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 *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
55  unsigned);
56 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
57  const SimplifyQuery &, unsigned);
58 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
59  unsigned);
60 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
61  const SimplifyQuery &Q, unsigned MaxRecurse);
62 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
64 static Value *SimplifyCastInst(unsigned, Value *, Type *,
65  const SimplifyQuery &, unsigned);
67  unsigned);
68 
69 static Value *foldSelectWithBinaryOp(Value *Cond, Value *TrueVal,
70  Value *FalseVal) {
71  BinaryOperator::BinaryOps BinOpCode;
72  if (auto *BO = dyn_cast<BinaryOperator>(Cond))
73  BinOpCode = BO->getOpcode();
74  else
75  return nullptr;
76 
77  CmpInst::Predicate ExpectedPred, Pred1, Pred2;
78  if (BinOpCode == BinaryOperator::Or) {
79  ExpectedPred = ICmpInst::ICMP_NE;
80  } else if (BinOpCode == BinaryOperator::And) {
81  ExpectedPred = ICmpInst::ICMP_EQ;
82  } else
83  return nullptr;
84 
85  // %A = icmp eq %TV, %FV
86  // %B = icmp eq %X, %Y (and one of these is a select operand)
87  // %C = and %A, %B
88  // %D = select %C, %TV, %FV
89  // -->
90  // %FV
91 
92  // %A = icmp ne %TV, %FV
93  // %B = icmp ne %X, %Y (and one of these is a select operand)
94  // %C = or %A, %B
95  // %D = select %C, %TV, %FV
96  // -->
97  // %TV
98  Value *X, *Y;
99  if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
100  m_Specific(FalseVal)),
101  m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
102  Pred1 != Pred2 || Pred1 != ExpectedPred)
103  return nullptr;
104 
105  if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
106  return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
107 
108  return nullptr;
109 }
110 
111 /// For a boolean type or a vector of boolean type, return false or a vector
112 /// with every element false.
113 static Constant *getFalse(Type *Ty) {
114  return ConstantInt::getFalse(Ty);
115 }
116 
117 /// For a boolean type or a vector of boolean type, return true or a vector
118 /// with every element true.
119 static Constant *getTrue(Type *Ty) {
120  return ConstantInt::getTrue(Ty);
121 }
122 
123 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
124 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
125  Value *RHS) {
126  CmpInst *Cmp = dyn_cast<CmpInst>(V);
127  if (!Cmp)
128  return false;
129  CmpInst::Predicate CPred = Cmp->getPredicate();
130  Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
131  if (CPred == Pred && CLHS == LHS && CRHS == RHS)
132  return true;
133  return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
134  CRHS == LHS;
135 }
136 
137 /// Does the given value dominate the specified phi node?
138 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
140  if (!I)
141  // Arguments and constants dominate all instructions.
142  return true;
143 
144  // If we are processing instructions (and/or basic blocks) that have not been
145  // fully added to a function, the parent nodes may still be null. Simply
146  // return the conservative answer in these cases.
147  if (!I->getParent() || !P->getParent() || !I->getFunction())
148  return false;
149 
150  // If we have a DominatorTree then do a precise test.
151  if (DT)
152  return DT->dominates(I, P);
153 
154  // Otherwise, if the instruction is in the entry block and is not an invoke,
155  // then it obviously dominates all phi nodes.
156  if (I->getParent() == &I->getFunction()->getEntryBlock() &&
157  !isa<InvokeInst>(I))
158  return true;
159 
160  return false;
161 }
162 
163 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
164 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
165 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
166 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
167 /// Returns the simplified value, or null if no simplification was performed.
169  Instruction::BinaryOps OpcodeToExpand,
170  const SimplifyQuery &Q, unsigned MaxRecurse) {
171  // Recursion is always used, so bail out at once if we already hit the limit.
172  if (!MaxRecurse--)
173  return nullptr;
174 
175  // Check whether the expression has the form "(A op' B) op C".
176  if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
177  if (Op0->getOpcode() == OpcodeToExpand) {
178  // It does! Try turning it into "(A op C) op' (B op C)".
179  Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
180  // Do "A op C" and "B op C" both simplify?
181  if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
182  if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
183  // They do! Return "L op' R" if it simplifies or is already available.
184  // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
185  if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
186  && L == B && R == A)) {
187  ++NumExpand;
188  return LHS;
189  }
190  // Otherwise return "L op' R" if it simplifies.
191  if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
192  ++NumExpand;
193  return V;
194  }
195  }
196  }
197 
198  // Check whether the expression has the form "A op (B op' C)".
199  if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
200  if (Op1->getOpcode() == OpcodeToExpand) {
201  // It does! Try turning it into "(A op B) op' (A op C)".
202  Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
203  // Do "A op B" and "A op C" both simplify?
204  if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
205  if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
206  // They do! Return "L op' R" if it simplifies or is already available.
207  // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
208  if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
209  && L == C && R == B)) {
210  ++NumExpand;
211  return RHS;
212  }
213  // Otherwise return "L op' R" if it simplifies.
214  if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
215  ++NumExpand;
216  return V;
217  }
218  }
219  }
220 
221  return nullptr;
222 }
223 
224 /// Generic simplifications for associative binary operations.
225 /// Returns the simpler value, or null if none was found.
227  Value *LHS, Value *RHS,
228  const SimplifyQuery &Q,
229  unsigned MaxRecurse) {
230  assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
231 
232  // Recursion is always used, so bail out at once if we already hit the limit.
233  if (!MaxRecurse--)
234  return nullptr;
235 
238 
239  // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
240  if (Op0 && Op0->getOpcode() == Opcode) {
241  Value *A = Op0->getOperand(0);
242  Value *B = Op0->getOperand(1);
243  Value *C = RHS;
244 
245  // Does "B op C" simplify?
246  if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
247  // It does! Return "A op V" if it simplifies or is already available.
248  // If V equals B then "A op V" is just the LHS.
249  if (V == B) return LHS;
250  // Otherwise return "A op V" if it simplifies.
251  if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
252  ++NumReassoc;
253  return W;
254  }
255  }
256  }
257 
258  // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
259  if (Op1 && Op1->getOpcode() == Opcode) {
260  Value *A = LHS;
261  Value *B = Op1->getOperand(0);
262  Value *C = Op1->getOperand(1);
263 
264  // Does "A op B" simplify?
265  if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
266  // It does! Return "V op C" if it simplifies or is already available.
267  // If V equals B then "V op C" is just the RHS.
268  if (V == B) return RHS;
269  // Otherwise return "V op C" if it simplifies.
270  if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
271  ++NumReassoc;
272  return W;
273  }
274  }
275  }
276 
277  // The remaining transforms require commutativity as well as associativity.
278  if (!Instruction::isCommutative(Opcode))
279  return nullptr;
280 
281  // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
282  if (Op0 && Op0->getOpcode() == Opcode) {
283  Value *A = Op0->getOperand(0);
284  Value *B = Op0->getOperand(1);
285  Value *C = RHS;
286 
287  // Does "C op A" simplify?
288  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
289  // It does! Return "V op B" if it simplifies or is already available.
290  // If V equals A then "V op B" is just the LHS.
291  if (V == A) return LHS;
292  // Otherwise return "V op B" if it simplifies.
293  if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
294  ++NumReassoc;
295  return W;
296  }
297  }
298  }
299 
300  // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
301  if (Op1 && Op1->getOpcode() == Opcode) {
302  Value *A = LHS;
303  Value *B = Op1->getOperand(0);
304  Value *C = Op1->getOperand(1);
305 
306  // Does "C op A" simplify?
307  if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
308  // It does! Return "B op V" if it simplifies or is already available.
309  // If V equals C then "B op V" is just the RHS.
310  if (V == C) return RHS;
311  // Otherwise return "B op V" if it simplifies.
312  if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
313  ++NumReassoc;
314  return W;
315  }
316  }
317  }
318 
319  return nullptr;
320 }
321 
322 /// In the case of a binary operation with a select instruction as an operand,
323 /// try to simplify the binop by seeing whether evaluating it on both branches
324 /// of the select results in the same value. Returns the common value if so,
325 /// otherwise returns null.
327  Value *RHS, const SimplifyQuery &Q,
328  unsigned MaxRecurse) {
329  // Recursion is always used, so bail out at once if we already hit the limit.
330  if (!MaxRecurse--)
331  return nullptr;
332 
333  SelectInst *SI;
334  if (isa<SelectInst>(LHS)) {
335  SI = cast<SelectInst>(LHS);
336  } else {
337  assert(isa<SelectInst>(RHS) && "No select instruction operand!");
338  SI = cast<SelectInst>(RHS);
339  }
340 
341  // Evaluate the BinOp on the true and false branches of the select.
342  Value *TV;
343  Value *FV;
344  if (SI == LHS) {
345  TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
346  FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
347  } else {
348  TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
349  FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
350  }
351 
352  // If they simplified to the same value, then return the common value.
353  // If they both failed to simplify then return null.
354  if (TV == FV)
355  return TV;
356 
357  // If one branch simplified to undef, return the other one.
358  if (TV && isa<UndefValue>(TV))
359  return FV;
360  if (FV && isa<UndefValue>(FV))
361  return TV;
362 
363  // If applying the operation did not change the true and false select values,
364  // then the result of the binop is the select itself.
365  if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
366  return SI;
367 
368  // If one branch simplified and the other did not, and the simplified
369  // value is equal to the unsimplified one, return the simplified value.
370  // For example, select (cond, X, X & Z) & Z -> X & Z.
371  if ((FV && !TV) || (TV && !FV)) {
372  // Check that the simplified value has the form "X op Y" where "op" is the
373  // same as the original operation.
374  Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
375  if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
376  // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
377  // We already know that "op" is the same as for the simplified value. See
378  // if the operands match too. If so, return the simplified value.
379  Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
380  Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
381  Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
382  if (Simplified->getOperand(0) == UnsimplifiedLHS &&
383  Simplified->getOperand(1) == UnsimplifiedRHS)
384  return Simplified;
385  if (Simplified->isCommutative() &&
386  Simplified->getOperand(1) == UnsimplifiedLHS &&
387  Simplified->getOperand(0) == UnsimplifiedRHS)
388  return Simplified;
389  }
390  }
391 
392  return nullptr;
393 }
394 
395 /// In the case of a comparison with a select instruction, try to simplify the
396 /// comparison by seeing whether both branches of the select result in the same
397 /// value. Returns the common value if so, otherwise returns null.
399  Value *RHS, const SimplifyQuery &Q,
400  unsigned MaxRecurse) {
401  // Recursion is always used, so bail out at once if we already hit the limit.
402  if (!MaxRecurse--)
403  return nullptr;
404 
405  // Make sure the select is on the LHS.
406  if (!isa<SelectInst>(LHS)) {
407  std::swap(LHS, RHS);
408  Pred = CmpInst::getSwappedPredicate(Pred);
409  }
410  assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
411  SelectInst *SI = cast<SelectInst>(LHS);
412  Value *Cond = SI->getCondition();
413  Value *TV = SI->getTrueValue();
414  Value *FV = SI->getFalseValue();
415 
416  // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
417  // Does "cmp TV, RHS" simplify?
418  Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
419  if (TCmp == Cond) {
420  // It not only simplified, it simplified to the select condition. Replace
421  // it with 'true'.
422  TCmp = getTrue(Cond->getType());
423  } else if (!TCmp) {
424  // It didn't simplify. However if "cmp TV, RHS" is equal to the select
425  // condition then we can replace it with 'true'. Otherwise give up.
426  if (!isSameCompare(Cond, Pred, TV, RHS))
427  return nullptr;
428  TCmp = getTrue(Cond->getType());
429  }
430 
431  // Does "cmp FV, RHS" simplify?
432  Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
433  if (FCmp == Cond) {
434  // It not only simplified, it simplified to the select condition. Replace
435  // it with 'false'.
436  FCmp = getFalse(Cond->getType());
437  } else if (!FCmp) {
438  // It didn't simplify. However if "cmp FV, RHS" is equal to the select
439  // condition then we can replace it with 'false'. Otherwise give up.
440  if (!isSameCompare(Cond, Pred, FV, RHS))
441  return nullptr;
442  FCmp = getFalse(Cond->getType());
443  }
444 
445  // If both sides simplified to the same value, then use it as the result of
446  // the original comparison.
447  if (TCmp == FCmp)
448  return TCmp;
449 
450  // The remaining cases only make sense if the select condition has the same
451  // type as the result of the comparison, so bail out if this is not so.
452  if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
453  return nullptr;
454  // If the false value simplified to false, then the result of the compare
455  // is equal to "Cond && TCmp". This also catches the case when the false
456  // value simplified to false and the true value to true, returning "Cond".
457  if (match(FCmp, m_Zero()))
458  if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
459  return V;
460  // If the true value simplified to true, then the result of the compare
461  // is equal to "Cond || FCmp".
462  if (match(TCmp, m_One()))
463  if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
464  return V;
465  // Finally, if the false value simplified to true and the true value to
466  // false, then the result of the compare is equal to "!Cond".
467  if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
468  if (Value *V =
469  SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
470  Q, MaxRecurse))
471  return V;
472 
473  return nullptr;
474 }
475 
476 /// In the case of a binary operation with an operand that is a PHI instruction,
477 /// try to simplify the binop by seeing whether evaluating it on the incoming
478 /// phi values yields the same result for every value. If so returns the common
479 /// value, otherwise returns null.
481  Value *RHS, const SimplifyQuery &Q,
482  unsigned MaxRecurse) {
483  // Recursion is always used, so bail out at once if we already hit the limit.
484  if (!MaxRecurse--)
485  return nullptr;
486 
487  PHINode *PI;
488  if (isa<PHINode>(LHS)) {
489  PI = cast<PHINode>(LHS);
490  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
491  if (!valueDominatesPHI(RHS, PI, Q.DT))
492  return nullptr;
493  } else {
494  assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
495  PI = cast<PHINode>(RHS);
496  // Bail out if LHS and the phi may be mutually interdependent due to a loop.
497  if (!valueDominatesPHI(LHS, PI, Q.DT))
498  return nullptr;
499  }
500 
501  // Evaluate the BinOp on the incoming phi values.
502  Value *CommonValue = nullptr;
503  for (Value *Incoming : PI->incoming_values()) {
504  // If the incoming value is the phi node itself, it can safely be skipped.
505  if (Incoming == PI) continue;
506  Value *V = PI == LHS ?
507  SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
508  SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
509  // If the operation failed to simplify, or simplified to a different value
510  // to previously, then give up.
511  if (!V || (CommonValue && V != CommonValue))
512  return nullptr;
513  CommonValue = V;
514  }
515 
516  return CommonValue;
517 }
518 
519 /// In the case of a comparison with a PHI instruction, try to simplify the
520 /// comparison by seeing whether comparing with all of the incoming phi values
521 /// yields the same result every time. If so returns the common result,
522 /// otherwise returns null.
524  const SimplifyQuery &Q, unsigned MaxRecurse) {
525  // Recursion is always used, so bail out at once if we already hit the limit.
526  if (!MaxRecurse--)
527  return nullptr;
528 
529  // Make sure the phi is on the LHS.
530  if (!isa<PHINode>(LHS)) {
531  std::swap(LHS, RHS);
532  Pred = CmpInst::getSwappedPredicate(Pred);
533  }
534  assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
535  PHINode *PI = cast<PHINode>(LHS);
536 
537  // Bail out if RHS and the phi may be mutually interdependent due to a loop.
538  if (!valueDominatesPHI(RHS, PI, Q.DT))
539  return nullptr;
540 
541  // Evaluate the BinOp on the incoming phi values.
542  Value *CommonValue = nullptr;
543  for (Value *Incoming : PI->incoming_values()) {
544  // If the incoming value is the phi node itself, it can safely be skipped.
545  if (Incoming == PI) continue;
546  Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
547  // If the operation failed to simplify, or simplified to a different value
548  // to previously, then give up.
549  if (!V || (CommonValue && V != CommonValue))
550  return nullptr;
551  CommonValue = V;
552  }
553 
554  return CommonValue;
555 }
556 
558  Value *&Op0, Value *&Op1,
559  const SimplifyQuery &Q) {
560  if (auto *CLHS = dyn_cast<Constant>(Op0)) {
561  if (auto *CRHS = dyn_cast<Constant>(Op1))
562  return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
563 
564  // Canonicalize the constant to the RHS if this is a commutative operation.
565  if (Instruction::isCommutative(Opcode))
566  std::swap(Op0, Op1);
567  }
568  return nullptr;
569 }
570 
571 /// Given operands for an Add, see if we can fold the result.
572 /// If not, this returns null.
573 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
574  const SimplifyQuery &Q, unsigned MaxRecurse) {
575  if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
576  return C;
577 
578  // X + undef -> undef
579  if (match(Op1, m_Undef()))
580  return Op1;
581 
582  // X + 0 -> X
583  if (match(Op1, m_Zero()))
584  return Op0;
585 
586  // If two operands are negative, return 0.
587  if (isKnownNegation(Op0, Op1))
588  return Constant::getNullValue(Op0->getType());
589 
590  // X + (Y - X) -> Y
591  // (Y - X) + X -> Y
592  // Eg: X + -X -> 0
593  Value *Y = nullptr;
594  if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
595  match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
596  return Y;
597 
598  // X + ~X -> -1 since ~X = -X-1
599  Type *Ty = Op0->getType();
600  if (match(Op0, m_Not(m_Specific(Op1))) ||
601  match(Op1, m_Not(m_Specific(Op0))))
602  return Constant::getAllOnesValue(Ty);
603 
604  // add nsw/nuw (xor Y, signmask), signmask --> Y
605  // The no-wrapping add guarantees that the top bit will be set by the add.
606  // Therefore, the xor must be clearing the already set sign bit of Y.
607  if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
608  match(Op0, m_Xor(m_Value(Y), m_SignMask())))
609  return Y;
610 
611  // add nuw %x, -1 -> -1, because %x can only be 0.
612  if (IsNUW && match(Op1, m_AllOnes()))
613  return Op1; // Which is -1.
614 
615  /// i1 add -> xor.
616  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
617  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
618  return V;
619 
620  // Try some generic simplifications for associative operations.
621  if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
622  MaxRecurse))
623  return V;
624 
625  // Threading Add over selects and phi nodes is pointless, so don't bother.
