LLVM  13.0.0git
InstCombineCompares.cpp
Go to the documentation of this file.
1 //===- InstCombineCompares.cpp --------------------------------------------===//
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 the visitICmp and visitFCmp functions.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APSInt.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/Statistic.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/KnownBits.h"
28 
29 using namespace llvm;
30 using namespace PatternMatch;
31 
32 #define DEBUG_TYPE "instcombine"
33 
34 // How many times is a select replaced by one of its operands?
35 STATISTIC(NumSel, "Number of select opts");
36 
37 
38 /// Compute Result = In1+In2, returning true if the result overflowed for this
39 /// type.
40 static bool addWithOverflow(APInt &Result, const APInt &In1,
41  const APInt &In2, bool IsSigned = false) {
42  bool Overflow;
43  if (IsSigned)
44  Result = In1.sadd_ov(In2, Overflow);
45  else
46  Result = In1.uadd_ov(In2, Overflow);
47 
48  return Overflow;
49 }
50 
51 /// Compute Result = In1-In2, returning true if the result overflowed for this
52 /// type.
53 static bool subWithOverflow(APInt &Result, const APInt &In1,
54  const APInt &In2, bool IsSigned = false) {
55  bool Overflow;
56  if (IsSigned)
57  Result = In1.ssub_ov(In2, Overflow);
58  else
59  Result = In1.usub_ov(In2, Overflow);
60 
61  return Overflow;
62 }
63 
64 /// Given an icmp instruction, return true if any use of this comparison is a
65 /// branch on sign bit comparison.
66 static bool hasBranchUse(ICmpInst &I) {
67  for (auto *U : I.users())
68  if (isa<BranchInst>(U))
69  return true;
70  return false;
71 }
72 
73 /// Returns true if the exploded icmp can be expressed as a signed comparison
74 /// to zero and updates the predicate accordingly.
75 /// The signedness of the comparison is preserved.
76 /// TODO: Refactor with decomposeBitTestICmp()?
77 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
78  if (!ICmpInst::isSigned(Pred))
79  return false;
80 
81  if (C.isNullValue())
82  return ICmpInst::isRelational(Pred);
83 
84  if (C.isOneValue()) {
85  if (Pred == ICmpInst::ICMP_SLT) {
86  Pred = ICmpInst::ICMP_SLE;
87  return true;
88  }
89  } else if (C.isAllOnesValue()) {
90  if (Pred == ICmpInst::ICMP_SGT) {
91  Pred = ICmpInst::ICMP_SGE;
92  return true;
93  }
94  }
95 
96  return false;
97 }
98 
99 /// This is called when we see this pattern:
100 /// cmp pred (load (gep GV, ...)), cmpcst
101 /// where GV is a global variable with a constant initializer. Try to simplify
102 /// this into some simple computation that does not need the load. For example
103 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
104 ///
105 /// If AndCst is non-null, then the loaded value is masked with that constant
106 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
107 Instruction *
109  GlobalVariable *GV, CmpInst &ICI,
110  ConstantInt *AndCst) {
111  Constant *Init = GV->getInitializer();
112  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
113  return nullptr;
114 
115  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
116  // Don't blow up on huge arrays.
117  if (ArrayElementCount > MaxArraySizeForCombine)
118  return nullptr;
119 
120  // There are many forms of this optimization we can handle, for now, just do
121  // the simple index into a single-dimensional array.
122  //
123  // Require: GEP GV, 0, i {{, constant indices}}
124  if (GEP->getNumOperands() < 3 ||
125  !isa<ConstantInt>(GEP->getOperand(1)) ||
126  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
127  isa<Constant>(GEP->getOperand(2)))
128  return nullptr;
129 
130  // Check that indices after the variable are constants and in-range for the
131  // type they index. Collect the indices. This is typically for arrays of
132  // structs.
133  SmallVector<unsigned, 4> LaterIndices;
134 
135  Type *EltTy = Init->getType()->getArrayElementType();
136  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
137  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
138  if (!Idx) return nullptr; // Variable index.
139 
140  uint64_t IdxVal = Idx->getZExtValue();
141  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
142 
143  if (StructType *STy = dyn_cast<StructType>(EltTy))
144  EltTy = STy->getElementType(IdxVal);
145  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
146  if (IdxVal >= ATy->getNumElements()) return nullptr;
147  EltTy = ATy->getElementType();
148  } else {
149  return nullptr; // Unknown type.
150  }
151 
152  LaterIndices.push_back(IdxVal);
153  }
154 
155  enum { Overdefined = -3, Undefined = -2 };
156 
157  // Variables for our state machines.
158 
159  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
160  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
161  // and 87 is the second (and last) index. FirstTrueElement is -2 when
162  // undefined, otherwise set to the first true element. SecondTrueElement is
163  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
164  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
165 
166  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
167  // form "i != 47 & i != 87". Same state transitions as for true elements.
168  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
169 
170  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
171  /// define a state machine that triggers for ranges of values that the index
172  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
173  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
174  /// index in the range (inclusive). We use -2 for undefined here because we
175  /// use relative comparisons and don't want 0-1 to match -1.
176  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
177 
178  // MagicBitvector - This is a magic bitvector where we set a bit if the
179  // comparison is true for element 'i'. If there are 64 elements or less in
180  // the array, this will fully represent all the comparison results.
181  uint64_t MagicBitvector = 0;
182 
183  // Scan the array and see if one of our patterns matches.
184  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
185  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
186  Constant *Elt = Init->getAggregateElement(i);
187  if (!Elt) return nullptr;
188 
189  // If this is indexing an array of structures, get the structure element.
190  if (!LaterIndices.empty())
191  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
192 
193  // If the element is masked, handle it.
194  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
195 
196  // Find out if the comparison would be true or false for the i'th element.
198  CompareRHS, DL, &TLI);
199  // If the result is undef for this element, ignore it.
200  if (isa<UndefValue>(C)) {
201  // Extend range state machines to cover this element in case there is an
202  // undef in the middle of the range.
203  if (TrueRangeEnd == (int)i-1)
204  TrueRangeEnd = i;
205  if (FalseRangeEnd == (int)i-1)
206  FalseRangeEnd = i;
207  continue;
208  }
209 
210  // If we can't compute the result for any of the elements, we have to give
211  // up evaluating the entire conditional.
212  if (!isa<ConstantInt>(C)) return nullptr;
213 
214  // Otherwise, we know if the comparison is true or false for this element,
215  // update our state machines.
216  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
217 
218  // State machine for single/double/range index comparison.
219  if (IsTrueForElt) {
220  // Update the TrueElement state machine.
221  if (FirstTrueElement == Undefined)
222  FirstTrueElement = TrueRangeEnd = i; // First true element.
223  else {
224  // Update double-compare state machine.
225  if (SecondTrueElement == Undefined)
226  SecondTrueElement = i;
227  else
228  SecondTrueElement = Overdefined;
229 
230  // Update range state machine.
231  if (TrueRangeEnd == (int)i-1)
232  TrueRangeEnd = i;
233  else
234  TrueRangeEnd = Overdefined;
235  }
236  } else {
237  // Update the FalseElement state machine.
238  if (FirstFalseElement == Undefined)
239  FirstFalseElement = FalseRangeEnd = i; // First false element.
240  else {
241  // Update double-compare state machine.
242  if (SecondFalseElement == Undefined)
243  SecondFalseElement = i;
244  else
245  SecondFalseElement = Overdefined;
246 
247  // Update range state machine.
248  if (FalseRangeEnd == (int)i-1)
249  FalseRangeEnd = i;
250  else
251  FalseRangeEnd = Overdefined;
252  }
253  }
254 
255  // If this element is in range, update our magic bitvector.
256  if (i < 64 && IsTrueForElt)
257  MagicBitvector |= 1ULL << i;
258 
259  // If all of our states become overdefined, bail out early. Since the
260  // predicate is expensive, only check it every 8 elements. This is only
261  // really useful for really huge arrays.
262  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
263  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
264  FalseRangeEnd == Overdefined)
265  return nullptr;
266  }
267 
268  // Now that we've scanned the entire array, emit our new comparison(s). We
269  // order the state machines in complexity of the generated code.
270  Value *Idx = GEP->getOperand(2);
271 
272  // If the index is larger than the pointer size of the target, truncate the
273  // index down like the GEP would do implicitly. We don't have to do this for
274  // an inbounds GEP because the index can't be out of range.
275  if (!GEP->isInBounds()) {
276  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
277  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
278  if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
279  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
280  }
281 
282  // If the comparison is only true for one or two elements, emit direct
283  // comparisons.
284  if (SecondTrueElement != Overdefined) {
285  // None true -> false.
286  if (FirstTrueElement == Undefined)
287  return replaceInstUsesWith(ICI, Builder.getFalse());
288 
289  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
290 
291  // True for one element -> 'i == 47'.
292  if (SecondTrueElement == Undefined)
293  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
294 
295  // True for two elements -> 'i == 47 | i == 72'.
296  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
297  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
298  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
299  return BinaryOperator::CreateOr(C1, C2);
300  }
301 
302  // If the comparison is only false for one or two elements, emit direct
303  // comparisons.
304  if (SecondFalseElement != Overdefined) {
305  // None false -> true.
306  if (FirstFalseElement == Undefined)
307  return replaceInstUsesWith(ICI, Builder.getTrue());
308 
309  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
310 
311  // False for one element -> 'i != 47'.
312  if (SecondFalseElement == Undefined)
313  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
314 
315  // False for two elements -> 'i != 47 & i != 72'.
316  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
317  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
318  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
319  return BinaryOperator::CreateAnd(C1, C2);
320  }
321 
322  // If the comparison can be replaced with a range comparison for the elements
323  // where it is true, emit the range check.
324  if (TrueRangeEnd != Overdefined) {
325  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
326 
327  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
328  if (FirstTrueElement) {
329  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
330  Idx = Builder.CreateAdd(Idx, Offs);
331  }
332 
333  Value *End = ConstantInt::get(Idx->getType(),
334  TrueRangeEnd-FirstTrueElement+1);
335  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
336  }
337 
338  // False range check.
339  if (FalseRangeEnd != Overdefined) {
340  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
341  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
342  if (FirstFalseElement) {
343  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
344  Idx = Builder.CreateAdd(Idx, Offs);
345  }
346 
347  Value *End = ConstantInt::get(Idx->getType(),
348  FalseRangeEnd-FirstFalseElement);
349  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
350  }
351 
352  // If a magic bitvector captures the entire comparison state
353  // of this load, replace it with computation that does:
354  // ((magic_cst >> i) & 1) != 0
355  {
356  Type *Ty = nullptr;
357 
358  // Look for an appropriate type:
359  // - The type of Idx if the magic fits
360  // - The smallest fitting legal type
361  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
362  Ty = Idx->getType();
363  else
364  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
365 
366  if (Ty) {
367  Value *V = Builder.CreateIntCast(Idx, Ty, false);
368  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
369  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
370  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
371  }
372  }
373 
374  return nullptr;
375 }
376 
377 /// Return a value that can be used to compare the *offset* implied by a GEP to
378 /// zero. For example, if we have &A[i], we want to return 'i' for
379 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
380 /// are involved. The above expression would also be legal to codegen as
381 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
382 /// This latter form is less amenable to optimization though, and we are allowed
383 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
384 ///
385 /// If we can't emit an optimized form for this expression, this returns null.
386 ///
388  const DataLayout &DL) {
390 
391  // Check to see if this gep only has a single variable index. If so, and if
392  // any constant indices are a multiple of its scale, then we can compute this
393  // in terms of the scale of the variable index. For example, if the GEP
394  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
395  // because the expression will cross zero at the same point.
396  unsigned i, e = GEP->getNumOperands();
397  int64_t Offset = 0;
398  for (i = 1; i != e; ++i, ++GTI) {
399  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
400  // Compute the aggregate offset of constant indices.
401  if (CI->isZero()) continue;
402 
403  // Handle a struct index, which adds its field offset to the pointer.
404  if (StructType *STy = GTI.getStructTypeOrNull()) {
405  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
406  } else {
407  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
408  Offset += Size*CI->getSExtValue();
409  }
410  } else {
411  // Found our variable index.
412  break;
413  }
414  }
415 
416  // If there are no variable indices, we must have a constant offset, just
417  // evaluate it the general way.
418  if (i == e) return nullptr;
419 
420  Value *VariableIdx = GEP->getOperand(i);
421  // Determine the scale factor of the variable element. For example, this is
422  // 4 if the variable index is into an array of i32.
423  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
424 
425  // Verify that there are no other variable indices. If so, emit the hard way.
426  for (++i, ++GTI; i != e; ++i, ++GTI) {
427  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
428  if (!CI) return nullptr;
429 
430  // Compute the aggregate offset of constant indices.
431  if (CI->isZero()) continue;
432 
433  // Handle a struct index, which adds its field offset to the pointer.
434  if (StructType *STy = GTI.getStructTypeOrNull()) {
435  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
436  } else {
437  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
438  Offset += Size*CI->getSExtValue();
439  }
440  }
441 
442  // Okay, we know we have a single variable index, which must be a
443  // pointer/array/vector index. If there is no offset, life is simple, return
444  // the index.
445  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
446  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
447  if (Offset == 0) {
448  // Cast to intptrty in case a truncation occurs. If an extension is needed,
449  // we don't need to bother extending: the extension won't affect where the
450  // computation crosses zero.
451  if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
452  IntPtrWidth) {
453  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
454  }
455  return VariableIdx;
456  }
457 
458  // Otherwise, there is an index. The computation we will do will be modulo
459  // the pointer size.
460  Offset = SignExtend64(Offset, IntPtrWidth);
461  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
462 
463  // To do this transformation, any constant index must be a multiple of the
464  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
465  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
466  // multiple of the variable scale.
467  int64_t NewOffs = Offset / (int64_t)VariableScale;
468  if (Offset != NewOffs*(int64_t)VariableScale)
469  return nullptr;
470 
471  // Okay, we can do this evaluation. Start by converting the index to intptr.
472  if (VariableIdx->getType() != IntPtrTy)
473  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
474  true /*Signed*/);
475  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
476  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
477 }
478 
479 /// Returns true if we can rewrite Start as a GEP with pointer Base
480 /// and some integer offset. The nodes that need to be re-written
481 /// for this transformation will be added to Explored.
482 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
483  const DataLayout &DL,
484  SetVector<Value *> &Explored) {
485  SmallVector<Value *, 16> WorkList(1, Start);
486  Explored.insert(Base);
487 
488  // The following traversal gives us an order which can be used
489  // when doing the final transformation. Since in the final
490  // transformation we create the PHI replacement instructions first,
491  // we don't have to get them in any particular order.
492  //
493  // However, for other instructions we will have to traverse the
494  // operands of an instruction first, which means that we have to
495  // do a post-order traversal.
496  while (!WorkList.empty()) {
498 
499  while (!WorkList.empty()) {
500  if (Explored.size() >= 100)
501  return false;
502 
503  Value *V = WorkList.back();
504 
505  if (Explored.contains(V)) {
506  WorkList.pop_back();
507  continue;
508  }
509 
510  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
511  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
512  // We've found some value that we can't explore which is different from
513  // the base. Therefore we can't do this transformation.
514  return false;
515 
516  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
517  auto *CI = cast<CastInst>(V);
518  if (!CI->isNoopCast(DL))
519  return false;
520 
521  if (Explored.count(CI->getOperand(0)) == 0)
522  WorkList.push_back(CI->getOperand(0));
523  }
524 
525  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
526  // We're limiting the GEP to having one index. This will preserve
527  // the original pointer type. We could handle more cases in the
528  // future.
529  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
530  GEP->getType() != Start->getType())
531  return false;
532 
533  if (Explored.count(GEP->getOperand(0)) == 0)
534  WorkList.push_back(GEP->getOperand(0));
535  }
536 
537  if (WorkList.back() == V) {
538  WorkList.pop_back();
539  // We've finished visiting this node, mark it as such.
540  Explored.insert(V);
541  }
542 
543  if (auto *PN = dyn_cast<PHINode>(V)) {
544  // We cannot transform PHIs on unsplittable basic blocks.
545  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
546  return false;
547  Explored.insert(PN);
548  PHIs.insert(PN);
549  }
550  }
551 
552  // Explore the PHI nodes further.
553  for (auto *PN : PHIs)
554  for (Value *Op : PN->incoming_values())
555  if (Explored.count(Op) == 0)
556  WorkList.push_back(Op);
557  }
558 
559  // Make sure that we can do this. Since we can't insert GEPs in a basic
560  // block before a PHI node, we can't easily do this transformation if
561  // we have PHI node users of transformed instructions.
562  for (Value *Val : Explored) {
563  for (Value *Use : Val->uses()) {
564 
565  auto *PHI = dyn_cast<PHINode>(Use);
566  auto *Inst = dyn_cast<Instruction>(Val);
567 
568  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
569  Explored.count(PHI) == 0)
570  continue;
571 
572  if (PHI->getParent() == Inst->getParent())
573  return false;
574  }
575  }
576  return true;
577 }
578 
579 // Sets the appropriate insert point on Builder where we can add
580 // a replacement Instruction for V (if that is possible).
