred == ICmpInst::ICMP_ULE)
2791  return ConstantInt::getTrue(RHS->getContext());
2792  }
2793  }
2794  }
2795 
2796  if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2797  LBO->getOperand(1) == RBO->getOperand(1)) {
2798  switch (LBO->getOpcode()) {
2799  default:
2800  break;
2801  case Instruction::UDiv:
2802  case Instruction::LShr:
2803  if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2804  !Q.IIQ.isExact(RBO))
2805  break;
2806  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2807  RBO->getOperand(0), Q, MaxRecurse - 1))
2808  return V;
2809  break;
2810  case Instruction::SDiv:
2811  if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2812  !Q.IIQ.isExact(RBO))
2813  break;
2814  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2815  RBO->getOperand(0), Q, MaxRecurse - 1))
2816  return V;
2817  break;
2818  case Instruction::AShr:
2819  if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2820  break;
2821  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2822  RBO->getOperand(0), Q, MaxRecurse - 1))
2823  return V;
2824  break;
2825  case Instruction::Shl: {
2826  bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
2827  bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
2828  if (!NUW && !NSW)
2829  break;
2830  if (!NSW && ICmpInst::isSigned(Pred))
2831  break;
2832  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2833  RBO->getOperand(0), Q, MaxRecurse - 1))
2834  return V;
2835  break;
2836  }
2837  }
2838  }
2839  return nullptr;
2840 }
2841 
2842 /// Simplify integer comparisons where at least one operand of the compare
2843 /// matches an integer min/max idiom.
2845  Value *RHS, const SimplifyQuery &Q,
2846  unsigned MaxRecurse) {
2847  Type *ITy = GetCompareTy(LHS); // The return type.
2848  Value *A, *B;
2850  CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2851 
2852  // Signed variants on "max(a,b)>=a -> true".
2853  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2854  if (A != RHS)
2855  std::swap(A, B); // smax(A, B) pred A.
2856  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2857  // We analyze this as smax(A, B) pred A.
2858  P = Pred;
2859  } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2860  (A == LHS || B == LHS)) {
2861  if (A != LHS)
2862  std::swap(A, B); // A pred smax(A, B).
2863  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2864  // We analyze this as smax(A, B) swapped-pred A.
2865  P = CmpInst::getSwappedPredicate(Pred);
2866  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2867  (A == RHS || B == RHS)) {
2868  if (A != RHS)
2869  std::swap(A, B); // smin(A, B) pred A.
2870  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2871  // We analyze this as smax(-A, -B) swapped-pred -A.
2872  // Note that we do not need to actually form -A or -B thanks to EqP.
2873  P = CmpInst::getSwappedPredicate(Pred);
2874  } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2875  (A == LHS || B == LHS)) {
2876  if (A != LHS)
2877  std::swap(A, B); // A pred smin(A, B).
2878  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2879  // We analyze this as smax(-A, -B) pred -A.
2880  // Note that we do not need to actually form -A or -B thanks to EqP.
2881  P = Pred;
2882  }
2883  if (P != CmpInst::BAD_ICMP_PREDICATE) {
2884  // Cases correspond to "max(A, B) p A".
2885  switch (P) {
2886  default:
2887  break;
2888  case CmpInst::ICMP_EQ:
2889  case CmpInst::ICMP_SLE:
2890  // Equivalent to "A EqP B". This may be the same as the condition tested
2891  // in the max/min; if so, we can just return that.
2892  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2893  return V;
2894  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2895  return V;
2896  // Otherwise, see if "A EqP B" simplifies.
2897  if (MaxRecurse)
2898  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2899  return V;
2900  break;
2901  case CmpInst::ICMP_NE:
2902  case CmpInst::ICMP_SGT: {
2904  // Equivalent to "A InvEqP B". This may be the same as the condition
2905  // tested in the max/min; if so, we can just return that.
2906  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2907  return V;
2908  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2909  return V;
2910  // Otherwise, see if "A InvEqP B" simplifies.
2911  if (MaxRecurse)
2912  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2913  return V;
2914  break;
2915  }
2916  case CmpInst::ICMP_SGE:
2917  // Always true.
2918  return getTrue(ITy);
2919  case CmpInst::ICMP_SLT:
2920  // Always false.
2921  return getFalse(ITy);
2922  }
2923  }
2924 
2925  // Unsigned variants on "max(a,b)>=a -> true".
2927  if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2928  if (A != RHS)
2929  std::swap(A, B); // umax(A, B) pred A.
2930  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2931  // We analyze this as umax(A, B) pred A.
2932  P = Pred;
2933  } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2934  (A == LHS || B == LHS)) {
2935  if (A != LHS)
2936  std::swap(A, B); // A pred umax(A, B).
2937  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2938  // We analyze this as umax(A, B) swapped-pred A.
2939  P = CmpInst::getSwappedPredicate(Pred);
2940  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2941  (A == RHS || B == RHS)) {
2942  if (A != RHS)
2943  std::swap(A, B); // umin(A, B) pred A.
2944  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2945  // We analyze this as umax(-A, -B) swapped-pred -A.
2946  // Note that we do not need to actually form -A or -B thanks to EqP.
2947  P = CmpInst::getSwappedPredicate(Pred);
2948  } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2949  (A == LHS || B == LHS)) {
2950  if (A != LHS)
2951  std::swap(A, B); // A pred umin(A, B).
2952  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2953  // We analyze this as umax(-A, -B) pred -A.
2954  // Note that we do not need to actually form -A or -B thanks to EqP.
2955  P = Pred;
2956  }
2957  if (P != CmpInst::BAD_ICMP_PREDICATE) {
2958  // Cases correspond to "max(A, B) p A".
2959  switch (P) {
2960  default:
2961  break;
2962  case CmpInst::ICMP_EQ:
2963  case CmpInst::ICMP_ULE:
2964  // Equivalent to "A EqP B". This may be the same as the condition tested
2965  // in the max/min; if so, we can just return that.
2966  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2967  return V;
2968  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2969  return V;
2970  // Otherwise, see if "A EqP B" simplifies.
2971  if (MaxRecurse)
2972  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2973  return V;
2974  break;
2975  case CmpInst::ICMP_NE:
2976  case CmpInst::ICMP_UGT: {
2978  // Equivalent to "A InvEqP B". This may be the same as the condition
2979  // tested in the max/min; if so, we can just return that.
2980  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2981  return V;
2982  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2983  return V;
2984  // Otherwise, see if "A InvEqP B" simplifies.
2985  if (MaxRecurse)
2986  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2987  return V;
2988  break;
2989  }
2990  case CmpInst::ICMP_UGE:
2991  // Always true.
2992  return getTrue(ITy);
2993  case CmpInst::ICMP_ULT:
2994  // Always false.
2995  return getFalse(ITy);
2996  }
2997  }
2998 
2999  // Variants on "max(x,y) >= min(x,z)".
3000  Value *C, *D;
3001  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3002  match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3003  (A == C || A == D || B == C || B == D)) {
3004  // max(x, ?) pred min(x, ?).
3005  if (Pred == CmpInst::ICMP_SGE)
3006  // Always true.
3007  return getTrue(ITy);
3008  if (Pred == CmpInst::ICMP_SLT)
3009  // Always false.
3010  return getFalse(ITy);
3011  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3012  match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3013  (A == C || A == D || B == C || B == D)) {
3014  // min(x, ?) pred max(x, ?).
3015  if (Pred == CmpInst::ICMP_SLE)
3016  // Always true.
3017  return getTrue(ITy);
3018  if (Pred == CmpInst::ICMP_SGT)
3019  // Always false.
3020  return getFalse(ITy);
3021  } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3022  match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3023  (A == C || A == D || B == C || B == D)) {
3024  // max(x, ?) pred min(x, ?).
3025  if (Pred == CmpInst::ICMP_UGE)
3026  // Always true.
3027  return getTrue(ITy);
3028  if (Pred == CmpInst::ICMP_ULT)
3029  // Always false.