626  // Threading over the select in "A + select(cond, B, C)" means evaluating
627  // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
628  // only if B and C are equal. If B and C are equal then (since we assume
629  // that operands have already been simplified) "select(cond, B, C)" should
630  // have been simplified to the common value of B and C already. Analysing
631  // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
632  // for threading over phi nodes.
633 
634  return nullptr;
635 }
636 
637 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
638  const SimplifyQuery &Query) {
639  return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
640 }
641 
642 /// Compute the base pointer and cumulative constant offsets for V.
643 ///
644 /// This strips all constant offsets off of V, leaving it the base pointer, and
645 /// accumulates the total constant offset applied in the returned constant. It
646 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
647 /// no constant offsets applied.
648 ///
649 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
650 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
651 /// folding.
653  bool AllowNonInbounds = false) {
655 
656  Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
657  APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
658 
659  // Even though we don't look through PHI nodes, we could be called on an
660  // instruction in an unreachable block, which may be on a cycle.
661  SmallPtrSet<Value *, 4> Visited;
662  Visited.insert(V);
663  do {
664  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
665  if ((!AllowNonInbounds && !GEP->isInBounds()) ||
666  !GEP->accumulateConstantOffset(DL, Offset))
667  break;
668  V = GEP->getPointerOperand();
669  } else if (Operator::getOpcode(V) == Instruction::BitCast) {
670  V = cast<Operator>(V)->getOperand(0);
671  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
672  if (GA->isInterposable())
673  break;
674  V = GA->getAliasee();
675  } else {
676  if (auto *Call = dyn_cast<CallBase>(V))
677  if (Value *RV = Call->getReturnedArgOperand()) {
678  V = RV;
679  continue;
680  }
681  break;
682  }
683  assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
684  } while (Visited.insert(V).second);
685 
686  Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
687  if (V->getType()->isVectorTy())
689  OffsetIntPtr);
690  return OffsetIntPtr;
691 }
692 
693 /// Compute the constant difference between two pointer values.
694 /// If the difference is not a constant, returns zero.
696  Value *RHS) {
697  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
698  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
699 
700  // If LHS and RHS are not related via constant offsets to the same base
701  // value, there is nothing we can do here.
702  if (LHS != RHS)
703  return nullptr;
704 
705  // Otherwise, the difference of LHS - RHS can be computed as:
706  // LHS - RHS
707  // = (LHSOffset + Base) - (RHSOffset + Base)
708  // = LHSOffset - RHSOffset
709  return ConstantExpr::getSub(LHSOffset, RHSOffset);
710 }
711 
712 /// Given operands for a Sub, see if we can fold the result.
713 /// If not, this returns null.
714 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
715  const SimplifyQuery &Q, unsigned MaxRecurse) {
716  if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
717  return C;
718 
719  // X - undef -> undef
720  // undef - X -> undef
721  if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
722  return UndefValue::get(Op0->getType());
723 
724  // X - 0 -> X
725  if (match(Op1, m_Zero()))
726  return Op0;
727 
728  // X - X -> 0
729  if (Op0 == Op1)
730  return Constant::getNullValue(Op0->getType());
731 
732  // Is this a negation?
733  if (match(Op0, m_Zero())) {
734  // 0 - X -> 0 if the sub is NUW.
735  if (isNUW)
736  return Constant::getNullValue(Op0->getType());
737 
738  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
739  if (Known.Zero.isMaxSignedValue()) {
740  // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
741  // Op1 must be 0 because negating the minimum signed value is undefined.
742  if (isNSW)
743  return Constant::getNullValue(Op0->getType());
744 
745  // 0 - X -> X if X is 0 or the minimum signed value.
746  return Op1;
747  }
748  }
749 
750  // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
751  // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
752  Value *X = nullptr, *Y = nullptr, *Z = Op1;
753  if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
754  // See if "V === Y - Z" simplifies.
755  if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
756  // It does! Now see if "X + V" simplifies.
757  if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
758  // It does, we successfully reassociated!
759  ++NumReassoc;
760  return W;
761  }
762  // See if "V === X - Z" simplifies.
763  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
764  // It does! Now see if "Y + V" simplifies.
765  if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
766  // It does, we successfully reassociated!
767  ++NumReassoc;
768  return W;
769  }
770  }
771 
772  // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
773  // For example, X - (X + 1) -> -1
774  X = Op0;
775  if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
776  // See if "V === X - Y" simplifies.
777  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
778  // It does! Now see if "V - Z" simplifies.
779  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
780  // It does, we successfully reassociated!
781  ++NumReassoc;
782  return W;
783  }
784  // See if "V === X - Z" simplifies.
785  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
786  // It does! Now see if "V - Y" simplifies.
787  if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
788  // It does, we successfully reassociated!
789  ++NumReassoc;
790  return W;
791  }
792  }
793 
794  // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
795  // For example, X - (X - Y) -> Y.
796  Z = Op0;
797  if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
798  // See if "V === Z - X" simplifies.
799  if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
800  // It does! Now see if "V + Y" simplifies.
801  if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
802  // It does, we successfully reassociated!
803  ++NumReassoc;
804  return W;
805  }
806 
807  // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
808  if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
809  match(Op1, m_Trunc(m_Value(Y))))
810  if (X->getType() == Y->getType())
811  // See if "V === X - Y" simplifies.
812  if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
813  // It does! Now see if "trunc V" simplifies.
814  if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
815  Q, MaxRecurse - 1))
816  // It does, return the simplified "trunc V".
817  return W;
818 
819  // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
820  if (match(Op0, m_PtrToInt(m_Value(X))) &&
821  match(Op1, m_PtrToInt(m_Value(Y))))
822  if (Constant *Result = computePointerDifference(Q.DL, X, Y))
823  return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
824 
825  // i1 sub -> xor.
826  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
827  if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
828  return V;
829 
830  // Threading Sub over selects and phi nodes is pointless, so don't bother.
831  // Threading over the select in "A - select(cond, B, C)" means evaluating
832  // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
833  // only if B and C are equal. If B and C are equal then (since we assume
834  // that operands have already been simplified) "select(cond, B, C)" should
835  // have been simplified to the common value of B and C already. Analysing
836  // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
837  // for threading over phi nodes.
838 
839  return nullptr;
840 }
841 
842 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
843  const SimplifyQuery &Q) {
844  return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
845 }
846 
847 /// Given operands for a Mul, see if we can fold the result.
848 /// If not, this returns null.
849 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
850  unsigned MaxRecurse) {
851  if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
852  return C;
853 
854  // X * undef -> 0
855  // X * 0 -> 0
856  if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
857  return Constant::getNullValue(Op0->getType());
858 
859  // X * 1 -> X
860  if (match(Op1, m_One()))
861  return Op0;
862 
863  // (X / Y) * Y -> X if the division is exact.
864  Value *X = nullptr;
865  if (Q.IIQ.UseInstrInfo &&
866  (match(Op0,
867  m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
868  match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
869  return X;
870 
871  // i1 mul -> and.
872  if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
873  if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
874  return V;
875 
876  // Try some generic simplifications for associative operations.
877  if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
878  MaxRecurse))
879  return V;
880 
881  // Mul distributes over Add. Try some generic simplifications based on this.
882  if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
883  Q, MaxRecurse))
884  return V;
885 
886  // If the operation is with the result of a select instruction, check whether
887  // operating on either branch of the select always yields the same value.
888  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
889  if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
890  MaxRecurse))
891  return V;
892 
893  // If the operation is with the result of a phi instruction, check whether
894  // operating on all incoming values of the phi always yields the same value.
895  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
896  if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
897  MaxRecurse))
898  return V;
899 
900  return nullptr;
901 }
902 
905 }
906 
907 /// Check for common or similar folds of integer division or integer remainder.
908 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
909 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
910  Type *Ty = Op0->getType();
911 
912  // X / undef -> undef
913  // X % undef -> undef
914  if (match(Op1, m_Undef()))
915  return Op1;
916 
917  // X / 0 -> undef
918  // X % 0 -> undef
919  // We don't need to preserve faults!
920  if (match(Op1, m_Zero()))
921  return UndefValue::get(Ty);
922 
923  // If any element of a constant divisor vector is zero or undef, the whole op
924  // is undef.
925  auto *Op1C = dyn_cast<Constant>(Op1);
926  if (Op1C && Ty->isVectorTy()) {
927  unsigned NumElts = Ty->getVectorNumElements();
928  for (unsigned i = 0; i != NumElts; ++i) {
929  Constant *Elt = Op1C->getAggregateElement(i);
930  if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
931  return UndefValue::get(Ty);
932  }
933  }
934 
935  // undef / X -> 0
936  // undef % X -> 0
937  if (match(Op0, m_Undef()))
938  return Constant::getNullValue(Ty);
939 
940  // 0 / X -> 0
941  // 0 % X -> 0
942  if (match(Op0, m_Zero()))
943  return Constant::getNullValue(Op0->getType());
944 
945  // X / X -> 1
946  // X % X -> 0
947  if (Op0 == Op1)
948  return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
949 
950  // X / 1 -> X
951  // X % 1 -> 0
952  // If this is a boolean op (single-bit element type), we can't have
953  // division-by-zero or remainder-by-zero, so assume the divisor is 1.
954  // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
955  Value *X;
956  if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
957  (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
958  return IsDiv ? Op0 : Constant::getNullValue(Ty);
959 
960  return nullptr;
961 }
962 
963 /// Given a predicate and two operands, return true if the comparison is true.
964 /// This is a helper for div/rem simplification where we return some other value
965 /// when we can prove a relationship between the operands.
966 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
967  const SimplifyQuery &Q, unsigned MaxRecurse) {
968  Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
969  Constant *C = dyn_cast_or_null<Constant>(V);
970  return (C && C->isAllOnesValue());
971 }
972 
973 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
974 /// to simplify X % Y to X.
975 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
976  unsigned MaxRecurse, bool IsSigned) {
977  // Recursion is always used, so bail out at once if we already hit the limit.
978  if (!MaxRecurse--)
979  return false;
980 
981  if (IsSigned) {
982  // |X| / |Y| --> 0
983  //
984  // We require that 1 operand is a simple constant. That could be extended to
985  // 2 variables if we computed the sign bit for each.
986  //
987  // Make sure that a constant is not the minimum signed value because taking
988  // the abs() of that is undefined.
989  Type *Ty = X->getType();
990  const APInt *C;
991  if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
992  // Is the variable divisor magnitude always greater than the constant
993  // dividend magnitude?
994  // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
995  Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
996  Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
997  if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
998  isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
999  return true;
1000  }
1001  if (match(Y, m_APInt(C))) {
1002  // Special-case: we can't take the abs() of a minimum signed value. If
1003  // that's the divisor, then all we have to do is prove that the dividend
1004  // is also not the minimum signed value.
1005  if (C->isMinSignedValue())
1006  return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1007 
1008  // Is the variable dividend magnitude always less than the constant
1009  // divisor magnitude?
1010  // |X| < |C| --> X > -abs(C) and X < abs(C)
1011  Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1012  Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1013  if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1014  isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1015  return true;
1016  }
1017  return false;
1018  }
1019 
1020  // IsSigned == false.
1021  // Is the dividend unsigned less than the divisor?
1022  return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1023 }
1024 
1025 /// These are simplifications common to SDiv and UDiv.
1027  const SimplifyQuery &Q, unsigned MaxRecurse) {
1028  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1029  return C;
1030 
1031  if (Value *V = simplifyDivRem(Op0, Op1, true))
1032  return V;
1033 
1034  bool IsSigned = Opcode == Instruction::SDiv;
1035 
1036  // (X * Y) / Y -> X if the multiplication does not overflow.
1037  Value *X;
1038  if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1039  auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1040  // If the Mul does not overflow, then we are good to go.
1041  if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1042  (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1043  return X;
1044  // If X has the form X = A / Y, then X * Y cannot overflow.
1045  if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1046  (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1047  return X;
1048  }
1049 
1050  // (X rem Y) / Y -> 0
1051  if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1052  (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1053  return Constant::getNullValue(Op0->getType());
1054 
1055  // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1056  ConstantInt *C1, *C2;
1057  if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1058  match(Op1, m_ConstantInt(C2))) {
1059  bool Overflow;
1060  (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1061  if (Overflow)
1062  return Constant::getNullValue(Op0->getType());
1063  }
1064 
1065  // If the operation is with the result of a select instruction, check whether
1066  // operating on either branch of the select always yields the same value.
1067  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1068  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1069  return V;
1070 
1071  // If the operation is with the result of a phi instruction, check whether
1072  // operating on all incoming values of the phi always yields the same value.
1073  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1074  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1075  return V;
1076 
1077  if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1078  return Constant::getNullValue(Op0->getType());
1079 
1080  return nullptr;
1081 }
1082 
1083 /// These are simplifications common to SRem and URem.
1085  const SimplifyQuery &Q, unsigned MaxRecurse) {
1086  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1087  return C;
1088 
1089  if (Value *V = simplifyDivRem(Op0, Op1, false))
1090  return V;
1091 
1092  // (X % Y) % Y -> X % Y
1093  if ((Opcode == Instruction::SRem &&
1094  match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1095  (Opcode == Instruction::URem &&
1096  match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1097  return Op0;
1098 
1099  // (X << Y) % X -> 0
1100  if (Q.IIQ.UseInstrInfo &&
1101  ((Opcode == Instruction::SRem &&
1102  match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1103  (Opcode == Instruction::URem &&
1104  match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1105  return Constant::getNullValue(Op0->getType());
1106 
1107  // If the operation is with the result of a select instruction, check whether
1108  // operating on either branch of the select always yields the same value.
1109  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1110  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1111  return V;
1112 
1113  // If the operation is with the result of a phi instruction, check whether
1114  // operating on all incoming values of the phi always yields the same value.
1115  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1116  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1117  return V;
1118 
1119  // If X / Y == 0, then X % Y == X.
1120  if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1121  return Op0;
1122 
1123  return nullptr;
1124 }
1125 
1126 /// Given operands for an SDiv, see if we can fold the result.
1127 /// If not, this returns null.
1128 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1129  unsigned MaxRecurse) {
1130  // If two operands are negated and no signed overflow, return -1.
1131  if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1132  return Constant::getAllOnesValue(Op0->getType());
1133 
1134  return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1135 }
1136 
1139 }
1140 
1141 /// Given operands for a UDiv, see if we can fold the result.
1142 /// If not, this returns null.
1143 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1144  unsigned MaxRecurse) {
1145  return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1146 }
1147 
1150 }
1151 
1152 /// Given operands for an SRem, see if we can fold the result.
1153 /// If not, this returns null.
1154 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1155  unsigned MaxRecurse) {
1156  // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1157  // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1158  Value *X;
1159  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1160  return ConstantInt::getNullValue(Op0->getType());
1161 
1162  // If the two operands are negated, return 0.
1163  if (isKnownNegation(Op0, Op1))
1164  return ConstantInt::getNullValue(Op0->getType());
1165 
1166  return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1167 }
1168 
1171 }
1172 
1173 /// Given operands for a URem, see if we can fold the result.
1174 /// If not, this returns null.
1175 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1176  unsigned MaxRecurse) {
1177  return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1178 }
1179 
1182 }
1183 
1184 /// Returns true if a shift by \c Amount always yields undef.
1185 static bool isUndefShift(Value *Amount) {
1186  Constant *C = dyn_cast<Constant>(Amount);
1187  if (!C)
1188  return false;
1189 
1190  // X shift by undef -> undef because it may shift by the bitwidth.
1191  if (isa<UndefValue>(C))
1192  return true;
1193 
1194  // Shifting by the bitwidth or more is undefined.
1195  if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1196  if (CI->getValue().getLimitedValue() >=
1197  CI->getType()->getScalarSizeInBits())
1198  return true;
1199 
1200  // If all lanes of a vector shift are undefined the whole shift is.
1201  if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1202  for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1204  return false;
1205  return true;
1206  }
1207 
1208  return false;
1209 }
1210 
1211 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1212 /// If not, this returns null.
1214  Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1215  if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1216  return C;
1217 
1218  // 0 shift by X -> 0
1219  if (match(Op0, m_Zero()))
1220  return Constant::getNullValue(Op0->getType());
1221 
1222  // X shift by 0 -> X
1223  // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1224  // would be poison.
1225  Value *X;
1226  if (match(Op1, m_Zero()) ||
1227  (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1228  return Op0;
1229 
1230  // Fold undefined shifts.
1231  if (isUndefShift(Op1))
1232  return UndefValue::get(Op0->getType());
1233 
1234  // If the operation is with the result of a select instruction, check whether
1235  // operating on either branch of the select always yields the same value.
1236  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1237  if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1238  return V;
1239 
1240  // If the operation is with the result of a phi instruction, check whether
1241  // operating on all incoming values of the phi always yields the same value.
1242  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1243  if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1244  return V;
1245 
1246  // If any bits in the shift amount make that value greater than or equal to
1247  // the number of bits in the type, the shift is undefined.
1248  KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1249  if (Known.One.getLimitedValue() >= Known.getBitWidth())
1250  return UndefValue::get(Op0->getType());
1251 
1252  // If all valid bits in the shift amount are known zero, the first operand is
1253  // unchanged.
1254  unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1255  if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1256  return Op0;
1257 
1258  return nullptr;
1259 }
1260 
1261 /// Given operands for an Shl, LShr or AShr, see if we can
1262 /// fold the result. If not, this returns null.
1264  Value *Op1, bool isExact, const SimplifyQuery &Q,
1265  unsigned MaxRecurse) {
1266  if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1267  return V;
1268 
1269  // X >> X -> 0
1270  if (Op0 == Op1)
1271  return Constant::getNullValue(Op0->getType());
1272 
1273  // undef >> X -> 0
1274  // undef >> X -> undef (if it's exact)
1275  if (match(Op0, m_Undef()))
1276  return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1277 
1278  // The low bit cannot be shifted out of an exact shift if it is set.
1279  if (isExact) {
1280  KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1281  if (Op0Known.One[0])
1282  return Op0;
1283  }
1284 
1285  return nullptr;
1286 }
1287 
1288 /// Given operands for an Shl, see if we can fold the result.
1289 /// If not, this returns null.
1290 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1291  const SimplifyQuery &Q, unsigned MaxRecurse) {
1292  if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1293  return V;
1294 
1295  // undef << X -> 0
1296  // undef << X -> undef if (if it's NSW/NUW)
1297  if (match(Op0, m_Undef()))
1298  return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1299 
1300  // (X >> A) << A -> X
1301  Value *X;
1302  if (Q.IIQ.UseInstrInfo &&
1303  match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1304  return X;
1305 
1306  // shl nuw i8 C, %x -> C iff C has sign bit set.
