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  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
1849  return V;
1850 
1851  // Try some generic simplifications for associative operations.
1852  if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1853  MaxRecurse))
1854  return V;
1855 
1856  // And distributes over Or. Try some generic simplifications based on this.
1857  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1858  Q, MaxRecurse))
1859  return V;
1860 
1861  // And distributes over Xor. Try some generic simplifications based on this.
1862  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1863  Q, MaxRecurse))
1864  return V;
1865 
1866  // If the operation is with the result of a select instruction, check whether
1867  // operating on either branch of the select always yields the same value.
1868  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1869  if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1870  MaxRecurse))
1871  return V;
1872 
1873  // If the operation is with the result of a phi instruction, check whether
1874  // operating on all incoming values of the phi always yields the same value.
1875  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1876  if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1877  MaxRecurse))
1878  return V;
1879 
1880  // Assuming the effective width of Y is not larger than A, i.e. all bits
1881  // from X and Y are disjoint in (X << A) | Y,
1882  // if the mask of this AND op covers all bits of X or Y, while it covers
1883  // no bits from the other, we can bypass this AND op. E.g.,
1884  // ((X << A) | Y) & Mask -> Y,
1885  // if Mask = ((1 << effective_width_of(Y)) - 1)
1886  // ((X << A) | Y) & Mask -> X << A,
1887  // if Mask = ((1 << effective_width_of(X)) - 1) << A
1888  // SimplifyDemandedBits in InstCombine can optimize the general case.
1889  // This pattern aims to help other passes for a common case.
1890  Value *Y, *XShifted;
1891  if (match(Op1, m_APInt(Mask)) &&
1892  match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
1893  m_Value(XShifted)),
1894  m_Value(Y)))) {
1895  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1896  const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
1897  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1898  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1899  if (EffWidthY <= ShftCnt) {
1900  const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
1901  Q.DT);
1902  const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
1903  const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
1904  const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
1905  // If the mask is extracting all bits from X or Y as is, we can skip
1906  // this AND op.
1907  if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
1908  return Y;
1909  if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
1910  return XShifted;
1911  }
1912  }
1913 
1914  return nullptr;
1915 }
1916 
1919 }
1920 
1921 /// Given operands for an Or, see if we can fold the result.
1922 /// If not, this returns null.
1923 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1924  unsigned MaxRecurse) {
1925  if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1926  return C;
1927 
1928  // X | undef -> -1
1929  // X | -1 = -1
1930  // Do not return Op1 because it may contain undef elements if it's a vector.
1931  if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1932  return Constant::getAllOnesValue(Op0->getType());
1933 
1934  // X | X = X
1935  // X | 0 = X
1936  if (Op0 == Op1 || match(Op1, m_Zero()))
1937  return Op0;
1938 
1939  // A | ~A = ~A | A = -1
1940  if (match(Op0, m_Not(m_Specific(Op1))) ||
1941  match(Op1, m_Not(m_Specific(Op0))))
1942  return Constant::getAllOnesValue(Op0->getType());
1943 
1944  // (A & ?) | A = A
1945  if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1946  return Op1;
1947 
1948  // A | (A & ?) = A
1949  if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1950  return Op0;
1951 
1952  // ~(A & ?) | A = -1
1953  if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1954  return Constant::getAllOnesValue(Op1->getType());
1955 
1956  // A | ~(A & ?) = -1
1957  if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1958  return Constant::getAllOnesValue(Op0->getType());
1959 
1960  Value *A, *B;
1961  // (A & ~B) | (A ^ B) -> (A ^ B)
1962  // (~B & A) | (A ^ B) -> (A ^ B)
1963  // (A & ~B) | (B ^ A) -> (B ^ A)
1964  // (~B & A) | (B ^ A) -> (B ^ A)
1965  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1966  (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1967  match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1968  return Op1;
1969 
1970  // Commute the 'or' operands.
1971  // (A ^ B) | (A & ~B) -> (A ^ B)
1972  // (A ^ B) | (~B & A) -> (A ^ B)
1973  // (B ^ A) | (A & ~B) -> (B ^ A)
1974  // (B ^ A) | (~B & A) -> (B ^ A)
1975  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1976  (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1977  match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1978  return Op0;
1979 
1980  // (A & B) | (~A ^ B) -> (~A ^ B)
1981  // (B & A) | (~A ^ B) -> (~A ^ B)
1982  // (A & B) | (B ^ ~A) -> (B ^ ~A)
1983  // (B & A) | (B ^ ~A) -> (B ^ ~A)
1984  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1985  (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1986  match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1987  return Op1;
1988 
1989  // (~A ^ B) | (A & B) -> (~A ^ B)
1990  // (~A ^ B) | (B & A) -> (~A ^ B)
1991  // (B ^ ~A) | (A & B) -> (B ^ ~A)
1992  // (B ^ ~A) | (B & A) -> (B ^ ~A)
1993  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1994  (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1995  match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1996  return Op0;
1997 
1998  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
1999  return V;
2000 
2001  // Try some generic simplifications for associative operations.
2002  if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
2003  MaxRecurse))
2004  return V;
2005 
2006  // Or distributes over And. Try some generic simplifications based on this.
2007  if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2008  MaxRecurse))
2009  return V;
2010 
2011  // If the operation is with the result of a select instruction, check whether
2012  // operating on either branch of the select always yields the same value.
2013  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2014  if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2015  MaxRecurse))
2016  return V;
2017 
2018  // (A & C1)|(B & C2)
2019  const APInt *C1, *C2;
2020  if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2021  match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2022  if (*C1 == ~*C2) {
2023  // (A & C1)|(B & C2)
2024  // If we have: ((V + N) & C1) | (V & C2)
2025  // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2026  // replace with V+N.
2027  Value *N;
2028  if (C2->isMask() && // C2 == 0+1+
2029  match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2030  // Add commutes, try both ways.
2031  if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2032  return A;
2033  }
2034  // Or commutes, try both ways.
2035  if (C1->isMask() &&
2036  match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2037  // Add commutes, try both ways.
2038  if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2039  return B;
2040  }
2041  }
2042  }
2043 
2044  // If the operation is with the result of a phi instruction, check whether
2045  // operating on all incoming values of the phi always yields the same value.
2046  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2047  if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2048  return V;
2049 
2050  return nullptr;
2051 }
2052 
2055 }
2056 
2057 /// Given operands for a Xor, see if we can fold the result.
2058 /// If not, this returns null.
2059 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2060  unsigned MaxRecurse) {
2061  if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2062  return C;
2063 
2064  // A ^ undef -> undef
2065  if (match(Op1, m_Undef()))
2066  return Op1;
2067 
2068  // A ^ 0 = A
2069  if (match(Op1, m_Zero()))
2070  return Op0;
2071 
2072  // A ^ A = 0
2073  if (Op0 == Op1)
2074  return Constant::getNullValue(Op0->getType());
2075 
2076  // A ^ ~A = ~A ^ A = -1
2077  if (match(Op0, m_Not(m_Specific(Op1))) ||
2078  match(Op1, m_Not(m_Specific(Op0))))
2079  return Constant::getAllOnesValue(Op0->getType());
2080 
2081  // Try some generic simplifications for associative operations.
2082  if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2083  MaxRecurse))
2084  return V;
2085 
2086  // Threading Xor over selects and phi nodes is pointless, so don't bother.
2087  // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2088  // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2089  // only if B and C are equal. If B and C are equal then (since we assume
2090  // that operands have already been simplified) "select(cond, B, C)" should
2091  // have been simplified to the common value of B and C already. Analysing
2092  // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2093  // for threading over phi nodes.
2094 
2095  return nullptr;
2096 }
2097 
2100 }
2101 
2102 
2104  return CmpInst::makeCmpResultType(Op->getType());
2105 }
2106 
2107 /// Rummage around inside V looking for something equivalent to the comparison
2108 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2109 /// Helper function for analyzing max/min idioms.
2111  Value *LHS, Value *RHS) {
2113  if (!SI)
2114  return nullptr;
2115  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2116  if (!Cmp)
2117  return nullptr;
2118  Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2119  if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2120  return Cmp;
2121  if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2122  LHS == CmpRHS && RHS == CmpLHS)
2123  return Cmp;
2124  return nullptr;
2125 }
2126 
2127 // A significant optimization not implemented here is assuming that alloca
2128 // addresses are not equal to incoming argument values. They don't *alias*,
2129 // as we say, but that doesn't mean they aren't equal, so we take a
2130 // conservative approach.
2131 //
2132 // This is inspired in part by C++11 5.10p1:
2133 // "Two pointers of the same type compare equal if and only if they are both
2134 // null, both point to the same function, or both represent the same
2135 // address."
2136 //
2137 // This is pretty permissive.
2138 //
2139 // It's also partly due to C11 6.5.9p6:
2140 // "Two pointers compare equal if and only if both are null pointers, both are
2141 // pointers to the same object (including a pointer to an object and a
2142 // subobject at its beginning) or function, both are pointers to one past the
2143 // last element of the same array object, or one is a pointer to one past the
2144 // end of one array object and the other is a pointer to the start of a
2145 // different array object that happens to immediately follow the first array
2146 // object in the address space.)
2147 //
2148 // C11's version is more restrictive, however there's no reason why an argument
2149 // couldn't be a one-past-the-end value for a stack object in the caller and be
2150 // equal to the beginning of a stack object in the callee.
2151 //
2152 // If the C and C++ standards are ever made sufficiently restrictive in this
2153 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2154 // this optimization.
2155 static Constant *
2157  const DominatorTree *DT, CmpInst::Predicate Pred,
2158  AssumptionCache *AC, const Instruction *CxtI,
2159  const InstrInfoQuery &IIQ, Value *LHS, Value *RHS) {
2160  // First, skip past any trivial no-ops.
2161  LHS = LHS->stripPointerCasts();
2162  RHS = RHS->stripPointerCasts();
2163 
2164  // A non-null pointer is not equal to a null pointer.
2165  if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2166  IIQ.UseInstrInfo) &&
2167  isa<ConstantPointerNull>(RHS) &&
2168  (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2169  return ConstantInt::get(GetCompareTy(LHS),
2170  !CmpInst::isTrueWhenEqual(Pred));
2171 
2172  // We can only fold certain predicates on pointer comparisons.
2173  switch (Pred) {
2174  default:
2175  return nullptr;
2176 
2177  // Equality comaprisons are easy to fold.
2178  case CmpInst::ICMP_EQ:
2179  case CmpInst::ICMP_NE:
2180  break;
2181 
2182  // We can only handle unsigned relational comparisons because 'inbounds' on
2183  // a GEP only protects against unsigned wrapping.
2184  case CmpInst::ICMP_UGT:
2185  case CmpInst::ICMP_UGE:
2186  case CmpInst::ICMP_ULT:
2187  case CmpInst::ICMP_ULE:
2188  // However, we have to switch them to their signed variants to handle
2189  // negative indices from the base pointer.
2190  Pred = ICmpInst::getSignedPredicate(Pred);
2191  break;
2192  }
2193 
2194  // Strip off any constant offsets so that we can reason about them.
2195  // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2196  // here and compare base addresses like AliasAnalysis does, however there are
2197  // numerous hazards. AliasAnalysis and its utilities rely on special rules
2198  // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2199  // doesn't need to guarantee pointer inequality when it says NoAlias.
2200  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2201  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2202 
2203  // If LHS and RHS are related via constant offsets to the same base
2204  // value, we can replace it with an icmp which just compares the offsets.
2205  if (LHS == RHS)
2206  return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2207 
2208  // Various optimizations for (in)equality comparisons.
2209  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2210  // Different non-empty allocations that exist at the same time have
2211  // different addresses (if the program can tell). Global variables always
2212  // exist, so they always exist during the lifetime of each other and all
2213  // allocas. Two different allocas usually have different addresses...