3030  return getFalse(ITy);
3031  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3032  match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3033  (A == C || A == D || B == C || B == D)) {
3034  // min(x, ?) pred max(x, ?).
3035  if (Pred == CmpInst::ICMP_ULE)
3036  // Always true.
3037  return getTrue(ITy);
3038  if (Pred == CmpInst::ICMP_UGT)
3039  // Always false.
3040  return getFalse(ITy);
3041  }
3042 
3043  return nullptr;
3044 }
3045 
3046 /// Given operands for an ICmpInst, see if we can fold the result.
3047 /// If not, this returns null.
3048 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3049  const SimplifyQuery &Q, unsigned MaxRecurse) {
3050  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3051  assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3052 
3053  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3054  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3055  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3056 
3057  // If we have a constant, make sure it is on the RHS.
3058  std::swap(LHS, RHS);
3059  Pred = CmpInst::getSwappedPredicate(Pred);
3060  }
3061  assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3062 
3063  Type *ITy = GetCompareTy(LHS); // The return type.
3064 
3065  // For EQ and NE, we can always pick a value for the undef to make the
3066  // predicate pass or fail, so we can return undef.
3067  // Matches behavior in llvm::ConstantFoldCompareInstruction.
3068  if (isa<UndefValue>(RHS) && ICmpInst::isEquality(Pred))
3069  return UndefValue::get(ITy);
3070 
3071  // icmp X, X -> true/false
3072  // icmp X, undef -> true/false because undef could be X.
3073  if (LHS == RHS || isa<UndefValue>(RHS))
3074  return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3075 
3076  if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3077  return V;
3078 
3079  if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3080  return V;
3081 
3082  if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3083  return V;
3084 
3085  // If both operands have range metadata, use the metadata
3086  // to simplify the comparison.
3087  if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3088  auto RHS_Instr = cast<Instruction>(RHS);
3089  auto LHS_Instr = cast<Instruction>(LHS);
3090 
3091  if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3092  Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3093  auto RHS_CR = getConstantRangeFromMetadata(
3094  *RHS_Instr->getMetadata(LLVMContext::MD_range));
3095  auto LHS_CR = getConstantRangeFromMetadata(
3096  *LHS_Instr->getMetadata(LLVMContext::MD_range));
3097 
3098  auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3099  if (Satisfied_CR.contains(LHS_CR))
3100  return ConstantInt::getTrue(RHS->getContext());
3101 
3102  auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3103  CmpInst::getInversePredicate(Pred), RHS_CR);
3104  if (InversedSatisfied_CR.contains(LHS_CR))
3105  return ConstantInt::getFalse(RHS->getContext());
3106  }
3107  }
3108 
3109  // Compare of cast, for example (zext X) != 0 -> X != 0
3110  if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3111  Instruction *LI = cast<CastInst>(LHS);
3112  Value *SrcOp = LI->getOperand(0);
3113  Type *SrcTy = SrcOp->getType();
3114  Type *DstTy = LI->getType();
3115 
3116  // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3117  // if the integer type is the same size as the pointer type.
3118  if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3119  Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3120  if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3121  // Transfer the cast to the constant.
3122  if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3123  ConstantExpr::getIntToPtr(RHSC, SrcTy),
3124  Q, MaxRecurse-1))
3125  return V;
3126  } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3127  if (RI->getOperand(0)->getType() == SrcTy)
3128  // Compare without the cast.
3129  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3130  Q, MaxRecurse-1))
3131  return V;
3132  }
3133  }
3134 
3135  if (isa<ZExtInst>(LHS)) {
3136  // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3137  // same type.
3138  if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3139  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3140  // Compare X and Y. Note that signed predicates become unsigned.
3142  SrcOp, RI->getOperand(0), Q,
3143  MaxRecurse-1))
3144  return V;
3145  }
3146  // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3147  // too. If not, then try to deduce the result of the comparison.
3148  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3149  // Compute the constant that would happen if we truncated to SrcTy then
3150  // reextended to DstTy.
3151  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3152  Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3153 
3154  // If the re-extended constant didn't change then this is effectively
3155  // also a case of comparing two zero-extended values.
3156  if (RExt == CI && MaxRecurse)
3158  SrcOp, Trunc, Q, MaxRecurse-1))
3159  return V;
3160 
3161  // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3162  // there. Use this to work out the result of the comparison.
3163  if (RExt != CI) {
3164  switch (Pred) {
3165  default: llvm_unreachable("Unknown ICmp predicate!");
3166  // LHS <u RHS.
3167  case ICmpInst::ICMP_EQ:
3168  case ICmpInst::ICMP_UGT:
3169  case ICmpInst::ICMP_UGE:
3170  return ConstantInt::getFalse(CI->getContext());
3171 
3172  case ICmpInst::ICMP_NE:
3173  case ICmpInst::ICMP_ULT:
3174  case ICmpInst::ICMP_ULE:
3175  return ConstantInt::getTrue(CI->getContext());
3176 
3177  // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3178  // is non-negative then LHS <s RHS.
3179  case ICmpInst::ICMP_SGT:
3180  case ICmpInst::ICMP_SGE:
3181  return CI->getValue().isNegative() ?
3182  ConstantInt::getTrue(CI->getContext()) :
3183  ConstantInt::getFalse(CI->getContext());
3184 
3185  case ICmpInst::ICMP_SLT:
3186  case ICmpInst::ICMP_SLE:
3187  return CI->getValue().isNegative() ?
3188  ConstantInt::getFalse(CI->getContext()) :
3189  ConstantInt::getTrue(CI->getContext());
3190  }
3191  }
3192  }
3193  }
3194 
3195  if (isa<SExtInst>(LHS)) {
3196  // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3197  // same type.
3198  if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3199  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3200  // Compare X and Y. Note that the predicate does not change.
3201  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3202  Q, MaxRecurse-1))
3203  return V;
3204  }
3205  // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3206  // too. If not, then try to deduce the result of the comparison.
3207  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3208  // Compute the constant that would happen if we truncated to SrcTy then
3209  // reextended to DstTy.
3210  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3211  Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3212 
3213  // If the re-extended constant didn't change then this is effectively
3214  // also a case of comparing two sign-extended values.
3215  if (RExt == CI && MaxRecurse)
3216  if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3217  return V;
3218 
3219  // Otherwise the upper bits of LHS are all equal, while RHS has varying
3220  // bits there. Use this to work out the result of the comparison.
3221  if (RExt != CI) {
3222  switch (Pred) {
3223  default: llvm_unreachable("Unknown ICmp predicate!");
3224  case ICmpInst::ICMP_EQ:
3225  return ConstantInt::getFalse(CI->getContext());
3226  case ICmpInst::ICMP_NE:
3227  return ConstantInt::getTrue(CI->getContext());
3228 
3229  // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3230  // LHS >s RHS.
3231  case ICmpInst::ICMP_SGT:
3232  case ICmpInst::ICMP_SGE:
3233  return CI->getValue().isNegative() ?
3234  ConstantInt::getTrue(CI->getContext()) :
3235  ConstantInt::getFalse(CI->getContext());
3236  case ICmpInst::ICMP_SLT:
3237  case ICmpInst::ICMP_SLE:
3238  return CI->getValue().isNegative() ?
3239  ConstantInt::getFalse(CI->getContext()) :
3240  ConstantInt::getTrue(CI->getContext());
3241 
3242  // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3243  // LHS >u RHS.
3244  case ICmpInst::ICMP_UGT:
3245  case ICmpInst::ICMP_UGE:
3246  // Comparison is true iff the LHS <s 0.
3247  if (MaxRecurse)
3248  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3249  Constant::getNullValue(SrcTy),
3250  Q, MaxRecurse-1))
3251  return V;
3252  break;
3253  case ICmpInst::ICMP_ULT:
3254  case ICmpInst::ICMP_ULE:
3255  // Comparison is true iff the LHS >=s 0.