582  bool Before = true) {
583  if (auto *PHI = dyn_cast<PHINode>(V)) {
584  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
585  return;
586  }
587  if (auto *I = dyn_cast<Instruction>(V)) {
588  if (!Before)
589  I = &*std::next(I->getIterator());
590  Builder.SetInsertPoint(I);
591  return;
592  }
593  if (auto *A = dyn_cast<Argument>(V)) {
594  // Set the insertion point in the entry block.
595  BasicBlock &Entry = A->getParent()->getEntryBlock();
596  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
597  return;
598  }
599  // Otherwise, this is a constant and we don't need to set a new
600  // insertion point.
601  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
602 }
603 
604 /// Returns a re-written value of Start as an indexed GEP using Base as a
605 /// pointer.
606 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
607  const DataLayout &DL,
608  SetVector<Value *> &Explored) {
609  // Perform all the substitutions. This is a bit tricky because we can
610  // have cycles in our use-def chains.
611  // 1. Create the PHI nodes without any incoming values.
612  // 2. Create all the other values.
613  // 3. Add the edges for the PHI nodes.
614  // 4. Emit GEPs to get the original pointers.
615  // 5. Remove the original instructions.
616  Type *IndexType = IntegerType::get(
617  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
618 
620  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
621 
622  // Create the new PHI nodes, without adding any incoming values.
623  for (Value *Val : Explored) {
624  if (Val == Base)
625  continue;
626  // Create empty phi nodes. This avoids cyclic dependencies when creating
627  // the remaining instructions.
628  if (auto *PHI = dyn_cast<PHINode>(Val))
629  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
630  PHI->getName() + ".idx", PHI);
631  }
632  IRBuilder<> Builder(Base->getContext());
633 
634  // Create all the other instructions.
635  for (Value *Val : Explored) {
636 
637  if (NewInsts.find(Val) != NewInsts.end())
638  continue;
639 
640  if (auto *CI = dyn_cast<CastInst>(Val)) {
641  // Don't get rid of the intermediate variable here; the store can grow
642  // the map which will invalidate the reference to the input value.
643  Value *V = NewInsts[CI->getOperand(0)];
644  NewInsts[CI] = V;
645  continue;
646  }
647  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
648  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
649  : GEP->getOperand(1);
651  // Indices might need to be sign extended. GEPs will magically do
652  // this, but we need to do it ourselves here.
653  if (Index->getType()->getScalarSizeInBits() !=
654  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
655  Index = Builder.CreateSExtOrTrunc(
656  Index, NewInsts[GEP->getOperand(0)]->getType(),
657  GEP->getOperand(0)->getName() + ".sext");
658  }
659 
660  auto *Op = NewInsts[GEP->getOperand(0)];
661  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
662  NewInsts[GEP] = Index;
663  else
664  NewInsts[GEP] = Builder.CreateNSWAdd(
665  Op, Index, GEP->getOperand(0)->getName() + ".add");
666  continue;
667  }
668  if (isa<PHINode>(Val))
669  continue;
670 
671  llvm_unreachable("Unexpected instruction type");
672  }
673 
674  // Add the incoming values to the PHI nodes.
675  for (Value *Val : Explored) {
676  if (Val == Base)
677  continue;
678  // All the instructions have been created, we can now add edges to the
679  // phi nodes.
680  if (auto *PHI = dyn_cast<PHINode>(Val)) {
681  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
682  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
683  Value *NewIncoming = PHI->getIncomingValue(I);
684 
685  if (NewInsts.find(NewIncoming) != NewInsts.end())
686  NewIncoming = NewInsts[NewIncoming];
687 
688  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
689  }
690  }
691  }
692 
693  for (Value *Val : Explored) {
694  if (Val == Base)
695  continue;
696 
697  // Depending on the type, for external users we have to emit
698  // a GEP or a GEP + ptrtoint.
699  setInsertionPoint(Builder, Val, false);
700 
701  // If required, create an inttoptr instruction for Base.
702  Value *NewBase = Base;
703  if (!Base->getType()->isPointerTy())
704  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
705  Start->getName() + "to.ptr");
706 
707  Value *GEP = Builder.CreateInBoundsGEP(
708  Start->getType()->getPointerElementType(), NewBase,
709  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
710 
711  if (!Val->getType()->isPointerTy()) {
712  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
713  Val->getName() + ".conv");
714  GEP = Cast;
715  }
716  Val->replaceAllUsesWith(GEP);
717  }
718 
719  return NewInsts[Start];
720 }
721 
722 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
723 /// the input Value as a constant indexed GEP. Returns a pair containing
724 /// the GEPs Pointer and Index.
725 static std::pair<Value *, Value *>
727  Type *IndexType = IntegerType::get(V->getContext(),
728  DL.getIndexTypeSizeInBits(V->getType()));
729 
731  while (true) {
732  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
733  // We accept only inbouds GEPs here to exclude the possibility of
734  // overflow.
735  if (!GEP->isInBounds())
736  break;
737  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
738  GEP->getType() == V->getType()) {
739  V = GEP->getOperand(0);
740  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
742  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
743  continue;
744  }
745  break;
746  }
747  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
748  if (!CI->isNoopCast(DL))
749  break;
750  V = CI->getOperand(0);
751  continue;
752  }
753  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
754  if (!CI->isNoopCast(DL))
755  break;
756  V = CI->getOperand(0);
757  continue;
758  }
759  break;
760  }
761  return {V, Index};
762 }
763 
764 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
765 /// We can look through PHIs, GEPs and casts in order to determine a common base
766 /// between GEPLHS and RHS.
769  const DataLayout &DL) {
770  // FIXME: Support vector of pointers.
771  if (GEPLHS->getType()->isVectorTy())
772  return nullptr;
773 
774  if (!GEPLHS->hasAllConstantIndices())
775  return nullptr;
776 
777  // Make sure the pointers have the same type.
778  if (GEPLHS->getType() != RHS->getType())
779  return nullptr;
780 
781  Value *PtrBase, *Index;
782  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
783 
784  // The set of nodes that will take part in this transformation.
785  SetVector<Value *> Nodes;
786 
787  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
788  return nullptr;
789 
790  // We know we can re-write this as
791  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
792  // Since we've only looked through inbouds GEPs we know that we
793  // can't have overflow on either side. We can therefore re-write
794  // this as:
795  // OFFSET1 cmp OFFSET2
796  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
797 
798  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
799  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
800  // offset. Since Index is the offset of LHS to the base pointer, we will now
801  // compare the offsets instead of comparing the pointers.
802  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
803 }
804 
805 /// Fold comparisons between a GEP instruction and something else. At this point
806 /// we know that the GEP is on the LHS of the comparison.
809  Instruction &I) {
810  // Don't transform signed compares of GEPs into index compares. Even if the
811  // GEP is inbounds, the final add of the base pointer can have signed overflow
812  // and would change the result of the icmp.
813  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
814  // the maximum signed value for the pointer type.
816  return nullptr;
817 
818  // Look through bitcasts and addrspacecasts. We do not however want to remove
819  // 0 GEPs.
820  if (!isa<GetElementPtrInst>(RHS))
821  RHS = RHS->stripPointerCasts();
822 
823  Value *PtrBase = GEPLHS->getOperand(0);
824  // FIXME: Support vector pointer GEPs.
825  if (PtrBase == RHS && GEPLHS->isInBounds() &&
826  !GEPLHS->getType()->isVectorTy()) {
827  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
828  // This transformation (ignoring the base and scales) is valid because we
829  // know pointers can't overflow since the gep is inbounds. See if we can
830  // output an optimized form.
831  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
832 
833  // If not, synthesize the offset the hard way.
834  if (!Offset)
835  Offset = EmitGEPOffset(GEPLHS);
837  Constant::getNullValue(Offset->getType()));
838  }
839 
840  if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
841  isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
842  !NullPointerIsDefined(I.getFunction(),
843  RHS->getType()->getPointerAddressSpace())) {
844  // For most address spaces, an allocation can't be placed at null, but null
845  // itself is treated as a 0 size allocation in the in bounds rules. Thus,
846  // the only valid inbounds address derived from null, is null itself.
847  // Thus, we have four cases to consider:
848  // 1) Base == nullptr, Offset == 0 -> inbounds, null
849  // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
850  // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
851  // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
852  //
853  // (Note if we're indexing a type of size 0, that simply collapses into one
854  // of the buckets above.)
855  //
856  // In general, we're allowed to make values less poison (i.e. remove
857  // sources of full UB), so in this case, we just select between the two
858  // non-poison cases (1 and 4 above).
859  //
860  // For vectors, we apply the same reasoning on a per-lane basis.
861  auto *Base = GEPLHS->getPointerOperand();
862  if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
863  auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
864  Base = Builder.CreateVectorSplat(EC, Base);
865  }
866  return new ICmpInst(Cond, Base,
868  cast<Constant>(RHS), Base->getType()));
869  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
870  // If the base pointers are different, but the indices are the same, just
871  // compare the base pointer.
872  if (PtrBase != GEPRHS->getOperand(0)) {
873  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
874  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
875  GEPRHS->getOperand(0)->getType();
876  if (IndicesTheSame)
877  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
878  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
879  IndicesTheSame = false;
880  break;
881  }
882 
883  // If all indices are the same, just compare the base pointers.
884  Type *BaseType = GEPLHS->getOperand(0)->getType();
885  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
886  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
887 
888  // If we're comparing GEPs with two base pointers that only differ in type
889  // and both GEPs have only constant indices or just one use, then fold
890  // the compare with the adjusted indices.
891  // FIXME: Support vector of pointers.
892  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
893  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
894  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
895  PtrBase->stripPointerCasts() ==
896  GEPRHS->getOperand(0)->stripPointerCasts() &&
897  !GEPLHS->getType()->isVectorTy()) {
898  Value *LOffset = EmitGEPOffset(GEPLHS);
899  Value *ROffset = EmitGEPOffset(GEPRHS);
900 
901  // If we looked through an addrspacecast between different sized address
902  // spaces, the LHS and RHS pointers are different sized
903  // integers. Truncate to the smaller one.
904  Type *LHSIndexTy = LOffset->getType();
905  Type *RHSIndexTy = ROffset->getType();
906  if (LHSIndexTy != RHSIndexTy) {
907  if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
908  RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
909  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
910  } else
911  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
912  }
913 
914  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
915  LOffset, ROffset);
916  return replaceInstUsesWith(I, Cmp);
917  }
918 
919  // Otherwise, the base pointers are different and the indices are
920  // different. Try convert this to an indexed compare by looking through
921  // PHIs/casts.
922  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
923  }
924 
925  // If one of the GEPs has all zero indices, recurse.
926  // FIXME: Handle vector of pointers.
927  if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
928  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
930 
931  // If the other GEP has all zero indices, recurse.
932  // FIXME: Handle vector of pointers.
933  if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
934  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
935 
936  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
937  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
938  // If the GEPs only differ by one index, compare it.
939  unsigned NumDifferences = 0; // Keep track of # differences.
940  unsigned DiffOperand = 0; // The operand that differs.
941  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
942  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
943  Type *LHSType = GEPLHS->getOperand(i)->getType();
944  Type *RHSType = GEPRHS->getOperand(i)->getType();
945  // FIXME: Better support for vector of pointers.
946  if (LHSType->getPrimitiveSizeInBits() !=
947  RHSType->getPrimitiveSizeInBits() ||
948  (GEPLHS->getType()->isVectorTy() &&
949  (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
950  // Irreconcilable differences.
951  NumDifferences = 2;
952  break;
953  }
954 
955  if (NumDifferences++) break;
956  DiffOperand = i;
957  }
958 
959  if (NumDifferences == 0) // SAME GEP?
960  return replaceInstUsesWith(I, // No comparison is needed here.
962 
963  else if (NumDifferences == 1 && GEPsInBounds) {
964  Value *LHSV = GEPLHS->getOperand(DiffOperand);
965  Value *RHSV = GEPRHS->getOperand(DiffOperand);
966  // Make sure we do a signed comparison here.
967  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
968  }
969  }
970 
971  // Only lower this if the icmp is the only user of the GEP or if we expect
972  // the result to fold to a constant!
973  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
974  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
975  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
976  Value *L = EmitGEPOffset(GEPLHS);
977  Value *R = EmitGEPOffset(GEPRHS);
978  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
979  }
980  }
981 
982  // Try convert this to an indexed compare by looking through PHIs/casts as a
983  // last resort.
984  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
985 }
986 
988  const AllocaInst *Alloca,
989  const Value *Other) {
990  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
991 
992  // It would be tempting to fold away comparisons between allocas and any
993  // pointer not based on that alloca (e.g. an argument). However, even
994  // though such pointers cannot alias, they can still compare equal.
995  //
996  // But LLVM doesn't specify where allocas get their memory, so if the alloca
997  // doesn't escape we can argue that it's impossible to guess its value, and we
998  // can therefore act as if any such guesses are wrong.
999  //
1000  // The code below checks that the alloca doesn't escape, and that it's only
1001  // used in a comparison once (the current instruction). The
1002  // single-comparison-use condition ensures that we're trivially folding all
1003  // comparisons against the alloca consistently, and avoids the risk of
1004  // erroneously folding a comparison of the pointer with itself.
1005 
1006  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1007 
1009  for (const Use &U : Alloca->uses()) {
1010  if (Worklist.size() >= MaxIter)
1011  return nullptr;
1012  Worklist.push_back(&U);
1013  }
1014 
1015  unsigned NumCmps = 0;
1016  while (!Worklist.empty()) {
1017  assert(Worklist.size() <= MaxIter);
1018  const Use *U = Worklist.pop_back_val();
1019  const Value *V = U->getUser();
1020  --MaxIter;
1021 
1022  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1023  isa<SelectInst>(V)) {
1024  // Track the uses.
1025  } else if (isa<LoadInst>(V)) {
1026  // Loading from the pointer doesn't escape it.
1027  continue;
1028  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1029  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1030  if (SI->getValueOperand() == U->get())
1031  return nullptr;
1032  continue;
1033  } else if (isa<ICmpInst>(V)) {
1034  if (NumCmps++)
1035  return nullptr; // Found more than one cmp.
1036  continue;
1037  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1038  switch (Intrin->getIntrinsicID()) {
1039  // These intrinsics don't escape or compare the pointer. Memset is safe
1040  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1041  // we don't allow stores, so src cannot point to V.
1042  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1043  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1044  continue;
1045  default:
1046  return nullptr;
1047  }
1048  } else {
1049  return nullptr;
1050  }
1051  for (const Use &U : V->uses()) {
1052  if (Worklist.size() >= MaxIter)
1053  return nullptr;
1054  Worklist.push_back(&U);
1055  }
1056  }
1057 
1058  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1059  return replaceInstUsesWith(
1060  ICI,
1062 }
1063 
1064 /// Fold "icmp pred (X+C), X".
1066  ICmpInst::Predicate Pred) {
1067  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1068  // so the values can never be equal. Similarly for all other "or equals"
1069  // operators.
1070  assert(!!C && "C should not be zero!");
1071 
1072  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1073  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1074  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1075  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1076  Constant *R = ConstantInt::get(X->getType(),
1077  APInt::getMaxValue(C.getBitWidth()) - C);
1078  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1079  }
1080 
1081  // (X+1) >u X --> X <u (0-1) --> X != 255
1082  // (X+2) >u X --> X <u (0-2) --> X <u 254
1083  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1084  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1085  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1086  ConstantInt::get(X->getType(), -C));
1087 
1088  APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1089 
1090  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1091  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1092  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1093  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1094  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1095  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1096  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1097  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1098  ConstantInt::get(X->getType(), SMax - C));
1099 
1100  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1101  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1102  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1103  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1104  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1105  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1106 
1107  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1108  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1109  ConstantInt::get(X->getType(), SMax - (C - 1)));
1110 }
1111 
1112 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1113 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1114 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1116  const APInt &AP1,
1117  const APInt &AP2) {
1118  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1119 
1120  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1121  if (I.getPredicate() == I.ICMP_NE)
1122  Pred = CmpInst::getInversePredicate(Pred);
1123  return new ICmpInst(Pred, LHS, RHS);
1124  };
1125 
1126  // Don't bother doing any work for cases which InstSimplify handles.
1127  if (AP2.isNullValue())
1128  return nullptr;
1129 
1130  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1131  if (IsAShr) {
1132  if (AP2.isAllOnesValue())
1133  return nullptr;
1134  if (AP2.isNegative() != AP1.isNegative())
1135  return nullptr;
1136  if (AP2.sgt(AP1))
1137  return nullptr;
1138  }
1139 
1140  if (!AP1)
1141  // 'A' must be large enough to shift out the highest set bit.
1142  return getICmp(I.ICMP_UGT, A,
1143  ConstantInt::get(A->getType(), AP2.logBase2()));
1144 
1145  if (AP1 == AP2)
1146  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1147 
1148  int Shift;
1149  if (IsAShr && AP1.isNegative())
1150  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1151  else
1152  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1153 
1154  if (Shift > 0) {
1155  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1156  // There are multiple solutions if we are comparing against -1 and the LHS
1157  // of the ashr is not a power of two.
1158  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1159  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1160  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1161  } else if (AP1 == AP2.lshr(Shift)) {
1162  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1163  }
1164  }
1165 
1166  // Shifting const2 will never be equal to const1.