1307  if (isNUW && match(Op0, m_Negative()))
1308  return Op0;
1309  // NOTE: could use computeKnownBits() / LazyValueInfo,
1310  // but the cost-benefit analysis suggests it isn't worth it.
1311 
1312  return nullptr;
1313 }
1314 
1315 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1316  const SimplifyQuery &Q) {
1317  return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1318 }
1319 
1320 /// Given operands for an LShr, see if we can fold the result.
1321 /// If not, this returns null.
1322 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1323  const SimplifyQuery &Q, unsigned MaxRecurse) {
1324  if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1325  MaxRecurse))
1326  return V;
1327 
1328  // (X << A) >> A -> X
1329  Value *X;
1330  if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1331  return X;
1332 
1333  // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1334  // We can return X as we do in the above case since OR alters no bits in X.
1335  // SimplifyDemandedBits in InstCombine can do more general optimization for
1336  // bit manipulation. This pattern aims to provide opportunities for other
1337  // optimizers by supporting a simple but common case in InstSimplify.
1338  Value *Y;
1339  const APInt *ShRAmt, *ShLAmt;
1340  if (match(Op1, m_APInt(ShRAmt)) &&
1341  match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1342  *ShRAmt == *ShLAmt) {
1343  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1344  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1345  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1346  if (ShRAmt->uge(EffWidthY))
1347  return X;
1348  }
1349 
1350  return nullptr;
1351 }
1352 
1353 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1354  const SimplifyQuery &Q) {
1355  return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1356 }
1357 
1358 /// Given operands for an AShr, see if we can fold the result.
1359 /// If not, this returns null.
1360 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1361  const SimplifyQuery &Q, unsigned MaxRecurse) {
1362  if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1363  MaxRecurse))
1364  return V;
1365 
1366  // all ones >>a X -> -1
1367  // Do not return Op0 because it may contain undef elements if it's a vector.
1368  if (match(Op0, m_AllOnes()))
1369  return Constant::getAllOnesValue(Op0->getType());
1370 
1371  // (X << A) >> A -> X
1372  Value *X;
1373  if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1374  return X;
1375 
1376  // Arithmetic shifting an all-sign-bit value is a no-op.
1377  unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1378  if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1379  return Op0;
1380 
1381  return nullptr;
1382 }
1383 
1384 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1385  const SimplifyQuery &Q) {
1386  return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1387 }
1388 
1389 /// Commuted variants are assumed to be handled by calling this function again
1390 /// with the parameters swapped.
1392  ICmpInst *UnsignedICmp, bool IsAnd) {
1393  Value *X, *Y;
1394 
1395  ICmpInst::Predicate EqPred;
1396  if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1397  !ICmpInst::isEquality(EqPred))
1398  return nullptr;
1399 
1400  ICmpInst::Predicate UnsignedPred;
1401  if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1402  ICmpInst::isUnsigned(UnsignedPred))
1403  ;
1404  else if (match(UnsignedICmp,
1405  m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1406  ICmpInst::isUnsigned(UnsignedPred))
1407  UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1408  else
1409  return nullptr;
1410 
1411  // X < Y && Y != 0 --> X < Y
1412  // X < Y || Y != 0 --> Y != 0
1413  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1414  return IsAnd ? UnsignedICmp : ZeroICmp;
1415 
1416  // X >= Y || Y != 0 --> true
1417  // X >= Y || Y == 0 --> X >= Y
1418  if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1419  if (EqPred == ICmpInst::ICMP_NE)
1420  return getTrue(UnsignedICmp->getType());
1421  return UnsignedICmp;
1422  }
1423 
1424  // X < Y && Y == 0 --> false
1425  if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1426  IsAnd)
1427  return getFalse(UnsignedICmp->getType());
1428 
1429  return nullptr;
1430 }
1431 
1432 /// Commuted variants are assumed to be handled by calling this function again
1433 /// with the parameters swapped.
1435  ICmpInst::Predicate Pred0, Pred1;
1436  Value *A ,*B;
1437  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1438  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1439  return nullptr;
1440 
1441  // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1442  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1443  // can eliminate Op1 from this 'and'.
1444  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1445  return Op0;
1446 
1447  // Check for any combination of predicates that are guaranteed to be disjoint.
1448  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1449  (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1450  (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1451  (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1452  return getFalse(Op0->getType());
1453 
1454  return nullptr;
1455 }
1456 
1457 /// Commuted variants are assumed to be handled by calling this function again
1458 /// with the parameters swapped.
1460  ICmpInst::Predicate Pred0, Pred1;
1461  Value *A ,*B;
1462  if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1463  !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1464  return nullptr;
1465 
1466  // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1467  // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1468  // can eliminate Op0 from this 'or'.
1469  if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1470  return Op1;
1471 
1472  // Check for any combination of predicates that cover the entire range of
1473  // possibilities.
1474  if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1475  (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1476  (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1477  (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1478  return getTrue(Op0->getType());
1479 
1480  return nullptr;
1481 }
1482 
1483 /// Test if a pair of compares with a shared operand and 2 constants has an
1484 /// empty set intersection, full set union, or if one compare is a superset of
1485 /// the other.
1487  bool IsAnd) {
1488  // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1489  if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1490  return nullptr;
1491 
1492  const APInt *C0, *C1;
1493  if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1494  !match(Cmp1->getOperand(1), m_APInt(C1)))
1495  return nullptr;
1496 
1497  auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1498  auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1499 
1500  // For and-of-compares, check if the intersection is empty:
1501  // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1502  if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1503  return getFalse(Cmp0->getType());
1504 
1505  // For or-of-compares, check if the union is full:
1506  // (icmp X, C0) || (icmp X, C1) --> full set --> true
1507  if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1508  return getTrue(Cmp0->getType());
1509 
1510  // Is one range a superset of the other?
1511  // If this is and-of-compares, take the smaller set:
1512  // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1513  // If this is or-of-compares, take the larger set:
1514  // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1515  if (Range0.contains(Range1))
1516  return IsAnd ? Cmp1 : Cmp0;
1517  if (Range1.contains(Range0))
1518  return IsAnd ? Cmp0 : Cmp1;
1519 
1520  return nullptr;
1521 }
1522 
1524  bool IsAnd) {
1525  ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1526  if (!match(Cmp0->getOperand(1), m_Zero()) ||
1527  !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1528  return nullptr;
1529 
1530  if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1531  return nullptr;
1532 
1533  // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1534  Value *X = Cmp0->getOperand(0);
1535  Value *Y = Cmp1->getOperand(0);
1536 
1537  // If one of the compares is a masked version of a (not) null check, then
1538  // that compare implies the other, so we eliminate the other. Optionally, look
1539  // through a pointer-to-int cast to match a null check of a pointer type.
1540 
1541  // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1542  // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1543  // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1544  // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1545  if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1547  return Cmp1;
1548 
1549  // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1550  // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1551  // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1552  // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1553  if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1555  return Cmp0;
1556 
1557  return nullptr;
1558 }
1559 
1561  const InstrInfoQuery &IIQ) {
1562  // (icmp (add V, C0), C1) & (icmp V, C0)
1563  ICmpInst::Predicate Pred0, Pred1;
1564  const APInt *C0, *C1;
1565  Value *V;
1566  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1567  return nullptr;
1568 
1569  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1570  return nullptr;
1571 
1572  auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1573  if (AddInst->getOperand(1) != Op1->getOperand(1))
1574  return nullptr;
1575 
1576  Type *ITy = Op0->getType();
1577  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1578  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1579 
1580  const APInt Delta = *C1 - *C0;
1581  if (C0->isStrictlyPositive()) {
1582  if (Delta == 2) {
1583  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1584  return getFalse(ITy);
1585  if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1586  return getFalse(ITy);
1587  }
1588  if (Delta == 1) {
1589  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1590  return getFalse(ITy);
1591  if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1592  return getFalse(ITy);
1593  }
1594  }
1595  if (C0->getBoolValue() && isNUW) {
1596  if (Delta == 2)
1597  if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1598  return getFalse(ITy);
1599  if (Delta == 1)
1600  if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1601  return getFalse(ITy);
1602  }
1603 
1604  return nullptr;
1605 }
1606 
1608  const InstrInfoQuery &IIQ) {
1609  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1610  return X;
1611  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1612  return X;
1613 
1614  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1615  return X;
1616  if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1617  return X;
1618 
1619  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1620  return X;
1621 
1622  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1623  return X;
1624 
1625  if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ))
1626  return X;
1627  if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, IIQ))
1628  return X;
1629 
1630  return nullptr;
1631 }
1632 
1634  const InstrInfoQuery &IIQ) {
1635  // (icmp (add V, C0), C1) | (icmp V, C0)
1636  ICmpInst::Predicate Pred0, Pred1;
1637  const APInt *C0, *C1;
1638  Value *V;
1639  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1640  return nullptr;
1641 
1642  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1643  return nullptr;
1644 
1645  auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1646  if (AddInst->getOperand(1) != Op1->getOperand(1))
1647  return nullptr;
1648 
1649  Type *ITy = Op0->getType();
1650  bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1651  bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1652 
1653  const APInt Delta = *C1 - *C0;
1654  if (C0->isStrictlyPositive()) {
1655  if (Delta == 2) {
1656  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1657  return getTrue(ITy);
1658  if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1659  return getTrue(ITy);
1660  }
1661  if (Delta == 1) {
1662  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1663  return getTrue(ITy);
1664  if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1665  return getTrue(ITy);
1666  }
1667  }
1668  if (C0->getBoolValue() && isNUW) {
1669  if (Delta == 2)
1670  if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1671  return getTrue(ITy);
1672  if (Delta == 1)
1673  if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1674  return getTrue(ITy);
1675  }
1676 
1677  return nullptr;
1678 }
1679 
1681  const InstrInfoQuery &IIQ) {
1682  if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1683  return X;
1684  if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1685  return X;
1686 
1687  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1688  return X;
1689  if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1690  return X;
1691 
1692  if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1693  return X;
1694 
1695  if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1696  return X;
1697 
1698  if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ))
1699  return X;
1700  if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, IIQ))
1701  return X;
1702 
1703  return nullptr;
1704 }
1705 
1707  FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1708  Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1709  Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1710  if (LHS0->getType() != RHS0->getType())
1711  return nullptr;
1712 
1713  FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1714  if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1715  (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1716  // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1717  // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1718  // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1719  // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1720  // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1721  // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1722  // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1723  // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1724  if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1725  (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1726  return RHS;
1727 
1728  // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1729  // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1730  // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1731  // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1732  // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1733  // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1734  // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1735  // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1736  if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1737  (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1738  return LHS;
1739  }
1740 
1741  return nullptr;
1742 }
1743 
1745  Value *Op0, Value *Op1, bool IsAnd) {
1746  // Look through casts of the 'and' operands to find compares.
1747  auto *Cast0 = dyn_cast<CastInst>(Op0);
1748  auto *Cast1 = dyn_cast<CastInst>(Op1);
1749  if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1750  Cast0->getSrcTy() == Cast1->getSrcTy()) {
1751  Op0 = Cast0->getOperand(0);
1752  Op1 = Cast1->getOperand(0);
1753  }
1754 
1755  Value *V = nullptr;
1756  auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1757  auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1758  if (ICmp0 && ICmp1)
1759  V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q.IIQ)
1760  : simplifyOrOfICmps(ICmp0, ICmp1, Q.IIQ);
1761 
1762  auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1763  auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1764  if (FCmp0 && FCmp1)
1765  V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1766 
1767  if (!V)
1768  return nullptr;
1769  if (!Cast0)
1770  return V;
1771 
1772  // If we looked through casts, we can only handle a constant simplification
1773  // because we are not allowed to create a cast instruction here.
1774  if (auto *C = dyn_cast<Constant>(V))
1775  return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1776 
1777  return nullptr;
1778 }
1779 
1780 /// Given operands for an And, see if we can fold the result.
1781 /// If not, this returns null.
1782 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1783  unsigned MaxRecurse) {
1784  if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1785  return C;
1786 
1787  // X & undef -> 0
1788  if (match(Op1, m_Undef()))
1789  return Constant::getNullValue(Op0->getType());
1790 
1791  // X & X = X
1792  if (Op0 == Op1)
1793  return Op0;
1794 
1795  // X & 0 = 0
1796  if (match(Op1, m_Zero()))
1797  return Constant::getNullValue(Op0->getType());
1798 
1799  // X & -1 = X
1800  if (match(Op1, m_AllOnes()))
1801  return Op0;
1802 
1803  // A & ~A = ~A & A = 0
1804  if (match(Op0, m_Not(m_Specific(Op1))) ||
1805  match(Op1, m_Not(m_Specific(Op0))))
1806  return Constant::getNullValue(Op0->getType());
1807 
1808  // (A | ?) & A = A
1809  if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1810  return Op1;
1811 
1812  // A & (A | ?) = A
1813  if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1814  return Op0;
1815 
1816  // A mask that only clears known zeros of a shifted value is a no-op.
1817  Value *X;
1818  const APInt *Mask;
1819  const APInt *ShAmt;
1820  if (match(Op1, m_APInt(Mask))) {
1821  // If all bits in the inverted and shifted mask are clear:
1822  // and (shl X, ShAmt), Mask --> shl X, ShAmt
1823  if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1824  (~(*Mask)).lshr(*ShAmt).isNullValue())
1825  return Op0;
1826 
1827  // If all bits in the inverted and shifted mask are clear:
1828  // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1829  if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1830  (~(*Mask)).shl(*ShAmt).isNullValue())
1831  return Op0;
1832  }
1833 
1834  // A & (-A) = A if A is a power of two or zero.
1835  if (match(Op0, m_Neg(m_Specific(Op1))) ||
1836  match(Op1, m_Neg(m_Specific(Op0)))) {
1837  if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1838  Q.DT))
1839  return Op0;
1840  if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1841  Q.DT))
1842  return Op1;
1843  }
1844 
1845  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
1846  return V;
1847 
1848  // Try some generic simplifications for associative operations.
1849  if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1850  MaxRecurse))
1851  return V;
1852 
1853  // And distributes over Or. Try some generic simplifications based on this.
1854  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1855  Q, MaxRecurse))
1856  return V;
1857 
1858  // And distributes over Xor. Try some generic simplifications based on this.
1859  if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1860  Q, MaxRecurse))
1861  return V;
1862 
1863  // If the operation is with the result of a select instruction, check whether
1864  // operating on either branch of the select always yields the same value.
1865  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1866  if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1867  MaxRecurse))
1868  return V;
1869 
1870  // If the operation is with the result of a phi instruction, check whether
1871  // operating on all incoming values of the phi always yields the same value.
1872  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1873  if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1874  MaxRecurse))
1875  return V;
1876 
1877  // Assuming the effective width of Y is not larger than A, i.e. all bits
1878  // from X and Y are disjoint in (X << A) | Y,
1879  // if the mask of this AND op covers all bits of X or Y, while it covers
1880  // no bits from the other, we can bypass this AND op. E.g.,
1881  // ((X << A) | Y) & Mask -> Y,
1882  // if Mask = ((1 << effective_width_of(Y)) - 1)
1883  // ((X << A) | Y) & Mask -> X << A,
1884  // if Mask = ((1 << effective_width_of(X)) - 1) << A
1885  // SimplifyDemandedBits in InstCombine can optimize the general case.
1886  // This pattern aims to help other passes for a common case.
1887  Value *Y, *XShifted;
1888  if (match(Op1, m_APInt(Mask)) &&
1889  match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
1890  m_Value(XShifted)),
1891  m_Value(Y)))) {
1892  const unsigned Width = Op0->getType()->getScalarSizeInBits();
1893  const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
1894  const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1895  const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1896  if (EffWidthY <= ShftCnt) {
1897  const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
1898  Q.DT);
1899  const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
1900  const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
1901  const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
1902  // If the mask is extracting all bits from X or Y as is, we can skip
1903  // this AND op.
1904  if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
1905  return Y;
1906  if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
1907  return XShifted;
1908  }
1909  }
1910 
1911  return nullptr;
1912 }
1913 
1916 }
1917 
1918 /// Given operands for an Or, see if we can fold the result.
1919 /// If not, this returns null.
1920 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1921  unsigned MaxRecurse) {
1922  if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1923  return C;
1924 
1925  // X | undef -> -1
1926  // X | -1 = -1
1927  // Do not return Op1 because it may contain undef elements if it's a vector.
1928  if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1929  return Constant::getAllOnesValue(Op0->getType());
1930 
1931  // X | X = X
1932  // X | 0 = X
1933  if (Op0 == Op1 || match(Op1, m_Zero()))
1934  return Op0;
1935 
1936  // A | ~A = ~A | A = -1
1937  if (match(Op0, m_Not(m_Specific(Op1))) ||
1938  match(Op1, m_Not(m_Specific(Op0))))
1939  return Constant::getAllOnesValue(Op0->getType());
1940 
1941  // (A & ?) | A = A
1942  if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1943  return Op1;
1944 
1945  // A | (A & ?) = A
1946  if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1947  return Op0;
1948 
1949  // ~(A & ?) | A = -1
1950  if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1951  return Constant::getAllOnesValue(Op1->getType());
1952 
1953  // A | ~(A & ?) = -1
1954  if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1955  return Constant::getAllOnesValue(Op0->getType());
1956 
1957  Value *A, *B;
1958  // (A & ~B) | (A ^ B) -> (A ^ B)
1959  // (~B & A) | (A ^ B) -> (A ^ B)
1960  // (A & ~B) | (B ^ A) -> (B ^ A)
1961  // (~B & A) | (B ^ A) -> (B ^ A)
1962  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1963  (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1964  match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1965  return Op1;
1966 
1967  // Commute the 'or' operands.