3256  if (MaxRecurse)
3257  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3258  Constant::getNullValue(SrcTy),
3259  Q, MaxRecurse-1))
3260  return V;
3261  break;
3262  }
3263  }
3264  }
3265  }
3266  }
3267 
3268  // icmp eq|ne X, Y -> false|true if X != Y
3269  if (ICmpInst::isEquality(Pred) &&
3270  isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3271  return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3272  }
3273 
3274  if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3275  return V;
3276 
3277  if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3278  return V;
3279 
3280  // Simplify comparisons of related pointers using a powerful, recursive
3281  // GEP-walk when we have target data available..
3282  if (LHS->getType()->isPointerTy())
3283  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3284  Q.IIQ, LHS, RHS))
3285  return C;
3286  if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3287  if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3288  if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3289  Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3290  Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3291  Q.DL.getTypeSizeInBits(CRHS->getType()))
3292  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3293  Q.IIQ, CLHS->getPointerOperand(),
3294  CRHS->getPointerOperand()))
3295  return C;
3296 
3297  if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3298  if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3299  if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3300  GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3301  (ICmpInst::isEquality(Pred) ||
3302  (GLHS->isInBounds() && GRHS->isInBounds() &&
3303  Pred == ICmpInst::getSignedPredicate(Pred)))) {
3304  // The bases are equal and the indices are constant. Build a constant
3305  // expression GEP with the same indices and a null base pointer to see
3306  // what constant folding can make out of it.
3307  Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3308  SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3310  GLHS->getSourceElementType(), Null, IndicesLHS);
3311 
3312  SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3314  GLHS->getSourceElementType(), Null, IndicesRHS);
3315  return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3316  }
3317  }
3318  }
3319 
3320  // If the comparison is with the result of a select instruction, check whether
3321  // comparing with either branch of the select always yields the same value.
3322  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3323  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3324  return V;
3325 
3326  // If the comparison is with the result of a phi instruction, check whether
3327  // doing the compare with each incoming phi value yields a common result.
3328  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3329  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3330  return V;
3331 
3332  return nullptr;
3333 }
3334 
3335 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3336  const SimplifyQuery &Q) {
3337  return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3338 }
3339 
3340 /// Given operands for an FCmpInst, see if we can fold the result.
3341 /// If not, this returns null.
3342 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3343  FastMathFlags FMF, const SimplifyQuery &Q,
3344  unsigned MaxRecurse) {
3345  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3346  assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3347 
3348  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3349  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3350  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3351 
3352  // If we have a constant, make sure it is on the RHS.
3353  std::swap(LHS, RHS);
3354  Pred = CmpInst::getSwappedPredicate(Pred);
3355  }
3356 
3357  // Fold trivial predicates.
3358  Type *RetTy = GetCompareTy(LHS);
3359  if (Pred == FCmpInst::FCMP_FALSE)
3360  return getFalse(RetTy);
3361  if (Pred == FCmpInst::FCMP_TRUE)
3362  return getTrue(RetTy);
3363 
3364  // Fold (un)ordered comparison if we can determine there are no NaNs.
3365  if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3366  if (FMF.noNaNs() ||
3367  (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3368  return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3369 
3370  // NaN is unordered; NaN is not ordered.
3372  "Comparison must be either ordered or unordered");
3373  if (match(RHS, m_NaN()))
3374  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3375 
3376  // fcmp pred x, undef and fcmp pred undef, x
3377  // fold to true if unordered, false if ordered
3378  if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3379  // Choosing NaN for the undef will always make unordered comparison succeed
3380  // and ordered comparison fail.
3381  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3382  }
3383 
3384  // fcmp x,x -> true/false. Not all compares are foldable.
3385  if (LHS == RHS) {
3386  if (CmpInst::isTrueWhenEqual(Pred))
3387  return getTrue(RetTy);
3388  if (CmpInst::isFalseWhenEqual(Pred))
3389  return getFalse(RetTy);
3390  }
3391 
3392  // Handle fcmp with constant RHS.
3393  // TODO: Use match with a specific FP value, so these work with vectors with
3394  // undef lanes.
3395  const APFloat *C;
3396  if (match(RHS, m_APFloat(C))) {
3397  // Check whether the constant is an infinity.
3398  if (C->isInfinity()) {
3399  if (C->isNegative()) {
3400  switch (Pred) {
3401  case FCmpInst::FCMP_OLT:
3402  // No value is ordered and less than negative infinity.
3403  return getFalse(RetTy);
3404  case FCmpInst::FCMP_UGE:
3405  // All values are unordered with or at least negative infinity.
3406  return getTrue(RetTy);
3407  default:
3408  break;
3409  }
3410  } else {
3411  switch (Pred) {
3412  case FCmpInst::FCMP_OGT:
3413  // No value is ordered and greater than infinity.
3414  return getFalse(RetTy);
3415  case FCmpInst::FCMP_ULE:
3416  // All values are unordered with and at most infinity.
3417  return getTrue(RetTy);
3418  default:
3419  break;
3420  }
3421  }
3422  }
3423  if (C->isNegative() && !C->isNegZero()) {
3424  assert(!C->isNaN() && "Unexpected NaN constant!");
3425  // TODO: We can catch more cases by using a range check rather than
3426  // relying on CannotBeOrderedLessThanZero.
3427  switch (Pred) {
3428  case FCmpInst::FCMP_UGE:
3429  case FCmpInst::FCMP_UGT:
3430  case FCmpInst::FCMP_UNE:
3431  // (X >= 0) implies (X > C) when (C < 0)
3432  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3433  return getTrue(RetTy);
3434  break;
3435  case FCmpInst::FCMP_OEQ:
3436  case FCmpInst::FCMP_OLE:
3437  case FCmpInst::FCMP_OLT:
3438  // (X >= 0) implies !(X < C) when (C < 0)
3439  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3440  return getFalse(RetTy);
3441  break;
3442  default:
3443  break;
3444  }
3445  }
3446 
3447  // Check comparison of [minnum/maxnum with constant] with other constant.
3448  const APFloat *C2;
3449  if ((match(LHS, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_APFloat(C2))) &&
3450  C2->compare(*C) == APFloat::cmpLessThan) ||
3451  (match(LHS, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_APFloat(C2))) &&
3452  C2->compare(*C) == APFloat::cmpGreaterThan)) {
3453  bool IsMaxNum =
3454  cast<IntrinsicInst>(LHS)->getIntrinsicID() == Intrinsic::maxnum;
3455  // The ordered relationship and minnum/maxnum guarantee that we do not
3456  // have NaN constants, so ordered/unordered preds are handled the same.
3457  switch (Pred) {
3459  // minnum(X, LesserC) == C --> false
3460  // maxnum(X, GreaterC) == C --> false
3461  return getFalse(RetTy);
3463  // minnum(X, LesserC) != C --> true
3464  // maxnum(X, GreaterC) != C --> true
3465  return getTrue(RetTy);
3468  // minnum(X, LesserC) >= C --> false
3469  // minnum(X, LesserC) > C --> false
3470  // maxnum(X, GreaterC) >= C --> true
3471  // maxnum(X, GreaterC) > C --> true
3472  return ConstantInt::get(RetTy, IsMaxNum);
3475  // minnum(X, LesserC) <= C --> true
3476  // minnum(X, LesserC) < C --> true
3477  // maxnum(X, GreaterC) <= C --> false
3478  // maxnum(X, GreaterC) < C --> false
3479  return ConstantInt::get(RetTy, !IsMaxNum);
3480  default:
3481  // TRUE/FALSE/ORD/UNO should be handled before this.
3482  llvm_unreachable("Unexpected fcmp predicate");
3483  }
3484  }
3485  }
3486 
3487  if (match(RHS, m_AnyZeroFP())) {
3488  switch (Pred) {
3489  case FCmpInst::FCMP_OGE:
3490  case FCmpInst::FCMP_ULT:
3491  // Positive or zero X >= 0.0 --> true
3492  // Positive or zero X < 0.0 --> false
3493  if ((FMF.noNaNs() || isKnownNeverNaN(LHS, Q.TLI)) &&
3495  return Pred == FCmpInst::FCMP_OGE ? getTrue(RetTy) : getFalse(RetTy);
3496  break;
3497  case FCmpInst::FCMP_UGE:
3498  case FCmpInst::FCMP_OLT:
3499  // Positive or zero or nan X >= 0.0 --> true
3500  // Positive or zero or nan X < 0.0 --> false
3501  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3502  return Pred == FCmpInst::FCMP_UGE ? getTrue(RetTy) : getFalse(RetTy);
3503  break;
3504  default:
3505  break;
3506  }
3507  }
3508 
3509  // If the comparison is with the result of a select instruction, check whether
3510  // comparing with either branch of the select always yields the same value.