1167  // FIXME: This should always be handled by InstSimplify?
1168  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1169  return replaceInstUsesWith(I, TorF);
1170 }
1171 
1172 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1173 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1175  const APInt &AP1,
1176  const APInt &AP2) {
1177  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1178 
1179  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1180  if (I.getPredicate() == I.ICMP_NE)
1181  Pred = CmpInst::getInversePredicate(Pred);
1182  return new ICmpInst(Pred, LHS, RHS);
1183  };
1184 
1185  // Don't bother doing any work for cases which InstSimplify handles.
1186  if (AP2.isNullValue())
1187  return nullptr;
1188 
1189  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1190 
1191  if (!AP1 && AP2TrailingZeros != 0)
1192  return getICmp(
1193  I.ICMP_UGE, A,
1194  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1195 
1196  if (AP1 == AP2)
1197  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1198 
1199  // Get the distance between the lowest bits that are set.
1200  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1201 
1202  if (Shift > 0 && AP2.shl(Shift) == AP1)
1203  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1204 
1205  // Shifting const2 will never be equal to const1.
1206  // FIXME: This should always be handled by InstSimplify?
1207  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1208  return replaceInstUsesWith(I, TorF);
1209 }
1210 
1211 /// The caller has matched a pattern of the form:
1212 /// I = icmp ugt (add (add A, B), CI2), CI1
1213 /// If this is of the form:
1214 /// sum = a + b
1215 /// if (sum+128 >u 255)
1216 /// Then replace it with llvm.sadd.with.overflow.i8.
1217 ///
1219  ConstantInt *CI2, ConstantInt *CI1,
1220  InstCombinerImpl &IC) {
1221  // The transformation we're trying to do here is to transform this into an
1222  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1223  // with a narrower add, and discard the add-with-constant that is part of the
1224  // range check (if we can't eliminate it, this isn't profitable).
1225 
1226  // In order to eliminate the add-with-constant, the compare can be its only
1227  // use.
1228  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1229  if (!AddWithCst->hasOneUse())
1230  return nullptr;
1231 
1232  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1233  if (!CI2->getValue().isPowerOf2())
1234  return nullptr;
1235  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1236  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1237  return nullptr;
1238 
1239  // The width of the new add formed is 1 more than the bias.
1240  ++NewWidth;
1241 
1242  // Check to see that CI1 is an all-ones value with NewWidth bits.
1243  if (CI1->getBitWidth() == NewWidth ||
1244  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1245  return nullptr;
1246 
1247  // This is only really a signed overflow check if the inputs have been
1248  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1249  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1250  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1251  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1252  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1253  return nullptr;
1254 
1255  // In order to replace the original add with a narrower
1256  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1257  // and truncates that discard the high bits of the add. Verify that this is
1258  // the case.
1259  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1260  for (User *U : OrigAdd->users()) {
1261  if (U == AddWithCst)
1262  continue;
1263 
1264  // Only accept truncates for now. We would really like a nice recursive
1265  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1266  // chain to see which bits of a value are actually demanded. If the
1267  // original add had another add which was then immediately truncated, we
1268  // could still do the transformation.
1269  TruncInst *TI = dyn_cast<TruncInst>(U);
1270  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1271  return nullptr;
1272  }
1273 
1274  // If the pattern matches, truncate the inputs to the narrower type and
1275  // use the sadd_with_overflow intrinsic to efficiently compute both the
1276  // result and the overflow bit.
1277  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1279  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1280 
1282 
1283  // Put the new code above the original add, in case there are any uses of the
1284  // add between the add and the compare.
1285  Builder.SetInsertPoint(OrigAdd);
1286 
1287  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1288  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1289  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1290  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1291  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1292 
1293  // The inner add was the result of the narrow add, zero extended to the
1294  // wider type. Replace it with the result computed by the intrinsic.
1295  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1296  IC.eraseInstFromFunction(*OrigAdd);
1297 
1298  // The original icmp gets replaced with the overflow value.
1299  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1300 }
1301 
1302 /// If we have:
1303 /// icmp eq/ne (urem/srem %x, %y), 0
1304 /// iff %y is a power-of-two, we can replace this with a bit test:
1305 /// icmp eq/ne (and %x, (add %y, -1)), 0
1307  // This fold is only valid for equality predicates.
1308  if (!I.isEquality())
1309  return nullptr;
1310  ICmpInst::Predicate Pred;
1311  Value *X, *Y, *Zero;
1312  if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1313  m_CombineAnd(m_Zero(), m_Value(Zero)))))
1314  return nullptr;
1315  if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1316  return nullptr;
1317  // This may increase instruction count, we don't enforce that Y is a constant.
1318  Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1319  Value *Masked = Builder.CreateAnd(X, Mask);
1320  return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1321 }
1322 
1323 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1324 /// by one-less-than-bitwidth into a sign test on the original value.
1326  Instruction *Val;
1327  ICmpInst::Predicate Pred;
1328  if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1329  return nullptr;
1330 
1331  Value *X;
1332  Type *XTy;
1333 
1334  Constant *C;
1335  if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1336  XTy = X->getType();
1337  unsigned XBitWidth = XTy->getScalarSizeInBits();
1338  if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1339  APInt(XBitWidth, XBitWidth - 1))))
1340  return nullptr;
1341  } else if (isa<BinaryOperator>(Val) &&
1342  (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1343  cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1344  /*AnalyzeForSignBitExtraction=*/true))) {
1345  XTy = X->getType();
1346  } else
1347  return nullptr;
1348 
1349  return ICmpInst::Create(Instruction::ICmp,
1353 }
1354 
1355 // Handle icmp pred X, 0
1357  CmpInst::Predicate Pred = Cmp.getPredicate();
1358  if (!match(Cmp.getOperand(1), m_Zero()))
1359  return nullptr;
1360 
1361  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1362  if (Pred == ICmpInst::ICMP_SGT) {
1363  Value *A, *B;
1364  SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1365  if (SPR.Flavor == SPF_SMIN) {
1366  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1367  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1368  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1369  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1370  }
1371  }
1372 
1373  if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1374  return New;
1375 
1376  // Given:
1377  // icmp eq/ne (urem %x, %y), 0
1378  // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1379  // icmp eq/ne %x, 0
1380  Value *X, *Y;
1381  if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1382  ICmpInst::isEquality(Pred)) {
1383  KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1384  KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1385  if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1386  return new ICmpInst(Pred, X, Cmp.getOperand(1));
1387  }
1388 
1389  return nullptr;
1390 }
1391 
1392 /// Fold icmp Pred X, C.
1393 /// TODO: This code structure does not make sense. The saturating add fold
1394 /// should be moved to some other helper and extended as noted below (it is also
1395 /// possible that code has been made unnecessary - do we canonicalize IR to
1396 /// overflow/saturating intrinsics or not?).
1398  // Match the following pattern, which is a common idiom when writing
1399  // overflow-safe integer arithmetic functions. The source performs an addition
1400  // in wider type and explicitly checks for overflow using comparisons against
1401  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1402  //
1403  // TODO: This could probably be generalized to handle other overflow-safe
1404  // operations if we worked out the formulas to compute the appropriate magic
1405  // constants.
1406  //
1407  // sum = a + b
1408  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1409  CmpInst::Predicate Pred = Cmp.getPredicate();
1410  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1411  Value *A, *B;
1412  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1413  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1414  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1415  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1416  return Res;
1417 
1418  // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1419  Constant *C = dyn_cast<Constant>(Op1);
1420  if (!C)
1421  return nullptr;
1422 
1423  if (auto *Phi = dyn_cast<PHINode>(Op0))
1424  if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1425  Type *Ty = Cmp.getType();
1426  Builder.SetInsertPoint(Phi);
1427  PHINode *NewPhi =
1428  Builder.CreatePHI(Ty, Phi->getNumOperands());
1429  for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1430  auto *Input =
1431  cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1432  auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1433  NewPhi->addIncoming(BoolInput, Predecessor);
1434  }
1435  NewPhi->takeName(&Cmp);
1436  return replaceInstUsesWith(Cmp, NewPhi);
1437  }
1438 
1439  return nullptr;
1440 }
1441 
1442 /// Canonicalize icmp instructions based on dominating conditions.
1444  // This is a cheap/incomplete check for dominance - just match a single
1445  // predecessor with a conditional branch.
1446  BasicBlock *CmpBB = Cmp.getParent();
1447  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1448  if (!DomBB)
1449  return nullptr;
1450 
1451  Value *DomCond;
1452  BasicBlock *TrueBB, *FalseBB;
1453  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1454  return nullptr;
1455 
1456  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1457  "Predecessor block does not point to successor?");
1458 
1459  // The branch should get simplified. Don't bother simplifying this condition.
1460  if (TrueBB == FalseBB)
1461  return nullptr;
1462 
1463  // Try to simplify this compare to T/F based on the dominating condition.
1464  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1465  if (Imp)
1466  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1467 
1468  CmpInst::Predicate Pred = Cmp.getPredicate();
1469  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1470  ICmpInst::Predicate DomPred;
1471  const APInt *C, *DomC;
1472  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1473  match(Y, m_APInt(C))) {
1474  // We have 2 compares of a variable with constants. Calculate the constant
1475  // ranges of those compares to see if we can transform the 2nd compare:
1476  // DomBB:
1477  // DomCond = icmp DomPred X, DomC
1478  // br DomCond, CmpBB, FalseBB
1479  // CmpBB:
1480  // Cmp = icmp Pred X, C
1482  ConstantRange DominatingCR =
1483  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1485  CmpInst::getInversePredicate(DomPred), *DomC);
1486  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1487  ConstantRange Difference = DominatingCR.difference(CR);
1488  if (Intersection.isEmptySet())
1489  return replaceInstUsesWith(Cmp, Builder.getFalse());
1490  if (Difference.isEmptySet())
1491  return replaceInstUsesWith(Cmp, Builder.getTrue());
1492 
1493  // Canonicalizing a sign bit comparison that gets used in a branch,
1494  // pessimizes codegen by generating branch on zero instruction instead
1495  // of a test and branch. So we avoid canonicalizing in such situations
1496  // because test and branch instruction has better branch displacement
1497  // than compare and branch instruction.
1498  bool UnusedBit;
1499  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1500  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1501  return nullptr;
1502 
1503  if (const APInt *EqC = Intersection.getSingleElement())
1504  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1505  if (const APInt *NeC = Difference.getSingleElement())
1506  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1507  }
1508 
1509  return nullptr;
1510 }
1511 
1512 /// Fold icmp (trunc X, Y), C.
1514  TruncInst *Trunc,
1515  const APInt &C) {
1516  ICmpInst::Predicate Pred = Cmp.getPredicate();
1517  Value *X = Trunc->getOperand(0);
1518  if (C.isOneValue() && C.getBitWidth() > 1) {
1519  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1520  Value *V = nullptr;
1521  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1522  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1523  ConstantInt::get(V->getType(), 1));
1524  }
1525 
1526  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1527  SrcBits = X->getType()->getScalarSizeInBits();
1528  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1529  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1530  // of the high bits truncated out of x are known.
1531  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1532 
1533  // If all the high bits are known, we can do this xform.
1534  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1535  // Pull in the high bits from known-ones set.
1536  APInt NewRHS = C.zext(SrcBits);
1537  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1538  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1539  }
1540  }
1541 
1542  // Look through truncated right-shift of the sign-bit for a sign-bit check:
1543  // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1544  // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1545  Value *ShOp;
1546  const APInt *ShAmtC;
1547  bool TrueIfSigned;
1548  if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1549  match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1550  DstBits == SrcBits - ShAmtC->getZExtValue()) {
1551  return TrueIfSigned
1552  ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1553  ConstantInt::getNullValue(X->getType()))
1554  : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1555  ConstantInt::getAllOnesValue(X->getType()));
1556  }
1557 
1558  return nullptr;
1559 }
1560 
1561 /// Fold icmp (xor X, Y), C.
1563  BinaryOperator *Xor,
1564  const APInt &C) {
1565  Value *X = Xor->getOperand(0);
1566  Value *Y = Xor->getOperand(1);
1567  const APInt *XorC;
1568  if (!match(Y, m_APInt(XorC)))
1569  return nullptr;
1570 
1571  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1572  // fold the xor.
1573  ICmpInst::Predicate Pred = Cmp.getPredicate();
1574  bool TrueIfSigned = false;
1575  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1576 
1577  // If the sign bit of the XorCst is not set, there is no change to
1578  // the operation, just stop using the Xor.
1579  if (!XorC->isNegative())
1580  return replaceOperand(Cmp, 0, X);
1581 
1582  // Emit the opposite comparison.
1583  if (TrueIfSigned)
1584  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1585  ConstantInt::getAllOnesValue(X->getType()));
1586  else
1587  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1588  ConstantInt::getNullValue(X->getType()));
1589  }
1590 
1591  if (Xor->hasOneUse()) {
1592  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1593  if (!Cmp.isEquality() && XorC->isSignMask()) {
1594  Pred = Cmp.getFlippedSignednessPredicate();
1595  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1596  }
1597 
1598  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1599  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1600  Pred = Cmp.getFlippedSignednessPredicate();
1601  Pred = Cmp.getSwappedPredicate(Pred);
1602  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1603  }
1604  }
1605 
1606  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1607  if (Pred == ICmpInst::ICMP_UGT) {
1608  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1609  if (*XorC == ~C && (C + 1).isPowerOf2())
1610  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1611  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1612  if (*XorC == C && (C + 1).isPowerOf2())
1613  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1614  }
1615  if (Pred == ICmpInst::ICMP_ULT) {
1616  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1617  if (*XorC == -C && C.isPowerOf2())
1618  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1619  ConstantInt::get(X->getType(), ~C));
1620  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1621  if (*XorC == C && (-C).isPowerOf2())
1622  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1623  ConstantInt::get(X->getType(), ~C));
1624  }
1625  return nullptr;
1626 }
1627 
1628 /// Fold icmp (and (sh X, Y), C2), C1.
1630  BinaryOperator *And,
1631  const APInt &C1,
1632  const APInt &C2) {
1633  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1634  if (!Shift || !Shift->isShift())
1635  return nullptr;
1636 
1637  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1638  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1639  // code produced by the clang front-end, for bitfield access.
1640  // This seemingly simple opportunity to fold away a shift turns out to be
1641  // rather complicated. See PR17827 for details.
1642  unsigned ShiftOpcode = Shift->getOpcode();
1643  bool IsShl = ShiftOpcode == Instruction::Shl;
1644  const APInt *C3;
1645  if (match(Shift->getOperand(1), m_APInt(C3))) {
1646  APInt NewAndCst, NewCmpCst;
1647  bool AnyCmpCstBitsShiftedOut;
1648  if (ShiftOpcode == Instruction::Shl) {
1649  // For a left shift, we can fold if the comparison is not signed. We can
1650  // also fold a signed comparison if the mask value and comparison value
1651  // are not negative. These constraints may not be obvious, but we can
1652  // prove that they are correct using an SMT solver.
1653  if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1654  return nullptr;
1655 
1656  NewCmpCst = C1.lshr(*C3);
1657  NewAndCst = C2.lshr(*C3);
1658  AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1659  } else if (ShiftOpcode == Instruction::LShr) {
1660  // For a logical right shift, we can fold if the comparison is not signed.
1661  // We can also fold a signed comparison if the shifted mask value and the
1662  // shifted comparison value are not negative. These constraints may not be
1663  // obvious, but we can prove that they are correct using an SMT solver.
1664  NewCmpCst = C1.shl(*C3);
1665  NewAndCst = C2.shl(*C3);
1666  AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1667  if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1668  return nullptr;
1669  } else {
1670  // For an arithmetic shift, check that both constants don't use (in a
1671  // signed sense) the top bits being shifted out.
1672  assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1673  NewCmpCst = C1.shl(*C3);
1674  NewAndCst = C2.shl(*C3);
1675  AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1676  if (NewAndCst.ashr(*C3) != C2)
1677  return nullptr;
1678  }
1679 
1680  if (AnyCmpCstBitsShiftedOut) {
1681  // If we shifted bits out, the fold is not going to work out. As a
1682  // special case, check to see if this means that the result is always
1683  // true or false now.
1684  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1685  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1686  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1687  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1688  } else {
1689  Value *NewAnd = Builder.CreateAnd(
1690  Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1691  return new ICmpInst(Cmp.getPredicate(),
1692  NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1693  }
1694  }
1695 
1696  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1697  // preferable because it allows the C2 << Y expression to be hoisted out of a
1698  // loop if Y is invariant and X is not.
1699  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1700  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1701  // Compute C2 << Y.
1702  Value *NewShift =
1703  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1704  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1705 
1706  // Compute X & (C2 << Y).
1707  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1708  return replaceOperand(Cmp, 0, NewAnd);
1709  }
1710 
1711  return nullptr;
1712 }
1713 
1714 /// Fold icmp (and X, C2), C1.
1716  BinaryOperator *And,
1717  const APInt &C1) {
1718  bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1719 
1720  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1721  // TODO: We canonicalize to the longer form for scalars because we have
1722  // better analysis/folds for icmp, and codegen may be better with icmp.