2214  //
2215  // However, if there's an @llvm.stackrestore dynamically in between two
2216  // allocas, they may have the same address. It's tempting to reduce the
2217  // scope of the problem by only looking at *static* allocas here. That would
2218  // cover the majority of allocas while significantly reducing the likelihood
2219  // of having an @llvm.stackrestore pop up in the middle. However, it's not
2220  // actually impossible for an @llvm.stackrestore to pop up in the middle of
2221  // an entry block. Also, if we have a block that's not attached to a
2222  // function, we can't tell if it's "static" under the current definition.
2223  // Theoretically, this problem could be fixed by creating a new kind of
2224  // instruction kind specifically for static allocas. Such a new instruction
2225  // could be required to be at the top of the entry block, thus preventing it
2226  // from being subject to a @llvm.stackrestore. Instcombine could even
2227  // convert regular allocas into these special allocas. It'd be nifty.
2228  // However, until then, this problem remains open.
2229  //
2230  // So, we'll assume that two non-empty allocas have different addresses
2231  // for now.
2232  //
2233  // With all that, if the offsets are within the bounds of their allocations
2234  // (and not one-past-the-end! so we can't use inbounds!), and their
2235  // allocations aren't the same, the pointers are not equal.
2236  //
2237  // Note that it's not necessary to check for LHS being a global variable
2238  // address, due to canonicalization and constant folding.
2239  if (isa<AllocaInst>(LHS) &&
2240  (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2241  ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2242  ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2243  uint64_t LHSSize, RHSSize;
2244  ObjectSizeOpts Opts;
2245  Opts.NullIsUnknownSize =
2246  NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2247  if (LHSOffsetCI && RHSOffsetCI &&
2248  getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2249  getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2250  const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2251  const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2252  if (!LHSOffsetValue.isNegative() &&
2253  !RHSOffsetValue.isNegative() &&
2254  LHSOffsetValue.ult(LHSSize) &&
2255  RHSOffsetValue.ult(RHSSize)) {
2256  return ConstantInt::get(GetCompareTy(LHS),
2257  !CmpInst::isTrueWhenEqual(Pred));
2258  }
2259  }
2260 
2261  // Repeat the above check but this time without depending on DataLayout
2262  // or being able to compute a precise size.
2263  if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2264  !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2265  LHSOffset->isNullValue() &&
2266  RHSOffset->isNullValue())
2267  return ConstantInt::get(GetCompareTy(LHS),
2268  !CmpInst::isTrueWhenEqual(Pred));
2269  }
2270 
2271  // Even if an non-inbounds GEP occurs along the path we can still optimize
2272  // equality comparisons concerning the result. We avoid walking the whole
2273  // chain again by starting where the last calls to
2274  // stripAndComputeConstantOffsets left off and accumulate the offsets.
2275  Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2276  Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2277  if (LHS == RHS)
2278  return ConstantExpr::getICmp(Pred,
2279  ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2280  ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2281 
2282  // If one side of the equality comparison must come from a noalias call
2283  // (meaning a system memory allocation function), and the other side must
2284  // come from a pointer that cannot overlap with dynamically-allocated
2285  // memory within the lifetime of the current function (allocas, byval
2286  // arguments, globals), then determine the comparison result here.
2287  SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2288  GetUnderlyingObjects(LHS, LHSUObjs, DL);
2289  GetUnderlyingObjects(RHS, RHSUObjs, DL);
2290 
2291  // Is the set of underlying objects all noalias calls?
2292  auto IsNAC = [](ArrayRef<const Value *> Objects) {
2293  return all_of(Objects, isNoAliasCall);
2294  };
2295 
2296  // Is the set of underlying objects all things which must be disjoint from
2297  // noalias calls. For allocas, we consider only static ones (dynamic
2298  // allocas might be transformed into calls to malloc not simultaneously
2299  // live with the compared-to allocation). For globals, we exclude symbols
2300  // that might be resolve lazily to symbols in another dynamically-loaded
2301  // library (and, thus, could be malloc'ed by the implementation).
2302  auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2303  return all_of(Objects, [](const Value *V) {
2304  if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2305  return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2306  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2307  return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2308  GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2309  !GV->isThreadLocal();
2310  if (const Argument *A = dyn_cast<Argument>(V))
2311  return A->hasByValAttr();
2312  return false;
2313  });
2314  };
2315 
2316  if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2317  (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2318  return ConstantInt::get(GetCompareTy(LHS),
2319  !CmpInst::isTrueWhenEqual(Pred));
2320 
2321  // Fold comparisons for non-escaping pointer even if the allocation call
2322  // cannot be elided. We cannot fold malloc comparison to null. Also, the
2323  // dynamic allocation call could be either of the operands.
2324  Value *MI = nullptr;
2325  if (isAllocLikeFn(LHS, TLI) &&
2326  llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2327  MI = LHS;
2328  else if (isAllocLikeFn(RHS, TLI) &&
2329  llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2330  MI = RHS;
2331  // FIXME: We should also fold the compare when the pointer escapes, but the
2332  // compare dominates the pointer escape
2333  if (MI && !PointerMayBeCaptured(MI, true, true))
2334  return ConstantInt::get(GetCompareTy(LHS),
2336  }
2337 
2338  // Otherwise, fail.
2339  return nullptr;
2340 }
2341 
2342 /// Fold an icmp when its operands have i1 scalar type.
2344  Value *RHS, const SimplifyQuery &Q) {
2345  Type *ITy = GetCompareTy(LHS); // The return type.
2346  Type *OpTy = LHS->getType(); // The operand type.
2347  if (!OpTy->isIntOrIntVectorTy(1))
2348  return nullptr;
2349 
2350  // A boolean compared to true/false can be simplified in 14 out of the 20
2351  // (10 predicates * 2 constants) possible combinations. Cases not handled here
2352  // require a 'not' of the LHS, so those must be transformed in InstCombine.
2353  if (match(RHS, m_Zero())) {
2354  switch (Pred) {
2355  case CmpInst::ICMP_NE: // X != 0 -> X
2356  case CmpInst::ICMP_UGT: // X >u 0 -> X
2357  case CmpInst::ICMP_SLT: // X <s 0 -> X
2358  return LHS;
2359 
2360  case CmpInst::ICMP_ULT: // X <u 0 -> false
2361  case CmpInst::ICMP_SGT: // X >s 0 -> false
2362  return getFalse(ITy);
2363 
2364  case CmpInst::ICMP_UGE: // X >=u 0 -> true
2365  case CmpInst::ICMP_SLE: // X <=s 0 -> true
2366  return getTrue(ITy);
2367 
2368  default: break;
2369  }
2370  } else if (match(RHS, m_One())) {
2371  switch (Pred) {
2372  case CmpInst::ICMP_EQ: // X == 1 -> X
2373  case CmpInst::ICMP_UGE: // X >=u 1 -> X
2374  case CmpInst::ICMP_SLE: // X <=s -1 -> X
2375  return LHS;
2376 
2377  case CmpInst::ICMP_UGT: // X >u 1 -> false
2378  case CmpInst::ICMP_SLT: // X <s -1 -> false
2379  return getFalse(ITy);
2380 
2381  case CmpInst::ICMP_ULE: // X <=u 1 -> true
2382  case CmpInst::ICMP_SGE: // X >=s -1 -> true
2383  return getTrue(ITy);
2384 
2385  default: break;
2386  }
2387  }
2388 
2389  switch (Pred) {
2390  default:
2391  break;
2392  case ICmpInst::ICMP_UGE:
2393  if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2394  return getTrue(ITy);
2395  break;
2396  case ICmpInst::ICMP_SGE:
2397  /// For signed comparison, the values for an i1 are 0 and -1
2398  /// respectively. This maps into a truth table of:
2399  /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2400  /// 0 | 0 | 1 (0 >= 0) | 1
2401  /// 0 | 1 | 1 (0 >= -1) | 1
2402  /// 1 | 0 | 0 (-1 >= 0) | 0
2403  /// 1 | 1 | 1 (-1 >= -1) | 1
2404  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2405  return getTrue(ITy);
2406  break;
2407  case ICmpInst::ICMP_ULE:
2408  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2409  return getTrue(ITy);
2410  break;
2411  }
2412 
2413  return nullptr;
2414 }
2415 
2416 /// Try hard to fold icmp with zero RHS because this is a common case.
2418  Value *RHS, const SimplifyQuery &Q) {
2419  if (!match(RHS, m_Zero()))
2420  return nullptr;
2421 
2422  Type *ITy = GetCompareTy(LHS); // The return type.
2423  switch (Pred) {
2424  default:
2425  llvm_unreachable("Unknown ICmp predicate!");
2426  case ICmpInst::ICMP_ULT:
2427  return getFalse(ITy);
2428  case ICmpInst::ICMP_UGE:
2429  return getTrue(ITy);
2430  case ICmpInst::ICMP_EQ:
2431  case ICmpInst::ICMP_ULE:
2432  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2433  return getFalse(ITy);
2434  break;
2435  case ICmpInst::ICMP_NE:
2436  case ICmpInst::ICMP_UGT:
2437  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2438  return getTrue(ITy);
2439  break;
2440  case ICmpInst::ICMP_SLT: {
2441  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2442  if (LHSKnown.isNegative())
2443  return getTrue(ITy);
2444  if (LHSKnown.isNonNegative())
2445  return getFalse(ITy);
2446  break;
2447  }
2448  case ICmpInst::ICMP_SLE: {
2449  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2450  if (LHSKnown.isNegative())
2451  return getTrue(ITy);
2452  if (LHSKnown.isNonNegative() &&
2453  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2454  return getFalse(ITy);
2455  break;
2456  }
2457  case ICmpInst::ICMP_SGE: {
2458  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2459  if (LHSKnown.isNegative())
2460  return getFalse(ITy);
2461  if (LHSKnown.isNonNegative())
2462  return getTrue(ITy);
2463  break;
2464  }
2465  case ICmpInst::ICMP_SGT: {
2466  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2467  if (LHSKnown.isNegative())
2468  return getFalse(ITy);
2469  if (LHSKnown.isNonNegative() &&
2470  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2471  return getTrue(ITy);
2472  break;
2473  }
2474  }
2475 
2476  return nullptr;
2477 }
2478 
2480  Value *RHS, const InstrInfoQuery &IIQ) {
2481  Type *ITy = GetCompareTy(RHS); // The return type.
2482 
2483  Value *X;
2484  // Sign-bit checks can be optimized to true/false after unsigned
2485  // floating-point casts:
2486  // icmp slt (bitcast (uitofp X)), 0 --> false
2487  // icmp sgt (bitcast (uitofp X)), -1 --> true
2488  if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2489  if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2490  return ConstantInt::getFalse(ITy);
2491  if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2492  return ConstantInt::getTrue(ITy);
2493  }
2494 
2495  const APInt *C;
2496  if (!match(RHS, m_APInt(C)))
2497  return nullptr;
2498 
2499  // Rule out tautological comparisons (eg., ult 0 or uge 0).
2501  if (RHS_CR.isEmptySet())
2502  return ConstantInt::getFalse(ITy);
2503  if (RHS_CR.isFullSet())
2504  return ConstantInt::getTrue(ITy);
2505 
2506  ConstantRange LHS_CR = computeConstantRange(LHS, IIQ.UseInstrInfo);
2507  if (!LHS_CR.isFullSet()) {
2508  if (RHS_CR.contains(LHS_CR))
2509  return ConstantInt::getTrue(ITy);
2510  if (RHS_CR.inverse().contains(LHS_CR))
2511  return ConstantInt::getFalse(ITy);
2512  }
2513 
2514  return nullptr;
2515 }
2516 
2517 /// TODO: A large part of this logic is duplicated in InstCombine's
2518 /// foldICmpBinOp(). We should be able to share that and avoid the code
2519 /// duplication.
2521  Value *RHS, const SimplifyQuery &Q,
2522  unsigned MaxRecurse) {
2523  Type *ITy = GetCompareTy(LHS); // The return type.
2524 
2525  BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2526  BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2527  if (MaxRecurse && (LBO || RBO)) {
2528  // Analyze the case when either LHS or RHS is an add instruction.