1968  // (A ^ B) | (A & ~B) -> (A ^ B)
1969  // (A ^ B) | (~B & A) -> (A ^ B)
1970  // (B ^ A) | (A & ~B) -> (B ^ A)
1971  // (B ^ A) | (~B & A) -> (B ^ A)
1972  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1973  (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1974  match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1975  return Op0;
1976 
1977  // (A & B) | (~A ^ B) -> (~A ^ B)
1978  // (B & A) | (~A ^ B) -> (~A ^ B)
1979  // (A & B) | (B ^ ~A) -> (B ^ ~A)
1980  // (B & A) | (B ^ ~A) -> (B ^ ~A)
1981  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1982  (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1983  match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1984  return Op1;
1985 
1986  // (~A ^ B) | (A & B) -> (~A ^ B)
1987  // (~A ^ B) | (B & A) -> (~A ^ B)
1988  // (B ^ ~A) | (A & B) -> (B ^ ~A)
1989  // (B ^ ~A) | (B & A) -> (B ^ ~A)
1990  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1991  (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1992  match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1993  return Op0;
1994 
1995  if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
1996  return V;
1997 
1998  // Try some generic simplifications for associative operations.
1999  if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
2000  MaxRecurse))
2001  return V;
2002 
2003  // Or distributes over And. Try some generic simplifications based on this.
2004  if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2005  MaxRecurse))
2006  return V;
2007 
2008  // If the operation is with the result of a select instruction, check whether
2009  // operating on either branch of the select always yields the same value.
2010  if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2011  if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2012  MaxRecurse))
2013  return V;
2014 
2015  // (A & C1)|(B & C2)
2016  const APInt *C1, *C2;
2017  if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2018  match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2019  if (*C1 == ~*C2) {
2020  // (A & C1)|(B & C2)
2021  // If we have: ((V + N) & C1) | (V & C2)
2022  // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2023  // replace with V+N.
2024  Value *N;
2025  if (C2->isMask() && // C2 == 0+1+
2026  match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2027  // Add commutes, try both ways.
2028  if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2029  return A;
2030  }
2031  // Or commutes, try both ways.
2032  if (C1->isMask() &&
2033  match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2034  // Add commutes, try both ways.
2035  if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2036  return B;
2037  }
2038  }
2039  }
2040 
2041  // If the operation is with the result of a phi instruction, check whether
2042  // operating on all incoming values of the phi always yields the same value.
2043  if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2044  if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2045  return V;
2046 
2047  return nullptr;
2048 }
2049 
2052 }
2053 
2054 /// Given operands for a Xor, see if we can fold the result.
2055 /// If not, this returns null.
2056 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2057  unsigned MaxRecurse) {
2058  if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2059  return C;
2060 
2061  // A ^ undef -> undef
2062  if (match(Op1, m_Undef()))
2063  return Op1;
2064 
2065  // A ^ 0 = A
2066  if (match(Op1, m_Zero()))
2067  return Op0;
2068 
2069  // A ^ A = 0
2070  if (Op0 == Op1)
2071  return Constant::getNullValue(Op0->getType());
2072 
2073  // A ^ ~A = ~A ^ A = -1
2074  if (match(Op0, m_Not(m_Specific(Op1))) ||
2075  match(Op1, m_Not(m_Specific(Op0))))
2076  return Constant::getAllOnesValue(Op0->getType());
2077 
2078  // Try some generic simplifications for associative operations.
2079  if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2080  MaxRecurse))
2081  return V;
2082 
2083  // Threading Xor over selects and phi nodes is pointless, so don't bother.
2084  // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2085  // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2086  // only if B and C are equal. If B and C are equal then (since we assume
2087  // that operands have already been simplified) "select(cond, B, C)" should
2088  // have been simplified to the common value of B and C already. Analysing
2089  // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2090  // for threading over phi nodes.
2091 
2092  return nullptr;
2093 }
2094 
2097 }
2098 
2099 
2101  return CmpInst::makeCmpResultType(Op->getType());
2102 }
2103 
2104 /// Rummage around inside V looking for something equivalent to the comparison
2105 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2106 /// Helper function for analyzing max/min idioms.
2108  Value *LHS, Value *RHS) {
2110  if (!SI)
2111  return nullptr;
2112  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2113  if (!Cmp)
2114  return nullptr;
2115  Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2116  if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2117  return Cmp;
2118  if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2119  LHS == CmpRHS && RHS == CmpLHS)
2120  return Cmp;
2121  return nullptr;
2122 }
2123 
2124 // A significant optimization not implemented here is assuming that alloca
2125 // addresses are not equal to incoming argument values. They don't *alias*,
2126 // as we say, but that doesn't mean they aren't equal, so we take a
2127 // conservative approach.
2128 //
2129 // This is inspired in part by C++11 5.10p1:
2130 // "Two pointers of the same type compare equal if and only if they are both
2131 // null, both point to the same function, or both represent the same
2132 // address."
2133 //
2134 // This is pretty permissive.
2135 //
2136 // It's also partly due to C11 6.5.9p6:
2137 // "Two pointers compare equal if and only if both are null pointers, both are
2138 // pointers to the same object (including a pointer to an object and a
2139 // subobject at its beginning) or function, both are pointers to one past the
2140 // last element of the same array object, or one is a pointer to one past the
2141 // end of one array object and the other is a pointer to the start of a
2142 // different array object that happens to immediately follow the first array
2143 // object in the address space.)
2144 //
2145 // C11's version is more restrictive, however there's no reason why an argument
2146 // couldn't be a one-past-the-end value for a stack object in the caller and be
2147 // equal to the beginning of a stack object in the callee.
2148 //
2149 // If the C and C++ standards are ever made sufficiently restrictive in this
2150 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2151 // this optimization.
2152 static Constant *
2154  const DominatorTree *DT, CmpInst::Predicate Pred,
2155  AssumptionCache *AC, const Instruction *CxtI,
2156  const InstrInfoQuery &IIQ, Value *LHS, Value *RHS) {
2157  // First, skip past any trivial no-ops.
2158  LHS = LHS->stripPointerCasts();
2159  RHS = RHS->stripPointerCasts();
2160 
2161  // A non-null pointer is not equal to a null pointer.
2162  if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2163  IIQ.UseInstrInfo) &&
2164  isa<ConstantPointerNull>(RHS) &&
2165  (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2166  return ConstantInt::get(GetCompareTy(LHS),
2167  !CmpInst::isTrueWhenEqual(Pred));
2168 
2169  // We can only fold certain predicates on pointer comparisons.
2170  switch (Pred) {
2171  default:
2172  return nullptr;
2173 
2174  // Equality comaprisons are easy to fold.
2175  case CmpInst::ICMP_EQ:
2176  case CmpInst::ICMP_NE:
2177  break;
2178 
2179  // We can only handle unsigned relational comparisons because 'inbounds' on
2180  // a GEP only protects against unsigned wrapping.
2181  case CmpInst::ICMP_UGT:
2182  case CmpInst::ICMP_UGE:
2183  case CmpInst::ICMP_ULT:
2184  case CmpInst::ICMP_ULE:
2185  // However, we have to switch them to their signed variants to handle
2186  // negative indices from the base pointer.
2187  Pred = ICmpInst::getSignedPredicate(Pred);
2188  break;
2189  }
2190 
2191  // Strip off any constant offsets so that we can reason about them.
2192  // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2193  // here and compare base addresses like AliasAnalysis does, however there are
2194  // numerous hazards. AliasAnalysis and its utilities rely on special rules
2195  // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2196  // doesn't need to guarantee pointer inequality when it says NoAlias.
2197  Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2198  Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2199 
2200  // If LHS and RHS are related via constant offsets to the same base
2201  // value, we can replace it with an icmp which just compares the offsets.
2202  if (LHS == RHS)
2203  return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2204 
2205  // Various optimizations for (in)equality comparisons.
2206  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2207  // Different non-empty allocations that exist at the same time have
2208  // different addresses (if the program can tell). Global variables always
2209  // exist, so they always exist during the lifetime of each other and all
2210  // allocas. Two different allocas usually have different addresses...
2211  //
2212  // However, if there's an @llvm.stackrestore dynamically in between two
2213  // allocas, they may have the same address. It's tempting to reduce the
2214  // scope of the problem by only looking at *static* allocas here. That would
2215  // cover the majority of allocas while significantly reducing the likelihood
2216  // of having an @llvm.stackrestore pop up in the middle. However, it's not
2217  // actually impossible for an @llvm.stackrestore to pop up in the middle of
2218  // an entry block. Also, if we have a block that's not attached to a
2219  // function, we can't tell if it's "static" under the current definition.
2220  // Theoretically, this problem could be fixed by creating a new kind of
2221  // instruction kind specifically for static allocas. Such a new instruction
2222  // could be required to be at the top of the entry block, thus preventing it
2223  // from being subject to a @llvm.stackrestore. Instcombine could even
2224  // convert regular allocas into these special allocas. It'd be nifty.
2225  // However, until then, this problem remains open.
2226  //
2227  // So, we'll assume that two non-empty allocas have different addresses
2228  // for now.
2229  //
2230  // With all that, if the offsets are within the bounds of their allocations
2231  // (and not one-past-the-end! so we can't use inbounds!), and their
2232  // allocations aren't the same, the pointers are not equal.
2233  //
2234  // Note that it's not necessary to check for LHS being a global variable
2235  // address, due to canonicalization and constant folding.
2236  if (isa<AllocaInst>(LHS) &&
2237  (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2238  ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2239  ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2240  uint64_t LHSSize, RHSSize;
2241  ObjectSizeOpts Opts;
2242  Opts.NullIsUnknownSize =
2243  NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2244  if (LHSOffsetCI && RHSOffsetCI &&
2245  getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2246  getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2247  const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2248  const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2249  if (!LHSOffsetValue.isNegative() &&
2250  !RHSOffsetValue.isNegative() &&
2251  LHSOffsetValue.ult(LHSSize) &&
2252  RHSOffsetValue.ult(RHSSize)) {
2253  return ConstantInt::get(GetCompareTy(LHS),
2254  !CmpInst::isTrueWhenEqual(Pred));
2255  }
2256  }
2257 
2258  // Repeat the above check but this time without depending on DataLayout
2259  // or being able to compute a precise size.
2260  if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2261  !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2262  LHSOffset->isNullValue() &&
2263  RHSOffset->isNullValue())
2264  return ConstantInt::get(GetCompareTy(LHS),
2265  !CmpInst::isTrueWhenEqual(Pred));
2266  }
2267 
2268  // Even if an non-inbounds GEP occurs along the path we can still optimize
2269  // equality comparisons concerning the result. We avoid walking the whole
2270  // chain again by starting where the last calls to
2271  // stripAndComputeConstantOffsets left off and accumulate the offsets.
2272  Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2273  Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2274  if (LHS == RHS)
2275  return ConstantExpr::getICmp(Pred,
2276  ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2277  ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2278 
2279  // If one side of the equality comparison must come from a noalias call
2280  // (meaning a system memory allocation function), and the other side must
2281  // come from a pointer that cannot overlap with dynamically-allocated
2282  // memory within the lifetime of the current function (allocas, byval
2283  // arguments, globals), then determine the comparison result here.
2284  SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2285  GetUnderlyingObjects(LHS, LHSUObjs, DL);
2286  GetUnderlyingObjects(RHS, RHSUObjs, DL);
2287 
2288  // Is the set of underlying objects all noalias calls?
2289  auto IsNAC = [](ArrayRef<Value *> Objects) {
2290  return all_of(Objects, isNoAliasCall);
2291  };
2292 
2293  // Is the set of underlying objects all things which must be disjoint from
2294  // noalias calls. For allocas, we consider only static ones (dynamic
2295  // allocas might be transformed into calls to malloc not simultaneously
2296  // live with the compared-to allocation). For globals, we exclude symbols
2297  // that might be resolve lazily to symbols in another dynamically-loaded
2298  // library (and, thus, could be malloc'ed by the implementation).
2299  auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2300  return all_of(Objects, [](Value *V) {
2301  if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2302  return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2303  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2304  return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2305  GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2306  !GV->isThreadLocal();
2307  if (const Argument *A = dyn_cast<Argument>(V))
2308  return A->hasByValAttr();
2309  return false;
2310  });
2311  };
2312 
2313  if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2314  (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2315  return ConstantInt::get(GetCompareTy(LHS),
2316  !CmpInst::isTrueWhenEqual(Pred));
2317 
2318  // Fold comparisons for non-escaping pointer even if the allocation call
2319  // cannot be elided. We cannot fold malloc comparison to null. Also, the
2320  // dynamic allocation call could be either of the operands.
2321  Value *MI = nullptr;
2322  if (isAllocLikeFn(LHS, TLI) &&
2323  llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2324  MI = LHS;
2325  else if (isAllocLikeFn(RHS, TLI) &&
2326  llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2327  MI = RHS;
2328  // FIXME: We should also fold the compare when the pointer escapes, but the
2329  // compare dominates the pointer escape
2330  if (MI && !PointerMayBeCaptured(MI, true, true))
2331  return ConstantInt::get(GetCompareTy(LHS),
2333  }
2334 
2335  // Otherwise, fail.
2336  return nullptr;
2337 }
2338 
2339 /// Fold an icmp when its operands have i1 scalar type.
2341  Value *RHS, const SimplifyQuery &Q) {
2342  Type *ITy = GetCompareTy(LHS); // The return type.
2343  Type *OpTy = LHS->getType(); // The operand type.
2344  if (!OpTy->isIntOrIntVectorTy(1))
2345  return nullptr;
2346 
2347  // A boolean compared to true/false can be simplified in 14 out of the 20
2348  // (10 predicates * 2 constants) possible combinations. Cases not handled here
2349  // require a 'not' of the LHS, so those must be transformed in InstCombine.
2350  if (match(RHS, m_Zero())) {
2351  switch (Pred) {
2352  case CmpInst::ICMP_NE: // X != 0 -> X
2353  case CmpInst::ICMP_UGT: // X >u 0 -> X
2354  case CmpInst::ICMP_SLT: // X <s 0 -> X
2355  return LHS;
2356 
2357  case CmpInst::ICMP_ULT: // X <u 0 -> false
2358  case CmpInst::ICMP_SGT: // X >s 0 -> false
2359  return getFalse(ITy);
2360 
2361  case CmpInst::ICMP_UGE: // X >=u 0 -> true
2362  case CmpInst::ICMP_SLE: // X <=s 0 -> true
2363  return getTrue(ITy);
2364 
2365  default: break;
2366  }
2367  } else if (match(RHS, m_One())) {
2368  switch (Pred) {
2369  case CmpInst::ICMP_EQ: // X == 1 -> X
2370  case CmpInst::ICMP_UGE: // X >=u 1 -> X
2371  case CmpInst::ICMP_SLE: // X <=s -1 -> X
2372  return LHS;
2373 
2374  case CmpInst::ICMP_UGT: // X >u 1 -> false
2375  case CmpInst::ICMP_SLT: // X <s -1 -> false
2376  return getFalse(ITy);
2377 
2378  case CmpInst::ICMP_ULE: // X <=u 1 -> true
2379  case CmpInst::ICMP_SGE: // X >=s -1 -> true
2380  return getTrue(ITy);
2381 
2382  default: break;
2383  }
2384  }
2385 
2386  switch (Pred) {
2387  default:
2388  break;
2389  case ICmpInst::ICMP_UGE:
2390  if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2391  return getTrue(ITy);
2392  break;
2393  case ICmpInst::ICMP_SGE:
2394  /// For signed comparison, the values for an i1 are 0 and -1
2395  /// respectively. This maps into a truth table of:
2396  /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2397  /// 0 | 0 | 1 (0 >= 0) | 1
2398  /// 0 | 1 | 1 (0 >= -1) | 1
2399  /// 1 | 0 | 0 (-1 >= 0) | 0
2400  /// 1 | 1 | 1 (-1 >= -1) | 1
2401  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2402  return getTrue(ITy);
2403  break;
2404  case ICmpInst::ICMP_ULE:
2405  if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2406  return getTrue(ITy);
2407  break;
2408  }
2409 
2410  return nullptr;
2411 }
2412 
2413 /// Try hard to fold icmp with zero RHS because this is a common case.
2415  Value *RHS, const SimplifyQuery &Q) {
2416  if (!match(RHS, m_Zero()))
2417  return nullptr;
2418 
2419  Type *ITy = GetCompareTy(LHS); // The return type.
2420  switch (Pred) {
2421  default:
2422  llvm_unreachable("Unknown ICmp predicate!");
2423  case ICmpInst::ICMP_ULT:
2424  return getFalse(ITy);
2425  case ICmpInst::ICMP_UGE:
2426  return getTrue(ITy);
2427  case ICmpInst::ICMP_EQ:
2428  case ICmpInst::ICMP_ULE:
2429  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2430  return getFalse(ITy);
2431  break;
2432  case ICmpInst::ICMP_NE:
2433  case ICmpInst::ICMP_UGT:
2434  if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2435  return getTrue(ITy);
2436  break;
2437  case ICmpInst::ICMP_SLT: {
2438  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2439  if (LHSKnown.isNegative())
2440  return getTrue(ITy);
2441  if (LHSKnown.isNonNegative())
2442  return getFalse(ITy);
2443  break;
2444  }
2445  case ICmpInst::ICMP_SLE: {
2446  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2447  if (LHSKnown.isNegative())
2448  return getTrue(ITy);
2449  if (LHSKnown.isNonNegative() &&
2450  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2451  return getFalse(ITy);
2452  break;
2453  }
2454  case ICmpInst::ICMP_SGE: {
2455  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2456  if (LHSKnown.isNegative())
2457  return getFalse(ITy);
2458  if (LHSKnown.isNonNegative())
2459  return getTrue(ITy);
2460  break;
2461  }
2462  case ICmpInst::ICMP_SGT: {
2463  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2464  if (LHSKnown.isNegative())
2465  return getFalse(ITy);
2466  if (LHSKnown.isNonNegative() &&
2467  isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2468  return getTrue(ITy);
2469  break;
2470  }
2471  }
2472 
2473  return nullptr;
2474 }
2475 
2477  Value *RHS, const InstrInfoQuery &IIQ) {
2478  Type *ITy = GetCompareTy(RHS); // The return type.
2479 
2480  Value *X;
2481  // Sign-bit checks can be optimized to true/false after unsigned
2482  // floating-point casts:
2483  // icmp slt (bitcast (uitofp X)), 0 --> false
2484  // icmp sgt (bitcast (uitofp X)), -1 --> true
2485  if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2486  if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2487  return ConstantInt::getFalse(ITy);
2488  if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2489  return ConstantInt::getTrue(ITy);
2490  }
2491 
2492  const APInt *C;
2493  if (!match(RHS, m_APInt(C)))
2494  return nullptr;
2495 
2496  // Rule out tautological comparisons (eg., ult 0 or uge 0).