1723  if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1724  match(And->getOperand(1), m_One()))
1725  return new TruncInst(And->getOperand(0), Cmp.getType());
1726 
1727  const APInt *C2;
1728  Value *X;
1729  if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1730  return nullptr;
1731 
1732  // Don't perform the following transforms if the AND has multiple uses
1733  if (!And->hasOneUse())
1734  return nullptr;
1735 
1736  if (Cmp.isEquality() && C1.isNullValue()) {
1737  // Restrict this fold to single-use 'and' (PR10267).
1738  // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1739  if (C2->isSignMask()) {
1740  Constant *Zero = Constant::getNullValue(X->getType());
1741  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1742  return new ICmpInst(NewPred, X, Zero);
1743  }
1744 
1745  // Restrict this fold only for single-use 'and' (PR10267).
1746  // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1747  if ((~(*C2) + 1).isPowerOf2()) {
1748  Constant *NegBOC =
1749  ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1750  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1751  return new ICmpInst(NewPred, X, NegBOC);
1752  }
1753  }
1754 
1755  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1756  // the input width without changing the value produced, eliminate the cast:
1757  //
1758  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1759  //
1760  // We can do this transformation if the constants do not have their sign bits
1761  // set or if it is an equality comparison. Extending a relational comparison
1762  // when we're checking the sign bit would not work.
1763  Value *W;
1764  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1765  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1766  // TODO: Is this a good transform for vectors? Wider types may reduce
1767  // throughput. Should this transform be limited (even for scalars) by using
1768  // shouldChangeType()?
1769  if (!Cmp.getType()->isVectorTy()) {
1770  Type *WideType = W->getType();
1771  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1772  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1773  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1774  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1775  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1776  }
1777  }
1778 
1779  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1780  return I;
1781 
1782  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1783  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1784  //
1785  // iff pred isn't signed
1786  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1787  match(And->getOperand(1), m_One())) {
1788  Constant *One = cast<Constant>(And->getOperand(1));
1789  Value *Or = And->getOperand(0);
1790  Value *A, *B, *LShr;
1791  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1792  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1793  unsigned UsesRemoved = 0;
1794  if (And->hasOneUse())
1795  ++UsesRemoved;
1796  if (Or->hasOneUse())
1797  ++UsesRemoved;
1798  if (LShr->hasOneUse())
1799  ++UsesRemoved;
1800 
1801  // Compute A & ((1 << B) | 1)
1802  Value *NewOr = nullptr;
1803  if (auto *C = dyn_cast<Constant>(B)) {
1804  if (UsesRemoved >= 1)
1805  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1806  } else {
1807  if (UsesRemoved >= 3)
1808  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1809  /*HasNUW=*/true),
1810  One, Or->getName());
1811  }
1812  if (NewOr) {
1813  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1814  return replaceOperand(Cmp, 0, NewAnd);
1815  }
1816  }
1817  }
1818 
1819  return nullptr;
1820 }
1821 
1822 /// Fold icmp (and X, Y), C.
1824  BinaryOperator *And,
1825  const APInt &C) {
1826  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1827  return I;
1828 
1829  // TODO: These all require that Y is constant too, so refactor with the above.
1830 
1831  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1832  Value *X = And->getOperand(0);
1833  Value *Y = And->getOperand(1);
1834  if (auto *LI = dyn_cast<LoadInst>(X))
1835  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1836  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1837  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1838  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1839  ConstantInt *C2 = cast<ConstantInt>(Y);
1840  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1841  return Res;
1842  }
1843 
1844  if (!Cmp.isEquality())
1845  return nullptr;
1846 
1847  // X & -C == -C -> X > u ~C
1848  // X & -C != -C -> X <= u ~C
1849  // iff C is a power of 2
1850  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1851  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1853  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1854  }
1855 
1856  // (X & C2) == 0 -> (trunc X) >= 0
1857  // (X & C2) != 0 -> (trunc X) < 0
1858  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1859  const APInt *C2;
1860  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1861  int32_t ExactLogBase2 = C2->exactLogBase2();
1862  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1863  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1864  if (auto *AndVTy = dyn_cast<VectorType>(And->getType()))
1865  NTy = VectorType::get(NTy, AndVTy->getElementCount());
1866  Value *Trunc = Builder.CreateTrunc(X, NTy);
1867  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1869  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1870  }
1871  }
1872 
1873  return nullptr;
1874 }
1875 
1876 /// Fold icmp (or X, Y), C.
1878  BinaryOperator *Or,
1879  const APInt &C) {
1880  ICmpInst::Predicate Pred = Cmp.getPredicate();
1881  if (C.isOneValue()) {
1882  // icmp slt signum(V) 1 --> icmp slt V, 1
1883  Value *V = nullptr;
1884  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1885  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1886  ConstantInt::get(V->getType(), 1));
1887  }
1888 
1889  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1890  const APInt *MaskC;
1891  if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1892  if (*MaskC == C && (C + 1).isPowerOf2()) {
1893  // X | C == C --> X <=u C
1894  // X | C != C --> X >u C
1895  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1897  return new ICmpInst(Pred, OrOp0, OrOp1);
1898  }
1899 
1900  // More general: canonicalize 'equality with set bits mask' to
1901  // 'equality with clear bits mask'.
1902  // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1903  // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1904  if (Or->hasOneUse()) {
1905  Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1906  Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1907  return new ICmpInst(Pred, And, NewC);
1908  }
1909  }
1910 
1911  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1912  return nullptr;
1913 
1914  Value *P, *Q;
1915  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1916  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1917  // -> and (icmp eq P, null), (icmp eq Q, null).
1918  Value *CmpP =
1919  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1920  Value *CmpQ =
1921  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1922  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1923  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1924  }
1925 
1926  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1927  // a shorter form that has more potential to be folded even further.
1928  Value *X1, *X2, *X3, *X4;
1929  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1930  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1931  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1932  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1933  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1934  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1935  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1936  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1937  }
1938 
1939  return nullptr;
1940 }
1941 
1942 /// Fold icmp (mul X, Y), C.
1944  BinaryOperator *Mul,
1945  const APInt &C) {
1946  const APInt *MulC;
1947  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1948  return nullptr;
1949 
1950  // If this is a test of the sign bit and the multiply is sign-preserving with
1951  // a constant operand, use the multiply LHS operand instead.
1952  ICmpInst::Predicate Pred = Cmp.getPredicate();
1953  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1954  if (MulC->isNegative())
1955  Pred = ICmpInst::getSwappedPredicate(Pred);
1956  return new ICmpInst(Pred, Mul->getOperand(0),
1957  Constant::getNullValue(Mul->getType()));
1958  }
1959 
1960  // If the multiply does not wrap, try to divide the compare constant by the
1961  // multiplication factor.
1962  if (Cmp.isEquality() && !MulC->isNullValue()) {
1963  // (mul nsw X, MulC) == C --> X == C /s MulC
1964  if (Mul->hasNoSignedWrap() && C.srem(*MulC).isNullValue()) {
1965  Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
1966  return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1967  }
1968  // (mul nuw X, MulC) == C --> X == C /u MulC
1969  if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isNullValue()) {
1970  Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
1971  return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1972  }
1973  }
1974 
1975  return nullptr;
1976 }
1977 
1978 /// Fold icmp (shl 1, Y), C.
1980  const APInt &C) {
1981  Value *Y;
1982  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1983  return nullptr;
1984 
1985  Type *ShiftType = Shl->getType();
1986  unsigned TypeBits = C.getBitWidth();
1987  bool CIsPowerOf2 = C.isPowerOf2();
1988  ICmpInst::Predicate Pred = Cmp.getPredicate();
1989  if (Cmp.isUnsigned()) {
1990  // (1 << Y) pred C -> Y pred Log2(C)
1991  if (!CIsPowerOf2) {
1992  // (1 << Y) < 30 -> Y <= 4
1993  // (1 << Y) <= 30 -> Y <= 4
1994  // (1 << Y) >= 30 -> Y > 4
1995  // (1 << Y) > 30 -> Y > 4
1996  if (Pred == ICmpInst::ICMP_ULT)
1997  Pred = ICmpInst::ICMP_ULE;
1998  else if (Pred == ICmpInst::ICMP_UGE)
1999  Pred = ICmpInst::ICMP_UGT;
2000  }
2001 
2002  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2003  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2004  unsigned CLog2 = C.logBase2();
2005  if (CLog2 == TypeBits - 1) {
2006  if (Pred == ICmpInst::ICMP_UGE)
2007  Pred = ICmpInst::ICMP_EQ;
2008  else if (Pred == ICmpInst::ICMP_ULT)
2009  Pred = ICmpInst::ICMP_NE;
2010  }
2011  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2012  } else if (Cmp.isSigned()) {
2013  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2014  if (C.isAllOnesValue()) {
2015  // (1 << Y) <= -1 -> Y == 31
2016  if (Pred == ICmpInst::ICMP_SLE)
2017  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2018 
2019  // (1 << Y) > -1 -> Y != 31
2020  if (Pred == ICmpInst::ICMP_SGT)
2021  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2022  } else if (!C) {
2023  // (1 << Y) < 0 -> Y == 31
2024  // (1 << Y) <= 0 -> Y == 31
2025  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2026  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2027 
2028  // (1 << Y) >= 0 -> Y != 31
2029  // (1 << Y) > 0 -> Y != 31
2030  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2031  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2032  }
2033  } else if (Cmp.isEquality() && CIsPowerOf2) {
2034  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2035  }
2036 
2037  return nullptr;
2038 }
2039 
2040 /// Fold icmp (shl X, Y), C.
2042  BinaryOperator *Shl,
2043  const APInt &C) {
2044  const APInt *ShiftVal;
2045  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2046  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2047 
2048  const APInt *ShiftAmt;
2049  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2050  return foldICmpShlOne(Cmp, Shl, C);
2051 
2052  // Check that the shift amount is in range. If not, don't perform undefined
2053  // shifts. When the shift is visited, it will be simplified.
2054  unsigned TypeBits = C.getBitWidth();
2055  if (ShiftAmt->uge(TypeBits))
2056  return nullptr;
2057 
2058  ICmpInst::Predicate Pred = Cmp.getPredicate();
2059  Value *X = Shl->getOperand(0);
2060  Type *ShType = Shl->getType();
2061 
2062  // NSW guarantees that we are only shifting out sign bits from the high bits,
2063  // so we can ASHR the compare constant without needing a mask and eliminate
2064  // the shift.
2065  if (Shl->hasNoSignedWrap()) {
2066  if (Pred == ICmpInst::ICMP_SGT) {
2067  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2068  APInt ShiftedC = C.ashr(*ShiftAmt);
2069  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2070  }
2071  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2072  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2073  APInt ShiftedC = C.ashr(*ShiftAmt);
2074  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2075  }
2076  if (Pred == ICmpInst::ICMP_SLT) {
2077  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2078  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2079  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2080  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2081  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2082  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2083  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2084  }
2085  // If this is a signed comparison to 0 and the shift is sign preserving,
2086  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2087  // do that if we're sure to not continue on in this function.
2088  if (isSignTest(Pred, C))
2089  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2090  }
2091 
2092  // NUW guarantees that we are only shifting out zero bits from the high bits,
2093  // so we can LSHR the compare constant without needing a mask and eliminate
2094  // the shift.
2095  if (Shl->hasNoUnsignedWrap()) {
2096  if (Pred == ICmpInst::ICMP_UGT) {
2097  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2098  APInt ShiftedC = C.lshr(*ShiftAmt);
2099  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2100  }
2101  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2102  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2103  APInt ShiftedC = C.lshr(*ShiftAmt);
2104  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2105  }
2106  if (Pred == ICmpInst::ICMP_ULT) {
2107  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2108  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2109  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2110  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2111  assert(C.ugt(0) && "ult 0 should have been eliminated");
2112  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2113  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2114  }
2115  }
2116 
2117  if (Cmp.isEquality() && Shl->hasOneUse()) {
2118  // Strength-reduce the shift into an 'and'.
2120  ShType,
2121  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2122  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2123  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2124  return new ICmpInst(Pred, And, LShrC);
2125  }
2126 
2127  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2128  bool TrueIfSigned = false;
2129  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2130  // (X << 31) <s 0 --> (X & 1) != 0
2132  ShType,
2133  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2134  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2135  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2136  And, Constant::getNullValue(ShType));
2137  }
2138 
2139  // Simplify 'shl' inequality test into 'and' equality test.
2140  if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2141  // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2142  if ((C + 1).isPowerOf2() &&
2143  (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2144  Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2145  return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2147  And, Constant::getNullValue(ShType));
2148  }
2149  // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2150  if (C.isPowerOf2() &&
2151  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2152  Value *And =
2153  Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2154  return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2156  And, Constant::getNullValue(ShType));
2157  }
2158  }
2159 
2160  // Transform (icmp pred iM (shl iM %v, N), C)
2161  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2162  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2163  // This enables us to get rid of the shift in favor of a trunc that may be
2164  // free on the target. It has the additional benefit of comparing to a
2165  // smaller constant that may be more target-friendly.
2166  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2167  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2168  DL.isLegalInteger(TypeBits - Amt)) {
2169  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2170  if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2171  TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
2172  Constant *NewC =
2173  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2174  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2175  }
2176 
2177  return nullptr;
2178 }
2179 
2180 /// Fold icmp ({al}shr X, Y), C.
2182  BinaryOperator *Shr,
2183  const APInt &C) {
2184  // An exact shr only shifts out zero bits, so:
2185  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2186  Value *X = Shr->getOperand(0);
2187  CmpInst::Predicate Pred = Cmp.getPredicate();
2188  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2189  C.isNullValue())
2190  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2191 
2192  const APInt *ShiftVal;
2193  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2194  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2195 
2196  const APInt *ShiftAmt;
2197  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2198  return nullptr;
2199 
2200  // Check that the shift amount is in range. If not, don't perform undefined
2201  // shifts. When the shift is visited it will be simplified.
2202  unsigned TypeBits = C.getBitWidth();
2203  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2204  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2205  return nullptr;
2206 
2207  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2208  bool IsExact = Shr->isExact();
2209  Type *ShrTy = Shr->getType();
2210  // TODO: If we could guarantee that InstSimplify would handle all of the
2211  // constant-value-based preconditions in the folds below, then we could assert
2212  // those conditions rather than checking them. This is difficult because of
2213  // undef/poison (PR34838).
2214  if (IsAShr) {
2215  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2216  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2217  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2218  APInt ShiftedC = C.shl(ShAmtVal);
2219  if (ShiftedC.ashr(ShAmtVal) == C)
2220  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2221  }
2222  if (Pred == CmpInst::ICMP_SGT) {
2223  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2224  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2225  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2226  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2227  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2228  }
2229 
2230  // If the compare constant has significant bits above the lowest sign-bit,
2231  // then convert an unsigned cmp to a test of the sign-bit:
2232  // (ashr X, ShiftC) u> C --> X s< 0
2233  // (ashr X, ShiftC) u< C --> X s> -1
2234  if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2235  if (Pred == CmpInst::ICMP_UGT) {
2236  return new ICmpInst(CmpInst::ICMP_SLT, X,
2237  ConstantInt::getNullValue(ShrTy));
2238  }
2239  if (Pred == CmpInst::ICMP_ULT) {
2240  return new ICmpInst(CmpInst::ICMP_SGT, X,
2242  }
2243  }
2244  } else {
2245  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2246  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2247  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2248  APInt ShiftedC = C.shl(ShAmtVal);
2249  if (ShiftedC.lshr(ShAmtVal) == C)
2250  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2251  }
2252  if (Pred == CmpInst::ICMP_UGT) {
2253  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2254  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2255  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2256  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2257  }
2258  }
2259 
2260  if (!Cmp.isEquality())
2261  return nullptr;
2262 
2263  // Handle equality comparisons of shift-by-constant.
2264 
2265  // If the comparison constant changes with the shift, the comparison cannot
2266  // succeed (bits of the comparison constant cannot match the shifted value).
2267  // This should be known by InstSimplify and already be folded to true/false.
2268  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2269  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2270  "Expected icmp+shr simplify did not occur.");
2271 
2272  // If the bits shifted out are known zero, compare the unshifted value:
2273  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2274  if (Shr->isExact())
2275  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2276 
2277  if (C.isNullValue()) {
2278  // == 0 is u< 1.
2279  if (Pred == CmpInst::ICMP_EQ)
2280  return new ICmpInst(CmpInst::ICMP_ULT, X,
2281  ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2282  else
2283  return new ICmpInst(CmpInst::ICMP_UGT, X,
2284  ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2285  }
2286 
2287  if (Shr->hasOneUse()) {
2288  // Canonicalize the shift into an 'and':
2289  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2290  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2291  Constant *Mask = ConstantInt::get(ShrTy, Val);
2292  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2293  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2294  }
2295 
2296  return nullptr;
2297 }
2298 
2300  BinaryOperator *SRem,
2301  const APInt &C) {
2302  // Match an 'is positive' or 'is negative' comparison of remainder by a
2303  // constant power-of-2 value:
2304  // (X % pow2C) sgt/slt 0
2305  const ICmpInst::Predicate Pred = Cmp.getPredicate();
2306  if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2307  return nullptr;
2308 
2309  // TODO: The one-use check is standard because we do not typically want to
2310  // create longer instruction sequences, but this might be a special-case
2311  // because srem is not good for analysis or codegen.