2529  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2530  // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2531  bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2532  if (LBO && LBO->getOpcode() == Instruction::Add) {
2533  A = LBO->getOperand(0);
2534  B = LBO->getOperand(1);
2535  NoLHSWrapProblem =
2536  ICmpInst::isEquality(Pred) ||
2537  (CmpInst::isUnsigned(Pred) &&
2538  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2539  (CmpInst::isSigned(Pred) &&
2540  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2541  }
2542  if (RBO && RBO->getOpcode() == Instruction::Add) {
2543  C = RBO->getOperand(0);
2544  D = RBO->getOperand(1);
2545  NoRHSWrapProblem =
2546  ICmpInst::isEquality(Pred) ||
2547  (CmpInst::isUnsigned(Pred) &&
2548  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2549  (CmpInst::isSigned(Pred) &&
2550  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2551  }
2552 
2553  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2554  if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2555  if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2556  Constant::getNullValue(RHS->getType()), Q,
2557  MaxRecurse - 1))
2558  return V;
2559 
2560  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2561  if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2562  if (Value *V =
2564  C == LHS ? D : C, Q, MaxRecurse - 1))
2565  return V;
2566 
2567  // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2568  if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2569  NoRHSWrapProblem) {
2570  // Determine Y and Z in the form icmp (X+Y), (X+Z).
2571  Value *Y, *Z;
2572  if (A == C) {
2573  // C + B == C + D -> B == D
2574  Y = B;
2575  Z = D;
2576  } else if (A == D) {
2577  // D + B == C + D -> B == C
2578  Y = B;
2579  Z = C;
2580  } else if (B == C) {
2581  // A + C == C + D -> A == D
2582  Y = A;
2583  Z = D;
2584  } else {
2585  assert(B == D);
2586  // A + D == C + D -> A == C
2587  Y = A;
2588  Z = C;
2589  }
2590  if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2591  return V;
2592  }
2593  }
2594 
2595  {
2596  Value *Y = nullptr;
2597  // icmp pred (or X, Y), X
2598  if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2599  if (Pred == ICmpInst::ICMP_ULT)
2600  return getFalse(ITy);
2601  if (Pred == ICmpInst::ICMP_UGE)
2602  return getTrue(ITy);
2603 
2604  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2605  KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2606  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2607  if (RHSKnown.isNonNegative() && YKnown.isNegative())
2608  return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2609  if (RHSKnown.isNegative() || YKnown.isNonNegative())
2610  return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2611  }
2612  }
2613  // icmp pred X, (or X, Y)
2614  if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2615  if (Pred == ICmpInst::ICMP_ULE)
2616  return getTrue(ITy);
2617  if (Pred == ICmpInst::ICMP_UGT)
2618  return getFalse(ITy);
2619 
2620  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2621  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2622  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2623  if (LHSKnown.isNonNegative() && YKnown.isNegative())
2624  return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2625  if (LHSKnown.isNegative() || YKnown.isNonNegative())
2626  return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2627  }
2628  }
2629  }
2630 
2631  // icmp pred (and X, Y), X
2632  if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2633  if (Pred == ICmpInst::ICMP_UGT)
2634  return getFalse(ITy);
2635  if (Pred == ICmpInst::ICMP_ULE)
2636  return getTrue(ITy);
2637  }
2638  // icmp pred X, (and X, Y)
2639  if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2640  if (Pred == ICmpInst::ICMP_UGE)
2641  return getTrue(ITy);
2642  if (Pred == ICmpInst::ICMP_ULT)
2643  return getFalse(ITy);
2644  }
2645 
2646  // 0 - (zext X) pred C
2647  if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2648  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2649  if (RHSC->getValue().isStrictlyPositive()) {
2650  if (Pred == ICmpInst::ICMP_SLT)
2651  return ConstantInt::getTrue(RHSC->getContext());
2652  if (Pred == ICmpInst::ICMP_SGE)
2653  return ConstantInt::getFalse(RHSC->getContext());
2654  if (Pred == ICmpInst::ICMP_EQ)
2655  return ConstantInt::getFalse(RHSC->getContext());
2656  if (Pred == ICmpInst::ICMP_NE)
2657  return ConstantInt::getTrue(RHSC->getContext());
2658  }
2659  if (RHSC->getValue().isNonNegative()) {
2660  if (Pred == ICmpInst::ICMP_SLE)
2661  return ConstantInt::getTrue(RHSC->getContext());
2662  if (Pred == ICmpInst::ICMP_SGT)
2663  return ConstantInt::getFalse(RHSC->getContext());
2664  }
2665  }
2666  }
2667 
2668  // icmp pred (urem X, Y), Y
2669  if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2670  switch (Pred) {
2671  default:
2672  break;
2673  case ICmpInst::ICMP_SGT:
2674  case ICmpInst::ICMP_SGE: {
2675  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2676  if (!Known.isNonNegative())
2677  break;
2679  }
2680  case ICmpInst::ICMP_EQ:
2681  case ICmpInst::ICMP_UGT:
2682  case ICmpInst::ICMP_UGE:
2683  return getFalse(ITy);
2684  case ICmpInst::ICMP_SLT:
2685  case ICmpInst::ICMP_SLE: {
2686  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2687  if (!Known.isNonNegative())
2688  break;
2690  }
2691  case ICmpInst::ICMP_NE:
2692  case ICmpInst::ICMP_ULT:
2693  case ICmpInst::ICMP_ULE:
2694  return getTrue(ITy);
2695  }
2696  }
2697 
2698  // icmp pred X, (urem Y, X)
2699  if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2700  switch (Pred) {
2701  default:
2702  break;
2703  case ICmpInst::ICMP_SGT:
2704  case ICmpInst::ICMP_SGE: {
2705  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2706  if (!Known.isNonNegative())
2707  break;
2709  }
2710  case ICmpInst::ICMP_NE:
2711  case ICmpInst::ICMP_UGT:
2712  case ICmpInst::ICMP_UGE:
2713  return getTrue(ITy);
2714  case ICmpInst::ICMP_SLT:
2715  case ICmpInst::ICMP_SLE: {
2716  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2717  if (!Known.isNonNegative())
2718  break;
2720  }
2721  case ICmpInst::ICMP_EQ:
2722  case ICmpInst::ICMP_ULT:
2723  case ICmpInst::ICMP_ULE:
2724  return getFalse(ITy);
2725  }
2726  }
2727 
2728  // x >> y <=u x
2729  // x udiv y <=u x.
2730  if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2731  match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2732  // icmp pred (X op Y), X
2733  if (Pred == ICmpInst::ICMP_UGT)
2734  return getFalse(ITy);
2735  if (Pred == ICmpInst::ICMP_ULE)
2736  return getTrue(ITy);
2737  }
2738 
2739  // x >=u x >> y
2740  // x >=u x udiv y.
2741  if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2742  match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2743  // icmp pred X, (X op Y)
2744  if (Pred == ICmpInst::ICMP_ULT)
2745  return getFalse(ITy);
2746  if (Pred == ICmpInst::ICMP_UGE)
2747  return getTrue(ITy);
2748  }
2749 
2750  // handle:
2751  // CI2 << X == CI
2752  // CI2 << X != CI
2753  //
2754  // where CI2 is a power of 2 and CI isn't
2755  if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2756  const APInt *CI2Val, *CIVal = &CI->getValue();
2757  if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2758  CI2Val->isPowerOf2()) {
2759  if (!CIVal->isPowerOf2()) {
2760  // CI2 << X can equal zero in some circumstances,
2761  // this simplification is unsafe if CI is zero.
2762  //
2763  // We know it is safe if:
2764  // - The shift is nsw, we can't shift out the one bit.
2765  // - The shift is nuw, we can't shift out the one bit.
2766  // - CI2 is one
2767  // - CI isn't zero
2768  if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2769  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2770  CI2Val->isOneValue() || !CI->isZero()) {
2771  if (Pred == ICmpInst::ICMP_EQ)
2772  return ConstantInt::getFalse(RHS->getContext());
2773  if (Pred == ICmpInst::ICMP_NE)
2774  return ConstantInt::getTrue(RHS->getContext());
2775  }
2776  }
2777  if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2778  if (Pred == ICmpInst::ICMP_UGT)
2779  return ConstantInt::getFalse(RHS->getContext());
2780  if (Pred == ICmpInst::ICMP_ULE)
2781  return ConstantInt::getTrue(RHS->getContext());
2782  }
2783  }
2784  }
2785 
2786  if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2787  LBO->getOperand(1) == RBO->getOperand(1)) {
2788  switch (LBO->getOpcode()) {
2789  default:
2790  break;
2791  case Instruction::UDiv:
2792  case Instruction::LShr:
2793  if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2794  !Q.IIQ.isExact(RBO))
2795  break;
2796  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2797  RBO->getOperand(0), Q, MaxRecurse - 1))
2798  return V;
2799  break;
2800  case Instruction::SDiv:
2801  if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2802  !Q.IIQ.isExact(RBO))
2803  break;
2804  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2805  RBO->getOperand(0), Q, MaxRecurse - 1))
2806  return V;
2807  break;
2808  case Instruction::AShr:
2809  if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2810  break;
2811  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2812  RBO->getOperand(0), Q, MaxRecurse - 1))
2813  return V;
2814  break;
2815  case Instruction::Shl: {
2816  bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
2817  bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
2818  if (!NUW && !NSW)
2819  break;
2820  if (!NSW && ICmpInst::isSigned(Pred))
2821  break;
2822  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2823  RBO->getOperand(0), Q, MaxRecurse - 1))
2824  return V;
2825  break;
2826  }
2827  }
2828  }
2829  return nullptr;
2830 }
2831 
2832 /// Simplify integer comparisons where at least one operand of the compare
2833 /// matches an integer min/max idiom.
2835  Value *RHS, const SimplifyQuery &Q,
2836  unsigned MaxRecurse) {
2837  Type *ITy = GetCompareTy(LHS); // The return type.
2838  Value *A, *B;
2840  CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2841 
2842  // Signed variants on "max(a,b)>=a -> true".
2843  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2844  if (A != RHS)
2845  std::swap(A, B); // smax(A, B) pred A.
2846  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2847  // We analyze this as smax(A, B) pred A.
2848  P = Pred;
2849  } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2850  (A == LHS || B == LHS)) {
2851  if (A != LHS)
2852  std::swap(A, B); // A pred smax(A, B).
2853  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2854  // We analyze this as smax(A, B) swapped-pred A.
2855  P = CmpInst::getSwappedPredicate(Pred);
2856  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2857  (A == RHS || B == RHS)) {
2858  if (A != RHS)
2859  std::swap(A, B); // smin(A, B) pred A.
2860  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2861  // We analyze this as smax(-A, -B) swapped-pred -A.
2862  // Note that we do not need to actually form -A or -B thanks to EqP.
2863  P = CmpInst::getSwappedPredicate(Pred);
2864  } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2865  (A == LHS || B == LHS)) {
2866  if (A != LHS)
2867  std::swap(A, B); // A pred smin(A, B).
2868  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2869  // We analyze this as smax(-A, -B) pred -A.
2870  // Note that we do not need to actually form -A or -B thanks to EqP.
2871  P = Pred;
2872  }
2873  if (P != CmpInst::BAD_ICMP_PREDICATE) {
2874  // Cases correspond to "max(A, B) p A".
2875  switch (P) {
2876  default:
2877  break;
2878  case CmpInst::ICMP_EQ:
2879  case CmpInst::ICMP_SLE:
2880  // Equivalent to "A EqP B". This may be the same as the condition tested
2881  // in the max/min; if so, we can just return that.
2882  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2883  return V;
2884  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2885  return V;
2886  // Otherwise, see if "A EqP B" simplifies.
2887  if (MaxRecurse)
2888  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2889  return V;
2890  break;
2891  case CmpInst::ICMP_NE:
2892  case CmpInst::ICMP_SGT: {
2894  // Equivalent to "A InvEqP B". This may be the same as the condition
2895  // tested in the max/min; if so, we can just return that.
2896  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2897  return V;
2898  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2899  return V;
2900  // Otherwise, see if "A InvEqP B" simplifies.