2498  if (RHS_CR.isEmptySet())
2499  return ConstantInt::getFalse(ITy);
2500  if (RHS_CR.isFullSet())
2501  return ConstantInt::getTrue(ITy);
2502 
2503  ConstantRange LHS_CR = computeConstantRange(LHS, IIQ.UseInstrInfo);
2504  if (!LHS_CR.isFullSet()) {
2505  if (RHS_CR.contains(LHS_CR))
2506  return ConstantInt::getTrue(ITy);
2507  if (RHS_CR.inverse().contains(LHS_CR))
2508  return ConstantInt::getFalse(ITy);
2509  }
2510 
2511  return nullptr;
2512 }
2513 
2514 /// TODO: A large part of this logic is duplicated in InstCombine's
2515 /// foldICmpBinOp(). We should be able to share that and avoid the code
2516 /// duplication.
2518  Value *RHS, const SimplifyQuery &Q,
2519  unsigned MaxRecurse) {
2520  Type *ITy = GetCompareTy(LHS); // The return type.
2521 
2522  BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2523  BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2524  if (MaxRecurse && (LBO || RBO)) {
2525  // Analyze the case when either LHS or RHS is an add instruction.
2526  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2527  // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2528  bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2529  if (LBO && LBO->getOpcode() == Instruction::Add) {
2530  A = LBO->getOperand(0);
2531  B = LBO->getOperand(1);
2532  NoLHSWrapProblem =
2533  ICmpInst::isEquality(Pred) ||
2534  (CmpInst::isUnsigned(Pred) &&
2535  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2536  (CmpInst::isSigned(Pred) &&
2537  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2538  }
2539  if (RBO && RBO->getOpcode() == Instruction::Add) {
2540  C = RBO->getOperand(0);
2541  D = RBO->getOperand(1);
2542  NoRHSWrapProblem =
2543  ICmpInst::isEquality(Pred) ||
2544  (CmpInst::isUnsigned(Pred) &&
2545  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2546  (CmpInst::isSigned(Pred) &&
2547  Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2548  }
2549 
2550  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2551  if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2552  if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2553  Constant::getNullValue(RHS->getType()), Q,
2554  MaxRecurse - 1))
2555  return V;
2556 
2557  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2558  if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2559  if (Value *V =
2561  C == LHS ? D : C, Q, MaxRecurse - 1))
2562  return V;
2563 
2564  // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2565  if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2566  NoRHSWrapProblem) {
2567  // Determine Y and Z in the form icmp (X+Y), (X+Z).
2568  Value *Y, *Z;
2569  if (A == C) {
2570  // C + B == C + D -> B == D
2571  Y = B;
2572  Z = D;
2573  } else if (A == D) {
2574  // D + B == C + D -> B == C
2575  Y = B;
2576  Z = C;
2577  } else if (B == C) {
2578  // A + C == C + D -> A == D
2579  Y = A;
2580  Z = D;
2581  } else {
2582  assert(B == D);
2583  // A + D == C + D -> A == C
2584  Y = A;
2585  Z = C;
2586  }
2587  if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2588  return V;
2589  }
2590  }
2591 
2592  {
2593  Value *Y = nullptr;
2594  // icmp pred (or X, Y), X
2595  if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2596  if (Pred == ICmpInst::ICMP_ULT)
2597  return getFalse(ITy);
2598  if (Pred == ICmpInst::ICMP_UGE)
2599  return getTrue(ITy);
2600 
2601  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2602  KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2603  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2604  if (RHSKnown.isNonNegative() && YKnown.isNegative())
2605  return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2606  if (RHSKnown.isNegative() || YKnown.isNonNegative())
2607  return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2608  }
2609  }
2610  // icmp pred X, (or X, Y)
2611  if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2612  if (Pred == ICmpInst::ICMP_ULE)
2613  return getTrue(ITy);
2614  if (Pred == ICmpInst::ICMP_UGT)
2615  return getFalse(ITy);
2616 
2617  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2618  KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2619  KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2620  if (LHSKnown.isNonNegative() && YKnown.isNegative())
2621  return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2622  if (LHSKnown.isNegative() || YKnown.isNonNegative())
2623  return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2624  }
2625  }
2626  }
2627 
2628  // icmp pred (and X, Y), X
2629  if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2630  if (Pred == ICmpInst::ICMP_UGT)
2631  return getFalse(ITy);
2632  if (Pred == ICmpInst::ICMP_ULE)
2633  return getTrue(ITy);
2634  }
2635  // icmp pred X, (and X, Y)
2636  if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2637  if (Pred == ICmpInst::ICMP_UGE)
2638  return getTrue(ITy);
2639  if (Pred == ICmpInst::ICMP_ULT)
2640  return getFalse(ITy);
2641  }
2642 
2643  // 0 - (zext X) pred C
2644  if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2645  if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2646  if (RHSC->getValue().isStrictlyPositive()) {
2647  if (Pred == ICmpInst::ICMP_SLT)
2648  return ConstantInt::getTrue(RHSC->getContext());
2649  if (Pred == ICmpInst::ICMP_SGE)
2650  return ConstantInt::getFalse(RHSC->getContext());
2651  if (Pred == ICmpInst::ICMP_EQ)
2652  return ConstantInt::getFalse(RHSC->getContext());
2653  if (Pred == ICmpInst::ICMP_NE)
2654  return ConstantInt::getTrue(RHSC->getContext());
2655  }
2656  if (RHSC->getValue().isNonNegative()) {
2657  if (Pred == ICmpInst::ICMP_SLE)
2658  return ConstantInt::getTrue(RHSC->getContext());
2659  if (Pred == ICmpInst::ICMP_SGT)
2660  return ConstantInt::getFalse(RHSC->getContext());
2661  }
2662  }
2663  }
2664 
2665  // icmp pred (urem X, Y), Y
2666  if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2667  switch (Pred) {
2668  default:
2669  break;
2670  case ICmpInst::ICMP_SGT:
2671  case ICmpInst::ICMP_SGE: {
2672  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2673  if (!Known.isNonNegative())
2674  break;
2676  }
2677  case ICmpInst::ICMP_EQ:
2678  case ICmpInst::ICMP_UGT:
2679  case ICmpInst::ICMP_UGE:
2680  return getFalse(ITy);
2681  case ICmpInst::ICMP_SLT:
2682  case ICmpInst::ICMP_SLE: {
2683  KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2684  if (!Known.isNonNegative())
2685  break;
2687  }
2688  case ICmpInst::ICMP_NE:
2689  case ICmpInst::ICMP_ULT:
2690  case ICmpInst::ICMP_ULE:
2691  return getTrue(ITy);
2692  }
2693  }
2694 
2695  // icmp pred X, (urem Y, X)
2696  if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2697  switch (Pred) {
2698  default:
2699  break;
2700  case ICmpInst::ICMP_SGT:
2701  case ICmpInst::ICMP_SGE: {
2702  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2703  if (!Known.isNonNegative())
2704  break;
2706  }
2707  case ICmpInst::ICMP_NE:
2708  case ICmpInst::ICMP_UGT:
2709  case ICmpInst::ICMP_UGE:
2710  return getTrue(ITy);
2711  case ICmpInst::ICMP_SLT:
2712  case ICmpInst::ICMP_SLE: {
2713  KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2714  if (!Known.isNonNegative())
2715  break;
2717  }
2718  case ICmpInst::ICMP_EQ:
2719  case ICmpInst::ICMP_ULT:
2720  case ICmpInst::ICMP_ULE:
2721  return getFalse(ITy);
2722  }
2723  }
2724 
2725  // x >> y <=u x
2726  // x udiv y <=u x.
2727  if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2728  match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2729  // icmp pred (X op Y), X
2730  if (Pred == ICmpInst::ICMP_UGT)
2731  return getFalse(ITy);
2732  if (Pred == ICmpInst::ICMP_ULE)
2733  return getTrue(ITy);
2734  }
2735 
2736  // x >=u x >> y
2737  // x >=u x udiv y.
2738  if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2739  match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2740  // icmp pred X, (X op Y)
2741  if (Pred == ICmpInst::ICMP_ULT)
2742  return getFalse(ITy);
2743  if (Pred == ICmpInst::ICMP_UGE)
2744  return getTrue(ITy);
2745  }
2746 
2747  // handle:
2748  // CI2 << X == CI
2749  // CI2 << X != CI
2750  //
2751  // where CI2 is a power of 2 and CI isn't
2752  if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2753  const APInt *CI2Val, *CIVal = &CI->getValue();
2754  if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2755  CI2Val->isPowerOf2()) {
2756  if (!CIVal->isPowerOf2()) {
2757  // CI2 << X can equal zero in some circumstances,
2758  // this simplification is unsafe if CI is zero.
2759  //
2760  // We know it is safe if:
2761  // - The shift is nsw, we can't shift out the one bit.
2762  // - The shift is nuw, we can't shift out the one bit.
2763  // - CI2 is one
2764  // - CI isn't zero
2765  if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2766  Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2767  CI2Val->isOneValue() || !CI->isZero()) {
2768  if (Pred == ICmpInst::ICMP_EQ)
2769  return ConstantInt::getFalse(RHS->getContext());
2770  if (Pred == ICmpInst::ICMP_NE)
2771  return ConstantInt::getTrue(RHS->getContext());
2772  }
2773  }
2774  if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2775  if (Pred == ICmpInst::ICMP_UGT)
2776  return ConstantInt::getFalse(RHS->getContext());
2777  if (Pred == ICmpInst::ICMP_ULE)
2778  return ConstantInt::getTrue(RHS->getContext());
2779  }
2780  }
2781  }
2782 
2783  if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2784  LBO->getOperand(1) == RBO->getOperand(1)) {
2785  switch (LBO->getOpcode()) {
2786  default:
2787  break;
2788  case Instruction::UDiv:
2789  case Instruction::LShr:
2790  if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2791  !Q.IIQ.isExact(RBO))
2792  break;
2793  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2794  RBO->getOperand(0), Q, MaxRecurse - 1))
2795  return V;
2796  break;
2797  case Instruction::SDiv:
2798  if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2799  !Q.IIQ.isExact(RBO))
2800  break;
2801  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2802  RBO->getOperand(0), Q, MaxRecurse - 1))
2803  return V;
2804  break;
2805  case Instruction::AShr:
2806  if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2807  break;
2808  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2809  RBO->getOperand(0), Q, MaxRecurse - 1))
2810  return V;
2811  break;
2812  case Instruction::Shl: {
2813  bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
2814  bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
2815  if (!NUW && !NSW)
2816  break;
2817  if (!NSW && ICmpInst::isSigned(Pred))
2818  break;
2819  if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2820  RBO->getOperand(0), Q, MaxRecurse - 1))
2821  return V;
2822  break;
2823  }
2824  }
2825  }
2826  return nullptr;
2827 }
2828 
2829 /// Simplify integer comparisons where at least one operand of the compare
2830 /// matches an integer min/max idiom.
2832  Value *RHS, const SimplifyQuery &Q,
2833  unsigned MaxRecurse) {
2834  Type *ITy = GetCompareTy(LHS); // The return type.
2835  Value *A, *B;
2837  CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2838 
2839  // Signed variants on "max(a,b)>=a -> true".
2840  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2841  if (A != RHS)
2842  std::swap(A, B); // smax(A, B) pred A.
2843  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2844  // We analyze this as smax(A, B) pred A.
2845  P = Pred;
2846  } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2847  (A == LHS || B == LHS)) {
2848  if (A != LHS)
2849  std::swap(A, B); // A pred smax(A, B).
2850  EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2851  // We analyze this as smax(A, B) swapped-pred A.
2852  P = CmpInst::getSwappedPredicate(Pred);
2853  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2854  (A == RHS || B == RHS)) {
2855  if (A != RHS)
2856  std::swap(A, B); // smin(A, B) pred A.
2857  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2858  // We analyze this as smax(-A, -B) swapped-pred -A.
2859  // Note that we do not need to actually form -A or -B thanks to EqP.
2860  P = CmpInst::getSwappedPredicate(Pred);
2861  } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2862  (A == LHS || B == LHS)) {
2863  if (A != LHS)
2864  std::swap(A, B); // A pred smin(A, B).
2865  EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2866  // We analyze this as smax(-A, -B) pred -A.
2867  // Note that we do not need to actually form -A or -B thanks to EqP.
2868  P = Pred;
2869  }
2870  if (P != CmpInst::BAD_ICMP_PREDICATE) {
2871  // Cases correspond to "max(A, B) p A".
2872  switch (P) {
2873  default:
2874  break;
2875  case CmpInst::ICMP_EQ:
2876  case CmpInst::ICMP_SLE:
2877  // Equivalent to "A EqP B". This may be the same as the condition tested
2878  // in the max/min; if so, we can just return that.
2879  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2880  return V;
2881  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2882  return V;
2883  // Otherwise, see if "A EqP B" simplifies.
2884  if (MaxRecurse)
2885  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2886  return V;
2887  break;
2888  case CmpInst::ICMP_NE:
2889  case CmpInst::ICMP_SGT: {
2891  // Equivalent to "A InvEqP B". This may be the same as the condition
2892  // tested in the max/min; if so, we can just return that.
2893  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2894  return V;
2895  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2896  return V;
2897  // Otherwise, see if "A InvEqP B" simplifies.
2898  if (MaxRecurse)
2899  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2900  return V;
2901  break;
2902  }
2903  case CmpInst::ICMP_SGE:
2904  // Always true.
2905  return getTrue(ITy);
2906  case CmpInst::ICMP_SLT:
2907  // Always false.
2908  return getFalse(ITy);
2909  }
2910  }
2911 
2912  // Unsigned variants on "max(a,b)>=a -> true".
2914  if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2915  if (A != RHS)
2916  std::swap(A, B); // umax(A, B) pred A.
2917  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2918  // We analyze this as umax(A, B) pred A.
2919  P = Pred;
2920  } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2921  (A == LHS || B == LHS)) {
2922  if (A != LHS)
2923  std::swap(A, B); // A pred umax(A, B).
2924  EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2925  // We analyze this as umax(A, B) swapped-pred A.
2926  P = CmpInst::getSwappedPredicate(Pred);
2927  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2928  (A == RHS || B == RHS)) {
2929  if (A != RHS)
2930  std::swap(A, B); // umin(A, B) pred A.
2931  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2932  // We analyze this as umax(-A, -B) swapped-pred -A.
2933  // Note that we do not need to actually form -A or -B thanks to EqP.
2934  P = CmpInst::getSwappedPredicate(Pred);
2935  } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2936  (A == LHS || B == LHS)) {
2937  if (A != LHS)
2938  std::swap(A, B); // A pred umin(A, B).
2939  EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2940  // We analyze this as umax(-A, -B) pred -A.
2941  // Note that we do not need to actually form -A or -B thanks to EqP.
2942  P = Pred;
2943  }
2944  if (P != CmpInst::BAD_ICMP_PREDICATE) {
2945  // Cases correspond to "max(A, B) p A".
2946  switch (P) {
2947  default:
2948  break;
2949  case CmpInst::ICMP_EQ:
2950  case CmpInst::ICMP_ULE:
2951  // Equivalent to "A EqP B". This may be the same as the condition tested
2952  // in the max/min; if so, we can just return that.
2953  if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2954  return V;
2955  if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2956  return V;
2957  // Otherwise, see if "A EqP B" simplifies.
2958  if (MaxRecurse)
2959  if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2960  return V;
2961  break;
2962  case CmpInst::ICMP_NE:
2963  case CmpInst::ICMP_UGT: {
2965  // Equivalent to "A InvEqP B". This may be the same as the condition
2966  // tested in the max/min; if so, we can just return that.
2967  if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2968  return V;
2969  if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2970  return V;
2971  // Otherwise, see if "A InvEqP B" simplifies.
2972  if (MaxRecurse)
2973  if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2974  return V;
2975  break;
2976  }
2977  case CmpInst::ICMP_UGE:
2978  // Always true.
2979  return getTrue(ITy);
2980  case CmpInst::ICMP_ULT:
2981  // Always false.
2982  return getFalse(ITy);
2983  }
2984  }
2985 
2986  // Variants on "max(x,y) >= min(x,z)".
2987  Value *C, *D;
2988  if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2989  match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2990  (A == C || A == D || B == C || B == D)) {
2991  // max(x, ?) pred min(x, ?).
2992  if (Pred == CmpInst::ICMP_SGE)
2993  // Always true.
2994  return getTrue(ITy);
2995  if (Pred == CmpInst::ICMP_SLT)
2996  // Always false.
2997  return getFalse(ITy);
2998  } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2999  match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3000  (A == C || A == D || B == C || B == D)) {
3001  // min(x, ?) pred max(x, ?).
3002  if (Pred == CmpInst::ICMP_SLE)
3003  // Always true.
3004  return getTrue(ITy);
3005  if (Pred == CmpInst::ICMP_SGT)
3006  // Always false.
3007  return getFalse(ITy);
3008  } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3009  match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3010  (A == C || A == D || B == C || B == D)) {
3011  // max(x, ?) pred min(x, ?).
3012  if (Pred == CmpInst::ICMP_UGE)
3013  // Always true.
3014  return getTrue(ITy);
3015  if (Pred == CmpInst::ICMP_ULT)
3016  // Always false.
3017  return getFalse(ITy);
3018  } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3019  match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3020  (A == C || A == D || B == C || B == D)) {
3021  // min(x, ?) pred max(x, ?).
3022  if (Pred == CmpInst::ICMP_ULE)
3023  // Always true.
3024  return getTrue(ITy);
3025  if (Pred == CmpInst::ICMP_UGT)
3026  // Always false.
3027  return getFalse(ITy);
3028  }
3029 
3030  return nullptr;
3031 }
3032 
3033 /// Given operands for an ICmpInst, see if we can fold the result.
3034 /// If not, this returns null.
3035 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3036  const SimplifyQuery &Q, unsigned MaxRecurse) {
3037  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3038  assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3039 
3040  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3041  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3042  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3043 
3044  // If we have a constant, make sure it is on the RHS.
3045  std::swap(LHS, RHS);
3046  Pred = CmpInst::getSwappedPredicate(Pred);
3047  }
3048  assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3049 
3050  Type *ITy = GetCompareTy(LHS); // The return type.
3051 
3052  // For EQ and NE, we can always pick a value for the undef to make the
3053  // predicate pass or fail, so we can return undef.
3054  // Matches behavior in llvm::ConstantFoldCompareInstruction.