2312  if (!SRem->hasOneUse())
2313  return nullptr;
2314 
2315  const APInt *DivisorC;
2316  if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2317  return nullptr;
2318 
2319  // Mask off the sign bit and the modulo bits (low-bits).
2320  Type *Ty = SRem->getType();
2321  APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2322  Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2323  Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2324 
2325  // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2326  // bit is set. Example:
2327  // (i8 X % 32) s> 0 --> (X & 159) s> 0
2328  if (Pred == ICmpInst::ICMP_SGT)
2330 
2331  // For 'is negative?' check that the sign-bit is set and at least 1 masked
2332  // bit is set. Example:
2333  // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2334  return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2335 }
2336 
2337 /// Fold icmp (udiv X, Y), C.
2339  BinaryOperator *UDiv,
2340  const APInt &C) {
2341  const APInt *C2;
2342  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2343  return nullptr;
2344 
2345  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2346 
2347  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2348  Value *Y = UDiv->getOperand(1);
2349  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2350  assert(!C.isMaxValue() &&
2351  "icmp ugt X, UINT_MAX should have been simplified already.");
2352  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2353  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2354  }
2355 
2356  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2357  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2358  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2359  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2360  ConstantInt::get(Y->getType(), C2->udiv(C)));
2361  }
2362 
2363  return nullptr;
2364 }
2365 
2366 /// Fold icmp ({su}div X, Y), C.
2368  BinaryOperator *Div,
2369  const APInt &C) {
2370  // Fold: icmp pred ([us]div X, C2), C -> range test
2371  // Fold this div into the comparison, producing a range check.
2372  // Determine, based on the divide type, what the range is being
2373  // checked. If there is an overflow on the low or high side, remember
2374  // it, otherwise compute the range [low, hi) bounding the new value.
2375  // See: InsertRangeTest above for the kinds of replacements possible.
2376  const APInt *C2;
2377  if (!match(Div->getOperand(1), m_APInt(C2)))
2378  return nullptr;
2379 
2380  // FIXME: If the operand types don't match the type of the divide
2381  // then don't attempt this transform. The code below doesn't have the
2382  // logic to deal with a signed divide and an unsigned compare (and
2383  // vice versa). This is because (x /s C2) <s C produces different
2384  // results than (x /s C2) <u C or (x /u C2) <s C or even
2385  // (x /u C2) <u C. Simply casting the operands and result won't
2386  // work. :( The if statement below tests that condition and bails
2387  // if it finds it.
2388  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2389  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2390  return nullptr;
2391 
2392  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2393  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2394  // division-by-constant cases should be present, we can not assert that they
2395  // have happened before we reach this icmp instruction.
2396  if (C2->isNullValue() || C2->isOneValue() ||
2397  (DivIsSigned && C2->isAllOnesValue()))
2398  return nullptr;
2399 
2400  // Compute Prod = C * C2. We are essentially solving an equation of
2401  // form X / C2 = C. We solve for X by multiplying C2 and C.
2402  // By solving for X, we can turn this into a range check instead of computing
2403  // a divide.
2404  APInt Prod = C * *C2;
2405 
2406  // Determine if the product overflows by seeing if the product is not equal to
2407  // the divide. Make sure we do the same kind of divide as in the LHS
2408  // instruction that we're folding.
2409  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2410 
2411  ICmpInst::Predicate Pred = Cmp.getPredicate();
2412 
2413  // If the division is known to be exact, then there is no remainder from the
2414  // divide, so the covered range size is unit, otherwise it is the divisor.
2415  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2416 
2417  // Figure out the interval that is being checked. For example, a comparison
2418  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2419  // Compute this interval based on the constants involved and the signedness of
2420  // the compare/divide. This computes a half-open interval, keeping track of
2421  // whether either value in the interval overflows. After analysis each
2422  // overflow variable is set to 0 if it's corresponding bound variable is valid
2423  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2424  int LoOverflow = 0, HiOverflow = 0;
2425  APInt LoBound, HiBound;
2426 
2427  if (!DivIsSigned) { // udiv
2428  // e.g. X/5 op 3 --> [15, 20)
2429  LoBound = Prod;
2430  HiOverflow = LoOverflow = ProdOV;
2431  if (!HiOverflow) {
2432  // If this is not an exact divide, then many values in the range collapse
2433  // to the same result value.
2434  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2435  }
2436  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2437  if (C.isNullValue()) { // (X / pos) op 0
2438  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2439  LoBound = -(RangeSize - 1);
2440  HiBound = RangeSize;
2441  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2442  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2443  HiOverflow = LoOverflow = ProdOV;
2444  if (!HiOverflow)
2445  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2446  } else { // (X / pos) op neg
2447  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2448  HiBound = Prod + 1;
2449  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2450  if (!LoOverflow) {
2451  APInt DivNeg = -RangeSize;
2452  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2453  }
2454  }
2455  } else if (C2->isNegative()) { // Divisor is < 0.
2456  if (Div->isExact())
2457  RangeSize.negate();
2458  if (C.isNullValue()) { // (X / neg) op 0
2459  // e.g. X/-5 op 0 --> [-4, 5)
2460  LoBound = RangeSize + 1;
2461  HiBound = -RangeSize;
2462  if (HiBound == *C2) { // -INTMIN = INTMIN
2463  HiOverflow = 1; // [INTMIN+1, overflow)
2464  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2465  }
2466  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2467  // e.g. X/-5 op 3 --> [-19, -14)
2468  HiBound = Prod + 1;
2469  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2470  if (!LoOverflow)
2471  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2472  } else { // (X / neg) op neg
2473  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2474  LoOverflow = HiOverflow = ProdOV;
2475  if (!HiOverflow)
2476  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2477  }
2478 
2479  // Dividing by a negative swaps the condition. LT <-> GT
2480  Pred = ICmpInst::getSwappedPredicate(Pred);
2481  }
2482 
2483  Value *X = Div->getOperand(0);
2484  switch (Pred) {
2485  default: llvm_unreachable("Unhandled icmp opcode!");
2486  case ICmpInst::ICMP_EQ:
2487  if (LoOverflow && HiOverflow)
2488  return replaceInstUsesWith(Cmp, Builder.getFalse());
2489  if (HiOverflow)
2490  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2492  ConstantInt::get(Div->getType(), LoBound));
2493  if (LoOverflow)
2494  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2496  ConstantInt::get(Div->getType(), HiBound));
2497  return replaceInstUsesWith(
2498  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2499  case ICmpInst::ICMP_NE:
2500  if (LoOverflow && HiOverflow)
2501  return replaceInstUsesWith(Cmp, Builder.getTrue());
2502  if (HiOverflow)
2503  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2505  ConstantInt::get(Div->getType(), LoBound));
2506  if (LoOverflow)
2507  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2509  ConstantInt::get(Div->getType(), HiBound));
2510  return replaceInstUsesWith(Cmp,
2511  insertRangeTest(X, LoBound, HiBound,
2512  DivIsSigned, false));
2513  case ICmpInst::ICMP_ULT:
2514  case ICmpInst::ICMP_SLT:
2515  if (LoOverflow == +1) // Low bound is greater than input range.
2516  return replaceInstUsesWith(Cmp, Builder.getTrue());
2517  if (LoOverflow == -1) // Low bound is less than input range.
2518  return replaceInstUsesWith(Cmp, Builder.getFalse());
2519  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2520  case ICmpInst::ICMP_UGT:
2521  case ICmpInst::ICMP_SGT:
2522  if (HiOverflow == +1) // High bound greater than input range.
2523  return replaceInstUsesWith(Cmp, Builder.getFalse());
2524  if (HiOverflow == -1) // High bound less than input range.
2525  return replaceInstUsesWith(Cmp, Builder.getTrue());
2526  if (Pred == ICmpInst::ICMP_UGT)
2527  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2528  ConstantInt::get(Div->getType(), HiBound));
2529  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2530  ConstantInt::get(Div->getType(), HiBound));
2531  }
2532 
2533  return nullptr;
2534 }
2535 
2536 /// Fold icmp (sub X, Y), C.
2538  BinaryOperator *Sub,
2539  const APInt &C) {
2540  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2541  ICmpInst::Predicate Pred = Cmp.getPredicate();
2542  const APInt *C2;
2543  APInt SubResult;
2544 
2545  // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2546  if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2547  return new ICmpInst(Cmp.getPredicate(), Y,
2548  ConstantInt::get(Y->getType(), 0));
2549 
2550  // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2551  if (match(X, m_APInt(C2)) &&
2552  ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2553  (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2554  !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2555  return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2556  ConstantInt::get(Y->getType(), SubResult));
2557 
2558  // The following transforms are only worth it if the only user of the subtract
2559  // is the icmp.
2560  if (!Sub->hasOneUse())
2561  return nullptr;
2562 
2563  if (Sub->hasNoSignedWrap()) {
2564  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2565  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2566  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2567 
2568  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2569  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2570  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2571 
2572  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2573  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2574  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2575 
2576  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2577  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2578  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2579  }
2580 
2581  if (!match(X, m_APInt(C2)))
2582  return nullptr;
2583 
2584  // C2 - Y <u C -> (Y | (C - 1)) == C2
2585  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2586  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2587  (*C2 & (C - 1)) == (C - 1))
2588  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2589 
2590  // C2 - Y >u C -> (Y | C) != C2
2591  // iff C2 & C == C and C + 1 is a power of 2
2592  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2593  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2594 
2595  return nullptr;
2596 }
2597 
2598 /// Fold icmp (add X, Y), C.
2600  BinaryOperator *Add,
2601  const APInt &C) {
2602  Value *Y = Add->getOperand(1);
2603  const APInt *C2;
2604  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2605  return nullptr;
2606 
2607  // Fold icmp pred (add X, C2), C.
2608  Value *X = Add->getOperand(0);
2609  Type *Ty = Add->getType();
2610  CmpInst::Predicate Pred = Cmp.getPredicate();
2611 
2612  // If the add does not wrap, we can always adjust the compare by subtracting
2613  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2614  // are canonicalized to SGT/SLT/UGT/ULT.
2615  if ((Add->hasNoSignedWrap() &&
2616  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2617  (Add->hasNoUnsignedWrap() &&
2618  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2619  bool Overflow;
2620  APInt NewC =
2621  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2622  // If there is overflow, the result must be true or false.
2623  // TODO: Can we assert there is no overflow because InstSimplify always
2624  // handles those cases?
2625  if (!Overflow)
2626  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2627  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2628  }
2629 
2630  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2631  const APInt &Upper = CR.getUpper();
2632  const APInt &Lower = CR.getLower();
2633  if (Cmp.isSigned()) {
2634  if (Lower.isSignMask())
2635  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2636  if (Upper.isSignMask())
2637  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2638  } else {
2639  if (Lower.isMinValue())
2640  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2641  if (Upper.isMinValue())
2642  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2643  }
2644 
2645  if (!Add->hasOneUse())
2646  return nullptr;
2647 
2648  // X+C <u C2 -> (X & -C2) == C
2649  // iff C & (C2-1) == 0
2650  // C2 is a power of 2
2651  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2652  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2653  ConstantExpr::getNeg(cast<Constant>(Y)));
2654 
2655  // X+C >u C2 -> (X & ~C2) != C
2656  // iff C & C2 == 0
2657  // C2+1 is a power of 2
2658  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2659  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2660  ConstantExpr::getNeg(cast<Constant>(Y)));
2661 
2662  return nullptr;
2663 }
2664 
2666  Value *&RHS, ConstantInt *&Less,
2667  ConstantInt *&Equal,
2668  ConstantInt *&Greater) {
2669  // TODO: Generalize this to work with other comparison idioms or ensure
2670  // they get canonicalized into this form.
2671 
2672  // select i1 (a == b),
2673  // i32 Equal,
2674  // i32 (select i1 (a < b), i32 Less, i32 Greater)
2675  // where Equal, Less and Greater are placeholders for any three constants.
2676  ICmpInst::Predicate PredA;
2677  if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2678  !ICmpInst::isEquality(PredA))
2679  return false;
2680  Value *EqualVal = SI->getTrueValue();
2681  Value *UnequalVal = SI->getFalseValue();
2682  // We still can get non-canonical predicate here, so canonicalize.
2683  if (PredA == ICmpInst::ICMP_NE)
2684  std::swap(EqualVal, UnequalVal);
2685  if (!match(EqualVal, m_ConstantInt(Equal)))
2686  return false;
2687  ICmpInst::Predicate PredB;
2688  Value *LHS2, *RHS2;
2689  if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2690  m_ConstantInt(Less), m_ConstantInt(Greater))))
2691  return false;
2692  // We can get predicate mismatch here, so canonicalize if possible:
2693  // First, ensure that 'LHS' match.
2694  if (LHS2 != LHS) {
2695  // x sgt y <--> y slt x
2696  std::swap(LHS2, RHS2);
2697  PredB = ICmpInst::getSwappedPredicate(PredB);
2698  }
2699  if (LHS2 != LHS)
2700  return false;
2701  // We also need to canonicalize 'RHS'.
2702  if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2703  // x sgt C-1 <--> x sge C <--> not(x slt C)
2704  auto FlippedStrictness =
2706  PredB, cast<Constant>(RHS2));
2707  if (!FlippedStrictness)
2708  return false;
2709  assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check");
2710  RHS2 = FlippedStrictness->second;
2711  // And kind-of perform the result swap.
2712  std::swap(Less, Greater);
2713  PredB = ICmpInst::ICMP_SLT;
2714  }
2715  return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2716 }
2717 
2719  SelectInst *Select,
2720  ConstantInt *C) {
2721 
2722  assert(C && "Cmp RHS should be a constant int!");
2723  // If we're testing a constant value against the result of a three way
2724  // comparison, the result can be expressed directly in terms of the
2725  // original values being compared. Note: We could possibly be more
2726  // aggressive here and remove the hasOneUse test. The original select is
2727  // really likely to simplify or sink when we remove a test of the result.
2728  Value *OrigLHS, *OrigRHS;
2729  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2730  if (Cmp.hasOneUse() &&
2731  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2732  C3GreaterThan)) {
2733  assert(C1LessThan && C2Equal && C3GreaterThan);
2734 
2735  bool TrueWhenLessThan =
2736  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2737  ->isAllOnesValue();
2738  bool TrueWhenEqual =
2739  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2740  ->isAllOnesValue();
2741  bool TrueWhenGreaterThan =
2742  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2743  ->isAllOnesValue();
2744 
2745  // This generates the new instruction that will replace the original Cmp
2746  // Instruction. Instead of enumerating the various combinations when
2747  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2748  // false, we rely on chaining of ORs and future passes of InstCombine to
2749  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2750 
2751  // When none of the three constants satisfy the predicate for the RHS (C),
2752  // the entire original Cmp can be simplified to a false.
2753  Value *Cond = Builder.getFalse();
2754  if (TrueWhenLessThan)
2755  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2756  OrigLHS, OrigRHS));
2757  if (TrueWhenEqual)
2758  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2759  OrigLHS, OrigRHS));
2760  if (TrueWhenGreaterThan)
2761  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2762  OrigLHS, OrigRHS));
2763 
2764  return replaceInstUsesWith(Cmp, Cond);
2765  }
2766  return nullptr;
2767 }
2768 
2771  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2772  if (!Bitcast)
2773  return nullptr;
2774 
2775  ICmpInst::Predicate Pred = Cmp.getPredicate();
2776  Value *Op1 = Cmp.getOperand(1);
2777  Value *BCSrcOp = Bitcast->getOperand(0);
2778 
2779  // Make sure the bitcast doesn't change the number of vector elements.
2780  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2781  Bitcast->getDestTy()->getScalarSizeInBits()) {
2782  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2783  Value *X;
2784  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2785  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2786  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2787  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2788  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2789  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2790  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2791  match(Op1, m_Zero()))
2792  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2793 
2794  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2795  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2796  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2797 
2798  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2799  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2800  return new ICmpInst(Pred, X,
2801  ConstantInt::getAllOnesValue(X->getType()));
2802  }
2803 
2804  // Zero-equality checks are preserved through unsigned floating-point casts:
2805  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2806  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2807  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2808  if (Cmp.isEquality() && match(Op1, m_Zero()))
2809  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2810 
2811  // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2812  // the FP extend/truncate because that cast does not change the sign-bit.
2813  // This is true for all standard IEEE-754 types and the X86 80-bit type.
2814  // The sign-bit is always the most significant bit in those types.
2815  const APInt *C;
2816  bool TrueIfSigned;
2817  if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2818  InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2819  if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2820  match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2821  // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2822  // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2823  Type *XType = X->getType();
2824 
2825  // We can't currently handle Power style floating point operations here.
2826  if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2827 
2828  Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2829  if (auto *XVTy = dyn_cast<VectorType>(XType))
2830  NewType = VectorType::get(NewType, XVTy->getElementCount());
2831  Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2832  if (TrueIfSigned)
2833  return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2834  ConstantInt::getNullValue(NewType));
2835  else
2836  return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2837  ConstantInt::getAllOnesValue(NewType));
2838  }
2839  }
2840  }
2841  }
2842 
2843  // Test to see if the operands of the icmp are casted versions of other
2844  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2845  if (Bitcast->getType()->isPointerTy() &&
2846  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2847  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2848  // so eliminate it as well.