3511  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3512  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3513  return V;
3514 
3515  // If the comparison is with the result of a phi instruction, check whether
3516  // doing the compare with each incoming phi value yields a common result.
3517  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3518  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3519  return V;
3520 
3521  return nullptr;
3522 }
3523 
3524 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3525  FastMathFlags FMF, const SimplifyQuery &Q) {
3526  return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3527 }
3528 
3529 /// See if V simplifies when its operand Op is replaced with RepOp.
3530 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3531  const SimplifyQuery &Q,
3532  unsigned MaxRecurse) {
3533  // Trivial replacement.
3534  if (V == Op)
3535  return RepOp;
3536 
3537  // We cannot replace a constant, and shouldn't even try.
3538  if (isa<Constant>(Op))
3539  return nullptr;
3540 
3541  auto *I = dyn_cast<Instruction>(V);
3542  if (!I)
3543  return nullptr;
3544 
3545  // If this is a binary operator, try to simplify it with the replaced op.
3546  if (auto *B = dyn_cast<BinaryOperator>(I)) {
3547  // Consider:
3548  // %cmp = icmp eq i32 %x, 2147483647
3549  // %add = add nsw i32 %x, 1
3550  // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3551  //
3552  // We can't replace %sel with %add unless we strip away the flags.
3553  if (isa<OverflowingBinaryOperator>(B))
3554  if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3555  return nullptr;
3556  if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3557  return nullptr;
3558 
3559  if (MaxRecurse) {
3560  if (B->getOperand(0) == Op)
3561  return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3562  MaxRecurse - 1);
3563  if (B->getOperand(1) == Op)
3564  return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3565  MaxRecurse - 1);
3566  }
3567  }
3568 
3569  // Same for CmpInsts.
3570  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3571  if (MaxRecurse) {
3572  if (C->getOperand(0) == Op)
3573  return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3574  MaxRecurse - 1);
3575  if (C->getOperand(1) == Op)
3576  return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3577  MaxRecurse - 1);
3578  }
3579  }
3580 
3581  // Same for GEPs.
3582  if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3583  if (MaxRecurse) {
3584  SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3585  transform(GEP->operands(), NewOps.begin(),
3586  [&](Value *V) { return V == Op ? RepOp : V; });
3587  return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3588  MaxRecurse - 1);
3589  }
3590  }
3591 
3592  // TODO: We could hand off more cases to instsimplify here.
3593 
3594  // If all operands are constant after substituting Op for RepOp then we can
3595  // constant fold the instruction.
3596  if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3597  // Build a list of all constant operands.
3598  SmallVector<Constant *, 8> ConstOps;
3599  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3600  if (I->getOperand(i) == Op)
3601  ConstOps.push_back(CRepOp);
3602  else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3603  ConstOps.push_back(COp);
3604  else
3605  break;
3606  }
3607 
3608  // All operands were constants, fold it.
3609  if (ConstOps.size() == I->getNumOperands()) {
3610  if (CmpInst *C = dyn_cast<CmpInst>(I))
3611  return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3612  ConstOps[1], Q.DL, Q.TLI);
3613 
3614  if (LoadInst *LI = dyn_cast<LoadInst>(I))
3615  if (!LI->isVolatile())
3616  return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3617 
3618  return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3619  }
3620  }
3621 
3622  return nullptr;
3623 }
3624 
3625 /// Try to simplify a select instruction when its condition operand is an
3626 /// integer comparison where one operand of the compare is a constant.
3627 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3628  const APInt *Y, bool TrueWhenUnset) {
3629  const APInt *C;
3630 
3631  // (X & Y) == 0 ? X & ~Y : X --> X
3632  // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3633  if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3634  *Y == ~*C)
3635  return TrueWhenUnset ? FalseVal : TrueVal;
3636 
3637  // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3638  // (X & Y) != 0 ? X : X & ~Y --> X
3639  if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3640  *Y == ~*C)
3641  return TrueWhenUnset ? FalseVal : TrueVal;
3642 
3643  if (Y->isPowerOf2()) {
3644  // (X & Y) == 0 ? X | Y : X --> X | Y
3645  // (X & Y) != 0 ? X | Y : X --> X
3646  if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3647  *Y == *C)
3648  return TrueWhenUnset ? TrueVal : FalseVal;
3649 
3650  // (X & Y) == 0 ? X : X | Y --> X
3651  // (X & Y) != 0 ? X : X | Y --> X | Y
3652  if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3653  *Y == *C)
3654  return TrueWhenUnset ? TrueVal : FalseVal;
3655  }
3656 
3657  return nullptr;
3658 }
3659 
3660 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3661 /// eq/ne.
3663  ICmpInst::Predicate Pred,
3664  Value *TrueVal, Value *FalseVal) {
3665  Value *X;
3666  APInt Mask;
3667  if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3668  return nullptr;
3669 
3670  return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3671  Pred == ICmpInst::ICMP_EQ);
3672 }
3673 
3674 /// Try to simplify a select instruction when its condition operand is an
3675 /// integer comparison.
3676 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3677  Value *FalseVal, const SimplifyQuery &Q,
3678  unsigned MaxRecurse) {
3679  ICmpInst::Predicate Pred;
3680  Value *CmpLHS, *CmpRHS;
3681  if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3682  return nullptr;
3683 
3684  if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3685  Value *X;
3686  const APInt *Y;
3687  if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3688  if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3689  Pred == ICmpInst::ICMP_EQ))
3690  return V;
3691 
3692  // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
3693  Value *ShAmt;
3694  auto isFsh = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(),
3695  m_Value(ShAmt)),
3696  m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X),
3697  m_Value(ShAmt)));
3698  // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
3699  // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
3700  if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt &&
3701  Pred == ICmpInst::ICMP_EQ)
3702  return X;
3703  // (ShAmt != 0) ? X : fshl(X, *, ShAmt) --> X
3704  // (ShAmt != 0) ? X : fshr(*, X, ShAmt) --> X
3705  if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt &&
3706  Pred == ICmpInst::ICMP_NE)
3707  return X;
3708 
3709  // Test for a zero-shift-guard-op around rotates. These are used to
3710  // avoid UB from oversized shifts in raw IR rotate patterns, but the
3711  // intrinsics do not have that problem.
3712  // We do not allow this transform for the general funnel shift case because
3713  // that would not preserve the poison safety of the original code.
3714  auto isRotate = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X),
3715  m_Deferred(X),
3716  m_Value(ShAmt)),
3717  m_Intrinsic<Intrinsic::fshr>(m_Value(X),
3718  m_Deferred(X),
3719  m_Value(ShAmt)));
3720  // (ShAmt != 0) ? fshl(X, X, ShAmt) : X --> fshl(X, X, ShAmt)
3721  // (ShAmt != 0) ? fshr(X, X, ShAmt) : X --> fshr(X, X, ShAmt)
3722  if (match(TrueVal, isRotate) && FalseVal == X && CmpLHS == ShAmt &&
3723  Pred == ICmpInst::ICMP_NE)
3724  return TrueVal;
3725  // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
3726  // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
3727  if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
3728  Pred == ICmpInst::ICMP_EQ)
3729  return FalseVal;
3730  }
3731 
3732  // Check for other compares that behave like bit test.
3733  if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3734  TrueVal, FalseVal))
3735  return V;
3736 
3737  // If we have an equality comparison, then we know the value in one of the
3738  // arms of the select. See if substituting this value into the arm and
3739  // simplifying the result yields the same value as the other arm.