2901  if (MaxRecurse)
2902  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2903  return V;
2904  break;
2905  }
2906  case CmpInst::ICMP_SGE:
2907  // Always true.
2908  return getTrue(ITy);
2909  case CmpInst::ICMP_SLT:
2910  // Always false.
2911  return getFalse(ITy);
2912  }
2913  }
2914 
2915  // Unsigned variants on "max(a,b)>=a -> true".
2917  if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2918  if (A != RHS)
2919  std::swap(A, B); // umax(A, B) pred A.
2920  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2921  // We analyze this as umax(A, B) pred A.
2922  P = Pred;
2923  } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2924  (A == LHS || B == LHS)) {
2925  if (A != LHS)
2926  std::swap(A, B); // A pred umax(A, B).
2927  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2928  // We analyze this as umax(A, B) swapped-pred A.
2929  P = CmpInst::getSwappedPredicate(Pred);
2930  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2931  (A == RHS || B == RHS)) {
2932  if (A != RHS)
2933  std::swap(A, B); // umin(A, B) pred A.
2934  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2935  // We analyze this as umax(-A, -B) swapped-pred -A.
2936  // Note that we do not need to actually form -A or -B thanks to EqP.
2937  P = CmpInst::getSwappedPredicate(Pred);
2938  } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2939  (A == LHS || B == LHS)) {
2940  if (A != LHS)
2941  std::swap(A, B); // A pred umin(A, B).
2942  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2943  // We analyze this as umax(-A, -B) pred -A.
2944  // Note that we do not need to actually form -A or -B thanks to EqP.
2945  P = Pred;
2946  }
2947  if (P != CmpInst::BAD_ICMP_PREDICATE) {
2948  // Cases correspond to "max(A, B) p A".
2949  switch (P) {
2950  default:
2951  break;
2952  case CmpInst::ICMP_EQ:
2953  case CmpInst::ICMP_ULE:
2954  // Equivalent to "A EqP B". This may be the same as the condition tested
2955  // in the max/min; if so, we can just return that.
2956  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2957  return V;
2958  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2959  return V;
2960  // Otherwise, see if "A EqP B" simplifies.
2961  if (MaxRecurse)
2962  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2963  return V;
2964  break;
2965  case CmpInst::ICMP_NE:
2966  case CmpInst::ICMP_UGT: {
2968  // Equivalent to "A InvEqP B". This may be the same as the condition
2969  // tested in the max/min; if so, we can just return that.
2970  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2971  return V;
2972  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2973  return V;
2974  // Otherwise, see if "A InvEqP B" simplifies.
2975  if (MaxRecurse)
2976  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2977  return V;
2978  break;
2979  }
2980  case CmpInst::ICMP_UGE:
2981  // Always true.
2982  return getTrue(ITy);
2983  case CmpInst::ICMP_ULT:
2984  // Always false.
2985  return getFalse(ITy);
2986  }
2987  }
2988 
2989  // Variants on "max(x,y) >= min(x,z)".
2990  Value *C, *D;
2991  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2992  match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2993  (A == C || A == D || B == C || B == D)) {
2994  // max(x, ?) pred min(x, ?).
2995  if (Pred == CmpInst::ICMP_SGE)
2996  // Always true.
2997  return getTrue(ITy);
2998  if (Pred == CmpInst::ICMP_SLT)
2999  // Always false.
3000  return getFalse(ITy);
3001  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3002  match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3003  (A == C || A == D || B == C || B == D)) {
3004  // min(x, ?) pred max(x, ?).
3005  if (Pred == CmpInst::ICMP_SLE)
3006  // Always true.
3007  return getTrue(ITy);
3008  if (Pred == CmpInst::ICMP_SGT)
3009  // Always false.
3010  return getFalse(ITy);
3011  } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3012  match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3013  (A == C || A == D || B == C || B == D)) {
3014  // max(x, ?) pred min(x, ?).
3015  if (Pred == CmpInst::ICMP_UGE)
3016  // Always true.
3017  return getTrue(ITy);
3018  if (Pred == CmpInst::ICMP_ULT)
3019  // Always false.
3020  return getFalse(ITy);
3021  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3022  match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3023  (A == C || A == D || B == C || B == D)) {
3024  // min(x, ?) pred max(x, ?).
3025  if (Pred == CmpInst::ICMP_ULE)
3026  // Always true.
3027  return getTrue(ITy);
3028  if (Pred == CmpInst::ICMP_UGT)
3029  // Always false.
3030  return getFalse(ITy);
3031  }
3032 
3033  return nullptr;
3034 }
3035 
3036 /// Given operands for an ICmpInst, see if we can fold the result.
3037 /// If not, this returns null.
3038 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3039  const SimplifyQuery &Q, unsigned MaxRecurse) {
3040  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3041  assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3042 
3043  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3044  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3045  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3046 
3047  // If we have a constant, make sure it is on the RHS.
3048  std::swap(LHS, RHS);
3049  Pred = CmpInst::getSwappedPredicate(Pred);
3050  }
3051  assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3052 
3053  Type *ITy = GetCompareTy(LHS); // The return type.
3054 
3055  // For EQ and NE, we can always pick a value for the undef to make the
3056  // predicate pass or fail, so we can return undef.
3057  // Matches behavior in llvm::ConstantFoldCompareInstruction.
3058  if (isa<UndefValue>(RHS) && ICmpInst::isEquality(Pred))
3059  return UndefValue::get(ITy);
3060 
3061  // icmp X, X -> true/false
3062  // icmp X, undef -> true/false because undef could be X.
3063  if (LHS == RHS || isa<UndefValue>(RHS))
3064  return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3065 
3066  if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3067  return V;
3068 
3069  if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3070  return V;
3071 
3072  if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3073  return V;
3074 
3075  // If both operands have range metadata, use the metadata
3076  // to simplify the comparison.
3077  if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3078  auto RHS_Instr = cast<Instruction>(RHS);
3079  auto LHS_Instr = cast<Instruction>(LHS);
3080 
3081  if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3082  Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3083  auto RHS_CR = getConstantRangeFromMetadata(
3084  *RHS_Instr->getMetadata(LLVMContext::MD_range));
3085  auto LHS_CR = getConstantRangeFromMetadata(
3086  *LHS_Instr->getMetadata(LLVMContext::MD_range));
3087 
3088  auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3089  if (Satisfied_CR.contains(LHS_CR))
3090  return ConstantInt::getTrue(RHS->getContext());
3091 
3092  auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3093  CmpInst::getInversePredicate(Pred), RHS_CR);
3094  if (InversedSatisfied_CR.contains(LHS_CR))
3095  return ConstantInt::getFalse(RHS->getContext());
3096  }
3097  }
3098 
3099  // Compare of cast, for example (zext X) != 0 -> X != 0
3100  if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3101  Instruction *LI = cast<CastInst>(LHS);
3102  Value *SrcOp = LI->getOperand(0);
3103  Type *SrcTy = SrcOp->getType();
3104  Type *DstTy = LI->getType();
3105 
3106  // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3107  // if the integer type is the same size as the pointer type.
3108  if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3109  Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3110  if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3111  // Transfer the cast to the constant.
3112  if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3113  ConstantExpr::getIntToPtr(RHSC, SrcTy),
3114  Q, MaxRecurse-1))
3115  return V;
3116  } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3117  if (RI->getOperand(0)->getType() == SrcTy)
3118  // Compare without the cast.
3119  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3120  Q, MaxRecurse-1))
3121  return V;
3122  }
3123  }
3124 
3125  if (isa<ZExtInst>(LHS)) {
3126  // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3127  // same type.
3128  if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3129  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3130  // Compare X and Y. Note that signed predicates become unsigned.
3132  SrcOp, RI->getOperand(0), Q,
3133  MaxRecurse-1))
3134  return V;
3135  }
3136  // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3137  // too. If not, then try to deduce the result of the comparison.
3138  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3139  // Compute the constant that would happen if we truncated to SrcTy then
3140  // reextended to DstTy.
3141  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3142  Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3143 
3144  // If the re-extended constant didn't change then this is effectively
3145  // also a case of comparing two zero-extended values.
3146  if (RExt == CI && MaxRecurse)
3148  SrcOp, Trunc, Q, MaxRecurse-1))
3149  return V;
3150 
3151  // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3152  // there. Use this to work out the result of the comparison.
3153  if (RExt != CI) {
3154  switch (Pred) {
3155  default: llvm_unreachable("Unknown ICmp predicate!");
3156  // LHS <u RHS.
3157  case ICmpInst::ICMP_EQ:
3158  case ICmpInst::ICMP_UGT:
3159  case ICmpInst::ICMP_UGE:
3160  return ConstantInt::getFalse(CI->getContext());
3161 
3162  case ICmpInst::ICMP_NE:
3163  case ICmpInst::ICMP_ULT:
3164  case ICmpInst::ICMP_ULE:
3165  return ConstantInt::getTrue(CI->getContext());
3166 
3167  // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3168  // is non-negative then LHS <s RHS.
3169  case ICmpInst::ICMP_SGT:
3170  case ICmpInst::ICMP_SGE:
3171  return CI->getValue().isNegative() ?
3172  ConstantInt::getTrue(CI->getContext()) :
3173  ConstantInt::getFalse(CI->getContext());
3174 
3175  case ICmpInst::ICMP_SLT:
3176  case ICmpInst::ICMP_SLE:
3177  return CI->getValue().isNegative() ?
3178  ConstantInt::getFalse(CI->getContext()) :
3179  ConstantInt::getTrue(CI->getContext());
3180  }
3181  }
3182  }
3183  }
3184 
3185  if (isa<SExtInst>(LHS)) {
3186  // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3187  // same type.
3188  if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3189  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3190  // Compare X and Y. Note that the predicate does not change.
3191  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3192  Q, MaxRecurse-1))
3193  return V;
3194  }
3195  // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3196  // too. If not, then try to deduce the result of the comparison.
3197  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3198  // Compute the constant that would happen if we truncated to SrcTy then
3199  // reextended to DstTy.
3200  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3201  Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3202 
3203  // If the re-extended constant didn't change then this is effectively
3204  // also a case of comparing two sign-extended values.
3205  if (RExt == CI && MaxRecurse)
3206  if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3207  return V;
3208 
3209  // Otherwise the upper bits of LHS are all equal, while RHS has varying
3210  // bits there. Use this to work out the result of the comparison.
3211  if (RExt != CI) {
3212  switch (Pred) {
3213  default: llvm_unreachable("Unknown ICmp predicate!");
3214  case ICmpInst::ICMP_EQ:
3215  return ConstantInt::getFalse(CI->getContext());
3216  case ICmpInst::ICMP_NE:
3217  return ConstantInt::getTrue(CI->getContext());
3218 
3219  // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3220  // LHS >s RHS.
3221  case ICmpInst::ICMP_SGT:
3222  case ICmpInst::ICMP_SGE:
3223  return CI->getValue().isNegative() ?
3224  ConstantInt::getTrue(CI->getContext()) :
3225  ConstantInt::getFalse(CI->getContext());
3226  case ICmpInst::ICMP_SLT:
3227  case ICmpInst::ICMP_SLE:
3228  return CI->getValue().isNegative() ?
3229  ConstantInt::getFalse(CI->getContext()) :
3230  ConstantInt::getTrue(CI->getContext());
3231 
3232  // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3233  // LHS >u RHS.
3234  case ICmpInst::ICMP_UGT:
3235  case ICmpInst::ICMP_UGE:
3236  // Comparison is true iff the LHS <s 0.
3237  if (MaxRecurse)
3238  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3239  Constant::getNullValue(SrcTy),
3240  Q, MaxRecurse-1))
3241  return V;
3242  break;
3243  case ICmpInst::ICMP_ULT:
3244  case ICmpInst::ICMP_ULE:
3245  // Comparison is true iff the LHS >=s 0.