2849  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2850  Op1 = BC2->getOperand(0);
2851 
2852  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2853  return new ICmpInst(Pred, BCSrcOp, Op1);
2854  }
2855 
2856  // Folding: icmp <pred> iN X, C
2857  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2858  // and C is a splat of a K-bit pattern
2859  // and SC is a constant vector = <C', C', C', ..., C'>
2860  // Into:
2861  // %E = extractelement <M x iK> %vec, i32 C'
2862  // icmp <pred> iK %E, trunc(C)
2863  const APInt *C;
2864  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2865  !Bitcast->getType()->isIntegerTy() ||
2866  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2867  return nullptr;
2868 
2869  Value *Vec;
2871  if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2872  // Check whether every element of Mask is the same constant
2873  if (is_splat(Mask)) {
2874  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2875  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2876  if (C->isSplat(EltTy->getBitWidth())) {
2877  // Fold the icmp based on the value of C
2878  // If C is M copies of an iK sized bit pattern,
2879  // then:
2880  // => %E = extractelement <N x iK> %vec, i32 Elem
2881  // icmp <pred> iK %SplatVal, <pattern>
2882  Value *Elem = Builder.getInt32(Mask[0]);
2883  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2884  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2885  return new ICmpInst(Pred, Extract, NewC);
2886  }
2887  }
2888  }
2889  return nullptr;
2890 }
2891 
2892 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2893 /// where X is some kind of instruction.
2895  const APInt *C;
2896  if (!match(Cmp.getOperand(1), m_APInt(C)))
2897  return nullptr;
2898 
2899  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2900  switch (BO->getOpcode()) {
2901  case Instruction::Xor:
2902  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2903  return I;
2904  break;
2905  case Instruction::And:
2906  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2907  return I;
2908  break;
2909  case Instruction::Or:
2910  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2911  return I;
2912  break;
2913  case Instruction::Mul:
2914  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2915  return I;
2916  break;
2917  case Instruction::Shl:
2918  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2919  return I;
2920  break;
2921  case Instruction::LShr:
2922  case Instruction::AShr:
2923  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2924  return I;
2925  break;
2926  case Instruction::SRem:
2927  if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2928  return I;
2929  break;
2930  case Instruction::UDiv:
2931  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2932  return I;
2934  case Instruction::SDiv:
2935  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2936  return I;
2937  break;
2938  case Instruction::Sub:
2939  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2940  return I;
2941  break;
2942  case Instruction::Add:
2943  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2944  return I;
2945  break;
2946  default:
2947  break;
2948  }
2949  // TODO: These folds could be refactored to be part of the above calls.
2950  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2951  return I;
2952  }
2953 
2954  // Match against CmpInst LHS being instructions other than binary operators.
2955 
2956  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2957  // For now, we only support constant integers while folding the
2958  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2959  // similar to the cases handled by binary ops above.
2960  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2961  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2962  return I;
2963  }
2964 
2965  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2966  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2967  return I;
2968  }
2969 
2970  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2971  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2972  return I;
2973 
2974  return nullptr;
2975 }
2976 
2977 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2978 /// icmp eq/ne BO, C.
2980  ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
2981  // TODO: Some of these folds could work with arbitrary constants, but this
2982  // function is limited to scalar and vector splat constants.
2983  if (!Cmp.isEquality())
2984  return nullptr;
2985 
2986  ICmpInst::Predicate Pred = Cmp.getPredicate();
2987  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2988  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2989  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2990 
2991  switch (BO->getOpcode()) {
2992  case Instruction::SRem:
2993  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2994  if (C.isNullValue() && BO->hasOneUse()) {
2995  const APInt *BOC;
2996  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2997  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2998  return new ICmpInst(Pred, NewRem,
3000  }
3001  }
3002  break;
3003  case Instruction::Add: {
3004  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3005  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3006  if (BO->hasOneUse())
3007  return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3008  } else if (C.isNullValue()) {
3009  // Replace ((add A, B) != 0) with (A != -B) if A or B is
3010  // efficiently invertible, or if the add has just this one use.
3011  if (Value *NegVal = dyn_castNegVal(BOp1))
3012  return new ICmpInst(Pred, BOp0, NegVal);
3013  if (Value *NegVal = dyn_castNegVal(BOp0))
3014  return new ICmpInst(Pred, NegVal, BOp1);
3015  if (BO->hasOneUse()) {
3016  Value *Neg = Builder.CreateNeg(BOp1);
3017  Neg->takeName(BO);
3018  return new ICmpInst(Pred, BOp0, Neg);
3019  }
3020  }
3021  break;
3022  }
3023  case Instruction::Xor:
3024  if (BO->hasOneUse()) {
3025  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3026  // For the xor case, we can xor two constants together, eliminating
3027  // the explicit xor.
3028  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3029  } else if (C.isNullValue()) {
3030  // Replace ((xor A, B) != 0) with (A != B)
3031  return new ICmpInst(Pred, BOp0, BOp1);
3032  }
3033  }
3034  break;
3035  case Instruction::Sub:
3036  if (BO->hasOneUse()) {
3037  // Only check for constant LHS here, as constant RHS will be canonicalized
3038  // to add and use the fold above.
3039  if (Constant *BOC = dyn_cast<Constant>(BOp0)) {
3040  // Replace ((sub BOC, B) != C) with (B != BOC-C).
3041  return new ICmpInst(Pred, BOp1, ConstantExpr::getSub(BOC, RHS));
3042  } else if (C.isNullValue()) {
3043  // Replace ((sub A, B) != 0) with (A != B).
3044  return new ICmpInst(Pred, BOp0, BOp1);
3045  }
3046  }
3047  break;
3048  case Instruction::Or: {
3049  const APInt *BOC;
3050  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3051  // Comparing if all bits outside of a constant mask are set?
3052  // Replace (X | C) == -1 with (X & ~C) == ~C.
3053  // This removes the -1 constant.
3054  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3055  Value *And = Builder.CreateAnd(BOp0, NotBOC);
3056  return new ICmpInst(Pred, And, NotBOC);
3057  }
3058  break;
3059  }
3060  case Instruction::And: {
3061  const APInt *BOC;
3062  if (match(BOp1, m_APInt(BOC))) {
3063  // If we have ((X & C) == C), turn it into ((X & C) != 0).
3064  if (C == *BOC && C.isPowerOf2())
3065  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3066  BO, Constant::getNullValue(RHS->getType()));
3067  }
3068  break;
3069  }
3070  case Instruction::UDiv:
3071  if (C.isNullValue()) {
3072  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3073  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3074  return new ICmpInst(NewPred, BOp1, BOp0);
3075  }
3076  break;
3077  default:
3078  break;
3079  }
3080  return nullptr;
3081 }
3082 
3083 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3085  ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3086  Type *Ty = II->getType();
3087  unsigned BitWidth = C.getBitWidth();
3088  switch (II->getIntrinsicID()) {
3089  case Intrinsic::abs:
3090  // abs(A) == 0 -> A == 0
3091  // abs(A) == INT_MIN -> A == INT_MIN
3092  if (C.isNullValue() || C.isMinSignedValue())
3093  return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3094  ConstantInt::get(Ty, C));
3095  break;
3096 
3097  case Intrinsic::bswap:
3098  // bswap(A) == C -> A == bswap(C)
3099  return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3100  ConstantInt::get(Ty, C.byteSwap()));
3101 
3102  case Intrinsic::ctlz:
3103  case Intrinsic::cttz: {
3104  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3105  if (C == BitWidth)
3106  return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3108 
3109  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3110  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3111  // Limit to one use to ensure we don't increase instruction count.
3112  unsigned Num = C.getLimitedValue(BitWidth);
3113  if (Num != BitWidth && II->hasOneUse()) {
3114  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3115  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3116  : APInt::getHighBitsSet(BitWidth, Num + 1);
3117  APInt Mask2 = IsTrailing
3119  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3120  return new ICmpInst(Cmp.getPredicate(),
3121  Builder.CreateAnd(II->getArgOperand(0), Mask1),
3122  ConstantInt::get(Ty, Mask2));
3123  }
3124  break;
3125  }
3126 
3127  case Intrinsic::ctpop: {
3128  // popcount(A) == 0 -> A == 0 and likewise for !=
3129  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3130  bool IsZero = C.isNullValue();
3131  if (IsZero || C == BitWidth)
3132  return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3134 
3135  break;
3136  }
3137 
3138  case Intrinsic::uadd_sat: {
3139  // uadd.sat(a, b) == 0 -> (a | b) == 0
3140  if (C.isNullValue()) {
3141  Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3142  return new ICmpInst(Cmp.getPredicate(), Or, Constant::getNullValue(Ty));
3143  }
3144  break;
3145  }
3146 
3147  case Intrinsic::usub_sat: {
3148  // usub.sat(a, b) == 0 -> a <= b
3149  if (C.isNullValue()) {
3150  ICmpInst::Predicate NewPred = Cmp.getPredicate() == ICmpInst::ICMP_EQ
3152  return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3153  }
3154  break;
3155  }
3156  default:
3157  break;
3158  }
3159 
3160  return nullptr;
3161 }
3162 
3163 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3165  IntrinsicInst *II,
3166  const APInt &C) {
3167  if (Cmp.isEquality())
3168  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3169 
3170  Type *Ty = II->getType();
3171  unsigned BitWidth = C.getBitWidth();
3172  ICmpInst::Predicate Pred = Cmp.getPredicate();
3173  switch (II->getIntrinsicID()) {
3174  case Intrinsic::ctpop: {
3175  // (ctpop X > BitWidth - 1) --> X == -1
3176  Value *X = II->getArgOperand(0);
3177  if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
3178  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
3180  // (ctpop X < BitWidth) --> X != -1
3181  if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
3182  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
3184  break;
3185  }
3186  case Intrinsic::ctlz: {
3187  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3188  if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3189  unsigned Num = C.getLimitedValue();
3190  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3191  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3192  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3193  }
3194 
3195  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3196  if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3197  unsigned Num = C.getLimitedValue();
3198  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3199  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3200  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3201  }
3202  break;
3203  }
3204  case Intrinsic::cttz: {
3205  // Limit to one use to ensure we don't increase instruction count.
3206  if (!II->hasOneUse())
3207  return nullptr;
3208 
3209  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3210  if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3211  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3212  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3213  Builder.CreateAnd(II->getArgOperand(0), Mask),
3215  }
3216 
3217  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3218  if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3219  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3220  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3221  Builder.CreateAnd(II->getArgOperand(0), Mask),
3223  }
3224  break;
3225  }
3226  default:
3227  break;
3228  }
3229 
3230  return nullptr;
3231 }
3232 
3233 /// Handle icmp with constant (but not simple integer constant) RHS.
3235  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3236  Constant *RHSC = dyn_cast<Constant>(Op1);
3237  Instruction *LHSI = dyn_cast<Instruction>(Op0);
3238  if (!RHSC || !LHSI)
3239  return nullptr;
3240 
3241  switch (LHSI->getOpcode()) {
3242  case Instruction::GetElementPtr:
3243  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3244  if (RHSC->isNullValue() &&
3245  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3246  return new ICmpInst(
3247  I.getPredicate(), LHSI->getOperand(0),
3249  break;
3250  case Instruction::PHI:
3251  // Only fold icmp into the PHI if the phi and icmp are in the same
3252  // block. If in the same block, we're encouraging jump threading. If
3253  // not, we are just pessimizing the code by making an i1 phi.
3254  if (LHSI->getParent() == I.getParent())
3255  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3256  return NV;
3257  break;
3258  case Instruction::Select: {
3259  // If either operand of the select is a constant, we can fold the
3260  // comparison into the select arms, which will cause one to be
3261  // constant folded and the select turned into a bitwise or.
3262  Value *Op1 = nullptr, *Op2 = nullptr;
3263  ConstantInt *CI = nullptr;
3264  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3265  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3266  CI = dyn_cast<ConstantInt>(Op1);
3267  }
3268  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3269  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3270  CI = dyn_cast<ConstantInt>(Op2);
3271  }
3272 
3273  // We only want to perform this transformation if it will not lead to
3274  // additional code. This is true if either both sides of the select
3275  // fold to a constant (in which case the icmp is replaced with a select
3276  // which will usually simplify) or this is the only user of the
3277  // select (in which case we are trading a select+icmp for a simpler
3278  // select+icmp) or all uses of the select can be replaced based on
3279  // dominance information ("Global cases").
3280  bool Transform = false;
3281  if (Op1 && Op2)
3282  Transform = true;
3283  else if (Op1 || Op2) {
3284  // Local case
3285  if (LHSI->hasOneUse())
3286  Transform = true;
3287  // Global cases
3288  else if (CI && !CI->isZero())
3289  // When Op1 is constant try replacing select with second operand.
3290  // Otherwise Op2 is constant and try replacing select with first
3291  // operand.
3292  Transform =
3293  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3294  }
3295  if (Transform) {
3296  if (!Op1)
3297  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3298  I.getName());
3299  if (!Op2)
3300  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3301  I.getName());
3302  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3303  }
3304  break;
3305  }
3306  case Instruction::IntToPtr:
3307  // icmp pred inttoptr(X), null -> icmp pred X, 0
3308  if (RHSC->isNullValue() &&
3309  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3310  return new ICmpInst(
3311  I.getPredicate(), LHSI->getOperand(0),
3313  break;
3314 
3315  case Instruction::Load:
3316  // Try to optimize things like "A[i] > 4" to index computations.
3317  if (GetElementPtrInst *GEP =
3318  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3319  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3320  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3321  !cast<LoadInst>(LHSI)->isVolatile())
3322  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3323  return Res;
3324  }
3325  break;
3326  }
3327 
3328  return nullptr;
3329 }
3330 
3331 /// Some comparisons can be simplified.
3332 /// In this case, we are looking for comparisons that look like
3333 /// a check for a lossy truncation.
3334 /// Folds:
3335 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3336 /// Where Mask is some pattern that produces all-ones in low bits:
3337 /// (-1 >> y)
3338 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3339 /// ~(-1 << y)
3340 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3341 /// The Mask can be a constant, too.
3342 /// For some predicates, the operands are commutative.
3343 /// For others, x can only be on a specific side.
3346  ICmpInst::Predicate SrcPred;
3347  Value *X, *M, *Y;
3348  auto m_VariableMask = m_CombineOr(
3350  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3353  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3354  if (!match(&I, m_c_ICmp(SrcPred,
3356  m_Deferred(X))))
3357  return nullptr;
3358 
3359  ICmpInst::Predicate DstPred;
3360  switch (SrcPred) {
3361  case ICmpInst::Predicate::ICMP_EQ:
3362  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3363  DstPred = ICmpInst::Predicate::ICMP_ULE;
3364  break;
3365  case ICmpInst::Predicate::ICMP_NE:
3366  // x & (-1 >> y) != x -> x u> (-1 >> y)
3367  DstPred = ICmpInst::Predicate::ICMP_UGT;
3368  break;
3369  case ICmpInst::Predicate::ICMP_ULT:
3370  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3371  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3372  DstPred = ICmpInst::Predicate::ICMP_UGT;
3373  break;
3374  case ICmpInst::Predicate::ICMP_UGE:
3375  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3376  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3377  DstPred = ICmpInst::Predicate::ICMP_ULE;
3378  break;
3379  case ICmpInst::Predicate::ICMP_SLT:
3380  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3381  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3382  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3383  return nullptr;
3384  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3385  return nullptr;
3386  DstPred = ICmpInst::Predicate::ICMP_SGT;
3387  break;
3388  case ICmpInst::Predicate::ICMP_SGE:
3389  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3390  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3391  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3392  return nullptr;
3393  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3394  return nullptr;
3395  DstPred = ICmpInst::Predicate::ICMP_SLE;
3396  break;
3397  case ICmpInst::Predicate::ICMP_SGT:
3398  case ICmpInst::Predicate::ICMP_SLE:
3399  return nullptr;
3400  case ICmpInst::Predicate::ICMP_UGT:
3401  case ICmpInst::Predicate::ICMP_ULE:
3402  llvm_unreachable("Instsimplify took care of commut. variant");
3403  break;
3404  default:
3405  llvm_unreachable("All possible folds are handled.");
3406  }
3407 
3408  // The mask value may be a vector constant that has undefined elements. But it
3409  // may not be safe to propagate those undefs into the new compare, so replace
3410  // those elements by copying an existing, defined, and safe scalar constant.
3411  Type *OpTy = M->getType();
3412  auto *VecC = dyn_cast<Constant>(M);
3413  auto *OpVTy = dyn_cast<FixedVectorType>(OpTy);
3414  if (OpVTy && VecC && VecC->containsUndefOrPoisonElement()) {
3415  Constant *SafeReplacementConstant = nullptr;
3416  for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3417  if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3418  SafeReplacementConstant = VecC->getAggregateElement(i);
3419  break;
3420  }
3421  }
3422  assert(SafeReplacementConstant && "Failed to find undef replacement");
3423  M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3424  }
3425 
3426  return Builder.CreateICmp(DstPred, X, M);
3427 }
3428 
3429 /// Some comparisons can be simplified.