3055  if (isa<UndefValue>(RHS) && ICmpInst::isEquality(Pred))
3056  return UndefValue::get(ITy);
3057 
3058  // icmp X, X -> true/false
3059  // icmp X, undef -> true/false because undef could be X.
3060  if (LHS == RHS || isa<UndefValue>(RHS))
3061  return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3062 
3063  if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3064  return V;
3065 
3066  if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3067  return V;
3068 
3069  if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3070  return V;
3071 
3072  // If both operands have range metadata, use the metadata
3073  // to simplify the comparison.
3074  if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3075  auto RHS_Instr = cast<Instruction>(RHS);
3076  auto LHS_Instr = cast<Instruction>(LHS);
3077 
3078  if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3079  Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3080  auto RHS_CR = getConstantRangeFromMetadata(
3081  *RHS_Instr->getMetadata(LLVMContext::MD_range));
3082  auto LHS_CR = getConstantRangeFromMetadata(
3083  *LHS_Instr->getMetadata(LLVMContext::MD_range));
3084 
3085  auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3086  if (Satisfied_CR.contains(LHS_CR))
3087  return ConstantInt::getTrue(RHS->getContext());
3088 
3089  auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3090  CmpInst::getInversePredicate(Pred), RHS_CR);
3091  if (InversedSatisfied_CR.contains(LHS_CR))
3092  return ConstantInt::getFalse(RHS->getContext());
3093  }
3094  }
3095 
3096  // Compare of cast, for example (zext X) != 0 -> X != 0
3097  if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3098  Instruction *LI = cast<CastInst>(LHS);
3099  Value *SrcOp = LI->getOperand(0);
3100  Type *SrcTy = SrcOp->getType();
3101  Type *DstTy = LI->getType();
3102 
3103  // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3104  // if the integer type is the same size as the pointer type.
3105  if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3106  Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3107  if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3108  // Transfer the cast to the constant.
3109  if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3110  ConstantExpr::getIntToPtr(RHSC, SrcTy),
3111  Q, MaxRecurse-1))
3112  return V;
3113  } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3114  if (RI->getOperand(0)->getType() == SrcTy)
3115  // Compare without the cast.
3116  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3117  Q, MaxRecurse-1))
3118  return V;
3119  }
3120  }
3121 
3122  if (isa<ZExtInst>(LHS)) {
3123  // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3124  // same type.
3125  if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3126  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3127  // Compare X and Y. Note that signed predicates become unsigned.
3129  SrcOp, RI->getOperand(0), Q,
3130  MaxRecurse-1))
3131  return V;
3132  }
3133  // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3134  // too. If not, then try to deduce the result of the comparison.
3135  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3136  // Compute the constant that would happen if we truncated to SrcTy then
3137  // reextended to DstTy.
3138  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3139  Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3140 
3141  // If the re-extended constant didn't change then this is effectively
3142  // also a case of comparing two zero-extended values.
3143  if (RExt == CI && MaxRecurse)
3145  SrcOp, Trunc, Q, MaxRecurse-1))
3146  return V;
3147 
3148  // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3149  // there. Use this to work out the result of the comparison.
3150  if (RExt != CI) {
3151  switch (Pred) {
3152  default: llvm_unreachable("Unknown ICmp predicate!");
3153  // LHS <u RHS.
3154  case ICmpInst::ICMP_EQ:
3155  case ICmpInst::ICMP_UGT:
3156  case ICmpInst::ICMP_UGE:
3157  return ConstantInt::getFalse(CI->getContext());
3158 
3159  case ICmpInst::ICMP_NE:
3160  case ICmpInst::ICMP_ULT:
3161  case ICmpInst::ICMP_ULE:
3162  return ConstantInt::getTrue(CI->getContext());
3163 
3164  // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3165  // is non-negative then LHS <s RHS.
3166  case ICmpInst::ICMP_SGT:
3167  case ICmpInst::ICMP_SGE:
3168  return CI->getValue().isNegative() ?
3169  ConstantInt::getTrue(CI->getContext()) :
3170  ConstantInt::getFalse(CI->getContext());
3171 
3172  case ICmpInst::ICMP_SLT:
3173  case ICmpInst::ICMP_SLE:
3174  return CI->getValue().isNegative() ?
3175  ConstantInt::getFalse(CI->getContext()) :
3176  ConstantInt::getTrue(CI->getContext());
3177  }
3178  }
3179  }
3180  }
3181 
3182  if (isa<SExtInst>(LHS)) {
3183  // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3184  // same type.
3185  if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3186  if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3187  // Compare X and Y. Note that the predicate does not change.
3188  if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3189  Q, MaxRecurse-1))
3190  return V;
3191  }
3192  // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3193  // too. If not, then try to deduce the result of the comparison.
3194  else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3195  // Compute the constant that would happen if we truncated to SrcTy then
3196  // reextended to DstTy.
3197  Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3198  Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3199 
3200  // If the re-extended constant didn't change then this is effectively
3201  // also a case of comparing two sign-extended values.
3202  if (RExt == CI && MaxRecurse)
3203  if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3204  return V;
3205 
3206  // Otherwise the upper bits of LHS are all equal, while RHS has varying
3207  // bits there. Use this to work out the result of the comparison.
3208  if (RExt != CI) {
3209  switch (Pred) {
3210  default: llvm_unreachable("Unknown ICmp predicate!");
3211  case ICmpInst::ICMP_EQ:
3212  return ConstantInt::getFalse(CI->getContext());
3213  case ICmpInst::ICMP_NE:
3214  return ConstantInt::getTrue(CI->getContext());
3215 
3216  // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3217  // LHS >s RHS.
3218  case ICmpInst::ICMP_SGT:
3219  case ICmpInst::ICMP_SGE:
3220  return CI->getValue().isNegative() ?
3221  ConstantInt::getTrue(CI->getContext()) :
3222  ConstantInt::getFalse(CI->getContext());
3223  case ICmpInst::ICMP_SLT:
3224  case ICmpInst::ICMP_SLE:
3225  return CI->getValue().isNegative() ?
3226  ConstantInt::getFalse(CI->getContext()) :
3227  ConstantInt::getTrue(CI->getContext());
3228 
3229  // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3230  // LHS >u RHS.
3231  case ICmpInst::ICMP_UGT:
3232  case ICmpInst::ICMP_UGE:
3233  // Comparison is true iff the LHS <s 0.
3234  if (MaxRecurse)
3235  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3236  Constant::getNullValue(SrcTy),
3237  Q, MaxRecurse-1))
3238  return V;
3239  break;
3240  case ICmpInst::ICMP_ULT:
3241  case ICmpInst::ICMP_ULE:
3242  // Comparison is true iff the LHS >=s 0.
3243  if (MaxRecurse)
3244  if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3245  Constant::getNullValue(SrcTy),
3246  Q, MaxRecurse-1))
3247  return V;
3248  break;
3249  }
3250  }
3251  }
3252  }
3253  }
3254 
3255  // icmp eq|ne X, Y -> false|true if X != Y
3256  if (ICmpInst::isEquality(Pred) &&
3257  isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3258  return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3259  }
3260 
3261  if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3262  return V;
3263 
3264  if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3265  return V;
3266 
3267  // Simplify comparisons of related pointers using a powerful, recursive
3268  // GEP-walk when we have target data available..
3269  if (LHS->getType()->isPointerTy())
3270  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3271  Q.IIQ, LHS, RHS))
3272  return C;
3273  if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3274  if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3275  if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3276  Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3277  Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3278  Q.DL.getTypeSizeInBits(CRHS->getType()))
3279  if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3280  Q.IIQ, CLHS->getPointerOperand(),
3281  CRHS->getPointerOperand()))
3282  return C;
3283 
3284  if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3285  if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3286  if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3287  GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3288  (ICmpInst::isEquality(Pred) ||
3289  (GLHS->isInBounds() && GRHS->isInBounds() &&
3290  Pred == ICmpInst::getSignedPredicate(Pred)))) {
3291  // The bases are equal and the indices are constant. Build a constant
3292  // expression GEP with the same indices and a null base pointer to see
3293  // what constant folding can make out of it.
3294  Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3295  SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3297  GLHS->getSourceElementType(), Null, IndicesLHS);
3298 
3299  SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3301  GLHS->getSourceElementType(), Null, IndicesRHS);
3302  return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3303  }
3304  }
3305  }
3306 
3307  // If the comparison is with the result of a select instruction, check whether
3308  // comparing with either branch of the select always yields the same value.
3309  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3310  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3311  return V;
3312 
3313  // If the comparison is with the result of a phi instruction, check whether
3314  // doing the compare with each incoming phi value yields a common result.
3315  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3316  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3317  return V;
3318 
3319  return nullptr;
3320 }
3321 
3322 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3323  const SimplifyQuery &Q) {
3324  return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3325 }
3326 
3327 /// Given operands for an FCmpInst, see if we can fold the result.
3328 /// If not, this returns null.
3329 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3330  FastMathFlags FMF, const SimplifyQuery &Q,
3331  unsigned MaxRecurse) {
3332  CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3333  assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3334 
3335  if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3336  if (Constant *CRHS = dyn_cast<Constant>(RHS))
3337  return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3338 
3339  // If we have a constant, make sure it is on the RHS.
3340  std::swap(LHS, RHS);
3341  Pred = CmpInst::getSwappedPredicate(Pred);
3342  }
3343 
3344  // Fold trivial predicates.
3345  Type *RetTy = GetCompareTy(LHS);
3346  if (Pred == FCmpInst::FCMP_FALSE)
3347  return getFalse(RetTy);
3348  if (Pred == FCmpInst::FCMP_TRUE)
3349  return getTrue(RetTy);
3350 
3351  // Fold (un)ordered comparison if we can determine there are no NaNs.
3352  if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3353  if (FMF.noNaNs() ||
3354  (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3355  return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3356 
3357  // NaN is unordered; NaN is not ordered.
3359  "Comparison must be either ordered or unordered");
3360  if (match(RHS, m_NaN()))
3361  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3362 
3363  // fcmp pred x, undef and fcmp pred undef, x
3364  // fold to true if unordered, false if ordered
3365  if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3366  // Choosing NaN for the undef will always make unordered comparison succeed
3367  // and ordered comparison fail.
3368  return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3369  }
3370 
3371  // fcmp x,x -> true/false. Not all compares are foldable.
3372  if (LHS == RHS) {
3373  if (CmpInst::isTrueWhenEqual(Pred))
3374  return getTrue(RetTy);
3375  if (CmpInst::isFalseWhenEqual(Pred))
3376  return getFalse(RetTy);
3377  }
3378 
3379  // Handle fcmp with constant RHS.
3380  // TODO: Use match with a specific FP value, so these work with vectors with
3381  // undef lanes.
3382  const APFloat *C;
3383  if (match(RHS, m_APFloat(C))) {
3384  // Check whether the constant is an infinity.
3385  if (C->isInfinity()) {
3386  if (C->isNegative()) {
3387  switch (Pred) {
3388  case FCmpInst::FCMP_OLT:
3389  // No value is ordered and less than negative infinity.
3390  return getFalse(RetTy);
3391  case FCmpInst::FCMP_UGE:
3392  // All values are unordered with or at least negative infinity.
3393  return getTrue(RetTy);
3394  default:
3395  break;
3396  }
3397  } else {
3398  switch (Pred) {
3399  case FCmpInst::FCMP_OGT:
3400  // No value is ordered and greater than infinity.
3401  return getFalse(RetTy);
3402  case FCmpInst::FCMP_ULE:
3403  // All values are unordered with and at most infinity.
3404  return getTrue(RetTy);
3405  default:
3406  break;
3407  }
3408  }
3409  }
3410  if (C->isNegative() && !C->isNegZero()) {
3411  assert(!C->isNaN() && "Unexpected NaN constant!");
3412  // TODO: We can catch more cases by using a range check rather than
3413  // relying on CannotBeOrderedLessThanZero.
3414  switch (Pred) {
3415  case FCmpInst::FCMP_UGE:
3416  case FCmpInst::FCMP_UGT:
3417  case FCmpInst::FCMP_UNE:
3418  // (X >= 0) implies (X > C) when (C < 0)
3419  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3420  return getTrue(RetTy);
3421  break;
3422  case FCmpInst::FCMP_OEQ:
3423  case FCmpInst::FCMP_OLE:
3424  case FCmpInst::FCMP_OLT:
3425  // (X >= 0) implies !(X < C) when (C < 0)
3426  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3427  return getFalse(RetTy);
3428  break;
3429  default:
3430  break;
3431  }
3432  }
3433  }
3434  if (match(RHS, m_AnyZeroFP())) {
3435  switch (Pred) {
3436  case FCmpInst::FCMP_OGE:
3437  if (FMF.noNaNs() && CannotBeOrderedLessThanZero(LHS, Q.TLI))
3438  return getTrue(RetTy);
3439  break;
3440  case FCmpInst::FCMP_UGE:
3441  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3442  return getTrue(RetTy);
3443  break;
3444  case FCmpInst::FCMP_ULT:
3445  if (FMF.noNaNs() && CannotBeOrderedLessThanZero(LHS, Q.TLI))
3446  return getFalse(RetTy);
3447  break;
3448  case FCmpInst::FCMP_OLT:
3449  if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3450  return getFalse(RetTy);
3451  break;
3452  default:
3453  break;
3454  }
3455  }
3456 
3457  // If the comparison is with the result of a select instruction, check whether
3458  // comparing with either branch of the select always yields the same value.
3459  if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3460  if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3461  return V;
3462 
3463  // If the comparison is with the result of a phi instruction, check whether
3464  // doing the compare with each incoming phi value yields a common result.
3465  if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3466  if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3467  return V;
3468 
3469  return nullptr;
3470 }
3471 
3472 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3473  FastMathFlags FMF, const SimplifyQuery &Q) {
3474  return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3475 }
3476 
3477 /// See if V simplifies when its operand Op is replaced with RepOp.
3478 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3479  const SimplifyQuery &Q,
3480  unsigned MaxRecurse) {
3481  // Trivial replacement.
3482  if (V == Op)
3483  return RepOp;
3484 
3485  // We cannot replace a constant, and shouldn't even try.
3486  if (isa<Constant>(Op))
3487  return nullptr;
3488 
3489  auto *I = dyn_cast<Instruction>(V);
3490  if (!I)
3491  return nullptr;
3492 
3493  // If this is a binary operator, try to simplify it with the replaced op.
3494  if (auto *B = dyn_cast<BinaryOperator>(I)) {
3495  // Consider:
3496  // %cmp = icmp eq i32 %x, 2147483647
3497  // %add = add nsw i32 %x, 1
3498  // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3499  //
3500  // We can't replace %sel with %add unless we strip away the flags.
3501  if (isa<OverflowingBinaryOperator>(B))
3502  if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3503  return nullptr;
3504  if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3505  return nullptr;
3506 
3507  if (MaxRecurse) {
3508  if (B->getOperand(0) == Op)
3509  return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3510  MaxRecurse - 1);
3511  if (B->getOperand(1) == Op)
3512  return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3513  MaxRecurse - 1);
3514  }
3515  }
3516 
3517  // Same for CmpInsts.
3518  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3519  if (MaxRecurse) {
3520  if (C->getOperand(0) == Op)
3521  return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3522  MaxRecurse - 1);
3523  if (C->getOperand(1) == Op)
3524  return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3525  MaxRecurse - 1);
3526  }
3527  }
3528 
3529  // Same for GEPs.
3530  if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3531  if (MaxRecurse) {
3532  SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3533  transform(GEP->operands(), NewOps.begin(),
3534  [&](Value *V) { return V == Op ? RepOp : V; });
3535  return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3536  MaxRecurse - 1);
3537  }
3538  }
3539 
3540  // TODO: We could hand off more cases to instsimplify here.
3541 
3542  // If all operands are constant after substituting Op for RepOp then we can
3543  // constant fold the instruction.
3544  if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3545  // Build a list of all constant operands.
3546  SmallVector<Constant *, 8> ConstOps;
3547  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3548  if (I->getOperand(i) == Op)
3549  ConstOps.push_back(CRepOp);
3550  else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3551  ConstOps.push_back(COp);
3552  else
3553  break;
3554  }
3555 
3556  // All operands were constants, fold it.
3557  if (ConstOps.size() == I->getNumOperands()) {
3558  if (CmpInst *C = dyn_cast<CmpInst>(I))
3559  return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3560  ConstOps[1], Q.DL, Q.TLI);
3561 
3562  if (LoadInst *LI = dyn_cast<LoadInst>(I))
3563  if (!LI->isVolatile())
3564  return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3565 
3566  return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3567  }
3568  }
3569 
3570  return nullptr;
3571 }
3572 
3573 /// Try to simplify a select instruction when its condition operand is an
3574 /// integer comparison where one operand of the compare is a constant.
3575 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3576  const APInt *Y, bool TrueWhenUnset) {
3577  const APInt *C;
3578 
3579  // (X & Y) == 0 ? X & ~Y : X --> X
3580  // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3581  if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3582  *Y == ~*C)
3583  return TrueWhenUnset ? FalseVal : TrueVal;
3584 
3585  // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3586  // (X & Y) != 0 ? X : X & ~Y --> X
3587  if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3588  *Y == ~*C)
3589  return TrueWhenUnset ? FalseVal : TrueVal;
3590 
3591  if (Y->isPowerOf2()) {
3592  // (X & Y) == 0 ? X | Y : X --> X | Y
3593  // (X & Y) != 0 ? X | Y : X --> X
3594  if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3595  *Y == *C)
3596  return TrueWhenUnset ? TrueVal : FalseVal;
3597 
3598  // (X & Y) == 0 ? X : X | Y --> X
3599  // (X & Y) != 0 ? X : X | Y --> X | Y
3600  if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3601  *Y == *C)
3602  return TrueWhenUnset ? TrueVal : FalseVal;
3603  }
3604 
3605  return nullptr;
3606 }
3607 
3608 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3609 /// eq/ne.
3611  ICmpInst::Predicate Pred,
3612  Value *TrueVal, Value *FalseVal) {
3613  Value *X;
3614  APInt Mask;
3615  if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3616  return nullptr;
3617 
3618  return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3619  Pred == ICmpInst::ICMP_EQ);
3620 }
3621 
3622 /// Try to simplify a select instruction when its condition operand is an
3623 /// integer comparison.