3246  if (MaxRecurse)
3247  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3248  Constant::getNullValue(SrcTy),
3249  Q, MaxRecurse-1))
3250  return V;
3251  break;
3252  }
3253  }
3254  }
3255  }
3256  }
3257 
3258  // icmp eq|ne X, Y -> false|true if X != Y
3259  if (ICmpInst::isEquality(Pred) &&
3260  isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3261  return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3262  }
3263 
3264  if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3265  return V;
3266 
3267  if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3268  return V;
3269 
3270  // Simplify comparisons of related pointers using a powerful, recursive
3271  // GEP-walk when we have target data available..
3272  if (LHS->getType()->isPointerTy())
3273  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3274  Q.IIQ, LHS, RHS))
3275  return C;
3276  if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3277  if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3278  if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3279  Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3280  Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3281  Q.DL.getTypeSizeInBits(CRHS->getType()))
3282  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3283  Q.IIQ, CLHS->getPointerOperand(),
3284  CRHS->getPointerOperand()))
3285  return C;
3286 
3287  if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3288  if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3289  if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3290  GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3291  (ICmpInst::isEquality(Pred) ||
3292  (GLHS->isInBounds() && GRHS->isInBounds() &&
3293  Pred == ICmpInst::getSignedPredicate(Pred)))) {
3294  // The bases are equal and the indices are constant. Build a constant
3295  // expression GEP with the same indices and a null base pointer to see
3296  // what constant folding can make out of it.
3297  Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3298  SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3300  GLHS->getSourceElementType(), Null, IndicesLHS);
3301 
3302  SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3304  GLHS->getSourceElementType(), Null, IndicesRHS);
3305  return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3306  }
3307  }
3308  }
3309 
3310  // If the comparison is with the result of a select instruction, check whether
3311  // comparing with either branch of the select always yields the same value.
3312  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3313  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3314  return V;
3315 
3316  // If the comparison is with the result of a phi instruction, check whether
3317  // doing the compare with each incoming phi value yields a common result.
3318  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3319  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3320  return V;
3321 
3322  return nullptr;
3323 }
3324 
3325 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3326  const SimplifyQuery &Q) {
3327  return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3328 }
3329 
3330 /// Given operands for an FCmpInst, see if we can fold the result.
3331 /// If not, this returns null.
3332 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3333  FastMathFlags FMF, const SimplifyQuery &Q,
3334  unsigned MaxRecurse) {
3335  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3336  assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3337 
3338  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3339  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3340  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3341 
3342  // If we have a constant, make sure it is on the RHS.
3343  std::swap(LHS, RHS);
3344  Pred = CmpInst::getSwappedPredicate(Pred);
3345  }
3346 
3347  // Fold trivial predicates.
3348  Type *RetTy = GetCompareTy(LHS);
3349  if (Pred == FCmpInst::FCMP_FALSE)
3350  return getFalse(RetTy);
3351  if (Pred == FCmpInst::FCMP_TRUE)
3352  return getTrue(RetTy);
3353 
3354  // Fold (un)ordered comparison if we can determine there are no NaNs.
3355  if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3356  if (FMF.noNaNs() ||
3357  (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3358  return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3359 
3360  // NaN is unordered; NaN is not ordered.
3362  "Comparison must be either ordered or unordered");
3363  if (match(RHS, m_NaN()))
3364  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3365 
3366  // fcmp pred x, undef and fcmp pred undef, x
3367  // fold to true if unordered, false if ordered
3368  if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3369  // Choosing NaN for the undef will always make unordered comparison succeed
3370  // and ordered comparison fail.
3371  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3372  }
3373 
3374  // fcmp x,x -> true/false. Not all compares are foldable.
3375  if (LHS == RHS) {
3376  if (CmpInst::isTrueWhenEqual(Pred))
3377  return getTrue(RetTy);
3378  if (CmpInst::isFalseWhenEqual(Pred))
3379  return getFalse(RetTy);
3380  }
3381 
3382  // Handle fcmp with constant RHS.
3383  // TODO: Use match with a specific FP value, so these work with vectors with
3384  // undef lanes.
3385  const APFloat *C;
3386  if (match(RHS, m_APFloat(C))) {
3387  // Check whether the constant is an infinity.
3388  if (C->isInfinity()) {
3389  if (C->isNegative()) {
3390  switch (Pred) {
3391  case FCmpInst::FCMP_OLT:
3392  // No value is ordered and less than negative infinity.
3393  return getFalse(RetTy);
3394  case FCmpInst::FCMP_UGE:
3395  // All values are unordered with or at least negative infinity.
3396  return getTrue(RetTy);
3397  default:
3398  break;
3399  }
3400  } else {
3401  switch (Pred) {
3402  case FCmpInst::FCMP_OGT:
3403  // No value is ordered and greater than infinity.
3404  return getFalse(RetTy);
3405  case FCmpInst::FCMP_ULE:
3406  // All values are unordered with and at most infinity.
3407  return getTrue(RetTy);
3408  default:
3409  break;
3410  }
3411  }
3412  }
3413  if (C->isNegative() && !C->isNegZero()) {
3414  assert(!C->isNaN() && "Unexpected NaN constant!");
3415  // TODO: We can catch more cases by using a range check rather than
3416  // relying on CannotBeOrderedLessThanZero.
3417  switch (Pred) {
3418  case FCmpInst::FCMP_UGE:
3419  case FCmpInst::FCMP_UGT:
3420  case FCmpInst::FCMP_UNE:
3421  // (X >= 0) implies (X > C) when (C < 0)
3422  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3423  return getTrue(RetTy);
3424  break;
3425  case FCmpInst::FCMP_OEQ:
3426  case FCmpInst::FCMP_OLE:
3427  case FCmpInst::FCMP_OLT:
3428  // (X >= 0) implies !(X < C) when (C < 0)
3429  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3430  return getFalse(RetTy);
3431  break;
3432  default:
3433  break;
3434  }
3435  }
3436 
3437  // Check comparison of constant with minnum with smaller constant.
3438  const APFloat *C2;
3439  if ((match(LHS, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_APFloat(C2))) &&
3440  C2->compare(*C) == APFloat::cmpLessThan) ||
3441  (match(LHS, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_APFloat(C2))) &&
3442  C2->compare(*C) == APFloat::cmpGreaterThan)) {
3443  bool IsMaxNum =
3444  cast<IntrinsicInst>(LHS)->getIntrinsicID() == Intrinsic::maxnum;
3445  // The ordered relationship and minnum/maxnum guarantee that we do not
3446  // have NaN constants, so ordered/unordered preds are handled the same.
3447  switch (Pred) {
3449  // minnum(X, LesserC) == C --> false
3450  // maxnum(X, GreaterC) == C --> false
3451  return getFalse(RetTy);
3453  // minnum(X, LesserC) != C --> true
3454  // maxnum(X, GreaterC) != C --> true
3455  return getTrue(RetTy);
3458  // minnum(X, LesserC) >= C --> false
3459  // minnum(X, LesserC) > C --> false
3460  // maxnum(X, GreaterC) >= C --> true
3461  // maxnum(X, GreaterC) > C --> true
3462  return ConstantInt::get(RetTy, IsMaxNum);
3465  // minnum(X, LesserC) <= C --> true
3466  // minnum(X, LesserC) < C --> true
3467  // maxnum(X, GreaterC) <= C --> false
3468  // maxnum(X, GreaterC) < C --> false
3469  return ConstantInt::get(RetTy, !IsMaxNum);
3470  default:
3471  // TRUE/FALSE/ORD/UNO should be handled before this.
3472  llvm_unreachable("Unexpected fcmp predicate");
3473  }
3474  }
3475  }
3476 
3477  if (match(RHS, m_AnyZeroFP())) {
3478  switch (Pred) {
3479  case FCmpInst::FCMP_OGE:
3480  if (FMF.noNaNs() && CannotBeOrderedLessThanZero(LHS, Q.TLI))
3481  return getTrue(RetTy);
3482  break;
3483  case FCmpInst::FCMP_UGE:
3484  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3485  return getTrue(RetTy);
3486  break;
3487  case FCmpInst::FCMP_ULT:
3488  if (FMF.noNaNs() && CannotBeOrderedLessThanZero(LHS, Q.TLI))
3489  return getFalse(RetTy);
3490  break;
3491  case FCmpInst::FCMP_OLT:
3492  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3493  return getFalse(RetTy);
3494  break;
3495  default:
3496  break;
3497  }
3498  }
3499 
3500  // If the comparison is with the result of a select instruction, check whether
3501  // comparing with either branch of the select always yields the same value.
3502  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3503  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3504  return V;
3505 
3506  // If the comparison is with the result of a phi instruction, check whether
3507  // doing the compare with each incoming phi value yields a common result.
3508  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3509  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3510  return V;
3511 
3512  return nullptr;
3513 }
3514 
3515 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3516  FastMathFlags FMF, const SimplifyQuery &Q) {
3517  return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3518 }
3519 
3520 /// See if V simplifies when its operand Op is replaced with RepOp.
3521 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3522  const SimplifyQuery &Q,
3523  unsigned MaxRecurse) {
3524  // Trivial replacement.
3525  if (V == Op)
3526  return RepOp;
3527 
3528  // We cannot replace a constant, and shouldn't even try.
3529  if (isa<Constant>(Op))
3530  return nullptr;
3531 
3532  auto *I = dyn_cast<Instruction>(V);
3533  if (!I)
3534  return nullptr;
3535 
3536  // If this is a binary operator, try to simplify it with the replaced op.
3537  if (auto *B = dyn_cast<BinaryOperator>(I)) {
3538  // Consider:
3539  // %cmp = icmp eq i32 %x, 2147483647
3540  // %add = add nsw i32 %x, 1
3541  // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3542  //
3543  // We can't replace %sel with %add unless we strip away the flags.
3544  if (isa<OverflowingBinaryOperator>(B))
3545  if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3546  return nullptr;
3547  if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3548  return nullptr;
3549 
3550  if (MaxRecurse) {
3551  if (B->getOperand(0) == Op)
3552  return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3553  MaxRecurse - 1);
3554  if (B->getOperand(1) == Op)
3555  return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3556  MaxRecurse - 1);
3557  }
3558  }
3559 
3560  // Same for CmpInsts.
3561  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3562  if (MaxRecurse) {
3563  if (C->getOperand(0) == Op)
3564  return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3565  MaxRecurse - 1);
3566  if (C->getOperand(1) == Op)
3567  return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3568  MaxRecurse - 1);
3569  }
3570  }
3571 
3572  // Same for GEPs.
3573  if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3574  if (MaxRecurse) {
3575  SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3576  transform(GEP->operands(), NewOps.begin(),
3577  [&](Value *V) { return V == Op ? RepOp : V; });
3578  return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3579  MaxRecurse - 1);
3580  }
3581  }
3582 
3583  // TODO: We could hand off more cases to instsimplify here.
3584 
3585  // If all operands are constant after substituting Op for RepOp then we can
3586  // constant fold the instruction.
3587  if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3588  // Build a list of all constant operands.
3589  SmallVector<Constant *, 8> ConstOps;
3590  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3591  if (I->getOperand(i) == Op)
3592  ConstOps.push_back(CRepOp);
3593  else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3594  ConstOps.push_back(COp);
3595  else
3596  break;
3597  }
3598 
3599  // All operands were constants, fold it.
3600  if (ConstOps.size() == I->getNumOperands()) {
3601  if (CmpInst *C = dyn_cast<CmpInst>(I))
3602  return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3603  ConstOps[1], Q.DL, Q.TLI);
3604 
3605  if (LoadInst *LI = dyn_cast<LoadInst>(I))
3606  if (!LI->isVolatile())
3607  return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3608 
3609  return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3610  }
3611  }
3612 
3613  return nullptr;
3614 }
3615 
3616 /// Try to simplify a select instruction when its condition operand is an
3617 /// integer comparison where one operand of the compare is a constant.