3740  if (Pred == ICmpInst::ICMP_EQ) {
3741  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3742  TrueVal ||
3743  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3744  TrueVal)
3745  return FalseVal;
3746  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3747  FalseVal ||
3748  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3749  FalseVal)
3750  return FalseVal;
3751  } else if (Pred == ICmpInst::ICMP_NE) {
3752  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3753  FalseVal ||
3754  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3755  FalseVal)
3756  return TrueVal;
3757  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3758  TrueVal ||
3759  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3760  TrueVal)
3761  return TrueVal;
3762  }
3763 
3764  return nullptr;
3765 }
3766 
3767 /// Try to simplify a select instruction when its condition operand is a
3768 /// floating-point comparison.
3770  FCmpInst::Predicate Pred;
3771  if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
3772  !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
3773  return nullptr;
3774 
3775  // TODO: The transform may not be valid with -0.0. An incomplete way of
3776  // testing for that possibility is to check if at least one operand is a
3777  // non-zero constant.
3778  const APFloat *C;
3779  if ((match(T, m_APFloat(C)) && C->isNonZero()) ||
3780  (match(F, m_APFloat(C)) && C->isNonZero())) {
3781  // (T == F) ? T : F --> F
3782  // (F == T) ? T : F --> F
3783  if (Pred == FCmpInst::FCMP_OEQ)
3784  return F;
3785 
3786  // (T != F) ? T : F --> T
3787  // (F != T) ? T : F --> T
3788  if (Pred == FCmpInst::FCMP_UNE)
3789  return T;
3790  }
3791 
3792  return nullptr;
3793 }
3794 
3795 /// Given operands for a SelectInst, see if we can fold the result.
3796 /// If not, this returns null.
3797 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3798  const SimplifyQuery &Q, unsigned MaxRecurse) {
3799  if (auto *CondC = dyn_cast<Constant>(Cond)) {
3800  if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3801  if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3802  return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3803 
3804  // select undef, X, Y -> X or Y
3805  if (isa<UndefValue>(CondC))
3806  return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3807 
3808  // TODO: Vector constants with undef elements don't simplify.
3809 
3810  // select true, X, Y -> X
3811  if (CondC->isAllOnesValue())
3812  return TrueVal;
3813  // select false, X, Y -> Y
3814  if (CondC->isNullValue())
3815  return FalseVal;
3816  }
3817 
3818  // select ?, X, X -> X
3819  if (TrueVal == FalseVal)
3820  return TrueVal;
3821 
3822  if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3823  return FalseVal;
3824  if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3825  return TrueVal;
3826 
3827  if (Value *V =
3828  simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3829  return V;
3830 
3831  if (Value *V = simplifySelectWithFCmp(Cond, TrueVal, FalseVal))
3832  return V;
3833 
3834  if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3835  return V;
3836 
3837  Optional<bool> Imp = isImpliedByDomCondition(Cond, Q.CxtI, Q.DL);
3838  if (Imp)
3839  return *Imp ? TrueVal : FalseVal;
3840 
3841  return nullptr;
3842 }
3843 
3844 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3845  const SimplifyQuery &Q) {
3846  return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3847 }
3848 
3849 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3850 /// If not, this returns null.
3852  const SimplifyQuery &Q, unsigned) {
3853  // The type of the GEP pointer operand.
3854  unsigned AS =
3855  cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3856 
3857  // getelementptr P -> P.
3858  if (Ops.size() == 1)
3859  return Ops[0];
3860 
3861  // Compute the (pointer) type returned by the GEP instruction.
3862  Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3863  Type *GEPTy = PointerType::get(LastType, AS);
3864  if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3865  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3866  else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3867  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3868 
3869  if (isa<UndefValue>(Ops[0]))
3870  return UndefValue::get(GEPTy);
3871 
3872  if (Ops.size() == 2) {
3873  // getelementptr P, 0 -> P.
3874  if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3875  return Ops[0];
3876 
3877  Type *Ty = SrcTy;
3878  if (Ty->isSized()) {
3879  Value *P;
3880  uint64_t C;
3881  uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3882  // getelementptr P, N -> P if P points to a type of zero size.
3883  if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3884  return Ops[0];
3885 
3886  // The following transforms are only safe if the ptrtoint cast
3887  // doesn't truncate the pointers.
3888  if (Ops[1]->getType()->getScalarSizeInBits() ==
3889  Q.DL.getIndexSizeInBits(AS)) {
3890  auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3891  if (match(P, m_Zero()))
3892  return Constant::getNullValue(GEPTy);
3893  Value *Temp;
3894  if (match(P, m_PtrToInt(m_Value(Temp))))
3895  if (Temp->getType() == GEPTy)
3896  return Temp;
3897  return nullptr;
3898  };
3899 
3900  // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3901  if (TyAllocSize == 1 &&
3902  match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3903  if (Value *R = PtrToIntOrZero(P))
3904  return R;
3905 
3906  // getelementptr V, (ashr (sub P, V), C) -> Q
3907  // if P points to a type of size 1 << C.
3908  if (match(Ops[1],
3909  m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3910  m_ConstantInt(C))) &&
3911  TyAllocSize == 1ULL << C)
3912  if (Value *R = PtrToIntOrZero(P))
3913  return R;
3914 
3915  // getelementptr V, (sdiv (sub P, V), C) -> Q
3916  // if P points to a type of size C.
3917  if (match(Ops[1],
3918  m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3919  m_SpecificInt(TyAllocSize))))
3920  if (Value *R = PtrToIntOrZero(P))
3921  return R;
3922  }
3923  }
3924  }
3925 
3926  if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3927  all_of(Ops.slice(1).drop_back(1),
3928  [](Value *Idx) { return match(Idx, m_Zero()); })) {
3929  unsigned IdxWidth =
3930  Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3931  if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3932  APInt BasePtrOffset(IdxWidth, 0);
3933  Value *StrippedBasePtr =
3934  Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3935  BasePtrOffset);
3936 
3937  // gep (gep V, C), (sub 0, V) -> C
3938  if (match(Ops.back(),
3939  m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3940  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3941  return ConstantExpr::getIntToPtr(CI, GEPTy);
3942  }
3943  // gep (gep V, C), (xor V, -1) -> C-1
3944  if (match(Ops.back(),
3945  m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3946  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3947  return ConstantExpr::getIntToPtr(CI, GEPTy);
3948  }
3949  }
3950  }
3951 
3952  // Check to see if this is constant foldable.
3953  if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3954  return nullptr;
3955 
3956  auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3957  Ops.slice(1));
3958  if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3959  return CEFolded;
3960  return CE;
3961 }
3962 
3964  const SimplifyQuery &Q) {
3965  return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3966 }
3967 
3968 /// Given operands for an InsertValueInst, see if we can fold the result.
3969 /// If not, this returns null.
3971  ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3972  unsigned) {
3973  if (Constant *CAgg = dyn_cast<Constant>(Agg))
3974  if (Constant *CVal = dyn_cast<Constant>(Val))
3975  return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3976 
3977  // insertvalue x, undef, n -> x
3978  if (match(Val, m_Undef()))
3979  return Agg;
3980 
3981  // insertvalue x, (extractvalue y, n), n
3982  if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3983  if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3984  EV->getIndices() == Idxs) {
3985  // insertvalue undef, (extractvalue y, n), n -> y
3986  if (match(Agg, m_Undef()))
3987  return EV->getAggregateOperand();
3988 
3989  // insertvalue y, (extractvalue y, n), n -> y
3990  if (Agg == EV->getAggregateOperand())
3991  return Agg;
3992  }
3993 
3994  return nullptr;
3995 }
3996 
3998  ArrayRef<unsigned> Idxs,
3999  const SimplifyQuery &Q) {
4001 }
4002 
4004  const SimplifyQuery &Q) {
4005  // Try to constant fold.
4006  auto *VecC = dyn_cast<Constant>(Vec);
4007  auto *ValC = dyn_cast<Constant>(Val);
4008  auto *IdxC = dyn_cast<Constant>(Idx);
4009  if (VecC && ValC && IdxC)
4010  return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
4011 
4012  // Fold into undef if index is out of bounds.