3430 /// In this case, we are looking for comparisons that look like
3431 /// a check for a lossy signed truncation.
3432 /// Folds: (MaskedBits is a constant.)
3433 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3434 /// Into:
3435 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3436 /// Where KeptBits = bitwidth(%x) - MaskedBits
3437 static Value *
3440  ICmpInst::Predicate SrcPred;
3441  Value *X;
3442  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3443  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3444  if (!match(&I, m_c_ICmp(SrcPred,
3446  m_APInt(C1))),
3447  m_Deferred(X))))
3448  return nullptr;
3449 
3450  // Potential handling of non-splats: for each element:
3451  // * if both are undef, replace with constant 0.
3452  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3453  // * if both are not undef, and are different, bailout.
3454  // * else, only one is undef, then pick the non-undef one.
3455 
3456  // The shift amount must be equal.
3457  if (*C0 != *C1)
3458  return nullptr;
3459  const APInt &MaskedBits = *C0;
3460  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3461 
3462  ICmpInst::Predicate DstPred;
3463  switch (SrcPred) {
3464  case ICmpInst::Predicate::ICMP_EQ:
3465  // ((%x << MaskedBits) a>> MaskedBits) == %x
3466  // =>
3467  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3468  DstPred = ICmpInst::Predicate::ICMP_ULT;
3469  break;
3470  case ICmpInst::Predicate::ICMP_NE:
3471  // ((%x << MaskedBits) a>> MaskedBits) != %x
3472  // =>
3473  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3474  DstPred = ICmpInst::Predicate::ICMP_UGE;
3475  break;
3476  // FIXME: are more folds possible?
3477  default:
3478  return nullptr;
3479  }
3480 
3481  auto *XType = X->getType();
3482  const unsigned XBitWidth = XType->getScalarSizeInBits();
3483  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3484  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3485 
3486  // KeptBits = bitwidth(%x) - MaskedBits
3487  const APInt KeptBits = BitWidth - MaskedBits;
3488  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3489  // ICmpCst = (1 << KeptBits)
3490  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3491  assert(ICmpCst.isPowerOf2());
3492  // AddCst = (1 << (KeptBits-1))
3493  const APInt AddCst = ICmpCst.lshr(1);
3494  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3495 
3496  // T0 = add %x, AddCst
3497  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3498  // T1 = T0 DstPred ICmpCst
3499  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3500 
3501  return T1;
3502 }
3503 
3504 // Given pattern:
3505 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3506 // we should move shifts to the same hand of 'and', i.e. rewrite as
3507 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3508 // We are only interested in opposite logical shifts here.
3509 // One of the shifts can be truncated.
3510 // If we can, we want to end up creating 'lshr' shift.
3511 static Value *
3514  if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3515  !I.getOperand(0)->hasOneUse())
3516  return nullptr;
3517 
3518  auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3519 
3520  // Look for an 'and' of two logical shifts, one of which may be truncated.
3521  // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3522  Instruction *XShift, *MaybeTruncation, *YShift;
3523  if (!match(
3524  I.getOperand(0),
3525  m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3527  m_AnyLogicalShift, m_Instruction(YShift))),
3528  m_Instruction(MaybeTruncation)))))
3529  return nullptr;
3530 
3531  // We potentially looked past 'trunc', but only when matching YShift,
3532  // therefore YShift must have the widest type.
3533  Instruction *WidestShift = YShift;
3534  // Therefore XShift must have the shallowest type.
3535  // Or they both have identical types if there was no truncation.
3536  Instruction *NarrowestShift = XShift;
3537 
3538  Type *WidestTy = WidestShift->getType();
3539  Type *NarrowestTy = NarrowestShift->getType();
3540  assert(NarrowestTy == I.getOperand(0)->getType() &&
3541  "We did not look past any shifts while matching XShift though.");
3542  bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3543 
3544  // If YShift is a 'lshr', swap the shifts around.
3545  if (match(YShift, m_LShr(m_Value(), m_Value())))
3546  std::swap(XShift, YShift);
3547 
3548  // The shifts must be in opposite directions.
3549  auto XShiftOpcode = XShift->getOpcode();
3550  if (XShiftOpcode == YShift->getOpcode())
3551  return nullptr; // Do not care about same-direction shifts here.
3552 
3553  Value *X, *XShAmt, *Y, *YShAmt;
3554  match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3555  match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3556 
3557  // If one of the values being shifted is a constant, then we will end with
3558  // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3559  // however, we will need to ensure that we won't increase instruction count.
3560  if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3561  // At least one of the hands of the 'and' should be one-use shift.
3562  if (!match(I.getOperand(0),
3563  m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3564  return nullptr;
3565  if (HadTrunc) {
3566  // Due to the 'trunc', we will need to widen X. For that either the old
3567  // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3568  if (!MaybeTruncation->hasOneUse() &&
3569  !NarrowestShift->getOperand(1)->hasOneUse())
3570  return nullptr;
3571  }
3572  }
3573 
3574  // We have two shift amounts from two different shifts. The types of those
3575  // shift amounts may not match. If that's the case let's bailout now.
3576  if (XShAmt->getType() != YShAmt->getType())
3577  return nullptr;
3578 
3579  // As input, we have the following pattern:
3580  // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3581  // We want to rewrite that as:
3582  // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3583  // While we know that originally (Q+K) would not overflow
3584  // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3585  // shift amounts. so it may now overflow in smaller bitwidth.
3586  // To ensure that does not happen, we need to ensure that the total maximal
3587  // shift amount is still representable in that smaller bit width.
3588  unsigned MaximalPossibleTotalShiftAmount =
3589  (WidestTy->getScalarSizeInBits() - 1) +
3590  (NarrowestTy->getScalarSizeInBits() - 1);
3591  APInt MaximalRepresentableShiftAmount =
3593  if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3594  return nullptr;
3595 
3596  // Can we fold (XShAmt+YShAmt) ?
3597  auto *NewShAmt = dyn_cast_or_null<Constant>(
3598  SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3599  /*isNUW=*/false, SQ.getWithInstruction(&I)));
3600  if (!NewShAmt)
3601  return nullptr;
3602  NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3603  unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3604 
3605  // Is the new shift amount smaller than the bit width?
3606  // FIXME: could also rely on ConstantRange.
3607  if (!match(NewShAmt,
3608  m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3609  APInt(WidestBitWidth, WidestBitWidth))))
3610  return nullptr;
3611 
3612  // An extra legality check is needed if we had trunc-of-lshr.
3613  if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3614  auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3615  WidestShift]() {
3616  // It isn't obvious whether it's worth it to analyze non-constants here.
3617  // Also, let's basically give up on non-splat cases, pessimizing vectors.
3618  // If *any* of these preconditions matches we can perform the fold.
3619  Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3620  ? NewShAmt->getSplatValue()
3621  : NewShAmt;
3622  // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3623  if (NewShAmtSplat &&
3624  (NewShAmtSplat->isNullValue() ||
3625  NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3626  return true;
3627  // We consider *min* leading zeros so a single outlier
3628  // blocks the transform as opposed to allowing it.
3629  if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3630  KnownBits Known = computeKnownBits(C, SQ.DL);
3631  unsigned MinLeadZero = Known.countMinLeadingZeros();
3632  // If the value being shifted has at most lowest bit set we can fold.
3633  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3634  if (MaxActiveBits <= 1)
3635  return true;
3636  // Precondition: NewShAmt u<= countLeadingZeros(C)
3637  if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3638  return true;
3639  }
3640  if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3641  KnownBits Known = computeKnownBits(C, SQ.DL);
3642  unsigned MinLeadZero = Known.countMinLeadingZeros();
3643  // If the value being shifted has at most lowest bit set we can fold.
3644  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3645  if (MaxActiveBits <= 1)
3646  return true;
3647  // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3648  if (NewShAmtSplat) {
3649  APInt AdjNewShAmt =
3650  (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3651  if (AdjNewShAmt.ule(MinLeadZero))
3652  return true;
3653  }
3654  }
3655  return false; // Can't tell if it's ok.
3656  };
3657  if (!CanFold())
3658  return nullptr;
3659  }
3660 
3661  // All good, we can do this fold.
3662  X = Builder.CreateZExt(X, WidestTy);
3663  Y = Builder.CreateZExt(Y, WidestTy);
3664  // The shift is the same that was for X.
3665  Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3666  ? Builder.CreateLShr(X, NewShAmt)
3667  : Builder.CreateShl(X, NewShAmt);
3668  Value *T1 = Builder.CreateAnd(T0, Y);
3669  return Builder.CreateICmp(I.getPredicate(), T1,
3670  Constant::getNullValue(WidestTy));
3671 }
3672 
3673 /// Fold
3674 /// (-1 u/ x) u< y
3675 /// ((x * y) u/ x) != y
3676 /// to
3677 /// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3678 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3679 /// will mean that we are looking for the opposite answer.
3681  ICmpInst::Predicate Pred;
3682  Value *X, *Y;
3683  Instruction *Mul;
3684  bool NeedNegation;
3685  // Look for: (-1 u/ x) u</u>= y
3686  if (!I.isEquality() &&
3687  match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3688  m_Value(Y)))) {
3689  Mul = nullptr;
3690 
3691  // Are we checking that overflow does not happen, or does happen?
3692  switch (Pred) {
3693  case ICmpInst::Predicate::ICMP_ULT:
3694  NeedNegation = false;
3695  break; // OK
3696  case ICmpInst::Predicate::ICMP_UGE:
3697  NeedNegation = true;
3698  break; // OK
3699  default:
3700  return nullptr; // Wrong predicate.
3701  }
3702  } else // Look for: ((x * y) u/ x) !=/== y
3703  if (I.isEquality() &&
3704  match(&I, m_c_ICmp(Pred, m_Value(Y),
3706  m_Value(X)),
3707  m_Instruction(Mul)),
3708  m_Deferred(X)))))) {
3709  NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3710  } else
3711  return nullptr;
3712 
3714  // If the pattern included (x * y), we'll want to insert new instructions
3715  // right before that original multiplication so that we can replace it.
3716  bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3717  if (MulHadOtherUses)
3718  Builder.SetInsertPoint(Mul);
3719 
3721  I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3722  CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3723 
3724  // If the multiplication was used elsewhere, to ensure that we don't leave
3725  // "duplicate" instructions, replace uses of that original multiplication
3726  // with the multiplication result from the with.overflow intrinsic.
3727  if (MulHadOtherUses)
3728  replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3729 
3730  Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3731  if (NeedNegation) // This technically increases instruction count.
3732  Res = Builder.CreateNot(Res, "umul.not.ov");
3733 
3734  // If we replaced the mul, erase it. Do this after all uses of Builder,
3735  // as the mul is used as insertion point.
3736  if (MulHadOtherUses)
3737  eraseInstFromFunction(*Mul);
3738 
3739  return Res;
3740 }
3741 
3743  CmpInst::Predicate Pred;
3744  Value *X;
3745  if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3746  return nullptr;
3747 
3748  if (ICmpInst::isSigned(Pred))
3749  Pred = ICmpInst::getSwappedPredicate(Pred);
3750  else if (ICmpInst::isUnsigned(Pred))
3751  Pred = ICmpInst::getSignedPredicate(Pred);
3752  // else for equality-comparisons just keep the predicate.
3753 
3754  return ICmpInst::Create(Instruction::ICmp, Pred, X,
3755  Constant::getNullValue(X->getType()), I.getName());
3756 }
3757 
3758 /// Try to fold icmp (binop), X or icmp X, (binop).
3759 /// TODO: A large part of this logic is duplicated in InstSimplify's
3760 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3761 /// duplication.
3763  const SimplifyQuery &SQ) {
3764  const SimplifyQuery Q = SQ.getWithInstruction(&I);
3765  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3766 
3767  // Special logic for binary operators.
3768  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3769  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3770  if (!BO0 && !BO1)
3771  return nullptr;
3772 
3773  if (Instruction *NewICmp = foldICmpXNegX(I))
3774  return NewICmp;
3775 
3776  const CmpInst::Predicate Pred = I.getPredicate();
3777  Value *X;
3778 
3779  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3780  // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3781  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3782  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3783  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3784  // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3785  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3786  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3787  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3788 
3789  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3790  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3791  NoOp0WrapProblem =
3792  ICmpInst::isEquality(Pred) ||
3793  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3794  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3795  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3796  NoOp1WrapProblem =
3797  ICmpInst::isEquality(Pred) ||
3798  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3799  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3800 
3801  // Analyze the case when either Op0 or Op1 is an add instruction.
3802  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3803  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3804  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3805  A = BO0->getOperand(0);
3806  B = BO0->getOperand(1);
3807  }
3808  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3809  C = BO1->getOperand(0);
3810  D = BO1->getOperand(1);
3811  }
3812 
3813  // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3814  // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3815  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3816  return new ICmpInst(Pred, A == Op1 ? B : A,
3817  Constant::getNullValue(Op1->getType()));
3818 
3819  // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3820  // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3821  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3822  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3823  C == Op0 ? D : C);
3824 
3825  // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3826  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3827  NoOp1WrapProblem) {
3828  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3829  Value *Y, *Z;
3830  if (A == C) {
3831  // C + B == C + D -> B == D
3832  Y = B;
3833  Z = D;
3834  } else if (A == D) {
3835  // D + B == C + D -> B == C
3836  Y = B;
3837  Z = C;
3838  } else if (B == C) {
3839  // A + C == C + D -> A == D
3840  Y = A;
3841  Z = D;
3842  } else {
3843  assert(B == D);
3844  // A + D == C + D -> A == C
3845  Y = A;
3846  Z = C;
3847  }
3848  return new ICmpInst(Pred, Y, Z);
3849  }
3850 
3851  // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3852  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3853  match(B, m_AllOnes()))
3854  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3855 
3856  // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3857  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3858  match(B, m_AllOnes()))
3859  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3860 
3861  // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3862  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3863  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3864 
3865  // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3866  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3867  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3868 
3869  // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3870  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3871  match(D, m_AllOnes()))
3872  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3873 
3874  // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3875  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3876  match(D, m_AllOnes()))
3877  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3878 
3879  // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3880  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3881  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3882 
3883  // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3884  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3885  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3886 
3887  // TODO: The subtraction-related identities shown below also hold, but
3888  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3889  // wouldn't happen even if they were implemented.
3890  //
3891  // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3892  // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3893  // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3894  // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3895 
3896  // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3897  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3898  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3899 
3900  // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3901  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3902  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3903 
3904  // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3905  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3906  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3907 
3908  // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3909  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3910  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3911 
3912  // if C1 has greater magnitude than C2:
3913  // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3914  // s.t. C3 = C1 - C2
3915  //
3916  // if C2 has greater magnitude than C1:
3917  // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3918  // s.t. C3 = C2 - C1
3919  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3920  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3921  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3922  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3923  const APInt &AP1 = C1->getValue();
3924  const APInt &AP2 = C2->getValue();
3925  if (AP1.isNegative() == AP2.isNegative()) {
3926  APInt AP1Abs = C1->getValue().abs();
3927  APInt AP2Abs = C2->getValue().abs();
3928  if (AP1Abs.uge(AP2Abs)) {
3929  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3930  bool HasNUW = BO0->hasNoUnsignedWrap() && C3->getValue().ule(AP1);
3931  bool HasNSW = BO0->hasNoSignedWrap();
3932  Value *NewAdd = Builder.CreateAdd(A, C3, "", HasNUW, HasNSW);
3933  return new ICmpInst(Pred, NewAdd, C);
3934  } else {
3935  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3936  bool HasNUW = BO1->hasNoUnsignedWrap() && C3->getValue().ule(AP2);
3937  bool HasNSW = BO1->hasNoSignedWrap();
3938  Value *NewAdd = Builder.CreateAdd(C, C3, "", HasNUW, HasNSW);
3939  return new ICmpInst(Pred, A, NewAdd);
3940  }
3941  }
3942  }
3943 
3944  // Analyze the case when either Op0 or Op1 is a sub instruction.