3618 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3619  const APInt *Y, bool TrueWhenUnset) {
3620  const APInt *C;
3621 
3622  // (X & Y) == 0 ? X & ~Y : X --> X
3623  // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3624  if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3625  *Y == ~*C)
3626  return TrueWhenUnset ? FalseVal : TrueVal;
3627 
3628  // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3629  // (X & Y) != 0 ? X : X & ~Y --> X
3630  if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3631  *Y == ~*C)
3632  return TrueWhenUnset ? FalseVal : TrueVal;
3633 
3634  if (Y->isPowerOf2()) {
3635  // (X & Y) == 0 ? X | Y : X --> X | Y
3636  // (X & Y) != 0 ? X | Y : X --> X
3637  if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3638  *Y == *C)
3639  return TrueWhenUnset ? TrueVal : FalseVal;
3640 
3641  // (X & Y) == 0 ? X : X | Y --> X
3642  // (X & Y) != 0 ? X : X | Y --> X | Y
3643  if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3644  *Y == *C)
3645  return TrueWhenUnset ? TrueVal : FalseVal;
3646  }
3647 
3648  return nullptr;
3649 }
3650 
3651 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3652 /// eq/ne.
3654  ICmpInst::Predicate Pred,
3655  Value *TrueVal, Value *FalseVal) {
3656  Value *X;
3657  APInt Mask;
3658  if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3659  return nullptr;
3660 
3661  return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3662  Pred == ICmpInst::ICMP_EQ);
3663 }
3664 
3665 /// Try to simplify a select instruction when its condition operand is an
3666 /// integer comparison.
3667 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3668  Value *FalseVal, const SimplifyQuery &Q,
3669  unsigned MaxRecurse) {
3670  ICmpInst::Predicate Pred;
3671  Value *CmpLHS, *CmpRHS;
3672  if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3673  return nullptr;
3674 
3675  if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3676  Value *X;
3677  const APInt *Y;
3678  if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3679  if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3680  Pred == ICmpInst::ICMP_EQ))
3681  return V;
3682 
3683  // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
3684  Value *ShAmt;
3685  auto isFsh = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(),
3686  m_Value(ShAmt)),
3687  m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X),
3688  m_Value(ShAmt)));
3689  // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
3690  // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
3691  if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt &&
3692  Pred == ICmpInst::ICMP_EQ)
3693  return X;
3694  // (ShAmt != 0) ? X : fshl(X, *, ShAmt) --> X
3695  // (ShAmt != 0) ? X : fshr(*, X, ShAmt) --> X
3696  if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt &&
3697  Pred == ICmpInst::ICMP_NE)
3698  return X;
3699 
3700  // Test for a zero-shift-guard-op around rotates. These are used to
3701  // avoid UB from oversized shifts in raw IR rotate patterns, but the
3702  // intrinsics do not have that problem.
3703  // We do not allow this transform for the general funnel shift case because
3704  // that would not preserve the poison safety of the original code.
3705  auto isRotate = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X),
3706  m_Deferred(X),
3707  m_Value(ShAmt)),
3708  m_Intrinsic<Intrinsic::fshr>(m_Value(X),
3709  m_Deferred(X),
3710  m_Value(ShAmt)));
3711  // (ShAmt != 0) ? fshl(X, X, ShAmt) : X --> fshl(X, X, ShAmt)
3712  // (ShAmt != 0) ? fshr(X, X, ShAmt) : X --> fshr(X, X, ShAmt)
3713  if (match(TrueVal, isRotate) && FalseVal == X && CmpLHS == ShAmt &&
3714  Pred == ICmpInst::ICMP_NE)
3715  return TrueVal;
3716  // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
3717  // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
3718  if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
3719  Pred == ICmpInst::ICMP_EQ)
3720  return FalseVal;
3721  }
3722 
3723  // Check for other compares that behave like bit test.
3724  if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3725  TrueVal, FalseVal))
3726  return V;
3727 
3728  // If we have an equality comparison, then we know the value in one of the
3729  // arms of the select. See if substituting this value into the arm and
3730  // simplifying the result yields the same value as the other arm.
3731  if (Pred == ICmpInst::ICMP_EQ) {
3732  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3733  TrueVal ||
3734  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3735  TrueVal)
3736  return FalseVal;
3737  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3738  FalseVal ||
3739  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3740  FalseVal)
3741  return FalseVal;
3742  } else if (Pred == ICmpInst::ICMP_NE) {
3743  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3744  FalseVal ||
3745  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3746  FalseVal)
3747  return TrueVal;
3748  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3749  TrueVal ||
3750  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3751  TrueVal)
3752  return TrueVal;
3753  }
3754 
3755  return nullptr;
3756 }
3757 
3758 /// Try to simplify a select instruction when its condition operand is a
3759 /// floating-point comparison.
3761  FCmpInst::Predicate Pred;
3762  if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
3763  !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
3764  return nullptr;
3765 
3766  // TODO: The transform may not be valid with -0.0. An incomplete way of
3767  // testing for that possibility is to check if at least one operand is a
3768  // non-zero constant.
3769  const APFloat *C;
3770  if ((match(T, m_APFloat(C)) && C->isNonZero()) ||
3771  (match(F, m_APFloat(C)) && C->isNonZero())) {
3772  // (T == F) ? T : F --> F
3773  // (F == T) ? T : F --> F
3774  if (Pred == FCmpInst::FCMP_OEQ)
3775  return F;
3776 
3777  // (T != F) ? T : F --> T
3778  // (F != T) ? T : F --> T
3779  if (Pred == FCmpInst::FCMP_UNE)
3780  return T;
3781  }
3782 
3783  return nullptr;
3784 }
3785 
3786 /// Given operands for a SelectInst, see if we can fold the result.
3787 /// If not, this returns null.
3788 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3789  const SimplifyQuery &Q, unsigned MaxRecurse) {
3790  if (auto *CondC = dyn_cast<Constant>(Cond)) {
3791  if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3792  if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3793  return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3794 
3795  // select undef, X, Y -> X or Y
3796  if (isa<UndefValue>(CondC))
3797  return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3798 
3799  // TODO: Vector constants with undef elements don't simplify.
3800 
3801  // select true, X, Y -> X
3802  if (CondC->isAllOnesValue())
3803  return TrueVal;
3804  // select false, X, Y -> Y
3805  if (CondC->isNullValue())
3806  return FalseVal;
3807  }
3808 
3809  // select ?, X, X -> X
3810  if (TrueVal == FalseVal)
3811  return TrueVal;
3812 
3813  if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3814  return FalseVal;
3815  if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3816  return TrueVal;
3817 
3818  if (Value *V =
3819  simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3820  return V;
3821 
3822  if (Value *V = simplifySelectWithFCmp(Cond, TrueVal, FalseVal))
3823  return V;
3824 
3825  if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3826  return V;
3827 
3828  Optional<bool> Imp = isImpliedByDomCondition(Cond, Q.CxtI, Q.DL);
3829  if (Imp)
3830  return *Imp ? TrueVal : FalseVal;
3831 
3832  return nullptr;
3833 }
3834 
3835 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3836  const SimplifyQuery &Q) {
3837  return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3838 }
3839 
3840 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3841 /// If not, this returns null.
3843  const SimplifyQuery &Q, unsigned) {
3844  // The type of the GEP pointer operand.
3845  unsigned AS =
3846  cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3847 
3848  // getelementptr P -> P.
3849  if (Ops.size() == 1)
3850  return Ops[0];
3851 
3852  // Compute the (pointer) type returned by the GEP instruction.
3853  Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3854  Type *GEPTy = PointerType::get(LastType, AS);
3855  if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3856  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3857  else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3858  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3859 
3860  if (isa<UndefValue>(Ops[0]))
3861  return UndefValue::get(GEPTy);
3862 
3863  if (Ops.size() == 2) {
3864  // getelementptr P, 0 -> P.
3865  if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3866  return Ops[0];
3867 
3868  Type *Ty = SrcTy;
3869  if (Ty->isSized()) {
3870  Value *P;
3871  uint64_t C;
3872  uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3873  // getelementptr P, N -> P if P points to a type of zero size.
3874  if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3875  return Ops[0];
3876 
3877  // The following transforms are only safe if the ptrtoint cast
3878  // doesn't truncate the pointers.
3879  if (Ops[1]->getType()->getScalarSizeInBits() ==
3880  Q.DL.getIndexSizeInBits(AS)) {
3881  auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3882  if (match(P, m_Zero()))
3883  return Constant::getNullValue(GEPTy);
3884  Value *Temp;
3885  if (match(P, m_PtrToInt(m_Value(Temp))))
3886  if (Temp->getType() == GEPTy)
3887  return Temp;
3888  return nullptr;
3889  };
3890 
3891  // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3892  if (TyAllocSize == 1 &&
3893  match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3894  if (Value *R = PtrToIntOrZero(P))
3895  return R;
3896 
3897  // getelementptr V, (ashr (sub P, V), C) -> Q
3898  // if P points to a type of size 1 << C.
3899  if (match(Ops[1],
3900  m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3901  m_ConstantInt(C))) &&
3902  TyAllocSize == 1ULL << C)
3903  if (Value *R = PtrToIntOrZero(P))
3904  return R;
3905 
3906  // getelementptr V, (sdiv (sub P, V), C) -> Q
3907  // if P points to a type of size C.
3908  if (match(Ops[1],
3909  m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3910  m_SpecificInt(TyAllocSize))))
3911  if (Value *R = PtrToIntOrZero(P))
3912  return R;
3913  }
3914  }
3915  }
3916 
3917  if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3918  all_of(Ops.slice(1).drop_back(1),
3919  [](Value *Idx) { return match(Idx, m_Zero()); })) {
3920  unsigned IdxWidth =
3921  Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3922  if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3923  APInt BasePtrOffset(IdxWidth, 0);
3924  Value *StrippedBasePtr =
3925  Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3926  BasePtrOffset);
3927 
3928  // gep (gep V, C), (sub 0, V) -> C
3929  if (match(Ops.back(),
3930  m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3931  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3932  return ConstantExpr::getIntToPtr(CI, GEPTy);
3933  }
3934  // gep (gep V, C), (xor V, -1) -> C-1
3935  if (match(Ops.back(),
3936  m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3937  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3938  return ConstantExpr::getIntToPtr(CI, GEPTy);
3939  }
3940  }
3941  }
3942 
3943  // Check to see if this is constant foldable.
3944  if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3945  return nullptr;
3946 
3947  auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3948  Ops.slice(1));
3949  if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3950  return CEFolded;
3951  return CE;
3952 }
3953 
3955  const SimplifyQuery &Q) {
3956  return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3957 }
3958 
3959 /// Given operands for an InsertValueInst, see if we can fold the result.
3960 /// If not, this returns null.
3962  ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3963  unsigned) {
3964  if (Constant *CAgg = dyn_cast<Constant>(Agg))
3965  if (Constant *CVal = dyn_cast<Constant>(Val))
3966  return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3967 
3968  // insertvalue x, undef, n -> x
3969  if (match(Val, m_Undef()))
3970  return Agg;
3971 
3972  // insertvalue x, (extractvalue y, n), n
3973  if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3974  if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3975  EV->getIndices() == Idxs) {
3976  // insertvalue undef, (extractvalue y, n), n -> y
3977  if (match(Agg, m_Undef()))
3978  return EV->getAggregateOperand();
3979 
3980  // insertvalue y, (extractvalue y, n), n -> y
3981  if (Agg == EV->getAggregateOperand())
3982  return Agg;
3983  }
3984 
3985  return nullptr;
3986 }
3987 
3989  ArrayRef<unsigned> Idxs,
3990  const SimplifyQuery &Q) {
3992 }
3993 
3995  const SimplifyQuery &Q) {
3996  // Try to constant fold.
3997  auto *VecC = dyn_cast<Constant>(Vec);
3998  auto *ValC = dyn_cast<Constant>(Val);
3999  auto *IdxC = dyn_cast<Constant>(Idx);
4000  if (VecC && ValC && IdxC)
4001  return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
4002 
4003  // Fold into undef if index is out of bounds.
4004  if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4005  uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
4006  if (CI->uge(NumElements))
4007  return UndefValue::get(Vec->getType());
4008  }
4009 
4010  // If index is undef, it might be out of bounds (see above case)
4011  if (isa<UndefValue>(Idx))
4012  return UndefValue::get(Vec->getType());
4013 
4014  return nullptr;
4015 }
4016 
4017 /// Given operands for an ExtractValueInst, see if we can fold the result.