3624 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3625  Value *FalseVal, const SimplifyQuery &Q,
3626  unsigned MaxRecurse) {
3627  ICmpInst::Predicate Pred;
3628  Value *CmpLHS, *CmpRHS;
3629  if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3630  return nullptr;
3631 
3632  if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3633  Value *X;
3634  const APInt *Y;
3635  if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3636  if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3637  Pred == ICmpInst::ICMP_EQ))
3638  return V;
3639 
3640  // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
3641  Value *ShAmt;
3642  auto isFsh = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(),
3643  m_Value(ShAmt)),
3644  m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X),
3645  m_Value(ShAmt)));
3646  // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
3647  // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
3648  if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt &&
3649  Pred == ICmpInst::ICMP_EQ)
3650  return X;
3651  // (ShAmt != 0) ? X : fshl(X, *, ShAmt) --> X
3652  // (ShAmt != 0) ? X : fshr(*, X, ShAmt) --> X
3653  if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt &&
3654  Pred == ICmpInst::ICMP_NE)
3655  return X;
3656 
3657  // Test for a zero-shift-guard-op around rotates. These are used to
3658  // avoid UB from oversized shifts in raw IR rotate patterns, but the
3659  // intrinsics do not have that problem.
3660  // We do not allow this transform for the general funnel shift case because
3661  // that would not preserve the poison safety of the original code.
3662  auto isRotate = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X),
3663  m_Deferred(X),
3664  m_Value(ShAmt)),
3665  m_Intrinsic<Intrinsic::fshr>(m_Value(X),
3666  m_Deferred(X),
3667  m_Value(ShAmt)));
3668  // (ShAmt != 0) ? fshl(X, X, ShAmt) : X --> fshl(X, X, ShAmt)
3669  // (ShAmt != 0) ? fshr(X, X, ShAmt) : X --> fshr(X, X, ShAmt)
3670  if (match(TrueVal, isRotate) && FalseVal == X && CmpLHS == ShAmt &&
3671  Pred == ICmpInst::ICMP_NE)
3672  return TrueVal;
3673  // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
3674  // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
3675  if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
3676  Pred == ICmpInst::ICMP_EQ)
3677  return FalseVal;
3678  }
3679 
3680  // Check for other compares that behave like bit test.
3681  if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3682  TrueVal, FalseVal))
3683  return V;
3684 
3685  // If we have an equality comparison, then we know the value in one of the
3686  // arms of the select. See if substituting this value into the arm and
3687  // simplifying the result yields the same value as the other arm.
3688  if (Pred == ICmpInst::ICMP_EQ) {
3689  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3690  TrueVal ||
3691  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3692  TrueVal)
3693  return FalseVal;
3694  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3695  FalseVal ||
3696  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3697  FalseVal)
3698  return FalseVal;
3699  } else if (Pred == ICmpInst::ICMP_NE) {
3700  if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3701  FalseVal ||
3702  SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3703  FalseVal)
3704  return TrueVal;
3705  if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3706  TrueVal ||
3707  SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3708  TrueVal)
3709  return TrueVal;
3710  }
3711 
3712  return nullptr;
3713 }
3714 
3715 /// Try to simplify a select instruction when its condition operand is a
3716 /// floating-point comparison.
3718  FCmpInst::Predicate Pred;
3719  if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
3720  !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
3721  return nullptr;
3722 
3723  // TODO: The transform may not be valid with -0.0. An incomplete way of
3724  // testing for that possibility is to check if at least one operand is a
3725  // non-zero constant.
3726  const APFloat *C;
3727  if ((match(T, m_APFloat(C)) && C->isNonZero()) ||
3728  (match(F, m_APFloat(C)) && C->isNonZero())) {
3729  // (T == F) ? T : F --> F
3730  // (F == T) ? T : F --> F
3731  if (Pred == FCmpInst::FCMP_OEQ)
3732  return F;
3733 
3734  // (T != F) ? T : F --> T
3735  // (F != T) ? T : F --> T
3736  if (Pred == FCmpInst::FCMP_UNE)
3737  return T;
3738  }
3739 
3740  return nullptr;
3741 }
3742 
3743 /// Given operands for a SelectInst, see if we can fold the result.
3744 /// If not, this returns null.
3745 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3746  const SimplifyQuery &Q, unsigned MaxRecurse) {
3747  if (auto *CondC = dyn_cast<Constant>(Cond)) {
3748  if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3749  if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3750  return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3751 
3752  // select undef, X, Y -> X or Y
3753  if (isa<UndefValue>(CondC))
3754  return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3755 
3756  // TODO: Vector constants with undef elements don't simplify.
3757 
3758  // select true, X, Y -> X
3759  if (CondC->isAllOnesValue())
3760  return TrueVal;
3761  // select false, X, Y -> Y
3762  if (CondC->isNullValue())
3763  return FalseVal;
3764  }
3765 
3766  // select ?, X, X -> X
3767  if (TrueVal == FalseVal)
3768  return TrueVal;
3769 
3770  if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3771  return FalseVal;
3772  if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3773  return TrueVal;
3774 
3775  if (Value *V =
3776  simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3777  return V;
3778 
3779  if (Value *V = simplifySelectWithFCmp(Cond, TrueVal, FalseVal))
3780  return V;
3781 
3782  if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3783  return V;
3784 
3785  Optional<bool> Imp = isImpliedByDomCondition(Cond, Q.CxtI, Q.DL);
3786  if (Imp)
3787  return *Imp ? TrueVal : FalseVal;
3788 
3789  return nullptr;
3790 }
3791 
3792 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3793  const SimplifyQuery &Q) {
3794  return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3795 }
3796 
3797 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3798 /// If not, this returns null.
3800  const SimplifyQuery &Q, unsigned) {
3801  // The type of the GEP pointer operand.
3802  unsigned AS =
3803  cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3804 
3805  // getelementptr P -> P.
3806  if (Ops.size() == 1)
3807  return Ops[0];
3808 
3809  // Compute the (pointer) type returned by the GEP instruction.
3810  Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3811  Type *GEPTy = PointerType::get(LastType, AS);
3812  if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3813  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3814  else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3815  GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3816 
3817  if (isa<UndefValue>(Ops[0]))
3818  return UndefValue::get(GEPTy);
3819 
3820  if (Ops.size() == 2) {
3821  // getelementptr P, 0 -> P.
3822  if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3823  return Ops[0];
3824 
3825  Type *Ty = SrcTy;
3826  if (Ty->isSized()) {
3827  Value *P;
3828  uint64_t C;
3829  uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3830  // getelementptr P, N -> P if P points to a type of zero size.
3831  if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3832  return Ops[0];
3833 
3834  // The following transforms are only safe if the ptrtoint cast
3835  // doesn't truncate the pointers.
3836  if (Ops[1]->getType()->getScalarSizeInBits() ==
3837  Q.DL.getIndexSizeInBits(AS)) {
3838  auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3839  if (match(P, m_Zero()))
3840  return Constant::getNullValue(GEPTy);
3841  Value *Temp;
3842  if (match(P, m_PtrToInt(m_Value(Temp))))
3843  if (Temp->getType() == GEPTy)
3844  return Temp;
3845  return nullptr;
3846  };
3847 
3848  // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3849  if (TyAllocSize == 1 &&
3850  match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3851  if (Value *R = PtrToIntOrZero(P))
3852  return R;
3853 
3854  // getelementptr V, (ashr (sub P, V), C) -> Q
3855  // if P points to a type of size 1 << C.
3856  if (match(Ops[1],
3857  m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3858  m_ConstantInt(C))) &&
3859  TyAllocSize == 1ULL << C)
3860  if (Value *R = PtrToIntOrZero(P))
3861  return R;
3862 
3863  // getelementptr V, (sdiv (sub P, V), C) -> Q
3864  // if P points to a type of size C.
3865  if (match(Ops[1],
3866  m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3867  m_SpecificInt(TyAllocSize))))
3868  if (Value *R = PtrToIntOrZero(P))
3869  return R;
3870  }
3871  }
3872  }
3873 
3874  if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3875  all_of(Ops.slice(1).drop_back(1),
3876  [](Value *Idx) { return match(Idx, m_Zero()); })) {
3877  unsigned IdxWidth =
3878  Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3879  if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3880  APInt BasePtrOffset(IdxWidth, 0);
3881  Value *StrippedBasePtr =
3882  Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3883  BasePtrOffset);
3884 
3885  // gep (gep V, C), (sub 0, V) -> C
3886  if (match(Ops.back(),
3887  m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3888  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3889  return ConstantExpr::getIntToPtr(CI, GEPTy);
3890  }
3891  // gep (gep V, C), (xor V, -1) -> C-1
3892  if (match(Ops.back(),
3893  m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3894  auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3895  return ConstantExpr::getIntToPtr(CI, GEPTy);
3896  }
3897  }
3898  }
3899 
3900  // Check to see if this is constant foldable.
3901  if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3902  return nullptr;
3903 
3904  auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3905  Ops.slice(1));
3906  if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3907  return CEFolded;
3908  return CE;
3909 }
3910 
3912  const SimplifyQuery &Q) {
3913  return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3914 }
3915 
3916 /// Given operands for an InsertValueInst, see if we can fold the result.
3917 /// If not, this returns null.
3919  ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3920  unsigned) {
3921  if (Constant *CAgg = dyn_cast<Constant>(Agg))
3922  if (Constant *CVal = dyn_cast<Constant>(Val))
3923  return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3924 
3925  // insertvalue x, undef, n -> x
3926  if (match(Val, m_Undef()))
3927  return Agg;
3928 
3929  // insertvalue x, (extractvalue y, n), n
3930  if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3931  if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3932  EV->getIndices() == Idxs) {
3933  // insertvalue undef, (extractvalue y, n), n -> y
3934  if (match(Agg, m_Undef()))
3935  return EV->getAggregateOperand();
3936 
3937  // insertvalue y, (extractvalue y, n), n -> y
3938  if (Agg == EV->getAggregateOperand())
3939  return Agg;
3940  }
3941 
3942  return nullptr;
3943 }
3944 
3946  ArrayRef<unsigned> Idxs,
3947  const SimplifyQuery &Q) {
3949 }
3950 
3952  const SimplifyQuery &Q) {
3953  // Try to constant fold.
3954  auto *VecC = dyn_cast<Constant>(Vec);
3955  auto *ValC = dyn_cast<Constant>(Val);
3956  auto *IdxC = dyn_cast<Constant>(Idx);
3957  if (VecC && ValC && IdxC)
3958  return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3959 
3960  // Fold into undef if index is out of bounds.
3961  if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3962  uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3963  if (CI->uge(NumElements))
3964  return UndefValue::get(Vec->getType());
3965  }
3966 
3967  // If index is undef, it might be out of bounds (see above case)
3968  if (isa<UndefValue>(Idx))
3969  return UndefValue::get(Vec->getType());
3970 
3971  return nullptr;
3972 }
3973 
3974 /// Given operands for an ExtractValueInst, see if we can fold the result.
3975 /// If not, this returns null.
3977  const SimplifyQuery &, unsigned) {
3978  if (auto *CAgg = dyn_cast<Constant>(Agg))
3979  return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3980 
3981  // extractvalue x, (insertvalue y, elt, n), n -> elt
3982  unsigned NumIdxs = Idxs.size();
3983  for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3984  IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3985  ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3986  unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3987  unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3988  if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3989  Idxs.slice(0, NumCommonIdxs)) {
3990  if (NumIdxs == NumInsertValueIdxs)
3991  return IVI->getInsertedValueOperand();
3992  break;
3993  }
3994  }
3995 
3996  return nullptr;
3997 }
3998 
4000  const SimplifyQuery &Q) {
4002 }
4003 
4004 /// Given operands for an ExtractElementInst, see if we can fold the result.
4005 /// If not, this returns null.
4007  unsigned) {
4008  if (auto *CVec = dyn_cast<Constant>(Vec)) {
4009  if (auto *CIdx = dyn_cast<Constant>(Idx))
4010  return ConstantFoldExtractElementInstruction(CVec, CIdx);
4011 
4012  // The index is not relevant if our vector is a splat.
4013  if (auto *Splat = CVec->getSplatValue())
4014  return Splat;
4015 
4016  if (isa<UndefValue>(Vec))
4017  return UndefValue::get(Vec->getType()->getVectorElementType());
4018  }
4019 
4020  // If extracting a specified index from the vector, see if we can recursively
4021  // find a previously computed scalar that was inserted into the vector.
4022  if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4023  if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4024  // definitely out of bounds, thus undefined result
4025  return UndefValue::get(Vec->getType()->getVectorElementType());
4026  if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4027  return Elt;
4028  }
4029 
4030  // An undef extract index can be arbitrarily chosen to be an out-of-range
4031  // index value, which would result in the instruction being undef.
4032  if (isa<UndefValue>(Idx))
4033  return UndefValue::get(Vec->getType()->getVectorElementType());
4034 
4035  return nullptr;
4036 }
4037 
4039  const SimplifyQuery &Q) {
4041 }
4042 
4043 /// See if we can fold the given phi. If not, returns null.
4044 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4045  // If all of the PHI's incoming values are the same then replace the PHI node
4046  // with the common value.
4047  Value *CommonValue = nullptr;
4048  bool HasUndefInput = false;
4049  for (Value *Incoming : PN->incoming_values()) {
4050  // If the incoming value is the phi node itself, it can safely be skipped.
4051  if (Incoming == PN) continue;
4052  if (isa<UndefValue>(Incoming)) {
4053  // Remember that we saw an undef value, but otherwise ignore them.
4054  HasUndefInput = true;
4055  continue;
4056  }
4057  if (CommonValue && Incoming != CommonValue)
4058  return nullptr; // Not the same, bail out.
4059  CommonValue = Incoming;
4060  }
4061 
4062  // If CommonValue is null then all of the incoming values were either undef or
4063  // equal to the phi node itself.
4064  if (!CommonValue)
4065  return UndefValue::get(PN->getType());
4066 
4067  // If we have a PHI node like phi(X, undef, X), where X is defined by some
4068  // instruction, we cannot return X as the result of the PHI node unless it
4069  // dominates the PHI block.
4070  if (HasUndefInput)
4071  return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4072 
4073  return CommonValue;
4074 }
4075 
4076 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4077  Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4078  if (auto *C = dyn_cast<Constant>(Op))
4079  return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4080 
4081  if (auto *CI = dyn_cast<CastInst>(Op)) {
4082  auto *Src = CI->getOperand(0);
4083  Type *SrcTy = Src->getType();
4084  Type *MidTy = CI->getType();
4085  Type *DstTy = Ty;
4086  if (Src->getType() == Ty) {
4087  auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4088  auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4089  Type *SrcIntPtrTy =
4090  SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4091  Type *MidIntPtrTy =
4092  MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4093  Type *DstIntPtrTy =
4094  DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4095  if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4096  SrcIntPtrTy, MidIntPtrTy,
4097  DstIntPtrTy) == Instruction::BitCast)
4098  return Src;
4099  }
4100  }
4101 
4102  // bitcast x -> x
4103  if (CastOpc == Instruction::BitCast)
4104  if (Op->getType() == Ty)
4105  return Op;
4106 
4107  return nullptr;
4108 }
4109 
4110 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4111  const SimplifyQuery &Q) {
4112  return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4113 }
4114 
4115 /// For the given destination element of a shuffle, peek through shuffles to
4116 /// match a root vector source operand that contains that element in the same
4117 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4118 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4119  int MaskVal, Value *RootVec,
4120  unsigned MaxRecurse) {
4121  if (!MaxRecurse--)
4122  return nullptr;
4123 
4124  // Bail out if any mask value is undefined. That kind of shuffle may be
4125  // simplified further based on demanded bits or other folds.
4126  if (MaskVal == -1)
4127  return nullptr;
4128 
4129  // The mask value chooses which source operand we need to look at next.
4130  int InVecNumElts = Op0->getType()->getVectorNumElements();
4131  int RootElt = MaskVal;
4132  Value *SourceOp = Op0;
4133  if (MaskVal >= InVecNumElts) {
4134  RootElt = MaskVal - InVecNumElts;
4135  SourceOp = Op1;
4136  }
4137 
4138  // If the source operand is a shuffle itself, look through it to find the
4139  // matching root vector.
4140  if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4141  return foldIdentityShuffles(
4142  DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4143  SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4144  }
4145 
4146  // TODO: Look through bitcasts? What if the bitcast changes the vector element
4147  // size?
4148 
4149  // The source operand is not a shuffle. Initialize the root vector value for
4150  // this shuffle if that has not been done yet.
4151  if (!RootVec)
4152  RootVec = SourceOp;
4153 
4154  // Give up as soon as a source operand does not match the existing root value.
4155  if (RootVec != SourceOp)
4156  return nullptr;
4157 
4158  // The element must be coming from the same lane in the source vector
4159  // (although it may have crossed lanes in intermediate shuffles).
4160  if (RootElt != DestElt)
4161  return nullptr;
4162 
4163  return RootVec;
4164 }
4165 
4167  Type *RetTy, const SimplifyQuery &Q,
4168  unsigned MaxRecurse) {
4169  if (isa<UndefValue>(Mask))
4170  return UndefValue::get(RetTy);
4171 
4172  Type *InVecTy = Op0->getType();
4173  unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4174  unsigned InVecNumElts = InVecTy->getVectorNumElements();
4175 
4176  SmallVector<int, 32> Indices;
4177  ShuffleVectorInst::getShuffleMask(Mask, Indices);
4178  assert(MaskNumElts == Indices.size() &&
4179  "Size of Indices not same as number of mask elements?");
4180 
4181  // Canonicalization: If mask does not select elements from an input vector,
4182  // replace that input vector with undef.
4183  bool MaskSelects0 = false, MaskSelects1 = false;
4184  for (unsigned i = 0; i != MaskNumElts; ++i) {
4185  if (Indices[i] == -1)
4186  continue;
4187  if ((unsigned)Indices[i] < InVecNumElts)
4188  MaskSelects0 = true;
4189  else
4190  MaskSelects1 = true;
4191  }
4192  if (!MaskSelects0)
4193  Op0 = UndefValue::get(InVecTy);
4194  if (!MaskSelects1)
4195  Op1 = UndefValue::get(InVecTy);
4196 
4197  auto *Op0Const = dyn_cast<Constant>(Op0);
4198  auto *Op1Const = dyn_cast<Constant>(Op1);
4199 
4200  // If all operands are constant, constant fold the shuffle.
4201  if (Op0Const && Op1Const)
4202  return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4203 
4204  // Canonicalization: if only one input vector is constant, it shall be the
4205  // second one.