3945  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3946  A = nullptr;
3947  B = nullptr;
3948  C = nullptr;
3949  D = nullptr;
3950  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3951  A = BO0->getOperand(0);
3952  B = BO0->getOperand(1);
3953  }
3954  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3955  C = BO1->getOperand(0);
3956  D = BO1->getOperand(1);
3957  }
3958 
3959  // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3960  if (A == Op1 && NoOp0WrapProblem)
3961  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3962  // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
3963  if (C == Op0 && NoOp1WrapProblem)
3964  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3965 
3966  // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3967  // (A - B) u>/u<= A --> B u>/u<= A
3968  if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3969  return new ICmpInst(Pred, B, A);
3970  // C u</u>= (C - D) --> C u</u>= D
3971  if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3972  return new ICmpInst(Pred, C, D);
3973  // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
3974  if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
3975  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3976  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
3977  // C u<=/u> (C - D) --> C u</u>= D iff B != 0
3978  if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
3979  isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3980  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
3981 
3982  // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
3983  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
3984  return new ICmpInst(Pred, A, C);
3985 
3986  // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
3987  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
3988  return new ICmpInst(Pred, D, B);
3989 
3990  // icmp (0-X) < cst --> x > -cst
3991  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3992  Value *X;
3993  if (match(BO0, m_Neg(m_Value(X))))
3994  if (Constant *RHSC = dyn_cast<Constant>(Op1))
3995  if (RHSC->isNotMinSignedValue())
3996  return new ICmpInst(I.getSwappedPredicate(), X,
3997  ConstantExpr::getNeg(RHSC));
3998  }
3999 
4000  {
4001  // Try to remove shared constant multiplier from equality comparison:
4002  // X * C == Y * C (with no overflowing/aliasing) --> X == Y
4003  Value *X, *Y;
4004  const APInt *C;
4005  if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
4006  match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
4007  if (!C->countTrailingZeros() ||
4008  (BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
4009  (BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
4010  return new ICmpInst(Pred, X, Y);
4011  }
4012 
4013  BinaryOperator *SRem = nullptr;
4014  // icmp (srem X, Y), Y
4015  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
4016  SRem = BO0;
4017  // icmp Y, (srem X, Y)
4018  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
4019  Op0 == BO1->getOperand(1))
4020  SRem = BO1;
4021  if (SRem) {
4022  // We don't check hasOneUse to avoid increasing register pressure because
4023  // the value we use is the same value this instruction was already using.
4024  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4025  default:
4026  break;
4027  case ICmpInst::ICMP_EQ:
4028  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4029  case ICmpInst::ICMP_NE:
4030  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4031  case ICmpInst::ICMP_SGT:
4032  case ICmpInst::ICMP_SGE:
4033  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4035  case ICmpInst::ICMP_SLT:
4036  case ICmpInst::ICMP_SLE:
4037  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4038  Constant::getNullValue(SRem->getType()));
4039  }
4040  }
4041 
4042  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4043  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4044  switch (BO0->getOpcode()) {
4045  default:
4046  break;
4047  case Instruction::Add:
4048  case Instruction::Sub:
4049  case Instruction::Xor: {
4050  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4051  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4052 
4053  const APInt *C;
4054  if (match(BO0->getOperand(1), m_APInt(C))) {
4055  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4056  if (C->isSignMask()) {
4057  ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4058  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4059  }
4060 
4061  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4062  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4063  ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4064  NewPred = I.getSwappedPredicate(NewPred);
4065  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4066  }
4067  }
4068  break;
4069  }
4070  case Instruction::Mul: {
4071  if (!I.isEquality())
4072  break;
4073 
4074  const APInt *C;
4075  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
4076  !C->isOneValue()) {
4077  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4078  // Mask = -1 >> count-trailing-zeros(C).
4079  if (unsigned TZs = C->countTrailingZeros()) {
4081  BO0->getType(),
4082  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4083  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4084  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4085  return new ICmpInst(Pred, And1, And2);
4086  }
4087  }
4088  break;
4089  }
4090  case Instruction::UDiv:
4091  case Instruction::LShr:
4092  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4093  break;
4094  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4095 
4096  case Instruction::SDiv:
4097  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4098  break;
4099  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4100 
4101  case Instruction::AShr:
4102  if (!BO0->isExact() || !BO1->isExact())
4103  break;
4104  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4105 
4106  case Instruction::Shl: {
4107  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4108  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4109  if (!NUW && !NSW)
4110  break;
4111  if (!NSW && I.isSigned())
4112  break;
4113  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4114  }
4115  }
4116  }
4117 
4118  if (BO0) {
4119  // Transform A & (L - 1) `ult` L --> L != 0
4120  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4121  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4122 
4123  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4124  auto *Zero = Constant::getNullValue(BO0->getType());
4125  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4126  }
4127  }
4128 
4129  if (Value *V = foldUnsignedMultiplicationOverflowCheck(I))
4130  return replaceInstUsesWith(I, V);
4131 
4133  return replaceInstUsesWith(I, V);
4134 
4136  return replaceInstUsesWith(I, V);
4137 
4139  return replaceInstUsesWith(I, V);
4140 
4141  return nullptr;
4142 }
4143 
4144 /// Fold icmp Pred min|max(X, Y), X.
4146  ICmpInst::Predicate Pred = Cmp.getPredicate();
4147  Value *Op0 = Cmp.getOperand(0);
4148  Value *X = Cmp.getOperand(1);
4149 
4150  // Canonicalize minimum or maximum operand to LHS of the icmp.
4151  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4152  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4153  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4154  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4155  std::swap(Op0, X);
4156  Pred = Cmp.getSwappedPredicate();
4157  }
4158 
4159  Value *Y;
4160  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4161  // smin(X, Y) == X --> X s<= Y
4162  // smin(X, Y) s>= X --> X s<= Y
4163  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4164  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4165 
4166  // smin(X, Y) != X --> X s> Y
4167  // smin(X, Y) s< X --> X s> Y
4168  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4169  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4170 
4171  // These cases should be handled in InstSimplify:
4172  // smin(X, Y) s<= X --> true
4173  // smin(X, Y) s> X --> false
4174  return nullptr;
4175  }
4176 
4177  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4178  // smax(X, Y) == X --> X s>= Y
4179  // smax(X, Y) s<= X --> X s>= Y
4180  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4181  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4182 
4183  // smax(X, Y) != X --> X s< Y
4184  // smax(X, Y) s> X --> X s< Y
4185  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4186  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4187 
4188  // These cases should be handled in InstSimplify:
4189  // smax(X, Y) s>= X --> true
4190  // smax(X, Y) s< X --> false
4191  return nullptr;
4192  }
4193 
4194  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4195  // umin(X, Y) == X --> X u<= Y
4196  // umin(X, Y) u>= X --> X u<= Y
4197  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4198  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4199 
4200  // umin(X, Y) != X --> X u> Y
4201  // umin(X, Y) u< X --> X u> Y
4202  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4203  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4204 
4205  // These cases should be handled in InstSimplify:
4206  // umin(X, Y) u<= X --> true
4207  // umin(X, Y) u> X --> false
4208  return nullptr;
4209  }
4210 
4211  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4212  // umax(X, Y) == X --> X u>= Y
4213  // umax(X, Y) u<= X --> X u>= Y
4214  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4215  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4216 
4217  // umax(X, Y) != X --> X u< Y
4218  // umax(X, Y) u> X --> X u< Y
4219  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4220  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4221 
4222  // These cases should be handled in InstSimplify:
4223  // umax(X, Y) u>= X --> true
4224  // umax(X, Y) u< X --> false
4225  return nullptr;
4226  }
4227 
4228  return nullptr;
4229 }
4230 
4232  if (!I.isEquality())
4233  return nullptr;
4234 
4235  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4236  const CmpInst::Predicate Pred = I.getPredicate();
4237  Value *A, *B, *C, *D;
4238  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4239  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4240  Value *OtherVal = A == Op1 ? B : A;
4241  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4242  }
4243 
4244  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4245  // A^c1 == C^c2 --> A == C^(c1^c2)
4246  ConstantInt *C1, *C2;
4247  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4248  Op1->hasOneUse()) {
4249  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4250  Value *Xor = Builder.CreateXor(C, NC);
4251  return new ICmpInst(Pred, A, Xor);
4252  }
4253 
4254  // A^B == A^D -> B == D
4255  if (A == C)
4256  return new ICmpInst(Pred, B, D);
4257  if (A == D)
4258  return new ICmpInst(Pred, B, C);
4259  if (B == C)
4260  return new ICmpInst(Pred, A, D);
4261  if (B == D)
4262  return new ICmpInst(Pred, A, C);
4263  }
4264  }
4265 
4266  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4267  // A == (A^B) -> B == 0
4268  Value *OtherVal = A == Op0 ? B : A;
4269  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4270  }
4271 
4272  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4273  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4274  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4275  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4276 
4277  if (A == C) {
4278  X = B;
4279  Y = D;
4280  Z = A;
4281  } else if (A == D) {
4282  X = B;
4283  Y = C;
4284  Z = A;
4285  } else if (B == C) {
4286  X = A;
4287  Y = D;
4288  Z = B;
4289  } else if (B == D) {
4290  X = A;
4291  Y = C;
4292  Z = B;
4293  }
4294 
4295  if (X) { // Build (X^Y) & Z
4296  Op1 = Builder.CreateXor(X, Y);
4297  Op1 = Builder.CreateAnd(Op1, Z);
4298  return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4299  }
4300  }
4301 
4302  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4303  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4304  ConstantInt *Cst1;
4305  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4306  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4307  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4308  match(Op1, m_ZExt(m_Value(A))))) {
4309  APInt Pow2 = Cst1->getValue() + 1;
4310  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4311  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4312  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4313  }
4314 
4315  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4316  // For lshr and ashr pairs.
4317  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4318  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4319  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4320  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4321  unsigned TypeBits = Cst1->getBitWidth();
4322  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4323  if (ShAmt < TypeBits && ShAmt != 0) {
4324  ICmpInst::Predicate NewPred =
4326  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4327  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4328  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4329  }
4330  }
4331 
4332  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4333  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4334  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4335  unsigned TypeBits = Cst1->getBitWidth();
4336  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4337  if (ShAmt < TypeBits && ShAmt != 0) {
4338  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4339  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4340  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4341  I.getName() + ".mask");
4342  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4343  }
4344  }
4345 
4346  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4347  // "icmp (and X, mask), cst"
4348  uint64_t ShAmt = 0;
4349  if (Op0->hasOneUse() &&
4350  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4351  match(Op1, m_ConstantInt(Cst1)) &&
4352  // Only do this when A has multiple uses. This is most important to do
4353  // when it exposes other optimizations.
4354  !A->hasOneUse()) {
4355  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4356 
4357  if (ShAmt < ASize) {
4358  APInt MaskV =
4360  MaskV <<= ShAmt;
4361 
4362  APInt CmpV = Cst1->getValue().zext(ASize);
4363  CmpV <<= ShAmt;
4364 
4365  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4366  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4367  }
4368  }
4369 
4370  // If both operands are byte-swapped or bit-reversed, just compare the
4371  // original values.
4372  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4373  // and handle more intrinsics.
4374  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4375  (match(Op0, m_BitReverse(m_Value(A))) &&
4376  match(Op1, m_BitReverse(m_Value(B)))))
4377  return new ICmpInst(Pred, A, B);
4378 
4379  // Canonicalize checking for a power-of-2-or-zero value:
4380  // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4381  // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4382  if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4383  m_Deferred(A)))) ||
4384  !match(Op1, m_ZeroInt()))
4385  A = nullptr;
4386 
4387  // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4388  // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4389  if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4390  A = Op1;
4391  else if (match(Op1,
4392  m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4393  A = Op0;
4394 
4395  if (A) {
4396  Type *Ty = A->getType();
4397  CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4398  return Pred == ICmpInst::ICMP_EQ
4399  ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4400  : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4401  }
4402 
4403  return nullptr;
4404 }
4405 
4408  assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
4409  auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4410  Value *X;
4411  if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4412  return nullptr;
4413 
4414  bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4415  bool IsSignedCmp = ICmp.isSigned();
4416  if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4417  // If the signedness of the two casts doesn't agree (i.e. one is a sext
4418  // and the other is a zext), then we can't handle this.
4419  // TODO: This is too strict. We can handle some predicates (equality?).
4420  if (CastOp0->getOpcode() != CastOp1->getOpcode())
4421  return nullptr;
4422 
4423  // Not an extension from the same type?
4424  Value *Y = CastOp1->getOperand(0);
4425  Type *XTy = X->getType(), *YTy = Y->getType();
4426  if (XTy != YTy) {
4427  // One of the casts must have one use because we are creating a new cast.
4428  if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4429  return nullptr;
4430  // Extend the narrower operand to the type of the wider operand.
4431  if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4432  X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4433  else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4434  Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4435  else
4436  return nullptr;
4437  }
4438 
4439  // (zext X) == (zext Y) --> X == Y
4440  // (sext X) == (sext Y) --> X == Y
4441  if (ICmp.isEquality())
4442  return new ICmpInst(ICmp.getPredicate(), X, Y);
4443 
4444  // A signed comparison of sign extended values simplifies into a
4445  // signed comparison.
4446  if (IsSignedCmp && IsSignedExt)
4447  return new ICmpInst(ICmp.getPredicate(), X, Y);
4448 
4449  // The other three cases all fold into an unsigned comparison.
4450  return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4451  }
4452 
4453  // Below here, we are only folding a compare with constant.
4454  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4455  if (!C)
4456  return nullptr;
4457 
4458  // Compute the constant that would happen if we truncated to SrcTy then
4459  // re-extended to DestTy.
4460  Type *SrcTy = CastOp0->getSrcTy();
4461  Type *DestTy = CastOp0->getDestTy();
4462  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4463  Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4464 
4465  // If the re-extended constant didn't change...
4466  if (Res2 == C) {
4467  if (ICmp.isEquality())
4468  return new ICmpInst(ICmp.getPredicate(), X, Res1);
4469 
4470  // A signed comparison of sign extended values simplifies into a
4471  // signed comparison.
4472  if (IsSignedExt && IsSignedCmp)
4473  return new ICmpInst(ICmp.getPredicate(), X, Res1);
4474 
4475  // The other three cases all fold into an unsigned comparison.
4476  return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4477  }
4478 
4479  // The re-extended constant changed, partly changed (in the case of a vector),
4480  // or could not be determined to be equal (in the case of a constant
4481  // expression), so the constant cannot be represented in the shorter type.
4482  // All the cases that fold to true or false will have already been handled
4483  // by SimplifyICmpInst, so only deal with the tricky case.
4484  if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4485  return nullptr;
4486 
4487  // Is source op positive?
4488  // icmp ult (sext X), C --> icmp sgt X, -1
4489  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4491 
4492  // Is source op negative?
4493  // icmp ugt (sext X), C --> icmp slt X, 0
4494  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4495  return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4496 }
4497 
4498 /// Handle icmp (cast x), (cast or constant).
4500  auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4501  if (!CastOp0)
4502  return nullptr;
4503  if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4504  return nullptr;
4505 
4506  Value *Op0Src = CastOp0->getOperand(0);
4507  Type *SrcTy = CastOp0->getSrcTy();
4508  Type *DestTy = CastOp0->getDestTy();
4509 
4510  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4511  // integer type is the same size as the pointer type.
4512  auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4513  if (isa<VectorType>(SrcTy)) {
4514  SrcTy = cast<VectorType>(SrcTy)->getElementType();
4515  DestTy = cast<VectorType>(DestTy)->getElementType();
4516  }
4517  return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4518  };
4519  if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4520  CompatibleSizes(SrcTy, DestTy)) {
4521  Value *NewOp1 = nullptr;
4522  if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4523  Value *PtrSrc = PtrToIntOp1->getOperand(0);
4524  if (PtrSrc->getType()->getPointerAddressSpace() ==
4525  Op0Src->getType()->getPointerAddressSpace()) {
4526  NewOp1 = PtrToIntOp1->getOperand(0);
4527  // If the pointer types don't match, insert a bitcast.
4528  if (Op0Src->getType() != NewOp1->getType())
4529  NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4530  }
4531  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4532  NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4533  }
4534 
4535  if (NewOp1)
4536  return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4537  }
4538 
4539  return foldICmpWithZextOrSext(ICmp, Builder);
4540 }
4541 
4542 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4543  switch (BinaryOp) {
4544  default:
4545  llvm_unreachable("Unsupported binary op");
4546  case Instruction::Add:
4547  case Instruction::Sub:
4548  return match(RHS, m_Zero());
4549  case Instruction::Mul:
4550  return match(RHS, m_One());
4551  }
4552 }
4553 
4556  bool IsSigned, Value *LHS, Value *RHS,
4557  Instruction *CxtI) const {
4558  switch (BinaryOp) {
4559  default:
4560  llvm_unreachable("Unsupported binary op");
4561  case Instruction::Add:
4562  if (IsSigned)
4563  return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4564  else
4565  return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4566  case Instruction::Sub:
4567  if (IsSigned)
4568  return computeOverflowForSignedSub(LHS, RHS, CxtI);
4569  else
4570  return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4571  case Instruction::Mul:
4572  if (IsSigned)
4573  return computeOverflowForSignedMul(LHS, RHS, CxtI);
4574  else
4575  return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4576  }
4577 }
4578 
4579 bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
4580  bool IsSigned, Value *LHS,
4581  Value *RHS, Instruction &OrigI,
4582  Value *&Result,
4583  Constant *&Overflow) {
4584  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4585  std::swap(LHS, RHS);
4586 
4587  // If the overflow check was an add followed by a compare, the insertion point
4588  // may be pointing to the compare. We want to insert the new instructions
4589  // before the add in case there are uses of the add between the add and the
4590  // compare.
4591  Builder.SetInsertPoint(&OrigI);
4592 
4593  Type *OverflowTy = Type::getInt1Ty(LHS->getContext());
4594  if (auto *LHSTy = dyn_cast<VectorType>(LHS->getType()))
4595  OverflowTy = VectorType::get(OverflowTy, LHSTy->getElementCount());
4596 
4597  if (isNeutralValue(BinaryOp, RHS)) {
4598  Result = LHS;
4599  Overflow = ConstantInt::getFalse(OverflowTy);
4600  return true;
4601  }
4602 
4603  switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4605  return false;
4608  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4609  Result->tak