4018 /// If not, this returns null.
4020  const SimplifyQuery &, unsigned) {
4021  if (auto *CAgg = dyn_cast<Constant>(Agg))
4022  return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4023 
4024  // extractvalue x, (insertvalue y, elt, n), n -> elt
4025  unsigned NumIdxs = Idxs.size();
4026  for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4027  IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4028  ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4029  unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4030  unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4031  if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4032  Idxs.slice(0, NumCommonIdxs)) {
4033  if (NumIdxs == NumInsertValueIdxs)
4034  return IVI->getInsertedValueOperand();
4035  break;
4036  }
4037  }
4038 
4039  return nullptr;
4040 }
4041 
4043  const SimplifyQuery &Q) {
4045 }
4046 
4047 /// Given operands for an ExtractElementInst, see if we can fold the result.
4048 /// If not, this returns null.
4050  unsigned) {
4051  if (auto *CVec = dyn_cast<Constant>(Vec)) {
4052  if (auto *CIdx = dyn_cast<Constant>(Idx))
4053  return ConstantFoldExtractElementInstruction(CVec, CIdx);
4054 
4055  // The index is not relevant if our vector is a splat.
4056  if (auto *Splat = CVec->getSplatValue())
4057  return Splat;
4058 
4059  if (isa<UndefValue>(Vec))
4060  return UndefValue::get(Vec->getType()->getVectorElementType());
4061  }
4062 
4063  // If extracting a specified index from the vector, see if we can recursively
4064  // find a previously computed scalar that was inserted into the vector.
4065  if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4066  if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4067  // definitely out of bounds, thus undefined result
4068  return UndefValue::get(Vec->getType()->getVectorElementType());
4069  if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4070  return Elt;
4071  }
4072 
4073  // An undef extract index can be arbitrarily chosen to be an out-of-range
4074  // index value, which would result in the instruction being undef.
4075  if (isa<UndefValue>(Idx))
4076  return UndefValue::get(Vec->getType()->getVectorElementType());
4077 
4078  return nullptr;
4079 }
4080 
4082  const SimplifyQuery &Q) {
4084 }
4085 
4086 /// See if we can fold the given phi. If not, returns null.
4087 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4088  // If all of the PHI's incoming values are the same then replace the PHI node
4089  // with the common value.
4090  Value *CommonValue = nullptr;
4091  bool HasUndefInput = false;
4092  for (Value *Incoming : PN->incoming_values()) {
4093  // If the incoming value is the phi node itself, it can safely be skipped.
4094  if (Incoming == PN) continue;
4095  if (isa<UndefValue>(Incoming)) {
4096  // Remember that we saw an undef value, but otherwise ignore them.
4097  HasUndefInput = true;
4098  continue;
4099  }
4100  if (CommonValue && Incoming != CommonValue)
4101  return nullptr; // Not the same, bail out.
4102  CommonValue = Incoming;
4103  }
4104 
4105  // If CommonValue is null then all of the incoming values were either undef or
4106  // equal to the phi node itself.
4107  if (!CommonValue)
4108  return UndefValue::get(PN->getType());
4109 
4110  // If we have a PHI node like phi(X, undef, X), where X is defined by some
4111  // instruction, we cannot return X as the result of the PHI node unless it
4112  // dominates the PHI block.
4113  if (HasUndefInput)
4114  return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4115 
4116  return CommonValue;
4117 }
4118 
4119 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4120  Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4121  if (auto *C = dyn_cast<Constant>(Op))
4122  return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4123 
4124  if (auto *CI = dyn_cast<CastInst>(Op)) {
4125  auto *Src = CI->getOperand(0);
4126  Type *SrcTy = Src->getType();
4127  Type *MidTy = CI->getType();
4128  Type *DstTy = Ty;
4129  if (Src->getType() == Ty) {
4130  auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4131  auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4132  Type *SrcIntPtrTy =
4133  SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4134  Type *MidIntPtrTy =
4135  MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4136  Type *DstIntPtrTy =
4137  DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4138  if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4139  SrcIntPtrTy, MidIntPtrTy,
4140  DstIntPtrTy) == Instruction::BitCast)
4141  return Src;
4142  }
4143  }
4144 
4145  // bitcast x -> x
4146  if (CastOpc == Instruction::BitCast)
4147  if (Op->getType() == Ty)
4148  return Op;
4149 
4150  return nullptr;
4151 }
4152 
4153 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4154  const SimplifyQuery &Q) {
4155  return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4156 }
4157 
4158 /// For the given destination element of a shuffle, peek through shuffles to
4159 /// match a root vector source operand that contains that element in the same
4160 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4161 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4162  int MaskVal, Value *RootVec,
4163  unsigned MaxRecurse) {
4164  if (!MaxRecurse--)
4165  return nullptr;
4166 
4167  // Bail out if any mask value is undefined. That kind of shuffle may be
4168  // simplified further based on demanded bits or other folds.
4169  if (MaskVal == -1)
4170  return nullptr;
4171 
4172  // The mask value chooses which source operand we need to look at next.
4173  int InVecNumElts = Op0->getType()->getVectorNumElements();
4174  int RootElt = MaskVal;
4175  Value *SourceOp = Op0;
4176  if (MaskVal >= InVecNumElts) {
4177  RootElt = MaskVal - InVecNumElts;
4178  SourceOp = Op1;
4179  }
4180 
4181  // If the source operand is a shuffle itself, look through it to find the
4182  // matching root vector.
4183  if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4184  return foldIdentityShuffles(
4185  DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4186  SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4187  }
4188 
4189  // TODO: Look through bitcasts? What if the bitcast changes the vector element
4190  // size?
4191 
4192  // The source operand is not a shuffle. Initialize the root vector value for
4193  // this shuffle if that has not been done yet.
4194  if (!RootVec)
4195  RootVec = SourceOp;
4196 
4197  // Give up as soon as a source operand does not match the existing root value.
4198  if (RootVec != SourceOp)
4199  return nullptr;
4200 
4201  // The element must be coming from the same lane in the source vector
4202  // (although it may have crossed lanes in intermediate shuffles).
4203  if (RootElt != DestElt)
4204  return nullptr;
4205 
4206  return RootVec;
4207 }
4208 
4210  Type *RetTy, const SimplifyQuery &Q,
4211  unsigned MaxRecurse) {
4212  if (isa<UndefValue>(Mask))
4213  return UndefValue::get(RetTy);
4214 
4215  Type *InVecTy = Op0->getType();
4216  unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4217  unsigned InVecNumElts = InVecTy->getVectorNumElements();
4218 
4219  SmallVector<int, 32> Indices;
4220  ShuffleVectorInst::getShuffleMask(Mask, Indices);
4221  assert(MaskNumElts == Indices.size() &&
4222  "Size of Indices not same as number of mask elements?");
4223 
4224  // Canonicalization: If mask does not select elements from an input vector,
4225  // replace that input vector with undef.
4226  bool MaskSelects0 = false, MaskSelects1 = false;
4227  for (unsigned i = 0; i != MaskNumElts; ++i) {
4228  if (Indices[i] == -1)
4229  continue;
4230  if ((unsigned)Indices[i] < InVecNumElts)
4231  MaskSelects0 = true;
4232  else
4233  MaskSelects1 = true;
4234  }
4235  if (!MaskSelects0)
4236  Op0 = UndefValue::get(InVecTy);
4237  if (!MaskSelects1)
4238  Op1 = UndefValue::get(InVecTy);
4239 
4240  auto *Op0Const = dyn_cast<Constant>(Op0);
4241  auto *Op1Const = dyn_cast<Constant>(Op1);
4242 
4243  // If all operands are constant, constant fold the shuffle.
4244  if (Op0Const && Op1Const)
4245  return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4246 
4247  // Canonicalization: if only one input vector is constant, it shall be the
4248  // second one.
4249  if (Op0Const && !Op1Const) {
4250  std::swap(Op0, Op1);
4251  ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4252  }
4253 
4254  // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4255  // value type is same as the input vectors' type.
4256  if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4257  if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4258  OpShuf->getMask()->getSplatValue())
4259  return Op0;
4260 
4261  // Don't fold a shuffle with undef mask elements. This may get folded in a
4262  // better way using demanded bits or other analysis.
4263  // TODO: Should we allow this?
4264  if (find(Indices, -1) != Indices.end())
4265  return nullptr;
4266 
4267  // Check if every element of this shuffle can be mapped back to the
4268  // corresponding element of a single root vector. If so, we don't need this
4269  // shuffle. This handles simple identity shuffles as well as chains of
4270  // shuffles that may widen/narrow and/or move elements across lanes and back.
4271  Value *RootVec = nullptr;
4272  for (unsigned i = 0; i != MaskNumElts; ++i) {
4273  // Note that recursion is limited for each vector element, so if any element
4274  // exceeds the limit, this will fail to simplify.
4275  RootVec =
4276  foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4277 
4278  // We can't replace a widening/narrowing shuffle with one of its operands.
4279  if (!RootVec || RootVec->getType() != RetTy)
4280  return nullptr;
4281  }
4282  return RootVec;
4283 }
4284 
4285 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4287  Type *RetTy, const SimplifyQuery &Q) {
4288  return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4289 }
4290 
4292  Value *&Op, const SimplifyQuery &Q) {
4293  if (auto *C = dyn_cast<Constant>(Op))
4294  return ConstantFoldUnaryOpOperand(Opcode, C, Q.DL);
4295  return nullptr;
4296 }
4297 
4298 /// Given the operand for an FNeg, see if we can fold the result. If not, this
4299 /// returns null.
4301  const SimplifyQuery &Q, unsigned MaxRecurse) {
4302  if (Constant *C = foldConstant(Instruction::FNeg, Op, Q))
4303  return C;
4304 
4305  Value *X;
4306  // fneg (fneg X) ==> X
4307  if (match(Op, m_FNeg(m_Value(X))))
4308  return X;
4309 
4310  return nullptr;
4311 }
4312 
4314  const SimplifyQuery &Q) {
4316 }
4317 
4319  // If the input is a vector with undef elements, just return a default NaN.
4320  if (!In->isNaN())
4321  return ConstantFP::getNaN(In->getType());
4322 
4323  // Propagate the existing NaN constant when possible.
4324  // TODO: Should we quiet a signaling NaN?
4325  return In;
4326 }
4327 
4328 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4329  if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4330  return ConstantFP::getNaN(Op0->getType());
4331 
4332  if (match(Op0, m_NaN()))
4333  return propagateNaN(cast<Constant>(Op0));
4334  if (match(Op1, m_NaN()))
4335  return propagateNaN(cast<Constant>(Op1));
4336 
4337  return nullptr;
4338 }
4339 
4340 /// Given operands for an FAdd, see if we can fold the result. If not, this
4341 /// returns null.
4343  const SimplifyQuery &Q, unsigned MaxRecurse) {
4344  if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4345  return C;
4346 
4347  if (Constant *C = simplifyFPBinop(Op0, Op1))
4348  return C;
4349 
4350  // fadd X, -0 ==> X
4351  if (match(Op1, m_NegZeroFP()))
4352  return Op0;
4353 
4354  // fadd X, 0 ==> X, when we know X is not -0
4355  if (match(Op1, m_PosZeroFP()) &&
4356  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4357  return Op0;
4358 
4359  // With nnan: -X + X --> 0.0 (and commuted variant)
4360  // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4361  // Negative zeros are allowed because we always end up with positive zero:
4362  // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4363  // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4364  // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4365  // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4366  if (FMF.noNaNs()) {
4367  if (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4368  match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0))))
4369  return ConstantFP::getNullValue(Op0->getType());
4370 
4371  if (match(Op0, m_FNeg(m_Specific(Op1))) ||
4372  match(Op1, m_FNeg(m_Specific(Op0))))
4373  return ConstantFP::getNullValue(Op0->getType());
4374  }
4375 
4376  // (X - Y) + Y --> X
4377  // Y + (X - Y) --> X
4378  Value *X;
4379  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4380  (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4381  match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4382  return X;
4383 
4384  return nullptr;
4385 }
4386 
4387 /// Given operands for an FSub, see if we can fold the result. If not, this
4388 /// returns null.