4013  if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4014  uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
4015  if (CI->uge(NumElements))
4016  return UndefValue::get(Vec->getType());
4017  }
4018 
4019  // If index is undef, it might be out of bounds (see above case)
4020  if (isa<UndefValue>(Idx))
4021  return UndefValue::get(Vec->getType());
4022 
4023  // Inserting an undef scalar? Assume it is the same value as the existing
4024  // vector element.
4025  if (isa<UndefValue>(Val))
4026  return Vec;
4027 
4028  // If we are extracting a value from a vector, then inserting it into the same
4029  // place, that's the input vector:
4030  // insertelt Vec, (extractelt Vec, Idx), Idx --> Vec
4031  if (match(Val, m_ExtractElement(m_Specific(Vec), m_Specific(Idx))))
4032  return Vec;
4033 
4034  return nullptr;
4035 }
4036 
4037 /// Given operands for an ExtractValueInst, see if we can fold the result.
4038 /// If not, this returns null.
4040  const SimplifyQuery &, unsigned) {
4041  if (auto *CAgg = dyn_cast<Constant>(Agg))
4042  return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4043 
4044  // extractvalue x, (insertvalue y, elt, n), n -> elt
4045  unsigned NumIdxs = Idxs.size();
4046  for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4047  IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4048  ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4049  unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4050  unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4051  if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4052  Idxs.slice(0, NumCommonIdxs)) {
4053  if (NumIdxs == NumInsertValueIdxs)
4054  return IVI->getInsertedValueOperand();
4055  break;
4056  }
4057  }
4058 
4059  return nullptr;
4060 }
4061 
4063  const SimplifyQuery &Q) {
4065 }
4066 
4067 /// Given operands for an ExtractElementInst, see if we can fold the result.
4068 /// If not, this returns null.
4070  unsigned) {
4071  if (auto *CVec = dyn_cast<Constant>(Vec)) {
4072  if (auto *CIdx = dyn_cast<Constant>(Idx))
4073  return ConstantFoldExtractElementInstruction(CVec, CIdx);
4074 
4075  // The index is not relevant if our vector is a splat.
4076  if (auto *Splat = CVec->getSplatValue())
4077  return Splat;
4078 
4079  if (isa<UndefValue>(Vec))
4080  return UndefValue::get(Vec->getType()->getVectorElementType());
4081  }
4082 
4083  // If extracting a specified index from the vector, see if we can recursively
4084  // find a previously computed scalar that was inserted into the vector.
4085  if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4086  if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4087  // definitely out of bounds, thus undefined result
4088  return UndefValue::get(Vec->getType()->getVectorElementType());
4089  if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4090  return Elt;
4091  }
4092 
4093  // An undef extract index can be arbitrarily chosen to be an out-of-range
4094  // index value, which would result in the instruction being undef.
4095  if (isa<UndefValue>(Idx))
4096  return UndefValue::get(Vec->getType()->getVectorElementType());
4097 
4098  return nullptr;
4099 }
4100 
4102  const SimplifyQuery &Q) {
4104 }
4105 
4106 /// See if we can fold the given phi. If not, returns null.
4107 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4108  // If all of the PHI's incoming values are the same then replace the PHI node
4109  // with the common value.
4110  Value *CommonValue = nullptr;
4111  bool HasUndefInput = false;
4112  for (Value *Incoming : PN->incoming_values()) {
4113  // If the incoming value is the phi node itself, it can safely be skipped.
4114  if (Incoming == PN) continue;
4115  if (isa<UndefValue>(Incoming)) {
4116  // Remember that we saw an undef value, but otherwise ignore them.
4117  HasUndefInput = true;
4118  continue;
4119  }
4120  if (CommonValue && Incoming != CommonValue)
4121  return nullptr; // Not the same, bail out.
4122  CommonValue = Incoming;
4123  }
4124 
4125  // If CommonValue is null then all of the incoming values were either undef or
4126  // equal to the phi node itself.
4127  if (!CommonValue)
4128  return UndefValue::get(PN->getType());
4129 
4130  // If we have a PHI node like phi(X, undef, X), where X is defined by some
4131  // instruction, we cannot return X as the result of the PHI node unless it
4132  // dominates the PHI block.
4133  if (HasUndefInput)
4134  return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4135 
4136  return CommonValue;
4137 }
4138 
4139 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4140  Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4141  if (auto *C = dyn_cast<Constant>(Op))
4142  return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4143 
4144  if (auto *CI = dyn_cast<CastInst>(Op)) {
4145  auto *Src = CI->getOperand(0);
4146  Type *SrcTy = Src->getType();
4147  Type *MidTy = CI->getType();
4148  Type *DstTy = Ty;
4149  if (Src->getType() == Ty) {
4150  auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4151  auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4152  Type *SrcIntPtrTy =
4153  SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4154  Type *MidIntPtrTy =
4155  MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4156  Type *DstIntPtrTy =
4157  DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4158  if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4159  SrcIntPtrTy, MidIntPtrTy,
4160  DstIntPtrTy) == Instruction::BitCast)
4161  return Src;
4162  }
4163  }
4164 
4165  // bitcast x -> x
4166  if (CastOpc == Instruction::BitCast)
4167  if (Op->getType() == Ty)
4168  return Op;
4169 
4170  return nullptr;
4171 }
4172 
4173 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4174  const SimplifyQuery &Q) {
4175  return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4176 }
4177 
4178 /// For the given destination element of a shuffle, peek through shuffles to
4179 /// match a root vector source operand that contains that element in the same
4180 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4181 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4182  int MaskVal, Value *RootVec,
4183  unsigned MaxRecurse) {
4184  if (!MaxRecurse--)
4185  return nullptr;
4186 
4187  // Bail out if any mask value is undefined. That kind of shuffle may be
4188  // simplified further based on demanded bits or other folds.
4189  if (MaskVal == -1)
4190  return nullptr;
4191 
4192  // The mask value chooses which source operand we need to look at next.
4193  int InVecNumElts = Op0->getType()->getVectorNumElements();
4194  int RootElt = MaskVal;
4195  Value *SourceOp = Op0;
4196  if (MaskVal >= InVecNumElts) {
4197  RootElt = MaskVal - InVecNumElts;
4198  SourceOp = Op1;
4199  }
4200 
4201  // If the source operand is a shuffle itself, look through it to find the
4202  // matching root vector.
4203  if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4204  return foldIdentityShuffles(
4205  DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4206  SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4207  }
4208 
4209  // TODO: Look through bitcasts? What if the bitcast changes the vector element
4210  // size?
4211 
4212  // The source operand is not a shuffle. Initialize the root vector value for
4213  // this shuffle if that has not been done yet.
4214  if (!RootVec)
4215  RootVec = SourceOp;
4216 
4217  // Give up as soon as a source operand does not match the existing root value.
4218  if (RootVec != SourceOp)
4219  return nullptr;
4220 
4221  // The element must be coming from the same lane in the source vector
4222  // (although it may have crossed lanes in intermediate shuffles).
4223  if (RootElt != DestElt)
4224  return nullptr;
4225 
4226  return RootVec;
4227 }
4228 
4230  Type *RetTy, const SimplifyQuery &Q,
4231  unsigned MaxRecurse) {
4232  if (isa<UndefValue>(Mask))
4233  return UndefValue::get(RetTy);
4234 
4235  Type *InVecTy = Op0->getType();
4236  unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4237  unsigned InVecNumElts = InVecTy->getVectorNumElements();
4238 
4239  SmallVector<int, 32> Indices;
4240  ShuffleVectorInst::getShuffleMask(Mask, Indices);
4241  assert(MaskNumElts == Indices.size() &&
4242  "Size of Indices not same as number of mask elements?");
4243 
4244  // Canonicalization: If mask does not select elements from an input vector,
4245  // replace that input vector with undef.