4206  if (Op0Const && !Op1Const) {
4207  std::swap(Op0, Op1);
4208  ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4209  }
4210 
4211  // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4212  // value type is same as the input vectors' type.
4213  if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4214  if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4215  OpShuf->getMask()->getSplatValue())
4216  return Op0;
4217 
4218  // Don't fold a shuffle with undef mask elements. This may get folded in a
4219  // better way using demanded bits or other analysis.
4220  // TODO: Should we allow this?
4221  if (find(Indices, -1) != Indices.end())
4222  return nullptr;
4223 
4224  // Check if every element of this shuffle can be mapped back to the
4225  // corresponding element of a single root vector. If so, we don't need this
4226  // shuffle. This handles simple identity shuffles as well as chains of
4227  // shuffles that may widen/narrow and/or move elements across lanes and back.
4228  Value *RootVec = nullptr;
4229  for (unsigned i = 0; i != MaskNumElts; ++i) {
4230  // Note that recursion is limited for each vector element, so if any element
4231  // exceeds the limit, this will fail to simplify.
4232  RootVec =
4233  foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4234 
4235  // We can't replace a widening/narrowing shuffle with one of its operands.
4236  if (!RootVec || RootVec->getType() != RetTy)
4237  return nullptr;
4238  }
4239  return RootVec;
4240 }
4241 
4242 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4244  Type *RetTy, const SimplifyQuery &Q) {
4245  return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4246 }
4247 
4249  // If the input is a vector with undef elements, just return a default NaN.
4250  if (!In->isNaN())
4251  return ConstantFP::getNaN(In->getType());
4252 
4253  // Propagate the existing NaN constant when possible.
4254  // TODO: Should we quiet a signaling NaN?
4255  return In;
4256 }
4257 
4258 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4259  if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4260  return ConstantFP::getNaN(Op0->getType());
4261 
4262  if (match(Op0, m_NaN()))
4263  return propagateNaN(cast<Constant>(Op0));
4264  if (match(Op1, m_NaN()))
4265  return propagateNaN(cast<Constant>(Op1));
4266 
4267  return nullptr;
4268 }
4269 
4270 /// Given operands for an FAdd, see if we can fold the result. If not, this
4271 /// returns null.
4273  const SimplifyQuery &Q, unsigned MaxRecurse) {
4274  if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4275  return C;
4276 
4277  if (Constant *C = simplifyFPBinop(Op0, Op1))
4278  return C;
4279 
4280  // fadd X, -0 ==> X
4281  if (match(Op1, m_NegZeroFP()))
4282  return Op0;
4283 
4284  // fadd X, 0 ==> X, when we know X is not -0
4285  if (match(Op1, m_PosZeroFP()) &&
4286  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4287  return Op0;
4288 
4289  // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4290  // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4291  // Negative zeros are allowed because we always end up with positive zero:
4292  // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4293  // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4294  // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4295  // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4296  if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4297  match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4298  return ConstantFP::getNullValue(Op0->getType());
4299 
4300  // (X - Y) + Y --> X
4301  // Y + (X - Y) --> X
4302  Value *X;
4303  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4304  (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4305  match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4306  return X;
4307 
4308  return nullptr;
4309 }
4310 
4311 /// Given operands for an FSub, see if we can fold the result. If not, this
4312 /// returns null.
4314  const SimplifyQuery &Q, unsigned MaxRecurse) {
4315  if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4316  return C;
4317 
4318  if (Constant *C = simplifyFPBinop(Op0, Op1))
4319  return C;
4320 
4321  // fsub X, +0 ==> X
4322  if (match(Op1, m_PosZeroFP()))
4323  return Op0;
4324 
4325  // fsub X, -0 ==> X, when we know X is not -0
4326  if (match(Op1, m_NegZeroFP()) &&
4327  (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4328  return Op0;
4329 
4330  // fsub -0.0, (fsub -0.0, X) ==> X
4331  Value *X;
4332  if (match(Op0, m_NegZeroFP()) &&
4333  match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4334  return X;
4335 
4336  // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4337  if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4338  match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4339  return X;
4340 
4341  // fsub nnan x, x ==> 0.0
4342  if (FMF.noNaNs() && Op0 == Op1)
4343  return Constant::getNullValue(Op0->getType());
4344 
4345  // Y - (Y - X) --> X
4346  // (X + Y) - Y --> X
4347  if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4348  (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4349  match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4350  return X;
4351 
4352  return nullptr;
4353 }
4354 
4355 /// Given the operands for an FMul, see if we can fold the result
4357  const SimplifyQuery &Q, unsigned MaxRecurse) {
4358  if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4359  return C;
4360 
4361  if (Constant *C = simplifyFPBinop(Op0, Op1))
4362  return C;
4363 
4364  // fmul X, 1.0 ==> X
4365  if (match(Op1, m_FPOne()))
4366  return Op0;
4367 
4368  // fmul nnan nsz X, 0 ==> 0
4369  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4370  return ConstantFP::getNullValue(Op0->getType());
4371 
4372  // sqrt(X) * sqrt(X) --> X, if we can:
4373  // 1. Remove the intermediate rounding (reassociate).
4374  // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4375  // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4376  Value *X;
4377  if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4378  FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4379  return X;
4380 
4381  return nullptr;
4382 }
4383 
4385  const SimplifyQuery &Q) {
4386  return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4387 }
4388 
4389 
4391  const SimplifyQuery &Q) {
4392  return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4393 }
4394 
4396  const SimplifyQuery &Q) {
4397  return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4398 }
4399 
4401  const SimplifyQuery &Q, unsigned) {
4402  if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4403  return C;
4404 
4405  if (Constant *C = simplifyFPBinop(Op0, Op1))
4406  return C;
4407 
4408  // X / 1.0 -> X
4409  if (match(Op1, m_FPOne()))
4410  return Op0;
4411 
4412  // 0 / X -> 0
4413  // Requires that NaNs are off (X could be zero) and signed zeroes are
4414  // ignored (X could be positive or negative, so the output sign is unknown).
4415  if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4416  return ConstantFP::getNullValue(Op0->getType());
4417 
4418  if (FMF.noNaNs()) {
4419  // X / X -> 1.0 is legal when NaNs are ignored.
4420  // We can ignore infinities because INF/INF is NaN.
4421  if (Op0 == Op1)
4422  return ConstantFP::get(Op0->getType(), 1.0);
4423 
4424  // (X * Y) / Y --> X if we can reassociate to the above form.
4425  Value *X;
4426  if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4427  return X;
4428 
4429  // -X / X -> -1.0 and
4430  // X / -X -> -1.0 are legal when NaNs are ignored.
4431  // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4432  if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4433  match(Op1, m_FNegNSZ(m_Specific(Op0))))
4434  return ConstantFP::get(Op0->getType(), -1.0);
4435  }
4436 
4437  return nullptr;
4438 }
4439 
4441  const SimplifyQuery &Q) {
4442  return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4443 }
4444 
4446  const SimplifyQuery &Q, unsigned) {
4447  if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4448  return C;
4449 
4450  if (Constant *C = simplifyFPBinop(Op0, Op1))
4451  return C;
4452 
4453  // Unlike fdiv, the result of frem always matches the sign of the dividend.
4454  // The constant match may include undef elements in a vector, so return a full
4455  // zero constant as the result.
4456  if (FMF.noNaNs()) {
4457  // +0 % X -> 0
4458  if (match(Op0, m_PosZeroFP()))
4459  return ConstantFP::getNullValue(Op0->getType());
4460  // -0 % X -> -0
4461  if (match(Op0, m_NegZeroFP()))
4462  return ConstantFP::getNegativeZero(Op0->getType());
4463  }
4464 
4465  return nullptr;
4466 }
4467 
4469  const SimplifyQuery &Q) {
4470  return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4471 }
4472 
4473 //=== Helper functions for higher up the class hierarchy.
4474 
4475 /// Given operands for a BinaryOperator, see if we can fold the result.
4476 /// If not, this returns null.
4477 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4478  const SimplifyQuery &Q, unsigned MaxRecurse) {
4479  switch (Opcode) {
4480  case Instruction::Add:
4481  return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4482  case Instruction::Sub:
4483  return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4484  case Instruction::Mul:
4485  return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4486  case Instruction::SDiv:
4487  return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4488  case Instruction::UDiv:
4489  return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4490  case Instruction::SRem:
4491  return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4492  case Instruction::URem:
4493  return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4494  case Instruction::Shl:
4495  return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4496  case Instruction::LShr:
4497  return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4498  case Instruction::AShr:
4499  return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4500  case Instruction::And:
4501  return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4502  case Instruction::Or:
4503  return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4504  case Instruction::Xor:
4505  return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4506  case Instruction::FAdd:
4507  return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4508  case Instruction::FSub:
4509  return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4510  case Instruction::FMul:
4511  return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4512  case Instruction::FDiv:
4513  return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4514  case Instruction::FRem:
4515  return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4516  default:
4517  llvm_unreachable("Unexpected opcode");
4518  }
4519 }
4520 
4521 /// Given operands for a BinaryOperator, see if we can fold the result.
4522 /// If not, this returns null.
4523 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4524 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4525 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4526  const FastMathFlags &FMF, const SimplifyQuery &Q,
4527  unsigned MaxRecurse) {
4528  switch (Opcode) {
4529  case Instruction::FAdd:
4530  return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4531  case Instruction::FSub:
4532  return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4533  case Instruction::FMul:
4534  return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4535  case Instruction::FDiv:
4536  return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4537  default:
4538  return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4539  }
4540 }
4541 
4542 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4543  const SimplifyQuery &Q) {
4544  return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4545 }
4546 
4547 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4548  FastMathFlags FMF, const SimplifyQuery &Q) {
4549  return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4550 }
4551 
4552 /// Given operands for a CmpInst, see if we can fold the result.
4553 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4554  const SimplifyQuery &Q, unsigned MaxRecurse) {
4556  return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4557  return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4558 }
4559 
4560 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4561  const SimplifyQuery &Q) {
4562  return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4563 }
4564 
4566  switch (ID) {
4567  default: return false;
4568 
4569  // Unary idempotent: f(f(x)) = f(x)
4570  case Intrinsic::fabs:
4571  case Intrinsic::floor:
4572  case Intrinsic::ceil:
4573  case Intrinsic::trunc:
4574  case Intrinsic::rint:
4575  case Intrinsic::nearbyint:
4576  case Intrinsic::round:
4577  case Intrinsic::canonicalize:
4578  return true;
4579  }
4580 }
4581 
4583  const DataLayout &DL) {
4584  GlobalValue *PtrSym;
4585  APInt PtrOffset;
4586  if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4587  return nullptr;
4588 
4589  Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4591  Type *Int32PtrTy = Int32Ty->getPointerTo();
4592  Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4593 
4594  auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4595  if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4596  return nullptr;
4597 
4598  uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4599  if (OffsetInt % 4 != 0)
4600  return nullptr;
4601 
4603  Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4604  ConstantInt::get(Int64Ty, OffsetInt / 4));
4605  Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4606  if (!Loaded)
4607  return nullptr;
4608 
4609  auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4610  if (!LoadedCE)
4611  return nullptr;
4612 
4613  if (LoadedCE->getOpcode() == Instruction::Trunc) {
4614  LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4615  if (!LoadedCE)
4616  return nullptr;
4617  }
4618 
4619  if (LoadedCE->getOpcode() != Instruction::Sub)
4620  return nullptr;
4621 
4622  auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4623  if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4624  return nullptr;
4625  auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4626 
4627  Constant *LoadedRHS = LoadedCE->getOperand(1);
4628  GlobalValue *LoadedRHSSym;
4629  APInt LoadedRHSOffset;
4630  if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4631  DL) ||
4632  PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4633  return nullptr;
4634 
4635  return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4636 }
4637 
4639  auto *ConstMask = dyn_cast<Constant>(Mask);
4640  if (!ConstMask)
4641  return false;
4642  if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4643  return true;
4644  for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4645  ++I) {
4646  if (auto *MaskElt = ConstMask->getAggregateElement(I))
4647  if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4648  continue;
4649  return false;
4650  }
4651  return true;
4652 }
4653 
4655  const SimplifyQuery &Q) {
4656  // Idempotent functions return the same result when called repeatedly.
4657  Intrinsic::ID IID = F->getIntrinsicID();
4658  if (IsIdempotent(IID))
4659  if (auto *II = dyn_cast<IntrinsicInst>(Op0))
4660  if (II->getIntrinsicID() == IID)
4661  return II;
4662 
4663  Value *X;
4664  switch (IID) {
4665  case Intrinsic::fabs:
4666  if (SignBitMustBeZero(Op0, Q.TLI)) return Op0;
4667  break;
4668  case Intrinsic::bswap:
4669  // bswap(bswap(x)) -> x
4670  if (match(Op0, m_BSwap(m_Value(X)))) return X;
4671  break;
4672  case Intrinsic::bitreverse:
4673  // bitreverse(bitreverse(x)) -> x
4674  if (match(Op0, m_BitReverse(m_Value(X)))) return X;
4675  break;
4676  case Intrinsic::exp:
4677  // exp(log(x)) -> x
4678  if (Q.CxtI->hasAllowReassoc() &&
4679  match(Op0, m_Intrinsic<Intrinsic::log>(m_Value(X)))) return X;
4680  break;
4681  case Intrinsic::exp2:
4682  // exp2(log2(x)) -> x
4683  if (Q.CxtI->hasAllowReassoc() &&
4684  match(Op0, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) return X;
4685  break;
4686  case Intrinsic::log:
4687  // log(exp(x)) -> x
4688  if (Q.CxtI->hasAllowReassoc() &&
4689  match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) return X;
4690  break;
4691  case Intrinsic::log2:
4692  // log2(exp2(x)) -> x
4693  if (Q.CxtI->hasAllowReassoc() &&
4694  (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) ||
4695  match(Op0, m_Intrinsic<Intrinsic::pow>(m_SpecificFP(2.0),
4696  m_Value(X))))) return X;
4697  break;
4698  case Intrinsic::log10:
4699  // log10(pow(10.0, x)) -> x
4700  if (Q.CxtI->hasAllowReassoc() &&
4701  match(Op0, m_Intrinsic<Intrinsic::pow>(m_SpecificFP(10.0),
4702  m_Value(X)))) return X;
4703  break;
4704  case Intrinsic::floor:
4705  case Intrinsic::trunc:
4706  case Intrinsic::ceil:
4707  case Intrinsic::round:
4708  case Intrinsic::nearbyint:
4709  case Intrinsic::rint: {
4710  // floor (sitofp x) -> sitofp x
4711  // floor (uitofp x) -> uitofp x
4712  //
4713  // Converting from int always results in a finite integral number or
4714  // infinity. For either of those inputs, these rounding functions always
4715  // return the same value, so the rounding can be eliminated.
4716  if (match(Op0, m_SIToFP(m_Value())) || match(Op0, m_UIToFP(m_Value())))
4717  return Op0;
4718  break;
4719  }
4720  default:
4721  break;
4722  }
4723 
4724  return nullptr;
4725 }
4726 
4728  const SimplifyQuery &Q) {
4729  Intrinsic::ID IID = F->getIntrinsicID();
4730  Type *ReturnType = F->getReturnType();
4731  switch (IID) {
4732  case Intrinsic::usub_with_overflow:
4733  case Intrinsic::ssub_with_overflow:
4734  // X - X -> { 0, false }
4735  if (Op0 == Op1)
4736  return Constant::getNullValue(ReturnType);
4737  // X - undef -> undef
4738  // undef - X -> undef
4739  if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4740  return UndefValue::get(ReturnType);
4741  break;
4742  case Intrinsic::uadd_with_overflow:
4743  case Intrinsic::sadd_with_overflow:
4744  // X + undef -> undef
4745  if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4746  return UndefValue::get(ReturnType);
4747  break;
4748  case Intrinsic::umul_with_overflow:
4749  case Intrinsic::smul_with_overflow:
4750  // 0 * X -> { 0, false }
4751  // X * 0 -> { 0, false }
4752  if (match(Op0, m_Zero()) || match(Op1, m_Zero()))
4753  return Constant::getNullValue(ReturnType);
4754  // undef * X -> { 0, false }
4755  // X * undef -> { 0, false }
4756  if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
4757  return Constant::getNullValue(ReturnType);
4758  break;
4759  case Intrinsic::uadd_sat:
4760  // sat(MAX + X) -> MAX
4761  // sat(X + MAX) -> MAX
4762  if (match(Op0, m_AllOnes()) || match(Op1, m_AllOnes()))
4763  return Constant::getAllOnesValue(ReturnType);
4765  case Intrinsic::sadd_sat:
4766  // sat(X + undef) -> -1
4767  // sat(undef + X) -> -1
4768  // For unsigned: Assume undef is MAX, thus we saturate to MAX (-1).
4769  // For signed: Assume undef is ~X, in which case X + ~X = -1.
4770  if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
4771  return Constant::getAllOnesValue(ReturnType);
4772 
4773  // X + 0 -> X
4774  if (match(Op1, m_Zero()))
4775  return Op0;
4776  // 0 + X -> X
4777  if (match(Op0, m_Zero()))
4778  return Op1;
4779  break;
4780  case Intrinsic::usub_sat:
4781  // sat(0 - X) -> 0, sat(X - MAX) -> 0
4782  if (match(Op0, m_Zero()) || match(Op1, m_AllOnes()))
4783  return Constant::getNullValue(ReturnType);
4785  case Intrinsic::ssub_sat:
4786  // X - X -> 0, X - undef -> 0, undef - X -> 0
4787  if (Op0 == Op1 || match(Op0, m_Undef()) || match(Op1, m_Undef()))
4788  return Constant::getNullValue(ReturnType);
4789  // X - 0 -> X
4790  if (match(Op1, m_Zero()))
4791  return Op0;
4792  break;
4793  case Intrinsic::load_relative:
4794  if (auto *C0 = dyn_cast<Constant>(Op0))
4795  if (auto *C1 = dyn_cast<Constant>(Op1))
4796  return SimplifyRelativeLoad(C0, C1, Q.DL);
4797  break;
4798  case Intrinsic::powi:
4799  if (auto *Power = dyn_cast<ConstantInt>(Op1)) {
4800  // powi(x, 0) -> 1.0
4801  if (Power->isZero())
4802  return ConstantFP::get(Op0->getType(), 1.0);
4803  // powi(x, 1) -> x
4