4390  const SimplifyQuery &Q, unsigned MaxRecurse) {
4391  if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4392  return C;
4393 
4394  if (Constant *C = simplifyFPBinop(Op0, Op1))
4395  return C;
4396 
4397  // fsub X, +0 ==> X
4398  if (match(Op1, m_PosZeroFP()))
4399  return Op0;
4400 
4401  // fsub X, -0 ==> X, when we know X is not -0
4402  if (match(Op1, m_NegZeroFP()) &&
4403  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4404  return Op0;
4405 
4406  // fsub -0.0, (fsub -0.0, X) ==> X
4407  Value *X;
4408  if (match(Op0, m_NegZeroFP()) &&
4409  match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4410  return X;
4411 
4412  // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4413  // fsub 0.0, (fneg X) ==> X if signed zeros are ignored.
4414  if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4415  (match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))) ||
4416  match(Op1, m_FNeg(m_Value(X)))))
4417  return X;
4418 
4419  // fsub nnan x, x ==> 0.0
4420  if (FMF.noNaNs() && Op0 == Op1)
4421  return Constant::getNullValue(Op0->getType());
4422 
4423  // Y - (Y - X) --> X
4424  // (X + Y) - Y --> X
4425  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4426  (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4427  match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4428  return X;
4429 
4430  return nullptr;
4431 }
4432 
4433 /// Given the operands for an FMul, see if we can fold the result
4435  const SimplifyQuery &Q, unsigned MaxRecurse) {
4436  if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4437  return C;
4438 
4439  if (Constant *C = simplifyFPBinop(Op0, Op1))
4440  return C;
4441 
4442  // fmul X, 1.0 ==> X
4443  if (match(Op1, m_FPOne()))
4444  return Op0;
4445 
4446  // fmul nnan nsz X, 0 ==> 0
4447  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4448  return ConstantFP::getNullValue(Op0->getType());
4449 
4450  // sqrt(X) * sqrt(X) --> X, if we can:
4451  // 1. Remove the intermediate rounding (reassociate).
4452  // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4453  // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4454  Value *X;
4455  if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4456  FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4457  return X;
4458 
4459  return nullptr;
4460 }
4461 
4463  const SimplifyQuery &Q) {
4464  return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4465 }
4466 
4467 
4469  const SimplifyQuery &Q) {
4470  return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4471 }
4472 
4474  const SimplifyQuery &Q) {
4475  return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4476 }
4477 
4479  const SimplifyQuery &Q, unsigned) {
4480  if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4481  return C;
4482 
4483  if (Constant *C = simplifyFPBinop(Op0, Op1))
4484  return C;
4485 
4486  // X / 1.0 -> X
4487  if (match(Op1, m_FPOne()))
4488  return Op0;
4489 
4490  // 0 / X -> 0
4491  // Requires that NaNs are off (X could be zero) and signed zeroes are
4492  // ignored (X could be positive or negative, so the output sign is unknown).
4493  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4494  return ConstantFP::getNullValue(Op0->getType());
4495 
4496  if (FMF.noNaNs()) {
4497  // X / X -> 1.0 is legal when NaNs are ignored.
4498  // We can ignore infinities because INF/INF is NaN.
4499  if (Op0 == Op1)
4500  return ConstantFP::get(Op0->getType(), 1.0);
4501 
4502  // (X * Y) / Y --> X if we can reassociate to the above form.
4503  Value *X;
4504  if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4505  return X;
4506 
4507  // -X / X -> -1.0 and
4508  // X / -X -> -1.0 are legal when NaNs are ignored.
4509  // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4510  if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4511  match(Op1, m_FNegNSZ(m_Specific(Op0))))
4512  return ConstantFP::get(Op0->getType(), -1.0);
4513  }
4514 
4515  return nullptr;
4516 }
4517 
4519  const SimplifyQuery &Q) {
4520  return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4521 }
4522 
4524  const SimplifyQuery &Q, unsigned) {
4525  if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4526  return C;
4527 
4528  if (Constant *C = simplifyFPBinop(Op0, Op1))
4529  return C;
4530 
4531  // Unlike fdiv, the result of frem always matches the sign of the dividend.
4532  // The constant match may include undef elements in a vector, so return a full
4533  // zero constant as the result.
4534  if (FMF.noNaNs()) {
4535  // +0 % X -> 0
4536  if (match(Op0, m_PosZeroFP()))
4537  return ConstantFP::getNullValue(Op0->getType());
4538  // -0 % X -> -0
4539  if (match(Op0, m_NegZeroFP()))
4540  return ConstantFP::getNegativeZero(Op0->getType());
4541  }
4542 
4543  return nullptr;
4544 }
4545 
4547  const SimplifyQuery &Q) {
4548  return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4549 }
4550 
4551 //=== Helper functions for higher up the class hierarchy.
4552 
4553 /// Given the operand for a UnaryOperator, see if we can fold the result.
4554 /// If not, this returns null.
4555 static Value *simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q,
4556  unsigned MaxRecurse) {
4557  switch (Opcode) {
4558  case Instruction::FNeg:
4559  return simplifyFNegInst(Op, FastMathFlags(), Q, MaxRecurse);
4560  default:
4561  llvm_unreachable("Unexpected opcode");
4562  }
4563 }
4564 
4565 /// Given the operand for a UnaryOperator, see if we can fold the result.
4566 /// If not, this returns null.
4567 /// In contrast to SimplifyUnOp, try to use FastMathFlag when folding the
4568 /// result. In case we don't need FastMathFlags, simply fall to SimplifyUnOp.
4569 static Value *simplifyFPUnOp(unsigned Opcode, Value *Op,
4570  const FastMathFlags &FMF,
4571  const SimplifyQuery &Q, unsigned MaxRecurse) {
4572  switch (Opcode) {
4573  case Instruction::FNeg:
4574  return simplifyFNegInst(Op, FMF, Q, MaxRecurse);
4575  default:
4576  return simplifyUnOp(Opcode, Op, Q, MaxRecurse);
4577  }
4578 }
4579 
4581  const SimplifyQuery &Q) {
4582  return ::simplifyFPUnOp(Opcode, Op, FMF, Q, RecursionLimit);
4583 }
4584 
4585 /// Given operands for a BinaryOperator, see if we can fold the result.
4586 /// If not, this returns null.
4587 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4588  const SimplifyQuery &Q, unsigned MaxRecurse) {
4589  switch (Opcode) {
4590  case Instruction::Add:
4591  return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4592  case Instruction::Sub:
4593  return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4594  case Instruction::Mul:
4595  return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4596  case Instruction::SDiv:
4597  return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4598  case Instruction::UDiv:
4599  return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4600  case Instruction::SRem:
4601  return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4602  case Instruction::URem:
4603  return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4604  case Instruction::Shl:
4605  return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4606  case Instruction::LShr:
4607  return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4608  case Instruction::AShr:
4609  return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4610  case Instruction::And:
4611  return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4612  case Instruction::Or:
4613  return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4614  case Instruction::Xor:
4615  return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4616  case Instruction::FAdd:
4617  return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4618  case Instruction::FSub:
4619  return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4620  case Instruction::FMul:
4621  return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4622  case Instruction::FDiv:
4623  return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4624  case Instruction::FRem:
4625  return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4626  default:
4627  llvm_unreachable("Unexpected opcode");
4628  }
4629 }
4630 
4631 /// Given operands for a BinaryOperator, see if we can fold the result.
4632 /// If not, this returns null.
4633 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4634 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4635 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4636  const FastMathFlags &FMF, const SimplifyQuery &Q,
4637  unsigned MaxRecurse) {
4638  switch (Opcode) {
4639  case Instruction::FAdd:
4640  return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4641  case Instruction::FSub:
4642  return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4643  case Instruction::FMul:
4644  return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4645  case Instruction::FDiv:
4646  return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4647  default:
4648  return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4649  }
4650 }
4651 
4652 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4653  const SimplifyQuery &Q) {
4654  return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4655 }
4656 
4657 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4658  FastMathFlags FMF, const SimplifyQuery &Q) {
4659  return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4660 }
4661 
4662 /// Given operands for a CmpInst, see if we can fold the result.
4663 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4664  const SimplifyQuery &Q, unsigned MaxRecurse) {
4666  return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4667  return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4668 }
4669 
4670 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4671  const SimplifyQuery &Q) {
4672  return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4673 }
4674 
4676  switch (ID) {
4677  default: return false;
4678 
4679  // Unary idempotent: f(f(x)) = f(x)
4680  case Intrinsic::fabs:
4681  case Intrinsic::floor:
4682  case Intrinsic::ceil:
4683  case Intrinsic::trunc:
4684  case Intrinsic::rint:
4685  case Intrinsic::nearbyint:
4686  case Intrinsic::round:
4687  case Intrinsic::canonicalize:
4688  return true;
4689  }
4690 }
4691 
4693  const DataLayout &DL) {
4694  GlobalValue *PtrSym;
4695  APInt PtrOffset;
4696  if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4697  return nullptr;
4698 
4699  Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4701  Type *Int32PtrTy = Int32Ty->getPointerTo();
4702  Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4703 
4704  auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4705  if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4706  return nullptr;
4707 
4708  uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4709  if (OffsetInt % 4 != 0)
4710  return nullptr;
4711 
4713  Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4714  ConstantInt::get(Int64Ty, OffsetInt / 4));
4715  Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4716  if (!Loaded)
4717  return nullptr;
4718 
4719  auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4720  if (!LoadedCE)
4721  return nullptr;
4722 
4723  if (LoadedCE->getOpcode() == Instruction::Trunc) {
4724  LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4725  if (!LoadedCE)
4726  return nullptr;
4727  }
4728 
4729  if (LoadedCE->getOpcode() != Instruction::Sub)
4730  return nullptr;
4731 
4732  auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4733  if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4734  return nullptr;
4735  auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4736 
4737  Constant *LoadedRHS = LoadedCE->getOperand(1);
4738  GlobalValue *LoadedRHSSym;
4739  APInt LoadedRHSOffset;
4740  if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4741  DL) ||
4742  PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4743  return nullptr;
4744 
4745  return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4746 }
4747 
4749  const SimplifyQuery &Q) {
4750  // Idempotent functions return the same result when called repeatedly.
4751  Intrinsic::ID IID = F->getIntrinsicID();
4752  if (IsIdempotent(IID))
4753  if (auto *II = dyn_cast<IntrinsicInst>(Op0))
4754  if (II->getIntrinsicID() == IID)
4755  return II;
4756 
4757  Value *X;
4758  switch (IID) {
4759  case Intrinsic::fabs:
4760  if (SignBitMustBeZero(Op0, Q.TLI)) return Op0;
4761  break;
4762  case Intrinsic::bswap:
4763  // bswap(bswap(x)) -> x
4764  if (match(Op0, m_BSwap(m_Value(X)))) return X;
4765  break;
4766  case Intrinsic::bitreverse:
4767  // bitreverse(bitreverse(x)) -> x
4768  if (match(Op0, m_BitReverse(m_Value(X)))) return X;
4769  break;
4770  case Intrinsic::exp:
4771  // exp(log(x)) -> x
4772  if (Q.CxtI->hasAllowReassoc() &&
4773  match(Op0, m_Intrinsic<Intrinsic::log>(m_Value(X)))) return X;
4774  break;
4775  case Intrinsic::exp2:
4776  // exp2(log2(x)) -> x
4777  if (Q.CxtI->hasAllowReassoc() &&
4778  match(Op0, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) return X;
4779  break;
4780  case Intrinsic::log:
4781  // log(exp(x)) -> x
4782  if (Q.CxtI->hasAllowReassoc() &&
4783  match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) return X;
4784  break;
4785  case Intrinsic::log2:
4786  // log2(exp2(x)) -> x
4787  if (Q.CxtI->