4246  bool MaskSelects0 = false, MaskSelects1 = false;
4247  for (unsigned i = 0; i != MaskNumElts; ++i) {
4248  if (Indices[i] == -1)
4249  continue;
4250  if ((unsigned)Indices[i] < InVecNumElts)
4251  MaskSelects0 = true;
4252  else
4253  MaskSelects1 = true;
4254  }
4255  if (!MaskSelects0)
4256  Op0 = UndefValue::get(InVecTy);
4257  if (!MaskSelects1)
4258  Op1 = UndefValue::get(InVecTy);
4259 
4260  auto *Op0Const = dyn_cast<Constant>(Op0);
4261  auto *Op1Const = dyn_cast<Constant>(Op1);
4262 
4263  // If all operands are constant, constant fold the shuffle.
4264  if (Op0Const && Op1Const)
4265  return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4266 
4267  // Canonicalization: if only one input vector is constant, it shall be the
4268  // second one.
4269  if (Op0Const && !Op1Const) {
4270  std::swap(Op0, Op1);
4271  ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4272  }
4273 
4274  // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4275  // value type is same as the input vectors' type.
4276  if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4277  if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4278  OpShuf->getMask()->getSplatValue())
4279  return Op0;
4280 
4281  // Don't fold a shuffle with undef mask elements. This may get folded in a
4282  // better way using demanded bits or other analysis.
4283  // TODO: Should we allow this?
4284  if (find(Indices, -1) != Indices.end())
4285  return nullptr;
4286 
4287  // Check if every element of this shuffle can be mapped back to the
4288  // corresponding element of a single root vector. If so, we don't need this
4289  // shuffle. This handles simple identity shuffles as well as chains of
4290  // shuffles that may widen/narrow and/or move elements across lanes and back.
4291  Value *RootVec = nullptr;
4292  for (unsigned i = 0; i != MaskNumElts; ++i) {
4293  // Note that recursion is limited for each vector element, so if any element
4294  // exceeds the limit, this will fail to simplify.
4295  RootVec =
4296  foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4297 
4298  // We can't replace a widening/narrowing shuffle with one of its operands.
4299  if (!RootVec || RootVec->getType() != RetTy)
4300  return nullptr;
4301  }
4302  return RootVec;
4303 }
4304 
4305 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4307  Type *RetTy, const SimplifyQuery &Q) {
4308  return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4309 }
4310 
4312  Value *&Op, const SimplifyQuery &Q) {
4313  if (auto *C = dyn_cast<Constant>(Op))
4314  return ConstantFoldUnaryOpOperand(Opcode, C, Q.DL);
4315  return nullptr;
4316 }
4317 
4318 /// Given the operand for an FNeg, see if we can fold the result. If not, this
4319 /// returns null.
4321  const SimplifyQuery &Q, unsigned MaxRecurse) {
4322  if (Constant *C = foldConstant(Instruction::FNeg, Op, Q))
4323  return C;
4324 
4325  Value *X;
4326  // fneg (fneg X) ==> X
4327  if (match(Op, m_FNeg(m_Value(X))))
4328  return X;
4329 
4330  return nullptr;
4331 }
4332 
4334  const SimplifyQuery &Q) {
4336 }
4337 
4339  // If the input is a vector with undef elements, just return a default NaN.
4340  if (!In->isNaN())
4341  return ConstantFP::getNaN(In->getType());
4342 
4343  // Propagate the existing NaN constant when possible.
4344  // TODO: Should we quiet a signaling NaN?
4345  return In;
4346 }
4347 
4348 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4349  if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4350  return ConstantFP::getNaN(Op0->getType());
4351 
4352  if (match(Op0, m_NaN()))
4353  return propagateNaN(cast<Constant>(Op0));
4354  if (match(Op1, m_NaN()))
4355  return propagateNaN(cast<Constant>(Op1));
4356 
4357  return nullptr;
4358 }
4359 
4360 /// Given operands for an FAdd, see if we can fold the result. If not, this
4361 /// returns null.
4363  const SimplifyQuery &Q, unsigned MaxRecurse) {
4364  if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4365  return C;
4366 
4367  if (Constant *C = simplifyFPBinop(Op0, Op1))
4368  return C;
4369 
4370  // fadd X, -0 ==> X
4371  if (match(Op1, m_NegZeroFP()))
4372  return Op0;
4373 
4374  // fadd X, 0 ==> X, when we know X is not -0
4375  if (match(Op1, m_PosZeroFP()) &&
4376  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4377  return Op0;
4378 
4379  // With nnan: -X + X --> 0.0 (and commuted variant)
4380  // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4381  // Negative zeros are allowed because we always end up with positive zero:
4382  // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4383  // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4384  // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4385  // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4386  if (FMF.noNaNs()) {
4387  if (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4388  match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0))))
4389  return ConstantFP::getNullValue(Op0->getType());
4390 
4391  if (match(Op0, m_FNeg(m_Specific(Op1))) ||
4392  match(Op1, m_FNeg(m_Specific(Op0))))
4393  return ConstantFP::getNullValue(Op0->getType());
4394  }
4395 
4396  // (X - Y) + Y --> X
4397  // Y + (X - Y) --> X
4398  Value *X;
4399  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4400  (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4401  match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4402  return X;
4403 
4404  return nullptr;
4405 }
4406 
4407 /// Given operands for an FSub, see if we can fold the result. If not, this
4408 /// returns null.
4410  const SimplifyQuery &Q, unsigned MaxRecurse) {
4411  if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4412  return C;
4413 
4414  if (Constant *C = simplifyFPBinop(Op0, Op1))
4415  return C;
4416 
4417  // fsub X, +0 ==> X
4418  if (match(Op1, m_PosZeroFP()))
4419  return Op0;
4420 
4421  // fsub X, -0 ==> X, when we know X is not -0
4422  if (match(Op1, m_NegZeroFP()) &&
4423  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4424  return Op0;
4425 
4426  // fsub -0.0, (fsub -0.0, X) ==> X
4427  // fsub -0.0, (fneg X) ==> X
4428  Value *X;
4429  if (match(Op0, m_NegZeroFP()) &&
4430  match(Op1, m_FNeg(m_Value(X))))
4431  return X;
4432 
4433  // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4434  // fsub 0.0, (fneg X) ==> X if signed zeros are ignored.
4435  if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4436  (match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))) ||
4437  match(Op1, m_FNeg(m_Value(X)))))
4438  return X;
4439 
4440  // fsub nnan x, x ==> 0.0
4441  if (FMF.noNaNs() && Op0 == Op1)
4442  return Constant::getNullValue(Op0->getType());
4443 
4444  // Y - (Y - X) --> X
4445  // (X + Y) - Y --> X
4446  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4447  (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4448  match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4449  return X;
4450 
4451  return nullptr;
4452 }
4453 
4454 /// Given the operands for an FMul, see if we can fold the result
4456  const SimplifyQuery &Q, unsigned MaxRecurse) {
4457  if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4458  return C;
4459 
4460  if (Constant *C = simplifyFPBinop(Op0, Op1))
4461  return C;
4462 
4463  // fmul X, 1.0 ==> X
4464  if (match(Op1, m_FPOne()))
4465  return Op0;
4466 
4467  // fmul nnan nsz X, 0 ==> 0
4468  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4469  return ConstantFP::getNullValue(Op0->getType());
4470 
4471  // sqrt(X) * sqrt(X) --> X, if we can:
4472  // 1. Remove the intermediate rounding (reassociate).
4473  // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4474  // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4475  Value *X;
4476  if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4477  FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4478  return X;
4479 
4480  return nullptr;
4481 }
4482 
4484  const SimplifyQuery &Q) {
4485  return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4486 }
4487 
4488 
4490  const SimplifyQuery &Q) {
4491  return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4492 }
4493 
4495  const SimplifyQuery &Q) {
4496  return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4497 }
4498 
4500  const SimplifyQuery &Q, unsigned) {
4501  if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4502  return C;
4503 
4504  if (Constant *C = simplifyFPBinop(Op0, Op1))
4505  return C;
4506 
4507  // X / 1.0 -> X
4508  if (match(Op1, m_FPOne()))
4509  return Op0;
4510 
4511  // 0 / X -> 0
4512  // Requires that NaNs are off (X could be zero) and signed zeroes are
4513  // ignored (X could be positive or negative, so the output sign is unknown).
4514  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4515  return ConstantFP::getNullValue(Op0->getType());
4516 
4517  if (FMF.noNaNs()) {
4518  // X / X -> 1.0 is legal when NaNs are ignored.
4519  // We can ignore infinities because INF/INF is NaN.
4520  if (Op0 == Op1)
4521  return ConstantFP::get(Op0->getType(), 1.0);
4522 
4523  // (X * Y) / Y --> X if we can reassociate to the above form.
4524  Value *X;
4525  if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4526  return X;
4527 
4528  // -X / X -> -1.0 and
4529  // X / -X -> -1.0 are legal when NaNs are ignored.
4530  // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4531  if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4532  match(Op1, m_FNegNSZ(m_Specific(Op0))))
4533  return ConstantFP::get(Op0->getType(), -1.0);
4534  }
4535 
4536  return nullptr;
4537 }
4538 
4540  const SimplifyQuery &Q) {
4541  return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4542 }
4543 
4545  const SimplifyQuery &Q, unsigned) {
4546  if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4547  return C;
4548 
4549  if (Constant *C = simplifyFPBinop(Op0, Op1))
4550  return C;
4551 
4552  // Unlike fdiv, the result of frem always matches the sign of the dividend.
4553  // The constant match may include undef elements in a vector, so return a full
4554  // zero constant as the result.
4555  if (FMF.noNaNs()) {
4556  // +0 % X -> 0
4557  if (match(Op0, m_PosZeroFP()))
4558  return ConstantFP::getNullValue(Op0->getType());
4559  // -0 % X -> -0
4560  if (match(Op0, m_NegZeroFP()))
4561  return ConstantFP::getNegativeZero(Op0->getType());
4562  }
4563 
4564  return nullptr;
4565 }
4566 
4568  const SimplifyQuery &Q) {
4569  return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4570 }
4571 
4572 //=== Helper functions for higher up the class hierarchy.
4573 
4574 /// Given the operand for a UnaryOperator, see if we can fold the result.
4575 /// If not, this returns null.
4576 static Value *simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q,
4577  unsigned MaxRecurse) {
4578  switch (Opcode) {
4579  case Instruction::FNeg:
4580  return simplifyFNegInst(Op, FastMathFlags(), Q, MaxRecurse);
4581  default:
4582  llvm_unreachable("Unexpected opcode");
4583  }
4584 }
4585 
4586 /// Given the operand for a UnaryOperator, see if we can fold the result.
4587 /// If not, this returns null.
4588 /// In contrast to SimplifyUnOp, try to use FastMathFlag when folding the
4589 /// result. In case we don't need FastMathFlags, simply fall to SimplifyUnOp.
4590 static Value *simplifyFPUnOp(unsigned Opcode, Value *Op,
4591  const FastMathFlags &FMF,
4592  const SimplifyQuery &Q, unsigned MaxRecurse) {
4593  switch (Opcode) {
4594  case Instruction::FNeg:
4595  return simplifyFNegInst(Op, FMF, Q, MaxRecurse);
4596  default:
4597  return simplifyUnOp(Opcode, Op, Q, MaxRecurse);
4598  }
4599 }
4600 
4601 Value *llvm::SimplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q) {
4602  return ::simplifyUnOp(Opcode, Op, Q, RecursionLimit);
4603 }
4604 
4606  const SimplifyQuery &Q) {
4607  return ::simplifyFPUnOp(Opcode, Op, FMF, Q, RecursionLimit);
4608 }
4609 
4610 /// Given operands for a BinaryOperator, see if we can fold the result.
4611 /// If not, this returns null.
4612 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4613  const SimplifyQuery &Q, unsigned MaxRecurse) {
4614  switch (Opcode) {
4615  case Instruction::Add:
4616  return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4617  case Instruction::Sub:
4618  return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4619  case Instruction::Mul:
4620  return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4621  case Instruction::SDiv:
4622  return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4623  case Instruction::UDiv:
4624  return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4625  case Instruction::SRem:
4626  return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4627  case Instruction::URem:
4628  return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4629  case Instruction::Shl:
4630  return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4631  case Instruction::LShr:
4632  return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4633  case Instruction::AShr:
4634  return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4635  case Instruction::And:
4636  return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4637  case Instruction::Or:
4638  return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4639  case Instruction::Xor:
4640  return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4641  case Instruction::FAdd:
4642  return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4643  case Instruction::FSub:
4644  return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4645  case Instruction::FMul:
4646  return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4647  case Instruction::FDiv:
4648  return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4649  case Instruction::FRem:
4650  return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4651  default:
4652  llvm_unreachable("Unexpected opcode");
4653  }
4654 }
4655 
4656 /// Given operands for a BinaryOperator, see if we can fold the result.
4657 /// If not, this returns null.
4658 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4659 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4660 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4661  const FastMathFlags &FMF, const SimplifyQuery &Q,
4662  unsigned MaxRecurse) {
4663  switch (Opcode) {
4664  case Instruction::FAdd:
4665  return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4666  case Instruction::FSub:
4667  return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4668  case Instruction::FMul:
4669  return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4670  case Instruction::FDiv:
4671  return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4672  default:
4673  return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4674  }
4675 }
4676 
4677 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4678  const SimplifyQuery &Q) {
4679  return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4680 }
4681 
4682 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4683  FastMathFlags FMF, const SimplifyQuery &Q) {
4684  return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4685 }
4686 
4687 /// Given operands for a CmpInst, see if we can fold the result.
4688 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4689  const SimplifyQuery &Q, unsigned MaxRecurse) {
4691  return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4692  return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4693 }
4694 
4695 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4696  const SimplifyQuery &Q) {
4697  return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4698 }
4699 
4701  switch (ID) {
4702  default: return false;
4703 
4704  // Unary idempotent: f(f(x)) = f(x)
4705  case Intrinsic::fabs:
4706  case Intrinsic::floor:
4707  case Intrinsic::ceil:
4708  case Intrinsic::trunc:
4709  case Intrinsic::rint:
4710  case Intrinsic::nearbyint:
4711  case Intrinsic::round:
4712  case Intrinsic::canonicalize:
4713  return true;
4714  }
4715 }
4716 
4718  const DataLayout &DL) {
4719  GlobalValue *PtrSym;
4720  APInt PtrOffset;
4721  if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4722  return nullptr;
4723 
4724  Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4726  Type *Int32PtrTy = Int32Ty->getPointerTo();
4727  Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4728 
4729  auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4730  if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4731  return nullptr;
4732 
4733  uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4734  if (OffsetInt % 4 != 0)
4735  return nullptr;
4736 
4738  Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4739  ConstantInt::get(Int64Ty, OffsetInt / 4));
4740  Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4741  if (!Loaded)
4742  return nullptr;
4743 
4744  auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4745  if (!LoadedCE)
4746  return nullptr;
4747 
4748  if (LoadedCE->getOpcode() == Instruction::Trunc) {
4749  LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4750  if (!LoadedCE)
4751  return nullptr;
4752  }
4753 
4754  if (LoadedCE->getOpcode() != Instruction::Sub)
4755  return nullptr;
4756 
4757  auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4758  if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4759  return nullptr;
4760  auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4761 
4762  Constant *LoadedRHS = LoadedCE->getOperand(1);
4763  GlobalValue *LoadedRHSSym;
4764  APInt LoadedRHSOffset;
4765  if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4766  DL) ||
4767  PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4768  return nullptr;
4769 
4770  return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4771 }
4772 
4774  const SimplifyQuery &Q) {
4775  // Idempotent functions return the same result when called repeatedly.
4776  Intrinsic::ID IID = F->getIntrinsicID();
4777  if (IsIdempotent(IID))
4778  if (auto *II = dyn_cast<IntrinsicInst>(Op0))
4779  if (II